content
stringlengths 1
15.9M
|
|---|
\section{Introduction}
\label{S:intro}
This paper concerns the Monge minimization problem in metric spaces with geodesic structure:
given two Borel probability measure $\mu,\nu \in \mathcal{P}(X)$, where $(X,d)$ is a Polish space, i.e. complete and separable metric space, we study the minimization of the functional
\[
\mathcal{I}(T) = \int d_{N}(x,T(x)) \mu(dy)
\]
where $T$ varies over all Borel maps $T:X \to X$ such that $T_{\sharp}\mu = \nu$ and
$d_{N}$ is a Borel distance that makes $(X,d_{N})$ a possibly branching geodesic space.
We will apply the results to the obstacle problem: let $C \subset \mathbb{R}^{d}$ be a convex set with $\partial C = M$ smooth,
$(d-1)$-dimensional compact submanifold of $\mathbb{R}^{d}$.
Let $X = (\mathbb{R}^{d} \setminus C) \cup M$, $\mu,\nu \in \mathcal{P}(X)$ and $d_{M}(x,y)$ be the infimum among
all the Lipschitz curves in $X$ connecting to $x$ to $y$ of the euclidean length of such curves. We will prove the existence of a solution for
\[
\min_{T : T_{\sharp} \mu = \nu} \int d_{M}(x,T(x)) \mu(dx),
\]
provided $\mu \ll \mathcal{L}^{d}$.
Before describing our investigation, we present a little bit of the existing literature referring to \cite{villa:topics} and \cite{villa:Oldnew}
for a deeper insight into optimal transportation.
In the original formulation given by Monge in 1781 the problem was settled in $\mathbb{R}^{d}$,
with the cost given by the Euclidean norm and the measures $\mu, \nu$ were supposed to be absolutely
continuous and supported on two disjoint compact sets.
The original problem remained unsolved for a long time.
In 1978 Sudakov \cite{sudak} claimed to have a solution for any distance cost function induced
by a norm: an essential ingredient in the proof was that if $\mu \ll \mathcal{L}^{d}$ and $\mathcal{L}^{d}$-a.e. $\mathbb{R}^{d}$ can be decomposed into
convex sets of dimension $k$, then
then the conditional probabilities are absolutely continuous with respect to the $\mathcal{H}^{k}$ measure of the correct dimension.
But it turns out that when $d>2$, $0<k<d-1$ the property claimed by Sudakov is not true. An example with $d=3$, $k=1$ can be found in \cite{larm}.
The Euclidean case has been correctly solved only during the last decade. L. C. Evans and W. Gangbo in \cite{evagangbo}
solved the problem under the assumptions that
$\textrm{spt}\,\mu \cap \textrm{spt}\,\nu = \emptyset$, $\mu,\nu \ll \mathcal{L}^{d}$ and their densities are Lipschitz function with compact support.
The first existence results for general absolutely continuous measures $\mu,\nu$ with compact support have been independently obtained by
L. Caffarelli, M. Feldman and R.J. McCann in \cite{caffafeldmc} and by N. Trudinger and X.J. Wang in \cite{trudiwang}.
Afterwards M. Feldman and R.J. McCann \cite{feldcann:mani} extended the results to manifolds with geodesic cost.
The case of a general norm as cost function on $\mathbb{R}^{d}$, including also the case with non strictly convex unitary ball,
has been solved first in the particular case of crystalline norm by L. Ambrosio, B. Kirchheim and A. Pratelli in
\cite{ambprat:crist}, and then in fully generality by T. Champion and L. De Pascale in \cite{champdepasc:Monge}.
The case with $(X,d_{N})$ non-branching geodesic space has been studied by S. Bianchini and the author in \cite{biacava:streconv}.
Concerning optimal transportation around a convex obstacle,
it is worth noting that the case with transport cost $d_{M}^{2}$ is studied in \cite{jimenez:obstacle}.
\subsection{Overview of the paper}
\label{Ss:over}
Let $(X,d_N)$ be a geodesic space, not necessarily Polish. To assure that standard measure theory can be used,
there exists a second distance $d$ on $X$ which makes $(X,d)$ Polish and $d_N$ Borel on $X \times X$ with respect to the metric $d \times d$.
We will prove that given a $d_N$-cyclically monotone transference plan $\pi \in \Pi(\mu,\nu)$,
under appropriate assumptions on the first marginal and on the plan $\pi$, there exists an admissible
map $T:X \to X$ with the same transference cost of $\pi$.
Since we do not require $d_N$ to be l.s.c.,
the existence of an optimal transference plan is not guaranteed and our strategy doesn't rely on a possible optimality of $\pi$.
Moreover it is worth notice that due to the lack of regularity of $d_{N}$ we will not use the existence of optimal potentials $(\phi,\psi)$.
Our strategy to cope with the Monge problem with branching distance cost is the following:
\begin{enumerate}
\item reduce the problem, via Disintegration Theorem, to transportation problems in sets where,
under a regularity assumption on the first marginal and on $\pi$, we know how to produce an optimal map;
\item show that the disintegration of the first marginal $\mu$ on each of this sets verifies this regularity assumption;
\item find a transport map on each of these sets and piece them together.
\end{enumerate}
In the easier case of $d_{N}$ non-branching, given a $d_{N}$-cyclically monotone transference plan it is always possible
to reduce the problem on single geodesics. The reduced problem becomes essentially one dimensional and there
the precise regularity assumption is that the first marginal has no atoms (is continuous).
If $d_{N}$ is a branching geodesic distance this reduction can't be done anymore
and there is not another reference set where the existence of Monge minimizer is known.
The reduction set will be a concatenation of more geodesics and to produce an optimal map
we will need a regularity assumption also on the shape of this set.
As in the non-branching case, the reduction sets come from the class of geodesics used by a $d_{N}$-monotone plan $\pi$.
This class can be obtained
from a $d_{N}$-cyclical monotone set $\Gamma$ on which $\pi$ is concentrated: one can construct the set of transport rays $R$,
the transport set $\mathcal{T}_{e}$, i.e. the set of geodesics used by $\pi$, and from them construct
\begin{itemize}
\item the set $\mathcal T$ made of inner points of geodesics,
\item the set $a \cup b := \mathcal T_e \setminus \mathcal T$ of initial points $a$ and end points $b$.
\end{itemize}
Since branching of geodesics is admitted, $R$ is not a partition on $\mathcal{T}$. To obtain an equivalence relation
we have to consider the set $H$ of chain of transport rays:
it is the set of couples $(x,y)$ such that we can go from $x$ to $y$ with a
finite number of transport rays such that their common points are not final or initial points.
Hence $H$ will provide the partition of the transport set $\mathcal{T}$ and each equivalence class, $H(y)$ for $y$ in the quotient space, will be a reduction set.
Even if a partition is given, the reduction to transport problems on the equivalence classes is not straightforward: a necessary and sufficient condition is that the disintegration of the measure $\mu$
w.r.t. the partition $H$ is strongly consistent.
This is equivalent to the fact that there exists a $\mu$-measurable quotient map $f : \mathcal T \to \mathcal T$ of the equivalence relation induced by the partition. \\
Since this partition is closely related to the geodesics of $d_{N}$, the strong consistency will follow from
a topological property of the geodesic as set in $(X,d)$ and from a metric property of $d_{N}$ as a function:
\begin{enumerate}[label=(1.\alph*),ref=(1.\alph*)]
\item \label{I:comegeo} each chain of transport rays $H(y)$ restricted to a $d_{N}$ closed ball is $d$-closed;
\item \label{I:comelocpt} $d_{N}(x,\cdot)$ restricted to $H(x)$ is bounded on $d$-bounded sets.
\end{enumerate}
Observe that these conditions on $H$ and $d_{N}$ are the direct generalization of the ones on geodesics used in \cite{biacava:streconv}
(continuity and local compactness) and they depend on the particular choice of the transference plan.
This assumptions permit to disintegrate $\mu$ restricted to $\mathcal T$.
Hence one can write
\[
\mu\llcorner_{\mathcal{T}} = \int \mu_y m(dy), \quad m := f_\sharp \mu, \quad \mu_y(f^{-1}(y)) = 1,
\]
i.e. the conditional probabilities $\mu_y$ are concentrated on the counterimages $f^{-1}(y)$ (which is an equivalence class).
The reduced problems are obtained by disintegrating $\pi$ w.r.t. the partition $H \times (X \times X)$,
\[
\pi\llcorner_{\mathcal{T}\times \mathcal{T}} = \int \pi_y m(dy), \quad \nu = \int \nu_y m(dy) \quad \nu_y := (P_2)_\sharp \pi_y,
\]
and considering the problems on the sets $H(y)$ with marginals $\mu_y$, $\nu_y$ and cost $d_{N}$.
To next step is study the continuity of the conditional probabilities $\mu_y$
and whether $\mu\llcorner_{\mathcal{T}_{e}}=\mu\llcorner_{\mathcal{T}}$ holds true.
To pursue this aim we consider a natural operation on sets: the translation along geodesics. If $A$ is a subset of $\mathcal T$, we denote by $A_t$ the set translated by $t$ in the direction determined by $\pi$. A rigorous definition of the translation of sets along geodesic will be given during the paper.
It turns out that $\mu(a \cup b) = 0$ and the continuity of $\mu_y$ both depend on how the function $t \mapsto \mu(A_t)$ behaves.
Indeed assuming that:
\begin{enumerate}[label=(2)]
\item \label{I:NDEatom} for all $A$ Borel there exists a sequence $\{t_{n}\}\subset \mathbb{R}$ and $C>0$ such that $\mu(A_{t_{n}})\geq C \mu(A)$ as $t_{n}\to 0$,
\end{enumerate}
we have the following.
\begin{theorem}[Proposition \ref{P:puntini} and Proposition \ref{P:nonatoms}]
\label{T:-1}
If Assumption \ref{I:NDEatom} holds,
then $\mu(a \cup b) = 0$ and the conditional probabilities $\mu_y$ are continuous.
\end{theorem}
At this level of generality we don't know how to obtain a $d_{N}$-monotone admissible map
for the restricted problem even if the marginal $\mu_{y}$ satisfies some regularity assumptions.
Therefore we need to assume that $H(y)$ has a particular structure:
\begin{enumerate}[label=(3), ref=(3)]
\item \label{I:clessidra3} for $m$-a.e. $y$, the chain of transport rays $H(y)$ is contained, up to set of $\mu_{y}$-measure zero, in an uncountable ``increasing'' family of measurable sets.
\end{enumerate}
A rigorous formulation of Assumption \ref{I:clessidra3} and of ``increasing'' will be given during the paper.
If $H(y)$ satisfies Assumption \ref{I:clessidra3}, then we can perform a disintegration of $\mu_{y}$ with respect to the partition induced
by the uncountable ``increasing'' family of sets.
Then if the quotient measure and the marginal measures of $\mu_{y}$ are continuous,
we prove the existence of an optimal map between $\mu_{y}$ and $\nu_{y}$.
\begin{theorem}[Proposition \ref{P:alternativa} and Theorem \ref{T:finale}]\label{T:0}
Let $\pi\in \Pi(\mu,\nu)$ be a $d_{N}$-monotone plan concentrated on a set $\Gamma$.
Assume that Assumptions \ref{I:comegeo}, \ref{I:comelocpt}, \ref{I:NDEatom}, \ref{I:clessidra3} holds and
that the quotient measure and the marginal measures of $\mu_{y}$ are continuous for $m$-a.e. $y$.
Then there exists an admissible map with the same transference cost of $\pi$.
\end{theorem}
It follows immediately that if we also assume that $\pi$ is optimal in the hypothesis of Theorem \ref{T:0}, then the Monge minimization problem admits a solution.
Before presenting an application of Theorem \ref{T:0}, we summarize the theoretical results.
Let $\pi \in \Pi(\mu,\nu)$ be a $d_{N}$-cyclically monotone transference plan concentrated on a set $\Gamma$.
We consider the corresponding family of chain of transport rays and,
if assumptions \ref{I:comegeo} and \ref{I:comelocpt} are satisfied, we can perform, neglecting the set of initial points, a disintegration
of $\mu,\nu$ and $\pi$ with respect to the partition induced by the chain of transport rays.
Then if assumption \ref{I:NDEatom} is satisfied it follows that the set of initial points is $\mu$-negligible and
the conditional probabilities $\mu_{y}$ are continuous.
Since the geometry of $H(y)$ can be wild, we need another assumption to build a $d_{N}$-monotone transference map between $\mu_{y}$
and $\nu_{y}$.
If $H(y)$ satisfies assumption \ref{I:clessidra3} we can perform another disintegration and, under additional regularity of
the conditional probabilities of $\mu_{y}$ and of the quotient measure of $\mu_{y}$, we prove the existence of a
$d_{N}$-monotone transference map between $\mu_{y}$ and $\nu_{y}$.
Applying the same reasoning for $m$-a.e. $y$ we prove the existence of a transport map $T$ between $\mu$ and $\nu$ that has the
same transference cost of the given $d_{N}$-cyclically monotone plan $\pi$.
In the last part of the paper we show an application of Theorem \ref{T:0}.
Consider a hyper-surface $M \subset \mathbb{R}^{d}$ that is the boundary of a convex and compact set $C$.
Let $X$ be the closure, in the euclidean topology, of $\mathbb{R}^{d} \setminus C$ and take as cost function
$d_{M}$: the minimum of the euclidean length among all Lipschitz curves in $X$ that do not cross $M$.
We will study the Monge minimization problem with $C$ as convex obstacle.
We will prove that if $\mu$ is absolutely continuous w.r.t. $\mathcal{L}^{d}$, the Monge minimization problem with cost $d_{M}$ admits a solution.
It is worth noting that the hypothesis of Theorem \ref{T:0}, namely Assumptions \ref{I:comegeo},
\ref{I:comelocpt}, \ref{I:NDEatom}, \ref{I:clessidra3}, are all about the behavior of a given $d_{N}$-cyclically monotone transference plan.
For the obstacle problem we will prove that any $d_{N}$-cyclically monotone transference plan satisfies this hypothesis.
Therefore this concrete example provide also a confirmation of the validity of the proposed strategy and, in particular, of the
non artificiality of the assumptions.
\subsection{Structure of the paper}
\label{Ss:structure}
The paper is organized as follows.
In Section \ref{S:preli}, we recall the mathematical tools we use in this paper.
In Section \ref{Ss:univmeas} the fundamental results of projective set theory are listed.
In Section \ref{S:disintegrazione} we recall the Disintegration Theorem.
Next, the basic results of selection principles are listed in Section \ref{Ss:sele}, and in Section \ref{ss:Metric} we define the geodesic structure $(X,d,d_N)$ which is studied in this paper. Finally, Section \ref{Ss:General Facts} recalls some fundamental results in optimal transportation.
Section \ref{S:Optimal} shows how using only the $d_N$-cyclical monotonicity of a set $\Gamma$ we can obtain a partial order relation $G \subset X \times X$ as follows (Lemma \ref{L:analGR} and Proposition \ref{P:equiv}): $xGy$ iff there exists $(w,z) \in \Gamma$ and a geodesic $\gamma$, passing trough $w$ and $z$ and with direction $w \to z$, such that $x$, $y$ belongs to $\gamma$ and $\gamma^{-1}(x) \leq \gamma^{-1}(y)$. This set $G$ is analytic, and allows to define
\begin{itemize}
\item the transport rays set $R$ (\ref{E:Rray}),
\item the transport sets $\mathcal T_e$, $\mathcal T$ (with and without and points) (\ref{E:TR0}),
\item the set of initial points $a$ and final points $b$ (\ref{E:endpoint0}).
\end{itemize}
Even if this part of Section \ref{S:Optimal} contains the same results of the first part of Section 3 of \cite{biacava:streconv},
for being as self contained as possible, we state this results and show their proofs again.
The main difference with the non-branching case is that here $R$ is not an equivalence relation.
Therefore the approach proposed in \cite{biacava:streconv} doesn't work anymore and indeed
the only common part with \cite{biacava:streconv} is the first part of Section \ref{S:Optimal}.
To obtain an equivalence relation $H \subset X \times X$
we have to consider the set of couples $(x,y)$ for $x,y\in \mathcal{T}$ such that
there is a continuous path from $x$ to $y$, union of a
finite number of transport rays never passing through $a\cup b$, Definition \ref{D:concat}.
In Proposition \ref{P:equiv} we prove that $H$ is an equivalence relation.
Section \ref{S:partition} proves that the compatibility conditions
\ref{I:comegeo} and \ref{I:comelocpt}
between $d_{N}$ and $d$ imply that the disintegration induced by $H$ on
$\mathcal T$ is strongly consistent (Proposition \ref{P:sicogrF}). Using this fact we can reduce the analysis on $H(y)$ for $y$ in the quotient set.
In Section \ref{S:pt iniziali e finali} we prove Theorem \ref{T:-1}. We first introduce the operation $A \mapsto A_t$, the translation along geodesics (\ref{E:At}), and show that $t \mapsto \mu(A_t)$ is a $\mathcal{A}$-measurable function if $A$ is analytic (Lemma \ref{L:measumuAt}). \\ Next, we show that under the assumption
\[
\mu(A) > 0 \quad \Longrightarrow \quad \mu(A_{t_{n}}) \geq C \mu(A)
\]
for an infinitesimal sequence $t_{n}$ and $C>0$,
the set of initial points $a$ is $\mu$-negligible (Proposition \ref{P:puntini}) and the conditional probabilities $\mu_y$ are continuous.
In Section \ref{S:Solution} we prove Theorem \ref{T:0}.
First in Theorem \ref{T:finale} we prove that gluing all the $d_{N}$-cyclically monotone maps defined on $H(y)$
we obtain a measurable transference map $T$ from $\mu$ to $\nu$ $d_{N}$-cyclically monotone.
Then the assumption on the structure of $\Gamma$ is stated (Assumption \ref{A:clessidra3})
and in Proposition \ref{P:alternativa} we show that on the equivalence class $H(y)$ satisfying Assumption \ref{A:clessidra3} there exists an
optimal transference map $T_{y}$ from $\mu_{y}$ to $\nu_{y}$, provided the quotient measure and the marginal probabilities of $\mu_{y}$
induced by the partition given by Assumption \ref{A:clessidra3} are continuous.
Section \ref{S:application} gives an application of Theorem \ref{T:0}: $M$ is a connected smooth hyper-surface of $\mathbb{R}^{d}$ that
is the boundary of a convex and compact set $C$. Let $X=cl (\mathbb{R}^{d}\setminus C)$.
The distance $d_{M}$ is the minimum of the euclidean length among all the Lipschitz curves in $X$ (\ref{E:distostc}).
Hence $C$ is to be intended as an obstacle for euclidean geodesics.
The geodesic space $(X,d_{M})$ fits into the setting of Theorem \ref{T:0} (Lemma \ref{L:cont} and Remark \ref{O:sempre}).
If $\mu \ll \mathcal{L}^{d}$ then the $\mu$-measure of the set of initial points is zero and the marginal $\mu_{y}$ are continuous (Lemma \ref{L:appl}).
Finally we show in Proposition \ref{P:fondamentale} and Proposition \ref{P:mappaostacolo1} that any $d_{M}$-cyclically monotone set and $\mu$ satisfy the hypothesis of Proposition \ref{P:alternativa}.
It follows the existence of a solution for the Monge minimization problem.
\section{Preliminaries}
\label{S:preli}
In this section we recall some general facts about projective classes, the Disintegration Theorem for measure, measurable selection principles, geodesic spaces and optimal transportation problems.
\subsection{Borel, projective and universally measurable sets}
\label{Ss:univmeas}
The \emph{projective class $\Sigma^1_1(X)$} is the family of subsets $A$ of the Polish space $X$ for which there exists $Y$ Polish and $B \in \mathcal{B}(X \times Y)$ such that $A = P_1(B)$. The \emph{coprojective class $\Pi^1_1(X)$} is the complement in $X$ of the class $\Sigma^1_1(X)$. The class $\Sigma^1_1$ is called \emph{the class of analytic sets}, and $\Pi^1_1$ are the \emph{coanalytic sets}.
The \emph{projective class $\Sigma^1_{n+1}(X)$} is the family of subsets $A$ of the Polish space $X$ for which there exists $Y$ Polish and $B \in \Pi^1_n(X \times Y)$ such that $A = P_1(B)$. The \emph{coprojective class $\Pi^1_{n+1}(X)$} is the complement in $X$ of the class $\Sigma^1_{n+1}$.
If $\Sigma^1_n$, $\Pi^1_n$ are the projective, coprojective pointclasses, then the following holds (Chapter 4 of \cite{Sri:courseborel}):
\begin{enumerate}
\item $\Sigma^1_n$, $\Pi^1_n$ are closed under countable unions, intersections (in particular they are monotone classes);
\item $\Sigma^1_n$ is closed w.r.t. projections, $\Pi^1_n$ is closed w.r.t. coprojections;
\item if $A \in \Sigma^1_n$, then $X \setminus A \in \Pi^1_n$;
\item the \emph{ambiguous class} $\Delta^1_n = \Sigma^1_n \cap \Pi^1_n$ is a $\sigma$-algebra
and $\Sigma^1_n \cup \Pi^1_n \subset \Delta^1_{n+1}$.
\end{enumerate}
We will denote by $\mathcal{A}$ the $\sigma$-algebra generated by $\Sigma^1_1$: clearly $\mathcal{B} = \Delta^1_1 \subset \mathcal{A} \subset \Delta^1_2$.
We recall that a subset of $X$ Polish is \emph{universally measurable} if it belongs to all completed $\sigma$-algebras of all Borel measures on $X$:
it can be proved that every set in $\mathcal{A}$ is universally measurable.
We say that $f:X \to \mathbb{R} \cup \{\pm \infty\}$ is a \emph{Souslin function} if $f^{-1}(t,+\infty] \in \Sigma^{1}_{1}$.
\begin{lemma}
\label{L:measuregP}
If $f : X \to Y$ is universally measurable, then $f^{-1}(U)$ is universally measurable if $U$ is.
\end{lemma}
See \cite{biacava:streconv} for the proof.
\subsection{Disintegration of measures}
\label{S:disintegrazione}
We follow the approach of \cite{biacar:cmono}.
Given a measurable space $(R, \mathscr{R})$ and a function $r: R \to S$, with $S$ generic set, we can endow $S$ with the \emph{push forward $\sigma$-algebra} $\mathscr{S}$ of $\mathscr{R}$:
$$
Q \in \mathscr{S} \iff r^{-1}(Q) \in \mathscr{R},
$$
which could be also defined as the biggest $\sigma$-algebra on $S$ such that $r$ is measurable. Moreover given a measure space
$(R,\mathscr{R},\rho)$, the \emph{push forward measure} $\eta$ is then defined as $\eta := (r_{\sharp}\rho)$.
Consider a probability space $(R, \mathscr{R},\rho)$ and its push forward measure space $(S,\mathscr{S},\eta)$ induced by a map $r$. From the above definition the map $r$ is clearly measurable and inverse measure preserving.
\begin{definition}
\label{defi:dis}
A \emph{disintegration} of $\rho$ \emph{consistent with} $r$ is a map $\rho: \mathscr{R} \times S \to [0,1]$ such that
\begin{enumerate}
\item $\rho_{s}(\cdot)$ is a probability measure on $(R,\mathscr{R})$, for all $s\in S$,
\item $\rho_{\cdot}(B)$ is $\eta$-measurable for all $B \in \mathscr{R}$,
\end{enumerate}
and satisfies for all $B \in \mathscr{R}, C \in \mathscr{S}$ the consistency condition
$$
\rho\left(B \cap r^{-1}(C) \right) = \int_{C} \rho_{s}(B) \eta(ds).
$$
A disintegration is \emph{strongly consistent with r} if for all $s$ we have $\rho_{s}(r^{-1}(s))=1$.
\end{definition}
We say that a $\sigma$-algebra $\mathcal{A}$ is \emph{essentially countably generated} with respect to a measure $m$, if there exists a countably generated $\sigma$-algebra $\hat{\mathcal{A}}$ such that for all $A \in \mathcal{A}$ there exists $\hat{A} \in \hat{\mathcal{A}}$ such that $m (A \vartriangle \hat{A})=0$.
We recall the following version of the theorem of disintegration of measure that can be found on \cite{Fre:measuretheory4}, Section 452.
\begin{theorem}[Disintegration of measure]
\label{T:disintr}
Assume that $(R,\mathscr{R},\rho)$ is a countably generated probability space, $R = \cup_{s \in S}R_{s}$ a decomposition of R, $r: R \to S$ the quotient map
and $\left( S, \mathscr{S},\eta \right)$ the quotient measure space. Then $\mathscr{S}$ is essentially countably generated w.r.t. $\eta$ and there exists a unique disintegration $s \to \rho_{s}$ in the following sense:
if $\rho_{1}, \rho_{2}$ are two consistent disintegration then $\rho_{1,s}(\cdot)=\rho_{2,s}(\cdot)$ for $\eta-$a.e. $s$.
If $\left\{ S_{n}\right\}_{n\in \mathbb{N}}$ is a family essentially generating
$\mathscr{S}$ define the equivalence relation:
$$
s \sim s' \iff \ \{ s \in S_{n} \iff s'\in S_{n}, \ \forall\, n \in \mathbb{N}\}.
$$
Denoting with p the quotient map associated to the above equivalence relation and with $(L,\mathscr{L}, \lambda)$ the quotient measure space, the following properties hold:
\begin{itemize}
\item $R_{l}:= \cup_{s\in p^{-1}(l)}R_{s} = (p \circ r)^{-1}(l)$ is $\rho$-measurable and $R= \cup_{l\in L}R_{l}$;
\item the disintegration $\rho = \int_{L}\rho_{l} \lambda(dl)$ satisfies $\rho_{l}(R_{l})=1$, for $\lambda$-a.e. $l$. In particular there exists a
strongly consistent disintegration w.r.t. $p \circ r$;
\item the disintegration $\rho = \int_{S}\rho_{s} \eta(ds)$ satisfies $\rho_{s}= \rho_{p(s)}$, for $\eta$-a.e. $s$.
\end{itemize}
\end{theorem}
In particular we will use the following corollary.
\begin{corollary}
\label{C:disintegration}
If $(S,\mathscr{S})=(X,\mathcal{B}(X))$ with $X$ Polish space, then the disintegration is strongly consistent.
\end{corollary}
\subsection{Selection principles}
\label{Ss:sele}
Given a multivalued function $F: X \to Y$, $X$, $Y$ metric spaces, the \emph{graph} of $F$ is the set
\begin{equation}
\label{E:graphF}
\textrm{graph}(F) := \big\{ (x,y) : y \in F(x) \big\}.
\end{equation}
The \emph{inverse image} of a set $S\subset Y$ is defined as:
\begin{equation}
\label{E:inverseF}
F^{-1}(S) := \big\{ x \in X\ :\ F(x)\cap S \neq \emptyset \big\}.
\end{equation}
For $F \subset X \times Y$, we denote also the sets
\begin{equation}
\label{E:sectionxx}
F_x := F \cap \{x\} \times Y, \quad F^y := F \cap X \times \{y\}.
\end{equation}
In particular, $F(x) = P_2(\textrm{graph}(F)_x)$, $F^{-1}(y) = P_1(\textrm{graph}(F)^y)$. We denote by $F^{-1}$ the graph of the inverse function
\begin{equation}
\label{E:F-1def}
F^{-1} := \big\{ (x,y): (y,x) \in F \big\}.
\end{equation}
We say that $F$ is \emph{$\mathcal{R}$-measurable} if $F^{-1}(B) \in \mathcal{R}$ for all $B$ open. We say that $F$ is \emph{strongly Borel measurable} if inverse images of closed sets are Borel. A multivalued function is called \emph{upper-semicontinuous} if the preimage of every closed set is closed: in particular u.s.c. maps are strongly Borel measurable.
In the following we will not distinguish between a multifunction and its graph. Note that the \emph{domain of $F$} (i.e. the set $P_1(F)$) is in general a subset of $X$. The same convention will be used for functions, in the sense that their domain may be a subset of $X$.
Given $F \subset X \times Y$, a \emph{section $u$ of $F$} is a function from $P_1(F)$ to $Y$ such that $\textrm{graph}(u) \subset F$. We recall the following selection principle, Theorem 5.5.2 of \cite{Sri:courseborel}, page 198.
\begin{theorem}[Von Neumann]
\label{T:vanneuma}
Let $X$ and $Y$ be Polish spaces, $A \subset X \times Y$ analytic, and $\mathcal{A}$ the $\sigma$-algebra generated by the analytic subsets of X. Then there is an $\mathcal{A}$-measurable section $u : P_1(A) \to Y$ of $A$.
\end{theorem}
A \emph{cross-section of the equivalence relation $E$} is a set $S \subset E$ such that the intersection of $S$ with each equivalence class is a singleton. We recall that a set $A \subset X$ is saturated for the equivalence relation $E \subset X \times X$ if $A = \cup_{x \in A} E(x)$.
The next result is taken from \cite{Sri:courseborel}, Theorem 5.2.1.
\begin{theorem}
\label{T:KRN}
Let $Y$ be a Polish space, $X$ a nonempty set, and $\mathcal{L}$ a $\sigma$-algebra of subset of $X$.
Every $\mathcal{L}$-measurable, closed value multifunction $F:X \to Y$ admits an $\mathcal{L}$-measurable section.
\end{theorem}
A standard corollary of the above selection principle is that if the disintegration is strongly consistent in a Polish space, then up to a saturated set of negligible measure there exists a Borel cross-section.
In particular, we will use the following corollary.
\begin{corollary}
\label{C:weelsupprr}
Let $F \subset X \times X$ be $\mathcal{A}$-measurable, $X$ Polish, such that $F_x$ is closed and define the equivalence relation $x \sim y \ \Leftrightarrow \ F(x) = F(y)$. Then there exists a $\mathcal{A}$-section $f : P_1(F) \to X$ such that $(x,f(x)) \in F$ and $f(x) = f(y)$ if $x \sim y$.
\end{corollary}
\begin{proof}
For all open sets $G \subset X$, consider the sets $F^{-1}(G) = P_1(F \cap X \times G) \in \mathcal{A}$, and let $\mathcal{R}$ be the $\sigma$-algebra generated by $F^{-1}(G)$. Clearly $\mathcal{R} \subset \mathcal{A}$.
If $x \sim y$, then
\[
x \in F^{-1}(G) \quad \Longleftrightarrow \quad y \in F^{-1}(G),
\]
so that each equivalence class is contained in an atom of $\mathcal{R}$, and moreover by construction $x \mapsto F(x)$ is $\mathcal{R}$-measurable.
We thus conclude by using Theorem \ref{T:KRN} that there exists an $\mathcal{R}$-measurable section $f$: this measurability condition implies that $f$ is constant on atoms, in particular on equivalence classes.
\end{proof}
\subsection{Metric setting}
\label{ss:Metric}
In this section we refer to \cite{burago}.
\begin{definition}
\label{D:lengthstr}
A \emph{length structure} on a topological space $X$ is a class $\mathtt{A}$ of admissible paths, which is a subset of all continuous paths in X, together with a map $L: \mathtt{A} \to [0,+\infty]$: the map $L$ is called \emph{length of path}. The class $\mathtt{A}$ satisfies the following assumptions:
\begin{description}
\item[closure under restrictions] if $\gamma : [a,b] \to X$ is admissible and $a \leq c\leq d \leq b$, then $\gamma \llcorner_{[c,d]}$ is also admissible.
\item[closure under concatenations of paths] if $\gamma : [a,b] \to X$ is such that its restrictions $\gamma_1, \gamma_2$ to $[a,c]$ and $[c,b]$ are both admissible, then so is $\gamma$.
\item[closure under admissible reparametrizations] for an admissible path $\gamma : [a,b] \to X$ and a for $\varphi: [c,d]\to[a,b]$, $\varphi \in B$, with $B$ class of admissible homeomorphisms that includes the linear one, the composition $\gamma(\varphi(t))$ is also admissible.
\end{description}
The map $L$ satisfies the following properties:
\begin{description}
\item[additivity] $L(\gamma \llcorner_{[a,b]}) = L(\gamma \llcorner_{[a,c]}) + L(\gamma \llcorner_{[c,b]})$ for any $c\in [a,b]$.
\item[continuity] $L(\gamma \llcorner_{[a,t]})$ is a continuous function of $t$.
\item[invariance] The length is invariant under admissible reparametrizations.
\item[topology] Length structure agrees with the topology of $X$ in the following sense: for a neighborhood $U_x$ of a point $x \in X$, the length of paths connecting $x$ with points of the complement of $U_x$ is separated from zero:
\[
\inf \big\{ L(\gamma) : \gamma(a)=x, \gamma(b) \in X\setminus U_{x} \big\} >0.
\]
\end{description}
\end{definition}
Given a length structure, we can define a distance
\[
d_N(x,y) = \inf \Big\{ L(\gamma): \gamma:[a,b]\to X, \gamma \in \mathtt{A}, \gamma(a)=x, \gamma(b)=y \Big\},
\]
that makes $(X,d_{N})$ a metric space (allowing $d_{N}$ to be $+\infty$). The metric $d_{N}$ is called \emph{intrinsic}.
It follows from Proposition 2.5.9 of \cite{burago} that every admissible curve of finite length admits a constant speed parametrization, i.e.
$\gamma$ defined on $[0,1]$ and $L(\gamma\llcorner[t,t'])= v (t'-t)$, with $v$ velocity.
\begin{definition}
A length structure is said to be \emph{complete} if for every two points $x,y$ there exists an admissible path joining them whose length $L(\gamma)$ is equal to $d_{N}(x,y)$.
\end{definition}
Observe that in the previous definition we do no require $d_{N}(x,y)<+\infty$.
Intrinsic metrics associated with complete length structure are said to be \emph{strictly intrinsic}. The metric space $(X,d_N)$ with $d_N$ strictly intrinsic is called a \emph{geodesic space}. A curve whose length equals the distance between its end points is called \emph{geodesic}.
From now on we assume the following: \label{P:assumpDL}
\medskip
\begin{enumerate}
\item $(X,d)$ Polish space;
\item $d_N : X \times X \to [0,+\infty]$ is a Borel distance;
\item $(X,d_N)$ is a geodesic space;
\end{enumerate}
\medskip
Since we have two metric structures on $X$, we denote the quantities relating to $d_N$ with the subscript $N$: for example
\[
B_r(x) = \big\{ y : d(x,y) < r \big\}, \quad B_{r,N}(x) = \big\{ y : d_N(x,y) < r \big\}.
\]
In particular we will use the notation
\[
D_N(x) = \big\{ y : d_N(x,y) < + \infty \big\},
\]
$(\mathcal{K},d_H)$ for the compact sets of $(X,d)$ with the Hausdorff distance $d_H$ and $(\mathcal{K}_N,d_{H,N})$ for the compact sets of $(X,d_N)$ with the Hausdorff distance $d_{H,N}$. We recall that $(\mathcal{K},d_H)$ is Polish.
\begin{lemma}
\label{L:measclos}
If $A$ is analytic in $(X,d)$, then $\{x : d_N(A,x) < \varepsilon \}$ is analytic for all $\varepsilon > 0$.
\end{lemma}
\begin{proof}
Observe that
\[
\big\{ x : d_N(A,x) < \varepsilon \big\} = P_1 \Big( X \times A \cap \big\{ (x,y) :d_N(x,y) < \varepsilon \big\} \Big),
\]
so that the conclusion follows from the invariance of the class $\Sigma^1_1$ w.r.t. projections.
\end{proof}
In particular, $\overline{A}^{_{d_N}}$, the closure of $A$ w.r.t. $d_N$, is analytic if $A$ is analytic.
\subsection{General facts about optimal transportation}
\label{Ss:General Facts}
Let $(X,\Omega,\mu)$ and $(Y,\Sigma,\nu)$ be two probability spaces and $c:X \times Y \to \mathbb{R}^{+}$ be a $\Omega \times \Sigma$ measurable function. Consider the set of transference plans
\[
\Pi (\mu,\nu) : =\Big\{ \pi \in \mathcal{P}(X\times Y) : (P_{1})_{\sharp}\pi = \mu, (P_{2})_{\sharp}\pi = \nu \Big\},
\]
where $P_{i}(x_{1},x_{2})= x_{i}, i=1,2$. Define the functional
\begin{equation}
\label{E:Ifunct}
\begin{array}{ccccl}
\mathcal{I} &:& \Pi(\mu,\nu) &\longrightarrow& \mathbb{R}^{+} \cr
& & \pi & \longmapsto &\displaystyle \mathcal{I}(\pi):=\int_{X\times Y} c \pi.
\end{array}
\end{equation}
The \emph{Monge-Kantorovich minimization problem} is the minimization of $\mathcal{I}$ over all transference plans.
If we consider a map $T : X \to Y$ such that $T_{\sharp}\mu=\nu$, the functional (\ref{E:Ifunct}) becomes
\[
\mathcal{I}(T):= \mathcal{I}( (Id \times T)_{\sharp}\mu ) = \int_{X} c(x,T(x)) \mu(dx).
\]
The minimization problem over all $T$ is called \emph{Monge minimization problem}.
The Kantorovich problem admits a (pre) dual formulation: before stating it, we introduce two definitions.
\begin{definition}
A map $\varphi : X \to \mathbb{R} \cup \left\{ -\infty \right\} $ is said to be \emph{$c$-concave} if it is not identically $-\infty$ and there exists $\psi : Y \to \mathbb{R} \cup \left\{ - \infty \right\}$, $\psi \not\equiv{-\infty}$, such that
\[
\varphi (x) = \inf_{y \in Y} \left\{ c(x,y) - \psi (y) \right\}.
\]
The \emph{$c$-transform} of $\varphi$ is the function
\begin{equation}
\label{E:ctransf}
\varphi^{c}(y):= \inf_{x\in X} \left\{ c(x,y) - \varphi (x) \right\}.
\end{equation}
The \emph{$c$-superdifferential $\partial^{c}\varphi$} of $\varphi$ is the subset of $X \times Y$ defined by
\begin{equation}
\label{E:csudiff}
(x,y) \in \partial^{c}\varphi \iff\ c(x,y) - \varphi(x) \leq c(z,y) - \varphi(z) \quad \forall z \in X.
\end{equation}
\end{definition}
\begin{definition}
A set $\Gamma \subset X\times Y$ is said to be
\emph{c-cyclically monotone} if, for any $n \in \mathbb{N}$ and for any family $(x_{1},y_{1}), \dots, (x_{n},y_{n})$ of points of $\Gamma$, the following inequality holds
\[
\sum_{i=0}^{n}c(x_{i},y_{i}) \leq \sum_{i=0}^{n}c(x_{i+1},y_{i}),
\]
with $x_{n+1}=x_{1}$. A transference plan is said to be \emph{$c$-cyclically monotone} (or just \emph{$c$-monotone}) if it is concentrated on a $c$-cyclically monotone set.
\end{definition}
Consider the set
\begin{equation}
\label{E:Phicset}
\Phi_c := \Big\{ (\varphi,\psi) \in L^1(\mu) \times L^1(\nu): \varphi(x) + \psi(y) \leq c(x,y) \Big\}.
\end{equation}
Define for all $(\varphi,\psi)\in \Phi_c$ the functional
\begin{equation}
\label{E:Jfunct}
J(\varphi,\psi) := \int \varphi \mu + \int \psi \nu.
\end{equation}
The following is a well known result (see Theorem 5.10 of \cite{villa:Oldnew}).
\begin{theorem}[Kantorovich Duality]
\label{T:kanto}
Let X and Y be Polish spaces, let $\mu \in \mathcal{P}(X)$ and $\nu \in \mathcal{P}(Y)$, and let $c: X\times Y \to \mathbb{R}^{+}\cup \left\{ +\infty \right\}$ be lower semicontinuous.
Then the following holds:
\begin{enumerate}
\item Kantorovich duality:
\[
\inf_{\pi \in \Pi[\mu,\nu]} \mathcal{I} (\pi) = \sup _{(\varphi,\psi)\in \Phi_{c}} J(\varphi,\psi).
\]
Moreover, the infimum on the left-hand side is attained and the right-hand side is also equal to
\[
\sup _{(\varphi,\psi)\in \Phi_{c}\cap C_{b}} J(\varphi,\psi),
\]
where $C_{b}= C_{b}(X, \mathbb{R}) \times C_{b}(Y,\mathbb{R})$.
\item If $c$ is real valued and the optimal cost
\[
C(\mu,\nu):=\inf _{\pi \in \Pi(\mu,\nu)} I (\pi)
\]
is finite, then there is a measurable $c$-cyclically monotone set $\Gamma \subset X\times Y$, closed if $c$ is continuous, such that for any $\pi \in \Pi(\mu,\nu)$ the following statements are equivalent:
\begin{enumerate}
\item $\pi$ is optimal;
\item $\pi$ is $c$-cyclically monotone;
\item $\pi$ is concentrated on $\Gamma$;
\item there exists a $c$-concave function $\varphi$ such that $\pi$-a.s. $\varphi(x)+\varphi^{c}(y)=c(x,y)$.
\end{enumerate}
\item If moreover
\[
c(x,y) \leq c_{X}(x) + c_{Y}(y), \quad \ c_{X}\ \mu\textrm{-measurable}, \ c_{Y}\ \nu\textrm{-measurable},
\]
then there exist a couple of potentials and the optimal transference plan $\pi$ is concentrated on the set
\[
\big\{ (x,y) \in X\times Y\, | \, \varphi (x) + \psi (y) = c(x,y) \big\}.
\]
Finally if $(c_{X},c_{Y})\in \mathcal{L}^{1}(\mu) \times \mathcal{L}^{1}(\nu)$ then the supremum is attained
\[
\sup_{\Phi_{c}} J = J(\varphi, \varphi^{c}).
\]
\end{enumerate}
\end{theorem}
We recall also that if $-c$ is Souslin, then every optimal transference plan $\pi$ is concentrated on a $c$-cyclically monotone set \cite{biacar:cmono}.
\section{Optimal transportation in geodesic spaces}
\label{S:Optimal}
Let $\mu, \nu \in \mathcal{P}(X)$ and consider the transportation problem with cost $c(x,y)= d_N(x,y)$, and let $\pi \in \Pi(\mu,\nu)$ be a $d_N$-cyclically monotone transference plan with finite cost. By inner regularity, we can assume that the optimal transference plan is concentrated on a $\sigma$-compact $d_N$-cyclically monotone set $\Gamma \subset \{d_N(x,y) < +\infty\}$.
Consider the set
\begin{eqnarray}
\label{E:gGamma}
\Gamma' & :=&~ \bigg\{ (x,y) : \exists I \in \mathbb{N}_0, (w_i,z_i) \in \Gamma \ \textrm{for}
\ i = 0,\dots,I, \ z_I = y \crcr
& &~ \qquad \qquad w_{I+1} = w_0 = x, \ \sum_{i=0}^I d_N(w_{i+1},z_i) - d_N(w_i,z_i) = 0 \bigg\}.
\end{eqnarray}
In other words, we concatenate points $(x,z), (w,y) \in \Gamma$ if they are initial and final point of a cycle with total cost $0$.
\begin{lemma}
\label{L:gGamma}
The following holds:
\begin{enumerate}
\item $\Gamma \subset \Gamma' \subset \{d_N(x,y) < +\infty\}$;
\item if $\Gamma$ is analytic, so is $\Gamma'$;
\item if $\Gamma$ is $d_N$-cyclically monotone, so is $\Gamma'$.
\end{enumerate}
\end{lemma}
\begin{proof}
For the first point, set $I=0$ and $(w_{n,0},z_{n,0}) = (x,y)$ for the first inclusion. If $d_N(x,y) = +\infty$, then $(x,y) \notin \Gamma$ and all finite set of points in $\Gamma$ are bounded.
For the second point, observe that
\begin{eqnarray*}
\Gamma' &=&~ \bigcup_{I \in \mathbb{N}_0} P_{1,2I+1} (A_I) \crcr
&=&~ \bigcup_{I \in \mathbb{N}_0} P_{1,2I+1} \bigg( \prod_{i=0}^I \Gamma \cap \bigg\{ \prod_{i=0}^I (w_i,z_i) : \sum_{i=0}^I d_N(w_{i+1},z_i) - d_N(w_i,z_i) = 0, w_{I+1} = w_0 \bigg\} \bigg).
\end{eqnarray*}
For each $I \in \mathbb{N}_0$, since $d_N$ is Borel, it follows that
\begin{eqnarray*}
\bigg\{ \prod_{i=0}^I (w_i,z_i) : \sum_{i=0}^I d_N(w_{i+1},z_i) - d_N(w_i,z_i) = 0, w_{I+1} = w_0 \bigg\}
\end{eqnarray*}
is Borel in $\prod_{i=0}^I (X \times X)$, so that for $\Gamma$ analytic each set $A_{n,I}$ is analytic. Hence $P_{1,2I+1}(A_I)$ is analytic, and since the class $\Sigma^1_1$ is closed under countable unions and intersections it follows that $\Gamma'$ is analytic.
For the third point, observe that for all $(x_j,y_j) \in \Gamma'$, $j=0,\dots,J$, there are $(w_{j,i},z_{j,i}) \in \Gamma$, $i = 0,\dots,I_j$, such that
\begin{eqnarray*}
d_N(x_j,y_j) + \sum_{i=0}^{I_j-1} d_N(w_{j,i+1},z_{j,i}) - \sum_{i=0}^{I_j} d_N(w_{j,i},z_{j,i}) = 0.
\end{eqnarray*}
Hence we can write for $x_{J+1} = x_0$, $w_{j,I_j+1} = w_{j+1,0}$, $w_{J+1,0} = w_{0,0}$
\begin{eqnarray*}
\sum_{j=0}^J d_N(x_{j+1},y_j) - d_N(x_j,y_j) =&~ \sum_{j=0}^J \sum_{i=0}^{I_j} d_N(w_{j,i+1},z_{j,i}) - d_N(w_{j,i},z_{j,i}) \geq 0,
\end{eqnarray*}
using the $d_N$-cyclical monotonicity of $\Gamma$.
\end{proof}
\begin{definition}[Transport rays]
\label{D:Gray}
Define the \emph{set of oriented transport rays}
\begin{equation}
\label{E:trG}
G := \Big\{ (x,y): \exists (w,z) \in \Gamma', d_N(w,x) + d_N(x,y) + d_N(y,z) = d_N(w,z) \Big\}.
\end{equation}
For $x \in X$, the \emph{outgoing transport rays from $x$} is the set $G(x)$ and the \emph{incoming transport rays in $x$} is the set $G^{-1}(x)$. Define the \emph{set of transport rays} as the set
\begin{equation}
\label{E:Rray}
R := G \cup G^{-1}.
\end{equation}
\end{definition}
The set $G$ is the set of all couples of points on oriented geodesics with endpoints in $\Gamma'$.
In $R$ the couples are non oriented.
\begin{lemma}
\label{L:analGR}
The following holds:
\begin{enumerate}
\item $G$ is $d_N$-cyclically monotone;
\item $\Gamma' \subset G \subset \{d_N(x,y) < +\infty\}$;
\item the sets $G$, $R := G \cup G^{-1}$ are analytic.
\end{enumerate}
\end{lemma}
\begin{proof}
The second point follows by the definition: if $(x,y) \in \Gamma'$, just take $(w,z) = (x,y)$ in the r.h.s. of (\ref{E:trG}).
The third point is consequence of the fact that
\[
G = P_{34} \Big( \big( \Gamma' \times X \times X \big) \cap \Big\{ (w,z,x,y) : d_N(w,x) + d_N(x,y) + d_N(y,z) = d_N(w,z) \Big\} \Big),
\]
and the result follows from the properties of analytic sets.
The first point follows from the following observation: if $(x_i,y_i) \in \gamma_{[w_i,z_i]}$, then from triangle inequality
\begin{eqnarray*}
d_N(x_{i+1},y_i) - d_N(x_i,y_i) &\geq &~ d_N(x_{i+1},z_i) - d_N(z_i,y_i) - d_N(x_i,y_i) \crcr
&=&~ d_N(x_{i+1},z_i) - d_N(x_i,z_i) \crcr
&\geq&~ d_N(w_{i+1},z_i) - d_N(w_{i+1},x_{i+1}) - d_N(x_i,z_i) \crcr
&=&~ d_N(w_{i+1},z_i) - d_N(w_i,z_i) + d_N(w_i,x_{i}) - d_N(w_{i+1},x_{i+1}).
\end{eqnarray*}
Since $(w_{n+1},x_{n+1})=(w_{1},x_{1})$, it follows that
\[
\sum_{i=1}^{n} d_N(x_{i+1},y_i) - d_N(x_i,y_i) \geq \sum_{i=1}^{n} d_N(w_{i+1},z_i) - d_N(w_i,z_i) \geq 0.
\]
Hence the set $G$ is $d_N$-cyclically monotone.
\end{proof}
\begin{definition} Define the \emph{transport sets}
\begin{subequations}
\label{E:TR0}
\begin{eqnarray}
\label{E:TR}
\mathcal T :=&~ P_1 \big( G^{-1} \setminus \{x = y\} \big) \cap P_1 \big( G \setminus \{x = y\} \big), \\
\label{E:TRe}
\mathcal T_e :=&~ P_1 \big( G^{-1} \setminus \{x = y\} \big) \cup P_1 \big( G \setminus \{x = y\} \big).
\end{eqnarray}
\end{subequations}
\end{definition}
Since $G$ and $G^{-1}$ are analytic sets, $\mathcal{T}$, $\mathcal{T}_e$ are analytic. The subscript $e$ refers to the endpoints of the geodesics: clearly we have
\begin{equation}
\label{E:RTedef}
\mathcal{T}_e = P_1(R \setminus \{x = y\}).
\end{equation}
The following lemma shows that we have only to study the Monge problem in $\mathcal{T}_e$.
\begin{lemma}
\label{L:mapoutside}
It holds $\pi(\mathcal{T}_e \times \mathcal{T}_e \cup \{x = y\}) = 1$.
\end{lemma}
\begin{proof}
If $x \in P_1(\Gamma \setminus \{x=y\})$, then $x \in G^{-1}(y) \setminus \{y\}$. Similarly, $y \in P_2(\Gamma \setminus \{x=y\})$ implies that $y \in G(x) \setminus \{x\}$. Hence $\Gamma \setminus \mathcal{T}_e \times \mathcal{T}_e \subset \{x = y\}$.
\end{proof}
As a consequence, $\mu(\mathcal{T}_e) = \nu(\mathcal{T}_e)$ and any maps $T$ such that for $\nu \llcorner_{\mathcal{T}_e} = T_\sharp \mu \llcorner_{\mathcal{T}_e}$ can be extended to a map $T'$ such that $\nu = T_\sharp \mu$ with the same cost by setting
\begin{equation} \label{E:extere}
T'(x) =
\begin{cases}
T(x) & x \in \mathcal{T}_e \\ x & x \notin \mathcal{T}_e
\end{cases}
\end{equation}
\begin{definition}
\label{D:endpoint}
Define the multivalued \emph{endpoint graphs} by:
\begin{subequations}
\label{E:endpoint0}
\begin{eqnarray}
\label{E:endpointa}
a :=&~ \big\{ (x,y) \in G^{-1} : G^{-1}(y) \setminus \{y\} = \emptyset \big\}, \\
\label{E:endpointb}
b :=&~ \big\{ (x,y) \in G : G(y) \setminus \{y\} = \emptyset \big\}.
\end{eqnarray}
\end{subequations}
We call $P_2(a)$ the set of \emph{initial points} and $P_2(b)$ the set of \emph{final points}.
\end{definition}
\begin{proposition}
The following holds:
\begin{enumerate}
\item the sets
\[
a,b \subset X \times X, \quad a(A), b(A) \subset X,
\]
belong to the $\mathcal{A}$-class if $A$ analytic;
\item $a \cap b \cap \mathcal{T}_e \times X = \emptyset$;
\item $a(\mathcal{T}) = a(\mathcal{T}_e)$, $b(\mathcal{T}) = b(\mathcal{T}_e)$;
\item $\mathcal{T}_e = \mathcal{T} \cup a(\mathcal{T}) \cup b(\mathcal{T})$, $\mathcal{T} \cap (a(\mathcal{T}) \cup b(\mathcal{T})) = \emptyset$.
\end{enumerate}
\end{proposition}
\begin{proof}
Define
\[
C := \big\{ (x,y,z) \in \mathcal{T}_e \times \mathcal{T}_e \times \mathcal{T}_e: y \in G(x), z \in G(y) \big\} = (G \times X) \cap (X \times G)
\cap \mathcal{T}_e \times \mathcal{T}_e \times \mathcal{T}_e,
\]
that is clearly analytic. Then
\[
b = \big\{ (x,y) \in G : y \in G(x), G(y) \setminus \{y\} = \emptyset \} = G \setminus P_{12}(C \setminus X \times \{y=z\}),
\]
\[
b(A) = \big\{ y: y \in G(x), G(y) \setminus \{y\} = \emptyset, x \in A \} = P_2(G \cap A \times X) \setminus P_2(C \setminus X \times \{y=z\}).
\]
A similar computation holds for $a$:
\[
a = G^{-1} \setminus P_{23}(C \setminus \{x=y\} \times X), \quad a(A) = P_1(G_{S} \cap X \times A) \setminus P_1(C \setminus \{x=y\} \times X).
\]
Hence $a,b \in \mathcal{A}(X \times X)$, $a(A), b(A) \in \mathcal{A}(X)$, being the intersection of an analytic set with a coanalytic one.
If $x \in \mathcal{T}_e \setminus \mathcal{T}$, then it follows that $G(x)=\{x\}$ or $G^{-1}(x)=\{x\}$ hence $x \in a(x) \cup b(x)$.
The other points follow easily.
\end{proof}
\begin{definition}[Chain of transport rays]\label{D:concat}
Define the set of \emph{chain of transport rays}
\begin{eqnarray}
\label{E:partiz}
H &: = &~ \bigg\{ (x,y) \in \mathcal{T}_{e} \times \mathcal{T}_{e}: \exists I \in \mathbb{N}_{0}, z_{i}\in \mathcal{T} \ \textrm{for}\ 1 \leq i\leq I, \crcr
&&~ \qquad \qquad (z_{i},z_{i+1}) \in R, \ 0\leq i \leq I+1, \ z_{0}=x, \ z_{I+1}=y \bigg\}.
\end{eqnarray}
\end{definition}
Using similar techniques of Lemma \ref{L:gGamma} it can be shown that $H$ is analytic.
\begin{proposition}
\label{P:equiv}
The set $H \cap \mathcal{T} \times \mathcal{T}$ is an equivalence relation on $\mathcal{T}$. The set $G$ is a partial order relation on $\mathcal{T}_e$.
\end{proposition}
\begin{proof}
Using the definition of $H$, one has in $\mathcal{T}$:
\begin{enumerate}
\item $x \in \mathcal{T}$ clearly implies that $(x,x) \in H$;
\item since $R$ is symmetric, if $y \in H(x)$ then $x \in H(y)$;
\item if $y \in H(x)$, $z \in H(y)$, $x,y,z \in \mathcal{T}$.
Glue the path from $x$ to $y$ to the one from $y$ to $z$. Since $y \in \mathcal{T}$, $z \in H(x)$.
\end{enumerate}
The second part follows similarly:
\begin{enumerate}
\item $x \in \mathcal{T}_e$ implies that
\[
\exists (x,y) \in \big( G \setminus \{x=y\} \big) \cup \big( G^{-1} \setminus \{x=y\} \big),
\]
so that in both cases $(x,x) \in G$;
\item $(x,y), (y,z) \in G \setminus \{x=y\}$ implies by $d_N$-cyclical monotonicity that $(x,z) \in G$.
\end{enumerate}
\end{proof}
We finally show that we can assume that the $\mu$-measure of final points and the $\nu$-measure of the initial points are $0$.
\begin{lemma}
\label{L:finini0}
The sets $G \cap b(\mathcal{T}) \times X$, $G \cap X \times a(\mathcal{T})$ is a subset of the graph of the identity map.
\end{lemma}
\begin{proof}
From the definition of $b$ one has that
\[
x \in b(\mathcal{T}) \quad \Longrightarrow \quad G(x) \setminus \{x\} = \emptyset,
\]
A similar computation holds for $a$.
\end{proof}
Hence we conclude that
\[
\pi (b(\mathcal{T}) \times X) = \pi(G \cap b(\mathcal{T}) \times X) = \pi(\{x = y\})
\]
and following (\ref{E:extere}) we can assume that
\[
\mu(b(\mathcal{T})) = \nu(a(\mathcal{T})) = 0.
\]
\section{Partition of the transport set}
\label{S:partition}
To perform a disintegration we have to assume some regularity of the support $\Gamma$ of the transport plan $\pi \in \Pi(\mu,\nu)$.
From now on we will assume the following: \label{P:assumpDN}
\begin{assumption}\label{A:assu1}
We say that $\Gamma$ satisfies Assumption \ref{A:assu1} if
\begin{enumerate}[label=(\alph*), ref= (\alph*)]
\item \label{A:comegeo} for all $x\in \mathcal{T}$ and for all $r>0$ the set $H(x) \cap \overline{B_{r,N}(x)}^{d_N}$ is $d$-closed;
\item \label{A:comelocpt} for all $x\in \mathcal{T}$ there exists $r>0$ such that $d_{N}(x,\cdot)_{ \llcorner H(x) \cap \overline{B_{r}(x)}} $
is bounded.
\end{enumerate}
\end{assumption}
Note that points \ref{A:comegeo} and \ref{A:comelocpt} of Assumption \ref{A:assu1} were already introduced at page \pageref{I:comegeo}.
Let $\{x_i\}_{i \in \mathbb{N}}$ be a dense sequence in $(X,d)$.
\begin{lemma}
\label{L:reguloclco}
The sets
\[
W_{ijk} := \Big\{ x \in \mathcal{T} \cap \bar B_{2^{-j}}(x_i):
d_{N}(x,\cdot)_{\llcorner H(x)\cap \bar B_{2^{-j}}(x_i) } \leq k \Big\}
\]
form a countable covering of $\mathcal{T}$ of class $\mathcal{A}$.
\end{lemma}
\begin{proof}
We first prove the measurability. We consider separately the conditions defining $W_{ijk}$.
{\it Point 1.} The set
\[
A_{ij} := \mathcal{T} \cap \bar B_{2^{-j}}(x_i)
\]
is clearly analytic.
{\it Point 2.}
The set
\[
D_{ijk}: = \bigg\{ (x,y) \in H : d(x_{i},y)\leq 2^{-j}, d_{N}(x,y) > k \bigg\}
\]
is again analytic.
We finally can write
\begin{eqnarray*}
W_{ijk} = A_{ij} \cap P_{1}( D_{ijk})^{c},
\end{eqnarray*}
and the fact that $\mathcal{A}$ is a $\sigma$-algebra proves that $W_{ijk} \in \mathcal{A}$.
To show that it is a covering, notice that from \ref{A:comelocpt} of Assumption \ref{A:assu1} for all $x \in \mathcal{T}$ there exists $r>0$ such that, on the set $H(x)\cap \bar B_{r}(x)$,
$d_{N}(x,\cdot)$ is bounded. Choose $j$ and $i$ such that $2^{-j-1} \leq r$ and
$d(x_{i},x)\leq 2^{-j-1}$, hence
\[
\bar B_{2^{-j}}(x_{i}) \subset \bar B_{r}(x)
\]
and therefore for some $\bar k \in \mathbb{N}$ we obtain that $x \in W_{ijk}$.
\end{proof}
\begin{remark}
\label{R:compFF}
Observe that $\bar B_{2^{-j}}(x_i) \cap H(x)$ is closed for all $x \in W_{ijk}$.
Indeed take $\{y_{n}\}_{n\in \mathbb{N}} \subset \bar B_{2^{-j}}(x_i) \cap H(x)$ with $d(y_{n},y) \to 0$ as $n \to + \infty$, then since $x \in W_{ijk}$
it holds $d_{N}(x,y_{n}) \leq k$. By \ref{A:comegeo} of Assumption \ref{A:assu1}, $d_{N}(x,y)\leq k$ and $y \in \bar B_{2^{-j}}(x_i) \cap H(x)$.
\end{remark}
\begin{lemma}
\label{L:partitilW}
There exist $\mu$-negligible sets $N_{ijk} \subset W_{ijk}$ such that the family of sets
\begin{eqnarray*}
\mathcal{T}_{ijk} = H^{-1}(W_{ijk} \setminus N_{ijk}) \cap\mathcal{T}
\end{eqnarray*}
is a countable covering of $\mathcal{T} \setminus \cup_{ijk} N_{ijk}$ into saturated analytic sets.
\end{lemma}
\begin{proof}
First of all, since $W_{ijk} \in \mathcal{A}$, then there exists $\mu$-negligible set $N_{ijk} \subset W_{ijk}$ such that $W_{ijk} \setminus N_{ijk} \in \mathcal{B}(X)$. Hence $\{W_{ijk} \setminus N_{ijk}\}_{i,j,k \in \mathbb{N}}$ is a countable covering of $\mathcal{T} \setminus \cup_{ijk} N_{ijk}$. It follows immediately that $\{\mathcal{T}_{ijk}\}_{i,j,k \in \mathbb{N}}$ satisfies the lemma.
\end{proof}
From any analytic countable covering, we can find a countable partition into $\mathcal{A}$-class saturated sets by defining
\begin{equation}
\label{E:Zkije}
\mathcal{Z}_{m} := \mathcal{T}_{i_mj_mk_m} \setminus \bigcup_{m' = 1}^{m-1} \mathcal{T}_{i_{m'}j_{m'}k_{m'}},
\end{equation}
where
\[
\mathbb{N} \ni m \mapsto (i_m,j_m,k_m) \in \mathbb{N}^3
\]
is a bijective map.
Since $H$ is an equivalence relation on $\mathcal{T}$, we use this partition to prove the strong consistency.
On $\mathcal{Z}_m$, $m > 0$, we define the closed valued map
\begin{equation}
\label{E:mapTijkF}
\mathcal{Z}_m \ni x \mapsto F(x) := H(x) \cap \bar B_{2^{-j_m}}(x_{i_m}).
\end{equation}
\begin{proposition}
\label{P:sicogrF}
There exists a $\mu$-measurable cross section $f : \mathcal{T} \to \mathcal{T}$ for the equivalence relation $H$.
\end{proposition}
\begin{proof}
First we show that $F$ is $\mathcal{A}$-measurable: for $\delta > 0$,
\begin{eqnarray*}
F^{-1}(B_\delta(y)) =&~ \Big\{ x \in \mathcal{Z}_m: H(x) \cap B_{\delta}(y) \cap \bar B_{2^{-{j_m}}}(x_{i_m}) \not= \emptyset \Big\} \crcr
=&~ \mathcal{Z}_m \cap P_1 \Big( H \cap \big( X \times B_{\delta}(y) \cap \bar B_{2^{-{j_m}}}(x_{i_m}) \big) \Big).
\end{eqnarray*}
Being the intersection of two $\mathcal{A}$-class sets, $F^{-1}(B_\delta(y))$ is in $\mathcal{A}$.
In Remark \ref{R:compFF} we have observed that $F$ is a closed-valued map, hence,
from Lemma 5.1.4 of \cite{Sri:courseborel}, $\textrm{graph}(F)$ is $\mathcal{A}$-measurable.
By Corollary \ref{C:weelsupprr} there exists a $\mathcal{A}$-class section $f_m : \mathcal Z_m \to \bar B_{2^{-{j_m}}}(x_{i_m})$. The proposition follows by setting $f \llcorner_{\mathcal Z_m} = f_m$ on $\cup_m \mathcal Z_m$, and defining it arbitrarily on $\mathcal T \setminus \cup_m \mathcal Z_m$: the latter being
$\mu$-negligible, $f$ is $\mu$-measurable.
\end{proof}
Up to a $\mu$-negligible saturated set $\mathcal{T}_N$, we can assume it to have $\sigma$-compact range: just let $S \subset f(\mathcal{T})$ be a $\sigma$-compact set where $f_\sharp \mu$ is concentrated, and set
\begin{equation}
\label{E:TNngel}
\mathcal{T}_S := H^{-1}(S) \cap \mathcal{T}, \quad \mathcal{T}_N := \mathcal{T} \setminus \mathcal{T}_S, \quad \mu(\mathcal{T}_N) = 0.
\end{equation}
Hence we have a measurable cross-section
\[
\mathcal{S} := S \cup f(\mathcal{T}_N) = (\textrm{Borel}) \cup (f(\textrm{$\mu$-negligible})).
\]
Hence Disintegration Theorem \ref{T:disintr} yields
\begin{equation}
\label{E:disiparzmu}
\mu\llcorner_{\mathcal{T}} = \int_S \mu_y m(dy), \quad m = f_\sharp \mu\llcorner_{\mathcal{T}}, \ \mu_y \in \mathcal{P}(H(y))
\end{equation}
and the disintegration is strongly consistent since the quotient map $f : \mathcal T \to \mathcal T$ is $\mu$-measurable and $(\mathcal T,\mathcal B(\mathcal T))$ is countably generated.
Observe that $H$ induces an equivalence relation also on $\mathcal{T} \times X \cap \Gamma$ where the equivalence classes are $H(y)\cap \mathcal{T} \times X$ and
the quotient map is the $f$ of Proposition \ref{P:sicogrF}. Hence
\begin{equation}
\label{E:disiparzpi}
\pi\llcorner_{\mathcal{T} \times X \cap \Gamma} =
\int_S \pi_y m_{\pi}(dy), \quad m_{\pi} = f_\sharp \pi\llcorner_{\mathcal{T} \times X \cap \Gamma}, \ \pi_y \in \mathcal{P}(H(y)\cap \mathcal{T} \times X).
\end{equation}
Observe that $m = m_{\pi}$.
\section{Regularity of the disintegration}
\label{S:pt iniziali e finali}
In this Section we consider the translation of Borel sets by the optimal geodesic flow, we introduce the fundamental regularity assumption (Assumption \ref{A:NDEatom}) on the measure $\mu$ and we show that an immediate consequence is that the set of initial points is negligible and consequently we obtain a disintegration of $\mu$
on the whole space. A second consequence is that the disintegration of $\mu$ w.r.t. the $H$ has continuous conditions probabilities.
\subsection{Evolution of Borel sets}
\label{Ss:evolution}
Let $A \subset \mathcal{T}_e$ be an analytic set and define for $t \in \mathbb{R}$ the \emph{$t$-evolution $A_t$ of $A$} by:
\begin{equation}
\label{E:At}
A_t :=
\begin{cases}
P_{2}\big\{ (x,y)\in G \cap A \times X : d_{N}(x,y)=t \big\} & t\geq 0 \\
P_{2}\big\{ (x,y)\in G^{-1} \cap A \times X : d_{N}(x,y)=t \big\} & t<0.
\end{cases}
\end{equation}
It is clear from the definition that if $A$ is analytic, also $A_t $ is analytic .
We can show that $t \mapsto \mu(A_t)$ is measurable.
\begin{lemma}
\label{L:measumuAt}
Let $A$ be analytic. The function $t \mapsto \mu(A_t)$ is $\mathcal{A}$-measurable for $t \in \mathbb{R}$.
\end{lemma}
\begin{proof}
We divide the proof in three steps.
{\it Step 1.}
Define the subset of $X\times \mathbb{R}$
\[
\hat A:= \big\{ (x,t) : x \in A_{t} \big\}.
\]
Note that
\begin{eqnarray*}
\hat A & = &~ P_{13} \bigg\{ (x,y,t) \in X \times X \times \mathbb{R}^{+} : (x,y) \in G \cap A \times X, d_{N}(x,y)=t \bigg\} \crcr
&&~ \qquad \cup P_{13} \bigg\{ (x,y,t) \in X \times X \times \mathbb{R}^{-} : (x,y) \in G^{-1} \cap A \times X, d_{N}(x,y)= -t \bigg\},
\end{eqnarray*}
hence it is analytic. Clearly $A_{t} = \hat A (t)$.
{\it Step 2.} Define the closed set in $\mathcal{P}(X \times [0,1])$
$$
\Pi(\mu) : = \big\{ \pi \in \mathcal{P}(X\times [0,1]) : (P_{1})_{\sharp}(\pi)= \mu \big\}
$$
and let $B \subset X\times \mathbb{R} \times [0,1]$ be a Borel set such that $P_{12}(B)= \hat A$.
Consider the function
\[
\mathbb{R} \times \Pi(\mu) \ni (t, \pi) \mapsto \pi(B(t)).
\]
A slight modification of Lemma 4.12 in \cite{biacar:cmono} shows that this function is Borel.
{\it Step 3.}
Since supremum of Borel function are $\mathcal{A}$-measurable, pag. 134 of \cite{Sri:courseborel},
the proof is concluded once we show that
\[
\mu(A_{t}) = \mu (\hat A(t)) = \sup_{ \pi \in \Pi(\mu) } \pi(B(t)).
\]
Since $\hat A(t) \times [0,1] \supset B(t)$
\[
\mu (\hat A(t)) = \pi(\hat A(t) \times [0,1]) \geq \pi (B(t)).
\]
On the other hand from Theorem \ref{T:vanneuma}, there exists an $\mathcal{A}$-measurable section of the analytic set $B(t)$,
so we have $u: \hat A(t) \to B(t)$.
Clearly for $\pi_{u} = (\mathbb{I},u)_{\sharp}(\mu)$ it holds $\pi_{u}(B(t))= \mu (\hat A(t))$.
\end{proof}
The next assumption is the fundamental assumption of the paper.
\begin{assumption}[Non-degeneracy assumption]
\label{A:NDEatom} The measure $\mu$ satisfies Assumption \ref{A:NDEatom}
if for each analytic set $A\subset \mathcal{T}_{e}$ there exists a sequence $\{ t_{n} \}_{n\in \mathbb{N}} \subset \mathbb{R}$ and a strictly positive constant $C$ such that $t_{n} \to 0$ as $n\to +\infty$ and $\mu(A_{t_{n}}) \geq C \mu (A)$ for every $n \in \mathbb{N}$.
\end{assumption}
Note that Assumption \ref{A:NDEatom} was already introduced at page \pageref{I:NDEatom}.
Clearly it is enough to verify Assumption \ref{A:NDEatom} for $A$ compact set.
An immediate consequence of the Assumption \ref{A:NDEatom} is that the measure $\mu$ is concentrated on $\mathcal{T}$.
\begin{proposition}
\label{P:puntini}
If $\mu$ satisfies Assumption \ref{A:NDEatom} then
\[
\mu(\mathcal{T}_e\setminus \mathcal{T}) = 0.
\]
\end{proposition}
\begin{proof}
Let $A= \mathcal{T}_{e}\setminus \mathcal{T}$. Suppose by contradiction $\mu(A)>0$.
By the inner regularity there exists $\hat A \subset A$ closed with $\mu (\hat A)>0$. By Assumption \ref{A:NDEatom} there exist $C > 0$ and
$\{ t_{n} \}_{n\in \mathbb{N}}$ converging to 0 such that $\mu(\hat A_{t_{n}}) \geq C \mu(\hat A)$.
Define $ \hat A^{\varepsilon}:= \big\{ x : d_N(\hat A,x) < \varepsilon \big\}$.
Since $\hat A \subset A$, for all $n \in \mathbb{N}$ it holds $\hat A_{t_{n}} \cap A = \emptyset$.
Moreover for $t_{n}\leq\varepsilon$ we have $\hat A^{\varepsilon} \supset \hat A_{t_{n}}$. So we have
\[
\mu(\hat A) = \lim_{\varepsilon \to 0} \mu (\hat A^{\varepsilon}) \geq \mu(\hat A) + \mu(\hat A_{t_{n}}) \geq (1+ C) \mu (\hat A),
\]
that gives the contradiction.
\end{proof}
Once we know that $\mu(\mathcal{T}) = 1$, we can use the Disintegration Theorem \ref{T:disintr} to write
\begin{equation}
\label{E:disintT}
\mu = \int_S \mu_y m(dy), \quad m = f_\sharp \mu, \ \mu_y \in \mathcal{P}(H(y)).
\end{equation}
The disintegration is strongly consistent since the quotient map $f : \mathcal T \to \mathcal T$ is $\mu$-measurable and $(\mathcal T,\mathcal B(\mathcal T))$ is countably generated.
The second consequence of Assumption \ref{A:NDEatom} is that $\mu_y$ is continuous, i.e. $\mu_y(\{x\}) = 0$ for all $x \in X$.
\begin{proposition}
\label{P:nonatoms}
If $\mu$ satisfies Assumption \ref{A:NDEatom} then the conditional probabilities $\mu_y$ are continuous for $m$-a.e. $y \in S$.
\end{proposition}
\begin{proof}
From the regularity of the disintegration and the fact that $m(S) = 1$, we can assume that the map $y \mapsto \mu_y$ is weakly continuous on a compact set $K \subset S$ of comeasure $<\varepsilon$.
It is enough to prove the proposition on $K$.
{\it Step 1.} From the continuity of $K \ni y \mapsto \mu_y \in \mathcal{P}(X)$ w.r.t. the weak topology, it follows that the map
\[
y \mapsto A(y) := \big\{ x \in H(y): \mu_y(\{x\}) > 0 \big\} = \cup_n \big\{ x \in H(y): \mu_y(\{x\}) \geq 2^{-n} \big\}
\]
is $\sigma$-closed: in fact, if $(y_m,x_m) \to (y,x)$ and $\mu_{y_m}(\{x_m\}) \geq 2^{-n}$, then $\mu_y(\{x\}) \geq 2^{-n}$ by u.s.c. on compact sets.
Hence $A$ is Borel, where $A = \{ (y,A(y)) : y\in K \}$.
{\it Step 2.} The claim is equivalent to $\mu(P_{2}(A))=0$. Suppose by contradiction $\mu(P_{2}(A))>0$.
By Lusin Theorem (Theorem 5.8.11 of \cite{Sri:courseborel})
$A$ is the countable union of Borel graphs, $A = \cup_{n} A_{n}$.
Therefore we can take a Borel selection of $A$ just considering one of the Borel graphs, say $\hat A$.
Since at least one of $P_{2}(A_{n})$ must have positive $\mu$-measure, we can assume $\mu (P_{2}(\hat A))>0$.
By Assumption \ref{A:NDEatom}
$\mu( (P_{2}(\hat A))_{t_{n}} ) \geq C \mu(P_{2}(\hat A))$ for some $C>0$ and $t_{n}\to 0$.
Since $\hat A$ is a Borel graph, for every $y \in P_{1}(\hat A)$ the set $P_{2}( \{y\} \times X \cap \hat A)$ is a singleton.
Hence $(P_{2}(\hat A))_{t_{n}} \cap (P_{2}(\hat A))_{t_{m}} = \emptyset$. We have a contradiction with the fact that the measure is finite.
\end{proof}
\section{Solution to the Monge problem}
\label{S:Solution}
Throughout the section we assume $\mu$ to satisfy Assumption \ref{A:NDEatom}.
It follows from Disintegration Theorem \ref{T:disintr}, Proposition \ref{P:puntini} and Proposition \ref{P:nonatoms} that
\[
\mu = \int \mu_y m(dy), \ \pi = \int \pi_y m(dy), \quad \mu_y \ \textrm{continuous}, \ (P_1)_\sharp \pi_y = \mu_y,
\]
where $m = f_\sharp \mu$ and $\mu_y \in \mathcal{P}(H(y))$.
We write moreover
\[
\nu = \int \nu_y m(dy) = \int (P_2)_\sharp \pi_y m(dy).
\]
Note that $\pi_y \in \Pi(\mu_y,\nu_y)$ is $d_N$-cyclically monotone and (since $d_{N}\llcorner_{H(y) \times H(y)}$ is finite and,
from point \ref{A:comegeo} of Assumption \ref{A:assu1}, lower semi-continuous) optimal for $m$-a.e. $y$.
If $\nu(\mathcal{T}) = 1$, then the above formula is the disintegration of $\nu$ w.r.t. $H$.
\begin{theorem}\label{T:finale}
Assume that for all $y \in S$ there exists an optimal map $T_{y}$ from $\mu_{y}$ to $\nu_{y}$.
Then there exists a $\mu$-measurable map $T: X \to X$ such that
\[
\int d_{N}(x,T(x)) \mu(dx) = \int d_{N}(x,z) \pi(dxdz), \qquad T_{\sharp} \mu = \nu.
\]
\end{theorem}
Recall $S \subset \mathcal{T}$ introduced in (\ref{E:TNngel}).
\begin{proof}
The idea is to use Theorem \ref{T:vanneuma}.
{\it Step 1.} Let $\mathbf{T} \subset S \times \mathcal{P}(X^2)$ be the set: for $y \in S$, $\mathbf{T}_y$ is the family of optimal transference plans in $\Pi(\tilde \mu_y,\tilde \nu_y)$ concentrated on a graph,
\[
\mathbf{T} = \Big\{ (y, \pi) \in S \times \mathcal{P}(X^2): \pi \in \Pi(\mu_y, \nu_y) \ \textrm{optimal}, \exists T : X \to X, \pi(\textrm{graph}(T)) = 1 \Big\}.
\]
where for optimal in $\Pi(\mu_{y},\nu_{y})$ we mean
\[
\int d_{N} \pi = \min_{\pi \in \Pi(\mu_{y},\nu_{y})}\int d_{N} \pi.
\]
Note that, since $\pi$ is a Borel measure, in the definition of $\mathbf{T}$, $T$ can be taken Borel.
Moreover the $y$ section $\mathbf{T}_y = \mathbf{T} \cap \{y\} \times \mathcal{P}(X^2)$ is not empty.
{\it Step 2.} Since the projection is a continuous map, then the set
\[
\tilde \Pi = \Big\{ (y, \pi): (P_1)_\sharp \pi = \mu_y, (P_2)_\sharp \pi = \nu_y \Big\}
\]
is a Borel subsets of $S \times \mathcal{P}(X^2)$: in fact it is the counter-image of the Borel set $\textrm{graph}( (\mu_y, \nu_y)) \subset
S \times \mathcal{P}(X)^2$ w.r.t. the weakly continuous map $(y,\pi) \mapsto (y,(P_1)_\sharp \pi,(P_2)_\sharp \pi)$.
Define the Borel function
\[
S \times \mathcal{P}(X^2) \ni (y, \pi) \mapsto f(y, \pi) :=
\begin{cases}
\int d_{N} \pi & \pi \in \Pi( \mu_y, \nu_y) \crcr
+ \infty & \textrm{otherwise}
\end{cases}
\]
It follows that $y \mapsto g(y) := \inf_{\pi} f(y, \pi)$ is an $\mathcal{A}$-function: we can redefine it on a $m$-negligible set to make it Borel,
where $m$ is the quotient measure of $\mu$. Hence the set
\[
\tilde \Pi^{\textrm{opt}} = \bigg\{ (y, \pi): \pi \in \tilde \Pi( \mu_y, \nu_y), \int d_{N} \pi \leq g(y) \bigg\} = \tilde \Pi \cap \Big\{ (y, \pi): \int d_{N} \pi \leq g(y) \Big\}
\]
is Borel.
{\it Step 3.} Now we show that the set of $\pi \in \mathcal{P}(X^{2})$ concentrated on a graph is analytic. By Borel Isomorphism Theorem,
see \cite{Sri:courseborel} page 99, it is enough to prove the same statement for $\pi \in \mathcal{P}([0,1])^{2}$.
Consider the function
\[
\mathcal{P}([0,1]^{2}) \times C_b([0,1],[0,1]) \ni ( \pi,\phi) \mapsto h( \pi,\phi) := \pi(\textrm{graph}(\phi)) \in [0,1].
\]
Since $\textrm{graph}(\phi)$ is compact, $h$ is u.s.c.. Hence the set $B^n = h^{-1}([1-2^{-n},1])$ is closed, so that
\[
\mathscr{T} = \bigcap_n P_1(B^n) = \Big\{ \pi: \forall \varepsilon > 0\ \exists \phi_\varepsilon, \pi(\phi_\varepsilon) > 1 - \varepsilon \Big\}
\]
is an analytic set. It is easy to prove that $\pi \in \mathscr{T}$ iff $\pi$ is concentrated on a graph.
{\it Step 4.} It follows that
\[
\mathbf{T} = S \times \mathscr{T} \cap \tilde \Pi^{\textrm{opt}}
\]
is analytic and by Theorem \ref{T:vanneuma} there exists a $m$-measurable selection $y \mapsto \pi_y \in \mathbf{T}_y$. It is fairly easy to prove that $\int \pi_y m(dy)$ is concentrated on a graph, has the same transference cost of $\pi$ and belongs to $\Pi(\mu,\nu)$.
\end{proof}
It follows from Theorem \ref{T:finale}
that it is enough to solve for each $y \in S$ the Monge minimization problem with marginal $\mu_{y}$ and $\nu_{y}$ on the set $H(y)$.
In order to solve it, we introduce an assumption on the geometry of the set $H(y)$.
\begin{assumption}
\label{A:clessidra3}
For a given $y \in S$, $H(y)$ satisfies Assumption \ref{A:clessidra3} if
there exist two families of disjoint $\mathcal{A}$-measurable sets $\{ K_{t}\}_{t \in [0,1]}$ and $\{Q_{s} \}_{s\in [0,1]}$ such that
\begin{itemize}
\item $\mu_{y}( H(y) \setminus \cup_{t \in [0,1]} K_{t} ) = \nu_{y}( H(y) \setminus \cup_{s \in [0,1]} Q_{s} )=0$;
\item the associated quotient maps $\varphi_{K}$ and $\varphi_{Q}$ are respectively $\mu_{y}$-measurable and $\nu_{y}$-measurable;
\item for $t\leq s$, $K_{t} \times Q_{s}\subset G$.
\end{itemize}
\end{assumption}
Note that Assumption \ref{A:clessidra3} was already introduced at page \pageref{I:clessidra3}.
In the measurability condition of Assumption \ref{A:clessidra3}, the set $[0,1]$ is equipped with the Borel $\sigma$-algebra $\mathcal{B}([0,1])$.
If $H(y)$ satisfies Assumption \ref{A:clessidra3} we can disintegrate the marginal measures $\mu_{y}$ and $\nu_{y}$ respectively
w.r.t. the family $\{K_{t}\}$ and $\{ Q_{s} \}$:
\[
\mu_{y} = \int \mu_{y,t} m_{\mu_{y}}(dt), \quad \nu_{y} = \int \nu_{y,t} m_{\nu_{y}}(dt)
\]
where $m_{\mu_{y}} = \varphi_{K \, \sharp} \mu_{y}$, $m_{\nu_{y}} = \varphi_{Q \, \sharp} \nu_{y}$ and the disintegrations are strongly consistent.
\begin{proposition}\label{P:alternativa}
Suppose that $H(y)$ satisfies Assumption \ref{A:clessidra3} and that the following conditions hold true:
\begin{itemize}
\item $m_{\mu_{y}}$ is continuous;
\item $\mu_{y,t}$ is continuous for $m_{\mu_{y}}$-a.e. $t \in [0,1]$;
\item $m_{\mu_{y}} ([0,t]) \geq m_{\nu_{y}} ([0,t])$ for $m_{\mu_{y}}$-a.e. $t \in [0,1]$.
\end{itemize}
Then there exists a $d_{N}$-cyclically monotone $\mu_{y}$-measurable map $T_{y}$ such that $T_{y\, \sharp} \mu_{y} = \nu_{y}$ and
\[
\int d_{N}(x,T_{y}(x)) \mu_{y}(dx) = \int d_{N}(x,z) \pi_{y}(dxdz).
\]
\end{proposition}
\begin{proof}
{\it Step 1.}
Since $m_{\mu_{y}}$ is continuous and $m_{\mu_{y}} ([0,t]) \geq m_{\nu_{y}} ([0,t])$,
there exists an increasing map $\psi : [0,1] \to [0,1]$ such that $\psi_{\sharp} m_{\mu_{y}} = m_{\nu_{y}}$.
Moreover, since for $m_{\mu_{y}}$-a.e. $t \in [0,1]$ $\mu_{y,t}$ is continuous,
there exists a Borel map $T_{t} : K_{t} \to Q_{\psi(t)}$ such that
$T_{t\,\sharp}\mu_{y,t} = \nu_{y,\psi(t)}$ for $m_{\mu_{y}}$-a.e. $t \in [0,1]$.
Since $\psi(t) \geq t$ the map $T_{t}$ is $d_{N}$-cyclically monotone, hence optimal between $\mu_{y,t}$ and $\nu_{y,t}$.
{\it Step 2.} Reasoning as in the proof of Theorem \ref{T:finale}, one can prove the existence of a $\mu_{y}$-measurable map $T_{y}: H(y) \to H(y)$
that is the gluing of all the maps $T_{t}$ constructed in {\it Step 1.}.
Hence there exists a $\mu_{y}$-measurable map $T_{y}: H(y) \to H(y)$ such that $T_{y\,\sharp} \mu_{y,t} = \nu_{y,\psi(t)}$.
It follows from Assumption \ref{A:clessidra3} that $\textrm{graph}( T_{y}) \subset G$, hence $T_{y}$ is $d_{N}$-cyclically monotone and
\[
T_{\sharp} \mu_{y} = \int T_{\sharp}\mu_{y,t} m_{\mu_{y}}(dt) = \int \nu_{y,\psi(t)} m_{\mu_{y}}(dt) =
\int \nu_{y,t} (\psi_{\sharp}m_{\mu_{y}})(dt) = \nu_{y}.
\]
\end{proof}
The next corollary follows straightforwardly and it sums up all the results.
\begin{corollary}
Let $\pi \in \Pi(\mu,\nu)$ be concentrated on a $d_{N}$-cyclically monotone set $\Gamma$ satisfying Assumption \ref{A:assu1}.
Assume that $\mu$ satisfies Assumption \ref{A:NDEatom} and for $m$-a.e. $y \in S$ the set $H(y)$ satisfies Assumption \ref{A:clessidra3}.
If for $m$-a.e. $y \in S$ the hypothesis of Proposition \ref{P:alternativa} are verified,
then there exists an Borel map $T : X \to X$ such that
\[
\int d_{N}(x,T(x)) \mu(dx) = \int d_{N}(x,z) \pi(dxdz), \qquad T_{\sharp} \mu = \nu.
\]
If $\pi$ is also optimal, then $T$ solves the Monge minimization problem.
\end{corollary}
Let us summarize the theoretical results obtained so far.
Let $\pi \in \Pi(\mu,\nu)$ be a $d_{N}$-cyclically monotone transference plan concentrated on a set $\Gamma$.
Consider the corresponding family of chain of transport rays and assume
that $\Gamma$ satisfies Assumption \ref{A:assu1}.
Then the partition induced by $H$ permits to obtain a strongly consistent disintegration formula of $\mu,\nu$ and $\pi$ holds.
If $\mu$ satisfies Assumption \ref{A:NDEatom} then the set of initial points is $\mu$-negligible and
the conditional probabilities $\mu_{y}$ are continuous.
Since the geometry of $H(y)$ can be wild, we need another assumption to build a $d_{N}$-monotone transference map between $\mu_{y}$
and $\nu_{y}$.
If $H(y)$ satisfies Assumption \ref{A:clessidra3} we can perform another disintegration and, under additional regularity of
the conditional probabilities of $\mu_{y}$ and of the quotient measure of $\mu_{y}$, we prove the existence of a
$d_{N}$-monotone transference map between $\mu_{y}$ and $\nu_{y}$.
Applying the same reasoning for $m$-a.e. $y$ we prove the existence of a transport map $T$ between $\mu$ and $\nu$ that has the
same transference cost of the given $d_{N}$-cyclically monotone plan $\pi$.
\subsection{Example}\label{ss:esempio}
We conclude this Section with the analysis of a particular case in which the set $H(y)$ satisfies Assumption \ref{A:clessidra3}.
The hypothesis of Proposition \ref{P:alternativa} and Assumption \ref{A:clessidra3}
were partially inspired by this example.
What follows will be useful in the next Section,
however, since it is not only related to what will be proved in Section \ref{S:application}, we have decided to present it here.
Fix the following notation: a continuous curve $\gamma:[0,1] \to X$ is \emph{increasing} if for $t,s \in [0,1]$
\[
t\leq s \Longrightarrow (\gamma(t),\gamma(s)) \in G
\]
\begin{definition}[Hourglass sets]\label{D:clessidra}
For $z \in X$ define the \emph{hourglass set}
\[
K(z):= \Big\{ (x,y) \in X \times X : (x,z), (z,y) \in G \Big\}.
\]
\end{definition}
Assume that there exists an increasing curve $\gamma$ such that
\[
H(y) \times X \cap \Gamma \subset \bigcup_{t \in [0,1]} K(\gamma(t)) \cap \Gamma.
\]
Note that this assumption is equivalent to request
that on each chain of transport rays the branching structures can appear only along an increasing curve $\gamma$.
\begin{figure}
\label{Fi:clessidra}
\psfrag{g}{$\gamma$}
\psfrag{x}{$x$}
\psfrag{y}{$y$}
\psfrag{Z}{$z$}
\centerline{\resizebox{11cm}{3cm}{\includegraphics{clessidra4.eps}}}
\caption{The hourglass set $K(z)$.}
\end{figure}
Then $H(y)$ satisfies Assumption \ref{A:clessidra3}.
Indeed first notice that $K(z)$ is analytic, then define the family of sets
\[
K_{t} : = G^{-1}(\gamma(t)) \setminus \bigcap_{s < t} G^{-1}(\gamma(s)), \qquad
Q_{t} : = G(\gamma(t)) \setminus \bigcap_{t < s} G(\gamma(s)).
\]
Since $\gamma$ is increasing, $K_{t}$ and $Q_{s}$ are $\mathcal{A}$-measurable and the quotient maps are $\mathcal{A}$-measurable:
let $[a,b] \subset [0,1]$
\[
\varphi_{K}^{-1}([a,b]) = \bigcup_{t \in [a,b]} K_{t} = G^{-1}(\gamma(b)) \setminus G^{-1}(\gamma(a)) \cup K_{a}\, \in \mathcal{A}
\]
and the same calculation holds true for $\varphi_{Q}$. From the increasing property of $\gamma$ it follows that $K_{t} \times Q_{s} \in G$ for every
$0 \leq t \leq s \leq 1$.
Again from the increasing property of $\gamma$ it follows that $m_{\mu_{y}}([0,t]) \geq m_{\nu_{y}}([0,t])$.
\bigskip
\section{An application}\label{S:application}
Throughout this section $|\cdot|$ will be the euclidean distance of $\mathbb{R}^{d}$.
Let $C \subset \mathbb{R}^{d}$ be an open convex set such that $M:=\partial C$ is a smooth compact sub-manifold of $\mathbb{R}^{d}$ of dimension $d-1$.
Let $X: = \mathbb{R}^{d}\setminus C$. Clearly $X$ endowed with the euclidean topology is a Polish space.
Consider the following geodesic distance: $d_{M}:X\times X \to [0,+\infty]$:
\begin{equation}\label{E:distostc}
d_{M}(x,y) : = \inf \{ L(\gamma): \gamma \in \textrm{Lip} ([0,1], X), \gamma(0)=x,\gamma(1)=y \},
\end{equation}
where $L$ is the standard euclidean arc-length: $L(\gamma)=\int |\dot \gamma|$.
Hence $M$ can be seen as an obstacle for geodesics connecting points in $X$.
Note that any minimizing sequence has uniformly bounded Lipschitz constant, therefore in the definition of $d_{M}$ we can substitute $\inf$ with $\min$.
Hence $d_{M}$ is a geodesic distance on $X$.
We will show that given $\mu,\nu \in \mathcal{P}(\mathbb{R}^{d})$ with $\mu\ll \mathcal{L}^{d}$,
the Monge minimization problem with geodesic cost $d_{M}$ admits a solution.
From now on we will assume that $\mu \ll \mathcal{L}^{d}$.
and all the sets and structures introduced during the paper will be referred to this Monge problem.
The strategy to solve the Monge minimization problem is the one used in Section \ref{S:Solution}:
build an optimal map on each equivalence class $H(y)$ and then use Theorem \ref{T:finale}.
To prove the existence these optimal maps we will show that the geometry of the chain of transport rays $H(y)$
is the one presented in Example \ref{ss:esempio} and that the hypothesis of Proposition \ref{P:alternativa} are satisfied.
\begin{lemma}\label{L:cont}
The distance $d_{M}$ is a continuous map.
\end{lemma}
\begin{proof}
{\it Step 1.} Let $\{x_{n}\}_{n\in \mathbb{N}},\{y_{n}\}_{n\in \mathbb{N}} \in X$ such that $|x_{n} - x| \to 0$ $|y_{n} - y| \to 0$.
Since the boundary of $X$ is a smooth manifold,
for every $n \in \mathbb{N}$ there exist curves $\gamma_{1,n},\gamma_{2,n} \in \textrm{Lip}([0,1],X)$
such that
\begin{itemize}
\item $\gamma_{1,n}(0)=x, \gamma_{1,n}(1)=x_{n}$;
\item $\gamma_{2,n}(0)=y, \gamma_{2,n}(1)=y_{n}$;
\item $L(\gamma_{i,n}) \to 0$ as $n \to +\infty$, for $i=1,2$.
\end{itemize}
Consider $\gamma_{n}\in \textrm{Lip}([0,1],X)$ such that $\gamma_{n}(0)=x_{n}, \gamma_{n}(1)=y_{n}$ and $L(\gamma_{n})\leq d_{M}(x_{n},y_{n}) + 2^{-n}$.
Gluing $\gamma_{1,n}$ and $\gamma_{2,n}$ to $\gamma_{n}$ it follows
\[
d_{M}(x,y) \leq d_{M}(x_{n},y_{n}) + 2^{-n} + L(\gamma_{1,n}) + L(\gamma_{2,n}).
\]
Hence $d_{M}$ is l.s.c..
{\it Step 2.} Taking a minimizing sequence of admissible curves for $d_{M}(x,y)$ and gluing them with $\gamma_{i,n}$ as in \emph{Step 1.},
it is fairly easy to prove that $d_{M}$ is u.s.c. and therefore continuous.
\end{proof}
As a corollary we have the existence of an optimal transference plan $\pi$. Hence from now on $\pi$ will be an optimal transference plan
and all the structures defined during the paper starting from a generic $d_{N}$-cyclically monotone plan, are referred to it.
Moreover there exists $\varphi \in \textrm{Lip}_{d_{M}}(X,\mathbb{R})$ such that $\Gamma =\Gamma' = G = \{(x,y)\in X\times X: \varphi(x)-\varphi(y)=d_{M}(x,y)\}$.
Note that $\Gamma$ is closed.
The next result shows that the sets $H(y)$ have the structure of Example \ref{ss:esempio}.
The convex assumption on the obstacle is fundamental:
each transport rays is composed by a straight line, a geodesic on $M$ where branching structures are allowed and again a straight line.
\begin{lemma}\label{L:geoH}
For all $y \in S$, $H(y)$ has the geometry of Example \ref{ss:esempio}:
there exists an increasing curve $\gamma_{y}: [0,1] \to X$ such that
\[
H(y) \times X \cap \Gamma \subset \bigcup_{t \in [0,1]} K(\gamma_{y}(t)) \cap \Gamma.
\]
\end{lemma}
\begin{proof}
Since due to convexity and smoothness of the obstacle, the geodesics of $d_{M}$ are smooth and composed by a first straight line, a geodesic of the manifold and a final straight line, a branching structure can appear only on the manifold $M$.
If $H(y) \neq R(y)$, consider the following sets:
\[
Z:= \bigcap_{z\in H(y)\cap M} G^{-1}(z) \cap M, \quad W:= \bigcap_{z \in H(y)\cap M} G(z) \cap M.
\]
By $d_{M}$-monotonicity, smoothness and convexity of $M$, for all $z\in H(y)\cap M$ the set $ G^{-1}(z) \cap M$ is
always contained in the same geodesic of $M$.
Using the compactness of $M$, $Z= \{ z\}$ and $W = \{w \}$ and $(z,w) \in G$.
Consider the unique increasing geodesic $\gamma_{y} \in \gamma_{[z,w]}$ such that $\gamma_{y} = G(z) \cap G^{-1}(w)$.
Hence
\[
H(y) \times X \cap \Gamma \subset \bigcup_{t \in [0,1]} K(\gamma_{y}(t)) \cap \Gamma.
\]
\end{proof}
\begin{remark}\label{O:sempre}
From Lemma \ref{L:cont} and Lemma \ref{L:geoH} it follows that, for any transference plan $\pi$,
the set of chain of transport rays $H$ satisfies Assumption \ref{A:assu1}.
Indeed consider $H(y)$ and the corresponding geodesic $\gamma_{y}$ from Lemma \ref{L:geoH}.
Then take any sequence $x_{n} \in H(y)$ such that $|x_{n} - x| \to 0$ as $n \to +\infty$.
Note that there exist $s_{n} \in [0,1]$ and $t_{n}\in \mathbb{R}$ such that
$x_{n} = \gamma_{y}(s_{n}) + t_{n} \nabla\gamma_{y}(s_{n})$. Possibly passing to subsequences, $s_{n}\to s$, $t_{n} \to t$ with
$x = \gamma_{y}(s) + t \nabla\gamma_{y}(s)$.
Since $(\gamma_{y}(s_{n}), x_{n}) \in G$ and $G$ is closed it follows that $(\gamma_{y}(s), x) \in G$.
From $(y,\gamma_{y}(s)) \in R$ follows $x \in H(y)$. Hence point \ref{A:comegeo} of Assumption \ref{A:assu1} holds true.
Point \ref{A:comelocpt} of Assumption \ref{A:assu1} follows directly from the continuity of $d_{N}$.
\end{remark}
In the following Lemma we prove that the problem can be reduced to the equivalence classes $H(y)$.
We use the following notation: the quotient map induced by $H$ will be denoted by $f^{y}$ and the corresponding quotient measure $f^{y}_{\sharp}\mu$ by $m_{H}$.
\begin{lemma}\label{L:appl}
The $\mu$-measure of the set of initial points is zero, hence
\[
\mu = \int \mu_{y} m_{H}(dy).
\]
Moreover $\mu_{y}$ is continuous for $m_{H}$-a.e. $y$.
\end{lemma}
\begin{proof}
{\it Step 1.} Since $\mu\ll \mathcal{L}^{d}$, it is enough to prove that the set of initial points is $\mathcal{L}^{d}$-negligible and
that the disintegration w.r.t. $H$ of $\mathcal{L}^{d}$ restricted to any compact set has continuous conditional probabilities.
Indeed if $\mathcal{L}^{d}\llcorner_{K} = \int \eta_{y} m_{\mathcal{L}^{d}}(dy)$ and $\mu = \rho \mathcal{L}^{d}$ then $m_{\mu}\ll m_{\mathcal{L}^{d}}$ and
\[
\mu \llcorner_{K} = \int \rho \eta_{y} m_{\mathcal{L}^{d}}(dy) = \int \rho \frac{d m_{\mathcal{L}^{d}}}{d m_{\mu}} \eta_{y} m_{\mu}(dy),
\]
where $m_{\mu}$ is the quotient measure of $\mu\llcorner_{K}$.
It follows that the continuity of $\eta_{y}$ implies the continuity of conditional probabilities of $\mu$.
Hence the claim is to prove that $\mathcal{L}^{d}$ satisfies Assumption \ref{A:NDEatom}.
{\it Step 2.}
Let $K \subset X$ be any compact set with $\mathcal{L}^{d}(K)>0$.
Possibly intersecting $K$ with $B_{r}(x)$ for some $x \in \mathbb{R}^{d} \setminus C$ and $r>0$,
we can assume w.l.o.g. that $K\subset B_{\varepsilon}(x)$ and $B_{2\varepsilon}(x) \cap M = \emptyset$.
Since $d_{M}\geq d$, $K_{t}\subset B_{2\varepsilon}(x)$ for all $t\leq\varepsilon$.
Since $d_{M}= |\cdot|$ in $B_{2\varepsilon}(x)$,
it follows that inside $B_{2\varepsilon}(x) \times B_{2\varepsilon}(x)$ $d_{M}$-cyclically monotonicity is
equivalent to $|\cdot|$-cyclically monotonicity. It follows that the set $H \cap G(K) \times G^{-1}(K_{\varepsilon})$ is $|\cdot|$-cyclically monotone.
{\it Step 3.} The following is proved in \cite{biacava:streconv}: consider a metric measure space $(X,d,m)$ with $d$ non-branching geodesic distance,
$m \in \mathcal{P}(X)$ and assume that $(X,d,m)$ satisfies $MCP(K,N)$, for the definition of $MCP(K,N)$ we refer to \cite{Sturm:MGH2}. Let $\Gamma$ be a $d$-cyclically monotone set and consider the evolution of sets induced by $\Gamma$, then $m$ satisfies Assumption \ref{A:NDEatom} w.r.t. this evolution of sets.
Since $B_{2\varepsilon}(x)$ is a convex set, it follows that $(B_{2\varepsilon}(x),|\cdot|, \mathcal{L}^{d})$ satisfies $MCP(0,d)$. Therefore
$\mathcal{L}^{d}$ satisfies Assumption \ref{A:NDEatom} w.r.t. the evolution of sets induced by $H \cap G(K) \times G^{-1}(K_{\varepsilon})$.
The claim follows.
\end{proof}
Hence we can assume w.l.o.g. that $\mu(G^{-1}(M)) = \nu (G (M)) = 1$:
if $H(y)$ do not intersect the obstacle, it is a straight line and the marginal $\mu_{y}$
is continuous. Since the existence of an optimal transport map on a straight line with first marginal continuous is a standard fact in optimal transportation, the reduction follows.
Recall the two family of sets introduced in Example \ref{ss:esempio}:
\[
K_{y,t} : = G^{-1}(\gamma_{y}(t)) \setminus \bigcap_{s < t} G^{-1}(\gamma_{y}(s)), \qquad
Q_{y,t} : = G(\gamma_{y}(t)) \setminus \bigcap_{t < s} G(\gamma_{y}(s)).
\]
It follows from Lemma \ref{L:geoH} and Example \ref{ss:esempio} that
\[
\mu_{y} = \int \mu_{y,t} m_{\mu_{y}}(dt), \quad \nu_{y} = \int \nu_{y,t} m_{\nu_{y}}(dt).
\]
with $\mu_{y,t}(K_{y,t}) = \nu_{y,t}(Q_{y,t}) =1$.
Moreover using the increasing curve $\gamma_{y}$, we can assume that $m_{\mu_{y}} \in \mathcal{P}(M)$, indeed
\begin{equation}\label{E:primaM}
\mu_{y} = \int_{[0,1]} \mu_{y,t} m_{\mu_{y}}(dt) = \int_{\gamma_{y}([0,1])} \mu_{y, \gamma^{-1}_{y}(z)} (\gamma_{y \, \sharp} m_{\mu_{y}})(dz).
\end{equation}
And the same calculation holds true for $\nu_{y}$ and $m_{\nu_{y} } $.
Therefore in the following
\begin{equation}\label{E:primaM2}
\mu_{y} = \int_{M} \mu_{y,z} m_{\mu_{y}}(dz), \quad \nu_{y} = \int_{M} \nu_{y,z} m_{\nu_{y}}(dz)
\end{equation}
with $\mu_{y,z}( K_{y, \gamma_{y}^{-1}(z)} ) = \nu_{y,z}( Q_{y, \gamma_{y}^{-1}(z)} ) = 1$ and $m_{\mu_{y}}(\gamma_{y}([0,1])) = m_{\nu_{y}}(\gamma_{y}([0,1]))=1$.
Moreover w.l.o.g. we can assume that $S = f^{y}(\mathbb{R}^{d}) \subset M$, in particular we can assume that for all $y \in S$ there exists $t(y) \in [0,1]$
such that $y = \gamma_{y}(t(y))$.
According to Proposition \ref{P:alternativa},
to obtain the existence of an optimal map on $H(y)$ it is enough to prove that
$m_{\mu_{y}}$ is continuous and $\mu_{y,z}$ is continuous for $m_{\mu_{y}}$-a.e. $z \in M$.
Recall that $m_{\mu_{y}}( \gamma_{y}([0,t])) \geq m_{\nu_{y}} (\gamma_{y}([0,t]))$ is a straightforward consequence of
the increasing property of $\gamma_{y}$.
\begin{remark}\label{R:tuttoM}
Consider the following $\mathcal{A}$-measurable map:
\[
G^{-1}(M) \setminus (a(M) \cap M) \ni w \mapsto f^{M}(w): = \textrm{Argmin} \{ d(z,w) : z \in M \cap G(w) \} \in M.
\]
Consider the measure $m : = f^{M}_{\sharp} \mu \in \mathcal{P}(M)$.
Observing that $f^{M}(H(y)) = \gamma_{y}([0,1])$, it follows that
the support of $m$ is partitioned by a $d_{M}$-cyclically monotone equivalence relation:
\[
m \Big( \bigcup_{y \in S} \gamma_{y}([0,1]) \Big)=1, \qquad \bigcup_{y \in S} \gamma_{y}([0,1]) \times \bigcup_{y \in S} \gamma_{y}([0,1]) \cap G \ \ \textrm{is $d_{M}$-cyclically monotone}
\]
Moreover $f^{y}$ is a quotient map also for this equivalence relation.
Note that $f^{y}_{\sharp} m = m_{H}$: consider $I \subset S$
\begin{eqnarray*}
(f^{y}_{\sharp} m )(I) = &~ m \Big(\bigcup_{y \in I} \gamma_{y}([0,1]) \Big) = \mu \Big( G^{-1}(\bigcup_{y \in I} \gamma_{y} ([0,1]) ) \Big) \crcr
=&~ \mu \Big( \bigcup_{y \in I} H(y) \Big) = (f^{y}_{\sharp}\mu)(I) = m_{H}(I).
\end{eqnarray*}
It follows that
\[
m = \int_{S} ( f^{M}_{\sharp} \mu_{y} ) m_{H}(dy)
\]
and from (\ref{E:primaM2}) $f^{M}_{\sharp} \mu_{y} = m_{\mu_{y}}$.
Hence the final disintegration formula for $m$ is the following one:
\begin{equation}\label{E:dopoM}
m = \int_{S} m_{\mu_{y}} m_{H}(dy).
\end{equation}
\end{remark}
\begin{proposition}\label{P:fondamentale}
The measure $m$ is absolutely continuous w.r.t. the Hausdorff measure $\mathcal{H}^{d-1}$ restricted to M.
\end{proposition}
\begin{proof}
Recall that $\varphi \in \textrm{Lip}_{d_{M}}(\mathbb{R}^{d})$ is the potential associated to $\Gamma$ and consider the following set
\[
M_{2}: = P_{1} \Big( \{(x,y) \in M\times M : | \varphi(x)- \varphi(y) | = d_{M}(x,y) \} \setminus \{x=y \} \Big).
\]
{\it Step 1.} Define the following map: $M_{2} \ni w \mapsto \Xi(w) : = \textrm{Argmin}\{ \varphi(w) - \varphi(z) : z\in M_{2} \}$.
Then the function $\varphi$ is a potential for the Monge minimization problem on $M$ with cost the geodesic distance, that coincides with $d_{M}$,
with first marginal $m$ and as second marginal $\Xi_{\sharp} m$.
It follows from Proposition 15 of \cite{feldcann:mani}
that $\nabla \varphi$ is a Lipschitz function: for all $x,y \in M_{2}$
\[
| \nabla \varphi (x) - \nabla \varphi (y)| \leq L d_{M}(x,y).
\]
In \cite{feldcann:mani} the Lipschitz constant $L$ is uniform for $x,y$ belonging to sets uniformly far from the starting and ending points of the geodesics on $M$ of the transport set. Since in our setting the geodesics on $M$ do not intersect, $L$ is uniform on the whole $M$.
Moreover note that if $z = \gamma_{y}(t)$, then
\[
\nabla \varphi(z) = - \frac{\dot{\gamma}_{y}(t) }{ |\dot{\gamma}_{y}(t)|}.
\]
{\it Step 2.}
For $t \geq 0$, define the following map
\[
M_{2} \ni x \mapsto \psi_{t} (x) : = x + \nabla \varphi (x) t.
\]
Possibly restricting $\psi_{t}$ to a subset of $M$ of points coming from transport rays of uniformly positive length,
since $ t \mapsto \psi_{t}(x)$ is a parametrization of the transport ray touching $M$ in $x$,
by $d_{M}$-cyclical monotonicity of $\Gamma$, we can assume that $\psi_{t}$ is injective.
Moreover $\psi_{t}$ is bi-Lipschitz, provided $t$ is small enough: indeed
\[
|x + \nabla \varphi (x) t - y - \nabla \varphi (y) t| \geq |x- y | (1 - L t).
\]
It follows that
\[
M_{2} \times [-\delta,\delta] \ni (x,s) \mapsto \psi(x,s) : = x + \nabla \varphi (x) (t + s)
\]
is bi-Lipschitz and injective provided $\delta \leq 1/L + t$. Hence the Jacobian determinant of $\varphi$, $J d \varphi $, is uniformly positive.
{\it Step 3.}
Consider the following set
\[
B : = \{ x \in \mathbb{R}^{d} : t - \delta \leq d(M,x) \leq t + \delta \} \cap G^{-1}(M)
\]
where $d$ is the euclidean distance. Clearly $B$ is the range of $\psi$ and $\mathcal{L}^{d}(B)> 0$.
Since $M$ is a smooth manifold, we can pass to local charts: let $U_{\alpha} \subset \mathbb{R}^{d-1}$ be an open set and
$h_{\alpha} : U_{\alpha} \to M$ the corresponding parametrization map.
The map
\[
U_{\alpha}\times [-\delta,\delta]\ni (x,s) \mapsto \psi_{\alpha}(x,s) : = \psi (h_{\alpha}(x),s)
\]
is a bi-Lipschitz parametrization of the set $B_{\alpha} : = B \cap G^{-1}( h_{\alpha}(U_{\alpha}))$.
It follows directly from the Area Formula, see for example \cite{ambfuspal:bv}, that
\[
\mathcal{L}^{d}\llcorner_{B_{\alpha}} = \psi_{\alpha\, \sharp} \Big( J d \psi_{\alpha}( \mathcal{L}^{d-1}\times dt) \llcorner_{U_{\alpha} \times [-\delta,\delta]} \Big),
\]
hence $f^{M}_{\sharp} \mathcal{L}^{d}\llcorner_{B_{\alpha}} \ll \mathcal{H}^{d-1} \llcorner_{M}$.
Since $B$ can be covered with a finite number of $B_{\alpha}$ and $\mathcal{L}^{d}\llcorner_{B_{\alpha}}$ is equivalent to $m$, the claim
follows.
\end{proof}
Recall the following result.
Let $(M,g)$ be a $n$-dimensional compact Riemannian manifold, let $d_{M}$ be the geodesic distance induced by $g$
and $\eta$ the volume measure.
Then the disintegration of $\eta$ w.r.t. any $d_{M}$-cyclically monotone set is strongly consistent and the conditional probabilities are continuous.
This result is proved in \cite{biacava:streconv}, Theorem 9.5, in the more general setting of metric measure space satisfying the measure contraction property.
\begin{corollary}\label{C:boh}
For $m_{H}$-a.s. $y \in S$, the quotient measure $m_{\mu_{y}}$ is continuous.
\end{corollary}
\begin{proof}
We have proved in Remark \ref{R:tuttoM} that the measures $m_{\mu_{y}}$ are the conditional probabilities of the disintegration of $m$ w.r.t. the equivalence relation
given by the membership to geodesics $\gamma_{y}$ and $m_{H}$ is the corresponding quotient measure.
Hence the claim follows directly from Theorem 9.5 of \cite{biacava:streconv} and Proposition \ref{P:fondamentale}.
\end{proof}
\begin{proposition}\label{P:mappaostacolo1}
For $m_{H}$-a.e. $y \in S$, the measures
$\mu_{y,z}$ are continuous
for $m_{\mu_{y}}$-a.e. $z \in M$.
\end{proposition}
\begin{proof}
Recall that $f^{M}_{\sharp} \mu = m$.
{\it Step 1.}
The measure $\mu$ can be disintegrated w.r.t. the partition given by the family of pre-images of the $\mathcal{A}$-measurable map $f^{M}$:
$\{ (f^{M})^{-1}(p) \}_{p \in f^{M}(\mathbb{R}^{d})}$. Clearly $f^{M}$ is a possible quotient map, hence
\begin{equation}\label{E:conto}
\mu = \int \mu_{z} m(dz),
\end{equation}
The set $G^{-1}(M)\setminus a(M) \times G^{-1}(M)\setminus a(M) \cap G$ is $|\cdot |$-cyclically monotone and $\mu \ll \mathcal{L}^{d}$,
hence it follows that for $m$-a.e. $z \in f^{M}(\mathbb{R}^{d})$, $\mu_{z}$ is continuous.
{\it Step 2.}
From Lemma \ref{L:appl} $\mu = \int \mu_{y} m_{H}(dy)$, therefore
\[
m = f^{M}_{\sharp}\mu = \int (f^{M}_{\sharp}\mu_{y}) m_{H}(dy),
\]
hence using (\ref{E:conto}) and the uniqueness of the disintegration
\[
\mu = \int \bigg( \int \mu_{z} (f^{M}_{\sharp} \mu_{y})(dz) \bigg) m_{H}(dy), \qquad \mu_{y} = \int \mu_{z} (f^{M}_{\sharp} \mu_{y})(dz),
\]
where the last equality holds true for $m_{H}$-a.e. $y \in S$.
Hence for $m_{H}$-a.e. $y \in S$ the measures $\mu_{y,z}$ are continuous for $m_{\mu_{y}}$-a.e. $z \in M$.
\end{proof}
Finally we can prove the existence of an optimal map for the Monge minimization problem with obstacle.
\begin{theorem}
There exists a solution for the Monge minimization problem with cost $d_{M}$ and marginal $\mu, \nu$ with $\mu \ll \mathcal{L}^{d}$.
\end{theorem}
\begin{proof}
From Lemma \ref{L:appl} it follows that $\mu$ can be disintegrated w.r.t. the equivalence relation $H$.
From Theorem \ref{T:finale} it follows that to prove the claim it is enough to prove the existence of an optimal map on each equivalence class $H(y)$.
Hence we restrict the analysis to the classes $H(y)$ such that $H(y)\neq R(y)$ and for them we proved in
Lemma \ref{L:geoH} that Assumption \ref{A:clessidra3} holds true.
In Proposition \ref{P:fondamentale}, Corollary \ref{C:boh} and Proposition \ref{P:mappaostacolo1} we proved that
for $m_{H}$-a.e. $y \in S$ the measures $m_{\mu_{y}}$ and $\mu_{y,z}$ verify the hypothesis of Proposition \ref{P:alternativa}.
Therefore the claim follows.
\end{proof}
\medskip
\centerline{\sc Acknowledgements}
\vspace{2mm}
I would like to express my gratitude to Stefano Bianchini for his support during the preparation of this paper.
I wish to thank Luigi De Pascale for drawing my attention on the obstacle problem.
|
\section*{Introduction}
For the past 30 years, one of the most striking and influential
developments in combinatorial representation theory would be the
discovery of crystal bases for quantum groups and their
representations \cite{Kas90, Kas91}. Right after its discovery,
the crystal basis theory has attracted a lot of attention and
research activities because it has simple and explicit
combinatorial features and have many significant applications to a
wide variety of mathematical and physical theories. In particular,
crystal bases have an extremely nice behavior with respect to
tensor products, which leads to natural and exciting connections
with combinatorics of Young tableaux and Young walls
(\cite{Ka2003, KK08, KN, MM, N93}). Moreover, inspired by the
original works \cite{Kas90, Kas91, Kas93, Kas94}, many important
and deep results have been established for crystal bases for
quantum groups associated with symmetrizable Kac-Moody algebras
(see, for example, \cite{HK2002, JMMO, KMN1, KMN2, KS97, Li1, Li2,
Naka02, Sai}).
In \cite{Lus90, Lus91}, Lusztig provided a geometric approach to this subject.
On the other hand, not much has been known about crystal bases for
quantum groups corresponding to Lie superalgebras. A major difficulty
one encounters in the superalgebra case is the fact that the
category of finite-dimensional representations is in general not
semisimple. Fortunately, there is an interesting and natural
category of finite-dimensional $U_q(\mathfrak{g})$-modules which is
semisimple for the two super-analogues of the general linear Lie
algebra $\mathfrak{gl} (n)$: $\mathfrak{g} = \mathfrak{gl} (m | n)$
and $\mathfrak{g} = \mathfrak{q} (n)$. This is the category
$\Oint$ of so-called {\it tensor modules};
i.e., those that appear as submodules of tensor powers ${\bf
V}^{\otimes N}$ of the natural $U_q(\mathfrak{g})$-module ${\bf V}$.
The semisimplicity
of $\Oint$ is verified in
\cite{BKK} for the general linear Lie superalgebra $\mathfrak{g} =
\mathfrak{gl} (m|n)$ and in \cite{GJKK} for the queer Lie
superalgebra $\mathfrak{g} = \mathfrak{q}(n)$.
Furthermore, the
crystal basis theory of $\Oint$ for
$\mathfrak{g}= \mathfrak{gl} (m |n)$ was developed in \cite{BKK},
while the foundations of the highest weight representation theory of
$U_q (\qn)$ have been established in \cite{GJKK}.
In this paper, we develop the crystal basis theory for $\Uq$-modules
in the category $\Oint$. The (quantum) queer
superalgebra is interesting not only as the remaining case for which
$\Oint$ is semisimple, but also due to its
remarkable combinatorial properties. An example of such properties is
the queer analogue of the celebrated {\it Schur-Weyl duality}, often
referred to as {\it Schur-Weyl-Sergeev duality}, which was obtained
in \cite{Ser} for $U (\qn)$ and in \cite{Ol} for $U_q (\qn)$.
Being very interesting on the one hand, the representation theory of
(quantum) queer superalgebra faces numerous challenges on the other.
The queer Lie superalgebra is the only classical Lie superalgebra
whose Cartan subsuperalgebra has a nontrivial odd part. As a
result, the highest weight space of any finite-dimensional
$\qn$-module has a structure of a Clifford module and the
corresponding $\mathfrak{gl} (n)$-module appears with multiplicity
higher than one (in fact, a power of $2$). Also, as observed in
\cite{GJKK}, due to the different classification of Clifford modules
over $\C$ and $\C (q)$, the classical limit of an irreducible
highest weight $U_q (\qn)$-module is an irreducible highest weight
$U(\qn)$-module or a direct sum of two irreducible highest weight
$U(\qn)$-modules. On top of these and in contrast to the case of
$\mathfrak{gl} (m|n)$, the odd root generators $e_{\ol{i}}$ and
$f_{\ol{i}}$ of $U_q (\qn)$ are not nilpotent.
We overcome the challenges described above in several steps.
First, we set the ground field to be the field $\C ((q))$ of formal
Laurent power series. By enlarging the base field, we obtain an
equivalence of the two categories of Clifford modules, and in
particular, establish a standard version of the classical limit
theorem. As the next step, we introduce the {\it odd Kashiwara
operators} $\tilde{e}_{\ol{1}}$, $\tilde{f}_{\ol{1}}$, and
$\tilde{k}_{\ol{1}}$, where $\tilde{k}_{\ol{1}}$ corresponds to an
odd element in the Cartan subsuperalgebra of $\qn$. The definitions
of $\tilde{e}_{\ol{1}}$, $\tilde{f}_{\ol{1}}$ are new in the sense
that they are based solely on the comultiplication formulas for
$e_{\ol 1}$, $f_{\ol{1}}$ and lead to nilpotent operators on $L/qL$,
where $L$ is a crystal lattice. Furthermore, from these definitions,
we deduce a special {\it tensor product rule} for odd Kashiwara
operators.
Our definition of a {\it crystal basis} for a $U_q(\qn)$-module $M$
in the category $\Oint$ is also new: such a
basis is a triple $(L,B,(l_b)_{b \in B})$, where the crystal lattice
$L$ is a free $\C[[q]]$-submodule of $M$, $B$ is a finite
$\mathfrak{gl} (n)$-crystal, $(l_b)_{b \in B}$ is a family of
nonzero vector spaces such that $L / qL = \soplus_{b \in B} l_b$,
with a set of compatibility conditions for the action of the
Kashiwara operators imposed in addition. The definition of crystal
bases leads naturally to the notion of {\it abstract
$\qn$-crystals}, an example of which is the
$\mathfrak{gl}(n)$-crystal $B$ in any crystal basis $(L,B,(l_b)_{b
\in B})$. The modified notion of crystals allows us to consider the
multiple occurrence of $\mathfrak{gl} (n)$-crystals corresponding to
a highest weight $U_q(\qn)$-module $M$ in $\Oint$ as a single $\qn$-crystal.
As a result of this new setting, the existence and uniqueness
theorem for crystal bases is proved for any highest weight (not
necessarily irreducible) module $M$ in the category $\Oint$.
Moreover, the $\qn$-crystal $B$ of $M$ depends
only on the highest weight $\lambda$ of $M$ and hence we may write
$B = B(\lambda)$. In addition to the existence and uniqueness
theorem, the decompositions of the module ${\bf V} \otimes M$ and
the crystal ${\bf B} \otimes B(\lambda)$ are established, where
${\bf B}$ is the crystal of ${\bf V}$. These decompositions are
parametrized by the set of all $\lambda + \varepsilon_j$ such that
$\lambda + \varepsilon_j$ is a strict partition $(j=1, \ldots, n)$.
One of key ingredients of the proof of our main theorem is the
characterization of highest weight vectors in ${\bf B} \otimes
B(\lambda)$ in terms of even Kashiwara operators and the highest
weight vector of $B(\lambda)$. All these statements are verified
simultaneously by a series of interlocking inductive arguments.
This paper is organized as follows. In Section 1, we recall some of
the basic properties of $U_q(\qn)$-modules in the category
$\Oint$. Section 2 is devoted to the
definitions, examples, and some preparatory statements related to
crystal bases. In particular, we prove the tensor product rule. In
Section 3, we give algebraic and combinatorial characterizations of
highest weight vectors in $\B^{\otimes N}$. In Section 4, we prove
our main result: the existence and uniqueness theorem for crystal
bases.
\section{The quantum queer superalgebra}\label{sec:qn}
Let $\F=\C((q))$ be the field of formal
Laurent series in an indeterminate $q$
and let $\A=\C[[q]]$ be the subring of $\F$
consisting of formal power series in $q$. For $k \in \Z_{\ge 0}$, we
define
$$[k]= \frac{q^k - q^{-k}}{q - q^{-1}}, \quad [0]!=1, \quad [k]! =
[k] [k-1] \cdots [2][1].$$
For an integer $n \geq 2$,
let $P^{\vee} = \Z k_1 \oplus \cdots \oplus \Z k_n$ be a free
abelian group of rank $n$ and let ${\mathfrak h} = \C \otimes_{\Z}
P^{\vee}$ be its complexification. Define the linear functionals
$\epsilon_i \in \mathfrak{h}^*$ by $\epsilon_i(k_j) = \delta_{ij}$
$(i,j=1, \ldots, n)$ and set $P= \Z \epsilon_1 \oplus \cdots \oplus
\Z \epsilon_n$. We denote by $\alpha_i = \epsilon_i -
\epsilon_{i+1}$ the {\em simple roots} and by $h_i=k_i-k_{i+1}$
the {\em simple coroots}.
\Def The {\em quantum queer superalgebra $U_q(\mathfrak{q}(n))$} is
the superalgebra over $\F$ with $1$ generated by the symbols $e_i$,
$f_i$, $e_{\ol i}$, $f_{\ol i}$ $(i=1, \ldots, n-1)$, $q^{h}$ $(h\in
P^\vee)$, $k_{\ol j}$ $(j=1, \ldots, n)$ with the following defining
relations.
\begin{align}
\allowdisplaybreaks
\nonumber & q^{0}=1, \ \ q^{h_1} q^{h_2} = q^{h_1 + h_2} \ \ (h_1,
h_2 \in P^{\vee}), \\
\nonumber & q^h e_i q^{-h} = q^{\alpha_i(h)} e_i \ \ (h\in P^{\vee}), \displaybreak[1]\\
\nonumber & q^h f_i q^{-h} = q^{-\alpha_i(h)} f_i \ \ (h\in P^{\vee}), \displaybreak[1]\\
\nonumber & q^h k_{\ol j} = k_{\ol j} q^h, \displaybreak[1]\\
\nonumber & e_i f_j - f_j e_i = \delta_{ij} \dfrac{q^{k_i - k_{i+1}} - q^{-k_i
+ k_{i+1}}}{q-q^{-1}}, \displaybreak[1]\\
\nonumber & e_i e_j - e_j e_i = f_i f_j - f_j f_i = 0 \quad \text{if} \ |i-j|>1, \displaybreak[1]\\
\nonumber & e_i^2 e_j -(q+q^{-1}) e_i e_j e_i + e_j e_i^2= 0 \quad \text{if} \ |i-j|=1,\displaybreak[1]\\
\nonumber & f_i^2 f_j - (q+q^{-1}) f_i f_j f_i + f_j f_i^2 = 0 \quad \text{if} \ |i-j|=1,\displaybreak[1]\\
\nonumber & k_{\ol i}^2 = \dfrac{q^{2k_i} - q^{-2k_i}}{q^2 - q^{-2}}, \\
\nonumber & k_{\ol i} k_{\ol j} + k_{\ol j} k_{\ol i} =0 \ \ (i \neq j),
\\
\nonumber & k_{\ol i} e_i - q e_i k_{\ol i} = e_{\ol i} q^{-k_i}, \\
& k_{\ol i} f_i - q f_i k_{\ol i} = -f_{\ol i} q^{k_i}, \label{eq:defining relations}\\
\nonumber & e_i f_{\ol j} - f_{\ol j} e_i = \delta_{ij} (k_{\ol i}
q^{-k_{i+1}} - k_{\ol{i+1}} q^{-k_i}), \\
\nonumber & e_{\ol i} f_j - f_j e_{\ol i} = \delta_{ij} (k_{\ol i}
q^{k_{i+1}} - k_{\ol{i+1}} q^{k_i}), \\
\nonumber &e_i e_{\ol i} - e_{\ol i} e_i = f_i f_{\ol i} - f_{\ol i} f_i = 0, \\
\nonumber &e_i e_{i+1} - q e_{i+1}e_i =
e_{\overline{i}}e_{\overline{i+1}}+ q
e_{\overline{i+1}}e_{\overline{i}}, \\
\nonumber&q f_{i+1}f_i - f_i f_{i+1} =
f_{\overline{i}}f_{\overline{i+1}}+ q f_{\overline{i+1}}f_{\overline{i}}, \\
\nonumber & e_i^2 e_{\overline{j}} - (q+q^{-1})e_i e_{\overline{j}}
e_i + e_{\overline{j}} e_i^2= 0 \quad \text{if} \ |i-j|=1, \\
\nonumber & f_i^2 f_{\overline{j}} - (q+q^{-1})f_i f_{\overline{j}}
f_i + f_{\overline{j}} f_i^2=0 \quad \text{if} \ |i-j|=1.
\end{align}
\edf
The generators $e_i$, $f_i$ $(i=1, \ldots, n-1)$, $q^{h}$ ($h\in
P^\vee$) are regarded as {\em even} and $e_{\ol i}$, $f_{\ol i}$
$(i=1, \ldots, n-1)$, $k_{\ol j}$ $(j=1, \ldots, n)$ are {\em
odd}. From the defining relations, it is easy to see that the even
generators together with $k_{\ol 1}$ generate the whole algebra
$\Uq$.
\begin{remark}
The generators in \eqref{eq:defining relations}
are different from those in \cite[Theorem 2.1]{GJKK}.
The elements $e_i$, $f_i$, $e_{\ol i}$ and $f_{\ol i}$ in \eqref{eq:defining relations}
correspond to $q^{k_{i+1}} e_i $, $f_i q^{-k_{i+1}}$, $q^{k_{i+1}}e_{\ol i}$
and $f_{\ol i}q^{-k_{i+1}}$
in \cite[Theorem 2.1]{GJKK}, respectively.
We rewrite the whole defining relations in \cite[Theorem 2.1]{GJKK}
in terms of new generators and remove some relations which can be derived from the others.
\end{remark}
The superalgebra $U_q(\mathfrak{q}(n))$ is a bialgebra with the
comultiplication $\Delta\cl U_q(\mathfrak{q}(n)) \to
U_q(\mathfrak{q}(n)) \otimes U_q(\mathfrak{q}(n))$ defined by
\begin{equation}
\begin{aligned}
& \Delta(q^{h}) = q^{h} \otimes q^{h}\quad\text{for $h\in P^\vee$,} \\
& \Delta(e_i) = e_i \otimes q^{-k_i + k_{i+1}} + 1 \otimes e_i, \\
& \Delta(f_i) = f_i \otimes 1 + q^{k_i - k_{i+1}} \otimes f_i, \\
& \Delta(k_{\ol 1}) =k_{\ol 1}\otimes q^{k_1}+ q^{-k_1} \otimes k_{\ol 1}.
\end{aligned}
\end{equation}
Let $U^{+}$ (resp.\ $U^{-}$) be the subalgebra of
$U_q(\mathfrak{q}(n))$ generated by $e_i$, $e_{\ol i}$
$(i=1,\ldots, n-1)$ (resp.\ $f_i$, $f_{\ol i}$ ($i=1, \ldots, n-1$)), and
let $U^{0}$ be the subalgebra
generated by $q^{h}$ ($h\in P^\vee$) and $k_{\ol j}$ $(j=1, \ldots, n)$.
In \cite{GJKK}, it was
shown that the algebra $U_q(\mathfrak{q}(n))$ has the {\em
triangular decomposition}:
\begin{equation}
U^{-} \otimes U^{0} \otimes U^{+}\isoto
U_q(\mathfrak{q}(n)).
\end{equation}
Hereafter, a $U_q(\mathfrak{q}(n))$-module is understood as a
$U_q(\mathfrak{q}(n))$-supermodule.
A $U_q(\mathfrak{q}(n))$-module $M$ is called a {\em weight module}
if $M$ has a weight space decomposition $M=\soplus_{\mu \in P}
M_{\mu}$, where
$$M_{\mu}\seteq \set{ m \in M}{q^h m = q^{\mu(h)} m \ \ \text{for all} \ h
\in P^{\vee} }.$$ The set of weights of $M$ is defined to be
$$\wt(M) = \set{\mu \in P}{M_{\mu} \neq 0 }.$$
\Def A weight module $V$ is called a {\em highest weight module
with highest weight $\la \in P$} if $V_{\la}$ is finite-dimensional and
satisfies the following conditions:
\bna
\item $V$ is generated by $V_{\la}$,
\item $e_i v = e_{\ol i} v =0$ for all $v \in V_{\la}$,
$i=1, \ldots, n-1$,
\item $q^h v = q^{\la(h)} v$ for all $v \in V_{\la}$, $h \in
P^{\vee}$.
\ee \edf
As seen in \cite{GJKK}, there exists a unique irreducible highest weight module with highest
weight $\la \in P$ up to parity change, which will be denoted by
$V(\la)$.
Set
\begin{equation*}
\begin{aligned}
P^{\ge 0} = & \{ \la = \la_1 \epsilon_1 + \cdots + \la_n \epsilon_n
\in P\, ; \, \la_j \in \Z_{\ge 0} \ \ \text{for all} \ j=1, \ldots, n \}, \\
\La^{+} = & \{\la = \la_1 \epsilon_1 + \cdots + \la_n \epsilon_n \in
P^{\ge 0}\, ; \, \text{$\la_{i} \ge \la_{i+1}$ and $\la_{i}=\la_{i+1}$ implies}\\
& \hs{31ex}\text{$\la_{i} = \la_{i+1} = 0$ for all $i=1,
\ldots,n-1$}\}.
\end{aligned}
\end{equation*}
Note that each element $\la \in \La^{+}$ corresponds to a {\em
strict partition} $\la = (\la_1 > \la_2 > \cdots > \la_r >0)$.
Thus we will often call $\la \in \La^{+}$ a strict partition. With
the same reason, we call $\lambda=(\la_1,\la_2,\ldots,\la_n) \in
P^{\geq 0}$ a {\it partition} if $\la_1 \geq \la_2 \geq \cdots
\geq \la_r > \la_{r+1}=\cdots =\la_n=0$. We denote $r$ by
$\ell(\la)$.
\begin{example}
Let $$\V = \soplus_{j=1}^n \F v_{j} \oplus \soplus_{j=1}^n \F
v_{\ol j}$$ be the vector representation of $U_q(\mathfrak{q}(n))$.
The action of $\Uq$ on $\V$ is given as follows:
\begin{equation}
\ba{llll}
e_iv_j=\delta_{j,i+1}v_i, &e_iv_{\ol j}=\delta_{j,i+1}v_{\ol i},
&f_iv_j=\delta_{j,i}v_{i+1},&f_iv_{\ol j}=\delta_{j,i}v_{\ol{i+1}}, \\[1ex]
e_{\ol i}v_j=\delta_{j,i+1}v_{\ol{i}},&e_{\ol i}v_{\ol j}=\delta_{j,i+1}v_{i},&
f_{\ol i}v_j=\delta_{j,i}v_{\ol{i+1}},&f_{\ol i}v_{\ol j}=\delta_{j,i}v_{{i+1}}, \\[1ex]
q^h v_j=q^{\epsilon_j(h)} v_j, &q^h v_{\ol j}=q^{\epsilon_j(h)} v_{\ol j},
&k_{\ol i}v_j=\delta_{j,i}v_{\ol j},&k_{\ol i}v_{\ol j}=\delta_{j,i}v_{j}.
\ea
\end{equation}
Note that $\V$ is an irreducible highest weight module with highest weight $\epsilon_1$.
\end{example}
\Def We define $\Oint$ to be the category of
finite-dimensional weight modules $M$ satisfying the following conditions:
\bna
\item $\wt(M) \subset P^{\ge 0}$,
\item for any $\mu\in P^{\ge0}$ and $i \in \{1,\ldots, n\}$
such that $\lan k_i,\mu\ran=0$,
we have $k_{\ol i}\vert_{M_\mu}=0$.
\ee
\edf
\begin{remark}
By Lemma~\ref{le:invarianct under tildeki} below,
it is enough to assume $i=1$ in the
condition (b).
Note also that the condition (b) is equivalent to saying that
every weight space $M_\mu$ is completely reducible as a $U^0$-module.
\end{remark}
The fundamental properties of the category $\Oint$ are summarized in the following proposition.
\vskip 2ex
\Prop [\cite{GJKK}] \hfill
\bna
\item Every $U_q(\mathfrak{q}(n))$-module in
$\Oint$ is completely reducible.
\item Every irreducible object in $\Oint$ has the
form $V(\la)$ for some $\la \in \La^{+}$.
\item The category $\Oint$ is stable under
tensor products.
\ee
\enprop
In \cite{GJKK}, we employed the rational function field $\C(q)$
as the base field of $\uqqn$.
But here, we employ $\C((q))$ instead of $\C(q)$ as the base field of
$\uqqn$.
Note that when $m$ is a non-negative integer, the
$q$-integer $\dfrac{q^{2m}-q^{-2m}}{q^2-q^{-2}}$ has a square root
in $\C((q))$ but not in $\C(q)$.
This difference gives the following two statements,
which is simpler than the corresponding statements
in \cite{GJKK}.
\begin{prop}[{cf.\ \cite[Corollary 3.9]{GJKK})}]
Let ${\rm Cliff} _q(\la)$ be the associative superalgebra over
$\C((q))$ generated by odd generators
$\set{t_{\ol i}}{i=1,2,\ldots, n}$ with the
defining relations
$$
\begin{array}{cc}
t_{\ol i} t_{\ol j} + t_{\ol j} t_{\ol i} = \delta_{ij}
\dfrac{2(q^{2\la_i}-q^{-2\la_i})}{q^2-q^{-2}}, & i,j =
1,2,\ldots,n.
\end{array}
$$
Then ${\rm Cliff} _q(\la)$
has up to isomorphism
\bna
\item two simple modules $E^q(\la)$ and $\Pi(E^q(\la))$ of dimension $2^{k-1} | 2^{k-1}$ if $\ell(\la)=2k$,
\item one simple module $E^q(\la) \cong \Pi(E^q(\la))$ of dimension $2^k | 2^k$ if $\ell(\la)=2k+1$.
\end{enumerate}
\end{prop}
\begin{prop}[{cf.\ \cite[Theorem 5.14]{GJKK}}]
\label{prop:classical limit}
Let $V(\la)$ be an irreducible highest weight module with highest weight $\la \in \Lambda^+$.
Then we have
$$\ch V(\la) = \ch V_{\cls}(\la),$$
where $V_{\cls}(\la)$ is an irreducible highest weight module over $\q(n)$ with highest weight $\la$.
\end{prop}
In short, contrary to \cite{GJKK}, we have the same classification for
the modules over ${\rm Cliff} _q(\la)$ as that for the modules over
the Clifford algebra with the base field $\C$. Also we have the same characters
of the irreducible modules over $\uqqn$ as those of the
irreducible modules over $\q(n)$.
\begin{remark} \label{rem:Grothendieck rings}
Define $\mathcal O^{\geq 0}_{\inte,\cls}$ to be the category of finite-dimensional weight modules $M$ over $\q(n)$
such that i) $\wt(M) \subset P^{\geq 0}$,
ii) $k_{\ol i}|_{M_\mu}=0$ for $i \in \{1, \ldots, n \}$ and $\mu \in P^{\geq 0}$
satisfying $\langle k_i, \mu \rangle =0$.
Here $k_{\ol i}$ is the element of $\q(n)$ given by
$\left(
\begin{array}{cc} 0& E_{i,i} \\ E_{i,i} & 0
\end{array} \right) $, where $E_{i,i}$ is the $n \times n$-matrix having $1$ in the $(i,i)$-position and $0$ elsewhere.
Let us denote the Grothendieck rings of the categories by
$K(\mathcal O^{\geq 0}_{\inte})$ and $K(\mathcal O^{\geq 0}_{\inte,\cls})$, respectively.
Since $\mathcal O^{\geq 0}_{\inte,\cls}$
and $\Oint$ are semisimple categories,
by taking the classical limit (i.e., taking the reduction at $q=1$),
we have a ring isomorphism
$$K(\mathcal O^{\geq 0}_{\inte}) \isoto K(\mathcal O^{\geq 0}_{\inte,\cls})$$
which sends $V(\la) \mapsto V_{\cls}(\la)$.
\end{remark}
Now we give a decomposition of the tensor product of the natural
representation with a highest weight module.
\Th\label{th:decomposition} Let $M$
be a highest weight $\Uq$-module in $\Oint$ with
highest weight $\lambda \in \Lambda^+$. Then we have
$$\V \otimes M \simeq\soplus_{\stackrel{\la + \epsilon_j :}{ \text{strict partition}}}
M_ j,$$ where $M_j$ is a highest weight $\Uq$-module in the
category $\Oint$ with highest weight $\la +
\epsilon_j$ and $\dim (M_{j})_{\la + \epsilon_j} = 2 \dim
M_{\la}$.
\enth
\begin{proof}
We will prove that our assertion holds for finite-dimensional
highest weight modules over $\q(n)$. Then,
by Remark~\ref{rem:Grothendieck rings}, our assertion holds also for
finite-dimensional highest weight modules over $\uqqn$.
Let $\uqn$ be the universal enveloping algebra of $\q(n)$ and
let $U^{\geq 0}$ be the universal enveloping algebra of the standard
Borel subalgebra of $\q(n)$. Let $M$ be a highest weight
$\uqn$-module with highest weight $\la \in \Lambda^+$ and $\V_{\cls}
= \soplus_{i=1}^n (\C v_i \oplus \C v_{\ol i})$ be the natural
representation of $\uqn$. Consider a surjective homomorphism
$$\uqn \tensor_{U^{\geq 0}} \mathbf v_\la \twoheadrightarrow M,$$
where $\mathbf v_\la \simeq M_\la$ as a $U^{\geq 0}$-module.
Now we have
$$ \V_{\cls} \tensor \bigl(\uqn \tensor_{U^{\geq 0}} \mathbf v_\la\bigr) \simeq
\uqn \tensor_{U^{\geq 0}} (\V_{\cls} \tensor \mathbf v_\la).
$$
Then $
F_i(\V_{\cls} \tensor \mathbf v_\la) \seteq
\soplus_{j \leq i}(\C v_j \oplus \C v_{\ol j}) \tensor \mathbf v_\la$
is a $U^{\geq 0}$-module. We set
$$
N \seteq \uqn \tensor_{U^{\geq 0}} (\V_{\cls} \tensor \mathbf v_\la),
\quad
F_i(N) \seteq \uqn \tensor_{U^{\geq 0}} F_i(\V_{\cls} \otimes \mathbf v_\la).
$$
Since
$$
F_i(\V_{\cls} \tensor \mathbf v_\la) / F_{i-1}(\V_{\cls} \tensor
\mathbf v_\la) \simeq (\C v_i \oplus \C v_{\ol i}) \tensor \mathbf
v_\la,
$$
we see that
$$
F_i(N) / F_{i-1}(N) \simeq
\uqn \tensor_{U^{\geq 0}} \big(F_i(\V_{\cls} \tensor \mathbf v_\la) /
F_{i-1}(\V_{\cls} \tensor \mathbf v_\la)\big)
$$
is a highest weight module with highest weight $\la + \eps_i$.
\medskip
Now we shall show
\eq
N\simeq\soplus_{k \leq r} \big( F_k(N) / F_{k-1}(N)\big)\oplus
N / F_r(N),\quad \text{where $r=\ell(\la)$.}
\label{eq:fildec}
\eneq
First note that $F_i(N) / F_{i-1}(N)$ admits the central character
$$\chi_i \seteq \chi_{\la+\eps_i}\cl\mathcal Z \to \C,$$
where $\mathcal Z$ is the center of $\uqn$ and $\chi_\mu$ is the
central character afforded by the Weyl module $W(\mu)$ with
highest weight $\mu$ (see \cite[Section 1]{GJKK} for Weyl modules
and central characters). From \cite[Proposition 1.7]{GJKK}, we
know that $\chi_1,\ldots, \chi_r, \chi_{r+1}$ are different from
each other, and $\chi_{r+1} = \chi_{r+2} = \cdots = \chi_n$.
\noindent
Let us choose an element $a \in \mathcal Z$ such that $\chi_1(a)
= \cdots = \chi_r(a)=0$ and $\chi_{r+1}(a) \neq 0$. Then we have
$ a |_{F_i(N)/F_{i-1}(N)} =0$ and hence $a F_i(N) \subset
F_{i-1}(N)$ for $i \leq r$. It follows that $a^r F_r(N) =F_{-1}(N)
=0$. Hence $N \stackrel{a^r}{\rightarrow} N$ factors through $N
\rightarrow N/F_r(N) \stackrel{\psi}{\rightarrow} N$. Since $a^r\cl
N/F_r(N) \to N/F_r(N)$ is an isomorphism, we have the diagram
\begin{displaymath}
\xymatrix@R=2ex {
&N \ar[rd]\\
N/F_r(N) \ar[ru]^-{\psi} \ar[rr]_{a^r}^-{\sim} & & N/F_r(N)\,.}
\end{displaymath}
It follows that
$$N \simeq (N/F_r(N)) \oplus F_r(N).$$
Using a similar argument, we can conclude that
$$F_k(N) \simeq (F_k(N)/F_{k-1}(N)) \oplus F_{k-1}(N)$$
for $k \leq r$. Hence we obtain \eqref{eq:fildec}.
By \cite[Proposition 1.4 (3)]{GJKK}, we know that $F_i(N) /
F_{i-1}(N)$ admits a finite-dimensional quotient
if and only if $\la+\eps_i$ is a strict partition,
and $N/F_r(N)$ has only trivial finite-dimensional quotient.
Since $\V_{\cls} \tensor M$
is a largest finite-dimensional quotient of $N$, we get the desired result.
\end{proof}
\begin{corollary}\label{cor:Vtens}
Any irreducible $\uqqn$-module in
$\Oint$ appears as a direct summand of tensor products of\/ $\V$.
\end{corollary}
\Proof
It follows immediately
from Theorem~\ref{th:decomposition}.
\QED
\section{Crystal bases}
Let $M$ be a $U_q(\mathfrak{q}(n))$-module in the category
$\Oint$. For $i=1, 2, \ldots, n-1$, let $u \in
M_{\la}$ $(\la \in P)$ be a weight vector and consider the {\em
$i$-string decomposition} of $u$:
$$u =\sum_{k\ge 0} f_i^{(k)} u_k,$$
where $e_i u_k =0$ for all $k \ge 0$ and $f_i^{(k)} = f_i^{k} /
[k]!$. We define the {\em even Kashiwara operators} $\tei$, $\tfi$
$(i=1, \ldots, n-1)$ by
\begin{equation}
\begin{aligned}
& \tei u = \sum_{k \ge 1} f_i^{(k-1)} u_k, \qquad
\tfi u = \sum_{k \ge 0} f_i^{(k+1)} u_k.
\end{aligned}
\end{equation}
On the other hand, we define the {\em odd Kashiwara operators}
$\tilde{k}_{\ol {1}}$, $\tilde{e}_{\ol {1}}$, $\tilde{f}_{\ol
{1}}$ by
\begin{equation}
\begin{aligned}
\tkone & = q^{k_1-1}k_{\ol 1}, \\
\teone & = - (e_1 k_{\ol 1} - q k_{\ol 1} e_1) q^{k_1 -1}, \\
\tfone & = - (k_{\ol 1} f_1 - q f_1 k_{\ol 1}) q^{k_2-1}.
\end{aligned}
\end{equation}
The following lemma is obvious.
\begin{lemma}\label{com:evenodd}
The operators $\teone$ and $\tfone$ commute with
$\te_i$ and $\tf_i$ $(3\le i\le n-1)$.
\end{lemma}
Recall that an {\em abstract $\mathfrak{gl}(n)$-crystal} is a set
$B$ together with the maps $\tei, \tfi\cl B \to B \sqcup \{0\}$,
$\vphi_i, \eps_i \cl B \to \Z \sqcup \{-\infty\}$ $(i \in I =\{1,
\ldots, n-1\})$, and $\wt\cl B \to P$ satisfying the following
conditions (see \cite{Kas93}):
\begin{itemize}
\item[(i)] $\wt(\tei b) = \wt b + \alpha_i$ if $i\in I$ and
$\tei b \neq 0$,
\item[(ii)] $\wt(\tfi b) = \wt b - \alpha_i$ if $i\in I$ and
$\tfi b \neq 0$,
\item[(iii)] for any $i \in I$ and $b\in B$, $\vphi_i(b) = \eps_i(b) +
\langle h_i, \wt b \rangle$,
\item[(iv)] for any $i\in I$ and $b,b'\in B$,
$\tfi b = b'$ if and only if $b = \tei b'$,
\item[(v)] for any $i \in I$ and $b\in B$
such that $\tei b \neq 0$, we have $\eps_i(\tei b) = \eps_i(b) -
1$, $\vphi_i(\tei b) = \vphi_i(b) + 1$,
\item[(vi)] for any $i \in I$ and $b\in B$ such that $\tfi b \neq 0$,
we have $\eps_i(\tfi b) = \eps_i(b) + 1$, $\vphi_i(\tfi b) =
\vphi_i(b) - 1$,
\item[(vii)] for any $i \in I$ and $b\in B$ such that $\vphi_i(b) = -\infty$, we
have $\tei b = \tfi b = 0$.
\end{itemize}
In this paper, we say that an abstract $\mathfrak{gl}(n)$-crystal is
a {\em $\mathfrak{gl}(n)$-crystal} if it is realized as a crystal
basis of a finite-dimensional integrable
$U_q(\mathfrak{gl}(n))$-module. In particular, for any $b$ in a
$\mathfrak{gl}(n)$-crystal $B$, we have
$$\eps_i(b)=\max\{n\in\Z_{\ge0}\,;\,\tei^nb\not=0\}, \quad
\vphi_i(b)=\max\{n\in\Z_{\ge0}\,;\,\tfi^nb\not=0\}.$$
\Def \label{def:crystal base}
Let $M= \soplus_{\mu \in P^{\ge 0}} M_{\mu}$ be a
$U_q(\mathfrak{q}(n))$-module in the category
$\Oint$. A {\em crystal basis} of $M$ is a
triple $(L, B, l_{B}=(l_{b})_{b\in B})$, where
\bna
\item $L$ is a free $\A$-submodule of $M$ such that
\bni
\item $\F \otimes_{\A} L \isoto M$,
\item $L = \soplus_{\mu \in P^{\ge 0}} L_{\mu}$, where $L_{\mu} = L
\cap M_{\mu}$,
\item $L$ is stable under the Kashiwara operators $\tei$,
$\tfi$ $(i=1, \ldots, n-1)$, $\tkone$, $\teone$, $\tfone$.
\end{enumerate}
\item $B$ is a $\mathfrak{gl}(n)$-crystal together with
the maps $\teone, \tfone \cl B \to B \sqcup \{0\}$ such that
\bni
\item $\wt(\teone b) = \wt(b) + \alpha_1$, $\wt(\tfone b) = \wt(b) -
\alpha_1$,
\item for all $b, b' \in B$, $\tfone b = b'$ if and only if $b = \teone b'$.
\end{enumerate}
\item $l_{B}=(l_{b})_{b \in B}$ is a family of
non-zero $\C$-vector
subspaces of $L/qL$
such that
\bni
\item $l_{b} \subset (L/qL)_{\mu}$ for $b \in B_{\mu}$,
\item $L/qL = \soplus_{b \in B} l_{b}$,
\item $\tkone l_{b} \subset l_{b}$,
\item for $i=1, \ldots, n-1, \ol 1$, we have
\be[{\rm(1)}]
\item
if $\tei b=0$ then $\tei l_{b} =0$, and
otherwise $\tei$ induces an isomorphism
$l_{b}\isoto l_{\tei b}$,
\item
if $\tfi b=0$ then $\tfi l_{b} =0$,
and otherwise $\tfi$ induces an isomorphism $l_{b}\isoto l_{\tfi b}$.
\ee
\end{enumerate}
\end{enumerate}
\edf
\begin{prop}
Let $(L, B, l_{B})$ be a crystal basis of a $\uqqn$-module $M$.
Then we have
$$\teone^2 = \tfone^2 = 0$$ as endomorphisms on $L/qL$.
\begin{proof}
Since every $u \in L_{\la}$ has a $1$-string decomposition
$u=\sum_{k=0}^N f_1^{(k)}u_k$ with $e_1 u_k=0$ for $k=0, \ldots,
N$, it suffices to show that $\teone^2 u \equiv \tfone^2 u \equiv 0$ (mod $qL$) for
$u=f_1^{(s)}v$ with $e_1 v =0$ and $\wt(v) = \mu$ ($s \geq 0$).
We first show $\teone^2 u \equiv 0$ (mod $qL)$.
From the defining relations $k_{\ol 1} e_1 -q e_1 k_{\ol 1}=e_{\ol
1}q^{-k_1}$ and $e_1 e_{\ol 1}=e_{\ol 1}e_1$, we obtain
$$e_1 k_{\ol 1}e_1-qe_1^2k_{\ol 1}=e_1 e_{\ol 1}q^{-k_1} \quad \text{and} \quad k_{\ol
1}e_1^2-qe_1k_{\ol 1} e_1=q^{-1} e_{\ol 1} e_1 q^{-k_1}=q^{-1} e_1
e_{\ol 1} q^{-k_1}.$$ Then we have
$$ e_1 k_{\ol 1} e_1 -qe_1^2 k_{\ol 1} = q k_{\ol 1}e_1^2 - q^2
e_1 k_{\ol 1} e_1.$$ That is,
\eq
\ e_1 k_{\ol 1} e_1=e_1^{(2)} k_{\ol 1}
+k_{\ol 1 } e_1^{(2)}.\label{eq:k1serre}
\eneq
Using this formula, we obtain
\begin{equation*}
\begin{aligned}
\teone^{2} &=(e_1 k_{\ol 1} -q k_{\ol 1} e_1)^2 q^{2k_1-1} \\
&=( (e_1^{(2)}
k_{\ol 1} +k_{\ol 1 } e_1^{(2)}) k_{\ol 1} -q e_1 k_{\ol 1}^2 e_1
-q k_{\ol 1}
e_1^2 k_{\ol 1}+q^2 k_{\ol 1} (e_1^{(2)}
k_{\ol 1} +k_{\ol 1 } e_1^{(2)}))q^{2k_1-1} \\
&= \dfrac{q-q^{-1}}{q+q^{-1}} q^2 e_1^2 q^{4k_1}.
\end{aligned}
\end{equation*}
It follows that
\begin{equation*}
\begin{aligned}
\teone^2 u & = \dfrac{q-q^{-1}}{q+q^{-1}} q^{\langle 4k_1, \mu -s
\alpha_1 \rangle +2} e_1^2 f_1^{(s)}v \\
&=\dfrac{q-q^{-1}}{q+q^{-1}} q^{4\langle k_1, \mu \rangle -4s +2}
[\langle k_1-k_2, \mu \rangle -s+1 ] [\langle k_1-k_2, \mu \rangle
-s+2]f_1^{(s-2)}v.
\end{aligned}
\end{equation*}
Note that $q^{2\langle k_1-k_2, \mu \rangle -2s +1} [\langle
k_1-k_2, \mu \rangle -s+1 ] [\langle k_1-k_2, \mu \rangle -s+2]
\equiv 1 $ (mod $q \A)$. Since
\eqn
&&4 \langle k_1, \mu \rangle -4s +2 -(2 \langle k_1-k_2, \mu \rangle
-2s +1)\\
&&\hs{10ex} = 2 (\langle k_1-k_2, \mu \rangle -s)+4 \langle k_2, \mu \rangle
+1\ge1
\eneqn
we have
$$q^{4\langle k_1, \mu \rangle -4s +2}
[\langle k_1-k_2, \mu \rangle -s+1 ] [\langle k_1-k_2, \mu \rangle
-s+2] \in q \A,
$$
which implies $\teone^2 u \equiv 0$ (mod $qL$) as desired.
Now we show $\tfone^2 u \equiv 0$ (mod $qL$). By a similar argument
as above, we obtain
$$ f_1 k_{\ol 1} f_1 = f_1^{(2)}k_{\ol 1} +k_{\ol 1} f_1^{(2)}.$$
Then we have \begin{equation*}
\begin{aligned}
\tfone^{2}&=(k_{\ol 1}f_1 -q f_1 k_{\ol 1})^2 q^{2k_2 -1}\\
&=( k_{\ol{1}} (f_1^{(2)}k_{\ol 1} +k_{\ol 1} f_1^{(2)})
-q k_{\ol 1}f_1^2 k_{\ol 1} -q f_1 k_{\ol 1}^2 f_1 +q^2
(f_1^{(2)}k_{\ol 1} +k_{\ol 1} f_1^{(2)}) k_{\ol 1})q^{2k_2
-1}\\
&=\dfrac{q-q^{-1}}{q+q^{-1}}f_1^2q^{2k_1+2k_2-2} .
\end{aligned}
\end{equation*}
It follows that
$$
\begin{array}{ll}
\tfone^2 u&=
\dfrac{q-q^{-1}}{q+q^{-1}}f_1^2 q^{ \langle 2k_1+2k_2, \mu-s \alpha_1 \rangle -2 }f_1^{(s)}v \\
&= \dfrac{q-q^{-1}}{q+q^{-1}} q^{2 \langle k_1+ k_2 , \mu \rangle-2} [s+2] [s+1] f_1^{(s+2) } v.
\end{array}
$$
If $ \langle k_1-k_2, \mu \rangle < s+2$, then $f_1^{(s+2)}v=0$,
i.e., $\tfone^2 u \equiv 0 $ (mod $qL$). If $ \langle k_1-k_2, \mu
\rangle \geq s+2 $, we have
$$ 2\langle k_1+k_2, \mu
\rangle -2 \geq 2 \langle k_1-k_2, \mu \rangle -2 \geq 2s+2.$$
Since $q^{2s+1}[s+2][s+1] \equiv 1$ mod $q\A$, we have
$$q^{\langle 2k_1 +2k_2 , \mu \rangle -2} [s+2][s+1] \in
q \A,$$ which proves our assertion.
\end{proof}
\end{prop}
\vskip 3mm
\begin{example}
Let $\V = \soplus_{j=1}^n \F v_{j} \oplus \soplus_{j=1}^n \F
v_{\ol j}$ be the vector representation of $U_q(\mathfrak{q}(n))$.
Set $$\mathbf{L} = \soplus_{j=1}^n \A v_{j} \oplus
\soplus_{j=1}^n \A v_{\ol j}\quad \text{and }l_{j} = \C v_{j} \oplus \C
v_{\ol j} \subset \mathbf{L}/ q \mathbf{L},$$ and let $\B$ be the
$\mathfrak{gl}(n)$-crystal with the $\bar 1$-arrow given below.
$$\xymatrix@C=5ex
{*+{\young(1)} \ar@<0.1ex>[r]^-{1}
\ar@{-->}@<-0.9ex>[r]_{\ol 1} & *+{\young(2)} \ar[r]^2 & *+{\young(3)} \ar[r]^3 & \cdots \ar[r]^{n-1} & *+{\young(n)} }.$$
Here, the actions of $\tfi$ $(i=1, \ldots, n-1, \ol 1)$ are
expressed by $i$-arrows. Then $(\mathbf{L}, \B,
l_{\B}=(l_j)_{j=1}^n)$ is a crystal basis of $\V$.
\end{example}
\begin{remark} \label{rem:gln structure}
Let $M$ be a $\uqqn$-module in the category $\Oint$
with a crystal basis $(L,B,l_{B})$,
and let $B=\coprod_{k=1}^s B_k$ be the decomposition of $B$
into
connected $\mathfrak{gl}(n)$-crystals.
Then there exists a decomposition
$$M = \soplus_{k=1}^{s} \soplus_{j=1}^{m_k} M_{k,j} $$
of $M$ as a $U_q(\mathfrak{gl}(n))$-module, where \bna
\item $m_k=\dim l_b$ for some $b \in B_k$,
\item $M_{k,j}$ has a $U_q(\mathfrak{gl}(n))$-crystal basis $(L_{k,j}, B_{k,j})$ such that
\bni
\item $L=\soplus_{k,j}L_{k,j}$,
\item there exists a $\mathfrak{gl}(n)$-crystal isomorphism $\phi_{k,j} \cl B_k \isoto B_{k,j}$ so that
the vectors $\phi_{k,j}(b)$ $(j=1, \ldots, m_k)$ form a basis of
$l_b$ for each $b \in B_k$. \ee \ee
\end{remark}
\begin{remark}
Let $M$ be a $\uqqn$-module in the category $\Oint$
with a crystal basis $(L,B,l_{B})$.
For $i=1,\ldots,n-1,\ol1$ and $b$, $b'\in B$, if $b'=\tf_ib$ is satisfied,
then we have isomorphisms $\tf_i\cl l_b\isoto l_{b'}$ and
$\te_i\cl l_{b'} \isoto l_b$.
If $i=1,\ldots,n-1$, then they are inverses to each other
by Remark~\ref{rem:gln structure}.
However, when $i=\ol1$, they are not inverses to each other in general.
\end{remark}
The {\it tensor product rule} given in the following theorem is one
of the most important features of crystal basis theory.
\Th \label{th2:tensor product}
Let $M_j$ be a $\uqqn$-module in $\Oint$ with a crystal basis
$(L_j, B_j, l_{B_j})$ $(j=1,2)$. Set $B_1\otimes B_2 = B_1 \times B_2$ and
$l_{b_1\otimes b_2}=l_{b_1} \otimes l_{b_2}$ for $b_1\in B_1$ and $b_2\in B_2$.
Then $(L_1 \otimes_{\A} L_2, B_1 \otimes B_2,(l_b)_{b\in B_1 \otimes B_2})$
is a crystal basis of $M_1 \otimes_{\F}M_2$,
where the action of the Kashiwara operators on $B_1\otimes B_2$
are given as follows:
\eq
&&\begin{aligned}
\tei(b_1 \otimes b_2) & = \begin{cases} \tei b_1 \otimes b_2 \ &
\text{if} \ \vphi_i(b_1) \ge \eps_i(b_2), \\
b_1 \otimes \tei b_2 \ & \text{if} \ \vphi_i(b_1) < \eps_i(b_2),
\end{cases} \\
\tfi(b_1 \otimes b_2) & = \begin{cases} \tfi b_1 \otimes b_2 \
& \text{if} \ \vphi_i(b_1) > \eps_i(b_2), \\
b_1 \otimes \tfi b_2 \ & \text{if} \ \vphi_i(b_1) \le \eps_i(b_2),
\end{cases}
\end{aligned}\\[2ex]
\label{eq1:tensor product}
&&\begin{aligned}
\teone (b_1 \otimes b_2) & = \begin{cases} \teone b_1 \otimes b_2
& \text{if} \ \lan k_1, \wt b_2 \ran = \lan k_2, \wt b_2 \ran =0, \\
b_1 \otimes \teone b_2
& \text{otherwise,}
\end{cases} \\
\tfone(b_1 \otimes b_2) & = \begin{cases} \tfone b_1 \otimes b_2
& \text{if} \ \lan k_1, \wt b_2 \ran = \lan k_2, \wt b_2 \ran =0, \\
b_1 \otimes \tfone b_2
& \text{otherwise}.
\end{cases}\\
\end{aligned}
\label{eq2:tensor product}
\eneq
\enth
\begin{proof}
It is obvious that
\begin{equation*}
\begin{aligned}
& (L_1\tensor L_2)/q(L_1 \tensor L_2) =\soplus_{b_1 \in B_1, b_2
\in B_2} l_{b_1} \tensor l_{b_2},\\
& l_{b_1} \tensor l_{b_2} \subset ((L_1 \tensor L_2 )/ q(L_1
\tensor L_2) )_{\la+\mu} \ \ \text{for} \ b_1 \in (B_1)_{\la}, b_2
\in (B_2)_{\mu}.
\end{aligned}
\end{equation*}
For $i=1, 2, \ldots, n-1$, our assertions were already proved in
\cite{Kas90, Kas91}. Let us show the $i= {\ol 1}$ case. The following
comultiplication formulas can be checked easily:
\begin{equation*}
\left\{\begin{aligned}
& \Delta(\tkone) = \tkone \otimes q^{2 k_1} + 1 \otimes \tkone, \\
& \Delta(\teone) = \teone \otimes q^{k_1 + k_2} + 1 \otimes \teone
- (1-q^2) \tkone \otimes e_1 q^{2 k_1}, \\
& \Delta(\tfone) = \tfone \otimes q^{k_1 + k_2} + 1 \otimes \tfone
- (1-q^{2}) \tkone \otimes f_1 q^{k_1 + k_2-1}.
\end{aligned}
\right.
\end{equation*}
Clearly, $L_1 \tensor L_2$ and $l_{b_1} \tensor l_{b_2}$ are
stable under $\Delta (\tkone)$ for all $b_1 \in B_1, b_2 \in B_2$.
We will show that $L_1 \otimes L_2$ is stable under
$\Delta(\teone)$ and $\Delta(\tfone)$.
Let $u_1 \in L_1$ and $u_2 \in L_2$.
Then the comultiplication formula implies
$$\Delta(\teone)(u_1 \otimes u_2) =
\teone u_1 \otimes q^{k_1 + k_2} u_2 \pm u_1 \otimes
\teone u_2 - (1-q^2) \tkone u_1 \otimes e_1 q^{2 k_1} u_2,$$
where $\pm$ is according that $u_1$ is even or odd.
It is obvious that the first two terms belong to $L_1 \otimes L_2$.
For the last term, we may assume that $u_2 = f_1^{(s)} v$ with
$e_1 v=0$. Then we have
\begin{equation*}
\begin{aligned}
e_1 q^{2k_1} u_2 & = e_1 q^{2k_1} f_1^{(s)} v = q^{2 \langle k_1,
\wt(v) - s \alpha_1 \rangle} [\langle k_1 - k_2, \wt(v) \rangle -
s +1] f_1^{(s-1)} v \\
& = q^{2 \langle k_1, \wt(v) \rangle -2s } [\langle k_1 - k_2,
\wt(v) \rangle -s + 1] \te_1 u_2 \\
&=\frac{q^{\langle 3k_1 - k_2, \wt(v) \rangle -3s +2} - q^{\langle
k_1 + k_2, \wt(v) \rangle -s}}{q^2 -1} \te_1 u_2.
\end{aligned}
\end{equation*}
Since
\eqn
&&\langle 3k_1 - k_2, \wt(v) \rangle -3s +2 = 3 (\langle k_1-k_2,
\wt(v) \rangle -s)+2 \langle k_2, \wt(v) \rangle +2>0,\\
&&\langle k_1 + k_2 , \wt(v) \rangle -s =\lan k_1,\wt(u_2)\ran
+\lan k_2,\wt(v)\ran\ge\lan k_1,\wt(u_2)\ran\ge0,
\eneqn
If $\lan k_1,\wt(u_2)\ran=0$, then $f_1u_2=0$ and hence
$s=-\lan h_1,\wt(u_2)\ran= \lan k_2,\wt(u_2)\ran$.
Thus we conclude
\eq
&&\ba{l}
e_1 q^{2k_1} u_2\equiv \te_1u_2\pmod{L_2}\quad\text{if $\lan k_1,\wt(u_2)\ran=0$,}
\\[1ex]
e_1 q^{2k_1} u_2 \in qL_2\quad\text{if $\lan k_1,\wt(u_2)\ran>0$.}
\ea\label{eq:e1q}
\eneq
Hence $L_1 \otimes L_2$ is
stable under $\Delta(\teone)$.
Similarly, one can show that $f_1 q^{k_1 + k_2 -1} L_2 \subset
L_2$, which implies $L_1 \otimes L_2$ is stable under
$\Delta(\tfone)$. Thus we have shown that $L_1\otimes L_2$
is stable under the Kashiwara operators.
\medskip
We shall prove the tensor product rule. To prove the $\teone$-case,
let $u_1 \in l_{b_1}, u_2 \in l_{b_2}$,
and we consider the following three cases separately.
\vskip 3mm
\noindent
{\bf Case 1:} $\langle k_1, \wt(b_2) \rangle = \langle k_2,
\wt(b_2) \rangle =0$.
By the comultiplication formula, we have
$$\Delta(\teone)(u_1 \otimes u_2) = \teone u_1 \otimes u_2 \pm u_1
\otimes \teone u_2 - (1-q^2) \tkone u_1 \otimes e_1 u_2.$$
Since $\langle k_2, \wt(b_2) + \alpha_1 \rangle = \langle k_2, \wt(b_2)
+ \eps_1 - \eps_2 \rangle = -1 <0$, we must have $\teone u_2=e_1 u_2 =0$.
Hence $\Delta(\teone) (u_1 \otimes u_2) = \teone
u_1 \otimes u_2 $.
If $\teone =0$ on $l_{b_1}$, then $\teone \otimes 1 =0$ on $l_{b_1}
\otimes l_{b_2}$. If $\teone\cl l_{b_1} \rightarrow l_{\teone b_1}$ is
an isomorphism, then $\teone \otimes 1 \cl l_{b_1} \otimes l_{b_2}
\rightarrow l_{\teone b_1} \otimes l_{b_2}$ is also an isomorphism
as desired.
\medskip
\noindent
{\bf Case 2:} $\langle k_1, \wt(b_2) \rangle > 0$.
By the comultiplication formula and \eqref{eq:e1q}, we have
\begin{equation*}
\begin{aligned}
\Delta(\teone)(u_1 \otimes u_2) &
= \teone u_1 \otimes q^{\langle k_1 + k_2, \wt(b_2) \rangle} u_2
\pm u_1 \otimes\teone u_2 \\
& - (1-q^2) \tkone u_1 \otimes e_1q^{2k_1} u_2 \\
& \equiv \pm u_1 \otimes \teone u_2 \quad (\text{mod} \ q L_1 \otimes
L_2).\end{aligned}
\end{equation*}
\medskip
\noindent
{\bf Case 3:} $\langle k_1, \wt(b_2) \rangle =0, \ \ \langle k_2,
\wt(b_2) \rangle > 0$.
The comultiplication formula and \eqref{eq:e1q} yield
\begin{equation*}
\begin{aligned}
\Delta(\teone)(u_1 \otimes u_2) & = \teone u_1 \otimes q^{\langle
k_1 + k_2, \wt(b_2) \rangle} u_2 \pm u_1 \otimes \teone u_2 \\
& - (1-q^2) \tkone u_1 \otimes e_1q^{2k_1} u_2 \\
& \equiv \pm u_1 \otimes \teone u_2 -\tkone u_1 \otimes e_1 u_2 \quad
(\text{mod} \ q L_1 \otimes L_2).
\end{aligned}
\end{equation*}
Since $\langle k_1, \wt(b_2) \rangle =0$ and $\tkone^2=(1-q^4)^{-1}(1-q^{4k_1})$, we have
$$k_{\bar 1}u_2 =0, \ \
\tkone^2 e_1 u_2 = \dfrac{1 - q^{4 k_1}}{1-q^4} e_1 u_2 = e_1 u_2.$$
It follows that
$$\teone u_2 = -q^{-1} (e_1 k_{\bar 1} - q k_{\bar 1} e_1) q^{k_1}
u_2 = k_{\bar 1} e_1 q^{k_1}u_2 = k_{\bar 1} q^{k_1 -1} e_1 u_2 = \tkone
e_1 u_2.$$ Hence we obtain
$$\tkone \teone u_2 = \tkone^2 e_1 u_2 = e_1 u_2, $$
which implies
$$
\begin{aligned}
\Delta(\teone)(u_1 \otimes u_2) &\equiv\pm u_1 \otimes \teone u_2 - \tkone
u_1 \otimes \tkone \teone u_2 \\
&\equiv
(1 - \tkone \otimes \tkone) (1 \otimes \teone) (u_1 \otimes u_2).
\end{aligned}
$$
The operator $1 - \tkone \otimes \tkone$ on $l_{b_1}\otimes l_{\te_1b_2}$
is invertible because
$(\tkone \otimes \tkone)^2=-\tkone^2 \otimes \tkone^2=
-(1-q^4)^{-1}(1 - q^{4 k_1})\otimes \id$ acts on $l_{b_1\otimes \teone b_2}$ by
the multiplication of a scalar different from $1$.
Hence the map $\Delta(\teone) \cl l_{b_1} \otimes l_{b_2}
\rightarrow l_{b_1} \otimes l_{\teone b_2}$, which is either 0 or an
isomorphism according that $\teone b_2=0$ or not.
The assertions on $\tfone$ can be verified in a similar manner.
The remaining property (b) (ii) in Definition~\ref{def:crystal base}
follows immediately from the formula \eqref{eq2:tensor product}.
\end{proof}
\begin{comment}
$l_{b_1} \tensor l_{b_2} \subset ((L_1 \tensor L_2 )/ q(L_1
\tensor L_2) )_{\la+\mu}$, where $\wt (b_1)=\la, \wt (b_2)=\mu$
and $(L_1\tensor L_2)/q(L_1 \tensor L_2) =\soplus_{b_1 \in B_1,
b_2 \in B_2} l_{b_1} \tensor l_{b_2}$. By Lemma~\ref{le:e1q2k1}
below we know that $L_1 \tensor L_2$ is stable under $\Delta
(\teone)$ and $\Delta(\tfone)$. Now we shall show that for $b_1
\tensor b_2 \in B_1 \tensor B_2$, if $\teone (b_1 \tensor b_2)=0$
then $\teone l_{b_1 \tensor b_2} =0$, and otherwise $\teone$
induces an isomorphism $l_{b_1 \tensor b_2} \isoto l_{\teone (b_1
\tensor b_2)}$.
\begin{enumerate}[(a)]
\item $\langle k_1, \wt (b_2) \rangle= \langle k_2, \wt (b_2) \rangle = 0$
We have $\teone(b_1 \tensor b_2) = \teone b_1 \tensor b_2$ by the tensor
product rule given above, and
$\te_{\ol{1}} l_{b_2} = \te_1 l_{b_2}=0$, since $\langle k_2, \wt b_2 \rangle =0$.
By Lemma~\ref{le:e1q2k1}, we also have $e_1 q^{2k_1}$ acts as $\te_1$ on $l_{b_2}$.
Thus we get $\teone |_{l_{b_1} \tensor l_{b_2}}=\teone \tensor 1$,
which is $0$ if $\teone b_1=0$ (equivalently, $\teone (b_1 \tensor b_2)=0$), and
an isomorphism if $\teone b_1 \neq 0$ (equivalently, $\teone (b_1 \tensor b_2) \neq 0$).
\item $\langle k_1, \wt (b_2) \rangle > 0$
We have $\teone(b_1 \tensor b_2) = b_1 \tensor \teone b_2$ by the tensor product rule.
Since $\langle k_1, \wt (b_2) \rangle > 0$, $q^{k_1+k_2} $ acts trivially on $l_{b_2}$.
By Lemma~\ref{le:e1q2k1}, we also have $e_1 q^{2k_1}$ acts trivially on $l_{b_2}$.
Therefore $\teone |_{l_{b_1} \tensor l_{b_2}}=1 \tensor \teone$, and then we obtain the desired result.
\item $\langle k_1, \wt (b_2) \rangle = 0, \langle k_2, \wt (b_2) \rangle > 0$
We have $\teone (b_1 \tensor b_2) =b_1 \tensor \teone b_2$.
By Lemma~\ref{le:e1q2k1}, we also have
$$\teone |_{l_{b_1} \tensor l_{b_2}}=1 \tensor \teone-\tkone \tensor \te_1.$$
If $\tkone$ acts trivially on $l_{b_1}$, we are done.
Consider the case such that $\tkone$ does not act trivially on $l_{b_1}$.
Since $\tkone$ acts trivially on $l_{b_2}$, we obtain that
$$ \teone u = q^{-1} (e_1 k_{\ol 1} -qk_{\ol 1}e_1)q^{k_1} u=k_{\ol 1} e_1 u= k_{\ol 1}q^{k_1-1}e_1 q^{2k_1}u=\tkone \te_1 u$$
for $u \in l_{b_2}$. Since $\tkone^2 =1$ on $l_{\te_1 b_2}$, we have $\tkone \teone$ acts as $\te_1$ on $l_{b_2}$. It gives
$$\teone |_{l_{b_1} \tensor l_{b_2}}=1 \tensor \teone-\tkone \tensor \te_1= 1 \tensor \teone-\tkone \tensor \tkone \teone =(1-\tkone \tensor \tkone)(1 \tensor \teone).$$ Since $(\tkone \tensor \tkone) (\tkone \tensor \tkone) =- \tkone^2 \tensor \tkone^2=-1$, $1-\tkone \tensor \tkone$ induces an isomorphism. Therefore we get $\teone |_{l_{b_1} \tensor l_{b_2}}=0$ if $\teone b_2=0$ (equivalently, $\teone (b_1 \tensor b_2)=0$), and
an isomorphism if $\teone b_2 \neq 0$ (equivalently, $\teone (b_1 \tensor b_2) \neq 0$).
\end{enumerate}
\begin{lemma} \label{le:e1q2k1}
Let $(L, B, l_B)$ be a crystal basis of a module $M \in \Oint$.
Then $e_1 q^{2k_1} L \subset L $ and for $u \in l_b$, we have
\begin{equation*}
e_1 q^{2k_1} u \equiv
\begin{cases}
\te_1 u & \text{if} \ \langle k_1, \wt b \rangle =0, \\
0 & otherwise,
\end{cases}
\end{equation*}
mod $qL$. Similarly, we have $f_1q^{k_1+k_2-1} L \subset L$ and for $u \in l_b$,
\begin{equation*}
f_1 q^{k_1+k_2-1} u \equiv
\begin{cases}
\tf_1 u & \text{if} \ \langle k_1, \wt b \rangle =1, \\
0 & otherwise,
\end{cases}
\end{equation*}
mod $qL$.
\begin{proof}
Let $\wt u =\la$. We can assume that $u=f_1 ^{(s)} v$ for some $s \geq 0$ and some $v \in L_{\mu}$ such that $\la=\mu - s \alpha_1$ and $e_1 v=0$.
Then we have $\te_1 u = f_1^{(s-1)} v$ and
$$
e_1 q^{2k_1} u = q^{2(\langle k_1, \mu \rangle-s)}[1-s+\langle k_1-k_2, \mu \rangle] \te_1 u.
$$
Note that $q^{-s+\langle k_1-k_2, \mu \rangle} [1-s+\langle
k_1-k_2, \mu \rangle]=1$ mod $q \A$ and
$$2(\langle k_1, \mu \rangle-s) + s- \langle k_1-k_2, \mu \rangle =
\langle k_1, \mu \rangle+ \langle k_2, \mu \rangle-s \geq 0.$$
Since $\langle \mu, k_1 \rangle \geq \langle \mu, k_1-k_2 \rangle \geq s $, we know that
$$\langle k_1, \mu \rangle+ \langle k_2, \mu \rangle-s=0$$
if and only if
$\langle k_1, \la \rangle=0$. Thus we get the desired result.
The other assertion can be proved by the similar argument.
\end{proof}
\end{lemma}
\end{comment}
Motivated by the properties of crystal bases, we introduce the
notion of abstract crystals.
\Def An {\em abstract $\mathfrak{q}(n)$-crystal} is a
$\mathfrak{gl}(n)$-crystal together with the maps $\teone, \tfone\cl B
\to B \sqcup \{0\}$ satisfying the following conditions:
\bna
\item $\wt(B)\subset P^{\ge0}$,
\item $\wt(\teone b) = \wt(b) + \alpha_1$, $\wt(\tfone b) = \wt(b) -
\alpha_1$,
\item for all $b, b' \in B$, $\tfone b = b'$ if and only if $b = \teone b'$,
\item
if $3\le i\le n-1$, we have
\be[{\rm(i)}]
\item
the operators $\teone$ and $\tfone$ commute with
$\te_i$ and $\tf_i$ ,
\item if $\teone b\in B$, then
$\eps_i(\teone b)=\eps_i(b)$ and $\vphi_i(\teone b)=\vphi_i(b)$.
\ee
\end{enumerate}
\edf
Note that any crystal basis of $\uqqn$-modules in $\Oint$
satisfies the property (d) by Lemma~\ref{com:evenodd}.
Let $B_1$ and $B_2$ be abstract $\qn$-crystals. The {\em tensor
product} $B_1 \otimes B_2$ of $B_1$ and $B_2$ is defined to be the
$\mathfrak{gl}(n)$-crystal $B_1 \otimes B_2$ together with the maps
$\teone$, $\tfone$ defined by \eqref{eq2:tensor product}. Then it is
an abstract $\qn$-crystal.
The following associativity of the tensor product
is easily checked.
\begin{prop}
Let $B_1, B_2$ and $B_3$ be abstract $\qn$-crystals. Then we have
$$(B_1 \otimes B_2) \otimes B_3 \simeq B_1\otimes (B_2 \otimes B_3).$$
\end{prop}
\begin{comment}
\begin{proof}
We already know that the tensor product is associative under the
even Kashiwara operators. To show the associativity under the odd
Kashiwara operators $\tfone, \teone$, we check the following
cases.
\noi
{\bf Case 1:} $\langle k_1, \wt(b_3) \rangle =\langle k_2, \wt(b_3)
\rangle =0$.
If $\langle k_1, \wt(b_2) \rangle =\langle k_2,
\wt(b_2) \rangle =0$, we obtain
\begin{equation*}
\begin{aligned}
&\tfone((b_1 \tensor b_2) \tensor b_3)=\tfone(b_1 \tensor b_2)
\tensor b_3=(\tfone b_1 \tensor b_2) \tensor b_3, \\
&\tfone(b_1 \tensor( b_2 \tensor b_3))=\tfone b_1 \tensor (b_2
\tensor b_3).
\end{aligned}
\end{equation*}
If $\langle k_1, \wt(b_2) \rangle >0$ or $\langle k_2,
\wt(b_2) \rangle >0$, we have
\begin{equation*}
\begin{aligned}
&\tfone((b_1 \tensor b_2) \tensor b_3)=\tfone(b_1 \tensor b_2)
\tensor b_3=(b_1 \tensor \tfone b_2) \tensor b_3, \\
&\tfone(b_1 \tensor( b_2 \tensor b_3))=b_1 \tensor \tfone(b_2
\tensor b_3)=b_1 \tensor (\tfone b_2 \tensor b_3).
\end{aligned}
\end{equation*}
\noi
{\bf Case 2:} $\langle k_1, \wt(b_3) \rangle >0$ or $\langle k_2,
\wt(b_3) \rangle >0$.
In this case, we see that
\begin{equation*}
\begin{aligned}
&\tfone((b_1 \tensor b_2) \tensor b_3)=(b_1 \tensor b_2)
\tensor \tfone b_3, \\
&\tfone(b_1 \tensor( b_2 \tensor b_3))=b_1 \tensor \tfone(b_2
\tensor b_3)=b_1 \tensor ( b_2 \tensor \tfone b_3).
\end{aligned}
\end{equation*}
In all cases, we observe that the action $\tfone$ on $(b_1 \tensor
b_2) \tensor b_3$ is the same as that on $b_1 \tensor (b_2 \tensor
b_3 )$ as desired. The assertion on $\teone$ can be checked in the
same manner.
\end{proof}
\end{comment}
\begin{example} \hfill
\bna
\item If $(L, B, l_{B})$ is a crystal basis of a $\Uq$-module $M$ in the
category $\Oint$, then $B$ is an abstract
$\qn$-crystal.
\item The crystal graph $\B$ of the vector representation $\V$
is an abstract $\qn$-crystal.
\item By the tensor product rule, $\B^{\otimes N}$ is an abstract
$\qn$-crystal. When $n=3$, the $\qn$-crystal structure of $\B
\otimes \B$ is given below.
$$\xymatrix
{*+{\young(1) \otimes \young(1)} \ar[r]^1 \ar@{-->}[d]^{\ol 1} &
*+{\young(2) \otimes \young(1)} \ar@<-0.5ex>[d]_1 \ar@{-->}@<0.5ex>[d]^{\ol 1} \ar[r]^2&
*+{\young(3) \otimes \young(1)} \ar@<-0.5ex>[d]_1 \ar@{-->}@<0.5ex>[d]^{\ol 1} \\
*+{\young(1) \otimes \young(2)} \ar[d]^2 &
*+{\young(2) \otimes \young(2)} \ar[r]_2 &
*+{\young(3) \otimes \young(2)} \ar[d]^2 \\
*+{\young(1) \otimes \young(3)} \ar@{-->}@<-0.5ex>[r]_{\ol 1} \ar@<0.5ex>[r]^{1} &
*+{\young(2) \otimes \young(3)} &
*+{\young(3) \otimes \young(3)}
}$$
\item For a strict partition $\la = (\la_1 > \la_2 > \cdots > \la_r
>0)$,
let $Y_{\la}$ be the skew Young diagram having $\la_1$ many
boxes in the principal diagonal, $\la_2$ many boxes in the second
diagonal, etc.
For example, if $\la=(7 > 6 > 4 > 2 > 0)$, then we
have
$$Y_{\la} = \young(::::::\hfill,:::::\hfill\hfill,::::\hfill\hfill\hfill,:::\hfill\hfill\hfill\hfill,::\hfill\hfill\hfill\hfill,:\hfill\hfill\hfill,\hfill\hfill) \quad.$$
Let $\B(Y_{\la})$ be the set of all semistandard tableaux of shape
$Y_{\la}$ with entries from $1, 2, \ldots, n$. Then by an {\em
admissible reading} introduced in \cite{BKK}, $\B(Y_{\la})$ can be
embedded in $\B^{\otimes N}$, where $N=\la_1 + \cdots + \la_r$,
and it is stable under the Kashiwara operators $\tei,\tfi$ ($i=1,
\cdots, n-1,\ol1$). Hence it becomes an abstract
$\qn$-crystal. Moreover, the $\qn$-crystal structure thus obtained
does not depend on the choice of admissible reading.
Indeed, since $Y_\la$ is a skew Young diagram, it is stable under
the even Kashiwara operators, and the $\gl(n)$-crystal structure
does not depend on the choice of admissible reading. Let $T$ be a
semistandard tableau of shape $\la$ and let $\beta$ be the
lowest
box with entry $1$ in the principal diagonal of $T$. Since a
box with entry $1$ must lie in the principal diagonal of $T$, every
box with entry $1$ except $\beta$ lies in the northeast of
$\beta$. Let $\psi\cl \B(Y_\la) \rightarrow \B^{\tensor N}$ be an
admissible reading. It follows that $\beta$ is the rightmost box
with entry $1$ in $\psi(T)$. If there is a box, say $\gamma$, with
entry $2$ in the southwest of $\beta$ in $T$, then $\gamma$ must
appear after $\beta$ in $\psi(T)$. Thus we get $\tfone (\psi(T)) =0$. If
there is no box with entry $2$ in the southwest of $\beta$ in $T$,
then we know that every box with entry $2$ must lie in the
northeast of $\beta$ in $T$, and hence there is no box with entry
$2$ after $\beta$ in $\psi(T)$. Thus $\tfone$ acts on $\beta$.
Since the entry of the right box of $\beta$ in $T$ is greater than or
equal to $2$,
we have $\tfone (\psi(T))=\psi(T')$, where $T'$ is the semistandard tableau of shape $\la$ obtained from $T$
by replacing the entry of $\beta$ from $1$ to $2$.
It follows that $\B(Y_\la)$ is stable under the action of $\tfone$ and
it does not depend on the choice of admissible reading.
Let $\delta$ be the leftmost box with entry $2$ in $T$. If
$\delta$ lies in the second diagonal, the entry of the box lying
in the left of $\delta$ must be $1$. Then, for any admissible
reading $\psi$, $\teone \psi(T) = 0$. Thus we may assume that $\delta$ lies in
the principal diagonal of $T$,
and our assertion on $\teone$ follows from similar arguments as above.
In Figure~\ref{fi:B(Y_(3,1,0))}, we illustrate the crystal
$\B(Y_\la)$ for $n=3$ and $\la=(3>1>0)$. Note that it is
connected. However, in general, $\B(Y_\la)$ is not connected.
\ee
\end{example}
\begin{figure}[!h]
$$\scalebox{.8}{\xymatrix@R=1pc@H=1pc{ & & {\young(::1,:12,1)} \ar[dl]_1 \ar_2[d] \ar^{\ol 1}@{-->}[dr] & & & \\
& {\young(::1,:22,1)} \ar@<-0.5ex>_1[dl] \ar@<1ex>^{\ol 1}@{-->}[dl]
\ar_2[d] & {\young(::1,:13,1)} \ar_1[d]\ar^{\ol 1}@{-->}[dr] &
{\young(::1,:12,2)} \ar_2[d] & & \\
{\young(::1,:22,2)} \ar_2[d] &
{\young(::1,:23,1)}\ar@<-0.5ex>_1 [dl] \ar@<1ex>^{\ol 1}@{-->}[dl] \ar_2[d]
& {\young(::2,:13,1)} \ar_1 [d] \ar^{\ol 1}@{-->}[dr] & {\young(::1,:13,2)}
\ar_1 [d] \ar_2[dr] & & {\young(::1,:12,3)} \ar@<-0.5ex>_1 [d]
\ar@<1ex>^{\ol 1}@{-->}[d] \\
{\young(::1,:23,2)} \ar_2[d] & {\young(::1,:33,1)} \ar^{\ol 1}@{-->}[dl] \ar_1 [d] & {\young(::2,:23,1)} \ar@<-0.5ex>_1 [d]
\ar@<1ex>^{\ol 1}@{-->}[d] \ar_2[dl] & {\young(::2,:13,2)}
\ar_2[d] & {\young(::1,:13,3)} \ar_1 [dl] \ar^{\ol 1}@{-->}[dr] & {\young(::1,:22,3)} \ar_2[d] \\
{\young(::1,:33,2)} \ar_2[dr] & {\young(::2,:33,1)} \ar@<-0.5ex>@{-->}_{\ol 1}[dr] \ar@<1ex>^{1}[dr] &{\young(::2,:23,2)} \ar_2[d] & {\young(::2,:13,3)} \ar@<-0.5ex>_1 [d] \ar@<1ex>^{\ol 1}@{-->}[d] & & {\young(::1,:23,3)} \\
& {\young(::1,:33,3)} \ar@<-0.5ex>@{-->}_{\ol 1}[dr] \ar@<1ex>^{1}[dr] & {\young(::2,:33,2)} \ar_2[d] & {\young(::2,:23,3)} & & \\
& &{ \young(::2,:33,3)}&& &}}$$
\caption{${\mathbf B}(Y_\la)$ for $n=3$, $\la = (3>1>0)$} \label{fi:B(Y_(3,1,0))}
\end{figure}
Let $B$ be an abstract $\qn$-crystal. For $i = 1,2,\ldots, n-1$,
we define the automorphism $S_i$ on $B$ by
\eq
&&S_i b =
\begin{cases}
\tfi^{\langle h_i, \wt b \rangle} b & \text{if} \quad {\langle h_i, \wt b \rangle} \geq 0, \\
\tei^{-\langle h_i, \wt b \rangle} b & \text{if} \quad {\langle h_i, \wt b \rangle} \leq 0.
\end{cases}\label{def:Sicr}
\eneq
Let $w$ be an element of the Weyl group $W$ of $\gl(n)$. Then, as
shown in \cite{Kas94}, there exists a unique action $S_{w} \cl B \to
B$ of $W$ on $B$ such that $S_{s_i}=S_i$ for $i=1,2,\ldots,n-1$. Note that
$\wt(S_w b)=w(\wt(b))$ for any $w \in W$ and $b \in B$.
For $i=1, \ldots, n-1$, we set
\eq
&&w_i = s_2 \cdots s_{i} s_1 \cdots s_{i-1}.
\label{def:wi}
\eneq
Then $w_i$ is the shortest element in $W$ such that $w_i(\alpha_i) = \alpha_1$.
We define the {\em odd Kashiwara operators} $\teibar$, $\tfibar$ $(i=2, \ldots, n-1)$ by
$$\teibar = S_{w_i^{-1}} \teone
S_{w_i}, \ \ \tfibar = S_{w_i^{-1}} \tfone S_{w_i}.$$
We say that
$b \in B$ is a {\em highest weight vector} if $\tei b = \teibar b
=0$ for all $i=1, \ldots, n-1$.
\vs{2ex}
\begin{remark}\label{rem:teibar}
These actions can be lifted to actions on $\uqqn$-modules.
Let $M$ be a $\uqqn$-module in $\Oint$.
For each $i=1,\ldots, n-1$, we have
$$M=\soplus_{\substack{\ell\ge k\ge0,\\\la\in P,\; \lan h_i,\la\ran=\ell}}
f_i^{(k)}\bl(\Ker(e_i)_\la\br).
$$
Hence we can define the endomorphism $S_i$ of $M$ by
\eq
S_i(f_i^{(k)}u)=f_i^{(\ell-k)}u\quad\text{for $u\in\Ker(e_i)_\la$.}
\label{def:Si}
\eneq
Then $S_i^2=\id_M$ and we have
$S_i(M_\la)=M_{s_i\la}$.
If $(L,B,l_B)$ is a crystal basis of $M$, then
$L$ is stable under $S_i$, and
$S_i$ induces an action on $L$ and $L/qL$.
Obviously, we have $S_i(l_b)=l_{S_ib}$ for $b\in B$, where $S_ib$
is defined in \eqref{def:Sicr}.
We define the endomorphisms $\te_{\ol i}$ and $\tf_{\ol i}$ of $M$ by
\eq
\ba{rcl}
\te_{\ol i}&=&(S_2\cdots S_iS_1\cdots S_{i-1})^{-1}\circ
\teone\circ(S_2\cdots S_iS_1\cdots S_{i-1}),\\[1ex]
\tf_{\ol i}&=&(S_2\cdots S_iS_1\cdots S_{i-1})^{-1}\circ
\tfone\circ(S_2\cdots S_iS_1\cdots S_{i-1}).
\ea\label{def:efibar}
\eneq
Then we have
$$\text{$\te_{\ol i}M_\mu\subset M_{\mu+\alpha_i}$ and
$\tf_{\ol i}M_\mu\subset M_{\mu-\alpha_i}$ for every $\mu\in P^{\ge0}$.}$$
Let $(L,B,l_B)$ be a crystal basis of $M$.
Then $L$ is stable under the action of $\te_{\ol i}$, and
$\te_{\ol i}$ induces an action on $L/qL$,
and we have
\eqn
&&\left\{
\parbox{\mylength}{
(i) if $\te_{\ol i}b\not=0$, then $\te_{\ol i}$ induces an isomorphism
$l_{b}\isoto l_{\te_{\ol i}b}$,\\[1.5ex]
(ii) if $\te_{\ol i}b=0$, then $\te_{\ol i}(l_b)=0$.
}
\right.
\eneqn
Similar properties hold for $\tf_{\ol i}$. Note that
$$
\begin{aligned}
\Ker(\teibar\cl L/qL\to L/qL) =& \Ker (\teone S_{w_i}) = S_{w_i}^{-1}(\Ker \teone)
= S_{w_i^{-1}}(\Ker \teone)\\
=& S_{w_i^{-1}} \Big( \soplus_{\teone b=0} l_{b} \Big)
= \soplus_{\teone b =0} l_{S_{w_i^{-1}} b}
=\soplus_{\teone S_{w_i} b =0} l_b = \soplus_{\teibar b = 0} l_b .
\end{aligned}
$$
\end{remark}
\begin{example}
Let $\la$ be a strict partition.
Observe that $\B(Y_\la)$ has a unique element of weight $\la$, say $b_{Y_\la}$.
Since $\la + \alpha_i \notin \wt(\B(Y_\la))$ for any $i=1, 2, \ldots, n-1$,
$b_{Y_\la}$ is a highest weight vector.
Thus, for each admissible reading $\psi$, we see that $\psi(b_{Y_\la})$ is a highest weight vector
in $\B^{\tensor N}$.
\end{example}
\begin{lemma} \label{le:existence of h.w. vectors}
Every abstract $\q(n)$-crystal contains a highest weight vector.
\begin{proof}
Recall that $\la \in \wt(B)\seteq\set{\wt(b)}{b \in B}$
is called {\it maximal} if
$\la + \alpha_i \notin \wt(B)$ for $i=1,2,\ldots,n-1$.
Since $\wt(\teibar b)=\wt(b)+ \alpha_i $,
a vector in a crystal $B$ with a maximal weight is a highest weight vector.
Because $\wt(B)$ is a finite set,
there exists a maximal element $\lambda$ so that
we have an element $b \in B$ with a maximal weight $\lambda$.
\end{proof}
\end{lemma}
\begin{remark} \label{re:Clifford algebras}
\bna
\item Let $\la$ be a strict partition with $\ell(\la) =r$
and let $M$ be a highest weight module of
highest weight $\la$ in $\Oint$. Set $\tki =
q^{k_i-1} k_{\ol i}$ for $i=1,2,\ldots, n$.
Since $M \in \Oint$, we have $\tki=0$ on $M_\la$ for $i > r$.
Note that $\tki^2=\dfrac{1-q^{4\la_i}}{1-q^4}$ on $M_\la$
and
$\Big (\dfrac{1-q^{4\la_i}}{1-q^4} \Big)^{-\frac{1}{2}} \in \A \subset \F$.
Let
$$C_i \seteq \Big (\dfrac{1-q^{4\la_i}}{1-q^4} \Big)^{-\frac{1}{2}} \tki.$$
Then on $M_\la$, we have
\eq \label{defining relations of Clifford algebra}
C_i^2=1, \ C_i C_j + C_j C_i = 0 \ (i \neq j).
\eneq
Thus $M_\lambda$ can be regarded as a module over $\F[C_1,\ldots,C_r]$,
where $\F[C_1,\ldots,C_r]$ is the associative $\F$-algebra generated by $\set{C_i}{i=1,2,\ldots,r}$
with the defining relations \eqref{defining relations of Clifford algebra}.
\item
Let $\C[C_1,\ldots,C_r]$ and $\A[C_1,\ldots,C_r]$ be the
associative $\C$-algebra and $\A$-algebra, respectively,
generated by $\set{C_i}{i=1,2,\ldots,r}$ with the defining
relations \eqref{defining relations of Clifford algebra}. For a
superring
$R$, we define $\Mod(R)$ and $\SMod(R)$ to be the category of
$R$-modules and the category of $R$-supermodules, respectively.
If $r$ is odd, then we have the following commutative diagram :
$$
\xymatrix{
\Mod(\A) \ar[r]^(.34){\sim} \ar[d]^{\F \tensor_\A ( - )} & \SMod(\A[C_1,\ldots, C_r]) \ar[d]^{\F \tensor_\A ( - )}\\
\Mod(\F) \ar[r]^(.34){\sim} & \SMod(\F[C_1,\ldots, C_r]),
}
$$
If $r$ is even, then we have the following commutative diagram :
$$\xymatrix{
\SMod(\A) \ar[r]^(.36){\sim} \ar[d]^{\F \tensor_\A ( - )} & \SMod(\A[C_1,\ldots, C_r]) \ar[d]^{\F \tensor_\A ( - )}\\
\SMod(\F) \ar[r]^(.36){\sim} & \SMod(\F[C_1,\ldots, C_r]).
}
$$
In both cases, the horizontal arrows are given by
$$K \mapsto V \tensor_\C K$$
for each module $K$ in the left hand side,
where $V$ denotes an irreducible supermodule over $\C[C_1,\ldots,C_r]$.
\ee
\end{remark}
To summarize, we obtain the following proposition.
\begin{prop} \label{prop:uniqueness of lattice of highest weight space of an irreducible module}
\bna
\item For a strict partition $\la \in \Lambda^+$ with $l(\la)=r$, let $\rm HT(\la)$ be the category of highest weight modules with highest weight $\la$
in $\Oint$.
Then $\rm HT(\la)$ is equivalent
to $\SMod(\F[C_1,\ldots,C_r])$, where the equivalence is given by
$${\rm HT(\la)} \ni M \mapsto M_\la \in \SMod(\F [C_1,\ldots,C_r]).$$
In particular, the homomorphism
$\End_{\uqqn}(M)\to\End_{\F[C_1,\ldots,C_r]}(M_\la)$
is an isomorphism for any $M\in \rm HT(\la)$.
\item For a $\uqqn$-module $M \in \rm HT(\la)$, let $L$, $L'$
be finitely generated free $\A$-submodules of $M_\la$
such that
\quad {\rm (i)} $L$ and $L'$ are stable under $\tki$'s
$(i=1,2,\ldots, n)$,
\quad {\rm (ii)} $\F \otimes_\A L \simeq \F \otimes_\A L'\simeq M_\la$.
\ee
Then there exists a $\uqqn$-module automorphism $\vphi$ of $M$ such
that $\vphi L = L'$.
\end{prop}
\vskip 1cm
\section{Highest weight vectors in $\B^{\tensor N}$}
In this Section, we will give algebraic and combinatorial
characterizations of highest weight vectors in the abstract $\q(n)$-crystal
$\B^{\tensor N}$.
\begin{definition}
Let $B$ be an abstract $\q(n)$-crystal.
\bni
\item An element $b \in B$ is called a \emph{$\gl(a)$-highest weight vector} if $\te_i b = 0$ for $1 \leq i < a \leq n$.
\item An element $b \in B$ is called a \emph{$\q (a)$-highest weight vector} if $\te_i b = \teibar b = 0$ for $1 \leq i < a \leq n$.
\end{enumerate}
\end{definition}
\noindent
In particular, a highest weight vector in $B$ is a $\q(n)$-highest weight vector.\\
From now on, we denote by $\tensor_{j \geq m \geq i} (r_1 \ r_2
\cdots r_m )^{\tensor y_m} $ the following vector in $\B^{\tensor
N}$ :
$$\begin{array}{l}
\underbrace{(r_1 \tensor \cdots \tensor r_j) \tensor
\cdots \tensor (r_1 \tensor \cdots \tensor r_j)}_{y_j - \rm{times}}
\tensor
\underbrace{(r_1 \tensor \cdots \tensor r_{j-1}) \tensor
\cdots \tensor (r_1 \tensor \cdots \tensor r_{j-1})}_{y_{j-1} - \rm{times}}
\tensor \\[.5ex]
\cdots \tensor \underbrace{(r_1 \tensor \cdots \tensor r_{i+1}) \tensor
\cdots \tensor (r_1 \tensor \cdots \tensor r_{i+1})}_{y_{i+1} - \rm{times}}
\tensor \underbrace{(r_1 \tensor \cdots \tensor r_i) \tensor
\cdots \tensor (r_1 \tensor \cdots \tensor r_i)}_{y_i - \rm{times}},
\end{array}
$$
where $N = \sum^{j}_{m=i} m y_m$.
Let $b$ be an element of a $\gl(n)$-crystal $B$. We denote
by $C(b)$ the connected component of $B$ containing $b$.
\begin{definition}
Let $B_i$ be a $\gl(n)$-crystal and let $b_i \in B_i$
$(i=1,2)$. We say that $b_1$ is \emph{$\gl(n)$-crystal equivalent}
to $b_2$ if there exists an isomorphism of $\gl(n)$-crystals $C(b_1)
\isoto C(b_2)$ sending $b_1$ to $b_2$.
\end{definition}
Recall that $w_i = s_2 \cdots s_{i} s_1 \cdots s_{i-1}$.
\begin{lemma} \label{le:swib0}
Let $B$ be a $\gl(n)$-crystal.
\bna
\item A vector $b_0$ in $\B \tensor B$ is a
$\gl(n)$-highest weight vector if and only if
$b_0=1 \tensor \tf_1 \cdots \tf_{j-1} b$ for some $j \in \{1,2,\ldots, n\}$ and
some $\gl(n)$-highest weight vector $b \in B$ such that $\wt(b_0) =
\wt(b)+\eps_j$ is a partition.
\item Let $b$ be a $\gl(n)$-highest weight vector in $B$ and
$j \in \{1,2,\ldots, n\}$.
If $\wt(b)+\eps_j$ is a partition, then
$b_0=1\tensor \tf_1 \cdots \tf_{j-1} b$ is a $\gl(n)$-highest weight
vector in $\B \tensor B$ and we have
$$
S_{w_i} b_0 =
\begin{cases}
3 \tensor \tf_3 \cdots \tf_{j+1} S_{w_i} b & \text{if} \quad j+1 \leq i < n, \\
1 \tensor S_{w_i} b & \text{if} \quad i=j, \\
1 \tensor \tf_1 S_{w_i} b & \text{if} \quad i=j-1,
\end{cases}
$$
and
$$
S_{u_i}b_0 = 1 \tensor \tf_1 \tf_2 S_{u_i} b' \quad \text{if} \quad i \leq j-2,
$$
where $z_i = s_3 s_4 \cdots s_{i+1}$, $u_i = z_i w_i $ and $b' = \tf_{i+2} \cdots \tf_{j-1} b$.
\ee
\begin{proof}
{\rm (a)}
For a partition $\la$, let us denote by $B_{\gl(n)}(\la)$ the crystal graph of the highest weight $\gl(n)$-module
with highest weight $\la$.
It is enough to show that the assertion holds for $B=B_{\gl(n)}(\la)$ for any partition $\la$.
Let $b_0=1 \tensor \tf_1 \cdots \tf_{j-1} b$ for some
$\gl(n)$-highest weight
vector $b \in B$ such that $\wt(b_0)$ is a partition.
Since any two $\gl(n)$-highest weight vectors with the same
highest weight are $\gl(n)$-crystal equivalent,
by embedding $B$ to $\B^{\tensor N}$ for some $N$,
we may assume that
$b=\tensor_{n \geq m \geq 1} (12 \cdots m)^{\tensor x_m}$, where
$x_m=\langle k_m-k_{m+1} , \wt(b) \rangle$ for $1 \leq m \le n-1$.
Since $\wt(b)+\eps_j = \wt(b_0)$ is a partition, we have
$x_{j-1} \geq 1$. Thus we have
\begin{equation} \label{eq:gln h.w. vectors}
\begin{array}{ll}
&1 \tensor \tf_1 \tf_2 \cdots \tf_{j-1} b = \\
& \qquad1 \tensor \tensor_{m \geq j} (1 \cdots m)^{\tensor x_m}
\tensor (2 3 \cdots j) \tensor \tensor_{j-1 \geq m \geq 1}(1 \cdots
m)^{\tensor(x_m-\delta_{m,{j-1}})},
\end{array}
\end{equation}
which is a $\gl(n)$-highest weight vector in $\B \tensor B$.
Since we have
$$\B \tensor B \simeq \soplus_{ \la + \eps_j : \, \, partiton} B_{\gl(n)}(\la+\eps_j), $$
the number of highest weight vectors in $\B \tensor B$ is the same as
the number of vectors of the form in \eqref{eq:gln h.w. vectors}.
\bigskip
\noi
{\rm(b)}
We may assume that $b=\tensor_{n \geq m \geq 1} (12 \cdots m)^{\tensor x_m}$
as above.
Then by \eqref{eq:gln h.w. vectors}, we have
\begin{equation}\label{eq:b0}
b_0 = 1 \tensor \tensor_{m \geq j} (1 2\cdots m)^{\tensor x_m} \tensor (2 3 \cdots j) \tensor
\tensor_{j-1 \geq m \geq 1}(1 2 \cdots m)^{\tensor (x_m-\delta_{m,j-1})}.
\end{equation}
We also have
\begin{equation} \label{eq:swib}
\begin{aligned}
S_{w_i}b = & \tensor_{m \geq i+1} (1 2\cdots m)^{\tensor x_m} \tensor(1 3 4 \cdots i+1)^{\tensor x_i} \tensor \\
& \tensor_{i-1 \geq m \geq 1}(3 4 \cdots m+2)^{\tensor x_m}.
\end{aligned}
\end{equation}
Here we used the following facts :
\begin{enumerate}
\item For $w \in W$ and $\gl(n)$-highest weight vectors $b_1$ and $b_2$,
$$S_w(b_1 \tensor b_2) = S_w b_1 \tensor S_w b_2. $$
\item Suppose that $0 < a_1 < a_2 < \cdots < a_r \leq n $, $0 < x_1 < x_2 < \cdots < x_r \leq n$ and
$w(\{a_1, \ldots a_r \})=\{x_1, \ldots x_r\}$.
Then we have
$$S_w (a_1 \otimes \cdots \otimes a_r) = x_1 \otimes \cdots \otimes x_r.$$
\end{enumerate}
\noi
{\bf Case 1:} $j+1 \leq i < n$. \\
From \eqref{eq:b0}, we have
\begin{equation*}
\begin{array}{lll}
S_{w_i}b_0 =&3 \tensor \tensor_{m \geq i+1} (1 2\cdots m)^{\tensor x_m} \tensor(1 3 4 \cdots i+1)^{\tensor x_i} \tensor \\
& \tensor_{i-1 \geq m \geq j}(3 4 \cdots m+2)^{\tensor x_m} \tensor (4 5 \cdots j+2) \tensor \\
& \tensor_{j-1 \geq m \geq 1} (3 \cdots m+2)^{\tensor(x_m-\delta_{m, j-1})}.
\end{array}
\end{equation*}
On the other hand, from \eqref{eq:swib}, we have
\begin{equation*}
\begin{array}{lll}
\tf_3 \cdots \tf_{j+1} S_{w_i}b =&\tensor_{m \geq i+1} (1 2\cdots m)^{\tensor x_m} \tensor(1 3 4 \cdots i+1)^{\tensor x_i} \tensor \\
&\tensor_{i-1 \geq m \geq j}(3 4 \cdots m+2)^{\tensor x_m} \tensor (4 5 \cdots j+2) \tensor \\
&\tensor_{j-1 \geq m \geq 1}(3 4 \cdots m+2)^{\tensor(x_m-\delta_{m,j-1})}.
\end{array}
\end{equation*}
Thus we get $$S_{w_i} b_0 = 3 \tensor \tf_3 \cdots \tf_{j+1} S_{w_i} b.$$
\noi
{\bf Case 2:} $i=j$. \\
From \eqref{eq:b0} and \eqref{eq:swib}, we have
\begin{equation*}
\begin{array}{lll}
S_{w_i} b_0 &= S_{w_i}\big(1\tensor \tensor_{m \geq j}(1\cdots m)^{\tensor x_m} \tensor(2 \cdots j)
\tensor\tensor_{j-1 \geq m \geq 1}(1 \cdots m)^{\tensor (x_m-\delta_{m,j-1})}\big)\\
&= S_2 \cdots S_j \big(1\tensor \tensor_{m \geq j}(1\cdots m)^{\tensor x_m} \tensor
\tensor_{j-1 \geq m \geq 1}(2 \cdots m+1)^{\tensor x_m}\big) \\
&= 1\tensor \tensor_{m \geq j+1}(1\cdots m)^{\tensor x_m} \tensor(1 3 4 \cdots j+1)^{\tensor x_j} \tensor
\tensor_{j-1 \geq m \geq 1}(3 \cdots m+2)^{\tensor x_m} \\
&=1 \tensor S_{w_i} b.
\end{array}
\end{equation*}
\noi
{\bf Case 3:} $i=j-1$. \\
From \eqref{eq:b0}, we have
\begin{equation*}
\begin{array}{lll}
S_{w_i} b_0 &=& S_2 \cdots S_{j-1} \big(1\tensor \tensor_{m \geq j}(1\cdots m)^{\tensor x_m} \tensor
(2 \cdots j) \tensor (1 2 \cdots j-1)^{\tensor (x_{j-1} -1)} \tensor \\
& & \tensor_{j-2 \geq m \geq 1}(2 \cdots m+1)^{\tensor x_m}\big) \\
&=& 1\tensor \tensor_{m \geq j}(1\cdots m)^{\tensor x_m} \tensor (2 \cdots j) \tensor
\tensor(1 3 4 \cdots j)^{\tensor (x_{j-1}-1)} \tensor \\
&&\tensor_{j-2 \geq m \geq 1}(3 \cdots m+2)^{\tensor x_m}.
\end{array}
\end{equation*}
On the other hand, from \eqref{eq:swib}, we have
\begin{equation*}
\begin{array}{lll}
&1 \tensor \tf_1 S_{w_i} b \\
&= 1 \tensor \tf_1 \big( \tensor_{m \geq j}(1\cdots m)^{\tensor x_m} \tensor
(1 3 4 \cdots j)^{\tensor x_{j-1}} \tensor \tensor_{j-2 \geq m \geq 1}(3 \cdots m+2)^{\tensor x_m}\big) \\
&= 1\tensor \tensor_{m \geq j}(1\cdots m)^{\tensor x_m} \tensor (2 \cdots j)\otimes
(1 3 4 \cdots j)^{\tensor (x_{j-1}-1)} \tensor \tensor_{j-2 \geq m \geq 1}(3 \cdots m+2)^{\tensor x_m}.
\end{array}
\end{equation*}
Hence we get
\begin{equation*}
S_{w_i} b_0 = 1 \tensor \tf_1 S_{w_i} b.
\end{equation*}
\noi
{\bf Case 4:} $i \leq j-2$. \\
Note that
\begin{equation*}
u_i(m) =
\begin{cases}
m+3 & 1 \leq m < i, \\
1 & m= i, \\
2 & m=i+1 , \\
3 & m=i+2 , \\
m & m \geq i+3.
\end{cases}
\end{equation*}
We have
\begin{equation*}
\begin{array}{l}
S_{u_i} b_0 \\
= S_3 \cdots S_{i+1}
\big(1 \tensor \tensor_{m \geq j}(1 \cdots m)^{\tensor x_m}
\tensor (2 \cdots j) \tensor
\tensor_{j-1 \geq m \geq i+1} (1 \cdots m)^{\tensor (x_m -\delta_{m, j-1})} \\[1ex]
\hs{15ex} \tensor (1 3 \cdots i+1)^{\tensor x_i} \tensor
\tensor_{i-1 \geq m \geq 1} (3 \cdots m+2)^{\tensor x_m} \big) \\[1.5ex]
=1 \tensor \tensor_{m \geq j}(1 \cdots m)^{\tensor x_m} \tensor (2 \cdots j)
\tensor_{j-1 \geq m \geq i+2} (1 \cdots m)^{\tensor (x_m -\delta_{m, j-1})} \\[1ex]
\hs{10ex} \tensor (1 2 4 \cdots i+2)^{\tensor x_{i+1}} \tensor (1 4 \cdots i+2)^{\tensor x_i}
\tensor \tensor_{i-1 \geq m \geq 1} (4 \cdots m+3)^{\tensor x_m}.
\end{array}
\end{equation*}
On the other hand, we have
\begin{equation*}
\begin{array}{l}
\tf_1 \tf_2 ( S_{u_i} b') \\[1ex]
\hs{3ex}= \tf_1 \tf_2 S_{u_i} \big(\tensor_{m \geq j}(1 \cdots m)^{\tensor x_m}
\tensor (1 \cdots i+1 \ i+3 \cdots j) \\
\hs{25ex} \tensor
\tensor_{j-1 \geq m \geq 1} (1 \cdots m)^{\tensor (x_m -\delta_{m, j-1})} \big) \\[1ex]
\hs{3ex}=\tf_1 \tf_2 \big( \tensor_{m \geq j}(1 \cdots m)^{\tensor x_m}
\tensor (1 2 4 \cdots j) \tensor
\tensor_{j-1 \geq m \geq i+2} (1 \cdots m)^{\tensor (x_m -\delta_{m, j-1})} \\
\hs{15ex} \tensor (1 2 4 \cdots i+2)^{\tensor x_{i+1}}
\tensor (1 4 \cdots i+2)^{\tensor x_i} \tensor
\tensor_{i-1 \geq m \geq 1} (4 \cdots m+3)^{\tensor x_m} \big) \\[1ex]
\hs{3ex}=\tensor_{m \geq j}(1 \cdots m)^{\tensor x_m}
\tensor (2 3 4 \cdots j) \tensor
\tensor_{j-1 \geq m \geq i+2} (1 \cdots m)^{\tensor (x_m -\delta_{m, j-1})} \\
\hs{15ex} \tensor (1 2 4 \cdots i+2)^{\tensor x_{i+1}}
\tensor (1 4 \cdots i+2)^{\tensor x_i} \tensor
\tensor_{i-1 \geq m \geq 1} (4 \cdots m+3)^{\tensor x_m}.
\end{array}
\end{equation*}
Thus, we obtain
\begin{equation*}
S_{u_i} b_0 = 1 \tensor \tf_1 \tf_2 S_{u_i} b'.
\end{equation*}
\end{proof}
\end{lemma}
\begin{lemma}\label{lem:e1f1}
Assume that $b\in \B^{\tensor N}$ satisfies
$\tf_1b\not=0$ and $\teone \tf_1b=0$.
Then $\teone b=0$.
\end{lemma}
\Proof
If $b$ does not contain $2$, then it is trivial.
Assume that $b$ contains $2$ and $\teone b\not=0$. Then we can write
$b=b_1\otimes 2\otimes b_2$ such that $b_2$ contains neither $1$ nor $2$.
Since $\tf_1b\not=0$, we have
$\tf_1b=(\tf_1b_1)\otimes 2\otimes b_2$ and $\tf_1b_1\not=0$.
Therefore, $\teone \tf_1b=(\tf_1b_1)\otimes 1\otimes b_2$ does not vanish,
which is a contradiction.
\QED
\begin{theorem} \label{th:h.w. vectors}
Suppose that $b$ is a $\gl(n)$-highest weight vector in $\B^{\tensor (N-1)}$
and $b_0=1 \tensor \tf_1 \cdots \tf_{j-1} b$ is a highest
weight vector in $\B^{\tensor N}$. Then $b$ is
a highest weight vector in $\B^{\tensor(N-1)}$.
\end{theorem}
\begin{proof}
We shall prove $\teibar b =0 $ for $1 \leq i < n$.
\bigskip
\noindent
{\bf Case 1:} $j+1 \leq i <n$.
By Lemma~\ref{le:swib0}, we have $$S_{w_i} b_0 = 3 \tensor \tf_3 \cdots \tf_{j+1} S_{w_i} b.$$
Since $0=\te_{\ol 1} S_{w_i} b_0= \te_{\ol 1} (3 \tensor \tf_3 \cdots \tf_{j+1} S_{w_i} b)$,
we obtain $\te_{\ol 1} S_{w_i} b =0$. \\
\noindent
{\bf Case 2:} $i=j$.
We have
\begin{equation*}
S_{w_i} b_0 =1 \tensor S_{w_i} b.
\end{equation*}
Since $0=\te_{\ol 1} S_{w_i} b_0= \te_{\ol 1} (1 \tensor S_{w_i} b)$,
we get $\te_{\ol 1} S_{w_i} b =0$.
\bigskip
\noindent
{\bf Case 3:} $i=j-1$.
Since
\begin{equation*}
S_{w_i} b_0 = 1 \tensor \tf_1 S_{w_i} b,
\end{equation*}
we have
\begin{equation*}
\te_{\ol 1} \tf_1 S_{w_i} b = 0.
\end{equation*}
Hence Lemma~\ref{lem:e1f1} implies $\te_{\ol 1} S_{w_i} b = 0$.
\bigskip
\noindent
{\bf Case 4:} $i \leq j-2$.
Set $b' \seteq \tf_{i+2} \cdots \tf_{j-1} b$.
Then $\te_k b' =0$ for $k \leq i+1$.
Hence $b'$ is a $\gl(i+2)$-highest weight vector.
Since $u_i^{-1}(\alpha_1)$ and $u_i^{-1}(\alpha_2)$ are positive roots,
$S_{u_i} b'$ is a $\gl(3)$-highest weight vector.
Here we have used the fact:
\eq&&
\parbox{\mylength}{if $b$ is a $\gl(n)$-highest weight vector
and $w^{-1}(\alpha_i)$ is a positive root for $w\in W$
and $i$, then $\te_iS_wb=0$.}\label{eq:wh}
\eneq
For the same reason, $S_{u_i} b_0$ is a $\gl(n)$-highest weight vector.
By Lemma~\ref{le:swib0}, we have
\begin{equation*}
S_{u_i}b_0 = 1 \tensor \tf_1 \tf_2 S_{u_i} b'.
\end{equation*}
Since $\te_{\ol 1}$ commutes with $S_3, \ldots S_{n-1}$,
$\te_{\ol1}$ commutes with $S_{z_i}$.
Hence
\begin{equation*}
\te_{\ol 1} S_{u_i} b_0 = \te_{\ol 1} S_{z_i} S_{w_i} b_0 = S_{z_i} \te_{\ol 1} S_{w_i} b_0 = 0.
\end{equation*}
Since $w_2 u_i = z_i w_{i+1}$, we also have
\begin{equation*}
\te_{\ol 1} S_{w_2} S_{u_i} b_0 = \te_{\ol 1} S_{z_i} S_{w_{i+1}}
b_0 =S_{z_i} \te_{\ol 1} S_{w_{i+1}} b_0 =0.
\end{equation*}
Thus $S_{u_i} b_0$ is a $\q(3)$-highest weight vector.
By Lemma~\ref{le:q3hw} below, we have $\te_{\ol 1} S_{u_i} b' =0.$
Since $\te_{\ol 1}$ commutes with $S_{z_i}$, we get
$\te_{\ol1} S_{z_i} S_{w_i} b' = S_{z_i} \te_{\ol 1} S_{w_i} b'$, and hence we conclude
$\te_{\ol i} b'= 0$.
On the other hand,
$\te_{\ol i}$ commutes with $\te_{j-1}\cdots\te_{i+2}$,
because $\te_k$ ($k\ge i+2$) commutes with
$S_1,\ldots, S_i$ and $\teone$.
Hence $\te_{j-1}\cdots\te_{i+2}$ commutes with $\te_{\ol i}$.
Since $b=\te_{j-1}\cdots\te_{i+2}b'$,
we obtain $\te_{\ol i}b=\te_{\ol i}\te_{j-1}\cdots\te_{i+2}b'=
\te_{j-1}\cdots\te_{i+2}\te_{\ol i}b'=0$.
\end{proof}
\begin{lemma} \label{le:q3hw}
Suppose that $b$ is a $\gl(3)$-highest
weight vector in $\B^{\tensor (N-1)}$
and $b_0=1 \tensor \tf_1 \tf_2 b$ is a $\q(3)$-highest
weight vector in $\B^{\tensor N}$. Then $\te_{\ol 1} b =0$.
\begin{proof}
If $\te_{\ol1} b \neq 0$, then $b=b_1 \tensor 2 \tensor b_2$, where
$b_2$ contains neither $1$ nor $2$.
Since $\te_{\ol 1} b_0 =0$, we have $\te_{\ol 1} \tf_1 \tf_2 b=0$ and
hence Lemma~\ref{lem:e1f1} implies $\te_{\ol 1} \tf_2 b=0$.
It follows that $\tf_2 b = b_1\tensor 3 \tensor b_2$.
Hence $\te_{\ol 1}\tf_2 b=0$ implies $$\te_{\ol 1}b_1=0.$$
Moreover, $\tf_2(b_1 \tensor 2 \tensor b_2)=b_1\tensor 3 \tensor b_2$
implies that $\varphi_2(b_1)=0$ and $b_2$ does not contain $3$.
Since $\varepsilon_2(b_1)=0$, we conclude that
$b_1$ is $\gl(3)$-crystal equivalent to $1^{\tensor x}$ for some
positive integer $x$. Thus we get
\begin{equation}
\begin{array}{rl}
S_2 S_1 b_0 &= S_2 S_1 (1 \tensor \tf_1 \tf_2 b)
= S_2 S_1 (1 \tensor \tf_1 b_1 \tensor 3\otimes b_2) \\
&= S_2 (1 \tensor S_1 b_1 \tensor 3\otimes b_2)
= 1 \tensor \te_2 S_2 S_1 b_1 \tensor 3\otimes b_2.
\end{array}\label{eq:s2s1}
\end{equation}
Here the third equality follows from
$$S_1(1\tensor \tf_1 (1^{\tensor x})) = S_1(1 \tensor 2 \tensor 1^{\tensor (x-1)})
= 1 \tensor 2\tensor 2^{\tensor (x-1)} = 1 \tensor S_1 (1^{\tensor x}),$$
and the last equality follows from
$$S_2(1 \tensor S_1(1^{\tensor x}) \tensor 3) =
S_2(1\tensor 2^{\tensor x} \tensor 3)=1\tensor 3^{\tensor (x-1)}\tensor2\tensor 3
=1 \tensor \te_2S_2 S_1 (1^{\tensor x}) \tensor 3.$$
Since $\te_{\ol 2} b_0 = 0$ by the assumption, $\teone S_2 S_1 b_0=0$,
and \eqref{eq:s2s1} implies
$$\te_{\ol 1} \te_{2} S_2 S_1 b_1 = 0.$$
On the other hand, $\tf_1(b_1\otimes 3\otimes b_2)=\tf_1\tf_2b\not=0$ implies
$\tf_1b_1\not=0$.
Hence $b_1$ contains $1$, and
$\te_{\ol 1} b_1 =0 $ implies that
$b_1 = b_3 \tensor 1 \tensor b_4$ where
$b_4$ contains neither $1$ nor $2$.
Since $b_3$ is a $\gl(3)$-highest weight vector,
we have $S_1 b_1 = S_1 b_3 \tensor 2 \tensor b_4$.
Since $\te_2 S_1 b_1 =0$ by \eqref{eq:wh}, we have $\te_2 S_1 b_3 =0$.
Then we have
$\te_2 S_2 S_1 b_1 =b_5 \tensor 2 \tensor b_4$,
for some $b_5$
because
$\te_2S_2(2^{\otimes y} \otimes 2 \otimes 3^{\otimes z})
=\te_2 \bigl(3^{\otimes (y+1-z)} \otimes 2^{\otimes z} \otimes 3^{\otimes z} \bigr)
=3^{\otimes(y-z)} \otimes 2^{\otimes z} \otimes 2 \otimes 3^{\otimes z}$.
This contradicts $\te_{\ol 1} \te_2 S_2 S_1 b_1 = 0$.
Hence we get the desired result $\teone b=0$.
\end{proof}
\end{lemma}
\begin{lemma} \label{le:e1barbneq0}
If $\varepsilon_1 b = 0$ and $\langle k_1, \wt(b) \rangle = \langle k_2, \wt(b) \rangle > 0$,
then $\te_{\ol 1} b \neq 0$.
\begin{proof}
Assume that $\te_{\ol 1} b = 0$.
Then $b=b_1 \tensor 1 \tensor b_2$ for some $b_1$ and $b_2$,
where $b_2$ contains neither $1$ nor $2$.
Since $\varepsilon_1(b_1)=0$, we have
$$\langle k_1 , \wt(b) \rangle = 1+ \langle k_1 , \wt(b_1) \rangle \geq 1+\langle k_2 , \wt(b_1) \rangle = \langle k_2 , \wt(b) \rangle +1,$$
which is a contradiction.
\end{proof}
\end{lemma}
\begin{prop} \label{prop:strict partition}
If $b$ is a highest weight vector in $\B^{\tensor N}$, then $\wt(b)$ is a strict partition.
\end{prop}
\begin{proof}
Assuming that $\langle k_i, \wt(b) \rangle = \langle k_{i+1}, \wt(b)
\rangle > 0$, we shall derive a contradiction.
Set $b' \seteq S_{w_i} b$. Since $w_i^{-1}(\alpha_1)=\alpha_i$,
\eqref{eq:wh} implies $\te_1 b'=0$. Hence
Lemma~\ref{le:e1barbneq0} implies $\teone b'\not=0$, which is a contradiction.
\end{proof}
\begin{lemma} \label{le:A}
Let $b$ be a vector in $\B^{\tensor N}$.
{\rm (a)} If $\te_{\ol 1} b= \te_1 b=0$ and $\langle k_1, \wt(b)
\rangle \geq \langle k_2, \wt(b) \rangle +2$, then $\te_{\ol 1} (1
\tensor \tf_1 b) = 0$.
{\rm (b)}
If $\te_{\ol 1} b= \te_1 b = \te_2 b =0$ and $\langle k_2,
\wt(b) \rangle > \langle k_3, \wt(b) \rangle$, then $\te_{\ol
1} (1 \tensor \tf_1 \tf_2 b) = 0$.
\end{lemma}
\begin{proof}
(a) Since $\langle k_1, \wt(b) \rangle >0$ and $\te_{\ol 1} b=0$,
we can write $b=b_1 \tensor 1 \tensor b_2$ for some $b_1$ and $b_2$
such that $b_2$ contains neither $1$ nor $2$. Then we get
$$2 \leq \langle k_1, \wt(b) \rangle - \langle k_2, \wt(b) \rangle
= \langle k_1, \wt(b_1) \rangle - \langle k_2, \wt(b_1) \rangle + 1
=\varphi_1(b_1) - \varepsilon_1(b_1)+1.$$
Thus $\varphi_1(b_1) > 0 = \varepsilon_1(1)$ and hence $\tf_1 b= \tf_1 b_1 \tensor 1 \tensor b_2$
It follows that $\teone(1 \tensor \tf_1 b)=0$.
\bigskip
\noi
(b) Since $\langle k_1, \wt(b) \rangle \geq \langle k_2, \wt(b)
\rangle
>0$ and $\te_{\ol 1} b=0$, we can write $b=b_1 \tensor 1 \tensor
b_2$ for some $b_1$ and $b_2$ such that $b_2$ contains neither $1$
nor $2$. It follows that $\varepsilon_2(b_1) = \varphi_2(b_2)=0$.
Observe that
$$\varphi_2(b_1) = \langle k_2, \wt(b_1) \rangle - \langle k_3, \wt(b_1) \rangle
> \langle k_3, \wt(b_2) \rangle - \langle k_2, \wt(b_2) \rangle
= \varepsilon_2(b_2).$$
Hence we have $\tf_2 b = \tf_2 b_1 \tensor 1 \tensor b_2$.
Since $\varepsilon_1(\tf_2 b_1)=0$,
we deduce that
$$\varphi_1(\tf_2 b_1)= \langle k_1-k_2, \wt(\tf_2 b_1) \rangle
=\langle k_1-k_2, \wt(b_1) \rangle + 1
=\varphi_1(b_1) +1 > 0= \varepsilon_1(1 \tensor b_2),$$
and hence
$$\tf_1 \tf_2 b = \tf_1 \tf_2 b_1 \tensor 1 \tensor b_2.$$
Therefore we have $\te_{\ol 1} (1 \tensor \tf_1\tf_2 b)=0$.
\end{proof}
\begin{prop} \label{pro:h.w. vectors 2}
If $b \in \B^{\tensor (N-1)}$ is a highest weight vector with
$\langle k_{j-1}, \wt(b) \rangle \geq \langle k_j, \wt(b) \rangle +
2$, then $b_0 = 1 \tensor \tf_1 \cdots \tf_{j-1} b$ is a highest
weight vector in $\B^{\tensor N}$.
\begin{proof}
We will show $\te_{\ol i} b_0 = 0$ for $i = 1,2,\ldots, n-1$.
\vskip 3mm
\noi
{\bf Case 1:} $i \geq j+1$.
By Lemma~\ref{le:swib0}, we have
$$S_{w_i} b_0 = 3 \tensor \tf_3 \cdots \tf_{j+1} S_{w_i} b.$$
Thus we obtain
$$\te_{\ol 1} S_{w_i} b_0 = 3 \tensor \tf_3 \cdots \tf_{j+1} \te_{\ol 1} S_{w_i} b = 0.$$
\medskip
\noi
{\bf Case 2:} $i=j$.
Since $S_{w_i}b_0 = 1 \tensor S_{w_i} b$, we have
$\te_{\ol 1} S_{w_i} b_0=0$.\\
\medskip
\noi
{\bf Case 3:} $i=j-1$.
We have
$$S_{w_i} b_0 = 1 \tensor \tf_1 S_{w_i} b$$
and
$$\langle k_1, \wt(S_{w_i}b) \rangle = \langle k_{j-1}, \wt(b) \rangle
\geq \langle k_j, \wt(b) \rangle +2= \langle k_2, \wt(S_{w_i}b)
\rangle +2.$$
By Lemma~\ref{le:A}(a), we obtain $\te_{\ol 1} S_{w_i} b_0 = 0$. \\
\medskip
\noi
{\bf Case 4:} $i \leq j-2$.
Set $b' \seteq \tf_{i+2} \cdots \tf_{j-1} b$.
Here we understand $b'=b$ if $i=j-2$.
Then $b'$ is a $\gl(i+2)$-highest
weight vector and $\te_{\ol 1} b'=0$.
Because $\te_{\ol 1}$ commutes with $S_{u_i}$, we have $\te_{\ol 1} S_{u_i}b'=0$.
Since $u_i^{-1}(\alpha_1)$ and $u_i^{-1}(\alpha_2)$ are positive roots,
$S_{u_i} b'$ is a $\gl(3)$-highest weight vector by \eqref{eq:wh}.
By Lemma~\ref{le:swib0}, we have
$$S_{u_i} b_0 = 1 \tensor \tf_1 \tf_2 S_{u_i} b'.$$
Observe that
$$
\begin{aligned}
\langle k_2, \wt(S_{u_i}b') \rangle - \langle k_3, \wt(S_{u_i}b') \rangle
&= \langle k_{i+1}, \wt(b') \rangle - \langle k_{i+2}, \wt(b') \rangle \\
&= \langle k_{i+1}-k_{i+2} , \wt(b)-\eps_{i+2} + \eps_j \rangle \\
&= \langle k_{i+1}-k_{i+2}, \wt(b) \rangle + 1 - \delta_{j, i+2} \\
& \geq 1.
\end{aligned}
$$
By Lemma~\ref{le:A}(b), we get $\te_{\ol 1} S_{u_i} b_0 = 0$. Since
$S_{u_i}=S_{z_i} S_{w_i}$ and $\te_{\ol 1}$ commutes with $S_{z_i}$,
we obtain
$\te_{\ol 1} S_{w_i} b_0 =0$.
\end{proof}
\end{prop}
\begin{theorem} \label{th:char.h.w}
Assume that $b$ is a $\gl(n)$-highest
weight vector in $\B^{\tensor (N-1)}$
and $b_0\seteq1 \tensor \tf_1 \cdots \tf_{j-1} b$ is a $\gl(n)$-highest
weight vector in $\B^{\tensor N}$. Then $b_0$ is a highest
weight vector if and only if $b$ is a highest weight vector and
$\wt(b_0) = \wt(b)+\eps_j$ is a strict partition.
\begin{proof}
Note that $\wt(b)$ and $\wt(b_0)$ are partitions.
If $b_0$ is a highest weight vector, then by Theorem~\ref{th:h.w.
vectors} and Proposition~\ref{prop:strict partition}, $b$ is a
highest weight vector and
$\wt(b_0)$ is a strict partition.
Conversely, if $b$ is a highest weight vector so that $\wt(b)$ is a strict partition and $\wt(b) + \eps_j$ is still a strict partition,
then we have
$\langle k_{j-1} - k_j, \wt(b) \rangle \geq 2 $
and hence, by Proposition~\ref{pro:h.w. vectors 2}, $b_0$ is a
highest weight vector.
\end{proof}
\end{theorem}
\begin{comment}
If $M$ is an irreducible highest weight module in $\Oint$
of weight $\la$ for some strict partition $\la$ and $M_{\la}$ is generated by a free
$\A$-submodule $L_{\la}^{0}$ invariant under $\tki$ $(i=1, \ldots,n)$, then for any free $\A$-
submodule $L_{\la}$ of $M_\la$ which generates $M_\la$ and is stable under $\tki$'s
we have $L_{\la}^{0} = a L_{\la}$ for some $a \in \F$.
Assume that $r= \ell(\lambda)$ is odd. Since $M_\la$ is an irreducible over $\F[C_1,\ldots,C_r]$, it is given by
$$M_\la \isoto V \otimes_\C \F,$$
for an irreducible module $V$ over $\C[C_1,\ldots,C_r]$.
By Remark~\ref{re:Clifford algebras}, $L_{\la}^0$ is an $\A[C_1,\ldots,C_r]$-submodule of $V \otimes_\C \F$
given by $V \otimes_\C S^0$ for some $\A$-submodule $S^0$ of $\F$.
Since $L_{\la}^0$ is a finitely generated $\A$-module, $S^0$ is also a finitely generated $\A$-submodule of $\F$
and hence there exists $a \in \F$ such that $S^0=a \A$. One can also prove the same thing when $r$ is even.
\end{comment}
\vskip 1cm
\section{Existence and uniqueness}
In this section, we state and prove the main result of our paper:
the existence and uniqueness theorem for crystal bases. We first
prove several lemmas that are needed in the proof of our main
theorem.
We set
\eq
\tki=
q^{k_i-1} k_{\ol i}\quad\text{ for all $i=1, \ldots, n$.}
\eneq
\begin{lemma} \label{le:invarianct under tildeki}
Let $M$ be a $\Uq$-module in $\Oint$.
\bna
\item
For $\mu\in \wt(M)$ and $i\in\{1,\ldots,n-1\}$
such that $\mu+\alpha_i\not\in\wt(M)$, we have
$$\tk_{\ol{i+1}}=S_i\circ\tk_{\ol i}\circ S_i\quad\text{as endomorphisms of $M_\mu$,}$$
where $S_i$ is defined in {\rm Remark~\ref{rem:teibar}}.
\item Assume that $\la\in\wt(M)$ satisfies $\la+\alpha_i\not\in \wt(M)$
for all $i=1,\ldots,n-1$.
If $(L, B, l_{B})$ is a
crystal basis of $M$, then $L_\la$ is invariant under
$\tki$ for all $i=1,\ldots,n$.
\ee
\end{lemma}
\begin{proof}
(a) Set $\ell\seteq\lan h_i,\mu\ran\ge0$.
Then $S_i\cl M_\mu\isoto M_{s_i\mu}$
is given by $f_i^{(\ell)}$, and
its inverse is given by $e_i^{(\ell)}$.
Note that $e_iM_\mu=0$.
From the defining relation it follows that
\begin{equation*}
\begin{array}{lll}
e_{\ol i} f_i - f_i e_{\ol i} &= e_{\ol i} q^{-k_i} q^k_i f_i - f_i e_{\ol i} q^{-k_i} q^{k_i}\\
&= (k_{\ol i} e_i - q e_i k_{\ol i}) q^{k_i} f_i - f_i (k_{\ol i} e_i - q e_i k_{\ol i}) q^{k_i} \\
&= (k_{\ol i} e_i - q e_i k_{\ol i}) f_i q^{k_i -1} - f_i (k_{\ol i} e_i - q e_i k_{\ol i}) q^{k_i} \\
&= k_{\ol i} e_i f_i q^{k_i -1} - q e_i k_{\ol i} f_i q^{k_i -1} - f_i k_{\ol i} e_i q^{k_i} + q f_ie_i k_{\ol i} q^{k_i}\\
&= k_{\ol i}\left(f_i e_i + \dfrac{q^{h_i}-q^{-h_i}}{q - q^{-1}} \right)q^{k_i -1}
- q e_i k_{\ol i} f_i q^{k_i -1} - f_i k_{\ol i} e_i q^{k_i} + q f_ie_i k_{\ol i} q^{k_i}.
\end{array}
\end{equation*}
Thus on $M_{\mu}$, we have
\begin{equation*}
\begin{array}{lll}
e_{\ol i} f_i - f_i e_{\ol i} &= k_{\ol i}
\dfrac{q^{h_i}-q^{-h_i}}{q - q^{-1}} q^{k_i -1} - e_i k_{\ol i} f_i
q^{k_i},
\end{array}
\end{equation*}
which yields
\begin{equation} \label{eq:k i+1 bar}
\begin{array}{lll}
k_{\ol{i+1}} &= q^{-h_i} k_{\ol i} -(e_{\ol i} f_i - f_i e_{\ol i}) q^{-k_i} \\
&= q^{-h_i} k_{\ol i} - \left(k_{\ol i} \dfrac{q^{h_i}-q^{-h_i}}{q - q^{-1}} q^{k_i -1} - e_i k_{\ol i} f_i q^{k_i} \right) q^{-k_i} \\
&= q^{-h_i} k_{\ol i} - k_{\ol i} \dfrac{q^{h_i}-q^{-h_i}}{q - q^{-1}} q^{-1} + e_i k_{\ol i} f_i \\
&= -k_{\ol i} \left( \dfrac{q^{h_i -1}-q^{-h_i+1}}{q - q^{-1}} \right) + e_i k_{\ol i} f_i \\
&= -[\ell-1] k_{\ol i} + e_i k_{\ol i} f_i. \\
\end{array}
\end{equation}
On the other hand, we have, similarly to \eqref{eq:k1serre},
\begin{equation*}
\begin{array}{lll}
e_i ^{(2)} k_{\ol i} = e_i k_{\ol i} e_i- k_{\ol i} e_i^{(2)}. \\
\end{array}
\end{equation*}
By induction on $s$, we obtain
\begin{equation*}
\begin{array}{lll}
e_i ^{(s)} k_{\ol i} = e_i k_{\ol i} e_i^{(s-1)}-[s-1] k_{\ol i} e_i^{(s)} & (s \geq 1). \\
\end{array}
\end{equation*}
If $\ell >0$, we have on $M_{\mu}$
\begin{equation*}
\begin{array}{lll}
e_i^{(\ell)} q^{k_{i}-1} k_{\ol i} f_i^{(\ell)}
&= q^{-\ell} e_i^{(\ell)} k_{\ol i} f_i^{(\ell)} q^{k_i -1} \\
&= q^{-\ell} (e_i k_{\ol i} e_i^{(\ell-1)}-[\ell-1] k_{\ol i} e_i^{(\ell)}) f_i^{(\ell)} q^{k_i -1} \\
&= q^{-\ell} (e_i k_{\ol i} f_i - [\ell-1]k_{\ol i}) q^{k_i -1} \\
&= (e_i k_{\ol i} f_i - [\ell-1]k_{\ol i}) q^{k_{i+1} -1} \\
&= k_{\ol{i+1}} q^{k_{i+1} -1} = \tk_{\ol {i+1}}.
\end{array}
\end{equation*}
\noi
If $\ell=0$, then $f_i M_{\mu}=0$, and hence
\eqref{eq:k i+1 bar} implies $k_{\ol{i+1}} = k_{\ol i}$.
Therefore we have
$\tk_{\ol{i+1}}=k_{\ol{i+1}} q^{k_{i+1}-1}= k_{\ol i}q^{k_{i}-1}=\tk_{\ol{i}}$ on $M_{\mu}$.
In the both cases, we have $\tk_{\ol{i+1}} = S_i^{-1} \tki S_i$ on $M_{\mu}$.
\bigskip
\noi
(b)\quad
Let $M'=\uqqn M_\la \subset M$, and let $L'=L \cap M'$.
Set $\mu_j \seteq s_{j}\cdots s_{i-1} \la $ for $j=1,\ldots, i$.
Then $\langle h_{j}, \mu_{j+1} \rangle \ge0$, $s_j\mu_{j+1}=\mu_j$, and
$\mu_{j+1} + \alpha_j \notin \wt(M')$.
{}From (a) it follows that
$\tk_{\ol{j+1}}\vert_{M'_{\mu_{j+1}}}=S_j\circ\tk_{\ol j}\circ S_{j}\vert_{M'_{\mu_{j+1}}}$.
Hence, if $L'_{\mu_j}$ is stable under $\tk_{\ol{j}}$,
then $L'_{\mu_{j+1}}$ is stable under $\tk_{\ol{j+1}}$.
Since $L'_{\mu_1}$ is stable under $\tkone$,
$L'_{\mu_j}$ is stable under $\tk_{\ol{j}}$ for
all $j=1,\ldots, i$ by induction.
In particular, $L_{\la}=L'_\la$ is stable under $\tki$.
\end{proof}
\vskip 3mm
\begin{lemma} \label{le:uniqueness of crystal lattice}
Let $M$ be a $\uqqn$-module in $\Oint$,
and $\la\in P^{\ge0}$.
Let $(L, B, l_B)$ be a crystal basis of $M$ such that
any connected component of $B$ intersects with $B_\la$.
Let $L'$ be an $\A$-submodule of $M$
with the weight space decomposition $L'=\soplus_{\mu \in P^{\geq 0}} (L' \cap M_{\mu}) $,
which is stable under $\tei$, $\tfi$ $(i=1,\ldots,n-1,\ol1)$.
Then
\bna
\item $L'_{\lambda} \subset L_{\lambda}$ implies $L' \subset
L$,
\item $L'_{\lambda} \supseteq L_{\lambda}$ implies $L' \supseteq L$.
\end{enumerate}
\begin{proof}
(a)
Assume that $L'_{\lambda} \subset L_{\lambda}$.
Set $S \seteq (L \cap q L') / (qL \cap qL') $.
Then $S \subset L/qL$ and $S$ is stable under $\tei$, $\tfi$
($i=1,\ldots,n-1,\ol1$).
Note that
$$S_{\lambda} = S \cap (L/qL)_{\lambda} = (L_{\lambda} \cap q L'_{\lambda}) / (q L_{\lambda} \cap q L'_{\lambda}) = 0.$$
For each $b \in B$, let $P_b \cl L / qL \twoheadrightarrow l_b$ be the
canonical projection.
Since $S_{\lambda} = 0$, we have
$P_{b} (S) = 0$ for any $b \in B_\la$. If $\tei b \neq 0$ for some $i=1,\ldots, n-1,
\overline{1}$, then $\tei\circ P_b = P_{\tei b}\circ \tei$ implies
$\tei P_b (S) = P_{\tei b} \tei (S) \subset P_{\tei b} (S)$.
Therefore, if $P_{\tei b} (S) = 0$, then $P_b (S) = 0$.
The same property holds for $\tfi$.
Since any $b \in B$ can be connected with
an element of weight $\la$ by a sequence of
operators in $\tei$, $\tfi$ ($i=1,\ldots,n-1,\ol1$), we have
$P_b(S)=0$ for all $b \in B$. It follows that $S=0$ and hence $L
\cap q L' \subset q L$.
Since
$$L ' \cap q^{-m} L \subset q^{-(m-1)} (L' \cap q^{-1} L ) \subset q^{-(m-1)} L$$
for all $m\ge1$, we have $L ' \cap q^{-m} L \subset L' \cap q^{-(m-1)} L$.
Hence we obtain $L' \cap q^{-m} L \subset L$.
It follows that $L' \subset L$ as desired.
\bigskip
\noindent
(b) Assume that $L'_{\lambda} \supset L_{\lambda}$.
Set $S \seteq (L' \cap L) / (L' \cap qL)$.
Then $S \subset L/qL$ and $S$ is stable under $\tei$, $\tfi$.
Note that $l_{b}\subset S$ for any $b \in B_\la$.
If $\tei b \neq 0$ and $l_b \subset S$,
then $l_{\tei b} = \tei l_b \subset \tei S \subset S$.
The same is true for $\tfi$. Thus we have
$L/qL = \soplus_{b \in B} l_b \subset S$. By Nakayama's lemma, we
have $L' \cap L = L$.
\end{proof}
\end{lemma}
\begin{lemma} \label{le:crystal bases for irreducible summands}
Let $M$ be a highest weight $\uqqn$-module with highest weight $\la \in \La^+$
in the category $\Oint$.
Suppose that $M$ has a crystal basis $(L, B, l_B)$ such that
$B_\la = \{b_\la \}$ and $B$ is connected.
Let $L_\la = \soplus_{j=1}^s E_j$ be a decomposition into indecomposable modules over
$\A[C_1,\ldots, C_r]$ \ro see {\rm Remark~\ref{re:Clifford algebras}}\rf,
where $r=\ell(\la)$, and let
$$
\begin{aligned}
M_j \seteq \uqqn E_j, & \quad L_j \seteq M_j \cap L & \text{and} \quad l^{j}_b \seteq l_b \cap \big(L_j / qL_j \big).
\end{aligned}
$$
Then we have
\bna \item $M_j$ is irreducible over $\uqqn$, \item
$M=\soplus_{j=1}^s M_j$, $L = \soplus_{j=1}^s L_j$ and $l_b =
\soplus_{j=1}^s l^j_b$, \item $(L_j, B, (l^{j}_b)_{b \in B})$ is
a crystal basis of $M$. \ee
\begin{proof}
By Remark~\ref{re:Clifford algebras},
we see that
$(M_j)_\la \simeq\F \otimes_\A E_j$
is an irreducible module over $\F[C_1, \ldots, C_r]$ for each $j = 1,2, \ldots, s$.
Hence, Proposition~\ref{prop:uniqueness of lattice of highest weight space of an irreducible module} (a) implies
that $M_j$ is irreducible over $\uqqn$ and $M = \soplus_{j=1}^s M_j$.
Note that
$$L_j/ q L_j \subset L / qL \quad (j=1,2, \ldots, s).$$
Since $\soplus_{j=1}^s(L_j)_\la =\soplus_{j=1}^s \big(M_j \cap L_\la \big)=\soplus_{j=1}^s E_j=L_\la$,
we have
$$l_{b_{\la}} = L_\la / q L_\la = \soplus_{j=1}^s \big( (L_j)_\la / q(L_j)_\la \big)
=\soplus_{j=1}^s l^j_{b_{\la}}.$$
Consider $b_1, b_2 \in B$ such that $b_2=\tei b_1$ (equivalently,
$b_1=\tfi b_2$) for some $i =1,2,\ldots, n-1, \ol 1$.
Then we have injective maps
$$\tei|_{l^{j}_{b_1}} \cl l^{j}_{b_1} \mono l^{j}_{b_2}, \quad \tfi|_{l^{j}_{b_2}}
\cl l^{j}_{b_2} \mono l^{j}_{b_1}.$$
Hence comparing their dimensions, we conclude that
$$\tei \cl {l^{j}_{b_1}} \isoto l^{j}_{b_2}\quad\text{and}\quad
\tfi \cl {l^{j}_{b_2}} \isoto l^{j}_{b_1}
\quad\text{for all $j=1,2,\ldots, s$.}$$
Therefore $l_{b_1}=\soplus_{j=1}^s l^{j}_{b_1}$ if and only if
$l_{b_2}=\soplus_{j=1}^s l^{j}_{b_2}$.
Since $B$ is connected,
$\soplus_{j=1}^s l^{j}_{b} =l_{b}$
for all $b \in B$.
\noi
Since
$$L / q L = \soplus_{b \in B} l_b =
\soplus_{j=1}^s \soplus_{b \in B} l^{j}_b
\subset \soplus_{j=1}^s L_j /qL_j,$$
Nakayama's lemma implies that
$L=\soplus_{j=1}^s L_j$, and $(L_j, B, (l^{j}_b)_{b \in B})$
is a crystal basis of $M_j$.
\end{proof}
\end{lemma}
\begin{lemma} \label{le:isomorphic crystals}
Let $M$ be a $\Uq$-module in the category $\Oint$ and
let $(L_1, B_1, l^1_{B_1})$, $(L_2, B_2, l^2_{B_2})$ be
two crystal bases of $M$ such that $L_1=L_2$. If $B_1$ is a
connected abstract $\qn$-crystal and there exist $b_1 \in B_1$, $b_2
\in B_2$ such that $l^1_{b_1}=l^2_{b_2}$, then there exists a bijection
$\vphi \cl B_1 \to B_2$
which commutes with the Kashiwara operators and $l^1_b=l^2_{\vphi(b)}$
for all $b \in B_1$.
\begin{proof}
Let us set $S=\set{b\in B_1}{\text{there exists $b'\in B_2$ such that
$l^1_b=l^2_{b'}$}}$.
Then it is easy to see that it is stable under the Kashiwara operators
and it contains $b_1$.
Hence $S$ coincides with $B_1$.
Therefore for every $b\in B_1$,
there exists a $b'\in B_2$ such that $l^1_b=l^2_{b'}$.
Such a $b'$ is unique and we can define
$\vphi$ by $\vphi(b)=b'$.
Since $L_1/qL_1=\soplus\nolimits_{b\in B_1}l^1_b=\soplus\nolimits_{b\in B_2}l^2_b$,
$\vphi\cl B_1\to B_2$ is bijective.
\end{proof}
\end{lemma}
\begin{lemma} \label{le:existence of crystal base for a lattice}
Let $\la \in \Lambda^+$ and assume that $V(\la)$ has a crystal basis $(L_0,B_0,l_{B_0})$
such that $B_0$ is connected and $(B_0)_\la =\{b_\la\}$.
Let $M \in \Oint$ be a highest weight $\Uq$-module with highest weight
$\la \in \Lambda^+$.
If $E$ is a free $\A$-submodule of $M_{\la}$, which is stable under $\tki$
for $i=1,2,\ldots, n$ and generates $M_{\la}$ over $\F$,
then there exists a unique crystal basis $(L,B,l_B)$ such that
\bna
\item $L_{\la}=E$,
\item $B \simeq B_0$ as an abstract $\qn$-crystal,
\end{enumerate}
\end{lemma}
\begin{proof}
By Lemma~\ref{le:invarianct under tildeki}
and Proposition~\ref{prop:uniqueness of lattice of highest weight
space of an irreducible module}, there exists a finitely generated
free $\A$-module $K$ such that $M \simeq K \tensor_\A V(\la)$ and $E
\simeq K \tensor_\A (L_0)_\la$. Then $(K \tensor_\A (L_0), B_0, (K
\tensor l_b)_{b \in B_0})$ is a crystal basis for $M$. The
uniqueness follows from Lemma~\ref{le:uniqueness of crystal lattice}
and Lemma~\ref{le:isomorphic crystals}.
\end{proof}
\vskip 1ex
For a weight $\la = \la_1 \epsilon_1 + \cdots + \la_n
\epsilon_n \in P$, define $|\la| =\sum_{i=1}^{n} \la_i$.
Now we are ready to state our main theorem.
\Th \label{th:main theorem} \hfill
\bna
\item Let $M$ be an irreducible highest weight $\Uq$-module with highest weight
$\la \in \La^{+}$. Then there exists a crystal basis $(L, B, l_{B})$
of $M$ such that
\bni
\item $B_{\la} = \{b_{\la} \}$,
\item $B$ is connected.
\end{enumerate}
Moreover, such a crystal basis is unique up to an automorphism of
$M$. In particular, $B$ depends only on $\la$ as an abstract
$\qn$-crystal and we write $B=B(\la)$.
\item The $\qn$-crystal $B(\la)$ has a unique highest weight vector
$b_{\la}$.
\item A vector $b \in \B \otimes B(\la)$ is a highest weight vector
if and only if
$$b = 1 \otimes \tf_{1} \cdots \tf_{j-1} b_{\la}$$
for some $j$ such that $\la + \epsilon_j$ is a strict partition.
\item Let $M$ be a finite-dimensional highest weight $\Uq$-module
with highest weight
$\la \in \La^{+}$.
Assume that $M$ has a crystal basis $(L,B(\la), l_{B(\la)})$
such that $L_{\la} / q L_{\la} = l_{b_{\la}}$.
Then we have
\bni
\item $\V \otimes M = \soplus_{\la + \epsilon_j : \text{strict}}M_ j,$
where $M_j$ is a highest weight $\Uq$-module with highest
weight $\la + \epsilon_j$ and $\dim (M_{j})_{\la + \epsilon_j} = 2\dim M_{\la}$,
\item
if we set $L_{j} = (\mathbf{L} \otimes L) \cap M_{j}$ and
$B_j=\set{b\in\B \otimes B(\la)}{l_b\subset L_j/qL_j}$, then we have
$\B \otimes B(\la)=\coprod\limits_{\la+ \epsilon_j: \text{strict}} B_{j}$
and $L_j/qL_j = \soplus_{b \in B_j} l_{b}$,
\item $M_j$ has a crystal basis $(L_j, B_j, l_{B_j})$,
\item $B_{j}\simeq B(\la + \epsilon_j)$ as an abstract $\qn$-crystal.
\end{enumerate}
\end{enumerate}
\enth
\begin{proof}
We shall argue by induction on $|\la|$.
For a positive integer $k$, we denote by ${\rm (a)}_{k}$,
${\rm(b)}_{k}$, ${\rm(c)}_{k}$ and ${\rm(d)}_{k}$ the assertions
(a), (b), (c) and (d) for $\la$ with $|\la| = k$, respectively.
It is straightforward to check ${\rm (a)}_{1}$ and ${\rm (b)}_{1}$.
Assuming the assertions ${\rm (a)}_{k}$, ${\rm (b)}_{k}$
for $k \leq N$ and the assertions ${\rm (c)}_{k}$, ${\rm (d)}_{k}$ for $k < N $,
let us show ${\rm (a)}_{N+1}$, ${\rm (b)}_{N+1}$,
${\rm (c)}_{N}$ and ${\rm (d)}_{N}$.
\vs{2ex}
\noindent
\textbf{Step 1} :
We shall prove ${\rm (c)}_N$.
Let $\la$ be a strict partition with $|\la|=N$.
By choosing a sequence of strict partitions $\eps_1=\la_1, \
\la_2, \ \la_3, \ldots, \la_N=\la $ such that
$\la_{k+1} = \la_k + \eps_{j_k}$ for some $j_k$ and applying ${\rm (d)}_{k}$ on each $\la_k$ successively for $k < N$,
we can embed $B(\la)$ into $\B^{\tensor N}$.
It follows that $\B \otimes B(\lambda) \subset \B^{\tensor
(N+1)}$.
By ${\rm (b)}_N$, we know that there exists a unique
highest weight vector, say $b_\la$, in $B(\la)$.
By Theorem~\ref{th:char.h.w}, an element $b \in \B \otimes B(\la)$ is a highest weight vector if and only if
$$b = 1 \otimes \tf_{1} \cdots \tf_{j-1} b_{\la}$$
for some $j$ such that $\la+\eps_j$ is a strict partition. So
${\rm (c)}_{N}$ holds.
\vs{2ex}
\noindent
\textbf{Step 2} : We shall show that ${\rm(d)}_N$ holds except (iv).
Let $M$ be a finite-dimensional highest weight module with highest
weight $\la \in \La^{+}$ with $|\la|=N$ and let $(L, B(\la),
l_{B(\la)})$ be a crystal basis of $M$. By Theorem
\ref{th:decomposition}, we have a decomposition
$\V \otimes M = \soplus_{\la +\epsilon_j : \text{strict}} M_ j$, where $M_j$ is a
highest weight $\Uq$-module with highest weight
$\la + \epsilon_j$ and $\dim (M_{j})_{\la + \epsilon_j} = 2 \dim
M_{\la}$.
By Theorem~\ref{th2:tensor product}, $\V
\otimes M$ admits a crystal basis $(\tilde{L},
\ \B \otimes B(\la), \ l_{\B \otimes B(\la)})$
where $\tilde{L}\seteq \mathbf{L} \otimes L$.
Set $L_j \seteq M_j \cap \tilde{L}$.
Note that
$$\F \otimes_\A L_j \isoto M_j \quad \text{and} \quad
L_j = \soplus_{\mu \in P^{\geq 0}} L_j \cap (M_j)_{\mu}. $$
Then we have
$$L_j / q L_j \subset \tilde L / q \tilde L = \soplus_{b \in \B \otimes B(\la)} l_b.$$
Since $\tei (M_j)_{\la+\epsilon_j} = \tilde e_{\ol i} (M_j)_{\la+\epsilon_j}=0 $
for any $i=1,2, \ldots, n-1$ (see Remark~\ref{rem:teibar}),
we have as subspaces of $\tL/q\tL$
$$(L_j)_{\la+\epsilon_j } / q (L_j)_{\la+\epsilon_j } \subset
\Big( \bigcap_{i=1}^{n-1} \Ker \tei \bigcap \bigcap_{i=1}^{n-1} \Ker \tilde e_{\ol i} \Big)_{\lambda+\epsilon_j} =
\soplus _{\stackrel{\wt(b)=\la+\epsilon_j,}{\tei b = \tilde e_{\ol i} b =0 }} l_b = l_{b_{j}},$$
where $b_j = 1 \otimes \tilde f_1 \cdots \tilde f_{j-1} b_{\la}$ in $\B \otimes B(\la)$.
Here, the last equality follows from ${\rm(c)}_{N}$.
Because ${\rm rank}_\A(L_j)_\mu = \dim_\F (M_j)_\mu$ for any $\mu \in \wt(M_j)$, we have
$$
\begin{aligned}
\dim_{\C} \big((L_j)_{\la+\epsilon_j } / q (L_j)_{\la+\epsilon_j } \big)=& {\rm rank}_\A(L_j)_{\lambda+\epsilon_j}
=\dim_\F (M_j)_{\lambda+\epsilon_j} \\
=&2 \dim_{\F} M_{\la} = \dim_{\C} l_{b_j}
\end{aligned}$$
and hence
$$(L_j / q L_j)_{\la + \epsilon_j}=(L_j)_{\la+\epsilon_j } / q (L_j)_{\la+\epsilon_j } = l_{b_j}.$$
Let $B_j$ be the connected component containing $b_j$ in $\B \otimes B(\la)$.
By ${\rm(c)}_N$ and Lemma~\ref{le:existence of
h.w. vectors}, we obtain $\bigcup_j B_j = \B \otimes B(\la)$.
Since $L_j$ is stable under $\tei, \tfi, \teone$ and $\tfone$, we have
$$\soplus_{b \in B_j} l_b \subset L_j / q L_j. $$
It follows that \eq \nonumber \tilde L / q \tilde L = \soplus_{b
\in \B \otimes B(\la)} l_b =\soplus_{b \in \bigcup_j B_j} l_b
\subset \sum_j (L_j/ q L_j).
\eneq
By Nakayama's Lemma, we get \eq \label{eq:nakayama lemma} \tilde L
= \sum_j L_j. \eneq
Since $\sum_j L_j = \soplus\nolimits_j L_j$, we
obtain $\tilde L = \soplus\nolimits_j L_j$ and
\eq \nonumber
\soplus_{b \in \bigcup_j B_j} l_b = \tilde L / q \tilde L \simeq
\soplus_j (L_j/ q L_j) \supseteq \soplus_j \soplus_{b \in B_j} l_{b_j}.
\eneq
Therefore, we obtain
$$ L_j / q L_j = \soplus_{b \in B_j} l_b\quad\text{and}
\quad\B \otimes B(\la) = \coprod_j B_j.$$
Thus $(L_j, \ B_j, \ l_{B_j} =(l_b)_{b \in B_j})$ is a crystal basis of $M_j$.
Note that each $B_j$ has a unique highest weight vector $b_j$ and
that $B_j$ is connected.
\vs{2ex}
\noindent
\textbf{Step 3} :
We will show ${\rm (a)}_{N+1}$.
Since an irreducible highest weight module
is uniquely determined up to parity change, and since the crystal
structure dose not vary under the parity change functor, it is
enough to show that there exists an irreducible highest weight
module with a crystal basis which satisfies (i) and (ii) in ${\rm
(a)}$.
Let $\la$ be a strict partition with $|\la|=N+1$.
Choose a strict partition $\mu$ and $\ell=1,\ldots, n$
such that $\lambda=\mu +\epsilon_\ell$. By ${\rm (a)}_N$,
there exists an irreducible highest weight $\Uq$-module $M$
of highest weight $\mu$ which has a crystal basis $(L, B(\mu),
l_{B(\mu)})$. Then we have
$$\V \otimes M = \soplus_{\la + \epsilon_j : \text{strict}} M_j,$$
and each $M_j$ has a crystal basis as in {\bf Step 2}. Therefore
there exists a finite-dimensional highest weight $\Uq$-module $M$
with highest weight $\la$ which has a crystal basis $(L, B, l_{B})$
such that $B$ is connected and $B_\la=\{b_0\}$.
Moreover we see that in {\bf Step 2}, $B$ has a unique highest weight vector.
By Lemma~\ref{le:crystal bases for irreducible summands}, we conclude
that each irreducible summand of $M$ admits a crystal basis
with the abstract crystal $B$ which satisfies (i) and (ii) in ${\rm (a)}$
and $B$ has a unique highest weight vector.
The uniqueness in ${\rm (a)}_{N+1}$ immediately follows from Lemma~\ref{le:invarianct under tildeki}, Lemma~\ref{le:isomorphic crystals} and Proposition~\ref{prop:uniqueness of lattice of highest weight space of an irreducible module}.
The property ${\rm (b)}_{N+1}$ is obvious.
The remaining (iv) in ${\rm (d)}_{N}$ follows
from Lemma~\ref{le:existence of crystal base for a lattice}.
\end{proof}
\begin{corollary} \hfill
\bna
\item Every $U_q(\mathfrak{q}(n))$-module in the category
$\Oint$ has a crystal basis.
\item If $M$ is a $\Uq$-module in the category
$\Oint$ and $(L,B,l_B)$ is a crystal basis of
$M$, then there exist decompositions $M=\soplus_{a\in A}M_a$ as a
$U_q(\mathfrak{q}(n))$-module, $L=\soplus_{a\in A}L_a$ as an
$\A$-module, $B=\coprod_{a\in A}B_a$ as a $\qn$-crystal,
parametrized by a set $A$ such that the following
conditions are satisfied for any $a\in A$ :
\bni
\item
$M_a$ is a highest weight module with
highest weight $\la_a$ and $B_a \simeq B(\la_a)$ for some strict
partition $\la_a$,
\item
$L_a=L\cap M_a$,
$L_a/qL_a=\soplus_{b\in B_a}l_b$, \item $(L_a,B_a,l_{B_a})$ is a
crystal basis of $M_a$.
\ee
\ee
\begin{proof}
(a) Our assertion follows from the semisimplicity of the category
$\Oint$. Indeed, if $M = \soplus_\nu M_{\nu}$
is a decomposition of $M$ into irreducible $\uqqn$-modules, then
each $M_{\nu}$ is an irreducible highest weight module, and hence it
admits a crystal basis $(L_\nu, B_\nu, l_{B_\nu})$ by Theorem
\ref{th:main theorem}. Set
\begin{equation*}
\begin{array}{ccc}
L \seteq \soplus_{\nu} L_\nu, & B \seteq \coprod_\nu B_\nu, & l_B
\seteq (l_{b})_{b \in B}.
\end{array}
\end{equation*}
Then $(L, B, l_B)$ is a crystal basis of $M$.
\bigskip\noi
(b) Let $\la$ be a maximal element in $\wt(B)=\wt(M)$. Note that if
$\ell(\la)=r$ is odd, then we have the following commutative
diagram (see Remark~\ref{re:Clifford algebras} for notations):
$$
\xymatrix{
\Mod(\A) \ar[r]^(.34){\sim} \ar[d]^{\C \tensor_{\A/q\A} ( - )} & \SMod(\A[C_1,\ldots, C_r]) \ar[d]^{\C \tensor_{\A/q\A} ( - )}\\
\Mod(\C) \ar[r]^(.34){\sim} & \SMod(\C[C_1,\ldots, C_r]),
}
$$
and if $r$ is even, then we have the following commutative diagram:
$$\xymatrix{
\SMod(\A) \ar[r]^(.36){\sim} \ar[d]^{\C \tensor_{\A/q\A} ( - )} & \SMod(\A[C_1,\ldots, C_r]) \ar[d]^{\C \tensor_{\A/q\A} ( - )}\\
\SMod(\C) \ar[r]^(.36){\sim} & \SMod(\C[C_1,\ldots, C_r]).
}
$$
The horizontal arrows are given by
$$K \mapsto V \tensor_\C K$$
for each module $K$ in the left hand side,
where $V$ denotes an irreducible supermodule over $\C[C_1,\ldots,C_r]$.
Let $M^{(\la)}\seteq\uqqn M_\la$ be the isotypic component of $M$
that is a highest weight module of highest weight $\la$.
Let $B_{\la} = \set{b^\nu}{\nu =1,2, \ldots, s}$.
Then we have $L_\la/qL_\la=\soplus\nolimits_{\nu=1}^sl_{b^\nu}$.
Hence one can find an $\A[C_1,\cdots C_r]$-submodule $E_\nu$
of $L_\la$ for each $\nu =1,2, \ldots, s$ such that
$$E_\nu / qE_\nu =l_{b^{\nu}} \quad \text{and} \quad L_\la=\soplus_{\nu =1}^s E_\nu.$$
Setting $M^\nu\seteq\uqqn E_\nu$, we have
$$
M^{(\la)} = \soplus_{\nu =1}^s M^\nu.
$$
By Lemma~\ref{le:existence of crystal base for a lattice}, $M^{\nu}$
has a crystal basis
$$(L(M^\nu), B(\la), (l^\nu_b)_{b \in B(\la)})$$
such that $L(M^\nu)_\la=E_\nu$.
Hence their direct sum
$\soplus_{\nu=1}^s(L(M^\nu), B(\la), (l^\nu_b)_{b \in B(\la)})$ is a crystal basis
of $M^{(\la)}$.
Set $L(M^{(\la)})\seteq M^{(\la)}\cap L$.
Since $L(M^{(\la)})_\la=L_\la=\sum_\nu E_\nu=(\sum_\nu L(M^\nu))_\la$,
Lemma~\ref{le:uniqueness of crystal lattice}
implies $L(M^{(\la)})=\soplus\nolimits_{\nu=1}^sL(M^{\nu})$.
In particular, we have $L(M^\nu)=L\cap M^\nu$, and we can regard
$L(M^\nu)/qL(M^{\nu})$ as a subspace of $L/qL$.
\noi
The set
$\set{b \in B(\la)}{ l^\nu_{b} = l_{b'} \quad \text{for some } b' \in B}$
is stable under the Kashiwara operators and contains $b_\la$,
and hence it coincides with $B(\la)$. Therefore
the map $\phi_\nu \cl B(\la) \to B$ given by $l^\nu_b=l_{\phi_\nu(b)}$ ($b\in B(\la)$)
is injective and commutes with the Kashiwara operators.
Its image $C_\nu$ is therefore the connected component of
$b_\nu$ and we obtain
$$L(M^\nu) / q L(M^\nu) = \soplus_{b \in C_{\nu}} l_{b}.$$
Write $B=B_1 \sqcup B_2$, where $B_1=\coprod_{\nu=1}^s C_\nu$.
Then $(L(M^{(\la)}), B_1, l_{B_1})$ is a crystal basis of
$M^{(\la)}$ and coincides with the direct sum of the crystal bases
$(L(M^\nu), B(\la), l^\nu_{B(\la)})$ of $M^\nu$.
Let $M = M^{(\la)} \oplus \widetilde M$ be the decomposition as a $\uqqn$-module,
and set $\widetilde L \seteq L \cap \widetilde M$.
Set $S \seteq q^{-1}L(M^{(\la)}) \cap \bl(q^{-1} \widetilde L + L(M^{(\la)})\br)$.
Then $S$ is invariant under the Kashiwara operators and $S_\la =
L(M^{(\la)})_\la$. Hence by Lemma~\ref{le:uniqueness of crystal
lattice}, we have $S = L(M^{(\la)})$, which implies $L(M^{(\la)})
\cap (\widetilde{L}+qL(M^{(\la)}))=q L(M^{(\la)})$.
Hence we obtain
\eq
&&\big(L(M^{(\la)}) / q L(M^{(\la)})\big)
\cap \big(\widetilde L / q\widetilde L \big) =0
\quad\text{as a subspace of $L/qL$.}
\label{eq:LMtilde}
\eneq
By comparing dimensions, we have
$$L/q L =\big(L(M^{(\la)}) / q L(M^{(\la)}) \big) \oplus \big(\widetilde
L / q \widetilde L\big).$$ Therefore, by Nakayama's lemma, we obtain
\eq
L = L(M^{(\la)}) + \widetilde L = L(M^{(\la)}) \oplus \widetilde L.
\label{eq:LLL}
\eneq
Now, we shall show
\eq
&&\widetilde L / q \widetilde L = \soplus_{b \in B_2} l_{b}.
\eneq
For $b\in B$, let $P_b \cl L/qL \twoheadrightarrow l_b$ be the canonical
projection. Then, for $i=1,\ldots, n-1,\ol1$ satisfying $\te_i b \in B$,
we have a commutative diagram
$$\xymatrix{
L/qL\ar[r]^{{\te_i}}\ar[d]^{P_b}&L/qL\ar[d]^{P_{\te_ib}}\\
l_b\ar[r]^-\sim_-{\te_i}&\;l_{\te_ib}\;.
}
$$
Hence $P_{\te_ib}(\tL/q\tL)=0$ implies $P_b(\tL/q\tL)=0$.
Similarly, $P_{\tf_ib}(\tL/q\tL)=0$ implies
$P_b(\tL/q\tL)=0$.
Hence $S\seteq\{b\in B_1\,;\,P_b(\tL/q\tL)=0\}$
is stable under the Kashiwara operators.
Since $S_\la=B_\la$, we obtain $S=B_1$.
Hence $\tL/q\tL\subset \soplus_{b\in B_2}l_b$.
Then \eqref{eq:LLL} implies
the desired result $\tL/q\tL=\soplus_{b\in B_2}l_b$.
Therefore $(\widetilde L, B_2, (l_b)_{b \in B_2})$ is a crystal
basis of $\widetilde M$. Hence
the crystal basis $(L,B, l_B)$ of $M$
is the direct sum of a crystal basis of $\widetilde M$
and crystal bases of $M^\nu$ ($\nu=1,\ldots,s$).
Since $\dim \widetilde M < \dim M$,
our assertion follows by induction on $\dim M$.
\end{proof}
\end{corollary}
\vskip 1cm
|
\section{What is a tatami tiling?}
\input{introduction.tex}
\section{The structure of tatami tilings: $\mathsf T$-diagrams}
\input{structure.tex}
\section{Counting results}
\label{sec:results}
\input{maxmonomers.tex}
\subsection{Square tatami tilings}
\input{Tnnn/Tnnn.tex}
\subsection{Fixed height tatami tilings}
\label{sec:fixedheight}
\input{fixedheight.tex}
\section{More conjectures and further research}
\input{conjectures.tex}
\section{Acknowledgements}
Thanks to Donald Knuth for his comments on an earlier draft of this
paper and to Martin Matamala for pointing out the tomography problem.
\subsection{Rectangular regions}
\begin{conjecture}
For all $d \ge 0$ and $m \ge 1$ there is an $n_0$ such
that, for all $n \ge n_0$.
\begin{align*}
T(n,n+d,m) = T(n_0,n_0+d,m),
\end{align*}
whenever $n(n+d)$ has the same parity as $m$ (otherwise $T(n,n+d,m)
= 0$, by \tref{maxmon}).
\end{conjecture}
Experimentally, it appears that the smallest $n_0$ is $m+d+4$, if
$d\ge 1$.
The easiest case occurs when $d = 0$ and $m = 1$. It is
not hard to show that for all odd $n \ge 3$ we have $T(n,n,1) = 10$
(the single monomer must go at a corner or in the center).
In a subsequent paper we will show that for $m<n$,
\begin{align*}
T(n,n,m) = m2^m +(m+1)2^{m+1},
\end{align*}
whenever $m$ and $n$ have the same parity.
Returning to the subject of generating functions, ignoring signs, it
appears that the denominators of $T_r(z)$ in \sref{fixedheight} are
self-reciprocal. There must be a combinatorial explanation for this.
Similar questions in the non-tatami case are considered in
\cite{ABBP}.
Generating functions also appear in \conref{cyclotomic}, inspired by
conversations with Knuth. Let $T(n,z)$ be the generating polynomial
for the number of $n\times n$ tilings with $n$ monomers and $i$
vertical dimers. Once again, to count such tilings we consider
flipping diagonals, with the added precaution that the sum of the
number of tiles in the flipped diagonals is a given constant. The
relationship between this and subsets of $\{1, \ldots, n\}$ which have
a given sum is detailed in a subsequent publication.
Let $\phi_n(z)$ denote the $n$th cyclotomic polynomial. Recall that
the roots of $\phi_n(z)$ are the primitive roots of unity. One of
their more well-known properties is that
\begin{equation}
1-z^n = \prod_{d \mid n} \phi_d(z).
\end{equation}
Let $S_n(z)$ denote the ordinary generating function of subsets of
$\{1, \ldots, n\}$ which have a given sum. That is, $\langle z^k
\rangle S_n(z)$ is the number of subsets $A$ of \\
$\{1,2,\ldots,n\}$
such that the sum of the numbers in $A$ is $k$. It is not difficult
to see that
\begin{align}
S_n(z) = (1+z)(1+z^2) \cdots (1+z^n) = \prod_{k=1}^n (1+z^k).
\label{eq:Snz}
\end{align}
Let $\nu(n)$ denote the number of 2s in the prime factorization of $n$
and note that
\begin{align*}
1+z^n = \frac{1-z^{2n}}{1-z^n} = \frac{\prod_{d \mid 2n}
\phi_d(z)}{\prod_{d \mid n} \phi_d(z)} = \prod_{\substack{d \mid 2n \\ d
\nmid n}} \phi_d(z) = \prod_{\substack{d \mid n \\ d \text{ odd}}}
\phi_{2^{1+\nu(n)}d}(z).
\end{align*}
When this latter expression is used in $S_n(z)$ some interesting
simplification occurs.
\begin{lemma}
For all $n \ge 1$,
\begin{align*}
S_n(z) = \prod_{j=1}^n \left( \phi_{2j} (z)\right)^{\lfloor \frac{n+j}{2j} \rfloor}
\end{align*}
\label{lem:nj2}
\end{lemma}
\begin{proof}
The index $2j$ will occur for those $k$s in \eref{Snz} for which $j
= 2^{\nu(k)}d$ for some odd $d$ where $d \mid k$. This equation is
satisfied for $k = j,3j,5j,\ldots$. There are $\lfloor (n+j)/(2j)
\rfloor$ such $k$s that are less than or equal to $n$ (See
\fref{floor}).
\end{proof}
\begin{figure}[ht]
\centering
\input{nj2.eps_tex}
\caption{A visual aid for the last line of the proof of \lref{nj2}.
The pink dots represent the sequence, $j,3j,5j, \ldots$, with $ij\le
n$. Adding $j$ to $n$ shows that the number of dots is $\lfloor
(n+j)/(2j) \rfloor$. }
\label{fig:floor}
\end{figure}
\begin{conjecture}
\label{conj:cyclotomic}
The generating polynomial $T(n,z)$ has the factorization
\begin{align*}
T(n,z) = P(n,z)\prod_{j\ge 1} S_{\left\lfloor \frac{n-1}{2^j}
\right\rfloor}(z)
\end{align*}
where $P(n,z)$ is an irreducible polynomial.
\end{conjecture}
We return to the topic mentioned in the introduction: Tatami-tilings
of orthogonal regions.
\subsection{Orthogonal regions}
\begin{figure}[ht]
\centering
\subfloat[]{
\includegraphics[scale=0.45]{TatamiPuzzle2.eps}
\label{fig:solution:a}
}
\subfloat[]{
\includegraphics[scale=0.45]{TatamiPuzzle3.eps}
\label{fig:solution:b}
}
\caption{(a) The solution to the question posed in \fref{driveway}; no
monomers are required to tatami tile the region. (b) A legal
configuration of six magnetic water striders in an orthogonal
``pond''. Note that no further striders may be added.}
\label{fig:solution}
\end{figure}
We believe that the main structural components are the same as they
were for rectangles, but there are a few subtleties to be clarified at
inside corners, since a ray could begin at such a place.
What is the computational complexity of determining the least number
of monomers that can be used to tile an orthogonal region given the
segments that form the boundary of the region and the unit size of
each dimer/monomer? In the rectangular grid this is answerable in
polynomial time using $\mathsf T$-diagrams, however, it appears to be NP-hard
for an arbitrary number of segments.
The problem of minimizing the number of monomers in a tiling inspires
what we call the ``magnetic water strider problem''. This time the
orthogonal region is a pond populated by water striders. A water
strider is an insect that rides atop water in ponds by using surface
tension. Its 4 longest legs jut out at 45 degrees from its body. In
the fancifully named magnetic water strider problem, we require the
body to be aligned north-south. Furthermore its legs support it, not
by resting on the water, but by extending to the boundary of the pond.
Naturally, the legs of the striders are not allowed to intersect. A
legal configuration of magnetic water striders in an orthogonal pond
is shown \fref{solution:b}.
There are two problems and a game here. The first is a packing
problem: What is the largest number of magnetic waters striders that a
pond can support? On the other hand, one can ask what is the minimum
number that can be placed so that no more can be added. Placing and
packing striders can be tricky, which gives rise to an adversarial
game where players take turns placing striders in an orthogonal
region. Brian Wyvill has kindly implemented a version of this game,
available at \url{http://www.theory.cs.uvic.ca/~cos/tatami/}.
Interpreted as a matching problem on a subgraph of a grid graph $G$, a
tatami tiling is a matching $M$ with the property that $G-M$ contains
no 4-cycles. Note that there is always such a matching (e.g., take
the ``running bond'' layout on the infinite grid graph and then
restrict it to $G$). However, if we insist on a perfect matching,
then the problem is equivalent to our ``perfect'' driveway paving
problem from the introduction.
More generally, a matching whose removal destroys $k$-cycles is called
\emph{$C_k$-transverse}. Ross Churchley proved that finding a
$C_k$-transverse matching in an arbitrary graph is NP-hard when $k\ge 4$
(private communication \cite{churchley}).
\subsection{Combinatorial games}
Consider the following game. Given an orthogonal region, players take
turns placing dimers (or dimers and monomers); each placement must
satisfy the tatami constraint and the last player who can move wins.
This game, called Oku!, is reminiscent of the game called Nimm, in
which players also win by making the last move, however a winning
strategy for our game is unknown and there are grid sizes in which the
second player can force a win. The name is a phonetic spelling of the
Japanese word for ``put''.
Another game applies tomography to rectangular tilings.
Tiling tomography is a rich and open area of complexity theory to
which a good introduction can be found in \cite{chrobak}. The
relevant question is as follows: Given $r+c$ triples of numbers
$(h,v,m)$, one for each row and one for each column, is there a tatami
tiling which has $h$ horizontal dimers, $v$ vertical dimers, and $m$
monomers in the respective row or column?
Without the tatami condition this decision problem is NP-hard (Theorem
4, \cite{DGM}). Hard or not, the tatami condition gives considerable
information in practice, however, making the reconstruction of a
tatami tiling an entertaining challenge. Erickson, A., has created an
online computer game out of this called Tomoku. It is playable at
\url{http://tomokupuzzle.com}, complete with music, countdown timers
and high scores.
\begin{figure}[ht]
\centering
\includegraphics[width = 0.7\textwidth]{tomoku.eps}
\caption{The Tomoku web game. The player is shown which tiles are
completely contained in each row and column, and the object is to
reconstruct the tiling. Note that each monomer appears twice in the
projections.}
\label{fig:tomoku}
\end{figure}
Both of these games are at
\url{http://www.theory.cs.uvic.ca/~cos/tatami/}, and it should be
noted that they can also be played with a pencil and paper.
\subsection{Outline}
In Section 2 we determine the structure of tatami tilings in a
rectangle. Our structural characterization has important algorithmic
implications, for example, it reduces the size of the description of a
tiling from $\Theta(rc)$ to $O(\max\{r,c\})$ and may be used to
generate tilings quickly. The three theorems in \sref{results} are
the main results of the paper and are also stated here. The first of
these concerns the maximum possible number of monomers. Let
$T(r,c,m)$ be the number of tilings of the $r\times c$ grid, with $m$
monomers (and the other tiles being horizontal or vertical dimers).
\begin{varthm}
If $T(r,c,m)>0$, then $m$ has the same parity as $rc$ and
$m\le\max(r+1,c+1)$.
\end{varthm}
Following this we prove a counting result for
maximum-monomer tilings of square grids.
\begin{varthm}
The number of $n\times n$ tilings with $n$ monomers, $n2^{n-1}$.
\end{varthm}
Our final result concerns fixed-height tilings with an unrestricted number of monomers.
\begin{varthm}
For a fixed number of rows $r$, the ordinary generating function of
the number of tilings of an $r \times n$ rectangle is a rational
function.
\end{varthm}
We also provide an algorithm which outputs this generating function
for a given $r$ and explicitly give the generating function for $r=1,
2$ and $3$, along with the coefficients of the denominator for $1 \le
r \le 11$. In Section 4 we return to the question of tatami tiling
general orthogonal regions and introduce the ``magnetic water strider
problem'' along with additional conjectures and open problems.
|
\section{Introduction}
\label{IN}
The primary goal of this paper is the study of the behavior, as $\varepsilon \to 0$, of the solution $u^\varepsilon=u^\varepsilon(x,t,\omega)$ of the initial value problem
\begin{equation} \label{HJ}
\left\{ \begin{aligned}
&u^\varepsilon_t - \varepsilon \tr \!\left( A(\tfrac x\varepsilon, \omega) D^2u^\varepsilon \right) + H(Du^\varepsilon, \tfrac x\varepsilon, \omega) = 0 & \mbox{in} & \ \ensuremath{\mathds{R}^\d} \times \ensuremath{\mathds{R}}_+, \\
& u^\varepsilon = u_0 & \mbox{on} & \ \ensuremath{\mathds{R}^\d} \times \{ 0 \}.
\end{aligned} \right.
\end{equation}
Here $A=A(y,\omega)$ and $H=H(p,y,\omega)$ are random processes as they depend on $\omega$, an element of an underlying probability space $(\Omega, \mathcal F,\ensuremath{\mathds{P}})$, and the initial datum $u_0$ is a bounded, uniformly continuous function on $\ensuremath{\mathds{R}^\d}$. Precise notation and hypotheses on the coefficients are given in the following section, but we mention here that the diffusion matrix $A$ is degenerate elliptic and the Hamiltonian $H$ is convex in $p$ and, in general, unbounded from below, both in $y$ for fixed $(p,\omega)$ and in $\omega$ for fixed $(p,y)$. The random coefficients $A$ and $H$ are required to be stationary and ergodic, and, furthermore, the unboundedness of $H$ is controlled by a random variable which is assumed to be strongly mixing.
Model equations satisfying our hypotheses include the viscous Hamilton-Jacobi equation
\begin{equation} \label{PCP}
u_t^\varepsilon - \varepsilon \Delta u^\varepsilon + |Du^\varepsilon|^\gamma - V(\tfrac x\varepsilon, \omega) = 0,
\end{equation}
and the first-order Hamilton-Jacobi equation
\begin{equation} \label{PCPnv}
u_t^\varepsilon + |Du^\varepsilon|^\gamma - V(\tfrac x\varepsilon, \omega) = 0,
\end{equation}
where $\gamma > 1$ and $V$ is, for example, a Poissonian potential.
\medskip
The first of the main results (Theorem~\ref{MAIN}, which is stated in Section~\ref{MR}) of the paper is the identification of a deterministic effective nonlinearity $\overline H=\overline H(p)$ combined with the assertion that, as $\varepsilon \to 0$, the solutions $u^\varepsilon$ of \eqref{HJ} converge, locally uniformly in $(x,t)$ and almost surely in $\omega$, to a deterministic function $u=u(x,t)$, the unique solution of the initial-value problem
\begin{equation} \label{HJ-eff}
\left\{ \begin{aligned}
&u_t +\overline H(Du) = 0 & \mbox{in} & \ \ensuremath{\mathds{R}^\d} \times \ensuremath{\mathds{R}}_+, \\
& u = u_0 & \mbox{on} & \ \ensuremath{\mathds{R}^\d} \times \{ 0 \}.
\end{aligned} \right.
\end{equation}
To obtain almost sure convergence, we study the asymptotic behavior of the special problem
\begin{equation} \label{meteqn}
\left\{ \begin{aligned}
& -\varepsilon \tr\!\left(A(\tfrac x{\varepsilon},\omega)D^2m_\mu^\varepsilon\right) + H(p+Dm_\mu^\varepsilon,\tfrac x{\varepsilon},\omega) = \mu & \mbox{in} & \ \ensuremath{\mathds{R}^\d} \! \setminus \! D_\varepsilon, \\
& m_\mu^\varepsilon = 0 & \mbox{on} & \ \partial D_\varepsilon, \\
& \lefteqn{0 \leq \liminf_{|x|\to\infty} |x|^{-1} m_\mu^\varepsilon(x,\omega)}
\end{aligned} \right.
\end{equation}
where, depending on some parameters in the equation (e.g., the growth of $H$), either $D_\varepsilon=\{ 0\}$ or $D_\varepsilon = B_\varepsilon$. The problem \eqref{meteqn}, which we call the \emph{metric problem}, is of independent interest. We prove that it has a unique solution for each $\mu > \overline H(p)$ and no solution for $\mu < \overline H(p)$, thereby characterizing the effective Hamiltonian $\overline H$ in a new way. We also provide a new characterization of the $\min \overline H$. The functions $m^\varepsilon_\mu$ possess a subadditivity property sufficient to apply the subadditive ergodic theorem and obtain, in the limit $\varepsilon \to 0$, the almost sure convergence of $m_\mu^\varepsilon$ to the unique solution $\overline m_\mu$ of
\begin{equation} \label{meteqnbar}
\left\{ \begin{aligned}
& \overline H(p+D\overline m_\mu) = \mu \quad \mbox{in} \ \ensuremath{\mathds{R}^\d} \! \setminus \! \{ 0 \}, \\
& \overline m_\mu(0) = 0, \\
& 0 \leq \liminf_{|y|\to\infty} |y|^{-1} \overline m_\mu(y).
\end{aligned} \right.
\end{equation}
It turns out that the almost sure behavior of the solutions $u^\varepsilon$ of \eqref{HJ} are controlled by $m^\varepsilon_\mu$ and the characterization of $\min \overline H$, and this allows us to deduce the convergence of $u^\varepsilon$ almost surely.
\medskip
There has been much recent interest in the homogenization of partial differential equations in stationary ergodic random environments. While the linear case was settled long ago by Papanicolaou and Varadhan \cite{PV1,PV2} and Kozlov \cite{K}, and general variational problems were studied by Dal Maso and Modica \cite{DM2,DM1} (see also Zhikov, Kozlov, and Ole{\u\i}nik \cite{ZKO}), it was only relatively recently that nonlinear problems were considered (in bounded environments). Results for stochastic homogenization of Hamilton-Jacobi equations were first obtained by Souganidis \cite{S} (see also Rezakhanlou and Tarver \cite{RT}), and for viscous Hamilton-Jacobi equations by Lions and Souganidis \cite{LS2} and Kosygina, Rezakhanlou, and Varadhan \cite{KRV}. The homogenization of these equations in spatio-temporal media was studied by Kosygina and Varadhan \cite{KV} and Schwab \cite{Sch}. Davini and Siconolfi~\cite{DS} also proved, using some connections to Mather theory, some interesting results for Hamilton-Jacobi equations in bounded environments. We also mention the work of Caffarelli, Souganidis and Wang \cite{CSW} on the stochastic homogenization of uniformly elliptic equations of second-order, Caffarelli and Souganidis \cite{CS} who obtained a rate of convergence for the latter in strongly mixing environments, and Schwab~\cite{Sch2} on the homogenization of nonlocal equations. A new proof of the results of \cite{KRV,LS2,RT,S}, which yields only convergence in probability but does not rely on formulae, has recently been found by Lions and Souganidis \cite{LS3}. We adapt the approach of \cite{LS3} to our setting and upgrade the convergence to almost sure, using the ``metric problem'' introduced in Section~\ref{Q}.
\medskip
As far as unbounded environments are concerned, Sznitman \cite{Sz1,Sz2,Sz3, Szb} studied the behavior of Brownian motions in the presence of Poissonian obstacles and obtained very elegant and complete results concerning the asymptotic behavior and large deviations of such processes, both in the quenched and annealed settings
Although the results of Sznitman were not stated in terms of homogenization theory, some of them may be formulated in terms of the stochastic homogenization of \eqref{PCP} for $\gamma =2$. Part of our work can be seen as an extension of Sznitman's results to more general equations, including degenerate Bellman equations and Hamilton-Jacobi equations, as well as to more general random environments. In particular, we consider equations which cannot be written as a linear Schr\"odinger equation via the Hopf-Cole transformation, and general mixing rather than i.i.d. environments.
More recently, and after this paper was accepted, Rassoul-Agha, Sepp\"al\"ainen and Yilmaz posted a paper \cite{RSY} to the arXiv which considers quenched free energy and large deviations for a random walk in a random (unbounded) potential. This corresponds to a ``discrete'' version of the work of Sznitman, but with more general potentials, and to ours, in the case $\gamma=2$. The potential in \cite{RSY} is assumed to satisfy a strong mixing condition in order to compensate for its unboundedness, which is very similar (although not identical) to some of our hypotheses below.
\medskip
We remark that problems depending on the macroscopic variable~$x$ such as
\begin{equation} \label{HJx}
\left\{ \begin{aligned}
&u^\varepsilon_t - \varepsilon \tr \!\left( A(x,\tfrac x\varepsilon, \omega) D^2u^\varepsilon \right) + H(Du^\varepsilon,x, \tfrac x\varepsilon, \omega) = 0 & \mbox{in} & \ \ensuremath{\mathds{R}^\d} \times \ensuremath{\mathds{R}}_+, \\
& u^\varepsilon = u_0 & \mbox{on} & \ \ensuremath{\mathds{R}^\d} \times \{ 0 \},
\end{aligned} \right.
\end{equation}
can be handled by the methods in this paper, leading to an effective equation depending also on $x$. This requires proving a continuous dependence estimate (which is more sophisticated than Proposition~\ref{CDE-noxd}), so that we can work with a countable number of subsets of $\Omega$. This can be achieved by combining standard viscosity theoretic arguments (c.f. \cite{CIL}) with our Lemma~\ref{uppbndosc}. However, for clarity, we confine our attention in this paper to \eqref{HJ}, and pursue such generalizations in a future work.
\medskip
While in this work we focus on equations such as \eqref{PCP} and \eqref{PCPnv} with $V$ unbounded, our proof strategy can be applied to viscous and non-viscous Hamilton-Jacobi equations with other kinds of ``degenerate coercivity." For example, with minor modifications, our techniques can handle equations such as
\begin{equation} \label{perc}
u^\varepsilon_t - \varepsilon \delta \Delta u^\varepsilon + a\big( \tfrac x\varepsilon, \omega) |Du^\varepsilon|^\gamma = 1,
\end{equation}
where $\delta\in \{ 0,1\}$, and the function $a(y,\omega) > 0$ is stationary ergodic, but not necessarily bounded below. This lack of coercivity can be compensated for by adding a strong mixing hypothesis to $a$ as well as the assumption that a large moment of $a^{-1}$ is bounded, hypotheses which are similar to those imposed on $V$ in our model equation \eqref{PCP} and \eqref{PCPnv}. The problem \eqref{perc} with $\delta = 0$ corresponds to a continuum model for \emph{first passage percolation}, where the medium imposes a cost $a^{-1}(y,\omega)$ to travel near $y$, which may be arbitrarily large. See Kesten~\cite{Ke}.
\medskip
Finally we note that our proofs also yield the homogenization of of time independent versions of \eqref{HJ} and \eqref{HJx}, i.e., to equations of the general form
\begin{equation}\label{hjx}
u^\varepsilon - \varepsilon \tr \!\left( A(x,\tfrac x\varepsilon, \omega) D^2u^\varepsilon \right) + H(Du^\varepsilon,x, \tfrac x\varepsilon, \omega) = 0 \quad \mbox{in} \ U\end{equation}
for a domain $U$ with appropriate boundary conditions.
\medskip
The paper is organized as follows. In the next section, we review the notation, introduce the precise assumptions, and consider some motivating examples. In Section~\ref{MR} we state the main result and give an overview of its proof, which is completed in Section~\ref{PH}. A key ingredient of the proof is the study of the macroscopic (cell) problem, which is the content of Section~\ref{AMP}. The effective Hamiltonian is identified and its basic properties are studied in Section~\ref{EH}. In Section~\ref{Q} we consider the metric problem \eqref{meteqn} and obtain almost sure homogenization with the help of the subadditive ergodic theorem. In Section~\ref{P} we compare our results for \eqref{PCP} and $\gamma=2$ with those of Sznitman, explain some of the connections to probability theory, and give a new characterization of $\overline H$ in terms of the solvability metric problem. Several auxiliary lemmata are recorded in Appendix~\ref{A}.
\section{Preliminaries}
\label{PL}
We briefly review the notation and state precisely the assumptions on \eqref{HJ} before discussing some motivating examples.
\subsection{Notation and conventions}
The symbols $C$ and $c$ denote positive constants, which may vary from line to line and, unless otherwise indicated, do not depend on $\omega$. We work in the $\d$-dimensional Euclidean space $\ensuremath{\mathds{R}^\d}$ with $\d \geq 1$, and we write $\ensuremath{\mathds{R}}_+:=(0,\infty)$. The sets of rational numbers and positive integers are denoted respectively by $\ensuremath{\mathds{Q}}$ and $\ensuremath{\mathds{N}}$, while $\ensuremath{\mathcal{M}}^{m\times \d}$, and $\ensuremath{\mathcal{S}^\d} \subseteq \ensuremath{\mathcal{M}}^{\d\times \d}$ are respectively the sets of $m$-by-$\d$ matrices and $\d$-by-$\d$ symmetric matrices. If $v,w\in \ensuremath{\mathds{R}^\d}$, then $v\otimes w \in \ensuremath{\mathcal{S}^\d}$ is the symmetric tensor product which is the matrix with entries $\frac12(v_iw_j+ v_jw_i)$. For $y \in \ensuremath{\mathds{R}^\d}$, we denote the Euclidean norm of $y$ by $|y|$, while if $M\in \ensuremath{\mathcal{M}}^{d\times \d}$, $M^t$ is the transpose of $M$, $\tr(M)$ is the trace of $M$, and we write $|M| := \tr(M^t M)^{1/2}$. The identity matrix is $I_\d$. If $U \subseteq\ensuremath{\mathds{R}^\d}$, then $|U|$ is the Lebesgue measure of $U$. Open balls are written $B(y,r): = \{ x\in \ensuremath{\mathds{R}^\d} : |x-y| < r\}$, and we set $B_r : = B(0,r)$. The distance between two subsets $U,V\subseteq \ensuremath{\mathds{R}^\d}$ is denoi ted by $\dist(U,V) = \inf\{ |x-y|: x\in U, \, y\in V\}$. If $U\subseteq \ensuremath{\mathds{R}^\d}$ is open, then $\USC(U)$, $\LSC(U)$ and $\BUC(U)$ are the sets of upper semicontinuous, lower semicontinuous and bounded and uniformly continuous functions $U\to \ensuremath{\mathds{R}}$, respectively. If $f:U\to \ensuremath{\mathds{R}}$ is integrable, then $\fint_U f \, dy$ is the average of $f$ over $U$. If $f:U \to \ensuremath{\mathds{R}}$ is measurable, then we set $\osc_U f:= \esssup_U f - \essinf_U f$. The Borel $\sigma$-field on $\ensuremath{\mathds{R}^\d}$ is denoted by $\mathcal{B}(\ensuremath{\mathds{R}^\d})$. If $s,t\in \ensuremath{\mathds{R}}$, we write $s\wedge t : = \min\{ s,t\}$.
\medskip
We emphasize that, throughout this paper, all differential inequalities involving functions not known to be smooth are assumed to be satisfied in the viscosity sense. Wherever we refer to ``standard viscosity solution theory" in support of a claim, the details can always be found in the standard reference \cite{CIL}.
We abbreviate the phrase \emph{almost surely in} $\omega$ by ``a.s. in $\omega$." To keep the bookkeeping simple, $\Omega_1\supseteq\Omega_2 \supseteq \Omega_3 \supseteq \cdots$ is a decreasing sequence of subsets of $\Omega$ of full probability used to keep track of almost sure statements. Roughly, $\Omega_j$ is the subset of $\Omega$ of full probability on which all the almost sure statements appearing in the paper prior to the introduction of $\Omega_j$ hold.
\subsection{The random medium}
The random environment is described by a probability space $(\Omega, \mathcal F, \mathds P)$. A particular ``medium" is an element $\omega \in\Omega$. The probability space is endowed with an ergodic group $(\tau_y)_{y\in \ensuremath{\mathds{R}^\d}}$ of $\mathcal F$-measurable, measure-preserving transformations $\tau_y:\Omega\to \Omega$. Here \emph{ergodic} means that, if $D\subseteq \Omega$ is such that $\tau_z(D) = D$ for every $z\in \ensuremath{\mathds{R}^\d}$, then either $\ensuremath{\mathds{P}}[D] = 0$ or $\ensuremath{\mathds{P}}[D] = 1$. An $\mathcal F$-measurable process $f$ on $\ensuremath{\mathds{R}^\d} \times \Omega$ is said to be \emph{stationary} if the law of $f(y,\cdot)$ is independent of $y$. This is quantified in terms of $\tau$ by the requirement that
\begin{equation*}
f(y,\tau_z \omega) = f(y+z,\omega) \quad \mbox{for every} \ y,z\in \ensuremath{\mathds{R}^\d}.
\end{equation*}
Notice that if $\phi:\Omega \to S$ is a random process, then $\tilde \phi(y,\omega) : = \phi(\tau_y\omega)$ is stationary. Conversely, if $f$ is a stationary function on $\ensuremath{\mathds{R}^\d} \times \Omega$, then $f(y,\omega) = f(0,\tau_y\omega)$.
The expectation of a random variable $f$ with respect to $\mathds P$ is written $\ensuremath{\mathds{E}} f$, and we denote the variance of $f$ by $\Var(f): = \ensuremath{\mathds{E}}(f^2) - (\ensuremath{\mathds{E}} f)^2$. If $E \in \mathcal F$, then $\mathds{1}_E$ is the indicator random variable for $E$; i.e., $\mathds{1}_E(\omega) = 1$ if $\omega\in E$, and $\mathds{1}_E(\omega) = 0$ otherwise.
\medskip
Many times in this paper, we rely on the following multiparameter ergodic theorem, a proof of which can be found in Becker~\cite{B}.
\begin{prop} \label{ergthm}
Suppose that $f:\ensuremath{\mathds{R}^\d} \times \Omega \to \ensuremath{\mathds{R}}$ is stationary and $\ensuremath{\mathds{E}} |f(0,\cdot)| < \infty$. Then there is a subset $\widetilde \Omega \subseteq \Omega$ such that ${\mathds P}[\widetilde \Omega] =1$ and , for each bounded domain $V \subseteq \ensuremath{\mathds{R}^\d}$ and $\omega \in \widetilde \Omega$,
\begin{equation*}
\lim_{t\to \infty} \fint_{tV} f(y,\omega) \, dy = \ensuremath{\mathds{E}} f.
\end{equation*}
\end{prop}
We also make use of a subadditive ergodic theorem, the statement of which requires some additional notation. Let $\mathcal{I}$ denote the class of subsets of $[0,\infty)$ which consist of finite unions of intervals of the form $[a,b)$. Let $\{\sigma_t\}_{t\geq 0}$ be a semigroup of measure-preserving transformations on $\Omega$. A \emph{continuous subadditive process} on $(\Omega, \mathcal F, \ensuremath{\mathds{P}})$ with respect to $\sigma_y$ is a map
\begin{equation*}
Q: \mathcal I \rightarrow L^1(\Omega,\ensuremath{\mathds{P}})
\end{equation*}
which satisfies the following three conditions:
\begin{enumerate}
\item[(i)] $Q(I)(\sigma_t\omega) = Q( y+I)(\omega)$ for each $t>0$, $I \in \mathcal I$ and a.s. in $\omega$.
\item[(ii)] There exists a constant $C> 0$ such that, for each $I\in \mathcal I$,
\begin{equation*}
\ensuremath{\mathds{E}} \big| Q(I) \big| \leq C |I|.
\end{equation*}
\item[(iii)] If $I_1,\ldots I_k \in \mathcal I$ are disjoint and $I=\cup_j I_j$, then
\begin{equation*}
Q(I) \leq \sum_{j=1}^k Q(I_j).
\end{equation*}
\end{enumerate}
A proof of the following version of the subadditive ergodic theorem can be found in Akcoglu and Krengel~\cite{AK}.
\begin{prop} \label{SAergthm}
Suppose that $Q$ is a continuous subadditive process. Then there is a random variable $a(\omega)$ such that
\begin{equation*}
\frac1t Q([0,t))(\omega) \rightarrow a(\omega) \quad \mbox{a.s. in} \ \omega.
\end{equation*}
If, in addition, $\{ \sigma_t \}_{t>0}$ is ergodic, then $a$ is constant.
\end{prop}
\subsection{Assumptions on the coefficients}
The following hypotheses are in force throughout this article. The environment is given by a probability space $(\Omega, \ensuremath{\mathds{P}},\mathcal F)$, and
\begin{equation} \label{erghyp}
\tau_y : \Omega \to \Omega \mbox{is an ergodic group of measure-preserving transformations}
\end{equation}
as described above.
The following hypotheses we impose on the matrix~$A$ and the Hamiltonian~$H$ are taken to hold for \emph{every} $\omega \in \Omega$, rather than almost surely in $\omega$, since we lose no generality by initially removing an event of probability zero. We require $A$ and $H$ to be functions
\begin{equation*}
A:\ensuremath{\mathds{R}^\d}\times\Omega \to \ensuremath{\mathds{R}} \qquad \mbox{and} \qquad H:\ensuremath{\mathds{R}^\d}\times\ensuremath{\mathds{R}^\d}\times\Omega\to\ensuremath{\mathds{R}},
\end{equation*}
and we require that, for every $p \in \ensuremath{\mathds{R}^\d}$,
\begin{equation} \label{AHstat}
(y,\omega) \mapsto A(y,\omega) \quad \mbox{and} \quad (y,\omega) \mapsto H(p,y,\omega) \quad \mbox{are stationary.}
\end{equation}
We assume that, for each $\omega \in \Omega$,
\begin{equation} \label{AC11}
A(\cdot,\omega) \in C^{1,1}_\mathrm{loc}(\ensuremath{\mathds{R}^\d};\ensuremath{\mathcal{S}^\d}),
\end{equation}
and that $A$ has the form $A = \Sigma \Sigma^t$, where
\begin{equation}\label{Alip}
\Sigma(y, \omega) \in \ensuremath{\mathcal{M}}^{\d \times m} \quad \mbox{satisfies} \quad \| \Sigma(\cdot,\omega) \|_{C^{0,1}(\ensuremath{\mathds{R}^\d})} \leq C.
\end{equation}
As far as the Hamiltonian $H$ is concerned, we assume that, for each $(y,\omega)\in \ensuremath{\mathds{R}^\d}\times \Omega$,
\begin{equation} \label{Hconv}
p \mapsto H(p,y,\omega) \quad \mbox{is convex,}
\end{equation}
and that there exists $1 < \gamma < \infty$ such that, for every $(p,y,\omega) \in\ensuremath{\mathds{R}^\d}\times\ensuremath{\mathds{R}^\d}\times\Omega$,
\begin{equation} \label{Hbound}
H(p,y,\omega) \leq C \left(1+|p|^\gamma \right).
\end{equation}
The coercivity and Lipschitz regularity in $y$ of the Hamiltonian $H$ are governed by a nonnegative stationary process $V=V(y,\omega)$, as follows. We assume that there exist constants $c_0, C_0>0$ such that, for every $(p,y,\omega) \in \ensuremath{\mathds{R}^\d}\times \ensuremath{\mathds{R}^\d} \times \Omega$,
\begin{equation} \label{Hcoer}
H(p,y,\omega) \geq c_0 |p|^\gamma - V(y,\omega) - C_0,
\end{equation}
and we require that, for every $(p_1,y_1), (p_2, y_2) \in \ensuremath{\mathds{R}^\d} \times \ensuremath{\mathds{R}^\d}$ and $\omega \in \Omega$,
\begin{multline} \label{Hlip}
\left| H(p_1,y_1,\omega) - H(p_2,y_2,\omega) \right|\\ \leq C \Big( \left(1+ |p_1|+|p_2| \right)^{\gamma-1} |p_1-p_2| + \Big( \sup_{[y_1,y_2]} V(\cdot,\omega)\Big) |y_1-y_2| \Big).
\end{multline}
For fixed positive constants $\alpha, \beta > 0$, satisfying
\begin{equation} \label{alphabeta}
\alpha > \d \qquad \mbox{and} \qquad \beta > \d + \frac{4\d^2}{\gamma\alpha},
\end{equation}
the process $V=V(y,\omega) \geq 0$ is assumed to possess the following two properties:
\begin{equation} \label{Vmoment}
\ensuremath{\mathds{E}}[\sup_{y\in B_1} V(y,\cdot)^\alpha] < \infty,
\end{equation}
and
\begin{equation} \label{Vmix}
V \ \mbox{is strongly mixing with respect to} \ \tau \ \mbox{with an algebraic rate of} \ \beta.
\end{equation}
The mixing condition is an assumption on the rate of decay of the correlation between the random variables $V(y,\omega)$ and $V(z,\omega)$ as $|y-z|$ becomes large. To state this precisely, for each open set $U\subseteq \ensuremath{\mathds{R}^\d}$, we denote by $\mathcal{G}(U)$ the $\sigma$-algebra generated by the sets $\{ \omega : V(\omega,y) \in [0,a)\}$ ranging over $y\in U$ and $a>0$. The assumption \eqref{Vmix} is then written precisely as
\begin{equation} \label{mixing}
\sup\left\{ \left| \ensuremath{\mathds{P}}(E\cap F) - \ensuremath{\mathds{P}}(E)\ensuremath{\mathds{P}}(F) \right| :\ \dist(U,V) \geq r, \ E\in \mathcal{G}(U), \ F\in \mathcal{G}(V) \right\} \leq Cr^\beta.
\end{equation}
Observe that our hypotheses in the case of the model equations \eqref{PCP} and \eqref{PCPnv} allow the potential $V$ to be unbounded from above, but require it to be bounded from below. There is a very good reason for this; see Remark~\ref{invispot-easy} below.
\medskip
We emphasize that the assumptions \eqref{erghyp}, \eqref{AHstat}, \eqref{AC11}, \eqref{Alip}, \eqref{Hconv}, \eqref{Hbound}, \eqref{Hcoer}, \eqref{Hlip}, \eqref{alphabeta}, \eqref{Vmoment}, and \eqref{mixing} are in force throughout this article.
\medskip
As usual, measurability is an burdensome issue. To avoid an overly pedantic presentation, we are going to suppress demonstrations of the measurability of the various functions and processes we encounter in this paper. For the benefit of the concerned reader, however, we sketch here a proof that all of our random variables are measurable. First, the functions of which we need to check the measurability solve equations with random coefficients for which we have uniqueness. In fact, these solutions depend continuously on the coefficients, and the coefficients are measurable functions of $\omega$ combined with the other variables. Therefore, as continuous functions of measurable functions, they are measurable.
\subsection{Cloud point potentials}
We present a simple example of a model equation that fits our framework. Take $H$ to be a Hamiltonian of the form \begin{equation*}H(p,y,\omega) = \widetilde H(p) - V(y,\omega)\end{equation*} where $\widetilde H$ is an appropriate deterministic function, for instance $\widetilde H(p) = |p|^\gamma$ with $\gamma \geq 1$, and $V$ is a stationary potential satisfying \eqref{Vmoment} and \eqref{Vmix}.
\medskip
We construct some examples of unbounded potentials by first specifying a random process which generates a cloud of points (see Figure~1), and then attach to each point a smooth nonnegative function of compact support. We first briefly recall the notion of a point process on $\ensuremath{\mathds{R}^\d}$, referring to the books \cite{DVJ1,DVJ2} for more details.
\begin{figure}\label{cloudpic}
\centering
\subfigure[A Poisson cloud] {
\label{poissonfig}
\begin{tikzpicture}[scale=6]
\drawaxes;
\draw plot[dots=black] file{poisson.dat};
\end{tikzpicture}
}
\subfigure[A cluster point cloud] {
\label{clusterfig}
\begin{tikzpicture}[scale=6]
\drawaxes;
\draw plot[dots=black] file{cluster.dat};
\end{tikzpicture}
}
\caption{Point processes}
\end{figure}
Consider the set $\Omega$ of locally finite, simple pure point measures on $\ensuremath{\mathds{R}^\d}$. An element of $\Omega$ has the form
\begin{equation} \label{omeform}
\omega = \sum_{j=1}^\infty \updelta_{y_j}
\end{equation}
where the points $y_j$ are distinct and the set $\{ y_j\} \cap B_R$ is finite for each $R> 0$. Here $\updelta_y$ denotes the Dirac measure at $y \in \ensuremath{\mathds{R}^\d}$. Let $\mathcal F$ be the $\sigma$-field generated by the maps $\omega \mapsto \omega(U)$ ranging over all $U \in \mathcal{B}(\ensuremath{\mathds{R}^\d})$. It follows that $E + y :=\{ \omega(\cdot -y) : \omega\in E \}\in \mathcal F$ for any $E\in \mathcal F$.
Fix a smooth $W \in C(\ensuremath{\mathds{R}^\d} ;{\bar \ensuremath{\mathds{R}}}_+)$ with compact support such that $W(0)=1$ and define $V(y,\omega)$, for each $\omega \in \Omega$ of the form \eqref{omeform}, by
\begin{equation} \label{Vexam}
V(y,\omega) : = \int_{\ensuremath{\mathds{R}^\d}} W(y-z) \, d\omega(z) = \sum_{j=1}^\infty W(y-y_j).
\end{equation}
The law of $V$ is inherited from the probability measure $\ensuremath{\mathds{P}}$ we attach to the measurable space $(\Omega,\mathcal F)$. The canonical choice of the probability measure $\ensuremath{\mathds{P}}$ is to take it to be the Poisson law $\ensuremath{\mathds{P}}_\nu$
which is characterized uniquely by the properties:
\begin{enumerate}
\item[(i)] for every $U \in \mathcal B(\ensuremath{\mathds{R}^\d})$, $\ensuremath{\mathds{E}}_\nu[\omega(U) ] = \nu |U|$,
\item[(ii)] if $k\in \ensuremath{\mathds{N}}$ and $U_1, \ldots, U_k \in \mathcal{B}(\ensuremath{\mathds{R}^\d})$ are disjoint, then the random variables $\omega \mapsto \omega(U_j)$ are independent, and
\item[(iii)] $\ensuremath{\mathds{P}}_\nu[ E ] = \ensuremath{\mathds{P}}_\nu[ E+y]$ for each $E \in \mathcal F$.
\end{enumerate}
Under $\ensuremath{\mathds{P}}_\nu$, the canonical process $\omega$ is called a \emph{Poisson point process with intensity} $\nu$. It is easy to see that, with respect to $\ensuremath{\mathds{P}}_\nu$, the potential $V$ given by \eqref{Vexam} is unbounded in $\omega$ but satisfies \eqref{Vmoment} and \eqref{Vmix} for any $\alpha,\beta > 0$. In fact, it follows from (ii) that $V(y,\cdot)$ and $V(z,\cdot)$ are independent, provided that $|y-z| > \diam (\supt(W))$.
There are many examples of stationary point processes one may consider giving rise to potentials $V$ satisfying our hypotheses, and the reader is invited to consult \cite{DVJ1,DVJ2} for many more. We mention one other type, a \emph{cluster process}, in which \emph{cluster center points} are chosen according to a Poisson law and around each cluster center point a certain (possibly random) number of points are placed according to a predetermined distribution. For instance, in Figure~\ref{clusterfig} we have placed a Poisson number of points around each cluster center point according to a Gaussian distribution. If the clusters are not uniformly bounded, that is, points may land arbitrarily far from their cluster center with positive probability, then the process will not be i.i.d. but may be strongly mixing. While mixing conditions are difficult to check in practice, sufficient conditions for a cluster point process to satisfy a mixing condition such as \eqref{mixing} can be found in Laslett \cite{L}. Likewise, if the function $W$ in \eqref{Vexam} has unbounded support, then the potential $V$ is not i.i.d. but may nonetheless satisfy strong mixing hypothesis, depending on the rate of decay of $W(y)$ for large $|y|$.
\section{The main results and proof overview}
\label{MR}
With the hypotheses stated in the previous section in force, we now state the two main results.
\begin{thm} \label{MAIN}
There exist $\overline H:\ensuremath{\mathds{R}^\d}\to \ensuremath{\mathds{R}}$ and $\Omega_0 \subseteq \Omega$ of full probability such that, for each $\omega \in \Omega_0$ and $u_0\in \BUC(\ensuremath{\mathds{R}^\d})$, the unique solution $u^\varepsilon = u^\varepsilon(\cdot,\omega)$ of \eqref{HJ} given in Proposition~\ref{epWP} converges locally uniformly in $\ensuremath{\mathds{R}^\d} \times \ensuremath{\mathds{R}}_+$, as $\varepsilon \to 0$, to the unique solution $u$ of \eqref{HJ-eff} which belongs to $\BUC(\ensuremath{\mathds{R}^\d}\times [0,T])$ for every $T > 0$.
\end{thm}
\begin{thm}\label{MAIN1}
There exists a subset $\Omega_0 \subseteq \Omega$ of full probability such that, for each $\omega \in \Omega_0$ and $\mu > {\overline H}(p)$, with ${\overline H}:\ensuremath{\mathds{R}^\d}\to \ensuremath{\mathds{R}}$ as in Theorem~\ref{MAIN}, the unique solution $m^\varepsilon_\mu=m^\varepsilon_\mu(\cdot,\omega)$ of \eqref{meteqn} given in Proposition~\ref{metexistence}
converges, as $\varepsilon \to 0$, locally uniformly in $\ensuremath{\mathds{R}^\d}\setminus\{0\}$ to the unique solution ${\overline m}_\mu$ of \eqref{meteqnbar}. Moreover, ${\overline H}(p)$ is the unique constant such that \eqref{meteqnbar} has a unique solution for all $\mu > {\overline H}(p)$ and no solution if
$\mu < {\overline H}(p)$.
\end{thm}
The effective nonlinearity $\overline H$ is identified in Proposition~\ref{mainstep} below. Some of its basic properties are summarized in Proposition~\ref{effHam}. We remark that these properties yield that \eqref{HJ-eff} has a unique bounded uniformly continuous solution on $\ensuremath{\mathds{R}^\d}\times [0,T]$ for any $T> 0$. The well-posedness of \eqref{HJ} is discussed briefly in Section~\ref{PH}.
\medskip
As the proofs of Theorems~\ref{MAIN} and~\ref{MAIN1} are rather lengthy and involved, we devote the remainder of this section to summarizing the primary obstacles and main ideas needed to overcome them. To simplify our exposition, we consider the time-independent problem
\begin{equation} \label{stat}
u^\varepsilon - \varepsilon \tr \!\left( A(\tfrac x\varepsilon, \omega) D^2u^\varepsilon \right) + H(Du^\varepsilon, \tfrac x\varepsilon, \omega) = 0 \quad \mbox{in} \ \ensuremath{\mathds{R}^\d}. \\
\end{equation}
Assume that $u^\varepsilon$ admits the asymptotic expansion
\begin{equation} \label{asyexp}
u^\varepsilon(x,\omega) = u(x) + \varepsilon w(\tfrac x\varepsilon, \omega) + O(\varepsilon^2).
\end{equation}
Inserting this into \eqref{stat} and performing a formal computation, we arrive at
\begin{equation}\label{inserhr}
u - \tr \!\left( A (y, \omega) D^2_y w \right) + H(D_xu + D_yw,y, \omega) = 0.
\end{equation}
If the expression on the left of \eqref{inserhr} is independent of $(y,\omega)$, then we obtain an equation for~ $u$. That is, we suppose that, for each $p\in \ensuremath{\mathds{R}^\d}$, there exists a constant $\overline H = \overline H(p)$ and a function $w=w(y,\omega)$ such that
\begin{equation} \label{cell}
- \tr \!\left( A(y, \omega) D^2_y w \right) + H(p + D_yw,y, \omega) = \overline H(p) \quad \mbox{a.s. in} \ \omega.
\end{equation}
Substituting into \eqref{inserhr}, we thereby obtain the effective equation
\begin{equation} \label{effeq}
u + \overline H(Du) = 0 \quad \mbox{in} \ \ensuremath{\mathds{R}^\d}.
\end{equation}
Returning to \eqref{asyexp}, we see that the convergence $u^\varepsilon \to u$ locally uniformly in $\ensuremath{\mathds{R}^\d}$, as $\varepsilon \to 0$, is formally equivalent to $\varepsilon w(\tfrac x\varepsilon, \omega) \to0$, that is, $w$ must be strictly sublinear at infinity:
\begin{equation} \label{sslatinf}
|y|^{-1} w(y,\omega) \rightarrow 0 \quad \mbox{as} \ |y| \to \infty \quad \mbox{a.s. in} \ \omega.
\end{equation}
The PDE approach to homogenization lies in reversing this analysis: it is hoped that by studying the \emph{macroscopic problem} \eqref{cell}, one can justify the convergence of $u^\varepsilon$ to the solution $u$ of the effective equation \eqref{effeq}. A solution $w$ of \eqref{cell} satisfying \eqref{sslatinf} is typically called a \emph{corrector}, since it permits one to remove the influence of microscopic oscillations in the analysis of \eqref{stat}, as predicted by the asymptotic expansion \eqref{asyexp}. The effective Hamiltonian $\overline H$ is thereby identified via a compatibility condition, namely, the solvability of \eqref{cell} subject to condition \eqref{sslatinf}.
\medskip
In the periodic setting, \eqref{cell} is typically referred to as the \emph{cell problem} since it suffices to solve the equation on the unit cube (a single cell) with periodic boundary conditions. In this case, the condition \eqref{sslatinf} is obviously redundant since periodic functions are bounded.
\medskip
Solving the macroscopic problem requires considering an approximate problem, which typically is
\begin{equation} \label{aux}
\delta v^\delta - \tr\!\left( A(y,\omega) D^2_yv^\delta \right) + H(p+D_yv^\delta, y, \omega) = 0.
\end{equation}
Here $\delta > 0$ and $p\in \ensuremath{\mathds{R}^\d}$ are fixed. The auxiliary macroscopic problem \eqref{aux} is not merely an \emph{ad hoc} approximation to \eqref{cell}. Indeed, rescaling the latter by setting $u_\varepsilon(y) : = \varepsilon^{-1} u^\varepsilon(\varepsilon y) - p\cdot y$ and substituting $y=x/\varepsilon$ for the fast variable in \eqref{cell}, one obtains \eqref{stat} with $\delta=\varepsilon$.
In the case that $V$ is bounded, the fact that \eqref{aux} is \emph{proper} (i.e., strictly increasing in its dependence on $v^\delta$) permits one to obtain, from standard viscosity solution theory, a unique bounded solution $v^\delta\in \BUC(\ensuremath{\mathds{R}^\d})$ satisfying
\begin{equation} \label{locestout}
\| \delta v^\delta \|_{L^\infty(\ensuremath{\mathds{R}^\d})} + \| Dv^\delta \|_{L^\infty(\ensuremath{\mathds{R}^\d})} \leq C.
\end{equation}
The stationarity of $v^\delta$ is immediate from uniqueness. We then define $w^\delta (y,\omega) : = v^\delta(y,\omega) - v^\delta(0,\omega)$ and attempt to pass to the limits $-\delta v^\delta(0,\omega) \rightarrow \overline H$ and $w^\delta(y,\omega) \rightarrow w(y,\omega)$ as $\delta \to 0$, in the hope of obtaining \eqref{cell} from \eqref{aux}. The uniqueness of $\overline H$ then follows from an application of the comparison principle.
\medskip
A periodic environment possesses sufficient compactness to make this argument rigorous, and, as previously mentioned, \eqref{sslatinf} comes for free since the limit function $w$ is periodic and thus bounded.
To avoid messy measurability issues, in the random setting we must pass to limits in the variables $(y,\omega)$ together. Unfortunately we possess insufficient compactness in $\omega$ to accomplish this without further ado. What is more, unlike the periodic case, we cannot obtain \eqref{sslatinf} from \eqref{locestout}. These are not merely technical issues. In fact, Lions and Souganidis \cite{LS1} have shown that correctors do not exist in the general stationary ergodic case.
\medskip
Hence in the random setting we must concede our attempt to solve the macroscopic problem \eqref{cell} and instead return to the auxiliary problem \eqref{aux}. To obtain the homogenization result, it turns out to be sufficient to find a (necessarily unique) deterministic constant $\overline H$ such that, almost surely, the sequence $\delta v^\delta$ converges to $-\overline H$ uniformly in balls of radius $\sim1/\delta$ as $\delta \to 0$. In Proposition~\ref{bigstep} we show that, for every $r> 0$,
\begin{equation} \label{cplim}
\limsup_{\delta \to 0} \sup_{y\in B_{r/\delta}} \left| \delta v^\delta(y,\omega) + \overline H \right| = 0 \quad \mbox{a.s. in} \ \omega.
\end{equation}
An important idea in the proof of \eqref{cplim}, recently introduced in \cite{LS3}, is the observation that to homogenize it is nearly enough to construct a \emph{subcorrector} $w$ of \eqref{inserhr}, i.e., a subsolution of \eqref{inserhr} which is strictly sublinear at infinity. The subcorrector can be constructed by passing to weak limits in \eqref{aux} along a subsequence $\delta_j \to 0$ and using the convexity of $H$. It then follows from the ergodic theorem and the stationarity of the gradients $Dv^\delta$ and their weak limit that $w$ satisfies \eqref{sslatinf}. The comparison principle permits us to compare $w$ with the full sequence $v^\delta$. Combined with some elementary measure theory this yields that
\begin{equation} \label{inprob}
\ensuremath{\mathds{E}} \big| \delta v^\delta(0,\cdot) + \overline H \big| \rightarrow 0 \quad \mbox{as} \ \delta \to 0.
\end{equation}
Of course, this yields almost sure convergence along some subsequence-- and eventually homogenization almost surely-- but only along this particular subsequence.
Obtaining \eqref{cplim} from \eqref{inprob} along the full sequence $\delta \to 0$, without relying on explicit formulae, is nontrivial and requires some additional estimates and, more importantly, some new ideas.
We begin with the estimates. The first controls the oscillation of $\delta v^\delta$ in balls of radius $\sim1/\delta$. We show that there exists a deterministic constant $C> 0$, depending on $|p|$, such that, for each $y\in \ensuremath{\mathds{R}^\d}$ and $r > 0$,
\begin{equation} \label{preosc}
\limsup_{\delta \to 0} \osc_{B(y/\delta,r/\delta)} \delta v^\delta(\cdot,\omega) \leq Cr \quad \mbox{a.s. in} \ \omega.
\end{equation}
The second controls the size of $\delta v^\delta$ on balls of radius $\sim 1/\delta$, so that for each $y\in \ensuremath{\mathds{R}^\d}$,
\begin{equation}\label{uppbndosc1}
\sup_{R>0}\limsup_{\delta \to 0} \osc_{B(y/\delta,R/\delta)} \delta v^\delta(\cdot,\omega) \leq C \quad \mbox{a.s. in} \ \omega.
\end{equation}
The estimate \eqref{preosc} is useful for $r> 0$ small, while \eqref{uppbndosc1} is typically applied for $R>0$ very large. In the case of bounded $V$, we have global Lipschitz estimates on $v^\delta$ and $L^\infty$-bounds on the $\delta v^\delta$. Therefore \eqref{preosc} and \eqref{uppbndosc1} are immediate. In the unbounded setting, as we will see, proving \eqref{preosc} and \eqref{uppbndosc1} is a more delicate matter.
The next step to obtain \eqref{cplim} is to prove that
\begin{equation} \label{cplim0}
\delta v^\delta(0,\omega) \rightarrow -\overline H \quad \mbox{a.s. in} \ \omega.
\end{equation}
Once this is done, and we describe how it is proved below, we can get the convergence in balls of radius $\sim 1/\delta$ instead of just at the origin by relying again on \eqref{preosc}, a second application of the ergodic theorem and some elementary measure theory.
\medskip
Proving \eqref{cplim0} requires studying the behavior, as $\varepsilon \to 0$, of the solutions $m_\mu^\varepsilon$ of the metric problem \eqref{meteqn}. Incidentally, that $\mu > \overline H(p)$ is necessary and sufficient for the well-posedness of \eqref{meteqn} is itself another new result in the theory of viscosity solutions. As mentioned in the introduction, we use the subadditive ergodic theorem to conclude, after some work, that the functions $m^\varepsilon_\mu$ converge a.s. in $\omega$ and locally uniformly in $\ensuremath{\mathds{R}^\d}$ to solutions $\overline m_\mu$ of \eqref{meteqnbar}, for every $\mu > \overline H(p)$. With this in place, we can prove \eqref{cplim0}, using a new \emph{reverse} perturbed test function argument: if \eqref{cplim0} does not hold along a subsequence, then the convergence of the $m^\varepsilon_\mu$ must fail for every $p \in \ensuremath{\mathds{R}^\d}$ for which $\overline H(p) = \mu$. This argument works for all $p$ for which $H(p) > \min \overline H$, since we need $\mu > \overline H(q)$ for some $q\in \ensuremath{\mathds{R}^\d}$ for the existence of the functions $m^\varepsilon_\mu$. To conclude we need a separate argument for the case $\overline H(p) = \min \overline H$. For this we prove a new characterization of $\min \overline H$ which involves constructing subsolutions of \eqref{aux} which are permitted to grow linearly instead of strictly sublinearly at infinity.
\medskip
Several additional difficulties arise in unbounded environments. Their resolution has new and interesting implications for the theory of viscosity solutions. Firstly, obtaining stationary solutions $v^\delta$ of the auxiliary problem \eqref{aux} as well as the ``metric problem'' is nontrivial, as we see later. In particular, we cannot expect \eqref{aux} to have bounded solutions, and we must also prove a comparison principle which as far as we know is new in the context of viscosity solutions. Secondly, and this is a more serious problem, it is necessary, as already explained earlier, to have an independent of $\delta$ control over the modulus of continuity of $v^\delta$. We show in Lemma~\ref{impest} that
the mixing hypothesis on $V$ yields control of the oscillation of $\delta v^\delta$ on balls of radius $\sim 1/\delta$.
Finally, the absence of uniform Lipschitz estimates on $v^\delta$ causes an additional difficulty in the proof of \eqref{inprob}, using the comparison principle. This is overcome by relying again on the convexity of $H$.
\section{The auxiliary macroscopic problem}
\label{AMP}
We study here the auxiliary macroscopic problem \eqref{aux} setting the stage for the definition of $\overline H$ in the next section. The main goal is to show that \eqref{aux} has a unique (and therefore stationary) solution $v^\delta$, and that the function $x\mapsto \delta v^\delta(x/\delta)$ is uniformly Lipschitz on scales of order $O(1)$ as $\delta \to 0$. If $V$ is bounded, this is straightforward: a unique bounded solution $v^\delta$ exists for each $\delta > 0$, and we can easily obtain global Lipschitz estimates which are uniform in $\delta$; see \cite{LS2}.
The unbounded setting presents difficulties requiring us to utilize hypotheses \eqref{alphabeta}, \eqref{Vmoment} and \eqref{Vmix} on the potential $V$. Firstly, we cannot expect solutions $v^\delta$ of \eqref{aux} to be bounded from above. This complicates the well-posedness of \eqref{aux}, and, in particular, leads to difficulties with uniqueness, which we handle by using the convexity of $H$.
Secondly, the unboundedness of $V$ means that global Lipschitz estimates on $v^\delta$ and $L^\infty$ bounds on $\delta v^\delta$ cannot hold, and therefore obtaining \eqref{preosc} and \eqref{uppbndosc1} is a nontrivial matter. To deal with this difficulty, we use the mixing condition \eqref{Vmix} and some probability. We proceed by obtaining a local estimate on $|Dv^\delta|$ in terms of $V$ using Bernstein's method. By Morrey's inequality, this reduces the issue to controlling the average of a power of $V$ on large balls. The latter is precisely what a mixing condition allows us to estimate.
Finally, since we may only make countable intersects of subsets of $\Omega$, the proof of the homogenization theorem requires an estimate for the dependence of $v^\delta$ on the parameter $p$ in \eqref{aux}. This is dealt with in Lemma~\ref{CDE-noxd}, a by-product of which is the continuity of effective Hamiltonian $\overline H$, as we will see in Section~\ref{EH}. In the case of general $x$-dependent problems like \eqref{HJx}, this issue becomes much more complicated to resolve in the unbounded environment and requires a strengthening of the hypothesis on $V$; in particular we must take $\alpha$ and $\beta$ to be much larger than in \eqref{alphabeta}.
\subsection{The well-posedness of the auxiliary macroscopic problem} \label{AMP-wp}
We prove that, a.s. in $\omega$, \eqref{aux} has a unique bounded from below solution on $\ensuremath{\mathds{R}^\d}$. The first issue is to obtain strictly sublinear decay on subsolutions of \eqref{aux}.
The following lemma provides a local upper bound (depending on $\omega$) for subsolutions of \eqref{aux}. The proof is based on a barrier construction, following \cite{LL,LS2}.
\begin{lem} \label{locest}
Fix $\delta > 0$ and $(p,\omega) \in \ensuremath{\mathds{R}^\d}\times \Omega$. There exists a deterministic constant $C> 0$, independent of $\delta$, such that if $v\in \USC(\bar B_1)$ is a subsolution of \eqref{aux} in $B_1$, then
\begin{equation} \label{locesteq1}
\sup_{y\in B_{1/2}} \delta v(y) \leq \sup_{y\in B_1} V(y,\omega) + C(1+\delta).
\end{equation}
\end{lem}
\begin{proof}
We construct a simple barrier. To the extent that we rely on \eqref{Hcoer}, we may assume that $\gamma \leq 2$. Indeed, if \eqref{Hcoer} holds for $\gamma > 2$, then, for every $(p,y,\omega)$,
\begin{equation*}
H(p,y,\omega) \geq c_0|p|^2 - V(y,\omega) - (c_1+c_0).
\end{equation*}
With $a,b\geq 0$ to be selected below and $\eta := \frac{2-\gamma}{\gamma-1}$, consider the smooth function $w \in C^\infty(B_1)$ defined by
\begin{equation} \label{barr}
w(y) : = \begin{cases}
a + b(1-|y|^2 )^{-\eta} & \mbox{if} \ \gamma < 2, \\
a - b\log(1-|y|^2) & \mbox{if} \ \gamma=2.
\end{cases}
\end{equation}
It follows that
\begin{equation*}
|Dw(y)|^\gamma \geq c b^\gamma |y|^\gamma \!\left(1-|y|^2\right)^{-(\eta+2)},
\end{equation*}
and
\begin{equation*}
\tr\!\left( A(y,\omega) D^2w \right) \leq Cb \!\left( 1 - |y|^2\right)^{-(\eta+2)} \leq Cb \left( |y|^\gamma \!\left( 1 - |y|^2\right)^{-(\eta+2)} + C\right).
\end{equation*}
Inserting $w$ into the left side of \eqref{aux} yields
\begin{multline*}
\delta w - \tr\!\left( A(y,\omega) D^2w\right) + H(p+Dw, y, \omega) \\
\geq (a \delta -C(b+1) - V(y,\omega)) + (cb^\gamma - Cb) |y|^\gamma \!\left( 1 - |y|^2\right)^{-(\eta+2)} \geq 0 \quad \mbox{in} \ B_1,
\end{multline*}
provided that $b> 0$ is chosen sufficiently large in terms of the constants in \eqref{Alip} and \eqref{Hcoer}, and that $a$ is taken to be the random variable
\begin{equation*}
a(\omega): = \frac1\delta\! \left( \sup_{y\in B_1} V(y,\omega) + C(b+1) \right).
\end{equation*}
Since $w$ is smooth, it follows from the definition of viscosity subsolution that $v-w$ cannot have a local maximum at any point in the set $\{ v> w\}$. Since $w(y) \to +\infty$ as $|y|\to 1$, we deduce that $\{ v > w \}$ is empty and, therefore, $v \leq w$ in $B_{1}$. The bound \eqref{locesteq1} now follows.
\end{proof}
The hypothesis $\alpha > \d$ is needed to show that $V(\cdot,\omega)$ is strictly sublinear at infinity, a.s. in $\omega$. In light of Lemma~\ref{locest}, this ensures that any subsolution of \eqref{aux} is bounded from above, a.s. in $\omega$, by a function growing strictly sublinearly at infinity. The latter fact is needed to obtain a comparison principle for \eqref{aux}.
\begin{lem} \label{potests}
There exist a set $\Omega_1\subseteq \Omega$ of full probability and constants $\sigma < 1$ and $C_1> 0$, such that, for all $\omega \in \Omega_1$,
\begin{equation} \label{sublin}
\limsup_{R \to \infty} R^{-\sigma} \sup_{y\in B_R} V(y,\omega) = 0.
\end{equation}
\end{lem}
\begin{proof}
Using the stationarity of $V$ and \eqref{Vmoment} and by covering $B_R$ with at most $C R^\d$ balls of radius 1, we see that, for any $\mu > 0$,
\begin{equation} \label{crudecov}
\ensuremath{\mathds{P}}\left[ \sup_{y\in B_R} V(y,\cdot) \geq \mu \right] \leq C R^\d \mu^{-\alpha}.
\end{equation}
Taking $R=2^k$ and $\mu = R^\sigma\varepsilon$ for fixed $\varepsilon > 0$ and $\d/\alpha < \sigma < 1$, we find
\begin{equation} \label{crudity}
\ensuremath{\mathds{P}}\left[ \sup_{y\in B_{2^k}} V(y,\cdot) \geq 2^{\sigma k}\varepsilon \right] \leq C_\varepsilon 2^{(\d-\sigma\alpha) k}.
\end{equation}
Since $d-\sigma\alpha < 0$, we obtain
\begin{equation*}
\sum_{k=1}^\infty \ensuremath{\mathds{P}}\left[ \sup_{y\in B_{2^k}} V(y,\cdot) \geq 2^{\sigma k}\varepsilon \right] \leq C_\varepsilon \sum_{k=1}^\infty 2^{(\d-\sigma\alpha) k} < \infty.
\end{equation*}
Applying the Borel-Cantelli lemma, we get
\begin{equation*}
\limsup_{R\to \infty} R^{-\sigma} \sup_{y\in B_{R}} V(y,\omega) \leq 2^\sigma \limsup_{k\to \infty} \, 2^{-\sigma k} \sup_{y\in B_{2^k}} V(y,\omega)\leq 2^\sigma \varepsilon \quad \mbox{a.s. in} \ \omega.
\end{equation*}
Disposing of $\varepsilon >0$ yields \eqref{sublin} for all $\omega$ in a set of full probability.
\end{proof}
The following is an immediate consequence of Lemmata~\ref{locest} and~\ref{potests}.
\begin{cor} \label{globest}
Fix $\delta > 0$, $p\in\ensuremath{\mathds{R}^\d}$, $\omega\in \Omega_1$, and $\d / \alpha < \sigma < 1$. If $v\in \USC(\ensuremath{\mathds{R}^\d})$ is a subsolution of \eqref{aux} in $\ensuremath{\mathds{R}^\d}$, then
\begin{equation} \label{globesteq1}
\limsup_{|y|\to\infty} |y|^{-\sigma} v(y) \leq 0.
\end{equation}
\end{cor}
\medskip
Due to the unboundedness of $H$, we cannot apply standard comparison results from the theory of viscosity solutions to \eqref{aux}. However, the previous corollary, the convexity of $H$ and the one-sided bound \eqref{Hbound} suffice to prove the following comparison principle.
\begin{lem} \label{comparison}
Fix $\delta > 0$, $p\in \ensuremath{\mathds{R}^\d}$ and $\omega \in \Omega_1$. Suppose that $u\in \USC(\ensuremath{\mathds{R}^\d})$ and $v\in \LSC(\ensuremath{\mathds{R}^\d})$ are, respectively, a subsolution and supersolution of \eqref{aux} in $\ensuremath{\mathds{R}^\d}$. Assume also that
\begin{equation}\label{compsublin}
\liminf_{|y| \to \infty} \frac{v(y)}{|y|} \geq 0.
\end{equation}
Then $u \leq v$ in $\ensuremath{\mathds{R}^\d}$.
\end{lem}
\begin{proof}
It follows from Corollary~\ref{globest} that, for all $\omega \in \Omega_1$,
\begin{equation*}
\limsup_{|y| \to \infty} \frac{u(y)}{|y|} \leq 0.
\end{equation*}
Since $(p,\omega)$ play no further role in the argument, we omit them for the rest of the proof. Define the auxiliary function
\begin{equation} \label{linearphi}
\varphi(y): = -\left( 1 + |y|^2 \right)^{\frac12}.
\end{equation}
it is immediate that $|D\varphi| + |D^2\varphi| \leq C$, hence, using \eqref{Alip} and \eqref{Hbound}, we have
\begin{equation} \label{phieqbnd}
\left| \tr( A(y,\omega) D^2\varphi) \right| + H(p+D\varphi,y,\omega) \leq C.
\end{equation}
Fix $\varepsilon>0$ and define the function
\begin{equation*}
\hat u_\varepsilon (y): = (1-\varepsilon) u(y) + \varepsilon (\varphi(y) - k),
\end{equation*}
where $k> 0$ is taken sufficiently large (depending on $\delta$) that $\varphi-k$ is a subsolution of \eqref{aux} in $\ensuremath{\mathds{R}^\d}$. Formally, using the convexity of $H$, we see that the function $\hat u_\varepsilon$ is a subsolution of \eqref{aux}. This is made rigorous in the viscosity sense by appealing to Lemma~\ref{convtrick}, or by a more direct argument using that $\varphi-k$ is smooth. Owing to \eqref{globesteq1}, \eqref{compsublin}, and the definition of $\varphi$, we have
\begin{equation*}
\liminf_{|y| \to \infty} \frac{v(y) - \hat u_\varepsilon(y)}{|y|} \geq \varepsilon,
\end{equation*}
and therefore we may apply the standard comparison principle (see \cite{CIL}), yielding
\begin{equation*}
\hat u_\varepsilon \leq v \quad \mbox{in} \ \ensuremath{\mathds{R}^\d}.
\end{equation*}
We obtain the result upon sending $\varepsilon \to 0$.
\end{proof}
We next demonstrate the well-posedness of the auxiliary macroscopic problem \eqref{aux}, following the general Perron method outlined, for example, in \cite{CIL}.
\begin{prop} \label{auxsolve}
For each fixed $p\in \ensuremath{\mathds{R}^\d}$ and $\omega \in \Omega_1$, there exists a unique solution $v^\delta = v^\delta(\cdot,\omega; p) \in C(\ensuremath{\mathds{R}^\d})$ of \eqref{aux}, which is stationary and such that, for some $C> 0$,
\begin{equation}\label{delvestbel}
\delta v^\delta(\cdot,\omega;p) \geq -C\left(1+|p|^\gamma \right).
\end{equation}
\end{prop}
\begin{proof}
Observe that by \eqref{Hbound}, the constant function $-C\delta^{-1}(1+|p|)^\gamma$ is a subsolution of \eqref{aux}. It is easy to check using \eqref{sublin} that, for sufficiently large $k$ depending on $\delta$, the function $(1+|y|^2)^{1/2} + k$ is a supersolution of \eqref{aux}. Define
\begin{equation*}
v^\delta(y,\omega) : = \sup\left\{ w(y) : w\in \USC(\ensuremath{\mathds{R}^\d}) \ \mbox{is a subsolution of} \ \eqref{aux} \ \mbox{in} \ \ensuremath{\mathds{R}^\d} \right\}.
\end{equation*}
Since $\omega\in \Omega_1$, Lemma~\ref{comparison} yields that $v^\delta \leq (1+|y|^2)^{1/2} + k$ in $\ensuremath{\mathds{R}^\d}$. Thus $v^\delta(y,\omega)$ is well-defined and finite. It follows (see \cite[Lemma 4.2]{CIL}) that $(v^\delta)^*$ is a viscosity subsolution of \eqref{aux}, where $(v^\delta)^*$ denotes the upper semicontinuous envelope of $v^\delta$. Since $\omega \in \Omega_1$, we deduce that $(v^\delta)^*(y)$
satisfies \eqref{compsublin}.
\begin{equation*}
\limsup_{|y| \to \infty} \frac{(v^\delta)^*(y)}{|y|} \leq 0.
\end{equation*}
If the lower semicontinuous envelope $(v^\delta)_*$ of $v^\delta$ failed to be a supersolution of \eqref{aux}, then this would violate the definition of $v^\delta$, see \cite[Lemma 4.4]{CIL}.
Clearly \eqref{delvestbel} holds by definition, and therefore we have
Applying Lemma~\ref{comparison}, we conclude that $(v^\delta)^* \leq (v^\delta)_*$, and therefore $v^\delta = (v^\delta)^* = (v^\delta)_*$ and so $v^\delta \in C(\ensuremath{\mathds{R}^\d})$ is a solution of \eqref{aux}. Uniqueness is immediate from Lemma~\ref{comparison}, and stationarity follows from uniqueness.
\end{proof}
We conclude this subsection with a continuous dependence estimate, asserting that, a.s. in $\omega$, if $p_1$ and $p_2$ are close, then $\delta v^\delta(\cdot,\omega;p_1)$ and $\delta v^\delta(\cdot,\omega;p_2)$ are close in an appropriate sense. Once we have homogenization, this is equivalent to showing that $\overline H$ is continuous. We address this point now rather than later due to technical difficulties we encounter in the homogenization proof. In particular, we must obtain a single subset of full probability on which \eqref{cplim} holds for all $p\in\ensuremath{\mathds{R}^\d}$. To accomplish this, we first obtain \eqref{cplim}, a.s. in $\omega$, for each rational $p$ and then intersect the respective subsets of~$\Omega$. This yields a subset of $\Omega$ of full probability on which the limit \eqref{cplim} holds for all rational $p$. To argue that, in fact, \eqref{cplim} holds for all $\omega$ in this subset and all $p \in \ensuremath{\mathds{R}^\d}$ requires such a continuous dependence estimate. For exactly the same reason, we also need the following result in the next subsection, where we obtain a single set of full probability on which the estimate \eqref{preosc} holds for all $p$.
\medskip
The continuous dependence estimate is Lemma~\ref{CDE-noxd}, below. It is based on the following preliminary lemma, which will also be useful to us later.
\begin{lem} \label{movep}
Fix $\lambda > 0$ and $p\in \ensuremath{\mathds{R}^\d}$, and define $w^\delta(y,\omega): = \lambda v^\delta(y,\omega;p)$. If $\lambda < 1$, then $w^\delta$ satisfies, for any $q\in \ensuremath{\mathds{R}^\d}$,
\begin{equation}\label{movepdn}
\delta w^\delta -\tr\!\left( A(y,\omega)D^2w^\delta\right) + H( q + Dw^\delta ,y,\omega) \leq (1-\lambda) H\left( \frac{q-\lambda p}{1-\lambda},y,\omega\right) \quad \mbox{in} \ \ensuremath{\mathds{R}^\d}.
\end{equation}
Likewise, if $\lambda > 1$, then for any $q\in \ensuremath{\mathds{R}^\d}$,
\begin{equation}\label{movepup}
\delta w^\delta -\tr\!\left( A(y,\omega)D^2w^\delta\right) + H( q + Dw^\delta ,y,\omega) \geq (1-\lambda) H\left( \frac{q-\lambda p}{1-\lambda},y,\omega\right) \quad \mbox{in} \ \ensuremath{\mathds{R}^\d}.
\end{equation}
\end{lem}
\begin{proof}
Writing $w^\delta$ in the form
\begin{equation*}
w^\delta(y) = \lambda v^\delta(y,\omega;p) = \lambda \big( (q-p)\cdot y + v^\delta(y,\omega;p) \big) + (1-\lambda)\left( \frac{\lambda(q-p)}{1-\lambda} \cdot y \right)
\end{equation*}
and using the convexity of $H$, we find that, formally,
\begin{equation*}
H(q+Dw^\delta(y),y,\omega) \leq \lambda H(p+Dv^\delta(y,\omega;p) ) + (1-\lambda) H\left( \frac{q-\lambda p}{1-\lambda},y,\omega\right).
\end{equation*}
Therefore, formally we have
\begin{multline*}
\delta w^\delta -\tr\!\left( A(y,\omega)D^2w^\delta\right) + H( q + Dw^\delta ,y,\omega) -(1-\lambda) H\left( \frac{q-\lambda p}{1-\lambda},y,\omega\right) \\
\leq \lambda \left( \delta v^\delta(y,\omega;p) - \tr\!\left( A(y,\omega)D^2v^\delta(y,\omega;p) \right) + H( p + Dv^\delta(y,\omega;p) ,y,\omega) \right) = 0.
\end{multline*}
This inequality is easy to confirm in the viscosity sense by performing an analogous calculation with smooth test functions. We have proved \eqref{movepdn}.
The proof of \eqref{movepup} is similar. Expressing $v^\delta(\cdot; p)$ in terms of $w^\delta$ as
\begin{equation*}
v^\delta(y,\omega;p) = \lambda^{-1} \big( (p-q)\cdot y + w^\delta(y) \big) + (1-\lambda^{-1}) \left( \frac{\lambda(q-p)}{1-\lambda} \cdot y\right),
\end{equation*}
we use again the convexity of $H$ to find that, formally,
\begin{equation*}
H(p+Dv^\delta(y,\omega;p) ) \leq \lambda^{-1} H(q+Dw^\delta(y)) + (1-\lambda^{-1}) H\left( \frac{q-\lambda p}{1-\lambda},y,\omega\right).
\end{equation*}
From this we formally obtain \eqref{movepup}. The derivation is once again made rigorous with smooth test functions.
\end{proof}
\begin{lem} \label{CDE-noxd}
There exists $C > 0$ such that, for each $\delta > 0$, $p_1,p_2\in \ensuremath{\mathds{R}^\d}$ and $\omega \in \Omega_1$,
\begin{equation} \label{CDE-noxdeq}
\sup_{y\in \ensuremath{\mathds{R}^\d}} \left| \delta v^\delta(y,\omega; p_1) - \delta v^\delta(y,\omega; p_2) \right| \leq C (1+|p_1| + |p_2|)^{\gamma-1}|p_1-p_2|.
\end{equation}
\end{lem}
\begin{proof}
Fix $\delta> 0$, $\omega \in \Omega_1$, and, for $i=1,2$, write $v_i^\delta(y,\omega) := v^\delta(y,\omega;p_i)$. For $0 < \lambda < 1$ to be selected below define $w^\delta(y) : = \lambda v_2^\delta(y,\omega)$. According to Lemma~\ref{movep}, $w^\delta$ satisfies the inequality
\begin{equation*}
\delta w^\delta -\tr\!\left( A(y,\omega)D^2w^\delta\right) + H( p_1 + Dw^\delta ,y,\omega) \leq (1-\lambda) H\left( \frac{p_1-\lambda p_2}{1-\lambda},y,\omega\right) \quad \mbox{in} \ \ensuremath{\mathds{R}^\d}.
\end{equation*}
Set $\lambda : = 1 - (1+|p_1|+|p_2|)^{-1} |p_1-p_2|$. It follows that
\begin{equation*}
\frac{|p_1-\lambda p_2|}{1-\lambda} = \frac{\big| (1+|p_1|+|p_2|) (p_1 -p_2) + |p_1-p_2|p_2 \big| }{|p_1-p_2|} \leq 1+|p_1|+2|p_2|.
\end{equation*}
Thus by \eqref{Hbound},
\begin{align}
(1-\lambda) H\left( \frac{p_1-\lambda p_2}{1-\lambda},y,\omega\right) & \leq C \frac{|p_1-p_2|}{1+|p_1|+|p_2|} \left(1+|p_1|+|p_2| \right)^{\gamma} \nonumber \\ & = C \left(1+|p_1|+|p_2| \right)^{\gamma-1} |p_1-p_2|. \label{convHbar2}
\end{align}
Therefore, $\widetilde w^\delta(y) : = w^\delta(y) - \delta^{-1} C \left(1+|p_1|+|p_2| \right)^{\gamma-1} |p_1-p_2|$ is a subsolution of
\begin{equation*}
\delta \widetilde w^\delta - \tr\!\left( A(y,\omega)D^2 \widetilde w^\delta\right) + H( p_1 + D \widetilde w^\delta ,y,\omega) \leq 0 \quad \mbox{in} \ \ensuremath{\mathds{R}^\d}.
\end{equation*}
Applying Lemma~\ref{comparison}, we obtain
\begin{equation*}
\widetilde w^\delta \leq v_1^\delta \quad \mbox{in} \ \ensuremath{\mathds{R}^\d}.
\end{equation*}
Rearranging some terms and using the definition of $\lambda$ and \eqref{delvestbel} yields that, in $\ensuremath{\mathds{R}^\d}$,
\begin{align*}
v^\delta_2 - v^\delta_1 & = (\lambda - 1) v^\delta_2 + \widetilde w^\delta - v^\delta_1 + \delta^{-1} C\left(1+|p_1|+|p_2| \right)^{\gamma-1} |p_1-p_2| \\
& \leq \delta^{-1} C (1+|p_2|^\gamma) (1-\lambda) + \delta^{-1} C\left(1+|p_1|+|p_2| \right)^{\gamma-1} |p_1-p_2| \\
& \leq \delta^{-1} C\left(1+|p_1|+|p_2| \right)^{\gamma-1} |p_1-p_2|.
\end{align*}
Now we repeat the argument reversing the roles of $p_1$ and $p_2$ to obtain \eqref{CDE-noxdeq}.
\end{proof}
\subsection{Estimates on the oscillation of $\delta v^\delta$ in balls of radius $\sim 1/\delta$.}
We obtain important estimates controlling the oscillation of $\delta v^\delta$ in balls of radius $\sim1/\delta$, which plays a critical role in the proof of \eqref{cplim}. If $V$ is bounded uniformly in $\omega$, it is easy to obtain estimates which are independent of $\delta$. Indeed, Lemma~\ref{locest} and Proposition~\ref{auxsolve} provide an $L^\infty$ bound for $\delta v^\delta$ while Lemma~\ref{locestgrad} below provides uniform, global Lipschitz estimates for $v^\delta$, and there is nothing more to show. In the unbounded setting, the situation is subtle. The mixing hypothesis together with Morrey's inequality and the ergodic theorem is needed to obtain the required estimate.
\medskip
The idea is as follows. First, we use Bernstein's method to get a local Lipschitz estimate on $v^\delta$ in terms of the nearby behavior of $V$. Next, we use the mixing hypothesis to control the average of an appropriate power $V$ over large balls, thereby providing us with some control over the average of $|Dv^\delta|^q$ on large balls, for some $q> \d$. We then apply Morrey's inequality, which yields an estimate on the oscillation of $\delta v^\delta$ on balls of radius $\sim 1/\delta$, centered at the origin. A supplementary argument using Egoroff's theorem combined with the ergodic theorem upgrades the latter estimate to the one we need, which is contained in Lemma~\ref{impest}, below.
\medskip
We begin with the local Lipschitz estimate on $v^\delta$.
\begin{lem} \label{locestgrad}
Fix $\delta > 0$ and $(p,\omega) \in \ensuremath{\mathds{R}^\d}\times \Omega$. There exists, an independent of $\delta$, $C=C(|p|)> 0$ such that any solution $v \in C(\bar B_1)$ of \eqref{aux} in $B_1$ is Lipschitz on $B_{1/2}$, and
\begin{equation} \label{locesteq2}
\esssup_{y\in B_{1/2}} |Dv(y)|^\gamma \leq C\left( 1 + \sup_{y\in B_1} V(y,\omega) \right).
\end{equation}
\end{lem}
\begin{proof}
The estimate follows by the Bernstein method. By performing a routine regularization, smoothing the coefficients and adding, if necessary, a small viscosity term, we may assume that $v$ is smooth. Next we adapt here the arguments of \cite{LL,LS2}.
Let $0 < \theta <1$ be chosen below and select a cutoff function $\varphi\in C^\infty(B_1)$ satisfying
\begin{equation}
0 \leq \varphi \leq 1, \ \ \varphi \equiv 1 \ \mbox{on} \ B_{\frac12}, \ \ \varphi \equiv 0 \ \mbox{in} \ B_1 \setminus B_{\frac34}, \ \ |D\varphi|^2 \leq C\varphi^{1+\theta} \ \ \mbox{and} \ \ \left| D^2 \varphi \right| \leq C\varphi^{\theta}. \label{varphiest}
\end{equation}
For example, we may take $\varphi = \xi^{2/(1-\theta)}$ where $\xi \in C^\infty(B_1)$ is any smooth function satisfying the first three condition of \eqref{varphiest}.
Define $z : = |Dv|^2$ and $\xi : = \varphi |Dv|^2 = \varphi z$, and compute
\begin{equation} \label{easyderv}
D\xi = z D\varphi + 2\varphi D^2v Dv, \quad D^2\xi = z D^2\varphi + 2 D\varphi \! \otimes\! (D^2v Dv) + \varphi (D^3v Dv + D^2v D^2v).
\end{equation}
Differentiating \eqref{aux} with respect to $y$ and multiplying by $\varphi Dv$ yields
\begin{multline} \label{diffauxle}
\delta \xi - \varphi Dv \cdot \tr \!\left( D_yA(y,\omega) D^2v + A(y,\omega) D_yD^2v \right) + \varphi D_pH(p+Dv,y,\omega) \cdot D^2vDv \\ + \varphi Dv\cdot D_yH(p+Dv,y,\omega) = 0.
\end{multline}
Combining \eqref{easyderv} and \eqref{diffauxle} and performing some computation, we find that, at any point $y_0$ at which $\xi$ achieves a positive local maximum,
\begin{multline} \label{diffauxhell}
2\delta \xi - 2\varphi \tr \!\left( D_yA D^2v\right) \cdot Dv + \tr\!\left( A (zD^2\varphi + 2 \varphi^{-1} z D\varphi\otimes D\varphi + 2\varphi D^2vD^2v) \right) \\
- z D_pH(p+Dv,y_0) \cdot D\varphi +2 \varphi D_yH(p+Dv,y_0) \cdot Dv \leq 0.
\end{multline}
Writing $M := D^2v(y_0)$ and $q: =Dv(y_0)$ and using \eqref{Hlip}, we obtain
\begin{multline} \label{diffauxhell2}
\varphi\tr(AM^2) \leq \varphi |q| | \tr(D_yAM)| + C|A| |q|^2 |D^2\varphi| + C |A| |D\varphi|^2 \varphi^{-1} |q|^2 \\ + C|q|^2 |p+q|^{\nu-1} |D\varphi| + C |q| \varphi V(y_0,\omega).
\end{multline}
The Cauchy-Schwarz inequality in the form
\begin{equation} \label{CSineq}
(\tr(AM))^2\leq C |A| \tr(AM^2)
\end{equation}
and \eqref{Alip} imply that
\begin{equation*}
\left( \tr(D_yA M) \right)^2 = \left( 2\tr( D_y\Sigma\Sigma^t M) \right)^2 \leq 4 |D_y\Sigma|^2 \tr( A M^2 ) \leq C \tr(AM^2).
\end{equation*}
This inequality, \eqref{varphiest}, another use of \eqref{CSineq} and some elementary inequalites yield
\begin{equation} \label{diffauxhell3}
\varphi (\tr(AM))^2\leq C \varphi \tr(AM^2) \leq C\varphi^\theta |q|^2 + C|q|^2 |p+q|^{\nu-1} \varphi^{(1+\theta)/2} + C |q| \varphi V(y_0,\omega).
\end{equation}
By squaring the equation \eqref{aux} and using \eqref{diffauxhell3}, we obtain, at $y=y_0$,
\begin{equation*}
\varphi\left( \delta v + H(p+q,y_0) \right)^2 \leq C\varphi^\theta |q|^2 + C|q|^2 |p+q|^{\nu-1} \varphi^{(1+\theta)/2} + C |q| \varphi \sup_{y\in B_{1}} V(y,\omega).
\end{equation*}
The inequality above, \eqref{Hcoer} and \eqref{locesteq1} yield
\begin{equation*}
\varphi |q|^{2\gamma} \leq C\varphi \left( 1 + \sup_{y\in B_1} \ensuremath{V}(y,\omega) \right)^2 + C\varphi^\theta |q|^2 + C|q|^{\gamma+1} \varphi^{(1+\theta)/2},
\end{equation*}
for a constant $C> 0$ depending as well on an upper bound for $|p|$. Setting $\theta := 1 / \gamma$, we find
\begin{equation*}
(\varphi^\theta |q|^{2})^\gamma \leq C\varphi \left( 1 + \sup_{y\in B_1} \ensuremath{V}(y,\omega) \right)^2 + C\varphi^\theta |q|^2 + C(\varphi^\theta |q|^2)^{(1+\gamma)/2}.
\end{equation*}
Therefore,
\begin{equation*}
\xi^\gamma = (\varphi |q|^2)^\gamma \leq C + C\left( 1 + \sup_{y\in B_1} \ensuremath{V}(y,\omega) \right)^2.
\end{equation*}
The bound \eqref{locesteq2} follows.
\end{proof}
From Lemma~\ref{locestgrad} we see immediately that $v^\delta(\cdot,\omega;p) \in W^{1,\infty}_{\mathrm{loc}}(\ensuremath{\mathds{R}^\d})$ for each $\omega\in \Omega_1$ and $p \in \ensuremath{\mathds{R}^\d}$, with $|Dv^\delta|$ controlled locally by $V$ and an upper bound for $|p|$.
\medskip
The purpose of the mixing hypothesis is to prove the following estimate which, in light of Lemma~\ref{locest}, give some control over the average of $|Dv^\delta|^q$ in large balls.
\begin{lem} \label{potestsD}
There exist a set $\Omega_2 \subseteq \Omega_1$ of full probability and constants $q> \d$ and $C_2 > 0$ such that, for every $\omega \in \Omega_2$,
\begin{equation} \label{bigmix}
\limsup_{R\to \infty} \bigg( \fint_{B_R} \sup_{B(y,1)}V(\cdot,\omega)^{q/\gamma} \, dy \bigg)^{1/q} \leq C_2.
\end{equation}
\end{lem}
Before we present the proof of Lemma~\ref{potestsD}, we briefly describe the general idea of how the mixing condition is used to prove \eqref{bigmix}. For convenience, it is better to consider a cube $Q_K$ instead of the ball $B_R$. Here $K$ is large integer and $Q_K$ is the cube of side length $2K$ centered at the origin. The idea is then to subdivide $Q_K$ into smaller cubes, which are obtained by two successive partitions of $Q_K$. The smaller cubes are collected into groups in such a way that within each group, the cubes are sufficiently separated so that we may apply the mixing condition. This provides some decay on the probability that the average of $\sup_{B(y,1)}|V(\cdot,\omega)|^{q/\gamma}$ is large on the union of each group.
\begin{proof}[{Proof of Lemma~\ref{potestsD}}]
Let $K$ and $N$ be positive integers such that $K$ is a multiple of $N$. Partition $Q_K$ into $M : = (K/N)^\d$ subcubes $Q^1,\ldots, Q^M$ of side length $2N$. Partition each of the cubes $Q^i$ into $L:=N^\d$ subcubes $Q^{i1},\ldots, Q^{iL}$ of side length 2, in such a way that $Q^{ij}$ has the same position in $Q^{i}$ as does $Q^{1j}$ relative to $Q^1$. That is, the translation that takes $Q^i$ to $Q^1$ also takes $Q^{ij}$ to $Q^{1j}$ (the partitions are illustrated in Figure~\ref{cubepart}). If $N\geq 5$, this ensures that, for each $1 \leq i,j\leq M$ and $1 \leq k \leq L$, we have
\begin{equation} \label{dAij}
\dist(Q^{ik},Q^{jk}) \geq 2( N - \sqrt2) \geq N+2.
\end{equation}
\begin{figure} \label{cubepart}
\centering
\begin{tikzpicture}[>=stealth]
\pgftransformscale{.50}
\draw[very thick] (6,6) -- (6,-6) -- (-6,-6) -- (-6,6) -- (6,6);
\draw[thick] (-6,-2) -- (6,-2);
\draw[thick] (-6,2) -- (6,2);
\draw[thick] (-2,-6) -- (-2,6);
\draw[thick] (2,-6) -- (2,6);
\draw (-2,5) -- (-6,5);
\draw (-2,4) -- (-6,4);
\draw (-2,3) -- (-6,3);
\draw (-5,2) -- (-5,6);
\draw (-4,2) -- (-4,6);
\draw (-3,2) -- (-3,6);
\draw (2,1) -- (6,1);
\draw (2,0) -- (6,0);
\draw (2,-1) -- (6,-1);
\draw (5,2) -- (5,-2);
\draw (4,2) -- (4,-2);
\draw (3,2) -- (3,-2);
\fill[lightgray] (-4,5) -- (-3,5) -- (-3,4) -- (-4,4) -- (-4,5);
\fill[lightgray] (4,1) -- (5,1) -- (5,0) -- (4,0) -- (4,1);
\draw[thick,->] (3.7,2.4) node[anchor=south]{$Q^{ij}$} -- (4.5,0.5);
\draw[thick,->] (0.9,2.8) node[anchor=east]{$Q^{i}$} -- (2.2,1.8);
\draw[thick,->] (-1.6,4.2) node[anchor=west]{$Q^{1{j}}$} -- (-3.5,4.5);
\draw[thick,->] (-0.9,1.2) node[anchor=west]{$Q^{1}$} -- (-2.2,2.2);
\end{tikzpicture}
\caption{The partitioning of $Q_K$ into smaller cubes in the proof of Lemma~\ref{potestsD}}
\end{figure}
Observe that
\begin{multline} \label{breakdn}
\fint_{Q_K} \sup_{B(y,1)} |V(\cdot,\omega)|^{q/\gamma} \, dy = (2K)^{-\d} \sum_{j=1}^L \sum_{i=1}^M \int_{Q^{ij}} \sup_{B(y,1)} |V(\cdot,\omega)|^{q/\gamma} \, dy \\
\leq (2K)^{-\d} L \max_{1\leq j \leq L} \left( \sum_{i=1}^M\sup_{\widetilde Q^{ij}} |V(\cdot,\omega)|^{q/\gamma} \right) = 2^{-\d} \max_{1\leq j \leq L} \left( \frac{1}{M} \sum_{i=1}^M \sup_{\widetilde Q^{ij}} |V(\cdot,\omega)|^{q/\gamma} \right),
\end{multline}
where $\widetilde Q^{ij} : = \{ y \in \ensuremath{\mathds{R}^\d} : \dist(Q^{ij},y) \leq 1\}$ is the set of points within a unit distance of $Q^{ij}$. Note that $\diam(\widetilde Q^{ij}) \leq 2+ 2\sqrt\d$, and by \eqref{dAij},
\begin{equation} \label{disteps}
\dist(\widetilde Q^{ik}, \widetilde Q^{jk} ) \geq N.
\end{equation}
Fix $1 \leq J \leq L$ and define the random variable
\begin{equation*}
g_i(\omega) : = \sup_{\widetilde Q^{iJ}} \left( |V(\cdot,\omega)|^{q/\gamma} \wedge K^{q\sigma/\gamma} \right),
\end{equation*}
with $\d/\alpha< \sigma < 1$ to be selected below. Lemma~\ref{mix-decay}, the mixing hypothesis \eqref{mixing} and \eqref{disteps} imply that, for each $1 \leq i,j\leq M$,
\begin{equation*}
\left| \ensuremath{\mathds{E}}[g_ig_j] - \ensuremath{\mathds{E}}[g_i]\,\ensuremath{\mathds{E}}[g_j] \right| \leq CK^{2q\sigma/\gamma} N^{-\beta}.
\end{equation*}
Lemma~\ref{mix-EVest} yields the estimates
\begin{equation*}
\ensuremath{\mathds{E}} \bigg( \frac1M \sum_{j=1}^M g_j \bigg)^2 \leq C \left(1 + K^{2q\sigma/\gamma} N^{-\beta} \right)
\end{equation*}
and
\begin{equation*}
\Var \bigg( \frac1M \sum_{j=1}^M g_j \bigg) \leq \frac1M \Var(g_j) + CK^{2q\sigma/\gamma} N^{-\beta} \leq C\left( K^{-\d} N^{\d} + K^{2q\sigma/\gamma} N^{-\beta} \right).
\end{equation*}
According to \eqref{alphabeta}, it is possible to choose $s\in \ensuremath{\mathds{Q}}$ such that $0 < s < 1/2$, $q> \d$ and $\d/\alpha < \sigma < 1$ such that $s(\beta-\d) > 2\sigma q /\gamma$. Selecting $N = K^{s}$ (we show below that this is possible) and
\begin{equation*}
\varepsilon : = \min\{ (1-s)\d, s\beta - 2\sigma q / \gamma \}> s\d > 0,
\end{equation*}
we obtain
\begin{gather*}
\ensuremath{\mathds{E}} \bigg( \frac1M \sum_{j=1}^M g_j \bigg)^2 \leq C \quad \mbox{and} \quad \Var \bigg( \frac1M \sum_{j=1}^M g_j \bigg) \leq C K^{-\varepsilon}.
\end{gather*}
It follows by Chebyshev's inequality (see Remark~\ref{remcheb}) that, for some $C> 0$,
\begin{equation*}
\ensuremath{\mathds{P}} \bigg[ \frac1M \sum_{j=1}^M g_j > t \bigg] \leq (t^2 - C)^{-1} K^{-\varepsilon} \leq C K^{-\varepsilon} \quad \mbox{for all} \quad t > 2C.
\end{equation*}
Recalling \eqref{breakdn} and the choice of $N$, we see that, for each $t> 2C$,
\begin{align*}
\lefteqn{ \ensuremath{\mathds{P}}\bigg[ \bigg(\fint_{Q_K} \sup_{B(y,1)} |V(\cdot,\omega)|^{q/\gamma} \wedge R^{q\sigma/\gamma} \, dy\bigg)^{1/q} > t \bigg]} \qquad \qquad & \\
& \leq \ensuremath{\mathds{P}} \bigg[ \max_{1\leq j \leq L} \bigg( \frac{1}{M} \sum_{i=1}^M \sup_{\widetilde Q^{ij}} |V(\cdot,\omega)|^{q/\gamma} \wedge R^{q\sigma/\gamma} \bigg) > Ct^q \bigg] \\
& \leq L \ensuremath{\mathds{P}} \left[ \frac1M \sum_{j=1}^M g_j > C t^q \right] \leq C L K^{- \varepsilon} \leq C K^{s\d-\varepsilon}.
\end{align*}
Writing $h(y):= \sup_{B(y,1)} |V(\cdot,\omega)|^{q/\gamma}$, in view of \eqref{crudity}, we see that for all $t > C$,
\begin{align*}
\ensuremath{\mathds{P}}\bigg[ \bigg( \fint_{Q_K} h(y)\, dy\bigg)^{1/q} > t \bigg] & \leq \ensuremath{\mathds{P}}\bigg[ \bigg( \fint_{Q_K} h(y) \wedge R^{q\sigma/\gamma}\, dy \bigg)^{1/q} > t \bigg] + \ensuremath{\mathds{P}}\left[ \max_{y\in Q_K} h(y) > K^{q\sigma/\gamma} \right] \\
& \leq CK^{s\d-\varepsilon} + CK^{\d-\sigma\alpha}.
\end{align*}
Observe that the exponents $sd-\varepsilon$ and $d-\sigma\alpha$ are negative. The above applies to any $K\in\ensuremath{\mathds{N}}$ for which $K^s$ is an integer which divides $K$. In particular we can take $K = 2^m$ for any positive integer $m$ such that $ms\in\ensuremath{\mathds{N}}$. Writing $s=a/b$ for $a,b \in \ensuremath{\mathds{N}}$ and applying the Borel-Cantelli lemma to the estimate above, we obtain
\begin{equation*}
\limsup_{n\to \infty} \bigg( \fint_{Q_{2^{nb}}} h(y) \, dy \bigg)^{1/q} \leq C \quad \mbox{a.s. in} \ \omega.
\end{equation*}
Now \eqref{bigmix} follows, since for large $R > 0$, the positive integers $k := \lfloor \log_{2^b} (R/\sqrt\d) \rfloor $ and $\ell := \lceil \log_{2^b} R \rceil$ satisfy $Q_{2^{kb}} \subseteq B_R \subseteq Q_{2^{\ell b}}$ and $\left|Q_{2^{\ell b}} \right| \leq C \left| B_R \right| \leq C\left| Q_{2^{k b}} \right|$. Thus
\begin{equation*}
\fint_{Q_{2^{k b}}} h(y)\, dy \leq C \fint_{B_R} h(y)\,dy \leq C \fint_{Q_{2^{\ell b}}} h(y)\, dy. \qedhere
\end{equation*}
\end{proof}
In the next lemma, we prove \eqref{preosc}. The inequalities \eqref{locesteq2}, \eqref{bigmix}, and Morrey's inequality provides an estimate on the oscillation of $\delta v^\delta$ in the ball $B_{r/\delta}$. A supplementary argument combining the ergodic theorem and Egoroff's theorem extends this to the balls of the form $B(y/\delta, r/\delta)$.
\begin{lem} \label{impest}
There exist a set $\Omega_3 \subseteq \Omega_2$ of full probability and a constant $C = C(k) > 0$, such that, for each $y\in \ensuremath{\mathds{R}^\d}$, $r> 0$, $|p|<k$ and $\omega\in\Omega_3$,
\begin{equation} \label{oscest}
\limsup_{\delta \to 0} \osc_{B(y/\delta,r/\delta)} \delta v^\delta(\cdot,\omega) \leq Cr.
\end{equation}
\end{lem}
\begin{proof}
Fix $p \in \ensuremath{\mathds{R}^\d}$. It follows from Morrey's inequality, \eqref{locesteq2} and \eqref{bigmix} that, for each $r> 0$ and $\omega \in \Omega_2$,
\begin{multline} \label{oscatz}
\limsup_{\delta \to 0} \sup_{y\in B_{r/\delta}} \left| \delta v^\delta(y,\omega) - \delta v^\delta(0,\omega) \right| \leq C\limsup_{\delta \to 0}\, r \bigg( \fint_{B_{r/\delta}} |Dv^\delta(y,\omega)|^{q} \, dy \bigg)^{1/q} \\ \leq Cr \limsup_{\delta\to0} \bigg( \fint_{B_{r/\delta}} \sup_{B(y,1)} V(\cdot,\omega)^{q/\gamma} \, dy \bigg)^{1/q} \leq Cr,
\end{multline}
and \eqref{oscest} holds for $y=0$. Note that $C$ depends also on an upper bound for $|p|$.
\medskip
To obtain the full \eqref{oscest}, we combine \eqref{oscatz}, Egoroff's theorem and the ergodic theorem. Fix $r, s> 0$. According to Egoroff's theorem, for each $\eta > 0$, there exists $D_\eta \subseteq \Omega_2$ with $\ensuremath{\mathds{P}}[D_\eta] \geq 1-\eta$ such that, for each $\omega \in D_\eta$ and $\delta=\delta(\eta) > 0$ sufficiently small,
\begin{equation} \label{egor-lip}
\osc_{B(0,(r(1+\eta)/\delta)} \delta v^\delta(\cdot,\omega) \leq 2C(1 +2\eta) r.
\end{equation}
The ergodic theorem yields $E_\eta \subseteq \Omega$ such that $\ensuremath{\mathds{P}}[E_\eta]=1$ and, for all $\omega \in E_\eta$,
\begin{equation} \label{erglim}
\lim_{R \to \infty} \fint_{B_R} \mathds{1}_{D_\eta}(\tau_z\omega) \, dz = \ensuremath{\mathds{P}}[D_\eta] \geq 1-\eta.
\end{equation}
Fix $y\in \ensuremath{\mathds{R}^\d}$ such that $|y| \leq s$. It follows from \eqref{erglim} that, if $\delta=\delta(\eta,s) > 0$ is sufficiently small and $\omega \in E_\eta$, then
\begin{equation} \label{ergod-lip}
| \{ z \in B_{2|y|} : \tau_{z/\delta}\omega \in D_\eta \}| \geq (1-2\eta) |B_{2|y|}|.
\end{equation}
Define
\begin{equation*}
\Omega_{r,s} : = \bigcap_{j=1}^\infty \bigcup_{k=j}^\infty \left( D_{2^{-k}} \cap E_{2^{-k}} \right),
\end{equation*}
and observe that $\ensuremath{\mathds{P}}[\Omega_{r,s}] = 1$. Fix a positive integer $j$, $\omega \in \Omega_{r,s}$ and set $\varepsilon : = 2^{-j} > 0$. By making $j$ larger, if necessary, we may assume that $\omega \in D_{\varepsilon} \cap E_\varepsilon$. For all $\delta > 0$ sufficiently small, depending on $\omega$, $k$ and $j$, \eqref{ergod-lip} holds with $\eta = \varepsilon$. This implies that we can find $z\in B_{2|y|}$ such that $\tau_{z/\delta}\omega \in D_\varepsilon$ and $|z-y| \leq C\varepsilon |y|$, due to the fact that the ball $B(y,C\varepsilon |y|)$ is too large to lie in the complement of $\{ z \in \ensuremath{\mathds{R}^\d}: \tau_{z/\delta} \omega \in D_\varepsilon \}$. Then $B(y/\delta,r/\delta) \subseteq B(z/\delta, (r+C\varepsilon|y|)/\delta)$. By shrinking $\delta> 0$, again depending on $\omega$, $s$ and $j$, we may assume that \eqref{egor-lip} holds for $\eta = C\varepsilon|y|$. For such $\delta$, we have
\begin{multline*}
\osc_{B(y/\delta,r/\delta)} \delta v^\delta(\cdot, \omega) \leq \osc_{B(z/\delta,r(1+C\varepsilon|y|)/\delta)} \delta v^\delta(\cdot, \omega) \\ = \osc_{B(0,r(1+C\varepsilon |y|)/\delta)} \delta v^\delta(\cdot, \tau_{z/\delta}\omega) \leq 2C(1+2C\varepsilon s) r.
\end{multline*}
Sending first $\delta \to 0$ and then $j \to \infty$ so that $\varepsilon \to 0$, we deduce that, for every $|y| \leq s$ and $\omega \in \Omega_{r,s}$,
\begin{equation} \label{finliplim}
\limsup_{\delta\to 0} \osc_{B(y/\delta,r/\delta)} \delta v^\delta (\cdot, \omega) \leq 2Cr.
\end{equation}
Now define
\begin{equation*}
\Omega_3: = \bigcap_{r \in \ensuremath{\mathds{Q}}_+} \bigcap_{s\in \ensuremath{\mathds{N}}} \Omega_{r,s}.
\end{equation*}
Observe that $\ensuremath{\mathds{P}}[\Omega_3] = 1$ and \eqref{finliplim} holds for every rational $r> 0$, $y\in \ensuremath{\mathds{R}^\d}$ and $\omega\in \Omega_3$. By the monotonicity of the quantity on the left side of \eqref{finliplim} in $r$, the inequality \eqref{finliplim} must then hold for all $r> 0$, $y\in \ensuremath{\mathds{R}^\d}$ and $\omega\in \Omega_3$.
So far, we have obtained \eqref{oscest} only for a fixed $p\in \ensuremath{\mathds{R}^\d}$, that is, $\Omega_3$ depends on $p$. To complete the proof, we replace $\Omega_3$ by the intersection of such sets over all $p\in \ensuremath{\mathds{Q}}^\d$. An appeal to Lemma~\ref{CDE-noxd} then completes the proof.
\end{proof}
\begin{remark}
Notice that in the argument above we have proved slightly more than \eqref{oscest}, namely that, for each $p\in \ensuremath{\mathds{R}^\d}$, $r > 0$ and $\omega \in \Omega_3$,
\begin{equation} \label{super-oscest}
\sup_{R > 0} \limsup_{\delta\to 0} \sup_{y \in B_R} \frac1r\osc_{B(y/\delta, r/ \delta)} \delta v^\delta(\cdot,\omega) \leq C(|p|).
\end{equation}
We can do even better by observing that
\begin{equation*}
r\mapsto \sup_{y\in B_{R-r}} \frac1r \osc_{B(y/\delta,r/\delta)} \delta v^\delta(\cdot,\omega) \quad \mbox{is decreasing,}
\end{equation*}
from which it follows that, for every $\omega\in \Omega_3$ and $p\in \ensuremath{\mathds{R}^\d}$,
\begin{equation} \label{uber-oscest}
\sup_{s,R > 0} \limsup_{\delta\to 0} \sup_{y \in B_R} \sup_{s\leq r \leq R} \frac1r\osc_{B(y/\delta, r/ \delta)} \delta v^\delta(\cdot,\omega) \leq C(|p|).
\end{equation}
\end{remark}
The estimate \eqref{oscest} will typically be applied for $r>0$ very small. In certain situations we control of the oscillation of $\delta v^\delta$ in $B_{R/\delta}$ for large $R$, but require something better than $CR$. In bounded environments, of course, the quantity $\delta v^\delta$ is (essentially) bounded in $\ensuremath{\mathds{R}^\d}\times \Omega$. In the next lemma we prove that, roughly speaking, this is also true in our unbounded setting \emph{in balls of radius $\sim1/\delta$ as $\delta \to 0$}. The arguments strongly uses the mixing condition (again) as well as our previous oscillation bound \eqref{uber-oscest}.
\begin{lem} \label{uppbndosc}
There exists a set $\Omega_4 \subseteq\Omega_3$ of full probability and a constant $C=C(k) > 0$, such that, for each $|p| < k$ and $\omega\in \Omega_4$,
\begin{equation} \label{oscest-large}
\sup_{R>0} \limsup_{\delta \to 0} \osc_{B_{R/\delta}} \delta v^\delta(\cdot, \omega) \leq C.
\end{equation}
\end{lem}
\begin{proof}
It suffices to control $\max_{B_{R/\delta}} \delta v^\delta(\cdot,\omega)$, since by \eqref{delvestbel}, $\delta v^\delta (\cdot,\omega)$ is bounded from below uniformly in $\delta$ and a.s. in $\omega$.
We first estimate the quantity $\ensuremath{\mathds{P}}\big[ \inf_{y\in B_{r/\delta}} \sup_{z\in B(y,1)} \ensuremath{V}(z,\cdot) \geq t \big]$ for $r >0$, where $t>0$ is selected below. Find points $y_1,\ldots, y_{k_\delta} \in B_r$ such that $k_\delta \approx |\log \delta|$ and $|y_i - y_j| > r_\delta$ for $i\neq j$, where $r_\delta \approx k_\delta^{-1/\d} r$. For any $t> 0$,
\begin{align*}
\ensuremath{\mathds{P}}\Big[ \inf_{y\in B_{r/\delta}} \sup_{z\in B(y,1)} \ensuremath{V}(z,\cdot) \geq t \Big] & \leq \ensuremath{\mathds{P}}\Big[ \min_{1 \leq i \leq k_\delta} \sup_{z\in B(y_i/\delta,1)} \ensuremath{V}(z,\cdot) \geq t \Big]
\end{align*}
By the mixing hypothesis, the quantity on the right is at most
\begin{equation*}
\ensuremath{\mathds{P}}\Big[ \sup_{z\in B(y,1)} \ensuremath{V}(z,\cdot) \geq t \Big]^{k_\delta} + C(k_\delta-1) \Big( \frac{r_\delta}{\delta} - 2 \Big)^{-\beta}.
\end{equation*}
Choosing $t> 0$ so that $\ensuremath{\mathds{P}}\Big[\sup_{z\in B(y,1)} \ensuremath{V}(z,\cdot) \geq t \Big] \leq \tfrac 12$, we obtain, for some suitably small $\kappa > 0$,
\begin{equation} \label{oscestlg1}
\ensuremath{\mathds{P}}\Big[ \inf_{y\in B_{r/\delta}} \sup_{z\in B(y,1)} \ensuremath{V}(z,\cdot) \geq t \Big] \leq 2^{-k_\delta} + Ck_\delta^{1+\beta/d} r^{-\beta} \delta^{\beta} \leq C(1+r^{-\beta}) \delta^{\kappa}.
\end{equation}
\medskip
Fix $0 < r < R$ and choose $z_1,\ldots,z_\ell \in B_R$ with $\ell \approx (R/r)^d$ such that $B_R \subseteq \cup_{i=1}^\ell B(z_i,r/2)$. According to \eqref{oscestlg1}, the stationarity of $\ensuremath{V}$ and with $t> 0$ chosen as above, we have
\begin{multline*}
\ensuremath{\mathds{P}} \Big[ \sup_{z\in B_R} \inf_{y\in B(z/\delta, r/\delta)} \sup_{x\in B(y,1)} V(x,\cdot) \geq t\Big] \leq \ensuremath{\mathds{P}}\Big[ \sup_{1\leq i \leq \ell}\, \inf_{y\in B(z_i/\delta,r/2\delta)} \, \sup_{x\in B(y,1)} \ensuremath{V}(x,\cdot) \geq t \Big] \\ \leq \ell\, \ensuremath{\mathds{P}}\Big[ \inf_{y\in B_{r/\delta}}\, \sup_{z\in B(y,1)} \ensuremath{V}(z,\cdot) \geq t \Big] \leq C \ell (1+r^{-\beta}) \delta^\kappa.
\end{multline*}
Therefore, by the Borel-Cantelli lemma, along the diadic sequence $\delta_j : = 2^{-j}$,
\begin{equation} \label{upposc-est}
\limsup_{j\to \infty} \max_{z\in B_{R/\delta_j} } \inf_{y\in B(z,r/\delta_j)} \sup_{x\in B(y,1)} V(x,\omega) \leq t \quad \mbox{a.s. in} \ \omega.
\end{equation}
Since $t> 0$ is independent of both $R$ nor $r$, we deduce that \eqref{upposc-est} holds along the full sequence $\delta \rightarrow 0$, i.e.,
\begin{equation} \label{upposc-estfull}
\limsup_{\delta \to 0} \max_{z\in B_{R/\delta} } \inf_{y\in B(z,r/\delta)} \sup_{x\in B(y,1)} V(x,\omega) \leq t \quad \mbox{a.s. in} \ \omega.
\end{equation}
Combining this inequality and \eqref{locesteq1} we obtain, for a constant $C$ depending on the appropriate quantities,
\begin{equation*}
\limsup_{\delta \to 0} \sup_{y\in B_{R/\delta} } \inf_{y\in B(z,r/\delta)} \delta v^\delta(y,\omega) \leq C \quad \mbox{a.s. in} \ \omega.
\end{equation*}
The last inequality and \eqref{uber-oscest} imply that
\begin{equation*}
\limsup_{\delta \to 0} \sup_{y\in B_{R/\delta} } \delta v^\delta(y,\omega) \leq C(1+ C r) \quad \mbox{a.s. in} \ \omega.
\end{equation*}
We send $r \to 0$ to obtain the result.
\end{proof}
\begin{remark} \label{moregen}
The estimate \eqref{oscest-large} permits us to generalize our homogenization result to equations of the form \eqref{HJx}, with coefficients which depend also on the macroscopic variable $x$. The difficulty in generalizing to such equations lies in obtaining a continuous dependence result stating that $\delta v^\delta(0,\omega;p_1,x_1)$ is close to $\delta v^\delta(0,\omega;p_2,x_2)$ if $| p_1-p_2|+|x_1-x_2|$ is small, at least in the limit as $\delta \to 0$ (and this must hold a.s. in $\omega$). While we do not give details here, it is precisely \eqref{oscest-large} that permits us to obtain such a continuous dependence estimate using the classical viscosity solution-theoretic comparison machinery
\end{remark}
\section{The effective Hamiltonian}
\label{EH}
Here we define the effective Hamiltonian $\overline H$ and explore some of its basic properties. In the process, we perform much of the work for the proof of Theorem~\ref{MAIN}.
\subsection{Construction of the effective Hamiltonian $\overline H$}
The first step in the proof of Theorem~\ref{MAIN}, formulated in the following proposition, is the identification of the effective Hamiltonian $\overline H$. The argument is based on the method recently introduced in \cite{LS3}, although substantial modifications are necessary in the unbounded setting, as discussed above.
\begin{prop} \label{mainstep}
There exists $\Omega_5 \subseteq \Omega_4$ of full probability and a continuous function $\overline H:\ensuremath{\mathds{R}^\d} \to \ensuremath{\mathds{R}}$ such that, for every $R> 0$, $p \in \ensuremath{\mathds{R}^\d}$, and $\omega\in \Omega_5$,
\begin{equation} \label{mainstepeq}
\lim_{\delta \to 0} \ensuremath{\mathds{E}} \Big[ \, \sup_{y\in B_{R/\delta}} \left| \delta v^\delta (y,\omega;p) + \overline H(p)\right| \Big]=0
\end{equation}
and
\begin{equation} \label{oneside}
\liminf_{\delta \to 0} \delta v^\delta(0,\omega) = -\overline H(p).
\end{equation}
Moreover, for each $p \in \ensuremath{\mathds{R}^\d}$, there exists a function $w: \ensuremath{\mathds{R}^\d}\times \Omega \to \ensuremath{\mathds{R}}$ such that, a.s. in $\omega$, $w(\cdot,\omega) \in W^{1,\alpha}_{\mathrm{loc}} (\ensuremath{\mathds{R}^\d})$, $Dw$ is stationary, and for every $\omega \in \Omega_5$,
\begin{equation} \label{subcorr}
\left\{
\begin{aligned}
& -\tr (A(y,\omega) D^2w ) + H(p+Dw,y,\omega) \leq \overline H(p) \quad \mbox{in} \ \ \ensuremath{\mathds{R}^\d}, \\
& |y|^{-1} w(y,\omega) \rightarrow 0 \quad \mbox{as} \quad |y| \to \infty.
\end{aligned}
\right.
\end{equation}
\end{prop}
\begin{proof}
The (local Lipschitz) continuity of $\overline H$ follows from Lemma~\ref{CDE-noxd}, once we have shown \eqref{mainstepeq}. Moreover, according to Lemma~\ref{CDE-noxd}, we may argue for a fixed $p\in \mathbb{Q}^d$, and then obtain $\Omega_5$ by intersecting the relevant subsets of $\Omega$ obtained for each rational $p$. We therefore fix $p$ and omit all dependence on $p$.
\medskip
The proof is divided into three steps.
\medskip
\emph{Step 1: Construction of a subcorrector which is strictly sublinear at infinity.}
For each $\delta > 0$, define
\begin{equation*}
w^\delta(y,\omega) := v^\delta(y,\omega) - v^\delta(0,\omega).
\end{equation*}
Owing to \eqref{Vmoment}, \eqref{locesteq1}, \eqref{delvestbel} and \eqref{locesteq2}, we can find a subsequence $\delta_j \to 0$, a random variable $\overline H = \overline H(p,\omega) \in \ensuremath{\mathds{R}}$, a function $w\in L^\alpha_{\mathrm{loc}}(\ensuremath{\mathds{R}^\d}\times \Omega)$ and a field $\Phi \in L^\alpha_{\mathrm{loc}}(\ensuremath{\mathds{R}^\d}\times\Omega ; \ensuremath{\mathds{R}^\d})$ such that, for every $R> 0$, as $j\to \infty$,
\begin{equation} \label{weaklims}
\left\{ \begin{aligned}
& -\delta_j v^{\delta_j}(0,\cdot ) \rightharpoonup \overline H(p,\cdot) \quad \mbox{weakly} \ \mbox{in} \ L^\alpha(\Omega), \\
& w^{\delta_j} \rightharpoonup w \quad \mbox{weakly} \ \mbox{in} \ L^\alpha(B_R\times\Omega), \\
& Dv^{\delta_j} \rightharpoonup \Phi \quad \mbox{weakly} \ \mbox{in} \ L^{\alpha}(B_R\times\Omega;\ensuremath{\mathds{R}^\d}).
\end{aligned} \right.
\end{equation}
The stationarity of the functions $v^{\delta_j}$, the ergodicity hypothesis and Lemma~\ref{impest} imply that $\overline H$ is independent of $\omega$, i.e., $\overline H(p,\omega) = \overline H(p)$ a.s. in $\omega$. Indeed, it suffices to check that, for each $\mu \in \ensuremath{\mathds{R}}$, the event $\{ \omega \in \Omega : \overline H(p,\omega) \geq \mu \}$ is invariant under $\tau_y$, which follows immediately from \eqref{oscest}.
It is clear that $\Phi$ is stationary, a property inherited from the sequence $\{ Dv^{\delta_j}\}$, and that $\Phi = (\Phi^1,\ldots,\Phi^d)$ is gradient-like in the sense that for every compactly-supported smooth test function $\psi=\psi(y)$,
\begin{equation*}
\int_{\ensuremath{\mathds{R}^\d}} \left( \Phi^i(y,\omega) \psi_{y_j} (y) - \Phi^j(y,\omega) \psi_{y_i}(y) \right) \, dy = 0 \quad \mbox{a.s. in} \ \omega.
\end{equation*}
It follows (c.f. Kozlov \cite[Proposition 7]{K}) that $\Phi = Dw$, a.s. in $\omega$ and in the sense of distributions. Since $\alpha > d$, the Sobolev imbedding theorem yields, without loss of generality, that $w(\cdot,\omega) \in C(\ensuremath{\mathds{R}^\d})$ a.s. in $\omega$.
\medskip
The convexity hypothesis \eqref{Hconv} and the equivalence of distributional and viscosity solutions for linear inequalities (c.f. Ishii~\cite{I}) allow us to pass to weak limits in \eqref{aux}, obtaining that $w(\cdot,\omega)$ is a viscosity solution, a.s. in $\omega$, of
\begin{equation} \label{subcorreq}
- \tr\!\left( A(y,\omega) D^2w \right) + H(p+Dw, y, \omega) \leq \overline H(p) \quad \mbox{in} \ \ensuremath{\mathds{R}^\d}.
\end{equation}
Since
\begin{equation*}
\ensuremath{\mathds{E}}\Phi(0,\cdot) = \lim_{j\to \infty} \ensuremath{\mathds{E}} Dv^{\delta_j}(0,\cdot) = 0,
\end{equation*}
it follows from Lemma~\ref{koz} that
\begin{equation} \label{sublininfty}
\lim_{|y|\to\infty} |y|^{-1}w(y,\omega) = 0 \quad \mbox{a.s. in} \ \omega.
\end{equation}
Therefore $w$ is a subcorrector which is strictly sublinear at infinity.
\medskip
\emph{Step 2: $\overline H$ characterizes the full limit of $\delta v^\delta(0,\omega)$ in $L^1(\Omega,\mathbb P)$.}
The key step consists in showing that
\begin{equation} \label{cmpineq}
-\overline H\leq \liminf_{\delta \to 0} \delta v^\delta(0,\omega) \quad \mbox{a.s. in} \ \omega,
\end{equation}
which we prove by comparing the subcorrector $w$ to $v^\delta$. Let $\widetilde\Omega$ be a subset of $\Omega$ of full probability such that $\widetilde\Omega \subseteq \Omega_4$ and, for every $\omega \in \widetilde\Omega$, we have $\overline H(p,\omega) = \overline H(p)$ as well as \eqref{subcorreq} and \eqref{sublininfty}.
\medskip
Fix $\omega\in \widetilde\Omega$. We remark that the constants we introduce immediately below depend on $\omega$. Let $\varphi$ be defined by \eqref{linearphi}, and recall from \eqref{phieqbnd} that
\begin{equation} \label{msphibnd}
\left| \tr( A(y,\omega) D^2\varphi) \right| + H(p+D\varphi,y,\omega) \leq C.
\end{equation}
For each $\delta > 0$, define the function
\begin{equation*}
w^\delta(y) : = (1-\varepsilon) \!\left( w(y,\omega) - ( \overline H + \eta ) \delta^{-1} \right) + \varepsilon \varphi(y),
\end{equation*}
where $\eta > 0$ is a given small constant and $\varepsilon > 0$ will be chosen below in terms of $\eta$. The strategy for obtaining \eqref{cmpineq} lies in comparing $w^\delta$ and $v^\delta$ in the limit as $\delta \to 0$. Assuming that $w$ is smooth, in view of \eqref{subcorreq}, \eqref{msphibnd} and the convexity of $H$, we have
\begin{equation}
\delta w^\delta -\tr\!\left( A(y,\omega) D^2w^\delta \right) +H(p+Dw^\delta,y) \leq \delta w^\delta + (1-\varepsilon) \overline H + C \varepsilon.
\label{delwdel}
\end{equation}
In the case $w$ is not smooth, one can verify \eqref{delwdel} in the viscosity sense by using either that $\varphi$ is smooth, or by appealing to Lemma~\ref{convtrick}. According to \eqref{sublininfty},
\begin{equation*}
\sup_{B_R} w \leq C_\eta + \eta^3 R,
\end{equation*}
and so, by choosing $\varepsilon = \min\{ 1/4, \eta/4C\}$ with $C$ is as in \eqref{delwdel}, we estimate the right side of \eqref{delwdel} by
\begin{equation*}
\delta w^\delta + (1-\varepsilon) \overline H + C \varepsilon = (1-\varepsilon) (\delta w-\eta) + C\varepsilon \leq \delta C_{\eta} + \delta \eta^3 R - \frac12 \eta \quad \mbox{in} \ B_R.
\end{equation*}
Observe that \eqref{delvestbel} implies
\begin{equation*}
w^\delta - v^\delta \leq (1-\varepsilon) w +C \delta^{-1} - c\eta R \quad \mbox{on} \ \partial B_R.
\end{equation*}
By selecting $R : = C(\delta\eta)^{-1}$ for a large constant $C> 0$ and taking $\delta$ to be sufficiently small, depending on both $\omega$ and $\eta$, we have
\begin{equation}
\left\{ \begin{aligned}
& \delta w^\delta -\tr \! \left( A(x,y,\omega) D^2 w^\delta \right) + H(p+Dw^\delta,x,y,\omega) \leq 0 & \mbox{in} & \ B_R, \\
& w^\delta \leq v^\delta & \mbox{on} & \ \partial B_R.
\end{aligned} \right.
\end{equation}
We may now apply the comparison principle to deduce that $w^\delta(\cdot) \leq v^\delta(\cdot,\omega)$ in $B_R$, and in particular, $w^\delta(0) \leq v^\delta(0,\omega)$. Multiplying this inequality by $\delta$ and sending $\delta \to 0$ yields
\begin{equation*}
-\overline H -\eta \leq ( 1 - C \eta)^{-1} \liminf_{\delta \to 0} \delta v^\delta (0,\omega).
\end{equation*}
Recalling that $\omega \in \widetilde\Omega$ was arbitrary with $\ensuremath{\mathds{P}}\big[\widetilde\Omega\big] =1$ and disposing of $\eta > 0$ yields \eqref{cmpineq}.
\medskip
Since $-\overline H$ is the weak limit of the sequence $\delta_j v^{\delta_j}(0,\cdot)$, the reverse of \eqref{cmpineq} is immediate and we obtain \eqref{oneside}. It follows that the full sequence $\delta v^\delta(0,\cdot)$ converges weakly to $-\overline H$, that is, as $\delta \to 0$,
\begin{equation*}
\delta v^\delta (0,\cdot) \rightharpoonup -\overline H = \liminf_{\delta\to 0} \delta v^\delta(0,\cdot) \quad \mbox{weakly in} \ L^\alpha(\Omega).
\end{equation*}
Lemma~\ref{meastheor} yields
\begin{equation}\label{convprob}
\ensuremath{\mathds{E}} \big| \delta v^\delta(0,\cdot) + \overline H \big| \rightarrow 0 \quad \mbox{as} \quad \delta \to 0.
\end{equation}
\medskip
\emph{Step 3: Improvement of \eqref{convprob} to balls of radius $\sim 1/\delta$.} We claim that, for each $R> 0$,
\begin{equation} \label{convprob1del}
\lim_{\delta\to0} \ensuremath{\mathds{E}} \bigg[ \sup_{y\in B_{R/\delta}} \big| \delta v^\delta(0,\cdot) + \overline H \big| \bigg] = 0.
\end{equation}
Fix $R > 0$. Let $\rho > 0$ and select points $y_1,\ldots,y_k \in B_R$ such that
\begin{equation*}
B_R \subseteq \bigcup_{i=1}^k B(y_i,\rho) \quad \mbox{and} \quad k \leq C \bigg( \frac{R}{\rho}\bigg)^\d.
\end{equation*}
Using \eqref{uber-oscest}, we find
\begin{align*}
\lefteqn{ \limsup_{\delta \to 0} \ensuremath{\mathds{E}} \bigg[ \sup_{y\in B_{R/\delta}} \big| \delta v^\delta(0,\cdot) + \overline H \big| \bigg]} \qquad & \\
& \leq \sum_{i=1}^k \limsup_{\delta \to 0} \ensuremath{\mathds{E}} \big| \delta v^\delta(y_i/\delta ,\cdot) + \overline H \big| + \limsup_{\delta\to 0} \ensuremath{\mathds{E}}\bigg[ \max_{1\leq i \leq k} \osc_{z \in B(y_i/\delta, \rho/\delta)} \delta v^\delta(z,\cdot) \bigg] \\
& \leq k\limsup_{\delta\to 0}\ensuremath{\mathds{E}} \big| \delta v^\delta(0,\cdot) + \overline H \big| + C\rho \\
& = C\rho.
\end{align*}
Disposing of $\rho>0$ yields \eqref{convprob1del}.
\end{proof}
\begin{remark}
It is clear that the subcorrector $w$ is locally Lipschitz in $\ensuremath{\mathds{R}^\d}$ with a constant controlled by $V$. Indeed, recall from \eqref{weaklims} that since $\Phi=Dw$ is the weak limit of $Dv^{\delta_j}$, we deduce from \eqref{locesteq2} that for a.e. $z\in B(y,1/2)$,
\begin{equation}
|Dw(z,\omega)|^\gamma \leq \limsup_{j\to \infty} |Dv^{\delta_j}(z,\omega)|^\gamma \leq C\Big( 1+ \sup_{B(y,1)} V(\cdot,\omega) \Big) \quad \mbox{a.s. in} \ \omega.
\end{equation}
Hence
\begin{equation} \label{wLIP}
\esssup_{B(y,1/2)} |Dw(\cdot,\omega)|^\gamma \leq C\Big( 1+ \sup_{B(y,1)} V(\cdot,\omega) \Big) \quad \mbox{a.s. in} \ \omega.
\end{equation}
\end{remark}
\begin{remark} \label{assubseq}
From \eqref{mainstepeq} we deduce the existence of a subsequence $\delta_j \to 0$ and a subset $\Omega_6 \subseteq \Omega_5$ of full probability such that, for every $\omega\in \Omega_6$,
\begin{equation}\label{assubseqeq}
\lim_{j \to \infty} \sup_{y\in B_{R/\delta_j}} \left| \delta_j v^{\delta_j} (y,\omega;p) + \overline H(p)\right| =0.
\end{equation}
By a diagonalization procedure, and by intersecting the relevant subsets of $\Omega$, we may assume that \eqref{assubseqeq} holds for every $R> 0$, rational $p$ and $\omega\in \Omega_6$. Then by Lemma~\ref{CDE-noxd} we obtain \eqref{assubseqeq} for all $p\in \ensuremath{\mathds{R}^\d}$ and $\omega \in \Omega_6$.
\end{remark}
\begin{remark} \label{WDlimsup}
Proposition~\ref{mainstep} is a partial result in the direction of our ultimate goal, which is \eqref{cplim}. In light of \eqref{oneside}, the key step that remains is to show that $\limsup_{\delta\to 0} \delta v^\delta(0,\omega; p) = -\overline H(p)$. While we must postpone the proof of this fact until we have studied the metric problem, let us note here the quantity $\limsup_{\delta\to 0} \delta v^\delta(0,\omega; p)$ is at least deterministic, a.s. in $\omega$. Indeed, observe that, according to Lemma~\ref{impest} and the stationarity of the $v^\delta$'s, for each $\mu \in \ensuremath{\mathds{R}}$, the set
\begin{equation*}
\Big\{ \omega\in \Omega_3: \limsup_{\delta \to 0} \delta v^\delta(0,\omega) \geq \mu \Big\} = \Big\{ \tau_{y} \omega\in \Omega_3: \limsup_{\delta \to 0} \delta v^\delta(y,\omega) \geq \mu \Big\}
\end{equation*}
is invariant under $\tau_y$ for every $y\in \ensuremath{\mathds{R}^\d}$. Therefore, the ergodic hypothesis implies that each of these sets has probability 0 or 1. Letting $-\hat H=-\hat H(p)$ denote the supremum over $\mu$ for which the set above has full probability, we obtain
\begin{equation}\label{hatHdefeq}
\limsup_{\delta\to0} \delta v^\delta(0,\omega;p) = -\hat H(p) \quad \mbox{a.s. in} \ \omega.
\end{equation}
Select $\Omega_7 \subseteq \Omega_6$ such that \eqref{hatHdefeq} holds for all $\omega \in \Omega_7$.
\end{remark}
\subsection{Some properties of $\overline H$} \label{Hbarprop}
The effective nonlinearity $\overline H$ inherits the convexity, coercivity and continuity of $H$. We record these elementary observations in the next proposition. We conclude this subsection with a discussion of an interesting property $\overline H$ possesses in a particular case in the random setting, which strongly contrasts with the situation in periodic and almost periodic media.
\begin{prop} \label{effHam}
The effective nonlinearity $\overline H$ has the following properties:
\begin{enumerate}
\item[(i)] $p\mapsto \overline H(p)$ is convex,
\item[(ii)] with $c_0 > 0$ as in \eqref{Hcoer} and for all $p \in \ensuremath{\mathds{R}^\d}$
\begin{equation} \label{Hbarcoer}
\frac12 c_0 |p|^\gamma - C \leq \overline H(p) \leq C\left(1+|p|^\gamma\right),
\end{equation}
\item[(iii)] for all $p_1,p_2\in\ensuremath{\mathds{R}^\d}$,
\begin{equation} \label{contpeq}
\left| \overline H(p_1) - \overline H(p_2) \right| \leq C \! \left(1 + |p_1|+|p_2| \right)^{\gamma-1} |p_1-p_2|.
\end{equation}
\end{enumerate}
\end{prop}
\begin{proof}
Select $p_1,p_2,x\in \ensuremath{\mathds{R}^\d}$ and, for each $\delta > 0$, write $v_i^\delta (y,\omega) : = v^\delta(y,\omega; p_i)$. Set $p : = \frac12(p_1+p_2)$ and $\tilde v^\delta : = \frac12( v^\delta_1 + v^\delta_2)$. Applying Lemma~\ref{convtrick}, we observe that $\tilde v^\delta$ satisfies
\begin{equation*}
\delta \tilde v^\delta - \tr \! \left( A(y,\omega) D^2 \tilde v \right) + H( p+ D\tilde v, y, \omega) \leq 0 \quad \mbox{in} \ \ensuremath{\mathds{R}^\d}.
\end{equation*}
Lemma~\ref{comparison} yields $\tilde v^\delta(y,\omega) \leq v^\delta(y,\omega;p)$ a.s. in $\omega$. Multiplying this inequality by $-\delta$ and sending $\delta \to 0$, we obtain
\begin{equation*}
\overline H(p) \leq \tfrac12 \overline H(p_1) + \tfrac12 \overline H(p_2).
\end{equation*}
and (i) follows.
\medskip
The upper bound in \eqref{Hbarcoer} follows immediately from \eqref{delvestbel}. The coercivity of $\overline H$ is more subtle, requiring an integration by parts. To this end, we may assume, by regularizing the coefficients and making the diffusion matrix $A$ uniformly elliptic, that $v^\delta$ is smooth. Integrating \eqref{aux} over $B_1$ and taking expectations yields
\begin{equation} \label{coerint}
\ensuremath{\mathds{E}} \fint_{B_1} -\delta v^\delta(y) \, dy = \ensuremath{\mathds{E}} \fint_{B_1}\left( - \tr\!\left( A(y,\omega) D^2v^\delta \right) + H(p+Dv^\delta, y, \omega) \right) \, dy,
\end{equation}
while Proposition~\ref{mainstep} gives
\begin{equation} \label{Hcoerc1}
\lim_{\delta \to 0} \ensuremath{\mathds{E}} \fint_{B_1} -\delta v^\delta \, dy = \overline H(p).
\end{equation}
Integrating by parts and using \eqref{Alip} and \eqref{locesteq2}, we can estimate the first term on the right side of \eqref{coerint} pointwise in $\omega$ by
\begin{equation*}
\left| \int_{B_1} \tr \!\left( A(y,\omega) D^2v^\delta \right) \, dy \right| \leq C \int_{B_1} |Dv^\delta(y)|\, dy + C\sup_{y\in \partial B_1} |Dv^\delta(y)| \leq C \sup_{y\in B_2} V(y,\omega).
\end{equation*}
Taking expectations yields
\begin{equation} \label{diffest}
\ensuremath{\mathds{E}} \left| \fint_{B_1} \tr \!\left( A(y,\omega) D^2v^\delta \right) \, dy \right| \leq C \, \ensuremath{\mathds{E}} \sup_{B_2} V(y,\omega) \leq C.
\end{equation}
The second term on the right side of \eqref{coerint} is estimated from below using Jensen's inequality, \eqref{Hcoer} and \eqref{locesteq2}, as follows:
\begin{align}
\ensuremath{\mathds{E}} \fint_{B_1} H(p+Dv^\delta ,y,\cdot) \, dy & \geq \ensuremath{\mathds{E}} \fint_{B_1} \left( c_0 |p+Dv^\delta|^\gamma - V(y,\cdot) - C_0 \right) dy \nonumber \\
& \geq c_0 \left| \, p + \ensuremath{\mathds{E}} \fint_{B_1} Dv^\delta(y) \, dy\,\right|^\gamma - C \nonumber \\
& \geq \frac{1}{2} c_0 |p|^\gamma - \ensuremath{\mathds{E}} \sup_{y \in B_1} |Dv^\delta(y)|^\gamma \, dy - C \nonumber \\
& \geq \frac12 c_0 |p|^{\gamma} - C. \label{Hcoerc2}
\end{align}
Sending $\delta \to 0$ in \eqref{coerint} using \eqref{Hcoerc1}, \eqref{diffest} and \eqref{Hcoerc2} yields the lower bound in \eqref{Hbarcoer}.
\medskip
Finally, we note that (iii) follows from Lemma~\ref{CDE-noxd} by simply sending $\delta \to 0$ in \eqref{CDE-noxdeq}.
\end{proof}
\begin{remark} \label{effHcmp}
The properties of $\overline H$ enumerated in Proposition~\ref{effHam} imply, in view of standard viscosity solution theory, that the problem \eqref{HJ-eff} has a unique solution.
\end{remark}
We conclude this section with some observations about the effective Hamiltonian in the particular case that, for all $p,y \in \ensuremath{\mathds{R}^\d}$,
\begin{equation} \label{parcas}
H(p,y,\omega) \geq H(0,y,\omega) \quad \mbox{a.s. in} \ \omega.
\end{equation}
It is essentially known (at least in the bounded case, see for example \cite{LPV,LS2}) that for separated Hamiltonians, i.e., if
\begin{equation}\label{sepcase}
H(p,y,\omega) = \widetilde H(p) - V(y,\omega),
\end{equation}
where $\widetilde H(p) \geq \widetilde H(0)$ and $A\equiv 0$, we have
\begin{equation} \label{Hbarinv}
\overline H(p) \geq \overline H(0) = -\essinf_{\Omega} V(0,\cdot).
\end{equation}
Here we prove the following more general fact.
\begin{prop} \label{invispot-easy-prop}
Assume that $A\equiv 0$ and \eqref{parcas}. Then
\begin{equation} \label{aspacsas}
\overline H(p) \geq \overline H(0) = \esssup_{\omega \in \Omega} H(0,0,\omega).
\end{equation}
\end{prop}
\begin{proof}
We may assume that $\esssup_\Omega H(0,0,\cdot) = 0$. Fix $p\in \ensuremath{\mathds{R}^\d}$ and $\varepsilon > 0$, and define
\begin{equation*}
\widetilde\Omega_\varepsilon : = \{ \omega \in \Omega : H(0,0,\omega) > - \varepsilon \}.
\end{equation*}
Observe that $\ensuremath{\mathds{P}}\big[\widetilde\Omega_\varepsilon\big] > 0$. We claim that, for any $\delta > 0$ and $\omega \in \widetilde\Omega_\varepsilon \cap \Omega_1$, where $\Omega_1$ is as in Lemma~\ref{comparison}, we have
\begin{equation*}
\delta v^\delta(0,\omega; p) \leq 2\varepsilon.
\end{equation*}
To prove this, we construct a very simple barrier. In fact we may take any smooth function $w$ satisfying $w \geq w(0) = 2\delta^{-1} \varepsilon$ in $B_r$ and $w(y)\to \infty$ as $|y| \to r$. If $r > 0$ is small enough, depending on $\omega$, then, by continuity,
\begin{equation*}
\max_{B_r} H(0,0,\omega) > - 2\varepsilon.
\end{equation*}
It follows from this, $A\equiv 0$ and \eqref{parcas} that $w$ is a supersolution of \eqref{aux} in $B_r$. Thus the comparison principle gives $\delta v^\delta(0,\omega;p) \leq \delta w(0) = 2\varepsilon$.
Since $\ensuremath{\mathds{P}}\big[\widetilde\Omega_\varepsilon\cap \Omega_1 \big] > 0$, by \eqref{mainstepeq} we obtain $\overline H(p) \geq -2\varepsilon$.
Sending $\varepsilon \to 0$ yields $\overline H(p) \geq 0$. It follows that $\overline H(0) = 0$, since it is clear that
\begin{equation*}
\overline H(0) \leq \esssup_\Omega H(0,0,\cdot) = 0. \qedhere
\end{equation*}
\end{proof}
\begin{remark} \label{invispot-easy}
One observes from \eqref{Hbarinv} the reason we must require $V$ to be bounded from below, even as our assumptions allow it to be unbounded from above.
\end{remark}
The situation is different in the presence of diffusion, i.e., if $A\not\equiv 0$. For example, an easy calculation shows in the periodic setting, with $A = I_\d$ and $H$ separated as in \eqref{sepcase}, that \eqref{Hbarinv} holds if and only if $V$ is constant. We nonetheless observe in the next proposition that, in the latter case, \eqref{Hbarinv} holds if $V$ has arbitrarily large ``bare spots" (defined implicitly below) near its essential infimum. This phenomenon is special to the random setting. It is true for a Poissonian potential, for instance, while in contrast, any periodic or almost periodic potential having arbitrarily large bare spots is necessarily constant.
\begin{prop} \label{invispot}
Suppose that $H$ is of the form \eqref{sepcase} and, for all $p\in \ensuremath{\mathds{R}^\d}$,
\begin{equation} \label{tH}
\widetilde H(p) \geq \widetilde H(0) = 0,
\end{equation}
and, for each $\mu,R> 0$,
\begin{equation} \label{barespot}
\ensuremath{\mathds{P}}\left[ \sup_{y\in B_R} V(y,\cdot) \leq \mu \right] > 0.
\end{equation}
Then $\overline H \geq 0$, and, if, in addition, we assume that $V\geq 0$, then $\overline H(0) = 0$.
\end{prop}
\begin{proof}
Fix $\mu> 0$ and $R> 1$, and set $E: = \{ \omega \in \Omega : \sup_{B_R} V(y,\omega) \leq \mu \}$. The ergodic theorem yields an event $\widetilde\Omega\subseteq \Omega$ of full probability such that, for each $\omega \in \widetilde\Omega$,
\begin{equation} \label{eprtuoa}
\lim_{r\to \infty} \fint_{B_r} \mathds{1}_E (\tau_y\omega) \, dy = \ensuremath{\mathds{P}}[E] > 0.
\end{equation}
Fix $\omega \in \widetilde\Omega$. According to \eqref{eprtuoa}, for every sufficiently small $\delta$, depending on $\omega$, we can find $z \in B_{1/\delta}$ for which $\tau_z\omega \in E$. This implies that
\begin{equation} \label{barespotz}
\sup_{B(z,R)} V(\cdot ,\omega) = \sup_{B(0,R)} V(\cdot, \tau_z\omega) \leq \mu.
\end{equation}
Using a barrier function similar to the one in the proof of Lemma~\ref{locest}, we estimate $\delta v^\delta (z,\omega; p)$ from above. Translating the equation, we may assume that $z=0$. Notice that the hypotheses on $H$ imply that
\begin{equation} \label{Htform}
H(p,y,\omega) \geq \left( c_0 |p|^\gamma - C_0\right)_+ - V(y,\omega).
\end{equation}
Since the calculations below rely only on \eqref{Htform}, we assume that $\gamma < 2$. With positive constants $a,b> 0$ to be chosen, define
\begin{equation*}
w(y) : = a + b \left( R^2 - |y|^2 \right)^{-\eta} -bR^{-2\eta},
\end{equation*}
where $\eta : = (2-\gamma) / (\gamma-1)$, so that $\gamma(\eta+1) = \eta+2$. Routine calculations give
\begin{equation*}
|Dw(y)|^\gamma \geq c b^\gamma |y|^\gamma \left( R^2- |y|^2 \right)^{-(\eta+2)} \quad \mbox{and} \quad \tr\!\left(AD^2w(y)) \leq C b R^2 (R^2-|y|^2 \right)^{-(\eta+2)}.
\end{equation*}
Inserting $w$ into \eqref{aux} and using \eqref{barespotz} and \eqref{Htform}, we find
\begin{multline}
\delta w - \tr(AD^2w) + H(p+Dw,y,\omega)
\\ \geq \delta a - CbR^2 (R^2-|y|^2)^{-(\eta+2)} + \left( cb^\gamma |y|^\gamma \left( R^2- |y|^2 \right)^{-(\eta+2)} - C_0\right)_+ - \mu \quad \mbox{in} \ B_R.
\end{multline}
Taking $b:= CR^\eta$ for a sufficiently large constant $C$, we get
\begin{equation*}
\left( cb^\gamma |y|^\gamma \left( R^2- |y|^2 \right)^{-(\eta+2)} - C\right)_+ \geq CbR^2 \left( R^2- |y|^2 \right)^{-(\eta+2)} \quad \mbox{for all} \quad \frac R2 \leq |y| < R.
\end{equation*}
Of course,
\begin{equation*}
CbR^2 \left( R^2- |y|^2 \right)^{-(\eta+2)} \leq CbR^{-2(\eta+1)} \leq CR^{-2-\eta} \quad \mbox{for all} \quad |y | \leq \frac R2.
\end{equation*}
Therefore in $B_R$, we have
\begin{equation*}
\delta w - \tr(AD^2w) + H(p+Dw,y,\omega) \geq \delta a -\mu - CR^{-2-\eta}.
\end{equation*}
If $a := \delta^{-1}(\mu + CR^{-2-\eta})$, then $w$ is a supersolution of \eqref{aux}. Since $w(y) \rightarrow +\infty$ as $|y| \to R$, the comparison principle implies that
\begin{equation*}
\delta v^\delta(0,\omega;p) \leq \delta w(0) = \delta a = \mu + CR^{-2-\eta}.
\end{equation*}
Taking expectations and passing to the limit $\delta \to 0$, using \eqref{mainstepeq}, yields that
\begin{equation*}
\overline H(p) \geq -\mu - CR^{-2-\eta}.
\end{equation*}
Sending $\mu \to 0$ and $R\to \infty$ gives the conclusion.
\end{proof}
In the case that $\gamma=2$, $\overline H(0)$ is related to the bottom of the spectrum of random Schrodinger operators, and the previous proposition is known. See for instance Carmona and Lacroix \cite{CL} as well as Section 8 of \cite{LS2}.
Sznitman \cite{Sz2} proved more than Proposition~\ref{invispot} in the case that $A=I_\d$, $\widetilde H(p) = |p|^2$ and $V$ is a Poissonian potential. In particular, he showed that $\overline H$ has a ``flat spot" near $p=0$, that is, $\overline H(p) = 0$ for small $|p|$, a result which in his language he calls the ``nondegeneracy of the quenched Lyapunov exponents" (see \cite[Proposition 5.2.9]{Szb}). See also Section~\ref{P} below for a guide to translating between our notation and that of \cite{Szb}.
\medskip
\begin{remark}\label{approximation}
It is sometimes convenient to approximate $\overline H$ by effective Hamiltonians corresponding to bounded environments. For $c_0$ as in \eqref{Hcoer}, define
\begin{equation} \label{Hk}
H_k(p,y,\omega) : = \max\big\{ H(p,y,\omega), c_0|p|^\gamma - k \big\}.
\end{equation}
It is clear that $H_k$ satisfies the same hypotheses as $H$, and, in addition, $H_k$ is uniformly coercive, i.e., $H_k(p,y,\omega) \geq c_0|p|^\gamma - k$. Obviously $H_k \geq H$ and $H_k \downarrow H$ as $k\to \infty$. In fact, by Lemma~\ref{potests}, there is a constant $\sigma < 1$ such that
\begin{equation}
0 \leq H_k(p,y,\omega) - H(p,y,\omega) \leq C_0 + V(y,\omega) - k \leq C\big(1+|y|^\sigma\big) - k \quad \mbox{a.s. in} \ \omega.
\end{equation}
To each $H_k$ corresponds an ${\overline H}_k$ and, in view of the monotonicity of the $H_k$'s with respect to $k$, it is obvious that ${\overline H}_k \downarrow \hat H$ as $k\to \infty$. For bounded environments it is easy to check using standard viscosity arguments that ${\overline H} = \hat H$.
The argument is more complicated in the unbounded setting. We leave it up to the interested reader to fill in the details.
\end{remark}
We conclude this section by writing down the usual inf-sup formula for the effective Hamiltonian:
\begin{equation}\label{infsup}
\overline H(p) = \inf_{\Phi \in \mathcal S} \sup_{y\in \ensuremath{\mathds{R}^\d}} \big( -\tr A(y,\omega)D^2\Phi + H(p+D\Phi,y,\omega) \big).
\end{equation}
Here $\mathcal S$ is the set of $\Phi : \ensuremath{\mathds{R}^\d} \times \Omega \rightarrow \ensuremath{\mathds{R}}$ such that $\Phi(\cdot,\omega)$ is locally Lipschitz a.s. in $\omega$ and $D\Phi$ is stationary with $\ensuremath{\mathds{E}} [D\Phi ]= 0$. The supremum is to be understood in the viscosity sense, i.e., it is defined to be the smallest constant $k$ for which the inequality
\begin{equation*}
-\tr A(y,\omega)D^2\Phi + H(p+D\Phi,y,\omega) \leq k \quad \mbox{in} \ \ensuremath{\mathds{R}^\d}
\end{equation*}
holds in the viscosity sense.
The proof of \eqref{infsup} is simple. The inequality ``$\leq$" is clear from the existence of the subcorrector in Proposition~\ref{mainstep}. If this inequality were strict, however, we could repeat the argument used to obtain \eqref{cmpineq} in the proof of Proposition~\ref{mainstep}, with an appropriate choice of $\Phi$ in place of $w$, to obtain an improvement of \eqref{cmpineq}. But this would contradict \eqref{oneside}.
\section{The metric problem: another characterization of $\overline H$}
\label{Q}
\nocite{R,KMZ}
In this section we consider the special stationary equation
\begin{equation} \label{MP}
- \tr\! \left( A(y,\omega) D^2u\right) + H(p+Du, y, \omega) = \mu \quad \mbox{in} \ \ensuremath{\mathds{R}^\d}\!\setminus \! D
\end{equation}
where $D$ is a bounded, closed subset of $\ensuremath{\mathds{R}^\d}$ and $\mu \in \ensuremath{\mathds{R}}$. Subject to appropriate boundary and growth conditions, a solution $u$ of \eqref{MP} is related to the ``metric" (distance function) associated with the effective Hamiltonian.
\medskip
Our motivation for studying \eqref{MP} is threefold. First, as we will see, solutions of \eqref{MP} possess some subadditive structure. An application of the subadditive ergodic theorem therefore permits us to essentially homogenize a rescaled version of \eqref{MP}. Working backwards we are able to improve the convergence in $L^1$ obtained in \eqref{mainstepeq} to almost sure convergence, which is done later in Section~\ref{BS}. This is a critical step in the proof of Theorem~\ref{MAIN}. Second, we provide a characterization of $\overline H$ in terms of the solvability of \eqref{MP}, which is new even in the bounded setting. We hope that this formula will yield new information about the structure of the effective Hamiltonian, and we intend to return to this point in future work. Finally, the metric problem is natural from the probability point of view, and allows us to make a precise connection to the results of Sznitman in the case of the model equation \eqref{PCP} with $\gamma=2$ and a Poissonian potential $V$.
\medskip
We remark that, in this section, the subsets of $\Omega$ of full probability on which our ``almost sure" assertions hold depend on $p$. This is because, for the sake of brevity and because it is not required for the proof of our main theorems, we do not wish to trouble ourselves with proving a separate continuous dependence result.
\subsection{Well-posedness}
We begin by showing that \eqref{MP} is well-posed for each $\mu > \overline H(p)$, a.s. in $\omega$, subject to appropriate growth conditions at infinity and the boundary condition $u=0$ on $\partial D_1$, where $D=D_1$ is defined by
\begin{equation*}
D_1 : = \begin{cases} \{ 0 \} & \mbox{if} \ \gamma > 2 \ \mbox{or} \ A \equiv 0, \\ B_1 & \mbox{if} \ \gamma \leq 2 \ \mbox{and} \ A \not\equiv 0.\end{cases}
\end{equation*}
The reason for defining $D_1$ in this way, from the probability point of view, is that, in contrast to the case $\gamma\leq 2$, for $\gamma > 2$, it is possible to have Brownian bridges. That is, we may connect two points via a diffusion if $\gamma > 2$, while if $\gamma=2$ we may only connect a point to a small ball. From the pde point of view, this manifests itself in the kind of barrier functions we are able to build.
\medskip
In the next proposition, we prove a comparison result. Existence then follows from the Perron method once suitable barriers have been constructed.
\medskip
We prove the next proposition with a new argument that, as far as we know, has no analogue in the literature. What makes this results different from typical comparison results on unbounded domains is that we do \emph{not} assume that the subsolution $u$ and supersolution $v$ separate only sublinearly at infinity; see \eqref{meteqcmpgc}. In the proof, we ``lower" $u$ until it has strictly sublinear separation from $v$, and then argue that, if we needed to lower $u$ at all, we could have lowered it a little less. To accomplish this, first we perturb $u$ by subtracting a term $\varphi_R$ which is negligible in balls of radius $\sim R$ but grows linearly at infinity. Then we compare the resulting
function with $v$. We conclude sending $R \to \infty$. The fact that the parameter $\mu$ is strictly larger than $\overline H$ allows us to compensate for this perturbation by using the subcorrector.
\begin{prop} \label{metcomparison}
Fix $p\in \ensuremath{\mathds{R}^\d}$, $\mu > \overline H(p)$ and $\omega \in\Omega_7$. Suppose that $D\subseteq \ensuremath{\mathds{R}^\d}$ is closed and bounded. Assume that $u\in\USC(\overline{ \ensuremath{\mathds{R}^\d}\!\setminus \!D})$ and $v\in \LSC(\overline{ \ensuremath{\mathds{R}^\d}\!\setminus\! D})$ are, respectively, a subsolution and supersolution of \eqref{MP} such that $u \leq v$ on $\partial D$, and
\begin{equation} \label{meteqcmpgc}
\limsup_{|y| \to \infty} \frac{u(y)}{|y|} < \infty \quad \mbox{and} \quad \liminf_{|y| \to \infty} \frac{v(y)}{|y|} \geq 0.
\end{equation}
Then $u \leq v$ in $\ensuremath{\mathds{R}^\d} \!\setminus \! D$.
\end{prop}
\begin{proof}
We may assume that $\limsup_{|y| \to \infty} u(y)/|y| \geq 0$, since otherwise the result is immediate from the usual comparison principle (c.f. \cite{CIL}). We omit all dependence on $\omega$, since it plays no role in the argument. We may also assume that $p=0$. Define
\begin{equation*}
\Lambda : = \Big\{ 0 \leq \lambda \leq 1 : \liminf_{|y| \to \infty} \frac{v(y) - \lambda u(y)}{|y|} \geq 0 \Big\} \qquad \mbox{and} \qquad \overline \lambda: = \sup \Lambda.
\end{equation*}
We begin with the observation that $\Lambda = \big[0,\overline \lambda \big]$. This follows from \eqref{meteqcmpgc} and a continuity argument. The assumption \eqref{meteqcmpgc} implies $0 \in \Lambda$. To see that $\overline \lambda \in \Lambda$, select $\varepsilon > 0$ and $\lambda \in \Lambda$ such that $\overline \lambda \leq \lambda + \varepsilon$, and observe that, by \eqref{meteqcmpgc},
\begin{multline*}
\liminf_{|y| \to \infty} \frac{v(y) - \overline \lambda u(y)}{|y|} \geq \frac{\overline \lambda}{\lambda+\varepsilon} \liminf_{|y| \to \infty} \frac{v(y) - (\lambda+\varepsilon)u(y)}{|y|} \\ \geq \frac{\overline \lambda}{\lambda+\varepsilon} \left( -\varepsilon \limsup_{|y| \to\infty} \frac{u(y)}{|y|}\right) \geq -C\overline \lambda \varepsilon(\lambda+\varepsilon)^{-1}.
\end{multline*}
Sending $\varepsilon \to 0$ yields $\overline \lambda \in \Lambda$. If $\lambda \in \big(0,\overline \lambda\big)$, then using again \eqref{meteqcmpgc}, we have
\begin{equation*}
\liminf_{|y| \to \infty} \frac{v(y) - \lambda u(y)}{|y|} \geq \frac{\lambda}{\overline \lambda} \liminf_{|y| \to \infty} \frac{v(y) - \overline \lambda u(y)}{|y|} \geq 0,
\end{equation*}
and, hence, the claim.
\medskip
Next we show that $\overline \lambda = 1$. Select $\lambda < \overline \lambda \leq 1$. For each $R> 1$, define the auxiliary function
\begin{equation} \label{varphiR}
\varphi_R(y): = R- (R^2+|y|^2)^{1/2},
\end{equation}
and observe that, for a constant $C > 0$ independent of $R> 1$,
\begin{equation} \label{phidcn}
|D\varphi_R(y)| + |D^2\varphi_R(y)| \leq C.
\end{equation}
We have defined $\varphi_R$ in such a way that $-\varphi_R$ grows at a linear rate at infinity, which is independent of $R$, while $\varphi_R \to 0$ as $R \to \infty$. Indeed, it is easy to check that
\begin{equation} \label{phideath}
|\varphi_R(y)| \leq |y|^2 \left( R^2 + |y|^2 \right)^{-1/2}.
\end{equation}
Fix constants $0 < \eta < 1$ and $\theta > 1$ to be selected below. By \eqref{Hcoer} and \eqref{phidcn},
\begin{equation} \label{Ephi}
\left|\theta \tr\!\left(A(y) D^2\varphi_R\right)\right| + H(-\theta D\varphi_R,y) \leq C\theta^\gamma.
\end{equation}
Define
\begin{equation*}
\hat u : = \lambda(1+4\eta) u+ (1-\lambda(1+4\eta)) w
\end{equation*}
and
\begin{align*}
\hat u_R : = &\ (1-\eta) \hat u + \eta \theta \varphi_R \\ = &\ \lambda(1+4\eta)(1-\eta)u + (1-\eta)(1 - \lambda(1+4\eta))w + \eta \theta \varphi_R,
\end{align*}
where $w$ is the subcorrector which satisfies \eqref{subcorreq} and \eqref{sublininfty}, constructed in the proof of Proposition~\ref{mainstep}. By subtracting a constant from $w$, we may assume that $\sup_{D} w = 0$.
Since $\lambda < 1$, we may shrink $\eta$, if necessary, to ensure that
\begin{equation} \label{etachoice}
\lambda(1+4\eta) < 1 < (1+4\eta)(1-\eta).
\end{equation}
Select $\theta : = 1+4 \limsup_{|y| \to \infty} u(y) / |y|$, and observe that, by the previous inequality, \eqref{sublininfty}, $\lambda \in \Lambda$ and the definition of $\varphi_R$, we have, for every $R> 1$,
\begin{equation}
\liminf_{|y| \to \infty} \frac{v(y) - \hat u_R(y)}{|y|} \geq \liminf_{|y| \to \infty} \frac{v(y) - \lambda u(y)}{|y|} + \liminf_{|y| \to \infty} \frac{(3\eta-4\eta^2) u(y) + \eta\theta \varphi_R(y)}{|y|} > 0.
\end{equation}
\medskip
To get a differential inequality for $\hat u_R$, we apply Lemma~\ref{convtrick} twice. The first application, using \eqref{subcorreq} and that $u$ is a subsolution of \eqref{MP}, yields that
\begin{equation*}
-\tr\!\left(A(y) D^2 \hat u \right) + H(D\hat u,y) \leq \lambda(1+4\eta) \mu + (1-\lambda(1+4\eta)) \overline H(0) \quad \mbox{in} \ \ensuremath{\mathds{R}^\d}\! \setminus \! D.
\end{equation*}
Combining this with \eqref{Ephi}, we deduce that
\begin{equation*}
-\tr\!\left(A(y) D^2 \hat u_R \right) + H(D\hat u_R,y) \leq \widetilde\mu(\eta) \quad \mbox{in} \ \ensuremath{\mathds{R}^\d}\! \setminus \! D,
\end{equation*}
where the constant $\widetilde\mu(\eta)$ is given by
\begin{equation*}
\widetilde \mu(\eta) := \lambda(1+4\eta)(1-\eta)\mu + (1-\eta)(1-\lambda(1+4\eta))\overline H(0) + C\theta^\gamma \eta.
\end{equation*}
Since $\mu > \overline H(0)$, by selecting $\eta> 0$ sufficiently small we have $\widetilde\mu(\eta) < \mu$. We may therefore apply the comparison principle to $\hat u_R$ and $v$, obtaining that
\begin{equation*}
\hat u_R - v \leq \max_{\partial D} (\hat u_R - v) \quad \mbox{in} \ \ensuremath{\mathds{R}^\d}\!\setminus \! D.
\end{equation*}
By sending $R\to \infty$ and using the fact that $\varphi_R\rightarrow 0$ locally uniformly, we deduce that
\begin{equation} \label{hatuvinq}
(1-\eta) \hat u - v \leq \max_{\partial D} \big( (1-\eta) \hat u - v\big) \quad \mbox{in} \ \ensuremath{\mathds{R}^\d}\! \setminus\! D.
\end{equation}
Since $w$ is strictly sublinear at infinity, it follows that
\begin{equation*}
\liminf_{|y| \to \infty} \frac{v(y) - \lambda(1+4\eta)(1-\eta)u(y)}{|y|} \geq 0,
\end{equation*}
and, hence, $\overline \lambda \geq \lambda(1+4\eta)(1-\eta)$. If $\overline \lambda < 1$, then we may send $\lambda \to \overline \lambda$ to obtain that $\overline \lambda \geq \overline \lambda(1+4\eta)(1-\eta)$, a contradiction to \eqref{etachoice}. It follows that $\overline \lambda =1$.
\medskip
The preceding analysis therefore applies to any $0 <\lambda < 1$. Sending $\eta \to 0$ and then $\lambda \to 1$ in \eqref{hatuvinq} completes the proof.
\end{proof}
With a comparison principle in hand, the unique solvability of the metric problem \eqref{MP} for $\mu > \overline H(p)$ (subject to the boundary and growth conditions) follows from Perron's method. The proof in the case of bounded $V$ is accomplished more smoothly. In the unbounded case we approximate by bounded potentials.
\begin{prop} \label{metexistence}
For each fixed $p,z \in\ensuremath{\mathds{R}^\d}$ and $\mu > \overline H(p)$, there exists a unique solution ${m}_\mu={m}_\mu(\cdot,z,\omega) = {m}_\mu(\cdot,z,\omega;p) \in C(\overline{ \ensuremath{\mathds{R}^\d} \!\setminus\! D_1(z)})$ of \eqref{MP}, a.s. in $\omega$, with $D=D_1(z) : = z+D_1$, subject to the conditions
\begin{equation} \label{bngcond}
{m}_\mu(\cdot,z,\omega) = 0 \quad \mbox{on} \ \partial D_1(z) \quad \mbox{and} \quad 0 \leq \liminf_{|y|\to\infty} \frac{{m}_\mu(y,z,\omega)}{|y|} \ \quad \mbox{a.s. in} \ \omega.
\end{equation}
\end{prop}
\begin{proof}
Assume, without loss of generality, that $p=0$ and $z=0$. We first argue in the case that $V$ is bounded, i.e., $V(0,\omega) \leq C$ a.s. in $\omega$. According to the first step in the proof of Proposition~\ref{mainstep}, there exists a subcorrector $w$ which satisfies, a.s. in $\omega$, the inequality
\begin{equation*}
-\tr\!\left( A(y,\omega) D^2w \right) + H(Dw,y,\omega) \leq \overline H(0) \quad \mbox{in} \ \ensuremath{\mathds{R}^\d},
\end{equation*}
as well as
\begin{equation} \label{wsublinpmp}
\lim_{|y|\to\infty} \frac{w(y)}{|y|} = 0.
\end{equation}
By subtracting a constant from $w$, we may assume that $\inf_{D_1} w(\cdot,\omega) = 0$. Define
\begin{multline*}
m_\mu(y) : = \inf\Big\{ v(y) : v\in \LSC(\ensuremath{\mathds{R}^\d}\!\setminus \!D_1), \ v \ \mbox{is a supersolution of} \ \eqref{MP} \ \mbox{in} \ \ensuremath{\mathds{R}^\d}\!\setminus \! D_1,\\ v \geq 0 \ \mbox{on} \ \partial D_1, \ \mbox{and} \ \liminf_{|y| \to \infty} v(y) / |y| \geq 0 \Big\}.
\end{multline*}
To demonstrate that $m_\mu$ is well-defined, we must show that the admissible set is nonempty. Since $V$ is bounded, it is easy to check that, for $a> 0$ sufficiently large,
\begin{equation*}
v(y): =
\begin{cases}
a|y| & \mbox{if} \ A \equiv 0, \\
a(|y| + |y|^\theta) & \mbox{if} \ A \not\equiv 0 \ \ \mbox{and} \ \ \gamma > 2, \ \ \mbox{where} \ \ \theta: = (\gamma-2) / (\gamma-1),\\
a(|y|-1) & \mbox{if} \ A \not\equiv 0 \ \ \mbox{and} \ \ \gamma \leq 2,
\end{cases}
\end{equation*}
is an admissible supersolution, and moreover, $v = 0$ on $\partial D_1$ and $\limsup_{|y|\to \infty} |y|^{-1} v(y) = a < \infty$. Thus $m_\mu$ is well-defined, and we have $w \leq m_\mu \leq v$ in $\ensuremath{\mathds{R}^\d} \setminus D_1$. It follows from standard viscosity theory
that $(m_\mu)^*$ is a subsolution of \eqref{MP} and $(m_\mu)_*$ is a supersolution. Since these functions satisfy the hypotheses of Proposition~\ref{metcomparison}, we obtain that $(m_\mu)^*\leq (m_\mu)_*$, and therefore $(m_\mu)^* = (m_\mu)_*=m_\mu \in C(\ensuremath{\mathds{R}^\d} \!\setminus \! D_1)$ is a solution of \eqref{MP}-\eqref{bngcond}.
\medskip
By repeating the argument of Lemma~\ref{locest} we find that, for any $B(y_0,r+1/2) \subseteq \ensuremath{\mathds{R}^\d}\!\setminus \!D_1$,
\begin{equation} \label{MPLB}
\esssup_{B(y,r)} |Dm_\mu(y)|^\gamma \leq C \Big( 1 + \sup_{B(y,r+1/2)} V(y,\omega)\Big).
\end{equation}
Next, we mimic the proof of Lemma~\ref{impest}, using Morrey's inequality and Lemma~\ref{potests}, to obtain that, a.s in $\omega$, the oscillation of any solution of \eqref{MP} in a large ball grows at most like the radius of the ball. In particular,
\begin{equation} \label{metosc}
\limsup_{r\to \infty} \frac{1}{r} \osc_{B_r \! \setminus \! D_1} m_\mu \leq C,
\end{equation}
and $m_\mu$ grows at most like $C|y|$, where $C$ depends only the constants in our assumptions on $V$.
In the unbounded case, we approximate $H$ by bounded Hamiltonians as explained in Remark~\ref{approximation}. If $\mu > \hat H$, then for sufficiently large $k$ we have $\mu > \overline H_k$. For such $k$, there is a solution $m_{\mu,k}$ of the problem \eqref{MP}--\eqref{bngcond} for $H=H_k$, which satisfies $m_{\mu,k}(y) \leq C_1|y|$ for a constant $C_1$ independent of $k$. The comparison principle (Proposition~\ref{metcomparison}) implies that $m_{\mu,k}$ is monotone in $k$, so that $m_{\mu,k} \uparrow m_\mu$ as $k\to \infty$ to some function $m_\mu$. By the local Lipschitz estimate \eqref{MPLB}, $m_{\mu,k}$ converges locally uniformly (and a.s. in $\omega$) to $m_\mu$ which is therefore a solution of \eqref{MP}. We have $m_\mu(y) \leq C|y|$ a.s. in $\omega$, and by monotonicity we have that $\liminf_{|y| \to \infty} |y|^{-1}m_\mu(y) \geq 0$.
If $\hat H \geq \mu > \overline H$, we observe that the solution constructed above for any $\hat \mu >\hat H$ belongs to the admissible set used for the definition of $m_\mu$ at the beginning of the ongoing proof. Then we can repeat verbatim the arguments used for the bounded case to find a solution. In either case, uniqueness follows from Proposition~\ref{metcomparison}, and the proof is complete.
\end{proof}
\begin{remark} \label{MPstat}
The functions $m_\mu$ are jointly stationary in the sense that
\begin{equation*}
m_\mu(y,x,\tau_z\omega) = m_\mu(y+z,x+z,\omega).
\end{equation*}
This follows from uniqueness and the stationarity of the coefficients in the equation.
\end{remark}
\begin{remark} \label{lingrow}
It is easy to show that, for each $\mu > \overline H$,
\begin{equation*}
\essinf_{\omega\in \Omega}\liminf_{|y| \to \infty} \frac{{m}_\mu(y,0,\omega)}{|y|} > 0.
\end{equation*}
In brief, the idea is to choose $\overline H < \widetilde \mu < \mu$, add a small multiple of $\varphi_R$ (defined in \eqref{varphiR}) to ${m}_{\widetilde \mu}$, and compare it to ${m}_\mu$ using Proposition~\ref{metcomparison}. The room provided by taking $\widetilde \mu$ smaller than $\mu$ allows to compensate for the perturbation, following along the lines of the calculation in the proof of Proposition~\ref{metcomparison}. We will see in the next section that $t^{-1} {m}_\mu(ty,0,\omega)$ has a deterministic limit (a.s. in $\omega$) as $t\to \infty$, which is equivalent to the homogenization of the metric problem.
The argument above also implies that that ${m}_\mu$ is strictly increasing in $\mu$. This permits us to define
\begin{equation*}
{m} (y,x,\omega) : = \lim_{\mu \ssearrow \overline H} {m}_\mu(y,x,\omega).
\end{equation*}
\end{remark}
\begin{remark} \label{tildemetp}
By repeating the argument in the proof of Proposition~\ref{metexistence}, we obtain for each fixed $p,z \in \ensuremath{\mathds{R}^\d}$ and $\mu > \overline H(p)$ a unique solution $\widetilde m_\mu=\widetilde m_\mu(y,z,\omega;p)$ of the modified metric problem
\begin{equation} \label{meteqmod}
\left\{ \begin{aligned}
& -\tr\!\left( A(y , \omega) D^2\widetilde m_\mu \right) + H(p+D\widetilde m_\mu, y, \omega) = \mu \quad \mbox{in} \ \ \ensuremath{\mathds{R}^\d}\!\setminus \! D_1, \\
& \widetilde m_\mu(y,z,\omega;p) = w(y,\omega;p) - w(z,\omega;p) \ \ \mbox{on} \ \partial D_1, \\
& \liminf_{|y| \to \infty} |y|^{-1} \widetilde m_\mu(y,z,\omega) \geq 0 \ \ \mbox{a.s. in} \ \omega.
\end{aligned} \right.
\end{equation}
The only difference between $\widetilde m_\mu$ and the solution $m_\mu$ constructed in Proposition~\ref{metexistence} is the boundary conditions we impose on $\partial D_1$, which differ only in the case $A\not\equiv 0$ and $\gamma \leq 2$. To prove the existence of $\widetilde m_\mu$, there is an additional difficulty in the construction of barriers, but this is straightforward since $w$ is Lipschitz and $\partial D_1$ is the unit ball. We leave this detail to the reader. The function $\widetilde m_\mu$ is sometimes more convenient to work with than $m_\mu$, as we see below.
In the case $D_1=B_1$, in which $\widetilde m_\mu(y,z,\omega)$ is not defined for $|y-z| <1$, we extend $\widetilde m_\mu(\cdot,\omega)$ to $\ensuremath{\mathds{R}^\d}\times\ensuremath{\mathds{R}^\d}$ by defining $\widetilde m_\mu (y,z,\omega) : = w(y,\omega) - w(z,\omega)$ for $|y-z| < 1$.
\end{remark}
\subsection{The subadditivity of ${m}_\mu$}
\label{sadd}
An important property of ${m}_\mu$ is that, up to a modification in the case $\gamma \leq 2$ and $A\not\equiv 0$, it is \emph{subadditive}. This is easier to see in the case that $\gamma > 2$ or $A\equiv 0$, i.e., when $D_1 = \{ 0 \}$. With $w$ denoting the subcorrector, Proposition~\ref{metcomparison} implies that, for all $x,y\in \ensuremath{\mathds{R}^\d}$ and a.s. in $\omega$,
\begin{equation*}
w(y,\omega) - w(x,\omega) \leq {m}_\mu (y,x,\omega)
\end{equation*}
Reversing $x$ and $y$ and adding the two inequalities together, we obtain
\begin{equation} \label{subaddpr}
0 \leq {m}_\mu (y,x,\omega) + {m}_\mu (x,y,\omega).
\end{equation}
We next claim that, for every $x,y,z\in \ensuremath{\mathds{R}^\d}$ and a.s. in $\omega$,
\begin{equation} \label{msubadd}
{m}_\mu(y,x,\omega) \leq {m}_\mu (y,z,\omega) + {m}_\mu (z,x,\omega).
\end{equation}
Indeed, thinking of both sides of \eqref{msubadd} as a function of $y$ with $x$ and $z$ fixed, noting that the inequality holds at both $y=x$ and $y=z$ (the former is \eqref{subaddpr} and the latter is obvious), and applying Proposition~\ref{metcomparison} with $D = \{ x,z\}$, we obtain \eqref{msubadd}. The subadditivity property \eqref{msubadd} must be modified in the case that $D_1 = B_1$, which is recorded in the next lemma.
It is convenient to state the subadditivity in terms of $\widetilde m_\mu$ from Remark~\ref{tildemetp}. To this end, observe that there exists a random variable $c(y,\omega)$ satisfying $|y|^{-1} c(y,\omega) \to 0$ a.s. in $\omega$ and
\begin{equation} \label{cthingy}
\sup_{z\in B(y,1)} \big( |w(y,\omega) - w(z,\omega)| + |\widetilde m_\mu(y,\omega) - \widetilde m_\mu(z,\omega) | \big) \leq c(y,\omega).
\end{equation}
Indeed, according to \eqref{sublin}, \eqref{wLIP} and \eqref{MPLB}, it suffices to take
\begin{equation*}
c(y,\omega): = C\Big(1 + \sup_{B(y,2)} V(\cdot,\omega) \Big)^{1/\gamma}.
\end{equation*}
For future reference, we observe that $c$ is stationary and
\begin{equation} \label{cint}
\ensuremath{\mathds{E}} c(0,\cdot) < \infty.
\end{equation}
Define the quantity
\begin{equation*}
\hat m_\mu(y,x,\omega) : = \widetilde m_\mu(y,x,\omega) + \frac12 ( c(x,\omega) + c(y,\omega) )
\end{equation*}
We claim that $\hat m_\mu$ is subadditive.
\begin{lem}\label{subadditivity}
For each $p\in \ensuremath{\mathds{R}^\d}$ and $\mu > \overline H(p)$, the function $\widetilde m_\mu$ satisfies, for all $x,y,z\in \ensuremath{\mathds{R}^\d}$, and a.s. in $\omega$,
\begin{equation} \label{msubadd2}
\hat m_\mu(y,x,\omega) \leq \hat m_\mu(y,z,\omega) + \hat m_\mu(z,x,\omega)
\end{equation}
\end{lem}
\begin{proof}
The proof in the case $\gamma > 2$ or $A\equiv 0$ is given in the discussion preceding the lemma. The case $\gamma \leq 2$ and $A\not\equiv 0$, i.e., $D_1=B_1$, is handled by a modification of this argument. We fix $x$ and $z$, write \eqref{msubadd2} in the form
\begin{equation} \label{msubadd3}
\widetilde m_\mu(y,x,\omega) \leq \widetilde m_\mu(y,z,\omega) + \widetilde m_\mu(z,x,\omega) + 2c(z,\omega),
\end{equation}
and, thinking of both sides of \eqref{msubadd3} as functions of $y$, we observe that the desired inequality follows from Proposition~\ref{metcomparison} provided we can show that it holds in $B(x,1) \cup B(z,1)$.
According Proposition~\ref{metcomparison}, for every $y,z\in \ensuremath{\mathds{R}^\d}$, we have
\begin{equation} \label{subadds1}
w(y,\omega) - w(z,\omega) \leq \widetilde m_\mu (y,z,\omega).
\end{equation}
Using this inequality, we obtain, for all $|y-x| \leq 1$, $z\in \ensuremath{\mathds{R}^\d}$ and a.s. in $\omega$,
\begin{equation*}
\widetilde m_\mu(y,x,\omega) = w(y,\omega) - w(x,\omega) \leq \widetilde m_\mu(y,z,\omega) + \widetilde m_\mu(z,x,\omega).
\end{equation*}
For $|y-z| \leq 1$, we use \eqref{cthingy} to get
\begin{align*}
\widetilde m_\mu(y,x,\omega) & \leq c(z,\omega) + \widetilde m_\mu (z,x,\omega) + w(y,\omega) - w(z,\omega) \\& = \widetilde m_\mu (y,z,\omega) + \widetilde m_\mu (z,x,\omega) + c(z,\omega).
\end{align*}
This completes the proof.
\end{proof}
\subsection{The homogenization of the metric problem}
\label{mho}
We consider, for $\mu > \overline H(p)$ and
\begin{equation*}
D_\varepsilon : = \begin{cases} \{ 0 \} & \mbox{if} \ \gamma > 2 \ \mbox{or} \ A \equiv 0, \\ B_\varepsilon & \mbox{if} \ \gamma \leq 2 \ \mbox{and} \ A \not\equiv 0,\end{cases}
\end{equation*}
the boundary value problem
\begin{equation} \label{meteq}
\left\{ \begin{aligned}
& -\varepsilon \tr\!\left( A(\tfrac x\varepsilon , \omega) D^2m^\varepsilon_\mu \right) + H(p+Dm^\varepsilon_\mu, \tfrac x\varepsilon, \omega) = \mu \quad \mbox{in} \ \ \ensuremath{\mathds{R}^\d}\!\setminus \! D_\varepsilon, \\
& m^\varepsilon_\mu = 0 \ \ \mbox{on} \ \partial D_\varepsilon, \quad \liminf_{|y| \to \infty} |y|^{-1} m^\varepsilon_\mu(y,\omega) \geq 0 \ \ \mbox{a.s. in} \ \omega,
\end{aligned} \right.
\end{equation}
which is nothing but a rescaling of \eqref{MP}--\eqref{bngcond}. In particular, \eqref{meteq} has a unique solution given by
\begin{equation} \label{rescalemeteq}
{m}^\varepsilon_\mu(x,\omega) = \varepsilon {m}_\mu\left(\frac x\varepsilon,0,\omega\right).
\end{equation}
\medskip
We show that the $m^\varepsilon_\mu$'s converge, as $\varepsilon\to 0$, a.s. in $\omega$ and in $C(\ensuremath{\mathds{R}^\d} \setminus \{0\})$, to the unique solution $\overline m_\mu=\overline m_\mu(y) = \overline m_\mu(y;p)$ of
\begin{equation} \label{meteq-bar}
\left\{ \begin{aligned}
& \overline H(p+D\overline m_\mu ) = \mu \quad \mbox{in} \ \ \ensuremath{\mathds{R}^\d}\!\setminus \! \{ 0 \}, \\
& \overline m_\mu(0) = 0, \quad \liminf_{|y| \to \infty} |y|^{-1}\, \overline m_\mu(y) \geq 0,
\end{aligned} \right.
\end{equation}
which, in view of Proposition~\ref{metexistence}, is well-posed for every $\mu > \overline H(p)$.
\begin{lem}
For each $\mu > \overline H(p)$,
\begin{equation}\label{mbarform}
\overline m_\mu(y;p) = \sup \big\{ y\cdot q : \overline H(p+q) \leq \mu \big\}.
\end{equation}
\end{lem}
\begin{proof}
Since the right hand side of \eqref{mbarform} is a subsolution of \eqref{meteq-bar}, the inequality ``$\geq$" in \eqref{mbarform} follows at once from Proposition~\ref{metcomparison}.
To obtain the reverse inequality, notice first that $\overline m_\mu$ is positively 1-homogeneous in $y$, i.e., for every $t \geq 0$,
\begin{equation*}
\overline m_\mu(ty;p) = t \overline m_\mu(y;p).
\end{equation*}
The scaling invariance of \eqref{meteq-bar} and the uniqueness of $\overline m_\mu$. Likewise, the convexity of $\overline H$, Lemma~\ref{convtrick} and Proposition~\ref{metcomparison} are easily combined to yield that $\overline m_\mu(y;p)$ is convex in $y$. If \eqref{mbarform} fails, we can find a point $y\neq 0$ such that $\overline m_\mu(\cdot;p)$ is differentiable at $y$ and $\overline m_\mu(y;p) > \sup\{ y\cdot q: \overline H(p+q) \leq \mu \}$. Since $\overline m_\mu(y;p) = y \cdot D\overline m_\mu(y;p)$, it follows that $H(p+D\overline m_\mu(y;p)) > \mu$. This contradicts the elementary viscosity-theoretic fact that $\overline m_\mu$ must satisfy $H(p+D\overline m_\mu) = \mu$ at any point of differentiability.
\end{proof}
\begin{remark} \label{touch-mbar}
We can see from \eqref{meteq-bar} and the convexity of $\overline H$ that
\begin{equation*}
\overline m_\mu(y;p) = \sup\big\{ y\cdot q: \overline H(p+q) = \mu\big\}.
\end{equation*}
Moreover, if $\overline H(p+q_0) = \mu$, then $q_0$ belongs to the subdifferential $\partial \overline m_\mu(x_0;p)$ of $\overline m_\mu$ at some point $x_0\neq 0$. That is, the plane $x \cdot q_0$ must touch $\overline m_\mu$ from below at some nonzero $x_0 \in \ensuremath{\mathds{R}^\d}$. To see this, consider the largest value of $\lambda$ for which
\begin{equation*}
\overline m_\mu(y;p) \geq \lambda y\cdot q_0.
\end{equation*}
By the positive homogeneity of both sides, we need only check the above inequality on the unit sphere, and for the same reason, our claim follows if we can show that $\lambda =1$. But if $\lambda > 1$, then we would obtain a contradiction as follows: first, by convexity of $\overline H$ and $\lambda > 1$, we have $\overline H(p+\lambda q_0 ) > \mu$. Then, by the convexity of the sublevel sets of $\overline H$, we can choose a vector $z\in \ensuremath{\mathds{R}^\d}\setminus \{ 0 \}$ and $\varepsilon > 0$ such that $z\cdot (\lambda q_0) \geq z\cdot q+\varepsilon$ for every $\overline H(p+q)\leq \mu$. But this implies $z\cdot (\lambda q_0) > \overline m_\mu(z;p)$, a contradiction.
\end{remark}
From \eqref{rescalemeteq}, we see that the limit, as $\varepsilon\to 0$, of the $m_\mu^\varepsilon$'s is equivalent to studying the limit of $t^{-1} {m}_\mu(ty,0,\omega)$ as $t \to \infty$. Therefore we may state the result on the homogenization of \eqref{meteq} as follows.
\begin{prop} \label{homogMP}
For each $p,y\in \ensuremath{\mathds{R}^\d}$ and $\mu > \overline H(p)$,
\begin{equation} \label{homogMPeq}
\lim_{t\to \infty} \frac1t {m}_\mu(ty,0,\omega;p) = \overline m_\mu(y;p) \quad \mbox{a.s. in} \ \omega.
\end{equation}
\end{prop}
\begin{proof}
By Proposition~\ref{metcomparison}, the functions $\widetilde m(\cdot,0,\omega)$ and $m_\mu(\cdot, 0,\omega)$ differ by no more than $c(0,\omega)$. Likewise, $\hat m_\mu(y,0,\omega)$ and $\widetilde m_\mu(y,0,\omega)$ differ by at most $\frac12( c(0,\omega) + c(y,\omega))$, which is strictly sublinear in $y$, a.s. in $\omega$. In light of Lemma~\ref{subadditivity} and \eqref{cint}, and using the semigroup $\sigma_t:= \tau_{ty}$, we may apply the subadditive ergodic theorem (Proposition~\ref{SAergthm}) to obtain that, for each $y\in \ensuremath{\mathds{R}^\d}$ and a.s. in $\omega$,
\begin{equation} \label{preliMP}
\lim_{t\to \infty} t^{-1} m_\mu(ty,0,\omega) = \lim_{t\to \infty} t^{-1} \hat m_\mu(ty,0,\omega) = M_\mu(y,\omega),
\end{equation}
for some $M_\mu(y,\omega)$. At this point, we must allow for the possibility that $M_\mu$ is random because we do not assume $\{ \sigma_t \}$ is ergodic.
It remains to show that $M_\mu = \overline m_\mu$. We first show that $M_\mu$ is constant a.s. in $\omega$. This follows from the stationary and ergodic assumptions in the usual way. According to Lemma~\ref{potests}, \eqref{MPLB} and Remark~\ref{MPstat}, for each $\omega \in \Omega_1$ and $z\in \ensuremath{\mathds{R}^\d}$, we have
\begin{equation*}
\limsup_{t \to \infty} t^{-1} m_\mu(ty,0,\tau_z\omega) = \limsup_{t \to \infty} t^{-1} m_\mu(ty+z,z,\omega) = \limsup_{t \to \infty} t^{-1} m_\mu(ty,0,\omega),
\end{equation*}
i.e., the set $\{ \omega \in \Omega_1: \limsup_{t\to \infty} t^{-1} m_\mu(ty,0,\omega) \leq k \}$ is invariant under $\tau_z$, for each $k\in \ensuremath{\mathds{R}}$. Since $\Omega_1$ has full probability, the ergodic hypothesis implies that
\begin{equation*}
\limsup_{t \to \infty} t^{-1} m_\mu(ty,0,\tau_z\omega) \equiv C \quad \mbox{a.s. in} \ \omega.
\end{equation*}
Hence $M_\mu(y,\omega) = M_\mu(y)$ a.s. in $\omega$.
It is clear that $M_\mu(y)$ is positively homogeneous. According to \eqref{bngcond}, it is thus nonnegative. The estimates \eqref{oscest} and \eqref{MPLB} imply that $M_\mu(y)$ is Lipschitz in $y$.
To complete the proof that $M_\mu = \overline m_\mu$, we show that $M_\mu$ is the solution of \eqref{meteq-bar}. Suppose that $\varphi$ is a smooth function and $x_0 \neq 0$ are such that
\begin{equation} \label{Mmuslm}
x \mapsto M_\mu(x) - \varphi(x) \quad \mbox{has a strict local maximum at} \quad x=x_0.
\end{equation}
We show by a perturbed test function argument that
\begin{equation} \label{Mmuwts}
\overline H(p+D\varphi(x_0)) \leq \mu.
\end{equation}
Arguing by contradiction, we assume that $\theta: = \overline H(p+D\varphi(x_0)) - \mu > 0$. Take $\{ \delta_j \}$ to be the subsequence described in Remark~\ref{assubseq}, along which we have \eqref{assubseqeq}. Set $p_1 : = p+D\varphi(x_0)$, let $\lambda > 1$ to be selected below, and define the perturbed test function
\begin{equation*}
\varphi_j(x) : = \varphi(x) + \lambda\Big( \delta_j v^{\delta_j} \big(\frac x{\delta_j}, \omega; p_1 \big) + \overline H(p_1)\Big).
\end{equation*}
We claim that, for all sufficiently large $j$ and sufficiently small $r> 0$, $\varphi_j$ satisfies
\begin{equation} \label{vjclmsdf}
-\delta_j \tr(A(x/\delta_j,\omega) D^2\varphi_j) + H(p+D\varphi_j,x/\delta_j,\omega) \geq \mu + \frac12 \theta \quad \mbox{in} \ B(x_0,r).
\end{equation}
Since $\varphi_j$ is not smooth in general, we verify the inequality in the viscosity sense. To this end, select a smooth function $\psi$ and a point $x_1 \in B(x_0,r)$ at which $\varphi_j - \psi$ has a local minimum. It follows that
\begin{equation*}
y\mapsto \lambda v^{\delta_j}(y,\omega;p_1) - \delta_j^{-1}\big( \psi(\delta_jy) - \varphi(\delta_jy) \big) \quad \mbox{has a local minimum at} \quad y= x_1/\delta_j.
\end{equation*}
Using \eqref{movepup} with $q=p+ D\varphi(x_1)$, we obtain
\begin{multline} \label{ptfm1p}
\lambda \delta_j v^{\delta_j}(x_1/\delta_j,\omega;p_1) - \delta_j\tr\big( A(x_1/\delta_j,\omega) (D^2\psi(x_1) - D^2\varphi(x_1) \big) \\ + H(p+D\psi(x_1), x_1/\delta_j , \omega) \geq (1-\lambda) H\left( \frac{D\varphi(x_1) - \lambda D\varphi(x_0)}{1-\lambda},y,\omega \right).
\end{multline}
Fix $\lambda>1$ to be sufficiently close to $1$ so that, by \eqref{assubseqeq}, for all large $j$ we have
\begin{equation*}
-\lambda \delta_j v^{\delta_j} (x_1/\delta_j,\omega;p_1) \geq \overline H(p_1) - \frac{1}{8}\theta.
\end{equation*}
We may allow $r$ and $j$ to depend on $\varphi$, and so by taking $j$ larger still we may assume that
\begin{equation*}
\delta_j \big|\tr(A(x_1/\delta_j,\omega) D^2\varphi(x_1))\big| \leq \frac18 \theta.
\end{equation*}
Next, observe that by taking $r> 0$ to be small, depending on $\lambda$ and $\varphi$, then we obtain
\begin{equation*}
|\lambda-1|^{-1} \big| D\varphi(x_1) - \lambda D\varphi(x_0)\big| \leq |\lambda-1|^{-1} \big|D\varphi(x_1) - D\varphi(x_0) \big| + |D\varphi(x_0)| \leq 2 |D\varphi(x_0)|.
\end{equation*}
By shrinking $\lambda$, depending only on $\varphi$, it follows from \eqref{Hbound} that
\begin{equation*}
(1-\lambda) H\left( \frac{D\varphi(x_1) - \lambda D\varphi(x_0)}{1-\lambda},y,\omega \right) \geq -C(1-\lambda) \big( 1+ |D\varphi(x_0)|^\gamma \big) \geq -\frac14 \theta.
\end{equation*}
The observations above yield, for small $r>0$ and large $j$,
\begin{equation*}
-\delta_j \tr(A(x_1/\delta_j,\omega) D^2\psi) + H(p+D\psi,x_1/\delta_j,\omega) \geq \overline H(p_1) - \frac12 \theta = \mu + \frac12 \theta.
\end{equation*}
This confirms the claim \eqref{vjclmsdf} in the viscosity sense.
The comparison principle implies that $m^{\delta_j}_\mu - \varphi_j$ cannot have a local maximum in $B(x_0,r)$. Sending $j\to \infty$, we obtain a contradiction to \eqref{Mmuslm}. This completes the proof that $M_\mu$ is a subsolution of \eqref{meteq-bar}. The argument that $M_\mu$ is a supersolution of \eqref{meteq-bar} is very similar. The perturbed test function $\varphi_j$ is defined in the same way, but with $\lambda < 1$, and we use \eqref{movepdn} instead of \eqref{movepup}. We leave the details to the reader. We conclude that $M_\mu$ is a solution of \eqref{meteq-bar}. By uniqueness, $M_\mu = \overline m_\mu$.
\end{proof}
\begin{remark} \label{nomuassump}
In the proof above, the assumption that $\mu > \overline H(p)$ is only needed for the existence of $m^\varepsilon_\mu$, i.e., it is not used in the perturbed test function argument which identifies $M_\mu$ with $\overline m_\mu$.
\end{remark}
\begin{remark}
In the setting of Hamilton-Jacobi equations, the classical perturbed test function argument (see Evans~\cite{E}) typically requires uniform Lipschitz estimates so that the error induced in the perturbation inside the coercive $H$ can be controlled. In our setting, the unboundedness of the potential prevents us from having such Lipschitz estimates. To get around this technical glitch we have introduced a new, modified version of the perturbed test function argument which utilizes the convexity of $H$. This is the purpose of introducing the parameter $\lambda$ in the proof of Proposition~\ref{homogMP}, and this technical device will be repeated several times in this paper.
\end{remark}
\subsection{A characterization of $\min\overline{H}$}
Using the methods developed in the previous subsections, we notice that the minimum of $\overline H$ can be characterized by the solvability of the inequality
\begin{equation} \label{charmHeqcl}
-\tr \!\left( A(y, \omega) D^2 v \right) + H(Dv,y, \omega) \leq \mu \quad \mbox{in} \ \ensuremath{\mathds{R}^\d}
\end{equation}
where, instead of asking for sublinear growth at infinity, we impose the growth condition
\begin{equation} \label{charmHgc}
\limsup_{|y| \to \infty} |y|^{-1} v(y) < \infty.
\end{equation}
The observation plays an important role in the proof of Proposition~\ref{bigstep}.
\begin{prop} \label{charmH}
The following formula holds a.s. in $\omega$:
\begin{equation}
\min_{p\in \ensuremath{\mathds{R}^\d}} \overline H(p) = \inf \big\{ \mu \, : \mbox{there exists} \ \ v \in C(\ensuremath{\mathds{R}^\d}) \ \ \mbox{satisfying} \ \ \eqref{charmHeqcl} \ \ \mbox{and} \ \ \eqref{charmHgc} \big\}.\label{charmHeq}
\end{equation}
\end{prop}
\begin{proof}
Let $\widetilde H(\omega)$ denote the left side of \eqref{charmHeq}. It is easy to see, by the ergodic hypothesis and the stationarity of the coefficients, that $\widetilde H$ is constant; i.e., $\widetilde H(\omega) = \widetilde H$ a.s. in $\omega$. According to the existence of the subcorrector in Proposition~\ref{mainstep}, we have that $\widetilde H \leq \min_{p\in \ensuremath{\mathds{R}^\d}} \overline H(p)$. It remains to show the reverse inequality.
Select $\mu > \widetilde H$ and define
\begin{equation*}
s_\mu(y,x,\omega) : = \sup\big\{ v(y-x) : v\in C(\ensuremath{\mathds{R}^\d}) \ \mbox{satisfies} \ v \leq 0 \ \ \mbox{on} \ D_1, \ \eqref{charmHeqcl} \ \ \mbox{and} \ \ \eqref{charmHgc} \big\}.
\end{equation*}
Then using Proposition~\ref{metcomparison} and barrier functions built in Proposition~\ref{metexistence} (we may take for example $m_\mu(y,0,\omega)$ with any $p$), $s_\mu$ is well-defined, jointly stationary in the sense of Remark~\ref{MPstat}, and growing at most linearly at infinity. By similar arguments as in the proof of Proposition~\ref{metexistence}, we see that, for each fixed $z\in \ensuremath{\mathds{R}^\d}$, $s_\mu(y,z,\omega)$ is a solution of the equation
\begin{equation*}
-\tr \!\left( A(y, \omega) D^2 v \right) + H(Dv,y, \omega) = \mu \quad \mbox{in} \ \ensuremath{\mathds{R}^\d} \setminus D_1(z).
\end{equation*}
By repeating the arguments of Sections~\ref{sadd} and~\ref{mho}, we obtain the homogenization of $s_\mu$. We first observe that $s_\mu$ has sufficient properties to apply the subadditive ergodic theorem, and we are able to obtain, a.s. in $\omega$,
\begin{equation*}
\frac1t s_\mu(ty,0,\omega) \rightarrow \overline s_\mu(y) \quad \mbox{as} \ t\to \infty.
\end{equation*}
Then, by the proof of Proposition~\ref{homogMP} (see also Remark~\ref{nomuassump}), the function $\overline s_\mu$ is a solution of the equation
\begin{equation*}
\overline H(D\overline s_\mu) = \mu \quad \mbox{in} \ \ensuremath{\mathds{R}^\d}\setminus \{ 0 \}.
\end{equation*}
Therefore, $\min \overline H \leq \mu$. This holds for any $\mu > \widetilde H$, and so we conclude that $\min \overline H \leq \widetilde H$.
\end{proof}
\section{The proof of Homogenization}
\label{PH}
In this section we complete the proof of Theorem~\ref{MAIN}.
\subsection{The almost sure homogenization of the macroscopic problem} \label{BS}
We now use Propositions~\ref{homogMP} and~\ref{charmH} to obtain an improvement of Proposition~\ref{mainstep}. Informally, the limit \eqref{bigstepeq} says that the functions $\hat v^\delta(y,\omega): = v^\delta(\cdot,\omega) - v^\delta(0,\omega)$ are ``approximate correctors on balls of radius~$\sim1/\delta$." This is what we need to prove Theorem~\ref{MAIN}.
\begin{prop} \label{bigstep}
There is a subset $\Omega_8 \subseteq \Omega_7$ of full probability such that, for each $R> 0$ and $\omega\in \Omega_8$,
\begin{equation} \label{bigstepeq}
\lim_{\delta \to 0} \sup_{B_{R/\delta}} \big| \delta v^\delta(\cdot,\omega;p) + \overline H(p) \big| = 0.
\end{equation}
\end{prop}
\begin{proof}
Lemma~\ref{CDE-noxd} reduces the work to showing that \eqref{bigstepeq} holds, a.s. in $\omega$, for each fixed $p\in \ensuremath{\mathds{R}^\d}$. The general result is then obtained and $\Omega_8$ is defined by intersecting the relevant subsets, say $\Omega_p$, over a countable dense subset of $p$, and appealing to \eqref{CDE-noxdeq}. Therefore we fix $p\in \ensuremath{\mathds{R}^\d}$.
\medskip
We split the proof into two steps. The first is to use the homogenization of the metric problem to obtain almost sure convergence of $\delta v^\delta $ to $-\overline H$ at the origin. To accomplish this we use a perturbed test function argument \emph{in reverse}, arguing in effect that, if we improperly perturb our test function, then we could not have homogenization. In the second step, we combine the ergodic and Egoroff's theorems with Lemma~\ref{impest} to improve the almost sure convergence at the origin to balls of radius $\sim1/\delta$. Unfortunately, we must perform the two steps in reverse order, since we need the result of second step to perform the first.
\medskip
\emph{Step 1. Reducing the question to convergence at the origin.}
According to Proposition~\ref{mainstep} and Remark~\ref{WDlimsup},
\begin{equation} \label{limsorg}
\liminf_{\delta \to 0} \delta v^\delta (0,\omega) = -\overline H(p) \leq -\hat H(p) : = \limsup_{\delta \to 0} \delta v^\delta(0,\omega) \quad \mbox{a.s. in} \ \omega.
\end{equation}
We claim that, for each $R> 0$,
\begin{equation} \label{toballs1del}
\liminf_{\delta \to 0} \inf_{B_{R/\delta}} \delta v^\delta (\cdot,\omega) = -\overline H(p) \quad \mbox{and} \quad \limsup_{\delta \to 0} \sup_{B_{R/\delta}} \delta v^\delta(\cdot,\omega) = -\hat H(p) \quad \mbox{a.s. in} \ \omega.
\end{equation}
Since the arguments are nearly identical, we only prove the first identity of \eqref{toballs1del}. From \eqref{limsorg} and Egoroff's theorem, for every $\rho > 0$, there exists $\bar\delta(\rho)> 0$ and a set $E_\rho \subseteq \Omega$ such that $\ensuremath{\mathds{P}}[E_\rho] \geq 1-\rho$ and, for every $0 < \delta \leq \bar \delta(\rho)$,
\begin{equation*}
\inf_{\omega\in E_\rho} \delta v^\delta(0,\omega) + \overline H \geq -\rho.
\end{equation*}
The ergodic theorem provides, for each $\rho > 0$, a subset $F_\rho \subseteq \Omega$ of full probability such that, for every $\omega \in F_\rho$,
\begin{equation*}
\lim_{R \to \infty} \fint_{B_R} \mathds{1}_{E_\rho}(\tau_y\omega) \, dy = \ensuremath{\mathds{P}}[E_\rho] \geq 1-\rho.
\end{equation*}
Define $F_0 : = \cap_{j=1}^\infty (F_{2^{-j}}\cap \Omega_3)$ with $\Omega_3$ as in Lemma~\ref{impest}. Then $\ensuremath{\mathds{P}}[F_0] =1$. Fix $\omega \in F_0$ and $R, \rho > 0$, with $\rho = 2^{-j}$ for some $j\in \ensuremath{\mathds{N}}$. It follows that, if $\delta > 0$ is sufficiently small (depending on $\omega$, $R$ and $\rho$), then
\begin{equation} \label{fillingup}
| \{ y\in B_{R/\delta} : \tau_y \omega \in E_\rho \} | \geq (1-2\rho) |B_{R/\delta}|.
\end{equation}
Using \eqref{uber-oscest} and shrinking $\delta$, depending again on $\omega$, $\rho$ and $R$, we have, for $r\in (\rho R, R)$,
\begin{equation} \label{oscestapp}
\sup_{y\in B_R} \osc_{B(y/\delta, r/\delta)} \delta v^\delta(\cdot,\omega) \leq 2C r.
\end{equation}
Select any $z\in B_{R/\delta}$. In a similar way as in the proof of Lemma~\ref{impest}, we may use \eqref{fillingup} to find a point $y\in B_{R/\delta}$ with $|y-z|\leq C\rho R\delta^{-1}$ and $\tau_y\omega \in E_\rho$. In view of \eqref{oscestapp} and the stationarity of the $v^\delta$'s, we deduce that, for each $\delta$ sufficiently small, depending on $\omega$, $\rho$, and $R$,
\begin{align*}
\delta v^\delta(z,\omega) + \overline H & \geq -| \delta v^\delta(z,\omega) - \delta v^\delta(y,\omega)| + \delta v^\delta(y,\omega) + \overline H
\\ & \geq -2C_1(C\rho R) + \delta v^\delta(0,\tau_y\omega) + \overline H \\ & \geq -C\rho R - \rho,
\end{align*}
and, hence, for each $\omega \in F_0$ and $R>0$,
\begin{equation*}
\liminf_{\delta\to 0} \inf_{z\in B_{R/\delta}} \big( \delta v^\delta(z,\omega) + \overline H\big) \geq 0,
\end{equation*}
from which \eqref{toballs1del} follows.
\medskip
\emph{Step 2. The convergence at the origin in the case $\overline H(p) = \min \overline H$.}
We show that $\hat H(p) \geq \widetilde H$, the latter introduced in the proof of Proposition~\ref{charmH}. For each fixed $\omega \in \Omega_7$, so that \eqref{hatHdefeq} holds, observe that for we can pass to limit as $\delta\to 0$ along a subsequence of $\hat v^\delta(y,\omega): = v^\delta(\cdot,\omega) - v^\delta(0,\omega)$ to find a solution of
\begin{equation*}
-\tr \!\left( A(y, \omega) D^2 v \right) + H(p+Dv,y, \omega) = \hat H(p) \quad \mbox{in} \ \ensuremath{\mathds{R}^\d}.
\end{equation*}
It then follows from the definition of $\widetilde H$ that $\widetilde H \leq \hat H(p)$. According to our assumption and Proposition~\ref{charmH}, we have $\overline H(p) = \min_{\ensuremath{\mathds{R}^\d}} \overline H = \widetilde H$, and so we conclude that $\overline H(p) = \hat H(p)$. This completes the proof in the case that $p$ lies on the level set $\{ q\in \ensuremath{\mathds{R}^\d} : \overline H(q) = \min \overline H \}$.
\medskip
\emph{Step 3. The convergence at the origin in the case $\overline H(p) > \min \overline H$.} To complete the proof of \eqref{bigstepeq}, it remains to show that $\overline H(p) = \hat H(p)$. We may assume that
\begin{equation*}
\min_{\ensuremath{\mathds{R}^\d}} \overline H = \overline H(0) < \overline H(p).
\end{equation*}
Arguing by contradiction, we suppose that $\rho : = \overline H(p) - \hat H(p) > 0$. Fix $\omega \in\Omega_{7}$ and select a subsequence $\delta_j \to 0$, which may depend on $\omega$, such that
\begin{equation*}
\lim_{j \to \infty} {\delta_j} v^{\delta_j} (0,\omega;p) = -\hat H(p)= -\overline H(p) + \rho.
\end{equation*}
Put $\mu := \overline H(p)$. According to Remark~\ref{touch-mbar}, we can pick $x_0 \neq 0$ such that $\overline m_\mu (x_0; 0 ) = x_0 \cdot p$. With $\eta > 0$ to be selected below, define
\begin{equation*}
w_j(x):= (x+x_0)\cdot p + \delta_j v^{\delta_j} \Big(\frac x{\delta_j},\omega;p\Big) - \eta |x|^2.
\end{equation*}
It is clear that, for $r,\eta>0$ small enough and $j$ large enough, $w_j$ is a subsolution of
\begin{equation*}
-\delta_j \tr \!\Big( A\Big(\frac x{\delta_j}, \omega\Big) D^2w_j \Big) + H\Big(Dw_j,\frac x{\delta_j}, \omega\Big) \leq \mu - \frac12\rho \quad \mbox{in} \ B(0,r).
\end{equation*}
Denoting by $m^\varepsilon_\mu$ the unique solution of \eqref{meteq}, we see, from the comparison principle and the above inequality, that
\begin{equation} \label{nolocmin}
\min_{x\in B(0,r)} \big( m^{\delta_j}_\mu(x,-x_0,\omega;0) - w_j(x) \big) = \min_{x\in \partial B(0,r)}\big( m^{\delta_j}_\mu(x,-x_0,\omega;0) - w_j(x) \big).
\end{equation}
Letting $j\to \infty$, we deduce a contradiction as follows. Grouping the terms as
\begin{multline*}
m^{\delta_j}_\mu(x,-x_0,\omega;0) - w_j(x) = \big( m^{\delta_j}_\mu(x,-x_0,\omega;0) - \overline m_\mu(x+x_0;0) \big) + \big( \overline m_\mu(x+x_0;0) - p\cdot (x+x_0) \big) \\ + \big( -\delta_j v^{\delta_j}(x/\delta_j,\omega;p) + \eta|x|^2 \big),
\end{multline*}
we see that the first term converges uniformly to 0 in $B(0,r)$ as $j\to \infty$ by Proposition~\ref{homogMP}, the second term is nonnegative and vanishes at $x=0$, and the last term satisfies, by \eqref{toballs1del},
\begin{equation*}
\lim_{j\to \infty} \inf_{x \in B(0,r)} \big( -\delta_j v^{\delta_j}(x/\delta_j,\omega;p) + \eta|x|^2 \big) = \hat H(p) + \eta r^2 > \hat H(p) \geq \lim_{j\to \infty} \big( -\delta_j v^{\delta_j}(0,\omega;p) \big).
\end{equation*}
Therefore \eqref{nolocmin} is impossible for large $j$, and so we are forced to conclude that $\overline H(p) = \hat H(p)$. Now \eqref{toballs1del} yields \eqref{bigstepeq}.
\end{proof}
\subsection{The unique solvability of \eqref{HJ}}
We pause our march toward a proof of Theorem~\ref{MAIN} for a brief word regarding the well-posedness of \eqref{HJ}. It is not the main purpose of this paper to deal with such issues, and providing a complete proof of the unique solvability of \eqref{HJ} requires little more than repeating the arguments in Section~\ref{AMP-wp} or Section~\ref{AMP} and rearranging them to follow Section 6 of \cite{LS2}. For this reason, we omit the proof of the following proposition.
\begin{prop}\label{epWP}
There exists $\Omega_9 \subseteq \Omega_8$ such that, for every $u_0 \in \BUC(\ensuremath{\mathds{R}^\d})$, there exists a unique solution $u^\varepsilon=u^\varepsilon(\cdot,\omega) \in C(\ensuremath{\mathds{R}^\d}\times\ensuremath{\mathds{R}}_+)$ of \eqref{HJ} which is bounded below on $\ensuremath{\mathds{R}^\d} \times[0,T]$ for each $T>0$ and $\omega\in \Omega_9$. Moreover, there exists a constant $C >0$ such that, for each $(x_0,t_0)\in \ensuremath{\mathds{R}^\d}\times \ensuremath{\mathds{R}}_+$, $r> 0$ and $\omega\in \Omega_9$,
\begin{equation} \label{eposcbd}
\limsup_{\varepsilon\to 0} \osc_{B(x_0,r) \times (t_0,t_0+r)} u^\varepsilon(\cdot,\omega) \leq C r.
\end{equation}
\end{prop}
\subsection{The proof of the main result}
We now assemble the ingredients developed in the previous sections into a proof of the homogenization result. The argument is based on the classical perturbed test function method
and is also very similar to the argument in the proof of Proposition~\ref{homogMP}.
\begin{proof}[{Proof of Theorem~\ref{MAIN}}]
Fix an $\omega \in \Omega_9$ so that both \eqref{bigstepeq} and \eqref{eposcbd} hold. According to \eqref{eposcbd}, we may extract a subsequence $\varepsilon_j \to 0$ and $u\in (C(\ensuremath{\mathds{R}^\d} \times\ensuremath{\mathds{R}}_+) \cap \BUC(\ensuremath{\mathds{R}^\d} \times [0,T]))$, for every $T> 0$, such that, as $j\to \infty$,
\begin{equation*}
u^{\varepsilon_j} \rightarrow u \quad \mbox{locally uniformly in} \ \ensuremath{\mathds{R}^\d}\times\ensuremath{\mathds{R}}_+.
\end{equation*}
We claim that $u$ is the unique solution of the problem \eqref{HJ-eff}.
\medskip
Here we verify only that $u$ is a supersolution of \eqref{HJ-eff}, since the argument for showing $u$ is a subsolution is similar (and both are also similar to the proof Proposition~\ref{homogMP}. To this end we select a smooth test function $\varphi$ and a point $(x_0,t_0) \in \ensuremath{\mathds{R}^\d}\times\ensuremath{\mathds{R}}_+$ such that
\begin{equation} \label{touch1}
(x,t) \mapsto (u - \varphi)(x,t) \quad \mbox{has a strict local minimum at} \ (x_0,t_0).
\end{equation}
We must show that
\begin{equation*}
\varphi_t(x_0,t_0) + \overline H(D\varphi(x_0,t_0)) \geq 0.
\end{equation*}
Suppose on the contrary that
\begin{equation*}
-\theta : = \varphi_t(x_0,t_0) + \overline H(D\varphi(x_0,t_0)) < 0.
\end{equation*}
Set $p_0 : = D\varphi(x_0,t_0)$, let $0< \lambda < 1$ be a parameter to be selected below, and define the perturbed test function
\begin{equation*}
\varphi^\varepsilon(x,t) : = \varphi(x,t) + \lambda\varepsilon v^\varepsilon\!\left(\tfrac x\varepsilon, \omega; p_0 \right)\!,
\end{equation*}
where $v^\varepsilon$ is the solution of the auxiliary problem \eqref{aux} with $\delta = \varepsilon$ and $p=p_0$. While we cannot expect that $\varphi^\varepsilon$ is smooth, we claim that, for $r,\varepsilon> 0$ sufficiently small, $\varphi^\varepsilon$ is a viscosity solution of
\begin{equation} \label{corrclm}
\varphi^\varepsilon_t - \varepsilon \tr \!\left( A(\tfrac x\varepsilon, \omega) D^2\varphi^\varepsilon \right) + H(D\varphi^\varepsilon,\tfrac x\varepsilon, \omega) \leq -\frac12 \theta \quad \mbox{in} \ B(x_0,r) \times (t_0-r,t_0+r).
\end{equation}
To verify \eqref{corrclm}, select a smooth $\psi$ and a point $(x_1,t_1) \in B(x_0,r)\times (t_0-r,t_0+r)$ such that
\begin{equation*}
(x,t) \mapsto (\varphi^\varepsilon - \psi)(x,t) \quad \mbox{has a local maximum at} \ (x_1,t_1).
\end{equation*}
We may rewrite this as
\begin{equation*}
(y,t) \mapsto \lambda v^\varepsilon(y,\omega; p_0) - \frac1\varepsilon \left( \psi(\varepsilon y,t) - \varphi(\varepsilon y,t) \right) \quad \mbox{has a local maximum at} \ \left(\tfrac{x_1}\varepsilon,t_1\right).
\end{equation*}
In particular, $\varphi_t(x_1,t_1) = \psi_t(x_1,t_1)$. Using that $v^\varepsilon$ is a viscosity supersolution of \eqref{movepdn} with $q= D\varphi(x_1,t_1)$, we deduce that
\begin{multline}
\lambda \varepsilon v^\varepsilon \left( \tfrac {x_1}\varepsilon, \omega;p_0\right) - \varepsilon \tr \!\left( A\left(\tfrac {x_1}\varepsilon,\omega\right) \left(D^2\psi(x_1,t_1) - D^2\varphi(x_1,t_1) \right)\right) \\
+ H\!\left(D\psi(x_1,t_1), \tfrac {x_1}{\varepsilon}, \omega\right) \leq (1-\lambda) H\left( \frac{D\varphi(x_1,t_1) - \lambda D\varphi(x_0,t_0)}{1-\lambda}, y, \omega \right).
\end{multline}
In a similar way as in the proof of Proposition~\ref{homogMP}, we can take $r> 0$ small, $j$ large, and $\lambda$ close to $1$, all depending on $\varphi$ but not on $\psi$, so that
\begin{equation*}
\lambda \varepsilon v^\varepsilon(x_1/\varepsilon,\omega;p_0) \geq -\overline H(p_0) - \frac18 \theta,
\end{equation*}
\begin{equation*}
\big|\varphi_t(x_1,t_1) - \varphi_t(x_0,t_0)\big| + \varepsilon \big| \tr \!\left(A\left(x_1/\varepsilon,\omega\right) D^2\varphi(x_1,t_1) \right)\big| \leq \frac18\theta,
\end{equation*}
and
\begin{equation*}
(1-\lambda) H\left( \frac{D\varphi(x_1,t_1) - \lambda D\varphi(x_0,t_0)}{1-\lambda}, y, \omega \right) \leq \frac14 \theta
\end{equation*}
Combining the above inequalities yields
\begin{equation*}
\psi_t(x_1,t_1) - \varepsilon \tr \!\left( A\left(\tfrac {x_1}\varepsilon,\omega\right) D^2\psi(x_1,t_1) \right) + H\!\left(D\psi(x_1,t_1), \tfrac {x_1}{\varepsilon}, \omega\right) \leq - \frac12 \theta,
\end{equation*}
and hence \eqref{corrclm}.
\medskip
According to the comparison principle, the map $(x,t) \mapsto (u^\varepsilon - \varphi^\varepsilon)$ cannot have a local maximum in the set $B(x_0,r) \times (t_0-r,t_0+r)$. This is the contradiction to \eqref{touch1}.
\medskip
The solution of \eqref{HJ-eff} is unique, according to Remark~\ref{effHcmp}. This implies the local uniform convergence of the full sequence $u^\varepsilon(\cdot,\omega)$ to $u$ for each $\omega$ in a set of full probability, and completes the proof of the theorem.
\end{proof}
\section{Further properties of the metric problem} \label{P}
\subsection{A probabilistic interpretation}
The analysis of the metric problem allows us to make a precise connection to some of the results of Sznitman \cite{Sz1,Sz3, Sz2}. What we describe here is found in Chapter 5 of \cite{Szb}. For the reader's convenience, we deviate a bit from the notation in the rest of this paper in an effort to be more faithful to that of \cite{Szb}.
In what follows we denote by $Z=Z_t$ the canonical process on $\ensuremath{\mathds{R}^\d}$, $P_0$ is a Wiener measure, and $E_0$ is the expectation with respect to $P_0$. Sznitman studies the behavior of the Brownian motion $Z$ moving in a random environment composed of soft Poissonian obstacles. The latter are represented by a potential $V$ given by \eqref{Vexam}.
The obstacles have the interpretation of ``killing" the Brownian motion, that is, the medium is absorbing the diffusion at a rate $V$.
\medskip
Specifically, Sznitman studies the quantity
\begin{equation*}
e_\mu(y,\omega) : = E_0 \left[ \exp\left( -\int_0^{h(y)} (V(y+Z_s,\omega) + \mu) \, ds \right) : \ h(y) < \infty \right],
\end{equation*}
where $\mu \geq 0$ is a parameter and $h(y): = \inf\{ s> 0 : y+ Z_s \in B_1 \}$ is the first hitting time of the unit ball for the Brownian motion translated by $y$. The goal is to understand how $e_\mu$ behaves for very large $|y|$. According to the Feynman-Kac formula, $e_\mu$ is a solution, $\ensuremath{\mathds{P}}$-a.s., of
\begin{equation*}
-\Delta e_\mu +\left( V(y,\omega) +\mu \right) e_\lambda= 0 \quad \mbox{in} \ \ensuremath{\mathds{R}^\d}\!\setminus\! B_1,
\end{equation*}
and $0 < e_\mu \leq 1$ in $\ensuremath{\mathds{R}^\d} \!\setminus \! B_1$ and $e_\mu = 1$ on $\partial B_1$.
\medskip
The change of variables
\begin{equation*}
a_\mu(y,\omega): = -\log e_\mu(y,\omega),
\end{equation*}
yields that $a_\mu \geq 0$ solves
\begin{equation*}
\left\{ \begin{aligned}
& -\Delta a_\mu + |Da_\mu|^2 - V(y,\omega) = \mu & \mbox{in} & \ \ensuremath{\mathds{R}^\d}\!\setminus \! B_1,\\
& a_\mu = 0 & \mbox{on} & \ \partial B_1, \\
& a_\mu \geq 0 & \mbox{in} & \ \ensuremath{\mathds{R}^\d}\!\setminus \! B_1.
\end{aligned} \right.
\end{equation*}
Recalling from Proposition~\ref{invispot} that in this case $\overline H(0) = 0$, we see that, for $\mu > 0$, $a_\mu$ is the unique solution $m_\mu$ of the metric problem given by Proposition~\ref{metexistence} and, according to Proposition~\ref{homogMP},
\begin{equation*}
\lim_{|y| \to \infty} \frac{|a_\mu(y,\omega) - \alpha_\mu(y)|}{|y|} = 0 \quad \mbox{a.s. in} \ \omega,
\end{equation*}
where $\alpha_\mu(y): = \overline m_\mu(y)$ is the solution of \eqref{meteq-bar}. Using the fact that, according to Proposition~\ref{metcomparison}, the functions $\alpha_\mu$ and $a_\mu$ are monotone in $\mu$, we deduce that, for each $\Lambda > 0$,
\begin{equation} \label{shapethm}
\lim_{|y| \to \infty} \sup_{0 < \mu \leq \Lambda} \frac{|a_\mu(y,\omega) - \alpha_\mu(y)|}{|y|} = 0 \quad \mbox{a.s. in} \ \omega.
\end{equation}
The homogenization assertion \eqref{shapethm} is what Sznitman calls the ``shape theorem," which he proves using entirely probabilitstic methods (see \cite[Theorem 2.5 in Chapter 5]{Szb}). The functions $\alpha_\mu = \overline m_\mu$ are called the \emph{quenched Lyapunov exponents}.
\medskip
The homogenization result in Proposition~\ref{homogMP} is more general than Sznitman's shape theorem, as we consider more general Hamiltonians and diffusions, and even in the special case $H = |p|^2 + V(y,\omega)$, more general potentials $V$.
\subsection{New formulae for $\overline H$}
Since $\overline m_\mu$ is positively 1-homogeneous, $D\overline m_\mu$ is constant along rays starting from the origin, and the function $\overline m_\mu$ is determined by its values on the unit sphere. It follows that $\overline m_\mu \geq 0$. Moreover, since $\mu > \overline H(p)$, it is clear that $\overline m_\mu$ cannot be touched from below by a constant function at any point of $\ensuremath{\mathds{R}^\d} \!\setminus \!\{0\}$, and, therefore, $\overline m_\mu > 0$ in $\ensuremath{\mathds{R}^\d}\! \setminus \! \{0\}$. Then by homogeneity, we have $\overline m_\mu(y) \geq c|y|$ for some $c> 0$.
\medskip
The convexity of $\overline H$ and $\mu > \overline H(p)$ imply that the level set $\{ q \in \ensuremath{\mathds{R}^\d}: \overline H(p+q) = \mu \}$ has empty interior. It follow that the level set above equals the closure of the image of $D\overline m_\mu$. In this way, one may recover the effective Hamiltonian $\overline H$ from the functions $\overline m_\mu$, since we may fix $p_0 \in \ensuremath{\mathds{R}^\d}$ with $\overline H (p_0) = \min \overline H$ and write
\begin{equation} \label{barHbarm}
\overline H(p) = \inf\big\{ \mu > \min \overline H \, : \, \overline m_\mu(y;p_0) > (p-p_0)\cdot y\, \mbox{ for all } y\in \ensuremath{\mathds{R}^\d} \big\}.
\end{equation}
From the results and arguments above, we can characterize the effective Hamiltonian in terms of the solvability of the metric problem.
\begin{prop} The effective Hamiltonian is given by
\begin{equation}
\overline H(p) = \inf\!\left\{ \mu \in \ensuremath{\mathds{R}} : \mbox{there exists} \ u=u(y,\omega) \ \mbox{satisfying} \ \eqref{MP}\mbox{--}\eqref{bngcond}, \ \mbox{a.s. in} \ \omega \right\}. \label{Hformula}
\end{equation}
\end{prop}
\begin{proof}
We first argue that, for $\mu < \overline H(p)$, \eqref{meteq-bar} has no subsolution $u$ satisfying \eqref{meteq-bar}. Assuming to the contrary that, for some $\mu < \overline H(p)$, \eqref{MP} has a subsolution $u$, we argue that $u$ cannot satisfy the growth condition at infinity. To this end, for each $R> 1$, consider the smooth function $\phi_R:(0,\infty) \to \ensuremath{\mathds{R}}$
\begin{equation*}
\phi_R(t): = \left( 1 - (t-R)^2 \right)^{1/2} - (R^2+1)^{1/2},
\end{equation*}
which satisfies
\begin{equation*}
\phi_R(0) = 0, \quad \phi_R(t) \leq -ct \quad \mbox{on} \ [0,R/2] \quad \mbox{and} \quad |\phi'_R(t)| \leq C \quad \mbox{on} \ [0,\infty),
\end{equation*}
for constants $C,c> 0$ independent of $R> 1$, and, finally, $t^{-1}\phi_R(t) \rightarrow 1$ as $t\to \infty$. Define $\varphi_R(y): = \varepsilon \phi_R(|y|)$, and observe that $\varphi_R$ is smooth and that, by the continuity of $\overline H$, for sufficiently small $\varepsilon > 0$,
\begin{equation*}
\overline H(p+D\varphi_R) > \mu.
\end{equation*}
Since $\varphi_R(0) = 0$ and $\varphi_R(y) \sim \varepsilon |y|$ for large $|y|$, we deduce from Proposition~\ref{metcomparison} that $u < \varphi_R$ in $\ensuremath{\mathds{R}^\d}\setminus \{ 0 \}$. In particular, we obtain
\begin{equation*}
u(y) \leq -c\varepsilon |y| \quad \mbox{for all} \ |y| \leq R/2.
\end{equation*}
Sending $R \to \infty$, we see that $\limsup_{|y|\to \infty} u(y) / |y| \leq -c \varepsilon$.
Next, suppose there exists a function $u=u(y,\omega)$ satisfying \eqref{MP}--\eqref{bngcond}, a.s. in $\omega$. Then by repeating the arguments in the proof of Theorem~\ref{MAIN}, we can find an $\omega$ and a subsequence $\varepsilon_j\to 0$ for which
\begin{equation*}
\varepsilon_j u(\varepsilon_j^{-1} y, \omega) \rightarrow \widetilde u \quad \mbox{locally uniformly in} \ \ensuremath{\mathds{R}^\d}\!\setminus \! D_1,
\end{equation*}
and such that $\widetilde u$ satisfies \eqref{MP}. Hence $\mu \geq \overline H(p)$.
\end{proof}
We conclude by noting that the statements in Theorem~\ref{MAIN1} follow from the results in Section~\ref{Q} and the proof of the above proposition.
|
\section{Introduction}
This paper is a follow-up to \cite{BenYaacov-Berenstein-Melleray:TopometricGroups}, in which we introduced the notion of a \emph{Polish topometric group}, and defined a notion of \emph{ample generics} in that context. We first recall some basic terminology: a \emph{Polish metric structure} $\mathcal M$ is a complete, separable metric space $(M,d)$, along with a family $(R_i)_{i \in I}$ such that each $R_i$ is a uniformly continuous map from some $M^{n_i}$ to $\bR$.
The automorphism group $\Aut(\cM)$ is made up of all the isometries of $(M,d)$ which preserve all the relations $R_i$.
When endowed with the pointwise convergence topology $\tau$, $\Aut(\mathcal M)$ is a Polish group.
It is also natural to consider the metric of uniform convergence $\partial$ defined by
\begin{gather*}
\partial(g,h)= \sup \bigl\{ d(gm,hm) \colon m \in M \bigr\}.
\end{gather*}
This metric $\partial$ is complete, bi-invariant, and in general not separable.
It is also $\tau$-lower semi-continuous, i.e the sets $\{(g,h) \colon \partial(g,h) \le r\}$ are closed for all $r$.
The automorphism groups of Polish metric structures, when endowed with the topology $\tau$ and the metric $\partial$, provide the paradigm for Polish topometric groups, and are ubiquitous in analysis.
Of particular interest to us are the unitary group $\mathcal U(\ell_2)$ of a complex separable infinite-dimensional Hilbert space, and the isometry group of the Urysohn space $\mathbb U$.
These are highly homogeneous structures, and this leads to them having ``large'' diagonal conjugacy classes.
The exact notion of largeness we use involves an interplay of the topology and metric, since in both cases above it is known that each conjugacy class is topologically meagre.
Recall from \cite{BenYaacov-Berenstein-Melleray:TopometricGroups} that a Polish topometric group $(G,\tau,\partial)$ has \emph{(topometric) ample generics} if for any $n \in \mathbb N$ there exists an uple $(g_1,\ldots,g_n) \in G^n$ such that the $\partial$-closure of $\{(kg_1k^{-1},\ldots,kg_nk^{-1}) \colon k \in G\}$ is co-meagre in $G^n$.
This definition extends the usual notion of ample generics (obtained when $\partial$ above is the discrete metric).
As mentioned above, natural examples of groups with topometric ample generics do not have ample generics as Polish groups, and it is in fact an open question whether there exist Polish groups with ample generics which are not isomorphic to a closed subgroups of $\mathfrak S_{\infty}$, the group of permutations of the integers, whence the need for this relaxed version of ample generics.
Many automorphism group of highly homogeneous Polish metric structures with meagre conjugacy classes turn out to have ample generics when considered as Polish topometric group when endowed with the topology of pointwise convergence and the metric of uniform convergence.
For instance, $\mathcal U(\ell_2)$, $\Iso(\mathbb U)$ and the automorphism group $Aut([0,1],\lambda)$ of the unit interval endowed with the Lebesgue measure all have ample metric generics when endowed with their natural topometric structure.
We are especially interested in the link of ample generics with the \emph{automatic continuity property}; the following theorem (\cite[Theorem~4.6]{BenYaacov-Berenstein-Melleray:TopometricGroups}) extends to the topometric context an automatic continuity theorem proved by Kechris and Rosendal in \cite{Kechris-Rosendal:Turbulence}.
\begin{thm}
\label{thm:AutomaticContinuity}
Let $(G,\tau,\partial)$ be a Polish topometric group with ample generics, $H$ a topological group with uniform Suslin number strictly less than $2^{\aleph_0}$ (e.g a metrisable group of density character $<2^{\aleph_0}$), and $\varphi \colon G \to H$ a homomorphism such that $\varphi$ is continuous from $(G,\partial)$ to $H$.
Then $\varphi$ must be continuous from $(G,\tau)$ to $H$.
\end{thm}
In the context of subgroups of $\mathfrak S_{\infty}$, the \emph{small index property}, which says that subgroups of index $< 2^{\aleph_0}$ are open, has been extensively studied.
Kechris and Rosendal showed that a Polish group $G$ with ample generics always has the small index property; it is a consequence of the automatic continuity theorem above (when $\partial$ is discrete).
Our aim here is to establish the analogue of the small index property in the topometric context.
Since most of the groups in which we are interested do not have proper open subgroups, an ``analogous result'' involving subgroups in the ordinary sense (which can be proved) would seem too weak.
In order to state a stronger result, we replace subsets by functions with values in $[0,\infty]$, which we call \emph{graded subsets}.
Then, closed subsets correspond naturally to lower semi-continuous functions, open subsets to upper semi-continuous functions (hence clopen subsets correspond to continuous functions), and one is led to define the notion of a meagre graded subset.
We then obtain extensions of some classical elementary results of descriptive set theory (the Kuratowski-Ulam theorem, the Pettis theorem, etc.) to the ``graded context''.
It turns out that the analogue of a subgroup in that context (i.e, a \emph{graded subgroup}) is a \emph{semi-norm} on $G$, i.e a map $H$ such that
\begin{itemize}
\item $H(1)=0$;
\item $\forall g \in G \ H(g)=H(g^{-1})$;
\item $\forall g,g' \in G \ H(gg') \le H(g)+H(g')$.
\end{itemize}
Such a function $H$ naturally defines a left-invariant pseudo-metric $d_H$ on $G$, defined by $d_H(g,h)=H(g^{-1}h)$; the \emph{index} of $H$ is simply the density character of the metric space obtained when identifying points $g,h$ such that $d_H(g,h)=0$.
A ``graded'' analogue of the small index property, which at least on its face seems stronger than the classical version, would then be: whenever $H$ is a left-invariant pseudo-metric on $G$ with density character $< 2^{\aleph_0}$, $H$ must be continuous with respect to $\tau$.
A topometric version thereof must take $\partial$ into account (or else be too strong), and one is led to the following statement, which is our main result.
\begin{thm}
Assume that $(G,\tau,\partial)$ is a Polish topometric group with ample generics, and that $H$ is a semi-norm on $G$ which is lower semi-continuous with respect to $\partial$.
Then $H$ is continuous with respect to $\tau$.
\end{thm}
It is clear that this result extends the automatic continuity theorem \ref{thm:AutomaticContinuity} when the range group $H$ is metrisable; we say that a topometric group satisfying the conclusion of the above theorem has the \emph{small density property}.
Our hope is that, for certain Polish groups (first among them the unitary group $\mathcal U(\ell_2)$) one can obtain an outright automatic continuity theorem (i.e remove the assumption of $\partial$-lower semi-continuity of $H$).
In \cite{BenYaacov-Berenstein-Melleray:TopometricGroups} we showed that this is possible for $Aut([0,1],\lambda)$; we do not know if it is also possible for, say, $\mathcal U(\ell_2)$.
\section{Preliminaries}
We recall that the classical setting for descriptive set theory is that of \emph{Polish spaces}, i.e separable metrisable topological spaces whose topology is induced by a complete metric.
Our aim is to ``do some topology'', or descriptive set theory, where instead of considering subsets of a topological space $X$ we consider functions on $X$, say valued in $[0,\infty]$. For all the basic facts and theorems of descriptive set theory we use below, we refer the reader to \cite{Kechris:Classical}.
We start with a few general definitions and facts.
\begin{ntn}
For any two sets $A \subseteq X$ and two values $a,b$, we define $\langle A,a,b \rangle \colon X \to \{a,b\}$ by
\begin{gather*}
\langle A,a,b \rangle(x) =
\begin{cases}
a & x \in A, \\
b & x \notin A.
\end{cases}
\end{gather*}
\end{ntn}
\begin{ntn}
Given a function $\varphi\colon X \to [-\infty,\infty]$ and $r \in \bR$, we define $\varphi_{<r} = \bigl\{ x\colon \varphi(x) < r \bigr\}$, and similarly for $\varphi_{\leq r}$, $\varphi_{>r}$ and $\varphi_{\geq r}$.
\end{ntn}
Recall that a function $\varphi\colon X \to [-\infty,\infty]$ on a topological space $X$ is \emph{upper (lower) semi-continuous} if the set $\varphi_{<r}$ ($\varphi_{>r}$) is open for all $r \in \bR$.
\begin{lem}
\label{lem:InfUSC}
Let $\Phi$ be a family of upper semi-continuous functions on a topological space $X$.
Then $\inf \Phi \colon x \mapsto \inf \{f(x) \colon f \in \Phi\}$ is upper semi-continuous as well.
If $X$ admits a countable base then there exists a countable sub-family $\Phi_0 \subseteq \Phi$ such that $\inf \Phi = \inf \Phi_0$.
\end{lem}
\begin{proof}
The first assertion is quite standard.
For the second, let $\cB$ be a countable base for $X$.
We let $\Psi$ consist of all functions of the form $\langle U,q,\infty \rangle$, where $q \in \bQ$ and $U \in \cB$.
Then $\Psi$ is a countable family of upper semi-continuous functions on $X$, and every upper semi-continuous function $\varphi$ is the infimum of those $\psi \in \Psi$ greater than $\varphi$.
Now, choose $\Phi_0 \subseteq \Phi$ countable such that for each $\psi \in \Psi$, if there is $\varphi \in \Phi$ such that $\varphi \leq \psi$ then there is such $\varphi$ in $\Phi_0$ as well.
Then $\inf \Phi = \inf \Phi_0$.
\end{proof}
Whenever $A,B \subseteq X$ and $A \setminus B$ is meagre in $X$ we write $A \subseteq^* B$, and if $A \subseteq^* B \subseteq^* A$ then we write $A =^* B$.
A subset $A \subseteq X$ is called \emph{Baire measurable} if there exists an open set $U$ such that $A =^* U$.
The family of all Baire measurable sets forms a $\sigma$-algebra which contains all open sets and therefore all Borel sets.
When $X$ is Polish (or more generally, when it admits a countable base) we define
\begin{gather*}
U(A) = \bigcup \bigl\{ U \text{ open in } X \colon A \supseteq^* U \bigr\}.
\end{gather*}
The union is then equal to a countable sub-union, so $A \supseteq^* U(A)$, and $U(A)$ is the largest open set with this property.
The set $A$ is then Baire measurable if and only if $A =^* U(A)$, if and only if $A \subseteq^* U(A)$.
\begin{dfn}
Let $X$ be a Polish space, $\varphi\colon X \to [-\infty,\infty]$ any function.
We define
\begin{gather*}
\mathop{{\inf}^*}_{x \in X} \varphi(x) = \max \bigl\{ r\colon \forall^* x\, \varphi(x) \geq r \bigr\},
\qquad
\mathop{{\sup}^*}_{x \in X} \varphi(x) = \min \bigl\{ r\colon \forall^* x\, \varphi(x) \leq r \bigr\},
\end{gather*}
observing that the maximum and minimum are indeed attained.
\end{dfn}
\begin{prp}[Kuratowski-Ulam Theorem for functions]
Let $X$ and $Y$ be Polish spaces, $\varphi\colon X \times Y \to [-\infty,\infty]$ Baire measurable.
Then
\begin{gather*}
{\inf_{x,y}}^* \varphi(x,y) = {\inf_x}^* {\inf_y}^* \varphi(x,y) = {\inf_y}^* {\inf_x}^* \varphi(x,y),
\end{gather*}
and the sets
\begin{gather*}
\{x\colon y \mapsto \varphi(x,y)\text{ is Baire measurable} \},
\qquad
\{y\colon x \mapsto \varphi(x,y)\text{ is Baire measurable} \}
\end{gather*}
are co-meagre in $X$ and $Y$, respectively.
\end{prp}
\begin{proof}
The first assertion follows immediately from the set version of the Kuratowski-Ulam Theorem, applied to the Baire measurable set $\{(x,y)\colon \varphi(x,y) \geq r\}$.
For the second assertion, we know by the classical version that $A_{r,s} = \bigl\{ x\colon \{y\colon \varphi(x,y) \in (r,s) \} \text{ is Baire measurable} \bigl\}$ is co-meagre in $X$ for every interval $(r,s)$.
The intersection of all such sets, as $(r,s)$ varies over all open intervals with rational endpoints, is co-meagre as well, and is exactly the set $\{x\colon y \mapsto \varphi(x,y)\text{ is Baire measurable} \}$.
\end{proof}
\section{Graded sets}
Throughout, let $X$ denote a Polish space.
\begin{dfn}
By a \emph{graded subset} of $X$, denoted $\varphi \sqsubseteq X$, we mean a function $\varphi\colon X \to [0,\infty]$.
We say that a graded set $\varphi$ is \emph{open} ($\varphi \sqsubseteq_o X$) if it is upper semi-continuous as a function.
Similarly, it is \emph{closed} ($\varphi \sqsubseteq_c X$) if it is lower semi-continuous, and \emph{clopen} if it is both, namely continuous.
If $\varphi$ and $\psi$ are two graded subsets of $X$ then we say that $\varphi \sqsubseteq \psi$ if $\varphi \geq \psi$.
\end{dfn}
For example, one may identify an ordinary subset $A \subseteq X$ with its \emph{zero-indicator function} $\mathbold0_A = \langle A, 0, \infty \rangle$.
Notice that $\mathbold0_A$ is open, i.e., upper semi-continuous (respectively, closed, i.e., lower semi-continuous) if and only if $A$ is open (respectively, closed).
Also, we have $\varphi \sqsubseteq X$ if and only if $\varphi \sqsubseteq \mathbold0_X$.
One may further restrict graded subsets to values in $[0,1]$ (in which case we define $\mathbold0_A = \langle A,0,1 \rangle$), or extend to $[-\infty,\infty]$, since these are isomorphic ordered sets, equipped with the order topology, choosing one or the other has essentially no effect on most of our results.
\begin{lem}
Let $\varphi$ be a graded subset of $X$, and define
\begin{gather*}
U(\varphi) = \inf \bigl\{ \psi \sqsubseteq_o X\colon \varphi \leq^* \psi \bigr\}.
\end{gather*}
Then $\varphi \leq^* U(\varphi)$, $U(\varphi)$ is least u.s.c.\ with this property, and the following are equivalent:
\begin{enumerate}
\item We have $\varphi \geq^* U(\varphi)$.
\item We have $\varphi =^* U(\varphi)$.
\item There exists a u.s.c.\ function $\psi$ such that $\varphi =^* \psi$.
\item As a function, $\varphi$ is Baire measurable.
\end{enumerate}
\end{lem}
\begin{proof}
The first assertion is by \fref{lem:InfUSC} and the fact that a Polish space admits a countable base.
Then, implication from top to bottom is clear (since every upper semi-continuous function is Borel and therefore Baire measurable).
Assume now that $\varphi$ is Baire measurable, but for some $\varepsilon > 0$ the set $A = \{x\colon \varphi(x) < U(\varphi)(x) - \varepsilon\}$ is non meagre.
Then $A$ is Baire measurable, so $A =^* U(A)$, and since $X$ is a Baire space, $U(A) \neq \emptyset$.
Let $\psi = U(\varphi) - \varepsilon \mathbold1_{U(A)}$.
Then $\psi$ is also u.s.c.\ and $\varphi \leq^* \psi$, contradicting the minimality of $U(\varphi)$.
This contradiction implies that $\varphi \geq^* U(\varphi)$ concluding the proof.
\end{proof}
In other words, if we defined Baire measurable graded subsets as we did for ordinary subsets, this would coincide with the usual definition of Baire measurable functions. Accordingly, we will call a graded subset satisfying the equivalent conditions above a \emph{Baire-measurable graded subset} of $X$.
Throughout, let $G$ denote a Polish group, i.e a topological group whose topology is Polish. We now introduce two operations on graded subsets; the operations $\diamond$ reminds one of convolution.
\begin{dfn}
For two graded subsets $\varphi,\psi \sqsubseteq G$ we define $\varphi^{\diamond-1}, \varphi \diamond \psi \sqsubseteq G$ by
\begin{gather*}
\varphi^{\diamond-1}(x) = \varphi(x^{-1}), \qquad
\varphi \diamond \psi(x) = \inf_{h \in G} \varphi(h) + \psi(h^{-1}x).
\end{gather*}
\end{dfn}
Note that $\diamond$ is associative and has $\mathbold0_{\{1_G\}}$ as a neutral element.
We observe that for $A,B \subseteq G$, $\mathbold0_A^{\diamond-1} = \mathbold0_{A^{-1}}$ and $\mathbold0_A \diamond \mathbold0_B = \mathbold0_{A\cdot B}$.
Thus, ${}^{\diamond-1}$ and $\diamond$ extend the group operations of $G$, applied to subsets, to graded subsets (and should be thought of as operations on subsets, rather than as operations on group elements).
As expected, we have, for all graded subsets $\varphi,\psi$ of $G$, that
\begin{gather*}
\left(\varphi \diamond \psi \right)^{\diamond-1} = \psi^{\diamond-1} \diamond \varphi^{\diamond-1}.
\end{gather*}
By extension, for $g \in G$ we define $g\varphi = \mathbold0_{\{g\}} \diamond \varphi$, namely $(g\varphi)(x) = \varphi(g^{-1}x)$, so $g \mathbold0_A = \mathbold0_{g\cdot A}$.
We then obtain
\begin{gather*}
\varphi \diamond \psi(g) = \inf \varphi + g\psi^{\diamond-1}.
\end{gather*}
\begin{prp}[Pettis Theorem for graded subsets]
\label{prp:Pettis}
Let $G$ be a Polish group, $\varphi,\psi \sqsubseteq G$ graded subsets.
Then $U(\varphi) \diamond U(\psi) \sqsubseteq \varphi \diamond \psi$.
\end{prp}
\begin{proof}
It will be enough to show that if $U(\varphi)\diamond U(\psi)(g) < r$ for some $g \in G$ and $r \in \bR$ then $\varphi \diamond \psi(g) < r$.
Indeed, in this case the set $\bigl\{x\colon \bigl( U(\varphi) + gU(\psi)^{\diamond-1} \bigr)(x) < r \bigr\}$ is a non empty open subset of $G$, and in particular is not meagre.
Since $\varphi + g\psi^{\diamond-1} \leq^* U(\varphi) + gU(\psi)^{\diamond-1}$, we obtain that $\bigl\{ x \colon ( \varphi + g\psi^{\diamond-1} )(x) < r \bigr\}$ is non meagre, and in particular non empty, so $\varphi \diamond \psi(g) = \inf (\varphi + g \psi^{\diamond-1}) < r$, as desired.
\end{proof}
\begin{dfn}
Let $X$ be a Polish space, $\varphi \sqsubseteq X$ a graded subset.
\begin{enumerate}
\item We say that $\varphi$ is \emph{meagre} if there exists $r > 0$ such that $\varphi_{< r}$ is meagre, i.e., such that $\forall^*x \, \varphi(x) \geq r$.
This implies $U(\varphi) \geq r$, and for Baire measurable $\varphi$ the two are equivalent.
\item We define $\varphi^\circ$, the \emph{interior} of $\varphi$ to be the least u.s.c.\ $\psi$ greater than $\varphi$.
\item We define $\overline \varphi$, the \emph{closure} of $\varphi$ to be the greatest l.s.c.\ $\psi$ less than $\varphi$.
\end{enumerate}
One can check that
\begin{gather*}
\varphi^\circ(x) = \limsup_{y \to x} \varphi(y), \qquad \overline \varphi(x) = \liminf_{y \to x} \varphi(y).
\end{gather*}
\end{dfn}
\begin{lem}
\label{lem:NonMeagreBaireMeasruableGradedSubset}
Let $\varphi \sqsubseteq G$ be a non meagre Baire measurable graded subset.
Then $(\varphi \diamond \varphi^{\diamond-1})^\circ(1) = 0$
\end{lem}
\begin{proof}
Let $\psi = U(\varphi)$.
Since $\varphi$ is Baire measurable and non meagre, we have $\inf \psi = 0$, and thus $\psi \diamond \psi^{\diamond-1}(1) = 2\inf \psi = 0$.
By \fref{prp:Pettis} we have $\psi \diamond \psi^{\diamond-1} \sqsubseteq \varphi \diamond \varphi^{\diamond-1}$, and since $\psi \diamond \psi^{\diamond-1} \sqsubseteq_o G$, we obtain $\psi \diamond \psi^{\diamond-1} \sqsubseteq (\varphi \diamond \varphi^{\diamond-1})^\circ$.
Thus $(\varphi \diamond \varphi^{\diamond-1})^\circ(1) = 0$.
\end{proof}
We said that graded subsets of $X$ are a natural analogue of subsets of $X$; we pursue this analogy further by introducing the analogue of subgroups in this context.
\begin{dfn}
A \emph{graded subgroup} of a Polish group $G$ is a graded subset $H \sqsubseteq G$ satisfying the following properties:
\begin{itemize}
\item $H(1) = 0$,
\item $H^{\diamond-1} = H$ (i.e., $H(x) = H(x^{-1})$), and
\item $H \diamond H \sqsubseteq H$ (i.e., $H \diamond H \geq H$, or equivalently, $H(h) + H(g) \geq H(hg)$).
\end{itemize}
\end{dfn}
Notice that even if we did allow negative values for graded subsets, a graded subgroup would still always be non negative.
Also, $H(1) = 0$ implies that $H \diamond H \sqsupseteq H$, so in fact $H \diamond H = H$.
Finally, if $H \sqsubseteq G$ is a graded subgroup then so is $\overline H$, since
\begin{gather*}
\overline H(g) + \overline H(h) = \liminf_{g'\to g, h' \to h} H(g') + H(h') \geq \liminf_{g'\to g, h' \to h} H(g'h') = \overline H(gh).
\end{gather*}
\begin{rmk}
What we call a graded subgroup of $G$ is usually called a \emph{semi-norm} on $G$. These correpond to left-invariant pseudo-metrics on $G$ as follows: for a graded subgroup $H$ of $G$ and left-invariant pseudo-metric $d$ on $G$ define
\begin{gather*}
d_H(g,h) = H(g^{-1}h), \qquad H_d(g) = d(1,g).
\end{gather*}
Then $d \mapsto H_d$ and $H \mapsto d_H$ are inverses, yielding a natural bijection between graded subgroups of a group $G$ and left-invariant pseudo-distances on $G$ (allowing $+\infty$ as a value for the pseudo-distance, or excluding it for the graded subgroup).
Notice also that $H \sqsubseteq G$ is closed (open) if and only if $d_H$ is, and more generally, $\overline{d_H} = d_{\overline H}$ and $d_H^\circ = d_{H^\circ}$.
\end{rmk}
\begin{lem}
\label{lem:OpenGradedSubgroup}
Let $H \sqsubseteq G$ be a graded subgroup.
Then the following are equivalent
\begin{enumerate}
\item $\inf H^\circ = 0$.
\item $H$ is clopen in $G$.
\item $H$ is open in $G$.
\end{enumerate}
\end{lem}
\begin{proof}
It is enough to show that if $\inf H^\circ = 0$ then $H$ is clopen.
First, we observe that this implies that $H^\circ(1) = 0$. Indeed, if $H^\circ(y) < \varepsilon$ then one has $H(y^{-1})=H(y) < \varepsilon$, and then the fact that $H$ is a graded subgroup gives
$$H^{\circ}(1)=\limsup_{x \to 1} H(x) \le \limsup_{x \to 1} H(xy) +H(y^{-1})=H^\circ(y)+H(y^{-1}) < 2 \varepsilon \ . $$
Let us show that $H$ is open, i.e., that $H_{<r}$ is open for all $r$.
Indeed, assume that $H(g) = r - \delta$ for some $\delta > 0$.
Let $U = H^\circ_{<\delta}$, which is open and contains the identity by assumption, so $gU$ is a neighbourhood of $g$.
In addition, $H^\circ \sqsubseteq H$ implies that $H(h) < \delta$ for all $h \in U$.
Thus, if $h \in U$ then $H(gh) \leq H(g) + H(h) < H(g) + \delta = r$, so $H(x) < r$ for all $x \in gU$.
Now let us show that $H$ is closed, i.e., that $H_{>r}$ is open.
Assume this time that $H(g) = r + \delta$ for some $\delta > 0$, and let $U$ be as before.
If $h \in U$ then $H(g) \leq H(gh) + H(h^{-1}) < H(gh) + \delta$, so $H(x) > H(g)-\delta \ge r$ for all $x \in gU$, and the proof is complete.
\end{proof}
\begin{lem}
If $\varphi\sqsubseteq_o G$ and $\psi \sqsubseteq G$ then $\varphi^{\diamond-1} \sqsubseteq_o G$ and $\varphi \diamond \psi \sqsubseteq_o G$.
\end{lem}
\begin{proof}
The first assertion is obvious.
The second assertion is also immediate when one goes back to the definition of $\varphi \diamond \psi$: if $\varphi \diamond \psi (g) < r$ then there is some $x$ such that $\varphi(x) + \psi(x^{-1}g) =r- \delta < r$; the set $V=\{y \colon \psi(y) < \psi(x^{-1}g)+\delta\}$ is open and contains $x^{-1}g$, so $xV$ is open, contains $g$, and for $h \in xV$ one has $\varphi(x)+\psi(x^{-1}h) < r$.
This implies in particular that $\varphi \diamond \psi (h) < r$ for all $h \in xV$.
\end{proof}
\begin{lem}
\label{lem:NonMeagreBaireMeasruableGradedSubgroup}
Let $H \sqsubseteq G$ be a non meagre Baire measurable graded subgroup.
Then $H$ is open, and thus clopen.
\end{lem}
\begin{proof}
By definition, $H \diamond H^{\diamond-1} = H \diamond H \sqsubseteq H$, implying $(H \diamond H^{\diamond-1})^\circ \geq H^\circ$.
By \fref{lem:NonMeagreBaireMeasruableGradedSubset} this yields $H^\circ(1) = 0$, so $H$ is clopen by \fref{lem:OpenGradedSubgroup}.
\end{proof}
In usual terminology, this lemma says that if a semi-norm $H$ on a Polish group $G$ is Baire-measurable and for any $\varepsilon >0$ the set of $g$ such that $H(g) \le \varepsilon$ is non meagre, then $H$ must be continuous.
\section{The small density property}
\begin{dfn}
Let $G$ be a Polish group, $H \sqsubseteq G$ a graded subgroup.
Then we define the \emph{index} $[G:H]$ to be the density character of $d_H$, namely the least cardinal of a $d_H$-dense subset of $G$.
\end{dfn}
We observe that if $H$ is an ordinary subgroup then this agrees with the usual definition of index.
Here we are going to be interested in the condition $[G:H] < 2^{\aleph_0}$ (i.e., ``$H$ has small index'').
Since the cofinality of $2^{\aleph_0}$ is uncountable, the following are equivalent:
\begin{enumerate}
\item $[G:H] < 2^{\aleph_0}$.
\item For all $\varepsilon > 0$, $G$ can be covered by fewer than continuum many left translates of $H_{<\varepsilon}$.
\item For all $\varepsilon > 0$, a family of disjoint left translates of $H_{<\varepsilon}$ has cardinal smaller than the continuum.
\end{enumerate}
Recall from \cite{BenYaacov-Berenstein-Melleray:TopometricGroups} that a \emph{Polish topometric group} is a triple $(G,\tau,\partial)$, where $(G,\tau)$ is a Polish group, $\partial$ is a complete bi-invariant metric refining $\tau$ and $\tau$-lower semi-continuous, i.e such that for any $r$ the sets $\{(g,h) \colon \partial(g,h) \le r\}$ are closed in $G \times G$.
The canonical example one should have in mind when thinking of this is the isometry group of some Polish metric space $(X,d)$ (or, more generally, the automorphism group of some Polish metric structure), endowed with the topology of pointwise convergence and the supremum metric $\partial(g,h)= \sup \{d(gx,hx) \colon x \in X \}$.
\begin{dfn}
We say that a Polish topometric group $(G,\tau,\partial)$ has the \emph{small density property} if whenever $H \sqsubseteq G$ is a $\partial$-closed graded subgroup of index $< 2^{\aleph_0}$ then $H$ is open.
\end{dfn}
\begin{rmk}
We do not call this property the small index property, because even when $\partial$ is the discrete metric the small density property as defined above is stronger than the usual small index property (which corresponds to left-invariant \emph{ultrametrics} rather than left-invariant metrics).
\end{rmk}
\begin{thm}
\label{thm:SmallDensity}
Let $(G,\tau,\partial)$ be a Polish topometric group admitting ample generics.
Then $G$ has the small density property.
\end{thm}
\begin{proof}
If $A$ is a subset of $G$, we will denote by $(A)_{<\varepsilon}$ the et $\{g \colon \exists a \in A \partial(g,a) < \varepsilon\}$.
Let $H \sqsubseteq G$ be $\partial$-closed graded subgroup of small index.
For convenience, let us allow ourselves to consider $\partial$ as a unary function, $\partial(x) = \partial(x,1) = \partial(1,x)$.
For $n \in \bN$ define $H^{n\partial} = H \diamond n\partial$, namely (by invariance)
\begin{gather*}
H^{n\partial}(g) = \inf_h\, H(h) + n\partial(h,g).
\end{gather*}
We observe that by symmetry, $H^{n\partial} = n\partial \diamond H$ as well.
In addition, $H \diamond H^{n\partial} = H \diamond H \diamond n\partial = H \diamond n\partial = H^{n\partial}$.
Clearly $H^{n\partial} \leq H$; it is also easy to check that $\sup H^{n\partial} = \overline H^\partial$ (that $\sup_n H^{n\partial} \le \overline H^{\partial}$ comes from the existence of a sequence $(h_n)$ such that $\partial(g,h_n)< 1/n^2$ and $H(h_n) \to \overline H^\partial(g)$; the other direction is similar, using the fact that $\overline H^{\partial}(g)$ is the $\liminf$ of $H(h)$ as $h$ $\partial$-converges to $g$). From this fact and our assumption on $H$, we obtain $H = \sup H^{n\partial}$.
By the Kuratowski-Mycielski Theorem, $H$ cannot be meagre, that is, $H_{<\varepsilon}$ is not meagre for any $\varepsilon > 0$.
Assume, for a contradiction, that for some $\varepsilon > 0$ the set $G \setminus (H_{<\varepsilon})_\varepsilon$ is non meagre in every open subset of $G$. Below, for $h \in G$ and $A \subseteq G$ we denote by $A^h$ the set $h^{-1}Ah$.
By \cite[Lemma~3.4]{BenYaacov-Berenstein-Melleray:TopometricGroups} we can find a mapping $a \in 2^\omega \mapsto h_a \in G$ such that if $a,b \in 2^\omega$ are distinct then $\partial\bigl( H_{<\varepsilon/3}^{h_a},[G \setminus (H_{<\varepsilon})_\varepsilon]^{h_b} \bigr) < \varepsilon$, i.e., $\partial\bigl( H_{<\varepsilon/3}^{h_ah_b^{-1}},G \setminus (H_{<\varepsilon})_\varepsilon \bigr) < \varepsilon$.
It follows that $h_ah_b^{-1} \notin H_{<\varepsilon/3}$ for all $a \neq b$, implying that $[G:H] = 2^{\aleph_0}$, a contradiction.
Thus, for all $\varepsilon > 0$, the set $(H_{<\varepsilon})_\varepsilon$ is co-meagre in some non empty open set.
Applying Pettis' Theorem to $\varepsilon/2$ we see that $(H_{<\varepsilon})_\varepsilon$ contains a neighbourhood of the identity.
We remark that this last observation implies (in fact, is equivalent to) $(H^{n\partial})^\circ(1) = 0$ for all $n$. Indeed, if $V$ is an open set contained in $(H_{< \varepsilon/2n})_{\varepsilon/2n}$ then for all $v \in V$ there is $h$ such that $H(h)< \varepsilon /2n$ and $\partial(h,v) < \varepsilon/2n$, which implies that $H^{n\partial}(v) \le H(h) +n\partial(h,v)< \varepsilon$.
Now fix $n$, $\varepsilon > 0$ and a non empty open set $U \subseteq G$.
Then there exists a neighbourhood $1 \in V$ such that $H^{n\partial}(V) < \varepsilon/2$.
If $g \in G$ then $H^{n\partial}(gV) = (H \diamond H^{n\partial})(gV) < H(g) + \varepsilon/2$.
By definition of $\overline H$, the set $(H-\overline H)_{<\varepsilon/2}$ is dense, so we may choose $g \in (H-\overline H)_{<\varepsilon/2} \cap U$.
In particular, $g \in \overline H_{> H(g)-\varepsilon/2}$, and this last set is open.
Thus $W = gV \cap \overline H_{> H(g)-\varepsilon/2} \cap U$ is a non empty open set, and $H^{n\partial} < \overline H + \varepsilon$ on $W$.
Thus, for any open set $U$ there is some nonempty open subset $W \subseteq U$ such that $H^{n\partial}< \overline H+\varepsilon$ on $W$. Hence $H^{n\partial} < \overline H + \varepsilon$ on a dense open set for all $\varepsilon > 0$.
It follows that $H^{n\partial} \leq^* \overline H$ and therefore $H = \sup_n H^{n\partial} \leq^* \overline H$.
In particular, $U(H) = U(\overline H)$.
Now, $\overline H$ is closed and non meagre (since $H$ is non meagre), so $\overline H_{\leq \varepsilon}$ is closed and non meagre, and therefore of non empty interior, for all $\varepsilon > 0$.
It follows that $\inf \overline H^\circ = 0$ (recall that $\overline H^{\circ}(g)= \limsup_{h \to g} \overline H(h)$), so by Lemma \ref{lem:OpenGradedSubgroup} $\overline H$ is clopen.
In particular $U(\overline H) = \overline H$.
Thus $U(H) = \overline H$, and by Pettis' Theorem $\overline H = \overline H \diamond \overline H = U(H) \diamond U(H) \geq H \diamond H = H$.
Thus $H = \overline H$ is clopen.
\end{proof}
It is probably worthwhile to translate this result in the more common language of left-invariant metrics.
Assume that $(G,\tau,\partial)$ is a Polish topometric group with ample generics, and let $\rho$ be a left-invariant pseudo-metric on $G$.
Let $G_{\rho}$ denote the metric space obtained by identifying points $g,g'$ such that $\rho(g,g')=0$, and assume that the density character of $G_{\rho}$ is strictly less than the continuum.
The theorem above says that, under these assumptions, $\rho$ must be continuous (with regard to the topology $\tau$ of $G$) as soon as each set
$\{g \colon \rho(g,1) \le r\}$ is $\partial$-closed. Even without this assumption on the level sets of $\rho$, one can obtain some information: indeed, the function $\rho'$ defined by
$$\rho'(g,h)= \liminf_{g' \xrightarrow[\partial]{} g, h' \xrightarrow[\partial]{} h} \rho(g',h')$$
is still a left-invariant pseudo metric on $G$, and the density character of the associated metric space is still less than the continuum. The $\liminf$ ensures that the associated graded subgroup is $\partial$-closed, so our theorem says that $\rho'$ must be continuous with respect to the topology $\tau$.
The result above is insufficient to obtain a full automatic continuity theorem for, say, $\mathcal U(\ell_2)$, the unitary group of separable Hilbert space. This is a Polish topometric group with ample generics when endowed with the strong operator topology and the norm-operator metric, i.e
$$\partial(S,T)= \|S-T\|. $$
Actually, we do not even know whether $\mathcal U(\ell_2)$ admits a non-trivial homomorphism into $\mathfrak S_{\infty}$, the permutation group of the integers. Such a homomorphism would necessarily be discontinuous; it can exist if and only if $\mathcal U(\ell_2)$ admits a non-trivial subgroup of countable index.
Note that the results of \cite{Fong-Sourour:NormalSubgroups} imply that there are no normal subgroups of \emph{finite} index in $\mathcal U(\ell_2)$, so there are no nontrivial subgroups of finite index. This says nothing, however, about the existence of subgroups with countable index.
Assume that $\Gamma$ is such a subgroup, and let $H_{\Gamma}$ denote the graded subgroup that takes the value $0$ on $\Gamma$ and $1$ everywhere else.
Then, using the remark following \fref{thm:SmallDensity}, and the fact that the strong operator topology is connected, we see that for all $U \in \mathcal U(\ell_2)$ one has
\begin{gather*}
\liminf_{T \xrightarrow[\| \cdot \|]{} U} H_{\Gamma}(T)=0.
\end{gather*}
In other words, $\Gamma$ must be norm-dense in $\mathcal U(\ell_2)$. We do not know how to obtain more information about $\Gamma$; using a standard diagonal argument one may show that there must exist an infinite-dimensional subspace $\mathcal H_0$ of $\ell_2$ such that $\Gamma$ contains all unitary operators with support contained in $\mathcal H_0$, but this does not seem to be of much help either. So we are left with the following section to ponder:
\begin{quote}
\emph{Does there exist a nontrivial subgroup of $\mathcal U(\ell_2)$ which has a countable index?}
\end{quote}
All we said above would make sense also for the isometry group of the Urysohn space: any countable index subgroup must be dense for the uniform metric, and we do not know whether a nontrivial such subgroup exists.
We should also note here that R.\ Kallman has proved in \cite{Kallman:MatrixGroups} that there are plenty of nontrivial homomorphisms from a \emph{finite-dimensional} unitary group into $\mathfrak S_{\infty}$; actually, such homomorphisms separate points for any finite-dimensional linear group.
\section{A remark concerning graded subgroups of automorphism groups}
We conclude with a natural situation in which open graded subgroups arise.
We assume that the reader of this section is familiar with the formalism of continuous logic (\cite{BenYaacov-Usvyatsov:CFO,BenYaacov-Berenstein-Henson-Usvyatsov:NewtonMS}).
Let $\cM$ be a metric structure, $G = \Aut(\cM)$.
If $\cM$ is a classical structure then for every member $a \in M^{eq}$, the stabiliser $G_a$ is an open subgroup, and if $\cM$ is $\aleph_0$-categorical then every open subgroup is the stabiliser of some (real or imaginary) element.
In the metric case, on the other hand, the stabiliser $G_a$ is usually \emph{not} open in $G$, and again we encounter the need to consider graded subgroups.
\begin{dfn}
Let $\cM$ be a metric structure, $G = \Aut(\cM)$.
The \emph{(graded) stabiliser} of $a \in M$, still denoted $G_a$, is defined by $G_a(g) = d(a,ga)$.
\end{dfn}
The graded stabiliser is, by definition, an open graded subgroup, and the exact stabiliser is $G_{a,\leq 0}$.
As in classical logic, the converse holds for $\aleph_0$-categorical structures.
\begin{prp}
Let $\cM$ be an $\aleph_0$-categorical separable structure, $G = \Aut(\cM)$, and let $H \sqsubseteq_o G$ be a real-valued open graded subgroup (namely, we exclude $+\infty$ from the range).
Then there exists an imaginary $a \in M^{eq}$ such that $H = G_a$.
\end{prp}
\begin{proof}
We may consider $M^\bN$ as a sort, equipped with the distance
\begin{gather*}
d(b,c) = \bigvee_n 2^{-n} \wedge d(b_n,c_n).
\end{gather*}
We fix $a$ in this sort which enumerates a dense subset of $\cM$.
By homogeneity, the set of realisations of $p = \tp(a)$ in $\cM$ is exactly $\overline{Ga}$.
We observe that $\{G_{a,<\varepsilon}\}_{\varepsilon > 0}$ is a base of neighbourhoods of the identity, so for every $\varepsilon > 0$ there exists $\delta(\varepsilon) > 0$ such that $G_{a,<\delta(\varepsilon)} < H_{<\varepsilon}$.
Let $\varepsilon > 0$ and $g,g',h,h' \in G$, and assume that $d(ga,g'a)$ and $d(ha,h'a)$ are smaller than $\delta = \delta(\varepsilon/2)$.
Then $g^{-1}g', h^{-1}h' \in G_{a,<\delta} \subseteq H_{<\varepsilon/2}$, whereby $|d_H(g,h) - d_H(g',h')| < \varepsilon$.
Thus the map $\varphi\colon Ga \times Ga \to \bR$ sending $(ga,ha) \mapsto d_H(g,h)$ is well defined and uniformly continuous, and thus extends uniquely to a uniformly continuous function $\varphi\colon \overline{Ga} \times \overline{Ga} \to [0,\infty]$.
By the Ryll-Nardzewski Theorem \cite[Fact~1.14]{BenYaacov-Usvyatsov:dFiniteness}, since $\varphi$ is uniformly continuous and invariant by automorphism, it is a definable pseudo-metric on the set defined by $p$ (and therefore in particular bounded).
By \cite{BenYaacov:DefinabilityOfGroups}, and since by $\aleph_0$-categoricity the set defined by $p$ is definable, $\varphi$ extends to a definable pseudo-metric on all of $M^\bN$.
(To recall the argument, by the Tietze extension theorem we may extend $\varphi$ to something definable on all of $M^\bN \times M^\bN$, call it $\varphi_0(x,y)$, and then $\varphi_1(x,y) = \sup_{z \vDash p} |\varphi_0(x,z) - \varphi_0(y,z)|$ is a definable pseudo-metric which agrees with $\varphi$ on $p$.)
Finally, with $[b]$ denoting the canonical parameter for $\varphi(x,b)$,
\begin{gather*}
H(g) = d_H(g,1) = \varphi(ga,a) = d\bigl( [a], [ga] \bigr).
\end{gather*}
Thus $H$ is precisely the graded stabiliser of $[a]$.
\end{proof}
\bibliographystyle{begnac}
|
\section{Introduction}
Recently, Coulomb effects on polarization transfer from polarized
electrons or positrons to initially unpolarized protons or antiprotons
in elastic electromagnetic scattering have been studied in a distorted wave
approximation at low energies~\cite{Are07}. These studies were
motivated by the idea to polarize hadrons by their scattering on polarized
electrons or positrons in a storage ring~\cite{HoM94}. However, in
view of such a design it turned out that the considered observable,
i.e.\ the total cross section for the scattering of initially unpolarized
hadrons off polarized leptons to polarized final hadrons, the
polarization transfer $P_{z00z}$, cannot
contribute to a net polarization of the hadrons in a storage
ring. The reason for that is that this polarization observable does not
contain a genuine hadronic spin-flip process~\cite{WaA07,MiS08}, which is
necessary for a net polarization change. Moreover, our previous numerical results were
critizised by Milstein et al.~\cite{MiS08} who had taken a partial
wave expansion of the Coulomb scattering wave function instead of the
integral representation used in ref.~\cite{Are07}. Indeed, it turned out that
besides a minor error the main reason for the gross overestimation of the
polarisation transfer cross section was an accuracy problem in the
numerical evaluation, namely,
the relevant quantity was calculated as a difference of two
almost equal numbers multiplied by a huge factor~\cite{Are08,WaA08}.
For these reasons I have extended the previous study to the formal consideration of all
possible polarization observables in this scattering reaction including such
spin-flip transitions using again the distorted wave Born approximation.
In addition to the previously considered hyperfine interaction I have
included also the spin-orbit interactions of lepton and hadron. In the
next section the most general scattering cross section is introduced,
defining the various polarization observables in terms of bilinear
hermitean forms of the $T$-matrix elements.
For the nonrelativistic form of the $T$-matrix with
inclusion of hyperfine and spin-orbit interactions the detailed
expression of the general scattering cross section is given, allowing for the
polarization of all initial and final particles described by
corresponding spin density matrices.
In Section III I specialize to the case where the polarization of the
final lepton is not measured to the so-called triple polarization cross
section. For the numerical evaluation two different methods have been
applied, a partial wave expansion as in ref.~\cite{MiS08} and an
integral representation of the Coulomb wave function according to
ref.~\cite{LeA01}. Results for the structure functions and spin-flip
triple cross sections for the case of polarization along the incoming hadron
momentum are presented in Section IV and a summary
is given in Section V. Details for the evaluation of the hyperfine and
spin-orbit interactions are given in an appendix.
\section{The general differential cross section including
polarization of all particles}
Reviews on polarization phenomena may be found for lepton hadron
scattering in~\cite{Dom69}, for nuclear physics in~\cite{Ohl72} and
for nucleon nucleon scattering in~\cite{ByL78}.
I will consider hadron-lepton scattering in
the c.m.\ system, where hadron stands for proton or antiproton and
lepton for electron or positron,
\begin{equation}
h(\mathbf p \,) + l(-\mathbf p \,) \longrightarrow h(\mathbf
p^{\,\prime}) + l(-\mathbf p^{\,\prime}) \,,
\end{equation}
allowing for inital and final hadron and lepton polarization. The
hadron initial and final three momenta are denoted by $\mathbf p$ and
$\mathbf p^{\,\prime}$, respectively. All possible observables of this
reaction can be obtained from the ``quadruple polarization'' cross
section for which the spin states of all initial and final particles
are described by the corresponding general spin density matrices
$\rho^{l/h}({\mathbf P}^{i/f}_{l/h})$, where the initial density matrices
characterize the spin properties of target and beam and the final ones
those of the detectors. It is given by the general trace
\begin{eqnarray}
\frac{d\sigma^{quadruple}_{{\mathbf P}^{f}_{h},{\mathbf P}^{i}_{h},{\mathbf P}^{f}_{l},
{\mathbf P}^{i}_{l}}(\theta,\phi)}{d\Omega}&=&
{\cal O}({\mathbf P}^{f}_{h},{\mathbf P}^{f}_{l},{\mathbf P}^{i}_{h}
{\mathbf P}^{i}_{l};\theta,\phi) \nonumber\\
&=&\frac{M_l^2M_h^2}{\pi^2W^2(1+|{\mathbf P}^{f}_{l}|)(1+|{\mathbf P}^{f}_{h}|)}\,
\mbox{Trace}\Big[\widehat T^\dagger
\widehat \rho^{\, h}({\mathbf P}^{f}_{h})\widehat \rho^{\, l}({\mathbf P}^{f}_{l})\,
\widehat T\,\widehat \rho^{\, h}({\mathbf P}^{i}_{h})\widehat \rho^{\, l}({\mathbf P}^{i}_{l})\Big]\,,
\label{basic_xs}
\end{eqnarray}
where $\widehat T=\widehat T(\theta,\phi)$ denotes the T-matrix of the
scattering process with $(\theta,\phi)$ as scattering angles, and
$\rho(\mathbf P)$ the spin density matrix for a spin-$1/2$ particle,
with ${\mathbf P}$ characterizing the polarization of the
corresponding particle in the initial and final states, respectively.
The trace refers to the hadron and lepton spin degrees of freedom. The
factor in front takes into account the final phase space, the
incoming flux, and a normalization factor for the case of partially
polarized final states. The invariant energy of the hadron-lepton
system is denoted by $W=E_h+E_l$ and the
masses of hadron and lepton by $M_h$ and $M_l$, respectively.
In the c.m.\
frame I use as reference system the $z$-axis along the incoming
hadron momentum $\mathbf p$. The $x$- and $y$-axes are chosen to
form a right handed orthogonal system.
In view
of the fact, that in this work I am interested in the low energy
regime, a nonrelativistic framework is adopted.
The nonrelativistic density matrices for possible polarization of initial and
final states of a spin-$1/2$ particle have the standard form
\begin{eqnarray}
\widehat \rho\,({\mathbf P})&=&\frac{1}{2}(1+{\mathbf
P}\cdot {\boldsymbol\sigma})\,.
\end{eqnarray}
with the vector ${\mathbf P}$ describing the polarization of the
particle and $\boldsymbol\sigma$ denoting the Pauli spin vector. One
should note that in general $|\mathbf P^{i/f}_{h/l}|\le 1$.
From the basic equation
(\ref{basic_xs}) one obtains all possible polarization observables.
In detail they are:
\begin{description}
\item[(i)]
The unpolarized differential cross section:
\begin{equation}
\frac{d\sigma_0(\theta,\phi)}{d\Omega}={\cal O}
(\mathbf 0,\mathbf 0,\mathbf 0,\mathbf 0;\theta,\phi)
=S^0(\theta,\phi)\,.
\end{equation}
\item[(ii)]
Beam, target and beam-target asymmetries of the differential cross
section for unpolarized final states in the notation of
Bystricky et al.~\cite{ByL78}:
\begin{eqnarray}
\frac{d\sigma_{{\mathbf P}^{i}_{h},{\mathbf P}^{i}_{l}}(\theta,\phi)}{d\Omega}&=&
{\cal O}(\mathbf 0,\mathbf 0,{\mathbf P}^{i}_{h},{\mathbf
P}^{i}_{l};\theta,\phi)\nonumber\\
&=&\frac{d\sigma_0(\theta,\phi)}{d\Omega}\Big(1+\sum_j
P^{i}_{h,j}A_{00j0}(\theta,\phi) + \sum_k P^{i}_{l,k} A_{000k}(\theta,\phi)
+\sum_{j,k} P^{i}_{h,j} P^{i}_{l,k} A_{00jk} (\theta,\phi)\Big)\,,\label{beam-target-asy}
\end{eqnarray}
with the asymmetry vectors
\begin{eqnarray}
A_{00j0}(\theta,\phi)&=&\frac{1}{S^0}\frac{ \partial}{\partial {P}^{i}_{h,j}}
{\cal O}(\mathbf 0,\mathbf 0,{\mathbf P}^{i}_{h},\mathbf 0;\theta,\phi)\nonumber\\
&=&
\frac{1}{2S^0}\Big(
\frac{d\sigma_{{\mathbf P}^{i}_{h},{\mathbf P}^{i}_{l}}}{d\Omega}
-\frac{d\sigma_{-{\mathbf P}^{i}_{h},{\mathbf P}^{i}_{l}}}{d\Omega}
\Big)\Big|_{{P}^{i}_{h,k}=\delta_{jk}}\,,\\
A_{000j}(\theta,\phi)&=&\frac{1}{S^0}\frac{\partial}{\partial {P}^{i}_{l,j}}
{\cal O}(\mathbf 0,\mathbf 0,\mathbf 0,{\mathbf
P}^{i}_{l};\theta,\phi)\nonumber\\
&=&\frac{1}{2S^0}\Big(
\frac{d\sigma_{{\mathbf P}^{i}_{h},{\mathbf P}^{i}_{l}}}{d\Omega}
-\frac{d\sigma_{{\mathbf P}^{i}_{h},-{\mathbf P}^{i}_{l,}}}{d\Omega}
\Big)\Big|_{{P}^{i}_{l,k}=\delta_{jk}}\,,
\end{eqnarray}
and the hadron-lepton asymmetry tensor
\begin{eqnarray}
A_{00jk}(\theta,\phi)&=&\frac{1}{S^0}
\frac{\partial^2}{\partial {P}^{i}_{h,j}\partial {P}^{i}_{l,k}}
{\cal O}(\mathbf 0,\mathbf 0,{\mathbf P}^{i}_{h},{\mathbf
P}^{i}_{l};\theta,\phi)\nonumber\\
&=&\frac{1}{4S^0}\Big(
\frac{d\sigma_{{\mathbf P}^{i}_{h},{\mathbf P}^{i}_{l}}}{d\Omega}
+\frac{d\sigma_{-{\mathbf P}^{i}_{h},-{\mathbf P}^{i}_{l}}}{d\Omega}
-\frac{d\sigma_{-{\mathbf P}^{i}_{h},{\mathbf P}^{i}_{l}}}{d\Omega}
-\frac{d\sigma_{{\mathbf P}^{i}_{h},-{\mathbf P}^{i}_{l}}}{d\Omega}\Big)
\Big|_{P^{i}_{h,m}=\delta_{jm},\, P^{i}_{l,n}=\delta_{kn}}\,.
\end{eqnarray}
\item[(iii)]
Polarization of the final lepton or hadron for unpolarized beam and target:
\begin{eqnarray}
P_{0j00}(\theta,\phi)&=&\frac{1}{S^0}\frac{\partial}{\partial {P}^{f}_{l,j}}
{\cal O}(\mathbf 0,{\mathbf P}^{f}_{l},\mathbf 0,\mathbf 0;\theta,\phi)\,,\\
P_{j000}(\theta,\phi)&=&\frac{1}{S^0}\frac{\partial}{\partial {P}^{f}_{h,j}}
{\cal O}({\mathbf P}^{f}_{h},\mathbf 0,\mathbf 0,\mathbf 0;\theta,\phi)\,.
\end{eqnarray}
\item[(iv)]
Various correlations between the polarization of one outgoing particle and beam
and/or target polarizations. For example, the outgoing hadron
polarization for initial lepton polarization but unpolarized incoming
hadron, the lepton-hadron polarization transfer is given by
\begin{eqnarray}
P_{j00k}(\theta,\phi)&=&\frac{1}{S^0}
\frac{\partial^2}{\partial {P}^{f}_{h,j}\partial {P}^{i}_{l,k}}
{\cal O}({\mathbf P}^{f}_{h},\mathbf 0 ,\mathbf 0,{\mathbf P}^{i}_{l};\theta,\phi)\,.
\end{eqnarray}
\item[(v)]
Another interesting example is the hadron spin-flip of an initially
polarized hadron by the scattering on an initially polarized lepton. It is
a special case of the so-called ``triple polarization'' cross section
with all particles polarized except for the final lepton as defined by
\begin{equation}
\frac{d\sigma^{triple}_{{\mathbf P}^{f}_{h},{\mathbf P}^{i}_{h},{\mathbf
P}^{i}_{l}}(\theta,\phi)}{d\Omega}=
{\cal O}({\mathbf P}^{f}_{h},{\mathbf 0},{\mathbf P}^{i}_{h}
{\mathbf P}^{i}_{l};\theta,\phi)
\end{equation}
for the case ${\mathbf P}^{f}_{h}=-{\mathbf P}^{i}_{h}$,
i.e.\
\begin{equation}
\frac{d\sigma^{sf}_{{\mathbf P}^{i}_{h},{\mathbf P}^{i}_{l}}(\theta,\phi)}{d\Omega}=
\frac{d\sigma^{triple}_{-{\mathbf P}^{i}_{h},{\mathbf P}^{i}_{h},{\mathbf P}^{i}_{l}}(\theta,\phi)}{d\Omega}=
{\cal O}(-{\mathbf P}^{i}_{h},{\mathbf 0},{\mathbf P}^{i}_{h}
{\mathbf P}^{i}_{l};\theta,\phi)\,.
\end{equation}
This is the relevant quantity for the method of polarizing hadrons by
electromagnetic scattering on polarized leptons in a storage ring~\cite{WaA07,MiS08}.
\end{description}
\subsection{The nonrelativistic $T$-matrix}
For the explicit evaluation of the trace in eq.~(\ref{basic_xs}) one needs to know
the spin dependence of the $T$-matrix. In a nonrelativistic approach but
including contributions of the order $M^{-2}$, the
$T$-matrix contains the Coulomb, the lepton and hadron spin-orbit and
the lepton-hadron hyperfine interactions. Separating the various
contributions, the $T$-matrix is given in an obvious notation by
\begin{eqnarray}
\widehat T&=&\widehat T_C+\widehat T_{LS_l}+\widehat T_{LS_h}+\widehat
T_{SS}\,.\label{tmatrix}
\end{eqnarray}
In detail, one has the Coulomb contribution
\begin{eqnarray}
\widehat T_C=4\pi \alpha a_C\,,\label{tmatrix_C}
\end{eqnarray}
the spin-orbit interactions of lepton and hadron, respectively,
\begin{eqnarray}
\widehat T_{LS_{l/h}}&=&4\pi \alpha \mathbf b_{l/h}\cdot\boldsymbol\sigma^{l/h}\,,\label{tmatrix_so}
\end{eqnarray}
and the hyperfine interaction
\begin{eqnarray}
\widehat T_{SS}&=&4\pi \alpha
\boldsymbol\sigma^h\cdot\stackrel{\leftrightarrow}{d}\cdot\boldsymbol\sigma^l\,,
\label{tmatrix_SS}
\end{eqnarray}
where $\stackrel{\leftrightarrow}{d}$ denotes a symmetric rank-two tensor,
$\mathbf q=\mathbf p^{\,\prime}-\mathbf p$ the three-momentum transfer, and
$\alpha$ denotes
the Sommerfeld fine structure constant. The tensor
$\stackrel{\leftrightarrow}{d}$ can be decomposed
into a scalar and a spherical tensor of rank two, i.e., a symmetric
cartesian tensor with vanishing trace,
\begin{equation}
{\stackrel{\leftrightarrow}{d}}=d^{[0]}+d^{[2]}\,,\label{d-separation}
\end{equation}
where
\begin{eqnarray}
d_{ij}^{[0]}&=&d_0\delta_{ij},\quad\mbox{ and}\quad d_0=\frac{1}{3}
\mbox{Trace}({\stackrel{\leftrightarrow}{d}})\,,\\
d_{ij}^{[2]}&=&d_{ij}-d_0\delta_{ij}\,.
\end{eqnarray}
Furthermore, the parameters $a_C$, $\mathbf b_{l/h}$, and
${\stackrel{\leftrightarrow}{d}}$ depend on what kind of approximation is used.
These are:
\begin{description}
\item[(i)]
Plane wave approximation (PW), corresponding to a pure one-photon exchange; the
nonrelativistic reduction of the $T$-matrix including lowest order
relativistic contribution reads
\begin{eqnarray}
\widehat T^{PW}&=&\frac{4\pi\alpha }{q^{2}}
\Big\{Z_lZ_h\Big(1 +\frac{{\mathbf P}^2}{4M_lM_h}\Big)
-\frac{1}{8} \Big( Z_h\frac{2\mu_l-1}{M_l^2}
+Z_l\frac{2\mu_h-1}{8M_h^2}\Big)q^{2}
\nonumber\\&&
-\frac{Z_h}{8M_l}\Big(\frac{2\mu_l-1}{M_l}+\frac{2\mu_l}{M_h}\Big)
i(\boldsymbol\sigma_l\times\mathbf q\,)\cdot\mathbf P
-\frac{Z_l}{8M_h}\Big(\frac{2\mu_h-1}{M_h}+\frac{2\mu_h}{M_l}\Big)
i(\boldsymbol\sigma_h\times\mathbf q\,)\cdot\mathbf P
\nonumber\\ &&
+\frac{\mu_l\mu_h}{4M_lM_h}(\boldsymbol\sigma_l\cdot\mathbf q
\,\boldsymbol\sigma_h\cdot\mathbf q
-q^2\,\boldsymbol\sigma_l\cdot\boldsymbol\sigma_h
\,)\Big\}\,,
\end{eqnarray}
with $\mathbf P=\mathbf p+\mathbf p^{\,\prime}$, $Z_l$ and $Z_h$ as the lepton and hadron
charges, and $\mu_l$ and $\mu_h$ as their magnetic moments,
respectively. From this expression one reads off the
parameters, keeping in the spin independent term the lowest order
only,
\begin{eqnarray}
a_C^{PW}&=& \frac{Z_lZ_h}{q^2}\,,\\
\mathbf b^{\,PW}_{l/h}&=&i c_{l/h}^{LS}\,\frac{\mathbf p^{\,\prime}\times \mathbf p}{q^2}\,,\label{b-pw}\\
d_{ij}^{PW}&=&c^{SS}(\widehat q_i\widehat q_j -\delta_{ij})\,,
\end{eqnarray}
where $\widehat{q}$ denotes the unit vector along the three-momentum
transfer $\mathbf q$ and $q=|\mathbf q\,|$. The separation into a scalar and a traceless tensor according to
(\ref{d-separation}) reads
\begin{eqnarray}
d_{0}^{PW} &=& -\frac{2}{3}c^{SS}\,,\\
d_{ij}^{[2]\,PW}&=&c^{SS}(\widehat q_i\widehat q_j -\frac{1}{3}\delta_{ij})\,.
\end{eqnarray}
Furthermore, the strength parameters are
\begin{eqnarray}
c_l^{LS}&=&\frac{Z_h}{4M_l}\,\Big(\frac{2\mu_l-1}{M_l}+2\,
\frac{\mu_l}{M_h}\Big)\,,\\
c_h^{LS}&=&\frac{Z_l}{4M_h}\,\Big(\frac{2\mu_h-1}{M_h}+2\,
\frac{\mu_h}{M_l}\Big)\,,\\
c^{SS}&=&\frac{\mu_l \mu_h}{4M_l M_h}\,.
\end{eqnarray}
One should note that the strength parameter of the hadronic
spin-orbit interaction is about three orders of magnitude smaller than
the parameter of the leptonic one, because their ratio is
approximately
\begin{equation}
\frac{c_h^{LS}}{c_l^{LS}}\approx 2\mu_h\frac{M_l}{M_h}\approx 3\cdot 10^{-3}\,.
\end{equation}
\item[(ii)]
Distorted wave approximation (DW) using nonrelativistic Coulomb
scattering wave functions
\begin{equation}
\psi^{C(+)}_{\mathbf p}(\mathbf
r\,)=\sqrt{\frac{\pi\eta_C}{\sinh{\pi\eta_C}}}
\,e^{-\frac{\pi}{2}\eta_C}\, e^{i\mathbf p\cdot\mathbf r}
\,_1F_1(-i\eta_C,1;i(pr-\mathbf p\cdot\mathbf r\,))\label{c-scatt-wave}
\end{equation}
and
\begin{equation}
\psi^{C(-)}_{\mathbf p}(\mathbf r\,)=\Big(\psi^{C(+)}_{-\mathbf p}(\mathbf r\,) \Big)^*\,,
\end{equation}
where
$\psi^{C(\pm)}_{\mathbf p}$ denotes
the incoming and outgoing scattering waves~\cite{Mes69}, respectively.
Here, $_1F_1(a,b;z)$ denotes the confluent hypergeometric
function. In the expression for the scattering wave
in eq.~(\ref{c-scatt-wave}) I have already separated the constant Coulomb
phase factor $e^{i\sigma_C}$ with
\begin{eqnarray}
\sigma_C&=&\arg[\Gamma(1+i\eta_C)]\,,\label{coulomb-phase}
\end{eqnarray}
because it will disappear in the observables. The relevant quantity
for Coulomb effects is the Sommerfeld Coulomb parameter
\begin{equation}
\eta_C=\alpha Z_l Z_h/v
\end{equation}
with $v$ denoting the relative
hadron-lepton velocity.
Within this approach one finds
\begin{eqnarray}
a_C^{DW}&=& {e^{i\phi_C}}a_C^{PW}\,,\,\mbox{ with }\,
\phi_C(\theta)=-\eta_C \ln[\sin^2(\theta/2)]\,,\\
\mathbf b_{l/h}^{DW}&=&i\frac{c^{LS}_{l/h}}{4\pi}\int \frac{d^3r}{r^3}
\psi^{C(-)}_{\mathbf p^{\,\prime}}(\mathbf r\,)^*
\,(\mathbf r\times \mathbf \nabla)\,
\psi^{C(+)}_{\mathbf p}(\mathbf r\,) \,,\label{ls_dwba}\\
d^{DW}_{ij}&=&-\frac{c^{SS}}{4\pi}\int d^3r
\psi^{C(-)}_{\mathbf p^{\,\prime}}(\mathbf r\,)^*
\,\Big[\frac{1}{r^3}(3\hat{r}_i\,\hat{r}_j-\delta_{ij})+\frac{8\pi}{3}\delta_{ij}\delta(\mathbf
r\,) \Big]\,\psi^{C(+)}_{\mathbf p}(\mathbf r\,)\,.
\label{t_dwba}
\end{eqnarray}
Separating again the hyperfine contribution into a scalar and a
traceless tensor,
one obtains
\begin{eqnarray}
d^{DW}_{0}&=&-\frac{2}{3}\,c^{SS}N(\eta_C)^2 \,,\\
d^{[2]\,DW}_{ij}&=&\frac{c^{SS}}{4\pi}\int \frac{d^3r}{r^3}
\psi^{C(-)}_{\mathbf p^{\,\prime}}(\mathbf r\,)^*
\,(3\hat{r}_i\,\hat{r}_j-\delta_{ij})\,\psi^{C(+)}_{\mathbf p}(\mathbf r\,)\,.
\label{t_dwba_2}
\end{eqnarray}
One should note that the tensor
$d_{ij}^{DW}$ in eq.~(\ref{t_dwba}) is symmetric as well as
$d^{[2]\,DW}_{ij}$.
\end{description}
\subsection{The general scattering cross section and polarization
observables}
Evaluation of the trace in eq.~(\ref{basic_xs}) yields the following general
expression
\begin{eqnarray}\label{quadruplexs}
{\cal O}({\mathbf P}^{f}_{h},{\mathbf P}^{f}_{l},{\mathbf P}^{i}_{h}
{\mathbf P}^{i}_{l};\theta,\phi)
&=&\sum_{\alpha,\beta\in\{C,LS_l,LS_h,SS\}}
S_{\alpha,\beta}(\theta,\phi)\,,
\end{eqnarray}
where the various contributions are defined by
\begin{equation}
S_{\alpha,\beta}(\theta,\phi)=\frac{M_l^2M_h^2}
{\pi^2W^2(1+|{\mathbf P}^{f}_{l}|)(1+|{\mathbf P}^{f}_{h}|)}
\mbox{Trace}\Big[\widehat T^\dagger_\alpha
\rho_h^f({\mathbf P}^{f}_{h})\rho_l^f({\mathbf P}^{f}_{l})
\widehat T_\beta\rho_h^i({\mathbf P}^{i}_{h})\rho_l^i({\mathbf P}^{i}_{l})\Big]\,
\end{equation}
with $\widehat T_\alpha$ defined in eqs.~(\ref{tmatrix_C})-(\ref{tmatrix_SS}). One
should note the relation
\begin{equation}
S_{\alpha,\beta}=S_{\beta,\alpha}^*\,,
\end{equation}
from which follows that $S_\alpha:=S_{\alpha,\alpha}$ is
real.
Separating the diagonal contributions ($S_\alpha$) from the
interference terms ($S_{\alpha,\beta}$ for $\alpha\neq\beta$), one
obtains for the ``quadruple polarization'' cross section
\begin{eqnarray}
\frac{d\sigma_{{\mathbf P}^{f}_{h},{\mathbf P}^{f}_{l},{\mathbf P}^{i}_{h},
{\mathbf P}^{i}_{l}}(\theta,\phi)}{d\Omega}
&=&\sum_{\alpha\in\{C,LS_l,LS_h,SS\}}
S_{\alpha}(\theta,\phi)+
\sum_{\alpha<\beta\in\{C,LS_l,LS_h,SS\}}2\,\,Re\,
S_{\alpha,\beta} (\theta,\phi)\,.
\end{eqnarray}
Explicitly, one finds in terms of the different contributions to the
$T$-matrix in eq.~(\ref{tmatrix}) for the diagonal terms
\begin{eqnarray}
S_{C}(\theta,\phi)&=&V_0\,|a_C|^2\Pi_h^+\Pi_l^+\,,\label{S_C}\\
S_{LS_l}(\theta,\phi)&=&V_0\,\Pi_h^+\Big(\Pi_l^-\,\mathbf b^*_{l}\cdot \mathbf b_{l}
+2\,Re \Big[(\mathbf b_{l}\cdot\mathbf P^f_l)^*
(\mathbf b_{l}\cdot \mathbf P^i_l) \Big]\Big)\,,\label{S_LSl}\\
S_{LS_h}(\theta,\phi)&=&S_{LS_l}(\theta,\phi) |_{h\leftrightarrow l}\,,\label{S_LSh}\\
S_{SS}(\theta,\phi)&=&V_0\,\Big(\Pi_h^-\Pi_l^-D_0
-\mathbf P_{h}^-\cdot\stackrel{\leftrightarrow}{G} \cdot \mathbf P_{l}^-
+\Big[ \Pi_h^-\mathbf P^f_l\cdot{\stackrel{\leftrightarrow}{D}} \cdot\mathbf
P^i_l +(l\leftrightarrow h) \Big]\nonumber\\&&
-\,Im \Big[ \Big(\Pi_h^- (\mathbf P^{-}_l\cdot \mathbf H)
+2\sum_{jst} P_{l,j}^- P_{h,s}^{f} P_{h,t}^{i}\,E_{jst}\Big)
+(l\leftrightarrow h) \Big]\nonumber\\
&&
+2\,Re \Big[(\mathbf P_{l}^f\cdot{\stackrel{\leftrightarrow}{d}}^*\cdot\mathbf P^f_{h})\,
(\mathbf P_{l}^i\cdot{\stackrel{\leftrightarrow}{d}}\cdot\mathbf P^i_{h})
+(\mathbf P_{l}^i\cdot{\stackrel{\leftrightarrow}{d}}^* \cdot\mathbf P^f_{h})\,
(\mathbf P_{h}^i\cdot{\stackrel{\leftrightarrow}{d}}\cdot\mathbf P^f_{l}) \Big]
\Big)\,,\label{S_SS}
\end{eqnarray}
where
\begin{equation}
V_0=\frac{4\alpha^2M_l^2M_h^2}{W^2(1+|{\mathbf P}^{f}_{l}|)(1+|{\mathbf P}^{f}_{h}|)}\,. \label{vorfactor}
\end{equation}
The following quantities depend on the polarization parameters
\begin{eqnarray}
\Pi_{h/l}^{\pm}&=&1\pm\mathbf P^f_{h/l}\cdot \mathbf P^i_{h/l}\,,\label{pol-op1}\\
\mathbf P^{\pm}_{h/l}&=&\mathbf P^i_{h/l}\pm \mathbf P^f_{h/l}\,,\\
{\mathbf Q}_{h/l}&=&\mathbf P^{+}_{h/l}-i\mathbf P^f_{h/l}\times \mathbf P^i_{h/l}\label{pol-op3}\,.
\end{eqnarray}
Furthermore, I have introduced for convenience the following
quantities, which depend on the hyperfine interaction tensor $d_{ij}$,
\begin{eqnarray}
D_{ij}&=&2\,Re\Big(\sum_k d_{ik}^*
d_{kj}\Big)\,,\\
D_0&=&\sum_{ij} d_{ij}^*d_{ji}=
\frac{1}{2}\mbox{Trace}(\stackrel{\leftrightarrow}{D})\,,\\
E_{jst}&=&\sum_{kl}\varepsilon_{jkl}\, d_{ks}^*d_{lt}\,,\\
G_{ij}&=&
\sum_{lmst}\varepsilon_{ils}\,\varepsilon_{jmt}\,d_{lm}^*d_{st}\,,\\
H_{i}&=&
\sum_{klm}\varepsilon_{ikl}\,d_{km}^*d_{ml}\,,
\end{eqnarray}
where $\varepsilon_{ikl}$ denotes the totally antisymmetric
Levi-Civita tensor in three dimensions.
These quantities are functions which depend on the scattering angles
$(\theta,\phi)$.
Correspondingly, one finds for the interference terms
\begin{eqnarray}
S_{C,LS_l}(\theta,\phi)&=&V_0\,a_C^* \Pi_h^+\,\mathbf b_{l}\cdot\mathbf Q_l\,,\\
S_{C,LS_h}(\theta,\phi)&=&S_{C,LS_l}(\theta,\phi)|_{h\leftrightarrow l}\,,\\
S_{C,SS}(\theta,\phi)&=&V_0\,a_C^*\,\mathbf Q_h\cdot{\stackrel{\leftrightarrow}{d}} \cdot \mathbf Q_l\,,\\
S_{LS_l,LS_h}(\theta,\phi)&=&V_0\,(\mathbf b_{l}\cdot\mathbf Q_l) ^*\,(\mathbf b_{h}\cdot\mathbf Q_h)\,,\\
S_{LS_l,SS}(\theta,\phi)&=&V_0\,\mathbf Q_h\cdot{\stackrel{\leftrightarrow}{d}} \cdot\Big(\Pi_l^-\,\mathbf b_l^* -i\,\mathbf
b_l^*\times\mathbf P^{-}_l
+ (\mathbf b_{l}^*\cdot\mathbf P^i_l)\,\mathbf P^f_l+(\mathbf b_{l}^*\cdot\mathbf P^f_l)\,\mathbf P^i_l\Big)\,.\label{S_LS_l_SS}\\
S_{LS_h,SS}(\theta,\phi)&=&S_{LS_l,SS}(\theta,\phi)|_{h\leftrightarrow l}\,.
\end{eqnarray}
Using the separation of the tensor
$\stackrel{\leftrightarrow}{d}$
into a scalar and a traceless symmetric tensor according to
eq.~(\ref{d-separation}), one finds
\begin{eqnarray}
D_0&=&3|d_0|^2+\sum_ {i,k} d^{[2]*}_{ik} d^{[2]}_{ki}\,,\\
D_{ij}&=&2|d_0|^2\delta_{ij}+4\,Re(d_0^*d^{[2]}_{ij})+D_{ij}^{(2)}\,,\\
D_{ij}^{(2)} &=&2\,Re \sum_ k d^{[2]*}_{ik} d^{[2]}_{kj}
\,,\\
E_{jst}&=&\varepsilon_{jst}\,|d_0|^2+\sum_{kl}\varepsilon_{jkl}\,d_{ks}^{[2]\,*}d_{lt}^{[2]}
+\Big(\sum_{l}\varepsilon_{jsl}\,d_0^*
d^{[2]}_{lt}-(s\leftrightarrow t)^*\Big)
\,,\\
G_{ij}&=&2|d_0|^2\delta_{ij}-2\,Re
(d_0^*d^{[2]}_{ij})+G_{ij}^{(2)}\,,\\
H_{i}&=&
\sum_{klm}\varepsilon_{ikl}\,d^{[2]*}_{km} d^{[2]}_{ml}\,,\end{eqnarray}
where
\begin{eqnarray}
G_{ij}^{(2)}&=&
\sum_{lmst}\varepsilon_{ils}\varepsilon_{jmt}d_{lm}^{[2]*}d_{st}^{[2]}\,.
\end{eqnarray}
It is now easy to see that the vector $\mathbf H$ is purely imaginary and
that the tensor $G_{ij}$ is real and symmetric. Furthermore, one notes the
symmetry property
\begin{equation}
E_{jst}^*=-E_{jts}\,.
\end{equation}
It suffices to evaluate the spin-orbit vector $\mathbf b$ and the hyperfine
tensor $\stackrel{\leftrightarrow}{d}$ for $\phi=0$, because
then the values for an arbitrary $\phi$ can be generated by a rotation
around the $z$-axis exploiting their rotation properties. Examples
are given in the following section.
\section{The triple polarization cross section}
I will now specialize to the case where only the final lepton polarization
is not analyzed, i.e.\ $\mathbf P^f_l=0$, but all other particles are
completely polarized ($|\mathbf P^{i/f}_h|=1,\, |\mathbf P^{i}_l|=1$).
This case is of particular
interest for the polarization transfer in a storage ring~\cite{WaA07,MiS08}. The
corresponding ``triple polarization'' cross section has the form
\begin{eqnarray}\label{triplexs}
\frac{d\sigma^{triple}_{{\mathbf P}^{f}_{h},{\mathbf P}^{i}_{h}
{\mathbf P}^{i}_{l}}(\theta,\phi)}{d\Omega}&=&
{\cal O}({\mathbf P}^{f}_{h},{\mathbf 0},{\mathbf P}^{i}_{h}
{\mathbf P}^{i}_{l};\theta,\phi)\nonumber\\
&=&
\sum_{\alpha\in\{C,LS_l,LS_h,SS\}}
S_{\alpha}^{triple} (\theta,\phi)+
\sum_{\alpha<\beta\in\{C,LS_l,LS_h,SS\}}
2\,Re \,S_{\alpha,\beta}^{triple} (\theta,\phi)\,.
\end{eqnarray}
In this case, the lepton polarization
quantities in (\ref{pol-op1}) through (\ref{pol-op3}) become
\begin{equation}
\Pi_l^{\pm}=1,\quad \mathbf Q_l=\mathbf P_l^i,\quad \mathbf P_l^{\pm}=\mathbf P_l^i\,,
\end{equation}
and one finds for the diagonal terms
\begin{eqnarray}
S_C ^{triple}(\theta)&=&V\,\Pi_h^+|a_C (\theta)|^2\,,\\
S_{LS_l}^{triple} (\theta,\phi)&=&V\,\Pi_h^+\,\mathbf b^*_{l}\cdot \mathbf b_{l}\,,\\
S_{LS_h}^{triple} (\theta,\phi)&=&V\,\Big(\Pi_h^-\,\mathbf b^*_{h}\cdot \mathbf b_{h}
+2\,Re \Big[(\mathbf b_{h}\cdot\mathbf P^f_h)^*
(\mathbf b_{h}\cdot \mathbf P^i_h) \Big]\Big)\,,\\
S_{SS}^{triple} (\theta,\phi)&=&V\,\Big(\Pi_h^-D_0
-\mathbf P_{h}^-\cdot\stackrel{\leftrightarrow}{G} \cdot \mathbf P_{l}^i +\mathbf P^f_h\cdot{\stackrel{\leftrightarrow}{D}} \cdot\mathbf P^i_h
\nonumber\\&&
-\,Im \Big[\mathbf P^{-}_h\cdot \mathbf H +\Pi_h^-(\mathbf P^{i}_l\cdot \mathbf H)
+2\sum_{jst} P_{l,j}^i P_{h,s}^{f} P_{h,t}^{i}\,E_{jst}
\Big]\Big)\,,
\end{eqnarray}
and for the interference terms
\begin{eqnarray}
S_{C,LS_l}^{triple} (\theta,\phi)&=&V\,a_C^*\Pi_h^+\,\mathbf b_{l}\cdot\mathbf P_l^i\,,\\
S_{C,LS_h}^{triple} (\theta,\phi)&=&V\,a_C^*\,\mathbf b_{h}\cdot\mathbf Q_h\,,\\
S_{C,SS}^{triple}(\theta,\phi)&=&V\,a_C^*\,\mathbf Q_h\cdot{\stackrel{\leftrightarrow}{d}} \cdot \mathbf P_l^i\,,\\
S_{LS_l,LS_h}^{triple} (\theta,\phi)&=&V\,(\mathbf b_{l}\cdot\mathbf P_l^i) ^*\,(\mathbf b_{h}\cdot\mathbf Q_h)\,,\\
S_{LS_l,SS}^{triple} (\theta,\phi)&=&V\,\mathbf Q_h\cdot{\stackrel{\leftrightarrow}{d}} \cdot\Big(\mathbf b_l^* -i\,\mathbf
b_l^*\times\mathbf P^{i}_l \Big) \,,\\
S_{LS_h,SS}^{triple} (\theta,\phi)&=&V\,\mathbf
P_l^i\cdot{\stackrel{\leftrightarrow}{d}} \cdot\Big(\Pi_h^-\,\mathbf b_h^*
-i\,\mathbf b_h^*\times\mathbf P^{-}_h
+ (\mathbf b_{h}^*\cdot\mathbf P^i_h)\,\mathbf P^f_h+(\mathbf b_{h}^*\cdot\mathbf P^f_h)\,\mathbf P^i_h\Big)\,.
\end{eqnarray}
where
\begin{equation}
V=\frac{2\alpha^2M_l^2M_h^2}{W^2}\,. \label{vorfactor_a}
\end{equation}
From now on as a further specialization, I will consider
only polarization along the incoming direction which is chosen as
$z$-axis. Then with
\begin{equation}
\mathbf P_h^{i/f}=\lambda_h^{i/f}\hat z\,,\quad \mathbf P_h^{\pm}=
(\lambda_h^i\pm\lambda_h^f) \hat z\,,\,\,\,\mbox{ and }\,\,\, \Pi_h^{\pm}=1\pm
\lambda_h^i\lambda_h^f\,,
\end{equation}
where $\lambda_h^\pm=\lambda_h^i\pm\lambda_h^f$, one obtains
\begin{eqnarray}
\frac{d\sigma ^{triple}_{\lambda^f_h,\lambda^i_h,\lambda^i_l}(\theta,\phi)}{d\Omega}&=&
(1+\lambda_h^i\lambda_h^f)\Big[S_C(\theta)+S_0(\theta,\phi)+L_0^l(\theta,\phi)\Big]
+(1-\lambda_h^i\lambda_h^f)\Big[L_0^h(\theta,\phi)+\lambda_l^i \Big (S_1(\theta,\phi)+L_1^{h}(\theta,\phi) \Big)\Big]\nonumber\\
&&+\lambda_h^-\Big[S_1(\theta,\phi)+\lambda_l^i \Big (S_2^-(\theta,\phi)+L_2^{h}(\theta,\phi) \Big)\Big]
+\lambda_h^+\Big[\lambda_l^i \Big (S_{2}^+(\theta,\phi) +L_2^{l}(\theta,\phi)\Big)+L_1^{l}(\theta,\phi\Big] \nonumber\\&&
+\lambda_h^i\lambda_h^f\Big[ S_2(\theta,\phi)+\lambda_l^iS_3(\theta,\phi) \Big]
\,.\label{hadron}
\end{eqnarray}
The diagonal
contributions are
\begin{eqnarray}
L_0^{l/h}(\theta,\phi)&=&V|\mathbf b_{l/h} (\theta,\phi)|^2 \,,\label{L0_lh}\\
S_C(\theta)&=&V|a_C|^2\,,\\
S_0(\theta,\phi)&=&V\Big(3|d_0|^2+\sum_j D_{jj}(\theta,\phi) \Big)\,,\\
S_1(\theta,\phi)&=&iV\sum_{jk}\epsilon_{3jk}D_{jk}(\theta,\phi)
=i(D_{12}(\theta,\phi)-D_{21}(\theta,\phi))=-2\,Im(D_{12}(\theta,\phi))\,,\\
S_2^-(\theta,\phi)&=&2V\Big(\,Re\Big(d_0^*d^{[2]}_{33}(\theta,\phi)
-d^{[2]}_{11}(\theta,\phi)^*
d^{[2]}_{22}(\theta,\phi)\Big) -|d_0|^2+|d^{[2]}_{12}(\theta,\phi)|^2\Big)\,,\\
S_2(\theta,\phi)&=&2V\Big(2\,Re
\Big(d_0^*d^{[2]}_{33}(\theta,\phi)\Big)
-2|d_0|^2-D_{11}(\theta,\phi)-D_{22}(\theta,\phi)\Big)\,,\\
S_3(\theta,\phi)&=&-4V\,Im
\Big(d^{[2]}_{31}(\theta,\phi)^*d^{[2]}_{32}(\theta,\phi)\Big)\,,\label{S3}
\end{eqnarray}
and the interference terms
\begin{eqnarray}
L_1^{l/h}(\theta,\phi)&=&2V\,Re\Big(d^{[2]}_{31}(\theta,\phi)\,b^*_{l/h,1}
(\theta,\phi)+d^{[2]}_{32}(\theta,\phi)\,b^*_{l/h,2} (\theta,\phi)\Big) \,,\\
L_2^{l/h}(\theta,\phi)&=&2V\,Im\Big(d^{[2]}_{31}(\theta,\phi)\,b^*_{l/h,2}
(\theta,\phi)-d^{[2]}_{32}(\theta,\phi)\,b^*_{l/h,1} (\theta,\phi)\Big) \,,\\
S_2^+(\theta,\phi)&=&2V\,Re\Big[a_C^*(d_0+d^{[2]}_{33}(\theta,\phi))\Big]\,.\label{S2plus}
\end{eqnarray}
Here I have already used the fact that the spin-orbit vector $\mathbf
b_{l/h}$ is perpendicular to the $z$-axis, i.e.\ its third
component vanishes. Indeed, as shown explicitly in the appendix, it
has the form
\begin{equation}
\mathbf b_{l/h} (\theta,\phi)=i\,b_0^{l/h}(\theta)(-\sin{\phi},
\cos{\phi},0)=
i\,b^{l/h}_0 (\theta)\,\frac{\mathbf p^{\,\prime}\times\mathbf p}{|\mathbf p^{\,\prime}\times\mathbf p\,|}\,.\label{bls}
\end{equation}
However, not all of the contributions in eqs.~(\ref{L0_lh}) through
(\ref{S2plus}) are nonzero. In
fact, I now will show that the diagonal contributions
$S_1$ and $S_3$ and the interference terms $L_1^{l/h}$ vanish
identically. To this end I first will
consider the $\phi$-dependence of the tensors $d^{[2]}(\theta,\phi)$ and
$D(\theta,\phi)$. It suffices to evaluate them for $\phi=0$ from which one
obtains $ d^{[2]} (\theta,\phi)$ and $D(\theta,\phi)$ by a rotation around the
$z$-axis by an angle $\phi$. Introducing the notation
$d^{[2],0}(\theta)= d^{[2]} (\theta,\phi=0)$ and correspondingly
$D^0= D(\theta,0)$, and using furthermore the
fact $ d^{[2],0}_{12}= d^{[2],0}_{21}=0$, one finds
\begin{equation}
d^{[2]} (\theta,\phi)=\left(
\begin{matrix}
d^{[2],0}_{11}(\theta)\cos^2\phi+d^{[2],0}_{22}(\theta)\sin^2\phi &
\frac{1}{2}(d^{[2],0}_{11}(\theta)-d^{[2],0}_{22}(\theta))\sin 2\phi &
d^{[2],0}_{13}(\theta)\cos\phi\cr
& & \cr
\frac{1}{2}(d^{[2],0}_{11}(\theta)-d^{[2],0}_{22}(\theta))\sin 2\phi &
d^{[2],0}_{11}(\theta)\sin^2\phi
+d^{[2],0}_{22}(\theta)\cos^2\phi & d^{[2],0}_{13}(\theta)\sin\phi\cr
& & \cr
d^{[2],0}_{13}(\theta)\cos\phi & d^{[2],0}_{13}(\theta)\sin\phi & d^{[2],0}_{33}(\theta)\cr
\end{matrix}\right).\label{d2phi}
\end{equation}
From this representation one notes immediately that
\begin{equation}
d^{[2]}_{31}(\theta,\phi)^* d^{[2]}_{32}(\theta,\phi)= |d^{[2],0}_{13}(\theta)|\cos\phi\sin\phi
\end{equation}
is real and thus $S_3(\theta,\phi)=0$ according to eq.~(\ref{S3}).
Furthermore, with
\begin{equation}
D^0(\theta)=\left(
\begin{matrix}
|d^{[2],0}_{11}(\theta)|^2+|d^{[2],0}_{13}(\theta)|^2 & 0 &
d^{[2],0}_{11}(\theta)^*d^{[2],0}_{13}(\theta)+d^{[2],0}_{13}(\theta)^*d^{[2],0}_{33}(\theta)\cr
& & \cr
0 & |d^{[2],0}_{22}(\theta)| & 0 \cr
& & \cr
d^{[2],0}_{13}(\theta)^*d^{[2],0}_{11}(\theta)+d^{[2],0}_{33}(\theta)^*d^{[2],0}_{13}(\theta) &
0 & |d^{[2],0}_{33}(\theta)|^2+|d^{[2],0}_{13}(\theta)|^2\cr
\end{matrix}\right),\label{E0}
\end{equation}
which formally has the same structure as $d^{[2],0}(\theta)$, one finds
\begin{equation}
D(\theta,\phi)=\left(
\begin{matrix}
D^0_{11}(\theta)\cos^2\phi+D^0_{22}(\theta)\sin^2\phi &
\frac{1}{2}(D^0_{11}(\theta)-D^0_{22}(\theta))\sin
2\phi & D^0_{13}(\theta)\cos\phi\cr
& & \cr
\frac{1}{2}(D^0_{11}(\theta)-D^0_{22}(\theta))\sin 2\phi &
D^0_{11}(\theta)\sin^2\phi
+D^0_{22}(\theta)\cos^2\phi & D^0_{13}(\theta)\sin\phi\cr
& & \cr
D^0_{13}(\theta)\cos\phi & D^0_{13}(\theta)\sin\phi & D^0_{33}(\theta)\cr
\end{matrix}\right).\label{D0}
\end{equation}
Here again one notes that
\begin{equation}
D_{12}(\theta,\phi)= \frac{1}{2}(D^0_{11}(\theta)-D^0_{22}(\theta))\sin 2\phi
\end{equation}
is real because according to eq.~(\ref{E0}) $D^0_{11}(\theta)$ and
$D^0_{22}(\theta)$ are real and, therefore, $S_1=0$.
For the spin-orbit interaction vector $\mathbf b_{l/h}$ one finds from eq.~(\ref{bls})
$|\mathbf b_{l/h} (\theta,\phi)|^2=|b_0^{l/h}(\theta)|^2$, which means
that $L_0^{l/h}$ is $\phi$-independent. Moreover, using
eqs.~(\ref{bls}) and (\ref{d2phi}) one obtains
\begin{equation}
L_1^{l/h}(\theta,\phi)=4V\,Re\Big(d^{[2],0}_{13}(\theta)\cos\phi (-b_0^{l/h}(\theta)\sin{\phi})
+d^{[2],0}_{13}(\theta)\sin\phi \,b_0^{l/h}(\theta)\cos{\phi}\Big)=0\,.
\end{equation}
Finally, with the help of eqs.~(\ref{d2phi}) through (\ref{D0}) one finds that
the remaining structure functions become $\phi$-independent, which is
easy to understand since all polarizations are assumed to be along the
$z$-axis ruling out any $\phi$-dependence. Thus
also the cross section simplifies to
\begin{eqnarray}
\frac{d\sigma ^{triple}_{\lambda^f_h,\lambda^i_h,\lambda^i_l}(\theta,\phi)}{d\Omega}&=&
(1+\lambda_h^i\lambda_h^f)\Big[S_C(\theta)+S_0(\theta)+L_0^{l} (\theta) \Big]
+(1-\lambda_h^i\lambda_h^f)L_0^{h} (\theta)\nonumber\\
&&+\lambda_l^i\Big[ \lambda_h^- \Big(S_2^-(\theta,\phi)+L_2^{h}(\theta) \Big)
+\lambda_h^+\Big (S_{2}^+(\theta) +L_2^{l}(\theta) \Big)\Big
+\lambda_h^i\lambda_h^fS_2(\theta)\,,\label{hadron1}
\end{eqnarray}
where the structure functions are given by
\begin{eqnarray}
L_0^{l/h}(\theta)&=&V|b_0^{l/h} (\theta)|^2 \,,\\
L_2^{l/h}(\theta) &=& 2V\,Re\Big[d^{[2],0}_{13}(\theta) b_0^{l/h}(\theta)^* \Big] \,,\\
S_0(\theta)&=&V \Big (3|d_0|^2+\sum_{j=1}^3|d^{[2],0}_{jj}(\theta)|^2
+2|d^{[2],0}_{13}(\theta)|^2 \Big)\,,\\
S_2^+(\theta)&=&2V\,Re \Big [a_C^*\Big (d_0+d^{[2],0}_{33}(\theta) \Big) \Big]\,,\\
S_2^-(\theta)&=&2V\Big[\,Re \Big (d_0^*d^{[2],0}_{33}(\theta)-d^{[2],0}_{11}(\theta)^*
d^{[2],0}_{22}(\theta) \Big)-|d_0|^2\Big]\,,\\
S_2(\theta)&=&2V\Big[2\,Re \Big (d_0^*d^{[2],0}_{33}(\theta) \Big)
-2|d_0|^2-|d^{[2],0}_{11}(\theta)|^2
-|d^{[2],0}_{22}(\theta)|^2-|d^{[2],0}_{13}(\theta)|^2
\Big]\,.
\end{eqnarray}
The expression in eq.~(\ref{hadron1}) is an extension of the triple
polarization cross section given
in eq.~(16) of ref.~\cite{WaA07} by including the contributions from
the hadron and lepton spin-orbit interactions. (As a sideremark: The
corresponding cross section for final lepton polarization is obtained
from (\ref{hadron}) by the substitutions $\lambda_h^f\to\lambda_l^f$ and
$\lambda_h^i\leftrightarrow\lambda_l^i$.)
As a special case, one considers the
so-called hadronic spin-flip cross section for complete hadron
polarization, i.e.\
$\lambda_h^i=-\lambda_h^f=\lambda_h=\pm 1$, and the non-spin-flip
cross section with $\lambda_h^i=\lambda_h^f=\lambda_h=\pm 1$.
For the spin-flip cross section one finds
\begin{eqnarray}
\frac{d\sigma^\mathrm{sf}_{-\lambda_h,\lambda_h,\lambda^i_l}(\theta,\phi)}{d\Omega}&=&
2 L_0^{h} (\theta) -S_2(\theta)
+2\lambda_l^i\lambda_h\Big[S_2^-(\theta,\phi)+L_2^{h}(\theta) \Big]
\,.\label{hadron-flip}
\end{eqnarray}
It is governed by the hyperfine terms $S_2$ and $S_2^-$ and
the hadronic spin-orbit interaction via $L^h_0$ and $L^h_2$. On the other hand,
the non-spin-flip cross section is given by
\begin{eqnarray}
\frac{d\sigma^\mathrm{nsf}_{\lambda_h,\lambda_h,\lambda^i_l}(\theta,\phi)}{d\Omega}&=&
2 \Big[S_C(\theta)+S_0(\theta)+L_0^{l} (\theta) \Big] +S_2(\theta)
+2\lambda_l^i\lambda_h\Big[S_2^+(\theta,\phi)+L_2^{l}(\theta) \Big]
\,.\label{hadron-nonflip}
\end{eqnarray}
Its helicity independent part is overwhelmingly dominated by the
Coulomb term $S_C$ with additional tiny contributions
from the hyperfine and the leptonic spin-orbit interactions.
The difference of the non-spin-flip cross section for $\lambda_h=\pm 1$
\begin{eqnarray}
\frac{1}{2}\,\Big(\frac{d\sigma^\mathrm{nsf}_{1,1,\lambda}(\theta,\phi)}{d\Omega}
-\frac{d\sigma^\mathrm{nsf}_{-1,-1,\lambda}(\theta,\phi)}{d\Omega} \Big)
&=&2 \lambda\Big[S_2^+(\theta,\phi)+L_2^{l}(\theta) \Big]\,\label{diff-nonflip}
\end{eqnarray}
has been considered as lepton-hadron polarization transfer in~\cite{HoM94}.
This polarization transfer is dominated by the hyperfine structure
function $S_2^+$ because the additional spin-orbit contribution
$L_2^l$ is comparably small as shown in the next section. It differs
by a factor 2 and the presence of $L_2^l$ from the polarization
transfer $P_{z00z}$ for the scattering of unpolarized hadrons on
polarized leptons as considered in ref.~\cite{Are07}, where I had
considered only the leading term $S_2^+$,
the interference between Coulomb and hyperfine amplitudes, neglecting
higher order contributions. The more complete expression reads
\begin{eqnarray}
P_{z00z}\frac{d\sigma_0(\theta,\phi)}{d\Omega}&=&
\frac{\partial^2}{\partial\lambda^f_h \partial\lambda^i_l}
\frac{d\sigma^{triple}_{\lambda^f_h,\lambda^i_h,\lambda^i_l}(\theta,\phi)}{d\Omega}
\Big|_{\lambda^i_h=0}\nonumber\\
&=&S_2^+(\theta,\phi)-S_2^-(\theta,\phi)+L_2^l(\theta,\phi)-L_2^h(\theta,\phi)\,.
\end{eqnarray}
In addition to $S_2^+$, it includes $S_2^-$, which is quadratic
in the hyperfine amplitude, and $L_2^l$ and $L_2^h$, the contributions from the
interference of hyperfine and leptonic
and hadronic spin-orbit amplitudes,
respectively. However, the largest of these additional terms, $L_2^l$,
is still quite small if not negligible compared to $S_2^+$.
\section{Results for structure functions and polarization cross sections}
For the evaluation of the structure functions in eq.~(\ref{hadron1})
and the corresponding cross section I have used both
methods for the calculation of Coulomb distortion, the integral
representation as well as the partial
wave expansion. The integral representation has been used mainly in
order to check the convergence of the partial wave expansion as is
described in detail in the appendix. Thus all results presented in
this section are based on the partial wave expansion (PWE). For the
hyperfine amplitude it was found that an expansion up to a partial
wave with $l_{max}=2000$ was sufficient, but for the spin-orbit interaction,
beeing much slower convergent, $l_{max}=4000$ was taken.
\subsection{The structure functions}
\begin{figure}
\includegraphics[width=1.\columnwidth]{./structure_functions_S_0.eps}
\caption{Structure function $S_0(\theta)$ of the triple polarization
differential cross section
in plane wave (PW) and distorted wave approximation
for like charges (DW$^+$) and opposite charges (DW$^-$)
for various proton lab kinetic energies $T_h$. }
\label{fig_S_0}
\end{figure}
\begin{figure}
\includegraphics[width=1.\columnwidth]{./structure_functions_L_0_l_rel.eps}
\caption{The structure function $L^l_0(\theta)$ of the triple polarization
differential cross section
in plane wave (PW) and distorted wave approximation
for like charges (DW$^+$) and opposite charges (DW$^-$)
for various proton lab kinetic energies $T_h$. For comparison the Coulomb
structure function $S_C(\theta)$ is shown in addition reduced by a
factor 10$^{-n}$.}
\label{fig_L_0_l}
\includegraphics[width=1.\columnwidth]{./structure_functions_S_2.eps}
\caption{The structure function $S_2(\theta)$ of the triple polarization
differential cross section
in plane wave (PW) and distorted wave approximation
for like charges (DW$^+$) and opposite charges (DW$^-$)
for various proton lab kinetic energies $T_h$. One should note the
enhancement factors for PW and DW$^+$. }
\label{fig_S_2}
\end{figure}
\begin{figure}[h]
\includegraphics[width=1.\columnwidth]{./structure_functions_S_2_minus.eps}
\caption{The structure function $S_2^-(\theta)$ of the triple polarization
differential cross section
in plane wave (PW) and distorted wave approximation
for like charges (DW$^+$) and opposite charges (DW$^-$)
for various proton lab kinetic energies $T_h$. One should note the
enhancement factors for PW and DW$^+$. }
\label{fig_S_2_minus}
\includegraphics[width=1.\columnwidth]{./structure_functions_S_2_plus_rel.eps}
\caption{The structure function $S_2^+(\theta)$ of the triple polarization
differential cross section
scattering multiplied by $\sin^2(\theta/2)$
in plane wave (PW) and distorted wave approximation
for like charges (DW$^+$) and opposite charges (DW$^-$)
for various proton lab kinetic energies $T_h$. One should note the
enhancement factors for PW and DW$^+$. }
\label{fig_S_2_plus}
\end{figure}
\begin{figure}[h]
\includegraphics[width=1.\columnwidth]{./structure_functions_L_2_l_rel.eps}
\caption{The structure function $L_2^l(\theta)$ of the triple polarization
differential cross section
in plane wave (PW) and distorted wave approximation
for like charges (DW$^+$) and opposite charges (DW$^-$)
for various proton lab kinetic energies $T_h$. One should note the
enhancement factors for PW and DW$^+$ in the upper panels. }
\label{fig_L_2_l}
\end{figure}
First, I will discuss the various structure functions
which determine the triple polarization cross section of
eq.~(\ref{hadron1}). For the plane wave approximation, one finds
easily the following expressions
\begin{eqnarray}
L_0^{l/h}(\theta)&=&-\frac{1}{4}\,V(c^{LS}_{l/h})^2\cot^2(\theta/2)\,,\\
L_2^{l/h}(\theta) &=& -V c^{LS}_{l/h}c^{SS}\cos^2(\theta/2)\,,\\
S_0(\theta)&=&2\,V (c^{SS})^2\,,\\
S_2^+(\theta)&=&-\frac{Vc^{SS}}{2p^2}\,\cot^2(\theta/2)\,,\\
S_2^-(\theta)&=&-2\,V (c^{SS})^2\sin^2(\theta/2)\,,\\
S_2(\theta)&=&-2\,V (c^{SS})^2(1+\sin^2(\theta/2))\,.
\end{eqnarray}
The structure functions, evaluated in the c.m.~frame
for several lab kinetic energies, are shown in Figs.~\ref{fig_S_0} through
\ref{fig_L_2_l} for the various approximations, i.e.\ plane wave
approximation (PW) and Coulomb distortion for like (DW$^+$) and
opposite charges (DW$^-$).
The diagonal structure
function $S_0$ in Fig.~\ref{fig_S_0}, induced by the hyperfine
interaction, shows a rather flat, almost constant angular
behaviour. Its size scales roughly proportional to the inverse of the
kinetic energy $T_h$. Compared to the plane wave approximation (PW),
the distorted wave approximation is strongly enhanced for unlike
charges (DW$^-$) and strongly suppressed for like charges (DW$^+$) by
several orders of magnitude
for the lowest lab kinetic energy of $T_h=0.001$~MeV corrsponding to
$\eta_C\approx 5$. This enhancement resp.\ suppression is increasingly reduced
with growing kinetic energy $T_h$ and approaches the plane wave result
above $T_h=10$~MeV. The pure Coulomb
contribution $S_C$, however, is much larger by more than ten orders of
magnitude.
With respect to the other two diagonal structure
functions from the leptonic and hadronic spin-orbit interactions
$L_0^{l}$ and $L_0^{h}$, it suffices to show only the former one in
Fig.~\ref{fig_L_0_l}, because $L_0^{h}$ differs in magnitude only
by the factor $(c_{LS}^h/c_{LS}^l)^2$ being smaller by about five
orders of magnitude. One readily notes that $L_0^{l}$
exhibits a strong peaking in the
forward direction only and tends to oscillate at small angles for the
lowest $T_h$ considered here. Over the whole angular range,
especially in the forward direction, $L_0^{l}$ is much
larger by several orders of magnitude than $S_0$ but it is still
almost negligible compared to the size of $S_C$. The effect of
Coulomb distortion is qualitatively similar to what one observes in $S_0$.
Only these diagonal structure functions contribute to the
unpolarized cross section. However, as already mentioned,
their relative contribution is
extremely small as can be seen by comparison with the pure Coulomb
structure function $S_C$ which is also shown in Fig.~\ref{fig_L_0_l}
indicated by the large reduction factor applied to $S_C$.
The two hyperfine-hyperfine interference structure functions $S_2$ and
$S_2^-$, shown in Figs.~\ref{fig_S_2} and \ref{fig_S_2_minus}, which
both are negative througout, exhibit
a similar pattern, a smooth angular distribution with a slight
decrease in size at backward angles for the two lowest kinetic
energies, but with a slight increase for the two higher kinetic
energies like the PW result for all four cases. Again one notes
sizeable enhancements for opposite charges by Coulomb distortion
and suppression for like charges for the lowest
energies $T_h$. The distortion effect decreases with increasing $T_h$.
Also these two structure functions are quite small like $S_0$ because
they are quadratic in the hyperfine amplitudes.
Much larger
is the third interference structure function $S_2^+$ because it is an
interference between the Coulomb and the hyperfine amplitudes. Thus it
is strongly forward peaked. For this reason, it is displayed in
Fig.~\ref{fig_S_2_plus} multiplied by $\sin^2(\theta/2)$. Moreover,
Coulomb distortion induces a strong oscillatory behavior, again in
conjunction with a large enhancement for opposite charges and strong
suppression for like charges. As expected, the distortion effect
diminishes with increasing kinetic energy $T_h$. I would like
to point out that in the corresponding Figs.~1 and 2 of
ref.~\cite{Are07} erroneously a factor $\sin^2(2\theta)$ instead of
the indicated factor $\sin^2(\theta/2)$ has been applied.
Finally, the
spin-orbit-hyperfine interference structure function $L_2^l$ in
Fig.~\ref{fig_L_2_l} is comparable in size to $S_2$ and
$S_2^-$ but exhibits quite a different pattern. At the lowest
kinetic energy $T_h$ it is strongly enhanced by distortion for
opposite charges and
possesses a pronounced broad minimum around 100$^\circ$. It falls off
at forward and backward angles with many oscillations in the forward
direction. With increasing kinetic energy the minimum moves towards
smaller angles with fewer oscillations. Again the distortion effect
decreases strongly with increasing energy. Like the diagonal
structure function $L_0^h$, the hadronic interference structure function
$L_2^h$ is quite negligible.
\begin{figure}[ht]
\includegraphics[width=1.\columnwidth]{./diff-tripel-plus.eps}
\caption{Absolute value of spin-flip cross section $d\sigma_+^{sf}/d\Omega$ for
initial hadron polarization parallel to lepton polarization along
the initial relative momentum
in plane wave (PW) and distorted wave approximation
for like charges (DW$^+$) and opposite charges (DW$^-$)
for various proton lab kinetic energies $T_h$.}
\label{fig_spin_flip_plus}
\includegraphics[width=1.\columnwidth]{./diff-tripel-minus.eps}
\caption{Absolute value of spin-flip cross section $d\sigma_-^{sf}/d\Omega$ for
initial hadron polarization opposite to lepton polarization along
the initial relative momentum
in plane wave (PW) and distorted wave approximation
for like charges (DW$^+$) and opposite charges (DW$^-$)
for various proton lab kinetic energies $T_h$.}
\label{fig_spin_flip_minus}
\end{figure}
\begin{figure}[ht]
\includegraphics[width=1.\columnwidth]{./diff-tripel-plus_ss_log.eps}
\caption{Comparison of spin-flip cross section $d\sigma_+^{sf}/d\Omega$ for
initial hadron polarization parallel to lepton polarization along
the initial relative momentum with the hyperfine interaction alone
(DW(hfs)) and with inclusion of the spin-orbit contribution (DW)
for like charges (DW$^+$, left panels) and opposite charges (DW$^-$,
right panels) for the two lowest proton lab kinetic energies $T_h$.}
\label{fig_spin_flip_plus_ss}
\includegraphics[width=1.\columnwidth]{./diff-tripel-minus_ss_log.eps}
\caption{Comparison of spin-flip cross section $d\sigma_+^{sf}/d\Omega$ for
initial hadron polarization opposite to lepton polarization along
the initial relative momentum with the hyperfine interaction alone
(DW(hfs)) and with inclusion the of spin-orbit contribution (DW)
for like charges (DW$^+$, left panels) and opposite charges (DW$^-$,
right panels) for the two lowest proton lab kinetic energies $T_h$.}
\label{fig_spin_flip_minus_ss}
\end{figure}
\subsection{The triple polarization cross section}
Now I will discuss the triple polarization cross section of
eq.~(\ref{hadron1}). Previously, in ref.~\cite{WaA07} only the
hyperfine amplitude besides the Coulomb one has been considered,
whereas the hadron spin-orbit interaction has
been included already in ref.~\cite{MiS08}. However, as
mentioned above, its contribution to the helicity dependent part of
the spin-flip cross section in eq.~(\ref{hadron-flip}) is
negligible, whereas in the helicity independent part the diagonal contribution $L_0^h$ is
comparable in size to $S_2$ in the forward direction (compare
Figs.~\ref{fig_L_0_l} and \ref{fig_S_2}). Much more important
is the leptonic spin-orbit contribution which, however, appears in the
non-spin-flip cross section only (see eq.~(\ref{hadron-nonflip})) where
it is buried completely by the Coulomb contribution $S_C$.
The results for the spin-flip cross section for parallel initial spin orientations of
hadron and lepton is shown in Fig.~\ref{fig_spin_flip_plus} while the
one for the opposite spin orientation in
Fig.~\ref{fig_spin_flip_minus}. One notes again the strong influence
of Coulomb distortion. Furthermore, the leptonic spin-orbit
interaction plays a relatively important role in the region of the
minimum as can be seen in
Figs.~\ref{fig_spin_flip_plus_ss} and \ref{fig_spin_flip_minus_ss}
where a comparison is exhibited with the case for which the spin-orbit
interaction is switched off (curves labeled ``(hfs)''). One readily notes a substanial increase
when the spin-orbit part is included compared to the pure hyperfine
case. Furthermore, at the lowest two energies
the spin-orbit interaction induces oscillations, in particular in the
forward direction. The difference of the two spin-flip cross sections
determines the net hadron polarization in a storage ring of initially
unpolarized hadrons scattered at polarized leptons.
The non-spin-flip cross section in eq.~(\ref{hadron-nonflip}) is
overwhelmingly dominated by the Coulomb contribution $S_C$.
The small dependence on $\lambda_l^i$ and $\lambda_h$ leads to
different scattering strengths for the hadron polarization parallel or
antiparallel to the lepton polarization (see
eq.~(\ref{diff-nonflip})). The resulting lepton-hadron polarization
transfer $P_{z00z}$ is shown in Fig.~\ref{fig_pzz_plus_minus}. The
dominance of the hyperfine amplitude is clearly seen.
\begin{figure}[ht]
\includegraphics[width=1.\columnwidth]{./diff-pzz-plus-minus-log.eps}
\caption{Lepton-hadron polarization transfer cross section $P_{z00z} d\sigma_0/d\Omega$
with inclusion of spin-orbit contribution
for like charges (DW$^+$) and opposite charges (DW$^-$) for several
proton lab kinetic energies $T_h$. }
\label{fig_pzz_plus_minus}
\includegraphics[width=1.\columnwidth]{./int_structure_functions_S2plus.eps}
\caption{The integrated structure functions $\langle S_0 \rangle$ (left panel)
and $\langle S_2^+ \rangle$ (right panel) as function of the proton lab kinetic
energy $T_h$ for plane wave approximation (PW) and with Coulomb
distortion for like charges (DW$^+$) and opposite charges (DW$^-$).}
\label{Fig_int_s2plus}
\includegraphics[width=1.\columnwidth]{./int_structure_functions_S_minus.eps}
\caption{The integrated structure functions $\langle S_2^- \rangle$ (left panel)
and $\langle S_2 \rangle$ (right panel) as function of the proton lab kinetic
energy $T_h$ for plane wave approximation (PW) and with Coulomb
distortion for like charges (DW$^+$) and opposite charges (DW$^-$).}
\label{Fig_int_sminus}
\end{figure}
\begin{figure}[ht]
\includegraphics[width=1.\columnwidth]{./int_structure_functions_L.eps}
\caption{The integrated structure functions $\langle L_0^l \rangle$ (left panel)
and $\langle L_2^l \rangle$ (right panel) as function of the proton lab kinetic
energy $T_h$ for plane wave approximation (PW) and with Coulomb
distortion for like charges (DW$^+$) and opposite charges (DW$^-$).}
\label{Fig_int_L}
\includegraphics[width=1.\columnwidth]{./int_spin_flip_log_neu.eps}
\caption{The integrated spin-flip cross section $\langle
{\sigma^{sf}_{+}} \rangle$ (left panels)
and $\langle {\sigma^{sf}_{-}} \rangle$
(right panels) as function of the proton lab kinetic
energy $T_h$ for plane wave approximation (PW) and with Coulomb
distortion for like charges (DW$^+$, lower panels) and opposite charges
(DW$^-$, upper panels). Curves labeled ``hfs'' include the hyperfine
amplitude only, and in the upper panels the curves labeled
``MSS'' represent the results of ref.~\cite{MiS08}.}
\label{Fig_int_spin-flip_plus}
\end{figure}
\begin{figure}[ht]
\includegraphics[width=1.\columnwidth]{./int_spin_flip_b_dependence.eps}
\caption{Dependence of the integrated spin-flip cross sections for
opposite charges on the regularization parameter $b$.
}
\label{Fig_int_spin-flip-b-dep}
\end{figure}
\subsection{The integrated structure functions and cross sections}
Finally, I will present results for the integrated
structure functions and spin-flip cross sections which are the
relevant quantities for the polarization build-up in a storage
ring. They are defined by the integration over the solid angle except
for the small cone in the forward direction with $\theta < \theta_{min}$,
where the minimal scattering angle is defined by the requirement that
the impact parameter should not exceed a given value b
\begin{equation}
\theta_{min}=2\arctan(\eta_C/l)\,,
\end{equation}
with $l=bp$ as classical angular momentum. In the present work I have
chosen $b=10^{10}$. The choice of this value has been justified in
ref.~\cite{WaA07}. The dependence on this parameter is discussed
below. Thus for any structure
function or cross section ${\cal O}(\theta)$ I define
\begin{equation}
\langle {\cal O} \rangle= 2\pi\int_{\theta_{min}}^\pi d(\cos \theta)
{\cal O}(\theta)\,.
\end{equation}
For the plane wave approximation, one finds
easily the following expressions
\begin{eqnarray}
\langle L_0^{l/h}(\theta) \rangle&=&-2\pi\,V(c^{LS}_{l/h})^2
\Big(\ln(\sin(\theta_{min}/2))+\frac{1}{2}\cos^2(\theta_{min}/2)\Big) \,,\\
\langle L_2^{l/h}(\theta) \rangle&=& -2\pi\,V c^{LS}_{l/h}c^{SS}
\Big(1-\sin^2(\theta_{min}/2)(1+\cos^2(\theta_{min}/2))\Big)
\,,\\
\langle S_0(\theta) \rangle&=&8\pi\,V (c^{SS})^2\,,\\
\langle S_2^+(\theta) \rangle&=&4\pi\,V \frac{c^{SS}}{p^2}
\Big(\ln(\sin(\theta_{min}/2))-\frac{1}{2}\cos^2(\theta_{min}/2)\Big)
\,,\\
\langle S_2^-(\theta) \rangle&=&-4\pi\,V (c^{SS})^2
\cos^2(\theta_{min}/2)(1+\sin^2(\theta_{min}/2))
\,,\\
\langle S_2(\theta) \rangle&=&-4\pi\,V (c^{SS})^2(3-\sin^4(\theta_{min}/2))\,.
\end{eqnarray}
Thus in plane wave all of the integrated structure functions except
for $\langle S_2^+ \rangle$ are almost independent of $T_h$ except for a very weak
dependence via the minimal scattering angle. One should note the
logarithmic divergence for $\theta_{min} \to 0$ in $\langle
L_0^{l/h} \rangle$ and $\langle S_2^+ \rangle$. It
corresponds to the logarithmic divergence in the angular momentum $l$
of the partial wave expansion noted in
ref.~\cite{MiS08} which appears when integrating over the whole
range of scattering angles.
The results are exhibited in Figs.~\ref{Fig_int_s2plus} through
\ref{Fig_int_L} for the structure functions and in
Fig.~\ref{Fig_int_spin-flip_plus} for the spin-flip cross
sections. The integrated structure functions show a strong increasing
influence of Coulomb distortion with decreasing hadron kinetic energy
$T_h$ leading to large enhancements for opposite charges and strong
suppression for like charges compared to the plane wave case. The only
exception is $\langle S_2^+ \rangle$ for which Coulomb distortion results in a
reduction for both cases, however much stronger for like charges. The
reason for this feature is the strong oscillatory behavior in the
angular distribution of $\langle S_2^+ \rangle$ at low $T_h$ (see
Fig.~\ref{fig_S_2_plus} ).
The integrated spin-flip cross section is given by
\begin{equation}
\langle
{\sigma^{sf}_{\pm}} \rangle=
2 \langle L_0^{h} \rangle -\langle S_2 \rangle
+2\pm\Big[\langle S_2^-\rangle+\langle L_2^{h} \rangle \Big]
\,.\label{int-hadron-flip}
\end{equation}
The corresponding integrated spin flip cross section of Milstein et
al.~\cite{MiS08} reads according to their eq.~(21)
\begin{equation}
\langle
{\sigma^{sf}_{\pm}} \rangle_{MSS}=\pi\Big(\frac{\alpha\mu_p}{M_p}\Big)^2
\Big[(2\pi\eta_C)^2(\frac{11}{6}-\ln 2)+\ln(l_{max}/\eta_C)^2\mp (2\pi\eta_C)^2 \Big]\,.
\end{equation}
The appearing logarithmic divergence in the angular momentum $l$
is regularized by choosing a finite
$l_{max}$ determined by the classical relation $l_{max}=bp$ which
corresponds to the choice of a minimum scattering angle in the present
work. The integrated strength of the spin-flip cross sections in
Fig.~\ref{Fig_int_spin-flip_plus}
shows as expected with decreasing $T_h$ a growing strong influence
of Coulomb effects via the hyperfine and hadronic spin-orbit
interactions. The latter is only important in the spin-independent part
of the spin-flip cross section while its influence in the
spin-dependent part is negligible. The results of Milstein et
al.~\cite{MiS08}, shown in the upper
panels of Fig.~\ref{Fig_int_spin-flip_plus} for opposite charges, are
comparable to our results but display a slight overestimation which is
probably due to different approximations in ~\cite{MiS08}. The
dependence of the integrated spin-flip cross
section for opposite charges on the regularization parameter $b$
is exhibited in Fig.~\ref{Fig_int_spin-flip-b-dep} for $b=10^9$, $10^{10}$ and
$10^{11}$~fm. It appears to be quite weak.
The relevant quantity for a
polarization build-up in a storage ring is the ratio of the
spin-independent part over the spin-dependent part
\begin{equation}
R^{sf}=2\frac{\langle S_2^-\rangle+\langle L_2^{h}\rangle}
{2 \langle L_0^{h} \rangle -\langle S_2 \rangle }\,,
\end{equation}
which is shown in Fig.~\ref{Fig_int_spin-flip_ratio} . One readily
notes the reduction of this ratio by the hadronic spin-orbit
interaction, in particular quite strong at higher energies but only
about 12~\% at the lowest energy. This fact clearly shows the
importance of the hadronic spin-orbit interaction besides the
hyperfine contribution.
Finally, I show for completeness in Fig.~\ref{Fig_int_Pzz} the
integrated polarization transfer cross section $\langle
P_{z00z} \sigma_0 \rangle$. It is dominated by the hyperfine
interaction whereas the spin-orbit contribution is negligible.
\begin{figure}[h]
\includegraphics[width=.7\columnwidth]{./int_spin_flip_ratio.eps}
\caption{Ratio of the spin dependent part of the spin-flip cross
section over its spin-independent part as function of the proton lab kinetic
energy $T_h$ for plane wave approximation (PW) and with Coulomb
distortion for opposite charges (DW$^-$): present calculation and
the result of ref.~\cite{MiS08} (MSS). For the curves labeled
(hfs) only the hyperfine amplitude is included. }
\label{Fig_int_spin-flip_ratio}
\end{figure}
\begin{figure}[h]
\includegraphics[width=1.\columnwidth]{./int_Pzz.eps}
\caption{The integrated polarization transfer cross section $\langle
P_{z00z} \sigma_0 \rangle$
as function of the proton lab kinetic
energy $T_h$ for plane wave approximation (PW) and with Coulomb
distortion for like charges (DW$^+$, right panel) and opposite charges
(DW$^-$, left panel). For the curves labeled (hfs) only the hyperfine amplitude is
included. }
\label{Fig_int_Pzz}
\end{figure}
\section{Conclusions }
Formal expressions for polarization observables in electromagnetic
hadron-lepton scattering have been derived within a nonrelativistic
framework including the central Coulomb force as well as the lepton
and hadron spin-orbit and hyperfine interactions. The latter have
been treated in a distorted wave approximation.
Special emphasis has been
laid on the triple polarization cross section with polarizations of
the initial hadron and lepton and of the final
hadron along the incoming hadron momentum. The structure functions
which determine the differential
triple polarization cross section have been evaluated in plane and
distorted wave approximations for hadron lab kinetic energies between
1 keV and 100 MeV.
For the evaluation of the Coulomb distortion two
different methods have been employed: (i) an integral representation
of the nonrelativistic Coulomb scattering wave function and (ii) a
partial wave expansion. These two independent methods have served as a
mutual check for the numerical accuracy of the results.
As expected, the distortion effects are very important at low
energies in the small polarization observables, which are driven by
the spin-orbit and hyperfine interactions,
leading to sizeable enhancements for opposite charges and suppressions
for like charges according to the Coulomb attraction resp.\
repulsion. This is shown in detail for the structure functions of the
triple polarization cross section and for the special case of the
spin-flip differential cross section.
The leptonic spin-orbit interactions plays a
minor role in the non-spin-flip cross section in its spin-dependent
part, which, however, as a whole is by many orders of magnitude
smaller than the spin-independent part, dominated by the Coulomb term
$S_C$. The influence of the spin-orbit and hyperfine interactions
on the unpolarized cross section is almost negligible in the whole
range of energies studied here.
With respect to the integrated spin-flip cross sections our previous
work has been extended by the inclusion of the hadronic spin-orbit
interaction which shows a non-negligible effect in the
spin-independent part changing sizeably the ratio of the integrated
strength of its spin-dependent over the spin-independent part.
\section*{Acknowledgment}
I would like to thank Thomas Walcher for his continued interest in
this work and for many useful discussions. This work has been
supported by the SFB 443 of the Deutsche Forschungsgemeinschaft.
|
\section{Introduction}
Multiuser diversity can significantly improve system throughput when
the users suffer channel fluctuations. The performance gain of
multiuser diversity grows with the number of users when the scheduler
performs maximum throughput scheduling
\cite{Pv_Tse02}. For the broadcast channel, the dirty paper coding (DPC) \cite{M_H_M_Costa1983}
achieves a sum rate which was shown in \cite{J_Cioffi_04} to have a
growth rate $M\log\log(K)$, where $M$ is the number of transmit
antennas at the base station (BS) and $K$ is the number of users in
the system. However, this scheme has high complexity in
encoding/decoding and is difficult to be implemented. Therefore,
a suboptimal and low-complexity zero-forcing (ZF) beamforming
technique was proposed in \cite{TYoo2006} which also achieves the optimal growth
rate of the sum rate. The results both in DPC and ZF schemes were based on the full
channel state information (CSI) assumption at the transmitter and thus
the users are required to feedback perfect CSI to the BS. Although
the optimal sum rate can be achieved, the feedback load will
increase linearly with the number of users.
Various approaches have been proposed to limit the amount of
feedback load and investigate sum rate loss. In
\cite{Jindal06}\cite{Goldsmith07}, the quantized channel direction
information (CDI) was used to characterize the sum rate loss when the
ZF technique is considered. It was shown that the number of feedback bits of each
user needs to be increased linearly with the transmission power to maintain a constant sum rate loss. Another low-feedback-rate and practical scheme,
orthogonal random beamforming (ORB) was proposed in \cite{Hassibi05}.
In ORB, each user only feeds back the CSI and the beam index of its favorite beam to the BS. Therefore, the total amount of feedback
can be reduced from $MK$ CSI values to $K$ CSI values. Besides, the
sum rate loss is negligible when the number of users is
large \cite{M_Pugh10}. In an effort to further reduce feedback load,
a threshold based mechanism was proposed in \cite{Gesbert04} such that a user does not feed back when its CSI is below the threshold. In
that work, the design of
threshold does not take the scheduling algorithm into account. In
\cite{JH2010}, multiple thresholds were proposed. The design of multiple thresholds was based on the
order statistics of the signal-to-interference-noise-ratio (SINR) assuming that
greedy sum rate scheduling was performed in a {\it homogeneous} network where the users' channel gain distributions are the same.
Exhaustive search was used to obtain the thresholds, which
resulted in high computational complexity.
In this work, the closed form solution of the multiple thresholds in
\cite{JH2010} is found, and a more realistic assumption on the
channel distributions of the users is considered. We assume that the
distributions of the users' signal-to-noise rations (SNRs) are independent and
non-identical. Since the computational complexity is too high to
consider all the users' non-identical channel statistics, the
statistics of the users' channels are divided into multiple clusters.
The statistics in each cluster are used to calculate the corresponding threshold.
Each cluster corresponds to an SNR range and is quantized with a few
bits for differentiating the users' SNR levels falling in the same
cluster. In addition, the sum rate loss due to setting thresholds is
investigated. The performance of the proposed scheme is
compared with that of the conventional feedback scheme and single
threshold feedback scheme \cite{Gesbert04} in terms of sum rate,
feedback load and efficiency. Through simulations, it is shown that
the proposed cluster-based feedback scheme is more efficient than
conventional feedback schemes and achieves higher sum rate than the
single threshold feedback scheme.
\section{System Model}
The multiuser multiple-input multiple-out (MIMO) downlink system
is considered. The BS is equipped with $M$ antennas and there are
$K$ users, each having $N$ receive antennas. According to the ORB
strategy for multiuser transmission, the BS uses a precoding matrix
${\mathbf {W}}=[{\mathbf{w}}_1,{\mathbf{w}}_2,\ldots,{\mathbf{w}}_M]$
to simultaneously transmit signals, where ${\mathbf {w}}_{i}
\in{{\mathbb{C}}^{M\times1}}, i=1,2,\ldots,M$, are random orthogonal
vectors generated from isotropic distribution \cite{Marzetta1999}.
Let ${\mathbf{s}}=\sum_{m}{\mathbf{w}}_{m}s_{m}$ be the $M\times1$
vector of the transmitted signal, where $s_{m}$ is the $m$th transmitted
symbol.
The received signal for user $k$ is given by
\begin{eqnarray}
{\mathbf{y}}_k&=&{\mathbf{H}}_{k}{\mathbf{Ws}}+{\mathbf{n}}_k
\end{eqnarray}
where ${\mathbf{H}}_{k}$ denotes the channel matrix between the
BS and user $k$. The elements of the channel matrix ${\mathbf{H}}_{k}$
are assumed to be independent identically distributed (i.i.d.)
complex Gaussian random variables with zero mean and variance $\sigma_{k}^2$. The noise term
for user $k$, denoted ${\mathbf{n}}_{k}$, is modeled as
i.i.d. zero mean complex Gaussian with covariance
matrix
${\mathbb{E}}[{\mathbf{n}}_{k}{\mathbf{n}}_{k}^{H}]=\sigma_{N}^2{\mathbf{I}}, \forall k$, where ${\mathbb{E}}$ denotes the expectation operation and $(\cdot)^{H}$
represents the transpose conjugate.
User $k$ uses the ZF receiver \cite{Heath01} to perform channel
inversion to the received signal ${\mathbf{y}}_{k}$. Thus, the
received signal after ZF receiver is given by
\begin{eqnarray}
{({\mathbf{H}}_{k}{\mathbf{W}})}^{\dagger}{\mathbf{y}}_k={\mathbf{s}}+{({\mathbf{H}}_{k}{\mathbf{W}})}^{\dagger}{\mathbf{n}}_k
\end{eqnarray}
where
${\mathbf{H}}^{\dagger}=({\mathbf{H}}^{H}{\mathbf{H}})^{-1}{\mathbf{H}}^{H}$
is the pseudo-inverse of ${\mathbf{H}}$. Under the equal power
assumption, let the transmit energy of each antenna be
$\frac{P}{M}$, where $P$ is the total transmit power at the BS.
The SNR for the $k$th user at the $m$th spatial channel is denoted
by $X_{m,k}$
\begin{eqnarray}
X_{m,k}=\frac{P/M}{\sigma_{N}^2[(({\mathbf{H}}_{k}{\mathbf{W}})^{H}({\mathbf{H}}_{k}{\mathbf{W}}))^{-1}]_{m}}
& m=1,2,\ldots,M
\end{eqnarray}
where $[{\mathbf{A}}]_{m}$ denotes the $m$th diagonal element of
matrix ${\mathbf{A}}$. Assuming $N\geq M$, it is well known
that $X_{m,k}$ is a chi-square random variable with $2(N-M+1)$
degrees of freedom \cite{R_W_Heath02}\cite{TT2007}. Then the
probability density function (PDF) of $X_{m,k}, \forall m$, can be expressed as
\begin{equation}
f_{k}{(x)}=\frac{\lambda_{k}^{N-M+1}x^{N-M}e^{-\lambda_{k}x}}{(N-M)!}
\label{chi_square_distribution}\end{equation}
where $\lambda_{k}=\frac{M\sigma_{N}^2}{P\sigma_{k}^2}$. For
simplicity, we drop the index $m$, denote $X_{m,k}$ as $X_{k}$, and
restrict our analysis for the case $N=M$. Therefore, the distribution
of $X_{k}$ becomes the exponential distribution with
parameter $\lambda_{k}$.
Let ${X_{(1)},X_{(2)},\cdots, X_{(K)}}$, be the order statistics of
independent continuous random variables $X_{1},X_{2},\cdots,X_{K}$,
with the PDF ({\ref{chi_square_distribution}) in decreasing order,
i.e., ${X_{(1)}\geq X_{(2)}\geq \cdots \geq X_{(K)}}$. The
cumulative distribution function (CDF) of the largest order statistics
$X_{(1)}$ can be shown as
\begin{eqnarray}
F_{X_{(1)}}(x)=\prod_{i=1}^{K}(1-e^{-\lambda_{i}x})
\end{eqnarray}
and its corresponding PDF can be expressed as
\begin{eqnarray}
f_{X_{(1)}}(x)=\frac{1}{(K-1)!}\sum_{{\mathbf{T}}}F_{i_{1}}(x)\cdots
F_{i_{K-1}}(x)f_{i_{K}}(x)
\end{eqnarray}
where $\sum_{\mathbf{T}}$ denotes the summation over all $K!$
permutations $(i_{1},i_{2},\ldots,i_{K})$ of $(1,2,\ldots,K)$.
Applying the maximum sum rate scheduling algorithm, the sum rate of
the system can be written as follows:
\begin{eqnarray}
R&=&{\mathbb{E}}\left\{\sum_{m=1}^{M}\log_{2}\left(1+\max_{1\leq k\leq K}{X_{m,k}}\right)\right\} \nonumber \\
&=& {M}\int_{0}^{\infty}\log_{2}(1+x)f_{X_{(1)}}(x)dx.
\label{max_sum_rate_formula}
\end{eqnarray}
In order to achieve the sum rate described in
(\ref{max_sum_rate_formula}), each user should feed back the index of the precoding vector on which it sees the highest SNR,
and the corresponding SNR value.
\begin{figure}[!t]
\centering
\includegraphics[width=0.4\textwidth]{Cluster_Model.eps}
\caption{Cluster-based Model.} \label{fig:
Cluster_based_feedback_model}
\end{figure}
\section{Cluster-based Feedback Model}
The transmission and feedback procedure can be described as follows.
From the previous channel condition or location information feedbacks from the users, the BS derives the users' mean SNR. The BS then groups the users' mean SNRs into $N_c$ clusters according to their magnitudes.
The mean SNRs in each cluster are similar in quantity and are used to derive one SNR threshold, denoted $r_{c,i}$ for cluster $i$. Note that derivation of the users' mean SNRs and the cluster thresholds is done periodically. Derivation of the users' mean SNRs does not need to be very accurate. The BS broadcasts periodically the threshold
set $\{r_{c,1}\geq r_{c,2}\geq\cdots\geq r_{c,N_{c}}\}$ to the users.
At every feedback instant, each user compares its measured instantaneous SNR with the
thresholds to determine which cluster its instantaneous SNR belongs
to, and feeds back to the BS the cluster index and the quantization bits of the instantaneous SNR
for this cluster. When a user's SNR is smaller
than $r_{c,N_{c}}$, the user does not feedback to the BS. The proposed
procedure takes a little downlink bandwidth for the BS to
periodically broadcast the threshold set, in exchange of reduction
of the uplink feedback bandwidth.
To derive the thresholds, the overall statistics of users are
divided into multiple clusters according to the mean SNRs
$\frac{1}{\lambda_{1}}, \frac{1}{\lambda_{2}}, \cdots,
\frac{1}{\lambda_{K}}$. The means of the random variables ${X_{1},
X_{2},\cdots,X_{K}}$ are ranked with decreasing order as
$\frac{1}{\lambda_{(1)}}\geq \frac{1}{\lambda_{(2)}}\geq \cdots\geq
\frac{1}{\lambda_{(K)}}$ and uniformly divided into $N_{c}$ clusters
with size $L=K/N_{c}$ as
$\left\{\frac{1}{\lambda_{(1)}},\frac{1}{\lambda_{(2)}},\cdots,\frac{1}{\lambda_{(L)}}\right\}$,$\cdots$,
$\left\{\frac{1}{\lambda_{(L(N_{c}-1)+1)}},\frac{1}{\lambda_{(L(N_{c}-1)+1)}},\cdots,\frac{1}{\lambda_{(K)}}\right\}$.
To simplify the notation, we let random variable $Y_{m}^{n}$
represent the $m$th statistic in the $n$th cluster. Then,
$Y_{m}^{n}$ is exponentially distributed with mean value
$\frac{1}{\lambda_{(L(n-1)+m)}}$. The random variables in cluster
$n$ are $\left\{Y_{1}^{n},Y_{2}^{n},\cdots,Y_{L}^{n}\right\}$, and
let $Y_{(1)}^{n}\geq Y_{(2)}^{n}\geq\cdots\geq Y_{(L)}^{n}$ be the
order statistics of them. For a measured instantaneous SNR $y_i$ at
user $i$, we say that the rank of $y_i$ in cluster $n$ is $d$ if
$\underbrace{Y_{(1)}^{n}\geq Y_{(2)}^{n} \geq Y_{(3)}^{n}\cdots
}_{\text{$(d-1)$ variables}}\geq y_{i} \geq
\underbrace{Y_{(d+1)}^{n}\cdots \geq Y_{(K)}^{n}}_{\text{ $(L-d)$
variables}}$. In the following, we will discuss how to derive the
threshold in each cluster for heterogenous and homogeneous channel
distributions to reduce the feedback load.
\subsection{Heterogenous Case}
\subsubsection{Cluster-based Type-I}
With maximum sum rate scheduling, a user will be scheduled when its SNR on a particular precoding vector is the highest among the users. On the other hand, if a user has a low SNR, it is unlikely to be scheduled. For this user, feeding back CSI is wasteful of the uplink radio resource. The threshold of each cluster is designed according to the probability of a user's measured instantaneous SNR being a particular rank in that cluster.
Let $P_{m,n}^{i}(r)$ denote the probability that user $m$ is ranked $n$ among the $L$ users in
cluster $i$ when its instantaneous SNR is $r$.
\begin{eqnarray}
\begin{array}{l}
P_{m,n}^{i} (r)=P\{Y_{m}^{i}=Y_{(n)}^{i}|Y_{m}^{i}=r\} \nonumber \\
=\displaystyle{\sum_{\mathbf{S}}}P\{\underbrace{Y_{t_{1}}^{i}\geq
Y_{t_{2}}^{i}\geq \cdots \geq Y_{t_{n-1}}^{i}}_{\text{$(n-1)$
variables}}\geq r
\geq \underbrace{Y_{t_{n+1}}^{i} \cdots \geq Y_{t_{L}}^{i}}_{\text{ $(L-n)$ variables
}}\}\\
=\displaystyle{\frac{{\displaystyle{\sum_{\mathbf{S}}}e^{{-\displaystyle{\sum_{a=1}^{n-1}\lambda_{(L(i-1)+t_{a})}r}}}\displaystyle{\prod_{b=n+1}^{L}(1-e^{{-\lambda_{(L(i-1)+t_{b})}r}})}}}{(n-1)!(N_{c}-n)!}}\\
\end{array}
\end{eqnarray}
where $\sum_{\mathbf{S}}$ denotes the summation over all
permutations $(t_{1},\cdots,t_{n-1},t_{n+1},\cdots,t_{L})$ of
$(1,2,\cdots,m-1,m+1,\cdots,L)$ in the cluster $i$. For
example, when the instantaneous SNR of user $t$ in cluster $1$ is infinity,
the rank of user $t$ among the $L$ users in cluster $1$ is one
with probability one, i.e.,
$P_{t,1}^1 (\infty) = P\{Y_{t}^{1}=Y_{(1)}^{1}|Y_{t}^{1}=\infty\}=1$.
Being valid conditional probabilities, the $P_{m,n}^{i}(r)$'s satisfy
\begin{eqnarray}
\sum_{n=1}^{L}P_{m,n}^{i} (r)=1 ,&m=1,2,\cdots,L.
\end{eqnarray}
The most probable rank of user $m$ in cluster $i$ when its instantaneous SNR is $r$ can be obtained by
\begin{equation}
\hat{n}=\arg \max_{n\in \{1,2,\ldots,L \}} P_{m,n}^{i} (r).
\end{equation}
Let $Q_{m,n}^{i}$ be defined such that $P_{m,n}^{i} \left( Q_{m,n}^{i}\right)=P_{m,n+1}^{i} \left( Q_{m,n}^{i}\right)$. It can be seen that when the instantaneous SNR of user $m$ falls in the range $\left[Q_{m,n}^{i},Q_{m,n-1}^{i}\right)$, the most probable rank of user $m$ in cluster $i$ is $n$.
For maximum sum rate scheduling, we let only the users who are most likely to be rank one in each cluster to feedback.
Therefore, a threshold for each cluster $i$ is set as
\begin{eqnarray}
r_{c,i}=\max\{Q_{1,1}^{i},Q_{2,1}^{i},\cdots,Q_{L,1}^{i} \} &
i=1,2,\cdots,N_{c}
\end{eqnarray}
which satisfies the following inequality
\begin{eqnarray}
r_{c,1}\geq r_{c,2}\geq\cdots\geq r_{c,N_{c}}.
\end{eqnarray}
To reduce the computational complexity, another
simple and low complexity method is proposed in the next section.
\subsubsection{Cluster-based Type-II} \label{type-II}
We approximate by assuming that the random
variables $Y_{m}^{i}, m=1,2,\cdots,L$ in cluster $i$ have the same
exponential distribution with mean $\mu_{c_{i}}$ obtained by
\begin{eqnarray}
\mu_{c_{i}}=\frac{\sum_{m=1}^{L}\lambda_{(L(i-1)+m)}}{L}.
\end{eqnarray}
Then, $P_{m,n}^{i} (r)$ is the same for all users in a cluster, and can be obtained by
\begin{eqnarray}
P_{m,n}^{i} (r)=\frac{(L-1)!\exp{(-\mu_{c_{i}}r)}^{n-1}{(1-\exp(-\mu_{c_{i}}r))}^{L-n}}{(n-1)!(L-n)!}.
\end{eqnarray}
With this approximation, the closed-form solution of $r_{c,i}$ can be derived as
\begin{eqnarray}
r_{c,i}=Q_{1,1}^{i}=\frac{1}{\mu_{c_{i}}}\ln{L} , i=1,2,\cdots,
N_{c}. \label{closed_form_threshold}
\end{eqnarray}
\subsection{Homogeneous Case}
The homogeneous case can be viewed as a special case of the
cluster-based type-II scenario using only one cluster $(N_{c}=1,
K=L)$. The mean values of the random variables $Y_{m}^{1},
m=1,2,\cdots,K$ are the same, i.e.,
$1/\lambda=1/\lambda_{1}=\cdots=1/\lambda_{K}$. As in \cite{JH2010},
multiple thresholds are set according to the most probable rank as
\begin{eqnarray}
r_{c,p}=Q_{1,p}^{1}=\frac{1}{\lambda}\ln{\left(\frac{K}{p}\right)}, & p=1,2, \cdots,
N_{c}.
\end{eqnarray}
For all cases, at each feedback instant, when a user's instantaneous
SNR is greater than the smallest threshold $r_{c,N_{c}}$, the user needs to
feed back to the BS using $B_{C}=\log_{2}\lceil N_{c} \rceil $ bits
to indicate which region (between two adjacent thresholds) its instantaneous SNR belongs to. In addition, in order
to differentiate the users SNR in the same region $i$, the
region $i$ which is represented by $\mathfrak{C_{i}}$ is further
quantized with $b_{i}$ bits. Therefore, the feedback bits of each
user include two parts: one is the region index with $B_{C}$ bits
and the other is the quantization bits for that region.
\section{Analysis of Sum Rate Loss }
A natural question to ask
is how many clusters (thresholds) should be set.
Obviously, if many clusters are used, the number of region index bits $B_{C}$
is increased. On the other hand, when a small number of
clusters are used, the smallest threshold $r_{c,N_{c}}$ becomes
large. Then, the sum rate loss caused by the threshold $r_{c,N_{c}}$ may
increase. Therefore, an appropriate number of clusters is important
for the design. We now analyze the sum rate loss of the
system caused by the smallest threshold $r_{c,N_{c}}$. We define a
random variable $Z_{i}$ as follows:
\begin{eqnarray}
Z_{i}=\left\{\begin{array}{cc}
0 ,& X_{i}>r_{c,N_{c}} \\
X_{i} ,& X_{i}\leq r_{c,N_{c}} \\
\end{array}\right..
\end{eqnarray}
According to the maximum sum rate criterion, the exact SNR loss
for the system is $Z_{(1)}=\max\{Z_{1},Z_{2},\cdots,Z_{K}\}$. The
probability of the rate loss event can be obtained by
\begin{eqnarray}
P_{L}=P\{X_{(1)}\in
(0,r_{c,N_{c}})\}=\displaystyle{\prod_{i=1}^{K}}(1-\exp{(-\lambda_{i}r_{c,N_{c}})}).
\end{eqnarray}
Therefore, the sum rate loss can be expressed as
\begin{eqnarray}
\Delta R=ME\{\log_{2}(1+Z_{(1)})\}P_{L}.
\end{eqnarray}
\begin{theorem}
\cite{D_B2006} Let the means and variances of the
random variables $Z_1, Z_2, \ldots, Z_K$ be
${\mathbf{\mu}}=(\mu_{1},\mu_{2},\cdots,\mu_{K})$ and
${\mathbf{\sigma^{2}}}=(\sigma_{1}^2,\sigma_{2}^2,\cdots,\sigma_{K}^2)$, respectively.
The closed form upper bounds on the expected value of the largest
order statistic is:
\begin{eqnarray}
E\{Z_{(1)}\} &\leq& {\sum_{i=1}^{K}\left\{\begin{array}{c}
\displaystyle{\frac{\mu_{i}+\sqrt{(\mu_{i}-T)^{2}+\sigma_{i}^{2}}}{2}}\nonumber \\
\end{array}\right\}+\frac{(2-K)T}{2}} \nonumber \\ &\triangleq&
\mu_{Z_{(1)}^{U}}
\label{expacation_upper_bound}
\end{eqnarray}
where $T=\max_{1\leq j\leq
K}\{\mu_{j}+\frac{K-2}{2\sqrt{K-1}}\sigma_{j}\}$.
\end{theorem}
Applying Jensen's
inequality and the upper bound of $E\{Z_{(1)}\}$, the sum rate loss on a certain beam can be
bounded by
\begin{eqnarray}
\frac{\Delta R}{M}&=&E\{\log_{2}(1+Z_{(1)})\}P_{L}\leq
\log_{2}(1+E\{Z_{(1)}\})P_{L} \nonumber \\
&\leq&\log_{2}(1+\mu_{Z_{(1)}^{U}})P_{L}.
\label{rate_loss_upper_bound}
\end{eqnarray}
Using (\ref{rate_loss_upper_bound}), for a given tolerable sum rate loss $\Delta R_{U}$, the minimum
number of clusters (thresholds) required can be determined.
\begin{proposition}
When the total number of users approaches infinity,
the sum rate loss caused by the smallest finite threshold $r_{c,N_{c}}$ is negligible. Thus, the full multiuser diversity can be achieved.
\end{proposition}
\begin{proof}
When the total number of users $K$ approaches to infinity,
$\displaystyle{\lim_{K\rightarrow \infty}}P_{L}=0$. In
addition, $Z_{(1)}$ is bounded by $r_{c,N_{c}}$. The sum rate
loss on a certain beam becomes zero
\begin{eqnarray}
\lim_{K\rightarrow \infty}\frac{\Delta R}{M}&=&\lim_{K\rightarrow \infty}E\{\log_{2}(1+Z_{(1)})\}P_{L} \nonumber \\
&\leq&\lim_{K\rightarrow \infty}
\log_{2}(1+E\{Z_{(1)}\})P_{L} \nonumber \\
&\leq&\lim_{K\rightarrow \infty}\log_{2}(1+r_{c,N_{c}})P_{L} \nonumber \\
&=&0.
\end{eqnarray}
\end{proof}
\section{Bit Allocation and Feedback Load Analysis }
\subsection{Bit Allocation}
Let $r_{c,0} = \infty$. Assume that the SNR region $i$ $[r_{c,i}, r_{c,i-1})$, denoted
$\mathfrak{C}_{i}, i=1,2,\cdots,N_{c}$, is quantized with $b_{i}$ bits. In region $\mathfrak{C}_{i}$, the quantization
levels using $b_{i}$ bits are expressed by $q_{t}^{i}, t=1,2,\cdots,2^{b_{i}}$, obtained by a pdf quantizer \cite{Jose06}. Thus, each level will occur with the same probability.
Under the per user average feedback load constraint $C_{k}$ for user $k$, the available bits will be assigned to the regions to maximize sum rate. Let
$P_{\mathfrak{C}_{i}}=P\{X_{(1)}\in \mathfrak{C}_{i} \},
i=1,2,\cdots,N_{c}$. The bit allocation problem can be
described as follows:
\begin{eqnarray}
&\displaystyle{\max_{(b_{1},b_{2},\cdots,b_{N_{c}})}}& {M\sum_{i=1}^{N_{c}}\frac{P_{\mathfrak{C}_{i}}}{2^{b_i}}\sum_{t=1}^{2^{b_{i}}}\log(1+q_{t}^{i})} \nonumber \\
& \text{s.t}& \sum_{i=1}^{N_{c}}P_{\mathfrak{C}_{k,i}}b_{i} \leq
C_{k}, k=1,2,\cdots,K, \label{opt_bit}
\end{eqnarray}
where $P_{\mathfrak{C}_{k,j}}=P\{X_{k} \in
\mathfrak{C}_{j}\}=e^{-\lambda_{k}r_{c,j}}-e^{-\lambda_{k}r_{c,j-1}}$.
The closed-form integer solution for (\ref{opt_bit}) does not exist.
We use exhaustive search to find the optimal bit allocation set to
maximize the sum rate.
\subsection{Feedback Load Analysis}
When a user's instantaneous SNR is smaller than the smallest threshold
$r_{c,N_{c}}$, the user does not need to feedback.
Thus, the feedback probability for user $k$ is
$e^{-\lambda_{k}r_{c,N_{c}}}$. The average total feedback
load can be expressed as
\begin{eqnarray}
F_{b}&=&M\sum_{k=1}^{K}\{e^{(-\lambda_{k}r_{c,N_{c}})}{(B_{C}+C_{k})}\}
\end{eqnarray}
\section{Simulation Results}
In this section, we show the sum rate and feedback load performance
for the proposed feedback scheme. The BS is equipped with $M=4$
antennas and the total number of users is $K=10\sim100$. In order to
perform ZF beamforming, we let the number of receive antennas $N=4$.
The total transmit power $P$ is 10W, while the additive white Gaussian noise power at the receivers $\sigma_N^2$ is 1W. Note that these numbers are selected only for illustration purpose.
The elements of the channel matrix
$\mathbf{H}_{i}$ for the $i$th user are assumed to be i.i.d. complex
Gaussian distribution with zero and variance $\sigma_{i}^2$, where
$\sigma_{i}^2$ are drawn uniformly from the interval $[0,1]$ to
model heterogeneous channels. The number of clusters in type-I and
type-II schemes is set to four (thus $B_{C}=2$) according to the
tolerable sum rate loss $\Delta R_{U}=10^{-2}$ $bps/Hz$. The
feedback load constraint of user $k$ is $C_{k}=0.8, k=1,2,\cdots,K$.
In the simulation, the proposed feedback schemes are compared with
the conventional scheme and the single-threshold schemes proposed in
\cite{Gesbert04}. In the conventional feedback scheme, no matter
what the instantaneous SNR is, it is always quantized with $3$ bits.
In the single-threshold scheme, the region $(r_{th},\infty)$ of the
SNR is quantized using $3$ bits. The single threshold $r_{th}$ is
established according to the scheduling outage probability
$P_{out}$.
In Fig.~\ref{fig: sum_rate}, the type-I feedback scheme achieves the
highest sum rate, and the low-complexity type-II scheme
achieves almost the same rate as the type-I scheme. As shown in Fig.~\ref{fig: feedback_load}, the feedback
load of the conventional scheme increases linearly with the number of users. Using our
proposed schemes, the total number of feedback bits can be
dramatically reduced. Overall, the proposed schemes use fewer bits to achieve higher sum rate than the conventional scheme. The single-threshold scheme has lower feedback load, but suffers significantly in the sum rate performance.
In Fig.~\ref{fig: efficiency}, we plot the sum rate vs. the total
feedback load as an indication of the efficiency. The type-I scheme
is the most efficient (i.e., making best use of the feedback bits), but has high computational complexity.
The low-complexity type-II scheme not only achieves high sum rate
but also reduces the feedback load significantly. The single-threshold
feedback scheme only achieves a sum rate of about $11(bps/Hz)$ with a small feedback load.
\begin{figure}
\centering
\includegraphics[width=0.42\textwidth]{Sum_rate.eps}
\caption{Sum rate comparison between different feedback schemes.}\label{fig: sum_rate}
\end{figure}
\begin{figure}
\centering
\includegraphics[width=0.42\textwidth]{Feedback_load.eps}
\caption{Feedback load comparison between different feedback
schemes.}\label{fig: feedback_load}
\end{figure}
\begin{figure}
\centering
\includegraphics[width=0.42\textwidth]{Efficiency.eps}
\caption{The efficiencies of different feedback schemes.}\label{fig: efficiency}
\end{figure}
\section{Conclusion}
In this paper, we investigated the feedback load reduction problem in
multiuser MIMO broadcast system. We
proposed a cluster-based feedback scheme to reduce the feedback load
in heterogenous and homogeneous channels. The bit allocation problem
for the multiple-cluster feedback scheme was also discussed. The
simulation results showed that, compared to the existing feedback schemes, the cluster-based feedback scheme can make the best use of the feedback bits to achieve good feedback load reduction while maintaining good sum rate
performance.
\bibliographystyle{IEEEbib}
|
\section{Introduction}
Various systems in nature exhibit skew distributions, which are properly fit to
the Weibull distribution \cite{Weibull} as well as lognormal and power-law distributions;
relations between those skew distributions have been discussed recently~\cite{Choi}.
In particular, the Weibull distribution,
despite the simple mathematical form, particularly for the
cumulative distribution $F(x) = 1-\exp[-(x/\eta)^\beta]$,
has flexible shapes depending on the value of $\beta$
and is widely used to describe size distributions of, e.g.,
material strengths~\cite{Weibull, Kirchner}, cloud droplets~\cite{Liu},
biological tissues~\cite{Jo}, ocean wave heights~\cite{Muraleedharan},
and wind speeds~\cite{Takle}.
However, there still lacks an appropriate explanation of its ubiquitous emergence,
in sharp contrast with the Gaussian distribution, let aside the case-by-case
derivation such as material breaking with the weakest element~\cite{Weibull},
entropy maximization~\cite{Liu}, material fragmentation~\cite{Brown},
and extreme value statistics~\cite{Bertin,Moloney}.
It is well known that the binomial distribution results from
success events for given independent trials with the success
probability $p$ given. When the success is a rare event (i.e., $p$
is small), it reduces to the Poisson distribution. According to
the central limit theorem~\cite{Kallenberg}, (discrete) binomial
and Poisson distributions approach the (continuous) Gaussian
distribution in the limit of large trial numbers. In a similar
spirit, we here derive a continuous Weibull-like distribution from
the discrete Galton-Watson branching process, motivated by cell
replication in a tissue~\cite{Jo}. The branching process can serve
as a basic model to describe discrete events having two
possibilities, e.g., replication/non-replication or
nucleation/non-nucleation. The generating function for this
distribution was first obtained in the seminal work of general
branching processes~\cite{harris48,seneta69}. Specifically,
asymptotics were derived in the more general case of multiple
replicates and extinction processes at each stage of the process,
added to possible immigration events (see for example Ref.
\cite{biggins93}), but little is known about the shape of the
distribution itself relatively to other standard distributions,
except for few very specific cases where the limiting distribution
can be computed exactly through the use of a rational form for the
generating function at the first stage of the process and which
usually leads to a simple exponential function. Here we find that
it is approached by the Weibull distribution in rather a wide and
realistic range of the replication parameter $p$, making the two
distributions surprisingly indistinguishable in practice.
This paper consists of four sections and an appendix.
In Sec. II, cell replication is described in terms of a branching process.
The stationary distribution of the branching process is obtained and its general
properties are discussed. Results of Monte Carlo simulations are also presented.
Section III examines the relation between the distributions for different replication probabilities and probe the scaling with the help of an ansatz, which is justified from the exact series expansion.
Finally, Sec. IV discusses and summarizes the results.
In Appendix, all the moments of the distribution are obtained analytically from
the recurrence relation of the generating function.
\section{Cell replication and branching process}
For the binomial distribution, an independent event occurs at each time with given success probability.
In cell replication, on the other hand, the number of replication events in consideration depends on the current cell number of a tissue.
For example, even if there exists just a single mother cell initially, it may replicate from time to time, and
there may occur many replications of the mother and daughter cells.
Accordingly, we consider the probability distribution $f_n(l)$ of tissues with size
(i.e., the number of cells) $l$ at given time step $n$, which
satisfies the normalization condition $\sum_{l=1}^{2^n} f_n(l) =1$ with the
maximum possible cell number in the tissue after the $n$th replication given
by $2^n$.
Note that this process can be described in terms of a branching process with the
branching probability $p$, as illustrated in Fig. \ref{fig:branching}.
Each graph in the figure, where sites in the $n$th row represent cells
at the time step $n$, corresponds to one possible configuration
of cell growth for the given duration.
Each graph thus starts from a single site in the first row (i.e., a single mother cell initially);
sites may replicate or not, giving birth to new sites
at successive time steps (here the time step is fixed to be a constant).
\begin{figure
\centerline{\epsfig{file=fig1.eps,width=8.5cm}}
\caption{Cell replication graphs for a branching process. Cell number configurations at time steps $n = 1$ and $2$ are plotted with the replication probability at each step is given by $p$; $q \equiv 1-p$ corresponds to the probability that the cell does not replicate.}
\label{fig:branching}
\end{figure}
It is useful to consider the generating
function for the distribution $f_n$ at time $n$ in the branching
process \cite{harris48,Harris}:
\begin{eqnarray}
g_n(z)=
\sum_{l=1}^{2^n}f_n(l)z^l.
\end{eqnarray}
For example, the $k$th moment
at time $n$, defined to be $\sum_{l=1}^{2^n}f_n(l)\,l^k$, can be
computed by differentiating successively the generating function:
$\left (z\frac{d}{dz}\right )^kg_n(z)|_{z=1}$
with $g_n(1) =1$ for all $n$
(see Appendix for the derivation of all the moments).
In the following, for simplicity, we will impose $f_n(l)=0$ for $l>2^n$.
At the initial time ($n=0$) the system contains only
one element, leading to $g_0(z)=z$. Since the distribution $f_{n+1}$ is
related to the preceding one $f_n$ via combinatorial relations, it is easy to show
that the generating function satisfies the non-linear recursion equation,
$g_{n}(z) = g_1(g_{n-1}(z))$
for $n \geq 1$, where $g_1(z) = qz+pz^2$.
This equation provides a recursive function for the newly generated sites,
which are all independent,
with the generating function $g_1 (z)$.
>From this relation, we can deduce that the total number $N(n)$ of configurations or graphs
at (discrete) time $n$ satisfies the recurrence relation $N(n{+}1)=N(n)[1+N(n)]$,
with the initial condition $N(0)=1$, and grows rapidly in time.
Indeed this relation can be obtained easily from the observation that
$N(n)$ is equal to $g_n(1)$ with $p$ and $q$ replaced formally by unity.
Therefore $N(n)$ satisfies the same relation as $g_n(1)$ above.
It is also manifested from the physical point of view:
Given $N(n)$ graphs at time $n$,
there are two possible ways to generate graphs at time $(n{+}1)$.
(i) In the case of non-replication of the original site,
we simply have $N(n)$ graphs; (ii) in the case of replication of the same site,
we can attach to the two offsprings a total of $N(n)^2$ graphs.
As a result, we obtain $N(n)+N(n)^2$ possible configurations at time $(n{+}1)$.
This can be checked in Fig.~\ref{fig:branching}
for the first few graphs: $N(0)=1$, $N(1)=2$, $N(2)=6$, and so on.
Because a tissue of size $l$ results from $(l{-}1)$-times proliferation
starting from a single cell (see Fig.~\ref{fig:branching}), the recurrence relation
\begin{eqnarray}
g_{n+1}(z)=q g_{n}(z)+ p g_{n}^2(z)
\end{eqnarray}
leads to the recursive relation for the distribution $f_n(l)$
by simply identifying the coefficients of $z^l$ on the left and right sides of the last expression:
\begin{eqnarray} \label{fn}
f_{n+1}(l)=q f_n(l) + p \sum_{k=1}^{l-1} f_n(k) f_n(l-k).
\end{eqnarray}
Namely, a tissue of size $l$ at time $n+1$ can be generated
in the two ways:
(i) no replication at the first time step followed by producing $l$ descendants
at the following $n$ time steps
and (ii) replication at the first time step followed by
producing $k$ descendants from one offspring and
$l{-}k$ descendants from the other offspring at the following $n$ time steps.
The size distribution, computed from Eq.~(\ref{fn}), is exhibited in Fig.
\ref{fig:sizecomp}, together with that from Monte Carlo
simulations, manifesting perfect agreement. It is of interest that
Eq.~(\ref{fn}) can be mapped into a process of random aggregation
of clusters with the aggregation probability $p$. Using $q=1-p$
and $\sum_{k=1}^{2^n} f_n(k) = 1$, we thus obtain
\begin{eqnarray}
\label{aggregation} \Delta f_n(l) = -p \sum_{k=1}^{2^n} f_n(l)
f_n(k) + p \sum_{k=1}^{l-1} f_n(k) f_n(l-k)
\end{eqnarray}
with $\Delta
f_n(l) \equiv f_{n+1}(l) -f_n(l)$. Therefore a cluster of size $l$
can be formed from aggregation of a cluster of size $k$ and a
cluster of size $(l-k)$ with the aggregation probability $p$.
\begin{figure
\centerline{\epsfig{file=fig2.eps,width=8.5cm}}
\caption{(color online) Comparison of the tissue size distribution $f_n (l)$ at time $n=10$,
for the replication probability $p = 0.3, 0.5, 0.8$, and $0.9$.
Analytical (solid lines) and simulation (+ signs) results agree perfectly,
displaying multimodal shapes for large values of $p$.
In the Monte Carlo simulations of the branching process,
starting from a single cell, we let every cell replicate with a given replication probability
at each Monte Carlo step. Data have been obtained from $10^6$ trial moves.
}
\label{fig:sizecomp}
\end{figure}
Figure \ref{fig:weibull} shows the normalized size distribution
for $p=0.3$ at several time steps $n=10, 12$, and $14$.
Remarkably, when size $l$ is rescaled by the factor $(1+p)^{n}$,
the distributions collapse into a single curve independent of $n$,
suggesting the presence of a stationary distribution for the
branching process~\cite{harris48}. Indeed the average cell number
in a tissue after the $n$th replication with the replication
probability $p$ is given by $(1+p)^n=\sum_{l =1}^{2^n} l f_n (l)$.
Note that $f_n(l)$ may be regarded as a continuous function
$f_n(x)$ when $n$ is large (see Fig. \ref{fig:weibull}).
Since the average cell number after the $(n{-}1)$th replication
is $(1+p)^{n-1}$, we have the scaling relation
\begin{eqnarray}
\int dx \,x f_{n}(x)&=&(1+p) \int dx'
\,x' f_{n-1}(x')
\nonumber \\
&=&\int dx\,(1+p)^{-1} x f_{n-1}((1{+}p)^{-1} x),
\end{eqnarray}
which is consistent with the fact that the distribution in the long-time limit
can be described by a time-independent stationary function
$f(\tilde{x})$ with the rescaled size $\tilde{x} = x/\eta$
and the scale parameter $\eta= a (1+p)^n$.
The scale factor $a$ introduced here depends in particular on the replication
probability $p$ via boundary conditions, as discussed later.
\begin{figure
\centerline{\epsfig{file=fig3.eps,width=8.5cm}}
\caption{(color online) Weibull distribution of tissue sizes in the cell replication process.
Cell-number distribution for $p=0.3$ at three different time steps.
Distributions versus the rescaled size are plotted in the inset;
the collapse is fitted with a Weibull function with the shape parameter $\beta=1.37$ (black line).}
\label{fig:weibull}
\end{figure}
Finally, a quantity of interest is given by the Laplace transform
$\hat f(\lambda) \equiv \int_0^{\infty} d{\tilde x} \,e^{-\lambda {\tilde x}}
f({\tilde x})$, for which the recursive relation in Eq.~(\ref{fn}) reads
\cite{harris48} \begin{eqnarray}\label{laplac}
\hat f((1{+}p)\lambda)=q\hat f(\lambda)+p\hat f(\lambda)^2 . \end{eqnarray}
Equation (\ref{laplac}) takes the form of a Poincar\'e-type
equation \cite{lakner2008}, which is directly related in property
to Mahler functional equations \cite{mahler80} via an appropriate
change of variables \cite{poincare}.
In the limit of small $p$ where cells replicate very rarely, one
may expand Eq. (\ref{laplac}) as $\hat f((1{+}p)\lambda)\approx
\hat f(\lambda)+p\lambda \hat f'(\lambda)$, to obtain the differential
equation:
\begin{eqnarray}
\lambda
\hat f'(\lambda)=\hat f(\lambda)^2-\hat f(\lambda)
\end{eqnarray}
with the initial
conditions $\hat f(0)=1$ and $\hat f'(0)=-a^{-1}$. The solution reads
$\hat f(\lambda)=a(\lambda+a)^{-1}$, the inverse Laplace transform
of which is given by the simple exponential function
$f({\tilde x})=a\exp(-a{\tilde x})$. With the constraint $F(1)=1-e^{-1}$ on the
cumulative distribution
$F(\tilde{x}) \equiv \int_0^{{\tilde x}} d\tilde{x}' f(\tilde{x}')$, we
obtain the scaling factor $a=1$ and therefore $f({\tilde x})=\exp(-{\tilde x})$.
In the opposite case of $p=1$ where every cell replicates, we have
$\hat f(2\lambda)=\hat f(\lambda)^2$, with the simple solution
satisfying the initial conditions given by
$\hat f(\lambda)=\exp(-\lambda/a)$. This leads to the Dirac delta
distribution $f({\tilde x})=\delta({\tilde x}-a^{-1})$ and the Heaviside
cumulative distribution $F({\tilde x})=\theta({\tilde x}-a^{-1})$.
The constraint on $F(1)$ again imposes $a=1$.
\section{Scaling of the size distribution}
In this section, we consider the general case of $0<p<1$. As for the unique
stationary distribution $f(\tilde{x})$ for given $p$, one may
question whether there exists any relation between the
distribution $f(\tilde{x})$ corresponding to two different
replication probabilities $p$ and $p_0$, respectively. Since the
final stationary distributions result from the same branching
process, albeit with different branching probabilities, they are
expected to share qualitatively the same properties.
To probe the scaling of the tissue size in the replication process,
we display in Fig. \ref{fig:weibull2} the cumulative distribution for
the replication probability $p=0.1, 0.3,$ and $0.5$.
Note that the scale factor $a$ in the rescaling of the size has been adjusted
to satisfy the condition $F(\tilde{x}{=}1)=1-e^{-1}$.
To probe the functional relations between the cumulative distributions
for different values of $p$
under the constraints for $F$, we consider the change of variable
$\tilde{x}\rightarrow\tilde{x}^\beta$, as the simplest
possibility, where the exponent $\beta=\beta(p)$ is then adjusted
to make all curves for considered values of $p$ collapse onto a
single curve.
This ansatz indeed leads to the collapse of different cumulative
distributions into a unique distribution $F_0(\tilde{x})= 1 -
e^{-\tilde{x}}$, as shown in the inset. Therefore the new variable
$\tilde{x}^\beta$ determines the functional form of
$F(\tilde{x})$, at least for the numerical cases considered.
Indeed, using the known result $F(\tilde{x})=1-e^{-\tilde{x}}$ in
the limit $p\to 0$, we obtain
$F(\tilde{x})=1-e^{-\tilde{x}^\beta}$ with a good precision for
$p>0$, which leads to the Weibull distribution.
\begin{figure
\centerline{\epsfig{file=fig4.eps,width=8.5cm}}
\caption{(color online)
Cumulative distribution for $p=0.1, 0.3$, and $0.5$.
The rescaled size is given by $\tilde{x}= a^{-1}(1+p)^{-n} l$ with $n=20$, where
$a$ is the scale factor to adjust $F(1)=1-e^{-1}$.
Rescaled cumulative distribution functions are plotted in the inset,
disclosing the collapse into the function $F({\tilde x})=1-e^{-\tilde{x}^\beta}$ (black line).}
\label{fig:weibull2}
\end{figure}
The ansatz of the scaling ${\tilde x}^\beta$ can be justified from the exact series expansion
of the distribution $f({\tilde x})$.
Multiplying both sides of Eq. (\ref{laplac}) by $\exp(-i\lambda
{\tilde x})$, performing the rotation $\lambda \rightarrow i\lambda$, and
integrating over $\lambda$ along the real axis, we obtain
\begin{eqnarray} \label{func}
\frac{1}{1+p}f((1{+}p)^{-1}{\tilde x})=
qf({\tilde x})+p\int_0^{{\tilde x}}d{\tilde x}'f({\tilde x}')f({\tilde x}{-}{\tilde x}').
\end{eqnarray}
It can be
shown that $f({\tilde x})$ admits a series expansion in powers of ${\tilde x}$
consistent with the previous relation. In particular, $f({\tilde x})$
vanishes at the origin as $f({\tilde x}) \approx a_0{\tilde x}^{\beta-1}$, with
some constant $a_0$ and exponent $\beta= - [\log(1{+}p)]^{-1}
\log(1{-}p)\ge 1$ \cite{harris48}.
Here this analysis can be extended to consecutive terms to yield
the following expansion
\begin{eqnarray}\label{expf}
f({\tilde x})={\tilde x}^{\beta-1}\sum_{k\ge 0} a_k \,{\tilde x}^{k\beta} ,
\end{eqnarray}
where
identifying the powers in Eq. (\ref{func}) gives the recursion
relation for the coefficients:
\begin{eqnarray}\label{ak}
(q^{k+1}- q) a_k = p\sum_{l=0}^{k-1}
B(\beta(1{+}l),\beta(k{-}l)) a_l \,a_{k-1-l}
\end{eqnarray}
with the beta function
$B(x,y)=\int_0^1dt\,t^{x-1}(1-t)^{y-1}$.
Here $a_0$ is the only unknown parameter depending on boundary
conditions, since Eq. (\ref{ak}) implies the proportionality
relation $a_k\propto a_0^{1+k}$.
>From these results, it is easy to see that $f({\tilde x})$ can be cast
into the form
\begin{eqnarray}\label{scalingfunc}
f({\tilde x})=a_0\,{\tilde x}^{\beta-1}{\cal F}(a_0{\tilde x}^{\beta})
\end{eqnarray}
with the
unique regular expansion of the scaling function:
${\cal F}({\tilde x})=\sum_{k\ge 0}\tilde a_k {\tilde x}^{k}$, where $\tilde a_k$
satisfies the relation in Eq. (\ref{ak}) but with the initial term
$\tilde a_0=1$; this determines uniquely all the other
coefficients $\tilde a_{k}$ for $k\ge 1$. The cumulative
distribution $F({\tilde x})$ is equal to a scaling function of the
variable $a_0{\tilde x}^{\beta}$ alone since
\begin{eqnarray}\label{cumulativeF}
F({\tilde x})
=\frac{1}{\beta}\sum_{k\ge 0}\frac{\tilde a_k}{k+1}(a_0{\tilde x}^{\beta})^{k+1}
={\cal G}(a_0{\tilde x}^{\beta}),
\end{eqnarray}
where ${\cal G}$ is, like ${\cal F}$, uniquely defined by the coefficients $\tilde a_k$.
The parameter $a_0$ is defined according to the constraint
$F(1)=1-e^{-1}$, and can be related to $a$ via the equation for
the first moment $\int_0^{\infty}d{\tilde x} \,{\tilde x} f({\tilde x}) = a^{-1}$.
This relation simply gives
$a_0=a^{\beta}\left [\int_0^{\infty}u^{1/\beta}{\cal F}(u)du \right]^{\beta}$.
Note that the cumulative distribution $F$ is a function of the
variable ${\tilde x}^\beta$ up to a scaling factor, which is also true for
the Weibull distribution, $F({\tilde x}^\beta)=1-\exp(-{\tilde x}^\beta)$ with
${\tilde x}=x/\eta$. In the limit of small $p$, $\beta$ is close to unity
and one can show that the expansion coefficients satisfying Eq.
(\ref{ak}) are approximatively given by $\tilde a_k=(-1)^k/k!$.
Therefore ${\cal G}(a_0{\tilde x}^{\beta}) \approx 1-\exp(-a_0{\tilde x}^\beta)$ is
indeed close to the Weibull distribution.
The previous results show that the distribution can be expanded as
a series and vanishes as a power law with the exponent $\beta{-}1$
related to the replication probability $p$. In the opposite case
of large ${\tilde x}$, the integral equation (\ref{func}) can be
analyzed. Since we expect $f({\tilde x})$ to decrease with ${\tilde x}$ and
assume the stretched exponential behavior: $f({\tilde x})\approx
\exp(-a_{\infty}{\tilde x}^{\beta'})$ with $a_{\infty}$ constant, we observe that in
Eq. (\ref{func}) the left-hand-side term $f((1{+}p)^{-1}
{\tilde x})\propto \exp[-a_{\infty} (1+p)^{-\beta'}{\tilde x}^{\beta'}]$ is dominant
over the first term $f({\tilde x})$ on the right-hand side. The last term
can be analyzed by means of the saddle point analysis for the
function ${\tilde x}'^{\beta'}+({\tilde x}-{\tilde x}')^{\beta'}$ appearing in the
exponential contribution. The saddle point, obtained by taking the
extremum of this quantity with respect to ${\tilde x}'$, corresponds to
the middle point of the integration ${\tilde x}'={\tilde x}/2$. The overall
integral gives therefore a dominant contribution proportional to
$\exp[-2a_{\infty}({\tilde x}/2)^{\beta'}]$. The ansatz is consistent if the
two coefficients satisfy the relation
$(1+p)^{-\beta'}=2^{1-\beta'}$. This results in a new exponent
$\beta'= \log 2 [\log 2 -\log(1{+}p)]^{-1}$ valid in the
asymptotic limit; this was also obtained in Ref. \cite{harris48}.
\section{Discussion}
\begin{figure
\centerline{\epsfig{file=fig5.eps,width=8.5cm}}
\caption{(color online) Relation between exponent $\beta_{w}$ of the Weibull distribution and
the replication probability $p$ of the branching process.
The exponent $\beta$ is also plotted for comparison.}
\label{fig:exp}
\end{figure}
It has been shown that the replication process of cells with not
too large replication probability ($p \lesssim 0.5$) gives rise to
a distribution extremely close to the Weibull function. The
parameters of the Weibull distribution can then be related with
the first two moments of the distribution function $f_n(x)$:
$(1+p)^n = \eta \Gamma ( 1{+}\beta_{w}^{-1})$ and $2(1+p)^{2n-1} =
\eta^2 \Gamma (1{+}2\beta_{w}^{-1})$, where $\Gamma (x)$ is the Gamma function.
This leads to the following
relation between the replication probability $p$ and the shape
parameter $\beta_{w}$ of the Weibull distribution:
\begin{equation}
p = 2 \frac{\Gamma^2(1{+}\beta_{w}^{-1})}{\Gamma(1{+}2\beta_{w}^{-1})} - 1 ,
\label{pbeta}
\end{equation}
which is exhibited in Fig. \ref{fig:exp}.
In addition, the scale factor $a$ in the rescaling parameter
$\eta=a(1+p)^n$ is given by $a=\Gamma^{-1}(1{+}\beta_{w}^{-1})$. Note
that the exponents $\beta$ and $\beta_{w}$ are hardly distinguishable
for $p \lesssim 0.5$, where the scaling function ${\cal F}$ is
asymptotically similar to an exponential. This suggests that the
distribution in Eq. (\ref{scalingfunc}) belongs to the Weibull
class for small $p$. This regime applies to many cases in nature
that a certain event such as replication or nucleation occurs with
probability less than 50\,\% at a given time unit. On the other
hand, the replication process with a large value of $p$ results in
a different type of distribution, e.g., a multimodal distribution
(see Fig. \ref{fig:sizecomp}).
In conclusion, the branching process provides a general mechanism of the Weibull distribution
with $\beta \lesssim 2$, corresponding to the branching probability $p \lesssim 0.5$.
We have also found that the branching process can be mapped into a process of
aggregation of clusters.
A recent example includes the protein aggregation process with fission,
where the Weibull distribution with $\beta \sim 2$ emerges
as a stationary solution~\cite{Kunes}.
\begin{acknowledgments}
One of us (M.Y.C.) thanks Institut Jean Lamour, D\'epartement de Physique de la Mati\`ere et des Mat\'eriaux
at Universit\'e Henri Poincar\'e, where part of this work was carried out, for hospitality during his stay.
This work was supported by the intramural research program of NIH, NIDDK (J.J.)
and by NRF through the Basic Science Research Program (2009-0080791) (M.Y.C.).
\end{acknowledgments}
|
\section{Introduction}
Non-commutative gravity theories are receiving much attention in the
literature because they are part of a promising research program aimed
at developing an algebraic approach to the longstanding problem of
quantum gravity.$^{1}$ In particular, we are here concerned with the
developments described in Refs. 2 and 3.
The work in Ref. 2 has built a geometric theory
of non-commutative gravity where the Lagrangian is a globally defined
4-form, invariant under diffeomorphisms as well as $*$-diffeomorphisms
and where in the commutative limit only the classical field
degrees of freedom survive.
For pure gravity the action functional reads as
\begin{equation}
S=\int {\rm Tr}\Bigr(i \hat R \wedge_{*}\hat V
\wedge_{*}\hat V \gamma_{5}\Bigr),
\label{(1)}
\end{equation}
where $\hat V$ is the tetrad 1-form and $\hat R$ is the curvature 2-form
\begin{equation}
\hat R=d \hat\Omega-\hat\Omega \wedge_{*}\hat \Omega.
\label{(3)}
\end{equation}
As in Ref. 2 we expand the tetrad on the basis of
$\gamma$-matrices,
\begin{equation}
\hat V=\hat V^a\gamma_{a}+\widetilde {\hat V}{}^a\gamma_{a}\gamma_{5}.
\label{(2)}
\end{equation}
On denoting by $\eta_{ab}={\rm diag}(1,-1,-1,-1)$ the Minkowski
metric, and defining $\gamma_{ab} \equiv \gamma_{a}\gamma_{b}-I
\eta_{ab}$, one similarly expands the spin-connection 1-form
\begin{equation}
\hat\Omega={1\over 4}\hat\omega^{ab}\gamma_{ab}+i \hat\omega
I +{\widetilde {\hat\omega}}\gamma_{5},
\label{(4)}
\end{equation}
where $\hat \omega=\hat\omega_{\mu}dx^{\mu}$ and ${\widetilde {\hat\omega}}
={\widetilde {\hat \omega}}_{\nu}dx^{\nu}$ are 1-forms. From (3) and (4),
one has
\begin{equation}
\hat R={1\over 4}\hat R^{ab}\gamma_{ab}+i \hat r I +{\widetilde {\hat r}} \gamma_{5},
\label{(5)}
\end{equation}
where $\hat r$ and ${\widetilde {\hat r}}$ are 2-forms. We here correct some
numerical factors in Eqs. (5.2) and (5.3) of Ref. 2 and write the
following explicit formulae for the components of
the curvature in Eq. (5),
\begin{eqnarray}
\hat R^{ab}&=& d \hat\omega^{ab}-{1\over 2} \hat\omega_{\; c}^{a} \wedge_{*}
\hat\omega^{cb}+{1\over 2}\hat \omega_{\; c}^{b} \wedge_{*}\hat \omega^{ca}
-i(\hat\omega^{ab} \wedge_{*}\hat\omega+\hat\omega \wedge_{*}\hat\omega^{ab})
\nonumber \\
&-& {i\over 2}\varepsilon_{\; \; \; cd}^{ab}
\Bigr(\hat\omega^{cd}\wedge_{*}{\widetilde {\hat \omega}}
+{\widetilde {\hat\omega}} \wedge_{*}\hat\omega^{cd}\Bigr),
\label{(6)}
\end{eqnarray}
\begin{equation}
\hat r=d \hat\omega-i \left({1\over 8}\hat\omega^{ab}
\wedge_{*} \hat\omega_{ab}
+\hat\omega \wedge_{*} \hat\omega -{\widetilde {\hat\omega}} \wedge_{*}
{\widetilde {\hat\omega}} \right),
\label{(7)}
\end{equation}
\begin{equation}
{\widetilde {\hat r}}=d{\widetilde {\hat\omega}}
-i \Bigr(\hat\omega \wedge_{*}{\widetilde {\hat \omega}}
+{\widetilde {\hat\omega}}\wedge_{*}\hat\omega \Bigr)
+{i \over 16}\varepsilon_{abcd}\hat\omega^{ab} \wedge_{*}
\hat\omega^{cd}.
\label{(8)}
\end{equation}
We explicitly write in components the equations obtained by
varying the action (\ref{(1)}) with respect to
the tetrad components $\hat V^a$
and ${\widetilde {\hat V}}{}^a$, they respectively read as
\begin{eqnarray} &\;&
-\Bigr(\hat V^{d}\wedge_{*}\hat R^{ab}+\hat R^{ab}\wedge_{*}\hat V^{d}\Bigr)
\varepsilon_{abcd}+i(\eta_{bc}\eta_{ad}-\eta_{ac}\eta_{bd})
\Bigr(\hat R^{ab}\wedge_{*}{\widetilde {\hat V}}{}^{d}
-{\widetilde {\hat V}}{}^{d}
\wedge_{*}\hat R^{ab}\Bigr) \nonumber \\
& &~~+ 4i \eta_{dc}\Bigr({\widetilde {\hat r}}\wedge_{*}\hat V^{d}
-\hat V^{d}\wedge_{*}{\widetilde {\hat r}}\Bigr)
+4 \eta_{dc}\Bigr({\widetilde {\hat V}}{}^{d}\wedge_{*}{\hat r}
+{\hat r} \wedge_{*}{\widetilde {\hat V}}{}^{d}\Bigr)=0,
\label{(9)}
\end{eqnarray}
\begin{eqnarray}
&\;& -\Bigr({\widetilde {\hat V}}{}^{d}\wedge_{*}\hat R^{ab}
+\hat R^{ab}\wedge_{*}{\widetilde {\hat V}}{}^{d}\Bigr)
\varepsilon_{abcd}+i(\eta_{bc}\eta_{ad}-\eta_{ac}\eta_{bd})
\Bigr(\hat R^{ab}\wedge_{*}\hat V^{d}-\hat V^{d}
\wedge_{*}\hat R^{ab}\Bigr) \nonumber \\
& &~~+ 4i \eta_{dc}\Bigr({\widetilde {\hat r}}\wedge_{*}
{\widetilde {\hat V}}{}^{d}
-{\widetilde {\hat V}}{}^{d}\wedge_{*}{\widetilde {\hat r}}\Bigr)
+4 \eta_{dc}\Bigr(\hat V^{d}\wedge_{*} \hat r
+\hat r \wedge_{*}\hat V^{d}\Bigr)=0.
\label{(10)}
\end{eqnarray}
We notice that Eq. (10) can be obtained by the replacements $\hat V^{a}
\rightarrow {\widetilde {\hat V}}{}^{a}$ and ${\widetilde {\hat V}}{}^{b}
\rightarrow \hat V^{b}$ in Eq. (9).
The torsion 2-form is defined by
\begin{equation}
\hat T \equiv d\hat V-\Omega \wedge_{*}\hat V -\hat V \wedge_{*}\hat \Omega,
\label{(11)}
\end{equation}
and expanding $\hat T$ on the basis of $\gamma$-matrices as in
Eq. (2) we have $\hat T=\hat T^{a}\gamma_{a}
+{\widetilde {\hat T}}{}^{a}\gamma_{a}\gamma_{5}$.
Variation of the action with respect to the spin-connection gives
the remaining field equation $\left \{ \hat T \wedge_{*}\hat V-\hat V
\wedge_{*}\hat T,\gamma_{5} \right \}=0$, where we use curly brackets
for anticommutators. Since $\hat T \wedge_{*}\hat V$ is even in the
$\gamma$-matrices we eventually obtain
\begin{equation}
\hat T \wedge_{*}\hat V-\hat V \wedge_{*}\hat T=0, \label{(12)}
\end{equation}
which is satisfied by a vanishing torsion.
{\vskip.4cm}
The work in Ref. 3 has studied the
Seiberg--Witten map$^{4}$ which relates non-commutative
degrees of freedom for spin-connection, tetrad and gauge
parameter to their commutative counterparts. The relation is such that
non-commutative gauge transformations
(with non-commutative gauge parameter $\hat\Lambda$) correspond to
commutative gauge transformations (with commutative gauge parameter
$\Lambda$). For example for the tetrad we have
\begin{equation}
{\widehat V}_{\mu}+\delta_{\widehat \Lambda}{\widehat V}_{\mu}
={\widehat V}_{\mu}(V+\delta_{\Lambda}V,
\omega+\delta_{\Lambda}\omega),
\label{(17)}
\end{equation}
where $\hat V=\hat V_\mu dx^\mu$ is the non-commutative tetrad
1-form, while $V=V^a\gamma_a=V^a_\mu dx^\mu\gamma_a$ and
$\omega={1\over 4}\omega^{ab}\gamma_{ab}= {1\over
4}\omega_\mu^{ab}dx^\mu\gamma_{ab}$ are the usual commutative
tetrad and spin-connection. The non-commutativity we consider
is given by the Moyal--Weyl
$\star$-product associated with a constant antisymmetric matrix
$\theta^{\rho\sigma}$ in the (not necessarily Cartesian)
coordinates $x^\mu$; we have
${x}^{\rho}\star{ x}^{\sigma}-{x}^{\sigma}\star{ x}^{\rho}=i \theta^{\rho
\sigma}$.
Equation (\ref{(17)}) can be solved
order by order in perturbation theory. The solution of Eq. (\ref{(17)})
to first order in the non-commutativity $\theta^{\rho \sigma}$ has the
general structure
\begin{equation}
{\widehat V}_{\mu}={\widehat V}_{\mu}^{(0)a} \gamma_{a}
+{\widehat V}_{\mu 5}^{(1)a}\gamma_{5}\gamma_{a}
+{\rm O}(\theta^{2}).
\label{(18)}
\end{equation}
The zeroth- and first-order terms in $\theta^{\rho\sigma}$
in Eq. (\ref{(18)}) are found to be$^{3}$
\begin{equation}
{\widehat V}_{\mu}^{(0)a}=V_{\mu}^{a},
\label{(19)}
\end{equation}
\begin{equation}
{\widehat V}_{\mu 5}^{(1)a}\equiv-{\widetilde V}_{\mu}^{a}
={1\over 4}\theta^{\lambda \sigma}\omega_{\lambda}^{eb}
\left(\partial_{\sigma}V_{\mu}^{c}-{1\over 2}
\omega_{\sigma}^{cd}V_{\mu d} \right)
\varepsilon_{ebc}^{\; \; \; \; \; a}.
\label{(20)}
\end{equation}
Similarly, on writing for the spin-connection
\begin{equation}
{\widehat \omega}_{\mu}={1\over 2}
{\widehat \omega}_{\mu}^{(0)ab}\sigma_{ab}
+{\widehat a}_{\mu}^{(1)}
+i{\widehat b}_{\mu 5}^{(1)}\gamma_{5},
\label{(21)}
\end{equation}
one finds,$^{3}$ to first order in $\theta^{\rho \sigma}$,
\begin{equation}
{\widehat \omega}_{\mu}^{(0)ab}=\omega_{\mu}^{ab},
\label{(22)}
\end{equation}
\begin{equation}
{\widehat a}_{\mu}^{(1)}\equiv -i \omega_{\mu}
=-{1\over 16}\theta^{\lambda \sigma}\omega_{\lambda}^{ab}
\Bigr(\partial_{\sigma}\omega_{\mu}^{cd}
+R_{\sigma \mu}^{cd}\Bigr)
\eta_{ac}\eta_{bd},
\label{(23)}
\end{equation}
\begin{equation}
{\widehat b}_{\mu 5}^{(1)} \equiv i {\widetilde \omega}_{\mu}
=-{1\over 32}\theta^{\lambda \sigma}\omega_{\lambda}^{ab}
\Bigr(\partial_{\sigma}\omega_{\mu}^{cd}
+R_{\sigma \mu}^{cd}\Bigr)
\varepsilon_{abcd}.
\label{(24)}
\end{equation}
In this note we consider the non-commutative fields appearing in the
action (1) and in the corresponding field equations (9) and (10) as
dependent on the commutative ones via Seiberg--Witten map. Hence
we expand equations (9) and (10) in terms of the commutative tetrad and
spin-connection. This is done to first order in $\theta^{\mu \nu}$ by inserting
the first-order Seiberg--Witten map (15), (16) and (18)--(20). We have
done so when the underlying classical geometry is the spherically
symmetric solution of the classical Einstein equations in vacuum, i.e.
Schwarzschild.
\section{Expansion to first order of the non-commutative Einstein
equations}
By virtue of the wedge-$*$ product of forms defined in Ref. 2,
for any 1-form $\alpha^{1},\beta^{1}$ and any
2-form $\gamma^{2}$, one can write (provided that
$\partial_\mu x^\rho=\delta_\mu^\rho$),
\begin{equation}
\alpha^{1} \wedge_{*}\beta^{1}=\alpha^{1} \wedge \beta^{1}
+{i\over 2}\theta^{\rho \sigma}
\Bigr(\partial_{\rho}\alpha_{\mu}^{1}\Bigr)
\Bigr(\partial_{\sigma}\beta_{\nu}^{1}\Bigr)
dx^{\mu} \wedge dx^{\nu}+{\rm O}(\theta^{2}),
\label{(25)}
\end{equation}
\begin{equation}
\alpha^{1} \wedge_{*} \gamma^{2} = \alpha^{1} \wedge \gamma^{2}
+{i\over 2}\theta^{\rho \sigma}\Bigr(\partial_{\rho}\alpha_{\mu}^{1}\Bigr)
\Bigr(\partial_{\sigma}\gamma_{\nu \lambda}^{2}\Bigr)
dx^{\mu} \wedge dx^{\nu} \wedge dx^{\lambda}
+{\rm O}(\theta^{2}).
\label{(26)}
\end{equation}
Eq. (\ref{(26)}) can be applied repeatedly to the $\theta$-expansion of Eqs.
(9) and (10). For this purpose we need from Eq. (\ref{(26)}) the identities
\begin{equation}
\alpha^{1} \wedge_{*}\gamma^{2}+\gamma^{2}\wedge_{*}\alpha^{1}
=2 \alpha^{1} \wedge \gamma^{2}+{\rm O}(\theta^{2}),
\label{(27)}
\end{equation}
\begin{equation}
\alpha^{1}\wedge_{*}\gamma^{2}-\gamma^{2}\wedge_{*}\alpha^{1}
=i \theta^{\rho \sigma}\Bigr(\partial_{\rho}\alpha_{\mu}^{1}\Bigr)
\Bigr(\partial_{\sigma}\gamma_{\nu \lambda}^{2}\Bigr)
dx^{\mu} \wedge dx^{\nu} \wedge dx^{\lambda}+{\rm O}(\theta^{2}).
\label{(28)}
\end{equation}
Moreover, from the work in Ref. 2 we know that in non-commutative
field theory charge-conjugation conditions imply that
\begin{equation}
{\widetilde {\hat V}}{}^{a}(\theta)=-{\widetilde {\hat V}}{}^{a}(-\theta), \;
\hat \omega(\theta)=-\hat \omega(-\theta), \;
{\widetilde {\hat \omega}}(\theta)=-{\widetilde {\hat \omega}}(-\theta),
\label{(29)}
\end{equation}
and hence all non-commutative fields that are not present in the
commutative case are at least proportional to $\theta$ (and hence
vanish in the commutative limit),
\begin{equation}
{\widetilde{\hat V}}_{\mu}{}^{a}={\rm O}(\theta), \;
\omega={\rm O}(\theta), \;
{\widetilde {\hat \omega}}={\rm O}(\theta), \;
\hat r={\rm O}(\theta), \;
{\widetilde {\hat r}}={\rm O}(\theta).
\label{(30)}
\end{equation}
By virtue of (\ref{(27)}), (\ref{(28)}) and (\ref{(30)}),
Eq. (\ref{(9)}) reduces to
\begin{equation}
\varepsilon_{abcd}V^{d} \wedge \Bigr(R^{(0)ab}+R^{(1)ab}\Bigr)
+{\rm O}(\theta^{2})=0,
\label{(31)}
\end{equation}
where $R^{(n)ab}$ denotes the $n$-th order part of the curvature
2-form in powers of $\theta^{\rho \sigma}$, while Eq. (\ref{(10)}) becomes
\begin{equation}
\Bigr[-\varepsilon_{abcd}{\widetilde V}_{\mu}^{d}R_{\nu \lambda}^{(0)ab}
+\theta^{\rho \sigma}\Bigr(\partial_{\rho}V_{\mu}^{d}\Bigr)
\Bigr(\partial_{\sigma}R_{dc \; \nu \lambda}^{(0)}\Bigr)
+4V_{c \mu}r_{\nu \lambda}^{(1)}\Bigr]
dx^{\mu} \wedge dx^{\nu} \wedge dx^{\lambda}
+{\rm O}(\theta^{2})=0,
\label{(32)}
\end{equation}
having defined
\begin{equation}
r_{\nu \lambda}^{(1)} \equiv \theta^{\rho \sigma}
r_{\nu \lambda \rho \sigma}^{(1)}.
\label{(33)}
\end{equation}
We note that, in Eq. (\ref{(31)}), since $V^{d}$ is the classical tetrad,
the term $\varepsilon_{abcd}V^{d} \wedge R^{(0)ab}$ vanishes for
any solution of the vacuum Einstein equations.
\section{Non-commutative Einstein equations under the first-order
Seiberg--Witten map}
Now we consider the coordinates $x^1=t, x^2=r, x^3=\vartheta, x^4=\varphi$
and the tetrad $V^{a}=V_{\mu}^{a}dx^{\mu}$ given by
\begin{equation}
V^{(t)}=\sqrt{1-{2M \over r}}dt, \;
V^{(r)}={dr \over \sqrt{1-{2M\over r}}}, \;
V^{(\vartheta)}=r d\vartheta, \;
V^{(\varphi)}=r \sin \vartheta d\varphi.
\label{(34)}
\end{equation}
This tetrad with the associated spin-connection
(see formula (34) below)
describe a Schwarzschild geometry solution of the classical
Einstein equations in first-order formalism.
{}For this tetrad we have $R_{\mu \nu}^{(1)ab}=0$,
so that Eq. (\ref{(31)}) becomes an
identity. In order to try to solve the first-order non-commutative
equations (\ref{(32)}) we consider for the first-order expression of the
non-commutative tetrad the expression given by Eq. (\ref{(20)}),
which is the first-order Seiberg--Witten map for the tetrad.
This ansatz is supported by the following argument:
{To first order in $\theta$ the non-commutative Einstein action (1) is
the same as the commutative one if the non-comutative fields are
re-expressed in terms of the commutative ones
by using the Seiberg--Witten map (indeed the non-commutative fields
satisfy the charge conjugation conditions of Ref. 2).
Therefore it is expected that the non-commutative Einstein equations
are automatically satisfied if the non-commutative fields are expressed
via Seiberg--Witten map in terms of the commutative ones.}
We then substitute Eq. (\ref{(20)}) into Eq. (\ref{(32)}), we relabel
the indices summed over, and assuming (for ease of
calculations) that only the
non-commutativity component $\theta^{23}$
(that we rename $\theta$) is non-vanishing, i.e.
\begin{equation}
\theta^{\gamma \sigma} \frac{\partial}{\partial x^\gamma}\wedge
\frac{\partial}{\partial x^\sigma}= \theta
\biggr(\frac{\partial}{\partial r}\otimes \frac{\partial}{\partial
\vartheta} -\frac{\partial}{\partial \vartheta}\otimes
\frac{\partial} {\partial r}\biggr), \label{(35)}
\end{equation}
we re-express Eq. (\ref{(32)}) in the form (with our notation $dx^1=dt,
dx^2=dr, dx^3=d\vartheta, dx^4=d\varphi$)
\begin{eqnarray}
\; & \; & \biggr \{
\varepsilon_{abcd}\varepsilon_{epq}^{\; \; \; \; \; d}
R_{\nu \lambda}^{(0)ab}\biggr[\Bigr(\partial_{3}V_{\mu}^{q}\Bigr)
\omega_{2}^{ep}-\Bigr(\partial_{2}V_{\mu}^{q}\Bigr)\omega_{3}^{ep}
+{1\over 2}V_{\mu f}\Bigr(\omega_{2}^{qf}\omega_{3}^{ep}
-\omega_{3}^{qf}\omega_{2}^{ep}\Bigr)\biggr] \nonumber \\
&~&~+ 4 \biggr[\Bigr(\partial_{2}V_{\mu}^{q}\Bigr)
\Bigr(\partial_{3}R_{qc \; \nu \lambda}^{(0)}\Bigr)
-\Bigr(\partial_{3}V_{\mu}^{q}\Bigr)
\Bigr(\partial_{2}R_{qc \; \nu \lambda}^{(0)}\Bigr)\biggr] \nonumber \\
&~&~+ 16V_{c \mu}\Bigr(r_{\nu \lambda 23}^{(1)}-r_{\nu \lambda 32}^{(1)}\Bigr)
\biggr \}dx^{\mu}\wedge dx^{\nu}\wedge dx^{\lambda}=0.
\label{(36)}
\end{eqnarray}
At this stage, the identity
\begin{equation}
\varepsilon^{abcp}\varepsilon_{defp}=-{\rm det}
\pmatrix{\delta_{d}^{a} & \delta_{d}^{b} & \delta_{d}^{c} \cr
\delta_{e}^{a} & \delta_{e}^{b} & \delta_{e}^{c} \cr
\delta_{f}^{a} & \delta_{f}^{b} & \delta_{f}^{c} \cr}
\label{(37)}
\end{equation}
can be exploited, jointly with the standard evaluation of classical
curvature 2-form and classical spin-connection 1-form, the latter
being given by
\begin{equation}
\omega_{\mu}^{ab}={1\over 2}V^{a \nu}\Bigr(V_{\nu,\mu}^{b}
-V_{\mu,\nu}^{b}\Bigr)-{1\over 2}V^{b\nu}
\Bigr(V_{\nu,\mu}^{a}-V_{\mu,\nu}^{a}\Bigr)
+{1\over 2}V^{a\nu}V^{b \sigma}\Bigr(V_{\nu,\sigma}^{c}
-V_{\sigma,\nu}^{c}\Bigr)V_{c \mu}.
\label{(38)}
\end{equation}
Hence we find in a Schwarzschild background, bearing also in mind
Eq. (\ref{(7)}), the Seiberg--Witten map for the spin-connection
Eq. (\ref{(23)}), and the definiton (\ref{(33)}),
that the left-hand side of Eq. (\ref{(36)}) takes the form
$$
K_{c123}dx^{1}\wedge dx^{2} \wedge dx^{3}
+K_{c124}dx^{1}\wedge dx^{2} \wedge dx^{4}
+K_{c134}dx^{1}\wedge dx^{3} \wedge dx^{4}
+K_{c234}dx^{2}\wedge dx^{3} \wedge dx^{4},
$$
where each $K_{c \mu \nu \lambda}$ can be written as the sum of
$6$ terms. We obtain, after a lengthy calculation,
the simple formulae
\begin{equation}
K_{c123}={4M(5M-2r)\over r^{4}\sqrt{1-{2M \over r}}}\delta_{c1},
\label{(39)}
\end{equation}
\begin{equation}
K_{c124}=K_{c134}=0, \; \forall c=1,2,3,4,
\label{(40)}
\end{equation}
\begin{equation}
K_{c234}=4M {\sin \vartheta \over r^{2}}\delta_{c4}.
\label{(41)}
\end{equation}
Since $K_{1123}$ and $K_{4234}$ are non-vanishing, the
field configurations given in Eq. (\ref{(20)}), (\ref{(23)}) and
obtained by applying the Seiberg--Witten
map to the classical tetrad of Eq. (\ref{(34)}) and spin-connection of
Eq. (\ref{(38)}), are not solutions of the
non-commutative Einstein equations.
In order to search for solutions to non-commutative Einstein
equations that in the
commutative limit become Schwarzschild we have therefore to revert to
Eq. (\ref{(32)}), where no use of the Seiberg--Witten map is made, and look
for solutions of Eq. (\ref{(32)}) and also of the torsion
constraints in Eq. (\ref{(12)}).
\vskip.4cm
In conclusion, the performed calculation shows that there
is a mismatch between: (I) using the Seiberg--Witten map in the
non-commutative action (1) in order to express all non-commutative
fields in terms of the commutative tetrad and spin-connection, and
then solving the action (that in general will be a higher derivative
action) by varying with
respect to only the classical fields.\\
(II) obtaining the
non-commutative field equations by varying the action (1) with respect
to all non-commutative fields, and then trying to solve these equations by
expressing the non-commutative fields in terms of the commutative
ones via Seiberg--Witten map.
This mismatch can be due to the fact that in case (II), in order to obtain
the equations of motion we have to vary also with respect to the extra
fields ${\widetilde {\hat V}}$, ${\widetilde {\hat \omega}}$ and $\hat
\omega$.
Exactly these corresponding extra equations of motion are not
satisfied by considering the field configurations
${\widetilde {\hat V}}$, ${\widetilde {\hat \omega}}$ and $\hat \omega$
obtained by the Seiberg--Witten map with
$V^a$ and $\omega^{ab}$ the classical
black hole tetrad and spin-connection.
\vskip.4cm
We also notice that we have chosen the non-commutativity directions
$\frac{\partial}{\partial r}$ and $\frac{\partial}{\partial \vartheta}$
not to be Killing vector fields for our
classical black hole solution. This was done on purpose because,
similarly to Refs.$^{[5,6,7]}$, it is possible to show that,
when non-commutativity is (in part) obtained
by using Killing vector fields of a given classical solution to
Einstein equations, then this classical solution is also
a solution of the non-commutative field equations, where all
the extra non-commutative fields are taken to vanish.
\section*{Acknowledgments}
G. Esposito is grateful to the Dipartimento di Scienze Fisiche of
Federico II University, Naples, for hospitality and support; he dedicates
to Maria Gabriella his contribution to this work.
We are grateful to P. Vitale for scientific conversations.
|
\section{Introduction}
Root category was first introduced by D. Happel~\cite{Happel1987} for finite-dimensional
hereditary algebra, which was used to characterize a bijection
between the indecomposable objects of the root category for the path
algebra of Dynkin type and the root system of corresponding complex simple
Lie algebra.
Let $A$ be a finite-dimensional hereditary algebra over a field $k$.
Let $\der^b(\mod A)$ be the derived category of finitely generated
right $A$-modules. Then the root category $\cR_A$ of $A$ is defined
to be the $2$-periodic orbit category $\der^b(\mod A)/\Sigma^2$,
where $\Sigma$ is the suspension functor. It was proved by
Peng-Xiao~\cite{Peng-Xiao1997}, the root category $\cR_A$ is
triangulated via the
homotopy category of $2$-periodic complexes category of
$A$-modules. With this triangle structure, Peng and
Xiao~\cite{Peng-Xiao2000} constructed a so called Ringel-Hall Lie algebra
associated to each root category and realized all the symmetrizable
Kac-Moody Lie algebras. In fact, Peng-Xiao's construction is valid for
any $\Hom$-finite $2$-periodic triangulated category. In ~\cite{Lin-Peng2005},
Lin-Peng realized the elliptic Lie algebras of type $D_4^{(1,1)},
E_6^{(1,1)},E_7^{(1,1)},E_8^{(1,1)}$ via the $2$-periodic orbit categories
(which are triangulated) of corresponding tubular
algebras. However, in general, for arbitrary finite-dimensional $k$-algebra
$A$, the $2$-periodic orbit category $\der^b(\mod A)/\Sigma^2$ is
not triangulated with the inherited triangle structure from
$\der^b(\mod A)$({\it cf.} section ~\ref{e:minimal} or
~\cite{Keller2005} ).
Up to now, there are no suitable $\Hom$-finite $2$-periodic
triangulated categories to realize the other elliptic Lie algebras
via the Ringel-Hall Lie algebras approach.
Let $A$ be a finite-dimensional $k$-algebra with finite global
dimension. In ~\cite{Xiao-Xu-Zhang2006}, the authors propose to
study the homotopy category $\ck_2(\cp)$ of
$2$-periodic complexes category of finitely generated projective
$A$-modules and give a geometric construction of a Lie algebra over $\mathbb{C}$ directly instead
of over finite fields in \cite{Peng-Xiao2000}. In this paper, we
propose to study another $2$-periodic triangulated category $\cR_A$
called the root category of $A$ via Keller's
construction~\cite{Keller2005}. Then Peng-Xiao's construction~\cite{Peng-Xiao2000} gives a Lie
algebra $\ch(\cR_A)$ for arbitrary finite-dimensional $k$-algebra $A$ with finite global dimension.
We remark that the root category $\cR_A$ is invariant up to derived equivalence.
Note that by the $2$-universal property of root category, we
have an embedding $\cR_A\hookrightarrow \ck_2(\cp)$. When the algebra $A$ is hereditary,
$\cR_A\cong \ck_2(\cp)\cong \der^b(\mod A)/\Sigma^2$ and coincides with the original
definition of Happel.
We also remark that by using
$\cR_A$, one can easily construct $2$-periodic triangulated
categories such that the Grothendieck groups of these categories
characterize the root lattices for any elliptic Lie algebras ({\it cf.} section~\ref{elliptic
algebra}). This is one of the motivations to introduce the root category in this paper. The relation between the Ringel-Hall Lie algebras of these
categories and the corresponding elliptic Lie algebras would be interesting to study in future.
This paper is organized as follows: in section $2$, for
any finite-dimensional $k$-algebra
$A$ of finite global dimension, we introduce the root
category $\cR_A$ and study its basic properties.
It is $\Hom$-finite $2$-periodic triangulated category
and admits AR-triangles.
We also give an explicitly characterization of its Grothendieck group.
In section $3$, we study some motivating examples. In
particular, we give a minimal example such that the
$2$-periodic orbit category is not triangulated with the inherited triangle structure.
In section $4$, we consider the root categories
of representation-finite hereditary algebras, we show that
such root categories characterize the algebras up to derived
equivalence.
In the last section, we study the Ringel-Hall Lie
algebras
of a class of finite-dimensional $k$-algebras with global dimension $2$,
which turn out to give a negative answer for a question on GIM Lie algebra by
Slodowy. Let us mention that different counterexamples have been
discovered in \cite{Alpen1984} by using different approach.
In the appendix, we discuss the universal property of root category and study
recollement associated to root categories, which can be use to
construct various examples inductively such that the $2$-periodic orbit
category is not triangulated with the inherited triangle
structure from the bounded derived category.
Throughout this paper, we fix a field $k$ . All algebras are
finite-dimensional $k$-algebras with finite global dimension. All
modules are right modules. Let $\cc$ be a $k$-category. For any
$X,Y\in \cc$, we write $\cc(X,Y)$ for $\Hom_{\cc}(X,Y)$. For a
subcategory $\cm$ in a triangulated category $\ct$, we denote by
$\tria(\cm)$ the thick subcategory of $\ct$ contains $\cm$.
{\bf Acknowledgments.} I deeply thank my supervisor Liangang Peng for
his guidance and generous patience. Many thanks go to Bernhard Keller for kindly answering to my various questions and for his encouragement.
I would also like to thank Dong Yang for interesting and useful comments.
\section{root categories for finite-dimensional algebras}
\subsection{2-periodic orbit categories} Let $A$ be a finite-dimensional
$k$-algebra of finite global dimension. Let $\der^b(\mod A)$ be the
bounded derived category of finitely generated $A$-modules and
$\Sigma$ the suspension functor. Consider the left total derived
functor of $A\otimes_kA\op$-module $\Sigma^2 A$
\[\Sigma^2=?\lten_A\Sigma^2 A : \der^b(\mod A)\to \der^b(\mod A),
\]
which is an equivalence. For all $L,M$ in $\der^b(\mod A)$, the
group
\[\der^b(\mod A)(L,\Sigma^{2n}M)
\]
vanishes for all but finitely many $n\in \Z$. The {\it 2-periodic
orbit category}
\[\der^b(\mod A)/\Sigma^{2}
\]
of $A$ is defined as follows:
\begin{itemize}
\item[$\circ$] the objects are the same as those of $\der^b(\mod A)$;
\item[$\circ$] if $L$ and $M$ are in $\der^b(\mod A)$ the space of
morphisms is isomorphic to the space
\[\bigoplus_{n\in \Z}\der^b(\mod A)(L, \Sigma^{2n}M).
\]
\end{itemize}
The composition of morphisms is obviously. Suppose that $A$ is
hereditary. Then the orbit category is called {\it root category}
of $A$ which was first introduced by D. Happel in
\cite{Happel1987}.
A $\Hom$-finite $k$-additive triangulated category $\der$ is called
{\it $2$-periodic triangulated} if:
\begin{itemize}
\item[$\circ$] $\Sigma^2\cong \id$, where $\Sigma$ is the suspension
functor of $\der$;
\item[$\circ$] the endomorphism ring $\End(X)$ for
any indecomposable object $X$ is a finite-dimensional local
$k$-algebra.
\end{itemize}
In particular, the $2$-periodic orbit category of a hereditary
algebra $A$ is $2$-periodic triangulated with canonical triangle
structure proved by Peng-Xiao\cite{Peng-Xiao1997}. However, this is
not true in general. The first example is due to A.Neeman who
considers the algebra $A$ of dual numbers $k[X]/(X^2)$. Then the
$2$-periodic orbit category of $A$ is not triangulated ({\it cf.}
section $3$ of ~\cite{Keller2005}). No example of algebra with
finite global dimension seems to be known.
\subsection{Root category via Keller's construction}\label{keller's construction}
As shown by Keller in \cite{Keller2005}, if $\der^b(\mod A)$ is
triangulated equivalent to the bounded derived category of a
hereditary category, then the orbit category $\der^b(\mod
A)/\Sigma^2$ is triangulated. In general, the orbit category is not
triangulated. But a {\it triangulated hull} was defined in
~\cite{Keller2005} as the algebraic triangulated category $\cR_A$
with the following universal properties:
\begin{itemize}
\item[$\circ$] There exists an algebraic triangulated functor $\pi:\der^b(\mod A)\to
\cR_A$;
\item[$\circ$] Let $\cb$ be a dg category and $X$ an object of $\der(A^{op}\otimes
\cb)$. If there exists an isomorphism in $\der(A^{op}\otimes \cb)$
between $\Sigma^{2}A\lten_AX$ and $X$, then the triangulated
algebraic functor $?\lten_AX : \der^b(\mod A)\to \der(\cb)$
factorizes through $\pi$.
\end{itemize}
Consider $A$ as a dg
algebra concentrated in degree $0$. Let $\cs$ be the dg algebra with
underlying complex $A\oplus \Sigma A$, where the multiplication is
that of the trivial extension:
\[(a,b)(a', b')=(aa',ab'+ba').
\]
Let $\der(\cs)$ be the derived category of $\cs$ and $\der^b(\cs)$
the bounded derived category, {\it i.e.} the full subcategory of
$\der(\cs)$ formed by the dg modules whose homolgy has finite total
dimesion over $k$. Let $\per(\cs)$ be the perfect derived category
of $\cs$, {\it i.e.} the smallest subcategory of $\der(\cs)$
contains $\cs$ and stable under shift,extensions and passage to
direct factors. Clearly, the perfect derived category $\per(\cs)$ is
contained in $\der^b(\cs)$. Denote by $p: \cs\to A$ the canonical
projection. It induces a triangle functor $p_*: \der^b(\mod
A)\to \der^b(\cs)$. By composition we obtain a functor
\[\pi_A: \der^b(\mod A)\to \der^b(\cs)\to \der^b(\cs)/\per(\cs).
\]
Let $\tria(p_*A)$ be the thick subcategory of $\der^b(\cs)$
generated by the image of $p_*A$. By Theorem 2 of \cite{Keller2005},
the triangulated hull of the orbit category $\der^b(\mod
A)/\Sigma^{2}$ is the category
\[\cR_A=\tria(p_*(A))/\per(\cs).
\]
Moreover, there is an embedding $i: \der^b(\mod A)/\Sigma^{2}\hookrightarrow
<A>_{\cs}/\per(\cs)$. If $i$ is dense, then we say that the $2$-periodic orbit category $\der^b(\mod A)/\Sigma^2$ is triangulated with inherited triangle structure of $\der^b(\mod A)$.
If $A$ is an hereditary algebra, the embedding $i$ is essentially an equivalence of triangulated categories
\[\der^b(\mod A)/\Sigma^{2}\cong \tria(p_*(A))/\per(\cs).
\]
Since $\cs$ is a negative dg algebra. It is
well-known that there is a canonical $t$-structure $(\der^{\leq},
\der^{\geq})$ induced by homology over $\der(\cs)$. In particular,
$\der^{\leq}$ is the full subcategory of $\der(\cs)$ whose objects
are the dg modules $X$ such that the homology groups $H^p(X)$
vanishes for all $p>0$. Obviously, the $t$-structure restricts to
the subcategory $\der^b(\cs)$ of $\der(\cs)$. It is not hard to see
that $\der^b(\cs)=\tria(p_*A)$. Then the root category
$\cR_A=\der^b(\cs)/\per S$ in this case. In the following, we call
the triangulated hull $\cR_A$ the {\it root category} of $A$ and
$\pi_A:\der^b(\mod A)\to \der^b(\cs)/\per(\cs)=\cR_A$ the canonical
functor.
\begin{remark}
One can also consider the construction for the orbit category $\der^b(\mod
A)/\Sigma^{-2}$ which in fact the same as $\der^b(\mod A)/\Sigma^2$.
Then one replaces the dg algebra $\cs$ by $\cs'=A\oplus
\Sigma^{-3}A$. The root category defined as $\cR_A=\tria(p_*A)/\per
\cs'$.
\end{remark}
\subsection{Alternative description of $\cR_A$}\label{s:dg orbit category}
There is another description of $\cR_A$ in \cite{Keller2005}. Let $\ca$ be the dg category of bounded complexes of finitely generated projective $A$-modules. Naturally, the tensor product of $\Sigma^2 A$ define a dg functor from $\ca$ to $\ca$. Then one can form the {\it dg orbit category} $\cb$ as the dg category with the same objects of $\ca$ and such that for any $X, Y\in \cb$, we have
\[\cb(X,Y)\cong \bigoplus_{n \in \Z}\ca(X, \Sigma^{2n}Y).
\]
Now we have an equivalence of categories
\[\der^b(\mod A)/\Sigma^2\cong \ch^0(\cb).
\]
Let $\der(\cb)$ be the derived category of the dg category $\cb$.
Let the ambient triangulated category $\cm$ be the triangulated subcategory of
$\der(\cb)$ generated by the representable functors. Then theorem 2 of \cite{Keller2005} implies that $\cR_A\cong \cm$.
\begin{proposition}\label{2-periodic}
Let $A$ be a finite-dimensional $k$-algebra of finite global dimension.
Then the root category $\cR_A$ is a $\Hom$-finite $2$-periodic triangulated category.
\end{proposition}
\begin{proof}
The $\Hom$-finiteness follows from the description of $\cm$,
since the homomorphisms between representable functors of $\cb$ are finite-dimensional over $k$.
Consider the $\cb\otimes \cb\op$-module $X:$ $X(A, B)=\cb(A, B)$ for any $A, B\in \cb$, it induces the identity functor
\[\id: \der(\cb)\to \der(\cb).
\]
One can also consider the $\cb\otimes \cb\op$-module $Y(A, B)=\Sigma^2\cb(A,B)$ for any $A, B\in \cb$.
Clearly, the module $Y$ induces the triangle functor
\[\Sigma^2: \der(\cb)\to \der(\cb).
\]
By the definition of the dg orbit category $\cb$, we know that $X$ is isomorphic to $Y$ as
$\cb\otimes \cb\op$-modules, which will induce an invertible morphism
$\eta: \id\to \Sigma^2$ by Lemma 6.1 of \cite{Keller1994}.
Thus, to show that $\cR_A$ is $2$-periodic triangulated category,
it suffices to show that $\cR_A$ is Krull-Schmidt category.
It suffices to prove that each idempotent morphism of $\cR_A$ is split, {\it i.e.} $\cR_A$ is idempotent completed.
Recall that $\cR_A=\cm\subset \der(\cb)$ and $\der(\cb)$ is
idempotent complete since $\der(\cb)$ admits arbitrary direct
sums. Moreover, $\cm$ is closed under direct summands in
$\der(\cb)$. Now the result follows from the well-known fact that
if an additive category $\cc$ is idempotent completed,
then a full subcategory $\der$ of $\cc$ is idempotent completed if
and only if $\der$ is closed under direct summands.
\end{proof}
\subsection{Serre functor over $\cR_A$}
Keep the notations above. Let $D=\Hom_k(?, k)$ be the usual duality
over $k$. The
$\cs\otimes_k\cs\op$-module $D\cs$ induces a triangle functor
\[?\lten_{\cs}D\cs:\der(\cs)\to \der(\cs).
\]
We have the following well-known fact(see {\it e.g.} lemma 1.2.1 of
\cite{Amiotthesis}).
\begin{lemma}\label{l:non-degenerate}
There is a non-degenerate bilinear form
\[\alpha_{X,Y}:\der(\cs)(X, Y)\times \der(\cs)(Y, X\lten_{\cs}D\cs)\to k
\]
which is bifunctorial for $X\in \per(\cs)$ and $Y\in \der^b(\cs)$.
\end{lemma}
\begin{proposition}\label{p:auto-equivalence}
The functor $?\lten_{\cs}D\cs$ restricts to auto-equivalences
\[?\lten_{\cs}D\cs: \der^b(\cs)\to \der^b(\cs), ?\lten_{\cs}
D\cs:\per(\cs)\to \per(\cs).
\]
\end{proposition}
\begin{proof}
Since $A$ has finite global dimension, we know that $DA\in \per A$.
One can easily deduce that $D\cs\in \per \cs$. Similarly, we have
$\cs\in \tria(D\cs)\subseteq \der(\cs)$. This particular implies
that $D\cs$ is a small generator of $\der(\cs)$. It is not hard to
show that
\[\der(\cs)(\cs, \Sigma^n\cs)\cong \der(\cs)(D\cs,
\Sigma^nD\cs), n\in \Z.
\]
Thus by Lemma 4.2 of \cite{Keller1994}, we infer that
$?\lten_{\cs}D\cs$ is an equivalence over $\der(\cs)$. Now the
functor $?\lten_{\cs}D\cs$ restricts to $\per(\cs)$ follows from
$\tria(D\cs)=\per(\cs)$. Similarly, recall that we have
$\tria(p_*A)=\der^b(\cs)$ and $A\lten_{\cs}D\cs\cong
\Sigma^{-1}DA\in \der^b(\cs)$. Now again by the finite global
dimension of $A$, we have $A\in \tria(p_*(DA))\subseteq
\der^b(\cs)$. In particular, we have $\tria(p_*(DA))=\der^b(\cs)$.
Thus, $?\lten_{\cs}D\cs$ restricts to an equivalence
$?\lten_{\cs}D\cs:\der^b(\cs)\to \der^b(\cs)$.
\end{proof}
Before going to state the next result, we recall Amiot's
construction~\cite{Amiot2008} of bilinear form for quotient
category. Let $\ct$ be a triangulated category and $\cn\subset \ct$
a thick subcategory of $\ct$. Assume $\nu$ is an auto-equivalence of
$\ct$ such that $\nu(\cn)\subset \cn$. Moreover we assume that there
is a non degenerate bilinear form:
\[\beta_{N,X}:\ct(N,X)\times \ct(X,\nu N)\to k
\]
which is bifunctorial in $N\in \cn$ and $X\in \ct$. Let $X,Y\in
\ct$. A morphism $p: N\to X$ is called a {\it local $\cn$-cover of
$X$ relative to $Y$} if $N$ is in $\cn$ and it induces an exact
sequence:
\[0\to \ct(X,Y)\xrightarrow{p^*} \ct(N,Y).
\]
The following theorem is due to Amiot~\cite{Amiot2008}.
\begin{theorem}\label{t:Amiot}
\begin{itemize}
\item[1)] The bilinear form $\beta$ naturally induces a bilinear
form:
\[\beta'_{X,Y}:\ct/\cn(X,Y)\times \ct/\cn
(Y,\nu \Sigma^{-1}X)\to k\] which is also bifunctorial for $X,Y\in
\ct/\cn$;
\item[2)] Assume further $\ct$ is $\Hom$-finite. If there exists a
local $\cn$-cover of $X$ relative to $Y$ and a local $\cn$-cover of
$\nu Y$ relative to $X$, then the bilinear form $\beta'_{X,Y}$ is
non-degenerate.
\end{itemize}
\end{theorem}
Recall that $\cR_A=\der^b(\cs)/\per(\cs)$. Now we have the
following
\begin{proposition}\label{non-degenerate}
\begin{itemize}
\item[1)] The bilinear form $\alpha$ induces a bifunctorial bilinear
form $\alpha'$:
\[\alpha'_{X,Y}:\cR_A(X, Y)\times
\cR_A(Y,\Sigma^{-1}X\lten_{\cs}D\cs)\to k;
\]
\item[2)] The bilinear form $\alpha'$ is non-degenerate over
$\cR_A$.
\end{itemize}
\end{proposition}
\begin{proof}
The first statement follows form lemma~\ref{l:non-degenerate},
proposition~\ref{p:auto-equivalence} and theorem~\ref{t:Amiot}
directly.
Let $P_A=\tot(\cdots\to \Sigma^{n}\cs\to\Sigma^{n-1}\cs\to \cdots\to
\Sigma^2\cs\to \Sigma \cs\to \cs\to 0\to \cdots)$, {\it i.e.} $P_A$
is the projective resolution of $\cs$-module $A$. Then one can
easily see that $\der^b(\cs)(A, \Sigma^m A)$ is finite dimension
over $k$ for any $m\in \Z$. In particular, we have
\[\der^b(\cs)(A, \Sigma^{2m}A)\cong A \ \ \text{and }\
\der^b(\cs)(A, \Sigma^{2m+1}A)=0,
\]
for $m\geq 0$ and $\der^b(\cs)(A,\Sigma^mA)=0$ for $m<0$. Since
$p_*A$ generates the category $\der^b(\cs)$, which implies that
$\der^b(\cs)$ is $\Hom$-finite, {\it i.e.} for any $X, Y\in
\der^b(\cs)$, we have $\dim_k\der^b(\cs)(X, Y)<\infty$. Since
the non-degeneracy is extension closed,
it suffices to show that $\alpha'_{\Sigma^nA, \Sigma^mA}$ is
non-degenerate. Equivalently, it suffices to show that $\alpha'_{A,
\Sigma^nA}$ is non-degenerate for any $n\in \Z$. By Theorem
~\ref{t:Amiot} $2)$, it suffices to show that there exists a local
$\per \cs$-cover of $A$ relative to $\Sigma^nA$ and a local $\per
\cs$-cover of $\Sigma^nA$ relative to $A\lten_{\cs}D\cs$. For $n<0$,
since $\der^b(\cs)(A,\Sigma^nA)=0$, one can take $p:\cs\to A$ be the
local $\per \cs$ of $A$ relative to $\Sigma^nA$.
Suppose that $n\geq 0$. Let
\[P_{A,\Sigma^nA}:=\tot(\cdots\to 0\to \Sigma^n\cs\to \Sigma^{n-1}\cs\to \cdots\to \Sigma \cs\to \cs\to 0\to
\cdots).
\]
Clearly $P_{A,\Sigma^nA}\in \per \cs$. One can easily to see that
$p: P_{A,\Sigma^nA}\to A$ is a local $\per \cs$-cover of $A$
relative to $\Sigma^nA$. Note that $A\lten_{\cs}D\cs\cong
\Sigma^{-1}DA$. A local $\per \cs$-cover of $\Sigma^n A$ relative to
$\Sigma^{-1}DA$ is equivalent to a local $\per \cs$-cover of $A$
relative to $\Sigma^{-n-1}DA$. If $n\geq 0$, one can easily show
that $\der^b(\cs)(A, \Sigma^{-n-1}DA)=0$. Suppose that $n< 0$. One
can show that
\[P_{A,\Sigma^{-n-1}DA}:=\tot(\cdots\to 0\to \Sigma^{-n-1}\cs\to \Sigma^{-n-2}\cs\to \cdots\to \Sigma \cs\to \cs\to 0\to
\cdots)\to A
\]
is a local $\per \cs$-cover of $A$ relative to $\Sigma^{-n-1}DA$.
\end{proof}
\begin{theorem}\label{t:Serre functor}
The root category $\cR_A$ admits Auslander-Reiten triangles.
\end{theorem}
\begin{proof}
By proposition \ref{non-degenerate}, we know that
$\Sigma^{-1}?\lten_{\cs}D\cs$ is the Serre functor of
$\cR_A=\der^b(\cs)/\per \cs$. Now the result follows form Reiten-Van
den Bergh's result in \cite{Reiten-Van den Bergh2002}.
\end{proof}
\subsection{The Grothendieck group of $\cR_A$}\label{s:grothendieck
group}
Suppose that the algebra $A$ has $n$ non-isomorphic simple modules,
say $S_1, \cdots. S_n$ and $P_1, \cdots, P_n$ the corresponding
projective covers. Since $\cs$ be the trivial extension of $A$ with
non-standard gradation, $S_1, \cdots, S_n$ are also non-isomorphic
simple modules for $\cs$. By the existence of $t$-structure over
$\der^b(\cs)$, it is not hard to see that the Grothendieck group
$\go(\der^b(\cs))$ of $\der^b(\cs) $ is isomorphic to
$\Z[S_1]+\cdots+\Z[S_n]$. Indeed, consider the inclusion algebra
homomorphism $i: A\to \cs$, which induces a triangle functor $i_*:
\der^b(\cs)\to \der^b(\mod A)$. Compose with the functor
$p_*:\der^b(\mod A)\to \der^b(\cs)$, the result follows from that
$\go(\der^b(\mod A))\cong \Z^n$.
We have the following exact
sequence of triangulated categories
\[\per \cs\rightarrowtail \der^b(\cs)\twoheadrightarrow \cR_A,
\]
which induces an exact sequence of Grothendieck groups
\[\go(\per \cs)\xrightarrow{i^*} \go(\der^b(\cs))\xrightarrow{\phi}\go(\cR_A)\to 0.
\]
In particular, we have $\go(\cR_A)\cong \go(\der^b(\cs))/\im i_*$.
Let $\wt{P_i}=P_i\oplus \Sigma P_i$. It is not hard to see that
$\wt{P_i}$ are all the indecomposable projective objects in
$\der(\cs)$ up to shifts. Note that $\cs$ is negative, as remark in
\cite{Keller1994}, each compact object is an extension of direct sum
of $\Sigma^n\wt{P_i}, n\in \Z$. In particular, in the Grothendieck
group $\go(\per \cs)$, for any $X\in \per \cs$, $[X]$ is a finite
sum of $[\wt{P_i}], i=1, \cdots, n$. It is easy to see that
$i^*([\wt{P_i}])=0\in \go(\der^b(\cs))$. Thus, the image of $i^*$ is
zero and the induced linear map $\phi: \go(\der^b(\cs))\to
\go(\cR_A)$ is an isomorphism. Compose $\phi$ with
$p^*:\go(\der^b(\mod A))\to \go(\der^b(\cs))$ induced by
$p_*:\der^b(\mod A)\to \der^b(\cs)$, which is exactly the induced
map $\pi_A^*:\go(\der^b(\mod A))\to \go(\cR_A)$ by the canonical
functor $\pi_A:\der^b(\mod A)\to \cR_A$. In particular, $\pi_A^*$ is
an linear isomorphism.
Define the Euler bilinear form $\chi_{\cR_A}(-,-)$on $\go(\cR_A)$
by
\[\chi_{\cR_A}([X], [Y])=\dim_k \cR_A(X, Y)-\dim_k\cR_A(X, \Sigma Y)
\]
for any $X, Y\in \cR_A$. Clearly, it is well-defined due to the
$2$-periodic property of $\cR_A$. Let $\chi_A(-,-)$ be the Euler
bilinear form over $\der^b(\mod A)$, {\it i.e.}
\[\chi_A([X],[Y])=\sum_{i\in \Z} (-1)^i\dim_k \der^b(\mod A)(X, \Sigma^i Y)
\]
for any $X, Y\in \der^b(\mod A)$. Since $\go(\cR_A)$ is generated by
$[\pi_AS_i], i=1, \cdots, n$, where $S_i$ are simple $A$-modules, we
have $\chi_{\cR_A}([\pi_AS_i], [\pi_AS_j])=\chi_A([S_i], [S_j])$ for
any $1\leq i,j\leq n$. Thus the symmetric bilinear form $(-|-)$ over
$\go(\cR_A)$ given by
\[([X]|[Y])=\chi([X],[Y])+\chi([Y],[X])
\]
is the same over $\go(\der^b(\mod A))$ via the isomorphism $\pi_A^*$.
In particular, we have proved the following
\begin{proposition}
The canonical functor $\pi_A:\der^b(\mod A)\to \cR_A$ induces an
isometry $\pi_A^*:\go(\der^b(\mod A))\to \go(\cR_A)$, {\it i.e.} a
linear map such that $\chi_{A}(x,y)=\chi_{\cR_A}(\pi_A^*x,\pi_A^*y)$
for any $x,y\in \go(\der^b(\mod A))$.
\end{proposition}
\begin{remark}
If $A$ is finite-dimensional hereditary $k$-algebra. We have
$\cR_A\cong \der^b(\mod A)/\Sigma^2$, this shows that
$\go(\der^b(\mod )/\Sigma^2)\cong \go(\der^b(\mod A))$. Let $\pi:
\der^b(\mod A)\to \cR_A$ which is dense in this case. Let
$\ol{\go(\cR_A)}$ be the Grothendieck group of $\cR_A$ induced by
the triangles of image $\pi$. We have $\go(\cR_A)\cong
\ol{\go(\cR_A)}$.
\end{remark}
Let $c_{\cR_A}$ be the automorphism of $\go(\cR_A)$ induced by the
Auslander-Reiten translation of $\cR_A$. Let $c_{A}$ be the
automorphism of $\go(\der^b(\mod A))$ induced by the AR translation
of $\der^b(\mod A)$.
\begin{proposition}\label{coexter}
$c_A$ identifies with $c_{\cR_A}$ via the isomorphism $\pi_A^*: \go(\der^b(\mod A))\to \go(\cR_A)$.
\end{proposition}
\begin{proof}
Let $P_i$, $i=1, \cdots, n$ be the non-isomorphic indecomposable
projective modules of $A$. Since $A$ has finite global dimension, we
know that $[P_i]$, $i=1, \cdots, n$ also form a
basis of $\go(\der^b(\mod A))$. Via the isomorphism $\pi_A^*$, $[\pi_A P_i]$ is also a basis of $\go(\cR_A)$.
Thus, it suffices to show that $c_{\cR_A}(\pi_A P_i)$ coincides with $\pi_A^*c_{A}(P_i)$, $i=1, \cdots, n$. By the definition of $c_{\cR_A}$ and $c_A$, we have
\[c_{A}([P_i])=[\Sigma^{-1}P_i\lten_{A}DA], \ c_{\cR_A}([\pi_AP_i])=[\Sigma^{-2} P_i\lten_{\cs}D\cs],
\]
One can easily check that $[\Sigma^{-1}P_i\lten_{A}DA]=[\Sigma^{-2} P_i\lten_{\cs}D\cs]$ in $\go(\cR_A)$.
\end{proof}
\section{Motivating Examples}
\subsection{14 exceptional unimodular singularities}
Inspired by the theory that the universal
deformation and simultaneous resolution of a simple singularity are
described by the corresponding simple Lie
algebras~\cite{Brieskorn1970}, K. Saito associated in
\cite{Saito1985}, a generalization of root system to any regular
weight systems \cite{Saito1987}, and asks to construct a suitable
Lie theory in order to reconstruct the primitive forms for the
singularities. This is well-done for simple singularities and simple
elliptic singularities. But, in general, it is not clear how to
construct a suitable Lie theory even for $14$ exceptional unimodular
singularities.
Based on the duality theory of
the weight systems \cite{Saito1998} and the homological mirror
symmetry, Kajiura-Saito-Takahashi~\cite{Kajiura-Saito-Takahashi2005}
(Takahashi~\cite{Takahashi2005}) propose to study
the triangulated
category $HMF_A^{gr}(f_W)$ of matrix factorizations of the
homogenous polynomial $f_W$ associated to a simple singularity $W$,
then the root system appears as the set of the isomorphism classes
of the exceptional objects via the Grothendieck group of the
triangulated category. This approach has been generalized to the
case of regular weight systems with smallest exponent $\epsilon=-1$
in \cite{Kajiura-Saito-Takahashi2007} which includes the 14
exceptional unimodular singularities. In
\cite{Kajiura-Saito-Takahashi2005}, the authors show that the
category $HMF_A^{gr}(f_W)$ is triangulated equivalent to the
bounded derived category of finitely generated modules over the path
algebra of the corresponding $ADE$-type. In
\cite{Kajiura-Saito-Takahashi2007}, they show that the category
$HMF_A^{gr}(f_W)$ is triangulated equivalent to the bounded derived
category of finitely generated modules of certain finite-dimensional
algebras $A_W$. Moreover, the Grothendieck groups of these
triangulated categories characterize the strange duality for the 14
exceptional unimodular singularities. In his survey article
\cite{Saito2008}, K. Saito proposes three methods to construct Lie
algebras for each exceptional singularity and asks which Lie
algebra satisfies some extra requirements, for more details see
\cite{Saito2008}:
\begin{itemize}
\item[i)] the Lie algebra defined by the Chevalley generators and
generalized Serre relations for the Cartan matrix associated to
algebra $A_W$;
\item[ii)] the Lie subalgebra comes form vertex operator algebra for
the Grothendieck group $K_0(\cd^b(\mod A_W))$;
\item[iii)] the algebra constructed by Ringel-Hall construction for
the derived category of $\cd^b(\mod A_W)$.
\end{itemize}
We remark that for the simple singularities which are self-dual, if
one consider the Lie algebra iii) in the sense of Peng-Xiao
\cite{Peng-Xiao2000} for the root category, then these three Lie
algebras are isomorphic to each other. For the case of
$\epsilon=-1$, we remark that the algebra $A_W$ has global dimension
$2$ and it is not derived equivalent to any hereditary category.
It is not clear that whether the $2$-periodic orbit
category $\cd^b(\mod A_W)/\Sigma^2$ is triangulated or not. Now the
root category $\cR_{A_W}$ seems to be a suitable consideration for
the Lie algebra $iii)$. Then Peng-Xiao's Theorem
\cite{Peng-Xiao2000}(see also \cite{XuFan2009} ) implies that there
is a Lie algebra $\ch(\cR_{A_W})$ associated with $\cR_{A_W}$. We
remark that we do not know whether the Grothendieck group of $\cR_A$
is proper or not. In this case, the automorphism $c_{\cR_{A_W}}$ has
finite order $h_W$, where $h_W$ is the order of the Milnor monodromy
of the corresponding singularity $W$.
\subsection{An algebra of global dimension 2 }\label{elliptic algebra}
Let $Q$ be the following quiver
\[\xymatrix{3\ar@/^/[r]^{\alpha_1}\ar@/_/[r]_{\alpha_2}&2\ar@/^/[r]^{\beta_1}\ar@/_/[r]_{\beta_2}&1.}
\]
Let $I$ be the ideal generated by the relations $\beta_i\circ\alpha_i=0$. Let $A=kQ/I$ be the quotient algebra. The global
dimension of $A$ is $2$. Let $S_i, i=1,2,3$, be the non-isomorphic
simple modules of $A$. Consider the Euler symmetric bilinear form of
$\go(\cR_A)$ given by
\[([X]|[Y])=\chi(X,Y)+\chi(Y,X),
\]
for any $X, Y\in \cR_A$.
One can check that $([S_1]|[S_2])=-2, ([S_1]|[S_3])=2,
([S_2]|[S_3])=-2$. The bilinear form $(-|-)_{\go(\cR_A)}$ is
degenerate over $\go(\cR_A)$, one extends $\go(\cR_A)$ to $\cl$
such that the bilinear form $(-|-)_{\cl}$ over $\cl$ is
non-degenerate, {\it i.e.} the restriction
$(-|-)_{\cl}|_{\go(\cR_A)}$ coincides with $(-|-)_{\go(\cR_A)}$.
Let $V_{\cl}$ be the lattice vertex operator algebra associated to
$\cl$. If we consider the Lie algebra $\mathfrak{g}_A$ generated by
the vertex operators $e^{\pm[S_i]}$ in the Lie algebra
$V_{L}/DV_{L}$, where $D$ is the derivative operator, then
$\mathfrak{g}_A$ is isomorphic to the elliptic
algebra~\cite{Yoshii2000} of type $A_1^{(1,1)}$ and also isomorphic
to the toroidal algebra~\cite{Rao-Moody1994} of $\mathfrak{sl}_2$.
Now consider the root category $\cR_A$ of $A$, there is a Lie algebra $\ch(\cR_A)$(Ringel-Hall Lie algebra)
associated with $\cR_A$. We would
like to know what is the relation between $\ch(\cR_A)$ and the elliptic
algebra $A_1^{(1,1)}$? At this moment, we can only show that $u_{[\Sigma S_1]}, u_{[S_2]},
u_{[S_3]}$ satisfy the GIM Lie algebra relations \cite{Slodowy1986}.
We also remark
that up to now, there is not any triangulated category to realize
the elliptic algebras of type $A$ and $D$(except for $D_4^{(1,1)}$) via the approach of Ringel-Hall
Lie algebras. Similar to the above example, by triangular extension of algebras, for any elliptic Lie algebra,
one can construct a $2$-periodic triangulated category (possibly not unique)
such that the Grothendieck group with the symmetric Euler bilinear form characterizes the root lattice for the corresponding elliptic Lie algebra.
\subsection{A minimal example}\label{e:minimal} Let $Q$ be the following quiver with
relation
\[\xymatrix{2\ar@/^/[r]^{\alpha}&1\ar@/^/[l]^{\beta}}
\]
where $\beta\circ \alpha=0$. Let $A$ be the quotient algebra of path
algebra $kQ$ by the ideal generated by $\beta\circ \alpha$. Then $A$
is representation-finite and has global dimension $2$. Let
$\der^b(\mod A)$ be the bounded derived category of finitely
generated right $A$-modules. Let $\ca$ be the dg enhance of
$\der^b(\mod A)$, i.e. the dg category of bounded complexes of
finite generated projective $A$-modules. Let $\Sigma^2:\ca\to \ca$
be the dg enhance of the square of suspension functor of
$\der^b(\mod A)$. Let $\cb$ be the dg orbit category of $\ca$
respect to $\Sigma^2$({\it cf.} section ~\ref{s:dg orbit category}).
The canonical dg functor $\pi:\ca\to \cb$ yields a
$\cb\otimes_k\ca\op$-module
\[(B,A)\to \cb(B,\pi A),
\]
which induce the standard functors
\[\xymatrix{\der(\ca)\ar@<1ex>[r]^{\pi_*}&\der(\cb)\ar@<1ex>[l]^{\pi_{\rho}}}.
\]
Note that $\cR_A$ is the triangulated subcategory of $\der(\cb)$
generated by the representable functors. We also have a triangle
equivalence $F:\der(\Mod A)\to \der(\ca)$. Now the composition
\[\der^b(\mod A)\hookrightarrow \der(\Mod A)\xrightarrow{F}\der(\ca)\xrightarrow{\pi_*}\der(\cb)
\]
gives the canonical functor $\pi_*:\der^b(\mod A)\to \cR_A$.
\begin{proposition}The canonical function $\pi_*: \der^b(\mod A)\to
\cR_A$ is not dense.
\end{proposition}
\begin{proof}
We will construct an object in $\cR_A$ which is not in the image of
$\pi_*$. Let $S_i, i=1,2$ be the simple $A$-modules associated to
the vertices $i$ and $P_i, i=1,2$ be the corresponding
indecomposable projective modules. Let $l:P_2\to P_1$ be the
embedding and $\gamma: P_1\twoheadrightarrow S_1\hookrightarrow
P_2$. Let X be the complex $\cdots\to 0\to
P_2\xrightarrow{(l,0)}P_1\oplus P_2\xrightarrow{(0,l)^t}P_1\to
0\cdots$, where $P_1\oplus P_2$ is in the $0$-th component. Let $Y$ be
the complex $\cdots\to 0\to 0\to P_2\xrightarrow{0}P_2\to 0\cdots$,
where the left $P_2$ is in the $0$-th component. Let $f\in
\Hom_{\der^b(\mod A)}(X,Y)$ and $g\in \Hom_{\der^b(\mod A)}(X,
\Sigma^2 Y)$ be the followings
\[\xymatrix{0\ar[r]&P_2\ar[d]^{0}\ar[r]^{(l,0)}&P_1\oplus P_2\ar[d]^{(\gamma, 1)^t}\ar[r]^{(0,l)^t}&P_1\ar[d]^{\gamma}\ar[r]&0\\
0\ar[r] &0\ar[r]&P_2\ar[r]^{0}&P_2\ar[r]&0}
\]
\[\xymatrix{0\ar[r]\ar[d]^{0}&P_2\ar[d]^{1}\ar[r]^{(l,0)}&P_1\oplus P_2\ar[d]^0\ar[r]^{(0,l)^t}&P_1\ar[r]&0\\
P_2\ar[r]^0&P_2\ar[r]&0}
\]
Consider the mapping cone of $\pi_*(f+g)$ in $\cR_A$, we claim that
the mapping cone of $\pi_*(f+g)$ is not in the image of $\pi_*$.
Consider the triangle
\[\pi_*(X)\xrightarrow{\pi_*(f+g)}\pi_*(Y)\to Z\to \Sigma \pi_*(X).
\]
Applying the functor $\pi_{\rho}$, we get a triangle in $\der(\Mod
A)$
\[\pi_{\rho}\pi_*(X)\xrightarrow{\pi_{\rho}\pi_*(f+g)}\pi_{\rho}\pi_*(Y)\to\pi_{\rho}Z\to \Sigma \pi_{\rho\pi_*( X)}
\]
Note that for any $X\in \der^b{\mod A}$, we have
$\pi_{\rho}\pi_*(X)\cong \oplus_{i\in\Z}\Sigma^{2i}X$. Thus,
$\pi_{\rho}Z$ is isomorphic to the mapping cone of the following
chain map of complexes
\[\xymatrix{\cdots\ar[r]&P_1\oplus P_2\ar[d]^{(\gamma, 1)^t}\ar[r]^{\left(\begin{array}{cc}0&0\\l&0\end{array}\right)}&
P_1\oplus P_2\ar[d]^{(\gamma, 1)^t}\ar[r]^{\left(\begin{array}{cc}0&0\\l&0\end{array}\right)} &P_1\oplus P_2\ar[d]^{(\gamma,1)^t}\ar[r]^{\left(\begin{array}{cc}0&0\\l&0\end{array}\right)}&P_1\oplus P_2\ar[d]^{(\gamma,1)^t}\ar[r]&\cdots\\
\cdots\ar[r]&P_2\ar[r]^0&P_2\ar[r]^0&P_2\ar[r]^0&P_2\ar[r]&\cdots}
\]
In particular, the mapping cone is
\[\xymatrix{\cdots\ar[r] &P_2\oplus P_1\oplus P_2\ar[r]^{\left(\begin{array}{ccc}0 &0& 0\\-\gamma&0&0\\-1&-l&0\end{array}\right)}&
P_2\oplus P_1\oplus P_2\ar[r]^{\left(\begin{array}{ccc}0 &0&
0\\-\gamma&0&0\\-1&-l&0\end{array}\right)}&P_2\oplus P_1\oplus
P_2\ar[r]&\cdots}
\]
Let $h:P_1\twoheadrightarrow S_1\hookrightarrow P_1$, consider the
complex $P:\cdots\to P_1\xrightarrow{h}P_1\xrightarrow{h}P_1\to
\cdots$, one can check that
\[\xymatrix{\cdots\ar[r] &P_2\oplus P_1\oplus P_2\ar[d]^{(-l,1,0)^t}\ar[r]^{\left(\begin{array}{ccc}0 &0& 0\\-\gamma&0&0\\-1&-l&0\end{array}\right)}&
P_2\oplus P_1\oplus
P_2\ar[d]^{(-l,1,0)^t}\ar[r]^{\left(\begin{array}{ccc}0 &0&
0\\-\gamma&0&0\\-1&-l&0\end{array}\right)}&P_2\oplus P_1\oplus
P_2\ar[d]^{(-l,1,0)^t}\ar[r]&\cdots\\
\cdots\ar[r]&P_1\ar[r]^h&P_1\ar[r]^h&P_1\ar[r]^h&\cdots}
\]
is a quasi-isomorphism. In particular, $\pi_{\rho} Z$ is
isomorphic to $P$ in $\der(\Mod A)$. If there exists $U\in \der^b(\mod A)$ such that
$\pi_*(U)=Z$, then $\pi_{\rho}Z\cong \oplus_{i\in Z}\Sigma^{2i}U$.
But one can easily show that $P$ is indecomposable in $\der(\Mod
A)$. This completes the proof.
\end{proof}
The above example implies that in general the orbit category
$\der^b(\mod A)/\Sigma^2$ is not triangulated even with small global
dimension. In the appendix, we propose a way to construct various
examples from a known one by using recollement associates to root
categories. It would be interesting to konw that whether one can give an
example without oriented cycles such that the orbit category
$\der^b(\mod A)/\Sigma^2$ is not triangulated with inherited triangle structure.
\section{The ADE root categories}
In this section, we focus on the root categories of
finite-dimensional hereditary algebras of Dynkin type. We will show
that such root categories characterize the algebras up to derived
equivalence.
\subsection{Separation of AR-components}
Let $A$ be a finite dimensional $k$-algebra with finite global
dimension. Let $\pi_A:\der^b(\mod A)\to \cR_A$ be the canonical
triangle functor. By theorem~\ref{t:Serre functor}, we know that
$\cR_A$ has Serre functor, equivalently, Auslander-Reiten
triangles(AR-triangles). When $\pi_A$ is dense, it is quite easy to
show that $\pi_A$ preserves the AR-triangles, {\it i.e.} each
AR-triangle of $\cR_A$ comes from an AR-triangle of $\der^b(\mod A)$
via the canonical functor $\pi_A$. In particular, the AR-quiver of $\der^b(\mod A)$
determines the AR-quiver of $\cR_A$. In general, we have the
following
\begin{theorem}\label{t:separated}
Let $A$ be a finite-dimensional $k$-algebra with finite global
dimension and $\pi_A:\der^b(\mod A)\to \cR_A$ the canonical functor.
Then the functor $\pi_A$ maps AR-triangles of $\der^b(\mod A)$ to
AR-triangles of $\cR_A$. As a consequence, there is no irreducible
morphism between $\im \pi_A$ and $\cR_A\backslash \im \pi_A$.
\end{theorem}
\begin{proof}
Recall that for arbitrary objects $X, Y\in \der^b(\mod A)$, we have
canonical isomorphism
\[\cR_A(\pi_A(X), \pi_A(Y))\cong \bigoplus_{i\in \Z}\der^b(\mod
A)(\Sigma^{2i}X, Y),
\]
and $\der^b(\mod A)(\Sigma^{2i}X, Y)$ vanishes for all but finitely
many $i$. Let $S$ and $\wt{S}$ be the Serre functors of $\der^b(\mod
A)$ and $\cR_A$ respectively. Firstly, we show that $\pi_AS(X)\cong
\wt{S}\pi_A(X)$ for any indecomposable object $X\in \der^b(\mod
A)$. Consider the functor $D\cR_A(?,\pi_AS(X))$ over $\cR_A$,
where $D=\Hom_k(?,k)$ is the usual duality of $k$. We have the
following canonical isomorphism
\begin{eqnarray*}
D\cR_A(\pi_AX, \pi_AS(X))&\cong& D(\bigoplus_{i\in
\Z}\der^b(\mod A)(\Sigma^{2i}X, S(X)))\\
&\cong &\bigoplus_{i\in \Z}D\der^b(\mod A)(\Sigma^{2i}X,
S(X))\\
&\cong&\bigoplus_{i\in \Z}\der^b(\mod A)(X, \Sigma^{2i}X)\\
&\cong&\cR_A(\pi_AX, \pi_AX)
\end{eqnarray*}
The indecomposable property implies that $\cR_A(\pi_AX, \pi_AX)$ is
a local $k$-algebra. Let $\eta\in D\cR_A(\pi_AX, \pi_AS(X))$ be the
image of $1_{\pi_AX}\in \cR_A(\pi_AX, \pi_AX)$ via the canonical
isomorphism. Let $\eta^*:\cR_A(\pi_AX,?)\to D\cR_A(?,\pi_AS(X))$ be
the natural transformation corresponding to $\eta$. It is clear that
$\eta^*|_{\im \pi_A}$ is an isomorphism. Since $\cR_A$ is the
triangulated hull of $\im\pi_A$, one deduces that $\eta^*$ is an
isomorphism over $\cR_A$. In particular, $D\cR_A(?,\pi_AS(X))$ is
representable. On the other hand, the Serre functor $\wt{S}$ implies
$D\cR_A(?,\wt{S}\pi_AX)$ is also represented by $\cR_A(\pi_AX,?)$.
Thus, we have $\pi_AS(X)\cong \wt{S}\pi_AX$.
Let $\Sigma^{-1}SX\xrightarrow{f} Y\xrightarrow{g}
X\xrightarrow{h}S(X)$ be an AR-triangle of $\der^b(\mod A)$. Let
$\pi_A(\Sigma^{-1}SX)\xrightarrow{u} W\to
\pi_AX\xrightarrow{v}\pi_AS(X)$ be the AR-triangle in $\cR_A$.
Clearly, $\pi_A(f)$ is not a split monomorphism. Thus, by the
definition of AR-triangle, there is a morphism $t:W\to \pi_AY$ such
that $\pi_A(f)=t\circ u$. Namely, we have the following commutative
diagram of triangles
\[\xymatrix{\pi_A(\Sigma^{-1}SX)\ar@{=}[d]\ar[r]^u&W\ar[d]^t\ar[r]&\pi_AX\ar[d]^s\ar[r]^{v}&\pi_AS(X)\ar@{=}[d]\\
\pi_A(\Sigma^{-1}SX)\ar[r]^{\pi_A(f)}&\pi_A(Y)\ar[r]^{\pi_A(g)}&\pi_AX\ar[r]^{\pi_A(h)}&\pi_AS(X)}
\]
We claim that $s$ is an isomorphism. Suppose not, then $s$ is
nilpotent by the indecomposable of $X$. Then $\pi_A(h)\circ s=0$
follows form that $\Sigma^{-1}SX\xrightarrow{f} Y\xrightarrow{g}
X\xrightarrow{h}S(X)$ is an AR-triangle, which implies $v=0$,
contradiction. Thus, $t$ is isomorphism. In particular, the image of
$\Sigma^{-1}SX\xrightarrow{f} Y\xrightarrow{g} X\xrightarrow{h}S(X)$
is indeed an AR-triangle of $\cR_A$.
Now one can easily deduce that there is no irreducible morphism
between $\im \pi_A$ and $\cR_A\backslash \im\pi_A$, which completes
the proof.
\end{proof}
\begin{remark}
Theorem~\ref{t:separated} has been proved for the generalized
cluster category in ~\cite{AmiotOppermann2010} by using different
approach. We remark that one can adapt a variant proof to deduce the
result for generalized cluster category. Indeed, by the
$2$-Calabi-Yau property of generalized cluster category, one can
deduce that the Serre functor of the derived category coincides with
the Serre functor of the generalized cluster category on the
objects. Then one shows that the functor $\pi_A$ preserves
AR-triangles. By the universal property of root category, the Serre
functor $S:\der^b(\mod A)\to \der^b(\mod A)$ will induce a functor
$\ol{S}:\cR_A\to \cR_A$. It would be interesting to compare it with
the Serre functor $\wt{S}$.
\end{remark}
\subsection{The ADE root categories}
Let $A$ and $B$ be finite-dimensional $k$-algebras with finite
global dimension. If $A$ and $B$ are derived equivalent, it is clear
that $\cR_A\cong \cR_B$. It would be interesting to characterize all
the algebras which have the same root category up to triangle
equivanlence. In general, this question seems to be very hard. In
the following we will characterize the algebras share the root
category with a path algebra of Dynkin quiver. Since the derived
category of Dynkin quiver is not dependent on the choice of
orientation, we assume $Q$ be the following quiver for simplicity.
\[\xymatrix@R=0.3cm{A_n:&1\ar[r]&2\ar[r]&\cdots\ar[r]&n-1\ar[r]&n\\
&1\ar[rd]\\D_n:&&3\ar[r]&4\ar[r]&\cdots\ar[r]&n\\&2\ar[ru]\\
&&&3\ar[d]\\
E_6:& 1\ar[r]&2\ar[r]&4\ar[r]&5\ar[r]&6\\
&&&3\ar[d]\\
E_7:&1\ar[r]&2\ar[r]&4\ar[r]&5\ar[r]&6\ar[r]&7\\
&&&3\ar[d]\\
E_8:&1\ar[r]&2\ar[r]&4\ar[r]&5\ar[r]&6\ar[r]&7\ar[r]&8}
\]
\begin{theorem}\label{t:ADE root category}
Let $A$ be a finite-dimensional $k$-algebra with finite global
dimension. If the root category $\cR_A\cong \cR_{kQ}$ for some
Dykin quiver $Q$, then $A$ is derived equivalent to $kQ$.
\end{theorem}
\begin{proof}
Since $Q$ is finite Dykin quiver, the AR-quiver of $\der^b(\mod kQ)$
is connected. The canonical functor $\pi_{kQ}:\der^b(\mod kQ)\to
\cR_{kQ}$ is dense, which implies that the AR-quiver of $\cR_{kQ}$
is connected. By theorem~\ref{t:separated}, we inform that the
functor $\pi_A:\der^b(\mod A)\to \cR_A$ is dense. In particular, any
$X\in \cR_A$ has preimage in $\der^b(\mod A)$. Let $P_i,
i=1,\cdots,n$ be the indecomposable projective $kQ$-modules. It is
clear that
\[\dim_k\cR_{kQ}(\pi_{kQ}P_i,\pi_{kQ}P_j)\leq 1 \ \text{for}\ i\leq j \ \text{and}\
\cR_{kQ}(\pi_{kQ}P_i,\pi_{kQ}P_j)=0\ \text{for}\ i>j.
\]
Let $F:\cR_{kQ}\to \cR_A$ be the triangle equivalent functor. We
claim that there is an object $M=M_1\oplus\cdots\oplus M_n$ in
$\der^b(\mod A)$ such that \[\pi_A(M)=F(\pi_{kQ}(kQ)) \text{and}\
\der^b(\mod A)(M,\Sigma^tM)=0\ \text{for}\ t\neq 0.\]
Let
$\{\Sigma^{2r}X_i|r\in \Z\}$ be the preimages of $F(\pi_{kQ}(P_i))$
in $\der^b(\mod A)$. Let us prove this claim for case $Q=A_n$, the
other cases are similar. Note that $n$ is a sink vertex, we can
choose $M_n=X_n$. Since $\cR_A(F(\pi_{kQ}(P_{n-1})),
F(\pi_{kQ}(P_n)))\cong k$, there is a unique $r_{n-1}\in \Z$ such
that $\der^b(\mod A)(\Sigma^{2r_{n-1}}X_{n-1}, M_n)\cong k$ and
$\der^b(\mod A)(\Sigma^{2t}X_{n-1}, X_n)=0$ for $t\neq r_{n-1}$. We
can take $M_{n-1}=\Sigma^{2r_{n-1}}X_{n-1}$. Replace $M_n$ by
$M_{n-1}$, one can construct $M_{n-2}$ uniquely. For any nonzero
$f:M_{n-2}\to M_{n-1}, g:M_{n-1}\to M_n$,the composition $g\circ
f\neq 0$. Inductively, one can construct $M_i$ for any $i=1, \cdots,
n$. Clearly, we have $\pi_A(M)\cong F(\pi_{kQ}kQ)$ and $\der^b(\mod
A)(M, \Sigma^{2r}M)=0$ for $r\neq 0$. $\der^b(\mod
A)(M,\Sigma^{2r+1}M)=0, r\in Z$ follows from
$\cR_{kQ}(\pi_{kQ}kQ,\Sigma\pi_{kQ}kQ)=0$. In particular, $M$ is a
(partial) tilting complex of $\der^b(\mod A)$. We have $\der^b(\mod
kQ)\cong \der^b(\mod \End_{\der^b(\mod A)}(M))\cong \tria(M)$, where
$\tria (M)$ is the thick subcategory of $\der^b(\mod A)$ contains
$M$. If we can show that $\tria(M)=\der^b(\mod A)$, then we are
done. Let $i:\der^b(\mod kQ)\xrightarrow{\sim}
\tria(M)\hookrightarrow \der^b(\mod A)$ be the composition. By the
universal property of root category, we have the following
commutative diagram
\[\xymatrix{\der^b(\mod kQ)\ar[d]^{\pi_{kQ}}\ar@{^{(}->}[r]^i&\der^b(\mod A)\ar[d]^{\pi_A}\\
\cR_{kQ}\ar[r]^{\ol{i}}&\cR_A}
\]
where $\ol{i}$ is induced by the full embedding $i$. It is clear
that $\ol{i}$ is also full and faithful, thus an equivalence, which
implies $i$ is dense and an equivalence.
\end{proof}
\subsection{Tame quiver of type $\wt{D}$ and $\wt{E}$}
Assume $Q$ be the following quiver
\[\xymatrix@R=0.3cm{&&2\ar[d]&&&n-1\ar[d]\\
\wt{D_n}:&1\ar[r]&3\ar[r]&\cdots\ar[r]&n-2\ar[r]&n\ar[r]&n+1\\
&&&3\ar[d]\\&&&4\ar[d]\\
\wt{E_6}:&1\ar[r]&2\ar[r]&5\ar[r]&6\ar[r]&7\\
&&&&4\ar[d]\\
\wt{E_7}:&1\ar[r]&2\ar[r]&3\ar[r]&5\ar[r]&6\ar[r]&7\ar[r]&8\\
&&&3\ar[d]\\
\wt{E_8}:&1\ar[r]&2\ar[r]&4\ar[r]&5\ar[r]&6\ar[r]&7\ar[r]&8\ar[r]&9}
\]
The theorem ~\ref{t:ADE root category} also holds for tame quiver of
type $\wt{D}$ and $\wt{E}$. One can adapt a variant proof of theorem
~\ref{t:ADE root category}.
\begin{proposition}
Let $A$ be a finite dimensional $k$-algebra with finite global
dimension. If the root category $\cR_A\cong \cR_{kQ}$ for some tame
quiver $Q$ of type $\wt{D}\wt{E}$, then $A$ is derived equivalent to
$kQ$.
\end{proposition}
\begin{proof}
It suffices to prove this proposition for $Q$ be the above quiver.
Clearly the canonical functor $\pi_{kQ}:\der^b(\mod kQ)\to \cR_{kQ}$
is dense. In this case, $\der^b(\mod kQ)$ is the union of
preprojective component, preinjective component and tubes up to
shift. If the image $\im \pi_A$ intersects with preprojective (resp.
preinjective) component nonempty, by theorem~\ref{t:separated},
every object in this component belongs to $\im \pi_A$. Then one can
adapt the proof of theorem ~\ref{t:ADE root category}. Now suppose
that $\im \pi_A$ intersects with both preprojective and
preinjective component empty. Let $T$ be the union of $kQ$-modules
in the tubes. It is clear that $T$ is a hereditary abelian
subcategory of $kQ$-modules. By theorem $9.1$ of ~\cite{Keller2005},
we know that $\der^b(T)/\Sigma^2$ is triangulated and we have the
following commutative diagram
\[\xymatrix{\der^b(T)\ar[d]^{\pi}\ar@{^{(}->}[r]^i&\der^b(\mod kQ)\ar[d]^{\pi_{kQ}}\\
\der^b(T)/\Sigma^2\ar[r]^{\ol{i}}&\cR_{kQ}}
\]
where $\ol{i}$ is induced by $i$. In particular, we know that
$\ol{i}$ is a full embedding. Now $\im\pi_A\subset T\cup \Sigma T$
implies that $\im\pi_A\subset \der^b(T)/\Sigma^2$, which contradicts
to $\tria(\im\pi_A)=\cR_A$.
\end{proof}
\section{Ringel-Hall Lie algebras and GIM Lie algebras}
Throughout this section, let $k$ be a field with $|k|=q$. We study
the Ringel-Hall Lie algebras of a class of finite-dimensional
$k$-algebras with global dimension $2$. Building on the
representation theory of these algebras, we will give a negative
answer for a question on GIM-Lie algebras by Slodowy
in~\cite{Slodowy1986}. We remark that different counterexamples of
this question have been discovered by Alpen~\cite{Alpen1984} by
considering fixed point subalgebras of certain Lie algebras.
\subsection{Generalized intersection matrix Lie algebras}
We recall the generalized intersection matrix Lie algebra (GIM-Lie
algebra for short) following Slodowy~\cite{Slodowy1986}. A matrix
$A\in M_l(\Z)$ is called a {\it generalized intersection matrix}, or
GIM for short, if the followings are satisfied
\begin{eqnarray*}
&&A_{ii}=2\\
&&A_{ij}<0 \Longleftrightarrow A_{ji}<0\\
&&A_{ij}>0 \Longleftrightarrow A_{ji}>0
\end{eqnarray*}
If moreover $A$ is symmetric, then $A$ is called an {\it
intersection matrix}. Given a GIM $A\in M_l(\Z)$, a {\it root basis}
associated to $A$ is a triplet $(H,\triangledown,\vartriangle)$
consisting of
\begin{itemize}
\item[$\circ$] a finite dimensional $\Q$-vector space $H$;
\item[$\circ$] a family $\triangledown=\{\alpha_1^{\vee}, \cdots, \alpha_l^{\vee}\}$, where $\alpha_i^{\vee}\in
H$;
\item[$\circ$] a family $\vartriangle=\{\alpha_1, \cdots, \alpha_l\}$,
where $\alpha_i\in H^*=\Hom_{\Q}(H,\Q)$
\end{itemize}
satisfy the following
\begin{itemize}
\item[1)] both sets $\vartriangle$ and $\triangledown$ are linearly
independent;
\item[2)] $\alpha_j(\alpha_i^{\vee})=A_{ij}$ for all $1\leq i,j\leq l$;
\item[3)] $\dim_{\Q}H=2l-\rank A$.
\end{itemize}
The {\it GIM-Lie algebra} $\mg=GIM(A)$ attached to the root basis $(H,
\triangledown, \vartriangle)$ is given by the generators
$\mh=H\otimes_{\Q}\C$ and $e_{\pm \alpha}, \alpha\in \vartriangle$
satisfying the following relations:
\begin{eqnarray*}
&(1)&[h,h']=0, h,h'\in \mh\\
&(2)&[h,e_{\alpha}]=\alpha(h)e_{\alpha}, h\in \mh, \alpha\in \pm \vartriangle\\
&(3)&[e_{\alpha}, e_{-\alpha}]=\alpha^{\vee}, \alpha\in \vartriangle\\
&(4)&ad(e_{\alpha})^{max(1,1-\beta(\alpha^{\vee}))}e_{\beta}=0,
\alpha\in \vartriangle, \beta\in \pm\vartriangle\\
&(5)&ad(e_{-\alpha})^{max(1,1-\beta(-\alpha^{\vee}))}e_{\beta}=0,
\alpha\in \vartriangle, \beta\in \pm\vartriangle.
\end{eqnarray*}
If $A$ is a symmetrizable
generalized Cartan matrix, then the $GIM(A)$ is essentially the
Kac-Moody algebras associated to $(H, \triangledown,\vartriangle)$.
Let $ad: \mg\to \End(\mg)$ be the adjoint representation of $\mg$.
Consider the restriction of $ad$ to $\mh$, the Lie algebra $\mg$
decomposes into a direct sum
\[\mg=\bigoplus_{\gamma\in \mh^*}\mg_{\gamma}
\]
of eigenspaces
\[\mg=\{x\in \mg|[h,x]=\gamma(h)x \ \text{for all}\ h\in \mh\}.
\]
Clearly, we have $\mh\subseteq \mg_{0}$. The following question has
been addressed in ~\cite{Slodowy1986} by Slodowy: Does equality
hold?
If we consider the derived subalgebra $[\mg, \mg]$ of $\mg$, the
above question is equivalent to the following: Do we have $\dim_{\C}
[\mg, \mg]_{0}=l$? We remark that the derived subalgebra $[\mg,
\mg]$ can be presented by generators $\alpha_i^{\vee}, 1\leq i\leq
l$ and $e_{\alpha}, \alpha\in \pm\vartriangle$ with the same
relations in $\mg$. In~\cite{Alpen1984}, Alpen has given a negative
answer for this question by using Lie theory. In the following, we
will give a negative answer for this question via
representation-theoretic approach.
\subsection{The Ringel--Hall Lie algebra}
We recall the definition of the Ringel--Hall Lie algebra of a
$2$-periodic triangulated category following~\cite{Peng-Xiao2000}.
Let $\mathcal{R}$ be a Hom-finite $k$-linear triangulated category
with suspension functor $\Sigma$. By $\ind \mathcal{R}$ we denote a
set of representatives of the isoclasses of all indecomposable
objects in $\mathcal{R}$.
Given any objects $X,Y,L$ in $\mathcal{R}$, we define
\begin{eqnarray*}W(X,Y;L)&=&\{(f,g,h)\in \Hom_{\mathcal{R}}(X,L)\times \Hom_{\mathcal{R}}(L,Y)\times \Hom_{\mathcal{R}}(Y,\Sigma X)|\\
&&X\xrightarrow{f}L\xrightarrow{g} Y\xrightarrow{h}\Sigma X \text{
is a triangle}\}.
\end{eqnarray*}
The action of $\Aut(X)\times \Aut(Y)$ on $W(X,Y;L)$ induces the
orbit space
\[V(X,Y;L)=\{(f,g,h)^{\wedge}|(f,g,h)\in W(X,Y;L)\}
\]
where
\[(f,g,h)^{\wedge}=\{(af, gc^{-1}, ch(\Sigma a)^{-1})|(a,c)\in \Aut(X)\times\Aut(Y)\}.
\]
Let $\Hom_{\mathcal{R}}(X,L)_Y$ be the subset of
$\Hom_{\mathcal{R}}(X,L)$ consisting of
morphisms $l:X\to L$ whose mapping cone $Cone(l)$ is isomorphic to $Y$. Consider the action of the
group $\Aut(X)$ on $\Hom_{\mathcal{R}}(X,L)_Y$ by $d\cdot l=dl$, the orbit is denoted by $l^*$ and the
orbit space is denoted by $\Hom_{\mathcal{R}}(X,L)_Y^*$.
Dually one can also consider the subset $\Hom_{\mathcal{R}}(L,Y)_{\Sigma X}$
of $\Hom_{\mathcal{R}}(L,Y)$ with the group action $\Aut(Y)$ and the orbit
space $\Hom_{\mathcal{R}}(L,Y)_{\Sigma X}^*$. The following proposition is an observation due to
~\cite{XiaoXu2008}.
\begin{lemma}\label{l:hall-num}
$|V(X,Y;L)|=|\Hom_{\mathcal{R}}(X,L)_Y^*|=|\Hom_{\mathcal{R}}(L,Y)_{\Sigma
X}^*|$.
\end{lemma}
We assume further that $\mathcal{R}$ is \emph{$2$-periodic}, \ie
$\mathcal{R}$ is Krull--Schmidt and $\Sigma^2\cong 1$.
Let $\go(\mathcal{R})$ be the Grothendieck group of $\mathcal{R}$
and $I_{\mathcal{R}}(-,-)$ the symmetric Euler form of
$\mathcal{R}$. For an object $M$ of $\mathcal{R}$, we denote by
$[M]$ the isoclass of $M$ and by $h_M=\dimv M$ the canonical image
of $[M]$ in $\go(\mathcal{R})$. Let $\mh$ be the subgroup of
$\go(\mathcal{R})\otimes_{\Z}\Q$ generated by $\frac{h_M}{d(M)},
M\in \ind \mathcal{R}$, where $d(M)=\dim_k(\End(X)/rad \End(X))$.
One can naturally extend the symmetric Euler form to $\mh\times
\mh$. Let $\mn$ be the free abelian group with basis $\{u_X|X\in
\ind \mathcal{R}\}$. Let
\[\mg(\mathcal{R})=\mh\oplus\mn,
\]
a direct sum of $\Z$-modules. Consider the quotient group
\[\mg(\mathcal{R})_{(q-1)}=\mg(\mathcal{R})/(q-1)\mg(\mathcal{R}).
\]
Let $F_{YX}^L=|V(X,Y;L)|$. Then by Peng and Xiao
~\cite{Peng-Xiao2000} we know that $\mg(\mathcal{R})_{(q-1)}$ is a
Lie algebra over $\Z/(q-1)\Z$, called the \emph{Ringel--Hall Lie
algebra} of $\mathcal{R}$. The Lie operation is defined as follows.
\begin{itemize}
\item[(1)] for any indecomposable objects $X,Y\in \mathcal{R}$,
\[
[u_X,u_Y]=\sum_{L\in \ind
\mathcal{R}}(F_{YX}^L-F_{XY}^L)u_L-\delta_{X,\Sigma
Y}\frac{h_X}{d(X)},
\]
where $\delta_{X,\Sigma Y}=1$ for $X\cong \Sigma Y$ and $0$ else.
\item[(2)] $[\mh, \mh]=0.$
\item[(3)] for any objects $X,Y\in \mathcal{R}$ with $Y$ indecomposable,
\[[h_X,u_Y]=I_{\mathcal{R}}(h_X,h_Y)u_Y,\qquad [u_Y,
h_X]=-[h_X,u_Y].
\]
\end{itemize}
A triangulated category $\ct$ is called {\it proper}, if for any nonzero indecomposable object $X\in \ct$, $\dimv X\neq 0 $ in the Grothendieck group $\go(\ct)$. If the $2$-periodic triangulated category $\cR$ is proper, then $[u_X, u_{\Sigma X}]=-\frac{h_X}{d(X)}$, which coincides the origin definition in~\cite{Peng-Xiao2000}. However, the proof in ~\cite{Peng-Xiao2000} is still valid for non-proper $2$-periodic triangulated category for the Lie bracket defined above ({\it cf.} \cite{Xiao-Xu-Zhang2006}).
\subsection{A class of finite-dimensional $k$-algebras}
Let $Q$ be the following quiver
\[\xymatrix{&&&& 0\ar@<1ex>[d]^{\alpha}\ar@<-1ex>[d]\\&n
&n-1\ar[l]&2\ar@{.>}[l]&1\ar[r]^\beta\ar[l]^{\gamma}&n+1\ar@{.>}[r]&\circ\ar[r]&n+m}
\]
We assume $m\geq 1, n\geq 2$. Let $A$ be the quotient algebra of
path algebra $kQ$ by the ideal generated by $\beta\circ \alpha,
\gamma\circ \alpha$. It has global dimension $2$.
Let $E$ be a field extension of $k$ and set $V^E=V\otimes_kE$ for
any $k$-space $V$. Then $A^E$ is an $E$-algebra and, for $M\in \mod
A$, $M^E$ has a canonical $A^E$-module structure. Clearly, $A^E$
still has global dimension $2$. Let $\cR_{A^E}$ be the root category
of $A^E$. Thus, one has the Ringel-Hall Lie algebra
$\mg(\cR_{A^E})_{(|E|-1)}$, which is a Lie algebra over
$\Z/(|E|-1)\Z$.
Let $\ol{k}$ be the algebraic closure of $k$ and set
\[\Omega=\{E|k\subseteq E\subseteq \ol{k}\ \text{ is a finite field
extension}\}.
\]
We consider the direct product $\prod_{E\in
\Omega}\mg(\cR_{A^E})_{(|E|-1)}$ of Lie algebras and let
$\cl\mg(\cR_A)$ be the Lie subalgebra of $\prod_{E\in
\Omega}\mg(\cR_{A^E})_{(|E|-1)}$ generated by
$u_{S_i}=(u_{S_i^E})_{E\in \Omega}$ and $u_{\Sigma S_i}=(u_{\Sigma
S_i^E})_{E\in \Omega}$ for all simple $A$-modules $S_i, 0\leq i\leq
m+n$. Clearly, $h_i=(h_{S_i^E})_{E\in \Omega}, 0\leq i\leq n+m$
belong to $\cl\mg(\cR_A)$. We call $\cl\mg(\cR_A)$ the {\it integral
Ringel-Hall Lie algebra} of $A$. Clearly, the algebra
$\cl\mg(\cR_A)$ has a grading by the Grothendieck group $\go(\cR_A)$
of $\cR_A$, namely,
\[\cl\mg(\cR_A)=\bigoplus_{\alpha\in
\go(\cR_A)}\cl\mg(\cR_A)_\alpha
\]
such that $\deg u_{S_i}=\dimv S_i, \deg (u_{\Sigma S_i})=\dimv
\Sigma S_i$. In particular, $h_i\in \cl\mg(\cR_A)_0$.
Let $(-,-)$ be the symmetric Euler form of $\cR_A$ ({\it cf.}
section~\ref{s:grothendieck group}). Then image of simple
$A$-modules $[S_i], 0\leq i\leq n+m$ form a $\Z$-basis of
$\go(\cR_A)$. Define the matrix $C=(c_{ij})$, $c_{ij}=([S_i],[S_j])$
for $0\leq i, j\leq n+m$. One can easily show that $C$ is an
intersection matrix. Let $(H,\triangledown, \vartriangle)$ be a root
basis of $C$. Thus one can form the GIM Lie algebra $\mg(C)=GIM(C)$
associated to $C$. We are interested in its derived subalgebra
$\mg(C)'=[\mg(C), \mg(C)]$.
\begin{theorem}~\label{t:Lie algebra homom} There is a surjective Lie algebra homomorphism $\phi:\mg(C)'\to
\cl\mg(\cR_A)\otimes_{\Z}\C$ defined by
\begin{eqnarray*}
&&\alpha_i^{\vee}\mapsto h_i,\\
&&e_{\alpha_i}\mapsto u_{S_i}\\
&&e_{-\alpha_i}\mapsto -u_{\Sigma S_i}, 0\leq i\leq n+m.
\end{eqnarray*}
Moreover, $\phi$ keeps the gradations and $\dim_{\C}
(\cl\mg(\cR_A)\otimes_{\Z}\C)_0\geq m+n+2$. As a consequence, we infer
that $\dim_{\C}\mg(C)'_0\geq m+n+2$.
\end{theorem}
The proof of this theorem carries throughout the rest of this
section.
\begin{lemma}\label{l:dimension of 0th gradation}
Let $M$ be the unique indecomposable $A$-module with composition
series $S_0,S_1,S_2,S_{n+1}$. Then $u_{M}=(u_{M^E})_{E\in \Omega}\in
\cl\mg(\cR_A)$ and $0\neq [u_{M}, u_{\Sigma M}]\not\in
\wt{\mh}\otimes_{\Z}\C$, where $\wt{\mh}$ is the subspace of
$\cl\mg(\cR_A)$ spanned by $h_i, 0\leq i\leq n+m$.
\end{lemma}
\begin{proof}
One can easily check that
$u_{M^E}=[[[u_{S_0^E},u_{S_1^E}],u_{S_2^E}],u_{S_{n+1}^E}]$ by using
lemma~\ref{l:hall-num} for any $E\in \Omega$. Thus, both $u_M,
u_{\Sigma M}$ belong to $\cl\mg(\cR_A)$. Let $P_i$ be the
indecomposable projective $A$-modules associated to each vertex $i$.
Let $\to P_0\to P_1\to P_2\oplus P_{n+1}\to M\to 0$ be the
projective cover of $M$. We infer that
$\Hom_{\cR_A}(M,M)=\Hom_{\cd^b(\mod A)}(M,M)\oplus \Hom_{\cd^b(\mod
A)}(M,\Sigma^2 M)$. Moreover, $\dim_k\Hom_{\cR_A}(M,M)=2$ and
$\dim_k\rad \Hom_{\cR_A}(M,M)=1$.
Now consider triangles in $\cR_A$
\[M\to L\to \Sigma M\xrightarrow{f} \Sigma M \ \text{and}\ \Sigma M\to N\to
M\xrightarrow{g}\Sigma^2 M,
\]
we can write $f=f_0+f_1$, where $f_0\in \Hom_{\cd^b(\mod A)}(\Sigma
M,\Sigma M)$ and $f_1\in \Hom_{\cd^b(\mod A)}(\Sigma M,\Sigma^3M)$.
If $f_0\neq 0$, then $f$ is an isomorphism and $L\cong 0$. Thus, it
suffices to consider for $f_0=0$, {\it i.e.} $0\neq f\in \rad
\Hom_{\cR_A}(M,M)$, and then the triangle $M\to L\to\Sigma
M\xrightarrow{f}\Sigma M$ is induced by a triangle in $\cd^b(\mod
A)$. By computing the mapping cone of $f$ in $\cd^b(\mod A)$, we
infer that $L$ isomorphic to the complex $\cdots \to 0\to
P_0\xrightarrow{(f,l)}M\oplus P_1\to P_2\oplus P_3\to 0\cdots$,
where $P_2\oplus P_3$ is in the $-1$-th component. We claim that $L$
is indecomposable in $\cd^b(\mod A)$. Indeed, suppose $L\cong
X\oplus Y$ in $\cd^b(\mod A)$. Then $H^{*}(L)\cong H^{*}(X)\oplus
H^*(Y)$, where $H^*(-)$ be the homology groups of corresponding
complex. Now the only nonzero homology groups of $L$ are
$H^{-1}(L)\cong H^{-2}(L)\cong M$, which are indecomposable
$A$-modules. Thus, we may assume $X\cong \Sigma^2 M$ and $Y\cong
\Sigma M$. Now in the root category $\cR_A$, we have $\Sigma^2M\cong
M$. In particular, we get a triangle $M\to M\oplus \Sigma M\to
\Sigma M\xrightarrow{f} \Sigma M$. By a well-known fact, the
triangle is split and $f=0$, contradiction. Since $\dim_{k}\rad
\Hom_{\cR_A}(M,M)=1$, for any $f, h\in \rad \Hom_{\cR_A}(M,M)$, the
mapping of $f$ and $h$ are isomorphic to each other.
Similarly, one can discuss for $g$ and show that $N$ is
indecomposable if and only if $0\neq g\in \Hom_{\cd^b(\mod
A)}(M,\Sigma^2 M)$. In this case, we have $N\cong \Sigma^{-1}L$. Now
by the definition of the Lie bracket, we have
\begin{eqnarray*}[u_M,u_{\Sigma M}]&=&-h_{M}+\sum_{L\in \ind \cR_A}(F_{\Sigma M,
M}^L-F_{M,\Sigma M}^L)u_L\\
&=&-h_M+F_{\Sigma M, M}^Lu_L-F_{M,\Sigma
M}^{\Sigma^{-1}L}u_{\Sigma^{-1} L}
\end{eqnarray*}
One can show that $\dim_k\Hom_{\cR_A}(M,L)=1$. Therefore, by
lemma~\ref{l:hall-num} we have $F_{\Sigma M, M}^L=F_{M,\Sigma
M}^{\Sigma^{-1}L}=1$. In particular, we have $[u_M, u_{\Sigma
M}]=-h_M+u_L-u_{\Sigma L}$ in $\mg(\cR_A)_{(q-1)}$. We remark that
the proof above is valid for any finite field extension of $k$. Thus,
in the integral Ringel-Hall Lie algebra $\cl\mg(\cR_A)$, we also
have $[u_M, u_{\Sigma M}]=-h_M+u_{L}-u_{\Sigma L}$, which implies
the desired result.
\end{proof}
Now we are in a position to prove the theorem~\ref{t:Lie algebra
homom}.
\begin{proof} The relations
$(1)(2)(3)$ follows from the definition of Lie bracket of
Ringel-Hall Lie algebra. It suffices to show $u_{S_i}, u_{\Sigma
S_j}, 0\leq i,j\leq m+n$ satisfy the relations $(4)$ and $(5)$. We
discuss for $i,j$ in $4$ cases.
\begin{itemize}
\item[Case 1:] $i,j\in \{0, 1\}$. We consider the quotient algebra
$B=A/A(e_{2}+e_3+\cdots +e_{n+m})A$, where $e_i$ is the idempotent
associated to the vertex $i$. Note that $B$ is projective as right
$A$-module. Then the derived functor $F=-\lten_BB_A: \cd^b(\mod
B)\to \cd^b(\mod A)$ is an embedding by theorem ~3.1 in
~\cite{Cline-Parshall-Scott}. Now, lemma~\ref{l:fully faithful}
implies the induced functor $\ol{F}:\cR_B\to \cR_A$ is also fully
faithful. In particular, we have a injective Lie algebra
homomorphism $\cl\mg(\cR_B)\to \cl\mg(\cR_A)$. Moreover, we can
identify the simple $B$-modules with simple $A$-modules via the
functor $F$. Thus, to check the a relations for $\cl\mg(\cR_A)$
involve $0\leq i,j\leq 1$, it suffices to check it in
$\cl\mg(\cR_B)$. Note that the algebra $B$ is hereditary of type $\wt{A_1}$, we
infer that $\cl\mg(\cR_B)\otimes_{\Z}\C$ is isomorphic to the affine Kac-Moody algebra of type $\wt{A_1}$ by the main theorem of ~\cite{Peng-Xiao2000}, which implies relations $(4)$ and $(5)$ hold.
\item[Case 2:] $i,j\in \{1, 2, \cdots, n+m\}$. Let $B=A/Ae_0A$. It
is easy to see that $\Ext_A(B_A,B_A)=0$. Again by
theorem~3.1~\cite{Cline-Parshall-Scott}, we have $F=-\lten_BB_A:
\cd^b(\mod B)\to \cd^b(\mod A)$ is an embedding. Note that in this
case $B$ is hereditary of Dynkin type $A_{m+n}$. Thus, the
Ringel-Hall algebra $\cl\mg(\cR_B)\otimes_{\Z}\C$ is isomorphic to
simple Lie algebra of type $A_{m+n}$. Now the result follows from
the proof of case $1$.
\item[Case 3:] $i=0, j\neq 1,2,n+1$. In particular, by the definition of Lie bracket we only need to show that $[u_{S_0},u_{S_j}]=0$ and $[u_{S_0}, u_{\Sigma S_j}]=0$. This follows from the fact that $S_j$ has projective dimension $2$ and the projective resolution of $S_j$ does not involve $P_0$.
\item[Case 4:] $i, j\in \{0,2, n+1\}$. For the case $i=0, j=2$, we consider the quotient algebra $B=A/A(e_3+\cdots+e_{m+n})A$, which turns out to be a tilted algebra of tame hereditary algebra of type $\wt{A_2}$. Thus the integral Ringel-Hall algebra $\cl\mg(\cR_B)\otimes_{\Z}\C$ is isomorphic to the Kac-Moody algebra of type $\wt{A_2}$. Now the result follows from the proof of case $1$, since we still have full embeddings $F:\der^b(\mod B)\to \der^b(\mod A)$ and $\ol{F}:\cR_B\to \cR_A$. For the case $i=0,j=n+1$, one considers the quotient algebra $B=A/A(e_2+\cdots +e_n+e_{n+2}+\cdots +e_{n+m})A$.
\end{itemize}
Thus $\phi$ is indeed a Lie algebra homomorphism. It is obviously
surjective and keeps the gradation. Clearly, $h_i=(h_{S_i^E})_{E\in
\Omega}$ is linearly independent in
$(\cl\mg(\cR_A)\otimes_{\Z}\C)_0$. By lemma~\ref{l:dimension of 0th
gradation}, we infer that $u_{M}-u_{\Sigma M}\in
(\cl\mg(\cR_A)\otimes_{\Z}\C)_0$, which is linearly independent to $h_0,h_1, \cdots, h_{m+n}$.
Thus, $\dim_{\C} (\cl\mg(\cR_A)\otimes_{\Z}\C)_0\geq m+n+2$.
This completes the proof.
\end{proof}
\begin{remark}
Firstly,
theorem~\ref{t:Lie algebra homom} essentially give a negative answer to Slodowy's question. If the equality holds for $\mg=GIM(C)$, {\it i.e.} $\dim_{\C}\mg_0=m+n+2$, then for the derived subalgebra $\mg'=[\mg,\mg]$, we must have $\dim_{\C}\mg'_0=m+n+1$. In fact, following the proof of lemma~\ref{l:dimension of 0th gradation}, one can even show that $\dim_{\C} (\cl\mg(\cR_A)\otimes_{\Z}\C)_0\geq (m+1)n+1$. Secondly, by lemma~\ref{l:dimension of 0th gradation}, we know that $(\dimv M,\dimv M)=4$ and $u_M\in \cl\mg(\cR_A)$. In particular, this also shows that the GIM Lie algebra $\mg$ has root with length greater than $2$. Thirdly, one can easily see that the root basis of $GIM(C)$ is braid equivalent to a root basis of affine Kac-Moody algebra of type $\wt{A}_{m+n}$. By theorem~\ref{t:Lie algebra homom}, we know that $GIM(C)$ is never isomorphic to an affine Kac-Moody algebra of type $\wt{A}_{m+n}$, this also show that the GIM Lie algebras are not invariant under braid equivalent in general.
\end{remark}
\begin{appendix}
\section{Recollement lives in root categories}
In the appendix, we show that a recollement of bounded derived
categories lives in the corresponding root categories under suitable
assumption. This can be use to construct various algebras
inductively such that the $2$-periodic orbit category is not
triangulated with the inherited triangle structure from the
bounded derived category.
\subsection{Derived category of dg category}
Let $\ca$ be a small differential graded (dg) $k$-category. We
identify a dg $k$-algebra with a dg category with one object. Let
$\Dif \ca$ be the dg category of right dg $\ca$-modules. A dg
$\ca$-module $P$ is called {\it $\ck$-projective} if $\Dif \ca(P,?)$
preserves acyclicity. For any dg category $\cb$, let
$\mathcal{Z}^0(\cb)$ be the category with the same objects of $\ca$
whose $\Hom$-space is given by
\[\mathcal{Z}^0(\cb)(X,Y)=Z^0(\cb(X,Y)),
\]
{\it i.e.} the 0th cocycle of dg $k$-module $\cb(X,Y)$. Let
$\ch^0(\cb)$ be the category with the same objects of $\cb$ whose
$\Hom$-space is given by
\[\ch^0(\cb)(X,Y)=H^0(\cb(X,Y)),
\]
{\it i.e.} the 0th homology of dg $k$-module $\cb(X,Y)$. For the dg
category $\Dif \ca$, we define $\cc(\ca):=\mathcal{Z}^0(\Dif \ca)$
and $\ch(\ca):=\ch^0(\Dif \ca)$. A morphism $L\to N$ in $\cc(\ca)$
is called quasi-isomorphism if it induces an isomorphism in
homology. Let $\der(\ca)$ be the derived category of $\ca$, {\it
i.e.} the localization of $\cc(\ca)$ with respect to the class of
quasi-isomorphism. A dg $\ca$-module $L$ is called {\it compact} if
$\der(\ca)(L,?)$ commutes with arbitrary direct sums. For instance,
the projective $\ca$-module $\ca(?,A), A\in \ca$ is both
$\ck$-projective and compact. Let $\per (\ca)$ be the perfect
derived category of $\ca$, {\it i.e.} the smallest subcategory of
$\der\ca$ contains $\ca$ and stable under shift,extensions and
passage to direct factors. For any subcategory $\cm\subseteq
\der(\ca)$, let $\tria(\cm)$ be the thick subcategory of $\der(\ca)$
contains $\cm$.
Let $X$ be a dg $\cb\otimes_k\ca\op$-module. It gives rise to a pair
of adjoint dg functors
\[\xymatrix{\Dif \ca\ar@<1ex>[r]^{T_X}&\Dif \cb.\ar@<1ex>[l]^{H_X}}
\]
Assume $X$ is $\ck$-projective as $\cb\otimes_k\ca\op$-module, then
$(T_X,H_X)$ induces an adjoint pair triangle functors
$(\lt_X,\rh_X)$ over the derived categories, where $\lt_X$ is the
left derived functor of $T_X$. If both $\ca$ and $\cb$ are dg
$k$-algebras, we also write $?\lten_{\ca}X_{\cb}$ for $\lt_X$.
\subsection{The universal property of root category}
Let $A$ and $B$ be finite-dimensional $k$-algebras with finite
global dimension. Let $F: \cd^b(\mod A)\to \cd^b(\mod B)$ be a
standard functor, {\it i.e.} $F\cong ?\lten_AX_B$ for some complex
of $A^{op}\otimes_k B$-module. Since for any triangle functor
$L:\der^b(\mod A)\to \der^b(\mod B)$, we have $L\circ
\Sigma^2_A\cong \Sigma^2_B\circ L$. By the universal property of dg
orbit category ({\it cf.} section $9.4$ in ~\cite{Keller2005}), $F$
naturally induces a triangle functor $\ol{F}:\cR_A\to \cR_B$ and we
have the following commutative diagram
\[\xymatrix{\der^b(\mod A)\ar[d]^{\pi_A}\ar[r]^F&\der^b(\mod B)\ar[d]^{\pi_B}\\
\cR_A\ar[r]^{\ol{F}}&\cR_B}
\]
where $\pi_A, \pi_B$ are the canonical functors. In the following,
we will study the induced functor $\ol{F}$ explicitly.
We may
assume $_AX_B$ is $\ck$-projective as $A^{op}\otimes_kB$-module.
Clearly, $X$ has finite total homology. Moreover, $_AX_B$ is compact
as left $A$-module and right $B$-module respectively due to the fact
$A$ and $B$ have finite global dimension. Then we have the canonical
isomorphism $\RHom_{B}(_AX_B,?)\cong ?\lten_B \RHom_B(_AX_B,B)_A$.
Let $_BY_A\to _B\RHom_B(_AX_B,B)_A$ be a $\ck$-projective resolution
of $_B\RHom_B(_AX_B,B)_A$ as $B^{op}\otimes_k A$ -module. Thus, the
right adjoint $G$ of $F$ naturally isomorphic to $?\lten_B Y_A$.
Let $\ca$ and $\cb$ be the dg category of bounded complexes of finitely
generated projective $A$-modules and $B$-modules respectively. The tensor product by $X$ and $Y$ define
dg functors $?\lten_A X:\ca\to \cb$ and $?\lten_BY:\cb\to
\ca$. By abuse of notation, we denote these dg functors by $F$ and $G$ as
well. Similarly, one can lift the square of the shift functors $\Sigma_A^2:\der^b(\mod A)\to \der^b(\mod
A)$ and $\Sigma_B^2:\der^b(\mod B)\to \der^b(\mod B)$ to dg
functors $\Sigma_A^2:\ca\to \ca$ and $\Sigma_B^2:\cb\to\cb$.
Let
$\cR_{\ca}$ be the dg orbit category ({\it cf.} section $5$ of \cite{Keller2005}) of $\ca$ respects to $\Sigma_A^2$.
Let $\cR_{\cb}$ be the dg orbit category of $\cb$ respects to
$\Sigma_B^2$. We have canonical dg functors $\pi_{\ca}: \ca\to
\cR_{\ca}$ and $\pi_{\cb}:\cb\to \cR_{\cb}$. We have natural
isomorphisms $\Sigma_B^2\circ F\cong F\circ \Sigma_A^2$ and $\Sigma_A^2\circ G\cong G\circ \Sigma_B^2$ of dg
functors. Thus, by the universal property of dg orbit categories,
$F$ and $G$ induce dg functors $\ol{F}:\cR_{\ca}\to \cR_{\cb}$
and $\ol{G}:\cR_{\cb}\to \cR_{\ca}$. Clearly, $\ol{F}$ yields a
$\cR_{\cb}\otimes_k\cR_{\ca}^{op}$-bimodule $X_{\ol{F}}$
\[X_{\ol{F}}(B,A)\mapsto \cR_{\cb}(B,\ol{F}(A)).
\]
Similarly, $\ol{G}$ induces an
$\cR_{\ca}\otimes_k\cR_{\cb}^{op}$-bimodule $Y_{\ol{G}}$
\[Y_{\ol{G}}(A,B)\mapsto \cR_{\ca}(A, \ol{G}(B)).
\]
Let $\lt_{X_{\ol{F}}}:\der(\cR_{\ca})\to \der(\cR_{\cb})$ be the
derived tensor functor of $X_{\ol{F}}$. Let
$\lt_{Y_{\ol{G}}}:\der(\cR_{\cb})\to \der(\cR_{\ca})$ be the derived
tensor functor of $Y_{\ol{G}}$. In the following, we identify the
objects of $\ca$ with $\cR_{\ca}$ and the objects of $\cb$ with
$\cR_{\cb}$ respectively.
\begin{lemma}\label{l:adjoint}
$\lt_{X_{\ol{F}}}$ is left adjoint to $\lt_{Y_{\ol{G}}}$.
\end{lemma}
\begin{proof}
Clearly, $X_{\ol{F}}^A$ is $\ck$-projective for any $A\in \ca$ and $\lt_{X_{\ol{F}}}$ is left adjoint to $\rh_{X_{\ol{F}}}$. It suffices to
show that $\lt_{Y_{\ol{G}}}\cong \rh_{X_{\ol{F}}}$. For any $\wt{A}\in
\ca$, $X_{\ol{F}}(?,\wt{A})\cong \cR_{\cb}(?,\ol{F}(\wt{A}))$ which is
compact in $\der(\cR_{\cb})$. By Lemma $6.2$ (a) in ~\cite{Keller1994}, we have
$\lt_{X_{\ol{F}}^T}\cong \rh_{X_{\ol{F}}}$, where $X_{\ol{F}}^T$ is
defined by
\[X_{\ol{F}}^T(\wt{A},\wt{B})=\Dif \cR_{\cb}(X_{\ol{F}}(?,\wt{A}), \wt{B}^{\wedge})
\]
Thus, it suffices to show that we have a quasi-isomorphism $Y_{\ol{G}}\to
X_{\ol{F}}^T$ as $\cR_{\ca}\otimes_k\cR_{\cb}^{op}$-bimodule.
For any $\wt{A}\in \ca$ and $\wt{B}\in \cb$, we have
\begin{eqnarray*}
X_{\ol{F}}^T(\wt{A}, \wt{B})&=&\Dif \cR_{\cb}(X_{\ol{F}}(?,{\wt{A}}),\wt{B}^{\wedge})\\
&=&
\Dif \cR_{\cb}({\ol{F}}(\wt{A})^{\wedge}, \wt{B}^{\wedge})\\
&\cong&
\cR_{\cb}(\ol{F}(\wt{A}), \wt{B})\\
&\cong&\bigoplus_{n\in \Z}\cb(F(\wt{A}), \Sigma_B^{2n}\wt{B})\\
&\cong&\bigoplus_{n\in \Z} \RHom_{B}(\wt{A}\otimes_AX_B,
\Sigma_B^{2n}\wt{B})\\
&\cong&\bigoplus_{n\in \Z} \RHom_A(\wt{A},
\RHom_B(X,\Sigma_B^{2n}\wt{B}))
\end{eqnarray*}
Recall that we have quasi-isomorphism $\Sigma_B^{2n}\wt{B}\lten_B
\RHom_B(X,B)\to \RHom_B(_AX_B,\wt{B})$ and $\wt{A}$ is
$\ck$-projective as right $A$-module. In particular, we have a
quasi-ismorphism
\[\bigoplus_{n\in \Z}\RHom_A(\wt{A},\Sigma_B^{2n}\wt{B}\lten_B
\RHom_B(X,B))\xrightarrow{q.is} \bigoplus_{n\in
\Z}\RHom_A(\wt{A},\RHom_B(_AX_B,\wt{B})).
\]
Again, we also have quasi-isomorphism
$\Sigma_B^{2n}\wt{B}\otimes_BY\to
\Sigma_B^{2n}\wt{B}\lten_B\RHom_B(X,B)$, which implies
\[\bigoplus_{n\in \Z} \RHom_{A}(\wt{A}, \Sigma_B^{2n}\wt{B}\otimes_B
Y)\xrightarrow{q.is}\bigoplus_{n\in \Z}\RHom_A(\wt{A},\Sigma_B^{2n}\wt{B}\lten_B
\RHom_B(X,B)).
\]
The first term
\begin{eqnarray*}
\bigoplus_{n\in \Z} \RHom_{A}(\wt{A}, \Sigma_B^{2n}\wt{B}\otimes_B
Y)&\cong& \bigoplus_{n\in \Z} \RHom_A(\wt{A},
\Sigma_A^{2n}(\wt{B}\otimes_BY))\\
&\cong& \bigoplus_{n\in \Z}\ca(\wt{A},\Sigma_A^{2n}G(\wt{B}))\\
&\cong &\cR_{\ca}(\wt{A}, \ol{G}(\wt{B}))\\
&=&Y_{\ol{G}}(\wt{A}, \wt{B})
\end{eqnarray*}
Thus, we have obtained a quasi-isomorphism $Y_{\ol{G}}(\wt{A},
\wt{B})\to X_{\ol{F}}^T(\wt{A}, \wt{B})$, which is natural in both
$\wt{A}$ and $\wt{B}$. This completes the proof.
\end{proof}
The following lemma is quite obviously.
\begin{lemma}\label{l:fully faithful}
If $F:\der^b(\mod A)\to \der^b(\mod B)$ is fully faithful, then
$\lt_{X_{\ol{F}}}: \der(\cR_{\ca})\to \der(\cR_{\cb})$ is fully
faithful.
\end{lemma}
\begin{proof}
It follows from the Lemma $4.2$ (a) and (b) of ~\cite{Keller1994}
directly.
\end{proof}
Let $\cR_A$ be the perfect derived category of $\cR_{\ca}$. Let
$\cR_B$ be the perfect derived category of $\cR_{\cb}$. In other
word, $\cR_A$ and $\cR_B$ are the root categories of $A$ and $B$
respectively. Clearly, the triangle functors $\lt_{X_{\ol{F}}}$ and
$\lt_{Y_{\ol{G}}}$ restrict to an adjoint pair of triangle functors
\[\xymatrix{\cR_A\ar@<1ex>[r]^{\lt_{X_{\ol{F}}}}&\cR_B.\ar@<1ex>[l]^{\lt_{Y_{\ol{G}}}}}
\]
For simplicity, we still denote $\ol{F}:=\lt_{X_{\ol{F}}}:\cR_A\to
\cR_B$ and $\ol{G}:=\lt_{Y_{\ol{G}}}:\cR_B\to \cR_A$.
\subsection{Recollement lives in root categories} Suppose we are
given triangulated categories $\der', \der, \der''$ with triangle
functors
\[\xymatrix@C=1.5cm{\der'\ar[r]^-{i_*=i_!} &
\der\ar@<3ex>[l]^{i^!} \ar@<-3ex>[l]_-{i^* } \ar[r]^-{j^* =j^!}&
\der''. \ar@<3ex>[l]^-{j_*} \ar@<-3ex>[l]_-{j_!}}
\]
such that
\begin{itemize}
\item[$\circ$] $(i^*,i_*,i^!)$ and $(j_!,j^*,j_*)$ are adjoint
triples;
\item[$\circ$] $i_*, j_!,j_*$ are fully faithful;
\item[$\circ$] $j^*\circ i_*=0$;
\item[$\circ$] any $X$ in $\der$, there are distinguished
triangles
\[i_!i^!X\to X\to j_*j^*X\to \Sigma i_!i^!X\to X, j_!j^!X\to X\to
i_*i^*X\to \Sigma j_!j^!X
\]
where the morphisms $i_!i^!X\to , X\to j_*j^*X$, etc. are adjunction
morphisms.
\end{itemize}
Then we say that $\der$ admits {\it recollement} relative to $\der'$
and $\der''$. This notation was first introduced by
Beilinson-Bersstein-Deligne~\cite{Beilinson-Bersstein-Deligne1982}
in geometric setting with the idea that $ \der$ can be viewed as
bing glued together from $\der'$ and $\der''$. It is not hard to
show that if both $\der'$ and $\der''$ are Krull-Schmidt categories,
so is $\der$. Recollement in algebraic setting was studied
extensively due to the close relation with tilting theory~\cite{Jorgensen}\cite{Koenig1991}, etc.
Let $A,B,C$ are finite-dimensional $k$-algebras with finite global
dimension. Suppose that the bounded derived category $\der^b(\mod
B)$ admits a recollement relative to $\der^b(\mod A)$ and
$\der^b(\mod C)$. In particular, we have the following diagram of
triangulated categories and triangle functors
\[\xymatrix@C=1.2cm{\cd^b(\mod A)\ar[r]^-{i_*=i_!} &
\cd^b(\mod B)\ar@<3ex>[l]^-{i^!} \ar@<-3ex>[l]_-{i^* } \ar[r]^-{j^*
=j^!}& \cd^b(\mod C). \ar@<3ex>[l]^-{j_*} \ar@<-3ex>[l]_-{j_!}}
\]
Assume further that both the functors $i^*$ and $j_!$ are standard.
Then we have the following
\begin{theorem}\label{t:recollement}
Keep the notations above. Let $A$, $B$ and $C$ be finite-dimensional
$k$-algebras with finite global dimension such that the derived
category $\der^b(\mod B)$ admits a recollement relative to
$\der^b(\mod A)$ and $\der^b(\mod C)$. Assume that the functor $i^*$
and $j_!$ are standard. The root category $\cR_B$ admits a
recollement relative to $\cR_A$ and $\cR_C$. Moreover, we have the
following commutative diagram of recollements
\[\xymatrix@R=1.2cm@C=1.2cm{\cd^b(\mod A)\ar[d]^{\pi_A}\ar[r]^{i_*=i_!} &
\cd^b(\mod B)\ar[d]^{\pi_B}\ar@<3ex>[l]^-{i^!} \ar@<-3ex>[l]_-{i^* }
\ar[r]^-{j^* =j^!}& \cd^b(\mod C). \ar[d]^{\pi_C}\ar@<3ex>[l]^-{j_*}
\ar@<-3ex>[l]_-{j_!}\\
\cR_A\ar[r]^-{\ol{i_*}=\ol{i_!}} & \cR_B\ar@<3ex>[l]^-{\ol{i^!}}
\ar@<-3ex>[l]_-{\ol{i^*} } \ar[r]^-{\ol{j^*} =\ol{j^!}}& \cR_C
\ar@<3ex>[l]^-{\ol{j_*}} \ar@<-3ex>[l]_-{\ol{j_!}}.}
\]
\end{theorem}
\begin{proof}
Since $i^*$ and $j_!$ are standard, then all the functors
$i_*,i^!,j^*,j_*$ are standard due to fact $A,B,C$ have finite
global dimension. Thus, we have the corresponding induced functors
$\ol{i^*}, \ol{i_*},\ol{i^!},\ol{j_!},\ol{j^*},\ol{j_*}$. The
commutativity of the above diagram follows from the universal
property of the root categories. It suffices to show that $\cR_B$
admits a recollement relative to $\cR_A$ and $\cR_C$ together with
the functors $\ol{i^*},
\ol{i_*},\ol{i^!},\ol{j_!},\ol{j^*},\ol{j_*}$. Clearly, $(\ol{i^*},
\ol{i_*},\ol{i^!})$ and $(\ol{j_!},\ol{j^*},\ol{j_*})$ are adjoint
triples follows Lemma ~\ref{l:adjoint}. By Lemma ~\ref{l:fully
faithful}, one infers that $\ol{i_*}, \ol{j_!},\ol{j_*}$ are fully
faithful. Since $\cR_A$ is generated by $\pi_A(A)$, to show that
$\ol{j^*}\circ \ol{i_*}=0$, it suffices to show $\ol{j^*}\circ
\ol{i_*}(\pi_A(A))=0$. By the commutativity of the above diagram,
this result follows from $j^*\circ i_*=0$. It remains to show that
for any $X\in \cR_B$ there are triangles
\[\ol{i_!}\ol{i^!}X\to X\to \ol{j_*}\ol{j^*}X\to \Sigma
\ol{i_!}\ol{i^!}X,\ \ol{j_!}\ol{j^!}X\to X\to \ol{i_*}\ol{i^*}X\to
\Sigma \ol{j_!}\ol{j^!}X.
\]
We prove the existence of the first triangle, the second one is
similar.
If $X\in \im \pi_B$, there is $Y\in \der^b(\mod B)$ such that
$X=\pi_B(Y)$. By the recollement of $\der^b(\mod B)$ relative to
$\der^b(\mod A)$ and $\der^b(\mod C)$, we have
\[i_!i^!Y\to Y\to j_*j^*Y\to \Sigma i_!i^!Y.
\]
Applying the triangle functor $\pi_B$, we get a triangle in $\cR_B$
\[\pi_B(i_!i^!Y)\to \pi_B(Y)\to \pi_B(j_*j^*Y)\to \Sigma
\pi_B(i_!i^!Y).
\]
By the commutativity of the functors, we have
\[\ol{i_!}\ol{i^!}\pi_B(Y)\to \pi_B(Y)\to \ol{j_*}\ol{j^*}\pi_B(Y)\to \Sigma
\ol{i_!}\ol{i^!}\pi_B(Y).
\]
Clearly, this triangle is isomorphic to
\[\ol{i_!}\ol{i^!}\pi_B(Y)\xrightarrow{\eta_X} \pi_B(Y)\xrightarrow{\epsilon_X}\ol{j_*}\ol{j^*}\pi_B(Y)\to \Sigma
\ol{i_!}\ol{i^!}\pi_B(Y)
\]
where $\eta_X, \epsilon_X$ are adjunction morphisms, which implies
the later one is a distinguished triangle.
Consider the triangle $X\xrightarrow{f} Y \to Z\to \Sigma X$, where
$X, Y\in \im \pi_B$. Consider the following commutative square
\[\xymatrix{\ol{i_!}\ol{i^!}X\ar[d]^{\ol{j_*}\ol{j^*}f}\ar[r]^{\eta_X}&X\ar[d]^f\\
\ol{j_*}\ol{j^*}Y\ar[r]^{\eta_Y}&Y}
\]
By nine lemma, one can embed the square to the following commutative
diagram of triangles
\[\xymatrix{\ol{i_!}\ol{i^!}X\ar[d]^{\ol{j_*}\ol{j^*}f}\ar[r]^{\eta_X}&X\ar[d]^f\ar[r]^{\epsilon_X}&\ol{j_*}\ol{j^*}X\ar[d]^{\ol{j_*}\ol{j^*}f}\\
\ol{i_!}\ol{i^!}Y\ar[d]^{\ol{i_!}g_1}\ar[r]^{\eta_Y}&Y\ar[d]^g\ar[r]^{\epsilon_Y}&\ol{j_*}\ol{j^*}Y\ar[d]^{\ol{j_*}g_2}\\
\ol{i_!}U_Z\ar[r]^u&Z\ar[r]^v&\ol{j_*}V_Z}
\]
Let $\phi(u):U_Z\to \ol{i^!}Z $ be the morphism corresponds to $u$
under the natural isomorphism. Let $\phi(v):\ol{j^*}Z\to V_Z$ be the
morphism corresponds to $v$. It is clear that $\phi(u)$ and
$\phi(v)$ are isomorphisms. Thus, one gets the following commutative
diagram
\[\xymatrix{\ol{i_!}U_Z\ar[d]^{\ol{i_!}\phi(u)}\ar[r]^u&Z\ar@{=}[d]\ar[r]^v&\ol{j_*}V_Z\ar[d]^{\ol{j_*}(\phi(v)^{-1})}\ar[r]^w&\Sigma \ol{i_!}U_Z\ar[d]^{\Sigma \ol{i_!}\phi(u)}\\
\ol{i_!}\ol{i^!}Z\ar[r]^{\eta_Z}&Z\ar[r]^{\epsilon_Z}&\ol{j_*}\ol{j^*}Z\ar[r]^\delta &\Sigma \ol{i_!}\ol{i^!}Z }
\]
where $\delta=\ol{j_*}\phi(v)\circ w\circ \Sigma \ol{i_!}\phi(u)$.
Thus, one informs that
\[\ol{i_!}\ol{i^!}Z\xrightarrow{\eta_Z} Z\xrightarrow{\epsilon_Z} \ol{j_*}\ol{j^*}Z\to \Sigma
\ol{i_!}\ol{i^!}Z
\]
is a distinguished triangle. Now this holds for any $Z\in \cR_B$ by
'devissage'.
\end{proof}
\begin{corollary}\label{c:dense}
Keep the assumptions in theorem~\ref{t:recollement} . If the canonical functor $\pi_B$ is
dense, then both $\pi_A$ and $\pi_C$ are dense.
\end{corollary}
\begin{proof}
For any $X\in \cR_A$, consider $\ol{i_*}X\in \cR_B$. By the dense of
$\pi_B$, there is a $Y\in \der^b(\mod B)$ such that $\pi_B(Y)\cong
\ol{i_*}X$. For $Y$, one have the canonical triangle $i_!i^!Y\to
Y\to j_*j^*Y\to$. Applying the functor $\pi_B$, we have
\[\pi_B(i_!i^!Y)\to \pi_B(Y)\to \pi_B(j_*j^*Y)\to
\]
which have to isomorphic to the canonical triangle
\[\ol{i_!}\ol{i^!}(\ol{i_!}X)\to X\to 0\to.
\]
One gets $X\cong \pi_A(i^! Y)$. In particular, $\pi_A$ is dense.
Similar proof implies that $\pi_C$ is also dense.
\end{proof}
\begin{remark}
If only one of $i^*$ and $j_!$ is standard, say $i^*$ is standard.
Then lemma ~\ref{l:adjoint} ~\ref{l:fully faithful} and a result of
~\cite{Parshall-Scott1988} imply that there is a recollement
\[\xymatrix@C=1.5cm{
\cR_A\ar[r]^-{\ol{i_*}=\ol{i_!}} & \cR_B\ar@<3ex>[l]^-{\ol{i^!}}
\ar@<-3ex>[l]_-{\ol{i^*} } \ar[r]^-{\ol{j^*} =\ol{j^!}}&
\cR_B/\ol{i_*}\cR_A \ar@<3ex>[l]^-{\ol{j_*}}
\ar@<-3ex>[l]_-{\ol{j_!}}.}
\]
The corollary ~\ref{c:dense} also holds in this case (one should
replace the functor $\pi_C$).
\end{remark}
The following is now quite obviously.
\begin{corollary}
Let $A$ and $B$ be finite-dimensional
$k$-algebras with finite global dimension. Assume the root category
$\cR_A$ is not triangulated with the inherited triangle structure.
For any finite dimensional $A\otimes_k B\op$-module $M$, the root
category of the triangular extension of $A$ and $B$ by $M$ is not
triangulated with the inherited triangle structure.
\end{corollary}
\end{appendix}
\def\cprime{$'$}
\providecommand{\bysame}{\leavevmode\hbox
to3em{\hrulefill}\thinspace}
\providecommand{\MR}{\relax\ifhmode\unskip\space\fi MR }
\providecommand{\MRhref}[2]
\href{http://www.ams.org/mathscinet-getitem?mr=#1}{#2}
} \providecommand{\href}[2]{#2}
|
\section{Dukhin and P\'{e}clet numbers in the presence of slip}
Surface conduction in presence of slip is characterized by full Dukhin number, which is given by~\cite{khair_squires2009}:
\begin{equation} \label{Du}
Du = \frac{4(1+m)}{\kappa L}\sinh^2\left(\frac{ze\zeta}{2k_B T}\right)+\frac{2mb}{L}\sinh^2\left(\frac{ze\zeta}{4k_B T}\right),
\end{equation}
where $m=2\varepsilon(k_B T/ze)^2/(\eta D)$, and $D$ is the ion diffusivity, which is assumed equal for both types of ions.
The first term in (\ref{Du}) is the Dukhin number for the no-slip surface, $Du_{b=0}$, while the second one, $Du_{b}$, is due to hydrodynamic slip. In the Debye-H\"{u}ckel model
\begin{equation} \label{DH_limit}
\frac{ze\zeta}{k_B T}\ll 1.
\end{equation}
This restriction (\ref{DH_limit}) allows the simplification:
\begin{equation} \label{Du_ns}
Du_{b=0}\approx \frac{4(1+m)}{\kappa L}\left(\frac{ze\zeta}{2k_B T}\right)^2,
\end{equation}
\begin{equation} \label{Du_s}
Du_{b}\approx \frac{2mb}{L}\left(\frac{ze\zeta}{4k_B T}\right)^2.
\end{equation}
The parameter $m\approx 100/z^2$, and $\kappa L \gg 1$ since EDL is thin. Whence $Du_{b=0}\approx 1/(z^2 \kappa L)\ll 1$ provided the potential is low. For a ``slippery part'' of Du we evaluate $ Du_b\approx 0.5 b/L$ if $e\zeta/(k_B T)\approx 0.1$, and $Du_b\approx 5\cdot 10^{-3} b/L$ for $e\zeta/(k_B T)\approx 0.01$. Therefore for increasing surface charge (potential) and $b/L$, the conductivity of the diffuse layer can become
comparable to the bulk, and surface condition must be considered.
The P\'{e}clet number, $Pe=UL/D$ , in presence of slip can be evaluated as
\begin{equation} \label{Pe}
Pe= \frac{q_2 E_t(1+b\kappa)L}{\kappa\eta D}.
\end{equation}
Typically, electroosmotic velocity is of order is
of order micrometers per second for no-slip surfaces~\cite{anderson.jl:1989}. For nano-scale patterns $L<1$ $\mu$m and typical ion diffusivities $D\approx 10^{-6}$ ${\rm cm}^2/{\rm s}$ this gives $Pe_{b=0}<0.01 \ll 1 $. The slip implies a correction factor $(1+b\kappa)$ , which suggests that the convective ion transport can safely be neglected only for $b\kappa < 10$. Larger values of $b\kappa$ should relax this standard approximation of small $Pe$.
\section{Electro-osmotic velocity in eigendirections}
{\bf Longitudinal stripes.}-- In this configuration only $x-$velocity component remains, and the Stokes equation takes the form
\begin{equation} \label{Stokes_par}
\left(\partial^2_y+\partial^2_z\right)u=\varepsilon \kappa^2 \psi E_t
\end{equation}
We expand surface charge density in a Fourier series, and the potential is then
\begin{equation} \label{psi_par}
\psi(y,z) = \frac{\langle q\rangle}{\varepsilon\kappa}e^{-\kappa y} + \sum\limits_{n=1}^\infty \frac{q_n}{\varepsilon \xi_n}e^{-\xi_n y} \cos{\lambda_n z},
\end{equation}
where $\xi_n=\sqrt{\kappa^2+\lambda_n^2}$, $\lambda_n = 2 n \pi /L$ , $\langle q \rangle = q_1 \phi_1 + q_2 \phi_2 $ is the mean surface charge, and
\begin{equation}
q_n=\frac{2(q_2-q_1)}{\pi n}\sin{\frac{\pi n\delta}{L}}.
\end{equation}
The general solution to (\ref{Stokes_par}) for $u(y,z)$ has the form
\begin{equation} \label{upar1}
u(y,z) = U^{\parallel} + \sum\limits_{n=1}^\infty U_n e^{-\lambda_n y} \cos{\lambda_n z} + \frac{\varepsilon E_t}{\eta}\psi,
\end{equation}
where $U_n$ and $U^{\parallel}$ are determined by the slip boundary conditions. Imposing them on (\ref{upar1}) in a \emph{thin EDL} limit yields a dual series
\begin{multline} \label{dual_series_par_thina}
U^{\parallel}
+ \sum\limits_{n=1}^\infty U_n(1+b\lambda_n)\cos{\lambda_n z} = {} \\
{} = -\frac{E_t q_2}{\kappa\eta}(1+\kappa b), \quad |z|\leq \delta/2,
\end{multline}
\begin{equation} \label{dual_series_par_thinb}
U^{\parallel}
+ \sum\limits_{n=1}^\infty U_n\cos{\lambda_n z}= - \frac{E_t q_1}{\kappa\eta}, \quad \delta/2<|z|\leq L/2,
\end{equation}
which can be solved exactly by using a technique~\cite{belyaev.av:2010a} to obtain the thin-EDL electro-osmotic velocity:
\begin{equation} \label{U_par_thin}
U^{\parallel} = - \frac{E_0}{\eta} \left(b^\parallel_{\rm{eff}} \frac{q_2-q_1+q_2\kappa b}{\kappa b}+\frac{q_1}{\kappa}\right).
\end{equation}
{\bf Transverse stripes.}--Although an external pressure gradient is equal to zero, local pressure variations contribute into a non-zero term $\nabla p $ in the Stokes equation, so that the flow is essentially two-dimensional. We first introduce a stream function $f(x,y)$
\begin{equation} \label{Ff}
\partial_y f = u, \quad \partial_x f = -v,
\end{equation}
which obeys inhomogeneous biharmonic equation:
\begin{equation} \label{Eq_Ff}
\nabla^2 \nabla^2 f = E_t \frac{\varepsilon\kappa^2}{\eta}\frac{\partial\psi}{\partial y}
\end{equation}
Here $u$ and $v$ are $x$ and $y$ velocity components, correspondingly. The general periodic solution to (\ref{Eq_Ff}) has the form
\begin{multline} \label{Ff_sol}
f(x,y)=U^\perp y+\sum\limits_{n=1}^\infty{\left(\frac{E_t q_n}{\kappa^2 \eta}+g_n y\right)e^{-\lambda_n y}\cos{\lambda_n x}} + {} \\
{} +\frac{E_t \varepsilon}{\kappa^2 \eta}\frac{\partial\psi}{\partial y}
\end{multline}
Here the potential $\psi(x,y)$ has exactly the form ($\ref{psi_par}$) with $z$ replaced by $x$.
The dual series problem in a \emph{thin EDL} limit can be written as
\begin{multline} \label{dual_series_ort1a}
U^{\perp}
+ \sum\limits_{n=1}^\infty a_n(1+2b\lambda_n)\cos{\lambda_n x} = {} \\
{}= -\frac{E_t q_2}{\kappa\eta}(1+\kappa b), \quad |x|\leq \delta/2,
\end{multline}
\begin{equation} \label{dual_series_ort1b}
U^{\perp}
+ \sum\limits_{n=1}^\infty a_n\cos{\lambda_n x}= - \frac{E_t q_1}{\kappa\eta}, \quad \delta/2<|x|\leq L/2,
\end{equation}
where
\[
a_n=g_n+\frac{E_t q_n}{\eta\kappa^2}(\xi_n-\lambda_n).
\]
These dual series can be solved exactly to obtain
\begin{equation} \label{U_ort_thin}
U^{\perp} = - \frac{E_0}{\eta} \left(b^\perp_{\rm{eff}} \frac{q_2-q_1+q_2\kappa b}{\kappa b}+\frac{q_1}{\kappa}\right)
\end{equation}
We emphasize that a comparison of Eqs.(\ref{U_par_thin}) and (\ref{U_ort_thin}) indicates that the EO flow is generally anisotropic, so that our results do not support an earlier conclusion~\cite{bahga:2009} that the electro-osmotic mobility tensor is isotropic in the thin EDL limit. This inconsistensy~\cite{bahga:2009} (due to an erroneous expression for a transverse electro-osmotic velocity, where factor of 2 was lost) has been corrected for a case $b_2=\infty$ in~\cite{vinogradova.oi:2011}.
\bibliographystyle{rsc}
|
\chapter{Publications}\label{ch:Pub}
Wherever published work was used in this thesis, this was clearly indicated. The following is a complete list of my work published in peer--reviewed journals or on preprint archives, in chronological order. Preprints of the peer-reviewed publications, marked with a `$\star$' in the following list, that are most relevant to this thesis are then reproduced in the same order over the following several pages.
\section{Peer-reviewed journal articles}
\noindent
\rlap{\hspace{-0.7em}$\star$}\years{2009}\textbf{Andr\'e Xuereb}, Peter Domokos, Janos Asb\'oth, Peter Horak, and Tim Freegarde; \emph{Scattering theory of heating and cooling in optomechanical systems}; Phys.\ Rev.\ A \textbf{79}, 053810 (2009); \href{http://www.arxiv.org/abs/0903.3132}{\texttt{arXiv:0903.3132}}. Given a synopsis in the APS journal \emph{Physics} and mentioned in the Research Highlights section of Nature Photonics \textbf{3}, 7 (2009)\\
\rlap{\hspace{-0.7em}$\star$}\years{2009}\textbf{Andr\'e Xuereb}, Peter Horak, and Tim Freegarde; \emph{Atom cooling using the dipole force of a single retroreflected laser beam}; Phys.\ Rev.\ A \textbf{80}, 013836 (2009); \href{http://www.arxiv.org/abs/0903.2945}{\texttt{arXiv:0903.2945}}\\
\rlap{\hspace{-0.7em}$\star$}\years{2009}Hamid Ohadi, Matthew Himsworth, \textbf{Andr\'e Xuereb}, and Tim Freegarde; \emph{Magneto-optical trapping and background-free imaging for atoms near nanostructured surfaces}; Opt.\ Express \textbf{17}, 25, 23003 (2009); \href{http://www.arxiv.org/abs/0910.5003}{\texttt{arXiv:0910.5003}}\\
\years{2010}\textbf{Andr\'e Xuereb}, Mathias Groth, Karl Krieger, Otto Asunta, Taina Kurki-Suonio, Jari Likonen, David P Coster, ASDEX Upgrade Team; \emph{DIVIMP-B2-EIRENE modelling of $^{13}$C migration and deposition in ASDEX Upgrade L-mode plasmas}; J.\ Nucl.\ Mater.\ \textbf{396}, 2--3, 228 (2010). Based on work done whilst on an IAESTE traineeship at the Helsinki University of Technology (TKK), July--August 2006\\
\rlap{\hspace{-0.7em}$\star$}\years{2010}James Bateman, \textbf{Andr\'e Xuereb}, and Tim Freegarde; \emph{Stimulated Raman transitions via multiple atomic levels}; Phys.\ Rev.\ A \textbf{81}, 043808 (2010); \href{http://www.arxiv.org/abs/0908.2389}{\texttt{arXiv:0908.2389}}\\
\rlap{\hspace{-0.7em}$\star$}\years{2010}\textbf{Andr\'e Xuereb}, Tim Freegarde, Peter Horak and Peter Domokos; \emph{Optomechanical cooling with generalized interferometers}; Phys.\ Rev.\ Lett.\ \textbf{105}, 013602 (2010); \href{http://www.arxiv.org/abs/1002.0463}{\texttt{arXiv:1002.0463}}\\
\years{2010}James Bateman, Richard Murray, Matthew Himsworth, Hamid Ohadi, \textbf{Andr\'e\linebreak Xuereb}, and Tim Freegarde; \emph{H\"ansch--Couillaud locking of Mach--Zehnder interferometer for carrier removal from a phase-modulated optical spectrum}; J.\ Opt.\ Soc.\ Am.\ B \textbf{27}, 1530 (2010); \href{http://www.arxiv.org/abs/0911.1695}{\texttt{arXiv:0911.1695}}\\
\rlap{\hspace{-0.7em}$\star$}\years{2010}Peter Horak, \textbf{Andr\'e Xuereb}, and Tim Freegarde; \emph{Optical cooling of atoms in microtraps by time--delayed reflection}; J.\ Comput.\ Theor.\ Nanosci.\ \textbf{7}, 1747 (2010); \href{http://www.arxiv.org/abs/0911.4805}{\texttt{arXiv:0911.4805}}\\
\rlap{\hspace{-0.7em}$\star$}\years{2010}\textbf{Andr\'e Xuereb}, Peter Domokos, Peter Horak, and Tim Freegarde; \emph{Scattering theory of multilevel atoms interacting with arbitrary radiation fields}; Phys.\ Scr.\ \textbf{T140}, 014010 (2010); \href{http://www.arxiv.org/abs/0910.0802}{\texttt{arXiv:0910.0802}}\\
\rlap{\hspace{-0.7em}$\star$}\years{In press}\textbf{Andr\'e Xuereb}, Peter Horak, and Tim Freegarde; \emph{Amplified optomechanics in a unidirectional ring cavity}; to appear in J.\ Mod.\ Opt.; \href{http://www.arxiv.org/abs/1101.0130}{\texttt{arXiv:1101.0130}}\\
\rlap{\hspace{-0.7em}$\star$}\years{In press}\textbf{Andr\'e Xuereb}, Peter Domokos, Peter Horak, and Tim Freegarde; \emph{Cavity cooling of atoms: Within and without a cavity}; to appear in Eur.\ Phys.\ J.\ D; \href{http://www.arxiv.org/abs/1101.2739}{\texttt{arXiv:1101.2739}}
\section{Conference proceedings}
\years{2011}Peter Domokos, \textbf{Andr\'e Xuereb}, Peter Horak, and Tim Freegarde; \emph{Efficient optomechanical cooling in one-dimensional
interferometers}; presented at SPIE Photonics West 2011 (invited contribution)
\oneside
\includepdf[addtotoc={1,section,2,Scattering theory of cooling and heating in opto-mechanical systems,preprint:TMM},pages=-,offset=1in -1in,scale=0.9]{Preprints/TMM.pdf}
\includepdf[addtotoc={1,section,2,Atom cooling using the dipole force of a single retroflected laser beam,preprint:MMC},pages=-,offset=1in -1in,scale=0.9]{Preprints/MMC.pdf}
\includepdf[addtotoc={1,section,2,Magneto-optical trapping and background-free imaging for atoms near nanostructured surfaces,preprint:LambdaMOT},pages=-,offset=1in -1in,scale=0.9]{Preprints/LambdaMOT.pdf}
\includepdf[addtotoc={1,section,2,Stimulated Raman transitions via multiple atomic levels,preprint:RamanLambda},pages=-,offset=1in -1in,scale=0.9]{Preprints/RamanLambda.pdf}
\includepdf[addtotoc={1,section,2,Optomechanical cooling with generalized interferometers,preprint:PRL},pages=-,offset=1in -1in,scale=0.9]{Preprints/PRL.pdf}
\includepdf[addtotoc={1,section,2,Optical cooling of atoms in microtraps by time-delayed reflection,preprint:MonteCarlo},pages=-,offset=1in -1in,scale=0.9]{Preprints/MonteCarlo.pdf}
\includepdf[addtotoc={1,section,2,Scattering theory of multilevel atoms interacting with arbitrary radiation fields,preprint:MultiLevel},pages=-,offset=1in -1in,scale=0.9]{Preprints/Multilevel.pdf}
\includepdf[addtotoc={1,section,2,Amplified optomechanics in a unidirectional ring cavity,preprint:AOM},pages=-,offset=1in -1in,scale=0.9]{Preprints/AOM.pdf}
\includepdf[addtotoc={1,section,2,Cavity cooling of atoms: Within and without a cavity,preprint:CavityCooling},pages=-,offset=1in -1in,scale=0.9]{Preprints/CavityCooling.pdf}
\twoside
\chapter{Posters and Presentations}\label{ch:Posters}
\section{Talks}
Parts of this thesis were presented orally in the conferences and seminars listed below.
\par
\years{2010, Invited}\emph{Cooling polarisable particles with an optical memory}; University of Malta (Malta, November)\\
\years{2010, Invited}---; University of Hannover (Hannover, November)\\
\years{2010, Invited}---; Institute for Quantum Optics and Quantum Information (IQOQI; Innsbruck, July)\\
\years{2010, Invited}---; University of Vienna (Vienna, June)\\
\years{2010}\emph{Laser cooling using the dissipative dipole force}; University of Southampton School of Physics and Astronomy weekly QLM Seminar (Southampton, February)\\
\years{2009}\emph{Scattering theory of light--matter interactions}; 2nd UK Atom--Cavity Network Meeting (Leeds, December)\\
\years{2009}\emph{Scattering theory of cooling in optomechanical systems}; CMMC09 Workshop (Obergurgl, February)\\
\years{2008}\emph{Cooling of atoms using nanostructured surfaces}; University of Southampton School of Phy\-sics and Astronomy weekly QLM Seminar (Southampton, December)
\section{Posters}
A number of posters were also presented, both at specialist conferences and at research showcases aimed at a lay audience. These posters are listed below and reprinted over the next several pages.
\par
\years{2010}\emph{Cooling atoms, particles and polarisable objects using the dissipative dipole force}; International Conference on Quantum Optics (Obergurgl, 2010), Final CMMC Meeting (Herrsching, 2010) and ICAP 2010 (Cairns, 2010)\\
\years{2010}\emph{Mirror-mediated cooling of a particle by coupling to its own reflection}; Final CMMC Meeting (Herrsching, 2010) and ICAP 2010 (Cairns, 2010)\\
\years{2009}\emph{Scattering theory of cooling in optomechanical systems}; ICOLS09 (Hokkaido, 2009)\\
\years{2009}\emph{Novel Optical Cooling Methods for Atoms and Molecules}; University of Southampton FESM Research Showcases 2009 and 2010 (Southampton, 2009 and 2010) and explains author's research to non-specialists\\
\years{2008}\emph{Novel Optical Cooling Methods for Atoms and Molecules}; Photon08 (Edinburgh, 2008) and Les Houches School (Les Houches, 2008)\\
\years{2008}\emph{Semiclassical Theory of Coherent Atom Cooling with a Single Mirror}; EuroQUAM Inaugural Conference (Barcelona, 2008)
\newpage
\begin{figure*}
\centering
\fbox{\includegraphics[width=\textwidth]{./Posters/Barcelona}}
\caption[Poster 1]{Explanation of the mirror-mediated cooling mechanism and some initial investigations into the model.}
\end{figure*}
\newpage
\begin{figure*}
\centering
\fbox{\includegraphics[width=\textwidth]{./Posters/Edinburgh}}
\caption[Poster 2]{A more in-depth treatment of mirror-mediated cooling and some speculation about possible experimental realisations.}
\end{figure*}
\newpage
\begin{figure*}
\centering
\fbox{\includegraphics[width=\textwidth]{./Posters/FESM}}
\caption[Poster 3]{An explanation of laser cooling intended for a lay audience.}
\end{figure*}
\newpage
\begin{figure*}
\centering
\fbox{\includegraphics[width=\textwidth]{./Posters/Hokkaido}}
\caption[Poster 4]{Introduction to our transfer matrix theory and its use in explaining both mirror-mediated cooling and external cavity cooling.}
\end{figure*}
\newpage
\begin{figure*}
\centering
\fbox{\includegraphics[width=\textwidth]{./Posters/Herrsching_a}}
\caption[Poster 5]{An exploration of mirror-mediated cooling from the semi-classical and transfer matrix approaches, as well as external cavity cooling. Also some comments on the experimental apparatus used to conduct these investigations.}
\end{figure*}
\newpage
\begin{figure*}
\centering
\fbox{\includegraphics[width=\textwidth]{./Posters/Herrsching_b}}
\caption[Poster 6]{Introduction of a classical, fully 3D and vectorial scattering theory to explain the interaction of a polarisable particle with its own reflection.}
\end{figure*}
\chapter*{Conclusions and outlook}\label{ch:Conclusion}
\addtotoc{Conclusions and outlook}
Cold atom experiments have progressed significantly over the $25$ years since the first cold atoms were observed in optical molasses. Increasingly, research is focussing more and more towards the use of cold atoms in applications such as sensing, metrology and information processing. This trend has made it ever more important to find means of cooling more general species of atoms and molecules; current methods are simply too species-selective. Throughout the course of this thesis I have shown how the retarded dipole--dipole interaction can be used to achieve this aim; by investing the dipole force with a non-conservative nature, one can transfer energy from a moving polarisable particle to or from the light field---the nature of the particle itself is largely irrelevant. Several key theoretical results were reported in this thesis; these are summarised below:
\begin{itemize}
\item By introducing a time delay, in a very general sense, into an optical system, one can endow the dipole interaction with a non-conservative nature.
\item This delayed dipole--dipole interaction is a very general mechanism and applies to anything that is acted upon by the dipole force.
\item In order to explore these interactions, we devised an extended formalism based on the transfer matrix method and applied it successfully to various cooling schemes.
\item The prototypical mirror-mediated cooling mechanism was explored in great detail and its shortcomings addressed through external cavity cooling and amplified optomechanics.
\item Apart from rendering the exploration of the above cooling mechanisms possible, the formalisms we extended also naturally lead to the unification of cooling mechanisms for atomic motion, and optomechanical cooling mechanisms for micromirror vibrations.
\end{itemize}
The basis for an experimental investigation of these cooling mechanisms was also laid in the latter parts of this thesis, the main results of which are
\begin{itemize}
\item an experimental apparatus that simplifies the investigation of the interactions between atoms and surfaces structured on the nano- or micro-scale;
\item a multi-level imaging system for exploring cold atom clouds near highly reflective and highly scattering surfaces; and
\item the exploration of a number of physical configurations that promise experimentally-observable effects under realisable conditions.
\end{itemize}
\par
Work is currently underway to better understand the nature of the retarded dipole--dipole interaction in three dimensions and in a more general geometric setting; it is hoped that this understanding will greatly facilitate the production of systems, perhaps of an integrated nature, that can be tailored to generate cold samples of any arbitrary species. Such systems would revolutionise sensing and metrology applications and also provide an unprecedented means of interfacing with the quantum nature of matter at the micro- and mesoscopic scales.\\
Several questions are raised by the work in this thesis that are as yet unanswered, but which can form the basis of theoretical work in the future, are:
\begin{itemize}
\item Can the vectorial electromagnetic field inside a possibly truncated, micro- or mesoscopic, hemispherical mirror be expressed succinctly? Small hemispherical mirrors violate the approximations usually assumed in optics, namely the paraxial and ray approximations. The resulting behaviour of the electric field inside such deeply curved surfaces is very intricate. Some work has been done using scattering theories (see, for example, Refs.~\cite{Balian1970} and~\cite{Balian1971}), but these theories lead to numerically-intensive computations and do not give the insight required to explore the system from the point of view of the retarded dipole--dipole interaction.
\item A natural generalisation of the transfer matrix method for static scatterers in three dimensions is the scattering matrix theory; can this latter theoretical framework be extended to account for the motion of scatterers? The transfer matrix method explored in this thesis provides a very general formalism for exploring light--matter interactions. It is, however, limited to systems that are inherently in one dimension, systems that can be reduced to one dimension (such as ring cavities), or systems interacting with an electric field that can be described in terms of plane waves along orthogonal dimensions. Systems such as colloidal crystals in two or three dimensions cannot be explored using the matrix theory described in this thesis, and yet potentially present highly interesting dynamics mediated by light.
\item The transfer matrix approach presented is based on the assumption of a quasi-static system. Within such a framework, the motion of any scatterer must be slow enough for the system to reach optical steady-state at every point. Thus, any motion must be slow on the timescale defined by the decay lifetime of any cavity interacting with the scatterer. This condition specifically rules out operation in the so-called `resolved sideband' regime which, in the case of both micromirrors~\cite{Aspelmeyer2010} and ions~\cite{Ohadi2008}, has been shown to lead to the most efficient cooling processes. In particular, it is not yet known whether the external cavity cooling mechanism would be as effective in this regime.
\end{itemize}
\chapter*{Introduction}\label{ch:Introduction}
\addtotoc{Introduction}
\section*{Motivation}
The fundamental quest to know more about the nature of the world around us and the universe of which it is both an incomprehensibly insignificant part, objectively speaking, and a rather important part, subjectively speaking, has driven mankind to the edge of sanity\footnote{I am somewhat fond, however unfairly, of mentioning string theory in this context. See Ref.~\cite{Shapiro2007} for an overview into the birth of string theory that suggests otherwise.} and sometimes beyond. Amidst all of this, cold atoms can be seen as the ideal prototypical system with which to explore nature. In contrast to the sledgehammer approach so typical of high-energy physics experiments, clouds of ultracold matter provide an almost blank canvas that one can use as a microscope to probe the behaviour of matter at the atomic, and even subatomic~\cite{Hudson2002}, scale. Optical cooling is perhaps the only direct way of producing these ultracold clouds.
\par
Therein lies the catch, however. Whereas the most sensitive probes of various physical laws are perhaps the more complicated molecules, most optical cooling methods in use today are applicable to only a small (albeit growing) fraction of the periodic table. Experiments are quickly advancing to the point where a select few ``normal''\footnote{Ultracold di-alkali molecules are reasonably common (see Refs.~\cite{Jones2006} and~\cite{Kohler2006} for reviews), but require ultracold atoms as building blocks.} molecules will soon be optically cooled on demand~\cite{Stuhl2008,Zeppenfeld2009} and the race is now on to find methods of cooling any given molecule to ultracold temperatures.\\
In the related field of optomechanics, and I will in due course examine how the relation with cold atoms runs deeper than the widespread use of the word ``cold'' in the literature, the aim of cooling a macroscopic vibrating object to its ground state has been achieved using cryogenic methods~\cite{OConnell2010}, but several experimental groups~\cite{Groblacher2009a,Schliesser2009,Kippenberg2007} are tirelessly working to achieve the same using optical cooling methods.
\par
New cooling paradigms thus have to be devised, perhaps borrowing a healthy dose of inspiration from those currently in use, to bridge the gaps between the cooling of simple atoms, the cooling of arbitrary molecules, and even the cooling of micro- and mesoscopic structures.
\section*{Brief overview of past work}
Optical cooling of the motion of particles was originally observed in the regime of small dielectric particles, with Ashkin hypothesising~\cite{Ashkin1970} that similar ideas might be applicable to the control of atomic and molecular motion. Wonderful progress was made, both in experiment~\cite{Chu1985,Lett1988} and in theory~\cite{Dalibard1989,Ungar1989} in extending these ideas to three dimensions and in exploiting the internal structure of simple atoms. Cloud temperatures of tens of $\upmu$K have long been considered routine in magneto--optical traps~\cite{Lu1996}. Cooling the motion of dielectric particles proved to be more challenging, especially because the initial explorations~\cite{Ashkin1970} were conducted with the particles in suspension whereas a more desirable configuration~\cite{Chang2009b} would be with the particle suspended in vacuum; cooling of a microscopic particle to millikelvin temperatures was indeed reported very recently~\cite{Li2011}.
\par
Sub-Doppler cooling mechanisms, of the type reported in Ref.~\cite{Dalibard1989}, shift the energy loss process from the decay of the excited state population to the non-adiabatic decay of the population of different hyperfine levels. In the search for more generally-applicable cooling methods, it was realised that one can similarly shift the energy loss process to the decay of a degree of freedom external to the atom. This led to the proposal for the cooling of atomic motion inside a driven cavity~\cite{Horak1997}---demonstrated recently in experiments by the Vuleti\'c~\cite{Leibrandt2009} and Rempe~\cite{Koch2010} groups---by extension of similar mechanisms for cooling atomic motion using the modified vacuum field present in a cavity~\cite{Lewenstein1993}. Following a separate line of research, Braginsky and co-workers~\cite{Braginsky1967} suggested using cavity fields to cool the motion of micro- or mesoscopic mirrors, initially in the context of gravitational wave detectors. Experimentally, this idea lay dormant until technology improved to the point where such effects could be unambiguously demonstrated~\cite{Cohadon1999,Arcizet2006}, for example by cooling the vibrational motion of a micromechanical oscillator from an occupation number of around $53\,000$ quanta ($2.25$\,K) to ca.~$32$ quanta ($1.3$\,mK)~\cite{Groblacher2009b}.
\par
These two streams can be seen as two limiting cases of a general `matter interacting with a cavity field' configuration, with the matter interacting weakly (atom) or strongly (microscopic mirror) with the cavity field. One of my aims in this thesis will be to combine these two fields through the use of a generic matrix-based theoretical approach. The virtues of this approach are that such models are solvable in the fully general case, and that no restriction is placed on the strength of the light--matter interaction.\\
On a more general level, these ideas can be united with Sisyphus--type cooling mechanisms that make use of internal atomic variables following external ones non-adiabatically. This link comes about because a cavity field also introduces a non-adiabatic `delay' element into the dynamics of the situation. This idea of a system with a delay, or memory, as a means to a generic optical cooling mechanism will also be one of the recurring themes throughout this thesis.
\section*{Outline of thesis}
This thesis will begin with a general, if somewhat brief, overview of atomic physics, which will allow me to discuss several of the ``standard'' cooling or trapping methods in current use. In \cref{ch:CoolingMethods:TrapCool} I will look at what can be viewed as a prototypical system for cavity cooling of atoms, which I call `mirror-mediated cooling', and explore it from the semiclassical perspective.\\
The discussion in \pref{part:TMM} will see me exploring two classical theories of light--matter interactions: one based on the transfer matrix method, in \cref{ch:TMM:TMM}, and one based on a delayed fully vectorial three-dimensional scattering theory, in \cref{ch:TMM:Scattering}. The transfer matrix theory is developed in \sref{sec:TMM:Model}, solved in the general case in \sref{sec:TMM:General}, and extended in \sref{sec:TMM:Multilevel} to describe multi-level atoms. It is then applied to the ‘mirror-mediated cooling’ paradigm and, in \cref{ch:TMMApplications}, to another system, `external cavity cooling', which promises to be very important in optomechanical experiments. As a final application of the matrix formalism, I will explore another configuration, in a unidirectional ring cavity with a gain medium, which allows one to dispense with certain (somewhat stringent) requirements in mirror-mediated cooling. The scattering theory developed in \cref{ch:TMM:Scattering} is also initially applied to the one-dimensional mirror-mediated cooling geometry, in order to verify the method, and is then used to describe mirror-mediated cooling in three dimensions and also the ``optical binding'' of a refractive particle to its own reflection.\\
In the final part, I will first discuss the experimental work that is currently underway in the laboratories at the University of Southampton to explore both the consequences of the above, and atom--surface interactions in a more general context. Finally, I will analyse several of the cooling configurations discussed throughout this thesis from an experimental point of view and give an estimate for the performance that can be expected from each configuration under realistic experimental conditions.
\subsection*{Logical connections between chapters}
\begin{figure}[h]
\centering
\includegraphics[width=\linewidth]{Diagrams/Outline}
\end{figure}
\chapter{Experimental setup}\label{ch:Experimental:Past}
\epigraph{Each piece, or part, of the whole nature is always an approximation to the complete truth, or the complete truth so far as we know it. In fact, everything we know is only some kind of approximation, because we know that we do not know all the laws as yet. Therefore, things must be learned only to be unlearned again or, more likely, to be corrected. [...] The test of all knowledge is experiment. Experiment is the sole judge of scientific ``truth''.}{R.\ Feynman, \emph{The Feynman Lectures on Physics} (1964)}
The mechanisms described in the previous chapters, especially mirror-mediated cooling and external cavity cooling, present several exciting avenues not only for theoretical, but also for experimental, research. This chapter presents an overview of the vacuum and laser systems employed by our group in our ongoing investigations into these mechanisms and into MOT miniaturisation and atomic trap arrays. In the first section I describe the physical makeup of the vacuum and laser systems; the second section describes the novel trap geometry and imaging process employed in our system.
\section{Vacuum and laser system}\label{sec:Experimental:Vacuum}
The aim of the current experiment is to investigate atom--surface interactions. A `clean' cloud of ultracold $^{85}$Rb atoms in a magneto-optical trap was chosen as the starting point for these investigations. In designing the experiment, the apparatus had to satisfy a number of criteria, chiefly:
\begin{itemize}
\item the cold atom cloud needs to be formed close to a surface and the vacuum chamber windows;
\item the surface, or `sample', may be simply plane or structured on the micro- or nano-scale;
\item rapid (on the order of one or two weeks) turnaround time for changing the sample; and
\item very good optical access.
\end{itemize}
Let us look at each of these criteria in turn to explore the design choices they impose on the system.
\subsection{Atom cloud close to surface}
The mirror magneto-optical trap~\cite{Clifford2001} (mirror MOT) was devised as a way to obtain cold atom samples close to a surface. In the present context, however, the standard mirror MOT has a number of disadvantages: the plane of the mirror lies obliquely to the coils that generate the magnetic field necessary to form the MOT, and the beam that forms the MOT illuminates a large cross-section of the mirror directly below the atom cloud---the former will be discussed below in the context of optical access, whereas the latter can be solved by using what we term the `$\Lambda$MOT' beam geometry, as will be discussed in \sref{sec:Experimental:Paper}.
\subsection{Structured surface}\label{sec:Experimental:Vacuum:Structured}
\begin{figure}
\centering
\includegraphics[width=1.5\figwidth]{./Diagrams/Templating}
\caption[Templating process used to make the structured surfaces]{Templating process used to make the structured surfaces. (a) A colloidal suspension of latex spheres in water is allowed to evaporate on a gold substrate within a masked region. (b) An ordered, close-packed monolayer of spheres is formed on the substrate. (c) Gold is electrodeposited and grows from the substrate upwards, around the latex spheres. (d) The spheres and mask are removed by using conventional solvents.}
\label{fig:Templating}
\end{figure}
\begin{figure}[t]
\centering
\fbox{\includegraphics[width=\figwidth]{./Diagrams/PlaneSample}}
\caption[First sample used in the vacuum chamber]{First sample used in the vacuum chamber, pictured being secured into the holder. This sample consisted of a microscope slide onto which a layer of gold was evaporated.}
\label{fig:PlaneSample}
\end{figure}
\begin{figure}[t]
\centering
\fbox{\includegraphics[width=\figwidth]{./Diagrams/TemplatedSample}}
\caption[Second sample used, mounted on the vacuum-compatible translation stage]{Second sample used, mounted on the vacuum-compatible translation stage. This sample was templated with hemispherical concave mirrors; the bright dot in the mirror is the reflection of light off the hemispheres.}
\label{fig:TemplatedSample}
\end{figure}
The interaction between atoms and plane surfaces has been investigated both in the context of near field~\cite{Drexhage1968} and far field effects~\cite{Eschner2001}. The use of structured surfaces allows the coupling of fluorescence from the atom to surface plasmons. A number of advances by Bartlett and co-workers~\cite{Bartlett2000,Bartlett2002} over the past decade have permitted the rapid production of two-dimensional arrays of hemispherical dimples, with radii of $0.1$--$100$\,$\upmu$m, on gold surfaces and with extremely small surface roughness. This process is illustrated schematically in \fref{fig:Templating}; the result is an array of dimples in a small area of an otherwise plane mirror.\\
The initial characterisation of the system in the Southampton Laboratory was conducted using a plane mirror sample, \fref{fig:PlaneSample}. This sample was then replaced by one templated with $100$\,$\upmu$m hemispherical dimples, \fref{fig:TemplatedSample}.
\subsection{Rapid changing of surface}
The physical mechanisms behind the dominant interaction between the atom and the surface depend on the length scale of the structure present on the surface. In the case of a plane sample, the dominant interaction is the Casimir--Polder force in the extreme near field~\cite{Casimir1948,Scheel2009} and the retarded dipole force in the extreme far field (see \sref{sec:CoolingMethods:MMC}; see also Ref.~\cite{Wilson2003}). For a surface with hemispherical dimples having radii of the order of the wavelength, plasmonic effects~\cite{Coyle2001} are expected to dominate. For larger dimples, say those with radii of $10$\,$\upmu$m or greater, the system is approximated better by geometrical optics and one expects the formation of a dipole trap at the focus of each hemisphere. In this regime, one can also envisage depositing magnetically polarised films on the surface that would allow the formation of a microscopic MOT at the focus of each dimple.\\
For these reasons, it is important to be able to switch samples, in order to vary the length scale of the structure, rapidly. This imposes restrictions on the size of the vacuum chamber used: it must be as small as possible without hindering optical access.
\subsection{Good optical access}
\begin{figure}
\centering
\includegraphics[width=\textwidth]{./Diagrams/MCF600-SO200800}
\caption[Technical drawing for the Kimball Physics spherical octagon]{Technical drawing for the Kimball Physics MCF600-SO200800 spherical octagon. All dimensions are in inches (and millimetres in parentheses). Image reproduced from Ref.~\cite{KimballPhysicsSphOct2010}.}
\label{fig:Chamber}
\end{figure}
A good compromise for a vacuum chamber that satisfies the above restrictions was found in the Kimball Physics Spherical Octagon (part number MCF600-SO200800) chamber, shown in \fref{fig:Chamber}. The two large ports and large aspect ratio enable the use of high numerical aperture fluorescence collection optics. The sample close to which the MOT is formed is placed in the centre of the chamber with its plane parallel to the large ports. This in turn allows the use of a purpose-built high-magnification microscope objective (spatial resolution:~ca.~$2$\,$\upmu$m) to be used with its focal plane coincident with the sample.\\
Two of the smaller ports on the chamber are used for the mount holding the sample and for connecting the chamber to the ion pump and $^{85}$Rb dispensers. The other six are therefore left free, allowing almost unrestricted optical access to the cold atom cloud. The sample, \fref{fig:TemplatedSample}, is mounted on a vacuum-compatible translation stage, allowing the templated surface to be positioned correctly with respect to the MOT beams.
\subsection{Laser system}
The MOT cooling and repump beams are generated by external cavity diode lasers similar to the design in Ref.~\cite{Himsworth2009}. More details of the frequency locking system used are given in the next section.
\section{The \texorpdfstring{$\Lambda$}{`Lambda'-}MOT and multiphoton imaging}\label{sec:Experimental:Paper}
\newcommand{$^{85}$Rb}{$^{85}$Rb}
\newcommand{\texorpdfstring{$\Lambda$MOT}{`Lambda'-MOT}}{\texorpdfstring{$\Lambda$MOT}{`Lambda'-MOT}}
The work in this section is published as Ohadi, H., Himsworth, M., Xuereb, A., \& Freegarde, T. Opt.\ Express \textbf{17}, 23003 (2010) and is reproduced \emph{verbatim}~\cite{Ohadi2009}.\footnote{All authors contributed equally to this paper. Hamid Ohadi and Matthew Himsworth performed the measurements; AX and HO processed the data and wrote the paper. Tim Freegarde supervised the project at all stages.} Here, we describe and characterise the combined magneto--optical trap and imaging system that we developed in our laboratory to be able to explore atom--surface interactions in great detail and with considerable experimental flexibility.
\par
We demonstrate a combined magneto--optical trap and imaging system that is suitable for the investigation of cold atoms near surfaces. In particular, we are able to trap atoms close to optically scattering surfaces and to image them with an excellent signal-to-noise ratio. We also demonstrate a simple magneto--optical atom cloud launching method. We anticipate that this system will be useful for a range of experimental studies of novel atom-surface interactions and atom trap miniaturisation.
\subsection{Introduction and motivation}
Over the past two decades, several configurations for magneto--optical traps have been demonstrated~\cite{Raab1987,Shimizu1991,Emile1992,Lee1996,Reichel1999}. The starting point for most geometries has been the original, `6-beam', configuration~\cite{Raab1987}, where the atom trap is created in the intersection of three counterpropagating laser beams. Despite it having the advantage that the atoms can be trapped far from any surface, thereby reducing spurious scatter in the imaging of such a trap, one cannot easily use this configuration for investigations into atom--surface interactions, for precisely the same reason. Another, more recent, configuration is the so-called `mirror MOT'~\cite{Reichel1999}, where the trap is formed a short distance away from a mirror, which also serves to reduce the number of necessary incident laser beam paths to two. The major drawback of such a configuration is its reduced optical access, due to the oblique angle of the field coils with respect to the mirror. The presence of a reflecting surface close to the trap also presents a problem of an entirely different nature. If the object of one's investigation is to observe the interaction between atoms and surfaces structured at the micrometre scale, for example hemispherical mirrors of the type investigated in~\cite{Coyle2001}, the signal from the atoms will almost certainly be lost due to unwanted scattering of light into the optical system. MOTs on the meso- and microscopic scale, in particular, have received some recent interest~\cite{Folman2000}, but the small atom numbers in such traps have so far hindered their imaging and characterisation~\cite{Pollock2009}. In this section we detail a modified configuration that we call the `\texorpdfstring{$\Lambda$MOT}{`Lambda'-MOT}' and implement an imaging system based on a two-stage excitation process~\cite{Nez1993}, which help us overcome each of these limitations and aid our exploration of different atom--surface interactions.
\par
This section is structured as follows. The next subsection is devoted to the description and characterisation of our trap geometry. We then discuss the mechanism behind our multilevel imaging system and show how it does indeed allow for practically background-free imaging of the atom trap. The subsequent subsection discusses surface loading by magneto-optic launching, which allows us to load atoms onto a surface with a three-dimensional range of motion.
\subsection{The \texorpdfstring{$\Lambda$MOT}{`Lambda'-MOT}}\label{sec:MOT}
\subsubsection{Description}
\begin{figure}[t]
\centering
\includegraphics[width=2\figwidth]{./Diagrams/MOT/VWMOT}
\caption[Schematic of one of the two beam paths involved in our MOT geometry]{Schematic of one of the two beam paths involved in our MOT geometry. S is the incoming beam; A, B, and C are mirrors. The component marked `$\lambda/4$' is a quarter-wave plate. The cold atom cloud forms in the intersection region, O. In this diagram we do not show a second, identical, beam, which provides trapping and cooling forces in the plane normal to the paper. The area of mirror A immediately adjacent to the trapped atoms is not illuminated, and can therefore be patterned or structured to explore atom--surface interactions. \emph{Inset:} The lower surface of mirror A, showing the MOT beams and the sample area, which is not illuminated by any of the beams.}
\label{fig:Ohadi2009:VWMOT}
\end{figure}
\begin{figure}[t]
\centering
\includegraphics[width=1.5\figwidth]{./Diagrams/MOT/MOT_Annotated}
\caption[Image of our MOT in operation]{Image of our MOT in operation, corresponding to \fref{fig:Ohadi2009:VWMOT}; mirror A is indicated in the picture.}
\label{fig:Ohadi2009:MOTPic}
\end{figure}
A single beam of circularly polarised light of the right helicity is split using a non-polarising beamsplitter, to generate the two beams that produce the trap, and a half-wave plate is inserted in one of the two resulting beams to achieve the correct polarisations. Each of these beams, denoted S, is then used to construct the geometry shown in \fref{fig:Ohadi2009:VWMOT}. Mirror C is set up so as to retroreflect the beam. Mirrors B and C, together with the quarter-wave plate, allow us to change the polarisation in the retroreflected branch independently of the incoming polarisation. In a normal mirror MOT, the polarisations cannot be modified independently of each other and the quadrupole axis has to be at an oblique angle to the mirror. The four beams travelling towards O thus have the correct polarisations to produce the trapping and cooling forces necessary to form a MOT in this plane. Combined with the second set of beams, this means that the MOT is formed in the intersection region of four pairs of counterpropagating beams. We note that alignment of mirror B such that the beam is retroreflected perfectly will recover the traditional mirror MOT beam geometry, albeit with the incorrect polarisations for a MOT cloud to form.
\par
Several advantages are apparent in the use of this geometry. The trapping volume is the entire overlap of the trapping beams, unlike that in a mirror MOT where half the trapping volume is rendered inaccessible by the presence of the mirror. Optical access is also much improved, both because the coils are oriented in such a way as to be less obstructive, and because we have removed the necessity of having a beam travelling in a plane parallel to mirror A. This allows us to use as much of the $360$\textdegree{} viewing angle in that plane as is necessary for imaging or manipulation beams. If this is not a requirement, a simpler set-up can alternatively be used, where only one set of beams is used in the double-`$\Lambda$' geometry, the trapping and cooling forces in the plane normal to the paper in \fref{fig:Ohadi2009:VWMOT} being produced by means of a separate pair of counterpropagating beams.
\\
An important advantage of this geometry is that the double-`$\Lambda$' shape of the MOT beams affords better imaging of the trap, allowing microscope objectives to be mounted very close to it. With a custom-made objective, we can achieve high-NA imaging ($\mathrm{NA}>0.5$) and a diffraction-limited resolution of $<2$\,$\upmu$m. While a similar degree of optical access may be possible in the traditional 6-beam configuration, we note that this latter configuration is unsuitable for atom--surface interaction studies. In contrast, mirror A in our geometry can be replaced by any other suitable reflecting surface. One candidate for such a reflecting surface would be one of the surfaces of a Dove prism, which could then be used to form a two-dimensional bichromatic evanescent-field trap~\cite{Ovchinnikov1991} close to the mirror surface. This trap would be loaded from the MOT cloud using such techniques as magneto-optic launching, which is explained in \Sref{sec:Launching}.
\par
Aside from this marked increase in optical access, our system is simple to set up and operate. In particular, it requires fewer beam paths than a traditional MOT (two rather than three) and alignment of the beams is also easy: a CCD camera looking up at the mirror can be used to align the beams coarsely; once this is done, optimisation of the cold atom signal provides the fine-tuning of the alignment.
\subsubsection{Characterisation}
A typical trap, shown in \fref{fig:Ohadi2009:MOTPic}, is ellipsoidal in shape with a $1/e$ diameter of the order of $400$\,$\upmu$m along its minor axes and contains around $4\times 10^4$ $^{85}$Rb{} atoms. Combined with a measured trap lifetime $\tau_0\approx 6$\,s, this allows us to infer the trap loading rate, $N_0/\tau_0\approx 6.7\times10^3$\,s$^{-1}$. We measured a cloud temperature of $110\pm 40$\,$\upmu$K, the large uncertainty being due to the imprecision in measuring the cloud size.
\par
Typical parameters for the operation of our trap are: a detuning of $-14.9$\,MHz, or $-5.0$\,$\Gamma$ ($\Gamma\approx 3.0$\,MHz~\cite{Schultz2008}), for the cooling laser and a power of $6$\,mW divided between the two trapping beams (beam diameter: $6$\,mm). The minimum power necessary to produce the MOT was found to be $\approx\!1.3$\,mW in each of the two beams. The trap was loaded from background gas of a natural isotopic mixture of rubidium at a pressure of $10^{-9}$\,mbar. The cooling and repump lasers were locked using the DAVLL technique~\cite{Corwin1998} for long-term stability and flexibility of operation.
\subsection{Multilevel imaging system}\label{sec:Imaging}
\begin{figure}[t]
\centering
\includegraphics[width=\figwidth]{./Diagrams/MOT/420_Levels}
\caption[The four-level system in $^{85}$Rb{} that we use to image our atoms]{The four-level system in $^{85}$Rb{} that we use to image our atoms. The MOT lasers ($780$\,nm) and a laser at $776$\,nm are used to induce a ladder transition. The population decays back to the ground state, via an intermediate state, and emits a $420$\,nm photon in the process. The hyperfine splitting of the excited states is not drawn for clarity.}
\label{fig:Ohadi2009:420_Levels}
\end{figure}
The most common method of imaging a cold atom cloud in a MOT is fluorescence imaging. When the cloud is close to a reflecting surface both the cloud and its reflections will be seen by the imaging system (see Ref.~\cite{Clifford2001}, for example). This situation is exacerbated by the presence of surfaces that reflect unwanted light into the imaging optics and thereby decreasing the signal-to-noise ratio of the imaging system. \fref{fig:Ohadi2009:MOTPic}, shows an example of the mirror in our system scattering the MOT beams into the imaging system.
\\
This problem may be overcome using two-stage excitation imaging. We make use of a four-level system in $^{85}$Rb{} (see \fref{fig:Ohadi2009:420_Levels} for details), similarly to Refs.~\cite{Sheludko2008} and~\cite{Vernier2010}; atoms in the $5$S$_{1/2}$ ground state are pumped to the $5$D$_{5/2}$ state via $780$\,nm and $776$\,nm radiation, the former being provided by one of the MOT beams, and then decay back to the ground state via an intermediate $6$P$_{3/2}$ state, emitting $420$\,nm radiation, which we detect. We note that a very similar system was recently used to produce a multiphoton MOT~\cite{SWu2009}. In our system, this process gives a significantly smaller signal than can be obtained through $780$\,nm fluorescence imaging. However, it has the benefit of being entirely background-free: in a well-shielded system, the entire $420$\,nm signal reaching the detector has its origin in the cold atom cloud. Off-the-shelf filters can then be used to remove the $780$\,nm radiation reaching the detector.
\subsubsection*{Generation of the $776$\,nm beam}
\begin{figure}[t]
\centering
\includegraphics[width=1.5\figwidth]{./Diagrams/MOT/776Locking}
\caption[$776$\,nm spectroscopy and locking system]{$776$\,nm spectroscopy and locking system. (P)BS: (polarising) beam splitter cube; $\lambda/4$: quarter-wave plate; $\lambda/2$: half-wave plate; VC: heated vapour cell; PD: filtered photodiode.}
\label{fig:Ohadi2009:776Locking}
\end{figure}
The $776$\,nm beam is produced using a Sanyo DL7140-201S diode and the same external cavity diode laser design used to produce the MOT cooling and trapping beams. Since $^{85}$Rb{} has no spectral features in this wavelength range that are suitable for locking the laser frequency, a multilevel locking system is used (see \fref{fig:Ohadi2009:776Locking}). $5$\,mW from each of the MOT cooling and repump beams ($\approx\!780$\,nm) and $1.5$\,mW from the $776$\,nm beam, all rendered circularly polarised by the quarter-wave plates, enter the heated vapour cell (VC) from opposite ends. A large-area UV-enhanced filtered silicon photodiode (PD), operating in photovoltaic mode, picks up the resulting Doppler-free fluorescence and is amplified by means of a LMP7721 amplifier chip. Magnetic coils surrounding the heated vapour cell control the Zeeman shift of the magnetic sublevels of the atoms inside the cell, shifting this signal, and therefore the lock point, as required. Around $4$\,mW of the $776$\,nm beam is then mixed in with the MOT cooling and repump beams and sent through an optical fibre to the MOT.
\\
\begin{figure}[t]
\centering
\includegraphics[width=1.5\figwidth]{./Diagrams/MOT/420}
\caption[$420$\,nm fluorescence from the vapour cell as a function of the detuning of the $776$\,nm beam]{$420$\,nm fluorescence from the vapour cell, observed on PD (see \fref{fig:Ohadi2009:776Locking}) as a function of the detuning of the $776$\,nm beam, with the cooling and repump beams locked and shifted by $80$\,MHz with respect to the frequencies required to make a MOT. The various peaks are due the hyperfine structure in $^{85}$Rb{}. To obtain these data, we removed the quarter-wave plates on either end of the vapour cell, thus having linearly polarised light entering the cell from both ends.}
\label{fig:Ohadi2009:420Spectrum}
\end{figure}
\begin{figure}[t]
\centering
\includegraphics[width=1.5\figwidth]{./Diagrams/MOT/420nm-PD_PMT_Inset}
\caption[$420$\,nm fluorescence observed on PD and on a PMT imaging the MOT cloud as a function of the detuning of the $776$\,nm beam]{$420$\,nm fluorescence observed on PD (solid black line, see \fref{fig:Ohadi2009:776Locking}) and on a PMT imaging the MOT cloud (solid red line) as a function of the detuning of the $776$\,nm beam. The zero on the frequency axis corresponds to the point at which the signal from the MOT cloud is highest; we lock to this point. The magnitude and sign of the shift between the two curves can be set arbitrarily by varying the magnetic field generated by the coils around the vapour cell. \emph{Inset:} MOT cloud imaged at $420$\,nm (scale in $10^3$ counts per second). This image is naturally background-free and has a spatial resolution, limited by the optics used, of ca.~$2$\,$\upmu$m.}
\label{fig:Ohadi2009:420nmPDPMT}
\end{figure}
\begin{figure}[t]
\centering
\includegraphics[width=\linewidth]{./Diagrams/MOT/MagLaunch}
\caption[A sequence of four false colour fluorescence images of the cloud before and after it has been given a magnetic impulse]{A sequence of four false colour fluorescence images, taken at $8$\,ms intervals, of the cloud before and after it has been given a magnetic impulse. The first shot (leftmost picture) shows the cloud just before the magnetic field is pulsed. The second, and subsequent, shots show the cloud at later times. The transfer efficiency after $24$\,ms is over $40$\%.}
\label{fig:Ohadi2009:MagLaunch}
\end{figure}
We show a sample $420$\,nm signal, as detected at the photodiode, in \fref{fig:Ohadi2009:420Spectrum}, where the hyperfine splitting of the $5$D$_{5/2}$ level in $^{85}$Rb{} is evident in the shoulders on the right-hand side of the peak in the figure. The $776$\,nm laser diode is locked to the side of one the main peak, at the point indicated by the dashed line in \fref{fig:Ohadi2009:420nmPDPMT}, using a conventional PID circuit. The lock point is found by manually and slowly adjusting the frequency offset of the $776$\,nm laser to maximise the fluorescence from the MOT cloud. Depending on the parameters chosen, the spectra corresponding to the latter figure may exhibit two well-resolved peaks due to the Autler--Townes splitting~\cite{Autler1955}; in the spectrum shown in \fref{fig:Ohadi2009:420nmPDPMT}, however, the presence of the second peak manifests itself as a slight shoulder on the photodiode signal. Locking at a detuning of around $6.5$\,MHz from the peak of absorption in the vapour cell gives the strongest signal in the MOT cloud, as recorded by the photomultiplier tube trace shown in the same figure.
\subsection{Surface loading by magneto-optic launching}\label{sec:Launching}
Transporting cold atoms from the region where the trap naturally forms to the sample is an essential part of many experiments investigating atom--surface effects. Several methods have been devised for moving cold atom clouds, including the use of push beams~\cite{Wohlleben2001} and moving magnetic coils~\cite{Lewandowski2003}. Push beams are easy to set up, requiring either the addition of one extra beam or the switching off of one of the counterpropagating beams, but cannot be used to push atom clouds towards highly reflective surfaces. Using moving magnetic coils requires a rather involved mechanical setup.
\par
We make use of a third method, which we call magneto-optic launching, for transport of the atom cloud by rapidly moving the trap centre and then releasing the cloud, thereby imparting momentum to it. An auxiliary coil is added to the system in~\fref{fig:Ohadi2009:VWMOT}, above the upper MOT coil. After the MOT cloud forms, a long current pulse is applied to this auxiliary coil, which launches the cloud upward with a speed determined by the size and duration of the current pulse, and then the cloud is released from the trap by switching off the MOT beams after $20$\,ms. \fref{fig:Ohadi2009:MagLaunch} shows a series of photographs of the cloud after being launched by a magnetic pulse. It can be seen that the pulse results in an approximately uniform vertical cloud speed of $0.063$\,m\,s$^{-1}$. The physical orientation of our system, with the mirror and sample being \emph{above} the trapping region, allows us to launch the cloud upwards with a much greater degree of control than would be possible if the cloud were merely dropped downwards.
\par
Finally, we note that the equilibrium distance of the MOT cloud from the mirror surface depends on the beam diameter and the size of the `sample area', i.e., the section of the mirror that acts as a sample and is not usable as a plane mirror. With a sample area diameter of $2$\,mm and beam diameter of $4$\,mm, the cloud can be made to form less than $4$\,mm away from the surface, allowing us to use the magneto-optic launching method to move the atoms closer to the surface for interaction studies.
\chapter{A guide for future experiments}\label{ch:Experimental:Future}
\epigraph{[...] [I]t is more important to have beauty in one's equations than to have them fit experiment.}{P.\ A.\ M.\ Dirac, Scientific American \textbf{208}, 5 (1963)}
This chapter aims to provide a guide for experimentalists seeking to observe the effects we predicted in earlier chapters. \sref{sec:Experimental:Future:Overview} presents an overview of the different geometries explored in the previous chapters. For each of these geometries, \sref{sec:Experimental:Future:Forces} calculates and compares the relevant friction forces acting on the particle. \sref{sec:Experimental:Future:Numbers} discusses a number of experimentally-accessible configurations and calculates cooling times and equilibrium temperatures that can be expected in each situation. The first appendix to this chapter is a technical note discussing electric fields inside dielectrics and the origin of the Clausius--Mossotti equation that describes the response of a bulk dielectric to an applied electric field. Finally, two appendices then follow that discuss, respectively, some problems encountered when calculating electric fields inside microscopic hemispherical mirrors, and general expressions for the force acting on an atom inside an arbitrary monochromatic field, ignoring delay effects.
\section{Overview of several different possibilities}\label{sec:Experimental:Future:Overview}
It has been outlined in the preceding chapters that several different geometries exist that allow a memory. Moreover, within each such geometry, one can choose to investigate cooling mechanisms on different classes of particle. The aim of this section is to briefly summarise these different possibilities, noting the advantages and disadvantages of each: \sref{sec:Experimental:Future:Overview:Ions} to \sref{sec:Experimental:Future:Overview:Dielectric} look at species that can be cooled, and \sref{sec:Experimental:Future:Overview:DipoleArrays} to \sref{sec:Experimental:Future:Overview:ConcaveMirror} at the geometries themselves.
\subsection{Trapped ions}\label{sec:Experimental:Future:Overview:Ions}
Ions can be trapped in radio-frequency traps, and laser cooled down to the ground vibrational state of such traps,\footnote{It must be pointed out that such ions would have a translational temperature lower than that which can be achieved through mirror-mediated cooling setups using typical experimental parameters. The aims of such experiments would be (i)~a proof-of-principle demonstration, and (ii)~an exploration of the wavelength-scale variations of the forces.} with remarkable ease. The lifetime of trapped ions can be of the order of hours~\cite{Herskind2008}, which is orders of magnitude longer than the comparable figure for neutral atoms. For these reasons, ions would make ideal test subjects for exploring forces that vary significantly over length scales of the order of a wavelength~\cite{Eschner2001,Hetet2010}. Interest in using trapped ions also arises from their potential applications in quantum information storage and processing~\cite{Cirac1995}.
\subsection{Neutral atoms}\label{sec:Experimental:Future:Overview:Atoms}
The ease of manipulation of trapped ions using electric fields is a double-edged sword, in the sense that this very feature also makes trapped ions highly sensitive to the environment they are immersed in. One can avoid these issues through the use of neutral atoms rather than ions. Neutral atoms, however, cannot easily be confined to sub-wavelength regions without complex experimental systems such as the one used in a recent proof-of-concept experiment presented in Ref.~\cite{Proite2010}.
\subsection{Optomechanics---Cantilevers and micromirrors}\label{sec:Experimental:Future:Overview:Optomechanics}
Recent years have seen a surge in the popularity of optomechanics experiments, mostly with the aim of reaching the ground vibrational state of a vibrating reflective cantilever~\cite{Groblacher2009a}, or of a reflective micro-membrane~\cite{Thompson2008}. The use of such optical elements introduces a number of interesting possibilities:
\begin{itemize}
\item Engineered internal resonances---Ref.~\cite{Karrai2008} looks at using resonances inside a photonic crystal as a means of controlling its motion, in much the same way as one uses atomic resonances in Doppler cooling. In contrast with the case of an atom or ion, however, one can engineer the system to have a wide range of different properties.
\item Strong mirror--field coupling---Single atoms or ions do not have a large polarisability unless the driving field is close to resonance; this is problematic because strong heating effects become important under such conditions. Mirrors, even microscopic ones, consist of vast numbers of atoms, each of which is essentially an individual dipole, and therefore experience correspondingly stronger effects.
\item Positioning---micromirrors enjoy the advantages of both ease of positioning, shared with ions, and the immunity to electrostatic forces, shared with neutral atoms. Moreover, the technology exists to make silicon nitride (SiN) membranes much thinner than an optical wavelength~\cite{Jayich2008}, so such mirrors can indeed be used to explore sub-wavelength structure in the forces.
\end{itemize}
\subsection{Dielectric particles}\label{sec:Experimental:Future:Overview:Dielectric}
Spherical dielectric particles, of sizes on the nanometre~\cite{Chang2009b,Barker2010} or micrometre~\cite{Barker2010b} scales, have been proposed as replacements for individual atoms in cooling experiments. On the small end of the scale, the particles can be suspended using purely optical forces and are therefore not coupled to any physical heat bath. Nevertheless, even such small particles exhibit polarisabilities much larger than that of an individual atom, and therefore correspondingly stronger interactions with the light field. The larger particles would need to be physically supported, perhaps by being mounted on the end of a tapered fibre. This could, in turn, be achieved by ablating the tapered end of the fibre to form a microsphere. The mechanical properties of such a fibre would ensure poor coupling of phonons between the microsphere and the bulk fibre.
\par
The response to the electric field of a dielectric particle on the nanometre scale is related to its complex permeability $\epsilon$ through the Clausius--Mossotti relation:
\begin{equation}
\label{eq:ClausiusMossotti}
\chi=3V\frac{\epsilon-1}{\epsilon+2}\,.
\end{equation}
The derivation of the Clausius--Mossotti equation itself has some interesting subtleties that are discussed in \aref{sec:Experimental:Future:CM}. $\epsilon$ is furthermore related to the real, $\eta$, and imaginary, $\kappa$, parts of the refractive index $n=\sqrt{\epsilon}$ by
\begin{align}
\re{\epsilon}&=\eta^2-\kappa^2\,\text{, and}\\
\im{\epsilon}&=2\eta\kappa\,.
\end{align}
Finally, $\kappa$ is related to the $1/e$ power absorption length of a substance, $1/\alpha$, by $\alpha=2k\kappa$ at wavenumber $k$. For the common dielectric PMMA [poly(methyl methacrylate)] $\eta=1.5$ and, conservatively, $\alpha=50$\,m$^{-1}$ at a wavelength $\lambda=1$\,$\upmu$m in vacuum \cite{Beyer2003}; i.e., $\epsilon=2.2+\bigl(1.2\times 10^{-5}\bigr)\i$. The imaginary part of $\epsilon$ is much smaller than the real part and can generally be neglected when calculating optical forces. It is, however, responsible for absorption of part of the incident light and its effects cannot be neglected when calculating the power absorbed by an illuminated dielectric.
\subsection{Dipole trap arrays}\label{sec:Experimental:Future:Overview:DipoleArrays}
\begin{figure}
\centering
\fbox{\includegraphics[width=\figwidth]{./Diagrams/Dimples}}
\caption[Optical micrograph of a section of a templated gold surface]{Optical micrograph of a section of a templated gold surface. The bright spot in each of the dimples is the focus. The centre--to--centre distance is $100$\,$\upmu$m and the depth of the electrodeposited gold is $15$\,$\upmu$m. (Courtesy Nathan Cooper.)}
\label{fig:Dimples}
\end{figure}
Two-dimensional arrays of dipole traps have been built using microfabricated lens arrays~\cite{Dumke2002} and used to site-selectively address trapped atoms~\cite{Kruse2010}. However, the use of refractive elements such as lens arrays brings with it a number of disadvantages, of which we mention two:
\begin{itemize}
\item Fabrication costs---each array has to be custom-made and it is difficult to mass-produce such optical elements.
\item Integration---lens arrays require optical access from both sides, leading to a larger apparatus and making it difficult to integrate them into so-called atom chips~\cite{Folman2000}.
\end{itemize}
One way to overcome both these issues is to use arrays of reflective concave mirrors, which can be manufactured easily (this was explored in \fref{fig:Templating}). The end product, as shown in \fref{fig:Dimples}, is indeed close to ideal in terms of periodicity and surface quality. These concave mirrors can be used to construct individual dipole traps for either neutral atoms or ions.\par
The interest in using such arrays of mirrors lies not only in their ease of manufacture but also in the surface plasmon resonances that are exhibited by the individual hemispherical mirrors~\cite{Coyle2001}. Such resonances couple to the incident light and give rise the possibility of mechanisms of the ``external cavity cooling'' type, \sref{sec:TMM:ECCO}, using not Fabry--P\'erot cavities but material resonances. The advantages of using such a system are immediately obvious; we mention only that such a setup introduces the possibility of two-dimensional arrays of individual optical resonant elements that require essentially no alignment. Coupled with the fact that external cavity cooling, as with any mechanism based on the dipole force, is not species-selective, this leads to the possibility of producing two-dimensional arrays of cold ions, atoms, or even micromirrors. Such arrays would potentially revolutionise quantum information processing by implementing a scalable two-dimensional register for quantum information.
\subsection{Plane mirror cooling}\label{sec:Experimental:Future:Overview:MMC}
\begin{figure}
\centering
\includegraphics[scale=0.2]{./Diagrams/Realistic_MMC}
\caption[Schematic of an experiment to explore mirror-mediated cooling]{Schematic of an experiment to explore the mirror-mediated cooling mechanism. Light is focussed onto an ion trapped in an endcap trap, then coupled into a fibre. This light is retroreflected back into the fibre by means of a mirror whose position can be adjusted through piezoelectric elements.}
\label{fig:Realistic_MMC}
\end{figure}
The mirror-mediated cooling setup would be the most basic proof-of-concept experiment of optical cooling using a memory, seeing as it involves merely one mirror. As has been discussed previously, the friction force in such a setup oscillates on a sub-wavelength scale and, moreover, is only sizeable for atom--mirror distances of the order of metres. A realistic approach to implementing this delay is to couple the light, after interacting with the atom, into a single-mode fibre. The light inside the fibre is then retroreflected and imaged back onto the atom itself. This setup, shown schematically in \fref{fig:Realistic_MMC}, is conceptually similar to the one used in Ref.~\cite{Eschner2001}. In \fref{fig:Realistic_MMC}, the species to be cooled is shown to be a trapped ion rather than a neutral atom. The reason for this is again the small length scale over which the friction force changes from a cooling to a heating force, necessitating very good localisation of the particle to be cooled. One also notes that the delay line length must be stabilised interferometrically.
\subsection{External cavity cooling}\label{sec:Experimental:Future:Overview:ECCO}
The use of an optical resonance to enhance the cooling effect of the retarded dipole--dipole interaction presents a novel way of enhancing the performance of current optomechanical experiments. Indeed, such setups would be less sensitive to misalignment of the mirror to be cooled than traditional optomechanical setups~\cite{Aspelmeyer2010}, which require micromirrors with extremely good optical and mechanical properties. The sub-wavelength modulation of the friction force is an issue with this mechanism too, which necessitates the use of particles---such as thin micromirrors, membranes, or trapped ions---that can be localised to within a small fraction of a wavelength.
\subsection{Ring cavity cooling}\label{sec:Experimental:Future:Overview:RELIC}
\begin{figure}
\centering
\includegraphics[width=\figwidth]{./Diagrams/3drelic}
\caption[Highly schematic illustration of three-dimensional ring cavity cooling]{Highly schematic illustration of a three-dimensional ring cavity cooling setup. The pump light does not couple directly into any cavity, allowing one to use gain media in the fibres similarly to the amplified optomechanics scheme (see \sref{sec:TMM:AmplifiedOptomechanics}).}
\label{fig:3DRELIC}
\end{figure}
One may wish to do away with the sub-wavelength localisation problem inherent in mirror-mediated cooling altogether. Ring cavity cooling, discussed under the guise of ``amplified optomechanics'' in \sref{sec:TMM:AmplifiedOptomechanics}, provides one way of achieving this aim. This mechanism works best with particles that are rather poorly reflective, otherwise the advantages of the amplifier gain are lost, and is therefore more suited towards the cooling of atoms rather than micromirrors. In this instance, both neutral atoms and trapped ions are good candidates; it must also be mentioned that in the case of ions, cooling down to a very low vibrational state is \emph{not} needed in this case. The constraint on the delay line length is not lifted, however:~this must still be stabilised interferometrically.\par
An extension of this scheme can be envisaged where the pump light is not injected into the cavity but is directed off-axis at the particle. This geometry would not require an isolator, since the pump light never enters the cavity, and can easily be extended to three dimensions; see \fref{fig:3DRELIC}.
\subsection{Concave mirror cooling}\label{sec:Experimental:Future:Overview:ConcaveMirror}
\begin{figure}[t]
\centering
\subfigure[]{
\includegraphics{Diagrams/ConcaveMirrorRacetrackNoDelay}
}\hspace{1cm}
\subfigure[]{
\includegraphics{Diagrams/ConcaveMirrorRacetrackDelay}
}
\caption[Origin of the friction force for motion of a particle around the centre of a hemispherical mirror]{Origin of the friction force for tangential ``racetrack'' motion of a particle around the centre of curvature of a hemispherical mirror. (a) In the absence of delay, the particle and its image are at diametrically opposite points in the circular locus of the motion of the particle. (b) Due to delay, the image trails behind its equilibrium position slowing the particle down due to the repulsive interaction between the two. Note that the length scales are greatly exaggerated; in reality, the particle, and the path it moves along, would be much smaller than depicted.}
\label{fig:ConcaveMirrorRacetrack}
\end{figure}
A final possibility is that of using the reflection from a concave hemispherical mirror itself, and not any material resonances supported by the mirror, to cool the motion of a particle at the centre of curvature of the mirror. One would expect that for the effect to be sizeable, the mirror radius would need to be much larger than a wavelength, perhaps of the order of millimetres or centimetres. A brief justification for why a cooling effect is expected to exist in such a geometry is possible using the ideas developed in \sref{sec:CoolingMethods:Retarded} in the case where the particle experiences a repulsive interaction with its image. Such a case could correspond to an atom illuminated by light tuned to the blue of its resonance.\par
When the particle is near the centre of the mirror, a real image is formed that in steady-state is at the same distance from, but on the opposite side of, the centre. Given the symmetry of the situation, it is enough to decompose the motion of the particle into two orthogonal motions, radially and tangentially in a polar coordinate system centred on the centre of curvature of the mirror. In the case of radial motion, we will appeal to \eref{eq:GeneralTimeDelayedF}. The repulsive potential $U$ seen by a motionless particle at a small distance $r$ from the centre can be described by a Gaussian\footnote{To a good approximation, a small particle close to the centre of a spherical mirror produces a spherical Gaussian image upon focussing by the mirror~\cite{Siegman1990}.}, having a peak at $r=0$ and tailing off to zero as $r$ increases. For concreteness, let us take
\begin{equation}
U(\tilde{r},r,v=0)=U_0\exp\bigl[-(r+\tilde{r})^2/w^2\bigr]\,,
\end{equation}
using the notation of \sref{sec:CoolingMethods:Retarded}, where $w$ is the width of the potential, taken to be the same in every direction, and $U_0$ is non-negative. Thus, using \eref{eq:GeneralTimeDelayedF}, we obtain the delayed friction force
\begin{equation}
\textsl{\textsf{F}}=2U_0\tau v\frac{8r^2-w^2}{w^4}\exp\bigl(-4r^2/w^2\bigr)\,,
\end{equation}
which is a cooling force for small $r$, and where $\tau=2R/c$ is the delay time for a mirror of radius of curvature $R$.\\
The origin of the cooling force for motion in the tangential direction is best explained pictorially. \fref{fig:ConcaveMirrorRacetrack} shows the particle moving in a circle of constant radius around the centre of the mirror. In the absence of delay, \fref{fig:ConcaveMirrorRacetrack}(a), the particle and its image would be in diametrically opposite positions along this circle. Due to the delay, however, the image lags behind its equilibrium position, as shown in \fref{fig:ConcaveMirrorRacetrack}(b). The repulsive interaction between the particle and its image thereby produces a force opposing the motion of the particle. This force can easily be seen to increase with both $v$ and $\tau$, since the distance between the particle and its image decreases with both these quantities, and therefore acts to cool the tangential motion of the particle. Putting these two arguments together implies that a particle undergoing any motion around the centre of curvature of a spherical mirror experiences a friction force mediated by the delayed dipole--dipole interaction.
\section{Cooling forces experienced in different geometries}\label{sec:Experimental:Future:Forces}
In this section we shall use the three-dimensional scattering theory developed in \cref{ch:TMM:Scattering} to obtain the cooling forces experienced by a point-like dipole in a number of different configurations, both (quasi-)one-dimensional and three-dimensional.
\subsection{Longitudinal mirror-mediated cooling}
In one dimension, the mirror-mediated cooling force is given, in agreement with \eref{eq:Scattering1DMMCForce}, as
\begin{equation}
\textsl{\textsf{F}}=-\frac{1}{2c}\epsilon_0\frac{\mathcal{E}_0^2}{\sigma_\text{L}}\chi^2k^3x\sin(4kx)v+\mathcal{O}\bigl(\chi^4\bigr)\,,
\end{equation}
where $x$ is the position of the particle with respect to the mirror, and $\chi$ is the polarisability of the particle. Note that this force is oscillatory in the position of the particle and therefore has a zero spatial average on the wavelength scale.\par
A close analogue to the one-dimensional case in three dimensions is that of longitudinal pumping, where the pump field forms a standing wave with the mirror; this is the case considered earlier. When the electric field is assumed to be circularly polarised in the $y$--$z$ plane, there arises a dominant (oscillatory) component of the friction force,
\begin{multline}
\label{eq:Future:MMCLongitudinalNoAverageFriction}
\left(\begin{array}{c}
\textsl{\textsf{F}}_x\\
\textsl{\textsf{F}}_y\\
\textsl{\textsf{F}}_z
\end{array}\right)=\frac{1}{32\pi c}\epsilon_0\mathcal{E}_0^2\chi^2k^4\frac{1}{kx}\\
\times\left[\begin{array}{ccc}
4kx\cos(4kx)+\sin(2kx)+3\sin(4kx) & 0 & 0\\
0 & \cos(kx)\sin^3(kx) & 0\\
0 & 0 & \cos(kx)\sin^3(kx)
\end{array}\right]\\
\cdot\left(\begin{array}{c}
v_x\\
v_y\\
v_z
\end{array}\right)+\mathcal{O}\bigl[\chi^2/(kx)^2\bigr]+\mathcal{O}\bigl(\chi^3\bigr)\,,
\end{multline}
which averages over a wavelength\footnote{The averaging is done by assuming $kx\gg 1$, which allows us to hold the $1/(kx)^n$ terms constant and average only over the periodic functions.} to give the non-oscillatory terms,
\begin{multline}
\label{eq:Future:MMCLongitudinalAverageFriction}
\left(\begin{array}{c}
\textsl{\textsf{F}}_x\\
\textsl{\textsf{F}}_y\\
\textsl{\textsf{F}}_z
\end{array}\right)=-\frac{1}{128\pi c}\epsilon_0\mathcal{E}_0^2\chi^2k^4\frac{1}{(kx)^2}\left[\begin{array}{ccc}
1 & 0 & 0\\
0 & 3 & 0\\
0 & 0 & 3
\end{array}\right]\cdot\left(\begin{array}{c}
v_x\\
v_y\\
v_z
\end{array}\right)\\
-\frac{1}{1024\pi c}\epsilon_0\mathcal{E}_0^2\chi^3k^7\frac{1}{(kx)^3}\left[\begin{array}{ccc}
4 & 0 & 0\\
0 & 3 & 0\\
0 & 0 & 3
\end{array}\right]\cdot\left(\begin{array}{c}
v_x\\
v_y\\
v_z
\end{array}\right)+\mathcal{O}\bigl(\chi^4\bigr)\,.
\end{multline}
It must be noted that these expressions are valid for $kx$ large enough that the point-dipole approximation holds ($kx\gg 1$). In order to explore the physical origin of the above forces, it is helpful to introduce some auxiliary notation. We denote the pump field $\v{\mathcal{E}}_0$, as before, and the polarisation it induces in the particle $\v{\mathcal{P}}_0$. $\v{\mathcal{P}}_0$ is responsible for producing an electric field $\v{\mathcal{E}}_1$. Upon reflection by the mirror, $\v{\mathcal{E}}_1$ is in turn responsible for inducing a polarisation $\v{\mathcal{P}}_1$ in the particle. For each index $i$, $\v{\mathcal{E}}_i$ is of the order $\chi^i$ and $\v{\mathcal{P}}_i$ of the order $\chi^{i+1}$. The force acting on the particle due to a term of the form $\v{\mathcal{P}}_i^\ast\cdot\v{\mathcal{E}}_j$ is thereby of order $\chi^{i+j+1}$. Implicit in each of the above friction force expressions is a factor of order $x$ due to the retardation effects. This has to be understood as being physically separate from the factors of $x$ introduced by the spreading of wavefronts in three dimensions; indeed, in one dimension $\textsl{\textsf{F}}\propto x$ despite the fact that wavefronts do not spread. We will factor this term out in the following.\\
Let us first decompose the friction force in \eref{eq:Future:MMCLongitudinalNoAverageFriction}: there are two sets of terms, of order $\chi^2/(kx)$ and $\chi^2/(kx)^2$, both of which arise from interactions of the form $\v{\mathcal{P}}_0^\ast\cdot\v{\mathcal{E}}_1$ and $\v{\mathcal{P}}_1^\ast\cdot\v{\mathcal{E}}_0$. Respectively, these interactions represent the interaction of the polarisation induced by the incident field with the retarded re-radiated field, and that of the polarisation induced by this retarded field with the incident electric field. The first term in the non-zero spatially averaged force, \eref{eq:Future:MMCLongitudinalAverageFriction}, is in this case entirely due to the geometrical spreading out of the wavefronts in the same terms. The second term, which arises from a term that is not written explicitly in \eref{eq:Future:MMCLongitudinalNoAverageFriction}, has a different origin:~it arises from the phase-locked interaction between $\v{\mathcal{P}}_1$ and $\v{\mathcal{E}}_1$, which has no sub-wavelength spatial dependence, since $\v{\mathcal{E}}_1$ is a travelling wave and therefore any sub-wavelength dependences are factored out of the product $\v{\mathcal{P}}_1^\ast\cdot\v{\mathcal{E}}_1$. It is this friction force that therefore arises from a delayed `self-binding' interaction.
\subsection{Transverse mirror-mediated cooling}
In three dimensions one is of course free to choose the direction of illumination. Let us again consider the usual mirror-mediated cooling geometry, but where the pump light is a circularly polarised wave in the $x$--$y$ plane travelling in the $z$-direction, i.e., propagating parallel to the mirror. The resulting friction forces can again be written down as
\begin{multline}
\label{eq:Future:MMCTransverseNoAverageFriction}
\left(\begin{array}{c}
\textsl{\textsf{F}}_x\\
\textsl{\textsf{F}}_y\\
\textsl{\textsf{F}}_z
\end{array}\right)=\frac{1}{32\pi c}\epsilon_0\mathcal{E}_0^2\chi^2k^4\frac{1}{kx}\\
\times\left[\begin{array}{ccc}
-2kx\cos(2kx) & 0 & -2kx\cos(2kx)-\sin(2kx)\\
0 & \sin(2kx) & 0\\
2kx\cos(2kx) & 0 & 2kx\cos(2kx)+\sin(2kx)
\end{array}\right]\\
\cdot\left(\begin{array}{c}
v_x\\
v_y\\
v_z
\end{array}\right)+\mathcal{O}\bigl[\chi^2/(kx)^2\bigr]+\mathcal{O}\bigl(\chi^3\bigr)\,,
\end{multline}
and averaged over a wavelength to give
\begin{equation}
\label{eq:Future:MMCTransverseAverageFriction}
\left(\begin{array}{c}
\textsl{\textsf{F}}_x\\
\textsl{\textsf{F}}_y\\
\textsl{\textsf{F}}_z
\end{array}\right)=-\frac{1}{512\pi c}\epsilon_0\mathcal{E}_0^2\chi^3k^7\frac{1}{(kx)^3}\left[\begin{array}{ccc}
2 & 0 & 0\\
0 & 1 & 0\\
0 & 0 & 0
\end{array}\right]\cdot\left(\begin{array}{c}
v_x\\
v_y\\
v_z
\end{array}\right)+\mathcal{O}\bigl(\chi^4\bigr)\,.
\end{equation}
The oscillatory forces have the same origin as those in the longitudinal case. Note, however, that because of the geometry of this situation the only non-oscillatory forces that survive to this order are due to the self-binding terms of the form $\v{\mathcal{P}}_1^\ast\cdot\v{\mathcal{E}}_1$. In a transverse-pumping geometry, it is this self-binding force that is the dominant effect.
\subsection{Ring cavity cooling}
The mechanism involved in ring cavity cooling is essentially the same as that involved in (one-dimensional) mirror-mediated cooling. Indeed, we can compare \eref{eq:TMMFrictionSimple} to \eref{eq:TMM:MirrorCoolForce} and deduce that the dominant friction force acting on a particle inside a ring cavity will be of the form
\begin{equation}
\textsl{\textsf{F}}=-\frac{1}{4c}\epsilon_0\frac{\mathcal{E}_0^2}{\sigma_\text{L}}\chi^2k^3\Lambda Lv\,,
\end{equation}
where $\Lambda$ is a factor due to the properties of the cavity, and $L$ the length of the cavity.
\subsection{Summary: Orders of magnitude}
The table below is a convenient reference for the forms of the dominant friction forces acting on a particle in the geometries discussed in this section. Symbols of the form $\textsl{\textsf{F}}_{xx}$, for example, denote the force in the $x$ direction proportional to the $x$-component of the velocity.\par
\begin{center}
\begin{tabular}{l l | c c}
\hline
\hline
& & Oscillatory terms & Spatial average\\
\hline
\multicolumn{2}{l|}{Mirror-mediated cooling (MMC; 1D)} & $\chi^2(kx)$ & 0\\
\multirow{2}{*}{MMC (3D)\hspace{1em}$\Biggl\{$} & Longitudinal & $\chi^2/(kx)$ [$\textsl{\textsf{F}}_{xx}$: $\chi^2$] & $\chi^2/(kx)^2$\\
& Transverse & $\chi^2$ [$\textsl{\textsf{F}}_{yy}$: $\chi^2/(kx)$] & $\chi^3/(kx)^3$ [$\textsl{\textsf{F}}_{zz}$: $\chi^2/(kx)^5$]\\
\multicolumn{2}{l|}{Ring cavity cooling (1D)} & --- & $\chi^2(kx)$\\
\hline
\end{tabular}
\end{center}
We reiterate that the $\chi^2$ and $\chi^3$ terms forming the spatial average of the friction forces have different origins, with the former being due to the spreading of wavefronts and the latter the phase-locked interaction between $\v{\mathcal{P}}_1$ and $\v{\mathcal{E}}_1$.
\section{Cooling times and base temperatures}\label{sec:Experimental:Future:Numbers}
In the previous sections we identified a number of configurations that may be used to show cooling effects mediated by the dipole--dipole force. The aim of this section is to evaluate the cooling forces and base temperatures for a few specific situations.
\subsection{One-dimensional mirror-mediated cooling: Trapped ion}
The `fine structure' in the friction force produced in the mirror-mediated cooling geometry in one dimension is perhaps best explored using trapped ions, as shown schematically in \fref{fig:Realistic_MMC}. We assume that the species being cooled is a Ba$^+$ ion, and that the pump beam is at a wavelength of $\lambda=493$\,nm, detuned by $\Delta=\pm 10\Gamma$ from resonance and focussed down to a $10$\,$\upmu$m spot at the position of the ion. The length of the fibre-based delay line is taken to be $L=2$\,m. In order to ensure operation in the low-saturation regime, we set the pump power to be $P=1$\,nW. The mass of the ion is $m=2.3\times 10^{-25}$\,kg. These numbers result in:
\begin{center}
$1/e$ velocity cooling time: $12$\,ms, and\\
steady-state temperature: $570$\,$\upmu$K.
\end{center}
Both of these numbers decrease linearly with increasing $L$ and include a factor of $\tfrac{1}{2}$ originating from the presence of the harmonic trap confining the motion of the ion, as explained in \sref{sec:CoolingMethods:MMC:Perturbative:Dipole}. Doppler cooling of the ion, after which the position spread of the particle would be smaller than a wavelength (specifically, ca.~$35$\,nm in the setup of Ref.~\cite{Eschner2001}), would be necessary to resolve the sub-wavelength features in the friction force.
\subsection{External cavity cooling: Transmissive membrane}
We have already remarked that external cavity cooling would act to enhance optomechanical cooling mechanisms, such as those used to cool the vibrational motion of reflective mirrors. Let us suppose we use a membrane with power transmissivity of $50$\% and couple the transmitted light into a Fabry-P\'erot cavity, whose mirrors have power transmissivities of $1$\%. A typical commercially-available SiN membrane would have an effective mass $m=5\times 10^{-14}$\,kg~\cite{Thompson2008} for the centre-of-mass mode and negligible absorption. Coupling in $1$\,mW of light at a wavelength $\lambda=780$\,nm thereby gives
\begin{center}
$1/e$ velocity cooling time: $1$\,ms, and\\
steady-state temperature: $35$\,$\upmu$K,
\end{center}
with the steady-state temperature being a lower limit in the absence of any absorbed power or coupling to a substrate.
\subsection{Amplified optomechanics: Neutral atom}
The friction force in amplified optomechanics, or ring cavity cooling in general, is not dependent on the position of the particle to be cooled. To explore this mechanism, we therefore suggest using neutral atoms. Our atom of choice is $^{85}$Rb (mass $m=1.4\times 10^{-25}$\,kg), which can be Doppler-cooled and confined to a small cloud beforehand. We assume that the pump beam, of wavelength $\lambda=780$\,nm and detuning $\Delta=-10\Gamma$, is focussed down to a $10$\,$\upmu$m spot at the position of the atom, and that the power at the input coupler is $13$\,nW to guarantee operation in the low-saturation regime. The fibre-based cavity is taken to be $300$\,m long, the power loss at each of the two couplers terminating the cavity is assumed to be $50$\%, and the input coupler to transmit only $1$\% of the incident power. A gain medium, with gain $1.75$, is assumed to form part of the cavity. We therefore obtain
\begin{center}
$1/e$ velocity cooling time: $4$\,ms, and\\
steady-state temperature: $40$\,$\upmu$K.
\end{center}
\appendicesstart
\section{Appendix: Electric fields inside dielectrics}\label{sec:Experimental:Future:CM}
Within a simple model of the dielectric of volume $V$ as a collection of closely spaced dipoles at random positions, its response to an electric field can be embodied entirely in the relative permittivity $\epsilon$, defined by the relation
\begin{equation}
\v{\mathcal{P}}=V(\epsilon-1)\epsilon_0\v{\mathcal{E}}\,,
\end{equation}
The Clausius--Mossotti relation, as quoted in \eref{eq:ClausiusMossotti}, connects the susceptibility $\chi$ of the dielectric to the bulk refractive index of the dielectric. Many derivations of this relation (see, for example, Ref.~\cite[\textsection 4.5]{Jackson1998}) divide the dielectric into two regions: a spherical section of the dielectric, large enough to contain several dipoles but small enough that the polarisation is practically constant within; and the rest of the dielectric. One objection to this argument is that it depends critically on the first region chosen as being spherical, an assertion that has no real physical justification and no immediate connection to the main assumption, mentioned below, present in the Clausius--Mossotti relation. Hannay~\cite{Hannay1983} presented an alternative derivation of this same expression that does not make use of this model. Let us briefly recapitulate Hannay's argument.\\
The electric field $\v{\mathcal{E}}$ produced by an ideal point-like dipole $\v{p}$ is, as a function of the displacement $\v{r}$ from the dipole,
\begin{equation}
\v{\mathcal{E}}=\frac{1}{4\pi\epsilon_0}\Biggl[\frac{3(\v{p}\cdot\v{r})\v{r}}{\lvert\v{r}\rvert^5}-\frac{\v{p}}{\lvert\v{r}\rvert^3}\Biggr]-\frac{\v{p}}{3\epsilon_0}\delta(\v{r})\,.
\end{equation}
The central assumption that leads to the Clausius--Mossotti equation is that a ``test'' dipole inserted at a random position in the dielectric experiences an electric field that is free from the influence of the $\delta$-spikes that occur at each of the dipoles making up the dielectric. These $\delta$-spikes, of which there are $N$, contribute a spatially-averaged field
\begin{equation}
-\frac{N\v{p}}{3V\epsilon_0}=-\frac{\v{\mathcal{P}}}{3V\epsilon_0}\,,
\end{equation}
defining the macroscopic polarisation of the medium by $\v{\mathcal{P}}\equiv N\v{p}$. Thus, the field experienced by the test dipole---and, therefore, the field inside the medium---is equal to the ``normalised'' field
\begin{equation}
\v{\mathcal{E}}_\text{norm}=\v{\mathcal{E}}+\frac{\v{\mathcal{P}}}{3V\epsilon_0}\,,
\end{equation}
where $\v{\mathcal{E}}$ is the local microscopic spatial average of the electric field that the dielectric is immersed in. Finally, $\v{\mathcal{E}}_\text{norm}$ is related to $\v{\mathcal{P}}$ through the definition of $\chi$~\cite[\textsection 4.5]{Jackson1998},
\begin{equation}
\v{\mathcal{P}}=\epsilon_0\chi\v{\mathcal{E}}_\text{norm}=\epsilon_0\chi\Biggl(\v{\mathcal{E}}+\frac{\v{\mathcal{P}}}{3V\epsilon_0}\Biggr)\,.
\end{equation}
Thus,
\begin{equation}
\v{\mathcal{P}}=V(\epsilon-1)\epsilon_0\v{\mathcal{E}}=\frac{\epsilon_0\chi\v{\mathcal{E}}}{1-\chi/\bigl(3V\bigr)}\,,
\end{equation}
which can be rearranged to give the Clausius--Mossotti equation:
\begin{equation}
\chi=3V\frac{\epsilon-1}{\epsilon+2}\,.
\end{equation}
For a particle of volume $V$ made from a typical dielectric with a refractive index $n=1.5$, $\chi=0.9V\approx V$. An intuitive understanding of $\chi$ is therefore possible as the volume of dielectric that is polarised by an incoming field. It is perhaps interesting to note that this means that each single molecular dipole in the dielectric has an effective susceptibility
\begin{equation}
\chi_\text{eff}\approx\frac{4}{3}\pi a^3\,,
\end{equation}
where $2a$ is the mean distance between the molecules making up the dielectric; $a\sim 10^{-10}$\,m is typically several orders of magnitude larger than the molecular radius itself. However, $\chi_\text{eff}$ is of the same order as the susceptibility of a free molecule~\cite{Miller1990}. In other words, the polarisation of an individual dipole due to an off-resonant electric field is of about the same order whether that dipole is isolated or in a bulk solid.
\par
This $\chi$ can be used to calculate the power dissipated by a small dielectric sphere, modelled as a single point dipole, due to blackbody radiation. At a temperature $T$ there are $n_k=1/\bigl\{\exp\bigr[\hbar ck/\bigl(k_\text{B}T\bigr)\bigr]-1\bigr\}$ photons with a wavevector $\v{k}$ and wavenumber $k$. These photons produce an equivalent electric field of optical power $P_k=\hbar kc^2n_k\sigma_\text{L}/V_\text{q}$, $V_\text{q}$ being the quantisation volume, and therefore lead to an absorbed power due to that mode
\begin{equation}
P_{\text{abs},k}\approx\frac{\hbar k^2c^2n_k}{V_\text{q}}\im{\chi}\,,
\end{equation}
for small $\lvert\chi\rvert$. We must now sum over every mode to obtain the total absorbed power $P_\text{abs}=\sum_{\v{k}}P_{\text{abs},k}$, which quickly becomes cumbersome since the number of modes becomes infinite as the quantisation volume grows indefinitely. By assuming that the modes are evenly distributed in $\v{k}$-space, with a density $(2\pi)^3/V_\text{q}$, we can transform this sum into a three-dimensional integral
\begin{align}
P_\text{abs}&=\frac{2V_\text{q}}{(2\pi)^3}\iiint\frac{\hbar k^2c^2n_k}{V_\text{q}}\im{\chi}\,\mathrm{d}^3\v{k}\nonumber\\
&=\frac{\hbar c^2}{4\pi^3}\im{\chi}\int_0^{2\pi}\int_0^\pi\int_0^\infty\frac{k^4}{\exp\bigr[\hbar ck/\bigl(k_\text{B}T\bigr)\bigr]-1}\sin(\theta)\,\mathrm{d} k\,\mathrm{d}\theta\,\mathrm{d}\phi\nonumber\\
&=\frac{24\zeta(5)}{\pi^2 c^3\hbar^4}\im{\chi}\bigl(k_\text{B}T\bigr)^5\,,
\end{align}
where the extra factor of $2$ accounts for the two polarisations, where $\chi$ was assumed to be independent of $k$, and where $\zeta(5)$ is the Riemann zeta function. The $k$-integral is performed by appealing to the definition of the $\zeta(z)\equiv\zeta(z,1)$~\cite[\textsection 9.51]{Gradshteyn1994}. Our final step is to note that in thermal equilibrium, $P_\text{abs}$ is equal to the power dissipated by the sphere, $P_\text{diss}$.
\begin{figure}
\centering
\includegraphics[width=\figwidth]{./Diagrams/10um}
\caption[Electric field intensity inside a $10$\,$\upmu$m diameter hemispherical void in an ideal metal substrate]{Finite-difference time-domain analysis of the electric field intensity on a 2D slice of a $10$\,$\upmu$m diameter hemispherical void in an ideal metal substrate. The incident field is a linearly polarised plane wave, with a wavelength of $780$\,nm, propagating downwards. Two sections, both intersecting the focus, are also shown: the vertical (horizontal) section runs along the vertical (horizontal) dashed line.}
\label{fig:10um}
\end{figure}
\section{Appendix: Calculating the electric field inside hemispherical mirrors}\label{sec:Future:Dipole}
Hemispherical voids templated on gold surfaces are a good system to work with experimentally: once made, they require no further alignment; and regular, close-packed arrays can be made with several tens (for large diameters) up to several hundreds (for diameters of the order of $1$\,$\upmu$m~\cite{Bartlett2004}) of dimples. However, the analysis of the electromagnetic fields inside the dimples presents a challenge. The most direct route to exploring these fields is through the numerical solution of Maxwell's equations. A large number of software packages are available, with several operating either on finite element method (FEM) or finite-difference time-domain (FDTD) principles or employing Mie theory (see Ref.~\cite{Parsons2010} for a recent review of such techniques). Analyses of scattering of electromagnetic radiation off spherical particles are usually performed using Mie theory~\cite{Jackson1998}, which exploits the fact that the vector spherical harmonics form a complete orthogonal set of modes on the sphere. In the case of a hemisphere, however, no such set of modes exists---and the situation is even worse for truncated hemispheres---and FEM or FDTD techniques are more desirable. It is interesting to note that this problem can be formally circumvented in certain situations. For example, the authors of Ref.~\cite{Hetet2010} implicitly assume knowledge of the field outside the spherical region defined by a perfect truncated hemispherical mirror to explore the behaviour of the vacuum field inside this same spherical region. In the absence of a compact analytical solution, we use an open-source software package called \textsc{MEEP}~\cite{Oskooi2010} for our analysis.\\
\begin{figure}
\centering
\includegraphics[width=\figwidth]{./Diagrams/100um}
\caption[Electric field intensity inside a $100$\,$\upmu$m diameter hemispherical void in an ideal metal substrate]{Electric field intensity inside a $100$\,$\upmu$m diameter hemispherical void in an ideal metal substrate. The size of the hemispherical template and the depth of the void ($15$\,$\upmu$m vertically from the bottom to the lip) match the surface in \fref{fig:Dimples}.}
\label{fig:100um}
\end{figure}
We show an example of such a simulated field in \fref{fig:10um}. Close to the geometrical focus of the dimple, a strong focus is found; the size and peak intensity of this focus can be used to explore the possibility of producing single- or few-atom dipole traps inside such cavities. 2D arrays of dipole traps were demonstrated~\cite{Dumke2002} a number of years ago, and recently used to perform site-selective manipulation of atoms~\cite{Kruse2010}---both these experiments relied on a purpose-built refractive micro-lens array. Using templated surfaces offers a number of advantages over the micro-lens array, not least ease of manufacture and the possibility of integration into so-called atom chips~\cite{Folman2000}; in this regard, the use of reflective rather than refractive optical elements is of paramount importance.\\
The dimples in the first surface produced for the experimental investigations in Southampton were not grown to full hemispheres. Rather, latex spheres with a diameter of $100$\,$\upmu$m were used and gold was only templated up to a depth of $15$\,$\upmu$m. Part of the resulting surface is shown in \fref{fig:Dimples}, with the field inside one such dimple simulated in \fref{fig:100um}. Note that aliasing artifacts are more apparent in this figure than in \fref{fig:10um}, the reason being that the $0.05$\,$\upmu$m resolution possible in the case of the latter was not possible in simulating the larger sample, due to computer memory constraints. In the case of \fref{fig:100um}, a resolution of $0.17$\,$\upmu$m was used. In both cases, the incoming field was a plane wave with a wavelength of $\lambda=780$\,nm.
\section{Appendix: Force acting on an atom inside an arbitrary monochromatic field}
In a remarkable piece of work dating to 1980, Gordon and Ashkin~\cite{Gordon1980} give several useful expressions for the force and diffusion experienced by an atom inside what they called a `radiation trap'---essentially an arbitrary (monochromatic) electric field. Implicit in their work is the assumption that the system has no `memory'; we cannot directly apply their equations to mirror-mediated or external cavity cooling systems, for example. Nonetheless, such a model is perhaps the easiest way of exploring the behaviour of atoms inside fields as complex as those in hemispherical voids on a metal surface. Unfortunately, the authors of Ref.~\cite{Gordon1980} do not give explicit general formulae for the velocity-dependent force acting on an atom; we will now generalise their expressions [specifically, Eqs.~(14) and~(15)] to the case when the atom is not motionless. Let us first briefly introduce the notation we will use in this section:\footnote{Our notation will be identical to Ref.~\cite{Gordon1980} wherever possible.} $\v{\textsl{\textsf{F}}}$ is the force acting on the atom; $D$ is the difference in the populations of the upper and lower states; $\sigma$ is the atomic lowering operator; $\Delta=\omega-\omega_\text{a}$, with $\omega$ the frequency of the driving field and $\omega_\text{a}$ the atomic resonance frequency; $\gamma=\Gamma-\i\Delta$; $\v{\mu}_{21}$ the atomic dipole operator; $\v{v}$ the atomic velocity; and $\mathcal{E} e^{-\i\omega t}$ the classical incident electric field. We also define
\begin{equation}
g=\tfrac{\i}{\hbar}\v{\mu}_{21}\cdot\mathcal{E}\,,
\end{equation}
and the saturation parameter $s=2\lvert g\rvert^2/\lvert\gamma\rvert^2$. It is also convenient to define the vectors $\v{\alpha}$ and $\v{\beta}$ by $\grad g=(\v{\alpha}+\i\v{\beta})g$. In practice, numerical simulation gives us knowledge of $\mathcal{E}$, assuming that the perturbation of the atom on the field is quite small. By specifying the magnetic field, we can determine $\omega_\text{a}$, and therefore $\Delta$. Given the atomic species, we also know $\Gamma$ and $\v{\mu}_{21}$; knowledge of $\gamma$, $g$, $s$, $\v{\alpha}$, $\v{\beta}$ and $\expt\sigma$ then follows, as we will see. Finally, this determines $\expt{\v{\textsl{\textsf{F}}}}$, the classical force acting on the atom.\\
To first order in $\v{v}$, we have
\begin{align}
\label{eq:ForceGordon1980}
\expt{\v{\textsl{\textsf{F}}}}&=-\i\hbar\bigl[\expt{\sigma}^\ast\grad g-\expt{\sigma}\grad g^\ast\bigr]\,,\\
\expt{\dot{\sigma}}+\gamma\expt{\sigma}&=\expt{D}g\,,\\
\expt{\dot{D}}+2\Gamma\expt{D}&=2\Gamma-2(g^\ast\expt{\sigma}+g\expt{\sigma}^\ast)\,,\\
\expt{\dot{D}}&=-\tfrac{2s}{1+s}(\v{v}\cdot\v{\alpha})\expt{D}\,,\\
\expt{\dot{\sigma}}&=\bigl[(\v{v}\cdot\v{\alpha})\tfrac{1-s}{1+s}+\i(\v{v}\cdot\v{\beta})\bigr]\expt{\sigma}\,\text{, and}\\
\dot{g}&=\v{v}\cdot(\v{\alpha}+\i\v{\beta})g\,,
\end{align}
directly from Ref.~\cite{Gordon1980}. Thus, to the same order,
\begin{equation}
\expt{D}=\bigl[1-\tfrac{1}{\Gamma}(g^\ast\expt{\sigma}+g\expt{\sigma}^\ast)\bigr]+\tfrac{1}{\Gamma}\tfrac{s}{1+s}(\v{v}\cdot\v{\alpha})\bigl[1-\tfrac{1}{\Gamma}(g^\ast\expt{\sigma}+g\expt{\sigma}^\ast)\bigr]\,,
\end{equation}
whereby
\begin{multline}
\expt{\sigma}\bigl[\gamma+(\v{v}\cdot\v{\alpha})\tfrac{1-s}{1+s}+\i(\v{v}\cdot\v{\beta})\bigr]=\bigl[1-\tfrac{1}{\Gamma}(g^\ast\expt{\sigma}+g\expt{\sigma}^\ast)\bigr]g\\+\tfrac{1}{\Gamma}\tfrac{s}{1+s}(\v{v}\cdot\v{\alpha})\bigl[1-\tfrac{1}{\Gamma}(g^\ast\expt{\sigma}+g\expt{\sigma}^\ast)\bigr]\,.
\end{multline}
We can solve this for $\expt{\sigma}$ to obtain
\begin{multline}
\expt{\sigma}=\frac{g}{\gamma(1+s)}+\frac{g}{\lvert\gamma\rvert^2(1+s)}\bigl(\tfrac{1}{2\Gamma}\gamma^\ast+\tfrac{1-s}{1+s}\bigr)(\v{v}\cdot\v{\alpha})\\
-2\bigl[\Gamma(1-s)+\tfrac{1}{\Gamma}\lvert g\rvert^2\bigr]\frac{g}{\gamma\lvert\gamma\rvert^2(1+s)^3}(\v{v}\cdot\v{\alpha})\\
+\frac{\bigl[2\Delta-\i\gamma(1+s)\bigr]g}{\gamma\lvert\gamma\rvert^2(1+s)^2}(\v{v}\cdot\v{\beta})\,,
\end{multline}
which we can plug into \eref{eq:ForceGordon1980}, together with $g$, to obtain the (velocity-dependent) force acting on the atom.
\appendicesend
\chapter{The transfer matrix model}\label{ch:TMM:TMM}
\epigraph{The reader might wonder why it is of interest, physically, to consider $n$-manifolds for which $n$ is larger than $4$, since ordinary spacetime has just four dimensions. In fact many modern theories [...]\ operate within a `spacetime' whose dimension is much larger than $4$.}{R.\ Penrose, \emph{The Road to Reality} (2004)}
\cref{ch:CoolingMethods:AFInt}, and in particular \sref{sec:CoolingMethods:PolTLA}, developed the necessary tools to describe the interaction of an atom with the electromagnetic field, as parametrised by the atom's characteristic polarisability. In one dimension, one can succinctly describe the fields interacting with a linear scatterer through what is called the transfer matrix approach~\cite{Deutsch1995}. Restricting ourselves to one spatial dimension is not an overly restrictive approximation, despite the quote at the beginning of this chapter; the formalism that is discussed in this chapter allows us to describe a wealth of physical situations. The purpose of this chapter is to extend this model significantly, enabling it to account for moving as well as static scatterers; this is done in \sref{sec:TMM:Model} and the model that results is solved generally in \sref{sec:TMM:General}. The extended model discussed here takes into account the first-order Doppler shift but not relativistic effects, and it is therefore correct only up to first order in the velocity of the scatterer.\par
The transfer matrix method is more general than an analysis based on modal decomposition, and is therefore used to describe the optomechanics of scatterers inside cavities in \sref{sec:TMM:Optomechanics}. This method is extended even further in \sref{sec:TMM:Multilevel}, where it is shown that the concept of polarisability can also be applied to atoms having a Zeeman manifold and interacting with circularly polarised light. The results of these sections are confirmed by showing that the standard results for optical molasses, mirror-mediated cooling and cooling of cavity mirrors, as well as sub-Doppler cooling mechanisms~\cite{Dalibard1989}, can be reproduced by simple applications of the transfer matrix method. In \cref{ch:TMMApplications}, the theory developed over the present chapter will then be applied to describe two novel cooling schemes, outside cavities (\sref{sec:TMM:ECCO}) and inside ring cavities (\sref{sec:TMM:AmplifiedOptomechanics}).
\section{An extended scattering theory}\label{sec:TMM:Model}
In this first section we develop and present a scattering theory for optomechanically coupled systems, allowing for the efficient description of the motion of arbitrary combinations of atoms and mirrors interacting through the radiation field. We will restrict the model to one-dimensional motion and small velocities. The main building block is the beamsplitter transfer matrix~\cite{Deutsch1995, Asboth2008}, i.e., the \emph{local relation} between light field amplitudes at the two sides of a scatterer. We will calculate the radiation force acting on a moving scatterer up to linear order in the velocity. The model is completed by including the quantum fluctuations of the radiation force which stem from the quantised nature of the field. We will determine the momentum diffusion coefficient corresponding to the minimum quantum noise level.\\
One system we will consider in some detail is composed of two mirrors; one of them is fixed in space, whilst the other one is mobile. This is the generic scheme for radiation pressure cooling of moving mirrors~\cite{WilsonRae2007, Marquardt2007, Genes2008}. At the same time, in the limit of low reflection the moving mirror can equally well represent the a single atomic dipole interacting with its mirror image in front of a highly reflecting surface (\sref{sec:CoolingMethods:MMC}; see also Refs.~\cite{Eschner2001, Bushev2004, Xuereb2009a}).
\par
The work in this chapter is published as Xuereb, A., Domokos, P., Asb\'oth, J., Horak, P., \& Freegarde, T. Phys.\ Rev.\ A \textbf{79}, 053810 (2009) and Xuereb, A., Freegarde, T., Horak, P., \& Domokos, P., Phys.\ Rev.\ Lett.\ \textbf{105}, 013602 (2010) and is in part reproduced \emph{verbatim}. After our initial exploration of the transfer matrix description of moving scatterers, we subsequently apply this description to several systems, and also extend it in several ways. The resulting model is very general and can be solved to give analytical formulations of the friction forces and momentum diffusion processes acting in a generic optomechanical system.
\subsection{Basic building blocks of the model}\label{sec:TMM:BasicBuildingBlocksModel}
\begin{figure}[t]
\centering
\includegraphics[scale=0.75]{./Diagrams/DiagramsXuereb2009b/BS}
\caption[The four different modes that interact through a point-like beamsplitter in one dimension]{The four different modes that interact through a point-like beamsplitter in one dimension.}
\label{fig:TMM:BS}
\end{figure}
Consider a point-like scatterer (or beamsplitter), $\text{BS}$, moving along the `$x$' axis on the trajectory $x_{\text{BS}}(t)$. Outside the scatterer, the electric field $\mathcal{E}$ can be expressed in terms of a discrete sum\footnote{This is a simplifying assumption and all our results also hold for a continuum of field modes.} of left- and right-propagating plane wave modes with different wave numbers, $k$, and hence different frequencies, $\omega=kc$:
\begin{equation}
\label{eq:TMM:Efield}
\mathcal{E}=\begin{cases}
\sum_k\big[A(k)e^{-\i kx-\i\omega t}+B(k)e^{\i kx-\i\omega t}\big]+\rm{c.c.}&x<x_\text{BS}\\
\sum_k\big[C(k)e^{-\i kx-\i\omega t}+D(k)e^{\i kx-\i\omega t}\big]+\rm{c.c.}&x>x_\text{BS}\,,
\end{cases}
\end{equation}
where $A(k)$ and $B(k)$ are the mode amplitudes on the left side, $x<x_{\text{BS}}(t)$, while $C(k)$ and $D(k)$ are the amplitudes on the right side, $x>x_{\text{BS}}(t)$, of $\text{BS}$, and where $\text{c.c.}$ denotes the complex conjugate. In accordance, the magnetic field is~\cite{Jackson1998}
\begin{equation}
\label{eq:TMM:Bfield}
c \mathcal{B}=\begin{cases}
\sum_k\big[-A(k)e^{-\i kx-\i\omega t}+B(k)e^{\i kx-\i\omega t}\big]+\rm{c.c.}&x<x_\text{BS}\\
\sum_k\big[-C(k)e^{-\i kx-\i\omega t}+D(k)e^{\i kx-\i\omega t}\big]+\rm{c.c.}&x>x_\text{BS}\,.
\end{cases}
\end{equation}
As depicted schematically in \fref{fig:TMM:BS}, the scatterer mixes these waves. Our first goal is the derivation of the transfer matrix $M$ connecting the field amplitudes on the right to those on the left side of a beamsplitter moving at a fixed velocity $v$. This relation is well-known~\cite{Deutsch1995} for an immobile scatterer. Therefore, let us first transform the electromagnetic field into a frame moving with the instantaneous velocity $v$ of the $\text{BS}$.
\subsubsection{Transfer matrix for an immobile beamsplitter}\label{sec:TMM:FixedBS}
In the frame co-moving with the $\text{BS}$, the interaction of the field with the scatterer at $x^\prime=0$ can be characterised by the single parameter $\zeta$ by means of the one-dimensional wave equation~\cite{Jackson1998, Deutsch1995},
\begin{equation}
\label{eq:TMM:WaveEqn}
\left(\partial_{x^\prime}^2-\frac{1}{c^2}\partial_{t^\prime}^2\right) \mathcal{E}^\prime(x^\prime,t^\prime) = \frac{2}{kc^2}\zeta\,\delta(x^\prime)\,\partial^2_{t^\prime} \mathcal{E}^\prime(x^\prime,t^\prime)\,.
\end{equation}
The electric field can be considered in a modal decomposition similar to~\eref{eq:TMM:Efield}. Since a fixed beamsplitter couples only the plane waves with identical frequency and wave number, the stationary scattering can be fully described within the closed set of modes
\begin{equation}
\mathcal{E}^\prime(x^\prime,t^\prime) =\begin{cases} \bigl(A' e^{-\i kx^\prime -\i\omega t^\prime} + B' e^{\i kx^\prime -\i \omega t^\prime}\bigr)+\rm{c.c.} &x^\prime < 0\\
\bigl(C' e^{-\i kx^\prime -\i\omega t^\prime} + D' e^{\i kx^\prime -\i \omega t^\prime}\bigr)+\rm{c.c.} &x^\prime > 0\,,
\end{cases}
\end{equation}
where the index $k$ has been dropped. A linear relation between the field amplitudes on the right of the scatterer and those on the left can be derived from the wave equation~\cite{Deutsch1995},
\begin{equation}
\label{eq:TMM:TM_fix}
\begin{pmatrix}
C^\prime\\
D^\prime
\end{pmatrix}=M\begin{pmatrix}
A^\prime\\
B^\prime
\end{pmatrix}\,\text{, with}
\end{equation}
\begin{equation}
\label{eq:TMM:M0}
M = \begin{bmatrix}
1-\i\zeta & -\i\zeta\\
\i\zeta & 1+\i\zeta
\end{bmatrix}
= \frac{1}{\mathfrak{t}}
\begin{bmatrix}
1 & -\mathfrak{r} \\
\mathfrak{r} & \mathfrak{t}^2 - \mathfrak{r}^2
\end{bmatrix}\,.
\end{equation}
In the second form of the transfer matrix $M$, we expressed it in terms of the reflectivity $\mathfrak{r}$ and transmissivity $\mathfrak{t}$ of the beamsplitter. This latter form is more convenient to describe moving mirrors, while for atoms the scattering strength parameter $\zeta$ can be readily expressed in terms of its polarisability $\alpha$ [see \sref{sec:CoolingMethods:PolTLA}, and especially \eref{eq:ZetaDefn}]. In this case the transfer matrix depends on the wave number $k$, which might lead to significant effects, e.g., Doppler cooling, close to resonance with the atom (see \Sref{sec:Force}).
\subsubsection{Transfer matrix for a moving beamsplitter}\label{sec:TMM:MovBS}
The transformation back into the laboratory-fixed frame involves the change of the coordinates, $x^\prime=x-vt$ and $t^\prime=t$, and the Lorentz-boost of the electric field up to linear order in $v/c$ \cite[\textsection 11.10]{Jackson1998}:
\begin{equation}
\mathcal{E} = \mathcal{E}^\prime + {v} \mathcal{B}^\prime\,,
\end{equation}
where we assumed that $\mathcal{E}$ and $\mathcal{E}^\prime$ are polarised in the `$y$' direction, $\mathcal{B}$ and $\mathcal{B}^\prime$ are polarised in the `$z$' direction, and the velocity is along the $x$ axis. The electric field in the laboratory frame becomes
\begin{align}
\mathcal{E}(x,t) &= \sum_{k^\prime} \Biggl\{A'(k^\prime) e^{-\i k^\prime (x- vt) -\i \omega' t} + B'(k^\prime) e^{\i k^\prime(x-vt) -\i \omega' t}\nonumber\\
&\phantom{=\ \sum_{k^\prime} \Biggl\{}- \frac{v}{c}\Bigl[ A'(k^\prime) e^{-\i k^\prime (x- vt) -\i \omega' t} - B'(k^\prime) e^{\i k^\prime(x-vt) -\i \omega' t}\Bigr]\Biggr\} + \text{c.c.}\nonumber\\
&=\sum_k \Bigl[\bigl(1-\tfrac{v}{c}\bigr) A'\left(k+kv/c\right) e^{-\i k (1+v/c) x -\i \omega t}\nonumber\\
&\phantom{=\ \sum_k \bigl(}+ \left(1+\tfrac{v}{c}\right) B'\left(k-kv/c\right) e^{\i k (1-v/c) x -\i \omega t}\Bigr]+\rm{c.c.}\,,
\end{align}
which can be expressed as a linear transformation $\hat L(v)$ of the amplitudes,
\begin{equation}
\begin{pmatrix}
A(k)\\
B(k)
\end{pmatrix}=\hat L(-v)\begin{pmatrix}
A^\prime(k)\\
B^\prime(k)
\end{pmatrix}\,\text{, with}
\end{equation}
\begin{equation}
\label{eq:TMM:Lmatrix}
\hat L(v) = \begin{bmatrix}
\left(1+ \frac{v}{c}\right) \hat{P}_{-v} & 0\\
0 & \left(1- \frac{v}{c}\right) \hat{P}_v
\end{bmatrix}\,.
\end{equation}
This construction is explored further in~\aref{sec:POper}. Here we defined the operator $\hat{P}_v:f(k)\mapsto f\left(k+k\tfrac{v}{c}\right)$, which represents the Doppler shift of the plane waves in a moving frame. Obviously, $\hat L^{-1}(v) = \hat L(-v)$ to first order in $v/c$. The total action of the moving $\text{BS}$,
\begin{equation}
\label{eq:TMM:ABCD}
\begin{pmatrix}
C(k)\\
D(k)
\end{pmatrix}= \hat{M}
\begin{pmatrix}
A(k)\\
B(k)
\end{pmatrix}\text{,}
\end{equation}
can then be obtained from
\begin{align}
\label{eq:TMM:LML}
\hat{M} &= \hat L(-v) M \hat L(v)\\
&=\frac{1}{\mathfrak{t}}\begin{bmatrix}
1 & -(1-2\tfrac{v}{c})\mathfrak{r}\hat{P}_{2v}\nonumber\\
(1+2\tfrac{v}{c})\mathfrak{r}\hat{P}_{-2v}& \mathfrak{t}^2-\mathfrak{r}^2
\end{bmatrix}\,,
\end{align}
where we have assumed that $\mathfrak{r}$ and $\mathfrak{t}$ do not depend on the wave number. Compared to $M$ in \eref{eq:TMM:M0}, the difference lies in the off-diagonal terms including the Doppler shift imposed by the reflection on a moving mirror. In other words, the coupled counter-propagating plane wave modes differ in wave number, i.e., $k\left(1+\frac{v}{c}\right)$ right-propagating waves couple to $- k\left(1-\frac{v}{c}\right)$ left-propagating waves. Furthermore, if the polarisability itself depends on the wave number $k$, e.g., as in~\eref{eq:ZetaDefn}, the Doppler shift operator acts also on it. To see this effect explicitly, to linear order in $v/c$, $\hat{M}$ can be written as
\begin{equation}
\label{eq:TMM:Mzeta}
\begin{bmatrix}
1-\i\zeta - \i \tfrac{v}{c}{\omega}\tfrac{\partial\zeta}{\partial k} & -\i\zeta\left[1-\tfrac{v}{c}\big(2 - \tfrac{k}{\zeta}\tfrac{\partial\zeta}{\partial k}\big)\right]\hat{P}_{2v}\\
\i\zeta\left[1+ \tfrac{v}{c}\big(2 - \tfrac{k}{\zeta}\tfrac{\partial\zeta}{\partial k}\big)\right]\hat{P}_{-2v}& 1+\i\zeta -\i\tfrac{v}{c} k\tfrac{\partial\zeta}{\partial k}
\end{bmatrix}\,.
\end{equation}
The transfer matrix in the laboratory frame can thus be conceived as a $2$-by-$2$ supermatrix acting also in the $k$-space. The amplitude $C$ at a given wave number $k$, i.e., $C(k)$, is combined with the amplitudes $A(k)$ and $B\big(k-2k\tfrac{v}{c}\big)$. A similar statement holds for $D(k)$.
Starting from the knowledge of the incoming field amplitudes, this transfer matrix allows for calculating the total electromagnetic field around a beamsplitter moving with a fixed velocity. In the next step, we derive the force on the moving scatterer through the Maxwell stress tensor.
\subsubsection{Force on a medium in an electromagnetic field}
The Maxwell stress tensor (see~\cite[\textsection 6.7]{Jackson1998}) is defined, for a homogeneous medium in one dimension, $x$, as
\begin{equation}
\boldsymbol{T}_{xx}=-\frac{\epsilon_0}{2}\Big(\big|\mathcal{E}\big|^2+{c^2} \big|\mathcal{B}\big|^2\Big)\,,
\end{equation}
where the electric field $\mathcal{E}$ and the magnetic field $\mathcal{B}$, \eref{eq:TMM:Efield} and \eref{eq:TMM:Bfield}, respectively, have no components along $x$. It is trivial, then, to see that after applying the rotating wave approximation, we obtain
\begin{equation}
\label{eq:TMM:MST}
\boldsymbol{T}_{xx}=-2\epsilon_0\Bigg[\Big|\sum_k A(k)e^{-\i kx-\i\omega t}\Big|^2+\Big|\sum_k B(k)e^{\i kx-\i\omega t}\Big|^2\Bigg]\,,
\end{equation}
since the cross terms in $|\mathcal{E}|^2$ and $|\mathcal{B}|^2$ have opposite signs. Note that $\boldsymbol{T}_{xx}$ varies on time scales of the order of the optical period. Let us now introduce a characteristic time, $\tau\gg 2\pi/\omega_0$, over which the variations in $\boldsymbol{T}_{xx}$ will be averaged, $\omega_0$ being the central frequency of the pump beam. At $x = 0$,
\begin{align}
\frac{1}{\tau}\int_0^\tau \Big|\sum_k A(k)e^{-\i\omega t}\Big|^2\mathrm{d} t&=\sum_k|A(k)|^2+\sum_{i\neq j}\frac{1}{\tau}\int_0^\tau A(k_i)\big[A(k_j)\big]^\ast e^{-\i(\omega_i-\omega_j)t}\mathrm{d} t\nonumber\\
&\approx \sum_k|A(k)|^2+\sum_{i\neq j}A(k_i)\big[A(k_j)\big]^\ast\nonumber\\
&=\Big|\sum_kA(k)\Big|^2\,.
\end{align}
In the approximation we assumed that the frequency bandwidth of the excited modes, $\Delta = \max\left\{ \omega_i-\omega_j\right\}$, around $\omega_0$ is so narrow that $\tau \ll 2 \pi/\Delta$. Since the broadening is due to the Doppler shift, $\Delta \sim 2\omega_0\tfrac{v}{c}$, where $v$ is the speed of the beamsplitter. For example, taking $v$ to be the typical speed of atoms in a magneto-optical trap, we require $\tau\ll \pi/\big(\omega_0\tfrac{v}{c}\big)\sim 10^{-4}~$s. The time needed to reach the stationary regime of scattering is typically much shorter and thus this condition imposed on the averaging time $\tau$ can be safely fulfilled.
The force on the medium is given by the surface integral of $\boldsymbol{T}_{xx}$ on the surface, $\mathcal{S}$, of a fictitious volume $V=\sigma_\text{L}\,\delta l$ enclosing the medium, where $\sigma_\text{L}$ is the mode area and $\delta l$ the infinitesimal length of the volume along the `$x$' axis. Then, this force is given by
\begin{align}
\label{eq:TMM:Forcedef}
\textsl{\textsf{F}}&=\oint_\mathcal{S}\boldsymbol{T}_{xx} n_x\mathrm{d}\mathcal{S}\nonumber\\
&=\sigma_\text{L}\big[\boldsymbol{T}_{xx}({x\rightarrow 0^+})-\boldsymbol{T}_{xx}({x\rightarrow 0^-})\big]\,,
\end{align}
where $n_x=\sgn(x)$ is the normal to $\mathcal{S}$. Substituting the relevant expressions for $\boldsymbol{T}_{xx}$ into the preceding formula gives
\begin{equation}
\label{eq:TMM:MSTForce}
\textsl{\textsf{F}}=\frac{\hbar\omega_0}{c}\Big(\big|A\big|^2 +\big|B\big|^2 -\big|C\big|^2-\big|D\big|^2 \Big)\,,
\end{equation}
where $A=[{\hbar\omega_0/(2\sigma_\text{L}\epsilon_0c)}]^{-1/2}\sum_k A(k)$ is the photo-current amplitude, and similarly for $B$, $C$ and $D$, their modulus square giving the number of photons per unit time. Although we considered first the electric field composed of independent modes, in the force expression only the sums of the mode amplitudes occur. An identical result holds when we replace the discrete sum over $k$ by an integral, defining $A=[{\hbar\omega_0/(2\sigma_\text{L}\epsilon_0c)}]^{-1/2}\int\!A(k)\mathrm{d} k$, etc.
\subsubsection{Quantum fluctuations of the force}
\label{sec:MSTDiffusion}
In the previous section the force was derived based on the assumption that the field amplitudes are c-numbers. In order to describe the inherent quantum fluctuations of the force, we need to resort to the quantum theory of fields and represent the mode amplitudes by operators: $A(k)\to\hat{A}(k)$. To leading order the fluctuations of the force acting on a beamsplitter amount to a momentum diffusion process~\cite{Dalibard1989, Castin1990}. The diffusion coefficient will be evaluated in the case of coherent-state fields~\cite{Glauber1963}.
\par
The diffusion coefficient can be deduced from the second-order correlation function of the force operator, \eref{eq:FDTDeltaDependence}. The evaluation of this quantum correlation is system-specific. Quantum correlations, i.e., the operator algebra of the mode amplitudes $\hat A(k)$, $\hat B(k)$, $\hat C(k)$, and $\hat D(k)$, are influenced by multiple scattering and thus depend on the total transfer matrix of the entire system. The simplest case is a single beamsplitter at rest where the ``input'' modes $\hat B(k)$ and $\hat C(k)$ have independent fluctuations. The calculation, delegated to \aref{sec:TMM:AppDiffusion}, includes all the steps needed for the treatment of a general system. The diffusion coefficient for a single beamsplitter is obtained as
\begin{equation}
\label{eq:TMM:DiffCoeff}
\textsl{\textsf{D}} = (\hbar k)^2 \Bigl(\big|A\big|^2 +\big|B\big|^2+\big|C\big|^2+\big|D\big|^2 + 2 \re{\mathfrak{r} A^* B -\mathfrak{t} A^* C}+ 2\re{\mathfrak{r} D^*C - \mathfrak{t} D^* B}\Bigr)\,,
\end{equation}
where $A=\expt{\hat{A}}, B=\expt{\hat{B}}, C=\expt{\hat{C}}, D=\expt{\hat{D}}$ are the photo-current amplitudes (their modulus square is of the units of 1/sec), obeying \eref{eq:TMM:ABCD} for $v=0$.
\par
As an example, let us consider the diffusion coefficient for a two-level atom illuminated by counter-propagating monochromatic light waves. Using the polarisability $\zeta$, the transmission and reflection coefficients can be expressed as $\mathfrak{t} = 1/(1-\i\zeta)$ and $\mathfrak{r}=\i\zeta/(1-\i\zeta)$, respectively [see~\eref{eq:TMM:M0}]. \eref{eq:TMM:DiffCoeff} can then be rewritten in the form
\begin{equation}
\label{eq:TMM:GeneralDiffnBC}
\textsl{\textsf{D}} = (\hbar k)^2 \bigg[\frac{2\im{\zeta}}{|1-\i\zeta|^2} \big|B-C\big|^2 + \frac{4|\zeta|^2}{|1-\i\zeta|^2} \Big(\big|B\big|^2 + \big|C\big|^2\Big)\bigg]\,,
\end{equation}
where the first term, apart from the factor $|1-\i\zeta|^2$, corresponds to the result well-known from laser cooling theory. Second- and higher-order terms in $\lvert\zeta\rvert$ correspond to the back-action of the atom on the field and are usually absent in treatments of laser cooling due to the fact that $\lvert\zeta\rvert\ll 1$ is generally implicit in such treatments. Note that the diffusion process due to the recoil accompanying the spontaneous emission of a photon (see Ref.~\cite{Gordon1980}) is missing from this result. The detailed modelling of absorption, i.e., scattering photons into the three-dimensional space, is not included in our approach.
\subsubsection{Example: Force on a moving beamsplitter\label{sec:Force}}
We will now use~\eref{eq:TMM:MSTForce} to derive a general expression for the force on a moving beamsplitter illuminated by two counterpropagating, monochromatic, plane waves with amplitudes $B_0$ and $C_0$. On using~\eref{eq:TMM:ABCD} to express the outgoing field modes in terms of the incoming ones, we note that the outgoing amplitudes comprise two monochromatic terms each:
\begin{equation}
\label{eq:TMM:AD}A= \frac{\i\zeta\left[1-\tfrac{v}{c}\big(2+\tfrac{k}{\zeta}\tfrac{\partial\zeta}{\partial k}\big)\right]}{1-\i\zeta\Big(1-\tfrac{v}{c}\tfrac{k}{\zeta}\tfrac{\partial\zeta}{\partial k}\Big)}B_0 + \frac{1}{1-\i\zeta\Big(1+\tfrac{v}{c}\tfrac{k}{\zeta}\tfrac{\partial\zeta}{\partial k}\Big)}C_0\,.
\end{equation}
and
\begin{equation}
D= \frac{1}{1-\i\zeta\Big(1-\tfrac{v}{c}\tfrac{k}{\zeta}\tfrac{\partial\zeta}{\partial k}\Big)}B_0 + \frac{\i\zeta\left[1+\tfrac{v}{c}\big(2+\tfrac{k}{\zeta}\tfrac{\partial\zeta}{\partial k}\big)\right]}{1-\i\zeta\Big(1+\tfrac{v}{c}\tfrac{k}{\zeta}\tfrac{\partial\zeta}{\partial k}\Big)}C_0\,.
\end{equation}
These relations are substituted into~\eref{eq:TMM:MSTForce}, giving
\begin{multline}
\label{eq:TMM:ForceTLA}
\textsl{\textsf{F}}=\frac{2\hbar\omega}{\bigl|1-\i\zeta\bigr|^2}\bigg\{\Big(\im{\zeta}+\big|\zeta\big|^2\Big)\Big(\big|B_0\big|^2-\big|C_0\big|^2\Big)-2\re{\zeta}\im{B_0C_0^\ast}\\
-\frac{v}{c}\biggl[\biggl(\omega\im{\frac{1+\i\zeta^\ast}{1-\i\zeta}\frac{\partial\zeta}{\partial\omega}}+2\big|\zeta\big|^2\biggr)\Big(\big|B_0\big|^2+\big|C_0\big|^2\Big)\\+2\im{\omega\frac{1+\i\zeta^\ast}{1-\i\zeta}\frac{\partial\zeta}{\partial\omega}+2\zeta}\re{B_0C_0^\ast}\biggr]\bigg\}\,,
\end{multline}
accurate to first order in $v/c$. For $v=0$ this result reduces to the one in Ref.~\cite{Asboth2008}. Most of the $v$-dependent terms arise from the frequency dependence of the polarisability. These are the dominant terms in the case of a quasi-resonant excitation of a resonant scatterer, such as a two-level atom, since the prefactor $\tfrac{k}{\zeta}\tfrac{\partial\zeta}{\partial k} \sim \tfrac{\omega}{\Gamma}$ expresses resonant enhancement. The $v$-dependent terms linear in the polarisability $\zeta$ are in perfect agreement with the friction forces known from standard laser cooling theory, both for propagating and for standing waves. For example, assuming identical laser powers from the two sides, giving a standing wave with wavenumber $k_0$, and averaging spatially gives
\begin{equation}
\label{eq:TMM:MolassesForce}
\textsl{\textsf{F}}=-4\hbar k_0^2\big|B_0\big|^2\im{\tfrac{\partial\zeta}{\partial\omega}}v\,,
\end{equation}
for small $\lvert\zeta\rvert$ and to first order in $v/c$, which can be immediately recognised as the friction force in ordinary Doppler cooling, \eref{eq:OMForce}, when one uses the definition of $\zeta$ in~\eref{eq:ZetaDefn}. Finally, by making similar substitutions into~\eref{eq:TMM:GeneralDiffnBC}, we obtain
\begin{equation}
\label{eq:TMM:MolassesDiffn}
\textsl{\textsf{D}}=8(\hbar k_0)^2\im{\zeta}\big|B_0\big|^2\sin^2(k_0 x)\,,
\end{equation}
which, excluding the diffusion effects due to spontaneous emission, matches the standard result in \eref{eq:OMDiffn}. Note, however, that the scattering theory leads to a more general result which is represented by the terms of higher order in $\zeta$. These terms describe the back-action of the scatterer on the field, which we re-iterate is generally neglected in free-space laser cooling theory. The general result in \eref{eq:TMM:ForceTLA} reveals that this velocity-dependent force also acts on a scatterer whose polarisability is independent of the frequency. This is a very general class and we will only focus on such scatterers in the following.
\subsection{General system of a fixed and a mobile scatterer}\label{sec:TMM:GeneralMMC}
\begin{figure}[tb]
\centering
\includegraphics[scale=0.75]{./Diagrams/DiagramsXuereb2009b/Model}
\caption[Physical parameters of a model with a fixed and a mobile scatterer]{Physical parameters of our model. $A$, $B$, etc.\ represent the field mode amplitudes.}
\label{fig:TMM:Model}
\end{figure}
Consider the model in~\fref{fig:TMM:Model} where the scatterer, or `atom', has a polarisability $\zeta$ constant over the frequency range of interest. Letting $M_{\text{a}}$, $M_{\text{p}}$ and $M_{\text{m}}$ be the transfer matrices for the atom, propagation and mirror, respectively, we obtain the relation:
\begin{equation}
\begin{pmatrix}
A(k)\\
B(k)
\end{pmatrix} = M_{\text{a}}M_{\text{p}}M_{\text{m}}\begin{pmatrix}
C(k)\\
D(k)
\end{pmatrix}\,,
\end{equation}
where
\begin{align}
M_{\text{a}} &= \begin{bmatrix}
1+\i\zeta & \i\zeta\left(1- 2\tfrac{v}{c}\right)\hat{P}_{2v}\nonumber\\
-\i\zeta\left(1+2\tfrac{v}{c}\right)\hat{P}_{-2v}& 1-\i\zeta
\end{bmatrix} \\
&= \begin{bmatrix}
M_{11} & M_{12}\hat{P}_{2v}\\
M_{21}\hat{P}_{-2v} & M_{22}
\end{bmatrix}\,,
\end{align}
\begin{equation}
M_{\text{p}} = \begin{bmatrix}
e^{\i kd} & 0\\
0 & e^{-\i kd}
\end{bmatrix}\,,\text{ and }
M_{\text{m}} = \frac{1}{\mathfrak{t}}\begin{bmatrix}
\mathfrak{t}^2-\mathfrak{r}^2 & \mathfrak{r}\\
-\mathfrak{r} & 1
\end{bmatrix}\,.
\end{equation}
The distance between the atom and the mirror is denoted by $d$. Note that the free-propagation transfer matrix $M_\text{p}$ is non-uniform in the $k$-space, and therefore the Doppler shift has an influence on the phase shift accumulated between two scattering events.
\par
The boundary condition is set as follows. We assume that there is no incoming field from the right; therefore $C(k)=0$ for all $k$. The incoming field from the left is assumed to be monochromatic, $B(k)=\mathcal{B}\,\delta(k-k_0)$, with $k_0$ being the pump wavenumber. The resulting field comprises modes with wavenumbers in a narrow region around $k_0$. In the laboratory frame the field mode $A(k)$ interacts with $B(k-2k\tfrac{v}{c})$ and $C^\prime(k)$ through the Doppler shift, and similarly for $D^\prime(k)$. From $C(k)=0$ it directly follows that
\begin{align}
\label{eq:TMM:InputOutput}
A(k)&=\Big(\mathfrak{r} M_{11}e^{\i kd}+M_{12}\hat{P}_{2v}e^{-\i kd}\Big)\Big(\mathfrak{r} M_{21}\hat{P}_{-2v}e^{\i kd}+M_{22}e^{-\i kd}\Big)^{\!-1}B(k)\nonumber\\
&=\frac{1}{M_{22}}\Big(\mathfrak{r} M_{11}e^{\i kd}+M_{12}\hat{P}_{2v}e^{-\i kd}\Big)e^{\i kd}\sum_{n=0}^{\infty}\left(-\mathfrak{r}\frac{M_{21}}{M_{22}}\right)^{\!n} e^{2in kd \left[1-(n+1)\tfrac{v}{c}\right]}B\big(k-2nk\tfrac{v}{c}\big)\,.
\end{align}
We will need the sum of amplitudes, $\mathcal{A}=\int A(k)\mathrm{d} k / \mathcal{B}$, defined relative to the incoming amplitude $\mathcal{B}=\int B(k) dk$. Note that $\int \hat{P}_vf(k)\mathrm{d} k=\int f(k)\mathrm{d} k$. Thus, to first order in $\tfrac{v}{c}$,
\begin{equation}
\label{eq:TMM:IntA}
\mathcal{A}=\frac{M_{12}}{M_{22}}+\left(\frac{M_{12}}{M_{22}}-\frac{M_{11}}{M_{21}}\right)\sum_{n=1}^{\infty}\left(-\mathfrak{r}\frac{M_{21}}{M_{22}}\right)^{\!n} \left[1+2in(n-1)k_0 d\tfrac{v}{c}\right]e^{2in k_0d}\,.
\end{equation}
It is worth introducing the reference point at a distance $L =2 N \pi/k_0$ from the fixed mirror, where the integer $N$ is such that the moving atom's position $x$ is within a wavelength of this reference point. Then the atom--mirror distance can be replaced by $d=L-x$, and $k_0 L$ drops from all the trigonometric functions. The solution, \eref{eq:TMM:IntA}, has a clear physical meaning, in that the reflected field, $\mathcal{A}$, can be decomposed into an interfering sum of fields: the first term is the reflection directly from the atom, whereas the summation is over the electric field undergoing successive atom--mirror round-trips. We can also write the preceding expression in closed form:
\begin{align}
\label{eq:TMM:ReflectedField}
\mathcal{A} =\frac{1}{1-\i\zeta}\Biggl\{&\i\zeta + \mathfrak{r} \frac{e^{-2 \i k_0x}}{1 - \i\zeta -\mathfrak{r} \i\zeta e^{-2 \i k_0x}}\nonumber\\
&-2 \i\frac{v}{c} \zeta \Biggl[1 - \frac{\mathfrak{r}^2 e^{-4 \i k_0x}}{\big(1 - \i\zeta -\mathfrak{r} \i\zeta e^{-2 \i k_0x}\big)^2}-2\i k_0(L-x) \frac{\mathfrak{r}^2 (1-\i\zeta) e^{-4 \i k_0x}}{\big(1 - \i\zeta -\mathfrak{r} \i\zeta e^{-2 \i k_0x}\big)^3}\Biggr] \Biggr\}\,.
\end{align}
This result is valid for arbitrary $\zeta$. The main virtue of our approach is clearly seen, in that we can smoothly move from $\zeta=0$, which indicates the absence of the mobile scatterer, to $|\zeta|\rightarrow\infty$, which corresponds to a perfectly reflecting mirror, i.e., a moving boundary condition for the electromagnetic field.
\par
Let us outline some of the generic features of the above calculation that would be encountered in a general configuration of scatterers. By using the formal Doppler shift operators, we benefit from the transfer matrix method in keeping the description of the system as a whole within $2\times 2$ matrices. The input-output relation for the total system is always obtained in a form similar to that of \eref{eq:TMM:InputOutput}. As long as the Doppler broadening is well below the transient time broadening of the system, the calculation of forces and diffusion requires solely the sum of the mode amplitudes. An important point is that the integrated action of the Doppler shift operator $\hat P_v$ on monochromatic fields is a shift in $k$-space. Therefore, by interchanging the order of terms and putting the $\hat P_v$ terms just to the left of the input field amplitudes, they can be eliminated, such as in \eref{eq:TMM:IntA}. Finally, up to first order in $v/c$, the resulting power series, a trace of multiple reflections, can be evaluated in a closed form, as shown in \eref{eq:TMM:ReflectedField}. In conclusion, the illustrated method lends itself for the description of more complex schemes, for example, the cooling of a moving, partially reflective mirror in a high-finesse Fabry-Perot resonator \cite{Bhattacharya2008}. This scheme, however, rapidly increases in complexity with the summations becoming potentially unmanageable. An alternative is possible, in that the matrix $\hat{M}$ is invertible to first order in $v/c$ so that $\textsl{\textsf{F}}$ and $\textsl{\textsf{D}}$ can be evaluated in closed form for a mobile scatterer in a generic optical system. We will investigate this solution in \sref{sec:TMM:General} after some illustrative examples.
\subsubsection{Force acting on the mobile scatterer}
To obtain the force on the moving scatterer, we also need to evaluate $C^{\prime}(k)$ and $D^{\prime}(k)$:
\begin{align}
\label{eq:TMM:Cprime}
\begin{pmatrix}
C^{\prime}(k)\\
D^{\prime}(k)
\end{pmatrix} &= \begin{bmatrix}
1-\i\zeta & -\i\zeta \left(1 - 2 \tfrac{v}{c}\right)\hat{P}_{2v}\\
\i\zeta \left(1 + 2 \tfrac{v}{c}\right)\hat{P}_{2v}^{-1} & 1+\i\zeta
\end{bmatrix} \begin{pmatrix}
A(k)\\
B(k)
\end{pmatrix}\,,
\end{align}
where we applied the inverse of the transfer matrix $M_\text{a}$. Next, we make the following definitions:
\begin{equation}
\mathbb{A}=|\mathcal{A}|^2\,,\quad \mathbb{B}=|\mathcal{B}|^2\,,
\mathbb{C}=\frac{1}{\mathbb{B}}\left|\int C^{\prime}(k)\mathrm{d} k\right|^2\,,\text{ and }
\mathbb{D}=\frac{1}{\mathbb{B}}\left|\int D^{\prime}(k)\mathrm{d} k\right|^2\,,
\end{equation}
and a simple calculation leads to
\begin{align}
\mathbb{C}=&\left|1-\i\zeta\right|^2\mathbb{A}+\left|\i\zeta\big(1-2\tfrac{v}{c}\big)\right|^2+2\re{\i\zeta^{\ast}(1-\i\zeta)\big(1-2\tfrac{v}{c}\big)\mathcal{A}}\text{,}\\
\mathbb{D}=&\left|\i\zeta\big(1+2\tfrac{v}{c}\big)\right|^2\mathbb{A}+\left|1+\i\zeta\right|^2+2\re{\i\zeta(1+\i\zeta^{\ast})\big(1+2\tfrac{v}{c}\big)\mathcal{A}}\,.
\end{align}
Thereby the force acting on the scatterer is obtained as
\begin{align}
\label{eq:TMM:GeneralForce}
\textsl{\textsf{F}} &= (\hbar\omega/c)\mathbb{B}\left(\mathbb{A}+1-\mathbb{C}-\mathbb{D}\right)\nonumber\\
&= -2\hbar k_0 \mathbb{B}\biggl\{\Bigl[\lvert\zeta\rvert^2\bigl(1+2\tfrac{v}{c}\bigr)+\im{\zeta}\Bigr]\mathbb{A}+\lvert\zeta\rvert^2\bigl(1-2\tfrac{v}{c}\bigr)-\im{\zeta}\nonumber\\
&\phantom{= -2\hbar k_0 \mathbb{B}\biggl\{\ }+2\re{\Bigl(\i\re{\zeta}+\lvert\zeta\rvert^2+2\tfrac{v}{c}\im{\zeta}\Bigr)\mathcal{A}}\biggr\}\,,
\end{align}
where $\mathcal{A}$ has to be substituted from \eref{eq:TMM:ReflectedField}.
\begin{figure}[t]
\centering
\includegraphics[width=1.5\figwidth]{Diagrams/TMM/friction}
\caption[The position dependence of the cooling coefficient]{The position dependence of the cooling coefficient $\varrho$ in the relation $\textsl{\textsf{F}}=-\varrho v$ for the velocity-dependent force acting on the mobile scatterer in \fref{fig:TMM:Model}, for various scattering parameters $\zeta$, evaluated by using \eref{eq:TMM:ReflectedField} and \eref{eq:TMM:GeneralForce} with $k_0 L=2\pi\times 100$. The fixed mirror is assumed to be a perfect mirror. In order to fit all the curves into the same range, they are divided by the factors indicated in the figure. Note that $x$ is defined differently from \fref{fig:MMC:Analytic-Spatial}; the mobile scatterer here is closer to the fixed mirror for \emph{increasing} $x$.}
\label{fig:TMM:FrictionPosnDependence}
\end{figure}
The cooling coefficient $\varrho$, defined through the relation $\textsl{\textsf{F}}=-\varrho v$, is plotted in \fref{fig:TMM:FrictionPosnDependence} as a function of the position $x$ in a half-wavelength range for various values of $\zeta$. When varying the coupling strength from $\zeta=0.01$ up to $\zeta=1$, the cooling coefficient transforms between two characteristic regimes. For small coupling the linear velocity dependence tends to a simple sinusoidal function while, for large coupling, the friction exhibits a pronounced resonance in a narrow range. This resonance arises from the increased number of reflections between the mobile scatterer and the fixed mirror. It can be observed that the resonance shifts towards $k_0 x =\pi$ on increasing $\zeta$. In the opposite limit of small $\zeta$, the maximum friction is obtained periodically at $\big(n-\tfrac{1}{4}\big)\pi/2$ according to the sinusoidal function. The position of the maximum friction is plotted in \fref{fig:TMM:Phase}, showing the transition from $7\pi/8$ to $\pi$. The maximum friction force is plotted in \fref{fig:TMM:Maxfric}, showing the two limiting cases of $\zeta^2$ behaviour, in the limit of small $\zeta$, and $\zeta^6$ behaviour, in the limit of large $\zeta$. These two cases are described in \sref{sec:MovingAtom} and \sref{sec:Mirrorcooling}, respectively.
\begin{figure}[t]
\centering
\subfigure[]{
\includegraphics[width=1.5\figwidth]{./Diagrams/DiagramsXuereb2009b/maxcool}
\label{fig:TMM:Phase}
}\\
\subfigure[]{
\includegraphics[width=1.5\figwidth]{Diagrams/TMM/maxfric}
\label{fig:TMM:Maxfric}
}
\caption[Position and magnitude of maximum force as a function of $\zeta$]{(a)~The position of the maximum friction force, $k_0x_m$, as a function of the dimensionless scattering parameter $\zeta$ (on a semilog scale) acting on the scatterer in \fref{fig:TMM:Model}, with the fixed mirror being a perfect mirror. This position shifts from $7\pi/8$ to $\pi$ on increasing $\zeta$. (b)~A similar plot, showing the maximum friction force as a function of $\zeta$ (on a log-log scale) with $k_0L=2\pi\times 100$. In the limit of small $\zeta$, the force scales as $\zeta^2$ [cf.~\eref{eq:TMM:MirrorCoolForce}; dashed line] whereas in the limit of large zeta it scales as $\zeta^6$ [cf.~\eref{eq:TMM:FPCoolForce}; dotted line].}
\end{figure}
\subsubsection{Diffusion coefficient}
The calculation of the diffusion coefficient proceeds along the same lines as that corresponding to a single beamsplitter, shown in \aref{sec:TMM:AppDiffusion}. The difference is that the modes $B(k)$ and $C'(k)$ around the mobile scatterer are not independent, for the reflection at the fixed mirror mixes them. Therefore, all the modes $A$, $B$, $C'$, and $D'$ have to be expressed in terms of the leftmost and rightmost incoming modes, $B(k)$ and $C(k)$, respectively. Instead of the derivation of such a general result for the diffusion, here we will restrict ourselves to the special case of $\mathfrak{r}=-1$ ($\Leftrightarrow$ perfect mirror) and real $\zeta$ ($\Leftrightarrow$ no absorption in the moving mirror). In this special case the diffusion calculation simplifies a lot, because (i) the perfect mirror prevents the modes $C$ from penetrating into the interaction region, and (ii) quantum noise accompanying absorption does not intrude in the motion of the scatterer.
\par
Only the modes $\hat B(k)$ impart independent quantum fluctuations. When all the amplitudes around the scatterer are expressed in terms of $\hat B(k)$, and are inserted into the force correlation function given in \eref{eq:FDTDeltaDependence}, the commutator $[\hat b(t),\hat b^\dagger(t')]$ appears in all the terms (see \aref{sec:TMM:AppDiffusion}). Straightforward algebra leads to
\begin{equation}
\label{eq:TMM:GeneralDiff}
\textsl{\textsf{D}}=\hbar^2 k_0 ^2\mathbb{B}(\mathbb{A} +1-\mathbb{C}-\mathbb{D})^2\,.
\end{equation}
We emphasise that the above result is not general: the diffusion is not necessarily proportional to the square of the force. This simple relation here follows from the assumptions, $\mathfrak{r}=-1$ and $\im{\zeta}=0$, declared above.
\par
To be consistent with the calculation of the friction force linear in velocity, the diffusion should be evaluated only for $v=0$. From the ratio of these two coefficients, the steady-state temperature can be deduced. The velocity-independent components of the modes obey the following relations: $\mathbb{A} = 1$ and $\mathbb{C'} = \mathbb{D'}$ (all incoming power is reflected). Therefore the diffusion coefficient further simplifies,
\begin{equation}
\textsl{\textsf{D}} = 4 \hbar^2 k_0 ^2\mathbb{B} \Bigg(1-\frac{1}{\big|1 - \i\zeta + \i\zeta e^{-2 \i k_0x}\big|^2}\Bigg)^2\,.
\end{equation}
\begin{figure}[tbp]
\centering
\includegraphics[width=1.5\figwidth]{Diagrams/TMM/temperatures}
\caption[Characteristic temperature for the two-scatterer system]{Characteristic temperature for the two-scatterer system of \fref{fig:TMM:Model}, given by the ratio of the diffusion and cooling coefficients in the points where the friction is maximum, as a function of the dimensionless scattering parameter $\zeta$ on a log-log scale. Constant and $1/\zeta^2$ dependence can be read off in the limits of small and large $\zeta$, respectively. The fixed mirror is a perfect mirror.}
\label{fig:TMM:Temperature}
\end{figure}
In \fref{fig:TMM:Temperature}, the temperature $k_\text{B} T= \textsl{\textsf{D}}/\varrho$, is plotted as a function of the scattering parameter $\zeta$. The friction and the diffusion coefficients are taken at the position where the friction is maximum, as shown in \fref{fig:TMM:Phase}.
The two limits of small and large scattering parameter $\zeta$ will be analysed in \sref{sec:MovingAtom} and \sref{sec:Mirrorcooling}, respectively.
\subsection{Atom in front of a perfect mirror}\label{sec:MovingAtom}
An atom pumped with a far off-resonance beam can be modelled as a moving mirror with small and real $\zeta$. In this section we accordingly truncate our expressions to second order in $\zeta$. We also assume that the fixed mirror is perfect; i.e., $\mathfrak{r}=-1$ and $\mathfrak{t}=0$. Thus,
\begin{equation}
\label{eq:TMM:Force}
\textsl{\textsf{F}}=2\hbar k_0\mathbb{B}\big[2\zeta\im{\mathcal{A}}-2\zeta^2\re{\mathcal{A}}-\zeta^2\big(1+\tfrac{v}{c}\big)\mathbb{A}-\zeta^2\big(1-\tfrac{v}{c}\big)\big]\,.
\end{equation}
To obtain $\textsl{\textsf{F}}$ to second order in $\zeta$, we need $\mathcal{A}$ to first order. Using~\eref{eq:TMM:IntA} and~\eref{eq:TMM:Force}, we obtain:
\begin{equation}
\label{eq:TMM:CurlyA}
\mathcal{A}=-e^{-2\i k_0x}+\zeta(\i-2\i e^{-2\i k_0x}+\i e^{-4\i k_0x})+\zeta\tfrac{v}{c}\big[-2\i+2\i e^{-4\i k_0x}-4 k_0 (L-x) e^{-4\i k_0 x}\big]\,,
\end{equation}
and
\begin{align}
\label{eq:TMM:MirrorCoolForce}
\textsl{\textsf{F}}=4\hbar k_0\mathbb{B}\big(&\zeta\sin(2 k_0x)-\zeta^2\big\{2\sin^2( k_0x)\big[4\cos^2( k_0x)-1\big]\big\}\nonumber\\
&-\zeta^2\tfrac{v}{c}\big[4\sin^2(2 k_0x)-4 k_0 (L-x) \sin(4 k_0x)\big]\big)\,,
\end{align}
in agreement with \eref{eq:MMC:AnalyticLongFriction}. In the far field ($x\gg\lambda$), the dominant friction term in the preceding expression is the last term, which renders the $\sin(4 k_0x)$ position dependence shown in \fref{fig:TMM:Friction} for $\zeta=0.01$.
\par
We are now in a position to derive the diffusion coefficient for this system. By substituting~\eref{eq:TMM:CurlyA} into~\eref{eq:TMM:GeneralDiff} and setting $v=0$, we obtain
\begin{equation}
\textsl{\textsf{D}}=8(\hbar k_0)^2\zeta^2\mathbb{B}\,.
\end{equation}
This allows us to estimate the equilibrium temperature for such a system at a position of maximum friction:
\begin{equation}
\label{eq:TMM:MMCTemp}
T\approx\frac{\hbar}{k_{\text{B}}\tau}\text{, where }\tau=2(L-x)/c\,,
\end{equation}
which we note is identical in form to the Doppler temperature for a two-level atom undergoing free-space laser cooling~\cite{Metcalf2003}, but where we have replaced the upper state lifetime, $1/\Gamma$, by the round-trip time delay between the atom and the mirror.\footnote{This expression lacks the geometrical factor in \eref{eq:SemiclassicalMMCTemperature}; the reason for this is that the semiclassical treatment of a TLA is, strictly speaking, inconsistent with the assumption $\im{\zeta}=0$; the dominant term of the diffusion is then the first term in \eref{eq:TMM:GeneralDiffnBC}.} Note that this temperature corresponds to the constant value presented in \fref{fig:TMM:Temperature} for $\zeta<0.1$.
\subsection{Optical resonator with mobile mirror}\label{sec:Mirrorcooling}
After the small polarisability case of the previous section, we will now consider the $|\zeta| \rightarrow \infty$ limit. We again assume that the fixed mirror of the resonator is perfect, with $\mathfrak{r}=-1$, and that $C=0$. For simplicity, we assume that the moving mirror has a real polarisability; i.e., it is lossless. We expand the field mode amplitudes as power series in $v/c$, such that $\mathcal{A}=\mathcal{A}_0+\tfrac{v}{c}\mathcal{A}_1+\dots$, and similarly for $\mathcal{C}^{\prime}$.
Let us first calculate the field in the resonator for $v=0$. It follows from \eref{eq:TMM:Cprime} that
\begin{equation}
\mathcal{C}_0^\prime = (1-\i \zeta) \mathcal{A}_0 -\i\zeta = -\frac{e^{-2\i\varphi}}{1-\i\zeta+\i\zeta e^{-2 \i \varphi}}\,,
\end{equation}
with $\varphi = k_0 d$, which has a maximum at $\varphi_0$ obeying
\begin{equation}
\tan(2\varphi_0) = - \frac{1}{\zeta}\,.
\end{equation}
In the limit of $\zeta\rightarrow \infty$, the resonance is Lorentzian:
\begin{equation}
\mathcal{C}_0^\prime = -\frac{e^{-2\i\varphi}}{2 \i (1-\i\zeta) \left[(\varphi-\varphi_0) - \i \tfrac{1}{4 \zeta^2} \right]}\,,
\end{equation}
with a width of $1/(4\zeta^2)$.
The perfect mirror reflects the total power incoming from the left, $\mathbb{B}$. Moreover, for real $\zeta$, there is no absorption in the moving mirror, so the outgoing intensity has to be equal to the incoming one: $\mathbb{A} = 1$. This is true if $v=0$; for $v\neq 0$, the field can do work on the mirror. The expansion of the back-reflected intensity to linear order in velocity reads $\mathbb{A} = 1 + 2\tfrac{v}{c} \re{\mathcal{A}_0^*\mathcal{A}_1}$. Extracting the velocity-dependent terms for the general form of the force in \eref{eq:TMM:GeneralForce}, it reduces to
\begin{equation}
\textsl{\textsf{F}}_1= \tfrac{v}{c} 4 \hbar k_0 \mathbb{B} \zeta\im{\mathcal{A}_1\big/\left(1 + \i \zeta -\i \zeta e^{2\i\varphi}\right)}\,,
\end{equation}
which, after some algebra, leads to
\begin{equation}
\label{eq:TMM:FPCoolForce}
\textsl{\textsf{F}}_1 = - \tfrac{1}{2} \tfrac{v}{c} \hbar k_0^2 L \frac{(\varphi-\varphi_0)}{\zeta^4\, \left[\left( \frac{1}{4\zeta^2}\right)^2 + (\varphi-\varphi_0)^2 \right]^3} \mathbb{B}\,.
\end{equation}
On substituting $\kappa = c/\big(4L\zeta^2\big)$, $\Delta_C = - c(\varphi-\varphi_0)/L$, $\eta^2/(2\kappa)=\mathbb{B}$, and $G=c^2k_0^2/L^2$, the friction force renders that derived from the usual radiation pressure Hamiltonian in \aref{sec:RadPressCool}.
Expressing the field modes interacting with the mobile mirror in terms of the input field mode and performing a calculation similar to that leading to \eref{eq:TMM:DiffCoeff} readily gives
\begin{equation}
\textsl{\textsf{D}} \approx 4(\hbar k_0)^2 |\mathcal{C}^\prime_0|^4 \mathbb{B} \approx \frac{(\hbar k_0)^2\mathbb{B}}{ 4\zeta^4 \left[ \left( \frac{1}{4\zeta^2}\right)^2 + (\varphi-\varphi_0)^2\right]^2}\,.
\end{equation}
The resulting temperature thereby attains a minimum at $4 \zeta^2 (\varphi-\varphi_0)=1$, i.e., $\Delta_C=-\kappa$, in analogy with free-space Doppler cooling, at which point we have
\begin{equation}
\label{eq:TMM:MirrorCoolingTemp}
k_{\text{B}} T \approx \frac{\hbar c}{4 \zeta^2 L} = \hbar \kappa\,.
\end{equation}
Again, this asymptotic behaviour is reflected in \fref{fig:TMM:Temperature} for large $\zeta$. We note the similarity of the preceding expression with the temperature of an atom cooled in a cavity, in the good-cavity limit~\cite{Horak2001}. We conjecture that this is due to the fact that both systems can be considered to involve the coupling of a laser with a system having a decay rate $\kappa$. This result also holds for the case of an atom undergoing mirror-mediated cooling, as can be seen in \eref{eq:TMM:MMCTemp}.
\par
It is also important to note that the above discussion only treats the effects of the light fields on the scatterer. As such, the temperature limit, \eref{eq:TMM:MirrorCoolingTemp}, is intrinsic to the light forces, and the mechanical damping and heating processes present in a real, macroscopic mirror-cooling setup are not taken into account. In practice, these heating processes may dominate over the heating induced by the quantum noise in the light field \cite{Saulson1990, Cohadon1999}. In such cases, radiation pressure cooling is a possible means to lower the equilibrium temperature owing to the additional, optical, damping process.
\subappendicesstart
\subsection{Appendix: The Doppler shift operator\label{sec:POper}}
\begin{figure}[t]
\centering
\includegraphics[scale=0.75]{./Diagrams/DiagramsXuereb2009b/POper}
\caption[Reflection and transmission of a moving scatterer]{Reflection and transmission of a moving scatterer. $B$ and $C$ are the input field modes, and $A$ is the output field mode. A further output field mode (`$D$') is not drawn because it is not relevant to our discussion here.}
\label{fig:TMM:POper}
\end{figure}
Consider the situation in~\fref{fig:TMM:POper}, in the laboratory frame, where $\text{S}$ is a scatterer, and suppose that $B$ and $C$ are known. $A(k)$ has contributions arising from both $B\big(k+2k\tfrac{v}{c}\big)$ and $C(k)$, where $k$ is any arbitrary wave number, written separately as:
\begin{align}
A_B(k)&=a_1 B(k+2k\tfrac{v}{c})\,\text{, and}\\
A_C(k)&=a_2 C(k)\,.
\end{align}
We can therefore express $A(k)$ as
\begin{equation}
A(k)=a_1 B(k+2k\tfrac{v}{c})+a_2 C(k)\,.
\end{equation}
Defining $\hat{P}_v$ by $\hat{P}_v:f(k)\mapsto f(k+k\tfrac{v}{c})$, we have
\begin{equation}
A(k)=\hat{P}_{2v}a_1 B(k)+a_2 C(k)\,.
\end{equation}
A similar expression, involving $\hat{P}_v^{-1}=\hat{P}_{-v}$, holds for $D(k)$. These two operators can then be introduced into~\eref{eq:TMM:TM_fix} as part of the Lorentz transformation, and thus into the transfer matrix for the moving scatterer, giving rise to the form shown in~\eref{eq:TMM:LML}. The resulting transformation, for the transfer matrix $M$, of a scatterer moving with velocity $v$ can be written as:
\begin{equation}
\begin{bmatrix}
(1-\tfrac{v}{c})\hat{P}_v & 0\\
0 & (1+\tfrac{v}{c})\hat{P}_v^{-1}
\end{bmatrix}M\begin{bmatrix}
(1+\tfrac{v}{c})\hat{P}_v^{-1} & 0\\
0 & (1-\tfrac{v}{c})\hat{P}_v
\end{bmatrix}\,,
\end{equation}
to first order in $\tfrac{v}{c}$, where the ordering of the elements of $M$ is as described in the text. Note that this relation is general, in the sense that the elements of $M$ can depend on $k$ (see~\sref{sec:Force}).
For any finite $v$, $\hat{P}_v$ is trivially a bounded operator, having unit norm. This property follows from the important relation $\int \hat{P}^m_v f(k)\mathrm{d} k = \int f(k)\mathrm{d} k\text{,}$ for any function $f(k)$ and any integer $m$.
This operation can be generalised to $n=2,3$ pairs of modes in $n$ orthogonal dimensions. In two dimensions, this could correspond to a beamsplitter oriented at $45^\circ$ to the $x$-axis; in three dimensions, the beamsplitter would be oriented at $45^\circ$ to all three coordinate axes. We define a new operator by $\hat{S}_i(\v{v}):f(\v{k})\mapsto f(\v{k}+k_i\tfrac{v_i}{c}\mathbf{e}_i)$, where $\mathbf{e}_i$ is the unit vector along the $i$th coordinate axis, $\v{v}$ is the velocity vector of the scatterer, and $\mathbf{x}=(x_1,x_2,\dots)$ for any vector $\mathbf{x}$. In particular, we have $\hat{P}_{v}=\hat{S}_1(v\mathbf{e}_1)$. Now, let $\hat{\m{L}}(\v{v})$ be the $2n\times 2n$ matrix operator:
\begin{equation}
\begin{bmatrix}
(1+\tfrac{v_1}{c})\hat{S}_1^{-1}(\v{v}) & 0 & 0 & \cdots\\
0 & (1-\tfrac{v_1}{c})\hat{S}_1(\v{v}) & 0 & \cdots\\
0 & 0 & (1+\tfrac{v_2}{c})\hat{S}_2^{-1}(\v{v}) & \cdots\\
\vdots & \vdots & \vdots & \ddots
\end{bmatrix}\,.
\end{equation}
Then, the transfer matrix for the scatterer moving with velocity $\v{v}$ is given by
\begin{equation}
\hat{\m{L}}(\mathbf{-v})\,M\,\hat{\m{L}}(\v{v})\,,
\end{equation}
where $M$ is the original transfer matrix for the scatterer, obtained in a manner such as that used to obtain~\eref{eq:TMM:M0}, for example. The ordering of the elements of $M$ is such that it acts on the vector $\big(A_1(\v{k}),B_1(\v{k}),A_2(\v{k}),\dots\big)$:
\begin{equation}
\begin{pmatrix}
C_1(\v{k})\\
D_1(\v{k})\\
C_2(\v{k})\\
\vdots
\end{pmatrix}=\hat{\m{L}}(\mathbf{-v})\,M\,\hat{\m{L}}(\v{v})\begin{pmatrix}
A_1(\v{k})\\
B_1(\v{k})\\
A_2(\v{k})\\
\vdots
\end{pmatrix}\,,
\end{equation}
with $A_i(\v{k})$ being the outgoing mode and $B_i(\v{k})$ the incoming mode along the $i$th axis in the negative half-space (assuming that the scatterer is at the origin); and $C_i(\v{k})$ the incoming mode and $D_i(\v{k})$ the outgoing mode in the positive half-space.
\subsection{Appendix: Quantum correlation function of the force operator}\label{sec:TMM:AppDiffusion}
In quantum theory, we need to replace the mode amplitudes $A(k)$ by operators $\hat A(k)$, and similarly for the $B$, $C$, and $D$ modes.
The cross-correlation of these operators is not trivial because of the boundary condition connecting the mode amplitudes $A(k)$, $B(k)$, $C(k)$ and $D(k)$. The input modes $\hat C(k)$ and $\hat B(k)$ can be considered independent, and the commutator is non-vanishing for the creation and annihilation operators of the same mode, e.g.,
\begin{align}
\Big[ \hat B(k), {\hat B}^\dagger(k^\prime) \Big] & = \left[ \hat C(k), {\hat C}^\dagger(k^\prime) \right] = \frac{\hbar \omega}{2 \epsilon_0 V} \delta_{k,k^\prime} \text{,}\\
\Big[ \hat B(k), {\hat C}^\dagger(k^\prime) \Big] & =0\,,
\end{align}
assuming a discrete mode index of $k$, and a quantisation volume $V=\sigma_\text{L} l$ with $\sigma_\text{L}$ being the mode area and $l$ a fictitious total length of the space in one dimension.
\par
In using \eref{eq:FDT}, expressions for $\textsl{\textsf{F}}$ correct to order $n$ in $v/c$ are compared to $\textsl{\textsf{D}}$ to order $(n-1)$ in this same parameter. Since our force expressions are accurate up to first order in $v/c$, therefore, we need only consider terms that contribute to the diffusion in the $v=0$ case. In the quantum description, the linear relation for the output modes is
\begin{align}
\label{eq:ADNoise}
A(k) = {}& \mathfrak{t} C(k) + \mathfrak{r} B(k) + \sqrt{\epsilon} E \\
D(k) = {}& \mathfrak{r} C(k) + \mathfrak{t} B(k) + \sqrt{\epsilon} E\,,
\end{align}
where the transmission $\mathfrak{t}=1/M_{22}=1/(1-\i\zeta)$, and reflection $\mathfrak{r}=M_{12}/M_{22} = \i\zeta/(1-\i\zeta)$, as above. The amplitude $E$ represents the quantum noise fed into the system by the absorption. For $\epsilon = 1-\big(|\mathfrak{r}|^2 + |\mathfrak{t}|^2\big)$, this noise ensures that the output modes obey the same commutation relations as the input ones, namely
\begin{align}
\Big[ \hat A(k), {\hat A}^\dagger(k^\prime) \Big] &= \Big[ \hat D(k), {\hat D}^\dagger(k^\prime) \Big] = \frac{\hbar \omega}{2 \epsilon_0 V} \delta_{k,k^\prime}\,,\\
\Big[ \hat A(k), {\hat D}^\dagger(k^\prime) \Big] & =0\,.
\end{align}
However, the linear dependence implies that commutators between input and output mode operators are
\begin{align}
\Big[ \hat A(k), {\hat B}^\dagger(k^\prime) \Big] = \mathfrak{r} \Big[ \hat B(k), {\hat B}^\dagger(k^\prime) \Big] \,,\\
\Big[ \hat A(k), {\hat C}^\dagger(k^\prime) \Big] = \mathfrak{t} \Big[ \hat C(k), {\hat C}^\dagger(k^\prime) \Big] \,,
\end{align}
and similar relations hold for the cross-commutators with $D(k)$.
The proper treatment of quantum fluctuations and the derivation of correlation functions require that the explicit time dependence be considered. Let us introduce the time-varying operators
\begin{equation}
\hat a(t) = \sum_k \hat A(k) e^{-\i\omega t}\,,
\end{equation}
and similarly for $ \hat b(t)$, $ \hat c(t)$ and $ \hat d(t)$. It follows that
\begin{equation}
\Big[ \hat a(t), {\hat a}^\dagger (t^\prime) \Big] = \frac{\hbar \omega}{2 \epsilon_0 V} \sum_k e^{-\i\omega (t-t^\prime)} \approx \frac{\hbar \omega}{2 c \epsilon_0 \sigma_\text{L}} \delta(t-t^\prime)\,.
\end{equation}
Here we made use of the fact that the non-excited vacuum modes, having a bandwidth that is much broader than that of the pump, also contribute to the summation, allowing us to approximate the summation as an integral over all frequencies. The Fourier-type integral then yields $\delta$-function on the much slower timescale of interest. A similar commutation relation applies to the operators $ \hat b(t)$, $ \hat c(t)$, and $ \hat d(t)$. The cross-commutators can be derived directly from those concerning the modes, e.g.,
\begin{equation}
\Big[ \hat a(t), {\hat b}^\dagger (t^\prime) \Big] = \mathfrak{r} \frac{\hbar \omega}{2 c \epsilon_0 \sigma_\text{L}} \delta(t-t^\prime)\,.
\end{equation}
The force operator is
\begin{equation}
\label{eq:TMM:ForceOp}
\hat \textsl{\textsf{F}} = \sigma_\text{L}\big[\hat \boldsymbol{T}_{xx}({x\rightarrow 0^+})-\hat \boldsymbol{T}_{xx}({x\rightarrow 0^-})\big]\,,
\end{equation}
as before, where
\begin{align}
\hat \boldsymbol{T}_{xx}(x\rightarrow 0) = \begin{cases}
-2\epsilon_0 \Big[\hat a^\dagger(t) \hat a(t) + \hat b^\dagger(t) \hat b(t)\Big]\vspace{0.5em}&\text{for }x\rightarrow 0^-\\
-2\epsilon_0 \Big[\hat c^\dagger(t) \hat c(t) + \hat d^\dagger(t) \hat d(t)\Big]&\text{for }x\rightarrow 0^+
\end{cases}
\end{align}
is the quantised stress tensor. Assuming that the field is in a coherent state, in all normally ordered products, the mode amplitude operators can be replaced by the corresponding coherent state amplitudes, which are c-numbers: e.g.,\ $\hat A(k) \rightarrow A(k)$ and $\hat A^\dagger(k) \rightarrow A^\ast(k)$. The force operator in \eref{eq:TMM:ForceOp} is normally ordered in this way; therefore coherent-state fields render, as a mean value of the quantum expressions, the force \eref{eq:TMM:MSTForce} derived from the classical theory based on the definition \eref{eq:TMM:Forcedef}. Non-trivial quantum effects arise from non-normally ordered products, such as the fourth-order product terms of the second order correlation function of the force \eref{eq:FDTDeltaDependence}. These terms can be evaluated straightforwardly by invoking the above-derived commutators to rearrange the product into normal order. As an example, consider
\begin{equation}
\big\langle \hat a^\dagger(t) \hat a(t) \hat a^\dagger(t^\prime) \hat a(t^\prime)\big\rangle =\big\langle \hat a^\dagger(t) \hat a^\dagger(t^\prime) \hat a(t) \hat a(t^\prime)\big\rangle- \big\langle \hat a^\dagger(t) \hat a(t^\prime)\big\rangle \frac{\hbar \omega}{2 c \epsilon_0 \sigma_\text{L}} \delta(t-t^\prime)\,.
\end{equation}
For radiation fields in coherent state, the first term is cancelled from the correlation function by the $\langle \hat a^\dagger(t) \hat a(t^\prime)\rangle^2$ term. The coefficient of $\delta (t-t^\prime)$ in the second term is in normal order and can be replaced by c-numbers and then calculated identically as the force in \sref{sec:Force},
\begin{equation}
\label{eq:TMM:NormalOrderMean}
\big\langle \hat a^\dagger(t) \hat a(t)\big\rangle \approx \Big| \sum A (k) \Big|^2 = \frac{\hbar \omega}{2 c \epsilon_0 \sigma_\text{L}} |A|^2\,,
\end{equation}
in terms of the photo-current intensity $|A|^2$.
Assembling all similar contributions, originating from the non-vanishing commutators $[b,b^\dagger]$, $[c,c^\dagger]$, $[d,d^\dagger]$, $[a,b^\dagger]$, etc., one obtains~\eref{eq:TMM:DiffCoeff} presented in \sref{sec:MSTDiffusion}.
\subsection{Appendix: Mirror cooling via the radiation pressure coupling Hamiltonian\label{sec:RadPressCool}}
We describe a generic optomechanical system composed of a single, damped-driven field mode coupled to the motion of a massive particle, whose Hamiltonian is given by~\cite{Courty2001,Vitali2003}
\begin{equation}
\hat{\mathcal{H}} =\hbar \omega_c \hat{a}^\dagger \hat{a} + \i \hbar \eta (\hat{a}^\dagger e^{-\i \omega t} - \hat{a} e^{\i \omega t}) + \frac{\hat{p}^2}{2m} +V(\hat{x}) +\hbar G \hat{a}^\dagger \hat{a} \hat{x}\,.
\end{equation}
where $\hat{a}$ and $\hat{a}^\dagger$ are the annihilation and creation operators, respectively, of the mode; $\hat{x}$ and $\hat{p}$ are the position and momentum operators associated with the motion; and where we drop the carets to signify expectation values. The mode is driven by a coherent field with an effective amplitude $\eta$ and frequency $\omega$. This Hamiltonian describes, for example, the radiation pressure coupling of a moving mirror to the field in a Fabry-Perot resonator. In this case the coupling constant is $G=\omega_c/L$, rendering the cavity mode frequency detuning $\omega_cx/L$ provided the mirror is shifted by an amount $x$. Since the cavity mode is lossy with a photon escape rate of $2 \kappa$, the total system is dissipative. Thereby, with a proper setting of the parameters, in particular the cavity detuning $\Delta_C=\omega-\omega_C$, the mirror motion can be cooled. We will determine the corresponding friction force linear in velocity.
In a frame rotating at frequency $\omega$, the Heisenberg equation of motion for the field mode amplitude reads
\begin{equation}
\dot{\hat{a}} = \left[ \i (\Delta_C -G \hat{x}) - \kappa\right] \hat{a} + \eta\,.
\end{equation}
where the noise term is omitted. We assume that the mirror moves along the trajectory $x(t) \approx x + v t$ with fixed velocity $v$ during the short time that is needed for the field mode to relax to its steady-state. The variation of $\hat{a}$ arises from the explicit time dependence and from the motion of the mirror. A steady-state solution is sought in the form of $\hat{a} \approx \hat{a}^{(0)}(x) + v \hat{a}^{(1)}(x)$. On replacing this expansion into the above equation, and using the hydrodynamic derivative $\tfrac{\mathrm{d}}{\mathrm{d} t} \rightarrow \tfrac{\partial}{\partial t} + v \tfrac{\partial}{\partial x}$, one obtains a hierarchy of equations of different orders of the velocity $v$. To zeroth order the adiabatic field is obtained as
\begin{equation}
a^{(0)} = \frac{\eta}{-\i (\Delta_C-G x) + \kappa}\,.
\end{equation}
The linear response of $a$ to the mirror motion is then
\begin{equation}
a^{(1)} = \frac{1}{\i (\Delta_C-G x) - \kappa} \frac{\partial}{\partial x} a^{(0)} = \frac{\i \eta G}{\big[-\i ( \Delta_C-G x) + \kappa\big]^3}\,.
\end{equation}
The force acting on the mirror derives from the defining equation $\dot{\hat{p}}= \tfrac{\i}{\hbar}[{\cal \hat{H}}, \hat{p}] = -\hbar G \hat{a}^\dagger \hat{a}$ and, up to linear order in velocity, is
\begin{equation}
\textsl{\textsf{F}}_1 = - 2v \hbar G \re{{a^{(0)\ast}} a^{(1)}} = 4 v \frac{\hbar \eta^2 G^2 \kappa \Delta_C}{ \big( \Delta_C^2 + \kappa^2\big)^3}\,,
\end{equation}
where we used $x=0$ without loss of generality. It can be seen that mirror cooling requires that $\Delta_C<0$, i.e., the cavity resonance frequency is above the pump frequency. In this case, for efficient excitation of the field in the resonator, the frequency of the pump photons is up-shifted at the expense of the mirror's kinetic energy. This cooling force has been derived in \sref{sec:Mirrorcooling}, as a limiting case of the more general scattering theory. To check the perfect agreement between the two results, the quantity corresponding to $\eta$ can be deduced from the total field energy in the resonator for an immobile mirror, which is $\hbar \omega_C {\hat{a}^{(0)\dagger}}\hat{a}^{(0)}$ here.
\subappendicesend
\section{General solution to the transfer matrix approach}\label{sec:TMM:General}
The approach used to find $\textsl{\textsf{F}}$ in the previous section illustrates the nature of this quantity, arising from the multiple round-trips of the light between the static and moving mirrors, very well. However, the series summation does not lend itself to easy direct evaluation in a mechanical fashion. By rewriting the Doppler shift operator, however, this limitation can be lifted and $\textsl{\textsf{F}}$ and $\textsl{\textsf{D}}$ evaluated for a fully general optical system. The basis of this current section was published as Xuereb, A., Freegarde, T., Horak, P., \& Domokos, P. Phys.\ Rev.\ Lett.\ \textbf{105}, 013602 (2010), and will be elaborated on in a manuscript that is currently being prepared. In this section, we will first summarise the general transfer matrix approach and show how solutions in closed form can be found to the friction force, in \Sref{sec:TMM:Force}, and momentum diffusion, in \Sref{sec:Diffusion}, acting on the scatterer. \Sref{sec:TMM:Optomechanics} subsequently explores the linear and nonlinear optomechanical coupling that can be achieved between the field of a cavity and the position of a micromirror placed inside it.\par
Given a static scatterer represented by the transfer matrix $\hat{\mat{M}}$, the effect of the motion of the scatterer on the fields it is interacting with is included by applying the transformation [cf. \eref{eq:TMM:LML}]
\begin{equation}
\hat{\mat{M}}\rightarrow\hat{\mat{L}}(-v)\hat{\mat{M}}\hat{\mat{L}}(v)\,,
\end{equation}
where the `Doppler shift' operator matrix $\hat{\mat{L}}(v)$ for a scatterer moving with velocity $v$ can be written, from \eref{eq:TMM:Lmatrix}, as
\begin{equation}
\label{eq:TMMFirstordervExplicit}
\hat{\mat{L}}(v)=\begin{bmatrix}
1+\tfrac{v}{c}\bigl(1-k_0\partial_k\bigr)&0\\
0&1-\tfrac{v}{c}\bigl(1-k_0\partial_k\bigr)\\
\end{bmatrix}\,,
\end{equation}
to linear order in $v/c$ and under the assumption that the pump beam has a very narrow spread of wavenumbers about a central $k_0$. We also use the shorthand notation $\partial_k\equiv\tfrac{\partial}{\partial k}$ throughout and will drop the label $k$ wherever this is not necessary.
\subsection{Force acting on moving scatterer}\label{sec:TMM:Force}
\begin{figure}[tb]
\centering
\includegraphics[scale=0.45]{Diagrams/followup_cavityassistedcooling/Full_Model}
\caption[Schematic model of a scatterer interacting with two general optical systems]{The model we consider in this section, drawn schematically. A scatterer $\text{S}$ interacts with two `general optical systems' in one dimension, composed of immobile linear optical elements, one to either side. $\text{S}$ and these two systems are each represented by a $2\times 2$ matrix.}
\label{fig:Schematic}
\end{figure}
Let us now apply this description to the general model represented by the system in \fref{fig:Schematic}. Our aim is to obtain the force acting on scatterer $\text{S}$; to do this we need to express the fields it interacts with, $A$, $B$, $C$, and $D$, in terms of the input fields $B_\text{l}$ and $C_\text{r}$. As in Ref.~\cite{Xuereb2010b}, we define
\begin{equation}
\hat{\mat{M}}=\mat{M}_1\times\hat{\mat{M}}_\text{S}\times \mat{M}_2\equiv\begin{bmatrix}
\hat{\gamma}&\hat{\alpha}\\
\hat{\delta}&\hat{\beta}
\end{bmatrix}\,\text{and}\,\bigl(\mat{M}_1\bigr)^{-1}\equiv\bigl[\theta_{ij}\bigr]\,.
\end{equation}
We also define the convenient velocity-independent quantities $\alpha_0$, $\alpha_1^{(0)}$, $\alpha_1^{(1)}$, etc., by
\begin{align}
\hat{\alpha}&\equiv\alpha_0+\frac{v}{c}\Bigl(\alpha_1^{(0)}+\alpha_1^{(1)}\partial_k\Bigr)\,,\\
\hat{\beta}&\equiv\beta_0+\frac{v}{c}\Bigl(\beta_1^{(0)}+\beta_1^{(1)}\partial_k\Bigr)\,,\\
\hat{\gamma}&\equiv\gamma_0+\frac{v}{c}\Bigl(\gamma_1^{(0)}+\gamma_1^{(1)}\partial_k\Bigr)\,,\text{ and}\\
\hat{\delta}&\equiv\delta_0+\frac{v}{c}\Bigl(\delta_1^{(0)}+\delta_1^{(1)}\partial_k\Bigr)\,.
\end{align}
A final assumption is needed to be able to express the fields in closed form---we assume that the two input fields are at the same frequency and have only a very small spread in wavenumber:
\begin{equation}
B_\text{l}=B_0\,\delta(k-k_0)\ \text{and}\ C_\text{r}=C_0\,\delta(k-k_0)\,.
\end{equation}
Then, the field amplitudes $\mathcal{A}=\int A(k)\,\mathrm{d} k$ and $\mathcal{B}=\int B(k)\,\mathrm{d} k$ are given, to first order in $v/c$, by:
\begin{multline}
\label{eq:Afield}
\mathcal{A}=\biggl(\theta_{11}\frac{\alpha_0}{\beta_0}+\theta_{12}+\frac{v}{c}\biggl\{\frac{\theta_{11}}{\beta_0^2}\Bigl(\alpha_1^{(0)}\beta_0-\alpha_0\beta_1^{(0)}\Bigr)-\frac{1}{\beta_0}\biggl[\partial_k\frac{\theta_{11}}{\beta_0}\Bigl(\alpha_1^{(1)}\beta_0-\alpha_0\beta_1^{(1)}\Bigr)\biggr]\biggr\}\biggr)B_0\\
+\biggl(\theta_{11}\frac{\gamma_0\beta_0-\alpha_0\delta_0}{\beta_0}+\frac{v}{c}\biggl\{\frac{\theta_{11}}{\beta_0^2}\Bigl[\beta_0^2\gamma_1^{(0)}-\alpha_0\beta_0\delta_1^{(0)}-\Bigl(\alpha_1^{(0)}\beta_0-\alpha_0\beta_1^{(0)}\Bigr)\delta_0\Bigr]\\-\biggl[\partial_k\frac{\theta_{11}}{\beta_0}\Bigl(\beta_0\gamma_1^{(1)}-\alpha_0\delta_1^{(1)}\Bigr)\biggr]+\frac{\delta_0}{\beta_0}\biggl[\partial_k\frac{\theta_{11}}{\beta_0}\Bigl(\alpha_1^{(1)}\beta_0-\alpha_0\beta_1^{(1)}\Bigr)\biggr]\biggr\}\biggr)C_0\,,
\end{multline}
and
\begin{multline}
\label{eq:Bfield}
\mathcal{B}=\biggl(\theta_{21}\frac{\alpha_0}{\beta_0}+\theta_{22}+\frac{v}{c}\biggl\{\frac{\theta_{21}}{\beta_0^2}\Bigl(\alpha_1^{(0)}\beta_0-\alpha_0\beta_1^{(0)}\Bigr)-\frac{1}{\beta_0}\biggl[\partial_k\frac{\theta_{21}}{\beta_0}\Bigl(\alpha_1^{(1)}\beta_0-\alpha_0\beta_1^{(1)}\Bigr)\biggr]\biggr\}\biggr)B_0\\
+\biggl(\theta_{21}\frac{\gamma_0\beta_0-\alpha_0\delta_0}{\beta_0}+\frac{v}{c}\biggl\{\frac{\theta_{21}}{\beta_0^2}\Bigl[\beta_0^2\gamma_1^{(0)}-\alpha_0\beta_0\delta_1^{(0)}-\Bigl(\alpha_1^{(0)}\beta_0-\alpha_0\beta_1^{(0)}\Bigr)\delta_0\Bigr]\\-\biggl[\partial_k\frac{\theta_{21}}{\beta_0}\Bigl(\beta_0\gamma_1^{(1)}-\alpha_0\delta_1^{(1)}\Bigr)\biggr]+\frac{\delta_0}{\beta_0}\biggl[\partial_k\frac{\theta_{21}}{\beta_0}\Bigl(\alpha_1^{(1)}\beta_0-\alpha_0\beta_1^{(1)}\Bigr)\biggr]\biggr\}\biggr)C_0\,,
\end{multline}
where the derivatives are all evaluated at $k=k_0$.
We shall find it useful to express these results in the form $\mathcal{A}=\mathcal{A}_0+\tfrac{v}{c}\mathcal{A}_1$ and $\mathcal{B}=\mathcal{B}_0+\tfrac{v}{c}\mathcal{B}_1$, with $\mathcal{A}_{0,1}$ and $\mathcal{B}_{0,1}$ being independent of $v$. For conciseness, let us now assume that $\zeta$ does not depend on $k$. Then, using the elements of $\hat{\mat{M}}_\text{S}$, we obtain
\begin{align}
\mathcal{C}&=(1-\i\zeta)\mathcal{A}-\bigl(1-2\tfrac{v}{c}\bigl)\i\zeta\mathcal{B}\nonumber\\&=\bigl[(1-\i\zeta)\mathcal{A}_0-\i\zeta\mathcal{B}_0\bigr]+\tfrac{v}{c}\bigl[(1-\i\zeta)\mathcal{A}_1+2\i\zeta\mathcal{B}_0-\i\zeta\mathcal{B}_1\bigr]\,,
\end{align}
and
\begin{align}
\mathcal{D}&=\bigl(1+2\tfrac{v}{c}\bigl)\i\zeta\mathcal{A}+(1+\i\zeta)\mathcal{B}\nonumber\\&=\bigl[\i\zeta\mathcal{A}_0+(1+\i\zeta)\mathcal{B}_0\bigr]+\tfrac{v}{c}\bigl[2\i\zeta\mathcal{A}_0-\i\zeta\mathcal{A}_1-(1+\i\zeta)\mathcal{B}_1\bigr]\,.
\end{align}
We denote the velocity-independent parts of $\mathcal{C}$ and $\mathcal{D}$ by $\mathcal{C}_0$ and $\mathcal{D}_0$, respectively. The \emph{friction} force acting on the scatterer can be finally written down as
\begin{align}
\label{eq:FullFriction}
\textsl{\textsf{F}}=-4\hbar k_0\frac{v}{c} \Bigl[&\lvert\zeta\rvert^2\bigl(\lvert\mathcal{A}_0\rvert^2-\lvert\mathcal{B}_0\rvert^2\bigr)+\bigl(\lvert\zeta\rvert^2+\im{\zeta}\bigr)\re{\mathcal{A}_0\mathcal{A}_1^\ast}-2\im{\zeta}\re{\mathcal{A}_0\mathcal{B}_0^\ast}\nonumber\\&+\bigl(\lvert\zeta\rvert^2-\im{\zeta}\bigr)\re{\mathcal{B}_0\mathcal{B}_1^\ast}+\im{\zeta}\re{\mathcal{A}_0\mathcal{B}_1^\ast}\nonumber\\&+\re{\Bigl(\lvert\zeta\rvert^2+\i\re{\zeta}\Bigr)\mathcal{A}_1\mathcal{B}_0^\ast}\Bigr]\,.
\end{align}
\subsection{Momentum diffusion experienced by scatterer}\label{sec:Diffusion}
Photon number fluctuations in the input fields $B_\text{l}$ and $C_\text{r}$ give rise to a stochastic force, added to $\textsl{\textsf{F}}$, that averages out to zero; this stochastic force is, however, responsible for preventing the momentum of the scatter from reaching zero under the action of a friction force~\cite{Metcalf2003}. A complete theory based on the input-output operator formalism can be built along the lines of the previous section. The results we have obtained so far correspond to assuming that the input fields are in a coherent state, see \sref{sec:MSTDiffusion}, whereby the fields $A$, $B$, etc., are the expectation values of analogous bosonic annihilation operators: $\langle\hat{A}\rangle=A$, $\langle\hat{A^\dagger}\rangle=A^\ast$, and so on. The two input modes are the only independent modes in our system, and their corresponding operators obey the usual bosonic commutation relations:
\begin{equation}
\comm{\hat{B}_\text{l}}{\hat{B}_\text{l}}=\comm{\hat{C}_\text{r}}{\hat{C}_\text{r}}=\frac{\hbar k_0}{2\epsilon_0 \sigma_\text{L}}\delta(t-t^\prime)\,,\text{ and }\comm{\hat{B}_\text{l}}{\hat{C}_\text{r}}=0\,,
\end{equation}
with $\omega=kc$ being the (angular) frequency. Since we are working to first order in $v/c$, we need to evaluate everything in this section to zeroth order in $v/c$.\par
Besides photon number fluctuations, lossy beamsplitters also induce noise into the system; this is equivalent to having a noise mode $\hat{E}$ that modifies the output fields interacting with that component. The noise input in the system by any of its components is independent of the noise input by any other component; in other words, if $\hat{E}_1$ and $\hat{E}_2$ are two noise modes that interact with the system,
\begin{equation}
\comm{\hat{E}_1}{\hat{E}_1}=\comm{\hat{E}_2}{\hat{E}_2}=\frac{\hbar k_0}{2\epsilon_0 \sigma_\text{L}}\delta(t-t^\prime)\,,\comm{\hat{E}_1}{\hat{E}_2}=0\,,
\end{equation}
and
\begin{equation}
\langle\hat{E}_1\rangle=\langle\hat{E}_2\rangle=0\,,
\end{equation}
with the last pair of equalities being true by construction. We can therefore treat each of the noise modes independently. Moreover, the noise modes are independent of the input modes; for any such noise mode $\hat{E}$:
\begin{equation}
\comm{\hat{E}}{\hat{B}_\text{l}}=\comm{\hat{E}}{\hat{C}_\text{r}}=0\,.
\end{equation}
In this section, we will generalise the treatment in \sref{sec:TMM:Model} to the situation where the loss-inducing beamsplitter is part of a generic optical system. Let us now assume, for concreteness, that the moving scatterer $\text{S}$ from \Sref{sec:TMM:Force} introduces the only noise mode, $\hat{E}$, in the system. In the case of an isolated scatterer, $\hat{B}$ and $\hat{C}$ represent independent input modes and the above equations guarantee that the two-time self-commutators of $\hat{A}$ and $\hat{D}$ behave as expected; see \aref{sec:TMM:AppDiffusion}. Let us now solve the general problem. One can rewrite \erefs{eq:ADNoise} using a matrix similar to the transfer matrix, obtaining:
\begin{align}
\begin{pmatrix}
\hat{A}\\
\hat{B}
\end{pmatrix}&=\frac{1}{\mathfrak{t}}\begin{bmatrix}\begin{array}{cc}
\bigl(\mathfrak{t}^2-\mathfrak{r}^2\bigr)&\mathfrak{r}\\
-\mathfrak{r}&1
\end{array}\end{bmatrix}
\begin{pmatrix}
\hat{C}\\
\hat{D}
\end{pmatrix}+\frac{\sqrt{\epsilon}}{\mathfrak{t}}\begin{pmatrix}
\hat{E}\\
\hat{E}
\end{pmatrix}\nonumber\\
&=\frac{1}{\mathfrak{t}}\begin{bmatrix}\begin{array}{cc|c}
\bigl(\mathfrak{t}^2-\mathfrak{r}^2\bigr)&\mathfrak{r}&\phantom{+}\sqrt{\epsilon}\\
-\mathfrak{r}&1&-\sqrt{\epsilon}
\end{array}\end{bmatrix}
\begin{pmatrix}
\hat{C}\\
\hat{D}\\
\hat{E}
\end{pmatrix}\nonumber\\
&=\begin{bmatrix}\begin{array}{c|c}\,\hat{\mat{M}}_\text{S}\,&\begin{matrix}
\phantom{+}\sqrt{\epsilon}/\mathfrak{t}\\
-\sqrt{\epsilon}/\mathfrak{t}
\end{matrix}\end{array}\end{bmatrix}
\begin{pmatrix}
\hat{C}\\
\hat{D}\\
\hat{E}
\end{pmatrix}\,,
\end{align}
where $\hat{\mat{M}}_\text{S}$ is included as a $2\times 2$ submatrix in the new $2\times 3$ matrix, which can in turn be embedded in a $3\times 3$ square matrix:\footnote{Note that \eref{eq:ABECDEMatrix} does \emph{not} imply violation of the principle of conservation of energy; indeed, the matrix is not a transfer matrix and the $\hat{E}$ mode is the same mode on either side of the equality.}
\begin{equation}
\label{eq:ABECDEMatrix}
\begin{pmatrix}
\hat{A}\\
\hat{B}\\
\hat{E}
\end{pmatrix}=\begin{bmatrix}\begin{array}{c|c}\hat{\mat{M}}_\text{S}&\begin{matrix}
\phantom{+}\sqrt{\epsilon}/\mathfrak{t}\\
-\sqrt{\epsilon}/\mathfrak{t}
\end{matrix}\\\hline
\begin{matrix}\,0&&0\,\end{matrix}&1
\end{array}\end{bmatrix}
\begin{pmatrix}
\hat{C}\\
\hat{D}\\
\hat{E}
\end{pmatrix}\,.
\end{equation}
This assumes that the noise mode is essentially unaffected by the presence of the scatterer. In summary, then, each transfer matrix in the optical system is replaced by one of two matrices:
\begin{equation}
\hat{\mat{M}}\rightarrow
\begin{bmatrix}\begin{array}{c|c}\hat{\mat{M}}&\begin{matrix}
0\\
0
\end{matrix}\\\hline
\begin{matrix}\,0&&0\,\end{matrix}&\,\,1\,\,
\end{array}\end{bmatrix}\,,
\end{equation}
if the scatterer introduces no noise into the system, or
\begin{equation}
\hat{\mat{M}}\rightarrow
\begin{bmatrix}\begin{array}{c|c}\hat{\mat{M}}&\begin{matrix}
\phantom{+}\sqrt{\epsilon}/\mathfrak{t}\\
-\sqrt{\epsilon}/\mathfrak{t}
\end{matrix}\\\hline
\begin{matrix}\,0&&0\,\end{matrix}&1
\end{array}\end{bmatrix}\,,
\end{equation}
if the scatterer introduces the noise mode; as before, the properties of the scatterer are fully specified by $\mathfrak{r}$ and $\mathfrak{t}$. The third row and column of \emph{every} such matrix refer to the same, singular, noise mode being treated. Applying the above transformation to the matrices representing the generic system in \fref{fig:Schematic}, one obtains the field operators to zeroth order in $v/c$:
\begin{equation}
\label{eq:AfieldNoise}
\hat{A}=\biggl(\theta_{11}\frac{\alpha_0}{\beta_0}+\theta_{12}\biggr)\hat{B}_\text{l}+\theta_{11}\frac{\gamma_0\beta_0-\alpha_0\delta_0}{\beta_0}\hat{C}_\text{r}+\theta_{11}\frac{\sqrt{\epsilon}}{\mathfrak{t}}\biggl[\bigl(m_{11}-m_{12}\bigr)+\bigl(m_{22}-m_{21}\bigr)\frac{\alpha_0}{\beta_0}\biggr]\hat{E}\,,
\end{equation}
and
\begin{equation}
\label{eq:BfieldNoise}
\hat{B}=\biggl(\theta_{21}\frac{\alpha_0}{\beta_0}+\theta_{22}\biggr)\hat{B}_\text{l}+\theta_{21}\frac{\gamma_0\beta_0-\alpha_0\delta_0}{\beta_0}\hat{C}_\text{r}+\theta_{21}\frac{\sqrt{\epsilon}}{\mathfrak{t}}\biggl[\bigl(m_{11}-m_{12}\bigr)+\bigl(m_{22}-m_{21}\bigr)\frac{\alpha_0}{\beta_0}\biggr]\hat{E}\,,
\end{equation}
where we have used the shorthand notation $\bigl[m_{ij}\bigr]\equiv \mat{M}_1$. Here, $\hat{E}$ is the noise mode introduced by the scatterer itself. Other noise modes would give rise to contributions that are computed analogously. Now, denote $\hat{A}=b_\text{a}\hat{B}_\text{l}+c_\text{a}\hat{C}_\text{r}+e_\text{a}\hat{E}$ and $\hat{B}=b_\text{b}\hat{B}_\text{l}+c_\text{b}\hat{C}_\text{r}+e_\text{b}\hat{E}$ to this order in $v/c$; $b_\text{a,b}$, $c_\text{a,b}$, and $e_\text{a,b}$ can be read off \erefs{eq:AfieldNoise} and~(\ref{eq:BfieldNoise}). Then
\begin{align}
\comm{\hat{A}}{\hat{A}}&=\frac{\hbar k_0}{2\epsilon_0 \sigma_\text{L}}\delta(t-t^\prime)\bigl(\lvert b_\text{a}\rvert^2+\lvert c_\text{a}\rvert^2+\lvert e_\text{a}\rvert^2\bigr)\,,\\
\comm{\hat{B}}{\hat{B}}&=\frac{\hbar k_0}{2\epsilon_0 \sigma_\text{L}}\delta(t-t^\prime)\bigl(\lvert b_\text{b}\rvert^2+\lvert c_\text{b}\rvert^2+\lvert e_\text{b}\rvert^2\bigr)\,,\text{ and}\\
\comm{\hat{A}}{\hat{B}}&=\frac{\hbar k_0}{2\epsilon_0 \sigma_\text{L}}\delta(t-t^\prime)\bigl(b_\text{a}b^\ast_\text{b}+c_\text{a}c^\ast_\text{b}+e_\text{a}e^\ast_\text{b}\bigr)\,.
\end{align}
We can also express $\hat{C}=b_\text{c}\hat{B}_\text{l}+c_\text{c}\hat{C}_\text{r}+e_\text{c}\hat{E}$ and $\hat{D}=b_\text{d}\hat{B}_\text{l}+c_\text{d}\hat{C}_\text{r}+e_\text{d}\hat{E}$, where
\begin{align}
&b_\text{c}=b_\text{a}/\mathfrak{t}-b_\text{b}\mathfrak{r}/\mathfrak{t}\,,\
c_\text{c}=c_\text{a}/\mathfrak{t}-c_\text{b}\mathfrak{r}/\mathfrak{t}\,,\text{ and}\
e_\text{c}=e_\text{a}/\mathfrak{t}-e_\text{b}\mathfrak{r}/\mathfrak{t}-\sqrt{\epsilon}/\mathfrak{t}\,;\\
&b_\text{d}=b_\text{a}\mathfrak{r}/\mathfrak{t}+b_\text{b}\bigl(\mathfrak{t}^2-\mathfrak{r}^2\bigr)/\mathfrak{t}\,,\
c_\text{d}=c_\text{a}\mathfrak{r}/\mathfrak{t}+c_\text{b}\bigl(\mathfrak{t}^2-\mathfrak{r}^2\bigr)/\mathfrak{t}\,,\text{ and}\nonumber\\
&e_\text{d}=e_\text{a}\mathfrak{r}/\mathfrak{t}+e_\text{b}\bigl(\mathfrak{t}^2-\mathfrak{r}^2\bigr)\mathfrak{r}/\mathfrak{t}+\sqrt{\epsilon}/\mathfrak{t}\,.
\end{align}
Therefore, e.g.,
\begin{align}
\comm{\hat{C}}{\hat{C}}&=\frac{\hbar k_0}{2\epsilon_0 \sigma_\text{L}}\delta(t-t^\prime)\bigl(\lvert b_\text{c}\rvert^2+\lvert c_\text{c}\rvert^2+\lvert e_\text{c}\rvert^2\bigr)\,,\text{ and }\\
\comm{\hat{D}}{\hat{D}}&=\frac{\hbar k_0}{2\epsilon_0 \sigma_\text{L}}\delta(t-t^\prime)\bigl(\lvert b_\text{d}\rvert^2+\lvert c_\text{d}\rvert^2+\lvert e_\text{d}\rvert^2\bigr)\,.
\end{align}
Several other nontrivial commutators need to be computed, due to the presence of $\hat{E}$; for example,
\begin{equation}
\comm{\hat{A}}{\hat{D}}=\frac{\hbar k_0}{2\epsilon_0 \sigma_\text{L}}\delta(t-t^\prime)\bigl(b_\text{a}b^\ast_\text{d}+c_\text{a}c^\ast_\text{d}+e_\text{a}e^\ast_\text{d}\bigr)\,.
\end{equation}
For an isolated scatterer in free space, $b_\text{b}=c_\text{c}=1$, $b_\text{c}=c_\text{b}=0$, $e_\text{b}=e_\text{c}=0$, etc., and one can show that the various commutators behave as expected. Finally, the momentum diffusion constant is given by
\begin{align}
\textsl{\textsf{D}}\,\delta(t-t^\prime)=2\epsilon_0\sigma_\text{L}\,\hbar k_0\Bigl(&\lvert\mathcal{A}_0\rvert^2\comm{\hat{A}}{\hat{A}}+\lvert\mathcal{B}_0\rvert^2\comm{\hat{B}}{\hat{B}}+\lvert\mathcal{C}_0\rvert^2\comm{\hat{C}}{\hat{C}}+\lvert\mathcal{D}_0\rvert^2\comm{\hat{D}}{\hat{D}}\nonumber\\
&+2\lre{\mathcal{A}_0^\ast\mathcal{B}_0\comm{\hat{A}}{\hat{B}}-\mathcal{A}_0^\ast\mathcal{C}_0\comm{\hat{A}}{\hat{C}}}\nonumber\\
&\phantom{2Re\qquad}\left.-\mathcal{A}_0^\ast\mathcal{D}_0\comm{\hat{A}}{\hat{D}}-\mathcal{B}_0^\ast\mathcal{C}_0\comm{\hat{B}}{\hat{C}}\right.\nonumber\\
&\phantom{2Re\qquad}\left.-\mathcal{B}_0^\ast\mathcal{D}_0\comm{\hat{B}}{\hat{D}}+\mathcal{C}_0^\ast\mathcal{D}_0\comm{\hat{C}}{\hat{D}}\right\}\Bigr)\,,
\end{align}
where, for example, $\mathcal{A}_0^\ast \mathcal{D}_0=\int\langle\hat{A}^\dagger\rangle\mathrm{d} k\int\langle\hat{D}\rangle\mathrm{d} k=b^\ast_\text{a}b_\text{d}\lvert B_0\rvert^2+c^\ast_\text{a}c_\text{d}\lvert C_0\rvert^2+b^\ast_\text{a}c_\text{d}B^\ast_0C_0+c^\ast_\text{a}b_\text{d}C^\ast_0B_0$, recalling that $\langle\hat{E}\rangle=0$, and analogously for the rest of the terms. This result again reduces to the expected one for an isolated scatterer in free space; it can alternatively be extended to include further noise modes.
\section{Optomechanics of a micromirror inside a cavity}\label{sec:TMM:Optomechanics}
\begin{figure}
\centering
\subfigure[\ $\zeta=-0.100$]{
\includegraphics[width=0.4\textwidth]{Diagrams/followup_cavityassistedcooling/Mirror_CMC_Scan_Friction__Zeta_-0_100}
}
\subfigure[\ $\zeta=-0.300$]{
\includegraphics[width=0.4\textwidth]{Diagrams/followup_cavityassistedcooling/Mirror_CMC_Scan_Friction__Zeta_-0_300}
}\\
\subfigure[\ $\zeta=-0.500$]{
\includegraphics[width=0.4\textwidth]{Diagrams/followup_cavityassistedcooling/Mirror_CMC_Scan_Friction__Zeta_-0_500}
}
\subfigure[\ $\zeta=-0.700$]{
\includegraphics[width=0.4\textwidth]{Diagrams/followup_cavityassistedcooling/Mirror_CMC_Scan_Friction__Zeta_-0_700}
}\\
\subfigure[\ $\zeta=-1.000$]{
\includegraphics[width=0.4\textwidth]{Diagrams/followup_cavityassistedcooling/Mirror_CMC_Scan_Friction__Zeta_-1_000}
}
\subfigure[\ $\zeta=-2.000$]{
\includegraphics[width=0.4\textwidth]{Diagrams/followup_cavityassistedcooling/Mirror_CMC_Scan_Friction__Zeta_-2_000}
}\\
\subfigure[\ $\zeta=-5.000$]{
\includegraphics[width=0.4\textwidth]{Diagrams/followup_cavityassistedcooling/Mirror_CMC_Scan_Friction__Zeta_-5_000}
}
\subfigure[\ $\zeta=-10.000$]{
\includegraphics[width=0.4\textwidth]{Diagrams/followup_cavityassistedcooling/Mirror_CMC_Scan_Friction__Zeta_-10_000}
}\\
\subfigure{
\includegraphics[scale=0.3]{Diagrams/followup_cavityassistedcooling/Colorbar_intensity}
}\hspace{1cm}
\subfigure{
\includegraphics[scale=0.3]{Diagrams/followup_cavityassistedcooling/Colorbar_friction}
}
\caption[Field intensity and cooling coefficient acting on a micromirror inside a cavity]{Field intensity (left panels) at and cooling coefficient (right panels) acting on the micromirror as the micromirror position ($x$) and cavity length ($L_\text{C}+\Delta L_\text{C}$) are scanned. The subfigures differ only in the polarisability of the mirror, as indicated. The cavity parameters are modelled from Ref.~\cite{Thompson2008}. In the series of left panels, we note the progression from an almost bare cavity situation (a) to a very strong perturbation by the micromirror, leading to avoided crossings (h). The white dashed line traces a cavity node, whereas the black dashed lines [\eref{eq:CavityResonances}] trace the cavity resonances. In the series of right panels, note that the cooling coefficient is---as expected---a cooling force (blue) for red cavity detuning and a heating force (red) for blue detuning. The colourbars are on a logarithmic scale and are for $1$\,W of input power; the large dynamic range is needed to bring out the detail in the panels, due to the very narrow features present.}
\label{fig:IaF}
\end{figure}
\begin{figure}[t]
\centering
\subfigure[]{
\includegraphics[width=1.5\figwidth]{Diagrams/followup_cavityassistedcooling/Nonlinear_1}
}\hspace{1cm}
\subfigure[]{
\includegraphics[width=1.5\figwidth]{Diagrams/followup_cavityassistedcooling/Nonlinear_10}
}
\caption[Linear and quadratic optomechanical couplings as a function of mirror position]{Linear and quadratic optomechanical couplings as a function of mirror position for a very good cavity and for (a)~$\zeta=-1$, and (b)~$\zeta=-10$. In each figure we show the linear (solid curve) and quadratic (dashed curve) couplings, from \erefs{eq:LinearCoupling}~and~(\ref{eq:QuadraticCoupling}). Note that the peak value of $\omega^{\prime\prime}$ is roughly proportional to $\zeta$ whereas $\omega^\prime$ is bounded.}
\label{fig:Couplings}
\end{figure}
\begin{figure}
\centering
\subfigure[\ $\zeta=-0.100$]{
\includegraphics[width=0.4\textwidth]{Diagrams/followup_cavityassistedcooling/Mirror_CMC_Scan_Static_Force__Zeta_-0_100}
}
\subfigure[\ $\zeta=-0.300$]{
\includegraphics[width=0.4\textwidth]{Diagrams/followup_cavityassistedcooling/Mirror_CMC_Scan_Static_Force__Zeta_-0_300}
}\\
\subfigure[\ $\zeta=-0.500$]{
\includegraphics[width=0.4\textwidth]{Diagrams/followup_cavityassistedcooling/Mirror_CMC_Scan_Static_Force__Zeta_-0_500}
}
\subfigure[\ $\zeta=-0.700$]{
\includegraphics[width=0.4\textwidth]{Diagrams/followup_cavityassistedcooling/Mirror_CMC_Scan_Static_Force__Zeta_-0_700}
}\\
\subfigure[\ $\zeta=-1.000$]{
\includegraphics[width=0.4\textwidth]{Diagrams/followup_cavityassistedcooling/Mirror_CMC_Scan_Static_Force__Zeta_-1_000}
}
\subfigure[\ $\zeta=-2.000$]{
\includegraphics[width=0.4\textwidth]{Diagrams/followup_cavityassistedcooling/Mirror_CMC_Scan_Static_Force__Zeta_-2_000}
}\\
\subfigure[\ $\zeta=-5.000$]{
\includegraphics[width=0.4\textwidth]{Diagrams/followup_cavityassistedcooling/Mirror_CMC_Scan_Static_Force__Zeta_-5_000}
}
\subfigure[\ $\zeta=-10.000$]{
\includegraphics[width=0.4\textwidth]{Diagrams/followup_cavityassistedcooling/Mirror_CMC_Scan_Static_Force__Zeta_-10_000}
}\\
\subfigure{
\includegraphics[scale=0.3]{Diagrams/followup_cavityassistedcooling/Colorbar_static}
}
\caption[Static friction force acting on a micromirror inside a cavity]{Static friction force (i.e., the force acting on the mirror when $v=0$) computed from the scattering model presented here (left panels) and a model based on a modal decomposition~\cite{Jayich2008} (right panels), showing only one pair of modes. Red and green regions represent forces pointing in opposite directions, as indicated on the colourbar. We note qualitative agreement between the two models for $x\approx 0.25\lambda$ and for $\Delta L_\text{C}$ close to the resonances, especially for large $\lvert\zeta\rvert$. The discrepancies between the two sets of data, that are more pronounced for small polarisability, have significant consequences for any theory based on a coupled-cavity modal decomposition model. The black dashed lines [\eref{eq:CavityResonances}] trace the cavity resonances in the scattering model. The absolute values on the colourbar relate to the left panels.}
\label{fig:SF}
\end{figure}
Placing a micromirror inside a cavity is actively being explored, both theoretically~\cite{Bhattacharya2007a,Bhattacharya2008} and experimentally~\cite{Thompson2008,Jayich2008,Sankey2010}, as a means of realising strong optomechanical coupling between a movable scatterer and the cavity field. In a recent experiment, Sankey and co-workers~\cite{Sankey2010} looked at both linear and nonlinear coupling between the mechanical degree of freedom of the micromirror and the cavity field. The physical mechanism by which optomechanical coupling, and subsequently cooling, is enhanced by the presence of the cavity can be looked at either using the standard theory of cavity-mediated cooling~\cite{Horak1997,Hechenblaikner1998}, in the case of weakly reflective micromirrors, or by considering the micromirror to form the boundary between two coupled cavities~\cite{Bhattacharya2007a,Bhattacharya2008}, in the case of more strongly reflective micromirrors. The transfer matrix approach detailed in the previous sections provides a natural way to explore this interaction no matter what the optical properties of the micromirror are.
\par
We begin by modelling the system in Ref.~\cite{Thompson2008}: a two--mirror Fabry-P\'erot cavity with a micromirror near its centre, operating at a wavelength $\lambda=1064$\,nm and having a length $L_\text{C}=6.7$\,cm. The micromirror is modelled by its polarisability $\zeta$ which, in light of the small losses observed in practice, is taken to be real and negative. Whereas the real experimental system corresponds to $\zeta\gtrsim-1$, we allow $\zeta$ to vary freely in our model. The two quantities of interest in this section are the intensity of the field close to the micromirror, and the cooling coefficient acting on the micromirror. The former gives us knowledge of the resonant frequencies of the cavity and, therefore, of the optomechanical coupling, to all orders, between the cavity field and the micromirror. The latter is useful in optomechanical cooling experiments; the interest here lies in the fact that cooling the motion of a micromirror is one way towards achieving higher sensitivity in sensing applications, most notably in gravitational-wave detectors~\cite{Braginsky2002}.\\
These quantities are summarised in \fref{fig:IaF}, with the left panels showing the intensity at the mirror and the right panels the cooling coefficient acting on the mirror. Each subfigure (a)--(h) explores a different value for $\zeta$. For $\zeta\approx 0$, the cavity field is similar to the bare-cavity field; in particular, the cavity resonances are only slightly perturbed by the presence and position of the micromirror. The opposite is true of the $\zeta\ll-1$ case, where there is strong coupling between pairs of cavity modes, typified by the avoided crossings in the spectra. The resonance frequencies can be obtained analytically, in the limit of a good bare cavity, as frequency shifts from the bare resonances:
\begin{equation}
\Delta\omega=\frac{c}{L_\text{C}}\tan^{-1}\Biggl\{\frac{\zeta\cos(2k_0x)\mp\zeta\bigl[1+\zeta^2\sin^2(2k_0x)\bigr]}{\zeta^2\cos(2k_0x)\pm\bigl[1+\zeta^2\sin^2(2k_0x)\bigr]}\Biggr\}\,,
\label{eq:CavityResonances}
\end{equation}
with $L_\text{C}$ being the length of the cavity, $x$ the position of the micromirror, and $k_0=2\pi/\lambda$ the wavenumber of the light inside the cavity. The two sets of solutions to \eref{eq:CavityResonances} are, in the $\zeta\rightarrow 0$ limit or at $x=\lambda/8$, separated by a free spectral range. These cavity resonances, plotted as detuned cavity lengths $\Delta L_\text{C}=\bigl(L_\text{C}/\omega\bigr)\Delta\omega$, are traced by means of the dashed black curves in the left panels of \fref{fig:IaF}.\\
In the standard optomechanical coupling Hamiltonian, the mirror--field coupling is represented by a term of the form
\begin{equation}
\hat{H}_\text{OM}^{(1)}\sim\hbar\omega^\prime\hat{x}\hat{a}^\dagger\hat{a}\,,
\end{equation}
where $\hat{x}$ is the position operator of the mirror, and $\omega^\prime\equiv\partial(\Delta\omega)/\partial x$. $\hat{a}$ is the annihilation operator of the field mode that has the dominant interaction with the micromirror; in the $\lvert\zeta\rvert\rightarrow 0$ limit, these field modes are the bare cavity modes of the whole cavity. However, as $\lvert\zeta\rvert$ increases, the micromirror effectively splits the main cavity into two coupled cavities, giving rise to symmetric and antisymmetric modes, seen as the higher (bright) and lower (dark) branches in \fref{fig:IaF}(h) for $0<x<\lambda/4$; in such cases $a$ is the annihilation operator belonging to one of these eigenmodes. We note that similar behaviour was observed in Ref.~\cite{Jayich2008}.\\
Certain effects, such as mechanical squeezing of the mirror position~\cite{Nunnenkamp2010} and quantum non-demolition measurements on the mirror~\cite{Clerk2010b}, require not \emph{linear coupling} to $\hat{x}$ but \emph{quadratic coupling} to $\hat{x}^2$:
\begin{equation}
\hat{H}_\text{OM}^{(2)}\sim\hbar\omega^{\prime\prime}\hat{x}^2\hat{a}^\dagger\hat{a}\,,
\end{equation}
with $\omega^{\prime\prime}\equiv\partial^2(\Delta\omega)/\partial x^2$. In our notation, we have
\begin{equation}
\label{eq:LinearCoupling}
\omega^\prime=\mp\frac{2k_0c}{L_\text{C}}\frac{\zeta\sin(2k_0x)}{\big[1+\zeta^2\sin^2(2k_0x)\bigr]^{1/2}}\,,
\end{equation}
and
\begin{equation}
\label{eq:QuadraticCoupling}
\omega^{\prime\prime}=\mp\frac{4k_0^2c}{L_\text{C}}\frac{\zeta\cos(2k_0x)}{\big[1+\zeta^2\sin^2(2k_0x)\bigr]^{3/2}}\,.
\end{equation}
One thing we note immediately is that there is no value for $x$ such that $\omega^\prime=\omega^{\prime\prime}=0$; in other words, the optomechanical coupling is restricted to be linear or quadratic, to lowest order. Higher-order nonlinearities may be achieved by coupling different transverse modes of the cavity (see, e.g., the experimental results in Ref.~\cite{Sankey2010}) but are overwhelmed by the linear or quadratic couplings in a single-mode cavity. Moreover, the linear coupling $\omega^\prime$ is bounded in the $\zeta\rightarrow\infty$ limit:
\begin{equation}
\lvert\omega^\prime\rvert\leq\frac{2k_0c}{L_\text{C}}\approx 2\pi\times8.42\,\text{MHz/nm}\,,
\end{equation}
with the numeric value corresponding to our parameters. In the same limit, $\omega^{\prime\prime}$ exhibits resonant behaviour (see \fref{fig:Couplings}), indicative of avoided crossings in the spectrum, peaking at a value of:
\begin{equation}
\lvert\omega^{\prime\prime}\rvert\rightarrow\frac{4k_0^2c}{L_\text{C}}\lvert\zeta\rvert\approx 2\pi\times 0.10\,\lvert\zeta\rvert\,\text{MHz/(nm)$^2$}\,.
\end{equation}
We plot the lower ($\mp\rightarrow-$) branches of \erefs{eq:LinearCoupling} and~(\ref{eq:QuadraticCoupling}) in \fref{fig:Couplings} for two values for $\zeta$: $\zeta=-1$, representative of realistic micromirrors, and $\zeta=-10$, representative of highly reflective micromirrors. These correspond to cases (e) and (h) in \fref{fig:IaF}, respectively. Coupling between the pairs of modes is not very strong for the $\zeta=-1$ case; this is manifested by means of the smooth variation with $x$ of $\omega^\prime$ and $\omega^{\prime\prime}$ in \fref{fig:Couplings}(a). The second case shows strong signs of the avoided crossing behaviour seen in \fref{fig:IaF}(h), with $\omega^\prime$ no longer behaving smoothly and $\omega^{\prime\prime}$ acquiring a resonance-like character. Note that, independently of the magnitude of $\zeta$, the strongest quadratic coupling always occurs at the points where $\omega^\prime=0$.\\
The linear frequency shifts, \eref{eq:LinearCoupling} and the solid curves in \fref{fig:Couplings}, describe the same quantity as do the curves in Fig.~5 of Ref.~\cite{Bhattacharya2008}. In other words, our results match those in Ref.~\cite{Bhattacharya2008} in the relevant limit of a high-reflectivity micromirror, but the self-consistent scattering solution presented here is also applicable to the general situation where $\zeta$ can take any complex value. Thus, our `micromirror' could represent an actual micromirror, a poorly reflective membrane, or even an atom, whereas in the latter two cases the decomposition of the system into two coupled cavities is not valid. \fref{fig:SF} presents a comparison of the static force acting on the micromirror, for various values of $\zeta$, as calculated both from the model presented in this work (left panels) and from a modal decomposition model (Ref.~\cite{Jayich2008}, right panels). We note that the two sets of data agree in the large $\lvert\zeta\rvert$ limit, where the `two coupled cavities' model is formally valid, but there are significant qualitative differences when $\lvert\zeta\rvert\ll 1$, where the model used most often is that of a single cavity mode interacting with a scatterer~\cite{Hechenblaikner1998}.
\par
The behaviour of the cooling coefficient acting on the micromirror, as shown on the right panels of \fref{fig:IaF}, also has a number of interesting properties. Its overall trend follows the structure of the field intensity closely, as can be seen from an inspection of this figure. As is expected from earlier investigations~\cite{Hechenblaikner1998,Jayich2008} optomechanical cooling inside cavities proceeds when the light pumped into the cavity is tuned below resonance. This is shown quite clearly in \fref{fig:IaF}, in that the friction force generally acts to cool the micromirror when $\Delta L_\text{C}$ is smaller than the resonant length, and it acts to heat when $\Delta L_\text{C}$ is larger. One notes that for weak mirror--field coupling [\fref{fig:IaF}(a)] the behaviour of the cooling coefficient at a given $\Delta L_\text{C}$ close to resonance is approximately well-defined: the motion of the mirror is either cooled or heated, almost irrespectively of the value of $x$. The situation is qualitatively different for very strong mirror--field coupling [\fref{fig:IaF}(h)], where the cooling coefficient changes sign very rapidly as $x$ is scanned over the resonance at a fixed $\Delta L_\text{C}$. This has profound implications for experimental explorations, since the localisation of the micromirror within the cavity becomes of critical importance to the qualitative behaviour of the system.
\section{Optical pumping and multilevel atoms}\label{sec:TMM:Multilevel}
The work in this section is published as Xuereb, A., Domokos, P., Horak, P., \& Freegarde, T. Phys.\ Scr.\ \textbf{T140}, 014010 (2010) and is reproduced essentially \emph{verbatim}~\cite{Xuereb2010a}. The transfer matrix method described previously is extended here to handle multilevel atoms and arbitrarily polarised incident fields.
\par
The extended model we describe in the present section allows us to treat multi-level atoms as classical scatterers in light fields modified by, in principle, arbitrarily complex optical components such as mirrors, resonators, dispersive or dichroic elements, or filters. After we introduce the general extension in the next section, we verify our formalism for two prototypical sub-Doppler cooling mechanisms---the $J=\tfrac{1}{2}\rightarrow J^\prime=\tfrac{3}{2}$ transition, leading to the Sisyphus cooling mechanism, and the $J=1\rightarrow J^\prime=2$ transition---in \Sref{sec:LinPerpLin} and \Sref{sec:SigmaSigma}, respectively, and show that it agrees with the standard literature.
\subsection{A transfer matrix relating Jones vectors}\label{sec:Extension}
\begin{figure}[t]
\centering
\includegraphics[width=0.4\figwidth]{./Diagrams/DiagramsXuereb2010a/system}
\caption[Moving scatterer interacting with four field modes]{Moving scatterer interacting with four field modes represented by the Jones vectors $\mathbf{A}$, $\mathbf{B}$, $\mathbf{C}$, and $\mathbf{D}$. The scatterer has velocity $v$ and is described by means of its polarisability tensor $\boldsymbol{\zeta}$. The field mode amplitudes are, in general, functions of the wavenumber $k$.}
\label{fig:Xuereb2010a:System}
\end{figure}
We investigate the interaction of atoms with light of different polarisations. To this end, we denote the two polarisation basis vectors by $\mu$ and $\nu$, whereby the standard circular polarisation basis is equivalent to setting $\mu=\sigma^+$ and $\nu=\sigma^-$. Starting from the transfer matrix model explored in \sref{sec:TMM:Model} and using the definitions in \fref{fig:Xuereb2010a:System}, we replace each of the field modes by a corresponding Jones vector, similar to the model used in Ref.~\cite{Spreeuw1992}. Thus, for example,
\begin{equation}
A(k)\rightarrow\mathbf{A}(k)=\begin{pmatrix}
A_{\mu}(k)\\
A_{\nu}(k)
\end{pmatrix}\,,
\end{equation}
and similarly for $B$, $C$ and $D$, which are the mode amplitudes in the positive frequency part of the electric field. The transfer matrix $M$, describing the effect of the scatterer on the four field modes by means of the relation
\begin{equation}
\label{eq:Xuereb2010a:StandardTMM}
\begin{pmatrix}
A(k)\\
B(k)
\end{pmatrix}=M
\begin{pmatrix}
C(k)\\
D(k)
\end{pmatrix}\,,
\end{equation}
is now transformed into an order $4$ tensor of the form
\begin{equation}
\boldsymbol{M}=\begin{bmatrix}
\boldsymbol{m_{11}} & \boldsymbol{m_{12}}\\
\boldsymbol{m_{21}} & \boldsymbol{m_{22}}
\end{bmatrix}\,,
\end{equation}
where each of $\boldsymbol{m_{\alpha\beta}}$ ($\alpha,\beta=1,2$) is a $2\times 2$ matrix relating the respective Jones vector components. A general recipe for transforming the formulae for the field mode amplitudes, as given in \sref{sec:TMM:BasicBuildingBlocksModel}, can be summarised by means of the two replacements
\begin{equation}
1\rightarrow\mathds{1}=\begin{bmatrix}
1 & 0\\
0 & 1
\end{bmatrix}\text{\ and\ }\zeta\rightarrow\boldsymbol{\zeta}\,,
\end{equation}
wherever necessary. In particular, then,
\begin{equation}
\label{eq:Xuereb2010a:NewMatrix}
M=\begin{bmatrix}
1-\i\zeta & -\i\zeta\\
\i\zeta & 1+\i\zeta
\end{bmatrix}\rightarrow\boldsymbol{M}=\begin{bmatrix}
\mathds{1}-\i\boldsymbol{\zeta} & -\i\boldsymbol{\zeta}\\
\i\boldsymbol{\zeta} & \mathds{1}+\i\boldsymbol{\zeta}
\end{bmatrix}\,.
\end{equation}
The polarisability tensor $\boldsymbol{\zeta}$ is defined in \eref{eq:ZetaTensorDefn}. We note that this new transfer matrix still allows us to model the interaction of the multilevel atom with an arbitrary system of immobile optical elements such as mirrors, cavities, waveplates, etc. As was done in \sref{sec:TMM:Model}, this interaction is accounted for by the multiplication of the various transfer matrices of the elements making up the system; this model is, in principle, applicable to systems of arbitrary complexity.
\par
Finally, we recall that the diagonal elements, $\langle i\rvert\m{\rho}^\text{st}\lvert i\rangle$, of the steady-state density matrix $\m{\rho}^\text{st}$ are the populations in each of the sublevels, whereas its off-diagonal elements, $\langle i\rvert\m{\rho}^\text{st}\lvert j\rangle$, are the respective coherences. The matrix elements of $\m{\rho}^\text{st}$ are obtained from the appropriate optical Bloch equations (see, for example, the procedure outlined in Ref.~\cite{CohenTannoudji1977b}). We note here that, through its dependence on $\m{\rho}^\text{st}$, $\boldsymbol{M}$ depends on the fields that it helps to determine, and thus \eref{eq:Xuereb2010a:StandardTMM} will in general become a set of nonlinear equations. In cases like the ones considered in the following sections, where only one multilevel atom is interacting with a linear optical system, this problem may be solved using a procedure similar to the one outlined below:~the fields surrounding the atom are obtained from the input fields through linear operations and then used with the optical Bloch equations to obtain the populations and coherences of the atom's various levels. Knowledge of these quantities then determines the fields, and hence the forces acting on the atom, completely.
\par
In the following sections we will restrict our discussion to the case where the input field is not modified by other transfer matrices. We will apply this mechanism to investigate the behaviour of atoms in two cases where the polarisation of the light varies in space on scales of the order of the wavelength to verify the validity of the model given by \eref{eq:Xuereb2010a:NewMatrix}, \eref{eq:ZetaTensorDefn}, and \eref{eq:PolarisabilityOperator}. In the first instance, we illuminate our atom with two counterpropagating linearly polarised beams. We choose the planes of polarisation of the two beams to be orthogonal to each other. The second configuration we will investigate involves illuminating the atom with two circularly polarised beams, choosing opposite handedness for the two beams. These two cases mirror those in Ref.~\cite{Dalibard1989}.
\subsection{Atoms in a gradient of polarisation}\label{sec:LinPerpLin}
\begin{figure}[t]
\centering
\includegraphics[scale=0.55]{./Diagrams/DiagramsXuereb2010a/1-2_to_3-2}
\caption[Clebsch-Gordan coefficients for a $J=\tfrac{1}{2}\rightarrow J^\prime=\tfrac{3}{2}$ transition]{Clebsch-Gordan coefficients for a $J=\tfrac{1}{2}\rightarrow J^\prime=\tfrac{3}{2}$ transition.}
\label{fig:Xuereb2010a:1-2_to_3-2}
\end{figure}
In this and the following sections, we will adopt the low-intensity hypothesis. This allows us to simplify the optical Bloch equations and resulting system considerably by neglecting the populations and coherences of the excited state sublevels. We can thus replace $\m{\rho}^\text{st}$ by the ground state steady-state density matrix, $\m{\rho}_\text{g}^\text{st}$. We denote the diagonal element $(i,i)$ of $\m{\rho}_\text{g}^\text{st}$, the population in sublevel $i$, by $\Pi_i$, and the off-diagonal element $(i,j)$, the coherence between sublevels $i$ and $j$, by $C_{i,j}$.
\par
Here we will discuss what is perhaps the simplest transition between two levels with multiple magnetic sublevels: the $J=\tfrac{1}{2}\rightarrow J^\prime=\tfrac{3}{2}$ transition. In this case, we have two ground sublevels so that $\m{\rho}_g^\text{st}$ is a $2\times 2$ matrix. \fref{fig:Xuereb2010a:1-2_to_3-2} tabulates the Clebsch-Gordan coefficients required to evaluate $\boldsymbol{\zeta}$. We thus have:
\begin{equation}
\m{\rho}_g^\text{st} = \begin{bmatrix}
\Pi_{-\tfrac{1}{2}} & C_{-\tfrac{1}{2},+\tfrac{1}{2}}\\
C_{+\tfrac{1}{2},-\tfrac{1}{2}} & \Pi_{+\tfrac{1}{2}}
\end{bmatrix}
\end{equation}
and
\begin{equation}
\hat{\boldsymbol{\zeta}}=\zeta_0\begin{pmatrix}
\begin{bmatrix}
\tfrac{1}{3} & 0\\
0 & 1
\end{bmatrix} &
\boldsymbol{0}\\
\boldsymbol{0}&
\begin{bmatrix}
1 & 0\\
0 & \tfrac{1}{3}
\end{bmatrix}
\end{pmatrix}\,,
\end{equation}
whereby
\begin{equation}
\boldsymbol{\zeta}=\zeta_0\Biggl(\begin{bmatrix}
\tfrac{1}{3} & 0\\
0 & 1
\end{bmatrix}\Pi_{-\tfrac{1}{2}}+
\begin{bmatrix}
1 & 0\\
0 & \tfrac{1}{3}
\end{bmatrix}\Pi_{+\tfrac{1}{2}}\Biggr)\,.
\end{equation}
Suppose, now, that we illuminate the atom with two counterpropagating beams having orthogonal linear polarisation and equal intensity. This can be represented by setting
\begin{equation}
\mathbf{B}(k)=\tfrac{B}{\sqrt{2}}\begin{pmatrix}
1\\
1
\end{pmatrix}\,\exp(\i kx-\i\pi/4)\end{equation}
and
\begin{equation}
\mathbf{C}(k)=\tfrac{\i B}{\sqrt{2}}\begin{pmatrix}
1\\
-1
\end{pmatrix}\exp(-\i kx+\i\pi/4)\,,
\end{equation}
where the shift in the $x$ coordinate is introduced to simplify our expressions. Using the optical Bloch equations, we can show that the steady state populations in the ground sublevels at zero atomic velocity are given by
\begin{equation}
\label{eq:Xuereb2010a:LinLinPops}
\Pi_{-\tfrac{1}{2}}=\cos^2(kx)\,\text{ and }\,\Pi_{+\tfrac{1}{2}}=\sin^2(kx)\,,
\end{equation}
noting that the populations do not depend on the field amplitudes in the low intensity regime.
\par
We work to lowest order in $\zeta_0$ and make use of the above relations to find the net force acting on the atom~\cite{Asboth2008}; cf.~\eref{eq:TMM:MSTForce}:
\begin{align}
\label{eq:Xuereb2010a:GeneralForce}
\textsl{\textsf{F}}&=\hbar k\Bigl(\lvert\mathbf{A}\rvert^2+\lvert\mathbf{B}\rvert^2-\lvert\mathbf{C}\rvert^2-\lvert\mathbf{D}\rvert^2\Bigr)\nonumber\\
&=2\hbar k\im{\bigl[\boldsymbol{\zeta}\bigl(\mathbf{B}+\mathbf{C}\bigr)\bigr]\cdot\bigl(\mathbf{B}-\mathbf{C}\bigr)^\ast}+4\tfrac{v}{c}\hbar k\im{\bigl(\boldsymbol{\zeta}\mathbf{B}\bigr)\cdot\mathbf{C}^\ast+\bigl(\boldsymbol{\zeta}\mathbf{C}\bigr)\cdot\mathbf{B}^\ast}\nonumber\\
&\phantom{=\ }-2\tfrac{v}{c}\hbar k^2\im{\biggl[\tfrac{\partial\boldsymbol{\zeta}}{\partial k}\bigl(\mathbf{B}+\mathbf{C}\bigr)\biggr]\cdot\bigl(\mathbf{B}+\mathbf{C}\bigr)^\ast}\nonumber\\
&\approx2\hbar k\im{\bigl[\boldsymbol{\zeta}\bigl(\mathbf{B}+\mathbf{C}\bigr)\bigr]\cdot\bigl(\mathbf{B}-\mathbf{C}\bigr)^\ast}+2\tfrac{v}{c}\hbar k^2\im{\biggl[\tfrac{\partial\boldsymbol{\zeta}}{\partial k}\bigl(\mathbf{B}+\mathbf{C}\bigr)\biggr]\cdot\bigl(\mathbf{B}+\mathbf{C}\bigr)^\ast}\,,
\end{align}
where we have assumed that $\lVert k\,\partial\boldsymbol{\zeta}/\partial k\rVert\gg \lVert\boldsymbol{\zeta}\rVert$. The velocity-dependent force terms in the above expression arise through the Doppler shifting of photons both between field modes in the same polarisation and between field modes in different polarisations; these mechanisms are accounted for by the diagonal and off-diagonal terms in $\boldsymbol{\zeta}$,\footnote{We note that, whilst \eref{eq:Xuereb2010a:GeneralForce} is a general expression, the form of $\boldsymbol{\zeta}$ in this section has no nonzero off-diagonal terms, and only the first type of term contributes.} respectively. These terms emerge through the velocity-dependent terms in the generalised transfer matrix.
\\
In the present case, \eref{eq:Xuereb2010a:GeneralForce} simplifies approximately to
\begin{equation}
\label{eq:Xuereb2010a:StaticLinLinForce}
\textsl{\textsf{F}}=\tfrac{4}{3}\hbar k\zeta_0\lvert B\rvert^2\sin(2kx)\Bigl(\Pi_{+\tfrac{1}{2}}-\Pi_{-\tfrac{1}{2}}\Bigr)=-\tfrac{2}{3}\hbar k\lvert B\rvert^2\zeta_0\sin(4kx)\,,
\end{equation}
assuming that $\zeta_0$ is real for simplicity.
\\
Apart from the velocity-dependent terms in \eref{eq:Xuereb2010a:GeneralForce}, a second type of friction force emerges from the dynamics of the populations in the ground-state sublevels. Upon solving the optical Bloch equations~\cite{Dalibard1989}, expressions for $\Pi_{\pm\tfrac{1}{2}}$ are obtained that have a velocity-dependent term due to the time $\tau_\text{p}$ it takes for an atom in one sublevel to be pumped to the other. This adds a further velocity-dependent term to~\eref{eq:Xuereb2010a:StaticLinLinForce}, giving an overall force
\begin{equation}
\textsl{\textsf{F}}=-\tfrac{2}{3}\hbar k\lvert B\rvert^2\zeta_0\sin(4kx)-\tfrac{8}{3}\hbar k^2\lvert B\rvert^2\zeta_0v\tau_\text{p}\sin^2(2kx),
\end{equation}
which agrees precisely with the standard literature [cf. Eqs.~(4.20) and~(4.23) in Ref.~\cite{Dalibard1989}].
\subsection{Atoms in a gradient of ellipticity}\label{sec:SigmaSigma}
\begin{figure}[t]
\centering
\includegraphics[scale=0.55]{./Diagrams/DiagramsXuereb2010a/1_to_2}
\caption[Clebsch-Gordan coefficients for a $J=1\rightarrow J^\prime=2$ transition]{Clebsch-Gordan coefficients for a $J=1\rightarrow J^\prime=2$ transition.}
\label{fig:Xuereb2010a:1_to_2}
\end{figure}
If we illuminate an atom with two counterpropagating beams of light in a $\sigma^+$--$\sigma^-$ configuration, rich dynamics are obtained not in the simplest ($J=\tfrac{1}{2}\rightarrow J^\prime=\tfrac{3}{2}$) case, but in the next simplest, where the ground state has three magnetic sublevels ($J=1$) and the excited state five ($J^\prime=2$). In this case, then, we can express $\m{\rho}_g^\text{st}$ and $\hat{\boldsymbol{\zeta}}$ as
\begin{equation}
\m{\rho}_g^\text{st} = \begin{bmatrix}
\Pi_{-1} & C_{-1,0} & C_{-1,+1}\\
C_{0,-1} & \Pi_0 & C_{0,+1}\\
C_{+1,-1} & C_{+1,0} & \Pi_{+1}
\end{bmatrix}
\end{equation}
and
\begin{equation}
\hat{\boldsymbol{\zeta}}=\zeta_0\begin{pmatrix}
\begin{bmatrix}
\tfrac{1}{6} & 0\\
0 & 1
\end{bmatrix} &
\boldsymbol{0} &
\begin{bmatrix}
0 & \tfrac{1}{6}\\
0 & 0
\end{bmatrix}\\
\boldsymbol{0} &
\begin{bmatrix}
\tfrac{1}{2} & 0\\
0 & \tfrac{1}{2}
\end{bmatrix} &
\boldsymbol{0}\\
\begin{bmatrix}
0 & 0\\
\tfrac{1}{6} & 0
\end{bmatrix}&
\boldsymbol{0} &
\begin{bmatrix}
1 & 0\\
0 & \tfrac{1}{6}
\end{bmatrix}
\end{pmatrix}\,,
\end{equation}
using the Clebsch-Gordan coefficients in \fref{fig:Xuereb2010a:1_to_2}. Together, these give
\begin{equation}
\boldsymbol{\zeta}=\zeta_0\Biggl(\begin{bmatrix}
\tfrac{1}{6} & 0\\
0 & 1
\end{bmatrix}\Pi_{-1}+
\begin{bmatrix}
\tfrac{1}{2} & 0\\
0 & \tfrac{1}{2}
\end{bmatrix}\Pi_0+
\begin{bmatrix}
1 & 0\\
0 & \tfrac{1}{6}
\end{bmatrix}\Pi_{+1}+\begin{bmatrix}
0 & \tfrac{1}{6}\\
0 & 0
\end{bmatrix}C+
\begin{bmatrix}
0 & 0\\
\tfrac{1}{6} & 0
\end{bmatrix}C^\ast\Biggr)\,,
\end{equation}
with $C=C_{+1,-1}=C_{-1,+1}^\ast=\langle+1\vert\m{\rho}_\text{g}^\text{st}\vert\!-\!\!1\rangle$ representing the nonzero coherence between the $m_J=+1$ and the $m_J=-1$ sublevels. Note that we again apply the low intensity hypothesis, thereby replacing $\m{\rho}^\text{st}$ with $\m{\rho}_\text{g}^\text{st}$.
\par
We now illuminate the atom with two counterpropagating beams of equal intensity, $\mathbf{B}$ and $\mathbf{C}$, possessing $\sigma^+$ and $\sigma^-$ polarisation, respectively:
\begin{equation}
\mathbf{B}(k)=B\begin{pmatrix}
1\\
0
\end{pmatrix}\,\exp(\i kx)\text{\ and\ }
\mathbf{C}(k)=B\begin{pmatrix}
0\\
1
\end{pmatrix}\exp(-\i kx)\,.
\end{equation}
We again use~\eref{eq:Xuereb2010a:GeneralForce} to derive the force acting on the atom, which is given by
\begin{align}
\label{eq:Xuereb2010a:SigmaSigmaForce}
\textsl{\textsf{F}}&=2\hbar k\lvert B\rvert^2\,\imag{\tfrac{5}{6}\zeta_0\bigl(\Pi_{+1}-\Pi_{-1}\bigr)+\tfrac{1}{6}\i\zeta_0\im{C\exp(-2\i kx)}}\nonumber\\
&\phantom{=\ }-2\tfrac{v}{c}\hbar k^2\lvert B\rvert^2\im{\partial\zeta_0/\partial k}\Bigl(\tfrac{7}{6}\bigl(\Pi_{+1}+\Pi_{-1}\bigr)+\Pi_0+\tfrac{1}{3}\re{C\exp(-2\i kx)}\Bigr)\,,
\end{align}
where the populations and coherences are again obtained from the optical Bloch equations, and can be found in Ref.~\cite{Dalibard1989}. By observing the natural correspondence between $\zeta_0$ and $s_\pm$ in this reference, we can see that our expression for the force acting on the atom again agrees with the standard literature to first order in $\tfrac{v}{c}$ [cf. Eq.~(5.9) in Ref.~\cite{Dalibard1989}].
The resulting friction force is thus due to both the Doppler shift, as evident in the terms shown explicitly in~\eref{eq:Xuereb2010a:SigmaSigmaForce}, as well as to the non-adiabatic following of the atomic sublevel populations.
\appendicesstart
\section{Appendix: Cavity properties from the transfer matrix model}
In this section we will derive a consistent set of relations used to describe cavities, based on the transfer matrix formalism. Specifically, we will relate the cavity HWHM linewidth $\kappa$, its finesse $\mathcal{F}$, and its $Q$-factor to one another and to the reflectivity of the cavity mirrors.
\subsection{Cavity finesse}\label{sec:TMM:CavityFinesse}
The finesse of a cavity is defined as
\begin{equation}
\label{eq:FinesseDefn}
\mathcal{F}=\frac{\Delta\lambda}{\delta\lambda}\,,
\end{equation}
where $\delta\lambda$ is the FWHM linewidth of the cavity transmission peak and $\Delta\lambda$ is the free spectral range (FSR) of the cavity. We model the cavity as a Fabry--P\'erot resonator having mirrors of reflectivity $r_1$ and $r_2$ and a length $L$. On resonance, we can find some integer $n$ such that
\begin{equation}
L=\tfrac{1}{2}n\lambda\,.
\end{equation}
The FSR is defined~\cite{Siegman1990} as the wavelength interval such that
\begin{equation}
L=\tfrac{1}{2}(n+1)(\lambda-\Delta\lambda)\,.
\end{equation}
We approximate $\Delta\lambda\ll\lambda$, whereby
\begin{equation}
\label{eq:FSR}
\Delta\lambda=\frac{\lambda^2}{2L}\,.
\end{equation}
The cavity is described by the transfer matrix equation:
\begin{equation}
\begin{pmatrix}
A\\
B
\end{pmatrix}=
\frac{1}{t_1t_2}
\begin{bmatrix}
t_1^2-r_1^2 & r_1\\
-r_1 & 1
\end{bmatrix}
\begin{bmatrix}
e^{-\i kL} & 0\\
0 & e^{\i kL}
\end{bmatrix}
\begin{bmatrix}
t_2^2-r_2^2 & r_2\\
-r_2 & 1
\end{bmatrix}
\begin{pmatrix}
C\\
D
\end{pmatrix}\,.
\end{equation}
We set $C=0$ and write the transmitted field $D$ in terms of the only input field, $B$:
\begin{equation}
\label{eq:TransmissionLorentzian}
\lvert D\rvert^2=\frac{\lvert t_1t_2\rvert^2}{\lvert 1-\lvert r_1r_2\rvert\exp(2\i kL)\rvert^2}\lvert B\rvert^2\,,
\end{equation}
where the phase shifts induced by $r_1$ and $r_2$ have been absorbed in $L$. For a reasonably good cavity ($\lvert t_{1,2}\rvert\ll 1$, $\delta\lambda\ll\lambda$), the transmission is therefore Lorentzian, with a FWHM linewidth
\begin{equation}
\label{eq:FWHMLinewidth}
\delta\lambda=\frac{\lambda}{kL}\frac{1-\lvert r_1r_2\rvert}{\sqrt{\lvert r_1r_2\rvert}}\,.
\end{equation}
Finally, we substitute \eref{eq:FSR} and \eref{eq:FWHMLinewidth} into \eref{eq:FinesseDefn} to obtain
\begin{equation}
\mathcal{F}=\frac{\pi\sqrt{\lvert r_1r_2\rvert}}{1-\lvert r_1r_2\rvert}\,.
\end{equation}
If the approximations $\lvert t_{1,2}\rvert\ll 1$ and $\delta\lambda\ll\lambda$ no longer hold, it can be similarly shown that \eref{eq:TransmissionLorentzian} implies
\begin{equation}
\mathcal{F}=\frac{\pi/2}{\sin^{-1}\biggl(\frac{1-\lvert r_1r_2\rvert}{2\sqrt{\lvert r_1r_2\rvert}}\biggr)}\,.
\end{equation}
\subsection{Physical meaning of the cavity finesse}\label{sec:TMM:CavityFinesseN}
The factor $\rho=\lvert r_1r_2\rvert$ present in the above relations is related to the power lost by the cavity after one round-trip, $1-\rho^2$. Let us set $N=\sqrt{\rho}/(1-\rho)$. The power remaining in the cavity after $N$ round-trips is then
\begin{equation}
\rho^{2N}=\rho^{\frac{2\sqrt{\rho}}{1-\rho}}\to\frac{1}{e^2}\,,
\end{equation}
where we have taken the good-cavity ($\rho\to 1$) limit. In other words, we can write
\begin{equation}
\mathcal{F}=\pi N\,,
\end{equation}
where $N$ is the number of round-trips the light makes inside the cavity before the intensity decays by a factor of $1/e^2$.
\subsection{Cavity linewidth and quality factor}\label{sec:TMM:CavityKappaQ}
The HWHM cavity linewidth in frequency space can be defined in terms of the FWHM linewidth in wavelength space by means of the relation
\begin{equation}
2\kappa=\frac{\omega}{\lambda}\,\delta\lambda\,,
\end{equation}
whereupon
\begin{equation}
\delta\lambda=\frac{\kappa\lambda^2}{\pi c}\,,
\end{equation}
and substituting this expression for the linewidth into \eref{eq:FinesseDefn} gives
\begin{equation}
\mathcal{F}=\frac{\pi c}{2\kappa L}\,,\text{ or }\mathcal{\kappa}=\frac{\pi c}{2\mathcal{F}L}\,.
\end{equation}
The quality factor, or $Q$-factor, is defined as the ratio of the cavity frequency to its FWHM linewidth in frequency space: $Q=\omega/(2\kappa)$, or
\begin{equation}
Q=\frac{2\mathcal{F}L}{\lambda}\,.
\end{equation}
\appendicesend
\chapter{Applications of transfer matrices}\label{ch:TMMApplications}
\epigraph{[...] [T]he sciences do not try to explain, they hardly even try to interpret, they mainly make models. By a model is meant a mathematical construct which, with the addition of certain verbal interpretations, describes observed phenomena. The justification of such a mathematical construct is solely and precisely that it is expected to work [...].}{J.\ von\ Neumann, \emph{Method in the Physical Sciences} (1955)}
In this chapter, I will apply the transfer matrix method developed in \cref{ch:TMM:TMM} to novel cooling geometries outside cavities (\sref{sec:TMM:ECCO} and \sref{sec:TMM:Comparison}), as well as inside active ring cavities (\sref{sec:TMM:AmplifiedOptomechanics}).
\section{External cavity cooling}\label{sec:TMM:ECCO}
The basis for this current section was published as Xuereb, A., Freegarde, T., Horak, P., \& Domokos, P. Phys.\ Rev.\ Lett.\ \textbf{105}, 013602 (2010). We will apply the general solution detailed above to the external cavity cooling configuration, \sref{sec:CoolingMethods:Other:ECCO}, in the case of a micro-mechanical mirror. As a reference system for the analysis of the cooling force in this setup, we also consider the mirror-mediated cooling configuration, which is {the optomechanical} cooling scheme used in many experiments~\cite{Metzger2004,Arcizet2006,Gigan2006,Schliesser2008}. Note that in the external cavity cooling scheme {with a near mirror of complex transmissivity $t$, the limits of small and large $\lvert t\rvert$ render the situation where the cavity is replaced respectively by the near mirror only or the far mirror only.} For intermediate {$t$ compared {with} the transmissivity of the far mirror, $T$}, the moving scatterer interacts with a field reflected back from the {cavity and} is subject to the interference created by the multiple reflections between the two mirrors. Throughout most of this section, although not initially, we {consider in particular} an object having low reflectivity, around $50$\%, which corresponds to a polarisability $\zeta=-1$ and is representative of typical experimental conditions~\cite{Metzger2004}. This ensures that a high-finesse resonator cannot be formed between the object and the near mirror, thereby guaranteeing a parameter range where the cavity formed between the immobile mirrors dominates the interaction. For the sake of simplicity, we restrict ourselves to the special case of scatterers that can be characterised by a real polarisability; this is equivalent to assuming that no absorption takes place in the scatterer. Similar results hold when $\zeta$ is not real.
\par
\begin{figure}
\centering
\includegraphics[width=1.5\figwidth]{./Diagrams/DiagramsXuereb2010b/detuning}
\caption[The amplitude of the friction force acting on the scatterer, for various near-mirror transmissivities]{The amplitude of the friction force acting on the scatterer, for various near-mirror transmissivities, is shown as a function of the mirror separation in the cavity. The different curves represent different near-mirror transmissivities: $\lvert t\rvert=0.45$ (dashed--dotted curve), $\lvert t\rvert=0.20$ (dotted), $\lvert t\rvert=0.10$ (dashed), $\lvert t\rvert=0.05$ (solid). (Scatterer polarisability $\zeta=-1$, scatterer--cavity separation $x\approx 400\lambda_0$, $\lvert T\rvert=0.01$, $\lambda_0=780$~nm.)}
\label{fig:Detuning}
\end{figure}
A numerical fit to \eref{eq:FullFriction} for $\lvert t\rvert\sim\lvert T\rvert$ and $\lvert\zeta\rvert\ll 1$ renders a friction force of the approximate form
\begin{equation}
\label{eq:ECCOFriction}
\textsl{\textsf{F}}\approx-8\hbar k_0^2\zeta^2\tfrac{v}{c}(2x+0.17\mathcal{F}L)\sin(4k_0x+\phi)\lvert B_0\rvert^2\,,
\end{equation}
where $\mathcal{F}$ is the cavity finesse, $L$ the cavity length (optimised as discussed below), $x$ the separation between the scatterer and the near mirror, and $\phi$ a phase factor. The gross spatial variation of the friction force is linear in both $L$ and $x$; this is simply because of the linear increase of the retardation time of the reflected field with {the} distance between the scatterer and the mirrors. This dependence is modulated by a wavelength-scale oscillation of the friction force, which thereby follows the same oscillatory dependence as mirror-mediated cooling [cf.~\eref{eq:TMM:MirrorCoolForce}] and constrains cooling to regions of the size of $\lambda_0/8$, where $\lambda_0=2\pi/k_0$. In the case of a micro-mechanical mirror, where the vibrational amplitude is naturally much less than the wavelength, this presents no problem. The form of \eref{eq:ECCOFriction} is dependent on the properties of the scatterer and of the mirrors; for realistic mirrors and $\zeta=-1$, the enhancement factor $0.17\mathcal{F}$ drops to $0.04\mathcal{F}$. With typical experimental parameters this results in an enhancement of $10^3$--$10^4$ over the standard setup.
\par
As shown in \fref{fig:Detuning}, the fine tuning of the cavity {length by varying $L$ on the wavelength scale shows a Lorentzian-like resonant enhancement} of the friction amplitude (i.e., the amplitude of the cooling coefficient), {following that of} the intracavity field intensity. If we denote the complex reflectivities of the near and far mirror by $r$ and $R$, respectively, we can show that the peaks of \fref{fig:Detuning} lie around the cavity resonances, at approximately $L=\tfrac{1}{2}m\lambda_0-\tfrac{1}{2k_0}\arg\big(rR\big)$, with $m$ being an integer, and have approximately the same full-width at half-maximum, $\bigl(1-\lvert r R\rvert\bigr)/\bigl(k_0\sqrt{\lvert r R\rvert}\bigr)=\lambda_0/(2\mathcal{F})$. The enhancement of the friction {force by} the cavity {is due to the multiplication of the retardation time by the number of round trips in the cavity, which thereby acts as a `distance folding' mechanism. For the chosen parameters, the optical path length is effectively $2x+0.04\mathcal{F}L$; i.e., determined predominantly by the cavity length $L$.}
\par
{The friction force depends not only upon the retardation but also upon the cavity reflectivity, which drops near resonance in the well-known behaviour of a Fabry--P\'erot resonator. \fref{fig:TMM:Friction} shows the friction amplitude as a function of the near mirror transmissivity $\lvert t \rvert$ for a fixed far mirror transmissivity, $T=1/(1+100\i)$.} {We note that this nonideal reflectivity of the far mirror could equivalently arise from absorption, of ca.\ $0.01$\% with the given parameters, of the incident power by the mirror.}
\begin{figure}
\centering
\includegraphics[width=1.5\figwidth]{./Diagrams/DiagramsXuereb2010b/friction_vs_t}
\caption[Amplitude of the friction acting on a scatterer of polarisability $\zeta=-1$]{Amplitude of the friction acting on a scatterer of polarisability $\zeta=-1$ interacting with a cavity tuned to achieve maximum friction, for varying transmissivity of the near mirror. The friction amplitude (solid curve) approaches that for mirror-mediated cooling using the far (dotted line, $t\rightarrow 1$) or the near (dashed--dotted line, $t\rightarrow 0$) mirror only in the appropriate limits. The arrow indicates the point at which the two cavity mirrors have the same reflectivity. Also shown is the intracavity field (dashed). ($x\approx 400\lambda_0$, $L\approx 2000\lambda_0$, $\lvert T\rvert=0.01$, $\lambda_0=780$~nm, finesse at peak friction $5.0\times 10^4$.)}
\label{fig:TMM:Friction}
\end{figure}
{For each value of $\lvert t \rvert$, the cavity length $L$ has been adjusted to maximise the friction force, according to curves such as those in \fref{fig:Detuning}. The calculated result follows the intracavity {field (shown {dashed)} except} where the cavity reflectivity drops near resonance [region (b)], and in the extremes of regions (a) and (c), where the geometry is dominated by the near ($\lvert t\rvert\rightarrow 0$) or far ($\lvert t\rvert\rightarrow 1$) mirrors, respectively. \fref{fig:CandR} shows the effect of the drop in reflectivity as the cavity is scanned through resonance for similar mirror reflectivities. When this causes a dip in the friction amplitude peak, the optimum values plotted in \fref{fig:TMM:Friction} occur to either side of the resonance, and the friction force in this region is effectively limited by this interference effect.} {We note that the friction amplitude is not maximised at the point of maximum intracavity field ($t=T$)} because more light is lost through the cavity for larger $\lvert t\rvert$.
\begin{figure}
\centering
\includegraphics[width=1.5\figwidth]{./Diagrams/DiagramsXuereb2010b/collapse_and_revival}
\caption[Attenuation of the cooling coefficient amplitude for certain parameters]{In region (b) of \fref{fig:TMM:Friction}, the cooling coefficient amplitude (solid curves) is attenuated due to the attenuation in the field reflected from the cavity (dashed). $\lvert T\rvert=0.01$ in every plot; $\lvert t\rvert$ is, from left to right, $6.7\times 10^{-3}$, $8.3\times 10^{-3}$, $1.0\times 10^{-2}$, $1.2\times 10^{-2}$, and $1.5\times 10^{-2}$. (Parameters as in~\fref{fig:TMM:Friction}.)}
\label{fig:CandR}
\end{figure}
\par
{{The external cavity cooling mechanism may prove particularly valuable when the scatterer is a small mirror or other micro-mechanical optical component. In such cases, the advantage gained by using the external cavity over the standard optomechanical cooling scheme depends heavily upon the polarisability or reflectivity of the moving scatterer, which in the above calculations have so far been taken to be modest ($\zeta=-1$; $\lvert r\rvert=0.7$) in comparison with those of the cavity mirrors.} For $\lvert\zeta\rvert\ll 1$, the friction force is enhanced by a factor approximately equal to $\mathcal{F}$} because of the distance folding argument explained above. For larger $\lvert\zeta\rvert$, the system turns into a three--mirror resonator and the advantage of external cavity cooling is not as big, but is still significant. For $\lvert\zeta\rvert\approx 1$ we find enhancement by a factor $0.04\mathcal{F}$, as discussed above. For even larger $\zeta$, when the reflectivity of the moving mirror becomes comparable to that of the fixed mirrors, the scheme behaves similarly to the mirror-mediated cooling {configuration}. The main heating process that counteracts the cooling effect in the case of micromirrors is thermal coupling to the environment, which depends on the geometry. In the case of isolated scatterers that undergo no absorption, the heating is due to quantum fluctuations in the fields (see \aref{sec:TMM:AppDiffusion}); the limit temperature here is $\Lapprox\hbar c/\bigl(0.34k_\text{B}\mathcal{F}L\bigr)=1.87\hbar\kappa/k_\text{B}$ when $\lvert\zeta\rvert\ll 1$, which evaluates to $\Lapprox 0.1$\,mK for the parameters in \fref{fig:TMM:Friction}. This expression for the temperature conforms to the form expected from \eref{eq:GeneralTForm}.
\par
The usual cavity-mediated cooling {mechanism}~\cite{Thompson2008,Favero2008}, where the moving scatterer is inside a two-mirror cavity, can also be described by our general framework in terms of \erefs{eq:Afield} and (\ref{eq:FullFriction}). Compared with this scheme, external cavity cooling has the advantage of always having a sinusoidal spatial dependence; the narrow resonances in the friction force for particles inside a cavity, as seen in the next section, impose more stringent positioning requirements.
\section{Cavity cooling of atoms: within and without a cavity}\label{sec:TMM:Comparison}
In \sref{sec:TMM:ECCO} we explored the cooling of a micromirror outside a high-finesse cavity as a means of enhancing the optomechanical interaction of the mirror with the field. Our formalism is general and, indeed, so are our results: the mechanism works similarly for atoms. Inside a resonator, the enhancement of the interaction is accompanied by an enhancement in the field itself, which precludes using high optical powers in order not to saturate the atom. Outside a cavity, the enhancement in the friction force is lower, but the field itself is not amplified by the presence of a cavity, and therefore one can use much higher powers. The {important question}, then, is whether these two effects compensate for one another in such a way as to render the cooling forces experienced by an atom outside a cavity similar to those it experiences inside.\\
To answer this question we will first {describe} the two models, in \Sref{sec:CMC} and \Sref{sec:ECCO}, respectively, using realistic parameters for state-of-the-art {optical devices}. Taking into account saturation effects, it is seen that the two different models result in similar cooling forces and equilibrium temperatures. An examination of scaling properties of the force acting on the atom {in the two schemes} then follows in \Sref{sec:Scaling}. The work in this section has been accepted for publication in the Eur.\ Phys.\ J.\ D topical issue on \emph{Cold Quantum Matter -- Achievements and Prospects}.
\subsection{Comparison of cavity cooling schemes}\label{sec:ComparisonAtoms}
\subsubsection{Cavity-mediated cooling: Atom inside the cavity}\label{sec:CMC}
\begin{figure}[t]
\centering
\subfigure[]{
\includegraphics[scale=0.5]{Diagrams/ischgl/CMC}
}\\
\subfigure[]{
\includegraphics[width=1.5\figwidth]{Diagrams/ischgl/CMC_Friction}
}
\caption[Schematic model for cavity cooling inside cavities]{(a)~{Model} of a {scatterer}, $\text{S}$, inside a symmetric Fabry--P\'erot cavity of length $L_\text{c}$. The {cavity mirrors have} reflection and transmission coefficients, $\mathfrak{r}$ and $\mathfrak{t}$, and $\zeta$ is the polarisability of $\text{S}$. (b)~Cooling coefficient $\varrho$ per unit input power, experienced by the scatterer at different positions in the cavity for realistic parameters (see text for details).}
\label{fig:CMC}
\end{figure}
Placing a scatterer---atom~\cite{Horak1997,Leibrandt2009,Koch2010}, micromirror~\cite{Bhattacharya2007a,Thompson2008}, or `point polarisable particle'~\cite{Domokos2003}---inside a cavity has long been pointed out to be a generic means of cooling the translational motion of that scatterer. Cooling of atoms inside resonators has been observed: first~\cite{Leibrandt2009} as a means of counteracting the heating of a trapped ion inside a cavity, leading to a steady-state occupation number for the motion of the atom of around $20$ quanta; and later~\cite{Koch2010} as an increase in the storage time for a neutral atom inside a cavity from $35$\,ms to $1100$\,ms. The generic layout of such an experiment is {shown} in \fref{fig:CMC}(a). For our purposes, we place the scatterer inside a symmetric Fabry--P\'erot cavity of length $L_\text{c}$, which we pump from one side; the dominant field inside the cavity is a standing {wave} field if the reflectivity of the mirrors, $\mathfrak{r}$, is high enough. For a numeric example, we use the same cavity properties as Ref.~\cite{Mucke2010}: {finesse} $\mathcal{F}=56\,000$ {modelled by using mirrors with $\mathfrak{t}=\mathfrak{r}+1=1/(1+133.5\i)$}, {cavity} length $495$\,$\upmu$m, and mode waist $30$\,$\upmu$m{; we use a wavelength $\lambda=780$\,nm. In contrast with Ref.~\cite{Mucke2010}, however}, let us reiterate that our cavity is pumped along its axis.\\
We also take the scatterer to be a {two-level} atom, {with} the cavity {field detuned} $10\Gamma$ to the red of the atom transition frequency. Thus, the polarisability of the atom is $\zeta=4.1\times 10^{-5}+4.1\times 10^{-6}\i$. The maximum cooling coefficient is found at a detuning of $-2.6$\,$\kappa$ from the cavity resonance. As expected~\cite{Domokos2003}, the optimal cooling coefficient occurs for a negative detuning of the pump from the \emph{bare} cavity resonance, but for a positive detuning from the {dressed atom--cavity} resonance.\par
The dependence of the friction force, \eref{eq:FullFriction}, on the position of the {scatterer, scanned over a wavelength, is shown in \fref{fig:CMC}(b)}. {The presence of the cavity manifests itself primarily through a strong enhancement of both the cooling coefficient $-\textsl{\textsf{F}}/v$ and the intracavity field intensity. The scattering model explored above is only valid in the limit of small saturation. For the $^{85}$Rb D$_2$ transition, assuming that the beam is circularly polarised, the saturation intensity is $1.67$\,mW\,cm$^{-2}$~\cite{Steck2008}. In order to avoid saturation effects, we restrict the power input into the cavity to $2$\,pW; this equates to an intracavity intensity of $23$\,mW\,cm$^{-2}$ and hence a saturation parameter $s=0.14$; this is because $s$ is inversely proportional to the square of the detuning, $-10\Gamma$ in this case, of the pump beam from resonance. In turn, this input power also yields a maximum cooling coefficient of $1.5\times 10^{-20}$\,N/(m\,s$^{-1})$, which corresponds to a $1/e$ velocity cooling time of $9$\,$\upmu$s for the same atom; averaging the friction force over a wavelength gives a cooling time of $37$\,$\upmu$s.}\par
The friction force and the momentum diffusion both scale linearly with the input power, {in the low-saturation regime}. {Therefore, the equilibrium temperature is independent of the pump power in this regime. For the parameters used above}, the equilibrium temperature predicted for a scatterer at the point of maximum friction is $56$\,$\upmu$K; averaging the cooling coefficient, as well as the diffusion coefficient, over a wavelength, gives a higher equilibrium temperature of $220$\,$\upmu$K.
\subsubsection{External cavity cooling: {Atom} outside the cavity}\label{sec:ECCO}
\begin{figure}[t]
\centering
\subfigure[]{
\includegraphics[scale=0.5]{Diagrams/ischgl/ECCO}
}\\
\subfigure[]{
\includegraphics[width=1.5\figwidth]{Diagrams/ischgl/ECCO_Friction}
}
\caption[Schematic model for external cavity cooling]{{External cavity cooling. (a)~Model, similar to \fref{fig:CMC}, but with the atom at a distance $L-x$ outside the cavity.} (b)~Cooling coefficient $\varrho$ per unit input power experienced by the scatterer as $x$ is varied for realistic parameters (see text for details). Note the change of scale, on the vertical axis, from \fref{fig:CMC}(b).}
\label{fig:ECCO}
\end{figure}
In \sref{sec:TMM:ECCO} above, we also proposed~\cite{Xuereb2010b} that even with the scatterer \emph{outside} the cavity, the cavity's resonance can be exploited to greatly enhance the optomechanical friction experienced by the scatterer \emph{vis-\`a-vis} standard optomechanical cooling setups~\cite{Braginsky1967,Metzger2004,Groblacher2009a}, which place the scatterer in front of a single mirror. It is the aim of this subsection to explore this cooling mechanism, using {experimental parameters similar to those in the previous subsection}, and compare it to the cavity-mediated cooling mechanism discussed {there}.\par
Our mathematical model, \fref{fig:ECCO}, represents the cavity as a standard, symmetric Fabry--P\'erot cavity. However, we emphasise that in principle what is required is simply an optical resonance: the cavity in the model can indeed be replaced by whispering gallery mode resonators~\cite{Schliesser2010} or even solid-state resonators. As a basis for numerical calculations{, and to enable direct comparison}, we model the same resonator as in the previous subsection. It is important to emphasise that the achievable quality factors of the resonators used for external cavity cooling can intrinsically be made larger than the ones in the previous subsection (see, e.g., Ref.~\cite{Rempe1992}) because there does not need to be any form of optical or mechanical access inside the resonator itself.
\par
The pump beam frequency is again taken to be detuned by $10\Gamma$ to the red of the atomic transition. By placing the atom outside the cavity, one is free to use high-numerical-aperture optics to produce a tighter focus than might be possible in a cavity with good optical and mechanical access. Having a tight focus strengthens the atom--field coupling because of the $1/w^2$ dependence of $\zeta$ on the beam waist $w$; whereas the friction force scales linearly with the input power, it also scales as $\zeta^2\sim 1/w^4$ [cf.~\eref{eq:TMM:MirrorCoolForce}]. Focussing the beam therefore increases the atom--field coupling more than the local intensity. Thus, it is now assumed that the beam is focussed down to $1$\,$\upmu$m{, which} gives $\zeta=3.7\times 10^{-2}+3.7\times 10^{-3}\i$.\\
In order to make a fair comparison between the two cases, we choose to set the saturation parameter $s=0.14$, as in the previous subsection. The maximum achievable cooling coefficient is then {$2.9\times 10^{-21}$\,N/(m\,s$^{-1}$) for $200$\,pW} of input power, which is an order of magnitude smaller than the previous result {and leads to a $1/e$ velocity cooling time of $50$\,$\upmu$s and an equilibrium temperature of $280$\,$\upmu$K. The magnitude of the force in this case results from the much smaller pumping beam mode waist and the use of much higher powers, subsequently leading to a} stronger atom--field interaction. {W}ith this beam waist and finesse we would be restricted to input powers several orders of magnitude smaller if the atom were inside such a cavity.
\par
In {summary}, whereas the friction force inside a cavity is much stronger \emph{per unit input power and for the same beam waist}, the restrictions imposed on the magnitude of these quantities {when the atom is inside the cavity} reduce the maximally achievable friction force to a figure comparable to {when it lies} outside the cavity.
\subsection{Scaling properties {of cavity} cooling forces}\label{sec:Scaling}
\subsubsection{Localisation issues}\label{sec:Localisation}
\begin{figure}[t]
\centering
\includegraphics[width=1.5\figwidth]{Diagrams/ischgl/CMC_Waists}
\caption[Spatial dependence of the friction force acting on an atom inside a cavity]{Spatial dependence of the friction force acting on an atom \emph{inside} a cavity with different mode waists (i.e., different polarisabilities) but equal detuning from resonance, $10\Gamma$ to the red. The smaller the mode waist the stronger the friction force, by several orders of magnitude, but the more significant localisation issues become. (Parameters as in \Sref{sec:CMC} but with $\partial\zeta/\partial k=0$.)}
\label{fig:CMC_Waists}
\end{figure}
\begin{figure}[t]
\centering
\includegraphics[width=1.5\figwidth]{Diagrams/ischgl/ECCO_Waists}
\caption[Spatial dependence of the friction force acting on an atom outside a cavity]{Spatial dependence of the friction force acting on an atom \emph{outside} a cavity, with different pumping field waists {but equal} detuning from resonance, $10\Gamma$ to the red. The friction force scales roughly as the inverse fourth power of the waist [cf.~\eref{eq:TMM:MirrorCoolForce}], but the length scale of the cooling and heating regions is unaffected. (Parameters as in \Sref{sec:ECCO} but with $\partial\zeta/\partial k=0$.)}
\label{fig:ECCO_Waists}
\end{figure}
{The} broad nature of the spatial variations in the force shown in \fref{fig:CMC}(b) is {a consequence} of the small polarisability of the atom in such a cavity. {This is in sharp contrast to the case of large polarisability, achieved by an atom at a tight beam focus, as shown} in \fref{fig:CMC_Waists}, or by a micromirror. A {scatterer of larger polarisability} would experience extremely narrow ($\ll\lambda$) peaks in the friction force inside a cavity but not outside it.\\
{Within the scattering model used in this section, the atom--cavity coupling can be tuned by varying either the beam waist or the laser detuning from atomic resonance. Experimentally, however, atom--cavity coupling is rarely investigated close to resonance, in order to minimise the effects of atomic decorehence through spontaneous emission. In such cases, this coupling can be increased by operating a cavity with a small mode waist; this may in turn be detrimental to the performance of the system due to the strong sub-wavelength nature of the interaction, as explored in \fref{fig:CMC_Waists}.} The net {effect of having a smaller mode waist is} that this not only demands extremely good localisation but also tends to decrease the effective cooling coefficient drastically{---by up to several orders of magnitude---}because of spatial averaging effects.\par
In \Sref{sec:ECCO}, no mention was made of the average friction force acting on the scatterer; indeed this average computes to approximately zero for any case involving far-detuned atoms, or other particles with an approximately constant polarisability, outside cavities. This, then, {also demands strong} localisation of the atom; whilst experimentally challenging this disadvantage is somewhat mitigated by the {easy} mechanical and optical access afforded by external cavity cooling schemes. {In \fref{fig:ECCO_Waists} it is shown that the polarisability of the atom can be varied over a very wide range without affecting the length scale of the cooling and heating regions. The friction force can be seen to vary as $1/w^4$ for a beam waist $w$; in turn, this originates from the $\zeta^2$ scaling of the friction force [cf.~\eref{eq:TMM:MirrorCoolForce}].\\
In contrast with the atomic situation}{, if the scatterer is a micromirror mounted on a cantilever, localisation {does not present a problem}, since such micromirrors naturally undergo small} oscillations {and can be positioned with sub-nm accuracy}.
\subsubsection{Scaling with cavity {finesse and linewidth}}\label{sec:CavityLength}
{Cavity-mediated cooling mechanisms are heavily dependent on the physical properties of the cavity, namely its linewidth $\kappa$ and finesse $\mathcal{F}$. These parameters can be tuned independently by changing the length of the cavity and the reflectivity of its mirrors. This subsection briefly explores how the two mechanisms we are considering scale with $\kappa$ and $\mathcal{F}$.
\par
Expressions for the force acting on an atom inside a good cavity are not simple to write down. Nevertheless, in the good-cavity limit one may obtain an analytical formulation for the limiting temperature~\cite{Horak1997}:
\begin{equation}
T=\frac{\hbar\kappa}{k_\text{B}}\,,
\end{equation}
i.e., making a cavity longer decreases the equilibrium temperature proportionately. This result can be justified by observing that whereas the diffusion constant depends only on the intensity inside the cavity ($\propto\mathcal{F}$), the friction force scales linearly with both the intensity and, if the intensity is kept constant, with the lifetime of the cavity field ($\propto 1/\kappa$). The friction force is therefore proportional to $\mathcal{F}/\kappa$, and the equilibrium temperature proportional to $\kappa$.}
\par
\begin{figure}[t]
\centering
\includegraphics[width=1.5\figwidth]{Diagrams/ischgl/ECCO_Lengths}
\caption[Scaling with cavity length of the friction force acting on an atom outside a cavity]{Spatial dependence of the friction force acting on an atom \emph{outside} a cavity, with detuning $10\Gamma$ to the red of resonance. Three different cavity lengths are shown; the friction force scales almost linearly with the cavity length. Note the different scaling factors and cavity lengths, given above each curve. (Parameters as in \Sref{sec:ECCO} but with $\partial\zeta/\partial k=0$.)}
\label{fig:ECCO_Lengths}
\end{figure}
As is known from \sref{sec:TMM:ECCO}, the friction force ($\textsl{\textsf{F}}$) acting on {an atom} outside a cavity scales {approximately} linearly with {both the length and the finesse} of the cavity. This is interpreted in terms of a `distance folding' mechanism: the lifetime of the light inside a cavity scales {inversely} with its {linewidth $\kappa\propto 1/\bigl(\mathcal{F}L_\text{C}\bigr)$} if all other parameters are kept fixed. Within the range of parameters where the light coupled into the cavity dominates the optomechanical effects, this implies that {$\textsl{\textsf{F}}\propto 1/\kappa$}. One can see this behaviour reproduced in \fref{fig:ECCO_Lengths}, where the cooling coefficient acting on an atom outside each of three cavities having different lengths is shown. This mechanism loses its importance if {the atomic polarisability} is too large, whereby the system behaves more like two coupled cavities, or if the cavity is too long. {The momentum diffusion affecting the atom outside a cavity is essentially independent of the length of the properties of a good cavity, since it depends on the local intensity surrounding the scatterer. Putting these two results together, then, gives (\sref{sec:TMM:ECCO})
\begin{equation}
T\approx 1.9\frac{\hbar\kappa}{k_\text{B}}\,,
\end{equation}
in the limiting case of small polarisability and at the point of maximum friction; i.e., the temperature scales in the same way as for an atom inside the cavity. The numeric factor in the preceding equation depends on $\zeta$ and is larger for $\lvert\zeta\rvert\sim 1$.}
\section{Amplified optomechanics in a ring cavity}\label{sec:TMM:AmplifiedOptomechanics}
In the limit of strong scatterers, friction forces in standing-wave cavities become increasingly position-dependent (cf.~\Sref{sec:TMM:Comparison}), which limits the overall, averaged cooling efficiency. This can be overcome by using ring cavities~\cite{Gangl2000a,Elsasser2003,Kruse2003,Nagy2006,Slama2007,Hemmerling2010,Schulze2010,Niedenzu2010} where the translational symmetry guarantees position-independent forces. On the other hand, ring cavities are usually much larger and of lower $Q$-factor than their standing-wave counterparts. Using a gain medium inside a ring cavity has been proposed~\cite{Vuletic2001a,Salzburger2006} {to offset these losses}, allowing one {to effectively `convert'} a low-$Q$ cavity into a high-$Q$ one, and {thus to} increase the effective optomechanical interaction by orders of magnitude. This {same concept} has also been discussed in theoretical proposals investigating the use of optical parametric amplifiers in standard optomechanical systems~\cite{Huang2009}, or nonlinear media inside cavities~\cite{Kumar2010} as a tool to control the dynamics of a micromechanical oscillator. {A further application of ring cavities is in the investigation of collective atomic recoil lasing (CARL)~\cite{Bonifacio1994}, which exploits the spontaneous self-organisation of an atomic ensemble within a ring cavity, induced by a strong pump beam, to amplify a probe beam through Doppler-shifted reflection of the pump. The gain medium is in this case the atomic ensemble itself.}
\par
{Let us now consider a different system that shares several features with the above mechanisms. In particular, we consider a scatterer inside a ring cavity that includes a gain medium, spatially separated from the scatterer. An isolator is also included in the ring cavity, in such a way as to prevent the pump beam from circulating in the cavity and being amplified; this ensures that the intensity of the field surrounding the scatterer is always low and thereby circumvents any problems caused by atomic saturation or mirror burning. The Doppler-shifted reflection of the pump from the scatterer is, on the other hand, allowed to circulate, and its amplification in turn enhances the velocity-dependent forces acting on the scatterer.}\\
{In such a situation}, one is able to take advantage of properties inherent to the ring cavity system, such as the fact that the forces acting on the particle do not exhibit any sub-wavelength spatial modulation; this is due to the translational symmetry present in the system~\cite{Gangl2000a}. Moreover, modest amplification allows one to use optical fibres to form the ring cavity, opening the door towards increasing the optical length of such cavities. It is necessary to have amplification in such cases to compensate for the losses introduced when coupling to a fibre; such losses would otherwise limit the quality factor of the cavity. The optomechanical force is, {as we will see and} in the parameter domain of interest, linearly dependent on the cavity length; lengthening the cavity thus provides further enhancement of the interaction.
\par
This section is structured as follows. We shall first introduce the physical model, which we proceed to solve using the transfer matrix method to obtain the friction force and momentum diffusion acting on the particle. In the good-cavity limit, \Sref{sec:GoodCavity}, simple expressions for these quantities can be obtained, yielding further insight into the system and allowing us to draw some parallels with traditional cavity cooling. {In this limit, our model becomes equivalent to one based on a standard master equation approach}~\cite{Gangl2000a} {as outlined} in \Sref{sec:Semiclassical}. Realistic numerical values for the various parameters are then used in \Sref{sec:Results} to explore the efficiency and limits of the cooling mechanism. Finally, we will conclude by mentioning some possible extensions to the scheme. The work in this section will be published in the J.\ Mod.\ Opt.\ topical issue on \emph{New cooling mechanisms for atoms and molecules}.
\begin{figure}[t]
\centering
\includegraphics[width=\linewidth]{Diagrams/amplified_optomechanics/Model}
\caption[Model for a polarisable particle interacting with the field in a unidirectional ring cavity]{(a)~Physical schematic of a polarisable particle in a unidirectional ring cavity, showing the input field $\hat{B}_\text{in}$. (b)~Equivalent `transfer matrix'--style (unfolded) model; the particle is drawn on both sides of this schematic to illustrate the recursive nature of the cavity. The various components are defined in \Sref{sec:Model}.}
\label{fig:Model}
\end{figure}
\subsection{General expressions and equilibrium behaviour}\label{sec:Model}
The mathematical model of the ring cavity system, schematically drawn in \fref{fig:Model}(a), is shown in \fref{fig:Model}(b). {A particle, characterised by its polarisability $\zeta$, is in a ring cavity of round-trip length $L$. $C_{1,3}$ are the couplers, between which lies the particle, that terminate the fibre-based cavity, and $C_2$ is the input coupler that injects the pump beam into the cavity. The amplifier, having a gain $\gamma\geq 1$, is assumed to also function as an optical isolator. $r_i$ and $t_i=r_i+1$ are the (amplitude) transmission and reflection coefficients of coupler $C_i$ ($i=1,2,3$). All the ``$\hat{E}$'' modes are noise modes introduced by specific elements. $\epsilon_\text{A}$, the amplitude of $\hat{E}_\text{A}$ introduced into the system, depends on $\gamma$; similarly, $\epsilon$ depends on $\zeta$. The values of $l_1>l_\text{A}>l_2>l_3>0$ are not important.} {One of the travelling wave modes of the cavity is pumped} by light with wavenumber $k_0${, but is prevented from circulating inside the cavity. This avoids resonant enhancement of the pumped mode in the cavity and thus avoids saturating the particle. The backscattered counterpropagating mode, on the other hand, is amplified on every round trip by a factor $\gamma$ by means of the optical amplifier.} The TMM is used to self-consistently solve for the four field amplitudes {at every point in the cavity in the presence of the pump field and the noise modes introduced by the coupler losses and by the amplifier. Note that in the limit where the amplifier is compensating for the ring cavity losses, the amplifier noise is also comparable to the loss-induced noise and must therefore be} taken into account in our model.
\par
Using the notation in \fref{fig:Model}(a) we can relate the expectation values of the amplitudes of the {two input and two scattered} field modes interacting with the particle {in a one-dimensional scheme}, {$A(k)=\langle\hat{A}(k)\rangle$, $B(k)=\langle\hat{B}(k)\rangle$, $C(k)=\langle\hat{C}(k)\rangle$, and $D(k)=\langle\hat{D}(k)\rangle$, to $B_\text{in}(k)=\langle\hat{B}_\text{in}(k)\rangle$} by means of the relations
\begin{equation}
\label{eq:BasicRelations}
B = r_2t_3e^{ik_0l_2}B_\text{in}\,,\ C=\alpha A\,,\ \text{and}\ \begin{pmatrix}A\\B\end{pmatrix}=\hat{M}\begin{pmatrix}C\\D\end{pmatrix}\,,
\end{equation}
where {$\alpha(k)=t_1t_2t_3\gamma\exp(ikL)$} is the factor multiplied to the field amplitude every round trip. {In the preceding equations, as well as in the following, we do not write the index $k$ for simplicity of presentation. The operators $\hat{A}(k)$, etc., denote the annihilation operators of the various field modes. The three equations \erefs{eq:BasicRelations}} have a readily apparent physical significance---respectively, they correspond to: the propagation of $\hat{B}_\text{in}$ to reach the particle; the feeding back of $\hat{A}$ to $\hat{C}$ through the ring cavity; and the usual transfer matrix relation for a particle interacting with the four fields surrounding it. The first two of these relations are substituted into the third, which subsequently simplifies to
\begin{equation}
\label{eq:BaseRelation}
\begin{pmatrix}A\\B_\text{in}\end{pmatrix}=\left(\begin{bmatrix}1&0\\0&r_2t_3e^{\i k_0l_2}\end{bmatrix}-\hat{M}\begin{bmatrix}\alpha&0\\0&0\end{bmatrix}\right)^{-1}\hat{M}\begin{pmatrix}0\\D\end{pmatrix}
\end{equation}
If we assume far off-resonant operation, i.e., $\partial\zeta/\partial k=0$, the velocity-dependent transfer matrix $\hat{M}$ can be written as, cf.~\eref{eq:TMMFirstordervExplicit},
\begin{equation}
\begin{bmatrix}1+\i\zeta&\i\zeta-2\i\zeta\tfrac{v}{c}+2\i k_0\zeta\tfrac{v}{c}\partial_k\\-\i\zeta-2\i\zeta\tfrac{v}{c}+2\i k_0\zeta\tfrac{v}{c}\partial_k&1-\i\zeta\end{bmatrix}\,.
\end{equation}
Note that the partial derivative $\partial_k$ acts not only on $\alpha(k)$ but also on the field mode amplitudes it precedes. \eref{eq:BaseRelation} can be inverted in closed form to first order in $v/c$, similarly to \sref{sec:TMM:General}, and can thus be used to find $\mathcal{A}=\sqrt{2\epsilon_0\sigma_\text{L}/(\hbar k_0)}\int A(k)\mathrm{d} k$, $\mathcal{B}=\sqrt{2\epsilon_0\sigma_\text{L}/(\hbar k_0)}\int B(k)\mathrm{d} k$, $\mathcal{C}=\sqrt{2\epsilon_0\sigma_\text{L}/(\hbar k_0)}\int C(k)\mathrm{d} k$, and $\mathcal{D}=\sqrt{2\epsilon_0\sigma_\text{L}/(\hbar k_0)}\int D(k)\mathrm{d} k$, {where the normalisation is with respect to} the pump beam mode area $\sigma_\text{L}$ and {where a monochromatic pump is assumed}: $B_\text{in}(k)=B_0\delta(k-k_0)${. Here, $\lvert\mathcal{A}\rvert^2$, $\lvert\mathcal{B}\rvert^2$, etc., are the photon currents in units of photons per second}. The expectation value of the force acting on the scatterer is then given by \eref{eq:TMM:MSTForce}:
\begin{equation}
\label{eq:RawForce}
\hbar k_0\bigl(\lvert\mathcal{A}\rvert^2+\lvert\mathcal{B}\rvert^2-\lvert\mathcal{C}\rvert^2-\lvert\mathcal{D}\rvert^2\bigr)\,.
\end{equation}
The values of $\mathcal{A}$, $\mathcal{B}$, etc., from the solution of \eref{eq:BaseRelation} are then substituted in \eref{eq:RawForce}, which {we evaluate to} first order in $v/c$, in terms of $B_0$. After some algebra, we obtain the first main result of this section---the friction force acting on the particle:
\begin{equation}
\label{eq:TMMFriction}
\textsl{\textsf{F}}=-8\hbar k_0^2\tfrac{v}{c}\re{\frac{\bigl(1-\alpha^\ast\bigr)\zeta\re{\zeta}+\i\alpha^\ast\zeta\lvert\zeta\rvert^2}{1-\alpha-\i\zeta}\frac{\partial\alpha}{\partial k}}\frac{\bigl|r_2t_3B_0\bigr|^2}{\bigl|1-\alpha-\i\zeta\bigr|^2}\,.
\end{equation}
By extending the TMM appropriately, one can keep track of the various noise modes interacting with the system. {\erefs{eq:BasicRelations} then become}
\begin{subequations}
\begin{equation}
\hat{A}=\frac{\i\zeta}{1-\i\zeta}\hat{B}+\frac{1}{1-\i\zeta}\hat{C}+\epsilon\hat{E}\,,
\end{equation}
\begin{equation}
\hat{B}=r_2t_3e^{\i k_0l_2}\hat{B}_\text{in}+t_2t_3e^{\i k_0l_\text{A}}\hat{E}_\text{isolator}+r_3e^{\i k_0l_3}\hat{E}_3^\prime\,,\
\end{equation}
\begin{equation}
\hat{C}=\alpha\hat{A}+r_1e^{\i k_0(L-l_1)}\hat{E}_1+t_1r_2\gamma e^{\i k_0(L-l_2)}\hat{E}_2+t_1t_2r_3\gamma e^{\i k_0(L-l_3)}\hat{E}_3+t_1\epsilon_\text{A}e^{\i k_0(L-l_\text{A})}\hat{E}_\text{A}\,, \text{ and}
\end{equation}
\begin{equation}
\hat{D}=\frac{1}{1-\i\zeta}\hat{B}+\frac{\i\zeta}{1-\i\zeta}\hat{C}+\epsilon\hat{E}\,,
\end{equation}
\end{subequations}
with $\epsilon=\sqrt{1-\bigl(1+\lvert\zeta\rvert\bigr)/\lvert 1-\i\zeta\rvert^2}$ (see \aref{sec:TMM:AppDiffusion}) and $\epsilon_\text{A}=\sqrt{1-1/\lvert\gamma\rvert^2}$~\cite{Gardiner2004}. These equations can be solved simultaneously for $\hat{A}$, $\hat{B}$, $\hat{C}$, and $\hat{D}$, and the solution used to evaluate the momentum diffusion constant, $\textsl{\textsf{D}}$, {defined as} the two-time autocorrelation function of the force operator (cf. \sref{sec:MSTDiffusion}), keeping in mind that most of the noise modes, as well as $\hat{B}_\text{in}$, obey the commutation relation $\bigl[\hat{E}(t),\hat{E}^\dagger(t^\prime)\bigr]=\hbar k_0/(2\epsilon_0 \sigma_\text{L})\,\delta(t-t^\prime)$. The sole exception is the noise introduced by the amplifier, $\hat{E}_\text{A}$, for which $\bigl[\hat{E}_\text{A}(t),\hat{E}_\text{A}^\dagger(t^\prime)\bigr]=-\hbar k_0/(2\epsilon_0 \sigma_\text{L})\,\delta(t-t^\prime)$; this is due to the model of the amplifier as a negative temperature heat-bath, whereby the creation and annihilation operators effectively switch r\^oles. Further discussion of this model can be found in Ref.~\cite[\textsection 7.2]{Gardiner2004}. All the noise modes are independent from one another and from $\hat{B}_\text{in}$, which simplifies the expressions considerably.
\subsubsection{The good-cavity limit as a simplified case}\label{sec:GoodCavity}
{Before discussing the result of \Sref{sec:Model}, we shall} make several approximations to obtain a transparent set of equations to briefly explore the equilibrium behaviour of the scatterer {and to compare with a standard master equation approach}. In particular, $\zeta$ is assumed to be real, which is tantamount to assuming that the scatterer suffers no optical absorption or, if it is an atom, that it is pumped far off-resonance. Moreover, the cavity is assumed to be very good ($\lvert t_{1,2,3}\rvert\to 1$) {and thus no gain medium is introduced in the cavity ($\gamma=1$)}. With these simplifications, \eref{eq:TMMFriction} reduces to
\begin{align}
\label{eq:TMMFrictionSimple}
\textsl{\textsf{F}}&\approx -8\hbar k_0^2\tfrac{v}{c}\frac{\zeta^2}{\lvert 1-\alpha\rvert^4}\re{\bigl(1-\alpha^\ast\bigr)^2\frac{\partial\alpha}{\partial k}}\bigl|r_2B_0\bigr|^2\nonumber\\
&\approx 16\hbar k_0^2\zeta^2 v\frac{\kappa\Delta_\text{C}}{\bigl(\Delta_\text{C}^2+\kappa^2\bigr)^2}\frac{1}{\tau}\bigl|r_2B_0\bigr|^2\,.
\end{align}
In the preceding equations, $\Delta_\text{C}$ is the detuning {of the pump} from cavity resonance, $\kappa$ is the HWHM cavity {linewidth,
\begin{equation}
\kappa=\frac{1}{\tau}\frac{1-\lvert t_1t_2t_3\rvert\gamma}{\sqrt{\lvert t_1t_2t_3\rvert\gamma}}\,,
\end{equation}
for $\lvert\zeta\rvert\ll 1$,} {and} $\tau = L/c$ is the round-trip time{.} Using the same approximations {as for \eref{eq:TMMFrictionSimple}}, we also obtain the diffusion constant
\begin{equation}
\label{eq:TMMDiffusionSimple}
\textsl{\textsf{D}}\approx 8\hbar^2 k_0^2\zeta^2\frac{\kappa}{\Delta_\text{C}^2+\kappa^2}\frac{1}{\tau}\bigl|r_2B_0\bigr|^2\,.
\end{equation}
{Note that} $\gamma=1$ {here and therefore} $\hat{E}_\text{A}$ does not contribute to the diffusion constant{.} {\erefs{eq:TMMFrictionSimple} and~(\ref{eq:TMMDiffusionSimple}) hold for the case where $\Delta_\text{C}/\kappa$ is not too large. The cavity can be fully described by means of $\kappa$ and the finesse $\mathcal{F}=\pi c/(2L\kappa)$. Let us now set $\Delta_\text{C}=-\kappa$ in \erefs{eq:TMMFrictionSimple} and~(\ref{eq:TMMDiffusionSimple}), whereby
\begin{equation}
\textsl{\textsf{F}}=-\tfrac{8}{\pi}\hbar k_0^2\zeta^2 v\frac{\mathcal{F}}{\kappa}\bigl|r_2B_0\bigr|^2\,,\text{ and }
\textsl{\textsf{D}}=\tfrac{8}{\pi}\hbar^2 k_0^2\zeta^2\mathcal{F}\bigl|r_2B_0\bigr|^2\,.
\end{equation}
These two expressions have a readily-apparent physical significance; at a constant finesse, decreasing the cavity linewidth is equivalent to making the cavity longer, whereupon the retardation effects that underlie this cooling mechanism lead to a stronger friction force. At the same time, this has no effect on the intracavity field strength and therefore does not affect the diffusion. On the other hand, improving the cavity finesse by reducing losses at the couplers increases the intracavity intensity, thereby increasing both the friction force and the momentum diffusion.}
\par
Using the above results, we obtain, for $\Delta_\text{C}<0$,
\begin{equation}
\label{eq:TMMTemperatureSimple}
T_\text{A}\approx \frac{\hbar}{k_\text{B}}\biggl(\frac{\lvert\Delta_\text{C}\rvert}{\kappa}+\frac{\kappa}{\lvert\Delta_\text{C}\rvert}\biggr)\frac{\kappa}{2}\geq\frac{\hbar}{k_\text{B}}\kappa\,,
\end{equation}
with the minimum temperature occurring at $\Delta_\text{C}=-\kappa$. One notes that this expression is identical to the corresponding one for standard cavity-mediated cooling~\cite{Horak1997}. Like the corresponding results in \sref{sec:MovingAtom} and \sref{sec:Mirrorcooling}, it can be interpreted in a similar light as the Doppler temperature, albeit with the energy dissipation process shifted from the decay of the atomic excited state to the decay of the cavity field{.}
\par
A particular feature to note in all the preceding expressions is that they are not spatial averages over the position of the particle, but they do not depend on this position either. As a result of this, the force, momentum diffusion and equilibrium temperature do not in any way depend on the position of the particle along the cavity field in a 1D model. The issue of sub-wavelength modulation of the friction force is a major limitation of cooling methods based on intracavity standing fields, in particular mirror-mediated cooling (\sref{sec:TMM:GeneralMMC}) and cavity-mediated cooling (\Sref{sec:TMM:Comparison}).
\subsubsection{Comparison with a semiclassical model}\label{sec:Semiclassical}
In the good-cavity limit {and without gain our TMM model is equivalent to a standard master equation approach} with the Hamiltonian
\begin{align}
\label{eq:Hamiltonian}
\hat{H}=&-\hbar\Delta_\text{a}\hat{\sigma}^+\hat{\sigma}^- -\hbar\Delta_\text{C}\hat{a}_\text{C}^\dagger \hat{a}_\text{C}\nonumber\\
&+\hbar g\bigl(\hat{a}_\text{C}^\dagger\hat{\sigma}^- e^{\i k_0x}+\hat{\sigma}^+\hat{a}_\text{C}e^{-\i k_0x}\bigr)+\hbar g\bigl(a_\text{P}^\ast\hat{\sigma}^- e^{-\i k_0x}+\hat{\sigma}^+a_\text{P}e^{\i k_0x}\bigr)\,,
\end{align}
and the Liouvillian terms
\begin{equation}
\label{eq:Liouvillian}
\mathcal{L}\hat{\rho}=-\Gamma\bigl(\hat{\sigma}^+\hat{\sigma}^-\hat{\rho}-2\hat{\sigma}^-\hat{\rho}\hat{\sigma}^++\hat{\rho}\hat{\sigma}^+\hat{\sigma}^-\bigr)-\kappa\bigl(\hat{a}_\text{C}^\dagger \hat{a}_\text{C}\hat{\rho}-2\hat{a}_\text{C}\hat{\rho} \hat{a}_\text{C}^\dagger+\hat{\rho} \hat{a}_\text{C}^\dagger \hat{a}_\text{C}\bigr)\,,
\end{equation}
as adapted from Ref.~\cite{Gangl2000a} and modified for a unidirectional cavity where only the unpumped mode is allowed to circulate{. Here,} $\hat{\rho}$ is the density matrix of the system, $g$ the atom--field coupling strength, $\hat{a}_\text{C}$ the annihilation operator of the cavity field, $\hat{\sigma}^+$ the atomic dipole raising operator, $\Delta_\text{a}$ the detuning from atomic resonance, $\Gamma$ the atomic upper state HWHM linewidth, and $x$ the coordinate of the atom inside the cavity. The pump field is assumed to be unperturbed by its interaction with the atom, and in the above is replaced by a c-number, $a_\text{P}$. {Calculating the friction force from this model leads again to \eref{eq:TMMFrictionSimple}, thus confirming our TMM results by a more standard technique. The advantage of the TMM approach lies in the simplicity and generality of expressions such as \eref{eq:RawForce}, and the ease with which more optical elements can be introduced into the system. As shown above, the momentum diffusion coefficient is easily calculated from the TMM.}
\begin{figure}[t]
\centering
\includegraphics[angle=-90,width=1.5\figwidth]{Diagrams/amplified_optomechanics/Friction}
\caption[Cooling rate for $^{85}$Rb pumped $-10\Gamma$ from resonance inside a ring cavity]{Cooling rate $-(\text{d}v/\text{d}t)/v$ for $^{85}$Rb pumped $-10\Gamma$ from D$_2$ resonance inside a ring cavity with a round-trip length $L=300$\,m, for two different values of the amplifier gain $\gamma$. Note that the curve for {$\gamma=1$}, as drawn, is scaled \emph{up} by a factor of $10$. The cavity waist is taken to be $10$\,$\upmu$m. ($\lvert t_1\rvert^2=\lvert t_3\rvert^2=0.5$, $\lvert t_2\rvert^2=0.99$, {$B_0$ is chosen such as to give an} atomic saturation $s=0.1$.)}
\label{fig:Friction}
\end{figure}
\begin{figure}[t]
\centering
\includegraphics[angle=-90,width=1.5\figwidth]{Diagrams/amplified_optomechanics/Temperature}
\caption[Equilibrium temperature predicted by the transfer matrix model for two values of the amplifier gain $\gamma$]{Equilibrium temperature predicted by the transfer matrix model for two values of the amplifier gain $\gamma$. The Doppler temperature for $^{85}$Rb is also indicated. The horizontal axis differs from that in \fref{fig:Friction} mainly because the temperature is only well-defined for regions where the friction force promotes cooling. (Parameters are as in \fref{fig:Friction}.)}
\label{fig:Temperature}
\end{figure}
\subsection{Numerical results and discussion}\label{sec:Results}
We can use the conversion factor $\lvert B_0\rvert^2=P/(\hbar k_0c)$, where $P$ is the power of the input beam, to evaluate the above equations numerically in a physically meaningful way. Specifically, the particle is now assumed to be a (two--level) \textsuperscript{85}Rb atom, pumped $-10\Gamma$ from D$_2$ resonance, {where} $\Gamma=2\pi\times 3.03$\,MHz {is} the HWHM linewidth of this same transition at a wavelength of ca.~$780$\,nm; {because} the detuning is much larger than the linewidth, {we simplify the calculations} by setting $\partial\zeta/\partial\omega=0$. The beam waist where the particle interacts with the field is taken to be $10$\,$\upmu$m. With the parameters in \fref{fig:Friction}, the power is reduced by a factor of $1/\lvert t_1t_2t_3\rvert^2=4.04$ with each round-trip, in the presence of no gain in the amplifier. We shall compare this case to the low-gain case; the gain of the amplifier we consider is constrained to be small enough that $\lvert\alpha\rvert^2=\lvert t_1t_2t_3\gamma|^2<1$. Under these conditions, there is no {exponential} build-up of intensity inside the cavity and the system is stable. A cavity with a large enough gain that $\lvert\alpha\rvert^2>1$ would effectively be a laser cavity. Such a system {would have} no stable state in our model, since we assume that the gain medium is not depleted, and will therefore not be considered further in the following.
\par
\fref{fig:Friction} shows the friction force acting on the particle, and \fref{fig:Temperature} the equilibrium temperature, as the length of the cavity is tuned on the scale of one wavelength. In each of these two figures two cases are shown, one representing no gain in the amplifier ({$\gamma=1$}) and one representing a low-gain amplifier ($\gamma=1.75$); note that in both cases the condition $\lvert\alpha\rvert^2<1$ is satisfied.
\par
In order to provide a fair comparison between these two cases, we choose the pump amplitude $B_\text{in}$ such that the saturation of the particle is the same in the two cases. This ensures that any difference in cooling performance is not due to a simple increase in intensity. Since the TMM as presented here is based on a \emph{linear} model of the particle, our results presented above are only valid in the limit of saturation parameter much smaller than $1$. Thus, as a basis for the numerical comparisons between the two different cases, we choose to set the saturation parameter to $0.1$. \fref{fig:Friction} shows that under these conditions the amplified system leads to a significant, approximately $25$-fold, enhancement of the maximum friction force. This can therefore be attributed unambiguously to the effective enhancement of the cavity $Q$-factor by the amplifier.
\par
However, for the parameters considered here, in particular for small particle polarisability $\zeta$ and for $\lvert\alpha\rvert^2<1$, the counterpropagating mode intensity is much smaller than that of the pumped mode, even if the former is amplified. Thus, the intracavity field is always dominated by the pump beam, whereas the friction force is mostly dependent on the Doppler-shifted reflection of the pump from the particle. Specifically, for the parameters used above we find that the total field intensity changes by less than 1\% when the gain is increased from 1 to 1.75. Hence, similar results to those of \fref{fig:Friction} are obtained even \emph{without} pump normalisation.
\par
The steady-state temperature, obtained by the ratio of diffusion and friction, is shown in \fref{fig:Temperature} for the same parameters as above. We observe that the broader resonance in the friction as a function of cavity detuning (i.e., of cavity length), shown in \fref{fig:Friction}, also leads to a wider range of lower temperatures compared to the amplified case. However, as expected, within the narrower resonance of the amplified system where the friction is significantly enhanced, the stationary temperature is also significantly reduced. We see that while the maximum friction force is increased by a factor of $25.4$, the lowest achievable temperature is decreased by a factor of $19.9$ when switching from $\gamma=1$ to $\gamma=1.75$. While the overall cavity intensity is dominated by the pump field, and is therefore hardly affected by the amplifier, the diffusion is actually dominated by the interaction of the weak counterpropagating field with the pump field. This can be seen most clearly by the strong detuning dependence of the analytical expression for $\textsl{\textsf{D}}$ in the good-cavity limit, \eref{eq:TMMDiffusionSimple}. As a consequence, the lowest achievable temperature is improved by a slightly smaller factor than the maximum cooling coefficient. This is consistent with the idea that the amplifier not only increases the cavity lifetime, but also adds a small amount of additional noise into the system. Nevertheless, a strong enhancement of the cooling efficiency is observed in the presence of the amplifier.
\chapter{Three-dimensional scattering with an optical memory}\label{ch:TMM:Scattering}
\epigraph{Homogeneal Rays which flow from several Points of any Object, and fall almost Perpendicularly on any reflecting or refracting Plane or Spherical Surface, shall afterwards diverge from so many other Points, or be Parallel to so many other Lines, or converge to so many other Points, either accurately or without any sensible Error. And the same thing will happen, if the Rays be reflected or refracted successively by two or three or more Plane or spherical Surfaces.\par
[...]\par
Wherever the Rays which come from all the Points of any Object meet again in so many Points after they have been made to converge by Reflexion or Refraction, there they will make a Picture of the Object upon any white Body on which they fall.}{I.\ Newton, \emph{Opticks} (1704)}
The scattering theory presented in the previous chapters can be used to describe a wide variety of one-dimensional, or quasi-one-dimensional, situations involving mobile scatterers and immobile optics. The three-dimensional nature of the electromagnetic field can, however, be exploited to give rise to a different type of retarded dipole-dipole interaction, one mediated not only by the relative phase difference between the successive reflections but also the spreading nature of spherical waves in three dimensions.
\par
I start this chapter with an extension of a self-consistent scattering theory that was used to describe optical binding phenomena~\cite{Depasse1994}. Our treatment essentially identifies the two particles described in Ref.~\cite{Depasse1994} such that the `binding' that takes place really is between a particle and itself, as mediated by a mirror or other delay element. After describing this extension to the theory in \Sref{sec:TMM:Scattering:Theory}, I shall apply it to the mirror-mediated cooling described earlier in this work, both in one dimension (cf.~\sref{sec:TMM:1DMMC}) and in three (\sref{sec:TMM:3DMMC}). I will use these results in \cref{ch:Experimental:Future} to discuss experimentally-accessible configurations for exploring the various mechanisms described in this thesis.
\begin{figure}[t]
\centering
\includegraphics{Diagrams/Retarded_Binding}
\caption[A particle bound to its retarded reflection in a surface]{Polarised by incident radiation $\v{\mathcal{E}}^0(t,\v{r})$, a particle moving with velocity $\v{v}$ is bound to its retarded reflection in a surface, characterised by round-trip time $\tau$ and field propagator $\m{g}$.}
\label{fig:TMM:RetardedBinding}
\end{figure}
\section{Optical self-binding of Rayleigh particles}\label{sec:TMM:Scattering:Theory}
The work in this section is being prepared for publication by James Bateman, AX, and Tim Freegarde. Its basis originated from TF and was subsequently elaborated upon by all three authors. The \emph{Mathematica} code used to solve the equations and verify the solutions was written largely by JB.
\par
We first analyse the retarded classical electrostatic interaction between the induced dipoles of a particle, moving along a path $\v{r}(t)$, and its reflection, as shown in \fref{fig:TMM:RetardedBinding}, following the ideas set forth in Ref.~\cite{Depasse1994}. The unperturbed illuminating field is denoted $\v{\mathcal{E}}^0(t,\v{r})$. A tensor propagator, $\m{g}$, is used to describe the propagation of the scattered field from the particle and back to itself through reflection in the mirror. In one dimension,
\begin{equation}
\label{eq:1DPropagator}
\m{g}(t,\v{r})=-\i e^{\i\omega\tau(t,\v{r})}\mathds{1}\,,
\end{equation}
where the pre-factor $-\i$ accounts for the phase shift upon reflection as well as the Gouy phase shift: in our 1D model we take the particle to be a point-like dipole at the focus of a tightly-focussed beam of mode area $\sigma_\text{L}$. In three dimensions,
\begin{align}
\label{eq:3DPropagator}
\m{g}(t,\mathbf{r})=\frac{\sigma_\text{L}}{R(t)\lambda}e^{\i\omega\tau(t,\v{r})}\Biggl\{&\frac{1-\i kR(t)-\bigl[kR(t)\bigr]^2}{\bigl[kR(t)\bigr]^2}\mathds{1}\nonumber\\
&-\frac{3-3\i kR(t)-\bigl[kR(t)\bigr]^2}{\bigl[kR(t)\bigr]^2}\v{n}(t)\otimes\v{n}(t)\Biggr\}\,.
\end{align}
In the preceding equations, $R(t)=\lVert\v{R}(t)\rVert$, $\v{n}(t)=\v{R}(t)/R(t)$, $\v{R}(t)$ is the position vector of a test particle at time $t$ relative to its image at time $t-\tau[t,\v{r}(t)]$ (see \fref{fig:TMM:RetardedBinding}), and $\otimes$ represents the vector outer product. $\m{g}$ in \eref{eq:3DPropagator} is the tensor Green's function for free-space propagation~\cite{Levine1950,Martin1995} of an electromagnetic wave, normalised to render it dimensionless.
\par
The interaction of the particle with the electric field is assumed to be through the dipole interaction, whereby the particle is described by means of its polarisability $\chi$; this polarisability allows us to define a dimensionless polarisability
\begin{equation}
\zeta=\chi k/(2\sigma_\text{L})\,,
\end{equation}
as before. The form for $\zeta$ we choose is identical for both one- and three-dimensional cases, despite arising from different considerations; in one dimension $\zeta$ is defined in \sref{sec:CoolingMethods:ForceTLA}, whereas in three dimensions we define $\zeta$ through a particular grouping of constants to yield dimensionless $\zeta$ and $\m{g}$. In this chapter we will, for simplicity, assume a real, scalar value for $\chi$. The resulting equations can be subsequently generalised for complex, vector or tensor polarisabilities. Multiple scattering between the particle and the mirror is taken into account self-consistently by solving for the total electric field $\v{\mathcal{E}}[t,\v{r}(t)]$, experienced by the particle at position $\v{r}(t)$ and time $t$:
\begin{equation}
\v{\mathcal{E}} = \v{\mathcal{E}}^0+\zeta\m{g}\cdot\v{\mathcal{E}}\,,
\end{equation}
where all the terms are evaluated at $[t,\v{r}(t)]$. The solution of this equation can be given analytically to lowest order in $\v{v}$ and $\tau$, and for the case when $(\mathds{1}-\zeta\m{g})$ is invertible (i.e., when $\lVert\zeta\m{g}\rVert\ll 1$)~\cite{Depasse1994},
\begin{equation}
\label{eq:ClassicalField}
\v{\mathcal{E}}=\left(\begin{array}{c}
\mathcal{E}_x\\
\mathcal{E}_y\\
\mathcal{E}_z
\end{array}\right)=\Bigl[\mathds{1}-\left(\mathds{1}-\zeta\m{g}\right)^{-1}\cdot\tau\zeta\m{g}\cdot\mathrm{D}_t\Bigr]\bigl(\mathds{1}-\zeta\m{g}\bigr)^{-1}\cdot\v{\mathcal{E}}^0\,,
\end{equation}
where $\mathrm{D}_t=\partial_t+\v{v}\cdot\nabla$ is the total time derivative. All the terms in the preceding two equations are evaluated in a frame rotating with the angular frequency $\omega$, in order to remove the fast time variation, at $[t,\v{r}(t)]$. The non-retarded result of Ref.~\cite{Depasse1994}, applied to the particle and its reflection, is thus modified by the appearance of an additional, time- and velocity-dependent, term. The dipole force experienced by the particle may be obtained as in Ref.~\cite{Depasse1994}:
\begin{equation}
\boldsymbol{\textsl{\textsf{F}}}=\left(\begin{array}{c}
\textsl{\textsf{F}}_x\\
\textsl{\textsf{F}}_y\\
\textsl{\textsf{F}}_z
\end{array}\right)\,;\ \text{where}\ \textsl{\textsf{F}}_i=\tfrac{1}{2}\epsilon_0\re{\chi\sum_j\mathcal{E}_j\partial_i\mathcal{E}_j^\ast}\,,
\end{equation}
with $i,j$ separately representing the three spatial dimensions $x,y,z$. A series expansion in powers of $\lVert\zeta\mathbf{g}\rVert$ reveals the leading terms for a stationary particle to be the dipole force from the unperturbed field, and then the dipole force upon the polarised particle due to the field propagated from the induced polarisation.
\section{Mirror-mediated cooling in one dimension}\label{sec:TMM:1DMMC}
For a one-dimensional geometry with the particle a distance $x$ from a perfect mirror, the incident illumination combines with its reflection to give an electric field $\v{\mathcal{E}}^0(x)=\mathcal{E}_0\hat{\v{y}}\sin(kx)$, and $\tau=2x/c$; $\hat{\v{y}}$ is a unit vector in the $y$ direction. The force in the $x$ direction upon the moving particle is therefore
\begin{align}
\label{eq:Scattering1DMMCForce}
\textsl{\textsf{F}}_x=\frac{1}{4}\epsilon_0\chi k\mathcal{E}_0^2\biggl\{&\sin(2kx)+\frac{\chi k}{\sigma_\text{L}}\Bigl(1-\frac{v}{c}\Bigr)\sin^2(kx)\bigl[4\cos^2(kx)-1\bigr]\nonumber\\
&-\frac{\chi k^2\tau v}{\sigma_\text{L}}\sin(4kx)\biggr\}\,.
\end{align}
The force thus comprises three terms. The first two are the dipole force exerted by the unperturbed field, and a Doppler-shifted optical binding force between the particle and its reflection: the Doppler shift here changes the wavelengths of the Fourier field components and hence the gradient of the field formed by their superposition. The third term, which depends upon the particle velocity, the electric field propagator and the round-trip retardation time, is the velocity-dependent force, and dominates the velocity-dependent part of the second term when the distance from the mirror is many wavelengths. When the sign of this component is such as to oppose the particle velocity, cooling ensues. This third term is qualitatively and quantitatively identical to the equivalent terms derived from a semiclassical approach (\Sref{sec:CoolingMethods:MMC:Friction}) and using a one-dimensional scattering theory (\Sref{sec:MovingAtom}).
\section{Self-binding: mirror-mediated cooling in three dimensions}\label{sec:TMM:3DMMC}
\begin{figure}[t]
\centering
\subfigure[Laboratory frame]{
\includegraphics{Diagrams/TwoFrames_Lab}
}\hspace{1cm}
\subfigure[Particle frame]{
\includegraphics{Diagrams/TwoFrames_Particle}
}
\caption[Retarded binding of a normally-illuminated particle to its own reflection]{Retarded binding of a normally-illuminated particle, moving with velocity $v$, to its own reflection, depicted (a)~in the laboratory frame, in which the image lags behind; (b)~in the rest frame of particle, whereby the `wake' trails behind.}
\label{fig:SelfBinding}
\end{figure}
When a particle or ensemble is strongly coupled to its reflection, we must consider higher-order terms in the expansion of \eref{eq:ClassicalField}, such as the interaction between the propagated particle polarisation and the further polarisation which that induces. This corresponds to the optical binding~\cite{Burns1989,Metzger2006a} of the particle to its reflection, due to the tweezing of the particle by light that it has focussed, as shown in \fref{fig:SelfBinding}(a). The finite time taken for light from the particle to return via the mirror causes the reflected image to trail behind the moving particle, providing a component of the binding force in the direction of the particle velocity and therefore a transverse force even when the geometry shows translational symmetry. \fref{fig:SelfBinding}(b) shows the same geometry in the frame of the particle: it is now the inclination of the transformed incident illumination that causes the focussed `wake' again to lie behind the particle. The sign of the frictional component again alternates with distance from the mirror, but it does so asymmetrically because the apparent field strength described by \eref{eq:ClassicalField} is also modulated. The result is a non-zero force, when averaging over the $x$ coordinate,
\begin{equation}
\left<\left(\begin{array}{c}
\textsl{\textsf{F}}_x\\
\textsl{\textsf{F}}_y\\
\textsl{\textsf{F}}_z
\end{array}\right)\right>=(-v)\frac{\epsilon_0\mathcal{E}_0^2k^2\chi^2}{128\pi cx^2}\left(\begin{array}{c}
1\\
3\\
3
\end{array}\right)\,,
\end{equation}
under circularly-polarised illumination and taking into account the near-field effects in $\m{g}$. This force becomes comparable with the amplitude of the position-dependent force when $x\lesssim\lambda$, and could therefore be particularly significant for refractive nanoparticles in, for example, colloidal photonic crystals.
\chapter{Atom--field interactions}\label{ch:CoolingMethods:AFInt}
\epigraph{[...] [T]he semiclassical theory, when extended to take into account both the effect of the field on the molecules and the effect of the molecules on the field, reproduces almost quantitatively the same laws of energy exchange and coherence properties as the quantised field theory, even in the limit of one or a few quanta in the field mode.}{E.\ T.\ Jaynes and F.\ W.\ Cummings, Proceedings of the IEEE \textbf{51}, 89 (1963)}
I begin this chapter with a very brief review of atomic structure, covering the fine and hyperfine structures as well as the magnetic sublevels within the hyperfine manifold; these ideas will be used later to discuss trapping and cooling of atoms. The following sections describe the density matrix approach and build up to a derivation of the Optical Bloch Equations. These equations are the tools necessary to examine the interaction between an atom and the electromagnetic field and, ultimately, to derive an expression for the forces acting on an atom, as parametrised by the polarisability of that atom. The chapter continues with a note on the fluctuation--dissipation theorem and shows how the calculation of the full force acting on the atom allows the prediction of the equilibrium temperature a population of such atoms will tend to, and concludes with a short discussion on multi-level atoms. The reader is referred to Refs.~\cite{CohenTannoudji1978a,CohenTannoudji1978b,Shore1990a,Shore1990b,Woodgate2000,CohenTannoudji2004,Foot2005}, and references therein, for a more in-depth and complete treatment of certain parts of this chapter.
\section{Atomic structure}\label{sec:CoolingMethods:AS}
It has been known for centuries that the spectrum of the light emitted by excited atomic gases is made up of a finite, and usually large, number of distinct spectral lines. These lines are not randomly distributed but can be divided into various closely-spaced groups. Each of these groups can be thought of as being due to the coarse structure of the energy levels of the atom; i.e., each group of lines in the coarse structure originates from a particular transition between two (electron) energy levels. Each energy level has an associated orbital angular momentum, $\qv{L}$.
The total angular momentum $\qv{J}$ of an electron includes a contribution from both $\qv{L}$ and the electron's intrinsic spin angular momentum $\qv{S}$:
\begin{equation}
\qv{J}=\qv{L}+\qv{S}\,,
\end{equation}
whereby the corresponding quantum number $J$ is an integer or half-integer in the range $\lvert L-S\rvert\leq J\leq L+S$. This is called the $LS$ coupling scheme~\cite{Woodgate2000}. We adopt the convention that a quantised momentum vector $\qv{v}$ has magnitude $\sqrt{v(v+1)}\hbar$. The different possible values of $J$ give rise to the fine structure in the spectrum.
\par
Drilling down further, each line in the fine structure spectrum can be subdivided. This arises from the coupling, or interaction, between the total angular momentum of the electron and the angular momentum $\qv{I}$ of the nucleus. We can then define an atomic angular momentum $\qv{F}$ as~\cite{Woodgate2000}
\begin{equation}
\qv{F}=\qv{J}+\qv{I}\,,
\end{equation}
where $F$ again takes any integer or half-integer value in the range $\lvert J-I\rvert\leq F\leq J+I$. Finally, we denote the quantisation axis of our system as $z$, whereby the projection of $\qv{F}$ along $z$ is denoted $F_z\equiv m_F\hbar$, with $-F\leq m_F\leq F$, $m_F$ again being an integer or half-integer. In general, $F$ can take several values for each of the fine structure levels; this gives rise to the hyperfine structure in the observed spectrum.\\
We will temporarily constrain ourselves to the transition of perhaps largest experimental interest in cold atoms, the $L=0\rightarrow L=1$ transition in alkali atoms (the so-called D line), where $S=1/2$. First of all, we note that the ground state ($L=0$) has only one fine structure level ($J=1/2$) whereas the excited state ($L=1$) has two: $J=1/2$ and $J=3/2$. The D line is therefore composed of two separate lines: the D$_1$ line ($^2$S$_{1/2}\rightarrow$ $^2$P$_{1/2}$) and the D$_2$ line ($^2$S$_{1/2}\rightarrow$ $^2$P$_{3/2}$). In the preceding sentence we used spectroscopic notation, whereby the levels with $L=0,1,\dots$ are denoted by $\text{S},\text{P},\dots$, and where the superscript is equal to $2S+1$. The Hamiltonian that describes the hyperfine structure for the D line transitions is, to lowest order~\cite{Arimondo1977},
\begin{equation}
\hat{H}_\text{hfs}=A_\text{hfs}\qv{I}\cdot\qv{J}+B_\text{hfs}\frac{3(\qv{I}\cdot\qv{J})^2+\tfrac{3}{2}(\qv{I}\cdot\qv{J})-I(I+1)J(J+1)}{2I(2I-1)J(2J-1)}\,,
\end{equation}
with $A_\text{hfs}$ being the magnetic dipole constant and $B_\text{hfs}$ the electric quadrupole constant. These are experimentally determined and can be found in standard references, such as Ref.~\cite{Steck2008} for $^{85}$Rb. The two terms in the preceding equation are due to the magnetic dipole and electric quadrupole interaction between $\qv{I}$ and $\qv{J}$, respectively. Higher-order multipole components were ignored in the above expression. The hyperfine levels subsequently incur an energy shift
\begin{equation}
\Delta E_\text{hfs}=\tfrac{1}{2}A_\text{hfs}K+B_\text{hfs}\frac{\tfrac{3}{2}K(K+1)-2I(I+1)J(J+1)}{4I(2I-1)J(2J-1)}\,,
\end{equation}
where we have defined $K=F(F+1)-I(I+1)-J(J+1)$ for convenience.
\par
In the presence of a static external magnetic field $\v{\mathcal{B}}$, it is convenient to define the quantisation axis $z$ as being aligned with the applied field. The different $m_F$ levels (magnetic sublevels) undergo a shift depending on $\mathcal{B}=\lvert\v{\mathcal{B}}\rvert$. The effect on the atom of its interaction with $\v{\mathcal{B}}$ is qualitatively different for various regimes, depending on the size of the energy shifts compared to the splitting between the levels. We will only be concerned with the simplest of these cases, where the energy shift is small compared to the splitting between the hyperfine levels. In such cases, $F$ is a good quantum number (i.e., its value is well-defined as the system evolves in time) and the interaction Hamiltonian can be written
\begin{equation}
\label{eq:MagHamiltonian}
\hat{H}_\text{B}=\mu_\text{B}g_F\v{F}\cdot\v{\mathcal{B}}\,,
\end{equation}
where $\mu_\text{B}$ is the Bohr magneton and $g_F$ the Land\'e $g$-factor given approximately by\footnote{See Ref.~\cite{Woodgate2000,Steck2008}; we approximate $g_S\approx 2$ in their notation.}
\begin{equation}
g_F\approx\biggl[\frac{3J(J+1)+S(S+1)-L(L+1)}{2J(J+1)}\biggr]\times\biggl[\frac{F(F+1)-I(I+1)+J(J+1)}{2F(F+1)}\biggr]\,.
\end{equation}
The associated energy shift for each magnetic sublevel is then
\begin{equation}
\Delta E_\text{B}=\mu_\text{B}g_Fm_F\mathcal{B}\,,
\end{equation}
which is therefore different for each value of $m_F$, giving rise to the splitting between the magnetic sublevels known as the Zeeman effect.
\par
A similar effect, called the DC Stark effect, occurs in the presence of DC electric fields. The DC Stark effect is generally less pronounced than the Zeeman effect. To lowest order, the energy shift for each level is~\cite{Woodgate2000}
\begin{equation}
\Delta E_\text{E}\propto\mathcal{E}_z^2\,,
\end{equation}
where $\mathcal{E}_z$ is the electric field along the quantisation axis. It is only the higher order terms, omitted in the above expression, that lift the degeneracy for the different levels.
\section{The density matrix}\label{sec:CoolingMethods:DM}
Any physical Hamiltonian that acts on a system is Hermitian; i.e., $\hat{H}^\dagger=\hat{H}$. As such, it is diagonalisable and therefore affords a basis of \emph{eigenstates}. The system acted on by the Hamiltonian is said to be in a pure state if it can be represented as a weighted sum of the eigenstates of the Hamiltonian; in other words, if the Hamiltonian has eigenstates $\ket{\psi_0},\ket{\psi_1},\dots,\ket{\psi_n}$ ($n=0,1,\dots$), a pure state can be described by
\begin{equation}
\label{eq:WavefunctionDecomposition}
\ket{\Psi}=\sum_i{c_i\ket{\psi_i}}\,,
\end{equation}
with the $c_i$ being complex numbers normalised according to the condition $\sum_i{\lvert c_i\rvert^2}=1$. Often, the above eigenstates are represented as the basis vectors, with
\begin{equation}
\psi_i=\left(\begin{array}{l}
\left.\begin{matrix}
0\\
\vdots\\
0\\
\end{matrix}\right\}i-1\\
1\\
\left.\begin{matrix}
0\\
\vdots\\
0\\
\end{matrix}\right\}n-i\\
\end{array}
\right)
\end{equation}
being the $i$th basis vector in an $n$-dimensional complex Hilbert space. In general this description does not suffice since the state of the system may be a statistical mixture of the different states $\ket{\Psi_i}$, each of which is itself a weighted sum of the eigenstates of the Hamiltonian, as in \eref{eq:WavefunctionDecomposition}:
\begin{equation}
\ket{\Psi_i}=\sum_j{c_j^{(i)}\ket{\psi_j}}\,.
\end{equation}
We can conveniently express the state of such a system using the density matrix formalism~\cite{CohenTannoudji1978a}, in which we define the density matrix
\begin{equation}
\m{\rho}=\sum_i{p_i\ket{\Psi_i}\bra{\Psi_i}}\,,
\end{equation}
where each $p_i$ is interpreted as the probability of the system being in state $\ket{\Psi_i}$; i.e., $\sum_i{p_i}=1$ and $0\leq p_i\leq 1$. We note that the diagonal elements of $\m{\rho}$ are of the form $\m{\rho}_{ii}=\sum_jp_j\big\lvert c_j^{(i)}\big\rvert^2$, where $i$ runs from $1$ to $n$. $\m{\rho}_{ii}$ can be interpreted as the probability of finding the system in the eigenstate $\psi_i$. It is therefore called the population of this state. The off-diagonal elements of this matrix are related to interference effects between pairs of different eigenstates, and are called coherences.\\
The usefulness of the density matrix formalism is readily apparent when calculating expectation values of quantum mechanical operators, for let $\hat{A}$ be such an operator with expectation value $\expt{\hat{A}}$. Then,
\begin{equation}
\label{eq:DensityMatrixTraceRelation}
\expt{\hat{A}}=\Tr\bigl(\m{\rho}\m{A}\bigr)\,,
\end{equation}
with $\m{A}$ being the matrix representation of $\hat{A}$:
\begin{equation}
\m{A}\equiv\bigl[\bra{\psi_i}\hat{A}\ket{\psi_j}\bigr]_{i,j}\,.
\end{equation}
We will adopt this notation throughout: an operator $\hat{A}$ and a matrix $\m{A}$ correspond to the same operation in different representations; the $(i,j)$th matrix element of $\m{A}$ will be denoted $\m{A}_{ij}$. Let us see how \eref{eq:DensityMatrixTraceRelation} arises. For a system in pure state $\ket{\Psi}$, the expectation value of $\hat{A}$ is given by $\expt{\hat{A}}=\bra{\Psi}\hat{A}\ket{\Psi}=\sum_{i,j}c_i^\ast c_j\bra{\psi_i}\hat{A}\ket{\psi_j}=\sum_{i,j}\m{\rho}_{ji}\m{A}_{ij}=\Tr\bigl(\m{\rho}\m{A}\bigr)$, where $\m{\rho}=\ket{\Psi}\bra{\Psi}=\sum_{i,j}c_ic_j^\ast\ket{\psi_i}\bra{\psi_j}$. This relation similarly holds in the case of a statistical mixture of states.
\par
The time-dependent Schr\"odinger equation~\cite{Schrodinger1926} for a Hamiltonian $\hat{H}$ acting on a state $\ket{\psi}$, $\hat{H}\ket{\psi}=\i\hbar\tfrac{\partial}{\partial t}\ket{\psi}$, can be used to show that
\begin{equation}
\label{eq:DensityMatrixTimeEvolution}
\i\hbar\dot{\m{\rho}}=[\m{H},\m{\rho}]\equiv\m{H}\m{\rho}-\m{\rho}\m{H}\,,
\end{equation}
where we use the notation that a dot above a symbol represents its derivative with respect to time:
\begin{align}
\label{eq:QMEPrecursor}
\dot{\m{\rho}}&=\sum_i{p_i\bigl(\tfrac{\partial}{\partial t}\ket{\Psi_i}\bigr)\bra{\Psi_i}}+\sum_i{p_i\ket{\Psi_i}\bigl(\tfrac{\partial}{\partial t}\bra{\Psi_i}\bigr)}\nonumber\\
&=\sum_i{p_i\bigl(\tfrac{1}{\i\hbar}\hat{H}\ket{\Psi_i}\bigr)\bra{\Psi_i}}+\sum_i{p_i\ket{\Psi_i}\bigl(-\tfrac{1}{\i\hbar}\bra{\Psi_i}\hat{H}^\dagger\bigr)}\nonumber\\
&=\frac{1}{\i\hbar}\Biggl[\hat{H}\Biggl(\sum_i{p_i\ket{\Psi_i}\bra{\Psi_i}}\Biggr)-\Biggl(\sum_i{p_i\ket{\Psi_i}\bra{\Psi_i}}\Biggr)\hat{H}\Biggr]\nonumber\\
&=\tfrac{1}{\i\hbar}[\m{H},\m{\rho}]\,,
\end{align}
since $\hat{H}$ is Hermitian.\footnote{It is perhaps interesting to note that \eref{eq:DensityMatrixTimeEvolution} differs in sign from the Heisenberg equation of motion of an operator~\cite[Complement G$_\text{III}$]{CohenTannoudji1978a}. Mathematically, this is due to the operator ordering, of the form $\hat{U}^\dagger\hat{H}\hat{U}$, used to transform from the Schr\"odinger picture, where the state vector is time-dependent and observables correspond to time-independent operators, to the Heisenberg picture, where only the observables are time-dependent.} \eref{eq:DensityMatrixTimeEvolution} is valid for systems that have no dissipation. Incoherent processes---such as the decay of the electromagnetic field inside a cavity, spontaneous emission, etc.---are a result of the system coupling to an infinity of modes, for example the vacuum electromagnetic field. In most cases~\cite[\textsection 5.1.1]{Gardiner2004}, such processes can be most conveniently described not by directly including them in the Hamiltonian but by adding so-called \emph{Lindblad terms} to the preceding relation~\cite{Lindblad1976}.\footnote{Formally, the Lindblad terms arise upon tracing the master equation \eref{eq:QMEPrecursor} over the bath variables.} These terms are represented by a `superoperator' $\hat{\mathcal{L}}$ acting on $\m{\rho}$, giving the full quantum master equation~\cite[\textsection 5.4.2]{Gardiner2004}:
\begin{equation}
\label{eq:QME}
\dot{\m{\rho}}=\tfrac{1}{\i\hbar}[\m{H},\m{\rho}]+\hat{\mathcal{L}}\m{\rho}\,.
\end{equation}
If the system described by $\m{\rho}$ is coupled to a zero-temperature environment with a decay constant $\kappa$, then the Lindblad terms take the generic form
\begin{equation}
\label{eq:LindbladDissipativeSystem}
\hat{\mathcal{L}}\m{\rho} = -\kappa\bigl(\hat{c}^\dagger\hat{c}\m{\rho} + \m{\rho}\hat{c}^\dagger\hat{c} - 2\hat{c}\m{\rho}\hat{c}^\dagger\bigr)\,,
\end{equation}
with $\hat{c}$ being an operator that describes the system coupled to $\m{\rho}$. For example, the decay of a cavity field to which an atom is coupled is described by setting $\kappa$ to be the cavity field decay rate and $\hat{c}$ the annihilation operator, more commonly denoted $\hat{a}$, of the cavity field; $\hat{c}^\dagger$, or $\hat{a}^\dagger$, is then the corresponding creation operator. This description is used in \Sref{sec:CoolingMethods:MMC}, as adapted from the literature (see, in particular, Ref.~\cite{Domokos2001}), in our semiclassical exploration of the mirror-mediated cooling mechanism.\\
The result embodied in \erefs{eq:QME} and~(\ref{eq:LindbladDissipativeSystem}) is surprisingly independent of the properties of the \emph{bath}, or environment, that gives rise to the incoherent processes. The reader is referred to Ref.~\cite{Gardiner2004}, especially Ch.~5, for more discussion about this point. Throughout this thesis, we will assume:~(i)~weak coupling between the system and the bath, (ii)~that the dynamics is Markovian~\cite[Ch.~3]{Gardiner1996}, (iii)~that the rotating wave approximation (see \sref{sec:CoolingMethods:OBE:Inc}) is a valid approximation, and (iv)~that the bath is effectively at zero temperature. These assumptions are crucial in deriving \eref{eq:LindbladDissipativeSystem}.
\par
We now introduce what we shall refer to as the two-level atom (TLA). This is a system that is assumed to have two levels, a ground state $\ket{g}$ labelled `$\text{g}$' and an excited state $\ket{e}$ labelled `$\text{e}$'. In terms of the atomic model we introduced in \sref{sec:CoolingMethods:AS}, a TLA can be emulated by a $J=0\rightarrow J^\prime=1$ transition pumped by purely positive-handed circularly polarised light, such that only the $m_J=0$ and $m_{J^\prime}=1$ sublevels are coupled. In the so-called $D$ transitions of the alkali atoms, atoms can be made to cycle between two hyperfine levels in exactly the same fashion. Whilst the TLA is an idealisation, therefore, it can be approximated quite well in the laboratory.\\
A justification of \eref{eq:QME} is provided in Ref.~\cite[\textsection 1.5; see Eq.~(1.5.39)]{Gardiner2004} for the case of a TLA coupled to the radiation field. Such a description encompasses stimulated absorption and emission as well as spontaneous emission processes.
\section{The Optical Bloch Equations}\label{sec:CoolingMethods:OBE}
The Optical Bloch Equations (OBE) are a set of equations that explore the behaviour of the elements of $\m{\rho}$ when a TLA interacts with an incident harmonic electric field. To derive the OBE, we start off from the Hamiltonian of a TLA at rest in a radiation field~\cite{CohenTannoudji2004}:
\begin{equation}
\hat{H}=\hat{H}_\text{A}+\hat{H}_\text{R}+\hat{H}_\text{I}\,.
\end{equation}
The terms on the right-hand side of the above equation are the Hamiltonian of:~the atom, the quantised radiation field, and the interaction between the atom and the field, respectively. Let us now assume that the energy difference between $\ket{g}$ and $\ket{e}$ is equal to $\hbar\omega_0$, including any level shifts due to coupling to the quantised field. We can immediately write down $\hat{H}_\text{A}=\hbar\omega_0\ket{e}\bra{e}$, where we have taken the ground state as defining zero energy. The interaction Hamiltonian has two contributions, $\hat{H}_\text{I,i}$ arising from the incident electric field and $\hat{H}_\text{I,q}$ arising from the quantised field, which we will henceforth assume to be initially in the vacuum state, denoted $\ket{0}$. Let us consider these two contributions separately.
\subsection{Interaction with the quantised field}\label{sec:CoolingMethods:OBE:QEF}
The density matrix of the TLA interacting with the quantised field evolves according to
\begin{equation}
\dot{\m{\rho}}=\tfrac{1}{\i\hbar}[\m{H}_\text{A}+\m{H}_\text{R}+\m{H}_\text{I,q},\m{\rho}]\,,
\end{equation}
Following the standard procedure outlined in the literature (see, for example, Ref.~\cite[\textsection V.A]{CohenTannoudji2004}), we may derive the following equations for the density matrix elements:
\begin{subequations}
\label{eq:OBEQuantisedField}
\begin{align}
\dot{\m{\rho}}_\text{gg}&=2\Gamma\m{\rho}_\text{ee}\,,\\
\dot{\m{\rho}}_\text{ee}&=-2\Gamma\m{\rho}_\text{ee}\,,\\
\dot{\m{\rho}}_\text{ge}&=\i\omega_0\m{\rho}_\text{ge}-\Gamma\m{\rho}_\text{ge}\,,\text{ and}\\
\dot{\m{\rho}}_\text{eg}&=-\i\omega_0\m{\rho}_\text{eg}-\Gamma\m{\rho}_\text{eg}\,.
\end{align}
\end{subequations}
The first pair of these equations describes the decay, with a time constant $2\Gamma$,\footnote{This differs from standard notation by a factor of $2$; our $\Gamma$ is frequently denoted $\gamma$ in the literature. We use this notation for simplicity of presentation and consistency with Refs.~\cite{Xuereb2009a} and~\cite{Xuereb2009b}.} of atoms in the excited state to the ground state by means of spontaneous emission. The second pair describes the harmonic evolution, at a frequency $\omega_0$, and decay of the coherences of the atom.\par
Spontaneous emission as an irreversible process is a result of the coupling of the atomic dipole to the infinity of quantised vacuum modes in free space. It can therefore be described as an incoherent process using Lindblad terms. The decay terms in the above equations can be derived by adapting \eref{eq:LindbladDissipativeSystem} for the coupling between the atom and the vacuum field. In this case, $\kappa\rightarrow\Gamma$, and $\hat{c}\rightarrow\ket{g}\bra{e}$ is the lowering operator. \erefs{eq:OBEQuantisedField} then follow if we assume that the TLA, effectively, has infinite mass and suffers no momentum recoil from spontaneous emission.
\subsection{Interaction with the incident field}\label{sec:CoolingMethods:OBE:Inc}
Atoms and molecules, whose sizes are usually in the range of $10^{-10}$--$10^{-9}$\,m, are much smaller than the wavelength of visible electromagnetic radiation (ca.~$400$--$800\times10^{-9}$\,m) and the spatial variation of the electric field over the extent of the electron cloud can therefore be neglected. Several cases violate this criterion, such as the interaction of highly excited Rydberg atoms with visible radiation, but we will only consider cases where we can make this \emph{long-wavelength} assumption. We can thus write the interaction between the atom and the radiation field, by modelling the atom as a point dipole, as $\hat{H}_\text{I,i}=-\m{d}\cdot\v{\mathcal{E}}$~\cite{CohenTannoudji2004}, where $\m{d}$ is the dipole moment of the atom and $\v{\mathcal{E}}=\v{\mathcal{E}}_0\cos(\omega_\text{L}t)$ the incident electric field at the position of the atom. In the basis $\bigl\{\ket{g},\ket{e}\bigr\}$, $\m{d}$ is a Hermitian matrix having only off-diagonal elements.\footnote{The dipole operator has odd parity and its diagonal elements are therefore zero.} We define the Rabi frequency $\Omega$ of the interaction:
\begin{equation}
\label{eq:RabiFrequencyDefinition}
\Omega=-\tfrac{1}{\hbar}\m{d}_\text{eg}\cdot\v{\mathcal{E}}_0\,,
\end{equation}
with $\m{d}_\text{eg}=\bra{e}\m{d}\ket{g}$ as usual. By a suitable choice of phase we can ensure that $\Omega$ is real, in which case $\m{d}$ is a symmetric matrix; however, we will keep the rest of this chapter general and do not make this simplification. The evolution of the density matrix under the influence of this interaction is given by
\begin{equation}
\dot{\m{\rho}}=\tfrac{1}{\i\hbar}[-\m{d}\cdot\v{\mathcal{E}}_0\cos(\omega_\text{L}t),\m{\rho}]\,,
\end{equation}
which can be expanded to
\begin{subequations}
\label{eq:OBEIncNoRWA}
\begin{align}
\dot{\m{\rho}}_\text{gg}&=-\i\bigl(\Omega^\ast\m{\rho}_\text{eg}-\Omega\m{\rho}_\text{ge}\bigr)\cos(\omega_\text{L}t)\,,\\
\dot{\m{\rho}}_\text{ee}&=\i\bigl(\Omega^\ast\m{\rho}_\text{eg}-\Omega\m{\rho}_\text{ge}\bigr)\cos(\omega_\text{L}t)\,,\\
\dot{\m{\rho}}_\text{ge}&=-\i\Omega^\ast\bigl(\m{\rho}_\text{ee}-\m{\rho}_\text{gg}\bigr)\cos(\omega_\text{L}t)\,,\text{ and}\\
\dot{\m{\rho}}_\text{eg}&=\i\Omega\bigl(\m{\rho}_\text{ee}-\m{\rho}_\text{gg}\bigr)\cos(\omega_\text{L}t)\,.
\end{align}
\end{subequations}
In the same basis as before, we can write the atomic dipole $\m{d}=\m{d}_\text{eg}\ket{e}\bra{g}+\m{d}_\text{ge}\ket{g}\bra{e}$. We can also rewrite $\cos(\omega_\text{L}t)=\tfrac{1}{2}\bigl[\exp(\i\omega_\text{L}t)+\exp(-\i\omega_\text{L}t)\bigr]$. Then, we have
\begin{multline}
-\m{d}\cdot\v{\mathcal{E}}=\tfrac{1}{2}\hbar\bigl[\Omega\ket{e}\bra{g}\exp(-\i\omega_\text{L}t)+\Omega^\ast\ket{g}\bra{e}\exp(\i\omega_\text{L}t)\\+\Omega\ket{e}\bra{g}\exp(-\i\omega_\text{L}t)+\Omega^\ast\ket{g}\bra{e}\exp(\i\omega_\text{L}t)\bigr]\,.
\end{multline}
The first two terms inside the brackets in the above expression are resonant when $\omega_\text{L}$ is close to $\omega_0$. They describe the raising of the atomic energy level from $\ket{g}$ to $\ket{e}$ by absorption of a photon or the lowering of the energy level from $\ket{e}$ to $\ket{g}$ accompanied by the emission of a photon. The other two terms describe nonresonant processes and can be ignored when the incident radiation is sufficiently close to the atomic resonance. This is called the \emph{rotating-wave approximation} (RWA), and allows us to rewrite \erefs{eq:OBEIncNoRWA} as
\begin{subequations}
\label{eq:OBEIncRWA}
\begin{align}
\dot{\m{\rho}}_\text{gg}&=-\tfrac{1}{2}\i\bigl(\Omega^\ast\m{\rho}_\text{eg}e^{\i\omega_\text{L}t}-\Omega\m{\rho}_\text{ge}e^{-\i\omega_\text{L}t}\bigr)\,,\\
\dot{\m{\rho}}_\text{ee}&=\tfrac{1}{2}\i\bigl(\Omega^\ast\m{\rho}_\text{eg}e^{\i\omega_\text{L}t}-\Omega\m{\rho}_\text{ge}e^{-\i\omega_\text{L}t}\bigr)\,,\\
\dot{\m{\rho}}_\text{ge}&=-\tfrac{1}{2}\i\Omega^\ast\bigl(\m{\rho}_\text{ee}-\m{\rho}_\text{gg}\bigr)e^{\i\omega_\text{L}t}\,,\text{ and}\\
\dot{\m{\rho}}_\text{eg}&=\tfrac{1}{2}\i\Omega\bigl(\m{\rho}_\text{ee}-\m{\rho}_\text{gg}\bigr)e^{-\i\omega_\text{L}t}\,.
\end{align}
\end{subequations}
\ \\
We are now in a position to obtain the full time-independent OBE. The right-hand sides of \erefs{eq:OBEIncRWA} and~(\ref{eq:OBEQuantisedField}) can be added together, according to the approximation of independent rates,\footnote{Since the two systems of equations generally describe time evolution on two vastly different timescales, it is reasonable to assume that the physical processes they describe do not interfere, and that the time derivatives can simply be added.} to give the full time derivative of the elements of $\m{\rho}$. Our final manipulation is to set $\tilde{\m{\rho}}_\text{gg}=\m{\rho}_\text{gg}$, $\tilde{\m{\rho}}_\text{ee}=\m{\rho}_\text{ee}$, $\tilde{\m{\rho}}_\text{ge}=\m{\rho}_\text{ge}\exp(-\i\omega_\text{L}t)$, and $\tilde{\m{\rho}}_\text{eg}=\m{\rho}_\text{eg}\exp(\i\omega_\text{L}t)$, whereby
\begin{subequations}
\label{eq:OBE}
\begin{align}
\dot{\tilde{\m{\rho}}}_\text{gg}&=-\tfrac{1}{2}\i\bigl(\Omega^\ast\tilde{\m{\rho}}_\text{eg}-\Omega\tilde{\m{\rho}}_\text{ge}\bigr)+2\Gamma\tilde{\m{\rho}}_\text{ee}\,,\\
\label{eq:OBEEE}
\dot{\tilde{\m{\rho}}}_\text{ee}&=\tfrac{1}{2}\i\bigl(\Omega^\ast\tilde{\m{\rho}}_\text{eg}-\Omega\tilde{\m{\rho}}_\text{ge}\bigr)-2\Gamma\tilde{\m{\rho}}_\text{ee}\,,\\
\dot{\tilde{\m{\rho}}}_\text{ge}&=-\tfrac{1}{2}\i\Omega^\ast\bigl(\tilde{\m{\rho}}_\text{ee}-\tilde{\m{\rho}}_\text{gg}\bigr)-\bigl(\Gamma+\i\Delta_\text{L}\bigr)\tilde{\m{\rho}}_\text{ge}\,,\text{ and}\\
\dot{\tilde{\m{\rho}}}_\text{eg}&=\tfrac{1}{2}\i\Omega\bigl(\tilde{\m{\rho}}_\text{ee}-\tilde{\m{\rho}}_\text{gg}\bigr)-\bigl(\Gamma-\i\Delta_\text{L}\bigr)\tilde{\m{\rho}}_\text{eg}\,,
\end{align}
\end{subequations}
with $\Delta_\text{L}=\omega_\text{L}-\omega_0$. These are the Optical Bloch Equations for a TLA at rest~\cite{CohenTannoudji2004}.
\section{Polarisability of a two-level atom}\label{sec:CoolingMethods:PolTLA}
\erefs{eq:OBE} can be expressed in terms of three independent parameters, $u=\tfrac{1}{2}\bigl(\tilde{\m{\rho}}_\text{ge}+\tilde{\m{\rho}}_\text{eg}\bigr)$, $v=\tfrac{1}{2\i}\bigl(\tilde{\m{\rho}}_\text{ge}-\tilde{\m{\rho}}_\text{eg}\bigr)$ and $w=\tfrac{1}{2}\bigl(\tilde{\m{\rho}}_\text{ee}-\tilde{\m{\rho}}_\text{gg}\bigr)$, since the total population of atoms is fixed, i.e., $\tilde{\m{\rho}}_\text{gg}+\tilde{\m{\rho}}_\text{ee}=1$. These three parameters are all real, since $\m{\rho}$ is Hermitian, and can be used to describe the coherent evolution of the state of the atom geometrically, with the `Bloch' vector $(u,v,w)$ describing a path on the so-called Bloch Sphere. We will not be concerned with the Bloch Sphere in this thesis but will use $u$, $v$ and $w$ as a shortcut to deriving some results in this section.\par
If an atom is moving at not too fast a speed, say its velocity $\v{v}$ is such that $\tau_\text{i}\lvert\v{v}\rvert,\tau_\text{L}\lvert\v{v}\rvert\ll\lambda_\text{L}$, with $\tau_\text{L}=2\pi/\omega_\text{L}$ being the duration of an optical period of the incident radiation, $\lambda_\text{L}=2\pi c/\omega_\text{L}$ its wavelength, and $\tau_\text{i}=\min\bigl\{(2\Gamma)^{-1},\Omega^{-1}\bigr\}$ the timescale for the internal evolution of the atom, then we can assume that at each point in time the internal variables of the atom---represented by $\tilde{\m{\rho}}$ or, equivalently, $(u,v,w)$---will reach a state of equilibrium~\cite{CohenTannoudji2004}. In other words, we need only concern ourselves with the steady-state value of $\tilde{\m{\rho}}\equiv\tilde{\m{\rho}}^\text{st}$ (this process is known as \emph{adiabatic elimination}~\cite{Gardiner1984} of the external variables). Let us substitute $u$, $v$ and $w$ for the matrix elements of $\tilde{\m{\rho}}$ in \erefs{eq:OBE}, and let us set all the time derivatives to zero. Then we can write
\begin{subequations}
\begin{align}
\dot{u}&=0=-2\Gamma u^\text{st}+2\Delta_\text{L}v^\text{st}-\bigl(\Omega^\ast-\Omega\bigr)w^\text{st}\,,\\
\dot{v}&=0=-2\Delta_\text{L}u^\text{st}-2\Gamma v^\text{st}-\bigl(\Omega^\ast+\Omega\bigr)w^\text{st}\,,\text{ and}\\
\dot{w}&=0=\tfrac{1}{2}\i\bigl(\Omega^\ast-\Omega\bigr)u^\text{st}+\tfrac{1}{2}\bigl(\Omega^\ast+\Omega\bigr)v^\text{st}-2\Gamma w^\text{st}-\Gamma\,,
\end{align}
\end{subequations}
with the superscript `st' again denoting steady-state values. In particular, then,
\begin{subequations}
\label{eq:OBEFinalST}
\begin{align}
\tilde{\m{\rho}}_\text{gg}^\text{st}&=1-\frac{\lvert\Omega\rvert^2}{4}\frac{1}{\Delta_\text{L}^2+\Gamma^2+\lvert\Omega\rvert^2/2}=\frac{1}{2}\frac{2+s}{1+s}\,,\\
\label{eq:OBEFinalSTEE}
\tilde{\m{\rho}}_\text{ee}^\text{st}&=\frac{\lvert\Omega\rvert^2}{4}\frac{1}{\Delta_\text{L}^2+\Gamma^2+\lvert\Omega\rvert^2/2}=\frac{1}{2}\frac{s}{1+s}\,,\\
\tilde{\m{\rho}}_\text{ge}^\text{st}&=\frac{\Omega}{2}\frac{\Delta_\text{L}+\i\Gamma}{\Delta_\text{L}^2+\Gamma^2+\lvert\Omega\rvert^2/2}=\frac{\Omega/2}{\Delta_\text{L}-\i\Gamma}\frac{1}{1+s}\,,\text{ and}\\
\tilde{\m{\rho}}_\text{eg}^\text{st}&=\frac{\Omega^\ast}{2}\frac{\Delta_\text{L}-\i\Gamma}{\Delta_\text{L}^2+\Gamma^2+\lvert\Omega\rvert^2/2}=\frac{\Omega^\ast/2}{\Delta_\text{L}+\i\Gamma}\frac{1}{1+s}\,,
\end{align}
\end{subequations}
where we have defined the saturation parameter $s=\bigl(\lvert\Omega\rvert^2/2\bigr)\big/\bigl(\Delta_\text{L}^2+\Gamma^2\bigr)$, which is proportional to the intensity of the field interacting with the atom, since $\lvert\Omega\rvert\propto\lvert\v{\mathcal{E}}_0\rvert$. Much physical insight can be gained by exploring how $\tilde{\m{\rho}}^\text{st}$ behaves when $s$ is varied. For very large $s$, we have $\tilde{\m{\rho}}_\text{gg}^\text{st},\tilde{\m{\rho}}_\text{ee}^\text{st}\approx\tfrac{1}{2}$, in which case the population is evenly distributed between the two states, and $\tilde{\m{\rho}}_\text{ge}^\text{st},\tilde{\m{\rho}}_\text{eg}^\text{st}\approx 0$, so that the coherences between the two states essentially disappear. Conversely, for very small $s$, we have $\tilde{\m{\rho}}_\text{gg}^\text{st}\approx 1$ and $\tilde{\m{\rho}}_\text{ee}^\text{st}\approx 0$; i.e., practically the entire atomic population is in the ground state, and in the limit of small $s$ we have
\begin{align}
\tilde{\m{\rho}}_\text{ge}^\text{st}=\frac{\Omega/2}{\Delta_\text{L}-\i\Gamma}\,\text{ and }\,\tilde{\m{\rho}}_\text{eg}^\text{st}=\frac{\Omega^\ast/2}{\Delta_\text{L}+\i\Gamma}\,.
\end{align}
The average value of the (induced) atomic dipole moment in the steady state is given by
\begin{align}
\label{eq:AvgDip}
\expt{\m{d}}&=\Tr\bigl(\m{\rho}^\text{st}\m{d}\bigr)=2\real\bigl\{\m{\rho}^\text{st}_\text{eg}\m{d}_\text{ge}\bigr\}\nonumber\\
&=2\real\biggl\{\frac{\Omega^\ast/2}{\Delta_\text{L}+\i\Gamma}\frac{1}{1+s}\m{d}_\text{ge}e^{-\i\omega_\text{L}t}\biggr\}\,.
\end{align}
We can assume that the induced atomic dipole moment is aligned, in space, with the incident electric field, at which point we can set $\m{d}_\text{ge}=\lvert\m{d}_\text{ge}\rvert e^{i\phi}$, with $\phi$ being some phase, and $\m{d}_\text{ge}\cdot\v{\mathcal{E}}_0=\lvert\m{d}_\text{ge}\rvert\mathcal{E}_0 e^{i\phi}$, such that
\begin{equation}
\expt{\m{d}}=\real\biggl\{-\frac{1}{\hbar}\frac{\lvert\m{d}_\text{ge}\rvert^2e^{2i\phi}}{\Delta_\text{L}+\i\Gamma}\frac{1}{1+s}\mathcal{E}_0e^{-\i\omega_\text{L}t}\biggr\}\,.
\end{equation}
Let $\chi$ be the (complex) polarisability of the atom, whereby $\expt{\m{d}}=\epsilon_0\real\bigl\{\chi\mathcal{E}_0e^{-\i\omega_\text{L}t}\bigr\}$~\cite{Jackson1998}, $\epsilon_0$ being the vacuum permittivity. It can be verified by direct calculation that, within the RWA, $\real\{\v{d}\cdot\v{\mathcal{E}}\}=\real\{\v{d}\}\cdot\real\{\v{\mathcal{E}}\}$; this justifies the form of the equation in the preceding sentence. Thus,
\begin{equation}
\chi=-\frac{\lvert\bra{e}\m{d}\ket{g}\rvert^2}{\epsilon_0\hbar}\frac{1}{\Delta_\text{L}+\i\Gamma}\frac{1}{1+s}\,,
\end{equation}
where we have made the simplification $\phi=0$, corresponding to assuming that the matrix elements of $\m{d}$ are real, and have also used the relation $\lvert\m{d}_\text{ge}\rvert=\lvert\m{d}_\text{eg}\rvert=\lvert\bra{e}\m{d}\ket{g}\rvert$. Let us now simplify this expression. First, consider \eref{eq:OBEFinalSTEE}, which describes the population in the excited state, and note that atoms in the excited state decay at a rate $2\Gamma$, as also implied by \eref{eq:OBEEE}. The total scattering rate is therefore given by
\begin{equation}
\label{eq:ScatRate} R=2\Gamma\m{\rho}_\text{ee}^\text{st}=\Gamma\frac{\lvert\Omega\rvert^2/2}{\Delta_\text{L}^2+\Gamma^2+\lvert\Omega\rvert^2/2}=\Gamma\frac{I/I_\text{sat}}{1+\bigl(\Delta_\text{L}/\Gamma\bigr)^2+I/I_\text{sat}}\,,
\end{equation}
which will be justified in \Sref{sec:CoolingMethods:WD}, and where the second equality defines the saturation intensity
\begin{equation}
I_\text{sat}=\frac{2I\Gamma^2}{\lvert\Omega\rvert^2}\,,
\end{equation}
with $I=\tfrac{1}{2}c\epsilon_0\mathcal{E}_0^2$~\cite{Hecht2001}. We can use the definitions of the Rabi frequency $\Omega$, \eref{eq:RabiFrequencyDefinition}, and $I_\text{sat}$ to obtain
\begin{equation}
\lvert\bra{e}\m{d}\ket{g}\rvert^2=\frac{c\epsilon_0\Gamma^2\hbar^2}{I_\text{sat}}\,.
\end{equation}
The scattered power is given by $\hbar\omega_\text{L}R$. Let us now define $\sigma_\text{a}\equiv\hbar\omega_\text{L}R/I=\hbar\omega_\text{L}\Gamma/I_\text{sat}$ on resonance and in the limit of low incident power. Thus, $\sigma_\text{a}$ corresponds to the scattering, or radiative, cross-section (the scattered power divided by the incident energy flux) when $\Delta_\text{L}=0$ and $I\ll I_\text{sat}$. But then,
\begin{equation}
\frac{\lvert\bra{e}\m{d}\ket{g}\rvert^2}{\epsilon_0\hbar}=\frac{\sigma_\text{a}c}{\omega_\text{L}}\Gamma=\frac{\lambda}{\pi}\frac{\sigma_\text{a}}{2}\Gamma\,,\text{ or}
\end{equation}
\begin{equation}
\chi=-\frac{\sigma_\text{a}c}{\omega_\text{L}}\Gamma\frac{1}{\Delta_\text{L}+\i\Gamma}\frac{1}{1+s}=-\frac{\lambda}{\pi}\frac{\sigma_\text{a}}{2}\frac{\Gamma}{\Delta_\text{L}+\i\Gamma}\frac{1}{1+s}\,.
\end{equation}
It can also be shown that $\sigma_\text{a}=3\lambda^2/(2\pi)$~\cite[\textsection 68.6]{Drake2005}, which allows us to relate $\chi$ to known quantities. We will use the linear polarisability $\chi$ to define the dimensionless `scattering parameter'\footnote{We will refer to $\zeta$ as `polarisability' throughout---the linear relation between $\chi$ and $\zeta$, as well as the context, allows us to do this without giving rise to any ambiguity. $\chi$ is perhaps more correctly referred to as a `susceptibility'.} $\zeta$. Indeed, let us first look at the 1D wave equation, along the $z$ axis, for a monochromatic plane wave incident normally on a polarisable plane at $z=0$~\cite{Deutsch1995,Jackson1998}:
\begin{equation}
\Bigl(\tfrac{\partial^2}{\partial z^2}+k^2\Bigr)\v{\mathcal{E}}=-k^2\eta\chi\delta(z)\v{\mathcal{E}}\,,
\end{equation}
with $\delta(z)$ being the Dirac $\delta$-function, $k=2\pi/\lambda$ the wavenumber of the wave and $\eta$ the density of the particles making up the plane per unit area. If the wave interacts with one particle, then we can effectively set $\eta=1/\sigma_\text{L}$, $\sigma_\text{L}$ being the mode area of the (possibly focussed) wave. In situations where the atom is at the focus of an extremely tightly-focussed wave, however, this approximation breaks down~\cite{Tey2009} and we effectively have $\eta<1/\sigma_\text{L}$; we will henceforth assume that the atom is not in the centre of such a tightly-focussed beam.\\
By defining $\zeta=\tfrac{1}{2}k\eta\chi$, and making use of the boundary conditions~\cite{Deutsch1995}
\begin{subequations}
\begin{equation}
\v{\mathcal{E}}\big|_{z\rightarrow 0^-}=\v{\mathcal{E}}\big|_{z\rightarrow 0^+}\,, \text{ and}\\
\end{equation}
\begin{equation}
\bigl(\tfrac{\partial}{\partial z}\v{\mathcal{E}}\bigr)\big|_{z\rightarrow 0^-}-\bigl(\tfrac{\partial}{\partial z}\v{\mathcal{E}}\bigr)\big|_{z\rightarrow 0^+}=2k\zeta\Bigl(\v{\mathcal{E}}\big|_{z=0}\Bigr)\,,
\end{equation}
\end{subequations}
obtained by integrating the wave equation over a small interval centred at $z=0$, we can show that the complex (amplitude) reflection and transmission coefficients are
\begin{equation}
\mathfrak{r}=\frac{i\zeta}{1-\i\zeta}\,\text{ and }\,\mathfrak{t}=\frac{1}{1-\i\zeta}\,,
\end{equation}
respectively. The polarisability of a TLA can therefore be written~\cite{Xuereb2009b}
\begin{equation}
\label{eq:ZetaDefn}
\zeta=-\frac{\sigma_\text{a}}{2\sigma_\text{L}}\frac{\Gamma}{\Delta_\text{L}+\i\Gamma}\frac{1}{1+s}\,,\text{ or }\zeta=-\frac{\sigma_\text{a}}{2\sigma_\text{L}}\frac{\Gamma}{\Delta_\text{L}+\i\Gamma}\,,
\end{equation}
where the low-intensity ($s\rightarrow 0$) limit is taken at the end; we will be concerned with this limit in the subsequent work. This form of $\zeta$ is independent of the system of units (SI or CGS) used and expresses the polarisability of the TLA as a dimensionless number. One of the advantages of this formalism is that $\zeta$ can take, at the outset, any complex value. For far-detuned atoms, for example, $\zeta$ is approximately real, while on resonance it is purely imaginary. Going a step further, we can use this parameter as an abstract quantity representing a general 1D `scatterer' with (amplitude) reflectivity $\mathfrak{r}$ and transmissivity $\mathfrak{t}$, noting that for real $\zeta$ the scatterer absorbs none of the incident energy ($\lvert\mathfrak{r}\rvert^2+\lvert\mathfrak{t}\rvert^2=1$). In \pref{part:TMM} we will apply this concept to form a very physical link between the cooling of atoms in radiation fields and optomechanics, which concerns itself with the manipulation of mesoscopic optical elements using light.\par
Let us make some final comments about \eref{eq:ZetaDefn}. First of all, as can easily be verified, this definition for $\zeta$ obeys the Kramers--Kronig relations~\cite{Toll1956}. However, if we rewrite
\begin{equation}
\label{eq:ZetaFreq}
\zeta(\omega)=-\frac{\sigma_\text{a}}{2\sigma_\text{L}}\frac{\Gamma}{\omega-\omega_0+\i\Gamma}\,,
\end{equation}
whereby $\zeta=\zeta(\omega_\text{L})$ in \eref{eq:ZetaDefn}, we can immediately see that $\zeta(-\omega)\neq\bigl[\zeta(\omega)\bigr]^\ast$. This immediately leads to problems. For let us assume that, in the time domain, $\zeta(t)$ is real, which is equivalent to asserting that $\expt{\v{d}}$ in \eref{eq:AvgDip} is real. Then, using the definition of the Fourier transform,
\begin{equation}
\zeta(\omega)=\int_{-\infty}^\infty e^{-\i\omega t}\zeta(t)\,\mathrm{d} t\,,\text{ so that }\zeta(-\omega)=\int_{-\infty}^\infty e^{\i\omega t}\zeta(t)\,\mathrm{d} t=\bigl[\zeta(\omega)\bigr]^\ast\,;
\end{equation}
something is not right in our definition of $\zeta$. Indeed, a more rigorous treatment ignoring the RWA gives~\cite{Wang2009}
\begin{equation}
\label{eq:ZetaFreqFull}
\zeta(\omega)=-\frac{\sigma_\text{a}}{2\sigma_\text{L}}\biggl(\frac{\Gamma}{\omega-\omega_0+\i\Gamma}-\frac{\Gamma}{\omega+\omega_0+\i\Gamma}\biggr)\,,
\end{equation}
which satisfies the aforementioned equality. In all cases we will be concerned with, however, $\lvert\omega-\omega_0\rvert,\Gamma\lll\omega,\omega_0$---the RWA is valid---in which case the second term in the above expression can be safely neglected and \eref{eq:ZetaFreqFull} reduces to \eref{eq:ZetaFreq}.
\section{Energy balance: Work done on a two-level atom}\label{sec:CoolingMethods:WD}
Consider a small interval between a time $t$ and a time $t+\mathrm{d} t$, where $\mathrm{d} t$ is infinitesimally small. During this time interval, the electric field incident on the TLA can be taken to be constant, $\v{\mathcal{E}}_0\cos(\omega_\text{L}t)$, but the dipole moment of the TLA can change, say, due to the more rapid motion of the atomic electrons. The rate of work done, averaged over an optical cycle, by the field on the dipole is then~\cite{Jackson1998}
\begin{align}
\label{eq:AvgPower}
P&=\overline{\v{\mathcal{E}}_0\cos(\omega_\text{L}t)\cdot\expt{\dot{\v{d}}}}\nonumber\\
&=2\epsilon_0\omega_\text{L}\lvert\v{\mathcal{E}}_0\rvert^2\imag\{\chi\}\,.
\end{align}
By making use of the definition of $\chi$, the OBE, and the assumption that $\Omega$ is real, we can show that $P=2\hbar\omega_\text{L}\Gamma\m{\rho}_\text{ee}^\text{st}$. Physically, $P$ is the average power absorbed by the atom. For an atom at rest, this must be equal to the energy lost by the atom per unit time through scattering of photons. Dividing $P$ by the energy per incident photon, $\hbar\omega_\text{L}$, gives the number of photons scattered per unit time by the atom, $R=2\Gamma\m{\rho}_\text{ee}^\text{st}$. This justifies \eref{eq:ScatRate}.\par
Two facts emerge from \eref{eq:AvgPower} that are important for our discussion. First of all, a TLA with real $\chi$ will experience no photon scattering, i.e., scattering is significantly suppressed when $\Delta_\text{L}\gg2\Gamma$. Secondly, and more importantly, suppose that $\v{d}$ is delayed with respect to $\v{\mathcal{E}}$. Then an extra phase factor, let us say $\varphi$, appears in \eref{eq:AvgPower} such that $P=2\epsilon_0\omega_\text{L}\lvert\v{\mathcal{E}}_0\rvert^2\imag\bigl\{\chi e^{\i\varphi}\bigr\}$. For $\varphi\neq 0$, the \emph{real} component of $\chi$ can also lead to exchange of energy between the TLA and the electric field. This is exploited in cooling mechanisms that rely on optical pumping in polarisation gradients (see, for example, \Sref{sec:TMM:Multilevel} and Ref.~\cite{Dalibard1989} as well as the concluding remarks in the next section).
\par
A final remark can be made about the work done on a TLA in an optical potential. Indeed, the work done on a TLA undergoing small displacements in the field is not, in general, equal to the change in the potential energy of the atom in the field~\cite{Dalibard1989} because the energy in the electromagnetic field itself changes. The relation ``work done $=$ change of potential energy'' is only valid for closed systems, and because the system we are considering is patently open (the light leaving the system carries away energy) we can not expect these two quantities to be equal. Care must be taken, then, when deriving forces from gradients of optical potentials.
\section{Forces on a two-level atom}\label{sec:CoolingMethods:ForceTLA}
Under most circumstances, excluding of course Bose--Einstein condensation and related phenomena, we can assume that the spread of the wavefunction of a TLA is much smaller than the optical wavelength. The force on a TLA interacting with an incident field is $\v{\textsl{\textsf{F}}}=\v{\nabla}\bigl(\expt{\v{d}}\cdot\v{\mathcal{E}}\bigr)$, evaluated at the position of the atom ($\v{r}=\v{0}$); this follows from Ehrenfest's theorem~\cite{CohenTannoudji2004}. Let us set $\v{\mathcal{E}}=\v{\mathcal{E}}_0\cos\bigl[\omega_\text{L}t+\phi(\v{r})\bigr]$, with $\phi(\v{0})=0$. Then, we have
\begin{equation}
\v{\textsl{\textsf{F}}}=\Biggl[\sum_i\expt{\v{d}_i}\cdot\v{\nabla}\bigl({\v{\mathcal{E}}_0}\bigr)_i\Biggr]\cos\bigl(\omega_\text{L}t\bigr)-\bigl[\v{\nabla}\phi(\v{r})\bigr]\expt{\v{d}}\cdot\v{\mathcal{E}}_0\sin\bigl(\omega_\text{L}t\bigr)\,,
\end{equation}
where the subscripted $i$ denotes the spatial dimension and the sum runs over $i=x,y,z$. Moreover, \eref{eq:AvgDip} reduces to
\begin{equation}
\expt{\v{d}}=2\v{d}_\text{eg}\bigl[u^\text{st}\cos\bigl(\omega_\text{L}t\bigr)-v^\text{st}\sin\bigl(\omega_\text{L}t\bigr)\bigr]\,.
\end{equation}
Putting the last two equations together, we can express the time-averaged force acting on the TLA, assuming real $\Omega$, as
\begin{align}
\v{\textsl{\textsf{F}}}&=u^\text{st}\v{\nabla}\bigl(\v{d}_\text{eg}\cdot\v{\mathcal{E}}_0\bigr)+v^\text{st}\v{d}_\text{eg}\cdot\v{\mathcal{E}}_0\v{\nabla}\phi(\v{r})\nonumber\\
&=-\hbar u^\text{st}\v{\nabla}\Omega-\hbar\Omega v^\text{st}\v{\nabla}\phi(\v{r})\,,
\end{align}
since $\v{d}_\text{eg}$ has no spatial dependence on this scale. We can give this equation more physical meaning by using \erefs{eq:OBEFinalST} to obtain
\begin{equation}
u^\text{st}=\frac{\Omega}{2}\frac{\Delta_\text{L}}{\Delta_\text{L}^2+\Gamma^2+\Omega^2/2}\,\text{ and }\,v^\text{st}=\frac{\Omega}{2}\frac{\Gamma}{\Delta_\text{L}^2+\Gamma^2+\Omega^2/2}\,,
\end{equation}
whereby
\begin{subequations}
\begin{align}
\label{eq:FrictionReactDiss}
\v{\textsl{\textsf{F}}}&=\underbrace{-\frac{1}{2}\frac{\hbar\Omega\Delta_\text{L}}{\Delta_\text{L}^2+\Gamma^2+\Omega^2/2}\v{\nabla}\Omega}_\text{reactive force}\underbrace{-\frac{1}{2}\frac{\hbar\Omega^2\Gamma}{\Delta_\text{L}^2+\Gamma^2+\Omega^2/2}\v{\nabla}\phi(\v{r})}_\text{dissipative force}\\
\nonumber\\
&=\begin{cases}
-\dfrac{1}{2}\dfrac{\hbar\Omega\Delta_\text{L}}{\Delta_\text{L}^2+\Gamma^2}\v{\nabla}\Omega-\dfrac{1}{2}\dfrac{\hbar\Omega^2\Gamma}{\Delta_\text{L}^2+\Gamma^2}\v{\nabla}\phi(\v{r})&\text{for }s\rightarrow 0\\
\\
-\hbar\Delta_\text{L}\dfrac{\v{\nabla}\Omega}{\Omega}-\hbar\Gamma\v{\nabla}\phi(\v{r})&\text{for }s\rightarrow\infty\,.
\end{cases}
\end{align}
\end{subequations}
We note that, in the limit of high intensity ($s\rightarrow\infty$), the force acting on the TLA tends towards a limit of a dissipative force whose strength is independent of the intensity. In the following sections we will restrict ourselves to the limit of small $s$. This is useful in considering the interaction of a TLA with a standing wave. For small intensities, the TLA behaves as if it interacts with the two running waves making up the standing wave independently. For very high intensities, the process of absorption of photons from one running wave and stimulated emission into the other running wave can become important and modify the interaction significantly.\\
Another way of looking at this is to recall that the force is the gradient of the product of the induced dipole and the local electric field. The gradient operation is linear, so we need not concern ourselves with its effects. However, suppose that the electric field at the position of the TLA is a sum of two fields, $\v{\mathcal{E}}_1+\v{\mathcal{E}}_2$, with the corresponding induced dipole being equal to $\expt{\v{d}_1}+\expt{\v{d}_2}$. Then, the force on the TLA is $\v{\nabla}\bigl[\bigl(\expt{\v{d}_1}+\expt{\v{d}_2}\bigr)\cdot\bigl(\v{\mathcal{E}}_1+\v{\mathcal{E}}_2\bigr)\bigr]$, which has two extra terms in addition to the sum of the two independent forces, $\v{\nabla}\bigl(\expt{\v{d}_1}\cdot\v{\mathcal{E}}_1\bigr)+\v{\nabla}\bigl(\expt{\v{d}_2}\cdot\v{\mathcal{E}}_2\bigr)$. If the spatial (temporal) average of these extra two terms over a wavelength is zero (this happens, e.g., when there is no coherence between the two electric fields), the two forces can therefore be computed separately and added to give the total spatially (temporally) averaged force.
\par
Let us end this section with a note on the dipole force, which corresponds to the term in the force proportional to $\v{\nabla}\Omega$ (the reactive force). \eref{eq:AvgDip} and \eref{eq:AvgPower} tell us that only $v^\text{st}$ can give rise to a global exchange of energy between the TLA and the light field. The reactive component of $\v{\textsl{\textsf{F}}}$ above can, in fact, be derived from a potential. In other words, the dipole force is conservative. This is not true in general, however. Employing the same arguments as before, if $\expt{\m{d}}$ is delayed with respect to the field, a component of $v^\text{st}$ enters the dipole force and it is therefore no longer constrained to be strictly conservative.
\section{The Fluctuation--Dissipation Theorem}\label{ch:CoolingMethods:FDT}
In the case of the systems that we will investigate, the fluctuation--dissipation theorem is a relation between the systematic (dissipative) and fluctuating components of a force acting on a particle. We shall make extensive use of this theorem in the form of \eref{eq:FDT} below.\\
Consider a particle interacting with the radiation field, such that the force acting on the particle, as a function of time, is $\textsl{\textsf{F}}(t)=\textsl{\textsf{F}}_0(t)+\textsl{\textsf{F}}_\text{L}(t)$, where the first term describes a continuous force that, for example, acts to dampen the particle's motion and the second term is a Langevin force, which has a zero time average and describes the instantaneous fluctuations of the force about its average. In other words, $\expt{\textsl{\textsf{F}}(t)}=\expt{\textsl{\textsf{F}}_0(t)}$.\footnote{A note about notation is due: in this section we use the angle brackets, $\expt{\ \cdot\ }$, to represent the average for a classical force or the expectation value for a quantised force, depending on the nature of the force.} Now, consider
\begin{align}
\expt{\textsl{\textsf{F}}(t)\textsl{\textsf{F}}(t^\prime)}&=\expt{\textsl{\textsf{F}}_0(t)\textsl{\textsf{F}}_0(t^\prime)}+\expt{\textsl{\textsf{F}}_0(t)\textsl{\textsf{F}}_\text{L}(t^\prime)}+\expt{\textsl{\textsf{F}}_\text{L}(t)\textsl{\textsf{F}}_0(t^\prime)}+\expt{\textsl{\textsf{F}}_\text{L}(t)\textsl{\textsf{F}}_\text{L}(t^\prime)}\nonumber\\
&=\expt{\textsl{\textsf{F}}_0(t)}\expt{\textsl{\textsf{F}}_0(t^\prime)}+\expt{\textsl{\textsf{F}}_0(t)}\expt{\textsl{\textsf{F}}_\text{L}(t^\prime)}+\expt{\textsl{\textsf{F}}_\text{L}(t)}\expt{\textsl{\textsf{F}}_0(t^\prime)}+\expt{\textsl{\textsf{F}}_\text{L}(t)\textsl{\textsf{F}}_\text{L}(t^\prime)}\nonumber\\
&=\expt{\textsl{\textsf{F}}(t)}\expt{\textsl{\textsf{F}}(t^\prime)}+\expt{\textsl{\textsf{F}}_\text{L}(t)\textsl{\textsf{F}}_\text{L}(t^\prime)}\,,
\end{align}
where the time average is taken on a timescale that is short with respect to the timescale for variations in $\textsl{\textsf{F}}_0(t)$ but long with respect to that for $\textsl{\textsf{F}}_\text{L}(t)$, whereby $\expt{\textsl{\textsf{F}}_0(t)\textsl{\textsf{F}}_\text{L}(t^\prime)}=\expt{\textsl{\textsf{F}}_0(t)}\expt{\textsl{\textsf{F}}_\text{L}(t^\prime)}=0$. The correlation function for $\textsl{\textsf{F}}_\text{L}(t)$ on this timescale is assumed to have the form
\begin{equation}
\expt{\textsl{\textsf{F}}_\text{L}(t)\textsl{\textsf{F}}_\text{L}(t^\prime)}\approx\textsl{\textsf{D}}\,\delta(t-t^\prime)\,,
\end{equation}
where $\textsl{\textsf{D}}$ is some real number; i.e.,
\begin{equation}
\label{eq:FDTDeltaDependence}
\textsl{\textsf{D}}\,\delta(t-t^\prime)=\expt{\textsl{\textsf{F}}(t)\textsl{\textsf{F}}(t^\prime)}-\expt{\textsl{\textsf{F}}(t)}\expt{\textsl{\textsf{F}}(t^\prime)}\,,
\end{equation}
cf. Refs.~\cite{CohenTannoudji1992} and~\cite{Gordon1980}, or
\begin{equation}
\tfrac{1}{2}\textsl{\textsf{D}}=\int_0^\infty\Bigl(\expt{\textsl{\textsf{F}}(t)\textsl{\textsf{F}}(t^\prime)}-\expt{\textsl{\textsf{F}}(t)}\expt{\textsl{\textsf{F}}(t^\prime)}\Bigr)\mathrm{d} t^\prime\,.
\end{equation}
Let us now specify $\textsl{\textsf{F}}_0=-\varrho v$, where the \emph{cooling coefficient} $\varrho$ is some (positive) damping constant and $v$ the velocity of the particle (in 1D). The total force acting on the particle, say of mass $m$, is then
\begin{align}
\textsl{\textsf{F}}(t)=\frac{\mathrm{d}}{\mathrm{d} t}p(t)&=-\varrho v(t)+\textsl{\textsf{F}}_\text{L}(t)\nonumber\\
&=m\frac{\mathrm{d}}{\mathrm{d} t}v(t)\,.
\end{align}
We now multiply this equation by the integrating factor and integrate it from time $t=0$ to some general time $t$. Then:
\begin{align}
\int_0^t \biggl(\frac{\mathrm{d}}{\mathrm{d} t}v(t)\biggr)\biggr|_{t=t^\prime}e^{(\varrho/m)t^\prime}\mathrm{d} t^\prime&=-\int_0^t(\varrho/m)v(t^\prime)e^{(\varrho/m)t^\prime}\mathrm{d} t^\prime+\frac{1}{m}\int_0^t\textsl{\textsf{F}}_\text{L}(t^\prime)e^{(\varrho/m)t^\prime}\mathrm{d} t^\prime\nonumber\\
&=-\int_0^tv(t^\prime)\biggl(\frac{\mathrm{d}}{\mathrm{d} t}e^{(\varrho/m)t}\biggr)\biggr|_{t=t^\prime}\mathrm{d} t^\prime+\frac{1}{m}\int_0^t\textsl{\textsf{F}}_\text{L}(t^\prime)e^{(\varrho/m)t^\prime}\mathrm{d} t^\prime\,,
\end{align}
so that
\begin{align}
v(t)e^{(\varrho/m)t}&=v(0)+\frac{1}{m}\int_0^t\textsl{\textsf{F}}_\text{L}(t^\prime)e^{(\varrho/m)t^\prime}\mathrm{d} t^\prime\,,\text{ or}\nonumber\\
v(t)&=v(0)e^{-(\varrho/m)t}+\frac{1}{m}\int_0^t\textsl{\textsf{F}}_\text{L}(t^\prime)e^{-(\varrho/m)(t-t^\prime)}\mathrm{d} t^\prime\,.
\end{align}
Let us average this relation over the same timescale as above. Then $\expt{v(t)}=v(0)e^{-(\varrho/m)t}$, whereby the width of the velocity distribution, defined as
\begin{equation}
\sigma^2_v(t)=\expt{\bigl[v(t)-\expt{v(t)}\bigr]^2}\,,
\end{equation}
is given by
\begin{align}
\sigma^2_v(t)&=\frac{1}{m^2}\int_0^t\int_0^t\expt{\textsl{\textsf{F}}_\text{L}(t^{\prime\prime})\textsl{\textsf{F}}_\text{L}(t^\prime)}e^{-(\varrho/m)(t-t^\prime)}e^{-(\varrho/m)(t-t^{\prime\prime})}\mathrm{d} t^\prime\mathrm{d} t^{\prime\prime}\nonumber\\
&=\frac{1}{m^2}\int_0^t\int_0^t\textsl{\textsf{D}}\,\delta\bigl(t^\prime-t^{\prime\prime}\bigr)e^{-(\varrho/m)(t-t^\prime)}e^{-(\varrho/m)(t-t^{\prime\prime})}\mathrm{d} t^\prime\mathrm{d} t^{\prime\prime}\nonumber\\
&=\frac{\textsl{\textsf{D}}}{m^2}\int_0^te^{-2(\varrho/m)(t-t^\prime)}\mathrm{d} t^\prime=\frac{\textsl{\textsf{D}}}{2m\varrho}\Bigl[1-e^{-2(\varrho/m)t}\Bigr]\,.
\end{align}
For small values of $t$, $\sigma^2_v(t)\approx\textsl{\textsf{D}} t/m^2$. The width of the momentum distribution is similarly given by $\sigma^2_p(t)\approx\textsl{\textsf{D}} t$, whereby $\textsl{\textsf{D}}$ is, by definition, the momentum diffusion coefficient. For very long times we can make two observations. Firstly, $\expt{v(t)}=0$, so that the particle is cooled to zero mean velocity; secondly, $\sigma_v^2(t)=\textsl{\textsf{D}}/(2m\varrho)$. We will not go into any further depth here, but it can be shown~\cite{Risken1989} that all higher-order correlation functions vanish at long timescales. Moreover, this further implies that the velocity distribution of an ensemble of such particles is Gaussian, of the form $\exp\bigl[-v^2/(\sigma_v/2)^2\bigr]$, with the factor of $\tfrac{1}{2}$ arising due to the symmetry of the distribution. Comparing this with a Maxwell-Boltzmann distribution in thermal equilibrium at temperature $T$, for which $\sigma^2_v=k_\text{B}T/(2m)$ with $k_\text{B}$ being the Boltzmann constant, we can therefore calculate the equilibrium temperature of our particle:
\begin{equation}
\label{eq:FDT}
k_\text{B}T=\frac{\textsl{\textsf{D}}}{\varrho}\,.
\end{equation}
This is in fact of the same form as that obtained by Einstein~\cite{Einstein1905} in his discussion of the Brownian motion of particles suspended in a liquid. \eref{eq:FDT} links the fluctuating and dissipating forces experienced by the particle to its equilibrium temperature. Although the justification given here is not completely general, namely in terms of demanding a linear friction force and a constant momentum diffusion constant, \eref{eq:FDT} is a good approximation under most circumstances of interest. See Ref.~\cite{Metcalf1999} for a slightly more general treatment, and Ref.~\cite{Clerk2010} for a treatment based on the theory of quantum noise.\footnote{Considerations related to the general quantum case imply the failure of the Onsager hypothesis, which states that ``the average regression of fluctuations will obey the same laws as the corresponding macroscopic irreversible process''~\cite{Onsager1931}, in cases such as where strong coupling between the system and the noise bath is allowed. Put differently, the quantum regression theorem cannot be called a `theorem' in the mathematical sense. See the discussions in Refs.~\cite{Ford1996} and~\cite{Lax2000} for further details.} The results above can easily be adapted for forces acting in 3D by replacing the forces $\textsl{\textsf{F}}$, $\textsl{\textsf{F}}_0$, $\textsl{\textsf{F}}_\text{L}$ with their vector counterparts $\v{\textsl{\textsf{F}}}$, $\v{\textsl{\textsf{F}}}_0$, $\v{\textsl{\textsf{F}}}_\text{L}$; replacing products of the type $\mathcal{F}(t)\mathcal{F}^\prime(t^\prime)$ with $\v{\mathcal{F}}(t)\cdot\v{\mathcal{F}}^\prime(t^\prime)$ ($\mathcal{F},\mathcal{F}^\prime=\textsl{\textsf{F}},\textsl{\textsf{F}}_0,\textsl{\textsf{F}}_\text{L}$); and also replacing $v$ with $\v{v}$.
\section{Beyond two-level atoms}\label{sec:CoolingMethods:BeyondTLA}
\begin{figure}[t]
\centering
\includegraphics{Diagrams/RamanLambda}
\caption[Prototypical `$\Lambda$'--type system]{Prototypical `$\Lambda$'--type system used to describe stimulated Raman transitions. The transition between the ground states is taken to be dipole-forbidden.}
\label{fig:RamanLambdaSystem}
\end{figure}
The two-level atom is a useful, but rather crude, approximation. As noted by Dalibard, Reynaud and Cohen--Tannoudji~\cite{Dalibard1984}, amongst many others, several interesting physical mechanisms can only be explained when one includes the manifold of Zeeman sublevels, \sref{sec:CoolingMethods:AS}, that occur in the energy level structure of real atoms into the model. Indeed, this manifold of sublevels can be included into the definition of the $\zeta$, above, by generalising it into the polarisability tensor, defined as the steady-state expectation value of the polarisability operator $\hat{\boldsymbol{\zeta}}$,
\begin{equation}
\label{eq:ZetaTensorDefn}
\boldsymbol{\zeta}=\Tr\bigl(\tilde{\m{\rho}}^\text{st}\cdot\hat{\boldsymbol{\zeta}}\bigr)=\sum_{i,j}\langle j\rvert\tilde{\m{\rho}}^\text{st}\lvert i\rangle\langle i\rvert\hat{\boldsymbol{\zeta}}\lvert j\rangle\,,
\end{equation}
where $\tilde{\m{\rho}}^\text{st}$, as defined previously, is the steady-state density matrix describing the system and the summation runs over all the internal sublevels of the atom. The matrix elements of the polarisability operator $\hat{\boldsymbol{\zeta}}$, defined similarly to Eq.~(14.9-24) of Ref.~\cite{Shore1990b}, read in the general $\mu$, $\nu$ basis
\begin{equation}
\label{eq:PolarisabilityOperator}
\langle i\rvert\hat{\boldsymbol{\zeta}}\lvert j\rangle
=\zeta_0\sum_e\begin{bmatrix}
\langle i\rvert\hat{\boldsymbol{d}}_\mu\lvert e\rangle\langle e\rvert\hat{\boldsymbol{d}}_\mu\lvert j\rangle & \langle i\rvert\hat{\boldsymbol{d}}_\mu\lvert e\rangle\langle e\rvert\hat{\boldsymbol{d}}_\nu\lvert j\rangle\\
\langle i\rvert\hat{\boldsymbol{d}}_\nu\lvert e\rangle\langle e\rvert\hat{\boldsymbol{d}}_\mu\lvert j\rangle & \langle i\rvert\hat{\boldsymbol{d}}_\nu\lvert e\rangle\langle e\rvert\hat{\boldsymbol{d}}_\nu\lvert j\rangle
\end{bmatrix}\,,
\end{equation}
where $\zeta_0$ embodies the dimensional constants and the frequency dependence of the polarisability of the atom (see, for example, Ref.~\cite{Wang2009}):
\begin{equation}
\zeta_0=-\frac{k}{2\sigma_\text{L}}\frac{1}{\Delta_\text{L}+i\Gamma}\,,
\end{equation}
In the above equation, the dipole moment operator $\hat{\boldsymbol{d}}_\mu$ ($\hat{\boldsymbol{d}}_\nu$) is related to the $\mu$ ($\nu$) polarised light field and the sum runs over all the internal sublevels, $e$, of the atom. The matrix elements of $\hat{\boldsymbol{d}}_\mu$ ($\hat{\boldsymbol{d}}_\nu$) are given by the appropriate Clebsch--Gordan coefficients. We will explore the implications of this generalised polarisability tensor in \pref{part:TMM}, where it will be used to give an alternative derivation of Sisyphus cooling~\cite{Dalibard1989}.
\par
Stimulated Raman transitions~\cite{Shore1990a} also require multilevel atoms by their very definition. The basic model for an atom undergoing a stimulated Raman transition is a three-level `$\Lambda$'--type level system (\fref{fig:RamanLambdaSystem}): an atom having two `ground' states $\ket{g_1}$ and $\ket{g_2}$, and one excited state $\ket{e}$. The transition between $\ket{g_1}$ and $\ket{g_2}$ is always dipole-forbidden,\footnote{In other words, an atom in one of the $\ket{g_i}$ states will remain in that state, justifying their designation as `ground' states.} due to parity conservation, and the remaining two transitions are coupled by `Pump' ($\ket{g_1}\leftrightarrow\ket{e}$) and `Stokes' ($\ket{g_2}\leftrightarrow\ket{e}$) fields, characterised by strengths $\Omega_\text{P}$ and $\Omega_\text{S}$, respectively. For simplicity, we assume that both coupling fields have a detuning $\Delta$ from resonance. Under these conditions, Rabi oscillations occur between the two ground states at an effective frequency~\cite{Bateman2010b}
\begin{equation}
\frac{\lvert\Omega_\text{P}\Omega_\text{S}^\ast\rvert}{2\Delta}\,.
\end{equation}
It has often been assumed in the literature (see, for example, Ref.~\cite[\S2.1]{kasevich92}), that the hyperfine structure of $\ket{e}$ can easily be taken into account by summing over the appropriate multiple routes, giving an effective Rabi frequency
\begin{equation}
\sum_i\frac{\lvert\Omega_{\text{P};i}\Omega_{\text{S};i}^\ast\rvert}{2\Delta}\,,
\end{equation}
where the label $i$ denotes the coupling to the $i$th sublevel of $\ket{e}$, and $\Delta$ is assumed to be much larger than the energy splittings between the different sublevels. Indeed, it was only in a recent publication by Bateman, Xuereb and Freegarde~\cite{Bateman2010b} that this was proven mathematically to be the case in general, by diagonalising the appropriate multilevel Hamiltonian.\\
For completeness, we will now briefly go through the main arguments in Ref.~\cite{Bateman2010b}. This work was originally published as Bateman, J., Xuereb, A., \& Freegarde, T., Phys.\ Rev.\ A \textbf{81}, 043808 (2010).\footnote{JB set up the problem, solved the three-level system case, provided the interpretation and wrote the paper; AX produced the analytical formulation of the eigensystem of the general $(N+2)$-level Hamiltonian.} The energy structure of the atom we consider is a straightforward extension of that shown in \fref{fig:RamanLambdaSystem}, where we maintain the two ground states $\ket{g_1}$ and $\ket{g_2}$, but where a similar hyperfine structure is taken to exist for the excited state---the state $\ket{e}$ splits into $\ket{e_1},\ket{e_2},\dots,\ket{e_N}$, say. The coupling between the two beams and the different excited sublevels may not be identical, and we therefore set $\Omega_{\text{P};i}$ and $\Omega_{\text{S};i}$ to be, respectively, the coupling strengths for the transitions $\ket{g_1}\leftrightarrow\ket{e_i}$ and $\ket{g_2}\leftrightarrow\ket{e_i}$. The Hamiltonian for this system, after making the RWA and transforming to the interaction picture~\cite{Shore1990a},\footnote{The interaction picture, essentially, removes the fast time evolution from the state vectors and is accomplished by a transformation of the type in \eref{eq:RamanLambdaTransformation} where the transformation matrix is the diagonal matrix of eigenvalues of the time-independent part of the Hamiltonian. It lies in between the Schr\"odinger and the Heisenberg pictures. See Ref.~\cite[Complement G$_\text{III}$]{CohenTannoudji1978a}.} can be written out in state vector notation as
\begin{equation}
\label{eq:RamanLambdaHamiltonian}
\m{H}_\text{A}=\begin{bmatrix}
0 & 0 & \tfrac{1}{2}\Omega_{\text{P};1} & \tfrac{1}{2}\Omega_{\text{P};2} & \dots \\
0 & 0 & \tfrac{1}{2}\Omega_{\text{S};1}e^{\i\delta t} & \tfrac{1}{2}\Omega_{\text{S};3}e^{\i\delta t} & \dots \\
\tfrac{1}{2}\Omega_{\text{P};1}^\ast & \tfrac{1}{2}\Omega_{\text{S};1}^\ast e^{-\i\delta t} & -\Delta & 0 & \dots \\
\tfrac{1}{2}\Omega_{\text{P};2}^\ast & \tfrac{1}{2}\Omega_{\text{S};2}^\ast e^{-\i\delta t} & 0 & -\Delta & \dots \\
\vdots & \vdots & \vdots & \vdots & \ddots
\end{bmatrix}\,,
\end{equation}
which acts on the state vector
\begin{equation}
\begin{pmatrix}
\ket{g_1} \\
\ket{g_2} \\
\ket{e_1} \\
\ket{e_2} \\
\vdots
\end{pmatrix}\,.
\end{equation}
In \eref{eq:RamanLambdaHamiltonian} we made the approximation that the detuning of the Pump beam is $\Delta$ from each of the excited states, and that of the Stokes beam is $\Delta+\delta$. This approximation requires that $\Delta$ be much larger than the splittings between the excited state sublevels. We solve the Schr\"odinger equation with this Hamiltonian by using unitary transformations to find a basis where the time evolution of the states is simple and the transformed Hamiltonian is diagonal. When such a change of basis is done, say by using the transformation matrix $\m{O}_\text{BA}$, then the Hamiltonian in the new basis, $\m{H}_\text{B}$, reads~\cite{Shore1990a}
\begin{equation}
\label{eq:RamanLambdaTransformation}
\m{H}_\text{B}=\m{O}_\text{BA}\biggl(\m{H}_\text{A}\m{O}_\text{BA}^{-1}-\i\frac{\partial}{\partial t}\m{O}_\text{BA}^{-1}\biggr)\,.
\end{equation}
Choosing $\m{O}_\text{BA}$ to be the matrix of eigenvectors of $\m{H}_\text{A}$, the first term in this equation is simply the matrix of its eigenvalues. Upon diagonalising $\m{H}_\text{A}$, it turns out that two of the eigenvectors are superpositions of the two ground states and are decoupled from the rest of the levels in the limit of large detuning. The system can therefore be described as an effective two-level system.
\par
Indeed, we set
\begin{equation}
\lVert\m{H}_\text{B}-\lambda\m{I}\rVert=0\,,
\end{equation}
where $\m{I}$ is the $(N+2)\times(N+2)$ identity matrix and $\lambda$ a parameter. This equation simplifies to the fourth-order polynomial equation
\begin{equation}
\label{eq:RamanLambdaCharacteristicPoly}
16\lambda^2(\Delta-\lambda)^2+4\lambda(\Delta-\lambda)\bigl(\lVert\v{\Omega}_\text{P}\rVert^2+\lVert\v{\Omega}_\text{S}\rVert^2\bigr)+\lVert\v{\Omega}_\text{P}\rVert^2\lVert\v{\Omega}_\text{S}\rVert^2-\lvert\v{\Omega}_\text{P}\cdot\v{\Omega}_\text{S}^\ast\rvert^2=0\,,
\end{equation}
where we have defined the vectors
\begin{equation}
\v{\Omega}_\text{P}=\begin{pmatrix}
\Omega_{\text{P;1}} \\
\Omega_{\text{P;2}} \\
\vdots
\end{pmatrix}\,\text{ and }\,
\v{\Omega}_\text{S}=\begin{pmatrix}
\Omega_{\text{S;1}} \\
\Omega_{\text{S;2}} \\
\vdots
\end{pmatrix}\,.
\end{equation}
The constant term, which is independent of $\lambda$, in \eref{eq:RamanLambdaCharacteristicPoly} is also independent of $\Delta$, so that the product of all the solutions, $\prod_i\lambda_i$, is also independent of $\Delta$. Since not all the $\lambda_i$ are independent of $\Delta$, because the equation is also a second-order polynomial in $\Delta$, then \emph{at least one} eigenvalue must disappear as $\lvert\Delta\rvert\to\infty$. In this limit, we can therefore make the approximation $(\Delta-\lambda)\to\Delta$. The resulting equation has roots
\begin{equation}
\lambda_\pm=\frac{-\bigl(\lVert\v{\Omega}_\text{P}\rVert^2+\lVert\v{\Omega}_\text{S}\rVert^2\bigr)\pm\sqrt{\bigl(\lVert\v{\Omega}_\text{P}\rVert^2-\lVert\v{\Omega}_\text{S}\rVert^2\bigr)^2+4\lvert\v{\Omega}_\text{P}\cdot\v{\Omega}_\text{S}^\ast\rvert^2}}{8\Delta}\,;
\end{equation}
the respective eigenvectors are superpositions of $\ket{g_1}$ and $\ket{g_2}$. We now return to $\m{H}_\text{B}$ and constrain ourselves to the Hilbert space of the two ground states, whereupon we can write
\begin{equation}
\label{eq:RamanLambdaHamiltonian2LevelBare}
\m{H}_\text{B}=\begin{bmatrix}
-\delta\sin^2(\theta) & -\delta e^{\i\delta t}\cos(\theta)\sin(\theta) \\
-\delta e^{-\i\delta t}\cos(\theta)\sin(\theta) & \delta\sin^2(\theta)+\tilde{\Omega}_\text{B}
\end{bmatrix}\,;
\end{equation}
in this restricted Hilbert space, we have
\begin{equation}
\m{O}_\text{BA}=\begin{bmatrix}
\cos(\theta) & e^{\i\delta t}\sin(\theta) \\
-e^{-\i\delta t}\sin(\theta) & \cos(\theta)
\end{bmatrix}\,.
\end{equation}
In the preceding Hamiltonian, we have made use of the definitions, analogous to the two-level Rabi system,
\begin{equation}
\tilde{\Omega}_\text{B}=\sqrt{\Omega_\text{B}^2+\Delta_\text{B}^2}\,,\text{ and }\tan(\theta)=\frac{\Delta_\text{B}-\tilde{\Omega}_\text{B}}{\Omega_\text{B}}\,,
\end{equation}
where
\begin{equation}
\label{eq:RamanLambdaOmegaDelta}
\Omega_\text{B}=\frac{\lvert\v{\Omega}_\text{P}\cdot\v{\Omega}_\text{S}^\ast\rvert}{2\Delta}\,,\text{ and }\Delta_\text{B}=\frac{\lVert\v{\Omega}_\text{P}\rVert^2-\lVert\v{\Omega}_\text{S}\rVert^2}{4\Delta}\,.
\end{equation}
Indeed, we immediately notice that the first relation in \eref{eq:RamanLambdaOmegaDelta} implies that the naive summing over states is formally correct in this limit. In the second relation, $\Delta_\text{B}$ can be identified as the light shift, as justified below.
\par
We further transform this system into a simplified basis by first switching to a time-independent Hamiltonian by means of the matrix
\begin{equation}
\m{O}_\text{CB}=\begin{bmatrix}
1 & 0 \\
0 & e^{\i\delta t}
\end{bmatrix}\,,
\end{equation}
followed by the rotation through an angle $\phi$, through a matrix $\m{O}_\text{DC}$, defined by
\begin{equation}
\tan(2\phi)=\frac{\delta\sin(2\theta)}{\tilde{\Omega}_\text{B}-\delta\cos(2\theta)}\,.
\end{equation}
The difference between the two diagonal elements of the resulting Hamiltonian gives the oscillation frequency for phase evolution of the two dressed states, $\tilde{\Omega}_\text{D}=\sqrt{\Omega_\text{B}^2+\Delta_\text{D}^2}$, where $\Delta_\text{D}=\Delta_\text{B}-\delta$ is the modified detuning, justifying our identification of $\Delta_\text{B}$ with the light shift.
\par
By concatenating the unitary transformations, we can rewrite the bare states in terms of these final dressed states, giving an expression of the form
\begin{equation}
\cos(\theta+\phi)d_1-\sin(\theta+\phi)d_2e^{\i\tilde{\Omega}_\text{D} t}
\end{equation}
for the amplitude of $\ket{g_1}$, where $d_1$ and $d_2$ are the initial amplitudes of the two dressed states. Immediately, we note that the oscillation between $\ket{g_1}$ and $\ket{g_2}$ has a peak-to-peak amplitude that is bounded by
\begin{equation}
\sin[2(\theta+\phi)]=\frac{\Omega_\text{B}}{\sqrt{\Omega_\text{B}^2+\Delta_\text{D}^2}}
\end{equation}
This envelope function describes a power-broadened Lorentzian, centred on the light-shifted frequency difference between the two ground states. It represents the maximum possible population transfer, and any oscillation will be contained within this envelope.
\chapter{Trapping and cooling atoms}\label{ch:CoolingMethods:TrapCool}
\epigraph{There is room for one further general remark. [...] [I]n general one restricts oneself to a discussion of the \emph{energy} exchange, without taking the \emph{momentum} change into account. One feels easily justified in this, because the smallness of the impulses transmitted by the radiation field implies that these can almost always be neglected in practice [...].}{A.\ Einstein, Physikalische Zeitschrift \textbf{18}, 121 (1917)}
The general description given previously of the forces acting on two-level atoms allows the exploration of a number of laser trapping and cooling configurations currently used. In particular, I will look at dipole traps in \Sref{sec:CoolingMethods:DT}, optical molasses in \Sref{sec:CoolingMethods:OM}, and the magneto--optical trap (MOT) in \Sref{sec:CoolingMethods:MOT}. Following these, I will discuss a more recent attempt at a generally applicable laser cooling method, so-called `mirror-mediated cooling', \Sref{sec:CoolingMethods:MMC}, which naturally lends itself to being extended in various ways, as shall be seen in \Sref{sec:CoolingMethods:Other} and \pref{part:TMM}.
\section{Dipole traps}\label{sec:CoolingMethods:DT}
We have already remarked that the `reactive force' in \eref{eq:FrictionReactDiss} is often called the dipole force. The dipole force is proportional to the gradient of the magnitude of the Rabi frequency and is therefore sensitive to spatial nonuniformities in the electric field intensity that the atom is interacting with. In other words, an atom immersed in a tightly focussed light field will experience a force that attracts it to, or repels it from, the focus. Specifically, if $\Delta_\text{L}<0$ the force will point towards increasing light intensity; conversely if $\Delta_\text{L}>0$, the force points away from the focus. In the former case, the atom can be trapped at the focus of the beam, whereas in the latter case configurations can be found having a field minimum at some point in space~\cite{Freegarde2002}, raising the possibility of trapping the atom in such regions.\\
The dipole force is a very general mechanism; it applies not only to TLAs, as explained in the preceding paragraph, but to any object that has a nonzero polarisability. Let us consider an object with an induced dipole moment $\v{d}$ interacting with an electric field $\v{\mathcal{E}}$. Then, the object will experience a (dipole) potential $U=-\v{d}\cdot\v{\mathcal{E}}$. This potential then gives rise to the dipole force $\v{\textsl{\textsf{F}}}=-\v{\nabla}U$ in a manner similar to the force experienced by a TLA. We note that this derivation of $\v{\textsl{\textsf{F}}}$ implies that the dipole force is conservative; essentially identical arguments to those used when describing the TLA in \Sref{sec:CoolingMethods:ForceTLA} hold in the general case too, whereby a delay between $\v{d}$ and $\v{\mathcal{E}}$ can give rise to a nonconservative term in the dipole force.\par
This universality inherent in the dipole force makes it a very versatile experimental tool. It has been used to trap atoms in free space~\cite{Dumke2002} and cavities~\cite{Mucke2010}, manipulate microspheres~\cite{Rodrigo2006}, viruses and even living cells~\cite{Ashkin1997}; it is also essential in achieving fully optical Bose--Einstein condensation~\cite{Barrett2001}.
\section{Optical molasses}\label{sec:CoolingMethods:OM}
Optical molasses historically provided the first proof that purely optical mechanisms can be used to slow down the motion of ensembles of atoms~\cite{Chu1985}. Let us see how optical molasses work by looking at the behaviour of a TLA inside a standing wave composed of two weak ($s\ll 1$) counterpropagating travelling waves. \eref{eq:FrictionReactDiss} is again the key to exploring this interaction. For each travelling wave, having a propagation vector $\pm\v{k}$ [i.e., $\phi(\v{r})=\pm\v{k}\cdot\v{r}$], we can easily see that $\v{\nabla}\Omega=0$---we are thus only concerned with the dissipative force---and $\v{\nabla}\phi(\v{r})=\pm\v{k}$. Adding the two forces together thus seems to give a zero net force, but this is only because the Doppler shift has not yet been taken into account. Indeed, the frequency for the $\pm\v{k}$ wave is seen by the atom to be shifted by $\pm\v{k}\cdot\v{v}$, i.e., $\Delta_\text{L}\rightarrow\Delta_\text{L}\pm\v{k}\cdot\v{v}$. In effect, then, the total force seen by the atom is therefore
\begin{align}
\label{eq:OMForce}
\v{\textsl{\textsf{F}}}_\text{OM}&=-\frac{\hbar\Omega^2\Gamma/2}{\bigl(\Delta_\text{L}+\v{k}\cdot\v{v}\bigr)^2+\Gamma^2}\v{k}+\frac{\hbar\Omega^2\Gamma/2}{\bigl(\Delta_\text{L}-\v{k}\cdot\v{v}\bigr)^2+\Gamma^2}\v{k}\nonumber\\
&\approx\frac{2\hbar\Omega^2\Delta_\text{L}\Gamma}{\bigl(\Delta_\text{L}^2+\Gamma^2\bigr)^2}(\v{k}\cdot\v{v})\v{k}\nonumber\\
&\phantom{\approx}\ =\frac{2\hbar\Omega^2\Delta_\text{L}\Gamma}{\bigl(\Delta_\text{L}^2+\Gamma^2\bigr)^2}(\v{k}\otimes\v{k})\v{v}\,,
\end{align}
with the approximation holding only up to linear order in $\v{k}\cdot\v{v}\Gamma\big/\bigl(\Delta_\text{L}^2+\Gamma^2\bigr)$, and the vector outer product being defined as
\begin{equation}
\v{k}\otimes\v{k}=\begin{pmatrix}
k_1\\
k_2\\
k_3
\end{pmatrix}\otimes\begin{pmatrix}
k_1\\
k_2\\
k_3
\end{pmatrix}=\begin{bmatrix}
k_1^2&k_1k_2&k_1k_3\\
k_1k_2&k_2^2&k_2k_3\\
k_1k_3&k_2k_3&k_3^2
\end{bmatrix}\,.
\end{equation}
Let us now constrain ourselves to one dimension, whereby $\v{\textsl{\textsf{F}}},\v{v},\v{k}\rightarrow\textsl{\textsf{F}},v,k$. For every photon absorption or emission event that the atom undergoes, it experiences a momentum change $\hbar k$; this process occurs at a rate $2R$, with the factor of $2$ arising because of the presence of the two identical beams, and $R$ given by \eref{eq:ScatRate}. Thus, over small times $t$, we have $(\delta p)^2/t=2\hbar^2k^2R$. Identifying $\delta p=\sigma_p(t)$, we therefore have the diffusion and cooling coefficients,
\begin{equation}
\label{eq:OMDiffn}
D=2\hbar^2k^2R=\frac{\hbar^2k^2\Omega^2}{\Delta_\text{L}^2+\Gamma^2}\,,\text{ and }\varrho=-\frac{2\hbar k^2\Omega^2\Delta_\text{L}\Gamma}{\bigl(\Delta_\text{L}^2+\Gamma^2\bigr)^2}\,,
\end{equation}
respectively (cf. Ref.~\cite{Gordon1980}), whereby the equilibrium temperature of the atom is given by
\begin{equation}
T=\frac{\hbar\Gamma}{2k_\text{B}}\biggl(\frac{\lvert\Delta_\text{L}\rvert}{\Gamma}+\frac{\Gamma}{\lvert\Delta_\text{L}\rvert}\biggr)\,,
\end{equation}
which only makes sense for $\Delta_\text{L}<0$. $T$ attains its lower bound, known as the Doppler temperature $T_\text{D}$, when $\Delta_\text{L}=-\Gamma$, at which point
\begin{equation}
T_\text{D}=\frac{\hbar\Gamma}{k_\text{B}}\,.
\end{equation}
Remarkably, this limiting temperature is independent of the atom's mass $m$ or the intensity of the light beam (recall, however, that we have assumed $s\ll 1$).\par
A more fundamental limit than the Doppler temperature is due to what is called the recoil limit; this gives rise to the recoil temperature $T_\text{R}$. Any time an atom absorbs or emits a photon, it undergoes a momentum change of magnitude $\Delta p=\hbar k$. As such, then, the momentum of the atom can be described as a discrete one-dimensional random walk of step size $\pm\hbar k$. After $n$ steps in such a walk, the momentum $p=S_n\hbar k$ of the atom is described by the quantity $S_n=\sum_{i=1}^n c_i$, where each $c_i=\pm 1$. The statistics of $S_n$ obeys
\begin{equation}
\expt{S_n}=\sum_{i=1}^n \expt{\pm 1}=0\,\text{ and }\,\expt{S_n^2}=\sum_{i=1}^n \expt{(\pm 1)^2}=n\,,
\end{equation}
whereby
\begin{equation}
\sigma_p^2=\bigl(\expt{S_n^2}-\expt{S_n}^2\bigr)(\hbar k)^2=n(\hbar k)^2\,,
\end{equation}
such that the momentum diffusion experienced by the atom is $\textsl{\textsf{D}}=(\hbar k)^2(n/t)$; the quantity $(n/t)$ has to be interpreted as the rate at which photons interact with the atom. The recoil limit applies to any cooling process involving scattering of photons, and therefore at least one scattering event has to occur on average during such a process. The minimum value for $\sigma_p$ is therefore achieved when $n=1$; $\sigma_p=(\hbar k)^2$. Assuming the momentum distribution is Maxwell--Boltzmann,\footnote{This assumption is not very well-founded; ensembles close to, or below, the recoil limit generally obey different statistics (see, for example, Refs.~\cite{Anderson1995} and~\cite{Davis1995}, and Fig.~2 in Ref.~\cite{Aspect1988}). Nonetheless, in the absence of well-definition, assuming Maxwell--Boltzmann statistics allows us to assign a `temperature' to such an ensemble.} this corresponds to a temperature
\begin{equation}
T_\text{R}=\frac{\hbar^2k^2}{2k_\text{B}m}\,.
\end{equation}
For a \textsuperscript{85}Rb atom cycling on the D$_2$ transition, $T_\text{D}=146$\,$\upmu$K and $T_\text{R}=185$\,nK (see Ref.~\cite{Steck2008}, but their definition of the recoil temperature is a factor of $2$ larger than ours; we use a definition consistent with, e.g., Ref.~\cite{Aspect1988}). Not even the recoil limit is a hard limit, though, with several schemes---such as evaporative cooling~\cite{Leanhardt2003}, and even all-optical schemes like velocity-selective coherent population trapping~\cite{Aspect1988}---being devised to overcome it. Such schemes generally work by ensuring that the target state is subject to much less than one scattering event, on the average.
\section{Magneto--optical traps}\label{sec:CoolingMethods:MOT}
To describe MOTs, we have to give up the TLA model we have pursued throughout most of the preceding sections and appeal to the multilevel structure of realistic atoms. In particular, let us consider a $J=1\rightarrow J^\prime=2$ transition.\footnote{We use the standard notation here, with $J$ denoting the ground state and $J^\prime$ the excited state.} Our description of the cooling mechanism that dominates in a MOT will be cursory here---to present a unified approach to laser cooling, we will in \Sref{sec:TMM:Multilevel} treat this type of cooling using an extended form of the transfer matrix method that will be introduced in \pref{part:TMM}.\par
The `sub-Doppler' cooling mechanism in a MOT relies on the phenomenon of optical pumping between the multiple magnetic sublevels in the ground ($J=1$) state. Let us assume that the atom is interacting with two counterpropagating travelling waves of identical frequencies but opposite circular polarisations; this is often referred to as the $\sigma^+$--$\sigma^-$ configuration. The resulting standing wave pattern is linearly polarised everywhere, but this linear polarisation rotates in space on the wavelength scale. Consider now a ground-state static atom in such a field. Optical pumping processes between different magnetic sublevels of the $J=1$ level proceed at rates given by the various Clebsch--Gordan coefficients for the transitions involved. We note that this mechanism relies on pumping between these magnetic sublevels and therefore needs a ground state with $J\geq 1$. The physics behind cooling in this configuration~\cite{Dalibard1989} is rather involved but we can summarise it as follows. In its rest frame, an atom moving with (constant) velocity sees a linear polarisation rotating with a constant angular velocity which, by Larmor's theorem, acts like a fictitious magnetic field parallel to the motion of the atom. This magnetic field acts to effectively couple the magnetic sublevels of the ground state. The motion of the atom also induces an imbalance in the populations of these sublevels, causing it to preferentially scattering photons from one of the two running waves. It can be shown that, under the right conditions, this preferential scattering can lead to a significant friction force acting on the atom, enough to cool a population of such atoms to well below the Doppler temperature~\cite{Lett1988,Dalibard1989,Ungar1989}.\\
We can now describe the theory of operation of MOTs. A MOT is formed at the zero of a quadrupole magnetic field; let us use this magnetic field to define our coordinate system: the $z$-axis points along the axis of symmetry of this magnetic field and the $x$ and $y$ axes define what we will call the `symmetry plane'. The origin of our coordinate system will be taken at the zero of the magnetic field. Traditionally, a MOT requires three pairs of counterpropagating red-detuned circularly-polarised beams, one pair along each of the coordinate axes. The beams in the symmetry plane all have the same helicity, and opposite to that of the beams along the $z$-axis. The polarisation of the beams is as yet undetermined; sub-Doppler cooling occurs for either configuration, depending on the sign of the detuning of the cooling light from the atomic transition. This degree of freedom is therefore fixed by choosing whether to operate above or below atomic resonance.
\par
\begin{figure}
\centering
\includegraphics[scale=0.5]{Diagrams/MOT_Trapping.pdf}
\caption[Trapping in a magneto--optical trap]{Close to the centre of a magneto--optical trap, the magnetic field has a constant gradient, taken here to be positive along the positive $x$ direction. An atom to the right of the origin will see the $\sigma^+$ beam as further detuned than the $\sigma^-$ beam, even though both have the same frequency $\omega_\text{L}$, and therefore preferentially scatter from the latter, causing it to experience a restoring force towards $x=0$. A similar principle operates when the atom is to the left of the origin, whereby it will preferentially scatter from the $\sigma^+$ beam. We depict a $J=0\to J^\prime=1$ transition here. The designation $\sigma^\pm$ relates to the effective circular polarisation of the light, and \emph{not} its helicity.}
\label{fig:MOT_Trapping}
\end{figure}
We have thus far discussed only the cooling mechanism in a MOT. Trapping in a MOT operates as follows. Let us suppose that the magnetic field is such that the gradient is positive in the $x$ direction, and consider a static atom with a positive $x$ coordinate. Due to the nonzero magnetic field, the energy levels of the atom will be split. We will now, for simplicity, consider an atom with a $J=0$ ground state (similar conclusions will hold for $J\neq 0$). The atom is in a region where the magnetic field strength is positive; in other words, the $\Delta m_J=-1$ ($\Delta m_J=+1$) transition is shifted to a lower (higher) energy; this statement follows from \eref{eq:MagHamiltonian}. If the counterpropagating beam (i.e., the beam from the \emph{positive} $x$ direction) is $\sigma^-$ polarised, whereby the opposite beam is $\sigma^+$ polarised, the atom will preferentially scatter photons from the positive $x$ direction and therefore experience a force towards the origin; see \fref{fig:MOT_Trapping}. Conversely, if the magnetic field gradient is negative in the $x$ direction, we require the beam from the positive $x$ direction to be $\sigma^+$ polarised.
\section[Memory-based approach to cooling in laser light]{Memory-based approach to cooling in laser light:~the dipole force delayed\footnote{With sincere apologies to J.\ Dalibard and C.\ Cohen-Tannoudji.}}\label{sec:CoolingMethods:Retarded}
We have stated previously that the dipole force, in its `bare' form, is a conservative force, but if we introduce a delay between the dipole moment of our particle and the field that the particle interacts with, this restriction evaporates and the dipole force gains a non-conservative character~\cite{Braginsky1977}. In the atomic domain, this is generally referred to as a `Sisyphus'--type cooling mechanism, following the nomenclature first applied by Aspect and co-workers~\cite{Aspect1986} with reference to the ancient Greek myth. We will now generalise this idea and see that it leads us to several potentially promising cooling configurations that rely solely upon the dipole force.
\par
In the previous section we considered `standard' Sisyphus cooling. A similar mechanism can be seen to act for atoms inside optical resonators~\cite{Hechenblaikner1998}. Let us consider a simplified model of a resonant optical resonator having almost perfect mirrors. The electric field inside the resonator is, to a very good approximation, a sinusoidal standing wave. If one introduces an atom into the resonator, the effect that the atom has on the resonator will depend greatly on its position in this standing wave. An atom at a node of the field will not affect the field, whereas an atom at an antinode will affect the field rather more strongly.\footnote{A similar point, with implications for photonic crystals, is made about atoms in an optical lattice in Ref.~\cite{Deutsch1995}.} In such a picture, the atom is constantly being pumped into a dressed-state from which it preferentially decays in such a way as to lose kinetic energy. We note that a very similar mechanism is in operation here as in traditional Sisyphus cooling but---crucially---the delay process has been shifted from an internal delay mechanism (optical pumping between the different magnetic sublevels of the atom) to an external one (optical pumping between the atom--cavity dressed states), since the response of the cavity field is necessarily a viscous one.
\par
To illustrate the generality of this idea, let us consider the potential energy $U[x,x_\text{a}(t),v]$ at time $t$ on a test particle at position $x$ due to an atom following a path $x_\text{a}(t)=x_0+vt$. We suppose that the dependence of $U$ on the motion of the atom is mediated entirely through the action of a `memory' in the system, or another time-delayed effect, that can be represented by
\begin{equation}
U[x,x_\text{a}(t),v]=\int_0^\infty U[x,x_\text{a}(t-T),v=0]\,M(T)\,\mathrm{d} T\,,
\end{equation}
where $M$ is some function that represents the memory of the system. We can Taylor-expand $U$ in the second variable around $x_\text{a}(t)$ to first order in $v$ and write
\begin{align}
U[x,x_\text{a}(t),v]&=U[x,x_\text{a}(t),v=0]\int_0^\infty M(T)\,\mathrm{d} T\nonumber\\
&\phantom{=}\quad-v\biggl[\frac{\partial U(x,y,v=0)}{\partial y}\biggr|_{y=x_0}\biggr]\int_0^\infty T\,M(T)\,\mathrm{d} T\,.
\end{align}
Finally, we assume that $M$ is normalised, $\int_0^\infty M(T)\,\mathrm{d} T=1$, and set
\begin{equation}
\tau=\int_0^\infty T\,M(T)\,\mathrm{d} T\,,
\end{equation}
which has the units of time. Thus, we obtain
\begin{equation}
\label{eq:MemoryFunctionforU}
U[x,x_\text{a}(t),v]=U[x,x_\text{a}(t),v=0]-\tau v\biggl[\frac{\partial U(x,y,v=0)}{\partial y}\biggr|_{y=x_\text{a}(t)}\biggr]\,.
\end{equation}
A similar relation holds for any distributive functional of $U$.\footnote{That is, any $\mathcal{F}$ such that $\mathcal{F}(\alpha U+\beta V)=\alpha\mathcal{F}(U)+\beta\mathcal{F}(V)$ for real $\alpha$ and $\beta$.} In the prototypical Sisyphus mechanism, the system memory lies in the delayed populations. Indeed, if we concentrate on one sublevel having population $\Pi(x,v)\propto -U[x,x_\text{a}(t),v]$, then
\begin{equation}
\Pi[x_\text{a}(t),v]=\Pi^\text{st}[x_\text{a}(t)]-\tau_\text{p}v\biggl[\frac{\partial \Pi^\text{st}(y)}{\partial y}\biggr|_{y=x_\text{a}(t)}\biggr]\,,
\end{equation}
cf. Eq.~(4.22) in Ref.~\cite{Dalibard1989}, with $\tau_\text{p}$ being the \emph{pumping time} associated with our choice of memory function $M(T)=\delta(T-\tau_\text{p})$, or equivalently $M(T)=\exp(-t/\tau_\text{p})/\tau_\text{p}$, and $\Pi^\text{st}(y)\propto -U[x_\text{a}(t),y,v=0]$ the steady-state population when the atom is static.\\
Consider now a situation where the interaction between the particle and the field is mediated through the dipole force, such that the force acting on the particle is
\begin{equation}
\textsl{\textsf{F}}[x_\text{a}(t),v]=-\frac{\partial U[x,x_\text{a}(t),v]}{\partial x}\biggr|_{x=x_\text{a}(t)}\,.
\end{equation}
By using this relation in \eref{eq:MemoryFunctionforU}, we obtain
\begin{equation}
\label{eq:GeneralTimeDelayedF}
\textsl{\textsf{F}}[x_\text{a}(t),v]=\textsl{\textsf{F}}_0[x_\text{a}(t)]+\tau v\biggl[\frac{\partial^2U(x,y,v=0)}{\partial x\,\partial y}\biggr|_{x=x_\text{a}(t),y=x_\text{a}(t)}\biggr]\,,
\end{equation}
with $\textsl{\textsf{F}}_0$ being the force acting on the particle when it is static. \eref{eq:GeneralTimeDelayedF} may not be trivial to evaluate in practice, and in other circumstances the model itself may not apply; for a given situation $\tau$ may be ill-defined, for example. Nevertheless, it allows us to make a general and powerful prediction that applies to a wide range of systems:~by simply endowing a system with a memory, the dipole force can be used to cool the motion of the particle interacting with the system. It also provides a physical link between models that use an external memory, such as a cavity field, and those that use an internal memory, such as populations in different hyperfine levels.
\section{Cavity fields and atomic motion:~A brief review of current work}
The first few sections of this chapter discussed various ways of capturing atoms or slowing their motion down. In the sections following this, we will explore novel methods that use an optical memory, cf. \sref{sec:CoolingMethods:Retarded}, to achieve cooling. At this stage, therefore, it would be good to briefly summarise the existing work that uses cavities to slow atoms down or otherwise control their motion. Our later work will borrow heavily from the ideas presented here.
\subsection{Cavity-mediated cooling}
The electromagnetic field inside a cavity differs from that in free space in several respects. On the most basic level, the zeroth-order approximation of a (resonant) cavity field is a region in physical space where the electromagnetic field is, for a given driving power, stronger than in other locations. Such a coarse approximation, however, has limited applicability.\footnote{This point is, of course, arguable: the strong coupling of atoms to the electromagnetic field is an interesting field of research with several applications~\cite{Teo2010}.} A first-order approximation to a cavity field would include memory effects which, as we have seen above, imply that the cavity field can be used to control atomic motion. Quantum mechanically, however, the field inside a good cavity is modified in a highly nontrivial way, for even the vacuum fluctuations themselves are modified. This phenomenon is the origin of the Purcell effect~\cite{Purcell1946} and was recognised in the early 1990's as a way of inducing cooling forces on atoms~\cite{Lewenstein1993}. The theory of cavity-mediated cooling of atoms subsequently developed in the direction of driven high-$Q$ cavities~\cite{Cirac1993}. Ref.~\cite{Horak1997} was the first to point out evidence for a novel cooling process taking place inside such cavities that is reminiscent of Sisyphus--type sub-Doppler cooling mechanisms (cf. Ref.~\cite{Dalibard1989}, but see also above):~the basis of the mechanism here being that the photon number in a cavity is, in the strong coupling regime, intimately connected with the position of the atom inside the cavity. By choosing the right system parameters, this connection can be used to slow down the atomic motion. Cavity-mediated cooling of atoms was observed only in recent experiments~\cite{Leibrandt2009,Koch2010}. The potential application of this mechanism to the cooling of the motion of dielectric particles has also been noted~\cite{Barker2010,RomeroIsart2011} and progress is rapidly being made towards the experimental realisation of cavity cooling of microscopic dielectric `particles', both in the form of microspheres~\cite{Barker2010} and in the form of micromirrors~\cite{Kippenberg2008,Aspelmeyer2010}, with profound implications for the study of the foundations of quantum mechanics and even relativity~\cite{Li2011}.\par
Our work in later sections, in particular \sref{sec:TMM:ECCO}, will explore the use of cavities in a different way. By placing a particle outside a cavity, we can still benefit from the presence of the cavity in lengthening the time delay, but without subjecting the particle to the strong field present inside a resonant cavity.
\subsection{Ring cavity cooling}
Investigations of cavity cooling inside Fabry--P\'erot cavities have focussed on the `good-cavity' limit, where the bare cavity has a very high finesse or $Q$-factor. Ring cavities, by their very topology, tend to be larger and with a lower $Q$-factor. This has, perhaps, been one reason why cooling of atoms inside ring cavities has not been explored as thoroughly as cavity-mediated cooling. One distinct advantage of ring cavities is that the system is translationally invariant with respect to the position of the atom; any friction forces developed on the atom are subsequently not dependent on the position of the atom (this will be seen to be a serious limitation of cavity-mediated cooling in \sref{sec:Localisation}). The theory for cooling atoms inside ring cavities~\cite{Gangl2000a,Elsasser2003,Nagy2006,Hemmerling2010,Schulze2010,Niedenzu2010} is similar to that for cavity-mediated cooling. Experimental work on atoms inside ring cavities is rather sparse and tends to focus on recoil-induced effects~\cite{Kruse2003,Slama2007}, which will be discussed next. It has also been proposed (see Refs.~\cite{Vuletic2001a} and~\cite{Salzburger2006}; see also \sref{sec:TMM:AmplifiedOptomechanics}) that the practical limitation of ring cavities as low-$Q$ devices can be lifted by the use of a gain medium inside the ring cavity. This approach holds promise towards improving the performance of such cavity systems to the point that they may be useful in practical realisations of optomechanics experiments.
\subsection{Self-organisation of atoms inside cavities}
The strong feedback between atoms and cavity fields may, under the right conditions, result in self-organisation of an atomic ensemble inside a cavity. We mention here two distinct effects. In Fabry--P\'erot cavities, it was proposed~\cite{Domokos2002b} and subsequently observed~\cite{Baumann2010} that above a threshold pump power, an ensemble of atoms will organise into a checkerboard pattern. Two possible configurations are possible, with the atoms occupying either set of sites in the checkerboard, and transitions occur between the two due to the presence of noise in the system.\par
Collective atomic recoil lasing (CARL)~\cite{Bonifacio1994,Kruse2003,Zimmermann2004,Slama2007} is a related phenomenon that occurs inside ring cavities. CARL occurs when a pump beam inside a ring cavity is reflected off an ensemble of atoms into the counterpropagating mode. The motion of the atoms will Doppler-shift this reflection. An exponential build-up of this latter mode can then occur, accompanied by a spontaneous self-organisation of the ensemble into a grating-like structure. In such a setup, the atomic ensemble acts as a gain medium for the counterpropagating wave, and in its reliance on a gain medium CARL is related to the `amplified optomechanics' idea we discuss in \sref{sec:TMM:AmplifiedOptomechanics}. However, amplified optomechanics uses a gain medium that is spatially separated from the atomic ensemble, in a similar way to the ideas discussed in Ref.~\cite{Vuletic2001a}, whereas in CARL the two are one and the same.
\section{Mirror-mediated cooling}\label{sec:CoolingMethods:MMC}
\begin{figure}
\centering
\subfigure[Cavity-Mediated Cooling]{
\includegraphics[scale=0.5]{Diagrams/SpaceTime_CMC.pdf}
}\hspace{1cm}
\subfigure[Mirror-Mediated Cooling]{
\includegraphics[scale=0.5]{Diagrams/SpaceTime_MMC.pdf}
}
\caption[Space--time diagrams for cavity- and mirror-mediated cooling]{Space--time diagrams showing the interaction of an atom with light when the atom is (a)~inside an optical resonator, and (b)~in front of a mirror. It is clear that in case (a) one gets multiple interactions between the two and a `memory time' is not well-defined, whereas in case (b) a definite memory time can be assigned. The blank arrow indicates the pump beam.}
\label{fig:SpaceTime}
\end{figure}
As explained \sref{sec:CoolingMethods:Retarded}, the dipole force can be invested with a dissipative character by endowing the system as a whole with a memory. Exploring the cooling of atoms inside cavities using this terminology is fine in principle, since a cavity may be viewed as the archetypal optical memory element, but is not easy in practice. Indeed, in such a system, a well-defined ``delay time'' does not exist: a cavity has several memories, each one arising from a separate interaction of \emph{the same light} with the atom. By simplifying the system and removing one of the cavity mirrors, we can once again talk about such a concept as a delay time $\tau$. In this configuration, light interacts with the atom twice, at times $t=0$ and $t=\tau$, say, and no more.\footnote{This is a very `classical' picture, where we describe the light by means of point-like photons. This description is used here to illustrate the physics of the situation. Quantitative calculations can only be made by properly describing the light in terms of the solutions to Maxwell's equations.} These two pictures are illustrated schematically in \fref{fig:SpaceTime}. Whereas we can readily define the system memory time
\begin{equation}
\tau=t_2-t_1
\end{equation}
for \fref{fig:SpaceTime}(b), no such intuitive definition is possible in \fref{fig:SpaceTime}(a) because (i)~if the atom is not at the centre of the cavity, then $t_2-t_1\neq t_3-t_2\neq\dots$, and (ii)~each interaction has a different weighting due to the imperfect reflectivity of the mirrors; it is however possible to assign an effective memory time $\tau=1/(2\kappa)$, where $\kappa$ is the HWHM linewidth of the cavity. Indeed, it will be shown in \aref{sec:TMM:CavityKappaQ} that the finesse of a cavity can be written as
\begin{equation}
\mathcal{F}=\pi N\,,
\end{equation}
where $N$ is the number of round-trips made by the light inside the cavity before the intracavity intensity decays by a factor $1/e^2$. The dipole energy $U$ is proportional to the delayed electric field; so one would expect that the effect of the cavity field on $U$ decays by a factor $1/e$ over $N$ round-trips. It seems reasonable, therefore, to set
\begin{equation}
M(T)=\bigl(1-e^{-1/N}\bigr)\sum_{n\ \text{even}}\delta\bigl(T-nL_1/c)e^{-n/N}+\bigl(1-e^{-1/N}\bigr)\sum_{n\ \text{odd}}\delta\bigl(T-nL_2/c)e^{-n/N}\,,
\end{equation}
with $L_1$ and $L_2$ being the distances from the atom to the two cavity mirrors and $L=L_1+L_2$ the cavity length. This memory function represents an infinite number of discrete interactions, each one being weaker than the previous, of the cavity field with the atom. Thus, in the good-cavity limit of large $N$,
\begin{equation}
\tau=\frac{L_1+L_2}{c}\frac{1}{e^{1/N}-1}\to\frac{LN}{c}=\frac{\mathcal{F}L}{\pi c}\,.
\end{equation}
It will also be shown (see \aref{sec:TMM:CavityKappaQ}) that $\kappa=\pi c/(2\mathcal{F}L)$. It thus follows that $\tau=1/(2\kappa)$ is a good definition for the effective memory time of the system in this model. Throughout the rest of this chapter and in later chapters, we shall see how this memory time $\tau$ governs the interaction of the particle with the field.
\par
The remainder of this section will be devoted to the mathematical description of mirror-mediated cooling and follows closely Ref.~\cite{Xuereb2009a}. It was published as Xuereb, A., Horak, P. \& Freegarde, T. Phys.\ Rev.\ A \textbf{80}, 013836 (2009) and several parts are reproduced \emph{verbatim}. This paper was the first to introduce the `mirror-mediated cooling' mechanism, but was accompanied by another paper~\cite{Horak2010a} that discusses the more computationally-oriented aspects of the Monte--Carlo simulations performed.\\
A semiclassical model of the situation is presented in \Sref{sec:CoolingMethods:MMC:Model}. The model is then explored analytically in a perturbative manner (\Sref{sec:CoolingMethods:MMC:Perturbative}), and numerically using Monte--Carlo simulations (\Sref{sec:CoolingMethods:MMC:Numeric}). Finally, a few brief comments are made about some of the approximations made to describe the system.
\subsection{Mathematical model}\label{sec:CoolingMethods:MMC:Model}
We shall use the TLA model explored in detail in \cref{ch:CoolingMethods:AFInt}. The atom will be assumed to have a transition frequency $\omega_0$ and upper-state decay rate $2\Gamma$, as before, and be coupled to a continuum of quantised electromagnetic modes with frequencies $\omega$ and standing-wave mode functions
\begin{equation}
f(\omega,x)=\sin(\omega x/c)\,,
\end{equation}
with the mirror being at $x=0$. The field modes have annihilation and creation operators $\hat{a}(\omega)$ and $\hat{a}^\dagger(\omega)$, respectively, and the atom--field coupling is described by a single, frequency-independent, coupling coefficient $g$. The mode at frequency $\omega=\omega_\text{L}$ is pumped in a coherent field by a laser, and we assume that the detuning, defined as
\begin{equation}
\Delta_\text{L}=\omega_\text{L}-\omega_0\,,
\end{equation}
obeys the condition $\lvert\Delta_\text{L}\rvert\gg\Gamma$; i.e., we assume far off-resonant operation. The numerical examples we give will be for a \textsuperscript{85}Rb atom cycling in the $5\text{S}_{1/2}\leftrightarrow 5\text{P}_{3/2}$ transition.\\
The starting point for our description is the quantum master equation, \eref{eq:QME}, where the Hamiltonian now reads
\begin{equation}
\label{eq:MMC:Hamiltonian}
\hat{H}=\frac{\hat{p}^2}{2m}-\hbar\Delta\hat{\sigma}^+\hat{\sigma}^-+\int\hbar
(\omega-\omega_\text{L})\hat{a}^{\dagger}(\omega)\hat{a}(\omega)\,\mathrm{d}\omega-\i\hbar g\!\!\int[\hat{\sigma}^+\hat{a}(\omega)f(\omega,\hat{x}) - \text{H.c}]\,\mathrm{d}\omega\,,
\end{equation}
where $\hat{x}$ and $\hat{p}$ are the position and momentum operators of the atom and `$\text{H.c.}$' denotes the Hermitian conjugate. We also use the Liouvillian terms
\begin{equation}
\label{eq:MMC:Liouvillian}
\mathcal{L}\hat{\rho}=-\Gamma\bigg[\hat{\sigma}^+\hat{\sigma}^-\hat{\rho} +
\hat{\rho}\hat{\sigma}^+\hat{\sigma}^- -2\int_{-1}^{1}N(u)\hat{\sigma}^-e^{-\i u\hat{x}}\hat{\rho}\text{e}^{\i u\hat{
x}}\hat{\sigma}^+\,\mathrm{d} u\bigg]\,.
\end{equation}
In this expression, $N(u)$ describes the $1$D projection of the spontaneous emission pattern of the atomic dipole. In the low-saturation regime, we can adiabatically eliminate the internal atomic dynamics and formally express the dipole operator as
\begin{equation}
\hat{\sigma}^-=-\frac{\i\Delta+\Gamma}{\Delta^2+\Gamma^2}\,g\int f(\omega,\hat{x})\hat{a}(\omega)\,\mathrm{d}\omega+\hat{\xi}^-\,,
\end{equation}
where $\hat{\xi}^-$ is a noise term~\cite{Gardiner1984}.
\subsection{A perturbative approach to exploring the model}\label{sec:CoolingMethods:MMC:Perturbative}
\subsubsection{Force on a static atom}
The aim of this section is to derive an analytical expression for the friction force acting on the atom when the latter is not moving. To achieve this, we treat the atomic position classically, effecting the replacement $\hat{x}\to x_0$ in \eref{eq:MMC:Hamiltonian} to obtain
\begin{multline}
\hat{H}=\int\hbar(\omega-\omega_\text{L})\hat{a}^{\dagger}(\omega)\hat{a}(\omega)\,
\mathrm{d}\omega\\
+\hbar
\frac{g^2\Delta_\text{L}}{\Delta_\text{L}^2+\Gamma^2}\iint\sin(\omega_1x_0/c)\sin(\omega_2x_0/c)\hat{a}^{\dagger}(\omega_1)\hat{a}(\omega_2)\,\mathrm{d}\omega_1\,
\mathrm{d}\omega_2\,.
\end{multline}
A consequence of assuming $\lvert\Delta_\text{L}\rvert\gg\Gamma$, and operating in the low-saturation regime, is that the population of the excited state is negligible, and therefore last term in \eref{eq:MMC:Liouvillian} does not contribute to the dynamics. With this simplification, the master equation for the annihilation operators reduces to
\begin{equation}
\frac{\mathrm{d}}{\mathrm{d} t}\hat{a}(\omega,t)=\frac{\i}{\hbar}\bigl[\hat{H},\hat{a}(\omega,t)\bigr]\,.
\end{equation}
By substituting \eref{eq:MMC:Hamiltonian} into this equation, we obtain the integro-differential equation
\begin{equation}
\label{eq:MMC:StartingDE}
\frac{\mathrm{d}}{\mathrm{d} t}\hat{a}(\omega,t)=-\i(\omega-\omega_\text{L})\hat{a}(\omega,t)-\i\frac{g^2\Delta_\text{L}}{\Delta_\text{L}^2+\Gamma^2}\sin(\omega x_0/c)\int\sin(\omega_1x_0/c)\hat{a}(\omega_1,t)\,\mathrm{d}\omega_1\,.
\end{equation}
We now assume coherent states at all times for the fields and replace the operators with their respective expectation values.
Since we are pumping the atom at a single frequency, we take the initial condition $a(\omega,0)=A\,\delta(\omega-\omega_\text{L})$, where $A$ is the amplitude of the pump field, such that $|A|^2$ is the pump power in units of photons per second, and $\delta$ is the Dirac $\delta$-function. We now expand the fields $a(\omega,t)$ in the weak-coupling limit in powers of the coupling constant,
\begin{equation}
a(\omega,t) = \sum_n a_n(\omega,t)\biggl(\frac{g^2\Delta_\text{L}}{\Delta_\text{L}^2+\Gamma^2}\biggr)^n\,,
\end{equation}
with $a_n(\omega,t)$ being the $n$th coefficient of the series expansion. Solving \eref{eq:MMC:StartingDE} to successive orders in $g^2\Delta_\text{L}/\bigl(\Delta_\text{L}^2+\Gamma^2\bigr)$ yields the zeroth order term in this parameter,
\begin{equation}
\label{eq:MMC:a_0}
a_0(\omega,t)=A\delta(\omega-\omega_\text{L})\,,
\end{equation}
and the first order term
\begin{equation}
a_1(\omega,t)=A\frac{\exp\left[-\i(\omega-\omega_\text{L})t\right]-1}{\omega-\omega_\text{L}}\sin(\omega x_0/c)\sin(\omega_\text{L}x_0/c)\,.
\end{equation}
To the same level of approximation, we can now find the force acting on the atom to second order in $g^2\Delta_\text{L}/\bigl(\Delta_\text{L}^2+\Gamma^2\bigr)$.\footnote{Mathematically, the fact that the force is at a higher order arises from $\hat{H}$ not having any zeroth order terms [see \eref{eq:MMC:Hamiltonian}]; physically, the force is an interaction of an electric field with an induced dipole moment, and the electric field in this case is itself caused by the induced dipole at an earlier time. Both the induced dipole and this electric field are then, to lowest order, linear in $g^2\Delta_\text{L}/\bigl(\Delta_\text{L}^2+\Gamma^2\bigr)$. This term in the force is therefore quadratic in the same parameter.} The (classical) force acting on the atom is, then,
\begin{equation}
\textsl{\textsf{F}}(x_0)=-\frac{\partial H}{\partial x}\,,\text{ or}
\end{equation}
\begin{align}
\textsl{\textsf{F}}(x_0)&=\frac{\hbar}{c}\left|A\right|^2 \frac{g^2\Delta_\text{L}}{\Delta_\text{L}^2+\Gamma^2}\omega_\text{L}\bigg\{\sin(2\omega_\text{L} x_0/c)\nonumber\\
&\ \phantom{=\frac{\hbar}{c}\left|A\right|^2 \frac{g^2\Delta_\text{L}}{\Delta_\text{L}^2+\Gamma^2}\omega_\text{L}\bigg\{}-\frac{\pi}{2}\frac{g^2\Delta_\text{L}}{\Delta_\text{L}^2+\Gamma^2}\sin^2(\omega_\text{L} x_0/c)\big[4\cos^2(\omega_\text{L} x_0/c)-1
\big]\bigg\}\,.
\end{align}
The two terms making up this force have different origins. The first term represents the interaction between the dipole induced in the atom and the (unperturbed) pump field. The second term is the lowest-order correction to the force when the back-action of the atom on the light field is taken into account. The latter term therefore represents the interaction between the dipole induced in the atom by the pump field and the electric field propagated from this same dipole. What we shall see, in the next subsection, is that this propagated electric field can, after being reflected by the mirror, re-interact with the atomic dipole. The motion of the atom between these two interactions gives rise to a velocity-dependent force, i.e., a friction force, on the atom.
\subsubsection{Force on a moving atom: Friction forces}\label{sec:CoolingMethods:MMC:Friction}
Let us now assume that the atom is moving at a constant velocity, $v$. $\hat{x}$ is now a function of time, such that $\hat{x}\to x(t)=x_0+vt$, assuming that by $t=0$ the system has already reached steady-state; in other words, the pump beam has been on for a time longer than $\tau=2x_0/c$, which is the time taken for a disturbance to travel from the position of the atom to the mirror and back again. With this replacement for $\hat{x}$, we proceed to solve \eref{eq:MMC:StartingDE}, keeping terms up to first order in $g^2\Delta_\text{L}/\bigl(\Delta_\text{L}^2+\Gamma^2\bigr)$, as before. The friction force (in the longitudinal direction, i.e., along $x$) can finally be obtained:
\begin{equation}
\label{eq:MMC:AnalyticLongFrictionAll}
\textsl{\textsf{F}}_{\parallel}(x_0)=2\pi\hbar k_\text{L}\frac{v}{c}|A|^2\biggl(\frac{g^2\Delta_\text{L}}{\Delta_\text{L}^2+\Gamma^2}\biggr)^2\sin^2(2k_\text{L} x_0)
-\pi\hbar k_\text{L}^2v\tau|A|^2\biggl(\frac{g^2\Delta_\text{L}}{\Delta_\text{L}^2+\Gamma^2}\biggr)^2\sin(4k_\text{L} x_0)\,,
\end{equation}
with $k_\text{L}=\omega_\text{L}/c$ being the wavenumber of the pump field; this force lacks the usual `optical molasses' friction force that is produced by the interaction between the atomic dipole and the Doppler-shifted, unperturbed, pump beam because the atomic excited state has been adiabatically eliminated. The first term in the friction force expression is the velocity-dependent analogue of the second term in $\textsl{\textsf{F}}(x_0)$ above: it represents the interaction between the induced dipole of the atom and the instantaneous Doppler-shifted field produced by that same dipole. It is the second, `delayed', term, however, that we will be concerned with in the following. Indeed, this term represents the interaction between the atomic dipole and the reflected field it itself produces; the Doppler shift here can effectively be looked at as changing the phase accrued by the reflected field in arriving back at the atom.\\
The delayed friction force is larger than the instantaneous friction force by a factor $k_\text{L}x_0$, and we are therefore fully justified in setting
\begin{equation}
\label{eq:MMC:AnalyticLongFriction}
\textsl{\textsf{F}}_{\parallel}(x_0)=-\pi\hbar k_\text{L}^2v\tau|A|^2\biggl(\frac{g^2\Delta_\text{L}}{\Delta_\text{L}^2+\Gamma^2}\biggr)^2\sin(4k_\text{L} x_0)
\end{equation}
if we constrain ourselves to the far-field limit, $k_\text{L}x_0\ggg 1$.
\par
Supposing that the species we are cooling is $^{85}$Rb, and setting $|A|^2=62.5 \Gamma/(2\pi)$, $\Delta_\text{L}=-10\Gamma$,
$\tau=0.5/\Gamma$, and Gaussian beam waist $w=1.4\,\upmu$m, \eref{eq:MMC:AnalyticLongFriction} predicts $1/e$ cooling times of the order of $2$\,ms. The value for $\tau$ that we use implies a separation between the atom and the mirror of the order of several metres. We suggest that this problem can be overcome through the coupling of the light into an optical fibre, thereby avoiding the effects of diffraction. A recent experiment making use of a similar technique is described in Ref.~\cite{Kurtsiefer2009}.
\par
\eref{eq:MMC:AnalyticLongFriction} indicates an exponential decay or increase in velocity. We define the \emph{cooling coefficient} as $\varrho=-\textsl{\textsf{F}}/v$, whereby we obtain the relation $\mathrm{d} p^2/\mathrm{d} t=-2\varrho p^2$, which thus depends on position as $\sin(4k_\text{L}x_0)$. Moreover, since $p^2\propto T$ for a thermal ensemble, we also have $\mathrm{d} T/\mathrm{d} t=-2\varrho T$. \fref{fig:MMC:Analytic-Spatial} shows a plot of $\varrho$ against atomic position, where we introduced the coordinate $x_0^\prime$ relative to the nearest node of the standing wave pump. It is only in certain intervals that we expect the longitudinal force to be a damping force, as indicated in this figure by the shaded regions.
\begin{figure}[t]
\centering
\includegraphics[width=1.5\figwidth]{Diagrams/MMC/Analytic_x}
\caption[Longitudinal cooling coefficient in mirror-mediated cooling]{Spatial dependence of the longitudinal cooling coefficient $\varrho$ (thick solid line). The shaded areas promote cooling in the longitudinal direction. Also drawn is the transverse heating coefficient (thin solid line) and the field amplitude (dotted line). Parameters are for $^{85}$Rb atoms and $|A|^2=62.5 \Gamma/(2\pi)$, $\Delta_\text{L}=-10\Gamma$, $\tau=0.5/\Gamma$, $w=1.4~\upmu$m.}
\label{fig:MMC:Analytic-Spatial}
\end{figure}
\par
In order to complete the picture, we will now derive the friction force acting in the transverse direction (i.e., in a direction orthogonal to the pump beam). This force arises from the spatial variation of the coupling constant $g$ when the pump field is assumed to be a tightly-focussed Gaussian beam. In this case the coupling constant $g$ becomes a function $g(r)$, where $r$ is the coordinate in the transverse direction. For an atom moving at small constant velocity, we may write $g(r_0+vt)\approx g(r_0)+vtg^{\prime}(r_0)$, where $g^{\prime}(r)=\mathrm{d} g/\mathrm{d} r$ and time $t=0$, at which $r=r_0$, is defined as before. Substituting this into \eref{eq:MMC:StartingDE} we can derive an expression for the friction force, $\textsl{\textsf{F}}_{\perp}(x_0)=-\partial H/\partial r$, in the direction of $r$:
\begin{equation}
\textsl{\textsf{F}}_{\perp}(x_0)=-2\pi\hbar v\tau|A|^2\left(\frac{2gg^{\prime}\Delta_\text{L}}{\Delta_\text{L}^2+\Gamma^2}\right)^{\!\!2}\sin^3(k_\text{L}x_0)\cos(k_\text{L}x_0)\,,
\end{equation}
with $g$ and $g^\prime$ being evaluated at $r=r_0$. This transverse friction force is also shown in \fref{fig:MMC:Analytic-Spatial}, for comparison with the longitudinal friction force, assuming a Gaussian mode function of waist $w=1.4$\,$\upmu$m. Note that $\textsl{\textsf{F}}_{\parallel}$ and $\textsl{\textsf{F}}_{\perp}$ are comparable in magnitude for the parameters chosen here, i.e., where the mode waist is comparable to the optical wavelength. Moreover, we can see that there exist regions where both these forces promote cooling. In the remainder of this section, however, we will concentrate on a one-dimensional treatment of the problem and therefore only consider the longitudinal friction force.
\par
In terms of more familiar parameters, we can rewrite~\eref{eq:MMC:AnalyticLongFriction} in the limit $|\Delta_\text{L}|\gg\Gamma$ as
\begin{equation}
\label{eq:MMC:AnalyticLongFrictionFamiliar}
\textsl{\textsf{F}}_{\parallel}(x_0)=-vs\Gamma\frac{\sigma_{\text{a}}}{\sigma_\text{L}}\hbar k_\text{L}^2\tau\sin(4k_\text{L}x_0)\,,
\end{equation}
where $s=g^2|A|^2/(\Delta_\text{L}^2+\Gamma^2)$ is the maximum saturation parameter of the atom in the standing wave, $\sigma_{\text{a}}=3\lambda_\text{L}^2/(2\pi)$ is the atomic radiative cross-section at a wavelength $\lambda_\text{L}=2\pi/k_\text{L}$, as defined in \sref{sec:CoolingMethods:PolTLA}, and where we used the relation $2\pi g^2/\Gamma=4 \sigma_a/(2\sigma_\text{L})$, with $\sigma_\text{L}=\pi w^2/8$ being the mode area of the pump beam of waist $w$, and where the factor of $4$ arises from the standing wave amplitude.
\par
Aside from allowing us to make predictions of cooling times, \erefs{eq:MMC:AnalyticLongFriction}
and~(\ref{eq:MMC:AnalyticLongFrictionFamiliar}) also highlight the dependence of this cooling effect on the variation of important physical parameters. In particular, $\textsl{\textsf{F}}_{\parallel}$ depends on the square of the detuning, which means that it is possible to obtain cooling with both positive and negative detuning, in stark contrast with the standard Doppler cooling force. The friction force also scales with $w^{-4}$ and $|A|^2$. Hence, for a fixed laser intensity, proportional to $|A|^2/w^2$, i.e., for a fixed atomic saturation, the friction still scales with $w^{-2}$ and thus a tight focus is needed in order to have a sizeable effect. The physical reason for this dependence is that the dipole moment induced in the atom by the pump light is proportional to $|A|^2 g^2\propto|A|^2/w^2$; this polarisation couples to the electric field and, after being reflected by the mirror, polarises the atom again. The size of the force therefore scales as $|A|^2 g^4$, or $|A|^2/w^4$.\\
A very promising feature of these two equations is the linear dependence of the cooling rate on $\tau$: by increasing the distance between the atom and the mirror, we can increase the strength of the friction force acting on the atom. In \Sref{sec:CoolingMethods:MMC:Numeric} we further analyze the dependence of the cooling rate on the various parameters and support the validity of the analytical solution by comparing it with the results of simulations.
\subsubsection{Localising the atom: The effects of adding a harmonic trap}\label{sec:CoolingMethods:MMC:Perturbative:Dipole}
\eref{eq:MMC:AnalyticLongFriction} shows that, in order to observe any cooling effects, we need to localise the particle within around $\lambda_\text{L}/8$. This may be achieved, for example, by an additional far off-resonant and tightly focussed laser beam propagating parallel to the mirror and forming a dipole trap centred at a point $x_0$. In this section, we aim to characterise the effects of this dipole trap, or indeed any harmonic trap, on the atom--field interaction. We characterise this trap by means of its spring constant $k_\text{t}$, such that the trapping force is given by $\textsl{\textsf{F}}_\text{t}=-k_\text{t}(x-x_0)$, or equivalently by the harmonic oscillator frequency $\omega_\text{t}=\sqrt{k_\text{t}/m}$, where $m$ is the mass of the atom.
\par
With the atom being in a harmonic trap, we can now write down its time-dependent velocity as $v(t)=v_\text{m}\cos(\omega_\text{t}t)$, with a maximum velocity of the atom $v_\text{m}$. Using this expression, it is possible to derive a modified cooling coefficient by performing a perturbative expansion in the dimensionless parameter $v_\text{m}/c$. Proceeding along the lines of the preceding section, we arrive at a modified expression for the friction force,
\begin{equation}
\label{eq:MMC:AnalyticLongFrictionHarm}
\textsl{\textsf{F}}_{\parallel}(x_0)=-\pi\hbar
k_\text{L}^2v_\text{m}\tau\sinc(\omega_{\text{t}}\tau)|A|^2\biggl(\frac{g^2\Delta}{\Delta^2+\Gamma^2}\biggl)^2\sin(4k_\text{L}x_0)\,.
\end{equation}
By way of confirmation, it is easy to see that this expression reduces to \eref{eq:MMC:AnalyticLongFriction} in the limit of small $\omega_\text{t}$; i.e., in the free-particle limit for the atomic motion. The sinusoidal dependence on $\omega_\text{t}\tau$ can be justified in an intuitive manner: the effect on the particle is unchanged if the particle undergoes an integer number of oscillations in the round-trip time $\tau$.
\par
While \eref{eq:MMC:AnalyticLongFrictionHarm} was derived for an oscillating atom, it only accounts for the sinusoidal variation of the velocity, and therefore does not include the effect of the finite spatial distribution of the position of the atom. In order to obtain an estimate for the friction force in the presence of this spatial broadening, we calculate the overall energy loss rate experienced by the particle in terms of the time average of \eref{eq:MMC:AnalyticLongFriction}:
\begin{multline}
\label{eq:MMC:AnalyticFrictionTimeAvg}
\left\langle\frac{\mathrm{d} p^2}{\mathrm{d} t}\right\rangle=-\frac{\hbar
k_\text{L}^2p_0^2}{m}\tau|A|^2\biggr(\frac{g^2\Delta}{\Delta^2+\Gamma^2}\biggr)^2\int_0^{2\pi}\sin\bigl[4k_\text{L}x_0+4k_\text{L}x_\text{m}\sin(T)\bigr]\cos^2(T)\,\mathrm{d} T\,,
\end{multline}
where $p_0=m v_\text{m}$ is the maximum momentum of the particle in the trap, given by $p_0=x_\text{m}\sqrt{m k_\text{t}}$. The value of the integral in \eref{eq:MMC:AnalyticFrictionTimeAvg} can be expressed as
\begin{equation}
\frac{2\pi}{4k_\text{L}x_\text{m}}\bigl[\sin(4k_\text{L}x_0)J_1(4k_\text{L}x_\text{m})+\cos(4k_\text{L}x_0)H_1(4k_\text{L}x_\text{m})\bigr]\,,
\end{equation}
where $J_1$ is the order-$1$ Bessel function of the first kind and $H_1$ is the order-$1$ Struve function~\cite{Gradshteyn1994}. At the point of maximum friction, which occurs at $x_0=-3\lambda_\text{L}/16+n\lambda_\text{L}$ for some integer $n$, the integral in the above equation reduces to $2\pi J_1(4k_\text{L}x_\text{m})/(4k_\text{L}x_\text{m})$; this function can be readily evaluated numerically for a given trap frequency.
For small values of the trap frequency, the effect of the above averaging process is to introduce a factor of $\tfrac{1}{2}$ into \eref{eq:MMC:AnalyticLongFrictionHarm}. The result can be seen as being physically equivalent to the effect of cooling only one degree of freedom when the atom is in a harmonic trap. Finally, we combine the above two ideas to include both the effects of harmonic oscillation and the spatial extent of the atomic motion into \eref{eq:MMC:AnalyticFrictionTimeAvg}. This can be done, effectively, by replacing $\tau\to\sin(\omega_{\text{t}}\tau)/\omega_{\text{t}}$. The resulting approximate expression for the friction force, taking into account the periodicity in the time delay as well as spatial averaging effects, becomes
\begin{multline}
\label{eq:MMC:AnalyticLongFrictionHarmReduced}
\left\langle \textsl{\textsf{F}}_{\parallel}(x_0)\right\rangle=-\tfrac{1}{2}\hbar k_\text{L}^2v_\text{m}\tau|A|^2\biggl(\frac{g^2\Delta}{\Delta^2+\Gamma^2}\biggr)^2\sinc(\omega_\text{t}\tau)\\
\times\int_0^{2\pi}\sin\bigl[4k_\text{L}x_0+4k_\text{L}x_\text{m}\sin(T)\bigr]\cos^2(T)\,\mathrm{d}
T\,.
\end{multline}
\subsubsection{Momentum capture range}\label{sec:CoolingMethods:MMC:Perturbative:Cutoff}
\begin{figure}[t]
\centering
\includegraphics[width=1.5\figwidth]{Diagrams/MMC/CutoffExpln}
\caption[Momentum cut-off in mirror-mediated cooling]{Dependence of $\mathrm{d} p^2/\mathrm{d} t$ on the square of the initial momentum, $p_0^2$, for $\omega_\text{t}=0.45\times 2\pi\Gamma$ and $x_0^\prime=3\lambda_\text{L}/16$. Other parameters are as in \fref{fig:MMC:Analytic-Spatial}. Cooling is achieved only for a finite range of initial momenta.}
\label{fig:MMC:CutoffExpln}
\end{figure}
As discussed above, the addition of the dipole trap introduces several features into the friction force. Plotting the variation of the friction force in \eref{eq:MMC:AnalyticFrictionTimeAvg} with the initial momentum, $p_0$, of the atom, as in \fref{fig:MMC:CutoffExpln}, shows that the amplitude of the force is not monotonic. In fact, it increases from zero, for increasing $p_0$, achieves a maximum, and then decreases again until it reaches zero at some value for $p_0$. Physically, this is due to the broader spatial distribution for faster particles in the harmonic trap. For fast enough velocities, the particle oscillates into the heating regions, as can be deduced from \fref{fig:MMC:Analytic-Spatial}, even if the trap is centred at the position of maximum cooling. This defines a range of initial momenta, starting from zero, within which a particle is cooled by this mechanism; faster particles are heated and ejected from the trap. Note that this result was derived from the friction to lowest order in velocity $v$, and higher-order terms are expected to affect the capture range for high values of $p_0$.
\par
At particular values of $x_0^\prime$, e.g. at $-3\lambda_\text{L}/16$, this capture range can be conveniently estimated by using the location of the first zero of the Bessel function, giving a momentum capture range
\begin{equation}
\label{eq:MMC:CaptureRangeBesselZero}
p_\text{m}\approx 0.958\sqrt{m k_\text{t}}/k_\text{L} = 0.958 m \omega_\text{t}/k_\text{L}\,.
\end{equation}
Thus, $p_\text{m}^2 \propto \omega_\text{t}^2$, for values of $\omega_\text{t}$ that are not too large, and the capture range as defined in \fref{fig:MMC:CutoffExpln} is expected to scale with the square of the trap frequency. We will compare this later, in \Sref{sec:CoolingMethods:MMC:Numeric}, with the results of Monte--Carlo numerical simulations.
\subsubsection{Diffusion and steady-state temperature}\label{sec:CoolingMethods:MMC:SST}
\begin{figure}[t]
\centering
\includegraphics[width=1.5\figwidth]{Diagrams/MMC/SST_vs_Posn}
\caption[Steady-state temperature in mirror-mediated cooling]{Calculated steady-state temperature $T_\text{M}$ for an atom confined in a harmonic trap as a function of position whilst keeping the detuning and pump field constant. $\omega_\text{t}=0.1\times 2\pi\Gamma$; other parameters are as in
\fref{fig:MMC:Analytic-Spatial}.}
\label{fig:MMC:SST_vs_Posn}
\end{figure}
The main result of the preceding discussion was a friction force, which of course cools an atom towards zero momentum in the absence of any other effects. Momentum diffusion due to spontaneous scattering, in a manner similar to that discussed in \sref{ch:CoolingMethods:FDT}, introduces a constant in the equation for $\mathrm{d} p^2/\mathrm{d} t$ and results in a constant upward shift of the curve in~\fref{fig:MMC:CutoffExpln}. This slightly reduces the capture range for fast particles, but its main effect is to introduce a specific value of the momentum where friction and diffusion exactly compensate each other in the small $p_0$ range. This point corresponds to the steady-state temperature achievable through the cooling mechanism discussed here.
\par
To lowest order in the coupling frequency $g^2$, the momentum diffusion is given by the interaction of the atom with the unperturbed, standing-wave pump field. In this limit, diffusion in our system is therefore identical to that encountered in the usual Doppler cooling mechanism~\cite{Gordon1980,Cook1980,CohenTannoudji1992,BergSorensen1992}, where the diffusion coefficient $\textsl{\textsf{D}}$ is given to lowest order in $s$ by
\begin{equation}
\label{eq:MMC:Diffn}
\textsl{\textsf{D}}=\hbar^2k_\text{L}^2\Gamma s\Big[\cos^2(k_\text{L}x_0)+\tfrac{2}{5}\sin^2(k_\text{L}x_0)\Big]\,.
\end{equation}
The steady-state temperature $T_\text{M}$ of mirror-mediated cooling is then obtained from \eref{eq:FDT} using $\varrho=-\textsl{\textsf{F}}_{\parallel}(x_0)/v_\text{m}$, with $\textsl{\textsf{F}}_{\parallel}(x_0)$ given by \eref{eq:MMC:AnalyticLongFrictionHarm}. For $|\Delta_\text{L}|\gg\Gamma$ we thus find
\begin{equation}
\label{eq:MMC:TempM}
T_\text{M}=\frac{1}{5\pi}\frac{\hbar}{k_B}\frac{\omega_\text{t}\Gamma}{g^2}\frac{2+3\cos^2(k_\text{L}x_0)}{\sin(\omega_\text{t}\tau)\sin(4k_\text{L}x_0)}\,.
\end{equation}
An example of the dependence of $T_\text{M}$ on the trap position is shown in~\fref{fig:MMC:SST_vs_Posn}, predicting a minimum temperature of the order of $400$\,$\upmu$K. Whilst this may seem large in comparison to the Doppler temperature of $141$\,$\upmu$K, one has to keep in mind that $T_\text{M}$, given by \eref{eq:MMC:TempM}, is insensitive to detuning and, for far off-resonant operation of the order of tens of linewidths, it will be the dominant mechanism. This is further discussed in \Sref{sec:CoolingMethods:MMC:Beyond}. We also note that \fref{fig:MMC:SST_vs_Posn} further highlights the importance of the requirement for localising the particle.\\
Using \eref{eq:MMC:AnalyticLongFrictionFamiliar} we can approximate the steady-state temperature at the point of maximum friction by
\begin{equation}
\label{eq:SemiclassicalMMCTemperature}
T_{\text{M}}\approx\frac{\hbar}{2k_\text{B}\tau}\frac{\sigma_\text{L}}{\sigma_{\text{a}}}\,.
\end{equation}
It is interesting to note that this expression is closely related to the expression for the limiting temperature in Doppler cooling, $k_BT=\hbar\Gamma$, but where the upper state lifetime $1/(2\Gamma)$ is replaced by the atom--mirror delay time, $\tau$, and where a geometrical factor equal to the mode area divided by the atomic radiative cross-section is included.\par
This last point will turn out to be a general trend in the cooling methods we discuss. The steady-state temperature is in general described by a function of the form
\begin{equation}
\label{eq:GeneralTForm}
T=\frac{\hbar}{2k_\text{B}\tilde{\tau}}\phi\,,
\end{equation}
where $\tilde{\tau}$ is the characteristic time---e.g., the memory time---of the system and $\phi$ depends on the geometry of the situation. In Doppler cooling, the characteristic time is the lifetime of the upper state, $\tilde{\tau}=1/(2\Gamma)$, whereas in mirror-mediated cooling it is naturally the delay time $\tau$. In cavity-mediated cooling, where the atom is inside a cavity, the energy loss mechanism is due to the decay of the cavity field, and we thereby have $\tilde{\tau}=1/(2\kappa)$~\cite{Horak1997}, where $\kappa$ is the linewidth of the cavity field.
\subsection{Numerical analysis of mirror-mediated cooling}\label{sec:CoolingMethods:MMC:Numeric}
In this section we now investigate a more accurate numerical model to corroborate the simplified analytical results obtained above.\footnote{The Fokker--Planck equation was obtained, through the use of a suitably extended Wigner transform, by Peter Horak. The basis for the Monte--Carlo code was also written by PH, and then extended by AX. The simulations and data analysis were performed by AX.} In order to render the problem numerically tractable, the continuum of modes is replaced by a discrete set of modes with frequencies $\omega_k$, with $k=1,\dots,N$. The master equation \eref{eq:QME} is then converted by use of the Wigner transform~\cite{Gardiner2004} into a Fokker--Planck equation for the atomic and field variables. Applying a semiclassical approximation and restricting the equation of motion to second-order derivatives, one arrives at an equivalent set of stochastic differential equations for a single atom with momentum $p$ and position $x$ in a discrete multimode field with mode amplitudes $\alpha_k$ \cite{Horak2001},
\begin{subequations}
\label{eq:MMC:SDE}
\begin{align}
\mathrm{d} x=\ &\frac{p}{m}\mathrm{d} t\,,\\
\mathrm{d} p=\ &\i\gamma_0\big[\mathcal{E}(x)\tfrac{\mathrm{d}}{\mathrm{d} x}\mathcal{E}^{\ast}(x)-\mathcal{E}^{\ast}(x)\tfrac{\mathrm{d}}{\mathrm{d} x}\mathcal{E}(x)\big]\mathrm{d} t-U_0\big[\mathcal{E}(x)\tfrac{\mathrm{d}}{\mathrm{d} x}\mathcal{E}^{\ast}(x)+\mathcal{E}^{\ast}(x)\tfrac{\mathrm{d}}{\mathrm{d} x}\mathcal{E}(x)\big]\mathrm{d} t\nonumber\\
&-k_{\text{t}}(x-x_{\text{t}})\mathrm{d} t+\mathrm{d} P\,,\text{ and}\\
\mathrm{d}\alpha_k=\ &\i\Delta_k\alpha_k\mathrm{d} t-(\i U_0+\gamma_0)\mathcal{E}(x)f_k^{\ast}(x)\mathrm{d} t+\mathrm{d} A_k\,,
\end{align}
\end{subequations}
where $f_k(x)=\sin(\omega_k x/c)$ are the individual mode functions, $\mathcal{E}(x)=\sum_k \alpha_k f_k(x)$ is the total electric field, $\Delta_k=\omega_k-\omega_\text{L}$ is the detuning of each mode from the pump, $U_0$ is the light shift per photon, and $\gamma_0$ is the photon scattering rate. The terms $\mathrm{d} P$ and $\mathrm{d} A_k$ are correlated noise terms \cite{Horak2001} responsible for momentum and
field diffusion.
\par
In the following, we set the trap centre to $x_\text{t}=-3\lambda_\text{L}/16$, modulo $\lambda_\text{L}$, which is the point where the analytical solution predicts the maximum of the damping force. We use $N=256$ field modes with a mode spacing of $\Gamma/10$. At the start of every simulation, all field modes are empty with the exception of the pump mode which is initialised at $625$ photons, corresponding to a laser power of around $50$\,pW for our chosen parameters.
\par
The simulations were performed in runs of several thousand trajectories. Each such run was performed at a well-defined initial temperature, with the starting momenta of the particles chosen from a Gaussian distribution, and the starting position being the centre of the trap.
\subsubsection{Friction force and capture range\label{sec:CoolingMethods:MMC:Numeric:Friction}}
\begin{figure}[t]
\centering
\subfigure[$\omega_\text{t}=0.3\times 2\pi\Gamma$]{
\includegraphics[width=1.5\figwidth]{Diagrams/MMC/Comparison}
}\\
\subfigure[$\omega_\text{t}=0.5\times 2\pi\Gamma$]{
\includegraphics[width=1.5\figwidth]{Diagrams/MMC/Comparison_Stiff}
}
\caption[Comparison of mirror-mediated cooling simulations with analytical approximation]{Comparison of $\mathrm{d} p^2/\mathrm{d} t$ for the simulations without noise (data points) with the analytical approximation, \eref{eq:MMC:AnalyticLongFrictionHarmReduced}, including the harmonic trap (solid line). (a)~Weak harmonic
trap, $\omega_\text{t}=0.3\times 2\pi\Gamma$, showing also the linear dependence in the limit of small momenta, \eref{eq:MMC:AnalyticLongFrictionHarm} (dotted line). (b)~Stiff trap, $\omega_\text{t}=0.5\times 2\pi\Gamma$. The trap position $x_0^\prime=-3\lambda/16$ and other parameters are as in \fref{fig:MMC:Analytic-Spatial}.}
\label{fig:MMC:Comparison}
\end{figure}
\fref{fig:MMC:Comparison} presents the results of a set of simulations performed when setting the noise terms $\mathrm{d} P$ and $\mathrm{d} A_k$ in equations \eref{eq:MMC:SDE} to zero, i.e., neglecting momentum and photon number diffusion. The simulation data are compared with the result of the perturbative calculations \eref{eq:MMC:AnalyticLongFrictionHarmReduced}. For modest values of $\omega_{\text{t}}$, \fref{fig:MMC:Comparison}(a) justifies the averaging process used to derive \eref{eq:MMC:AnalyticFrictionTimeAvg}, which was based on spatial averaging but neglecting higher order terms in $v$. In contrast, for larger trap frequencies, the numerical simulations diverge significantly from the analytical result, as can be seen in \fref{fig:MMC:Comparison}(b). We expect that the terms in higher powers of the initial speed, which were dropped in the perturbative solution, are responsible for this discrepancy.
\begin{figure}[t]
\centering
\includegraphics[width=1.5\figwidth]{Diagrams/MMC/Cutoff}
\caption[Momentum capture range mirror-mediated cooling]{Capture range extracted from the simulations (data points) as compared to the analytical solution (dotted line) for various values of $\omega_\text{t}$. The solid line is a quadratic fit to the data for $\omega_\text{t}\geq 0.3\times 2\pi\Gamma$ and is only intended as a guide to the eye. Other parameters are as in \fref{fig:MMC:Comparison}.}
\label{fig:MMC:Cutoff}
\end{figure}
\par
We have already seen, in \eref{eq:MMC:CaptureRangeBesselZero}, that the capture range is expected to scale as $\omega_{\text{t}}^2$. For weak traps, as shown in~\fref{fig:MMC:Cutoff}, the numerical simulations agree well with these expectations. For stiffer traps, however, the capture range is consistently larger than that predicted; in fact, the simulations predict a capture range of around $450$\,mK for a trap frequency of $0.5\times 2\pi\Gamma$
\subsubsection{Steady-state temperature\label{sec:CoolingMethods:MMC:Numeric:SST}}
\begin{figure}[t]
\centering
\includegraphics[width=1.5\figwidth]{Diagrams/MMC/Heating_vs_Temp}
\caption[Cooling rate for a number of initial temperatures]{Cooling rate ($-\mathrm{d} T/\mathrm{d} t$) extracted from the simulations starting at a number of initial temperatures. The solid line represents a quadratic fit to the data. $\omega_\text{t}=0.5 \times 2\pi\Gamma$; other parameters are as in \fref{fig:MMC:Comparison}.}
\label{fig:MMC:Heating-vs-Temp}
\end{figure}
The next step in our investigation was to run simulations involving the full dynamics given by \erefs{eq:MMC:SDE} including the
diffusion terms. Because of the discrete nature of the field modes with uniform frequency spacing used in the simulations, the numerically modelled behaviour is always periodic in time with a periodicity given by the inverse of the frequency spacing. The simulations therefore cannot follow each trajectory to its steady-state unless an unfeasibly large number of modes is used. Instead, simulations were performed in several groups of trajectories, each group forming a thermal ensemble at a well-defined initial temperature. For each such group of trajectories the initial value of $\mathrm{d} T/\mathrm{d} t$ was calculated. The results for $\omega_\text{t}=0.5\times 2\pi\Gamma$ are shown in \fref{fig:MMC:Heating-vs-Temp}, where the error bars are due to statistical fluctuations for a finite number of stochastic integrations. The steady-state temperature is that temperature at which $\mathrm{d} T/\mathrm{d} t=0$ as clearly illustrated in this figure. For the chosen parameters, our data suggest a steady-state temperature of $722\pm 54$\,$\upmu$K with a $1/e$ cooling time of around $3.0$\,ms. This compares reasonably well with the steady-state temperature of $597$\,$\upmu$K predicted by \eref{eq:MMC:TempM}.
\begin{figure}[t]
\centering
\includegraphics[width=1.5\figwidth]{Diagrams/MMC/SST_vs_wtrap}
\caption[Mirror-mediated cooling steady-state temperature: comparison of simulations with analytical approximation]{Steady-state temperature for a number of simulations (circles) compared to the analytical formula \eref{eq:MMC:TempM} (solid line). The solid square represents the equivalent data from~\fref{fig:MMC:Heating-vs-Temp}, resulting from a much larger number of simulations. Parameters are as in \fref{fig:MMC:Comparison}.}
\label{fig:MMC:SST}
\end{figure}
\par
We finally performed a large number of simulations to investigate the dependence of the steady-state temperature on the trap frequency. \eref{eq:MMC:TempM} indicates that as one decreases $\omega_\text{t}$ the steady state temperature decreases. This is clearly seen in \fref{fig:MMC:SST}, which compares the prediction of \eref{eq:MMC:TempM} with a set of numerical simulations. The trend in the data is reproduced well by the analytical expression. However, the simulated steady-state temperature is consistently a little higher than predicted. We expect that this discrepancy is due mainly to two reasons:~(i)~\eref{eq:MMC:TempM} was derived from the friction \eref{eq:MMC:AnalyticLongFrictionHarm}, i.e., without the spatial averaging of \eref{eq:MMC:AnalyticLongFrictionHarmReduced} which would reduce the friction force; and (ii)~higher order terms in the velocity $v$ are also expected to reduce friction compared to the lowest order analytical result. For both these reasons, therefore, the analytical expression is expected to overestimate the friction force and thus to predict equilibrium temperatures that are too low.
\subsection{Beyond adiabatic theory}\label{sec:CoolingMethods:MMC:Beyond}
\begin{figure}[t]
\centering
\includegraphics[width=1.5\figwidth]{Diagrams/MMC/Doppler_vs_Mirror}
\caption[Limiting temperatures for Doppler cooling and mirror-mediated cooling]{Comparison between the calculated steady-state temperatures for mirror-mediated cooling $T_\text{M}$ (dash-dotted line), Doppler cooling $T_\text{D}$ (dashed), and in the presence of both effects $T$ (solid), drawn as a function of detuning whilst keeping the saturation parameter constant. $\omega_\text{t}=0.1\times 2\pi\Gamma$; other parameters are as in \fref{fig:MMC:Comparison}.}
\label{fig:MMC:Doppler_vs_Mirror}
\end{figure}
All the theoretical analysis and simulations discussed so far have been based on adiabatic elimination of the internal atomic degrees of freedom, and therefore neglected Doppler cooling. In~\fref{fig:MMC:Doppler_vs_Mirror}, we explore the variation of
$T_\text{M}$ and the Doppler temperature, $T_\text{D}$, as a function of detuning from resonance when the particle is at the point of greatest friction ($x_0^\prime=-3\lambda/16$), where $T_\text{D}$ is given by
\begin{equation}
T_\text{D} = \hbar\Gamma\,\frac{\Delta_\text{L}^2+\Gamma^2}{2\bigl(-\Delta_\text{L}\bigr)}\,,
\end{equation}
for $\Delta_\text{L}<0$. In the presence of both cooling effects, and assuming that the momentum diffusion terms are identical for both mechanisms, the stationary temperature achieved by the system is given by
\begin{equation}
T = \biggl(\frac{1}{T_\text{M}}+\frac{1}{T_\text{D}}\biggr)^{-1}\,.
\label{eq:MMC:fulltemp}
\end{equation}
Thus, for the parameters of \fref{fig:MMC:SST}, the calculated steady-state temperature $T$ reduces to $250$\,$\upmu$K in the limit of vanishing $\omega_\text{t}$.
\par
From \fref{fig:MMC:Doppler_vs_Mirror} one can see that the mirror-mediated force, for our tightly focussed pump, is stronger than the Doppler force for detunings larger than around $10\Gamma$ in magnitude. In practice this has two implications: for large negative detunings, we expect the steady-state temperature of the system to be significantly lower than that predicted by Doppler cooling; whereas for large \emph{positive} detunings, we still predict equilibrium temperatures of the order of mK.
\par
Both our perturbative expressions and our simulations are calculated to lowest orders in the atomic saturation. However, it is well known that in the limit of very large detunings also higher order terms in the saturation parameter $s$ become significant. Using the full expression for the diffusion constant~\cite{Gordon1980}, we can estimate the detuning for which we expect minimum diffusion and temperature. For the value of the saturation parameter $s\lesssim 0.1$ used throughout this section, it can be shown that $T_\text{M}$ attains a minimum at detunings of up to several tens of linewidths. Our chosen parameters are therefore within the range of validity of the model.
\subsection{Concluding remarks}\label{sec:CoolingMethods:MMC:Conclusions}
In conclusion, we have presented a mechanism for cooling particles by optical means which is based fundamentally on the dipole interaction of a particle with a light beam and therefore does not rely on spontaneous emission. The particle is assumed to be trapped and is simultaneously driven by an off-resonant laser beam. After the interaction with the particle the beam is reflected back onto the particle by a distant mirror. The time-delay incurred during the light round-trip to the mirror and back is exploited to create a non-conservative cooling force.
\par
The system was analysed using stochastic simulations of the semiclassical equations of motion representing a single two-level atom coupled to a continuum of electromagnetic modes. The results of these computations were found to agree with the expectations of a perturbative analysis. Our models predict sub-mK steady-state temperatures for \textsuperscript{85}Rb atoms interacting with a tightly focussed laser beam several metres from the mirror, in an arrangement similar to that of Ref.~\cite{Eschner2001}. While most of the theory is presented for a one-dimensional model, results for the friction force in the transverse direction suggest that three-dimensional cooling is possible with this scheme.
\par
The model presented here requires a large separation between the atom and the mirror, of the order of several metres, for an observable cooling effect. This limitation can be overcome in several ways. First, the light could propagate in an optical fibre between the atom and the mirror to avoid the effects of diffraction. Second, the required delayed reflection could be achieved through the use of a cavity instead of a mirror; in contrast to cavity-mediated cooling schemes~\cite{Horak1997,Vuletic2000,Maunz2004,Vilensky2007,Lev2008}, the atom would remain external to the cavity. For a time delay $\tau$ of order $1$\,ns one would require a cavity quality factor $Q=\omega\tau$~\cite{Rempe1992} of the order of $10^6-10^7$, which is achievable with present-day technology~\cite{Mabuchi1994}. This mechanism is explored heuristically in \sref{sec:CoolingMethods:Other:ECCO}, and subsequently investigated in greater depth in \sref{sec:TMM:ECCO}, after we have developed the necessary mathematical tools.
\subappendicesstart
\subsection{Appendix: A note on units}
The units used in the preceding work can perhaps best be called `quantum mechanical'. Here, we provide a number of useful conversions and numerical values, which could be of benefit to readers with a more experimental leaning.
\par
\begin{center}
\begin{tabular}{c | c}
\hline
\hline
Quantity in this work & Experimental value\\
\hline
$A$ & $2\sqrt{2\pi P/\bigl(\hbar k_\text{L}\bigr)}$\\
$g$ & $\sqrt{\Gamma\sigma_\text{a}/\bigl(\pi\sigma_\text{L}\bigr)}$\\
$s$ & $g^2|A|^2/\Delta_\text{L}^2=8\Gamma P\sigma_\text{a}/\bigl(\hbar k_\text{L}\sigma_\text{L}\Delta_\text{L}^2\bigr)$\\
\hline
\end{tabular}
\end{center}
\par
In the above, $P$ is the incident travelling-wave electromagnetic power, related to the incident electric field $\mathcal{E}$ by
\begin{equation}
P=\tfrac{1}{2}\epsilon_0c\sigma_\text{L}\lvert\mathcal{E}\rvert^2\,.
\end{equation}
$\Gamma$ is the HWHM linewidth of the upper state. For the D$_2$ line of some common alkali species we have:
\begin{center}
\begin{tabular}{c | c | c | c}
\hline
\hline
Quantity & $^{23}$Na~\cite{Steck2010b} & $^{85}$Rb~\cite{Steck2008}, $^{87}$Rb~\cite{Steck2010a} & $^{133}$Cs~\cite{Steck2010c}\\
\hline
$\Gamma$ & $2\pi\times 4.897$\,MHz & $2\pi\times 3.033$\,MHz & $2\pi\times 2.617$\,MHz\\
$\omega_\text{a}$ & $2\pi\times 508.848$\,THz & $2\pi\times 384.230$\,THz & $2\pi\times 351.726$\,THz\\
$\lambda_\text{a}=2\pi c/\omega_\text{a}$ & $589.158$\,nm & $780.241$\,nm & $852.347$\,nm\\
$k_\text{a}=2\pi/\lambda_\text{a}$ & $10.665\times 10^6$\,m$^{-1}$ & $8.055\times 10^6$\,m$^{-1}$ & $7.372\times 10^6$\,m$^{-1}$\\
$\sigma_\text{a}=3\lambda_\text{a}^2/\bigl(2\pi\bigr)$ & $1.657\times 10^{-13}$\,m$^2$ & $2.905\times 10^{-13}$\,m$^2$ & $3.469\times 10^{-13}$\,m$^2$\\
$T_\text{D}$ & $235.03$\,$\upmu$K & $145.57$\,$\upmu$K & $125.61$\,$\upmu$K\\
\hline
\end{tabular}
\end{center}
\subappendicesend
\section{Exploiting an optical memory in other geometries}\label{sec:CoolingMethods:Other}
\Sref{sec:CoolingMethods:MMC} provided us with a sound theoretical basis, in the case of the simplest possible geometry that permits a memory, for the arguments we put forward in \Sref{sec:CoolingMethods:Retarded}. In the present section, we will explore other geometries with which we can investigate the retarded dipole--dipole interaction. The ideas developed here will be fleshed out in later chapters, after we develop the necessary formalisms, but it is fairly instructive at this early stage to intuitively explore how the different mechanisms arise from very similar physical arguments.
\par
Let us first start by addressing the two major issues with the mirror-mediated cooling mechanism as described in the previous section. To recapitulate, the cooling force in this mechanism arises from the retarded dipole--dipole interaction of a particle with its own reflection in a mirror but (i)~requires a distance of `several metres' between the particle and the mirror for a sizeable effect, and (ii)~the friction force oscillates between cooling and heating on a sub-wavelength scale and has a zero spatial average.\\
Following the discussion of these two points, we will turn our attention to three-dimensional geometries to see how we can exploit the focussing properties of optics to achieve cooling.
\subsection{Lengthening the time delay: External cavity cooling}\label{sec:CoolingMethods:Other:ECCO}
From the general arguments in \Sref{sec:CoolingMethods:Retarded}, as well as the expressions in \Sref{sec:CoolingMethods:MMC}, we can see that the time delay is what governs the overall strength of the friction force in the retarded dipole--dipole interaction. This can be exploited in what we term `external cavity cooling' (\sref{sec:TMM:ECCO}), where the particle to be cooled interacts with a cavity field despite being outside the cavity. The nature of the interaction here is practically identical to mirror-mediated cooling; in particular, it suffers from the same drawback of having the friction force oscillate between a cooling and a heating force several times over the space of a single wavelength. Nevertheless, this is not a problem if what is to be cooled is not the motional energy of an atom but that of a micromirror since~(i)~a micromirror can be positioned, using piezoelectric actuators, with sub-nanometre-scale resolution~\cite{AttocubeSystemsPositioning2010}, and (ii)~the oscillation amplitude for the Brownian motion of a micromirror is also in the picometre to nanometre range. To illustrate the latter point, let us use the effective spring constant of a commercially-available micromirror as calculated in Ref.~\cite{Thompson2008}, $k_\text{M}=28$\,N\,m\textsuperscript{-1}. Then, the RMS displacement for the micromirror at a temperature $T=300$\,K, calculated as
\begin{equation}
\sqrt{\expt{x^2}}=\sqrt{\frac{k_\text{B}T}{k_\text{M}}}\,,
\end{equation}
is of the order of $10^{-11}$\,m.
\begin{figure}[t]
\centering
\includegraphics[scale=0.5]{Diagrams/SpaceTime_ECCO}
\caption[Space--time diagram explaining the distance folding mechanism]{Space--time diagram showing the distance folding argument used to explain the action of external cavity cooling. This figure should be compared to \fref{fig:SpaceTime}. The blank arrow indicates the pump beam.}
\label{fig:SpaceTimeECCO}
\end{figure}
\par
Having established the rationale, then, we can introduce `external cavity cooling' by means of the distance folding argument (see \fref{fig:SpaceTimeECCO}). For suppose we place a particle (atom or micromirror) in front of a cavity. The light that interacts with the particle is then allowed to couple into the cavity. This light then undergoes several round trips inside the cavity, and with each round-trip some light leaks out and re-interacts with the particle. An effective delay time can be defined that is related to the finesse (or, equivalently, the linewidth) of the cavity, and it is this concept that we show in \fref{fig:SpaceTimeECCO}. One notes that this delay time is, for a good cavity, orders of magnitude larger than that due to either mirror separately. Using a good cavity, with a finesse of the order of $10^5$, thereby allows us to squeeze the `several metres' into a sub-millimetre-scale device.\\
This opens the door to several important advances, of which we mention a few here. First of all, one is not constrained to use Fabry--P\'erot-type cavities, whereby the system can be constructed monolithically on a chip-scale device. Secondly, the cavity does not need to have good optical or mechanical access, which allows one to make significantly better and more stable cavities. Finally, having the micromirror or atom outside the cavity means that there is less chance of burning or saturation effects, since the local field surrounding the particle is not amplified by the cavity.
\subsection{Lifting the sub-wavelength dependence: Ring cavity cooling}
\begin{figure}[t]
\centering
\subfigure[Mirror-Mediated Cooling]{
\includegraphics[scale=0.5]{Diagrams/Invariance_MMC}
}\\[1cm]
\subfigure[Ring Cavity Cooling, position 1]{
\includegraphics[scale=0.5]{Diagrams/Invariance_RELIC_1}
}\hspace{1cm}
\subfigure[Ring Cavity Cooling, position 2]{
\includegraphics[scale=0.5]{Diagrams/Invariance_RELIC_2}
}
\caption[Delay times for atom different positions in mirror-mediated cooling and ring cavity cooling]{In (a)~mirror-mediated cooling, the time delay depends on the atomic position ($\tau_1\neq\tau_2$), whereas in (b,~c)~ring cavity cooling, it is independent of the atomic position ($\tau_1=\tau_2$). The blank arrow indicates the pump beam.}
\label{fig:Invariance}
\end{figure}
A cursory physical analysis of the mirror-mediated cooling mechanism will reveal that the origin of the $\sin(4k_\text{L}x_0)$ dependence lies in the fact that the atom is in a standing wave inside a system that is \emph{not} translationally invariant. To see this latter point, one simply needs to observe that, with the atom at two different places the time delay is of course different. One way of restoring translational invariance into the system is by using a ring cavity, rather than a plane mirror, to introduce a delay; see \fref{fig:Invariance}. The expectation of the physical mechanism being preserved, albeit without the position dependence in the friction force, is borne out when the system is examined in detail, as we shall see in \Sref{sec:TMM:AmplifiedOptomechanics} after we have developed the necessary mathematical model.
\subsection{Exploiting three-dimensional electromagnetism}
\begin{figure}[t]
\centering
\subfigure[Optical binding]{
\includegraphics[scale=0.5]{Diagrams/Tweezing_Two}
}\\[1cm]
\subfigure[Self--tweezing, static]{
\includegraphics[scale=0.5]{Diagrams/Tweezing_Static}
}\hspace{1cm}
\subfigure[Self--tweezing, moving]{
\includegraphics[scale=0.5]{Diagrams/Tweezing_Friction}
}
\caption[Three-dimensional retarded self-binding]{In (a)~optical binding, particle 2 is in a potential well caused by particle 1. Self--tweezing, (b), occurs when a particle is in a potential well caused by itself; if the particle is moving, (c), it will experience a force opposing its motion. The blank arrow indicates the pump beam and the thick black arrows the force.}
\label{fig:Tweezing}
\end{figure}
Perhaps the two biggest conceptual differences between electromagnetism in one dimension and that in three dimensions are orthogonal polarisations (which can nevertheless be mathematically mimicked in a quasi-one-dimensional geometry) and the spreading out of waves. Indeed, a travelling wave in one dimension can be represented as
\begin{equation}
\cos[\omega(t-x/c)]\,,
\end{equation}
whereby the amplitude of the wave is $1$, for any value of $x$. In three dimensions, however~\cite{Jackson1998} [see, for example, Eq.~(9.19) in this reference], the amplitude of the electric field is dependent on the distance $r$ from the source, and scales as $1/r$ for $r\gg\lambda$. Thus, suppose a particle has a dipole induced by a local electric field $E$. This dipole will subsequently emit a spherical wave, and let us suppose that we reflect this wave back onto the dipole. The potential energy mediated by the interaction between the instantaneous dipole polarisation and the reflected field (which has travelled through a distance $r$) is, up to some phase and constants,
\begin{equation}
U\sim-\frac{E^2}{r}\,.
\end{equation}
We can now apply the result in \eref{eq:GeneralTimeDelayedF}:~the dipole will experience a friction force proportional to $\tau c/r^3=1/r^2$.
\par
The spreading out of waves in three dimensions is but one instance of the more general concept of (de)focussing of an electric field. Let us examine a simple way in which we can exploit this idea, \emph{self-tweezing}. Optical binding~\cite{Karasek2006}, \fref{fig:Tweezing}(a), is an interesting phenomenon whereby light incident on a dielectric sphere is focussed by the sphere itself. This focus, in turn, produces a potential well for a second particle, which is thereby optically bound to the first. Suppose, then that we only have one particle, and that we put a plane mirror `downstream' of that particle, \fref{fig:Tweezing}(b). The mirror will, if the geometry is chosen properly, cause the focus to appear close enough to the particle that it will feel the potential well caused by this electric field. This situation, then is similar to the case with two particles, but the single particle is now optically bound to \emph{itself}. Any motion of the particle around this position will cause its image to lag behind it and a restoring, viscous, force to act on it, \fref{fig:Tweezing}(c).
|
\section{Introduction} \label{intro}
IC 443 (G189.1+3.0) is an extensively studied supernova remnant (SNR) and famous for its interaction with nearby molecular clouds.
The interaction has been widely observed over diverse wavelengths: the $\gamma$-ray from hardronic collisions \citep{Esposito(1996)ApJ_461_820,Albert(2007)ApJ_664_L87,Acciari(2009)ApJ_698_L133,Abdo(2010)ApJ_712_459,Tavani(2010)ApJ_710_L151}, the X-ray absorption by foreground clouds \citep{Troja(2006)ApJ_649_258}, the infrared forbidden lines \citep{Burton(1988)MNRAS_231_617,Burton(1990)ApJ_355_197,Inoue(1993)PASJ_45_539,Richter(1995)ApJ_454_277,Cesarsky(1999)A&A_348_945,Rho(2001)ApJ_547_885,Rosado(2007)AJ_133_89}, the enhanced CO line ratio \citep{Seta(1998)ApJ_505_286,Xu(2011)ApJ_727_81}, the broad molecular emission lines \citep{White(1987)A&A_173_337,Wang(1992)ApJ_386_158,Dickman(1992)ApJ_400_203,vanDishoeck(1993)A&A_279_541,Snell(2005)ApJ_620_758,Zhang(2009)arxiv_11_4815}, the bow-like feature in position-velocity diagrams of molecular line \citep{Tauber(1994)ApJ_421_570}, and the OH maser line \citep{Claussen(1997)ApJ_489_143,Hoffman(2003)ApJ_583_272,Hewitt(2006)ApJ_652_1288}.
Therefore, IC 443 is usually observed when studying the shock-cloud interaction.
Its age ranges from $\sim3-4$ kyr \citep{Petre(1988)ApJ_335_215,Wang(1992)PASJ_44_303,Troja(2008)A&A_485_777} to $\sim20-30$ kyr \citep{Chevalier(1999)ApJ_511_798,Olbert(2001)ApJ_554_L205,Gaensler(2006)ApJ_648_1037,Bykov(2008)ApJ_676_1050,Lee(2008)AJ_135_796}, and its distance is thought to be about 1.5 kpc based on several arguments, such as the contact with Gem OB 1 association \citep{Poveda(1968)AJ_73_65}, the empirical $\Sigma-D$ relation \citep{Milne(1979)AuJPh_32_83,Caswell(1979)MNRAS_187_201}, the total remnant energy \citep{Fesen(1980)ApJ_242_1023,Fesen(1984)ApJ_281_658}, and the high-velocity absorption lines observed against background stars \citep{Welsh(2003)A&A_408_545}.
It extends $\sim45'$ (cf.~\citealt{Gaensler(2006)ApJ_648_1037}) and overlaps in the sky with another more extended SNR G189.6+3.3 \citep{Asaoka(1994)A&A_284_573}.
IC 443 consists of two half shells, and another large shell of the SNR G189.6+3.3 overlaps with the former two shells; these three shells were named as A, B, and C, respectively, by \cite{Braun(1986)A&A_164_193}.
The overall picture of the remnant region is well outlined in \cite{Troja(2006)ApJ_649_258} and \cite{Lee(2008)AJ_135_796}.
In the middle of the two shells A and B, there is a W-shaped ridge which shows strong \Htwo{} emission lines \citep[cf.~Fig.~\ref{fig-slit} and][]{Rho(2001)ApJ_547_885}; this ridge is thought to be a torus-type molecular clouds overrun by the SNR shock.
Infrared \Htwo{} emission lines are useful to study shocked molecular clouds, since \Htwo{} is the most abundant molecule and its quantum levels cover a wide energy range enough to study shocked gas, whose temperature ranges from a few hundred to a few thousand kelvin.
Toward the W-shaped \Htwo{} ridge, infrared spectral observations have already been performed from ground \citep{Richter(1995)ApJ_454_277} and space by \isofull{} \citep[ISO,][]{Cesarsky(1999)A&A_348_945} and \spitzer{} \citep{Neufeld(2007)ApJ_664_890}.
However, the \Htwo{} emission lines which falls within $2.5-5.0$ \um{} have not been observed completely; the ground observations missed several lines because of the atmospheric absorption, and this wavelength range is not covered by \spitzer{} spectroscopy and was simply not observed by \isofull{} (\iso).
Besides, this wavelength range is worthwhile to observe because, for instance, we could obtain the population of high-$J$ $\upsilon=0$ levels, which is usually assumed to follow $\upsilon>0$ levels \citep[e.g.~][]{Rho(2001)ApJ_547_885,Giannini(2006)A&A_459_821}, but have not been thoroughly checked yet.
Here we present the results of near-infrared spectral observations over $2.5-5.0$ \um{} for the shocked \Htwo{} gas in the SNR IC 443.
The observations were performed with the satellite \akari{} (section \ref{obs-red}).
We detected many \Htwo{} emission lines toward shocked molecular gas (section \ref{ana-res-line}), and found that the population of the shocked \Htwo{} gas cannot be described by any combination of \Htwo{} gas in Local Thermodynamic Equilibrium (LTE), which have been usually used for (section \ref{ana-res-level} and \ref{ana-res-cmp}).
Instead, we employed a non-LTE \Htwo{} gas model (section \ref{ana-res-plmod}), and interpreted the results in terms of a shock combination (C-type and J-type) together with the diversity in shock velocities (section \ref{dis-plmod}).
The observed background emission, attributed to diffuse warm \Htwo{} gas, was discussed as well (section \ref{dis-bg}).
\section{Observations and Data Reduction} \label{obs-red}
The spectral observations were performed with the InfraRed Camera \citep[IRC,][]{Onaka(2007)PASJ_59_S401} onboard the Japanese satellite \akari{} \citep{Murakami(2007)PASJ_59_S369}, on 2008-Sep-26th and 27th during the post-Helium phase.
During this phase, only near-infrared observations were possible, since the cryogenic cooling with liquid Helium had been run out.
We used the $5''\times48''$ slit and the grism, whose resolving power and wavelength coverage are $\Delta\lambda\sim0.03$ \um{} and $2.5-5.0$ \um, respectively.
The observation mode, called as Astronomical Observation Template (AOT), is IRCZ4 which is designed for general spectroscopic observations.
It has an imaging observation sandwiched by spectroscopic observations of four frames (cf.~\citealt{Onaka(2009)mana}).
Comparing this reference image to the 2MASS catalog \citep{Skrutskie(2006)AJ_131_1163}, we corrected the default astrometry of the slit, given from the satellite attitude information.
We observed four regions, three of which are the shocked CO clumps and the rest is the background.
Figure \ref{fig-slit} shows the four slit positions over the 2MASS $K_s$ RGB image \citep{Skrutskie(2006)AJ_131_1163}, which displays a diffuse `W' feature that traces the \HtwolineK{} 2.12 \um{} line emissions \citep{Rho(2001)ApJ_547_885}.
We name three on-source positions as `B', `C', and `G' after the names of the shocked CO clumps \citep{Denoyer(1979)ApJ_232_L165,Huang(1986)ApJ_302_L63}, and the background as `BG.'
Table \ref{tbl-obs} summarizes our observations---the region name, RA-Dec position, observation ID, and AOT.
The data were reduced through the official pipeline, supported by the \akari{} team \citep[cf.~][]{Onaka(2009)mana}.
We used the new spectral response curve for the post-Helium phase data\footnote{http://www.ir.isas.jaxa.jp/AKARI/Observation/DataReduction/IRC/SpecResponse\_091113/}, which shows a degraded sensitivity, $\sim$70\% of the Helium phase sensitivity.
Columns of the two dimensional spectral images are occasionally saturated by the very bright sources in the imaging area of the detector, which cause the column pull-down effect.
This effect was corrected by masking out the relevant columns.
Hot pixels of the detector were also masked out.
During the data reduction, no smoothing and tilt-correction were applied to the two dimensional spectral images.
In order to extract spectra, we chose certain sections along the slit length, then averaged the pixel values within those sections.
The extraction sections were carefully chosen for the \Htwo{} emission lines to be nearly uniform within the sections.
The sections extend 10 pixels for the clumps B and C, 14 pixels for the clump G, 25 pixels for BG, where one pixel corresponds to $1.46''$ \citep{Onaka(2007)PASJ_59_S401}; for the data set 1420806-002 of BG, we only chose 5 pixels as an extraction section to avoid abnormal single pixel peak.
Figure \ref{fig-spec} displays the extracted spectra at each region.
For clarity, the error bars were omitted and the spectrum of BG was enlarged by a factor of 20.
The statistical error bars can be seen in Figures \ref{fig-fit-B}-\ref{fig-fit-bg}.
For the error estimation, we included the systematic error caused by the calibration source type, in addition to the statistical error.
Our target is a diffuse source, hence the flux calibration referred to standard point sources are not suitable for our source, since the aperture loss and the slit loss would vary with the source type; the official pipeline uses the calibration from point sources.
We considered this type of systematic error and adopted 10\% of the signal intensity as the systematic error (private communication with the \akari{} helpdesk).
It was squarely summed to the statistical error of the line intensity, $\sqrt{\sigma_{st}^2+\sigma_{sys}^2}$, after measuring the line intensities (cf.~section \ref{ana-res}).
\section{Analysis and Results} \label{ana-res}
\subsection{Line Identification and Intensity Measurement} \label{ana-res-line}
Figure \ref{fig-spec} shows many emission lines.
Since our concerns are the \Htwo{} emission lines, we first compared the wavelength of the observed emission lines with those of \Htwo{}.
For easier identification of single and blended lines, we made template spectra of \Htwo{} gas in LTE at diverse temperature from 1000 K to 4000 K.
This temperature range was adopted because the shocked \Htwo{} gas usually shows such a range of excitation temperatures at the upper levels of \EvJ{} $\sim5\times10^3-2.5\times10^4$ K \citep[e.g.~][]{Rosenthal(2000)A&A_356_705,Rho(2001)ApJ_547_885,Giannini(2006)A&A_459_821}, which includes the upper levels of the \Htwo{} emission lines we detected.
In this way, we identified all the detected emission lines as \Htwo{} emission lines.
Some lines are blended with nearby lines, hence we could not tag them as a single line.
The uniquely identified lines and the blended lines are indicated by `$\mid$' and `+', respectively, in Figure \ref{fig-spec}, and their line identifications are listed in Table \ref{tbl-result}.
For a cross-check, this identification was compared with that of the shocked \Htwo{} gas observed in the Orion Molecular Cloud-1 \citep[OMC-1,][]{Rosenthal(2000)A&A_356_705}, and our identification turned out to be reliable.
Some contribution from Br$\beta$ 2.63 \um{} can be blended with the 2.63 \um{} blended line (cf.~Fig.~\ref{fig-spec}); however, we think it is unlikely since Br$\alpha$ 4.05 \um, which should be stronger than Br$\beta$, was not detected; the spontaneous transition probabilities of Br$\alpha$ and Br$\beta$ are $A_{54}=2.7\times10^6$ s$^{-1}$ and $A_{64}=7.7\times10^5$ s$^{-1}$, respectively.
The features seen at the edge of the band ($<2.6$ and $>4.9$ \um) were ignored, since they are likely to be inadequate to analyze.
The 2.56 \um{} and 4.95 \um{} features seem to be the $\upsilon=1-0$ Q(9) blended with nearby lines and the $\upsilon=1-1$ S(9), respectively.
The line intensities were measured by fitting their line profiles with a continuum plus Gaussian whose full-width-at-half-maximum (FWHM) is fixed to the spectral resolution of the IRC ($\Delta\lambda=0.03$ \um).
Adjacent lines whose profiles are overlapped with each other were fitted simultaneously.
As a baseline continuum for the fitting, we used a median-smoothed spectrum of each region; the kernel width ranging from 0.20 to 0.54 \um{} was carefully chosen to be wide enough to erase out the line emission features.
The feature near the edge of the spectrum (e.g.~2.56 \um{} feature in Figure \ref{fig-fit-B}) remains unchanged after the median filtering, since the filter only works on those pixels that are away from the edge by more than a half of the filter width.
The fitted profiles are displayed in Figures \ref{fig-fit-B}--\ref{fig-fit-bg}, and their measured intensities are listed in Table \ref{tbl-result}.
The intensities of blended lines are shown with the symbol `$<$', since we cannot determine the individual contribution of each line blended.
The intensities of weak lines, whose signal-to-noise ratio are lower than 3.0, are expressed as 90\% confidence upper limits.
\subsection{Reddening Correction and \Htwo{} Level Population} \label{ana-res-level}
Over $2.5-5.0$ \um{} wavelengths, the extinction optical depth becomes greater than one with the hydrogen nuclei column density, \defNH, of $\gtrsim10^{22}$ \Ncm{} \citep{Draine(2003)ARA&A_41_241}.
Since the observed regions are pervaded with dense molecular gas, the measured intensity should be corrected for the reddening by the intervening interstellar dust.
We exploited the extinction curve of ``Milky Way, $R_V=3.1$'' \citep{Weingartner(2001)ApJ_548_296,Draine(2003)ARA&A_41_241}, and adopted proper hydrogen nuclei column densities for each region.
\Av=13.5 was adopted for the clumps C as \cite{Neufeld(2008)ApJ_678_974} did based on the results of \cite{Richter(1995)ApJ_454_277}.
We adopted the same \Av{} for the clump B, since its extinction is known to be similar with that of the clump C \citep{Burton(1988)MNRAS_231_617}.
No extinction measurement exist toward BG; thus, we assumed it to be the \Av{} of the clump C, the nearest clump from BG.
\Av=10.8 was adopted for the clump G; this was inferred from $A_{2.12}=1.3$, obtained by \cite{Richter(1995)ApJ_454_277}, employing the ``Milky Way, \Rv=3.1'' extinction curve.
The corresponding \NH{} was calculated with the equation \Av=\NH/($1.87\times10^{21}$ \Ncm) for \Rv=3.1 \citep{Bohlin(1978)ApJ_224_132}.
The \Htwo{} level population was derived from these reddening corrected intensities, assuming that the \Htwo{} emission lines are optically thin.
Since the infrared \Htwo{} emission is emanating from electric quadrupole transition, it is optically thin under a typical interstellar medium condition; for instance, the pure-rotational S(0) and S(1) lines become optically thick at the line center when \NHtwo{} $>10^{24}$ \Ncm{}, adopting the line width of 10 \kms.
We derived the reddening corrected level population of \Htwo{} gas from the following equation,
\begin{equation}
N_{rc}(\upsilon,J)=\frac{4\pi\lambda}{hc}\frac{I_{rc}(\upsilon,J\rightarrow\upsilon',J')}{A(\upsilon,J\rightarrow\upsilon',J')},
\end{equation}
where $I_{rc}(\upsilon,J\rightarrow\upsilon',J')$ and $A(\upsilon,J\rightarrow\upsilon',J')$ are the reddening corrected line intensity and the Einstein-A radiative transition probability of the transition from level $(\upsilon,J)$ to $(\upsilon',J')$, respectively.
The molecular data for \Htwo{} were obtained from the database provided by a simulation code, CLOUDY (version C08.00; \citealt{Ferland(1998)PASP_110_761}).
The results are listed in Table \ref{tbl-h2col} and their population diagrams are displayed in Figure \ref{fig-pop}.
$g_J$ is a weight factor which corresponds to $(2J+1)$ and $3(2J+1)$ for para (even $J$) and ortho (odd $J$) states, respectively; the population of LTE gas is appeared as a straight line in this diagram.
In Figure \ref{fig-pop}, the clumps B, C, and G show the population of $\upsilon=0,1,2$, while BG shows that of ($\upsilon,J$)=(0,11) only.
In the clumps B, C, and G, the $\upsilon=0$ population shows a similar shape and little zigzag pattern; when the ortho-to-para ratio is approaching to 3.0, the zigzag pattern disappears \citep[cf.~][]{Neufeld(2006)ApJ_649_816,Neufeld(2007)ApJ_664_890}.
The $\upsilon=1$ population, however, shows a similar shape and little zigzag pattern only in the clumps C and G; the clump B shows an evident zigzag pattern over the population of ($\upsilon,J$)=(1,1), (1,2), and (1,3).
Here we note that the $\upsilon=0$ and $\upsilon=1$ levels follow different branches.
This becomes clearer when plotted with the lower-$J$ $\upsilon=0$ population obtained from mid-infrared observations (see the following section).
BG show a much smaller but evident population of the ($\upsilon,J$)=(0,11) level.
It is about a factor of $12-38$ smaller than the clumps B, C, and G (cf.~Table \ref{tbl-h2col}).
\subsection{Comparison of \Htwo{} Populations with Previous Observations: the Clumps C and G} \label{ana-res-cmp}
As far as we know, the UKIRT CGS4 observation of \cite{Richter(1995)ApJ_454_277} is the only published near-infrared spectroscopic observation that targeted on the shocked \Htwo{} gas in IC 443; they observed the clumps C and G.
Besides, toward these two clumps, there are published results of mid-infrared spectroscopic observations for the shocked \Htwo{} gas, performed with \isofull{} \citep{Cesarsky(1999)A&A_348_945} and \spitzer{} \citep{Neufeld(2007)ApJ_664_890}.
For the clump B, mid-infrared spectroscopic observations were also performed with \spitzer{} \citep{Noriega-Crespo(2009)inproc}, however, they are under analysis and only covered three emission lines, $\upsilon=0-0$ S(0), S(1), and S(2).
Therefore, we concentrated on the data of the clumps C and G, and compared our \akari{} results with the ones from previous studies.
Before the comparison of the results, we compared the spectra extraction areas of \akari's and other's, to check whether we are comparing the same radiation sources.
The yellow circle and boxes in Figure \ref{fig-specext} indicate the extraction areas of the clumps C and G.
The details for the extraction area are given in the figure caption.
For the clump C, the centers of the extraction areas falls within $\sim20''$.
Especially, that of \spitzer{} data \citep{Neufeld(2007)ApJ_664_890}---the center of a Gaussian taper---well falls into the extraction area of the \akari{} data; hence, it is likely that the \Htwo{} emissions, observed by \akari{} and \spitzer{}, represent almost the same gas, unless the average physical property drastically varies over a few arcsec scale ($<0.07$ pc).
For the clump G, the centers of the extraction areas falls within about $30''$.
In this case, the extraction areas of \akari{} and \iso{} \citep{Cesarsky(1999)A&A_348_945} data overlap about 40 \% with a $\sim10''$ ($\sim0.07$ pc) separation.
Thus, as in the clump C, the \Htwo{} emissions, observed by \akari{} and \iso{}, likely represent almost the same gas, unless the average physical property drastically varies over a $\sim0.07$ pc scale.
Prior to the comparison of level population from different observations, we here note that the same extinction-correction curve and \Av{} values, used in section \ref{ana-res-level}, were used for the consistent comparison.
We first compared our results with those of mid-infrared observations \citep{Cesarsky(1999)A&A_348_945,Neufeld(2007)ApJ_664_890}.
The \emph{top} panels of Figure \ref{fig-popall} shows the comparison.
We adopted 15\% systematic calibration error, which is a dominant component, for both clumps, C and G \citep{Cesarsky(1999)A&A_348_945,Neufeld(2007)ApJ_664_890}.
We here note that, in both clumps, $\upsilon=0$ levels are more populated than $\upsilon=1$ levels at a fixed level energy with a 3-$\sigma$ significance at least.
This means that the \Htwo{} level population cannot be reproduced by any combination of \Htwo{} gas in LTE, which is usually adopted for the description of shocked \Htwo{} gas (e.g.~\citealt{Rho(2001)ApJ_547_885,Giannini(2006)A&A_459_821}).
This invalidity was previously pointed out in the study on the shocked \Htwo{} gas of the supernova remnant HB 21 \citep{Shinn(2010)AdSpR_45_445}.
Secondly, we compared our results with those of ground near-infrared observations by \cite{Richter(1995)ApJ_454_277}.
The extinction-corrected level populations were derived from their ``averaged'' intensities of ``position 1'' and ``position 3,'' each for the clumps C and G, respectively.
The comparisons are plotted in the \emph{middle} panels of Figure \ref{fig-popall}.
As the figures show, there is an inconsistency between the same ($\upsilon,J$) levels of two observations.
We attribute this inconsistency to the flux calibration difference; the $\upsilon=0,1$ levels show almost constant, vertical gaps in the population diagram, about a factor of $3-4$ in the column density.
Our \akari{} calibration is likely to be correct, because the \akari{} $\upsilon=0$ level populations are seamlessly merged with those obtained from previous mid-infrared observations (cf. the \emph{top} panels of Fig.~\ref{fig-popall}).
However, we cannot rule out that the inconsistency is caused by the difference of observed regions, since the average properties of the shocked \Htwo{} gas may change over $\sim30''\sim0.2$ pc scale (see Fig.~\ref{fig-specext}).
Lastly, we note the importance of the space observations, fully covering $2.5-5.0$ \um, for the study of shocked \Htwo{} gas.
With ground observations, the gap between $\upsilon=0$ and 1 levels is hardly inspectable, since only a few $\upsilon=0$ levels can be observable at largely separated \EvJ{} (see the \emph{grey} plots in \emph{middle} panels of Figure \ref{fig-popall}).
Hence, when using the results of ground observations, we are likely to think that a combination of \Htwo{} gas in LTE looks viable to reproduce the level populations up to \EvJ{} $\sim25,000$ K (see the \emph{bottom} panels of Figure \ref{fig-popall}).
However, such a gap can be immediately inspectable from the space observations, fully covering $2.5-5.0$ \um, since $\upsilon=0,1$ levels are covered continuously; the \akari{} near-infrared observation is a good example.
\subsection{Power-law Thermal Admixture Model of \Htwo{} Gas} \label{ana-res-plmod}
As seen in the previous section, the combination of \Htwo{} gas in LTE cannot reproduce the observed level population of the shocked \Htwo{} gas in the clump C and G, obtained from \akari{} and previous mid-infrared observations.
Therefore, we applied the power-law thermal admixture model of \Htwo{} gas, which successfully reproduced the level population of shocked \Htwo{} gas before \citep{Neufeld(2008)ApJ_678_974,Shinn(2009)ApJ_693_1883,Neufeld(2009)ApJ_706_170,Shinn(2010)AdSpR_45_445,Lee(2010)ApJ_709_L74,Takami(2010)ApJ_720_155,Yuan(2011)ApJ_726_76}.
The model configuration was the same with that used in the study of the supernova remnant HB 21 \citep{Shinn(2009)ApJ_693_1883,Shinn(2010)AdSpR_45_445} and the young stellar object (YSO) L 1251A \citep{Lee(2010)ApJ_709_L74}, except the inclusion of the H atom as an additional collider.
Our \akari{} spectra include several ro-vibrational transition lines (e.g.~$\upsilon=1\rightarrow0, 2\rightarrow1$), which are sensitive to the collision with H atoms \citep{Neufeld(2008)ApJ_678_974,Shinn(2009)ApJ_693_1883,Shinn(2010)AdSpR_45_445}.
Thus, we updated the previous model to reflect the collisions with H atoms.
The \Htwo{} column density was calculated from the following equation,
\begin{eqnarray}
&dN=a T^{-b} dT, \\
&\textrm{where, }a=\frac{\textrm{\scriptsize N}(H_2; T>100 \textrm{ \scriptsize K})(b-1)}{T_{min}^{1-b}-T_{max}^{1-b}} \nonumber \\
&T_{min}=\textrm{100 K}, \, T_{max}=\textrm{4,000 K}& \nonumber
\end{eqnarray}
$a$ and $b$ are constants, and $\textrm{N}(H_2; T>100 \textrm{ K})$ is a total column density of molecular hydrogen warmer than 100 K.
At each temperature, the statistical equilibrium was assumed, and the collisional partners were \Htwo, He, and H.
The collisional deexciation rates were obtained from \cite{LeBourlot(1999)MNRAS_305_802} for \Htwo{} and He, and from \cite{Wrathmall(2007)MNRAS_382_133} for H.
For the \Htwo{} collider, newer rates were calculated by \cite{Lee(2008)ApJ_689_1105}; however, their results are similar with \citeauthor{LeBourlot(1999)MNRAS_305_802}'s over $100-6000$ K and only include for those levels of $J_{up}\leq8$.
Hence, we kept using \citeauthor{LeBourlot(1999)MNRAS_305_802}'s.
The collisional excitation rates were calculated from the detailed balance relation.
The \Htwo{} density, \nHtwo, and the relative abundance of H atom to \Htwo, $X_H\equiv\textrm{log}\left[\frac{n(\textrm{\tiny H I})}{n(\textrm{\tiny H}_2)}\right]$, were set as free parameters, while \nHe{} was assumed to be $0.2\times$\nHtwo.
The ortho-to-para ratio was set to 3.0, since the level populations show little zigzag pattern (cf.~the \emph{top} panels of Fig.~\ref{fig-popall}).
First, we tried to fit with a single \nHtwo{} value, but we failed.
No single \nHtwo{} model could successfully reproduce the population of high energy-levels.
This result reaffirms the previously-noted tendency that the single \nHtwo{} model does not reproduce the level population over \EvJ{} $=0-25,000$ K (cf.~section \ref{dis-plmod-cmp}).
Neither the addition of another $b$ was successful.
Instead, we added an additional \nHtwo{} component, and then we obtained successful results.
The fitting results are displayed in Figure \ref{fig-mfit}.
For the clump G, the fewer data points from mid-infrared observations caused weaker constraints for the fitting parameters than the clump C; therefore, we fixed the $X_H$ value as $-1.7$ in view of the fitting results for the clump C.
Then, we scanned the $\chi^2$ space for fitting parameters and the results are displayed in Figure \ref{fig-mconf}; the step size for the scan was 0.1 for all parameters displayed in Figure \ref{fig-mconf}.
The fit parameters with their 90\% confidence intervals, together with the reduced chi-square values, are listed in Table \ref{tbl-mfit}.
As seen from Figure \ref{fig-mfit}, the reddening-corrected level populations are well reproduced by the power-law thermal admixture model with two different \nHtwo{}s: one low \nHtwo{} $\sim10^3-10^4$ \ncm{} and the other high \nHtwo{} $\sim10^5-10^6$ \ncm{} (cf.~Table \ref{tbl-mfit}).
The \emph{lower and higher} density \Htwo{} gases mainly contribute to the \emph{lower and higher} upper-energy levels, respectively.
This indicates that some model parameters may not be uniquely determined, since only a small portion of the modeled population is constrained by the observed population.
This point is more discussed in section \ref{dis-plmod-cmp}.
The two kinds of \Htwo{} gases share a common power-law index $b$; it is 1.6 and 3.5 for the clumps C and G, respectively.
The column density ratios of the lower to higher density gas are not much different; they are 40 and 13 for the clumps C and G, respectively (see section \ref{dis-plmod-cmp} for more notes about the column densities, however).
We obtained $X_H=-1.7$, which corresponds to \nHone/\nHtwo=0.02; this value is similar with the one where the collision with H atoms starts to dominate the collision with \Htwo{} for the rovibrational transition lines, as mentioned by \cite{Neufeld(2008)ApJ_678_974} and \cite{Shinn(2009)ApJ_693_1883,Shinn(2010)AdSpR_45_445}, and smaller than those obtained for protostellar outflows of LDN 1157, \nHone/\nHtwo{} $=0.1-0.3$ \citep{Nisini(2010)ApJ_724_69}.
The $\chi^2$ contour for the column densities show a rough correlation for both clumps, C and G (cf.~Fig.~\ref{fig-mconf}).
This seems to be caused by the fact as follows.
If both column densities increase or decrease together, then the $\chi^2$ can be decreased by adjusting the power-law index $b$ shared by the lower and higher density \Htwo{} gases.
However, if one column density increases and the other decreases, or vice versa, then the $\chi^2$ cannot be decreased by the same way as before, since the total \Htwo{} population changes its shape, which is related with the power-law index $b$.
The variation of \nHtwo{} and \NHtwo{} are not effective as of $b$ in adjusting the fitting, as the confidence intervals indicate (Table \ref{tbl-mfit}).
For the clump C, almost no correlation is shown between the densities and between $b$ and $X_H$.
For the clump G, the densities show a complex $\chi^2$ surface.
It may be caused by the absence of the lowest level data, ($\upsilon,J$)=(0,2) and (0,3), which imposes a weaker constraint for the fitting.
\section{Discussion} \label{dis}
\subsection{Shocked \Htwo{} Gas Described by the Power-law Thermal Admixture Model} \label{dis-plmod}
\subsubsection{Comparison Between Model Parameters from Previous and Our Studies} \label{dis-plmod-cmp}
The first attempt to describe the level population of shocked \Htwo{} gas with the power-law mixture of thermal \Htwo{} gas was by \cite{Brand(1988)ApJ_334_L103} (see also \citealt{Burton(1989)inproca}).
They assumed that the \Htwo{} gas, shocked by J-type shocks \citep{Draine(1993)ARA&A_31_373}, is in LTE and the postshock temperature profile is determined through the \Htwo{} radiative cooling, which is proportional to $T^{4.7}$.
In this way, they modeled the level population of thermally mixed \Htwo{} gas at the postshock region, which is equivalent to the power-law thermal admixture of \Htwo{} gas in LTE with the power-law index $b=4.7$ (cf. section \ref{ana-res-plmod}).
They applied this model to the observational data of OMC-1, and found that it is successful in describing the available data at that time.
However, the omission of important coolants---like CO, OH, H$_2$O, and grain---at dense environments as OMC-1 and of the magnetic field make the model assumption doubtful \citep{Chang(1991)ApJ_378_202,Draine(1993)ARA&A_31_373}.
Then, \cite{Oliva(1990)A&A_240_453} first tried to use the method of power-law thermal mixing as a phenomenological description tool for the \Htwo{} level population, with no background physics; they simply mixed \Htwo{} gas in LTE according to the power-law distribution, $dN(T)\sim T^{-b}dT$.
In this approach, they found that the \Htwo{} population of the SNR RCW 103, obtained from near-infrared observations, can be described with the models whose power-law indices are between $b=3.8$ and $b=4.7$.
This method was extended to the non-LTE case by \cite{Neufeld(2008)ApJ_678_974}.
Applying to the \spitzer{} IRAC data, they found that the shocked \Htwo{} gas of the supernova remnant IC 443 can be described with the power-law index $b$ of $\sim3-6$ and the molecular hydrogen density of \nHtwo{} $\sim10^6-10^7$ \ncm.
Afterwards, the same method was applied to the shocked \Htwo{} of the SNR HB 21 \citep{Shinn(2009)ApJ_693_1883,Shinn(2010)AdSpR_45_445} and of the outflows from YSOs \citep{Neufeld(2009)ApJ_706_170,Shinn(2010)AdSpR_45_445,Lee(2010)ApJ_709_L74,Takami(2010)ApJ_720_155,Nisini(2010)ApJ_724_69,Yuan(2011)ApJ_726_76}.
All of the non-LTE application results are summarized in Table \ref{tbl-nh2b}; we excluded the work of \cite{Nisini(2010)ApJ_724_69}, because they varied $T_{min}$ for the model application.
Table \ref{tbl-nh2b} shows that \nHtwo{} and $b$ vary with the estimated levels ($\upsilon=0,J$).
\nHtwo{} and $b$ are both tend to be smaller when the model applied to lower $J$ levels; \nHtwo{} shows a more drastic variation than $b$.
This tendency was already noted in \cite{Shinn(2010)AdSpR_45_445} through the model application to the data of the SNR HB 21 and OMC-1; now, the tendency is strengthened by additional data from other SNRs and outflows of YSOs.
This indicates that the level population of shocked \Htwo{} gas is not described by the power-law admixture model with \emph{single} \nHtwo{} and $b$.
Moreover, it suggests we should analyze the level population over as many levels as we can get; if not, we may have a limited picture on the shocked \Htwo{} gas.
In the point of extending the data coverage, our observations are important because we extended the $\upsilon=0$ level population obtained from previous mid-infrared studies, from $E(\upsilon,J)\lesssim7,000$ K up to $\lesssim22,000$ K (see Fig.~\ref{fig-popall}).
We could reproduce the estimated \Htwo{} population with two \nHtwo{} and one $b$ after including the H atom as an additional collision partner (Table \ref{tbl-mfit}).
The derived two \nHtwo{}s, \nHtwo=$10^{2.8-3.8}$ \ncm{} and \nHtwo=$10^{5.4-5.8}$ \ncm, fall into the range previously obtained, \nHtwo=$10^{2.7-7.0}$ \ncm{} (Table \ref{tbl-nh2b}).
As distinct from the previous applications (Table \ref{tbl-nh2b}), we included H atoms in the model, which efficiently excite high-$J$ $\upsilon=0$ levels; hence, we obtained smaller \nHtwo{} for the higher density component than we did without H atoms, by a factor of 0.8 dex and 1.7 dex for the clump C and G, respectively.
As mentioned in section \ref{ana-res-plmod}, the lower and higher density component mainly contribute to the lower and higher energy levels, respectively (Fig.~\ref{fig-mfit}).
This clearly show why we get different \nHtwo{}s depending on which levels are used for the estimation (cf. Table \ref{tbl-nh2b}).
This density property also indicates that there may be a non-uniqueness in the model parameter, because the observed population constrains only small portion of the modeled population (cf.~section \ref{ana-res-plmod}).
Indeed, the $T_{max}$ of the lower density gas and the $T_{min}$ of the higher density gas can be lower and higher, respectively, since small variations of those values may have a negligible effect on the fitting results.
We checked these possibilities by varying $T_{max}$ and $T_{min}$, with the obtained fitting parameters (Table \ref{tbl-mfit}).
In the case of clump C, we found that the $T_{max}$ of the lower density gas cannot be decreased from the initial model setting 4000 K, while the $T_{min}$ of the higher density gas can be relaxed up to 1000 K.
This result means that a small amount of warm, high density \Htwo{} gas can explain the high energy-level population, i.e., $N$(\Htwo; $T>1000$ K)$\sim4\times10^{18}$ \Ncm.
In the similar sense, the $X_H$ in the lower density gas can be lower, even down to the hydrogen-free case, \nHone/\nHtwo=0, since the excitation of the low-$J$ $\upsilon=0$ levels is dominantly determined by the collisions with \Htwo{} rather than \astH{I}.
\subsubsection{Nature of the Shocks} \label{dis-plmod-shock}
There have been many observational studies trying to identify the shock type at the clump C and G at diverse wavelengths---millimeter, sub-millimeter, and infrared.
Much of the observational results were compared with several J- and C-type shocks by \cite{Snell(2005)ApJ_620_758}, and they concluded that a combination of shocks (dissociative and non-dissociative) is required to explain all the observational results.
Later, \spitzer{} mid-infrared spectral observations showed the emission lines likely from such a combination of dissociative (J-type) and non-dissociative (C-type) shocks at the clump C \citep{Neufeld(2007)ApJ_664_890}.
Besides, previous studies on the shocked \Htwo{} gas of protostellar outflows which covered a similar \EvJ{} range with ours ($\sim0-25,000$ K) showed that a mixture of C- and J-type properties are required to explain the \Htwo{} level population \citep{Flower(2003)MNRAS_341_70,Giannini(2006)A&A_459_821,Gusdorf(2008)A&A_490_695}.
Under the shock-combination preference, the immediate question is which type of shocks the two component \Htwo{} gas originate from.
We first considered the lower density gas which occupies most mass of the shocked \Htwo{} gas (Fig.~\ref{fig-mfit} and Table \ref{tbl-mfit}).
If the lower density gas originates from J-type shocks, the recombinational lines of \astH{I} are also expected.
We checked the relative intensity of Br$\alpha$ 4.05 \um{} line and \Htwo{} $\upsilon=0-0$ S(3) line from a theoretical model \citep{Hollenbach(1989)ApJ_342_306}.
The ratio of Br$\alpha$/[\Htwo{} $\upsilon=0-0$ S(3)] is between $\sim1-10$ over the shock velocity $50-100$ \kms{} and preshock density $10^3-10^6$ \ncm.
The \Htwo{} density obtained from the model fitting likely represent the postshock density, since the \Htwo{} emissions are mainly emanating from the reformed \Htwo{} gas in J-type shocks \citep{Hollenbach(1989)ApJ_342_306}.
The typical J-type shock velocity is around 100 \kms{} under a general interstellar medium condition \citep{Draine(1993)ARA&A_31_373}, hence the density of the molecular reformation regions at far down stream would be $\gtrsim100$ times higher than the preshock density; this density ratio can be seen from many numerical studies \cite[e.g.~][]{Allen(2008)ApJS_178_20,Flower(2010)MNRAS_406_1745}.
Therefore, the preshock density should be \nHtwo{} $\lesssim10^{1.8}$ \ncm{} for the lower-density gas to be from J-type shocks.
This preshock density range is somewhat lower than considered in the above theoretical model ($10^3-10^6$ \ncm), but the similar ratio Br$\alpha$/[\Htwo{} $\upsilon=0-0$ S(3)] would be maintained since the postshock structure is insensitive to the preshock density in J-shocks.
Therefore, if the lower density gas is from J-shocks, then there should be Br$\alpha$ of $\sim2\times10^{-3}-2\times10^{-2}$ \luerg{}, based on the observed intensity of \Htwo{} $\upsilon=0-0$ S(3) \citep{Neufeld(2007)ApJ_664_890}.
However, no Br$\alpha$ line of such a high intensity was detected (Fig.~\ref{fig-spec}).
In addition, theoretically, a J-shock produce more abundant high-T ($\sim10^3$ K) \Htwo{} gas than a C-shock \citep{Wilgenbus(2000)A&A_356_1010,Flower(2010)MNRAS_406_1745}.
If the lower density gas is from J-shocks, the \Htwo{} population diagram must show a much up-turn curvature than we obtained (Fig.~\ref{fig-mfit}).
Overall, the lower density gas is not likely from J-type shocks.
We then checked whether the lower density gas originate from C-type shocks.
The postshock gas behind a planar C-type shock can be approximated as an isothermal gas whose density \nHtwo{} and temperature $T_s$ are as follows \citep{Neufeld(2006)ApJ_649_816},
\begin{eqnarray}
\textrm{\nHtwo}=1.5\,n_0 \\
T_s=375\,b_B^{-0.36}\,\upsilon_{s6}^{1.35} \textrm{ K,}
\end{eqnarray}
where $10\,\upsilon_{s6}$ \kms{} is the shock velocity and $b_B(n_0/\textrm{\ncm})^{1/2}$ $\mu$G is the assumed preshock magnetic field for a shock propagating in material of preshock \Htwo{} density $n_0$.
We here note the above approximation was tested over $n_0=10^4-10^6$ \ncm{} \citep{Neufeld(2006)ApJ_649_816}.
From this approximation, we can say the lower density gas is from C-shocks propagating into a preshock medium \nHtwo{} $\sim10^{2.6-3.6}$ \ncm.
Interestingly, this density is similar with the typical density of molecular clouds ($\gtrsim10^3$ \ncm), where the hydrogen dominantly exists in the molecular form \Htwo{} \citep{Snow(2006)ARA&A_44_367}.
Such a similarity gives a natural way to explain the mass dominance of the lower density gas (Fig.~\ref{fig-mfit}), since \Htwo{} survives C-type shocks.
Therefore, C-type shocks are more suitable to explain the lower density gas than J-type shocks.
If the lower density gas is from C-type shocks, the power-law index $b$ may reflect the multiple shocks of different velocities.
According to the approximation with $b_B=1$, it ranges from 3 to 58 \kms, which corresponds to $\sim100-4000$ K gas.
This velocity range is theoretically expected for a C-type shock whose preshock density is $\sim10^3$ \ncm{} and $b_B=1$ \citep{LeBourlot(2002)MNRAS_332_985}.
Also, such a diversity in the shock velocity is consistent with the results of \cite{Hewitt(2009)ApJ_694_1266} who found two C-shocks are preferred to explain the excitation of shocked \Htwo{} gas around SNRs over the range \EvJ{} $\sim0-8,000$ K.
However, we here note that this shock velocity range depends on the model parameters we fixed ($T_{min},T_{max}$); therefore, the shock velocity range mentioned above ($3-58$ \kms) is not observationally determined one.
Instead, it should be understood that multiple shocks of a few to a few tens of kilometer per second are required to explain the level population of the lower density gas.
The diversity in the shock velocity may be a natural result during the shock-cloud interaction, since the shock driving pressure would be higher at the head than the side.
This kind of property can be inferred from a numerical simulation for radiative clouds \citep{Nakamura(2006)ApJS_164_477}, which showed that the velocity dispersion of the shocked cloud is larger along the shock-normal (the blast wave direction; $z$-axis in \citealt{Nakamura(2006)ApJS_164_477}) than the shock-tangential ($\varpi$-axis in \citealt{Nakamura(2006)ApJS_164_477}); this velocity dispersion does not exactly corresponds to the dispersion of the shock velocity in the cloud, however, it shows the tendency which direction the shock would propagate faster and cause the larger velocity dispersion.
We assigned the same \astH{I} abundance both for the lower and higher density gas, and obtained $X_H=-1.7$, equivalent to \nHone/\nH{} $=0.01$ and \nHone/\nHtwo{} $=0.02$ (Table \ref{tbl-mfit}).
This value is higher for typical molecular clouds \citep[\nH{} $\gtrsim10^4$ \ncm, ][]{Draine(1983)ApJ_264_485}.
However, as mentioned in section \ref{dis-plmod-cmp}, we can achieve the population of the lower density gas even with zero \astH{I} abundance, since the population is dominantly determined by the collisions with \Htwo{} rather than \astH{I}.
Considering this fact, the C-shock interpretation for the lower density gas has no contradiction with the \astH{I} abundance.
Now we move to discuss which type of shock the higher density gas originate from, considering the \astH{I} abundance first.
The \astH{I} abundance, $X_H$, is mainly determined by the vertical gap between $\upsilon=0$ and $\upsilon>0$ levels in the population diagram, and the gap is dominantly described by the higher density gas (see Fig.~\ref{fig-mfit}).
The obtained $X_H=-1.7$, equivalent to \nHone/\nH{} $=0.01$ and \nHone/\nHtwo{} $=0.02$, is expected for a typical diffuse cloud (\nH{} $\sim10^2$ \ncm), but high for a typical molecular cloud (\nH{} $\gtrsim10^4$ \ncm) which has $X_H\lesssim-4.0$ \citep{Draine(1983)ApJ_264_485}.
Therefore, the obtained $X_H$ must be from one of the following cases: (1) non-dissociative (C-type) shocks propagating into molecular clouds of the high $X_H$ value; (2) non-dissociative (C-type) shocks propagating into diffuse clouds; (3) dissociative (J-type) shocks propagating into molecular clouds.
We checked the first case.
In view of the calculations of \cite{Solomon(1971)ApJ_165_41}, \cite{Draine(1983)ApJ_264_485}, and \cite{Goldsmith(2005)ApJ_622_938}, the total ionization rate for \Htwo{} must be $\zeta\sim10^{-13}-10^{-12}$ s$^{-1}$ to make the obtained abundance in the higher density \Htwo{} gas of \nHtwo{} $\sim10^{5.0}$ \ncm.
However, the measured rate is much smaller than the required: $\zeta\sim5.5\times10^{-16}$ s$^{-1}$ \citep{Hewitt(2009)ApJ_706_L270} and $\zeta\sim2.6\times10^{-15}$ s$^{-1}$ \citep{Indriolo(2010)ApJ_724_1357}.
The second case is negative either.
It was shown that the postshock gas behind a planar C-type shock can be approximated with a isothermal gas whose density is equal to $\sim1.5\,n_{preshock}$ \citep{Neufeld(2006)ApJ_649_816}.
If the shock is C-type, the obtained density of \nHtwo{} $=10^{5.4-5.8}$ \ncm{} is contradict with the requirement that the preshock medium must be a diffuse cloud (\nH{} $\sim10^2$ \ncm).
In the third case, however, the high $X_H$ of $-1.7$ is expected under appropriate conditions; such an $X_H$ can be achieved at the $T\sim1000-4000$ K region where the \Htwo{} emission is mainly emanating from reformations \citep{Hollenbach(1989)ApJ_342_306,Flower(2010)MNRAS_406_1745}.
If the higher density gas is from J-type shocks where the \Htwo{} emission dominantly comes from the postshock reformation regions, the obtained model parameters for the higher density gas (Table \ref{tbl-mfit}) can be regarded to represent \emph{characteristic} physical parameters of the reformation regions.
Based on this, the preshock density should be a factor of $\lesssim100$ smaller than the obtained density of \nHtwo{} $=10^{5.4-5.8}$ \ncm{} since the typical velocity of J-shocks is around 100 \kms.
Also, there should exist shocked \astH{I} gas, the raw material for the reformation of \Htwo.
The shocked \astH{I} gas has already been observed along the ``W'' ridge (Fig.~\ref{fig-slit}) we are studying \citep{Braun(1986)A&A_164_193,Lee(2008)AJ_135_796}.
\cite{Lee(2007)phdth} measured the column densities of the shocked \astH{I} gas at the clump C and G, which are about $\sim3\times10^{21}$ \Ncm.
If we assume that the shocked \astH{I} spreads over along our line-of-sight as long as the width of the ``W'' ridge ($\sim2'\sim1$ pc), then the mean \astH{I} density of $\sim10^3$ \ncm{} is obtained, which is smaller than the postshock \Htwo{} density \nHtwo{} $=10^{5.4-5.8}$ \ncm.
These \nHone{} and \nHtwo{} are consistent with the J-shock interpretation, since the \astH{I} recombines before the \Htwo{} and the postshock density increases along the distance from the shockfront.
We also checked the line profile of \astH{I} 21cm emission \citep{Lee(2007)phdth} and \HtwolineK{} emission \citep{Rosado(2007)AJ_133_89} at the clump C, and found that they both show the shocked component at the similar velocity $\sim-30$ \kms.
Therefore, the J-shock interpretation for the higher density gas seems plausible again.
We then considered the meaning of the power-law index $b$ in the higher density gas.
As mentioned in section \ref{dis-plmod-cmp}, $T_{min}$ of the higher density gas can be increased from 100 K to 1000 K without making any difference on the final level population of \EvJ{} $\gtrsim10^4$ K levels.
This means that the gas in $T=1000-4000$ K is necessary to describe the observed population of \EvJ{} $\gtrsim10^4$ K levels.
We thus checked whether the $b$ variation makes any significant change to the level population of \EvJ{} $\gtrsim10^4$ K, summed over $T=1000-4000$ K; the obtained \nHtwo{} and $X_H$ was fixed while the $b$ value was varied.
We found that the $b$ variation makes a negligible effect over $b=1.0-4.0$.
This is because the population of \EvJ{} $\gtrsim10^4$ K levels has an almost same slope over the temperature integration range ($1000-4000$ K) in the power-law thermal model.
Therefore, the $b$ value has almost no practical meaning for the higher density gas.
From all the above discussion, the lower and higher density gases are likely from the C-type and J-type shocks, respectively.
This conclusion is consistent with the conclusion from previous studies which claimed a mixture of dissociative and non-dissociative shocks for the description of the observed results \citep{Snell(2005)ApJ_620_758,Neufeld(2007)ApJ_664_890}.
If we interpret our results based on the hierarchical picture of molecular clouds \citep{Bergin(2007)ARA&A_45_339,Williams(2000)2000prpl.conf__97}, the C-type shocks propagate into ``clumps,'' while the J-type shocks propagate into ``clouds'' (interclump media); the typical densities of these two constituents (``clouds'' and ``clumps'') are $50-500$ \ncm{} and $10^3-10^4$ \ncm, respectively.
This interpretation is consistent with the requirement that the ram pressure ($\sim\rho v_{s}^{2}$) should be similar between C-type and J-type shocks.
In the above discussion, we claimed that C-type shocks propagate into a preshock medium of $\sim10^3$ \ncm{} with a shock velocity of a few 10 \kms, while J-type shocks propagate into a preshock medium of $<10^3$ \ncm{} with a shock velocity of $\sim100$ \kms.
Before closing this section, we note that the \HtwolineK{} is dominantly originating from J-type shocks rather than C-type shocks, according to our results.
Interestingly, it was observed that the [\astSi{II}] 34.8 \um{} emission well follows the \HtwolineK{} emission at the clump C \citep{Richter(1995)ApJ_454_277,Neufeld(2007)ApJ_664_890}; the [\astSi{II}] 34.8 \um{} emission is an efficient cooling line at $T\lesssim5000$ K where the \Htwo{} reformation also happens \citep{Hollenbach(1989)ApJ_342_306}.
The intensity of [\astSi{II}] and \HtwolineK{} are $0.97$ and $1.8$ in $10^{-4}$ \luerg, respectively (\citealt{Neufeld(2007)ApJ_664_890}, Table \ref{tbl-h2col}).
The ratio [\astSi{II}]/\HtwolineK{} is 0.54, which is significantly lower than the theoretical expectation of a 100 \kms{} J-shock into $10^3$ \ncm{} gas, $\sim5.0$ \citep{Hollenbach(1989)ApJ_342_306}.
In spite of this difference, the J-shock origin of \HtwolineK{} emission is worth to check further whether it is general in shocked regions, considering that the shocked \Htwo{} gas usually shows a similar shape in the level population over \EvJ{}$=0-25,000$ K (cf.~section \ref{dis-plmod-coher}).
\subsubsection{Embedded Coherence in the Obtained Model Parameters} \label{dis-plmod-coher}
Table \ref{tbl-mfit} shows some coherences of the obtained model parameters for the two-component gases, although the clump G has no data for the population of \EvJ=(0,2) and (0,3) which are important in the determination of $b$ and density.
For example, the density ratio and the column density ratio of the two-component gases are not much different at the each clump, C and G.
This may originate from some common properties of the preshock medium.
One step forward, \cite{Richter(1995)ApJ_454_277,Richter(1995)ApJ_449_L83} already noted that the similarity between the \Htwo{} level population of SNRs and outflows of YSOs.
We can study such similarities with the two-\nHtwo{} power-law thermal admixture model by comparing the model parameters, and may lead out the properties of the preshock medium.
Especially, the power-law index $b$ would give information about how shock velocity is diverse at the shock-cloud interaction regions, which may be related with the geometrical structure of the clouds.
The coming James Webb Space Telescope which can perform spectral observations over the $\sim1-30$ \um{} wavelength range would provide an excellent opportunity for such studies.
\subsection{Warm Diffuse Background \Htwo{} Gas} \label{dis-bg}
At the background region (BG in Fig.~\ref{fig-slit}), we detected the \Htwo{} $\upsilon=0-0$ S(9) line (Fig.~\ref{fig-fit-bg}).
The extinction corrected column density gives log[N(\Htwo; $\upsilon,J$=0,11)]=$15.08\pm0.11$ \Ncm{} (Table \ref{tbl-h2col}).
This emission may be related with the \emph{diffuse background \Htwo{} gas}, previously observed towards SNRs \citep{Neufeld(2007)ApJ_664_890,Hewitt(2009)ApJ_694_1266}.
In order to check whether the obtained ($\upsilon,J$=0,11) population is plausible for such an diffuse background \Htwo{} gas, the population was compared with those obtained towards other Galactic lines of sight.
Since the relative shape of the level population determines the physical condition of the \Htwo{} gas, we must compare such an relative shape, thus we need the population of other levels in addition to ($\upsilon,J$=0,11) level.
For this, we employed the ($\upsilon,J$=0,2) population of the diffuse background \Htwo{} gas, previously observed towards the clump C of IC 443 \citep{Neufeld(2007)ApJ_664_890}, as the ($\upsilon,J$=0,2) level population of our BG region (Fig.~\ref{fig-slit}).
We adopted $I_{diff}=1.5\times10^{-6}$ \luerg{} as the intensity for the \Htwo{} $\upsilon=0-0$ S(0) line \cite[cf.~Fig.~19 of][]{Neufeld(2007)ApJ_664_890}, and corrected the extinction with $A_V=13.5$ employing the ``Milky Way, \Rv=3.1'' extinction curve \citep{Weingartner(2001)ApJ_548_296,Draine(2003)ARA&A_41_241}.
The population is log[N(\Htwo; $\upsilon,J$=0,2)]=$19.05\pm0.07$ \Ncm.
After all, the weighted populations of the ($\upsilon,J$)=(0,2) and (0,11) levels for the BG region are log[N($\upsilon,J$)/g$_J$]=18.3 and 13.2, respectively.
We compared the populations of these two levels with those of ($\upsilon,J$)=($0,\le8$) obtained toward the ``translucent lines of sight'' to the Galactic background stars, from the far-ultraviolet observations of \Htwo{} absorption lines \citep{Jensen(2010)ApJ_711_1236}.
Since the population of ($\upsilon,J$)=($0,11$) is absent in the results of \cite{Jensen(2010)ApJ_711_1236}, the relative curvature in the population diagram was compared, by scaling up or down the populations of ($\upsilon,J$)=(0,2) and (0,11) levels.
From the comparison, we found that our populations make a more upturn curvature than expected from the excitation temperature of $J\ge2$ levels, $\sim200-550$ K \citep{Jensen(2010)ApJ_711_1236}.
This means the existence of \Htwo{} gas with higher temperature than 550 K, which is consistent with the suggestion of \cite{Jensen(2010)ApJ_711_1236}: more than two temperature components are probably necessary to describe the population.
\cite{Jensen(2010)ApJ_711_1236} tried to fit the population obtained toward two sample targets, with a model for photodissociation regions; however, the fits were poor.
Such a mismatch was also recognized by \cite{Gry(2002)A&A_391_675} and \cite{Falgarone(2005)A&A_433_997}, and they suggested that another collisional excitation mechanisms are required to explain the observed population, such as magnetohydrodynamic shocks or intermittent dissipation of turbulence.
The diffuse background \Htwo{} gas towards the SNR IC 443 may originate from these mechanisms.
\section{Conclusions} \label{concl}
We present near-infrared ($2.5-5.0$ \um) spectra toward three shocked \Htwo{} clumps (B, C, and G) of the SNR IC 443 and one background (BG) region (cf.~Fig.~\ref{fig-slit}).
The observations were performed with the satellite \akari{} during its post-Helium phase.
Only \Htwo{} emission lines were detected toward all four directions.
We measured the line intensities and, after the reddening-correction, obtained the relevant level populations.
The level populations were compared with the ones from previous near- and mid-infrared observations.
\akari{} level populations are well fitted with those from previous mid-infrared space observations, while there is a systematic difference with those from previous near-infrared ground observations.
We attributed this difference to the calibration error of previous near-infrared observations, although it may be caused by the different position of observed regions.
With the \akari{} near-infrared observations, we could extend the level population of shocked \Htwo{} gas in the clump C and G obtained from mid-infrared observations, from $\sim7000$ K to $\sim22000$ K continuously.
From these combined level populations, we found that the $\upsilon=0$ levels are more populated than the $\upsilon>0$ levels at a fixed level energy, which means the population cannot be reproduced with any combination of \Htwo{} gas in LTE, usually used for the description of shocked \Htwo{} gas.
The populations were well described two-\nHtwo{} power-law thermal admixture model, including the H atom as a collisional partner in addition to \Htwo{} and He.
The model parameters are two number densities \nHtwo, two column densities \NHtwo, one power-law index $b$, and the \astH{I} to \Htwo{} abundance $X_H$ (cf.~Table \ref{tbl-mfit}).
We attributed the lower (\nHtwo{} $=10^{2.8-3.8}$ \ncm) and higher (\nHtwo{} $=10^{5.4-5.8}$ \ncm) density gases to the shocked \Htwo{} gas behind C-type and J-type shocks, respectively, based on several arguments such as the preshock density, the amount of shocked gas, the obtained high \astH{I} abundance, the shape of level populations, the line profile, etc.
This interpretation is consistent with the hierarchical picture of molecular clouds whose constituents are classified into ``clouds,'' ``clumps,'' and ``cores'' \citep{Bergin(2007)ARA&A_45_339,Williams(2000)2000prpl.conf__97}.
The C-type shocks are likely propagating into the ``clumps'' ($\sim10^3$ \ncm), while the J-type shocks are propagating into the ``clouds'' (interclump media, $<10^3$ \ncm).
The power-law index $b$, mainly determined by the lower density gas, is attributed to the diversity in the shock velocity propagating into the clouds.
Such a velocity diversity may be a natural result during the shock-cloud interaction, since the shock driving pressure would be higher at the head than the side.
According to our results, the \HtwolineK{} emission is mainly from J-type shocks propagating into interclump media.
In addition, our power-law thermal admixture model would be useful to perform statistical studies on the shocked \Htwo{} gas observed around other SNRs and YSOs, which shows coherent excitation characteristics.
\Htwo{} emission was also detected at the BG region, and we attributed it to the diffuse \Htwo{} gas, pervaded in the Milky Way.
This gas may be excited by collisional processes, like shocks or turbulence dissipation, in addition to ultraviolet photon pumping.
\acknowledgments
This work is based on observations with \akari, a JAXA project with the participation of ESA.
The authors thank all the members of the \akari{} project.
J. H. S. is grateful to Jae-Joon Lee for the useful discussion on the two-density interpretation.
This research has made use of SAOImage DS9, developed by Smithsonian Astrophysical Observatory \citep{Joye(2003)inproc}.
{\it Facilities:} \facility{Akari}
\bibliographystyle{G:/Work/Publication/bibtex/astronat/apj/apj}
|
\section*{Introduction}
Under appropriate conditions, the resolvent of the Laplacian on an
asymptotically hyperbolic space continues analytically through the spectrum
\cite{Mazzeo-Melrose:Meromorphic}. In this paper we obtain estimates on
the analytic continuation of the resolvent for the Laplacian of a metric
that is a small perturbation of the Poincar\'e metric on hyperbolic
space. In particular we show for these perturbations of the metric, and
allowing in addition a real-valued potential, that there are only a finite
number of poles for the analytic continuation of the resolvent to any
half-plane containing the physical region and that the resolvent satisfies
polynomial bounds on appropriate weighted Sobolev spaces near infinity in
such a strip. This result, for a small strip, is then applied to the
twisted Laplacian which is the stationary part of the d'Alembertian on de
Sitter-Schwarzschild. In a companion paper
\cite{Melrose-SaBarreto-Vasy:Asymptotics} the decay of solutions to the
wave equation on de Sitter-Schwarzschild space is analyzed using these
estimates.
In the main part of this paper constructive semiclassical methods are used
to analyze the resolvent of the Laplacian, and potential perturbations of
it, for a complete, asymptotically hyperbolic, metric on the interior of the
ball, $\Bn,$
\begin{equation}
\begin{gathered}
g_{\delta}=g_0+\chi_\delta(z)H,\\
g_0=\frac{ 4 dz^2}{(1-|z|^2)^2},\
\chi_\delta(z)=\chi\left(\frac{(1-|z|)}{\delta}\right).
\end{gathered}
\label{SeClRe.14}\end{equation}
Here $g_{0}$ is the standard hyperbolic metric, $H=H(z,dz)$ is a symmetric
2-tensor which is smooth up to the boundary of the ball and $\chi\in
\CI(\bbR),$ has $\chi(s)=1$ if $|s|<\ha,$ $\chi(s)=0$ if $|s|>1.$ The
perturbation here is always the same at the boundary but is cut off closer
and closer to it as $\delta \downarrow0.$ For $\delta >0$ small enough we
show that the analytic continuation of the resolvent of the Laplacian is
smooth, so has no poles, in the intersection of the exterior of a
sufficiently large ball with any strip around the real axis in the
non-physical half-space as an operator between weighted $L^2$ or Sobolev
spaces, and obtain high-energy estimates for this resolvent in this strip.
A special case of these estimates is as follows: Let
$x=\frac{1-|z|}{1+|z|},$ $W\in\CI(\Bn)$ be real-valued and let $R_\delta
(\sigma)=(\Delta_{g_\delta}+x^2W-\sigma^2-n/2^2)^{-1}$ denote the resolvent
of $\Delta_{g_\delta}+x^2W.$ The spectral theorem shows that
$R_\delta(\sigma)$ is well defined as a bounded operator in $L^2(\Bn;dg)$
if $\im \sigma <<0,$ and the results of Mazzeo and the first author show
that it continues meromorphically to the upper half plane. Here we show
that there exists a strip about the real axis such that if $\delta$ is
small and $a$ and $b$ are suitably chosen $x^a R_\delta(\sigma) x^b $ has
no poles, provided $|\sigma|$ is large, and moreover we obtain a polynomial
bound for the norm of $x^a R_\delta(\sigma) x^b.$ More precisely, with
$H^k_0(\Bn)$ the $L^2$-based Sobolev space of order $k$, so for $k=0$,
$H^0_0(\Bn)=L^2(\Bn;dg):$
\begin{thm}(See Theorem~\ref{resolvent-bounds} for the full statement.)
\label{resolvent-bounds-simple} There exist $\delta_0>0,$ such that if
$0\leq \delta\leq \delta_0,$ then
$x^a R_{\delta}(\sigma) x^b$ continues holomorphically to the region $\im
\sigma < M,$ $M>0$ $| \sigma| >K(\delta,M),$ provided $ \im\sigma< b,$
and $a>\im \sigma.$ Moreover, there exists $C>0$ such that
\begin{gather}
\begin{gathered}
|| x^a R_\delta (\sigma) x^b v||_{H^k_0(\Bn)} \leq
C |\sigma|^{-1+\frac{n}{2} +k} ||v||_{L^2(\Bn)}, \;\ k=0,1,2, \\
|| x^a R_\delta (\sigma) x^b v||_{L^2(\Bn)} \leq
C |\sigma|^{-1+\frac{n}{2} +k} ||v||_{H_0^{-k}(\Bn)}, \;\ k=0,1,2, \\
\end{gathered} \label{sobolev1-simple}
\end{gather}
\end{thm}
This estimate is not optimal; the optimal bound is expected to be $O(
|\sigma|^{-1+k}).$ The additional factor of $|\sigma|^{\frac{n}{2}}$ results
from ignoring the oscillatory behavior of the Schwartz kernel of the
resolvent; a stationary phase type argument (which is made delicate by the
intersecting Lagrangians described later) should give an improved
result. However, for our application, which we now describe, the polynomial
loss suffices, as there are similar losses from the trapped geodesics in
compact sets.
As noted above, these results and the underlying estimates can be applied
to study the wave equation on $1+(n+1)$ dimensional de Sitter-Schwarzschild
space. This model is given by
\begin{equation}
M=\RR_t\times \stackrel{\circ}{X}, \;\ X=[r_\bH,r_{\sI}]_r\times\ms^{n}_\omega,
\label{model-ndim}
\end{equation}
with the Lorentzian metric
\begin{equation}\label{eq:dS-Sch-metric}
G=\alpha^2\,dt^2-\alpha^{-2}\,dr^2-r^2\,d\omega^2,
\end{equation}
where
$d\omega^2$ is the standard metric on $\ms^{n},$
\begin{equation}\label{eq:alpha-def}
\alpha=\left(1-\frac{2m}{r}-\frac{\Lambda r^2}{3}\right)^{1/2},
\end{equation}
with $\Lambda$ and $m$ positive constants satisfying
$0<9m^2\Lambda<1,$ and $r_\bH,r_{\sI}$ are the two positive roots of
$\alpha=0.$
The d'Alembertian associated to the metric $G$ is
\begin{equation}
\square= \alpha^{-2}(D_t^2-\alpha^2r^{-n}D_r(r^{n}\alpha^2
D_r)-\alpha^2r^{-2} \Delta_\omega).
\label{dal}\end{equation}
An important goal is then to give a precise description of the asymptotic
behavior, in all regions, of space-time, of the solution of the wave
equation, $\Box u=0$, with initial data which are is necessarily compactly
supported. The results given below can be used to attain this goal; see
the companion paper \cite{Melrose-SaBarreto-Vasy:Asymptotics}.
Since we are only interested here in the null space of the d'Alembertian,
the leading factor of $\alpha ^{-2}$ can be dropped. The results above can
be applied to the corresponding stationary
operator, which is a twisted Laplacian
\begin{equation}
\Delta_X=\alpha^2r^{-n}D_r(\alpha^2 r^{n}D_r)
+\alpha^2r^{-2}\Delta_\omega.
\label{modelX}
\end{equation}
In what follows we will sometimes consider $\alpha$ as a boundary defining
function of $X.$ This amounts to a change in the $\CI$ structure of
$X;$ we denote the new manifold by $X_{\ha}.$
The second order elliptic operator $\Delta_X$ in \eqref{modelX} is
self-adjoint, and non-negative, with respect to the measure
\begin{equation}
\Omega=\alpha^{-2} r^{n} \,d r \,d \omega,\label{measure}
\end{equation}
with $\alpha$ given by \eqref{eq:alpha-def}. So, by the spectral theorem,
the resolvent
\begin{gather}
R(\sigma)=(\Delta_X - \sigma^2 )^{-1} : L^2 ( X ; \Omega )
\longrightarrow L^2 ( X ; \Omega ) \label{resolv1}
\end{gather}
is holomorphic for $\im \sigma < 0 .$ In
\cite{Sa-Barreto-Zworski:Distribution} the second author and Zworski, using
methods of Mazzeo and the first author from
\cite{Mazzeo-Melrose:Meromorphic}, Sj\"ostrand and Zworski
\cite{Sjostrand-Zworski:ComplexScaling} and Zworski \cite{Zworski:LimitSet}
prove that the resolvent family has an analytic continuation.
\begin{thm}(S\'a Barreto-Zworski, see
\cite{Sa-Barreto-Zworski:Distribution})\label{mercont} As operators
\begin{gather*}
R ( \sigma ): \CIo( \stackrel{o}{X} )\longrightarrow\CI( \stackrel{o}{X} )
\end{gather*}
the family \eqref{resolv1} has a meromorphic continuation to $\mc$ with
isolated poles of finite rank. Moreover, there exists $\eps>0$ such that
the only pole of $R(\sigma)$ with $\im \sigma<\eps$ is at $\sigma=0;$ it
has multiplicity one.
\end{thm}
\noindent
Theorem \ref{mercont} was proved for $n+1=3,$ but its proof easily extends
to higher dimensions.
In order to describe the asymptotics of wave propagation precisely on $M$
via $R(\sigma)$ it is necessary to understand the action of $R(\sigma)$ on
weighted Sobolev spaces for $\sigma$ in a strip about the real axis as
$|\re\sigma|\to\infty$. The results of \cite{Mazzeo-Melrose:Meromorphic}
actually show that $R(\sigma)$ is bounded as a map between the weighted
spaces in question; the issue is uniform control of the norm at high energies. The strategy is to obtain bounds for $R(\sigma)$ for $\re(\sigma)$ large
in the interior of $X,$ then near the ends $r_{\bH}$ and $r_{\sI},$ and later glue those estimates. In the case $n+1=3,$ we can use the following result of Bony and H\"afner to obtain bounds for the resolvent in the interior
\begin{thm}(Bony-H\"afner, see
\cite{Bony-Haefner:Decay})\label{bhthm} There exists $\eps>0$
and $M\geq 0$ such that if $|\im\sigma|<\eps$ and $|\re \sigma|>1,$
then for any $\chi\in C_0^\infty(\stackrel{o}{X})$ there exists $C>0$ such that if $n=3$ in
\eqref{model-ndim}, then
\begin{equation}
||\chi R(\sigma) \chi f||_{L^2(X ;\Omega) } \leq C |\sigma|^M||f||_{L^2(X;\Omega)}.
\label{bonyhafnerest}\end{equation}
\end{thm}
This result is not known in higher dimensions (though the methods of
Bony and H\"afner would work even then),
and to prove our main theorem we use the results of Datchev and the third author \cite{Datchev-Vasy} and Wunsch and Zworski \cite{Wunsch-Zworski} to handle the general case. The advantage of the method of \cite{Datchev-Vasy} is that one does not need to obtain a bound for the exact resolvent in the interior and we may work with the approximate model of \cite{Wunsch-Zworski} instead. We decompose the manifold $X$ in two parts
\begin{gather}
\begin{gathered}
X= X_0 \cup X_1, \text{ where } \\
X_0=[r_{\bH}, r_\bH+4\del) \times \ms^n \cup (r_{\sI}-4\del, r_\sI] \times \ms^n \text{ and } \\
X_1=( r_\bH+\del, r_{\sI}-\del) \times \ms^n.
\end{gathered}\label{Xdecomp}
\end{gather}
If $\del$ is small enough and if $\gamma(t)$ is an integral curve of the Hamiltonian of $\Delta_X$ then
(see Section \ref{sec:black-hole} and either \cite{Datchev-Vasy} or \cite{Joshi-SaBarreto})
\begin{gather}
\text{ if } x(\gamma(t))<4\delta \text{ and } \frac{d x(t)}{dt}=0 \Rightarrow \frac{d^2 x(t)}{dt^2}<0.\label{bich-convexity1}
\end{gather}
We consider the operator $\Delta_X$ restricted to $X_1,$ and place it
into the setting of \cite{Wunsch-Zworski} as follows.
Let $X_1'$ be another Riemannian manifold extending $\tilde X_1=( r_\bH+\del/2,
r_{\sI}-\del/2, r_\sI) \times \ms^n$ (and thus $X_1$)
and which is Euclidean outside some compact set, and let
$\Delta_{X'_1}$ be a self-adjoint
operator extending $\Delta_X$ with principal symbol given by the
metric on $X'_1$ which is equal to the
Euclidean Laplacian on the ends. Let
\begin{gather}
P_1=h^2\Delta_{X_1'}-i\Upsilon, \;\ h \in (0,1) \label{operatorp1}
\end{gather}
where $\Upsilon\in \CI(X_1';[0,1])$ is such that $\Upsilon=0$ on $X_1$ and $\Upsilon=1$
outside $\tilde X_1$. Thus, $P_1-1$ is semiclassically elliptic on a
neighborhood of $X_1'\setminus\tilde X_1$. (In particular, this
implies that no bicharacteristic of $P_1-1$ leaves $X_1$ and returns
later, i.e.\ $X_1$ is bicharacteristically convex in $X_1'$,
since this is true for $X_1$ inside $\tilde X_1$, and $P_1-1$ is
elliptic outside $\tilde X_1$, hence has no characteristic set there.)
By Theorem 1 of \cite{Wunsch-Zworski} there exist positive constants $c,$ $C$ and $\eps$ independent of $h$ such that
\begin{gather}
||(P_1-\sigma^2)^{-1}||_{L^2\rightarrow L^2} \leq Ch^{-N}, \;\ \sigma
\in (1-c,1+c) \times (-c,\eps h)\subset\Cx. \label{wuzwestimate}
\end{gather}
Due to the fact that $0<9m^2\Lambda<1,$ the function $\beta(r)=\ha
\frac{d}{dr}\alpha^2(r)$ satisfies $\beta(r_{\bH})>0$ and
$\beta(r_{\sI})<0.$ Set $\beta_{\bH}=\beta(r_{\bH})$ and
$\beta_{\sI}=\beta(r_{\sI}).$ The weight function we consider,
$\tilde\alpha\in\CO([r_{\bH},r_{\sI}]),$ is positive in the interior and
satisfies
\begin{equation}
\tilde{\alpha}(r)= \begin{cases}
\alpha ^{1/\beta _{\bH}}&\Mnear r_{\bH}\\
\alpha ^{1/|\beta _{\sI}|}&\Mnear r_{\sI}
\end{cases}
\label{SeClRe.1}\end{equation}
We will prove
\begin{thm}\label{globalest}
If $n=2,$ let $\eps>0$ be such that \eqref{bonyhafnerest} holds.
In general, assume that $\delta$ is such that \eqref{bich-convexity1} is satisfied, and let $\eps>0$ be such that \eqref{wuzwestimate} holds. If $$0<\gamma<\min(\eps,\beta_{\bH},|\beta_{\sI}|,1),
$$
then for $b>\gamma$ there exist $C$ and $M$ such that if
$\im\sigma\le\gamma$ and $|\re\sigma|\ge1,$
\begin{equation}
||{\tilde{\alpha}}^b R(\sigma){\tilde{\alpha}}^bf||
_{L^2 ( X ;\Omega)}\leq C|\sigma|^M ||f||_{L^2 ( X ;\Omega) },
\label{mainest1}
\end{equation}
where $\Omega$ is defined in \eqref{measure}.
\end{thm}
\noindent
This result can be refined by allowing the power of the weight, on either
side, to approach $\im\sigma$ at the expense of an additional logarithmic term,
see Theorem~\ref{globalest-strong}.
Two proofs of Theorem \ref{globalest} are given below. The first, which is
somewhat simpler but valid only for $n+1=3,$ is given in sections \ref{sec:decomposition} and \ref{sec:black-hole-proof}. It uses techniques of Bruneau
and Petkov \cite{Bruneau-Petkov:Semiclassical} to glue the resolvent
estimates from Theorem~\ref{resolvent-bounds-simple} and the
localized estimate \eqref{bonyhafnerest}.
The second proof, valid in general dimension, uses the estimate \eqref{wuzwestimate} and the semiclassical
resolvent gluing techniques of Datchev and the third author
\cite{Datchev-Vasy}. This is carried out in section
\ref{sec:black-hole-n+1-proof}.
Related weighted $L^2$ estimates for the resolvent on asymptotically
hyperbolic and asymptotically Euclidean spaces have been proved by Cardoso
and Vodev in \cite{Cardoso-Vodev:Uniform}. However, such estimates,
combined with Theorem \ref{bhthm}, only give the holomorphic continuation
of the resolvent, as an operator acting on weighted $L^2$ spaces, for $|\re
\sigma|>1,$ to a region below a curve which converges polynomially to the
real axis. These weaker estimates do suffice to establish the asymptotic
behavior of solutions of the wave equation modulo rapidly decaying terms
(rather than exponentially decaying) and would give a different proof of
the result of Dafermos and Rodnianski \cite{Dafermos-Rodnianski:de-Sitter}.
In the case of a non-trapping asymptotically hyperbolic manifold which has
constant sectional curvature near the boundary, it was shown by Guillarmou
in \cite{Guillarmou:Absence} that there exists a strip about the real axis,
excluding a neighborhood of the origin, which is free of resonances. In
the case studied here, the sectional curvature of the metric associated to
$\Delta_X$ is not constant near the boundary, and there exist trapped
trajectories. However, see \cite{Sa-Barreto-Zworski:Distribution}, all
trapped trajectories of the Hamilton flow of $\Delta_X$ are hyperbolic, and
the projection of the trapped set onto the base space is contained in the
sphere $r=3m,$ which is known as the ergo-sphere. Since the effects of the
trapped trajectories are included in the estimates of Bony and H\"afner,
constructing the analytic continuation of the resolvent of a twisted
Laplacian that has the correct asymptotic behavior at infinity, uniformly
at high energies, allows one to obtain the desired estimates on weighted
Sobolev spaces via pasting techniques introduced by Bruneau and Petkov
\cite{Bruneau-Petkov:Semiclassical}. The main technical result is thus
Theorem~\ref{resolvent-bounds-simple}, and its strengthening,
Theorem~\ref{resolvent-bounds}.
Theorem~\ref{resolvent-bounds} is proved by the construction of
a high-energy parametrix for $\Delta_{g_\delta}+x^2W$. More precisely,
as customary, the problem is translated to the construction of a semiclassical
parametrix for
\begin{equation*}
P(h,\sigma)= h^2\Delta_g+h^2x^2 W-h^2\frac{n^2}{4}-\sigma^2
=h^2\left(\Delta_g+x^2 W-\frac{n^2}{4}-\left(\frac{\sigma}{h}\right)^2\right).
\end{equation*}
where now $\sigma\in (1-c,1+c)\times (-Ch,Ch)\subset\Cx$, $c,C>0$,
and $h\in (0,1)$, $h\to 0$, so the actual spectral parameter is
$$
\frac{n^2}{4}+\left(\frac{\sigma}{h}\right)^2,
$$
and $\im \frac{\sigma}{h}$ is bounded.
Note that for $\im\sigma<0$,
\begin{equation}\label{scresolvent}
R(h,\sigma)=P(h,\sigma)^{-1}:L^2(\Bn)\to H^2_0(\Bn)
\end{equation}
is meromorphic by the results of Mazzeo and the first author
\cite{Mazzeo-Melrose:Meromorphic}.
Moreover, while $\sigma$ is
not real, its imaginary part is $O(h)$ in the semiclassical sense, and is thus
not part of the semiclassical principal symbol of the operator.
The construction proceeds on the semiclassical resolution $M_{0,h}$ of the
product of the double space $\Bn\times_0\Bn$ introduced in
\cite{Mazzeo-Melrose:Meromorphic}, and the interval $[0,1)_h$ -- this space
is described in detail in
Section~\ref{semiclassical-double-space}. Recall that, for fixed $h>0$,
the results of \cite{Mazzeo-Melrose:Meromorphic} show that the Schwartz
kernel of $P(h,\sigma)^{-1}$, defined for $\im\sigma<0$, is well-behaved
(polyhomogeneous conormal) on $\Bn\times_0\Bn$, and it extends
meromorphically across $\im\sigma=0$ with a similarly polyhomogeneous
conormal Schwartz kernel. Thus, the space we are considering is a very
natural one. The semiclassical resolution is already needed away from all
boundaries; it consists of blowing up the diagonal at $h=0$. Note that
$P(h,\sigma)$ is a semiclassical differential operator which is elliptic
in the usual sense, but its semiclassical principal symbol
$g-(\re\sigma)^2$ is not elliptic (here $g$ is the dual metric
function). Ignoring the boundaries for a moment, when a semiclassical
differential operator is elliptic in both senses, it has a parametrix in
the small semiclassical calculus, i.e.\ one which vanishes to infinite
order at $h=0$ off the semiclassical front face (i.e.\ away from the
diagonal in $\Bn\times_0\Bn\times\{0\}$). However, as $P(h,\sigma)$ is
not elliptic semiclassically, semiclassical singularities (lack of decay
as $h\to 0$) flow out of the semiclassical front face.
It is useful to consider the flow in terms of Lagrangian geometry. Thus,
the small calculus of order $-\infty$ semiclassical pseudodifferential
operators consists of operators whose Schwartz kernels are
semiclassical-conormal to the diagonal at $h=0$. As $P(h,\sigma)$ is not
semiclassically elliptic (but is elliptic in the usual sense, so it behaves
as a semiclassical pseudodifferential operator of order $-\infty$ for our
purposes), in order to construct a parametrix for $P(h,\sigma)$, we need to
follow the flow out of semiclassical singularities from the conormal bundle
of the diagonal. For $P(h,\sigma)$ as above, the resulting Lagrangian
manifold is induced by the geodesic flow, and is in particular, up to a
constant factor, the graph of the differential of the distance function on
the product space. Thus, it is necessary to analyze the geodesic flow and
the distance function; here the presence of boundaries is the main issue.
As we show in Section~\ref{distance-function}, the geodesic flow is
well-behaved on $\Bn\times_0\Bn$ as a Lagrangian manifold of the
appropriate cotangent bundle. Further, for $\delta>0$ small (this is where
the smallness of the metric perturbation enters), its projection to the
base is a diffeomorphism, which implies that the distance function is also
well-behaved. This last step is based upon the precise description of the
geodesic flow and the distance function on hyperbolic space, see
Section~\ref{distance-function}.
In Section~\ref{full-sc-parametrix}, we then construct the parametrix by
first solving away the diagonal singularity; this is the usual elliptic
parametrix construction. Next, we solve away the small calculus error in
Taylor series at the semiclassical front face, and then propagate the
solution along the flow-out by solving transport equations. This is an
analogue of the intersecting Lagrangian construction of the first author
and Uhlmann \cite{Melrose-Uhlmann:Intersection}, see also the work of
Hassell and Wunsch \cite{Hassell-Wunsch:Semiclassical} in the semiclassical
setting. So far in this discussion the boundaries of $M_{0,\semi}$ arising
from the boundaries of $\Bn\times_0\Bn$ have been ignored; these enter into
the steps so far only in that it is necessary to ensure that the
construction is uniform (in a strong, smooth, sense) up to these
boundaries, which the semiclassical front face as well as the lift of
$\{h=0\}$ meet transversally, and only in the zero front face, i.e.\ at the
front face of the blow-up of $\Bn\times\Bn$ that created $\Bn\times_0\Bn$.
Next we need to analyze the asymptotics of the solutions of the transport
equations at the left and right boundary faces of $\Bn\times_0\Bn$; this is
facilitated by our analysis of the flowout Lagrangian (up to these boundary
faces). At this point we obtain a parametrix whose error is smoothing and
is $O(h^\infty)$, but does not, as yet, have any decay at the zero front
face. The last step, which is completely analogous to the construction of
Mazzeo and the first author, removes this error.
As a warm-up to this analysis, in Section~\ref{3D-parametrix} we present a
three dimensional version of this construction, with worse, but still
sufficiently well-behaved error terms. This is made possible by a
coincidence, namely that in $\RR^3$ the Schwartz kernel of the resolvent of
the Laplacian at energy $(\lambda-i0)^2$ is a constant multiple of
$e^{-i\lambda r} r^{-1}$, and $r^{-1}$ is a homogeneous function on
$\RR^3$, which enables one to blow-down the semiclassical front face at
least to leading order. Thus, the first steps of the construction are
simplified, though the really interesting parts, concerning the asymptotic
behavior at the left and right boundaries along the Lagrangian, are
unchanged. We encourage the reader to read this section first as it is more
explicit and accessible than the treatment of arbitrary dimensions.
In Section~\ref{sec:L2-bounds}, we obtain weighted $L^2$-bounds for the
parametrix and its error. In Section~\ref{sec:resolvent-bounds} we used these
to prove
Theorem~\ref{resolvent-bounds-simple} and
Theorem~\ref{resolvent-bounds}.
In Section~\ref{sec:black-hole} we describe in detail the
de Sitter-Schwarzschild set-up. Then in
Section~\ref{sec:decomposition}, in dimension $3+1$,
we describe the approach of Bruneau and Petkov \cite{Bruneau-Petkov:Semiclassical}
reducing the
necessary problem to the combination of analysis on the ends,
i.e.\ Theorem~\ref{resolvent-bounds}, and of the cutoff resolvent,
i.e.\ Theorem~\ref{bhthm}. Then, in Section~\ref{sec:black-hole-proof},
we use this method to prove Theorem~\ref{globalest}, and its strengthening,
Theorem~\ref{globalest-strong}. Finally, in
Section~\ref{sec:black-hole-n+1-proof} we give a different proof which
works in general dimension, and does not require knowledge of
estimates for the exact cutoff resolvent. Instead, it uses the results of Wunsch and Zworski
\cite{Wunsch-Zworski} for normally hyperbolic trapping in the presence
of Euclidean ends and of Datchev
and the third author \cite{Datchev-Vasy} which provide a method to combine these with our
estimates on hyperbolic ends. Since this method is described in detail
in \cite{Datchev-Vasy}, we keep this section fairly brief.
\paperbody
\section{Resolvent estimates for model operators}
In this section we state the full version of the main technical result,
Theorem~\ref{resolvent-bounds-simple}.
Let $g_0$ be the metric on $\Bn$ given by
\begin{gather}
g_{0}=\frac{4 dz^2}{(1-|z|^2)^2}. \label{metg0}
\end{gather}
We consider a one-parameter family of perturbations of $g_{0}$ supported in
a neighborhood of $\p \Bn$ of the form
\begin{gather}
g_\delta= g_{0} + \chi_\delta(z) H(z,dz), \label{metgeps}
\end{gather}
where $H$ is a symmetric 2-tensor, which is $\CI$ up to $\p \Bn,$
$\chi\in \CI(\mr),$ with $\chi(s)=1$ if $|s|<\ha,$ $\chi(s)=0$ if
$|s|>1,$ and $\chi_\delta(z)=\chi((1-|z|)\delta^{-1}).$
Let $x=\frac{1-|z|}{1+|z|},$ $W\in\CI(\Bn)$ and let $R_\delta (\sigma)=(\Delta_{g_\delta}+x^2W-\sigma^2-1^2)^{-1}$ denote the resolvent of $\Delta_{g_\delta}+x^2W.$
The spectral theorem gives that $R_\delta(\sigma)$ is well defined as a bounded operator in
$L^2(\Bn)=L^2(\Bn;dg)$ if $\im \sigma <<0.$
The results of \cite{Mazzeo-Melrose:Meromorphic}
show that $R_{\delta}(\sigma)$ continues meromorphically to $\mc\setminus i\mn/2$ as an operator mapping $\CI$ functions vanishing to infinite order at $\p \Bn$ to distributions in
$\Bn.$
We recall, see for example \cite{Mazzeo:Edge}, that for $k\in \mn,$
\begin{gather*}
H^k_0(\Bn)=\{ u \in L^2(\Bn): (x\p_x, \p_\omega)^m u \in L^2(\Bn), \;\ m \leq k\},
\end{gather*}
and
\begin{gather*}
H_0^{-k}=\{ v\in \mathcal{D}'(\Bn): \text{ there exists } u_{\beta} \in L^2(\Bn),
v= \sum_{|\beta|\leq k} (x\p_x, \p_\omega)^\beta u_\beta\}.
\end{gather*}
Our main result is the following theorem:
\begin{thm}\label{resolvent-bounds} There exist $\delta_0>0,$ such that if $0\leq \delta\leq \delta_0,$ then
$x^a R_{\delta}(\sigma) x^b$ continues holomorphically to $\im \sigma < M,$ $M>0,$ provided
$| \sigma| >K(\delta,M),$ $b> \im\sigma$ and $a>\im \sigma.$ Moreover,
there exists $C>0$ such that
\begin{gather}
\begin{gathered}
|| x^a R_\delta (\sigma) x^b v||_{H^k_0(\Bn)} \leq C |\sigma|^{-1+\frac{n}{2}+k} ||v||_{L^2(\Bn)}, \;\ k=0,1,2, \\
|| x^a R_\delta (\sigma) x^b v||_{L^2(\Bn)} \leq C |\sigma|^{-1+\frac{n}{2}+k} ||v||_{H_0^{-k}(\Bn)}, \;\ k=0,1,2, \\
\end{gathered} \label{sobolev1}
\end{gather}
If $a=\im \sigma,$ or $b=\im\sigma,$ or $a=b=\im\sigma,$ let $\phi_N(x)\in\CI((0,1)),$ $\phi_N\geq 1,$ $\phi_N(x)=|\log x|^{-N},$ if $x<\oq$ $\phi_N(x)=1$ if $x>\ha.$ Then in each case the operator
\begin{gather}
\begin{gathered}
T_{a,b,N}(\sigma)=:x^{\im\sigma} \phi_N(x) R_{\delta}(\sigma) x^b, \text{ if } b>\im \sigma, \\
T_{a,b,N=:}x^{a} R_{\delta}(\sigma) x^{\im \la} \phi_N(x), \text{ if } a> \im \sigma, \\
T_{a.b,N}=:x^{\im\sigma} \phi_N(x) R_{\delta}(\sigma) x^{\im \sigma} \phi_N(x),
\end{gathered}\label{deftabn}
\end{gather}
continues holomorphically to $\im \sigma < M,$ provided $N>\ha,$ $| \sigma| >K(\delta,M).$ Moreover
in each case there exists $C=C(M,N,\delta)$ such that
\begin{gather}
\begin{gathered}
|| T_{a,b,N}(\sigma) v||_{H^k_0(\Bn)} \leq C |\sigma|^{-1+\frac{n}{2}+k} ||v||_{L^2(\Bn)}, \;\ k=0,1,2, \\
|| T_{a,b,N}(\sigma) v||_{L^2(\Bn)} \leq C |\sigma|^{-1+\frac{n}{2}+k} ||v||_{H_0^{-k}(\Bn)}, \;\ k=0,1,2.
\end{gathered}\label{sobolev2}
\end{gather}
\end{thm}
\section{The distance Function}\label{distance-function}
In the construction of the uniform parametrix for the resolvent we will
make use of an appropriate resolution of the distance function, and
geodesic flow, for the metric $g_\delta.$ This in turn is obtained by
perturbation from $\delta =0,$ so we start with an analysis of the
hyperbolic distance, for which there is an explicit formula. Namely, the
distance function for the hyperbolic metric, $g_{0},$ is given in terms of
the Euclidean metric on the ball by
\begin{equation}
\begin{gathered}
\dist_{0}:\Bno \times \Bno \longrightarrow \mr\Mwhere\\
\cosh(\dist_{0}(z,z'))=1+\frac{2|z-z'|^2}{(1-|z|^2)(1-|z'|^2)}.
\end{gathered}
\label{distf}\end{equation}
We are particularly interested in a uniform description as one or both of the
points approach the boundary, i.e.\ infinity. The boundary behavior is
resolved by lifting to the `zero stretched product' as is implicit in
\cite{Mazzeo:Edge}. This stretched product, $\Bnc\times_0 \Bnc,$ is the
compact manifold with corners defined by blowing up the intersection of the
diagonal and the corner of $\Bnc\times\Bnc:$
\begin{equation}
\begin{gathered}
\beta:\Bnc\times_0\Bnc=[\Bnc\times\Bnc;\pa\Diag]\longrightarrow\Bnc\times\Bnc,
\\
\pa\Diag=\Diag\cap \; (\p\Bnc \times \p \Bnc)=\{(z,z);|z|=1\},\\
\Diag=\{(z,z')\in \Bnc\times\Bnc;z=z'\}.
\end{gathered}\label{zero-blow-up}
\end{equation}
See Figure~\ref{fig1} in which $(x,y)$ and $(x',y')$ are
local coordinates near a point in the center of the blow up, with boundary
defining functions $x$ and $x'$ in the two factors.
\begin{figure}[int1
\epsfxsize= 3.5in
\centerline{\epsffile{blow-up}}
\caption{The stretched product $\Bnc\times_0\Bnc.$}
\label{fig1}
\end{figure}
Thus $\Bnc\times_0\Bnc$ has three boundary hypersurfaces, the front face
introduced by the blow up and the left and right boundary faces which map
back to $\pa\Bnc\times\Bnc$ and $\Bnc\times\pa\Bnc$ respectively under
$\beta.$ Denote by $\Diag_0$ the lift of the diagonal, which in this case
is the closure of the inverse image under $\beta$ of the interior of the
diagonal in $\Bnc\times\Bnc.$
\begin{lemma}\label{dist0} Lifted to the interior of $\Bnc\times_0\Bnc$ the
hyperbolic distance function extends smoothly up to the interior of the
front face, in the complement of $\Diag_0,$ where it is
positive and, for an appropriate choice $\rho _L\in\CI(\Bnc\times_0\Bnc)$
of defining function for the left boundary and
with $\rho _R$ its reflection,
\begin{equation}
\begin{gathered}
\beta^*\dist_{0}(z,z')=
-\log(\rho_L\rho_R)+ F,\\
0<F\in\CI(\Bnc \times_0 \Bnc\setminus \Diag_0 ),\
F^2\in\CI(\Bnc \times_0 \Bnc),
\end{gathered}
\label{blowd}\end{equation}
with $F ^2$ a quadratic defining function for $\Diag_0.$
\end{lemma}
\begin{proof} We show first that the square of the Euclidean distance
function, $|z-z'|^2,$ lifts to be smooth on $\Bnc \times_0 \Bnc$ and to
vanish quadratically on $\Diag_0$ and on the front face produced by the
blow up
\begin{equation}
\beta ^*(|z-z'|^2)=R^2f,\ f\in\CI(\Bnc \times_0 \Bnc).
\label{SeClRe.12}\end{equation}
Here $f\ge0$ vanishes precisely at $\Diag_0$ and does so
quadratically. Indeed this is certainly true away from the front face
produced by the blow up. The spaces and the distance function are invariant
under rotational symmetry, which lifts under the blow up, so me may fix the
spherical position of one variable and suppose that $z'=(1-x',0),$ with
$x'>0$ and small, the blow up is then of $z=z',$ $x'=0.$ The fact that the
variables are restricted to the unit ball is now irrelevant, and using the
translation-invariance of the Euclidean distance we can suppose instead
that $z'=(x',0)$ and blow up $z=z',$ $x'=0.$ Since $|(x,0)-z|^2$ is
homogeneous in all variables it lifts to be the product of the square of a
defining function for the front face and a quadratic defining function for
the lift of the diagonal. This proves \eqref{SeClRe.12} after restriction
to the preimage of the balls and application of the symmetry.
The hyperbolic distance is given by \eqref{distf} where $1-|z|^2$ and
$1-|z'|^2$ are boundary defining functions on the two factors. If $R$ is
a defining function for the front face of $\Bnc\times_0\Bnc$ then these lift
to be of the form $\rho _LR$ and $\rho _RR$ so combining this with
\eqref{SeClRe.12}
\begin{equation}
\beta ^*\cosh(\dist_{0}(z,z'))=1+2\frac{f}{\rho _L\rho _R}.
\label{SeClRe.8}\end{equation}
Now $\exp(t)= \cosh t+ \left[ \cosh^2t-1\right]^\ha,$ for $t>0,$ and from
\eqref{SeClRe.8} it follows that
\begin{equation}
\exp (\dist_{0}(z,z'))=1+2\frac{f}{\rho _L\rho _R}+\left(2\frac{f}{\rho
_L\rho _R}+4\frac{f^2}{\rho^2 _L\rho^2 _R}\right)^{\frac12}.
\label{SeClRe.9}\end{equation}
Near $\Diag_0$ the square-root is dominated by the first part and near the
left and right boundaries by the second part, and is otherwise positive and
smooth. Taking logarithms gives the result as claimed, with the
defining function taken to be one near $\Diag_0$ and to be everywhere smaller
than a small positive multiple of $(1-|z|^2)/R.$
\end{proof}
This result will be extended to the case of a perturbation of the
hyperbolic metric by constructing the distance function directly from
Hamilton-Jacobi theory, i.e.\ by integration of the Hamilton vector field
of the metric function on the cotangent bundle. The presence of only simple
logarithmic singularities in \eqref{blowd} shows, perhaps somewhat
counter-intuitively, that the Lagrangian submanifold which is the graph of
the differential of the distance should be smooth (away from the diagonal)
in the b-cotangent bundle of $M^2_0.$ Conversely if this is shown for
the perturbed metric then the analogue of \eqref{blowd} follows except for
the possibility of a logarithmic term at the front face.
Since the metric is singular near the boundary, the dual metric function on
$T^*\Bnc$ is degenerate there. In terms of local coordinates near a
boundary point, $x,$ $y$ where the boundary is locally $x=0,$ and dual
variables $\xi,$ $\eta,$ the metric function for hyperbolic space is of the
form
\begin{equation}
2p_0=x^2\xi^2 + 4 x^2(1-x^2)^{-2} h_0(\omega,\eta)
\label{SeClRe.15}\end{equation}
where $h_0$ is the metric function for the induced metric on the boundary.
Recall that the $0$-cotangent bundle of a manifold with boundary $M,$
denoted $\To M,$ is a smooth vector bundle over $M$ which is a rescaled
version of the ordinary cotangent bundle. In local coordinates near, but
not at, the boundary these two bundles are identified by the (rescaling) map
\begin{equation}
T^* M\ni(x,y, \xi,\eta) \longmapsto (x,y,\la,\mu)=(x,y, x\xi,x \eta)\in \To M.
\label{SeClRe.40}\end{equation}
It is precisely this rescaling which makes the hyperbolic metric into a
non-degenerate fiber metric, uniformly up to the boundary, on this
bundle. On the other hand the b-cotangent bundle, also a completely natural
vector bundle, is obtained by rescaling only in the normal variable
\begin{equation}
T^* M\ni(x,y, \xi,\eta)\longmapsto (x,y,\la,\eta)=(x,y, x\xi,\eta)\in\Tb^*M.
\label{SeClRe.20}\end{equation}
Identification over the interior gives natural smooth vector bundle maps
\begin{equation}
\iota_{\text{b}0}:\oT M\longrightarrow \Tb M,\
\iota^t_{\text{b}0}:\Tb^* M\longrightarrow \To M.
\label{SeClRe.27}\end{equation}
The second scaling map can be constructed directly in terms of blow up.
\begin{lemma}\label{SeClRe.26} If $\To\pa M\subset\Tos{\pa M}M$ denotes
the annihilator of the null space, over the boundary, of $\iota
_{\text{b}0}$ in \eqref{SeClRe.27} then there is a canonical diffeomorphism
\begin{equation}
\Tb^*M \longrightarrow [\To M,\To\pa M]\setminus
\beta^{\#}\left(\Tos{\pa M}M\right),\
\beta:[\To M,\To\pa M]\longrightarrow \To M,
\label{SeClRe.28}\end{equation}
to the complement, in the blow up, of the lift of the boundary:
\begin{equation*}
\beta ^{\#}(\Tos{\pa M}M)=\overline{\beta^{-1}(\Tos{\pa M}M\setminus\To\pa M)}.
\label{SeClRe.29}\end{equation*}
\end{lemma}
\begin{proof} In local coordinates $x,$ $y_j,$ the null space of $\iota
_{\text{b}0}$ in \eqref{SeClRe.27} is precisely the span of the
`tangential' basis elements
$x\pa_{y_j}$ over each boundary point. Its annihilator, $\To\pa M$ is
given in the coordinates \eqref{SeClRe.20} by $\mu=0$ at $x=0.$
The lift of the `old boundary' $x=0$ is precisely the boundary
hypersurface near which $|\mu|$ dominates $x.$ Thus, $x$ is a valid
defining function for the boundary of the complement, on the right in
\eqref{SeClRe.28} and locally in this set, above the coordinate patch in $M,$
$\eta_j=\mu_j/x$ are smooth functions. The natural bundle map
$\Tb^*M\longrightarrow \To M$ underlying \eqref{SeClRe.28} is given in
these coordinates by $\lambda \frac{dx}x+\eta\cdot dy\longmapsto
\lambda \frac{dx}x+x\eta\cdot\frac{dy}x,$ which is precisely the same
map, $\mu=x\eta,$ as appears in \eqref{SeClRe.20}, so the result,
including naturality, follows.
\end{proof}
The symplectic form lifted to $\To M$ is
\begin{gather*}
\bo \omega=\frac{1}{x} d\la \wedge dx +
\frac{1}{x} d\mu \wedge dy - \frac{1}{x^2} dx \wedge (\mu\cdot dy)
\end{gather*}
whereas lifted to $\Tb^*M$ it is
\begin{equation}
{}^{\text{b}}\omega=\frac{1}{x} d\la \wedge dx+ d\eta \wedge dy .
\label{SeClRe.21}\end{equation}
Working, for simplicity of computation, in the non-compact upper half-space
model for hyperbolic space the metric function lifts to the non-degenerate
quadratic form on $\To M:$
\begin{equation}
2p_0=\la^2+h_0(\omega,\mu)
\label{SeClRe.16}\end{equation}
where $h_0=|\mu|^2$ is actually the Euclidean metric. The $0$-Hamilton
vector field of $p_0\in\CI(\To M),$ just the lift of the Hamilton
vector field over the interior, is determined by
\begin{equation}
\bo\omega(\cdot, \boH_p)=dp.
\label{SeClRe.44}\end{equation}
Thus
\begin{gather*}
\boH_p= x\frac{\p p}{\p \la} \p_x + x \frac{\p p}{\p \mu} \cdot \p_y -
\left( \mu\cdot \frac{\p p}{\p \mu} +
x \frac{\p p}{\p x}\right) \p_{\la} -
\left( -\frac{\p p}{\p \la}\mu + x \frac{\p p}{\p y}\right)\cdot \p_\mu
\end{gather*}
and hence
\begin{equation}
\boH_{p_0}=\la( x\p_x+\mu\p_\mu)-h_0\p_\la + \frac x2\sH_{h_0}
\label{SeClRe.17}\end{equation}
is tangent to the smooth (up to the boundary) compact sphere bundle given
by $p_0=1.$
Over the interior of $M=\bbB^n,$ the hyperbolic distance from any interior
point of the ball is determined by the graph of its differential, which is
the flow out inside $p_0=1,$ of the intersection of this smooth compact
manifold with boundary with the cotangent fiber to the initial
point. Observe that $\sH_{p_0}$ is also tangent to the surface $\mu=0$ over
the boundary, which is the invariantly defined subbundle $\To\pa M.$ Since the
coordinates can be chosen to be radial for any interior point, it follows
that all the geodesics from the interior arrive at the boundary at $x=0,$
$\mu=0,$ corresponding to the well-known fact that hyperbolic geodesics are
normal to the boundary. This tangency implies that $\sH_{p_0}$ lifts under
the blow up of $\To\pa M$ in \eqref{SeClRe.28} to a
smooth vector field on $\Tb M;$ this can also be seen by direct computation.
\begin{lemma}\label{SeClRe.19} The graph of the differential of the
distance function from any interior point, $p\in\Bno,$ of hyperbolic
space extends by continuity to a smooth Lagrangian submanifold of
$\Tb^*(\Bnc\setminus\{p\})$ which is transversal to the boundary, is a
graph over $\Bnc\setminus\{p\}$ and is given by integration of a
non-vanishing vector field up to the boundary.
\end{lemma}
\begin{proof} Observe the effect of blowing up $\mu=0,$ $x=0$ on the
Hamilton vector field in \eqref{SeClRe.17}. As noted above, near the
front face produced by this blow up valid coordinates are given by
$\eta=\mu/x,$ $\la$ and $y,$ with $x$ the boundary defining
function. Since this transforms $\bo\omega$ to ${}^{\text{b}}\omega$ it
follows that $\boH_{p_0}$ is transformed to
\begin{equation}
{}^{\text{b}}\sH_{p_0}=x(\la \p_x-xh_0\p_{\la}+x\sH_{h_0})
\label{SeClRe.22}\end{equation}
where now $\sH_{h_0}$ is the Hamilton vector field with respect to $y$ and
$\eta.$
The constant energy surface $p_0=1$ remains smooth, but non-compact, near
the boundary, which it intersects transversally in $\la=1.$ From this
the result follows -- with the non-vanishing smooth vector field being
${}^{\text{b}}\sH_{p_0}$ divided by the boundary defining function $x$ near
the boundary.
\end{proof}
The logarithmic behaviour of the distance function in \eqref{blowd} with
one point fixed in the interior is a consequence of Lemma~\ref{SeClRe.26},
since the differential of the distance must be of the form $adx/x+b\cdot
dy$ for smooth functions $a$ and $b,$ and since it is closed, $a$ is
necessarily constant on the boundary.
To examine the distance as a function of both variables a similar
construction for the product, in this case $M^2,$ $M=\Bnc$ can be used. The
graph, $\Lambda,$ of the differential of the distance $d(p,p')$ as a
function on $M^2$ is the joint flow out of the conormal sphere bundle to
the diagonal in $T^*M^2=T^*M\times T^*M,$ under the two Hamiltonian vector
fields of the two metric functions within the product sphere bundles. As
before it is natural to lift to $\To M\times\To M$ where the two sphere
bundles extend smoothly up to the boundary. However, one can make a
stronger statement, namely that the lifted Hamilton vector fields are
smooth on the b-cotangent bundle of $M^2_0,$ and indeed even on the
`partially b'-cotangent bundle of $M^2_0$, with `partially' meaning it is
the standard cotangent bundle over the interior of the front face. This is
defined and discussed in more detail below; note that the identification of
these bundles over the interior of $M^2_0$ extends to a smooth map from
these bundles to the lift of $\To M\times\To M$, as we show later,
explaining the `stronger' claim.
Smoothness of the Hamilton vector field together with transversality
conditions shows that the flow-out of the conormal bundle of the diagonal
is a smooth Lagrangian submanifold of the cotangent bundle under
consideration; closeness to a particular Lagrangian (such as that for
hyperbolic space) restricted to which projection to $M^2_0$ is a
diffeomorphism, guarantees that this Lagrangian is also a graph over
$M^2_0.$ Thus, over the interior, $(M^2_0)^\circ$, it is the graph of the
differential of the distance function, and the latter is smooth; the same
would hold globally if the Lagrangian were smooth on $T^*M^2_0.$ The latter
cannot happen, though the Lagrangian will be a graph in the b-, and indeed
the partial b-, cotangent bundles over $M^2_0.$ These give regularity of
the distance function, namely smoothness up to the front face (directly for
the partial b bundle, with a short argument if using the b bundle), and the
logarithmic behavior up to the other faces. Note that had we only showed
the graph statement in the pullback of $\To M\times\To M,$ one would obtain
directly only a weaker regularity statement for the distance function;
roughly speaking, the closer the bundle in which the Lagrangian is
described is to the standard cotangent bundle, the more regularity the
distance function has.
In fact it is possible to pass from the dual of the lifted product
0-tangent bundle to the dual of the b-tangent bundle, or indeed the partial
b-bundle, by blow-up, as for the single space above. Observe first that the
natural inclusion
\begin{equation}
\iota_{0\text{b}}\times \iota_{0\text{b}}:\oT M\times\oT M\longrightarrow
\Tb M^2=\Tb M\times\Tb M
\label{SeClRe.31}\end{equation}
identifies the sections of the bundle on the left with those sections of the
bundle on the right, the tangent vector fields on $M^2,$ which are also
tangent to the two fibrations, one for each boundary hypersurface
\begin{equation}
\phi_L:\pa M\times M\longrightarrow M,\ \phi_R:M\times\pa M\longrightarrow M.
\label{SeClRe.33}\end{equation}
\begin{lemma}\label{SeClRe.32} The fibrations \eqref{SeClRe.33}, restricted
to the interiors, extend by continuity to fibrations $\phi_L$, resp.\ $\phi_R$,
of the two `old'
boundary hypersurfaces of $M^2_0$ and the smooth sections of the lift of
$\oT M\times\oT M$ to $M^2_0$ are naturally identified with the subspace of the
smooth sections of $\Tb M^2_0$ which are tangent to these fibrations and
also to the fibres of the front face of the blow up,
$\beta_0:\ff(M^2_0)\longrightarrow \pa M\times\pa M.$
\end{lemma}
\begin{proof} It is only necessary to examine the geometry and vector
fields near the front face produced by the the blow up of the diagonal
near the boundary. Using the symmetry between the two factors, it
suffices to consider two types of coordinate systems. The first is valid
in the interior of the front fact and up to a general point in the interior
of the intersection with one of the old boundary faces. The second is
valid near a general point of the corner of the front face, which has
fibers which are quarter spheres.
For the first case let $x,$ $y$ and $x',$ $y'$ be the same local
coordinates in two factors. The coordinates
\begin{equation}
s=x/x',\ x',\ y\Mand Y=(y'-y)/x'
\label{SeClRe.35}\end{equation}
are valid locally in $M^2_0$ above the point $x=x'=0,$ $y=y',$ up to the
lift of the old boundary $x=0,$ which becomes locally $s=0.$ The fibration
of this hypersurface is given by the constancy of $y$
and the front face is $x'=0$ with fibration also given by the constancy of
$y.$ The vector fields
$$
x\pa_x,\ x\pa_y,\ x'\pa_{x'}\Mand x'\pa_{y'}
$$
lift to
$$
s\pa_s,\ sx'\pa_y-s\pa_Y,\ x'\pa_{x'}-s\pa_s-Y\cdot\pa_Y\Mand\pa_Y.
$$
The
basis $s\pa_s,$ $sx'\pa_y,$ $x'\pa_{x'}$ and $\pa_Y$ shows that these
vector fields are locally precisely the tangent vector fields also tangent
to both fibrations.
After relabeling the tangential variables as necessary, and possibly
switching their signs, so that $y'_1-y_1>0$
is a dominant variable, the coordinate system
\begin{equation}
t=y_1'-y_1,\ s_1=\frac{x}{y'_1-y_1},\
s_2=\frac{x'}{y'_1-y_1},\ Z_j=\frac{y'_j-y_j}{y'_1-y_1},\ j>1,\ y
\label{22.9.2010.1}\end{equation}
can be used at a point in the corner of the front face. The three boundary
hypersurfaces are locally $s_1=0,$ $s_2=0$ and $t=0$ and their respective
fibrations are given in these coordinates by
\begin{equation}
\begin{gathered}
s_1=0,\ y=\text{const.,}\\
s_2=0,\ y'_1=y_1+t=\text{const.},\ y'_j=y_j+tZ_j=\text{const.},\ j>1,\\
t=0,\ y=\text{const.}
\end{gathered}
\label{22.9.2010.2}\end{equation}
Thus, the intersections of fibres of the lifted left or right faces with
the front face are precisely boundary hypersurfaces of fibres there. On the
other hand within the intersection of the lifted left and right faces the
respective fibres are transversal except at the boundary representing the
front face. The lifts of the basis of the zero vector fields is easily
computed:
\begin{equation}
\begin{gathered}
x\pa_x\longmapsto s_1\pa_{s_1},\\
x\pa_{y_1}\longrightarrow
-s_1t\pa_t+s_1^2\pa_{s_1}+s_1s_2\pa_{s_2}+s_1Z\cdot\pa_Z+s_1t\pa_{y_1},\\
x\pa_{y_j}\longmapsto s_1t\pa_{y_j}-s_1\pa_{Z_j},\\
x'\pa_{x'}\longmapsto s_2\pa_{s_2},\\
x'\pa_{y'_1}\longrightarrow s_2t\pa_t-s_2s_1\pa_{s_1}-s_2^2\pa_{s_2}-s_2Z\cdot\pa_{Z},\\
x'\pa_{y'_j}\longmapsto s_2\pa_{Z_j}.
\end{gathered}
\label{22.9.2010.6}\end{equation}
The span, over $\CI(M^2_0),$ of these vector fields is also spanned by
$$
s_1\pa_{s_1},\ s_2\pa_{s_2},\ s_2\pa_{Z_j},\ j>1,\ s_1(t\pa_{y_j}-\pa_{Z_j}),\ j>1,
\ s_1(t\pa_t-t\pa_{y_1}-Z\cdot\pa_Z)\Mand s_2t\pa_t.
$$
These can be seen to locally
span the vector fields tangent to all three boundaries and corresponding
fibrations, proving the Lemma.
\end{proof}
With $\phi=\{\phi_L,\phi_R,\beta_0\}$
the collection of boundary fibrations, we denote by $\Tph M^2_0$ the
bundle whose smooth sections are exactly the smooth vector fields
tangent to all boundary fibrations. Thus, the content of the preceding lemma
is that
$$
\beta^*(\oT M\times \oT M)=\Tph M^2_0.
$$
These fibrations allow the reconstruction of $\Tb^*M^2_0$ as a blow up of
the lift of $\To M\times\To M$ to $M^2_0.$ It is also useful, for more
precise results later on, to consider the `partially b-' cotangent bundle
of $M^2_0$, $\Tbff^*M^2_0$; this is the dual space of the partially
b-tangent bundle, $\Tbff M^2_0,$ whose smooth sections are smooth vector
fields on $M^2_0$ which are tangent to the old boundaries, but not
necessarily to the front face, $\ff$. Thus, in coordinates
\eqref{SeClRe.35}, $s\pa_s$, $\pa_{x'}$, $\pa_y$ and $\pa_Y$ form a basis
of $\Tbff M^2_0$, while in coordinates \eqref{22.9.2010.1}, $s_1\pa_{s_1}$,
$s_2\pa_{s_2}$, $\pa_t$, $\pa_{Z_j}$ and $\pa_y$ do so. Let
$$
\iota^2_{0\bl}:\Tph M^2_0\to\Tb M^2_0,\ \iota^2_{0\bl,\ff}:\Tph M^2_0\to\Tbff M^2_0
$$
be the inclusion maps.
\begin{lemma}\label{SeClRe.34} The annihilators, in the lift of $\To M\times\To
M$ to $M^2_0,$ of the null space of either $\iota^2_{0\bl}$ or
$\iota^2_{0\bl,\ff}$ over the old boundaries, as in
Lemma~\ref{SeClRe.32}, form transversal embedded p-submanifolds. After
these are blown up, the closure of the annihilator of the nullspace of
$\iota^2_{0\bl}$, resp.\ $\iota^2_{0\bl,\ff}$, over
the interior of the front face of $M^2_0$ is a p-submanifold, the
subsequent blow up of which produces a manifold with corners with three
`old' boundary hypersurfaces; the complement of these three hypersurfaces
is canonically diffeomorphic to $\Tb^*M^2_0,$ resp.\ $\Tbff^* M^2_0$.
\end{lemma}
\begin{proof} By virtue of Lemma~\ref{SeClRe.26} and the product structure away
from the front face $\ff$ of $M^2_0$,
the statements here are trivially valid except possibly
near $\ff$. We may again use the coordinate systems
discussed in the proof of Lemma~\ref{SeClRe.32}. Consider the linear
variables in the fibres in which a general point is $l x\pa_x+v\cdot
x\pa_y+l'x'\pa_{x'}+v'\cdot x'\pa_{y'}.$
First consider the inclusion into $\Tb M^2_0$. In the interiors of $s_1=0$ and
$s_2=0$ and the front face respectively, the null bundles of the
inclusion into the tangent vector fields are
\begin{equation}
\begin{gathered}
l=l'=0,\ v'=0,\\
l=l'=0,\ v=0,\\
\begin{aligned}
l s\pa_s+v(sx'\pa_y-s\pa_Y)+l'(x'\pa_{x'}-s\pa_s-Y\pa_Y)+v'\pa_Y=0\Mat x'=0,\ s>0\\
\Longleftrightarrow
l=l'=0,\ v'=sv.
\end{aligned}
\end{gathered}
\label{22.9.2010.8}\end{equation}
The corresponding annihilator bundles, over the interiors of the boundary
hypersurfaces of $M^2_0,$ in the dual bundle, with basis
\begin{equation}
\la \frac{dx}x+\mu \frac{dy}x+\la '\frac{dx'}{x'}+\mu'\frac{dy'}{x'}
\label{22.9.2010.9}\end{equation}
are therefore, as submanifolds,
\begin{equation}
\begin{gathered}
s_1=0,\ \mu =0,\\
s_2=0,\ \mu '=0\text{ and}\\
x'=0,\ \mu+s\mu'=0\Mor t=0,\ s_2\mu +s_1\mu'=0.
\end{gathered}
\label{22.9.2010.10}\end{equation}
Here the annihilator bundle over the front face is given with respect to both
the coordinate system \eqref{SeClRe.35} and \eqref{22.9.2010.1}.
Thus, the two subbundles over the old boundary hypersurfaces meet
transversally over the intersection, up to the corner, as claimed and so
can be blown up in either order. In the complement of the lifts of the old
boundaries under these two blow ups, the variables $\mu/s_1$ and
$\mu'/s_2$ become legitimate; in terms of these the subbundle over the
front face becomes smooth up to, and with a product decomposition at, all
its boundaries. Thus, it too can be blown up. That the result is a
(painful) reconstruction of the b-cotangent bundle of the blown up manifold
$M^2_0$ follows directly from the construction.
It remains to consider the inclusion into $\Tbff M^2_0$. The only changes are
at the front face, namely the third line of \eqref{22.9.2010.8} becomes
\begin{equation}\begin{gathered}
l s\pa_s+v(sx'\pa_y-s\pa_Y)+l'(x'\pa_{x'}-s\pa_s-Y\pa_Y)+v'\pa_Y=0\Mat x'=0,\ s>0\\
\Longleftrightarrow
l=l',\ v'=sv+l'Y.
\end{gathered}\end{equation}
Correspondingly the third line of \eqref{22.9.2010.10} becomes
\begin{equation}\begin{gathered}\label{22.9.2010.10b}
x'=0,\ \lambda+\lambda'+\mu' Y=0,\ \mu+s\mu'=0\\
\Mor t=0,\ s_2\mu +s_1\mu'=0,
\ s_2(\lambda+\lambda')+\mu'_1+\sum_{j\geq 2}\mu'_j Z_j=0.
\end{gathered}\end{equation}
The rest of the argument is unchanged, except that the conclusion is that
$\Tbff^* M^2_0$ is being reconstructed.
\end{proof}
Note that for any manifold with corners, $X,$ the b-cotangent
bundle of any boundary hypersurface $H$ (or indeed any boundary face) includes
naturally as a subbundle $\Tb^*H\hookrightarrow \Tb^*_HX.$
\begin{lemma}
The Hamilton vector field of $g_\delta$ lifts from either the left or the right factor
of $M$ in $M^2$ to a smooth vector field, tangent to the boundary hypersurfaces,
on $\Tb^* M^2_0$, as well as on $\Tbff^* M^2_0$,
still denoted by $\sH^L_{g_\delta}$, resp.\ $\sH^R_{g_\delta}$. Moreover,
$\sH^L_{g_\delta}=\rho_L V^L$, $\sH^R_{g_\delta}=\rho_R V^R$, where
$V_L$, resp.\ $V_R$ are smooth vector fields tangent to all hypersurfaces
except the respective
cotangent bundles over the left, resp.\ right, boundaries, to which
they are transversal, and where $\rho_L$ and $\rho_R$ are defining functions of the
respective cotangent bundles over these boundaries.
\end{lemma}
\begin{proof}
Inserting the explicit form of the Euclidean metric, the Hamilton vector
field in \eqref{SeClRe.17} becomes
\begin{equation}
\boH_{p_0}=\la( x\p_x+\mu\p_\mu)-|\mu|^2\p_\la + x\mu\cdot\pa_y.
\label{SeClRe.37}\end{equation}
Consider the lift of this vector field to the product, $\To M\times\To M,$
from left and right, and then under the blow up of the diagonal in the
boundary. In the coordinate systems \eqref{SeClRe.35} and \eqref{22.9.2010.1}
\begin{equation}
\begin{gathered}
\boH_{p_0}^L=\lambda (s\pa_s+\mu\pa_\mu)-|\mu|^2\pa_{\la}+
s\mu(x'\pa_y-\pa_Y)\\
\boH_{p_0}^R=\lambda'(x'\pa_{x'}-s\pa_s-Y\pa_Y+\mu'\pa_{\mu'})-|\mu'|^2\pa_{\la'}+\mu'\pa_Y
\\
\begin{aligned}
\boH_{p_0}^L=\la(s_1\pa_{s_1}+\mu\pa_{\mu})&
-|\mu|^2\pa_{\la}+s_1\sum\limits_{j\ge2}\mu_j(t\pa_{y_j}-\pa_{Z_j})\\
+&s_1\mu_1(t\pa_{y_1}-t\pa_t+s_1\pa_{s_1}+s_2\pa_{s_2}
+\sum\limits_{j\ge2}Z_j\pa_{Z_j})
\end{aligned}\\
\begin{aligned}
\boH_{p_0}^R=\la'(s_2\pa_{s_2}+\mu'\pa_{\mu'})&
-|\mu'|^2\pa_{\la'}+s_2\sum\limits_{j\ge2}\mu'_j\pa_{Z_j}\\
+&s_2\mu'_1(t\pa_t-s_1\pa_{s_1}-s_2\pa_{s_2}-\sum\limits_{j\ge2}Z_j\pa_{Z_j}).
\end{aligned}
\end{gathered}
\label{SeClRe.38}\end{equation}
Note that the bundle itself is just pulled back here, so only
the base variables are changed.
Next we carry out the blow ups of Lemma~\ref{SeClRe.34}. The centers of
blow up are given explicitly, in local coordinates, in \eqref{22.9.2010.10}, with the
third line replaced by \eqref{22.9.2010.10b} in the case of $\iota^2_{0\bl,\ff}$.
We are only interested in the behaviour of the lifts of the vector
fields in \eqref{SeClRe.38} near the front faces introduced in the blow
ups.
Consider $\iota^2_{0\bl}$ first.
For the first two cases there are two blow-ups, first of $\mu=0$ in
$s=0$ and then of $\mu+s\mu'=0$ in $x'=0.$ Thus, near the front face of the
first blow up, the $\mu$ variables are replaced by $\tilde\mu=\mu/s$ and
then the center of the second blow up is $\tilde\mu+\mu'=0,$ $x'=0.$ Thus,
near the front face of the second blow up we can use as coordinates $s,$
$x',$ $\mu'$ and $\nu=(\tilde\mu+\mu')/x',$ i.e.\ substitute
$\tilde\mu=-\mu'+x'\nu.$ In the coordinate patch \eqref{SeClRe.38} the
lifts under the first blow up are
\begin{equation}
\begin{gathered}
\boH_{p_0}^L\longmapsto
\lambda s\pa_s-s^2|\tilde\mu|^2\pa_{\la}+
s^2\tilde\mu(x'\pa_y-\pa_Y)
\\
\boH_{p_0}^R\longmapsto
\lambda'(x'\pa_{x'}-s\pa_s-Y\pa_Y+\tilde\mu\pa_{\tilde\mu}+
\mu'\pa_{\mu'})-|\mu'|^2\pa_{\la'}+
\mu'\pa_Y.
\end{gathered}
\label{SeClRe.39}\end{equation}
Thus under the second blow up, the left Hamilton vector field lifts to
\begin{equation}
\begin{gathered}
\boH_{p_0}^L\longmapsto sT,\
T=\lambda \pa_s-s|\tilde\mu|^2\pa_{\la}+
s\tilde\mu(x'\pa_y-\pa_Y)
\end{gathered}
\label{SeClRe.41}\end{equation}
where $T$ is transversal to the boundary $s=0$ where $\lambda\not=0.$
A similar computation near the corner shows the lifts of the two Hamilton
vector fields under blow up the fibrations of $s_1=0$ and $s_2=0$ in terms
of the new coordinates $\tilde\mu=\mu/s_1$ and $\tilde\mu'=\mu'/s_2$ to be
\begin{equation}
\begin{gathered}
\begin{aligned}
\boH_{p_0}^L=&\la s_1\pa_{s_1}-s_1^2|\tilde\mu|^2\pa_{\la}+s_1^2\sum\limits_{j\ge2}\tilde\mu_j(t\pa_{y_j}-\pa_{Z_j})\\
+&s_1^2\tilde\mu_1
(t\pa_{y_1}-t\pa_t+s_1\pa_{s_1}-\tilde\mu\pa_{\tilde\mu}+s_2\pa_{s_2}
-\tilde\mu'\pa_{\tilde\mu'}+\sum\limits_{j\ge2}Z_j\pa_{Z_j}),
\end{aligned}\\
\begin{aligned}
\boH_{p_0}^R=&\la's_2\pa_{s_2}-s_2^2|\tilde\mu'|^2\pa_{\la'}+s_2^2\sum\limits_{j\ge2}\tilde\mu'_j\pa_{Z_j}\\
+&s_2^2\tilde\mu'_1
(t\pa_t-s_1\pa_{s_1}+{\tilde\mu}\pa_{\tilde\mu}-s_2\pa_{s_2}+
\tilde\mu'\pa_{\tilde\mu'}
-\sum\limits_{j\ge2}Z_j\pa_{Z_j}).
\end{aligned}
\end{gathered}
\label{SeClRe.42}\end{equation}
The final blow up is that of $t=0$, $\tilde\mu+\tilde\mu'=0$, near the front
face of this blow-up replacing $(t,\tilde\mu,\tilde\mu')$ by $(t,\tilde\mu,\tilde\nu)$,
$\tilde\nu=(\tilde\mu+\tilde\mu')/t$, as valid coordinates (leaving the others
unaffected). Then the vector fields above become
\begin{equation}
\begin{gathered}
\boH_{p_0}^L=s_1\tilde T^L,\ \boH_{p_0}^R=s_2\tilde T^R,\\
\begin{aligned}
\tilde T^L=&\la \pa_{s_1}-s_1|\tilde\mu|^2\pa_{\la}+s_1\sum\limits_{j\ge2}\tilde\mu_j(t\pa_{y_j}-\pa_{Z_j})\\
+&s_1\tilde\mu_1
(t\pa_{y_1}-t\pa_t+s_1\pa_{s_1}-\tilde\mu\pa_{\tilde\mu}+s_2\pa_{s_2}
+\sum\limits_{j\ge2}Z_j\pa_{Z_j}),
\end{aligned}\\
\begin{aligned}
\tilde T^R=&\la'\pa_{s_2}-s_2|\tilde\mu'|^2\pa_{\la'}+s_2\sum\limits_{j\ge2}\tilde\mu'_j\pa_{Z_j}\\
+&s_2\tilde\mu'_1
(t\pa_t-s_1\pa_{s_1}+{\tilde\mu}\pa_{\tilde\mu}-s_2\pa_{s_2}
-\sum\limits_{j\ge2}Z_j\pa_{Z_j}).
\end{aligned}
\end{gathered}
\label{SeClRe.42b}\end{equation}
Thus both left and right Hamilton vector fields are transversal to the
respective boundaries after a vanishing factor is removed, provided
$\lambda,$ $\lambda '\not=0.$
The final step is to show that the same arguments apply to the perturbed
metric. First consider the lift, from left and right, of
the perturbation to the Hamilton vector field arising from the
perturbation of the metric. By assumption, the perturbation $H$ is a
2-cotensor which is smooth up to the boundary. Thus, as a perturbation of
the dual metric function on $\To M$ it vanishes quadratically at the
boundary. In local coordinates near a boundary point it follows that the
perturbation of the differential of the metric function is of the form
\begin{equation}
dp-dp_0=x^2(a\frac{dx}x+b dy+cd\lambda +ed\mu)
\label{SeClRe.46}\end{equation}
From \eqref{SeClRe.44} it follows that the perturbation of the
Hamilton vector field is of the form
\begin{equation}
\sH_{p}-\sH_{p_0}=x^2(a'x\pa_x+b'x\pa_y+c'\pa_\lambda +e'\pa_\mu)
\label{SeClRe.45}\end{equation}
on $\To M.$ Lifted from the right or left factors to the product and then
under the blow-up of the diagonal to $M^2_0$ it follows that in the
coordinate systems \eqref{SeClRe.35} and \eqref{22.9.2010.1}, the
perturbations are of the form
\begin{equation}
\begin{gathered}
\sH_{p}^L-\sH_{p_0}^L=s^2(x')^2V^L,\ \sH_{p}^R-\sH_{p_0}^R=(x')^2V^R,\ V^L,\ V^R\in\Vb,\\
\sH_{p}^L-\sH_{p_0}^L=s_1^2t^2W^L ,\ \sH_{p}^R-\sH_{p_0}^R=s_2^2t^2W^R,\ W^R,\ W^R\in\Vb
\end{gathered}
\label{SeClRe.47}\end{equation}
where $\Vb$ denotes the space of smooth vector fields tangent to all
boundaries. Since the are lifted from the right and left factors, $V^L$ and
$V^R$ are necessarily tangent to the annihilator submanifolds of the right
and left boundaries. It follows that the vector fields $sx'V^L$ and $x'V^R$
are tangent to both fibrations above a coordinate patch as in
\eqref{SeClRe.35} and $s_1tV^L$ and $s_2tV^R$ are tangent to all three
annihilator submanifolds above a coordinate patch
\eqref{22.9.2010.1}. Thus, after the blow ups which reconstruct
$\Tb^*M^2_0,$ the perturbations lift to be of the form
\begin{equation}
\sH_{p}^L-\sH_{p_0}^L=\rho _L\rho _{\ff}U^L,\ \sH_{p}^R-\sH_{p_0}^R=\rho _R\rho _{\ff}U^R
\label{SeClRe.48}\end{equation}
where $U^L$ and $U^R$ are smooth vector fields on $\Tb^*M^2_0.$
From \eqref{SeClRe.48} it follows that the transversality properties in
\eqref{SeClRe.41} and \eqref{SeClRe.42b} persist.
Now consider $\iota^2_{0\bl,\ff}$. First, \eqref{SeClRe.39} is unchanged, since
the annihilators on the `old' boundary faces are the same in this case.
In particular, we still have $\tilde\mu=\mu/s$ as one of our coordinates
after the first blow up; the center of the second blow up is then $x'=0$,
$\lambda+\lambda'+\mu'\cdot Y=0$, $\tilde\mu+\mu'=0$.
Thus,
near the front face of the second blow up we can use as coordinates $s,$
$x',$ $\mu'$ and $\sigma=(\lambda+\lambda'+\mu'\cdot Y)/x'$,
$\nu=(\tilde\mu+\mu')/x',$ i.e.\ substitute
$\tilde\mu=-\mu'+x'\nu$, i.e.\ $\mu=-\mu's+x's\nu$,
and $\lambda=-\lambda'-\mu'\cdot Y+x'\sigma$.
Thus under the second blow up, the left Hamilton vector field lifts to
\begin{equation}
\begin{gathered}
\boH_{p_0}^L\longmapsto sT',\
T'=\lambda \pa_s+\mu\cdot(-\nu\pa_\sigma+x'\pa_y-\pa_Y),
\end{gathered}
\label{SeClRe.41b}\end{equation}
so $T'$ is transversal to the boundary $s=0$ where $\lambda\not=0.$
In the other coordinate chart, again, \eqref{SeClRe.42} is unchanged since the
annihilators on the `old' boundary faces are the same.
The final blow up is that of
$$
t=0,\ \tilde\mu+\tilde\mu'=0,
\ \lambda+\lambda'+\tilde\mu'_1+\sum_{j\geq 2}\tilde\mu'_j Z_j=0,
$$
near the front
face of this blow-up replacing $(t,\tilde\mu,\tilde\mu',\lambda,\lambda')$
by $(t,\tilde\mu,\tilde\nu,\lambda,\tilde\sigma)$,
$$
\tilde\nu=(\tilde\mu+\tilde\mu')/t,\ \tilde\sigma=(\lambda+\lambda'+\tilde\mu'_1+\sum_{j\geq 2}\tilde\mu'_j Z_j)/t,
$$
as valid coordinates (leaving the others
unaffected). Then the vector fields above become
\begin{equation}
\begin{gathered}
\boH_{p_0}^L=s_1\hat T^L,\ \boH_{p_0}^R=s_2\hat T^R,\\
\begin{aligned}
\hat T^L=&\la \pa_{s_1}-s_1|\tilde\mu|^2\pa_{\la}+s_1\sum\limits_{j\ge2}\tilde\mu_j(t\pa_{y_j}-\pa_{Z_j}-\tilde\nu_j\pa_{\tilde\sigma})\\
+&s_1\tilde\mu_1
(t\pa_{y_1}-t\pa_t+s_1\pa_{s_1}+\tilde\sigma\pa_{\tilde\sigma}-\tilde\nu\pa_{\tilde\nu}
+s_2\pa_{s_2}-\tilde\nu_1\pa_{\tilde\sigma}
+\sum\limits_{j\ge2}Z_j\pa_{Z_j}),
\end{aligned}\\
\begin{aligned}
\hat T^R=&\la'\pa_{s_2}+s_2\sum\limits_{j\ge2}\tilde\mu'_j\pa_{Z_j}\\
+&s_2\tilde\mu'_1
(t\pa_t-s_1\pa_{s_1}+{\tilde\mu}\pa_{\tilde\mu}-s_2\pa_{s_2}-\tilde\sigma\pa_{\tilde\sigma}
-\sum\limits_{j\ge2}Z_j\pa_{Z_j}).
\end{aligned}
\end{gathered}
\label{SeClRe.42c}\end{equation}
Again, both left and right Hamilton vector fields are transversal to the
respective boundaries after a vanishing factor is removed, provided
$\lambda,$ $\lambda '\not=0.$ The rest of the argument proceeds as above.
\end{proof}
\begin{prop}\label{SeClRe.23}\label{distance} The differential of the
distance function for the perturbed metric $g_\delta,$ for sufficiently
small $\delta,$ on $M=\Bnc$ defines a global smooth Lagrangian
submanifold $\Lambda_\delta,$ of $\Tb^*(M\times_0M),$ which is a smooth
section outside the lifted diagonal and which lies in $\Tb^*\ff$ over the
front face, $\ff(M^2_0)$ and in consequence there is a unique geodesic
between any two points of $\Bo,$ no conjugate points and \eqref{blowd}
remains valid for $\dist_{\delta}(z,z').$
\end{prop}
\begin{proof} For the unperturbed metric this already follows from
Lemma~\ref{dist0}. We first reprove this result by integrating the
Hamilton vector fields and then examine the effect of the metric
perturbation. Thus we first consider the lift of the Hamilton vector
field of the hyperbolic distance function from $\To M,$ from either the
left of the right, to $\Tbff^*M^2_0,$ using the preceding lemma.
Although the global regularity of the Lagrangian which is the graph of the
differential of the distance is already known from the explicit formula in
this case, note that it also follows from the form of these two vector
fields. The initial manifold, the unit conormal bundle to the diagonal,
becomes near the corner of $M^2$ the variety of those
\begin{equation}
\xi(dx-dx')+\eta(dy-dy')\text{ such that
}\xi^2+|\eta|^2=\frac1{2x^2},\ x=x'>0,\ y=y'.
\label{SeClRe.49}\end{equation}
In the blown up manifold $M^2_0$ the closure is smooth in $\Tb^*M^2_0$ and
is the bundle over the lifted diagonal given in terms of local coordinates
\eqref{SeClRe.35} by
\begin{equation}
\la\frac{ds}s+\mu dY,\ \la^2+|\mu|^2=\frac12,\ s=1,\ Y=0;
\label{SeClRe.50}\end{equation}
the analogous statement also holds in $\Tbff^*M^2_0$ where one has the `same'
expression.
Using the Hamilton flow in $\Tbff^*M^2_0$, we deduce that
the flow out is
a global smooth submanifold, where smoothness includes up to all
boundaries, of $\Tbff^*M^2_0$, and is also globally a graph away from the lifted
diagonal, as follows from the explicit form of the vector fields. Note that
over the interior of the front face, $\Tbff^*M^2_0$ is just the standard
cotangent bundle, so smoothness of the distance up to the front face
follows.
Over the
left and right boundaries the Lagrangian lies in $\lambda =1$ and $\lambda
'=1$ so the form \eqref{blowd} of the distance follows.
The analogous conclusion can also be obtained by using the flow in
$\Tb^*M^2_0$. In this setting, we need that
over the front face \eqref{SeClRe.50} is contained in the image of $\Tb^*\ff$ to which
both lifted vector fields are tangent. Thus it follows that the flow out is
a global smooth submanifold, where smoothness includes up to all
boundaries, of $\Tb^*M^2_0$ which is contained in the image of $\Tb^*\ff$
over $\ff(M^2_0).$ Again, it is globally a graph away from the lifted
diagonal, as follows from the explicit form of the vector fields. Over the
left and right boundaries the Lagrangian lies in $\lambda =1$ and $\lambda
'=1$ so the form \eqref{blowd} of the distance follows.
For small $\delta$ these
perturbation of the Hamilton vector fields are also small in supremum norm
and have the same tangency properties at the boundaries used above to
rederive \eqref{blowd}, from which the Proposition follows.
\end{proof}
\section{The semiclassical double space}\label{semiclassical-double-space}
In this section we construct the semiclassical double space, $M_{0,\semi}$, which
will be the locus of our parametrix construction. To motivate the construction, we
recall that
Mazzeo and the first author \cite{Mazzeo-Melrose:Meromorphic} have analyzed the resolvent $R(h,\sigma),$ defined in equation \eqref{scresolvent}, for $\sigma/h\in\Cx$, though the construction is
not uniform as $|\sigma/h|\to\infty$. They achieved this
by constructing the Schwartz kernel of the parametrix $G(h,\sigma)$ to
$R(h,\sigma)$ as a conormal distribution defined on the manifold
$$
M_0=\BB^{n+1} \times_0 \BB^{n+1}
$$
defined in
\eqref{zero-blow-up}, see also figure \ref{fig1}, with meromorphic dependence on
$\sigma/h$.
The manifold $ \BB^{n+1}\times \BB^{n+1}$ is a $2n+2$ dimensional manifold with corners. It contains two boundary components of codimension one, denoted as in \cite{Mazzeo-Melrose:Meromorphic} by
\begin{equation*}
\p_1^l( \BB^{n+1} \times \BB^{n+1})= \p \BB^{n+1} \times \BB^{n+1} \text{ and }
\p_1^r( \BB^{n+1} \times \BB^{n+1})= \BB^{n+1} \times \p \BB^{n+1},
\end{equation*}
which have a common boundary
$\p_2( \BB^{n+1}\times \BB^{n+1})=\p \BB^{n+1} \times \p \BB^{n+1}.$ The lift of
$\p_1^r( \BB^{n+1} \times \BB^{n+1})$ to $M_0$, which is the closure of
$$
\beta^{-1}(\p_1^r( \BB^{n+1} \times \BB^{n+1})\setminus \p_2( \BB^{n+1}\times \BB^{n+1})),
$$
with
$\beta:M_0\to\BB^{n+1}\times\BB^{n+1}$ the blow-down map, will be called the right face and denoted by $\mcr.$ Similarly,
the lift of $\p_1^l( \BB^{n+1} \times \BB^{n+1})$ will be called the left face and denoted by $\mcl.$ The lift of $\Diag \cap \p_2( \BB^{n+1}\times \BB^{n+1})$, which is its
inverse image under $\beta$, will be called the front face $\mcf,$ see figure \ref{fig1}.
We briefly recall the definitions of their classes of pseudodifferential operators, and refer the reader to
\cite{Mazzeo-Melrose:Meromorphic} for full details. First they define the class $\Psi_0^m(\BB^{n+1})$ which consists of those pseudodifferential operators of order
$m$ whose Schwartz kernels lift under the blow-down map $\beta$ defined in \eqref{zero-blow-up} to a distribution which is conormal (of order $m$) to the lifted diagonal and vanish to infinite order at all faces, with the exception of the front face, up to
which it is $\CI$ (with values in conormal distributions).
Here, and elsewhere in the paper, we trivialized the right density bundle using
a zero-density; we conveniently fix this as
$|dg_\delta(z')|$. Thus, the Schwartz kernel of $A\in\Psi_0^m(\Bn)$ is
$K_A(z,z') |dg_\delta(z')|$, with $K_A$ as described above, so in particular is $\CI$
up to the front face.
It then becomes necessary to introduce another class of operators whose kernels are singular at the right and left faces. This class will be denoted by $\Psi_0^{m,a,b}(\BB^{n+1}),$ $a,b \in \CC.$ An operator $P\in \Psi_0^{m,a,b}(\BB^{n+1})$ if it can be written as a sum $P=P_1+P_2,$ where $P_1\in \Psi_0^m(X)$ and the Schwartz kernel $K_{P_2}|dg_\delta(z')|$
of the operator $P_2$ is such that $K_{P_2}$
lifts under $\beta$ to a conormal distribution which is smooth up to the front face, and which satisfies the following conormal regularity with respect to the other faces
\begin{equation}
\mcv_b^{k} \beta^* K_{P_2} \in \rho_L^a \rho_R^b L^\infty(\BB^{n+1} \times_0 \BB^{n+1}), \;\ \forall \;\ k \in \mn, \label{boundary-regularity}
\end{equation}
where $\mcv_b$ denotes the space of vector fields on $M_0$ which are tangent to the right and left faces.
Next we define the semiclassical blow-up of
$$
\Bn\times\Bn\times[0,1)_h,
$$
and the corresponding classes of pseudodifferential operators associated with it that will be used in the construction of the parametrix. The semiclassical
double space
is constructed in two steps. First, as in \cite{Mazzeo-Melrose:Meromorphic}, we blow-up the intersection of the diagonal $\Diag \times [0,1)$ with $\p \BB^{n+1} \times \p \BB^{n+1} \times [0,1).$ Then we blow-up the intersection of the lifted diagonal times $[0,1)$ with $\{h=0\}.$ We define the manifold with corners
\begin{equation}
M_{0,\semi}=[\BB^{n+1}\times \BB^{n+1} \times [0,1]_{h }; \pa \Diag\times[0,1);\Diag_0\times\{0\}]. \label{scblowup1}
\end{equation}
See Figures~\ref{figscblow-up1} and \ref{figscblow-up2}. We will denote the blow-down map
\begin{equation}
\beta_{\semi}: M_{0,\semi} \longrightarrow \BB^{n+1} \times \BB^{n+1} \times [0,1). \label{betasc}
\end{equation}
\begin{figure}[int1
\epsfxsize= 3.5in
\centerline{\epsffile{scblowup1.eps}}
\caption{The stretched product $\BB^{n+1}\times_0 \BB^{n+1}\times [0,1).$}
\label{figscblow-up1}
\end{figure}
\begin{figure}[int1
\epsfxsize= 3.5in
\centerline{\epsffile{scblowup2.eps}}
\caption{The semiclassical blown-up space $M_{0,h}$ obtained by blowing-up $\BB^{n+1}\times_0 \BB^{n+1} \times [0,1)$ along $\Diag_0\cap \times \{h=0\}.$ }
\label{figscblow-up2}
\end{figure}
As above, we can define the right and left semiclassical faces as the lift of
\begin{gather*}
\p_1^l\left( \BB^{n+1} \times \BB^{n+1} \times [0,1)\right)= \p \BB^{n+1} \times \BB^{n+1} \times [0,1)\text{ and } \\
\p_1^r\left( \BB^{n+1} \times \BB^{n+1} \times [0,1))\right)= \BB^{n+1} \times \p \BB^{n+1} \times [0,1),
\end{gather*}
by the blow-down map $\beta_{\semi}.$ We will denote the lift of the diagonal under the map $\beta_{\semi}$ by $\Diag_{\semi},$ i.e.
\begin{gather*}
\Diag_{\semi}= \text { the closure of } \beta_{\semi}^{-1}\left( \Diag \times (0,1)\setminus (\Diag \cap (\p \Bn \times \p \Bn))\right).
\end{gather*}
The lift of $\pa \Diag \times [0,1)$ will be called the zero front face $\mcf,$ while the boundary obtained by the blow-up of $\Diag_0 \times [0,1]$ along $\Diag_0\times \{0\}$ will be called the semiclassical front face $\mathcal {S}.$ The face which is obtained by the lift of $\Bn \times \Bn \times \{0\}$ is the semiclassical face, and will be denoted by $\mathcal {A}.$
We wish to find a parametrix such that $P(h,\sigma)$ acting on the left
produces the identity plus an error which vanishes to high enough order on the right and left
faces, $\mcr$ and $\mcl,$ and to infinite order at the zero-front face $\mcf,$ the semiclassical front face $\mathcal {S}$ and the semiclassical face $\mathcal {A}.$ Thus
the error term is bounded as an operator acting between weighted $L^2(\Bn)$ spaces and its norm goes to zero as $h\downarrow 0.$
As in \cite{Mazzeo-Melrose:Meromorphic} we define the class of semiclassical pseudodifferential operators in two steps. First we define the space
$P\in\Psi_{0,\semi}^m(\Bn)$ which consists of operators whose kernel
$$
K_P(z,z',h)\,|dg_\delta(z')|
$$
lifts to a conormal distribution of order $m$ to the lifted diagonal and
vanishes to infinite order at all faces, except the zero front face, up to which it is
$\CI$ (with values in conormal distributions)
and the semiclassical front face, up to which it is $h^{-n-1}\CI$ (with values in
conormal distributions). We then define the space
\begin{gather}
\mck^{ a,b,c}(M_{0,\semi})=\{ K \in L^{\infty}(M_{0,\semi}): \mcv_b^m K \in \rho_L^a \rho_A^b \rho_R^c \rho_S^{-n-1}L^\infty(M_{0,\semi}), \;\ m\in \NN \}, \label{sc-con}
\end{gather}
where $\mcv_b$ denotes the Lie algebra of vector fields which are tangent to $\mcl,$ $\mathcal {A}$ and $\mcr.$ Again, as in \cite{Mazzeo-Melrose:Meromorphic}, we define the space $\Psi_{0,\semi}^{m,a,b,c}(\Bn)$ as the operators $P$ which can be expressed in the form $P=P_1+P_2,$ with $P_1\in \Psi_{0,\semi}^{m}(\Bn)$ and the kernel
$K_{P_2}\,|dg_\delta(z')|$ of $P_2$ is such
$\beta_{\semi}^* K_{P_2} \in \mck^{ a,b,c}(M_{0,\semi}).$
\section{A semiclassical parametrix for the resolvent in dimension three}\label{3D-parametrix}
In this section we construct a parametrix for the resolvent $R(h,\sigma)$ defined in \eqref{scresolvent} in dimension three. We do this case separately because it
is much simpler than in the general case, and one does not have to perform the semiclassical blow-up. Besides,
we will also need part of this construction in the general case. More precisely, if $n+1=3,$ and the metric $g$ satisfies the hypotheses of Proposition \ref{distance}, we will use Hadamard's method to construct the leading asymptotic term of the parametrix
$G(\sigma,h)$ at the diagonal, and the top two terms of the semiclassical
asymptotics. Our construction takes place on
$(\BB^3\times_0\BB^3)\times [0,1)_h$, instead of its semiclassical
blow up, i.e.\ the blow up of the zero-diagonal at $h=0$, as described above. This is made possible by a coincidence, namely
that in three dimensions, apart from an explicit exponential factor, the
leading term in the asymptotics lives on $\BB^3\times_0 \BB^3 \times [0,1)$. However, to obtain further
terms in the asymptotics would require the semiclassical blow up, as
it will be working in higher dimensions. For example, this method only give bounds for the resolvent of width 1, while the more general construction gives the bounds on any strip.
We recall that in three dimensions the resolvent of the Laplacian in hyperbolic space, $\Delta_{g_0},$
$R_0(\sigma)=(\Delta_{g_0}-\sigma^2-1)^{-1}$
has a holomorphic continuation to $\mc$ as an operator from functions vanishing to infinite order at
$\p \Bc$ to distributions in $\Bc.$
The Schwartz kernel of $R_0(\sigma)$ is given by
\begin{gather}
R_0(\sigma,z,z')= \frac{e^{-i\sigma r_0}}{4\pi \sinh r_0}, \label{resol0}
\end{gather}
where $r_0=r_0(z,z')$ is the geodesic distance between $z$ and $z'$ with respect to the metric
$g_{0},$ see for example \cite{Mazzeo-Melrose:Meromorphic}.
Since there are no conjugate points for the geodesic flow of $g,$ for each
$z'\in \Bo,$ the exponential map for the metric $g,$
$\exp_{g}: T_{z'} \Bo \longrightarrow \Bo,$ is a global diffeomorphism. Let
$(r,\theta)$ be geodesic normal coordinates for $g$ which are valid in $\Bo\setminus \{z'\};$
$r(z,z')=:d(z,z')$ is the distance function for the metric $g.$ Since $r(z,z')$ is globally defined, $g$ is a small perturbation of $g_0$ and the kernel of $R_0(\sigma)$ is given by \eqref{resol0}, it is reasonable to seek a parametrix of $R(h,\sigma)$ which has kernel of the form
\begin{gather}
G(h,\sigma,z,z')=e^{-i \frac{\sigma}{h} r} h^{-2}U(h,\sigma,z,z'), \label{parametrix}
\end{gather}
with $U$ properly chosen.
We now reinterpret this as a semiclassical Lagrangian distribution to relate it to the results
of Section~\ref{distance-function}. Thus, $-\sigma r=-\sigma d(z,z')$ is the phase function for semiclassical distributions corresponding
to the backward left flow-out of the conormal bundle of the diagonal inside
the characteristic set of $2p_\ep-\sigma^2$. This flowout is the same as
the forward right flow-out of the conormal bundle of the diagonal, and is
also the dilated version, by a factor of $\sigma$, in the fibers of the
cotangent bundle, of the flow-out of
in the characteristic set of $2p_\ep-1$, which we described in detail
in Section~\ref{distance-function}.
In view of the results of Section~\ref{distance-function},
for the characteristic set of $2p_\ep-1$ (the general case of
$2p_\ep-\sigma^2$ simply gives an overall additional factor of
$\sigma$ due to the dilation), the
lift $\beta^*\pa_r$ of $\pa_r$ to $\Bc\times_0\Bc$ satisfies
\begin{equation}\label{eq:pa_r-V_eps}
\beta^*\pa_r=\widetilde{\Pi_\eps}_*V_\eps^L,
\end{equation}
and thus is a $\CI$ vector field on
$(\Bc\times_0\Bc)\setminus\Diag_0$
which is tangent to all boundary faces, and at the left face
$L=B$,
\begin{equation}\label{eq:pa_r-at-left-face}
\pa_r=-\mathsf{R}_L+W_L,
\end{equation}
where
$\mathsf{R}_L$ is the radial vector field corresponding to the left face, and
$W_L\in\rho_L\Vb((\Bc\times_0\Bc)\setminus\Diag_0)$.
It is convenient to
blow up $\Diag_0$ and lift $\beta^*\pa_r$ further to this space.
We thus define
$\Bc \times_1 \Bc$ to be the manifold obtained from $\Bc \times_0 \Bc$ by blowing-up the diagonal $\Diag_0$ as shown in Fig. \ref{fig2}. Let $\beta_d : \Bc\times_1 \Bc\longmapsto \Bc \times_0 \Bc$ denote the blow-down map and let $\beta_d\circ \beta=\beta_{0d}.$
The vector field $\beta_{0d}^* \p_r$ is transversal to the
new face introduced by blowing up the diagonal; it still satisfies the lifted analogue of
\eqref{eq:pa_r-at-left-face}. Moreover, integral curves of $\beta_{0d}^* \p_r$ hit
the left face $L$ away from its intersection with the right face $R$ in finite time.
\begin{figure}[int1
\epsfxsize= 2.0in
\centerline{\epsffile{blow-up1.eps}}
\caption{The manifold $\Bn \times_1 \Bn.$}
\label{fig2}
\end{figure}
In coordinates $(r,\theta)$ the metric $g$ is given by
\begin{gather}
g = dr^2 + H(r, \theta, d\theta),\label{gnorm}
\end{gather}
where $H(r,\theta,d\theta)$ is a $\CI$ 1-parameter family of metrics on $\ms^2.$
The Laplacian with respect to $g$ in these coordinates is given by
\begin{gather*}
\Delta_g= -\p_r^2 -V \p_r + \Delta_H, \;\ V= \frac{1}{|g|^\ha} \p_r (|g|^\ha),
\end{gather*}
where $|g|^\ha$ is the volume element of the metric $g$ and $\Delta_H$ is the Laplacian
with respect to $H$ on $\ms^2.$ The volume element $|g|^\ha$ has the following expansion as $r\downarrow 0,$
\begin{gather}
|g|^{\ha}(r,\theta)= r^{2}(1+ r^2g_1(r,\theta)), \label{volume1}
\end{gather}
see for example page 144 of \cite{Gallot-Hulin-Lafontaine}. So
\begin{gather}
\Delta_g =-\p_r^2 -(\frac{2}{r}+rA) \p_r +\Delta_H \label{volume3}
\end{gather}
We want $U$ in \eqref{parametrix} to be of the form
\begin{gather}
U(h,\sigma,z,z')=U_0(\sigma,z,z')+ h U_1(\sigma,z,z'), \label{parametrix01}
\end{gather}
and so
\begin{equation}\begin{split}
&\left(h^2(\Delta_g+x^2W-1)-\sigma^2\right)
e^{-i\frac{\sigma}{h} r} h^{-2} U \\
&=e^{-i\frac{\sigma}{h}r}\Big( (\Delta_g+x^2 W-1) U_0+ 2i \frac{\sigma}{h} |g|^{-\oq} \p_r(|g|^\oq U_0)\\
&\qquad\qquad+
2i\sigma |g|^{-\oq} \p_r(|g|^\oq U_1) + h (\Delta_g+x^2 W-1)U_1\Big).
\end{split}\label{parametrix1}
\end{equation}
Here the leading term in $h$ as $h\to 0$ both overall, and as far as $U_0$ is concerned, is
\begin{equation}\label{eq:0th-transport}
2i \frac{\sigma}{h} |g|^{-\oq} \p_r(|g|^\oq U_0),
\end{equation}
and the leading term as far as $U_1$ is concerned is
\begin{equation}\label{eq:1st-transport}
2i\sigma |g|^{-\oq} \p_r(|g|^\oq U_1).
\end{equation}
In the interpretation as semiclassical Lagrangian distributions,
these are both the differential operators arising in the transport equations. For Hamiltonians
given by Riemannian metrics, these operators in the interior are well-known to be
Lie derivatives with respect to the Hamilton vector field, when interpreted as acting on half-densities.
This can also be read off directly from \eqref{eq:0th-transport}-\eqref{eq:1st-transport}, with
$|g|^\oq$ being the half-density conversion factor, $\sigma$ is due to working at energy
$\sigma$ (rather than $1$), and the factor of $2$ is due to the symbol of the Laplacian
being $2p_\eps$.
To get rid of the term in $h^{-1}$ we solve the 0th transport equation, i.e.\ we impose
\begin{gather*}
\p_r(|g|^\oq U_0)=0,
\end{gather*}
and we choose $U_0(r,\theta)= \frac{1}{4\pi} |g(r,\theta)|^{-\oq}.$ From \eqref{volume1} we have
\begin{gather}
|g|^{-\oq}(r,\theta)=r^{-1}(1+ r^2 g_2(r,\theta)) \text{ near } r=0. \label{expg}
\end{gather}
Therefore, near $r=0,$
\begin{gather}
\Delta_g \frac{1}{4\pi} |g(r,\theta)|^{-\oq}= \delta(z,z')+ \frac{1}{4\pi}A r^{-1} +\frac{1}{4\pi}\Delta_g (r g_2).
\label{expato}
\end{gather}
This only occurs in three dimensions, and makes this construction easier than in the general case. In higher dimensions the power or $r$ in \eqref{expg} is
$r^{-\frac{n}{2}},$ and does not coincide with the power of $r$ of the fundamental solution of the Laplacian, which, in dimension $n+1,$ is $r^{1-n},$ so one does not get the delta function in \eqref{expato}.
To get rid of the term independent of $h$ in \eqref{parametrix1} in $r>0$ we solve the
first transport equation,
\begin{gather*}
2i \sigma |g|^{-\oq}\p_r(|g|^\oq U_1)+ (\Delta_g+x^2W-1)U_0=0 \text{ in } r>0,\\
U_1=0 \text{ at } r=0.
\end{gather*}
So
\begin{gather}
U_1(r,\theta)= -\frac{1}{8i\sigma \pi}|g(r,\theta)|^{-\oq} \int_0^r |g|^\oq(s,\theta) \left(\Delta_g+x^2W-1\right)|g|^{-\oq}(s,\theta) \; ds.\label{formu1}
\end{gather}
Since $|g|^\oq$ is $\CI$ up to $r=0,$ and vanishes at $r=0,$ it follows from \eqref{expato} that
$|g|^\oq \Delta_g |g|^{-\oq}$ is $\CI$ up to $r=0.$ In particular the integrand in \eqref{formu1} is smooth up to $r=0.$ With these choices of $U_0$ and $U_1$ we obtain
\begin{gather}
\begin{gathered}
\left(h^2(\Delta_g+x^2W-1)-\sigma^2\right) e^{-i\frac{\sigma}{h} r} h^{-2}U(h,\sigma,z,z')\\
= \delta(z,z')+
h e^{-i\frac{\sigma}{h} r} (\Delta_g+x^2W-1) U_1(\sigma,z,z')
\end{gathered}\label{term1}
\end{gather}
This gives, in principle, a parametrix $G(h,\sigma,z,z')=e^{-i\frac{\sigma}{h} r} h^{-2}U(h,\sigma,z,z')$
in the interior of
$\Bc\times \Bc$ in the two senses that the diagonal singularity of $R(h,\sigma)$
is solved away to leading order, which in view of the ellipticity of the operator means that the error
$$
E(h,\sigma)=\left( h^2(\Delta_g+x^2W-1)-\sigma^2\right) G(h,\sigma)-\Id
$$
is a semiclassical pseudodifferential operator of order $-1$ (in a large
calculus, i.e.\ with non-infinite order vanishing off the semiclassical
front face, which did not even appear in our calculations), and the
top two terms of the
semiclassical parametrix of $\left(h^2(\Delta_g+x^2W-1)-\sigma^2\right)^{-1}$ are also found.
In fact, our parametrix is better than this. To understand the behavior of $G$ and the remainder
\begin{gather}
E(h,\sigma,z,z')=h e^{-i\frac{\sigma}{h} r}(\Delta_g +x^2W-1)U_1 \label{defE}
\end{gather}
near the boundary of $\Bc\times \Bc,$ we need to analyze the behavior of $U_0$
and $U_1$ at the left and right boundary faces.
We will do the computations for arbitrary dimensions, since we will need some of these estimates in the general case, but in this special situation we have $n=2.$
We start by noting that the asymptotics of $U_0$ and $U_1$ follow from the transport equation
which they satisfy. Indeed, much like we analyzed the flow-out of the conormal bundle of the
diagonal, we show now that
\begin{equation}\label{eq:U_0-U_1-asymp}
U_0\in \rho_D^{-1}\rho_L^{n/2}\rho_R^{n/2}\CI(\Bn\times_1\Bn),\ U_1\in R^2\rho_L^{n/2}\rho_R^{n/2}\CI(\Bn\times_1\Bn);
\end{equation}
here $\rho_D$ is the defining function of the front face of the blow-up creating $\Bn\times_1\Bn$.
Note that we have already shown this claim near this front face; the main content of the
statement is the precise behavior as $\rho_L,\rho_R\to 0$.
We start with $U_0$.
First, the conclusion away from the right face, $\rho_R=0$, follows immediately from
\eqref{eq:pa_r-at-left-face} since integral curves emanating from the lifted diagonal
hit this region at finite time, and solutions of the Lie derivative equation have this form
near the boundary. To have the analogous conclusion away from the left face, we
remark that solutions of the left transport equation automatically solve the right transport
equation; one can then argue by symmetry, or note
directly that as $-\pa_{r'}$ is the radial vector field at the right face, modulo an element
of $\rho_R\Vb(\Bn\times_0\Bn)$, and as integral curves of $\pa_{r'}$
emanating from the lifted diagonal
hit this region at finite time, and solutions of the Lie derivative equation have this form
near the boundary. It remains to treat the corner where both $\rho_L=0$ and $\rho_R=0$.
The conclusion here now follows immediately as integral curves of $\pa_r$ reach
this corner in finite time from a punctured neighborhood of this corner, and in this
punctured neighborhood we already have the desired regularity. This proves
\eqref{eq:U_0-U_1-asymp} for $U_0$.
To treat $U_1$, it suffices to prove that
\begin{equation}\label{eq:U_0-error-asymp}
(\Delta_g+x^2W-1)U_0\in R^2\rho_L^{n/2+2}\rho_R^{n/2}\CI(\Bn\times_0\Bn\setminus \Diag_0),
\end{equation}
for then
\begin{equation}\label{eq:U_1-asymp}
U_1\in R^2\rho_L^{n/2}\rho_R^{n/2}\CI(\Bn\times_0\Bn\setminus \Diag_0),
\end{equation}
by the same arguments as those giving the asymptotics of $U_0$, but now applied to
the inhomogeneous transport equation.
On the other hand, \eqref{eq:U_0-error-asymp} follows from
$$
\beta^*\pi_L^*(\Delta_g+x^2W-1)\in \Diffb^2(\Bn\times_0\Bn)
$$
with
$$
\beta^*\pi_L^*\Big((\Delta_g+x^2W-1)-(\Delta_{g_0}-1)\Big)\in R^2\rho_L^2\Diffb^2(\Bn\times_0\Bn);
$$
here $\pi_L$ is added to emphasize the lift is that of the differential operator acting on the
left factor; lifting the operator on the right factor results in an `error'
$R^2\rho_R^2\Diffb^2(\Bn\times_0\Bn)$. These two in turn follow immediately from
the form of the metric, namely $g_0-g_\ep\in x^2\CI(X;\zT^*X\otimes \zT^*X)$.
This completes the proof of \eqref{eq:U_0-U_1-asymp}, and also yields that, with
$n+1=3$,
\begin{gather}
\beta^*\left((\Delta_g+x^2W-1) U_1 \right)\in R^2\rho_L^{n/2+2}\rho_R^{n/2}\CI(\Bn\times_0\Bn\setminus \Diag_0).\label{boundrest}
\end{gather}
Therefore, in the case $n+1=3,$ we have proved the following
\begin{thm}\label{3Dparametrix-thm} There exists a pseudodifferential operator, $G(h,\sigma),$ $\sigma\not=0,$ whose kernel is of the form
\begin{gather*}
G(h,\sigma,z,z') = e^{-i\frac{\sigma}{h}}h^{-2}\left(U_0(h,\sigma,z,z')+ h U_1(h,\sigma,z,z') \right)
\end{gather*}
with $U_0$ and $U_1$ satisfying
\eqref{eq:U_0-U_1-asymp} and such that the error
$E(h,\sigma)= P(h,\sigma)G(h,\sigma)-\Id$ is given by \eqref{defE} and satisfies
\eqref{boundrest}.
\end{thm}
\section{The structure of the semiclassical resolvent }\label{resolvent-structure}\label{full-sc-parametrix}
In this section we construct the general right semiclassical parametrix $G(h,\sigma)$ for the resolvent. We will prove the following
\begin{thm} \label{nD-parametrix} There exists a pseudodifferential operator $G(h,\sigma),$ $\sigma\not=0,$ such that its kernel is of the form
\begin{gather}
G(h,\sigma,z,z')=e^{-i \frac{\sigma r}{h}} U(h,\sigma,z,z'),
\ U\in\Psi_{0,\semi}^{-2,\frac{n}{2},-\frac{n}{2}-1,\frac{n}{2}}(\Bn). \label{kernelofg}
\end{gather}
and that, using the notation of section \ref{semiclassical-double-space},
\begin{gather*}
P(h,\sigma) G(h,\sigma)-\Id \in \rho_{\mcf}^\infty \rho_{\mathcal {S}}^\infty \Psi_{0,\semi}^{-\infty, \infty,\infty,\frac{n}{2} +i\frac{\sigma}{h}}(\Bn).
\end{gather*}
\end{thm}
Now, $r^2$ is a $\CI$ function on the zero double space away from the
left and right faces and in a quadratic
sense it defines the diagonal non-degenerately; $r$ has
an additional singularity at the zero-diagonal. Correspondingly
$\left(\frac{\sigma r}{h}\right)^2$ is $\CI$ on $M_{0,\semi}$ away from $\mcl$, $\mcr$
and $\mathcal {A}$ and
defines the lifted diagonal non-degenerately in a quadratic sense;
$\frac{\sigma r}{h}$ has an additional singularity at the zero diagonal. In particular,
$e^{-i \frac{\sigma r}{h}}$ is $\CI$ on $M_{0,\semi}$ away from $\mcl$, $\mcr$,
$\mathcal {A}$ and the lifted diagonal, $\Diag_{\semi}$; at $\Diag_{\semi}$
it has the form of $1$ plus
a continuous conormal function vanishing there. Thus, its presence in any
compact subset of $M_{0,\semi}\setminus(\mcl\cup\mcr\cup\mathcal {A})$ is not
only artificial, but introduces an irrelevant singularity at $\Diag_{\semi}$, so
it is better to think of $G$ as
\begin{equation*}\begin{split}
G(h,\sigma,z,z')&=G'+e^{-i \frac{\sigma r}{h}} U'(h,\sigma,z,z'),\\
&G'\in\Psi_{0,\semi}^{-2},\ U'\in\Psi_{0,\semi}^{-\infty,\frac{n}{2},-\frac{n}{2}-1,\frac{n}{2}},
\end{split}\end{equation*}
where $G'$ is supported near $\Diag_{\semi}$ (i.e.\ its support intersects only the
boundary hypersurfaces $\mathcal {S}$ and $\mcf$, but not the other boundary hypersurfaces),
and $U'$ vanishes near $\Diag_{\semi}$.
Indeed,
the first step in the construction of $G$ is to construct a piece of $G'$, namely to
find $G_0\in\Psi_{0,\semi}^{-2}$ such that $E_0=P(h,\sigma)G_0(h,\sigma)-\Id$
has no
singularity at the lifted diagonal, $\Diag_{\semi},$ with $G_0$ (hence $E_0$)
supported in a neighborhood of $\Diag_{\semi}$ in $M_{0,\semi}$ that only intersects the boundary of $M_{0,\semi}$ at $\mathcal {S}$ and $\mcf$. Thus,
\begin{equation}
P(h,\sigma)G_0-\Id= E_0, \;\ \text{ with } E_0 \in \Psi_{0,\semi}^{-\infty}(\Bn). \label{consg0}
\end{equation}
Since $P(h,\sigma)$ is elliptic in the interior of $\Bn,$ the construction of $G_0$ in the interior follows from the standard Hadamard parametrix construction. We want to do this construction uniformly up to the zero front face and the semiclassical front face in the blown-up manifold $M_{0,\semi}.$ We notice that the lifted diagonal intersects the boundary of $M_{0,\semi}$ transversally at the zero and semiclassical front faces. Since $\Diag_{\semi}$ intersects the boundary of $M_{0,\semi}$ transversally at $\mathcal {S}$ and $\mcf,$ we proceed as in
\cite{Mazzeo-Melrose:Meromorphic}, and extend $\Diag_{\semi}$ across
the boundary of $M_{0,\semi},$ and want to extend the Hadamard parametrix construction across the boundary as well. To do that we have to make sure the operator $P(h,\sigma)$ lifts to be uniformly transversally
elliptic at the lift of $\Diag_{\semi}$ up to the boundary of $M_{0,\semi},$ i.e.\ it is
elliptic on the conormal bundle of this lift, and thus
it can be extended as a transversally (to the extension of the lifted diagonal)
elliptic operator across the boundaries of $M_{0,\semi}.$ Note that up to
$\mathcal {S}\setminus\mcf$, this is the standard semiclassical elliptic
parametrix construction,
while up to $\mcf\setminus\mathcal {S}$, this is the first step in the conformally compact
elliptic
parametrix construction of Mazzeo and the first author \cite{Mazzeo-Melrose:Meromorphic},
so the claim here is that these constructions are compatible with each other and extend
smoothly to the corner $\mathcal {S}\cap\mcf$ near $\Diag_{\semi}$.
To see the claimed ellipticity, and facilitate further calculations, we remark
that one can choose a defining function of the boundary $x$ such that the metric $g$ can be written in the form
\begin{equation*}
g=\frac{dx^2}{x^2} + \frac{H_\eps(x,\omega)}{x^2},
\end{equation*}
where $H_\eps$ is one-parameter family of $\CI$ metrics on $\ms^n.$ In these coordinates the operator $P(h,\sigma)$ is given by
\begin{equation}
P(h,\sigma)= h^2\left( -(x\p_x)^2+ n x\p_n + x^2 A(x,\omega) \p_x+ x^2 \Delta_{H_\eps(x,\omega)}+ x^2 W -\frac{n^2}{4}\right) -\sigma^2. \label{formulaofp}
\end{equation}
We then conjugate $P(h,\sigma)$ by $x^{\frac{n}{2}},$ and we obtain
\begin{equation}
Q(h,\sigma)= x^{-\frac{n}{2}} P(h,\sigma) x^{\frac{n}{2}}= h^2\left( -(x\p_x)^2 + x^2 A \p_x + x^2 \Delta_{H_\eps} + x^2 B\right) - \sigma^2,
\end{equation}
where $B=-\frac{n}{2}A+W.$ To analyze the lift of $Q(h,\sigma)$ under $\beta_{\semi}$ we work in projective coordinates for the blow-down map. We denote the coordinates on the left factor of $\BB^{n+1}$ by $(x,\omega),$ while the coordinates on the right factor will be denoted by $(x',\omega').$ Then
we define projective coordinates
\begin{equation}
x'=\rho, \;\ X=\frac{x}{x'}, \;\ Y=\frac{\omega-\omega'}{x'}, \label{projectivecoord}
\end{equation}
which hold away from the left face. The front face is given by $\mcf=\{\rho=0\}$ and the
lift of the diagonal is $\Diag_0=\{X=1, Y=0\}.$ The lift of $Q(h,\sigma)$ under the zero blow-down map $\beta$ is equal to
\begin{gather*}
Q_0(h,\sigma)=\beta^*Q(h,\sigma)= \\ h^2\left( -(X\p_X)^2 + X^2A \rho \p_X+ X^2 \Delta_{H_\eps(\rho X,\omega'+\rho Y)}(D_Y)
-X^2\rho^2B \right) -\sigma^2.
\end{gather*}
In this notation, the coefficients of $\Delta_{H_\eps(\rho X,\omega'+\rho Y)}(D_Y)$ depend on
$\rho,\omega'$ and $Y,$ but the derivatives are in $Y.$ This operator is transversally
elliptic in a neighborhood of $\{X=1,\ Y=0\}$, away from $h=0$.
The restriction of the lift of $Q_0(h,\sigma)$ to the front face $\mcf=\{\rho=0\},$ is given by
\begin{gather*}
N_\mcf (Q_0(h,\sigma))=h^2( -(X\p_X)^2 + X^2\Delta_{H_\eps(0,\omega')}(D_Y))-\sigma^2.
\end{gather*}
As in \cite{Mazzeo-Melrose:Meromorphic}, $N_{\mcf}(Q_0(h,\sigma))$
will be called the normal operator of $Q_0(h,\sigma)$ at the zero front face $\mcf.$ Notice that it can be identified with Laplacian with respect to the hyperbolic metric on the half-plane $\{X>0, Y\in \RR^n\}$ with metric $X^{-2}(dX^2+ H_\eps(0,\omega'))$ conjugated by $X^{\frac{n}{2}}.$
Now we blow-up the intersection of $\Diag_0\times [0,h)$ with $h=0.$ We define projective coordinates
\begin{equation}\label{eq:X-semi-def}
X_{\hbar}= \frac{X-1}{h}, \;\ \;\ Y_{\hbar}=\frac{Y}{h}.
\end{equation}
The lift of $Q(h,\sigma)$ under the semiclassical blow-down map $\beta_{\semi}$ is given in these coordinates by
\begin{equation*}\begin{split}
Q_{\semi}&=\beta_{\semi}^*Q(h,\sigma)\\
&= -((1+h X_{\hbar}) \p_{X_{\hbar}})^2 + (1+h X_{\hbar})^2 \Delta_{H_\eps}(D_{Y_{\hbar}})\\
&\qquad\qquad- (1+hX_{\hbar})^2 A \rho h \p_{X_{\hbar}}+ h^2\rho^2 (1+h X_{\hbar})^2 B -\sigma^2,
\end{split}\end{equation*}
where $H_\eps=H_\eps(\rho(1+h X_{\hbar}), \omega'+\rho h Y_{\hbar}).$ This operator is transversally elliptic to $\{X_{\hbar}=0,\ Y_{\hbar=0}\}$ near $h=0.$
The restriction of the lift of $Q_{\semi}$ to the semiclassical face, $\mathcal {S}=\{h=0\}$ will be called the normal operator of $Q_{\semi}$ at the semiclassical face, and is equal to
\begin{equation}
N_\mathcal {S}(Q_{\semi})=-\p_{X_{\hbar}}^2 +\Delta_{H_\eps(\rho,\omega')}(D_{Y_{\hbar}})-\sigma^2. \label{norm-op-semic-face}
\end{equation}
This is a family of differential operators on $\RR^{n+1}_{X_{\hbar},Y_{\hbar}}$
depending on $\rho$ and $\omega',$ and for each $\omega'$ and $\rho$ fixed, $N_\mathcal {S}(Q_{\semi})+\sigma^2$
is the Laplacian with respect to the metric
$$
\delta_{\semi}=dX_{\hbar}^2+ \sum H_{\eps,ij}(\rho,\omega) dY_{\hbar,i} dY_{\hbar,j},
$$
which is isometric to the Euclidean metric under a linear change of variables
$(X_{\hbar},Y_{\hbar})$ for fixed $(\rho,\omega')$, and the change of variables can
be done smoothly in $(\rho,\omega')$. Note that the fibers of the semiclassical
blow-down map $\beta_{\semi}$ on $\cS$ are given exactly by $(\rho,\omega')$
fixed.
Therefore, the operator $Q(h,\sigma)$ lifts under $\beta_{\semi}$ to an operator $Q_{\semi}$ which is elliptic in a neighborhood of the lifted diagonal uniformly up to the zero front face and the semiclassical front face. Since the diagonal meets the two faces transversally, one can extend it to a neighborhood of $\mcf$ and $\mathcal {S}$ in the double of the manifold $M_{0,\semi}$ across $\mathcal {S}$ and $\mcf,$ and one can also extend the operator $Q_{\semi}$ to be elliptic in that neighborhood. Now, using standard elliptic theory
(or, put somewhat differently, the standard theory of conormal distributions to an
embedded submanifold without boundary, in this case the extension of
the diagonal), one can find a $G_0\in \Psi_{0,\semi}^{-2}(\Bn)$ whose Schwartz kernel
lifts to a distribution supported in a neighborhood of $\Diag_{\semi}$ such that
\begin{equation}
Q(h,\sigma) G_0-\Id= E_0 \in \Psi_{0,\semi}^{-\infty}(\Bn). \label{term0}
\end{equation}
Next we will remove the error at the semiclassical front face. We will find an operator
$$
G_1= G'_1+e^{-i\frac{\sigma}{h}r }U'_1,
\ G'_1\in\Psi_{0,\semi}^{-\infty}(\Bn),
\ U'_1\in \Psi_{0,\semi}^{-\infty,\infty,-\frac{n}{2}-1,\infty}(\Bn)
$$
such that
$G'_1$ is supported near $\Diag_{\semi}$ while $U'_1$ is supported away from it, and
\begin{gather}
\begin{gathered}
Q(h,\sigma)G_1-E_0=E_1, \;\\
E_1=E'_1+ e^{-i \frac{\sigma}{h} r} F_1, \;\ E'_1\in\rho_S^\infty\Psi_{0,\semi},\ \ F_1
\in \rho_S^\infty \Psi_{0,\semi}^{-\infty,\infty,-\frac{n}{2}-1,\infty} \\
\text{ and with } \beta_{\semi}^* K_{E_1}\text{ supported away from} \ \mcl,\ \mcr,
\end{gathered}\label{term1n}
\end{gather}
and $K_{E'_1}$, resp.\ $K_{F_1}$ supported near, resp.\ away from, $\Diag_{\semi}$.
In other words, the the error term $E_1$ is such that the kernel of
$E'_1$ vanishes to infinite order at all boundary faces (hence from now on we can
regard it as trivial and ignore it), while
the kernel of $F_1$ lifts to a $\CI$ function which is supported near $\mathcal {S}$ (and in particular vanishes to infinite order at the right and left faces), vanishes to infinite order at the semiclassical front face, and also vanishes to order $\frac{n}{2}$ at the boundary face $\mathcal {A}.$
We will use the facts discussed above about the normal operator at the semiclassical face, $N_{\mathcal {S}}(Q_{\semi}).$
Notice that $\mathcal {S},$ the semiclassical front face, is itself a $\CI$ manifold with boundary which intersects the zero front face, $\mcf,$ transversally, and therefore it can be extended across $\mcf.$ Similarly, the operator $N_{\mathcal {S}}(Q_{\semi})$ can be extended to an elliptic operator across $\mcf.$
We deduce from \eqref{norm-op-semic-face} that for each $\rho$ and $\omega'$ fixed, and for $\im\sigma<0$, the inverse of
$N_{\mathcal {S}}(Q_{\semi})$ is essentially the resolvent of the Euclidean Laplacian at energy $\sigma^2,$ pulled back by the linear change of variables corresponding to
$H_{\eps}(\rho,\omega');$ for $\im\sigma\geq 0$ we use the analytic continuation
of the resolvent from $\im\sigma<0$.
Here is where we need to make a choice corresponding to
the analytic continuation of the resolvent of $P(h,\sigma)$
we wish to construct, i.e.\ whether
we proceed from $\im\sigma>0$ or $\im\sigma<0$; we need to make the
corresponding choice for the Euclidean resolvent.
Let $R_0$ denote the analytic continuation of the inverse $L^2\to H^2$
of the {\em family} (depending on $\rho,\omega'$)
$N_{\mathcal {S}}(Q_{\semi})$ from $\im\sigma<0$; it is thus (essentially, up to a linear change
of coordinates, depending smoothly on $\rho,\omega'$)
the analytic continuation of the resolvent of the Euclidean
Laplacian.
Since we are working with the analytic continuation of the resolvent, it is not
automatic that one can solve away exponentially growing errors which arise
in the construction below (i.e.\ that one can apply $R_0$ iteratively to errors that arise),
and thus it is convenient to make the following construction quite explicit order in $h$ we are merely in the
`limiting absorption principle' regime (i.e.\ with real spectral parameter),
thus the construction below is actually stronger than what is needed below.
Moreover, from this point of view the construction can be interpreted as
an extension of the semiclassical version of the intersecting
Lagrangian construction of \cite{Melrose-Uhlmann:Intersection}
extended to the 0-double space; from this perspective the method we present
is very `down to earth'.
Via the use of a partition of unity, we may assume that there is a coordinate patch
$U$ in $\Bn$ (on which the coordinates are denoted by $z$)
such that $E_0$ is supported in
$\beta_{\semi}^{-1}(U\times U\times [0,1))$. Note that coordinate charts of this form
cover a neighborhood of $\mathcal {S}$,
so in particular $E_0$ is in
$\dCI(\Bn\times_0\Bn\times[0,1))$ outside these charts, hence can already be regarded
as part of the final error term and we can ignore these parts henceforth.
Now, near $\mathcal {S}$,
$\beta_{\semi}^{-1}(U\times U\times [0,1))$ has a product
structure $\overline{\zT}U\times [0,\delta_0)= U\times\Bn\times[0,\delta_0)$,
where $\overline{\zT}U$
denotes the fiber-compactified zero tangent bundle, and $[0,\delta_0)$ corresponds
to the boundary defining function $\rho_S$. Indeed, the normal bundle of
$\Diag_0$ in $\Bn\times_0\Bn$ can be identified with $\zT\Bn$, via lifting 0-vector
fields from on $\Bn\times_0\Bn$ via the left projection, which are transversal
to $\Diag_0$, hence
the interior of the inward pointing spherical normal bundle of $\Diag_0\times\{0\}$
in $M_0=\Bn\times_0\Bn\times[0,1)$ can be identified with $\zT\Bn$,
while the inward pointing spherical normal bundle itself can be identified
with the radial compactification of $\zT\Bn$.
However, it is fruitful to choose the identification in a particularly convenient form
locally. Namely, away from
$\pa\Diag_0\times\{0\}$, coordinates $z$ on $U$ give coordinates
$$
z',Z_{\semi}=(z-z')/h,\rho_S
$$
near the interior of the front face $\mathcal {S}$
(here $Z_{\semi}$ is the coordinate
on the fiber of $T_U\Bn$ over $z'$), while near $\pa\Diag_0\times\{0\}$,
$(x',\omega',X_{\semi},Y_{\semi},\rho_S)$ (see \eqref{eq:X-semi-def}) are coordinates
near the interior of $\mathcal {S}$ (now $(X_{\semi},Y_{\semi})$ are the coordinates
on the fiber of $\zT_U\Bn$ over $(x',\omega')$). To obtain coordinates valid
near the corner, one simply
needs to radially compactify the fibers of $\zT_U\Bn$, i.e.\ replace the linear
coordinates $Z_{\semi}$, resp.\ $(X_{\semi},Y_{\semi})$ by radially compactified versions
such as $|Z_{\semi}|$ and $\hat Z_\semi=Z_{\semi}/|Z_{\semi}|\in \bbS^n$ in the former case.
Moreover, if a function, such as $E_0$,
is supported away from $\mathcal {A}$, then its support is compact in the {\em interior}
of the fibers $\Bn$ of the fiber-compactified tangent space.
Now, the interior of $\Bn$ is a vector space, $T_p U$, $p\in U$, and in particular
one can talk about fiberwise polynomials.
Over compact subsets of the fibers, the boundary defining function $\rho_S$ is
equivalent to $h$, and indeed we may choose boundary defining functions $\rho_A$
and $\rho_S$ such that
$$
h=\rho_A\rho_S.
$$
Note that $\rho_A$ is thus a boundary defining function of the compactified fibers
of the tangent bundle; it is convenient to make a canonical choice using
the metric $g_\ep$, which is an inner product on $\zT_p U$, hence a translation
invariant metric on the fibers of $\zT U$, namely to make the defining function
$\rho_A$ the reciprocal of the distance function from the zero section (i.e.\ the
diagonal under the identification), smoothed out at the zero section.
In particular, if $U$ is a coordinate chart near $\pa \Bn$ then
$\rho_{\mathcal {A}}=\left((X_1)^2+ |Y_1|_{H_\eps}^2\right)^{-\frac{1}{2}}.$
This is indeed consistent with our previous calculations since
$$
\rho_{\mathcal {A}}=h\left( (X-1)^2+|Y|^2\right)^{-\ha}
$$
and therefore it is, away from $\Diag_{\semi},$ a defining function of the semiclassical face $\mathcal {A}.$
If $v\in\CI(M_{0,\semi})$, then expanding $v$ in Taylor series around $\mathcal {S}$ up to
order $N$, we have
$$
v=\sum_{k\leq N} \rho_S^k v'_k+v',\ v'_k\in\CI(\overline{T}U),
\ v'\in\rho_S^{N+1}\CI(M_{0,\semi}).
$$
In terms of the local coordinates valid near the corner $\mathcal {S}\cap\mathcal {A}$ over an
interior coordinate chart $U$,
\begin{gather*}
\rho_S=|z-z'|,\ \frac{z-z'}{|z-z'|},\ \rho_A=\frac{h}{|z-z'|},\ z',
\end{gather*}
$v_k$ is a $\CI$ function of $|z-z'|,\ \frac{z-z'}{|z-z'|},\ z'$. It is convenient to
rewrite this as
\begin{equation}\label{eq:mod-TS-at-mcs}
v=\sum_{k\leq N} h^k |z-z'|^{-k} v'_k+v'=\sum_{k\leq N} h^k v_k+v',
\ v_k\in \rho_A^{-k}\CI(\overline{T}U),
\ v'\in\rho_S^{N+1}\CI(M_{0,\semi}),
\end{equation}
for the reason that the vector fields $D_{z_j}$ are tangent to the fibers given by
constant $h$,
i.e.\ commute with multiplication by $h$. One can rewrite $v$ completely
analogously,
$$
v=\sum_{k\leq N} h^k v_k+v',
\ v_k\in \rho_A^{-k}\CI(\overline{T}U),
\ v'\in\rho_S^{N+1}\CI(M_{0,\semi}),
$$
for coordinate charts $U$ at $\pa\Bn$.
In addition, if
$a\in\CI(\Bn\times_0\Bn\times[0,1))$, then expanding $a$ in Taylor series
around $\Diag_0\times\{0\}$, shows that for any $N$,
modulo $\rho_S^{N+1}\CI(U\times\Bn\times[0,\delta_0))$, it is of the form
\begin{gather*}
\sum_{|\alpha|+k\leq N}
a_{\alpha,k}(z') (z-z')^\alpha h^k
=\sum_{|\alpha|+k\leq N} a_{\alpha,k}(z') Z_\semi^\alpha h^{k+|\alpha|}\\
=\sum_{|\alpha|+k\leq N} a_{\alpha,k}(z') \hat Z_\semi^\alpha \rho_S^{k+|\alpha|}\rho_A^k,
\end{gather*}
where $\hat Z_\semi^\alpha=Z_\semi^{\alpha}/|Z_\semi|^{|\alpha|}$ (except near the zero section) is
$\CI$ on $\Bn$. While the last expression is the most geometric way of encoding
the asymptotics at $\pa\Bn$, it is helpful to take advantage of the stronger statement
on the previous line, which shows that the coefficients are polynomials in the fibers,
of degree $\leq N$.
The vector fields $hD_z$, resp.\ $hD_X$ and $hD_Y$, acting on a modified
Taylor series as in \eqref{eq:mod-TS-at-mcs}, become
$D_{Z_{\semi}}$, resp.\ $D_{X_{\semi}}$ and $D_{Y_{\semi}}$, i.e.\ act on the coefficients
$v_k$ only (and on $v'$, of course) so we obtain that,
modulo coefficients in
$\rho_S^{N+1}\CI(U\times\Bn\times[0,\delta_0))$, $P(h,\sigma)$ lifts to a differential
operator with polynomial coefficients on the fibers, depending smoothly on the
base variables, i.e.\ an operator of the form
$$
\sum_{|\alpha|+k\leq N,|\beta|\leq 2} a_{\alpha,k,\beta}(z') Z_\semi^\alpha h^{k+|\alpha|} D_{Z_{\semi}}^\beta,
$$
with an analogous expression in the $(X_{\semi},Y_{\semi})$ variables. Here the
leading term in $h$, corresponding to $h^0$, is $N_{\mathcal {S}}(Q_{\semi})=
\Delta_{g_{\ep}(x',\omega')}-\sigma^2$, i.e.\ we have
\begin{gather*}
P(h,\sigma)h^mv_m\\
=h^m(\Delta_{g_{\ep}(x',\omega')}-\sigma^2)v_m+
h^m\sum_{0<|\alpha|+k\leq N,|\beta|\leq 2} a_{\alpha,k,\beta}(z') Z_\semi^\alpha h^{k+|\alpha|} D_{Z_{\semi}}^\beta v_m.
\end{gather*}
Thus, one can iteratively solve away $E_0$ as follows. Write
$$
E_0=h^{-n-1}\sum h^mE_{0,m}
$$
as in \eqref{eq:mod-TS-at-mcs}, and note that each $E_{0,m}$ is compactly supported.
Let
$$
G_{1,0,m}=R_0 E_{0,m},
$$
so
\begin{gather*}
P(h,\sigma)h^{m-n-1} G_{1,0,m}\\
=h^mE_{0,m}+h^{m-n-1}\sum_{0<|\alpha|+k\leq N,|\beta|\leq 2} a_{\alpha,k,\beta}(z') Z_\semi^\alpha h^{k+|\alpha|-n-1} D_{Z_{\semi}}^\beta R_0 E_{0,m},
\end{gather*}
and thus we have replaced the error $h^mE_{0,m}$ by an error of the form
\begin{gather*}
\sum_{0<|\alpha|+k\leq N,|\beta|\leq 2} a_{\alpha,k,\beta}(z') Z_\semi^\alpha h^{m+k+|\alpha|-(n+1)} D_{Z_{\semi}}^\beta R_0 E_{0,m}\\
=\sum_{1\leq\ell\leq N} h^{m+\ell-(n+1)}
\sum_{|\alpha|\leq \ell}\sum_{|\beta|\leq 2} a_{\alpha,\ell-|\alpha|,\beta}(z') Z_\semi^\alpha D_{Z_{\semi}}^\beta R_0 E_{0,m}\\
=\sum_{1\leq\ell\leq N} h^{m+\ell-(n+1)} L_\ell R_0 E_{0,m}
\end{gather*}
which has the feature that not only does it vanish to (at least)
one order higher in $h$, but the $h^{m+\ell-(n+1)}$ term is given by
a differential operator $L_\ell$ with polynomial coefficients of degree $\leq \ell$
applied to $R_0E_{0,m}$,
with $E_{0,m}$ compactly supported. As the next lemma states,
one can apply $R_0$ to an expression
of the form $L_\ell R_0 E_{0,m}$, and thus iterate the construction.
\begin{lemma}
Suppose $M_j$, $j=1,2,\ldots,N$,
are differential operators with polynomial coefficients of degree $m_j$
on $\RR_w^{n+1}$. Then
$$
R_0 M_1 R_0 M_2\ldots R_0 M_N R_0:\CI_c(\RR^{n+1})\to\CmI(\RR^{n+1})
$$
has an analytic extension from $\im\sigma<0$ to $\re\sigma>1$,
$\im\sigma\in\bbR$, and
$$
R_0 M_1 R_0 M_2\ldots R_0 M_N R_0:\CI_c(\RR^{n+1})\to e^{-i\sigma\langle w\rangle}
\langle w\rangle^{-n/2+N+m}\CI(\Bn),
$$
with $m=\sum m_j$.
\end{lemma}
\begin{proof}
For $\im\sigma<0$, $D_{w_k}$ commutes with $\Delta-\sigma^2$, hence
with $R_0$, while commuting $D_{w_k}$ through a polynomial gives rise to a
polynomial of lower order, so we can move all derivatives to the right, and also assume
that $M_j=w^{\alpha^{(j)}}$, $\alpha^{(j)}\in\Nat^{n+1}$,
$|\alpha^{(j)}|\leq m_j$, so we are reduced to
examining the operator
$$
R_0 w^{\alpha^{(1)}} R_0 w^{\alpha^{(2)}}\ldots R_0 w^{\alpha^{(N)}} R_0.
$$
It is convenient to work in the Fourier transform representation.
Denoting the dual variable of $w$ by $\zeta$;
for $\im\sigma<0$,
$R_0$ is multiplication by $(|\zeta|^2-\sigma^2)^{-1}$, while $w^{\alpha^{(j)}}$ is the
operator $(-D_\zeta)^{\alpha^{(j)}}$. Rewriting
$\cF (R_0 w^{\alpha^{(1)}} R_0 w^{\alpha^{(2)}}\ldots R_0 w^{\alpha^{(N)}} R_0 f)$, the product rule
thus gives an expression of the form
\begin{gather*}
\sum_{|\beta|\leq \alpha^{(1)}+\ldots+\alpha^{(N)}}
(|\zeta|^2-\sigma^2)^{-(N+1+|\alpha^{(1)}|+\ldots+|\alpha^{(N)}|)} Q_{\alpha,N,\beta}(\zeta)
(-D_\zeta)^\beta
\Fr f\\
=\sum_{|\beta|\leq \alpha^{(1)}+\ldots+\alpha^{(N)}}
(|\zeta|^2-\sigma^2)^{-(N+1+|\alpha^{(1)}|+\ldots+|\alpha^{(N)}|)} \Fr Q_{\alpha,N,\beta}(D_{w})
w^\beta f,
\end{gather*}
where $Q_{\alpha,N,\beta}$ is a polynomial in $\zeta$. Since we are considering
compactly supported $f$, the differential operator $Q_{\alpha,N,\beta}(D_w)
w^\beta$ is harmless, and we only need to consider $R_0^{N+1+m}$ applied
to compactly supported functions. This can be further rewritten as
a constant multiple of $\pa_\sigma^{N+m} R_0$, so the well-known results
for the analytic continuation of the Euclidean resolvent yield the stated
analytic continuation and the form of the result; see
Proposition 1.1 of \cite{Melrose: Geometric Scattering}.
\end{proof}
Applying the lemma iteratively, we construct
$$
\tilde G_{1,m}=e^{-i\sigma\langle Z_\semi\rangle}
\langle Z_\semi\rangle^{-n/2+N+m}G'_{1,m},\ G'_{1,m}\in\CI(U\times\Bn),
$$
such that
$$
P(h,\sigma) \sum_{m=0}^\infty
h^{m-(n+1)} \tilde G_{1,m}\sim\sum_{m=0}^\infty h^{m-(n+1)} E_{0,m},
$$
where the series are understood as formal series (i.e.\ this is a statement of the
equality of coefficients). Borel summing $\langle z\rangle^{m}h^m G'_{1,m}
=\rho_S^m G'_{1,m}$, and obtaining $G_1\in\CI(U\times\Bn\times[0,\delta_0))$
as the result, which we may arrange to be supported where $\rho_S$ is small,
we deduce that \eqref{term1n} holds.
The next step is to remove the error at the semiclassical face $\mathcal {A}.$ We want to construct
$$
G_2=e^{-i\frac{\sigma}{h} r} U_2,\ U_2\in\rho_S^\infty\Psi_{0,\semi}^{-\infty,\frac{n}{2},
-\frac{n}{2}-1,\frac{n}{2}},
$$
such that
\begin{gather}
\begin{gathered}
P(h,\sigma) G_2-e^{-i \frac{\sigma}{h} r} F_1=E_2, \;\ E_2=e^{-i\frac{\sigma}{h} r} F_2, \;\\
\ F_2 \in h^\infty\rho_L^{n/2}\rho_R^{n/2}\CI(\Bn\times_0\Bn\times[0,1))=\rho_{\mathcal {S}}^\infty \Psi_{0,\semi}^{-\infty,\frac{n}{2}, \infty, \frac{n}{2}}.
\end{gathered}
\label{term2}
\end{gather}
In other words, we want the error to vanish to infinite order at the semiclassical face $\mathcal {A}$ and at the semiclassical front face $\mathcal {S},$ and to vanish to order
$\frac{n}{2}$ at the left and right faces; the infinite order vanishing
at $\mathcal {S}$ means that $\mathcal {S}$ can be blown-down (i.e.\ does not need to be
blown up), which, together with $h$ being the joint defining function
of $\mathcal {A}$ and $\mathcal {S}$, explains the equality of the indicated two spaces.
This construction is almost identical to the one carried out in section \ref{3D-parametrix}.
We begin by observing that the semiclassical face $\mathcal {A}$ consists of the stretched product $\Bn\times_0 \Bn$ with $\Diag_0$ blown-up, which is exactly the manifold $\Bn\times_1 \Bn$ defined in section \ref{3D-parametrix}. Moreover, as $F_1$ vanishes to infinite order at $\mathcal {S}$, see
\eqref{term1n}, the latter
can be blown down, i.e.\ $F_1$ can be regarded as being of the form
\begin{gather}
\begin{gathered}
F_1 \in h^{-n/2-1}\CI(\Bn\times_0\Bn\times[0,1)),
\ \text{ with } K_{F_1}\text{ supported away from} \ \mcl,\ \mcr,\\
\text{ and vanishing to infinite order at}\ \Diag_0\times[0,1).
\end{gathered}
\end{gather}
Now, $F_1$ has an asymptotic expansion at the boundary face $h=0$ of the form
\begin{gather*}
F_1\sim h^{-\frac{n}{2}-1} \sum_{j=0}^\infty h^j F_{1,j}, \;\ F_{1,j} \in \CI(\Bn\times_0\Bn), \\ F_{1,j} \text{ vanishing to infinite order at } \Diag_0\times[0,1),
\ \text{ supported near } \Diag_0\times[0,1).
\end{gather*}
So we think of $F_1$ as an element of $\Bn\times_1 \Bn \times [0,1),$ where the blow-up $\Bn\times_1 \Bn,$ was defined in section \ref{3D-parametrix},
see figure \ref{fig2}, with an expansion
\begin{gather*}
F_1\sim h^{-\frac{n}{2}-1} \sum_{j=0}^\infty h^j F_{1,j},\\
F_{1,j}\in \CI(\Bn\times_1 \Bn), \text{ vanishing to infinite order at } D.
\end{gather*}
So we seek $U_2 \sim h^{\frac{n}{2}} \sum_{j} h^j U_{2,j}$ with $U_{2,j}$ vanishing
to infinite order at $D,$ such that
\begin{gather*}
P(h,\sigma) e^{-i\frac{\sigma}{h} r} U_2-e^{-i \frac{\sigma}{h} r} F_1=e^{-i \frac{\sigma}{h} r} R, \\ R \in h^\infty \CI(\Bn\times_1 \Bn), \text{ vanishing to infinite order at } D.
\end{gather*}
Matching the coefficients of the expansions we get the following set of transport equations
\begin{gather*}
2i\sigma |g|^{-\oq} \p_r (|g|^{\oq} U_{2,0})=-F_{1,0}, \text{ if } r>0, \\
U_{2,0}=0 \text{ at } r=0,
\end{gather*}
and for $j\geq 1,$
\begin{gather*}
2i\sigma |g|^{-\oq} \p_r (|g|^{\oq} U_{2,j})= (\Delta+x^2 W-\frac{n^2}{4})U_{2,j-1}- F_{1,j}, \text{ if } r>0, \\
U_{2,j}=0 \text{ at } r=0.
\end{gather*}
Notice that $F_{j,0}$ is compactly supported and, as seen in equations \eqref{eq:U_0-U_1-asymp}, $U_{2,0}\in \rho_L^{\frac{n}{2}} \rho_R^{\frac{n}{2}} \CI(\Bn\times_1 \Bn).$ Moreover, as in \eqref{eq:U_0-error-asymp}, one gets that
$(\Delta+x^2 W-\frac{n^2}{4})U_{2,0}\in R^2 \rho_{R}^{\frac{n}{2}} \rho_L^{\frac{n}{2}+2} \CI(\Bn\times_1 \Bn),$
thus one can solve the transport equation for
$U_{2,1},$ and gets that $U_{2,1} \in \rho_L^{\frac{n}{2}} \rho_R^{\frac{n}{2}} \CI(\Bn\times_1 \Bn).$
One obtains by using induction that $U_{2,j}\in \rho_L^{\frac{n}{2}} \rho_R^{\frac{n}{2}} \CI(\Bn\times_1 \Bn),$ and
$(\Delta+x^2 W-\frac{n^2}{4})U_{2,j}\in R^2 \rho_{L}^{\frac{n}{2}+2} \rho_R^{\frac{n}{2}} \CI(\Bn\times_1 \Bn),$
for all $j.$ Then one sums the series asymptotically using Borel's lemma. This gives $U_2$ and proves \eqref{term2}.
The last step in the parametrix construction is to remove the error at the zero front face. So far we have
\begin{gather*}
P(h,\sigma) \left(G_0-G_1+G_2\right)-\Id=E_2, \;\
\end{gather*}
with
$$
E_2=e^{-i\frac{\sigma}{h} r} F_2, \;\ F_2 \in \rho_{\mathcal {S}}^\infty \Psi_{0,\semi}^{-\infty,\frac{n}{2}+2, \infty, \frac{n}{2}}.
$$
Now we want to construct $G_3$ such that
\begin{gather}
P(h,\sigma) G_3-E_2=E_3 \in e^{-i\frac{\sigma}{h} r}\rho_{\mathcal {S}}^\infty\rho_{\mcf}^\infty \Psi_{0,\semi}^{-\infty,2+\frac{n}{2}, \infty, \frac{n}{2}}(\Bn). \label{term3}
\end{gather}
We recall from Proposition \ref{distance}, that $r=-\log(\rho_R \rho_L)+F,$ $F>0.$ So,
$$
\exp(-i\frac{\sigma}{h} r)=(\rho_R \rho_L)^{i\frac{\sigma}{h} r} \exp(-i \frac{\sigma}{h} F).
$$
Therefore
the error term $E_2$ in \eqref{term2} satisfies
\begin{gather*}
\widetilde{E_2}=\beta_{\semi}^* E_2 \in \rho_{\mathcal {S}}^\infty \rho_{\mathcal {A}}^\infty \rho_{R}^{\frac{n}{2}+i \frac{\sigma}{h}} \rho_{L}^{2+\frac{n}{2}+i \frac{\sigma}{h}} \CI(M_{0,\semi}).
\end{gather*}
We write
\begin{gather*}
\widetilde{E_2}\sim \sum_{j=0}^\infty \rho_{\mcf}^j E_{2,j}, \;\ E_{2,j} \in \rho_\mathcal {S}^\infty \rho_\mathcal {A}^\infty \rho_{R}^{\frac{n}{2}+i \frac{\sigma}{h}} \rho_{L}^{2+\frac{n}{2}+i \frac{\sigma}{h}} \CI(\mcf),
\end{gather*}
and we want to construct $G_3$ such that
\begin{gather*}
\beta_{\semi}^* (G_3)\sim \sum_{j=0}^\infty \rho_{\mcf}^j G_{3,j},
\end{gather*}
and that
\begin{gather*}
N_{\mcf}(Q_{\semi}) G_{3,j}= E_{2,j}, \;\
G_{3,j} \in \rho_\mathcal {S}^\infty \rho_\mathcal {A}^\infty \rho_{R}^{\frac{n}{2}+i \frac{\sigma}{h}} \rho_{L}^{\frac{n}{2}+i \frac{\sigma}{h}} \CI(\mcf).
\end{gather*}
The asymptotic behavior in $\rho_R$ and $\rho_L$ follows from an application of Proposition 6.15 of \cite{Mazzeo-Melrose:Meromorphic}, and the fact that
$N_\mcf(Q_{\semi})$ can be identified with the Laplacian on the hyperbolic space. Now we just have to make sure, this does not destroy the asymptotics at the faces
$\mathcal {S}$ and $\mathcal {A}.$ But one can follow exactly the same construction we have used above, now restricted to the zero front face instead of $M_{0,\semi}$ to construct
$G_{3,j}$ vanishing to infinite order at the faces $\mathcal {S}$ and $\mathcal {A}.$
This gives a parametrix, $\widetilde{G}=G_0-G_1+G_2-G_3 \in \Psi_{0,\semi}^{-2, \frac{n}{2}+i\frac{\sigma}{h},-\frac{n}{2}-1,\frac{n}{2} +i\frac{\sigma}{h}}$ that satisfies
\begin{gather}
P(h,\sigma) \widetilde{G}-\Id=R=E_3+E_3' \in \rho_{\mcf}^\infty \rho_{\mathcal {S}}^\infty \Psi_{0,\semi}^{-\infty, 2+\frac{n}{2}+i\frac{\sigma}{h},\infty,\frac{n}{2}+i\frac{\sigma}{h}}(\Bn). \label{weakparametrix}
\end{gather}
Since
\begin{gather*}
E_3'\in h^\infty\Psi_{0,\semi}^{-\infty},\\
E_3=e^{-i\frac{\sigma}{h} r} F_3, \;\ F_3 \in \rho_{\mathcal {S}}^\infty \rho_{\mcf}^\infty
\Psi_{0,\semi}^{-\infty,2+\frac{n}{2}, \infty, \frac{n}{2}},
\end{gather*}
and $E_3$ is supported away from $\Diag_\semi$. The last step in the construction is to remove the error term at the left face and it will be done using the indicial operator as in section 7 of
\cite{Mazzeo-Melrose:Meromorphic}. Since in the region near the left face is away from the semiclassical face, this is in fact the same construction as in \cite{Mazzeo-Melrose:Meromorphic}, but with the parameter $h.$ Using equation \eqref{formulaofp} and the projective coordinates \eqref{projectivecoord}, we find that the operator $P(h,\sigma)$ lifts to
\begin{gather*}
P_0(h,\sigma)=h^2( -(X\p_X)^2+nX\p_X+\rho A X^2\p_X + X^2 \Delta_{H_\eps} +\rho^2 X^2 W -\frac{n^2}{4})-\sigma^2.
\end{gather*}
In these coordinates the left face is given by $\{X=0\}.$ Therefore, the kernel of the composition $K(P(h,\sigma) \widetilde{G})$ when lifted to $\Bn\times_0 \Bn$ is, near the left face, equal to
\begin{gather*}
K(P(h,\sigma) \widetilde G)=\left( h^2( -(X\p_X)^2+nX\p_X -\frac{n^2}{4})-\sigma^2\right) K(\widetilde{G}) + O(X^2).
\end{gather*}
The operator $I(P(h,\sigma))= h^2( -(X\p_X)^2+nX\p_X -\frac{n^2}{4})-\sigma^2$ is called the indicial operator of $P(h,\sigma).$
Since $\widetilde{G} \in \Psi_{0,\semi}^{-2, \frac{n}{2}+i\frac{\sigma}{h},-\frac{n}{2}-1,\frac{n}{2} +i\frac{\sigma}{h}}(\Bn),$ then near the left face
$K(\widetilde{G})\in \mck^{\frac{n}{2}+i\frac{\sigma}{h},-\frac{n}{2}-1,\frac{n}{2} +i\frac{\sigma}{h}}(M_{0,\semi}).$
But $I(P(h,\sigma)) X^{\frac{n}{2}+i\frac{\sigma}{h}}=0.$ So we deduce that near $\mcl,$
$K(P(h,\sigma) \widetilde G)\in \mck^{\frac{n}{2}+i\frac{\sigma}{h}+1,-\frac{n}{2}-1,\frac{n}{2} +i\frac{\sigma}{h}}(M_{0,\semi}).$ That is, we gain one order of vanishing at the left face. Since we already know from
\eqref{weakparametrix} that the kernel of the error vanishes to infinite order $\mathcal {A}$ and at the front face $\mcf,$ and $x=R \rho_R,$ $x'=R\rho_L,$
the kernel of the error $R,$ on the manifold $\Bn\times \Bn$ satisfies
\begin{gather*}
K(R)\in h^\infty x^{\frac{n}{2}+i\frac{\sigma}{h}+1} {x'}^{\frac{n}{2}+i\frac{\sigma}{h}} \CI(\Bn\times \Bn).
\end{gather*}
Then one can use a power series argument to find $G_4$ with Schwartz kernel in
$h^\infty x^{\frac{n}{2}+i\frac{\sigma}{h}+1}{x'}^{\frac{n}{2}+i\frac{\sigma}{h}} \CI(\Bn \times \Bn)$ such that
\begin{gather*}
P(h,\sigma) G_4-R \in \Psi^{-\infty, \infty,\infty,\frac{n}{2}+i\frac{\sigma}{h}}(\Bn).
\end{gather*}
So $G=G_0-G_1+G_2-G_3-G_4 \in \Psi_{0,\semi}^{-2, \frac{n}{2}+i\frac{\sigma}{h},-\frac{n}{2}-1,\frac{n}{2} +i\frac{\sigma}{h}}$ is the desired parametrix.
\section{$L^2$-bounds for the semiclassical resolvent}
\label{sec:L2-bounds}
We now prove bounds for the semiclassical resolvent,
$R(h,\sigma)=P(h,\sigma)^{-1}$:
\begin{thm}\label{semiclassical-resolvent-bounds} Let $M>0,$ $h>0$ and $\sigma\in \mc$ be such that $\frac{\im \sigma}{h}<M.$ Let $a,b\geq\max\{ 0, \frac{\im\sigma}{h} \}.$ Then there exists $h_0>0$ and $C>0$ independent of $h$ such that for $h\in (0,h_0),$
\begin{gather}
||x^a R(h,\sigma) x^b f||_{L^2(\Bn)} \leq C h^{-1-\frac{n}{2}}||f||_{L^2(\Bn)}. \label{eq:newbounds-res}
\end{gather}
\end{thm}
As usual, we prove Theorem \ref{semiclassical-resolvent-bounds} by obtaining bounds for the
parametrix $G$ and its error $E$ on
weighted $L^2$ spaces. As a preliminary remark, we recall that elements of
$\Psi^0_{\semi}$ on compact manifolds without boundary are $L^2$-bounded with
an $h$-independent bound;
the same holds for elements of $\Psi^0_{0,\semi}$. Thus, the diagonal singularity
can always be ignored (though in our setting, due to negative orders of the operators
we are interested in, Schur's lemma gives this directly in any case).
The following lemma follows from the argument of Mazzeo \cite[Proof of
Theorem~3.25]{Mazzeo:Edge}, since for the $L^2$ bounds, the proof
in that paper only utilizes estimates on the Schwartz kernel, rather than its
derivatives. Alternatively, it can be proved using Schur's lemma if one
writes the Schwartz kernel relative to a b-density.
\begin{lemma}\label{schurs}
Suppose that the Schwartz kernel of $B$ (trivialized by $|dg_\delta(z')|$) satisfies
\begin{gather*}
|B(z,z')| \leq C \rho_L^{\alpha}\rho_R^{\beta},
\end{gather*}
then we have four situations:
\begin{gather*}
\text{ If } \alpha,\beta>n/2, \text{ then } \|B\|_{\cL(L^2)}\leq C' C. \\
\text{ If } \alpha=n/2, \;\ \beta>n/2, \text{ then } \||\log x|^{-N} B\|_{\cL(L^2)}\leq C' C, \text{ for } N>\ha. \\
\text{ If } \alpha>n/2, \;\ \beta=n/2, \text{ then } \|| B|\log x|^{-N}\|_{\cL(L^2)}\leq C' C, \text{ for } N>\ha. \\
\text{ If } \alpha=\beta=n/2, \text{ then} \||\log x|^{-N} B |\log x|^{-N}\|_{\cL(L^2)}\leq C' C, \;\ N>\ha.
\end{gather*}
\end{lemma}
Now let $B(h,\sigma)$ have Schwartz kernel $B(z,z',\sigma,h)$
supported in $r>1$, and suppose that
$$
B(z,z',\sigma,h)=e^{-i\sigma r/h} h^k \rho_L^{n/2+\gamma}
\rho_R^{n/2}x^a (x')^{b} H,
\ H \in L^\infty, \text{ and } \frac{\im \sigma}{h}<N.
$$
Since from Proposition \ref{distance}, $r=-\log \rho_R \rho_L +F,$ $F\geq 0,$ $e^{-i \frac{\sigma}{h} r}=\rho_R^{i\frac{\sigma}{h}} \rho_L^{i\frac{\sigma}{h}} e^{-i\frac{\sigma}{h} F},$ and $\frac{\im \sigma}{h}<N,$
$|e^{-i\frac{\sigma}{h} F}|<C=C(N),$ it follows that
$$
|B(z,z',\sigma,h)|\leq C h^k \rho_L^{n/2+\gamma-\im\sigma/h+a}
\rho_R^{n/2-\im\sigma/h+b}R^{a+b}
$$
As an immediate consequence of Lemma \ref{schurs}, if $a+b\geq 0$, $\delta_0>0$ and
$\gamma-\im\sigma/h+a>\delta_0$ and
$-\im\sigma/h+b>\delta_0$, then
$$
\|B\|_{L^2}\leq C'C h^k.
$$
If either $\gamma-\im\sigma/h+a=0$ or $-\im\sigma/h+b=0$, we have to add the weight $|\log x|^{-N},$ with $N>\ha.$
On the other hand, suppose now that the Schwartz kernel of $B$ is supported
in $r<2$, and (again, trivialized by $|dg_\delta(z')|$)
$$
|B(z,z',\sigma,h)|\leq Ch^k\langle r/h\rangle^{-\ell}\langle h/r\rangle^s,\ s<n+1.
$$
Note that for fixed $z',\sigma,h$, $B$ is $L^1$ in $z$, and similarly with
$z'$ and $z$ interchanged. In fact,
since the volume form is bounded by $\tilde C r^n\,dr(z')\,d\omega'$ in $r<2$,
uniformly in $z$, Schur's lemma yields
$$
\|B\|_{L^2}\leq CC'' h^k\int_0^2 \langle h/r\rangle^s\langle r/h\rangle^{-\ell} r^{n}\,dr.
$$
But
\begin{gather*}
\int_0^2 \langle h/r\rangle^s\langle r/h\rangle^{-\ell} r^{n}\,dr\\
\leq C_0\left(\int_0^h (h/r)^s r^{n}\,dr
+\int_h^2 (r/h)^{-\ell} r^{n}\,dr\right)=C_1(h^{n+1}+h^{-\ell}),
\end{gather*}
so we deduce that
$$
\|B\|_{L^2}\leq CC' (h^{n+1+k}+h^{k-\ell}).
$$
First, if, with $n+1=3$, $G$ is the parametrix, with error $E$, constructed in
Section~\ref{3D-parametrix}, then, writing $G=G_1+G_2$, and provided $\frac{\im\sigma}{h}<1$ and $|\sigma|>1$ (recall that $U_1$ has a factor $\sigma^{-1}$),
with
$G_1$ supported in $r<2$, $G_2$ supported in $r>1$, then
$|G_1|\leq C h^{-2}r^{-1}$,
$|G_2|\leq C e^{\im\sigma r/h}\rho_L^{n/2}\rho_R^{n/2}$,
and thus, using Proposition \ref{distance},
\begin{gather*}
|x^a (x')^b G_1(z,z',\sigma,h)|\leq C h^{-n-1}\langle h/r\rangle^{n-1}
\langle r/h\rangle^{-n/2},\\
|x^a (x')^b G_2(z,z',\sigma,h)|\leq C h^{-n-1}x^a (x')^b\rho_L^{n/2-\im\sigma/h}
\rho_R^{n/2-\im\sigma/h}.
\end{gather*}
On the other hand,
\begin{gather}
|x^a (x')^b E(z,z',\sigma,h)|\leq C h x^a (x')^b\rho_L^{n/2+2-\im\sigma/h}\label{weakerror}
\rho_R^{n/2-\im\sigma/h}.
\end{gather}
Thus, we deduce the following bounds:
\begin{prop}\label{l2boundrest-3D}
Suppose $n+1=3$.
Let $G(h,\sigma)$ be the operator whose kernel is given by \eqref{parametrix},
and let $E(h,\sigma)=P(h,\sigma)G(h,\sigma)-\Id.$
Then for $|\sigma|>1,$ $\frac{\im\sigma}{h}<1,$ $a> \frac{\im \sigma}{h},$ $a\geq 0,$
and
$\frac{\im\sigma}{h}< b<2- \frac{\im \sigma}{h},$ $b\geq 0$,
we have, with $C$ independent of $h$,
\begin{gather}
\begin{gathered}
||x^{a} G(h,\sigma) x^b f||_{L^2(\Bn)} \leq C h^{-1-\frac{n}{2}} ||f||_{L^2(\Bn)} \text{ and } \\
||x^{-b} E(h,\sigma) x^b f||_{L^2(\Bn)} \leq C h ||f||_{L^2(\Bn)}.
\end{gathered}\label{l2bound1-3D}
\end{gather}
If either $a= \frac{\im \sigma}{h}$ or $b= \frac{\im \sigma}{h},$ or $a=b= \frac{\im \sigma}{h},$ one has to replace the factor $x^{\pm\frac{\im\sigma}{h} }$
in \eqref{l2bound1-3D} with $\left(x^{\frac{\im\sigma}{h}} |\log x|^{-N}\right)^{\pm 1},$ $N>\ha,$ to obtain the $L^2$ bounds.
\end{prop}
In view of \eqref{weakerror} this cannot be improved using the methods of section \ref{3D-parametrix}. To obtain bounds on any strip we need the sharper bounds on the error term given by
Theorem \ref{nD-parametrix}. We now turn to arbitrary $n$ and use the parametrix $G$ and error $E$,
constructed in Theorem \ref{nD-parametrix}. Writing $G=G_0+G_1+G_2$, with $G_0\in
\Psi_{0,\semi}^{-2}$, $G_1$ supported in $r<2$, $G_2$ supported in $r>1$,
$G_1\in e^{-i\sigma r/h}\Psi_{0,\semi}^{-\infty,\infty,\frac{n}{2}-1,\infty}$,
$G_2\in e^{-i\sigma r/h}\Psi_{0,\semi}^{-\infty,\frac{n}{2}+2,-\frac{n}{2}-1,\frac{n}{2}}$,
then for $|\sigma|>1$ and $\frac{\im\sigma}{h}<N,$
\begin{gather*}
x^aG_0(z,z',\sigma,h) (x')^b\in \Psi_{0,\semi}^{-2},\\
|x^a (x')^b G_1(z,z',\sigma,h)|\leq C h^{-n-1}\langle r/h\rangle^{-n/2},\\
|x^a (x')^b G_2(z,z',\sigma,h)|\leq C h^{-n-1}x^a (x')^b\rho_L^{n/2-\im\sigma/h}
\rho_R^{n/2+\im\sigma/h}.
\end{gather*}
On the other hand, writing
$E=E_1+E_2$, with $E_1$ supported in $r<2$, $E_2$ supported in $r>1$,
$E_1\in h^\infty\rho_{\mcf}^\infty\Psi_{0,\semi}^{-\infty,\infty,\frac{n}{2}-1,\infty}$,
$E_2\in h^\infty \rho_{\mcf}^\infty\Psi_{0,\semi}^{-\infty,\infty,-\frac{n}{2}-1,\frac{n}{2}}$,
then for any $k$ and $M$ (with $C=C(M,N)$),
\begin{gather*}
|x^{-b} (x')^b E_1(z,z',\sigma,h)|\leq C h^k,\\
|x^{-b} (x')^b E_2(z,z',\sigma,h)|\leq C h^k x^M (x')^b
\rho_R^{n/2-\im\sigma/h}.
\end{gather*}
In this case, we deduce the following bounds:
\begin{prop}\label{l2boundrest-gen}
Let $G(h,\sigma)$ be the operator whose kernel is given by \eqref{kernelofg},
and let $E(h,\sigma)=P(h,\sigma)G(h,\sigma)-\Id$.
Then for $|\sigma|>1,$ $\frac{\im\sigma}{h}<M,$ $M>0,$ $ a> \frac{\im \sigma}{h},$ $a\geq 0,$ and
$b>\frac{\im\sigma}{h},$ $b\geq 0,$ and $N$ arbitrary, we have, with $C$ independent of $h$,
\begin{gather}
\begin{gathered}
||x^{a} G(h,\sigma) x^b f||_{L^2(\Bn)} \leq C h^{-1-\frac{n}{2}} ||f||_{L^2(\Bn)} \text{ and } \\
||x^{-b} E(h,\sigma) x^b f||_{L^2(\Bn)} \leq C h^N ||f||_{L^2(\Bn)}.
\end{gathered}\label{l2bound1-gen}
\end{gather}
If either $a= \frac{\im \sigma}{h}$ or $b= \frac{\im \sigma}{h},$ or $a=b= \frac{\im \sigma}{h},$ one has to replace the factor $x^{\pm\frac{\im\sigma}{h} }$
in \eqref{l2bound1-gen} with $\left(x^{\frac{\im\sigma}{h}} |\log x|^{-k}\right)^{\pm 1},$ $k>\ha,$ to obtain the $L^2$ bounds.
\end{prop}
Now we can apply these estimates to prove Theorem \ref{semiclassical-resolvent-bounds}. We know that
\begin{gather*}
P(h,\sigma) G(h,\sigma)=I+ E(h,\sigma).
\end{gather*}
Since $R(h,\sigma)$ is bounded on $L^2(\Bn)$ for $\im\sigma<0$ we can write for $\im\sigma<0,$
\begin{gather*}
G(h,\sigma)= R(h,\sigma)( I+E(h,\sigma) ).
\end{gather*}
Therefore we have, still for $\im \sigma<0,$
\begin{gather*}
x^a G(h,\sigma) x^b= x^a R(h,\sigma)x^b (I + x^{-b} E(h,\sigma) x^{-b}).
\end{gather*}
For $a,b$ and $\sigma$ as in Proposition \ref{l2boundrest-gen} we can pick $h_0$ so that
$$
||x^{-b} E(h,\sigma) x^b f||_{L^2\rightarrow L^2} \leq \ha.
$$
In this case we have
\begin{gather*}
x^a G(h,\sigma) x^b (I + x^{-b} E(h,\sigma) x^{-b})^{-1}= x^a R(h,\sigma)x^b
\end{gather*}
and the result is proved.
\section{Proof of Theorem~\ref{resolvent-bounds}}
\label{sec:resolvent-bounds}
Now we are ready to prove Theorem \ref{resolvent-bounds}. To avoid using the same notation for different parameters, we will denote the spectral parameter in the statement of Theorem \ref{resolvent-bounds} by $\la,$ instead of $\sigma.$
We write, for $|\re\la|>1,$
\begin{equation*}
(\Delta_{g_\delta}+ x^2 W-\la^2-\frac{n^2}{4})= (\re \la)^2\left[ \frac{1}{(\re\la)^2}\left(\Delta_{g_\delta}+ x^2 W-\frac{n^2}{4}\right)-\frac{\la^2}{(\re \la)^2}\right].
\end{equation*}
Thus, if we denote $h=\frac{1}{\re \la}$ and $\sigma=\frac{\la}{\re \la},$ and $E(h,\sigma)$ and $G(h,\sigma)$ are the operators in Proposition \ref{l2boundrest-gen},
we have
\begin{gather*}
(\Delta_{g_\delta}+ x^2 W-\la^2-\frac{n^2}{4})G(\frac{1}{\re \la}, \frac{\la}{\re \la})=
(\re\la)^2\left(\Id + E(\frac{1}{\re\la},\frac{\la}{\re \la})\right).
\end{gather*}
Since $R_\del(\la)=\left( \Delta_{g_\delta}+ x^2 W-\la^2-\frac{n^2}{4}\right)^{-1}$ is a well defined bounded operator if $\im \la<0,$
we can write,
\begin{gather*}
G(\frac{1}{\re \la},\frac{\la}{\re \la})= (\re\la)^2 R_\delta(\la) \left(\Id + E(\frac{1}{\re\la},\frac{\la}{\re\la})\right), \text{ for } \im\la<0.
\end{gather*}
Therefore,
\begin{gather*}
x^aG(\frac{1}{\re \la},\frac{\la}{\re \la})x^b= (\re\la)^2 x^a R_\delta(\la) x^b\left(\Id + x^{-b} E(\frac{1}{\re\la},\frac{\la}{\re\la}) x^{b}\right).
\end{gather*}
According to Proposition \ref{l2boundrest-gen}, if $\im\la<M,$ we can pick $K$ such that if $|\re \la|>K,$ and $\im \la < b$ then,
\begin{equation*}
||x^{-b} E(\frac{1}{\re \la},\frac{\la}{\re \la} )x^b||<\ha.
\end{equation*}
Therefore
$(\Id + x^{-b} E(\frac{1}{\re \la},\frac{\la}{\re \la}) x^{b})^{-1}$ is holomorphic in $\im\la<M,$ and
bounded as an operator in $L^2(\Bc, g),$ with norm independent of $\re\la,$ provided $b>\im \la.$
On the other hand, if $a>\im \la,$ then from Proposition \ref{l2boundrest-gen},
\begin{gather*}
||x^{a} G(\frac{1}{\re \la},\frac{\la}{\re \la}) x^b f||_{L^2(\Bn)} \leq C (\re \la)^{1+\frac{n}{2}} ||f||_{L^2(\Bn)}.
\end{gather*}
Since,
\begin{gather*}
x^a R_\delta(\la) x^b =(\re \la)^{-2}x^a G(\frac{1}{\re \la},\frac{\la}{\re \la}) x^b\left( I+ x^{-b} E(\frac{1}{\re \la},\frac{\la}{\re \la}) x^{b}\right)^{-1},
\end{gather*}
then, for and for $a,b$ in this range, and $|\re \la|>K,$ $x^a R_\del(\la) x^b$ is holomorphic and
\begin{gather*}
||x^a R_\delta(\la) x^b f||_{L^2(\Bc,g)} \leq C (\re\la)^{\frac{n}{2}-1} ||f||_{L^2(\Bc,g)}.
\end{gather*}
When either $a=\im \la$ or $b=\im \la$ we have to introduce the logarithmic weight and in Proposition \ref{l2boundrest-gen}. This
concludes the proof of the
$L^2$ estimates of Theorem \ref{resolvent-bounds}.
The Sobolev estimates follow from these $L^2$ estimates and interpolation.
First we observe that the following commutator properties hold: There are $\CI(\Bc)$ functions $A_i$ and $B_j,$ $i=1, 2,$ and $1\leq j \leq 5,$ such that
\begin{gather*}
[\Delta_{g_\delta},x^a]= A_1 x^a + A_2 x^a xD_x, \\
[\Delta_{g_\delta}, x^a (\log x)^{-N}]= B_1 x^a (\log x)^{-N} + B_2 x^a (\log x)^{-N-1} + \\
B_3 x^a (\log x)^{-N-2} + \left( B_4 x^a (\log x)^{-N} + B_5 x^a (\log x)^{-N-1}\right) x D_x.
\end{gather*}
Hence
\begin{gather*}
\Delta_{g_\delta} x^a R_\delta(\sigma) x^b v= x^a \Delta_{g_\delta} R_\delta(\sigma) x^b v+
A_1 x^a R_\delta(\sigma) x^b v + A_2 x^a x D_x R_\delta(\sigma) x^b v.
\end{gather*}
Since $a,b\geq 0,$
$\Delta_{g_\delta}$ is elliptic, and $\Delta_{g_\delta} R_\delta(\sigma)=\Id + (\sigma^2+1-x^2 W(x) ) R_\delta$
it follows that, see for example \cite{Mazzeo:Edge},
that there exists a constant $C>0$ such that
\begin{equation}\begin{split}
&||x^a R_\delta(\sigma) x^b v||_{ H^2_0(\Bc)} \\
&\qquad\leq C \left( \sigma^2 ||x^b v||_{L^2(\Bc)} +
||x^a R_\delta(\sigma) x^b v||_{L^2(\Bc)} + ||x^a R_\delta(\sigma) x^b v||_{H^1_0(\Bc)} \right).\label{h1bound-res}
\end{split}\end{equation}
By interpolation between Sobolev spaces we know that there exists $C>0$ such that
\begin{gather}
||x^av||_{H^1_0(\Bc)}^2 \leq C ||x^a v||_{L^2(\Bc)} ||x^a v||_{H^2_0(\Bc)}\label{interp2}
\end{gather}
Therefore, for any $\eps>0,$
\begin{gather}
|| x^a v ||_{H^1_0(\Bc)} \leq C\left( \eps ||x^a v ||_{H^2_0(\Bc)} + \eps^{-1} ||x^a v ||_{L^2(\Bc)}\right), \label{interp1}
\end{gather}
and if one takes $\eps$ small enough, \eqref{h1bound-res} and \eqref{interp1} give
\eqref{sobolev1} with $k=2.$ If one uses \eqref{interp2} and \eqref{sobolev1} with $k=2$ one obtains
\eqref{sobolev1} with $k=1.$ The proof of \eqref{sobolev2} follows by the same argument.
This completes the proof of Theorem~\ref{resolvent-bounds}.
\section{Structure of $\Delta_X$ near the boundaries}
\label{sec:black-hole}
We begin the proof of Theorem \ref{globalest} by analyzing the structure of
the operator $\Delta_X$ near $r=r_{\bH}$ and $r=r_{\sI}.$ We recall that
$\beta(r)=\ha \frac{d}{dr} \alpha^2(r)$ and that $\beta_\bH=\beta(r_\bH)$
and $\beta_\sI=\beta(r_\sI).$
We show that near
these ends, after rescaling $\alpha,$ the operator $\alpha^{\frac{n}{2}} \Delta_X
\alpha^{-\frac{n}{2}}$ is a small perturbation of the Laplacian of the hyperbolic
metric of constant negative sectional curvature $-\beta_{\bH}^2$ near $r_{\bH}$ and
$-\beta_{\sI}^2$ near $r_{\sI}.$
Since $\alpha'(r)\not=0$ near $r=r_{\bH}$ and $r=r_{\sI},$ $\Delta_X$ can be
written in terms of $\alpha$ as a `radial' coordinate
\begin{equation*}
\Delta_X=
\beta r^{-n}\alpha D_\alpha( \beta r^n\alpha D_\alpha)+\alpha^2 r^{-2}
\Delta_\omega.
\end{equation*}
We define a $\CI$ function $x$ on $[r_{\bH}, r_{\sI}]$ which is positive in the interior of the interval and rescales $\alpha$ near the ends by
\begin{gather}
\alpha= 2r_{\bH}\beta_{\bH} x \text{ near } r=r_{\bH}\text{ and }
\alpha= 2r_{\sI}|\beta_{\sI}| x \text{ near }r=r_{\sI}.
\label{rescale}
\end{gather}
Using $x$ instead of $\alpha$ near the ends $r_\bH$ and $r_\sI,$ we obtain
\begin{gather}
\begin{gathered}
\Delta_X=
\beta r^{-n} x D_{x}( \beta r^n x D_{x})+4{x}^2\beta_{\bH}^2r_{\bH}^2 r^{-2}
\Delta_\omega, \text{ near } r=r_{\bH}, \\
\Delta_X=\beta r^{-n}xD_{x}(\beta r^n xD_{x})+
4x^2\beta_{\sI}^2r_{\sI}^2 r^{-2} \Delta_\omega, \text{ near } r=r_{\sI}.
\end{gathered}\label{conjugation}
\end{gather}
\begin{prop}\label{modelnearends} There exists $\del>0$ such that, if we identify each of the neighborhoods of $\{x=0\}$ given by $r\in [r_{\bH},r_{\bH}+\del)$ and $r\in (r_{\sI}-\del, r_{\sI}],$
with a neighborhood of the boundary of the ball
$\Bn,$ then there exist two $\CI$ functions,
$W_{\bH}(x)$ defined near $r=r_{\bH},$ and
$W_{\sI}(x)$ defined near $r= r_{\sI},$ such that
\begin{gather}
\begin{gathered}
\alpha^\frac{n}{2} \Delta_X \alpha^{-\frac{n}{2}}=x^\frac{n}{2} \Delta_X x^{-\frac{n}{2}} = \Delta_{g_{\bH}} + x^2 W_{\bH} -\beta_{\bH}^2\frac{n^2}{4},
\text{ near } r=r_{\bH} \text{ and } \\
\alpha^\frac{n}{2}\Delta_X \alpha^{-\frac{n}{2}}=x ^\frac{n}{2}\Delta_X {x}^{-\frac{n}{2}} = \Delta_{g_{\sI}} + x^2 W_{\sI} -\beta_{\sI}^2\frac{n^2}{4},
\text{ near } r=r_{\sI},
\end{gathered}\label{modelnearends:eq}
\end{gather}
where $g_{\bH}$ and $g_{\sI}$ are small perturbations of the hyperbolic metrics with sectional curvature
$-\beta_{\bH}^2$ and $-\beta_{\sI}^2$
respectively on the interior of $\Bn,$ i.e.
\begin{gather}
\begin{gathered}
g_{\bH}= \frac{ 4 dz^2}{ \beta_{\bH}^2(1-|z|^2)^2}+H_{\bH}, \text{ and }
g_{\sI}= \frac{ 4 dz^2}{ \beta_{\sI}^2(1-|z|^2)^2}+H_{\sI},
\end{gathered}\label{metform}
\end{gather}
where $H_{\bH}$ and $H_{\sI}$ are symmetric 2-tensors $\CI$ up to the
boundary of $\Bn.$
\end{prop}
\begin{proof} It is only necessary to prove the result near one of the ends. The computation near the other end is identical and one only needs to replace
the index $\bH$ by $\sI.$ From \eqref{conjugation} we we find that near $r=r_{\bH},$
\begin{equation*}\begin{split}
x^{\frac{n}{2}} \Delta_X {x}^{-\frac{n}{2}}= &\beta^2(x D_{x})^2 +in \beta^2 {x} D_{x}
+\beta r^{-n} (x D_{x}(\beta r^n) )x D_{x} +(2 \beta_{\bH} r_{\bH})^2 r^{-2} x^2 \Delta_\omega \\ & -i\frac{n}{2} \beta r^{-n} x D_{x}(\beta r^n)-\beta^2\frac{n^2}{4}.
\end{split}\end{equation*}
Let $g_\bH$ be the metric defined on a neighborhood of $\p \Bn$ given by
\begin{gather}
\begin{gathered}
g_{\bH}= \frac{dx^2}{\beta^2 x^2} + \la_{\bH}^{-2} {r^2}\frac{d\omega^2}{x^2}, \text{ where } \lambda_\bH= 2|\beta_\bH| r_\bH.
\end{gathered} \label{refmet}
\end{gather}
The Laplacian of this metric is
\begin{gather*}
\Delta_{g_\bH}= \beta^2(x D_x)^2 +in \beta^2 x D_x + \beta r^{-n} (x D_x(\beta r^n))x D_x +
\lambda_\bH^2 r^{-2} x^2\Delta_\omega.
\end{gather*}
Therefore we conclude that near the ends $r=r_{\bH}$
\begin{gather*}
x^\frac{n}{2} \Delta_X x^{-\frac{n}{2}} = \Delta_{g_\bH} -\beta^2\frac{n^2}{4} -i \frac{n}{2}\beta r^{-n} x D_x(\beta r^2).
\end{gather*}
Since $r=r(x^2),$ we can write near $r_{\bH}$ and $r_{\sI},$
\begin{gather*}
r=r_\bH + x^2 A_\bH(x^2) \text{ and }
\beta(r)= \beta_\bH + x^2 B_\bH(x^2)\text{ near } r=r_\bH.
\end{gather*}
Therefore, near $r=r_{\bH}$
\begin{gather*}
\frac{1}{\beta^2}= \frac{1}{\beta_\bH^2} + x^2 \widetilde{B}_\bH(x^2).
\end{gather*}
We conclude that there exist a symmetric 2-tensor $H_\bH(x^2,d x,d\omega)$ near
$r=r_\bH$ which is $\CI$ up to
$\{x=0\},$ and such that the metric $g_\bH$ given by \eqref{refmet} can be written
near $r=r_\bH$ as
\begin{gather}
\begin{gathered}
g_\bH= \frac{d x^2}{\beta_\bH^2 x^2} + \frac{ d\omega^2}{4\beta_\bH^2 x^2} + H_\bH \text{ near } r=r_\bH.
\end{gathered} \label{modelend}
\end{gather}
Let $\tilde{g}$ be the metric on the interior of $\Bn$ which is given by
\begin{gather*}
\tilde{g}= \frac{4 |dz|^2}{ c^2 (1-|z|^2)^2}.
\end{gather*}
We consider local coordinates valid for $|z|>0$ given by $(x,\omega),$ where $\omega=z/|z|,$ and
$x=\frac{1-|z|}{1+|z|}.$
The metric $\tilde{g}$ written in terms of these coordinates is given by
\begin{gather*}
\tilde{g}= \frac{ dx^2}{c^2 x^2} + (1-x^2)^2 \frac{d\omega^2}{4c^2 x^2}.
\end{gather*}
Therefore, near $x=0$
\begin{gather*}
\tilde{g}= \frac{ dx^2}{c^2 x^2} + \frac{ d\omega^2}{4 c^2 x^2} + H(x^2,\omega,dx,d\omega),
\end{gather*}
where $H$ is a symmetric 2-tensor smooth up to the boundary of $\Bn.$
This concludes the proof of the Proposition.
\end{proof}
\section{From cut-off and models to stationary resolvent}
\label{sec:decomposition}
Next we use the method of Bruneau and Petkov \cite{Bruneau-Petkov:Semiclassical} to decompose the
operator $R(\sigma)$ in terms of its cut-off part $\chi R(\sigma) \chi$ and the contributions from the ends, which are controlled by Theorem~\ref{resolvent-bounds}.
For that one needs to define some suitable cut-off functions. For $\delta>0$ let $\chi_j,$ $\chi_j^1,$ and $\tilde{\chi}_j,$ $j=1,2,$ defined by
\begin{gather*}
\chi_1(r)= 1 \text{ if } r>r_{\bH}+4\delta, \;\ \chi_1(\alpha)=0 \text{ if } r< r_{\bH} + 3\delta, \\
\chi_1^1(r)= 1 \text{ if } r>r_{\bH}+2\delta, \;\ \chi_1^1(\alpha)=0 \text{ if } r< r_{\bH} + \delta, \\
\tilde{\chi}_1(r)= 1 \text{ if } r>r_{\bH}+6\delta, \;\ \tilde{\chi_1}(\alpha)=0 \text{ if } r< r_{\bH} +5 \delta, \\
\chi_2(r)= 1 \text{ if } r<r_{\sI}-4\delta, \;\ \chi_2(\alpha)=0 \text{ if } r> r_{\sI} - 3\delta, \\
\chi_2^1(r)= 1 \text{ if } r<r_{\sI}-2\delta, \;\ \chi_2^1(\alpha)=0 \text{ if } r> r_{\sI} - \delta, \\
\tilde{\chi}_2(r)= 1 \text{ if } r<r_{\sI}-6\delta, \;\ \tilde{\chi}_2(\alpha)=0 \text{ if } r> r_{\sI} - 5\delta,
\end{gather*}
and let
\begin{gather*}
\chi_3(r)= 1-(1-\chi_1)(1-\chi_1^1)-(1-\chi_2)(1-\chi_2^1).
\end{gather*}
$\chi_3(r)$ is supported in $[r_{\bH}+\delta, r_{\sI}-\delta]$ and $\chi_3(r)=1$
if $r\in[r_{\bH}+2\delta, r_{\sI}-2\delta].$ Let $\chi \in C_0^\infty(r_{\bH},r_{\sI})$ with
$\chi(r)=1$ if $r\in [r_{\bH}+\delta/2, r_{\sI}-\delta/2].$
Now we will use Proposition \ref{modelnearends} for $\delta$ small enough. Let $g_{\bH}$ and $g_{\sI}$
be the metrics given on the interior of $\Bn$ given by \eqref{metform} and let
\begin{gather}
g_{\bH,\del}= \frac{ 4 dz^2}{ \beta_{\bH}^2 (1-|z|^2)^2}+ (1-\tilde{\chi}_1) H_{\bH}, \text{ and }
g_{\sI,\del}= \frac{ 4 dz^2}{ \beta_{\sI}^2 (1-|z|^2)^2}+(1-\tilde{\chi}_2)H_{\sI}. \label{g+andg++}
\end{gather}
Since $g_{\bH,\del}=g_{\bH}$ if $\tilde{\chi}_1=0,$ and $g_{\sI,\del}=g_{\sI}$ if $\tilde{\chi}_2=0,$ it follows from Proposition \ref{modelnearends} that, for $x$ given by equation \eqref{rescale}
\begin{gather*}
\alpha^\frac{n}{2} \Delta_X \alpha^{-\frac{n}{2}} (1-{\chi}_1) f= (\Delta_{g_{\bH,\del}} +x^2 W_{\bH}-\frac{n^2}{4} \beta_{\bH}^2-\sigma^2) (1-\chi_1) f, \text{ and } \\
\alpha^\frac{n}{2} \Delta_X \alpha^{-\frac{n}{2}} (1-\chi_1^1) f= (\Delta_{g_{\bH,\del}} +x^2 W_{\bH}-\frac{n^2}{4} \beta_{\bH}^2-\sigma^2) (1-\chi_1^1) f. \\
\alpha^\frac{n}{2} \Delta_X \alpha^{-\frac{n}{2}} (1-\chi_2) f= (\Delta_{g_{\sI,\del}} +x^2 W_{\sI}-\frac{n^2}{4}\beta_{\sI}^2-\sigma^2) (1-\chi_2) f, \text{ and } \\
\alpha^\frac{n}{2} \Delta_X \alpha^{-\frac{n}{2}} (1-\chi_2^1) f= (\Delta_{g_{\sI,\del}} +x^2 W_{\sI}-\frac{n^2}{4}\beta_{\sI}^2-\sigma^2) (1-\chi_2^1) f.
\end{gather*}
Let
\begin{gather*}
R_\alpha(\sigma)= \alpha^\frac{n}{2} R(\sigma) \alpha^{-\frac{n}{2}}=(\alpha^\frac{n}{2} \Delta_X \alpha^{-\frac{n}{2}}-\sigma^2)^{-1}, \;\
\im \sigma<0,
\end{gather*}
and let
\begin{gather*}
R_{\bH}(\sigma)=(\Delta_{g_{\bH,\del}} +x^2 W_{\bH} -\sigma^2-\frac{n^2}{4} \beta_{\bH}^2)^{-1} \text{ and } \\
R_{\sI}(\sigma)=(\Delta_{g_{\sI,\del}} +x^2 W_{\sI} -\sigma^2-\frac{n^2}{4} \beta_{\sI}^2)^{-1}
\end{gather*}
be operators acting on functions defined on $\Bn.$
If, as in Proposition \ref{modelnearends}, we identify neighborhoods of $r=r_{\bH}$ and $r=r_{\sI}$ with a neighborhood of the boundary of $\Bn,$ we obtain the following identity for the resolvent
\begin{gather}
\begin{gathered}
R_\alpha(\sigma)= R_\alpha(\sigma)\chi_3+ (1-\chi_1)R_{\bH}(\sigma)(1-\chi_1^1) +
(1-\chi_2) R_{\sI}(\sigma)(1-\chi_2^1) - \\
R_\alpha(\sigma)[\Delta_{g_{\bH}}, 1-\chi_1] R_{\bH}(\sigma)(1-\chi_1^1) -
R_\alpha(\sigma)[\Delta_{g_{\sI}}, 1-\chi_2] R_{\sI}(\sigma)(1-\chi_2^1).
\end{gathered}\label{residentity1}
\end{gather}
Similarly, one obtains
\begin{gather}
\begin{gathered}
R_\alpha(\sigma)= \chi_3 R_\alpha(\sigma)+ (1-\chi_1^1)R_{\bH}(\sigma)(1-\chi_1) +
(1-\chi_2^1) R_{\sI}(\sigma)(1-\chi_2) + \\
(1-\chi_1^1)R_{\bH}(\sigma)[\Delta_{g_{\bH}}, 1-\chi_1] R_\alpha(\sigma)+
(1-\chi_2^1)R_{\sI}(\sigma)[\Delta_{g_{\sI}},1- \chi_2] R_\alpha(\sigma).
\end{gathered}\label{residentity2}
\end{gather}
These equations can be verified by applying $\alpha^\frac{n}{2} \Delta_X \alpha^{-\frac{n}{2}}-\sigma^2$ on the left and on the right of both sides of the identities.
Substituting \eqref{residentity2} into \eqref{residentity1} we obtain
\begin{equation}
\begin{split}
R_\alpha(\sigma)&= M_1(\sigma) \chi R_\alpha(\sigma) \chi M_2(\sigma) +\\
&\qquad\qquad
(1-\chi_1)R_{\bH}(\sigma)(1-\chi_1^1)+ (1-\chi_2)R_{\sI}(\sigma)(1-\chi_2^1), \\
&\text{ where} \\
M_1(\sigma)&= \chi_3 + (1-\chi_1^1)R_{\bH}(\sigma) (1-\tilde{\chi}_1)[\Delta_{g_{\bH}},1-\chi_1]\\
&\qquad\qquad
+ (1-\chi_2^1)R_{\sI}(\sigma)(1-\tilde{\chi}_2)[\Delta_{g_{\sI}},1-\chi_2], \\
M_2(\sigma)&= \chi_3 -[\Delta_{g_{\bH}},1-\chi_1](1-\tilde{\chi}_1)R_{\bH}(\sigma)(1-\chi_1^1)\\
&\qquad\qquad-[\Delta_{g_{\sI}},1-\chi_2](1-\tilde{\chi}_2)R_{\sI}(\sigma)(1-\chi_2^1).
\end{split}\label{brpkid}
\end{equation}
This gives a decomposition of $R_\alpha(\sigma)$ in terms of the cutoff resolvent, studied by Bony and H\"afner \cite{Bony-Haefner:Decay}, and the resolvents of he Laplacian of a metric which are small perturbations of the Poincar\'e metric in $\Bn.$ The mapping properties of such operators were established in
Section \ref{sec:resolvent-bounds}. Next we will put the estimates together and finish the proof of our main result.
\section{The proof of Theorem~\ref{globalest} in $3$ dimensions}
\label{sec:black-hole-proof}
We now prove Theorem \ref{globalest} for $n+1=3$ using
Theorem \ref{resolvent-bounds} and Theorem \ref{bhthm}. We first restate a strengthened version of the theorem which includes the case where the weight
$b=\im\sigma.$
\begin{thm}\label{globalest-strong} Let $\eps>0$
be such that \eqref{bonyhafnerest}
holds and suppose
$$
0<\gamma<\min(\eps,\beta_{\bH},|\beta_{\sI}|,1).
$$
Then for $b>\gamma$
there exist $C$ and $M$ such that if $\im\sigma\le\gamma$ and $|\re\sigma
|\ge1,$
\begin{equation}
||{\tilde{\alpha}}^b R(\sigma){\tilde{\alpha}}^bf||
_{L^2 ( X ;\Omega)}\leq C|\sigma|^M ||f||_{L^2 ( X ;\Omega) },
\label{mainest1-r}
\end{equation}
where $\tilde{\alpha}$ was defined in \eqref{SeClRe.1} and the measure $\Omega$ was defined in \eqref{measure}.
Moreover, for $N>\ha$ and $0<\del<<1,$ choose $\psi_N(r)\in\CI(r_{\bH}, r_{\sI})$ with
$\psi_N(r)\ge1,$ such that
\begin{equation}
\psi_N=|\log \alpha|^{-N}\Mif r-r_{\bH}<\del\Mor r_{\sI}-r<\del.
\label{SeClRe.4}\end{equation}
Then with $\gamma$ as above, there exists $C>0$ and $M\geq 0$ such that for
$|\re\sigma|\ge1$ and $|\im\sigma|\le\gamma$
\begin{gather}
\begin{gathered}
||{\tilde{\alpha}}^{\im \sigma} \psi_N(\alpha) R(\sigma)
{\tilde{\alpha}}^{b} f||_{L^2 ( X ;\Omega )}
\leq C |\sigma|^M ||f||_{L^2 ( X ;\Omega)},
\\
||{\tilde{\alpha}}^{b} R(\sigma)
{\tilde{\alpha}}^{\im \sigma}\psi_N(\alpha) f||_{L^2 ( X ;\Omega)}
\leq C |\sigma|^M ||f||_{L^2 ( X ;\Omega) }\Mand
\\
||{\tilde{\alpha}}^{\im \sigma} \psi_N(\alpha) R(\sigma)
{\tilde{\alpha}}^{\im\sigma}\psi_N(\alpha) f||_{L^2 (X;\Omega)}
\leq C |\sigma|^M ||f||_{L^2 ( X ;\Omega)}.
\end{gathered}\label{mainest11}\end{gather}
\end{thm}
\begin{proof} Recall from \eqref{g+andg++} that
\begin{gather*}
g_{\bH,\delta}= \frac{1}{\beta_\bH^2} g_\delta \text{ and } g_{\sI,\delta}= \frac{1}{\beta_\sI^2} g_\delta,
\end{gather*}
where $g_\delta$ is of the form \eqref{metgeps}.
So we obtain $\Delta_{g_{\bH,\delta}}=\beta_\bH^2 \Delta_{g_\delta}$ and similarly
$\Delta_{g_{\sI,\delta}}=\beta_\sI^2 \Delta_{g_\delta}.$ Therefore,
\begin{gather}
\begin{gathered}
R_\bH(\sigma)=\left( \beta_\bH^2\Delta_{g_\delta}+ x^2 W_\bH-\sigma^2-\frac{n^2}{4} \beta_\bH^2\right)^{-1}= \\ \beta_\bH^{-2}\left(\Delta_{g_\delta}+ x^2 \beta_\bH^{-2} W_\bH-\sigma^2\beta_\bH^{-2}-\frac{n^2}{4}\right)^{-1}= \beta_\bH^{-2} R_\delta( \sigma|\beta_\bH|^{-1}).
\end{gathered}\label{bulletid}
\end{gather}
Therefore, by replacing $\sigma$ with $\sigma |\beta_\bH|^{-1}$ in \eqref{sobolev1} and
\eqref{sobolev2} setting $a=\frac{A}{|\beta_\bH|},$ and $b=\frac{B}{|\beta_\bH|},$
we deduce from Theorem \ref{resolvent-bounds} that there exists $\delta_0>0$ such that if $0<\delta<\delta_0,$
for $\im \sigma<A$ and $\im\sigma < B$ and $|\re \sigma|> K(\delta),$
\begin{gather}
\begin{gathered}
||x^{\frac{A}{|\beta_\bH|}} R_\bH(\sigma) x^{\frac{B}{|\beta_\bH|}} v||_{H^k(\Bc)} \leq
C |\sigma|^k || v||_{L^2(\Bc)}, \;\ k=0,1,2, \\
||x^{\frac{A}{|\beta_\bH|}} R_\bH\sigma) x^{\frac{B}{|\beta_\bH}} v||_{L^2(\Bc)} \leq
C |\sigma|^k || v||_{H_0^{-k}(\Bc)}, \;\ k=0,1,2.
\end{gathered}\label{sobolev3}
\end{gather}
Of course, the same argument applied to $R_{\sI}$ gives
\begin{gather}
\begin{gathered}
||x^{\frac{A}{|\beta_\sI|}} R_\sI(\sigma) x^{\frac{B}{|\beta_\sI|}} v||_{H^k(\Bc)} \leq
C |\sigma|^k || v||_{L^2(\Bc)}, \;\ k=0,1,2, \\
||x^{\frac{A}{|\beta_\sI|}} R_\sI\sigma) x^{\frac{B}{|\beta_\sI}} v||_{L^2(\Bc)} \leq
C |\sigma|^k || v||_{H_0^{-k}(\Bc)}, \;\ k=0,1,2.
\end{gathered}\label{sobolev3sI}
\end{gather}
When $A={\im\sigma},$ or $B={\im\sigma},$
we define $T_{\bH,A,B,N}$ and $T_{\sI,A,B,N}$ and
as in \eqref{deftabn} by replacing $R_\delta(\sigma)$ with either
$R_\bH(\sigma)$ or $R_\sI(\sigma),$ $a$ with $A$ and $b$ with $B.$ Using \eqref{sobolev2} we obtain
for $J=\bH$ or $J=\sI,$
\begin{gather}
\begin{gathered}
|| T_{J,A,B,N} (\sigma) v||_{H_0^k(\Bc)} \leq C |\sigma|^k ||v||_{L^2(\Bc)}, \;\ k=0,1,2, \\
|| T_{J,A,B,N} v||_{L^2(\Bc)} \leq C |\sigma|^k ||v||_{H_0^{-k}(\Bc)}, \;\ k=0,1,2.
\end{gathered}\label{sobolev4}
\end{gather}
Now we recall that $\alpha=2 r_{\bH}\beta_{\bH}$ near $r_{\bH}$ and similarly
$\alpha=2 r_{\sI}\beta_{\sI}$ near $r_{\sI}.$ We will use these estimates, identity \eqref{brpkid} and Theorem \ref{bhthm} to prove
Theorem \ref{globalest}. Indeed, in the case $a>\im\sigma,$ $b>\im \sigma$ we write
\begin{gather*}
{\tilde{\alpha}}^a R_\alpha {\tilde{\alpha}}^b= {\tilde{\alpha}}^a M_1(\sigma) {\tilde{\alpha}}^b
{\tilde{\alpha}}^{-b} \chi R_\alpha(\sigma) \chi {\tilde{\alpha}}^{-b} {\tilde{\alpha}}^b M_2(\sigma)
{\tilde{\alpha}}^b
+ \\ (1-\chi_1) {\tilde{\alpha}}^a R_{\bH}(\sigma){\tilde{\alpha}}^b (1-\chi_1^1) +(1-\chi_2)
{\tilde{\alpha}}^aR_{\sI}(\sigma){\tilde{\alpha}}^b (1-\chi_2^1).
\end{gather*}
Notice that the measure in $\Bn$ is $x^{-n-1}dxd\omega,$ which corresponds to
$\alpha^{-n-1}d\alpha d\omega$ which in turn corresponds to $\alpha^{-n-2} dr d\omega.$ In this case $n=2,$ but this part of the argument is the same for all dimensions, and we will not set $n=2.$
Thus, we deduce from Theorem \ref{resolvent-bounds} that
\begin{gather*}
|| (1-\chi_1) {\tilde{\alpha}}^a R_{\bH}(\sigma){\tilde{\alpha}}^b (1-\chi_1^1) ||_{L^2(X;\alpha^{-n-2} dr d\omega)} \leq C ||v||_{L^2(X;\alpha^{-n-2} dr d\omega)}, \\
||(1-\chi_2) {\tilde{\alpha}}^aR_{\sI}(\sigma){\tilde{\alpha}}^b (1-\chi_2^1) v||_{L^2(X;\alpha^{-n-2} dr d\omega)} \leq C||v||_{L^2(X;\alpha^{-n-2} dr d\omega)}.
\end{gather*}
Recall that $\Omega=\alpha^{-2} r^2 dr d\omega.$ Since $r\in [r_\bH,r_\sI],$ $r_\bH>0,$ this gives
\begin{gather}
\begin{gathered}
|| (1-\chi_1) {\tilde{\alpha}}^a \alpha^{-\frac{n}{2}}R_{\bH}(\sigma)\alpha^{\frac{n}{2}} {\tilde{\alpha}}^b (1-\chi_1^1) ||_{L^2(X;\Omega)} \leq C ||v||_{L^2(X;\Omega)}, \\
||(1-\chi_2) {\tilde{\alpha}}^a\alpha^{-\frac{n}{2}}R_{\sI}(\sigma)\alpha^\frac{n}{2} {\tilde{\alpha}}^b (1-\chi_2^1) v||_{L^2(X;\Omega)} \leq C||v||_{L^2(X;\alpha^{-n-2} \Omega)}.
\end{gathered}\label{finalestimate1}
\end{gather}
Similarly, using the Sobolev estimates in Theorem \ref{resolvent-bounds}, we obtain
\begin{gather}
\begin{gathered}
|| {\tilde{\alpha}}^a \alpha^{-\frac{n}{2}} M_1(\sigma) \alpha^\frac{n}{2} {\tilde{\alpha}}^b v||_{L^2(X;\Omega)} \leq
C |\sigma| ||v||_{L^2(X;\Omega)}, \\
|| {\tilde{\alpha}}^a \alpha^{-\frac{n}{2}} M_2(\sigma)\alpha^\frac{n}{2} {\tilde{\alpha}}^b v||_{L^2(X;\Omega)} \leq
C |\sigma| ||v||_{L^2(X;\Omega)}.
\end{gathered}\label{finalestimate2}
\end{gather}
Since $\chi$ is compactly supported in the interior of $X,$ it follows from Theorem \ref{bhthm} that
\begin{gather}
||{\tilde{\alpha}}^{-b} \chi R_\alpha(\sigma) \chi {\tilde{\alpha}}^{-b} v||_{L^2(X;\Omega)} \leq C||v||_{L^2(X;\Omega)}.\label{finalestimate3}
\end{gather}
Estimates \eqref{finalestimate1}, \eqref{finalestimate2} and \eqref{finalestimate3} imply that
\begin{gather*}
||{\tilde{\alpha}}^{-b} \alpha^{-\frac{n}{2}} R_\alpha(\sigma)\alpha^\frac{n}{2} \chi {\tilde{\alpha}}^{-b} v||_{L^2(X;\Omega)} \leq C||v||_{L^2(X;\Omega)}.
\end{gather*}
But $\alpha^{-\frac{n}{2}} R_\alpha(\sigma)\alpha^\frac{n}{2}=R(\sigma).$ This proves \eqref{mainest1-r}.
\end{proof}
\section{The proof of Theorem~\ref{globalest} in general dimension}\label{sec:black-hole-n+1-proof}
We will outline the main steps necessary to connect to
the results
of \cite{Datchev-Vasy}, and refer the reader to \cite{Datchev-Vasy} for more details.
First choose $\delta$ so small \eqref{bich-convexity1} holds.
Let $X_0$ and $X_1$ be as defined in \eqref{Xdecomp},
then we recall that for $\delta$ small,
\begin{gather*}
\alpha^{-\frac{n}{2}} P\alpha^{-\frac{n}{2}}|_{X_0}= P(h,\sigma),
\end{gather*}
where $P(h,\sigma)$ stands for model near either $r_{\bH}$ or near
$r_{\sI}.$ By
Theorem \ref{semiclassical-resolvent-bounds}
there exists $h_0>0$ such that for $h\in (0,h_0),$
\begin{gather*}
||x^a (h^2 P_0-\sigma)^{-1} x^b||_{L^2\rightarrow L^2} \leq C h^{-1-\frac{n}{2}} \;\ \sigma \in (1-c,1+c) \times (-c, \eps h)\subset\Cx.
\end{gather*}
Let $P_1$ be the operator defined in \eqref{operatorp1} and let $\eps>0$ be such that \eqref{wuzwestimate} holds. Then it follows from Theorem 2.1 of
\cite{Datchev-Vasy} that there exist $h_1>0,$ $C>0$ and $K>0$ such that for $h\in (0,h_1),$
\begin{equation}\begin{split}\label{eq:davaestimate}
||{\tilde{\alpha}}^b \alpha^{-\frac{n}{2}} (h^2 \Delta_X-\sigma^2)^{-1} &{\tilde{\alpha}}^a \alpha^{-\frac{n}{2}}||_{L^2\rightarrow L^2} \leq C h^{-K},\\
&\sigma \in (1-c,1+c) \times (-c, \eps h)\subset\Cx.
\end{split}\end{equation}
The estimate in Theorem \ref{globalest} follows by restating \eqref{eq:davaestimate}
in the non-semiclassical language, i.e.\ multiplying it by
$h^2$ and replacing $\sigma$ by $h^{-1}\sigma$.
\def$'$} \def\cprime{$'${$'$} \def$'$} \def\cprime{$'${$'$}
|
\section{Introduction}
A {\it Cullen number} is a number of the form $C_n=n2^n+1$ for some $n\ge 1$. They attracted attention of researchers since it seems that it is hard to find primes of this form. Indeed, Hooley \cite{Ho} showed that for most $n$ the number $C_n$ is composite. For more about testing $C_n$ for primality, see \cite{BF} and \cite{GO}. For an integer $a>1$, a {\it pseudoprime} to base $a$ is a compositive positive integer $m$ such that $a^m\equiv a\pmod m$. Pseudoprime Cullen numbers have also been studied. For example, in
\cite{LuSh} it is shown that for most $n$, $C_n$ is not a base $a$-pseudoprime. Some computer searchers up to several millions did not turn up any pseudo-prime $C_n$ to any base. Thus, it would seem that Cullen numbers which are pseudoprimes are very scarce. A {\it Carmichael number} is a positive integer $m$ which is a base $a$ pseudoprime for any $a$. A composite integer $m$ is called a {\it Lehmer number} if $\phi(m)\mid m-1$, where $\phi(m)$ is the Euler function of $m$. Lehmer numbers are Carmichael numbers; hence, pseudoprimes in every base. No Lehmer number is known, although it is known that there are no Lehmer numbers in certain sequences, such as the Fibonacci sequence (see \cite{Lu1}), or the sequence of repunits in base $g$ for any $g\in [2,1000]$ (see \cite{CiLu}). For other results on Lehmer numbers, see \cite{BaGu}, \cite{BaLu}, \cite{LuPo}, \cite{Po1}, \cite{Po2}.
\medskip
Our result here is that there is no Cullen number with the Lehmer property. Hence, if $\phi(C_n)\mid C_n-1$, then $C_n$ is prime.
\begin{theorem}
\label{thm:main}
Let $C_n$ be the $n$th Cullen number. If $\phi(C_n) \mid C_n-1$, then $C_n$ is prime.
\end{theorem}
\section{Proof of Theorem \ref{thm:main}}
Assume that $n\ge 30$, that $\phi(C_n)\mid C_n-1$, but that $C_n$ is not prime. Then $C_n$ is square-free. Write
$$
C_n=\prod_{i=1}^k p_i.
$$
So,
$$
\prod_{i=1}^k (p_i-1)\mid n2^n.
$$
Write $n=2^{\alpha}n_1$, where $n_1$ is odd. Then
$C_n=n_12^{n_2}+1$, where $n_2:=\alpha+n$. Let $p$ be any prime factor of $C_n$. Since $p-1\mid C_n-1$, it follows that $p=m_p2^{n_p}+1$ for some odd divisor $m_p$ of $n$ and some $n_p$ with
$$
n_p\le n_2=n+\alpha\le n+\frac{\log n}{\log 2}.
$$
Let us first show that in fact $n_p\le n$. Assume that $n_p>n$. Then,
\begin{equation}
\label{eq:11}
C_n=n2^n+1=p\lambda,
\end{equation}
for some positive integer $\lambda$, where $p\ge 2^{n+1}+1$. Observe that $\lambda>1$ because $C_n$ is not prime. Now
$$
\lambda=\frac{C_n}{p}\le \frac{n2^n+1}{2^{n+1}+1}<n.
$$
Reducing equation \eqref{eq:11} modulo $2^n$, we get that $2^n\mid \lambda-1$, so $2^{n}\le \lambda-1<n$, which is false for any $n>1$. Hence, $n_p\le n$.
\medskip
Next we look at $m_p$. If $m_p=1$, then $p=2^{n_p}+1$ is a Fermat prime. Hence, $n_p=2^{\gamma_p}$ for some nonnegative integer $\gamma$. Since $2^{\gamma_p}=n_p\le n$, we get that $\gamma_p< (\log n)/(\log 2)$. Hence, the prime $p$ can take at most
$1+(\log n)/(\log 2)$ values. Next, observe that since
\begin{equation}
\label{eq:mp}
\prod_{p\mid C_n} m_p\mid n,
\end{equation}
it follows that the number of prime factors $p$ of $C_n$ such that $m_p>1$ is $\le (\log n)/(\log 3)$. Hence, we arrived at the bound
\begin{equation}
\label{eq:k}
k<1+\frac{\log n}{\log 2}+\frac{\log n}{\log 3}<1+2.4\log n.
\end{equation}
\medskip
We next bound $n_p$. Put $N:=\lfloor {\sqrt{n/\log n}}\rfloor$, and consider pairs $(a,b)$ of integers in
$\{0,1,\ldots,N\}$. There are $(N+1)^2>n/\log n$ such pairs. For each such pair, consider the expression $L(a,b):=an+bn_p\in [0,2n^{3/2}/(\log n)^{1/2}]$. Thus, there exist two pairs $(a,b)\ne (a_1,b_1)$ such that
$$
|(a-a_1)n+(b-b_1)n_p|=|L(a,b)-L(a_1,b_1)|\le \frac{2n^{3/2}/(\log n)^{1/2}}{n/\log n-1}<3(n\log n)^{1/2}.
$$
Put $u:=a-a_1,~v:=b-b_1$. Then $(u,v)\ne (0,0)$ and
$$
|un+vn_p|<3(n\log n)^{1/2}.
$$
We may also assume that $u$ and $v$ are coprime, for if not, we replace the pair $(u,v)$ by the pair $(u_1,v_1)$, where $d:=\gcd(u,v),~u_1:=u/d,~v_1:=v/d$, and the properties that $\max\{|u_1|,|v_1|\}\le (n/\log n)^{1/2}$ and $|u_1n+v_1 n_p|<3(n\log n)^{1/2}$ are still fulfilled. Finally, up to replacing the pair $(u,v)$ by the pair $(-u,-v)$, we may assume that $u\ge 0$.
Now consider the congruences $n2^n\equiv -1\pmod p$ and $m_p2^{n_p}\equiv -1\pmod p$. Observe that $2,~n,~m_p$ are all three coprime to $p$. Raise the first congruence to $u$ and the second to $v$ and multiply them to get
$$
n^u m_p^v2^{nu+n_pv}\equiv (-1)^{u+v}\pmod p.
$$
Hence, $p$ divides the numerator of the rational number
\begin{equation}
\label{eq:xy}
A:=n^u m_p^v 2^{nu+n_p v}-(-1)^{u+v}.
\end{equation}
Let us show that $A\ne 0$. Assume that $A=0$. Recall that $C_n=n_12^{n_2}+1$. Thus, expression \eqref{eq:xy} is
$$
A=n_1^u m_p^v 2^{(n+\alpha)u+n_pv}-(-1)^{u+v}=0.
$$
Then
$n_1^u m_p^v=1$, $(n+\alpha)u+vn_p=0$, and $u+v$ is even. Since $u\ge 0$, it follows that $v\le 0$. Put $w:=-v$, so $w\ge 0$. There exists a positive integer $\rho$ which is odd such that $n_1=\rho^{w}$ and $m_p=\rho^u$. Since $u$ and $v$ are coprime and $u+v$ is even, it follows that $u$ and $v$ are both odd. Hence, $w$ is also odd. Also, since $m_p$ divides $n_1$, it follows that $u\le w$. We now get
$$
(2^{\alpha} \rho^{w}+\alpha) u-w n_p=0,
$$
so
$$
\frac{u}{n_p}=\frac{w}{2^{\alpha} \rho^w+\alpha}.
$$
The left--hand side is $\ge u/n=u/(2^{\alpha}\rho^u)$, because $n_p\le n=2^{\alpha}\rho^u$. Hence, we get that
$$
\frac{u}{2^{\alpha}\rho^u}\le \frac{u}{n_p}=\frac{w}{2^{\alpha} \rho^w+\alpha}\qquad {\text{\rm leading~to}}\qquad \frac{u}{\rho^{u}}\le
\frac{w}{\rho^{w}+(\alpha/2^{\alpha})}\le \frac{w}{\rho^{w}}.
$$
For $\rho\ge 3$, the function $s\mapsto s/\rho^s$ is decreasing for $s\ge 0$, so the above inequality together with the fact that $u\le w$ implies that $u=w$ (so both are $1$ because they are coprime), and that all the intermediary inequalities are also equalities. This means that $u=w=1$,
$\alpha=0$ and $n=n_p$, but all this is possible only when $C_n=p$, which is not allowed. If $\rho=1$, we then get that $n_1=1$, so
every prime factor $p$ of $C_n$ is a Fermat prime. Hence, we get
$$
C_n=2^{n_2}+1=\prod_{i=1}^{k} (2^{2^{\gamma_{p_i}}}+1)=\sum_{I\subseteq \{1,\ldots,k\}} 2^{\sum_{i\in I} 2^{\gamma_{p_i}}},
$$
and $k\ge 2$, but this is impossible by the unicity of the binary expansion of $C_n$.
Thus, it is not possible for the expression $A$ shown at \eqref{eq:xy} to be zero.
The size of the numerator of $A$ is at most
\begin{eqnarray*}
& & 2^{1+|nu+n_pv|} n^{u} m_p^{|v|}\le 2^{1+3(n\log n)^{1/2}} n^{2(n/\log n)^{1/2}}\\
& < & 2^{1+3(n\log n)^{1/2}+(2/\log 2)(n\log n)^{1/2}}<2^{6(n\log n)^{1/2}}.
\end{eqnarray*}
In the above chain of inequalities, we used the fact that $3+2/\log 2<5.9$, together with the fact that $(n\log n)^{1/2}>10$ for $n\ge 30$.
Thus, for $n\ge 30$, we have that the inequality
\begin{equation}
\label{eq:p}
p<2^{6(n\log n)^{1/2}}
\end{equation}
holds for all prime factors $p$ of $C_n$.
Thus, we get the inequality
\begin{equation*}
2^n<C_n=\prod_{i=1}^k p_i<\prod_{i=1}^k 2^{6(n\log n)^{1/2}}=2^{6k(n\log n)^{1/2}},
\end{equation*}
leading to
\begin{equation}
\label{eq:low}
k>\frac{n^{1/2}}{6(\log n)^{1/2}}.
\end{equation}
Comparing estimates \eqref{eq:k} and \eqref{eq:low}, we get
$$
\frac{n^{1/2}}{6(\log n)^{1/2}}<1+2.4\log n,
$$
implying $n<6\times 10^5$.
It remains to lower this bound. We first lower it to $n<93000$. Indeed, first note that since $n<6\times 10^5$, it follows that if $p=F_{\gamma}=2^{2^{\gamma}}+1$ is a Fermat prime dividing $C_n$, then $\gamma\le 18$. The only such Fermat primes are for $\gamma\in \{0,1,2,3,4\}$. Furthermore, $(\log n)/(\log 3)\le \log(6\times 10^5)/(\log 3)=12.1104\ldots$ Hence, $k\le 5+12=17$. It then follows, by equation \eqref{eq:low}, that
$$
\frac{n^{1/2}}{6(\log n)^{1/2}}<17,
$$
so $n<122000$. But then $(\log n)/(\log 3)<\log(122000)/(\log 3)=10.6605\ldots$, giving that in fact $k\le 15$. Inequality \eqref{eq:low} shows that
$$
\frac{n^{1/2}}{6(\log n)^{1/2}}<15,
$$
so $n<93000$. Next let us observe that if $n$ is not a multiple of $3$, then relation \eqref{eq:mp} leads easily to the conclusion that the number of prime factors $p$ of $C_n$ with $m_p>1$ is in fact $\le (\log n)/(\log 5)=7.15338\ldots$. Hence, the number of such primes is $\le 7$, giving that $k\le 12$, which contradicts a result of Cohen and Hagis \cite{CH} who showed that every number with the Lehmer property must have at least $14$ distinct prime factors. Hence, $3\mid n$, which shows that $C_n$ is not a multiple of $3$. An argument similar to one used before proves that $n$ is not a multiple of any prime $q>3$. Indeed, for if it were, then relation \eqref{eq:mp} would lead to the conclusion that the number of prime factors $p$ of $C_n$ with $m_p>1$ is $\le 1+\log(n/q)/(\log 3)\le 1+\log(93000/5)/(\log 3)=9.94849\ldots$, so there are at most $9$ such primes. Also, $C_n$ can be divisible with at most $4$ of the $5$ Fermat primes $F_{\gamma}$ with $\gamma\in \{0,1,2,3,4\}$, because $3=F_0$ does not divide $C_n$. Hence, $k\le 9+4=13$, which again contradicts the result from \cite{CH}. Thus, $n=2^{\alpha} 3^{\beta}$ and so all prime factors $p$ of $C_n$ are of the form $2^{\alpha_1}3^{\beta_1}+1$ for some nonnegative integers $\alpha_1$ and $\beta_1$.
Now write
\begin{equation}
\label{eq:a}
a=\frac{C_n-1}{\phi(C_n)}=\prod_{i=1}^k \left(1+\frac{1}{p_i-1}\right)
\end{equation}
for some integer $a\ge 2$. Since
$$
\prod_{\substack{\alpha_1\ge 0,~\beta_1\ge 0\\ 2^{\alpha_1}3^{\beta_1}+1~{\text{\rm prime}}}}\left(1+\frac{1}{2^{\alpha_1} 3^{\beta_1}}\right)<1.46,
$$
we get that $a<2$, which is a contradiction. This shows that in fact there are no numbers $C_n$ with the claimed property.
\medskip
We end with some challenges for the reader.
\medskip
{\bf Reearch problem.} {\it Prove that $C_n$ is not a Camichael number for any $n\ge 1$.}
\medskip
If this is too hard, can one at least give a sharp upper bound on the counting function of the set ${\mathcal C}$ of positive integers $n$ such that $C_n$ is a Carmichael number? We recall that Heppner \cite{Hep} proved that if $x$ is large then the number of positive integers $n\le x$ such that $C_n$ is prime is $O(x/\log x)$, whereas in \cite{LuSh} it was shown that if $a>1$ is a fixed integer then the number of positive integers $n\le x$ such that $C_n$ is base $a$-pseudoprime is $O(x(\log\log x)/\log x)$. Clearly, imposing that $C_n$ is Carmichael (which is a stronger condition) should lead to sharper upper bounds for the counting function of such indices $n$.
\medskip
Finally, here is a problem suggested to us by the referee. Theorem \ref{thm:main} shows that $\phi(C_n)/\gcd(C_n-1,\phi(C_n))$ exceeds $1$ for all $n$. Can one say something more about this ratio? For example, it is possible that a minor modification of the arguments in the paper would show that this function tends to infinity with $n$, but we have not worked out the details of such a deduction. It would be interesting to find a good (large) lower bound on this quantity which is valid for all $n$ and which tends to infinity with $n$. How about for most $n$? What about lower and upper bounds on the average value of this function when $n$ ranges in the interval $[1,x]$ and $x$ is a large real number? We leave these questions for further research.
\medskip
{\bf Acknowledgements.} We thank the referee for a careful reading of the paper and for suggesting some of the questions mentioned at the end. F.~L. was supported in part
by grants PAPIIT 100508 and SEP-CONACyT 79685,
|
\section{Introduction}
\label{sec:intro}
The study of radio sources inside galaxy clusters is becoming
increasingly important in the context of large-area Sunyaev-Zel'dovich
effect \citep[SZE,][]{1972A&A....20..189S, 1999PhR...310...97B}
cluster surveys \cite[e.g.][]{2003NewAR..47..933S,
2009ApJ...694.1610H, 2010ApJ...722.1180V, 2010arXiv1012.1610M},
especially in the case of single-frequency SZE surveys and where the
resolution is matched with the typical cluster size. Powerful radio
sources are generally associated with early type galaxies, and as the
latter preferentially reside in clusters, it is expected that radio
sources can alter the SZE decrement in both position an depth.
Apart from contamination of SZE surveys, a study of the radio-loud
population of AGN inside galaxy clusters sheds light on the
interaction between cluster cooling flows and AGN heating/feedback
scenarios \citep[e.g.][]{2009ApJ...701...66M}. Analyzing the radio
properties of galaxy clusters can provide direct evidence of heating
of the intra-cluster medium (ICM) by AGN through various stages of
cluster formation \citep[e.g.][]{2009ApJ...705..854H}.
Three pieces of information are crucial to assess the radio source
contamination of SZE observations: (1) the distribution of sources as
a function of projected distance from the cluster center; (2) the
brightness distribution of sources and its redshift evolution; and (3)
the spectral energy distributions (SEDs) of the sources. A convenient
way of dealing with the first two points is to construct a radio
luminosity function (RLF), i.e. the number density of sources as a
function of luminosity, averaged over the cluster volume. This is the
objective of the present paper.
\citet{1996AJ....112....9L} first constructed the RLF for radio
galaxies in clusters using a 20 cm VLA survey of Abell clusters.
Similar studies were carried out by \citet{2004ApJ...600..695R} for
seven nearby galaxy clusters, and by \citet{2004A&A...424..409M} for a
much larger sample of 951 Abell clusters at low
redshifts. \citet{1999AJ....117.1967S} used a sample of 19 X-ray
selected clusters at $0.2<z<0.8$ to constrain the evolution of radio
galaxies, and comparing their results with those of
\citet{1996AJ....112....9L}, they found no evidence for an evolution
of the radio-loud AGN population. Similar conclusions were drawn by
\citet{2007AJ....134..897C} using a sample of massive clusters
detected through the SZE decrement at 30 GHz. Although
\citet{2006A&A...446...97B} found indications of redshift evolution in
both slope and amplitude of the cluster RLF when comparing VLA
observations of a sample of 18 X-ray selected clusters with the
results of Stocke et al., no definitive conclusions were drawn on the
RLF redshift evolution.
Recent high-resolution X-ray imaging of clusters have provided
evidence that the AGN fraction inside clusters rapidly evolves with
redshift \citep{2007ApJ...664L...9E, 2009ApJ...701...66M}. These
results are affected by small number statistics, since the AGN
fraction in any single cluster is very small and can also vary with
other cluster properties, such as the velocity dispersion
\citep{2008ApJ...682..803S}. The volume-averaged radio luminosity
function is therefore expected to be a more robust indicator of
redshift evolution. Correlating X-ray selected clusters with optical
and IR selected AGN, \cite{2009ApJ...694.1309G} confirmed that the AGN
excess near cluster centers increases with redshift. A parametric
modeling of this positive evolution, which can be used to predict the
radio AGN contamination in SZE cluster surveys, was not considered in
these studies.
\citet[][hereafter LM07]{2007ApJS..170...71L} used a sample of 573
X-ray selected clusters to estimate masses and virial radii, yielding
a more physically meaningful cluster volume for the RLF than obtained
when using a constant cluster radius. We adopt that method in this
paper. LM07 found a narrow radial distribution of radio sources, which
we reproduce in this work.
This study focuses on the redshift evolution of the RLF and its
decoupling from a possible dependence on cluster mass. We use publicly
available radio source catalogs at 1.4 GHz to identify radio sources
associated with clusters of galaxies in a large sample of optically
selected clusters \citep{2007ApJ...660..239K} and a composite sample
of ROSAT selected X-ray clusters. Studies of radio point sources in
clusters at higher frequencies were carried out by, e.g.,
\cite{1998AJ....115.1388C}, \cite{2007AJ....134..897C} and
\cite{2009ApJ...694..992L}, and the extrapolation of the RLF from 1.4
GHz to frequencies relevant for SZE surveys was investigated by, e.g.,
\citet{2004A&A...424..409M} and by LM07. These two aspects are not
pursued in this work.
This paper is organized as follows: in \sect\ref{sec:samples} we
describe the samples of radio sources and cluster of galaxies. We
discuss the estimation of cluster masses from observable properties
for clusters selected from optical and X-ray cluster catalogs, and
describe how radio luminosities are derived from flux densities. The
radial density profile of radio sources associated with clusters of
galaxies is discussed in \sect\ref{sec:profiles}. In
\sect\ref{sec:mlcorr}, we investigate the correlation between cluster
mass and the luminosity of the brightest radio source in the central
region of the cluster. The main part of the paper is
\sect\ref{sec:lf}, where the 1.4 GHz volume-averaged radio luminosity
in galaxy clusters is derived from our sample. The method of computing
the RLF is described, and the statistical treatment of sources not
associated with the clusters is discussed. Confusion of sources is
discussed, and the redshift evolution of the RLF is derived. We
summarize our main results and offer our conclusions in
\sect\ref{sec:concl}.
Where not otherwise noted, we use the cosmological parameters from the WMAP
5-year cosmology \citep{2009ApJS..180..330K} with $h=0.705$, $\Omega_{m} h^2 =
0.136$ and $\Omega_{\Lambda}=0.726$.
\section{Cluster and radio source samples}
\label{sec:samples}
In this section we describe the galaxy cluster samples and 1.4 GHz radio
source catalogs used in this paper.
\subsection{Cluster samples}
\label{sec:samples:clusters}
To demonstrate a redshift evolution of the RLF in the best possible
way, we select clusters with a wide range of redshifts from publicly
available optical and X-ray surveys. To make robust estimates of the
volume averaged RLF, we derive cluster masses and radii using scaling
relations. The mass within a region whose mean density is 200 times
the critical density of the universe at the cluster redshift,
$M_{200}$, has been found to be a good estimator of the virial mass
\cite[e.g.][]{2001A&A...367...27W}, and is used here together with the
corresponding radius, $r_{200}$, which we use to define the typical
scale of the cluster.
\subsubsection{Description of the samples}
\label{sec:samples:clusters:descr}
Because of its large sample size, we use as our main sample the
optical maxBCG catalog \citep{2007ApJ...660..239K}, which contains a
total of 13,823 clusters extracted from the Sloan Digital Sky Survey
(SDSS). This is by far the largest homogeneous publicly available
cluster sample, and its redshift range $0.1 < z < 0.3$, corresponding
to approximately $2.5 \times 10^9$ years of cosmic time, is large
enough to allow an investigation of the redshift dependence of the
RLF.
To allow a comparison with X-ray selected clusters, we have gathered
in one large sample most of the published cluster detections from the
ROSAT mission. Data products from this satellite remain the reference
to date since only few new massive systems have been reported from
more recent observatories. Our source list includes the ROSAT all-sky
survey catalogs from NORAS and REFLEX \citep{2000ApJS..129..435B,
2004A&A...425..367B}, and some serendipitous catalogs extracted from
the archives of pointed observations: the 160deg$^2$
\citep{1998ApJ...502..558V}, the 400deg$^2$
\citep{2007ApJS..172..561B} and WARPSI/II \citep{2007ApJS..172..561B,
2008ApJS..176..374H} catalogs. We also added the complete sample of
12 high redshift very luminous systems of \cite{2007ApJ...661L..33E}.
In total this yields 1177 X-ray selected clusters. The number of
clusters from the respective catalogs and the corresponding flux
limits are given in Table~\ref{tab:xraysel}.
\begin{table*}[Ht!]
\caption{Our composite X-ray cluster sample. The third column is the effective flux limit, in the [0.5-2] keV energy band, estimated from our recomputed source parameters.
}
\label{tab:xraysel}
\centering
\begin{tabular}{lrr}
\hline
\hline
Catalog & Number of clusters & Limiting flux ($\text{erg} \, \text{s}^{-1} \, \text{cm}^{-2}$) \\
\hline
REFLEX & 447 & 2.1e-12 \\
NORAS & 371 & 1.2e-12 \\
160deg2 & 221 & 8.0e-14 \\
400deg2 & 242 & 1.4e-13 \\
WARPS1+2 & 124 & 8.0e-14 \\
MACS(z$>$0.5) & 12 & 7.0e-13 \\
\hline
\end{tabular}
\end{table*}
Mixing these catalogs enables us to sample a large fraction of the
$L_X-z$ plane, as can be seen in Fig.~\ref{fig:xraysel}.
\begin{figure}
\centering
\includegraphics[width=\columnwidth]{lx.z.dist.eps}
\caption{Distribution of X-ray selected clusters of
galaxies in the $L_X-z$ plane.}
\label{fig:xraysel}
\end{figure}
\subsubsection{Cluster masses and radii}
\label{sec:samples:clusters:scaling}
For the maxBCG sample, we estimate halo masses from the scaled
richness parameter $n_{\text{gal}}^{R200}$. \cite{2007ApJ...660..239K}
and \cite{2007ApJ...669..905B} find that this measure correlates well
with the galaxy velocity dispersion, and thus also with halo
mass. Indeed, \cite{2010MNRAS.404.1922A} found that optical richness
and X-ray luminosity perform similarly well in predicting cluster
masses. The relation between $n_{\text{gal}}^{R200}$ and $M_{200}$ has
been extensively studied, both using weak lensing
\citep{2007arXiv0709.1159J,2008JCAP...08..006M,2009ApJ...703.2217S,
2009ApJ...699..768R} and comparing to X-ray luminosities
\citep{2008MNRAS.387L..28R, 2008ApJ...675.1106R, 2009ApJ...703..601R}.
In this paper, we use the X-ray luminosity$-$optical richness relation
($L_X - n_{\text{gal}}^{R200}$) found by \cite{2008ApJ...675.1106R} to
derive the expected X-ray luminosities as
\begin{equation}
\label{eq:rykoff}
L_{x}(R_{200}) = e^{\alpha} \left(
\frac{n_{\text{gal}}^{R200}}{40} \right)^{\beta} \left( \frac{1+z}{1.23} \right)^{\gamma} \times 10^{42} \, h^{-2}
\,\, \text{ergs} \,\, \text{s}^{-1},
\end{equation}
where $\alpha = 3.90 \pm 0.04$, $\beta = 1.85 \pm 0.05$ and $\gamma =
6.0 \pm 0.8$. We scale the derived X-ray luminosities to mass using a
self-similar redshift evolution, in the same way as described below
for the X-ray sample.
For the X-ray sample, the heterogeneity of the data makes it difficult
to straightforwardly estimate the cluster masses in a uniform way. The
most basic observable, available for all the sub-samples although in
different bands ($[b_{min},b_{max}]$), is the total cluster X-ray
flux. We thus apply the scaling relation of
\cite{1999MNRAS.305..631A}, linking the average gas temperature $T$ to
the bolometric X-ray luminosity $L_{\text{bol}}$, with a self-similar
redshift evolution to define a unique mapping between
($z$,$F_X$,$b_{min}$,$b_{max}$) and ($L_{bol}$,$T$). For this, we
make use of the APEC spectral model \citep{2001ApJ...556L..91S} with
heavy element abundances set to 0.3 times the solar values. The
derived cluster temperatures are translated into $M_{200}$ following
the scaling law of \cite{2005A&A...441..893A}, again assuming
self-similar evolution.
From simple geometric considerations and the definition of over-density with
respect to the critical density, it follows that $r_{200}$ is related to
$M_{200}$ by
\begin{equation}
\label{eq:m200r200}
M_{200} \equiv (4\pi/3) r_{200}^3 \, 200 \, \rho_c(z),
\end{equation}
where $\rho_c(z)$ is the critical density of the universe at redshift
$z$. $r_{200}$ is typically in the range $1-3$ Mpc for the clusters
in our samples.
\subsubsection{Cluster sample selection criteria}
\label{sec:samples:clusters:cuts}
In order to limit our survey to massive systems, we exclude all
clusters less massive than $5 \times 10^{13} M_{\odot}$ from our
sample. We also exclude low-redshift systems with $z<0.1$, in order to
have the same redshift cutoff in both samples. In addition, we define
a low-redshift ($0.05 < z < 0.12 $) sub-sample of the X-ray sample,
with the same mass cut (see Table \ref{tab:ccuts}), to investigate the
mass dependence of the RLF (\sect\ref{sec:lf:model:mdep}). The lower
redshift limit of the sub-sample is chosen so as to avoid excessive
overlap of low-redshift cluster fields. The upper redshift limit is
chosen as high as possible without including redshifts where the mass
limit of the flux-limited X-ray cluster sample is greater than our
mass limit of $5 \times 10^{13} M_{\odot}$, in order to avoid a bias
in the the determination of a possible mass dependence in the RLF. The
fact that the low-redshift sub-sample has a slight overlap with the
high-redshift X-ray sample is of little consequence as the two samples
are never compared directly with one another.
Not all cluster fields in our samples are covered by the NVSS and
FIRST radio surveys that are used for this study. Apart from taking
this into account, we note that due to dynamic range limitations of
the radio interferometric data, in the FIRST and NVSS catalogs some
regions on the sky around strong radio sources are plagued by
abnormally high or low source counts, often with catalog entries not
corresponding to real objects.\footnote{Typical examples are the
fields centered on the quasars 3C273 and 3C295} We identify these
regions by counting the number of sources in a circular region with
radius $1^{\circ}$ around each cluster center, and excluding fields
where the source counts exceed or fall below the average counts in
cluster fields by more than three standard deviations.\footnote{The
$1^{\circ}$ radius corresponds to approximately ten times $r_{200}$
at the median redshift and mass of the maxBCG sample, and is chosen
so as to include all cluster sources and to have a sufficient number
of background and foreground sources to allow for robust statistics}
Note that using the average counts of FIRST and NVSS would have
underestimates the expected counts in cluster fields, where a local
surface over-density is expected. Counting sources within one degree
of maxBCG clusters (cf. \sect\ref{sec:profiles}), we find the average
density of FIRST sources in such a region to be 9\% higher than the
catalog average of 90 sources per square degree. In the case of NVSS,
only about 4\% of the cluster fields are affected by this cut
(determined from the difference of rows 4 and 5 of
Table~\ref{tab:ccuts} in the maxBCG case).
We also exclude pairs \mtx{(or multiplets)} of clusters
where the sum of the radii ($r_{200}$ projected on the sky plane) is
greater than the angular separation between the cluster centers
\mtx{(row 2 of Table~\ref{tab:ccuts})}.
Table \ref{tab:ccuts} summarizes the selection criteria applied to the
cluster sample, and lists how many clusters in each of the final
samples are covered by the FIRST and NVSS catalogs.
\begin{table*}[Ht!]
\begin{minipage}[Ht]{\textwidth}
\caption{\mtx{Cluster sample selection criteria}. The
selection is cumulative in the sense that in rows 3-4, each row
assumes all the conditions of the previous rows. Rows 5 and 6
assume all the conditions in rows 2-4. The low-$z$ X-ray sample is
used only to constrain the mass dependence of the RLF
(\sect\ref{sec:lf:model:mdep}). Where not explicitly stated
otherwise, the high-redshift samples are used in this
paper.}
\label{tab:ccuts}
\centering
\renewcommand{\footnoterule}{}
\begin{tabular}{c|r|c|c|c|}
\cline{2-5}
& \textbf{Main sample} & \textbf{maxBCG} & \multicolumn{2}{|c|}{\textbf{X-ray}} \\
\cline{2-5} \cline{2-5}
\textsl{1} & clusters in main sample & 13823 & \multicolumn{2}{|c|}{1177} \\
\textsl{2} & clusters with sufficient separation & 12846 & \multicolumn{2}{|c|}{1121} \\
\cline{2-5}
& \textbf{Sub-sample} & & \textbf{high-$z$} & \textbf{low-$z$} \\
& Redshift range & $0.1 \leq z \leq 0.3$ & $0.1 \leq z \leq 1.26$ & $0.05 < z < 0.12 $ \\
\cline{2-5}
\textsl{3} & Clusters within redshift range & 12846 & 690 & 292 \\
\textsl{4} & Clusters with $M > 5 \times 10^{13} M_{\odot}$ & 12522 & 674 & 275 \\
\cline{2-5}
\textsl{5} & clusters with NVSS coverage\footnote{Excluding regions with abnormally high or abnormally low source
counts (see text)} & 12475 & 596 & 218 \\
\textsl{6} & clusters with FIRST coverage$^{a}$ & 11812 & 273 & 75 \\
\cline{2-5}
\end{tabular}
\end{minipage}
\end{table*}
\subsection{Radio source samples}
\label{sec:samples:radio}
\subsubsection{Description of the samples}
\label{sec:samples:radio:descr}
To search for radio sources associated with clusters, we use the FIRST
\citep{1995ApJ...450..559B, 1997ApJ...475..479W} and NVSS
\citep{1998AJ....115.1693C} 1.4 GHz radio continuum surveys. The most
relevant properties of the NVSS and FIRST surveys for the present work
are summarized in Table \ref{tab:rcata}. We make direct use of
publicly available source catalogs from each survey. While the NVSS
catalog has the advantage of greater sky coverage
(cf. Table~\ref{tab:rcata}), thereby covering more of our cluster
fields, it has the disadvantage of a poorer resolution, resulting in
increased source confusion.
\begin{table}[Ht!]
\begin{minipage}[Ht]{\columnwidth}
\caption{Properties of the FIRST and NVSS radio continuum surveys}
\label{tab:rcata}
\centering
\renewcommand{\footnoterule}{}
\begin{tabular}{r c c}
\hline\hline
& FIRST & NVSS \\
\hline
effective resolution & 5$^{\prime\prime}$ & 45$^{\prime\prime}$ \\
completeness limit & 1 mJy \footnote{catalog detection threshold} & $\sim$2.5 mJy \\
positional uncertainty\footnote{for the brightest sources} & $ < 0.5 ^{\prime\prime}$ & $ < 1 ^{\prime\prime}$\\
positional uncertainty\footnote{at the completeness limit/detection threshold} &
~1$^{\prime\prime}$ & $\sim$7$^{\prime\prime}$ \\
total number of sources & 816,331\footnote{as of July 16, 2008} & 1,810,672\footnote{as of February, 2004} \\
sources per square degree & $\sim$90 & $\sim$\mtx{53} \\
area covered (deg$^2$) & 9055 & \mtx{33885} \\
\hline
\end{tabular}
\end{minipage}
\end{table}
We search for FIRST and NVSS sources within a projected radius corresponding
to $3 r_{200}$ around each cluster center, with $r_{200}$ computed from
Eq.~(\ref{eq:m200r200}). In order to facilitate a robust comparison of
NVSS and FIRST sources, we exclude all sources fainter than 5 mJy, which is
well above the completeness limit of both surveys.
\mtx{In principle the use of the FIRST catalog alone would be enough
for the purpose of this paper. We make use of the NVSS catalog at
all stages of the analysis because it serves as a test that our
results obtained with FIRST are robust. In particular, the NVSS data
are important for testing our method of adaptively accounting for
confusion, as described in \sect\ref{sec:lf:confusion}. Because
the two surveys have significant overlap, we do not need to ``mix''
them in the sense of combining source counts from the two catalogs
for any particular sample.}
\subsubsection{Radio source luminosities}
\label{sec:samples:radio:lum}
To compute radio luminosities, we assume isotropic emission, which is
not correct for any individual galaxy, but is well justified when
averaging over the entire sample. The luminosity (power) of a radio
source is
\begin{equation}
L_{\mathrm{1.4 \, GHz}} \ = \ (4\pi \ D_L^2) \ S_{\mathrm{1.4 \, GHz}}
\ \frac{ {\cal K}(z) } { (1+z) },
\label{eq:lumeq}
\end{equation}
where $S_{\mathrm{1.4 \, GHz}}$ is the angular integrated flux density
taken from the VLA catalog (FIRST or NVSS), $D_L$ is the cosmological
luminosity distance and ${\cal K}(z)$ is the $k$-correction.
\mtx{The radio sources are modeled with continuum spectra of
the form $S_{\nu} \propto \nu^{-\alpha}$, where $\alpha$ is the
spectral slope. When computing the luminosity from flux, we have to
account for the fact that due to the redshift, the observed flux
corresponds to a rest frequency higher than 1.4 GHz. For the
computation of this so-called $k$-correction, ${\cal K}(z) $, we
assign a spectral slope of $\alpha=0.72$ to all sources, as
determined by \cite{2007AJ....134..897C} in the range $1.4-30$
GHz. The effect of this correction is about $10\%$ for sources at
$z=0.2$ and $30\%$ at $z=0.6$.}
When binning the data by luminosity to compute the RLF, we determine a
redshift limit, $z_{\rm{cut}}$, for each luminosity bin to avoid
counting cluster fields where the bin luminosity corresponds to a flux
below our chosen threshold of 5 mJy. $z_{\rm{cut}}$ is determined from
Eq.~(\ref{eq:lumeq}) and only affects the lowest luminosities
considered in the analysis.
\section{Radial density distribution of radio sources}
\label{sec:profiles}
As we have no redshift information for individual radio continuum
sources, the over-density of sources toward clusters must be
quantified statistically. We construct the stacked radial profile of
radio sources around cluster centers. Following LM07, we take
$r/r_{200}$ as the radial coordinate. The resulting radial profile
includes both cluster galaxies and the field (background and
foreground) population. Although it cannot be decided whether
individual sources are cluster members, cluster and field sources can
be separated statistically, since the field population has a radially
constant contribution to the radial profile. \mtx{Besides determining
the stacked radial density distribution of all radio sources, we
also investigate whether the distribution depends on source
luminosity and host galaxy cluster mass.}
The radial profiles are used when constructing the luminosity function by
de-projection into volume number density as described in
\sect\ref{sec:lf:method:deproj}.
\subsection{Radial model}
\label{sec:profiles:model}
\mtx{The angular offset of radio sources with respect to the cluster
center (defined as the position of the brightest cluster galaxy
(BCG) in the maxBCG sample) is translated into projected physical
distance at the cluster redshift.} The physical radial distances are
stacked for all cluster fields and binned by radius.
We fit the radial profile of radio sources in each cluster sample by a
parametric model. In general, all our radial profiles are well fit by
a model of the form
\begin{equation}
\psi(\xi) = \psi_{\beta} \left(1+\frac{\xi^2}{\zeta^2}\right)^{-\frac{3}{2}\beta + \frac{1}{2}} \, + \,
\frac{1}{\xi}\psi_{G} e^{-(\frac{\xi}{2 \, \sigma})^2} \, + \,
\psi_f,
\label{eq:proffitfun}
\end{equation}
where $\psi(\xi)$ is the projected surface density and $\xi =
r/r_{200}$. The first term of (\ref{eq:proffitfun}) has the form of an
isothermal $\beta$-model \citep{1978A&A....70..677C}, the second term
is a Gaussian, corrected for the area in an annulus, and the third
term, $\psi_f$, is the constant field density.
The isothermal $\beta$-model is described by three parameters: the
(peak) normalization $\psi_{\beta}$, the scale radius $\zeta$, and the
power law index $\beta$. \mtx{The $\beta$-model is chosen because it
provides a good fit to our radially binned data outside $r/r_{200}
\simeq 0.006$.}
In our fractional radial coordinates, the constant field density,
$\psi_f$, is a function of the redshift distribution of the galaxy
cluster sample. It is not physically meaningful as we have
(incorrectly) placed all radio sources at the cluster redshifts, but
as we are not interested in the field population this component may
safely be ignored.
The second term in Eq.~(\ref{eq:proffitfun}) accounts for an
additional peaked feature in the radial distribution of FIRST sources
in the maxBCG galaxy cluster sample (Fig.~\ref{fig:prof:maxbcg}). It
is well fit by a Gaussian with normalization $\psi_{G}$ and variance
$\sigma^2$. The origin of this feature is the fact that the cluster
center is taken as the position of the BCG. As discussed further in
\sect\ref{sec:profiles:res}, the underlying distribution \mtx{derives
from the extended nature of the large majority of FIRST radio
sources associated with the BCG.} Note that the central Gaussian
component of $\psi(\xi)$ is considered only for the radial
distribution of FIRST sources in the maxBCG sample; in all other cases
we set $\psi_G = 0$.
For the purpose of constructing the RLF, the exact form of the fitting
function is not important; although the density profile is used in
constructing the RLF (as discussed in \sect\ref{sec:lf:method:deproj}), the
RLF is only very weakly sensitive to the parameterization of the source
density fitting function.
\subsection{Dependence on luminosity and cluster mass}
\label{sec:profiles:lmdep}
\mtx{Because the radial distribution of sources plays an important
role in determining cluster member radio source counts within
$r_{200}$, it is necessary to investigate whether bright radio
sources have a radial distribution different from the distribution
of faint sources. We compute the luminosities of FIRST sources in
maxBCG clusters according to Eq.~(\ref{eq:lumeq}) by placing all
radio sources at the cluster redshifts and binning the sample in
luminosity. We determine the radial source distribution in each
luminosity bin using the fitting function given by
Eq.~(\ref{eq:proffitfun}). Similarly, we investigate a possible
dependence on cluster mass by binning the cluster sample in mass.}
\mtx{Because the sample has been divided and uncertainties on individual
profiles are greater, we fix the power law parameter $\beta$ in
Eq.~(\ref{eq:proffitfun}) to the best-fit value $\beta=0.987$ from the
total fit (Table~\ref{tab:profres}) in order to get reliable estimates on
the scale radius $\zeta$.}
\subsection{Results}
\label{sec:profiles:res}
\subsubsection{Ensemble properties}
\label{sec:profiles:res:ensamble}
Figure~\ref{fig:prof:maxbcg} shows the source density profiles in
maxBCG clusters, fitted with the parametric form (\ref{eq:proffitfun})
and divided into its components.
\begin{figure}
\centering
\includegraphics[width=\columnwidth]{profiles.first.nvss.maxbcg.eps}
\caption{Radial distribution of 1.4 GHz radio sources brighter than 5
mJy in the maxBCG cluster sample. The normalization on the $y-$axis is
arbitrary. Radio sources selected from the FIRST catalog (triangles with
error bars) clearly indicate three components: a narrow central peak,
fitted with a Gaussian (dotted line); a broader distribution of sources,
fitted with a $\beta$-profile (short-dashed line) and a
background/foreground component (dash-dotted line). The sum of the
components is indicated by the solid line. The radial profile derived from
the NVSS catalog is indicated by the error bars without symbols. The fit
to these data (red long-dashed line) does not include a Gaussian
component.}
\label{fig:prof:maxbcg}
\end{figure}
Table \ref{tab:profres} indicates the most important parameters in the
radial fits of radio sources associated with clusters.
\begin{table*}
\caption{Results of $\beta$-model fits to the radial distribution of sources
associated with clusters of galaxies. The data are fitted using Eq.~(\ref{eq:proffitfun}). }
\label{tab:profres}
\centering
\begin{tabular}{l r r@{}l r@{}l r@{}l r@{}l}
\hline\hline
Cluster sample & Source catalog & $\beta$& & $\zeta$& & $\psi_G$& & $\sigma$& \\
\hline
maxBCG & FIRST & 0.987$\pm$&0.018 & 0.0115 $\pm$&0.0012 & 3.23$\pm$&0.18 & 0.00274$\pm$&0.00013 \\
maxBCG & NVSS & 1.018$\pm$&0.010 & 0.00257$\pm$&0.00019 & $-$& & $-$& \\
X-ray & FIRST & 0.899$\pm$&0.136 & 0.0191 $\pm$&0.0099 & $-$& & $-$& \\
X-ray & NVSS & 0.911$\pm$&0.083 & 0.0198 $\pm$&0.0063 & $-$& & $-$& \\
\hline
\end{tabular}
\end{table*}
\mtx{Note that the parameters ($\psi_G,\sigma$) of the
Gaussian were fit only for the FIRST sources associated with maxBCG
clusters, as this is the only combination of cluster and source
catalog that yields an additional centrally peaked feature. The
width of the peaked distribution ($\sim 3 \times 10^{-3} ~r_{200}$)
corresponds to 0.9$\asec$ at the median distance of the maxBCG
clusters. Although this is comparable to the maximum positional
uncertainty of FIRST point sources ($\sim1\asec$) at the
completeness limit, the FIRST positions are typically constrained to
a much better accuracy, in particular at higher flux levels. Thus,
positional uncertainties in the FIRST survey alone cannot explain
the width of the peaked distribution.}
\mtx{ A visual inspection of a large fraction of $\sim$1800 FIRST
sources with positional matches out to 15$\arcsec$ of the maxBCG
positions revealed that the very close matches ($<~0.3\asec$) are
dominated by compact radio sources, while the matches in the range
$~3\asec$--$10\asec$ can be classified into three broad categories:
(a) double sources with more or less symmetrical lobes straddling
the BCG position, (b) triple sources with weak cores (centered on
the BCG) and strong lobes, for which our threshold of 5~mJy picked
up only the lobes, and (c) complex sources, usually of FRII or
wide-angle tail type, some of which have truly spectacular radio
morphologies and sizes up to 1$\amin$ or more. From this inspection
we estimate that out to 10$\asec$ separation ($\sim 0.3 r_{200}$ at
the median redshift) at least 98\% of our matches correspond to the
BCGs (with at most ~10\% of the sources being duplicates, i.e.\
different components or lobes of the same complex source). While for
separations between 10 and 15$\arcsec$ (i.e.\ up to ~54 kpc at the
median redshift of the clusters of z=0.227) the fraction of
duplicates exceeds 30\%, our matches can be safely related to the
BCGs in at least 90\% of the cases. We conclude that the peaked
component in the radial distribution originates from mismatches in
SDSS and FIRST positions caused by extended radio emission.}
Because the BCG component can be separated from the general
distribution of FIRST sources in the maxBCG sample, the dimensionless
core radius, $\zeta$, for the remaining sources is much larger than
for the NVSS sources, where the peaked central distribution is
``hidden'' in the $\beta$-model component. The distribution of sources
in the X-ray cluster sample is less peaked (larger $\zeta$) because
the cluster center definition is based on the gas distribution and not
the position of the BCG.
We note that our results pertaining to FIRST sources in maxBCG
clusters verify the radial distribution derived by
\cite{2007ApJ...667L..13C} for the same sample. As indicated in
Fig.~\ref{fig:prof:maxbcg}, our narrow binning reveals the details of
the centrally peaked component in detail and confirm its origin as the
BCG population. \mtx{The relatively high redshifts in the
optical sample ($z \geq 0.1$) also imply that the population of star
forming galaxies in clusters -- with a broader distribution than
radio-loud AGN -- has been excluded by our flux limit of 5 mJy. As a
consequence our $\beta$-model fit to the radial profile is much
narrower compared to those obtained for the radio source
distribution in local clusters \citep[e.g.][]{2004ApJ...600..695R,
2004A&A...424..409M}. Similar narrow radial profiles consisting
mostly of radio-loud AGN were also found by LM07.}
\subsubsection{Luminosity and cluster mass dependence}
\label{sec:profiles:res:lmdep}
\mtx{From their X-ray sample of galaxy clusters, LM07 found that more
luminous radio sources are more centrally concentrated in galaxy
clusters. As indicated in Table \ref{tab:profres:lbins}, we are
unable to reproduce this result with our optical maxBCG sample;
although the highest luminosity bin has a slightly smaller scale
radius $\zeta$ of the outer profile, the result is not
significant. Similarly, there is no statistically significant
indication of a mass dependence of the radial density from the
maxBCG sample of galaxy clusters (Table \ref{tab:profres:mbins}). We
verify that the same is the case when using the NVSS sample. }
\begin{table}
\caption{Results of $\beta$-model fits to the radial distribution of FIRST
sources
associated with maxBCG clusters of galaxies, with the radio sources binned
by luminosity $L$. Only the parameters $\zeta$ and $\sigma$, pertaining to
the width of the profiles, are listed. $\beta$ has been fixed to the
best-fit value
0.987 from the total fit. The data are fitted using
Eq.~(\ref{eq:proffitfun}).}
\label{tab:profres:lbins}
\centering
\begin{tabular}{r@{}l r@{}l r@{}l}
\hline\hline
Luminosity& ~range & $\zeta$& & $\sigma$& \\
\hline
$\log(L \,[\text{W/Hz}])$& ~$<$ 23.8 & 0.0116 $\pm$& 0.0018 & 0.00249 $\pm$& 0.00067 \\
23.8 $\leq$ $\log(L \,[\text{W/Hz}])$& ~$<$ 24.2 & 0.0138 $\pm$& 0.0023 & 0.00319 $\pm$& 0.00054 \\
24.2 $\leq$ $\log(L \,[\text{W/Hz}])$& ~$<$ 24.6 & 0.0138 $\pm$& 0.0024 & 0.00343 $\pm$& 0.00073 \\
$\log(L \,[\text{W/Hz}])$& ~$ \geq $ 24.6 & 0.0097 $\pm$& 0.0018 & 0.00303 $\pm$& 0.00088 \\
\hline
\end{tabular}
\end{table}
\begin{table}
\caption{As table~\ref{tab:profres:lbins}, but with the sample divided into
mass bins.}
\label{tab:profres:mbins}
\centering
\begin{tabular}{r@{}l r@{}l r@{}l}
\hline\hline
Mass& ~range & $\zeta$& & $\sigma$& \\
\hline
$\log(M/M_{\odot})$& ~$<$ 13.90 & 0.0132 $\pm$& 0.0016 & 0.00334 $\pm$& 0.00060 \\
13.90 $\leq$ $\log(M/M_{\odot})$& ~$<$ 14.15 & 0.0139 $\pm$& 0.0017 & 0.00344 $\pm$& 0.00075 \\
14.15 $\leq$ $\log(M/M_{\odot})$& ~$<$ 14.40 & 0.0125 $\pm$& 0.0011 & 0.00260 $\pm$& 0.00053 \\
$\log(M/M_{\odot})$& ~$ \geq $ 14.40 & 0.0109 $\pm$& 0.0011 & 0.00209 $\pm$& 0.00049 \\
\hline
\end{tabular}
\end{table}
\section{Mass-luminosity correlation}
\label{sec:mlcorr}
In this section we investigate the correlation between cluster mass and the
radio luminosity of the brightest radio source in the central region of the
cluster. \mtx{We compare the results with similar studies done with the
optical luminosities of the brightest cluster galaxies, and discuss the
implications for our goal of comparing the volume averaged radio luminosity
function in clusters at different redshifts.}
\subsection{Background}
\label{sec:mlcorr:back}
\mtx{Brightest cluster galaxies (BCGs) tend to lie at the center of
the mass distribution and are more luminous than average
ellipticals. Moreover, their properties correlate strongly with
their host clusters, as seen from numerical simulations
\citep[e.g][]{2007MNRAS.375....2D} and optical/near-IR observations
\citep[e.g.][]{2004ApJ...617..879L,2008MNRAS.385L.103B}. It has been
shown that BCGs are about an order of magnitude more likely than
other ellipticals to host radio-loud AGN \citep{1991MNRAS.250..103E,
2004ApJ...617..879L, 2007MNRAS.379..894B}.} \mtx{Based on this,
it is likely that the most radio-luminous galaxies in clusters are
associated with BCGs, and in addition to the correlation between BCG
luminosity and cluster mass one can expect that 1.4 GHz radio
luminosities of the central radio-loud AGN are also correlated with
cluster mass.}
\subsection{Method}
\label{sec:mlcorr:method}
We investigate the correlation of BCG radio luminosity with cluster
mass by taking the luminosity of the single brightest source within a
radius of 50 kpc from the cluster center. We reiterate that the latter
is actually the BCG position in the maxBCG sample. In the X-ray
sample, the search radius allows for offsets between the X-ray center
and the BCG position; \cite{2007ApJ...660..239K} found typical offsets
of $~50 \, h^{-1}$ kpc by cross-correlating the maxBCG catalog with
NORAS and REFLEX data.
\mtx{ Note that this selection method leads to a different minimum flux
level for different redshift sub-samples, since we are disregarding clusters
that do not contain any radio source above a certain flux density. This
leads to a higher mean luminosity of the central radio source for clusters
at higher redshift, compared to clusters at lower redshifts within the same
mass bin (derived from the X-ray or optical mass observables). The actual
redshift evolution of the radio luminosities of the central BCGs is hidden
within this luminosity bias. Considering the large scatter in the
luminosity of the central radio source, we do not attempt to model the
redshift evolution from this method. }
\subsection{Results}
\label{sec:mlcorr:res}
Figure \ref{fig:mlcorr} shows the correlation of luminosities of FIRST sources
with cluster mass for both our cluster samples.
\begin{figure}[ht!]
\centering \includegraphics[width=\columnwidth]{./mlcorr.eps}
\caption{Correlation between radio luminosity of the brightest FIRST
source within a projected physical radius of 50 kpc from the
cluster center (BCG position in the maxBCG sample) and cluster
mass for three sub-samples of galaxy clusters. Diamonds indicate
maxBCG clusters in the redshift range $0.1 < z \leq 0.2$, and the
best-fit correlation is indicated by the dotted
line.
This is consistent with the X-ray sample in the same redshift
range (triangles). Fixing the slope, the correlation is
consistent with a local sub-sample of X-ray clusters ($0.01 < z
\leq 0.1$, squares), where the luminosity level is lower due to a
different luminosity cutoff in this sub-sample.}
\label{fig:mlcorr}
\end{figure}
Modeling the correlation with a power-law of the form $L \sim
M^{\alpha}$, the best-fit correlation from the maxBCG sample in the
redshift range $0.1 < z \leq 0.2$ is $\alpha = 0.31 \pm
0.12$. Although the X-ray sub-sample in the same redshift range is
essentially too small to constrain the power law, the data are
consistent with the maxBCG result, with $\alpha = 0.34 \pm 0.41$. The
low-redshift X-ray sub-sample yields $\alpha = 0.44 \pm 0.27$, again
consistent with the maxBCG relation.
\mtx{The found correlation between radio luminosity and cluster mass
is consistent with results from optical/near-IR observations of the
BCGs. \cite{2004ApJ...617..879L} compared the K-band near-IR
luminosities for a sample of X-ray selected clusters and found
$L_{\mathrm{BCG}} \propto M_{cl}^{0.26\pm
0.04}$. \cite{2010ApJ...713.1037H} reported a shallower
correlation from an X-ray selected sample; $L_{\mathrm{BCG}} \propto
M_{cl}^{0.18\pm 0.07}$. Thus, luminosities in other bands show
correlations with mass that are similar to our results. }
\mtx{For the optical cluster sample, the correlation of radio
luminosity and cluster mass strongly points towards increased radio
emission from the central AGN in the BCGs of more massive clusters.
\cite{2007ApJ...667L..13C} also computed the rest-frame 1.4 GHz
luminosities of the BCGs by cross-correlating the maxBCG catalog
with the FIRST survey, and noted the trend of increasing radio
luminosity with optical luminosity of the BCGs in the $r$-band. We
do not expand on this result further, as the radio luminosity of the
central brightest radio source is a poor indicator of the total
radio luminosity in clusters. \mtx{A cluster can have multiple BCGs
as a result of its merging history,}\footnote{\mtx{After a cluster
merger, it will take considerable time before the two initial
BCGs cease to dominate their subgroups and one of them comes to
dominate the merged cluster}} \mtx{and the radial source density
profile in Fig.~\ref{fig:prof:maxbcg} suggests that the radio
source population extends beyond the cluster core region. For
these reasons, we now turn to the construction and modeling of the
volume averaged radio luminosity function in galaxy clusters.}
\section{The radio luminosity function}
\label{sec:lf}
\mtx{ In this section, the radio luminosity function (RLF) is derived
for each of our samples. We start by discussing how accurate source
counts in the cluster volume are obtained
(\sect\ref{sec:lf:method}). In \sect\ref{sec:lf:confusion} we
describe how source confusion is accounted for. We describe a
parametric model for the RLF in \sect\ref{sec:lf:model}, and also
discuss how a redshift evolution is modeled. We give the results in
\sect\ref{sec:lf:res}. }
\subsection{Method}
\label{sec:lf:method}
The luminosity function $\phi(L)$ is defined as the average number of
radio sources per unit physical volume of a cluster and per
logarithmic luminosity bin (``magnitude'') . We use the estimated
values of $r_{200}$ from \sect\ref{sec:samples:clusters} to define
cluster volumes, expressed in $\rm{Mpc}^3$.
\subsubsection{Source counts}
\label{sec:lf:method:counts}
Figure \ref{fig:lf:cartoon}
\begin{figure}[ht!]
\vspace*{5mm}
\includegraphics[width=\columnwidth]{./lfdef.eps}
\caption{Illustration of the
method used to obtain source counts and volumes for the computation of the
luminosity function of radio sources inside clusters. For each cluster the
sampled volume is the ``line-of-sight cylinder''. The projected number
density in this volume is converted to volume density of sources inside the
cluster radius ($r_{200}$) by applying a model for the volume density of
sources.}
\label{fig:lf:cartoon}
\end{figure}
illustrates schematically how source counts are obtained. For each
cluster, a region within a projected physical radius $r = \eta \,
r_{200}$, from the center is searched for radio sources in a given
luminosity bin, using the cluster redshift to convert from flux to
luminosity. As explained below, the search radius does not have a
systematic effect on the RLF; however, it can be chosen such as to
minimize the uncertainties of the latter. The number of sources found
within the chosen radius is $N_{\eta}^{\rm{count}}$. We need to
convert this number into a de-projected number of sources within
$r_{200}$. For this, we need to know (i) the number of
foreground/background sources, and (ii) the radial source density
profile. In the following we discuss these issues in turn.
\subsubsection{Field subtraction}
\label{sec:lf:method:field}
To correctly model the field counts, we first need to understand the mean
surface number density of radio sources in the sky as a function of flux. For
this we bin the FIRST and NVSS catalogs in their entirety in logarithmic flux
bins to derive an estimate of the number of sources per flux interval and
solid angle, $\frac{dN}{dS \, d\Omega}$. Due to the large sample size of both
NVSS and FIRST, this quantity is well constrained except at very low and very
high flux densities. Because it is difficult to find a generally valid fitting
function, we model $\frac{dN}{dS \, d\Omega}$ using quadratic spline
interpolation between bins.
The number of foreground/background sources,
$N_{\eta}^{\mathrm{field}}$, inside the search radius
of a given cluster is determined by integrating over the \mtx{field
density}\footnote{\mtx{We approximate the \textit{field} as the sum
of all regions on the sky covered by the radio survey, minus
circular regions around the sample clusters, defined by the
respective radii ($r_{200}$) estimated from scaling relations.}}
as
\begin{equation}
N_{\eta}^{\mathrm{field}} = \Omega \int_{S_{\rm{min}}}^{S_{\rm{max}}} \frac{dN}{dS \, d\Omega} dS,
\end{equation}
where $\Omega$ is the angular area of the searched region, and $S_{\rm{min}}$
and $S_{\rm{max}}$ are determined from the limiting values of the luminosity
bin according to Eq.~(\ref{eq:lumeq}).
The total number of cluster sources inside the search radius is
\begin{equation*}
N_{\eta}=N_{\eta}^{\rm{count}}-N_{\eta}^{\mathrm{field}}.
\end{equation*}
\mtx{Uncertainties on $N_{\eta}$ are computed from Poisson statistics
as discussed by \cite{1986ApJ...303..336G}.}
For reasons of simplicity, the previous discussion has focused on source
counts in a single cluster field. However, it is straightforward to apply the
method discussed here to the entire stack of clusters by adding all individual
$N_{\eta}$.
\subsubsection{De-projection into cluster volume}
\label{sec:lf:method:deproj}
To convert the cluster source counts $N_{\eta}$ inside the ``line-of-sight
cylinder'' (cf. Fig.~\ref{fig:lf:cartoon}) limited by $\eta \, r_{200}$ to
counts $N$ inside $r_{200}$, we compute the expected ratio of sources inside
the sphere delimited by $r_{200}$ and inside the cylinder. We integrate the
spatial source density, as a function of de-projected radius, over the two
regions and take the ratio
\begin{equation}
C_N(\eta) = \frac{N_{\text{sphere}}}{N_{\text{cylinder}}} =
\frac{4 \pi \int \limits_{\Xi=0}^{1} \Psi(\Xi) \, \Xi^2 \, d\Xi}
{2 \pi \int \limits_{z=-\infty}^{+\infty} \int \limits_{\rho=0}^{\eta}
\Psi\left(\sqrt{\rho^2 + z^2} \, \right) \, \rho \, d\rho dz},
\label{eq:deproject}
\end{equation}
where $\Psi(\Xi)$ is the de-projected counterpart of $\psi(\xi)$
(cf. Eq.~\ref{eq:proffitfun}). Note that $\Xi = r/r_{200}$, with $r$
now being the de-projected (physical) coordinate. The integral over
the line-of-sight cylinder (in the denominator) is written in
cylindrical polar coordinates, with $\sqrt{\rho^2 + z^2} = \Xi$.
\mtx{The de-projected source distribution $\Psi(\Xi)$ is derived from
the projected source density distribution $\psi(\xi)$, discussed in
\sect\ref{sec:profiles}. Because no dependence on luminosity could
be determined for $\psi(\xi)$, we use the parameters derived from
the total sample for each combination of cluster and radio source
catalogs, as listed in Table~\ref{tab:profres}.}
Because the central Gaussian component of $\psi(\xi)$ in
Eq.~(\ref{eq:proffitfun}) does not correspond to a physical
distribution of sources (cf. \sect\ref{sec:profiles}), it is not
physically meaningful to de-project it to physical
coordinates. However, because the component is narrow, we can make the
approximation that all sources belonging to it (i.e. all BCGs) have
been counted inside the line-of-sight cylinder, provided $\eta$ is
chosen large enough ($\eta \gtrsim 0.1$). Thus, no correction in the
sense of Eq.~(\ref{eq:deproject}) is required for the central
component.
\mtx{We compute a correction factor only for the broader
$\beta$-model. In order to avoid over-correction where the narrow
component is present, we compute from (\ref{eq:proffitfun}) the
relative fraction of sources belonging to the broader profile and
multiply the correction factor by this number to yield a modified
correction factor, $\tilde{C}_N$}.
For the isothermal $\beta$-model, the physical (de-projected) density
of the sources is given by \citep[e.g.][]{1988xrec.book.....S}
\begin{equation}
\Psi(\Xi) = \Psi_{\beta} \left( 1+ \frac{\Xi^2}{\zeta^2} \right) ^ {-\frac{3}{2}\beta},
\label{betavol}
\end{equation}
which is the function we use when computing the integrals in
Eq.~(\ref{eq:deproject}). The parameters $(\beta,\zeta)$ are unique
for each galaxy cluster sample, as discussed in
\sect\ref{sec:profiles}. Because we are only interested in ratios, we
set $\Psi_{\beta}=1$.
The cluster volume is defined as a sphere with radius $r_{200}$. Thus, the RLF
in luminosity bin $\Delta L$ can be expressed as
\begin{equation}
\phi(L) \Delta L \ = \tilde{C}_N \frac{N_{\eta}}{V_{\rm{sph}}},
\end{equation}
where $V_{\rm{sph}}$ is the sum of physical cluster volumes (within
$r_{200}$), and $N_{\eta}$ is the total number of sources (in all
cluster fields) found within a radius $\eta \, r_{200}$ and having
luminosities in the luminosity bin $\Delta L$.
The uncertainty in the RLF is dominated by Poissonian noise in the
source counts. We compute the uncertainty by scaling from the
uncertainty $\delta N_{\eta}$ in $N_{\eta}$ as
\begin{equation*}
\frac{\delta (\phi(L) \Delta L)}{\phi(L) \Delta L} =
\frac{\delta N_{\eta}}{N_{\eta}}.
\end{equation*}
By construction, the RLF is insensitive to the chosen value of the projected
fractional radius $\eta$ after the correction factor $\tilde{C}_N$ has been applied.
However, uncertainties in the RLF can be minimized by carefully tuning this
parameter. Values too close to 1 will increase the background counts, and
thereby the error, since there are few cluster sources close to $r_{200}$. On
the other hand, too low values of $\eta$ will increase uncertainties due to
not providing enough cluster sources for good statistics. We find that
$\eta=0.5$ provides a good balance between cluster and background sources, and
use this value in computing the RLF.
\subsection{Confusion}
\label{sec:lf:confusion}
Confusion becomes important when the typical separation of sources is
comparable to the angular resolution of the radio survey from which the RLF is
derived. In particular, a larger beam tends to overestimate the RLF at the
high luminosity end as several confused low luminosity sources appear as fewer
sources with higher luminosity. This effect is redshift dependent through the
conversion of angular to physical distance, which complicates quantifying a
possible redshift evolution in the RLF.
Comparing the narrow radial distribution of sources (section
\ref{sec:profiles}) with the resolution of the FIRST survey, it is
apparent that confusion affects the counts also when using FIRST data.
This will affect the shape and normalization of the RLF in a way that
we have no direct way of quantifying. However, our main interest is
to constrain the redshift evolution of the RLF, and for this purpose
the absolute normalization of the RLF is of less importance.
Therefore, rather than attempting to correct for confusion, \mtx{we
can correct for the relative difference in confusion at different
redshifts.}
\subsubsection{Degrading the resolution of the radio data}
\label{sec:lf:confusion:degrade}
To address the confusion problem, we thus degrade the resolution of the
radio source data in an adaptive way to make sure that confusion effects are
the same at all redshifts (disregarding second order effects such as a number
density evolution in the central parts of clusters influencing the confusion
problem). In the following, we discuss how the resolution of the radio source
data can be degraded. To demonstrate the method, we apply it to the FIRST
catalog to enable a direct comparison of the RLF derived from FIRST to that
derived from NVSS, before discussing how the method is used in removing
first-order systematic effects of confusion in the redshift evolution of the
RLF.
The full-width half maxima (FWHM) of the FIRST and NVSS synthesized beams are
5$\arcsec$ and 45$\arcsec$, respectively. To what extent individual sources
can be distinguished in each of the surveys, however, largely depends on their
absolute and relative brightnesses. \mtx{Here we make the simplification
that given a resolution in terms of a beam FWHM, no two sources are resolved
if their mutual separation (irrespective of flux) is less than the FWHM.}
To degrade the resolution of the FIRST or NVSS data, we define a beam
with a FWHM larger than that of the original data. In a given field,
the position and flux density of each source are recorded. The
brightest source is located, and all sources within a radius equal to
the FWHM around this source are combined into one source. The position
of the new source is taken as a flux-weighted average of its parts,
and the total flux density is the sum of the integrated flux densities
of the parts. The next brightest source (excluding all sources that
have already been considered) is then located, and the procedure is
repeated until no two sources are separated by an angular distance
less than the FWHM. Note that this method is limited by the use of
source catalogs in the sense that sources fainter than a chosen limit
will not be considered. Here we use the completeness limit of the two
source catalogs to determine which sources to consider for summing.
As an improved method, one could instead use the raw maps of cluster
fields from the FIRST survey and degrade them to the NVSS resolution
before extracting sources. However, this requires a robust source
extraction in both maps and is quite complicated for faint sources
near the completeness limit of the surveys. \mtx{Moreover, neither
such a method nor the method described above can resolve the
inherent problem that some extended emission recovered by NVSS is
resolved out by FIRST.}
\subsubsection{NVSS and FIRST source counts}
\label{sec:lf:confusion:nvssfirst}
A simple way to visualize the effect of confusion on the RLF is to
compare the luminosity function as constructed from NVSS with that
constructed from FIRST, using the same completeness cutoff in both
surveys. For this comparison we select cluster fields in a narrow
redshift range, $0.1 \leq z \leq 0.16$, from the maxBCG catalog to
ensure a sufficiently large statistical sample. \mtx{This
sub-sample of the maxBCG catalog is used only for the purpose of
directly comparing luminosity functions constructed from FIRST and
NVSS in this section. It contains 2341 clusters with $M_{200} > 5
\times 10^{13} M_{\odot}$ which are covered by NVSS and FIRST.} As
indicated in Fig.~\ref{fig:nvss:first}, computing the RLF directly
from the two radio source samples yields inconsistent results. This is
expected due to confusion effects.
We attempt to re-create the NVSS based luminosity function from the
FIRST data by degrading the resolution as described above. \mtx{Note
that we do not expect a perfect agreement since the sensitivity to
spatial frequencies of FIRST is very different from that of NVSS. }
Because both surveys have many sources separated by distances much
smaller than the respective resolutions (due to the specific source
extraction methods applied to the synthesized images), we impose a
strict lower limit of 45$^{\prime\prime}$ on the separation of sources
in \textit{both} surveys for this comparison. The results are shown in
Fig.~\ref{fig:nvss:first}.
\begin{figure}[ht!]
\centering \includegraphics[width=\columnwidth]{./nvss.first.comparison.eps}
\caption{Change in the RLF resulting from degrading the resolution
of both the FIRST and NVSS catalogs to a minimum source separation
of 45$^{\prime\prime}$ (the FWHM of the synthesized beam of
NVSS). The shaded regions indicate the RLF computed from the
non-degraded NVSS (grey) and FIRST (hashed) data. The error bars
indicate the degraded versions; here the counts derived from the
FIRST catalog (blue, thin error bars) are in approximate agreement
with the NVSS counts (green, thick error bars), despite the large
difference in synthesized beam FWHM of the original data.}
\label{fig:nvss:first}
\end{figure}
Even using this simple method, the RLF from the degraded FIRST sample
is in good agreement with that derived from the NVSS counts. The RLF
constructed from the degraded FIRST data is systematically lower than
that constructed from the degraded NVSS data. \mtx{This is as
expected considering the different sensitivities to spatial
frequencies of the two surveys $-$ extended emission recovered by
NVSS is resolved out by FIRST, causing the amplitude of the RLF to
drop.} Note, however, that the difference in the luminosity function
at different redshifts caused by this effect will in fact be much
smaller than the residual difference seen in
Fig.~\ref{fig:nvss:first}, \mtx{since our relatively small
redshift range in the maxBCG sample results in a much smaller
relative difference in physical scales, as compared to the relative
difference between the FIRST and NVSS synthesized beams.}
\subsubsection{Accounting for confusion}
\label{sec:lf:confusion:accounting}
Given the above model, it is possible to adaptively introduce
confusion into our radio source sample to minimize systematic effects
in determining the redshift dependence of the RLF. In a given cluster
sample, the clusters with the greatest angular diameter distances will
be the most affected by confusion. Thus, we define a nominal
``confusion distance'', in physical units, by converting the radio
source survey resolution (45$^{\prime\prime}$ for NVSS or
5$^{\prime\prime}$ for FIRST) to a physical distance at the redshift
of the object with the greatest angular diameter distance in the
sample. Then, for every object in the sample, this physical distance
is converted back to angular units using the redshift of the object,
and the resulting angular scale is used to degrade the resolution of
this particular cluster field using the method described above.
To get an idea of the typical level of confusion, consider the maximum
redshift, $z=0.3$, of the maxBCG catalog. The FIRST resolution of
5$^{\prime\prime}$ corresponds to a physical distance of 22 kpc at
this redshift. We can translate this physical distance, the
``confusion distance'', into an angular distance at any lower
redshift; for example, at the low redshift limit of $z=0.1$ we obtain
an angular scale of 12$^{\prime\prime}$. Using this as a limit of
resolution, the FIRST source counts are reduced by approximately 25\%
in fields defined by $r_{200}$, while at $z=0.3$ the FIRST source
counts are reduced by around 10\%.\footnote{The counts are reduced
also at the resolution limit because, contrary to the FIRST catalog,
we enforce a strict limit of 5$^{\prime\prime}$ as the smallest
angular distance between two distinct sources.}
Note that we also have to adapt the background counts,
$N_{\eta}^{\mathrm{field}}$, when degrading the resolution. Because it is
time consuming to re-compute $\frac{dN}{dS \, d\Omega}$ for each angular
resolution used, we compute the field counts in several degraded versions (as
described in \sect\ref{sec:lf:confusion:degrade}) of the entire NVSS and
FIRST catalogs, using increments of 5$^{\prime\prime}$, and interpolate
between these results to derive estimates of $N_{\eta}^{\mathrm{field}}$
individually for all cluster fields.
\subsection{Modeling the luminosity function}
\label{sec:lf:model}
\subsubsection{Parametric model}
\label{sec:lf:model:param}
To allow for a quantitative estimate of the redshift evolution of the
RLF, we follow LM07 and use the parameterization
\begin{equation}
\log \phi = y - \left( b^2 + \left( \frac{\log L - x}{w} \right) ^2 \right) ^{1/2} -
1.5 \log L.
\label{eq:condonmodel}
\end{equation}
This fitting function was used by \cite{2002AJ....124..675C} to fit
the field RLF using the combined contributions from radio-loud AGN and
star-forming (SF) galaxies. LM07 used the two sets of values of
($b,x,w$) found by Condon et al. to make possible a separation of the
normalizations ($y$) of the radio-loud AGN and SF components in their
cluster RLF at low redshift.
\mtx{ It is beyond the scope of this paper to discuss the physical
interpretation of Eq.~(\ref{eq:condonmodel}) as our main concern is
the redshift evolution of the luminosity and number density. Refer
to \cite{2002AJ....124..675C} and references therein for details on
this hyperbolic fitting form of the RLF.}
\mtx{In contrast to the sample studied by LM07, our main samples do
not include systems with $z < 0.1$. This selection, combined with
our flux limit of 5 mJy, effectively ensures that we are only
sensitive to the high-luminosity radio-loud AGN population of the
RLF. Thus, a one-component model as described by
Eq.~(\ref{eq:condonmodel}) is sufficient for the present purpose.}
\subsubsection{Redshift evolution}
\label{sec:lf:model:zevol}
\cite{2000A&A...360..463M} quantified the redshift evolution of the
RLF \mtx{under the assumption that the overall shape remains constant,
as first suggested by \cite{1984ApJ...287..461C}. Under this
assumption, there can only be changes in overall luminosity and
overall number density}. The redshift dependence can be written as
\begin{equation}
\phi(L, z) \ = \ g(z) \ \phi \left[Lf(z), z\approx 0 \right],
\label{eq:lf:zevol}
\end{equation}
where $g(z)$ quantifies the number density evolution, corresponding to a
vertical shift $\Delta Y$ in $\phi(L)$, and $f(z)$ represents luminosity
evolution, corresponding to a horizontal shift $\Delta X$:
\begin{equation}
\log\left[ \phi(\log L,z)\right] \ = \ \log\left[\phi(\log L + \Delta X, z \approx 0)\right] \ + \ \Delta Y.
\label{eq:lf:zevol_shift}
\end{equation}
The vertical and horizontal shifts can be fitted to yield $f(z)$ and $g(z)$.
Shape preservation of the RLF in ($\log L, \log \phi$) space implies that
number density and luminosity scale as powers of redshift. For the luminosity,
\begin{equation}
\label{eq:lf:ppowerlaw}
L=L_0\left( \frac{1+z}{1+z_0} \right)^{\alpha_L},
\end{equation}
where $\alpha_L$ is the power law index, and correspondingly for the number
density
\begin{equation}
\label{eq:lf:phipowerlaw}
\phi = \phi_0\left( \frac{1+z}{1+z_0} \right)^{\alpha_{\phi}},
\end{equation}
with the power law index $\alpha_{\phi}$.
To constrain the redshift evolution in the RLF, we bin our samples by
redshift, and approximate the redshift in each bin by the median of
all cluster redshifts in the bin. We then apply
(\ref{eq:condonmodel}), shifted in $\phi$ and $L$ according to
(\ref{eq:lf:zevol_shift}), to constrain the power law indices
$\alpha_{\phi}$ and $\alpha_L$. Simultaneously, the four parameters of
(\ref{eq:condonmodel}) are constrained.
We let the parameter $y$ in (\ref{eq:condonmodel}) represent the
normalization of the RLF at $z=0$. We bin the data by redshift and
construct the RLF separately for each redshift bin. The thus
constructed luminosity function data are then used simultaneously to
constrain $y,b,x,w,\alpha_{\phi}$ and $\alpha_L$. \mtx{The fitting is
carried out by minimization of the $\chi^2$ statistic,
\begin{equation}
\chi^2 = \sum_{i=1}^{k}\left( \frac{X_i - \mu_i}{\sigma_i} \right) ^2,
\end{equation}
where $X_i$ is the data point with index $i$ (the number $k$ of data
points is the total number of luminosity bins in all redshift bins),
$\mu_i$ is the corresponding model value given a set of parameters
($y$, $b$, $x$, $w$, $\alpha_{\phi}$ and $\alpha_{L}$), and $\sigma_i$
is the uncertainty on data point $i$ estimated from Poisson
statistics. }
\mtx{Uncertainties on the fitted parameters are estimated
using a simple method. Given our data points, binned in redshift and
in luminosity, we vary the data according to Poisson statistics
within their estimated uncertainties, and carry out the $\chi^2$ fit
again. This process is repeated 10.000 times. As expected, the mean
of the distribution of a fitted parameter in any of the fits we
carry out (\sect\ref{sec:lf:res:zevol}) corresponds to its best-fit
value. The standard deviation of the distribution of a fitted
parameter is used as an estimate of the 1$\sigma$ uncertainty of
this parameter.}
\subsubsection{Mass dependence}
\label{sec:lf:model:mdep}
A possible mass dependence in the RLF can be investigated using the
same method as described above. By scaling the number density and
luminosity we can model the dependence to a first approximation as
power laws of the form $M_{200}^{\gamma}$. \mtx{Analogously with the
discussion in \sect\ref{sec:lf:model:zevol}, we make the assumption
of shape preservation of the RLF} and model the luminosity and
number density dependence on mass as (cf. Eqs.~(\ref{eq:lf:ppowerlaw})
and (\ref{eq:lf:phipowerlaw}))
\begin{equation}
\label{eq:mdeppower}
\begin{split}
L \sim (M_{200})^{\gamma_L}; \\
\phi \sim (M_{200})^{\gamma_{\phi}}.
\end{split}
\end{equation}
We use both the X-ray sample and the optical sample to constrain the
mass dependence. Because two new parameters are introduced into the
model, it is not possible to simultaneously constrain the mass
dependence and the redshift evolution, particularly given the limited
size of the X-ray sample. For this reason we use a sub-sample of the
X-ray selected clusters in a low redshift range of $0.05 \leq z \leq
0.12 $ (see Table \ref{tab:ccuts}). In this range, we divide the
sample into two mass bins, where care is taken that the lower mass
limit of the lower mass bin is above the completeness limit of the
X-ray sample at $z=0.12$. For the optical sample, we divide the
complete sample into three mass bins. In the next section it is shown
that in this range, our data are consistent with no mass dependence of
the RLF.
\subsection{Results}
\label{sec:lf:res}
\subsubsection{Mass dependence}
\label{sec:lf:res:mdep}
To separate a possible mass dependence in the RLF from a pure redshift
dependence, we construct RLF binned by mass as shown in
Figures~\ref{fig:mdep.xray} and \ref{fig:mdep.opt}.
\begin{figure}[Ht!]
\centering \includegraphics[width=\columnwidth]{./mdep.xray.eps}
\caption{RLF determined from the X-ray subsample in the redshift
range $0.05 \leq z \leq 0.12$, divided into two mass
bins. \mtx{Arrows indicate $1 \sigma$ upper limits
from Poisson statistics.} }
\label{fig:mdep.xray}
\end{figure}
\begin{figure}[Ht!]
\centering \includegraphics[width=\columnwidth]{./mdep.opt.eps}
\caption{RLF determined from the optical sample, divided into
three mass bins.}
\label{fig:mdep.opt}
\end{figure}
We then use a least-squares statistic to find the best-fit vertical and
horizontal shifts of the RLF, fixing the values of $b,w,x$ as described in
\sect\ref{sec:lf:model:zevol}, to interpret the shifts in terms of a
luminosity- and/or number density dependence on cluster mass. The results for
both our cluster catalogs are given in Table \ref{tab:mdepfits}, using
different priors on the power laws.
\begin{table}
\caption{Mass dependence of the RLF, parameterized in terms of power laws of
$M_{200}$ in luminosity and number density (Eq.~(\ref{eq:mdeppower})).}
\label{tab:mdepfits}
\centering
\begin{tabular}{l l r r@{}l r@{}l}
\hline\hline
Cluster & Source & Prior & $\gamma_{\phi}$& & $\gamma_{L}$& \\
sample & catalog & & & & & \\
\hline
X-ray & FIRST & & 0.25$\pm$&1.01 & -0.46$\pm$&0.55 \\
maxBCG & FIRST & & -0.36$\pm$&0.31 & 0.38$\pm$&0.65 \\
\hline
X-ray & FIRST & $\gamma_{\phi}=0$ & (0.0)~\,& & 0.13$\pm$&0.36 \\
maxBCG & FIRST & $\gamma_{\phi}=0$ & (0.0)~\,& & 0.27$\pm$&0.28 \\
\hline
X-ray & FIRST & $\gamma_{L}=0$ & 0.090$\pm$&0.16 & (0.0)~\,& \\
maxBCG & FIRST & $\gamma_{L}=0$ & 0.046$\pm$&0.15 & (0.0)~\,& \\
\hline
\end{tabular}
\end{table}
Note that all fits are consistent with $\gamma_{\phi}=0$ and $\gamma_{L}=0$ at
an approximate 1 $\sigma$ level.
Thus, contrary to LM07, who find that the amplitude of the RLF for
low-mass clusters is slightly larger at the low-luminosity end, we
find no evidence of a mass dependence in the RLF given our flux and
redshift cut-offs. As indicated in Figure \ref{fig:mdep.xray}, the
results are consistent with the most luminous radio sources only being
found in high-mass clusters, as was also found by LM07. However,
\mtx{due to the relatively small sample size, our data provides no
conclusive evidence of this.}
As there is no statistically significant indication of a mass
dependence from either cluster sample, we proceed to fit for a pure
redshift evolution in the next subsection.
\subsubsection{Redshift evolution}
\label{sec:lf:res:zevol}
Figure \ref{fig:zevol.opt}
\begin{figure}[Ht!]
\centering \includegraphics[width=\columnwidth]{./zevol.opt.eps}
\caption{Redshift evolution in the RLF determined from the optical
(maxBCG) sample, \mtx{without priors}. The data are shown as error
bars and the best simultaneous fit to luminosity and number count
evolution is represented by lines corresponding to the redshift
bins: $0.1 \leq z < 0.17$ (diamonds and solid line), $0.17 \leq z
< 0.24$ (triangles and dashed line), and $0.24 \leq z \leq 0.3$
(squares and dotted line). }
\label{fig:zevol.opt}
\end{figure}
shows the RLF computed from the maxBCG sample, using FIRST radio
sources, in three redshift bins chosen such that there are
approximately the same numbers of clusters in each bin ($0.1 \leq z <
0.17$, $0.17 \leq z < 0.24$ and $0.24 \leq z \leq 0.3$). We apply the
method outlined in \sect\ref{sec:lf:model:zevol} to constrain the
redshift evolution of the RLF by a simultaneous fit to the RLF
amplitude (parameterized by $y$ at $z=0$), the RLF shape
(parameterized by $b$, $x$ and $w$), and the power law evolution
(parameterized by $\alpha_{\phi}$ and $\alpha_{L}$ according to
Eqs.~(\ref{eq:lf:ppowerlaw}) and (\ref{eq:lf:phipowerlaw})).
We perform a number of fits to the data, with priors as listed in
Table~\ref{tab:zevol}.
\begin{table*}
\begin{center}
\caption{Redshift evolution in the RLF, quantified by simultaneous fits to the
RLF amplitude, shape and evolution. The goodness-of-fit is indicated by the
reduced chi-squared parameter, $\chi_{\text{red}}^2$. See section
\ref{sec:lf:model} for a description of the parameters.}
\label{tab:zevol}
\centering
\begin{tabular}{l l r r@{}l r@{}l r@{}l r@{}l r@{}l r@{}l r@{}l}
\hline\hline
Cluster & Source & Priors & $y$& & $b$& & $x$& & $w$& & $\alpha_{\phi}$& & $\alpha_{L}$& & $\chi_{\text{red}}^2$& \\
sample & catalog & & & & & & & & \\
\hline
maxBCG & FIRST & & 36.38$\pm$&1.02 & 1.05$\pm$&0.73 & 24.53$\pm$&0.18 & 0.66$\pm$&0.13 & $-$2.46$\pm$&1.58 & 6.20$\pm$&1.76 & 1.&07 \\
maxBCG & FIRST & $\alpha_{\phi}=0$ & 36.34$\pm$&0.92 & 0.91$\pm$&0.81 & 24.87$\pm$&0.14 & 0.72$\pm$&0.21 & (0.0)~\,& & 3.99$\pm$&1.24 & 1.&19 \\
maxBCG & FIRST & $\alpha_{L}=0$ & 36.74$\pm$&0.89 & 1.01$\pm$&0.55 & 25.11$\pm$&0.11 & 0.71$\pm$&0.19 & 1.03$\pm$&1.14 & (0.0)~\,& & 2.&25 \\
\hline
X-ray & FIRST & (a) & 36.19$\pm$&0.19 & (1.05)~\,& & (24.53)~\,& & (0.66)~\,& & 0.76$\pm$&1.86 & 8.12$\pm$&2.67 & 0.&94 \\
X-ray & FIRST & (a);~$\alpha_{\phi}=0$ & 36.26$\pm$&0.10 & (1.05)~\,& & (24.53)~\,& & (0.66)~\,& & (0.0)~\,& & 8.19$\pm$&2.66 & 0.&89 \\
X-ray & FIRST & (a);~$\alpha_{L}=0$ & 35.89$\pm$&0.18 & (1.05)~\,& & (24.53)~\,& & (0.66)~\,& & 9.40$\pm$&1.85 & (0.0)~\,& & 10.&48 \\
\hline
\end{tabular}
\end{center}
(a) The shape parameters $b$, $x$ and $w$ were fixed to the results of the maxBCG/FIRST
analysis with no priors. \\
\end{table*}
Although a simultaneous fit to both $\alpha_{\phi}$ and $\alpha_{L}$
yields the best fit (in the sense of the reduced $\chi^2$ parameter,
$\chi_{\text{red}}^2$, being the closest to unity), the data are
consistent with no number density evolution. Note that the best-fit
value $\alpha_{\phi}=-1.38$ implies \textit{fewer} radio-loud AGN at
higher redshift within the range $0.1 \leq z \leq 0.3$. \mtx{Though
not unphysical, such a negative evolution is unlikely, considering
the evidence of increased AGN activity in the field population
\citep{1990MNRAS.247...19D}, and the enhanced AGN fraction within
clusters as seen from X-ray observations
\citep{2007ApJ...664L...9E,2009ApJ...694.1309G}} \mtx{Therefore, we
carry out an additional fit in which we fix the number density power
law to zero} and fit for a pure luminosity evolution. \mtx{The
result is a positive luminosity evolution with $\alpha_L=3.99 \pm
1.24$}. While consistent with a pure luminosity evolution, the data
are inconsistent with a pure number density evolution, as indicated by
the $\chi_{\text{red}}^2$ of the third column in
Table~\ref{tab:zevol}.
The last three rows of Table~\ref{tab:zevol} indicate the results of
fitting the evolution parameters to the X-ray derived RLF, also binned
by redshift (redshift bins: $0.1 \leq z < 0.2$, $0.2 \leq z < 0.45$
and $z \geq 0.45$). Although the X-ray sample extends to greater
redshifts than the maxBCG sample, it has less leverage on the redshift
evolution because of its limited size. For this reason,
\mtx{we fit only for $y$, $\alpha_{\phi}$ and
$\alpha_{L}$. We keep the shape parameters fixed to the values
found in the maxBCG/FIRST fit, as this is the fit with the smallest
uncertainties. Because the smallest $\chi^2_{\text{red}}$ is
measured for the case of no priors (as expected) in the maxBCG/FIRST
case, we use the results of this fit for fixing the shape
parameters.} Again, we also carry out fits to pure luminosity
evolution and pure number density evolution. At one standard
deviation, the results from the X-ray sample are consistent with those
of the optical maxBCG sample. Again, a pure number density evolution
provides a poor fit to the data.
We note that in all cases the fit is especially sensitive to the
parameter $b$, which is explained by its quadratic dependence in
Eq.~(\ref{eq:condonmodel}).
\subsection{Systematic uncertainties}
\label{sec:lf:sys}
\mtx{ To estimate systematic uncertainties in the luminosity
and number density evolution of the RLF, we discuss below four
separate sources of error.
\mtx{The estimated systematic uncertainties for the maxBCG and X-ray
samples are summarized in Table \ref{tab:sys}. For each
uncertainty estimate, we keep all parameters but the RLF
normalization and the evolution parameter in question
($\alpha_{\phi}$ or $\alpha_{L}$) fixed, while carrying out the
complete RLF analysis with posterior assumptions altered as
described above. To estimate total systematic effects on
$\alpha_{\phi}$ and $\alpha_{L}$, we add all relevant components
in quadrature. } \mtx{We note that for this study, the statistical
errors on the fitted parameters (see Table~\ref{tab:zevol}) still
dominate over the systematics.}
\begin{table}[Ht]
\begin{minipage}[Ht]{\columnwidth}
\caption{Summary of systematic effects on the number density and
luminosity evolution of the RLF.}
\label{tab:sys}
\centering
\renewcommand{\footnoterule}{}
\begin{tabular}{l l l}
\hline\hline
Systematic & effect on $\alpha_{\phi}$ & effect on $\alpha_{L}$ \\
\hline
$L_{\text{bol}}-T$ scaling & $\pm0.30$ & \\
$L_{X}-n_{\text{gal}}^{R200}$ scaling\footnote{only for maxBCG clusters} & $\pm$0.80 & \\
$k$-correction & $^{+0.15}_{-0.51}$ & $^{+0.17}_{-0.16}$ \\
\hline
\hline
Totals\footnote{added in quadrature} & & \\
\hline
maxBCG & $^{+0.87}_{-1.00}$ & $^{+0.17}_{-0.16}$ \\
X-ray & $^{+0.33}_{-0.59}$ & $^{+0.17}_{-0.16}$ \\
\hline
\end{tabular}
\end{minipage}
\end{table}
\subsubsection{Scaling relations}
Uncertainties in the $z$-dependence of the scaling relations used to
derive cluster masses affects the derived number density evolution
through the indirect determination of cluster volumes. For the X-ray
sample, we estimate this systematic by replacing the self-similar
evolution of the $L_{\text{bol}}-T$ relation by (i) the strong
evolution measured by \cite{2005ApJ...633..781K} and (ii) by assuming
no evolution of the $L_{\text{bol}}-T$ relation. Re-computing the
$M_{200}$ for both cases results in a change in the number density
evolution parameter $\alpha_{\phi}$ of $\pm0.30$ when re-computing the
RLF for the X-ray sample. For the optical sample, we additionally need
to take the uncertainty in the $L_{X}-n_{\text{gal}}^{R200}$ relation
(Eq.~(\ref{eq:rykoff})) into account. We estimate the systematic
effects by using the extreme values (at 1$\sigma$ confidence) of the
redshift dependence from \cite{2008ApJ...675.1106R} to re-compute
$M_{200}$, resulting in an effect on $\alpha_{\phi}$ of $\pm 0.80$.
\subsubsection{$k$-correction}
Uncertainties in the $k$-correction used in converting from radio
source flux to luminosity will mainly affect the luminosity
evolution. It can also shift the number density evolution by assigning
the wrong luminosity bins to sources.
\mtx{As extreme values of the ensemble average spectral
slope, we use the 25th and 75th percentiles ($\alpha = 0.51$ and
$\alpha=0.92$, respectively) of the slope distribution of
\cite{2007AJ....134..897C}.} We re-compute the maxBCG RLF using
these extreme spectral slopes for all sources.
\subsubsection{Source counts}
Incorrect radio source counts can be caused either by counting
extended or complex sources (e.g. AGN with radio lobes) as several
separate sources, or by not resolving unrelated sources. This affects
the number density and luminosity evolution only if the effect is
varying with redshift; otherwise a mere offset in the normalization of
the RLF is introduced.
A complete treatment of these effects would require the careful visual
inspection of the radio maps in each cluster field, which is not
feasible for a large sample such as maxBCG. Instead we carried out a
visual inspection on parts of the sample to estimate the possible
effects. We visually inspected FIRST maps in the 500 fields
corresponding to the clusters with the lowest redshifts, and in the
500 fields corresponding to the clusters with the highest redshifts of
the maxBCG sample, taking weighted sums of the flux densities of FIRST
catalog entries deemed to be different components or lobes of the same
sources. Ambiguous cases (less than 2\% of the fields studied) were
removed. Comparing the RLFs constructed from the modified FIRST data
in the two sub-samples, we noted a drop in normalization of the RLF by
about 5\% in both the high-$z$ and the low-$z$ subsamples. There was
no indication of a significant change to the evolution of the RLF
within our statistical errors. Thus, we conclude that our adaptive
accounting of confusion with redshift (section \ref{sec:lf:confusion})
is sufficient to deal with this systematic, at least for the purpose
of constraining the redshift evolution of the RLF.
\subsubsection{Radio source flux}
Systematic offsets in radio source flux densities are expected to be a
problem mainly for the FIRST survey, where significant amounts of flux
of extended sources can be resolved out. Again, this problem only
affects the normalization of the RLF, unless there is a redshift
dependence in the amount of resolved-out flux. Since the results from
the NVSS and FIRST radio data for the redshift evolution of the RLF
from the maxBCG sample are consistent within statistical errors, we do
not pursue this point further.
\subsection{Comparison of the RLF with previous findings}
\mtx{To compare our findings with previously published results, we
consider the maxBCG/FIRST data in our lowest redshift bin, which
compares well to the low-redshift samples of clusters used in
previous works. Figure \ref{fig:lf.comparison}
\begin{figure}[Ht!]
\centering \includegraphics[width=\columnwidth]{./lf.comparison.eps}
\caption{\mtx{Comparison of our RLF derived from the
maxBCG/FIRST data in the lowest redshift bin ($0.1 < z < 0.17$,
$z_{\text{median}}=0.13$; black diamonds) with other relevant
volume-averaged radio luminosity functions. The RLF of
\cite{2007ApJS..170...71L} is indicated by shaded regions, while
the corresponding luminosity functions of
\cite{2004ApJ...600..695R} and \cite{2004A&A...424..409M} are
shown as green triangles and red squares, respectively.}}
\label{fig:lf.comparison}
\end{figure}
shows a comparison of our low-redshift RLF to the results of
\cite{2007ApJS..170...71L}, \cite{2004ApJ...600..695R} and
\cite{2004A&A...424..409M}.}
We note that the normalization of the RLF is consistent with the
results of LM07, \mtx{although the RLF from our optically selected
sample does not extend as high in luminosity ($\textrm{Log}(L)
\gtrsim 26 \, \textrm{W} \, \textrm{Hz}^{-1}$). The latter} is
likely a result of the fact that the optically selected sample
contains many more low-mass clusters, and that the more luminous radio
sources tend to reside in high-mass clusters (cf.
\sect\ref{sec:lf:res:mdep}).
\mtx{The RLF of \cite{2004A&A...424..409M} has a
significantly lower normalization than what we find,
\mtx{by approximately a factor of two}. This can be
expected due to the different definition of cluster volumes (using a
constant cluster radius across the complete sample) used in the
former work.}
\mtx{\cite{2006A&A...446...97B} used a definition of the
RLF as the number of radio sources per cluster rather than averaged
over a volume. A direct comparison with that work would require a
re-analysis of the \cite{2006A&A...446...97B} sample, which would be
only of moderate interest considering that the normalization would
rely on different assumptions on cluster volume and the sample is
selected differently from our maxBCG sample. Our main objective here
is to show the approximate agreement between different local
luminosity functions and to stress the difficulty in obtaining a
robust normalization of the RLF; although the normalization depends
strongly on the choice of cluster volume and is prone to systematic
effects due to a poor understanding of this volume, this does not
strongly affect our main conclusion, which is concerned with the
redshift evolution of the RLF independently of its overall
normalization.}
Early studies of the luminosity function of optically selected QSOs at $z \leq
2.2$ suggested a pure luminosity evolution with $L \sim (1+z)^{(3.5 \pm 0.3)}$
in the field population \citep{1987MNRAS.227..717B,1988MNRAS.235..935B}. A
later study of extragalactic radio sources by \cite{1990MNRAS.247...19D} at
2.7 GHz came to similar conclusions.
The pure luminosity evolution of the volume averaged RLF found in this study
is approximately consistent with more recent findings in the field population
of radio galaxies at low and intermediate redshift \citep{2000A&A...360..463M,
2001AJ....121.2381B}, although our best-fit evolutionary model with $L \sim
(1+z)^{(6.20 \pm 1.76)}$ is steeper than both the $L \sim (1+z)^{(4 \pm 1)}$
found by \cite{2001AJ....121.2381B} and the $L \sim (1+z)^{(3 \pm 1)}$ which
\cite{2000A&A...360..463M} found to be consistent with their data.
\section{Summary and conclusions}
\label{sec:concl}
We used the maxBCG optical sample of galaxy clusters and a composite
X-ray sample to construct the volume averaged radio luminosity
function (RLF) in galaxy clusters by cross-correlating cluster
positions with radio point source positions from the FIRST and NVSS
survey catalogs. Background and foreground counts were corrected for,
and variable confusion with redshift was accounted for. We
investigated the radial source density distribution of radio sources
associated with clusters, \mtx{and correlated the luminosity of the
brightest radio source with cluster mass}.
To fit the radial number density distributions, we used the
\mtx{functional form of the} isothermal $\beta$-model. \mtx{Combining
the maxBCG cluster sample with the FIRST catalog, we found an
additional narrow component which we identify with radio detections
of the brightest cluster galaxies (BCGs)}. All combinations of
cluster sample (maxBCG or X-ray) and radio source sample (NVSS or
FIRST) yield similar values of the power law index $\beta$. The core
radius in the maxBCG/FIRST radial distribution of sources compares
well to the distribution of sources in the X-ray sample. As it was
not possible to identify the narrow BCG component from the NVSS
survey, we found a much smaller core radius in the maxBCG/NVSS radial
distribution.
Our derived radial distributions are narrower than what has been found
in earlier studies \citep[e.g][]{2004ApJ...600..695R,
2004A&A...424..409M}, \mtx{even when disregarding the
central component associated with brightest cluster galaxies
(BCGs)}. A plausible explanation is that we are constructing the RLF
at higher redshifts than previous studies and thus, given our flux
limit of 5 mJy, we are sensitive only to radio-loud AGN. At lower
redshift there is a mixing with star forming galaxies, which have a
less \mtx{centralized} distribution.
\mtx{Unlike LM07, we do not find any evidence that bright radio sources
have a radial source density distribution different from that of faint sources.
Again, a likely explanation is that we are studying different populations of
radio sources through the redshift selection.}
We found that the luminosity of the most radio-luminous source within 50 kpc
from the cluster center scales with cluster mass following a power law with
slope 0.31$\pm$0.12 in the maxBCG sample. This is consistent with the results
of \cite{2004ApJ...617..879L} as well as with the results from our X-ray sample,
although the latter is also consistent with no correlation.
\mtx{We find the RLFs constructed from the optical and X-ray samples of
galaxy clusters to be in approximate agreement. The RLF from the optical
maxBCG sample is systematically lower at luminosities $L \gtrsim 3 \times
10^{25}$ W Hz$^{-1}$. This is likely a result of many more low-mass systems
being present in the optical sample. }
We provide the first evidence for a luminosity evolution of the
volume-averaged RLF in clusters of galaxies. The maxBCG/FIRST data
are consistent with a pure luminosity evolution, with power scaling
with redshift as $L \sim (1+z)^{\alpha_L}$, where $\alpha_L=6.20
^{+1.76 +0.19}_{-1.76 - 0.17}$ (statistical followed by systematic
uncertainties). There is no indication of a mass dependence in the
RLF from the present data. However, the results from the X-ray sample
are consistent with the findings of LM07, that the most luminous radio
sources reside in massive clusters. This is further corroborated by
the fact that the RLF constructed from the maxBCG sample (which
contains a smaller fraction of high-mass systems than both our X-ray
sample and the sample of LM07) is steeper at higher luminosities.
\mtx{Below $P \sim 10^{25} \, \text{W} \, \text{Hz}^{-1}$, the data are consistent with no mass
dependence in the RLF, as shown by constructing the RLF in the
low-redshift X-ray sample and binning by mass. Although we have found
that massive clusters have more luminous BCGs, this effect is
counteracted in the RLF by the fact that these massive systems also
have more volume.}
Both the X-ray sample and the maxBCG sample yield results
consistent with a pure luminosity evolution of the RLF. In addition,
the derived power laws are comparable, although the samples cover very
different redshift ranges.
\begin{acknowledgements}
We acknowledge partial support for this work from Priority
Programme 1177 and Transregio Programme TR33 of the German Research
Foundation (Deutsche Forschungsgemeinschaft). For the early stages of this
work, MWS acknowledges support through a stipend from the International Max
Planck Research School (IMPRS) for Radio and Infrared Astronomy at the
Universities of Bonn and Cologne. FP acknowledges support from grant
50\,OR\,1003 of the Deutsches Zentrum f\"ur Luft- und Raumfahrt. HA is
grateful to Mexican CONACyT for research grants 50921-F and 118295.
\end{acknowledgements}
\bibliographystyle{aa}
|
\section{Introduction}
\label{Intro}
The application of a resonant driving force is an efficient way of compensating dissipative losses in a soliton-bearing system.
If the dissipation coefficient and driving strength are weak,
and the driving frequency is just below the phonon band,
the amplitude of the arising oscillating soliton is
governed by the nonlinear
Schr\"odinger equation with damping and driving terms.
The damped-driven nonlinear
Schr\"odinger equations exhibit localised solutions with a variety
of temporal behaviours, from stationary to periodic and chaotic.
There is a whole range of analytical and numerical approaches to the study of
stationary and steadily travelling solitary waves.
As for the solitons with nontrivial time dependence, such as periodic,
the direct numerical simulation has remained an
exclusive means of obtaining these solutions and classifying their
stability.
The shortcoming of this method is that
simulations capture only {\it stable\/}
regimes.
This means that the actual mechanisms
and details of the
transformations of solitons (which are bifurcations involving both stable and
unstable solutions) remain inaccessible. Neither can simulations
be used to identify coexisting attractors in cases of bi-
or multistability.
In this paper we pursue a different approach to the analysis of these
hidden mechanisms. Instead of direct numerical simulations, the
time-periodic solitons are determined as solutions of a
boundary-value problem formulated on a two-dimensional
spatio-temporal domain.
The advantage of this approach is that it is potentially capable of
furnishing {\it all\/} solutions ---
all stable and all unstable.
The particular equation that we are concerned with here, is the
parametrically driven damped nonlinear Schr\"odinger,
\begin{equation}
i \psi_t + \psi_{xx} + 2 |\psi|^2 \psi - \psi = h \psi^* - i
\gamma \psi.
\label{NLS}
\end{equation}
In Eq.\eqref{NLS}, $\gamma>0$ is the damping
coefficient and $h$ the amplitude of the parametric driver, which
can also be assumed positive. Equation \eqref{NLS} was used to model a
variety of resonant phenomena in nonlinear dispersive media, including the
nonlinear Faraday resonance in a vertically oscillating water
trough \cite{Miles,LS,Faraday}, formation of oscillons in granular materials
and suspensions \cite{oscillons},
synchronisation in parametrically excited pendula arrays
\cite{pendula,Clerc1},
phase-sensitive amplification of light pulses in optical fibers
\cite{optics} and propagation of magnetization waves
in an easy-plane ferromagnet placed in a microwave field \cite{BBK,Clerc1,Clerc2}.
The same equation governs the
amplitude of breathers in
a variety of systems reducible to the
parametrically driven
damped sine-Gordon \cite{BBK}
and the $\phi^4$ \cite{JA} equation.
(For more contexts, see \cite{BZ3}.)
The studies of localised states in these mechanical,
hydrodynamical, magnetic and optical systems have been
focussing on structures oscillating {\it periodically}, i.e., without incommensurable frequencies
in their spectrum.
The {\it quasiperiodically\/} oscillating states
(e.g. states whose fundamental oscillation is modulated by
smaller frequencies) have been set aside as complex and atypical.
The periodically oscillating states in the physical systems reducible
to equation \eqref{NLS} are described by the stationary solutions of this equation.
Accordingly, all previous analyses of localised solutions of Eq.\eqref{NLS} have been
confined to its stationary solitons.
One of the conclusions of the present project, however, will be
that the time-periodic solitons of Eq.\eqref{NLS}
populate a significant part of its attractor chart. Therefore, the corresponding
{\it quasiperiodic\/} localised states should play
a much bigger role in all physical applications of Eq.\eqref{NLS} than it was assumed so far.
The aim of the present work is to follow the transformations
of temporally periodic solitons of equation \eqref{NLS} as its
parameters are varied, identify the arising bifurcations and
eventually verify and explain the attractor chart for this equation which was
compiled using direct numerical simulations in Ref.\cite{Bondila}.
In the second part of this project \cite{BZ3}
we will complement this one-soliton chart with a
chart of two-soliton attractors.
The paper is organised as follows. Section \ref{BG} (which continues this introduction)
contains background information on
stationary solitons and their transformations as the parameters of the equation are varied.
It is these transformations that we will be verifying and studying further in the subsequent
sections.
Next, sections \ref{P} and \ref{S} present mathematical techniques we employ in this project:
in section \ref{P} we outline our method of obtaining periodic solitons whereas
section \ref{S} describes our approach to the analysis of their stability.
Section \ref{Rad} introduces a theoretical framework for the treatment of radiation from
the oscillating soliton.
The central results of this study are obtained using numerical methods; these are
reported in section \ref{Num1}. Here we
present bifurcation diagrams for the time-periodic free-standing soliton
in various damping regimes.
The paper is concluded by section \ref{DC} where the results of the
numerical study are discussed and interpreted.
\section{Stationary and oscillatory solitons: the background}
\label{BG}
Localised stationary or periodic solutions of Eq.\eqref{NLS}
exist only if $h> \gamma$.
When $h> h_{\rm cont}(\gamma)$,
where
\begin{equation*}
h_{\rm cont} = \sqrt{1+ \gamma^2},
\end{equation*}
any localised solution is unstable to continuous-spectrum
perturbations. The evolution of this
instability leads to spatiotemporal chaos.
Two stationary soliton solutions of Eq.\eqref{NLS} are well known.
One soliton (denoted $\psi_-$ in what follows) exists in the parameter range
$\gamma \leq h \leq h_{\rm cont}(\gamma)$ and has the form
\begin{subequations}
\label{solmin}
\begin{equation}
\psi_-(x)= A_- \exp (- i \theta_-) \, {\rm sech\/} (A_- x), \label{solmin_a}
\end{equation}
where
\begin{equation}
A_- = \sqrt{1- \sqrt{h^2-\gamma^2}}, \quad
\theta_- = \frac{\pi}{2} - \frac12 \arcsin \frac{\gamma}{h}.
\label{solmin_b}
\end{equation}
\end{subequations}
This solution is unstable for all $h$ and $\gamma$ \cite{LS,BBK}.
We are mentioning this unstable object here because
it will reappear below as a constituent in
stationary multisoliton bound states. We will also be recalling
this soliton when interpreting complex temporal behaviour of
time-periodic solutions.
The other stationary soliton exists for all
$h \geq \gamma$; we denote it $\psi_+$:
\begin{subequations}
\label{solplus}
\begin{equation}
\psi_+(x)= A_+ \exp(- i \theta_+) \, {\rm sech\/} (A_+ x),
\label{solplus_a}
\end{equation}
where
\begin{equation}
A_+= \sqrt{1+\sqrt{h^2-\gamma^2}}, \quad
\theta_+ = \frac12 \arcsin \frac{\gamma}{h}.
\label{solplus_b}
\end{equation}
\end{subequations}
The stability properties of this soliton depend on $\gamma$ and $h$ \cite{BBK}.
When $\gamma>0.356$, the $\psi_+$ soliton is stable for
all $h$ in the range $\gamma < h< h_{\rm cont}(\gamma)$.
When $\gamma< 0.356$, on the other hand, the
soliton \eqref{solplus} is only stable for $\gamma< h < h_{\rm
Hopf}(\gamma)$, where the value $h_{\rm Hopf}(\gamma)$ lies
between $\gamma$ and $h_{\rm cont}(\gamma)$
(see the curve labelled $1$ in Fig.\ref{chart_from_PhysicaD}).
As we increase $h$ past $h_{\rm Hopf}(\gamma)$
keeping
$\gamma$ fixed, the stationary soliton loses its stability to a
time-periodic soliton \cite{BBK,ABP}. The transformation scenario
arising as $h$ is increased further depends on the choice of the (fixed) value of
$\gamma$.
The numerical simulations
\cite{Bondila} (also \cite{FLS}) indicate that
for $\gamma$ smaller than approximately $0.25$, the periodic
soliton follows a period-doubling route to temporal chaos. In
a wide region of
$h$ values
above the chaotic domain, the equation does not support any stable
spatially-localised solutions. In this ``desert" region,
the only attractor found in direct numerical
simulations was the trivial one, $\psi=0$. Finally, for even
larger values of $h$, the unstable soliton seeds the spatio-temporal
chaos \cite{Bondila} (see also \cite{FLS}).
As $h$ is increased for the fixed $\gamma$ greater than
approximately $0.275$,
the soliton follows a different
transformation scenario. Here, the period-doubling cascade does
not arise and the soliton death does not occur. The periodic
soliton remains stable until it yields directly to a
spatio-temporal chaotic state \cite{Bondila}.
\begin{figure}
\includegraphics[width =\linewidth]{fig1.pdf}
\caption{\label{chart_from_PhysicaD}The single-soliton
attractor chart for equation \eqref{NLS} compiled by direct numerical simulations \cite{Bondila}.
Stationary solitons serve as attractors in the blank area above the $h=\gamma$ line.
The area where only the trivial attractor is available is marked by the
empty triangles. Empty circles indicate stable periodic solitons;
the black triangles label spatio-temporal chaos.
Other symbols mark stable higher-periodic and
temporally chaotic solitons.
For details see \cite{Bondila}.
}
\end{figure}
In a short intermediate range of $\gamma$-values,
$0.25 < \gamma <0.275$, we have a combination of the
above two scenarios. The increase of $h$ for the fixed
$\gamma$ results in
the period-doubling of the soliton, culminating
in the
temporal-chaotic regime,
which is followed
by the soliton death. As we continue
to raise $h$, an inverse sequence of bifurcations is observed which
brings the stable single-periodic soliton back. On
further increase of $h$, it loses its stability to a
spatio-temporal chaotic state \cite{Bondila}.
The attractors arising for various
$h$ and $\gamma$ values are illustrated by Fig.\ref{chart_from_PhysicaD}
which we reproduce from Ref.\cite{Bondila}. We should emphasise that
this attractor chart has been compiled using direct numerical
simulations of equation \eqref{NLS}, with a particular choice of
initial conditions. (The initial condition was chosen
in the form of the unstable soliton $\psi_+$ --- in the parameter range where it
is unstable.) It is an open question, therefore,
how robust this chart is. Would it change if simulations started with
a different initial condition, or if one used a numerical scheme with different
parameters (such as the spatial interval length, the number of the Fourier modes,
the full time of simulation, temporal stepsize,
etc)? In particular, would qualitative
features of the chart survive, e.g. the coexistence of two
transformation scenarios, appearance of the desert region, and
the peculiar shape of the periodic-attractor domain?
One of the aims of the present work is to answer these questions.
\section{Periodic solitons as solutions of a boundary-value problem in 2D}
\label{P}
Instead of solving equation \eqref{NLS} with some initial
condition and determining the resulting attractor by running the
computation for a sufficiently long time, we were searching
for periodic solutions by solving Eq.\eqref{NLS}
as a boundary-value problem on a two-dimensional
domain $(-\infty, \infty) \times (0,T)$. The boundary
conditions were set as
%
\begin{equation}
\psi(x,t) \to 0 \quad \text{as} \ x \to \pm \infty, \label{bc}
\end{equation}
and
\begin{equation} \psi(x,t+T)= \psi(x,t). \label{per}
\end{equation}
The value of $T$ was not available beforehand;
the period was regarded as an
unknown, together with the solution $\psi(x,t)$.
The periodic solutions were continued (path-followed) in $h$
for the fixed $\gamma$. We employed a predictor-corrector continuation algorithm
\cite{continuation} with a fourth-order Newtonian iteration at
each $h$. A finite-difference discretization with the stepsize
$\Delta x=0.05$ was used on the interval $(-L,L)=(-50,50)$.
\section{Stability of periodic solutions}
\label{S}
\subsection{Floquet multipliers}
Let $\psi_0(x,t)= {\cal
R}(x,t)+ i {\cal I}(x,t)$ be a spatially localised, time-periodic solution.
Letting
$\psi(x,t)=\psi_0(x,t) + u(x,t)+i v(x,t)$ and linearising
\eqref{NLS} in the small perturbation $u+iv$, we obtain
\begin{equation}
J {\bf w}_t=({\cal H}- \gamma J){\bf w},
\label{linearised}
\end{equation}
where ${\bf w}={\bf w}(x,t)$ is a two-component column-vector
\[
{\bf w}= \left( \begin{array}{c} u \\ v \end{array} \right),
\]
$J$ is a skew-symmetric matrix
\[
J= \left(
\begin{array}{cc}
0 & -1 \\ 1 & 0
\end{array}
\right),
\]
and
$\cal H$ is a hermitian matrix-differential operator
\begin{widetext}
\begin{equation} {\cal H}= \left(
\begin{array}{cc}
-\partial_x^2 + 1+h - 6{\cal R}^2 - 2 {\cal I}^2 &
- 4 {\cal R}{\cal I} \\
- 4 {\cal R}{\cal I} &
-\partial_x^2 + 1-h - 2{\cal R}^2 - 6{\cal I}^2
\end{array} \right).
\label{Hcal}
\end{equation}
\end{widetext}
The solution to Eq.\eqref{linearised} with an initial condition
${\bf w}(x,0)$ can be written, formally, as
$
{\bf w}(x,t)= {\cal M}_t {\bf w}(x,0),
$
where the evolution operator ${\cal M}_t$ acts on (vector-)
functions of $x$ but depends, parametrically, on $t$. A
fundamental role is played by eigenvalues of the monodromy
operator ${\cal M}_T$:
\begin{equation} {\cal M}_T {\bf y}(x)= \mu {\bf y}(x). \label{EV} \end{equation}
Here ${\cal M}_T$ is the evolution operator ${\cal M}_t $
evaluated at $t=T$. The eigenvalues $\mu$ are usually referred
to as the Floquet multipliers and the exponents $\lambda$, where
$
\mu= e^{\lambda T},
$
as the Floquet exponents. According to the Floquet theory,
for each $\lambda$ there is a solution ${\bf w}(x,t)$ such that
\begin{equation} {\bf w}(x,t) = e^{\lambda t} {\bf p}(x,t),
\label{lp}
\end{equation}
where ${\bf p}(x,t)$ is periodic with the period
of the solution $\psi_0$: ${\bf p}(x,t+T)={\bf
p}(x,t)$ for all $t$.
It is useful to establish the relation between the
evolution operator ${\cal M}_t$ and symplectic maps.
Letting ${\bf w}(x,t)= e^{-\gamma t} {\tilde {\bf w}} (x,t)$,
Eq.\eqref{linearised} is cast in the form
\begin{equation}
J {\tilde {\bf w}}_t={\cal H}{\tilde {\bf w}}.
\label{hamsys}
\end{equation}
Since ${\cal H}$ is hermitian, equation \eqref{hamsys}
is a hamiltonian system (with a quadratic Hamilton
functional). The solution to \eqref{hamsys} with
an initial condition ${\bf w}(x,0)$ is, therefore,
\[
{\tilde {\bf w}}(x,t) = {\cal S}_t {\bf w}(x,0),
\]
where ${\cal S}_t$ is a symplectic map. Thus
the evolution operator ${\cal M}_t$ and, in particular,
the monodromy operator ${\cal M}_T$, are related to
symplectic maps: ${\cal M}_t = e^{-\gamma t} {\cal S}_t$
and
\begin{equation}
{\cal M}_T = e^{-\gamma T} {\cal S}_T.
\label{relation}
\end{equation}
We will use this relation below, in order to explain symmetries
of the set of eigenvalues of the operator ${\cal M}_T$.
\subsection{Spectrum structure}
Stability properties of {\it stationary\/} solutions are determined
by eigenvalues $\lambda$ of the operator $(J^{-1} {\cal H}-\gamma)$.
The corresponding Floquet multipliers are given by $e^{\lambda T}$
(where $T$ can be chosen arbitrarily for a stationary solution).
Therefore the stability eigenvalues $\lambda$ are nothing but the Floquet
exponents.
Each of the two stationary solitons $\psi_+$ and $\psi_-$ has just
one zero stability eigenvalue
(i.e. just one exponent $\lambda=0$). This zero mode originates
from the translation invariance of equation \eqref{NLS} ---
this is the only continuous symmetry the damped-driven
equation has.
At the point of the Hopf bifurcation of a stationary solution
(a free-standing soliton $\psi_+$ or a stationary bound state of solitons),
two complex-conjugate eigenvalues $\lambda, \lambda^*$
cross
into the ${\rm Re} \lambda >0$ half-plane. Irrespectively of
what one takes for the value of $T$
in this case, the corresponding Floquet multipliers
$\mu=e^{\lambda T}$ and $\mu^*= e^{\lambda^* T}$ cross through the
unit circle on the complex $\mu$-plane. If we, however, take $T$
to be equal to the period of the periodic solution
bifurcating off at this point, i.e., let $T= 2 \pi/ {\rm Im} \, \lambda$, the
Floquet multipliers cross through the unit circle exactly at $\mu=1$.
Adding the unit multiplier associated with the translation
invariance, we conclude that
any stationary solution has 3 unit Floquet multipliers
at the point of its Hopf bifurcation and, accordingly, the detaching
periodic solutions should also have 3 unit multipliers at that point.
As we continue the periodic solution
$\psi_0(x,t)$ away from the point where
it was born, two of the unit Floquet multipliers persist
(while the third one moves away along the real axis). These two
unit multipliers are associated with the translation invariance
in $x$ and periodicity in time, respectively.
Indeed, if we substitute $\psi_0(x,t)$ back in Eq.\eqref{NLS} and
differentiate the resulting identity with respect to $x$ and $t$,
we will obtain Eq.\eqref{linearised} with ${\bf w}=\partial_x
\boldsymbol{\psi}_0$ and ${\bf w}=\partial_t \boldsymbol{\psi}_0$,
respectively, where the two-component vector
\[
\boldsymbol{\psi}_0= \left( \begin{array}{c} {\cal R}(x,t)\\ {\cal
I}(x,t) \end{array} \right).
\]
This means that equation \eqref{linearised} has two {\it periodic\/}
solutions and the monodromy operator \eqref{EV} has two
eigenvalues $\mu=1$ with eigenfunctions
${\bf w}=\partial_x \boldsymbol{\psi}_0$ and ${\bf w}=\partial_t
\boldsymbol{\psi}_0$,
respectively. These two unit eigenvalues
will routinely arise in our stability analysis of periodic
solutions of Eq.\eqref{NLS}.
It is not difficult to show that real eigenvalues of ${\cal M}_T$ will
always arise in pairs whereas the complex eigenvalues will
always come in quadruplets. Indeed, the monodromy operator
is related to a symplectic map by the relation
\eqref{relation}. Real eigenvalues of symplectic maps are
known to come in pairs; if $\nu$ is a real eigenvalue
of ${\cal S}_T$, then so is $\nu^{-1}$. Complex eigenvalues
of ${\cal S}_T$ come in quadruplets; if $\nu$ is such an eigenvalue,
then so are $\nu^{-1}, \nu^*$, and $(\nu^*)^{-1}$ \cite{Arnold}.
The relation \eqref{relation} implies then that if
$\mu$ is a real eigenvalue of the evolution operator
${\cal M}_T$, then so is ${\hat \mu}$,
its inverse with respect to the circle of radius $e^{-\gamma T}$:
\[
\mu {\hat \mu} = e^{-2 \gamma T}.
\]
In a similar way, if $\mu$ is a complex eigenvalue of ${\cal M}_T$,
then $\mu^*$, ${\hat \mu}$, and ${\hat \mu}^*$ are eigenvalues
as well.
We will be referring to ${\hat \mu}$ simply
as the mirror image of $\mu$.
Out of the two Floquet multipliers $\mu$ and ${\hat \mu}$ only one
can cross the unit circle whereas its mirror image will always be confined
inside a smaller circle of the radius $e^{-\gamma T}$.
Consequently, when analysing the motion of multipliers on the
complex plane and resulting stability changes, we will be
focussing on $\mu$ with moduli greater than $e^{-\gamma T}$ and
ignoring those with $|\mu|<e^{-\gamma T}$.
Finally, we discuss the location of the continuous spectrum.
As $|x| \to \infty$, the operator $(J^{-1}{\cal H}-\gamma)$ becomes a matrix differential
operator with constant coefficients whose spectrum is easily
determined. Namely, when $h^2<1$, the spectrum consists of
all $\lambda$ of the form $\lambda=-\gamma \pm ip$, with
$\sqrt{1-h^2} \leq p \leq \infty$. The corresponding Floquet
multipliers fill in a circle of the radius $e^{-\gamma T}$ on the complex
$\mu$-plane. When $h^2>1$, the continuous spectrum of $(J^{-1}{\cal H}-\gamma)$
fills the vertical line ${\rm Re} \, \lambda=-\gamma$
and an interval on the real axis, $-\gamma-\sqrt{h^2-1}< \lambda
< -\gamma + \sqrt{h^2-1}$. The corresponding Floquet multipliers
fill the circle of the radius $e^{-\gamma T}$ and in addition,
an interval on the real axis: $e^{-(\gamma+\sqrt{h^2-1})T}< \mu
< e^{-(\gamma -\sqrt{h^2-1})T}$.
\subsection{Numerical stability analysis: the method}
To find out whether equation \eqref{linearised} admits solutions
of the form \eqref{lp} with $\mbox{Re} \lambda>0$, we expand
$u(x,t)$ and $v(x,t)$ in the Fourier series on the interval
$(-L,L)$.
The bulk of our eigenvalue calculations were done with $N=100$ but
we went up to $N=250$ when the eigenfunctions have shown
variations on a small scale.
\section{Numerical study}
\label{Num1}
\subsection{Strong damping: $\gamma=0.30$ and $\gamma=0.35$}
We explored $\gamma=0.30$ and $\gamma=0.35$
as two representative
sections of the attractor chart in its right-hand part,
where numerical simulations had detected no period-doubling
bifurcations. The transformation of the solution as it
is continued in $h$ is similar in the two cases; see Fig.\ref{ET_03}.
\begin{figure}[t]
\includegraphics[width = \linewidth]{fig2.pdf}
\caption{The period of the periodic
solutions with $\gamma=0.30$ and $\gamma=0.35$. The solid curves show
stable and
the dashed ones unstable branches.
\label{ET_03}}
\end{figure}
The top-end point of each of the two
curves in figure \ref{ET_03} corresponds
to the stationary single-soliton solution $\psi_+$.
The underlying value of $h$ equals $0.385$ for $\gamma=0.30$
and $0.7500$ for $\gamma=0.35$.
At these $h$, the stationary $\psi_+$ soliton undergoes a Hopf
bifurcation and a stable periodic solution is born.
At the starting point
of each curve, the spectrum
of the periodic soliton
includes three unit multipliers $\mu_{1,2,3}=1$.
As $h$ is increased,
the eigenvalue $\mu_3$ moves inside the unit circle
whereas $\mu_{1,2}$ remain at $1$.
Meanwhile, five real positive eigenvalues $\mu_{4,...,8}$
detach, one after another, from the continuous
spectrum.
One of these ($\mu_8$) later returns to the continuum while the other
four eigenvalues move
towards the unit circle.
At some point the eigenvalue
$\mu_4$ collides with $\mu_3$
producing a complex pair which, however, later returns to the positive
real axis. At the turning point $h_{\rm sn}$, two positive eigenvalues,
$\mu_3$ and $\mu_4$,
cross through the unit circle (almost simultaneously).
This is where the
periodic solution loses its stability.
Numerically, the turning-point value is $h_{\rm sn}=0.8761$
for $\gamma=0.30$ and $h_{\rm sn}=1.0186$
for $\gamma=0.35$.
On the unstable branch, the spectrum includes two
positive eigenvalues $\mu_{3,4}>1$, two unit eigenvalues $\mu_{1,2}=1$,
and
one positive eigenvalue close to unity ($\mu_5<1$).
In addition, two more positive eigenvalues approach the
unit circle from inside as we continue away from the turning point.
\begin{center}
\begin{figure}
\includegraphics[height = 1.7in, width = \linewidth]{fig3a.pdf}
\includegraphics[height = 1.7in, width = \linewidth]{fig3b.pdf}
\includegraphics[height = 1.7in, width = \linewidth]{fig3c.pdf}
\caption{\label{g03_3D} (Color online)
The absolute value of the periodic
solution with large $\gamma$. In this plot,
$\gamma=0.35$. ``Motion pictures" (a) and (c) have been
taken at the same value of $h$ but correspond to the different branches of
the diagram Fig.\ref{ET_03}: (a) $h=0..85$, $T=2.45$
(top branch) and (c) $h=0.85$, $T=2.34$ (bottom branch).
The evolution (b) corresponds to the turning point: $h=1.0185$, $T=2.21$.
In each case
several periods of oscillation are shown. }
\end{figure}
\end{center}
It is fitting to note here that the two turning point values, $h_{\rm sn}=0.8761$
and $h_{\rm sn}=1.0186$, are in a good agreement with the boundaries of
the periodic-attractor existence domain established previously.
Namely, the direct numerical simulation of Eq.\eqref{NLS}
gave values close to $0.86$ and $1.01$ for $\gamma=0.30$ and $0.35$,
respectively \cite{Bondila}.
The end point of the dashed curve ($h=0.61$ for $\gamma=0.30$ and
$h=0.760$
for $\gamma=0.35$)
corresponds to a stationary complex of solitons.
In the $\gamma=0.35$ case, this solution has three
separate humps in its real and imaginary part.
Comparing (the real and imaginary parts of)
each of the three humps to (the real and imaginary parts
of) the free-standing solitons $\psi_+$ and $\psi_-$,
we conclude that the complex consists of the $\psi_+$ soliton
in the middle and two solitons $\psi_-$ on its sides;
that is, the stationary complex should be interpreted as
$\psi_{(-+-)}$. The spectrum of discrete eigenvalues of this
stationary complex
is close to a union of the eigenvalues of the
soliton $\psi_+$ and those of two solitons $\psi_-$.
Namely, we have three unit eigenvalues $\mu_{1,2,5}=1$
(resulting from the translation modes of two $\psi_-$'s and one
$\psi_+$)
and two eigenvalues $\mu_{6,7}$ close to unity (contributed by the
$\psi_+$ soliton which is close to its Hopf bifurcation point).
In addition, the two
$\psi_-$ solitons contribute two real positive eigenvalues $\mu_{3,4}>1$.
In the case of $\gamma=0.30$,
the two side peaks in the real part of the end-point solution
are seen to have merged with the central peak (whose
height coincides with the height of the $\psi_+$
soliton in its real part).
On the other hand, the central peak in the
imaginary part of the solution is
seen to have disappeared while the two lateral peaks have
the same height as the soliton $\psi_-$ in its imaginary part.
This indicates that the solution can still be interpreted as
the $\psi_{(-+-)}$ complex,
albeit a tightly bound one.
Representative solutions are shown in figure \ref{g03_3D}.
Near the leftmost point of the curve pertaining to $\gamma=0.35$ in
Fig.\ref{ET_03}, the periodic solution looks like a single soliton
with a periodically oscillating amplitude and width
[Fig.\ref{g03_3D}(a)]. The soliton is emitting radiation waves; however,
the radiation is decaying rapidly as $|x| \to \infty$. As we move further along this curve in
Fig.\ref{ET_03}, the amplitude of oscillations as well as the
intensity of radiation increases [Fig.\ref{g03_3D}(b)]. The shape of the oscillating
solution evolves into a three-hump structure. Near the end point of the
curve, the amplitude of oscillations decreases
[Fig.\ref{g03_3D}(c)] and we arrive at the stationary
three-soliton complex.
Thus, periodic solitons with large $\gamma$ connect
(in the sense of paths in the parameter space) stationary
solitons and their complexes, with the connection points provided by the Hopf
bifurcations of the latter. It is appropriate to mention a recent
publication \cite{AW} where a similar organisation of the solution manifold
was reported for the Benjamin-Ono equation.
\subsection{Weak damping: $\gamma=0.1$ and $\gamma=0.2$}
\begin{figure}
\includegraphics[width = \linewidth]{fig4.pdf}
\caption{The period of the periodic solutions with
$\gamma = 0.10$ and $\gamma = 0.20$. The solid curves show stable and the dashed one unstable branches.
\label{weak_damping_fig}}
\end{figure}
In the weak damping regime we explored two representative
values of $\gamma$, $\gamma = 0.10$ and $\gamma = 0.20$. The two cases
exhibit similar bifurcation diagrams (Fig. \ref{weak_damping_fig}) which are, however, very
different from the strong-damping diagrams in Fig.\ref{ET_03}.
As before, the starting point (the left-end point)
of each branch corresponds to the stationary $\psi_{+}$ soliton.
At $h = 0.1250$ (for $\gamma = 0.10$) and $h = 0.2275$ (for $\gamma = 0.20$)
the stationary soliton undergoes a Hopf bifurcation and a periodic solution is born.
In Fig.\ref{weak_damping_fig}, this solution is marked $T1$.
The spectrum of the $T1$ solution at the starting point of the curve has three unit multipliers, $\mu_{1,2,3} = 1$.
As $h$ is increased, $\mu_{1}$ moves along the real axis inside the unit
circle and immerses in the continuous spectrum at
$\mu=e^{-\gamma T}$. For even a larger $h$, a pair of complex-conjugate multipliers
with real parts close to $-e^{-\gamma T}$ detaches from the continuum
before converging on the negative real axis. One
of the resulting real eigenvalues moves towards $-e^{-\gamma T}$ and rejoins the
continuum.
The other one moves away from the origin and eventually passes
through $-1$.
This is a signature of the period-doubling bifurcation: the original solution of period $T$ loses its stability and a stable
solution with a period $2T$ is born. (This happens at $h_{2T} = 0.1463$ and $h_{2T} = 0.2663$
for $\gamma = 0.10$ and $\gamma = 0.20$, respectively.)
The spectrum of the newly born $T2$ solution evolves similarly to the spectrum of $T1$.
It is also born with three discrete multipliers $\mu_{1,2,3}=1$.
As $h$ increases, $\mu_{1}$ moves along the real axis; immerses in the continuous spectrum;
reappears at $-e^{-\gamma T}$, and eventually
passes through $-1$ at some $h=h_{4T}$. (The value $h_{4T}$ equals $0.1511$ for $\gamma=0.10$
and $0.2725$ for $\gamma = 0.20$.)
At this point a new stable branch is born, with the period equal to two periods of $T2$.
This new, period-four branch is not shown in Fig. \ref{weak_damping_fig}.
As for the (unstable) $T1$ solution, the corresponding $T(h)$ curve is snaking
up, with the period growing without bound. The solution
approaches a homoclinic orbit, connecting the stationary $\psi_-$ soliton to
itself. In the next subsection, we will discuss this phenomenon in more detail.
We note that the bifurcation diagrams of Fig. \ref{weak_damping_fig}
are in agreement with direct numerical simulations
of Eq.\eqref{NLS} with $\gamma = 0.10$ and $\gamma = 0.20$ \cite{Bondila}
that revealed the period-doubling transition to temporal chaos in soliton's dynamics.
It is therefore natural to expect that our $T4$ branch
will also undergo a period-doubling bifurcation after a short interval of stability,
and similarly the branches that it births would undergo a sequence of period-doubling bifurcations.
We did not have computational capacity to verify this using our two-dimensional continuation approach.
It is worth mentioning, however, that the values of $h$
at which higher-periodic and temporally-chaotic attractors were
observed in simulations
($0.16$ and $0.28$ for $\gamma=0.1$ and $0.2$, respectively \cite{Bondila})
are close to the $h_{4T}$ given above.
Finally, we have not been able to perform
an accurate numerical continuation of solutions with $\gamma=0.05$ and
smaller.
In this case the solution consists of a finite-extent soliton
riding on a (second-harmonic) oscillatory background of a non-negligible amplitude,
which shows only a very slow spatial decay as
$|x| \to \infty$. [Fig.\ref{g03_3D}(b) gives an idea of the shape of the solution in that
case, for most $h$.]
In order to obtain this solution under the boundary conditions
$\psi(\pm L,t)=0$, one has to enlarge the length of the spatial
side of the domain of computation, $(-L,L) \times (0,T)$.
This quickly saturates the computational capacity available.
\subsection{Intermediate damping: $\gamma=0.265$.}
Finally, we investigate the transformation of the periodic
solutions with $\gamma$ lying between the strong- and weak-damping ranges.
As a representative value of such ``borderline" damping, we take
$\gamma=0.265$.
According to the numerical simulations, the {\it direct\/}
period-doubling cascade is followed by an {\it inverse\/} cascade here,
resulting in a peculiar shape of the
periodicity region on the ($\gamma,h$) plane (Fig.\ref{chart_from_PhysicaD}).
Our aim is to provide an explanation for this phenomenon
on the basis of the transformations of the periodic solutions as
$h$ is continuously varied.
\begin{figure}[t]
\includegraphics[width = \linewidth]{fig5.pdf}
\caption{\label{Fig_265}
The period of the
periodic solution for $\gamma=0.265$. Solid curves show stable
and the dashed ones unstable branches.}
\end{figure}
The results of our numerical continuation are summarised in Fig.\ref{Fig_265}.
The bifurcation diagram consists of three branches, the
second and third of which
arise as a result of the period-doubling bifurcation of the
solution on the first branch. The transformation of the solution
as we move along this period-one branch [the bottom branch in
Fig.\ref{Fig_265}, denoted $T1$] is similar to the transformation
of the solution along the curves shown in Fig.\ref{ET_03}. The
starting point of the curve (its leftmost point) corresponds to
the stationary solution $\psi_+$; the end point corresponds to the
stationary complex $\psi_{(-+-)}$.
The major difference from the
case of large $\gamma$ ($\gamma=0.3$ and $\gamma=0.35$) occurs
when a real eigenvalue crosses through
$-1$ as $h$ is increased through $h=0.44$. As $h$ is increased
past $0.44$, this negative eigenvalue continues to grow in
absolute value, reaches a maximum, and then starts to decrease. As
$h$ is increased through $h=0.75$, the negative eigenvalue crosses
through $-1$ once again, this time in the direction of decreasing
modulus. The described behaviour of the negative
eigenvalue corresponds to period-doubling bifurcations at $h=0.44$
and $h=0.75$, where periodic solutions with double period are
detached. For larger $h$ the solution is stable --- all the way to
the turning point,
where two real eigenvalues cross out
through the unit circle.
The spectrum of the first double-periodic solution (detaching at
$h=0.44$ and denoted $T2$ in Fig.\ref{Fig_265}) includes two unit eigenvalues,
$\mu_{1,2}=1$. When $h=0.44$, there is also a negative
eigenvalue $\mu_3=-1$. As $h$ is increased from $0.44$,
$\mu_3$ moves inside the unit circle,
then reverses and crosses through $ -1$
once again (at $h_{4T}=0.495$). At this point, a new periodic solution
$T4$
is born, with the period equal to double the period of the
solution $T2$ --- roughly four times the period of $T1$.
(This ``period-four" solution is not shown in Fig.\ref{Fig_265}.)
The value $h_{4T}=0.495$ coincides with the largest value of $h$
at which higher-periodic attractors were seen in simulations \cite{Bondila}.
Returning to the $T2$ solution,
it has two positive eigenvalues $\mu_{4,5}$
in addition to the eigenvalues $\mu_{1,2,3}$. As the $T2$ curve
in Fig.\ref{Fig_265} is traced
around the turning point at $h=0.543$, these two eigenvalues
cross, almost simultaneously, through the unit circle
so that $\mu_{4,5}>1$ on the upper
branch of the curve.
As we continue further along the upper
branch, the period of the solution grows:
the solution develops a long
epoch where it remains very close to the stationary soliton $\psi_-$.
Each period now consists of two phases: the solution performs a
rapid oscillation with its spatio-temporal profile close to that
of the $T1$ solution, followed by a slow passage
through the bottleneck near the $\psi_-$ soliton
[see Fig.\ref{Fig_2650}].
\begin{center}
\begin{figure}[t]
\includegraphics[height = 1.7in, width = \linewidth]{fig6a.pdf}
\vspace*{3mm}
\includegraphics[width = \linewidth]{fig6b.pdf}
\caption{\label{Fig_2650} (Color online)
(a) The absolute value of the double-periodic solution
found on the left upper branch in Fig.\ref{Fig_265}. Here $\gamma=0.265$, $h=0.502$ and
$T=22.985$. A rapid oscillation
is followed by a long quasistationary
epoch where the solution remains close to the $\psi_-$ soliton.
(b) The phase portrait of this solution taken at $x=0$. Shown is
${\rm Im} \, \psi(0,t)$ vs ${\rm Re} \, \psi(0,t)$.
The filled and open circle represent two fixed points:
the stationary solitons $\psi_+$ and $\psi_-$.
}
\end{figure}
\end{center}
The second period-2 solution
(denoted $\widetilde{T2}$ in Fig.\ref{Fig_265}), which
detaches at $h=0.75$, remains stable for
$0.731\leq h \leq 0.75$. As with the first double-periodic
solution $T2$, the period of $\widetilde{T2}$ grows as we move along the
branch. The solution changes similarly to
what we have described in the previous paragraph: a rapid
oscillation is followed by a long quasistationary epoch when the
solution is very close to the stationary soliton $\psi_-$.
Thus, periodic solutions on the $T2$ and $\widetilde{T2}$ branches approach a
homoclinic orbit: an infinite-period solution which tends to the stationary soliton
$\psi_-$ as $t \to -\infty$ and $t \to \infty$. The same transformation scenario was detected
in the weak-damping case ($\gamma<0.25$) where the $T1$ (and, presumably, the $T2$) branch was
seen to snake up to $T =\infty$ (Fig.\ref{weak_damping_fig}).
At the point $h=h_\infty$ where the period becomes infinite, the periodic solution
undergoes a homocilinic bifurcation which may serve as the source of chaos
(see section \ref{explosion} below).
\section{Radiation from the oscillating soliton}
\label{Rad}
\subsection{The long-range radiation indicator}
Like their stationary counterparts,
the oscillatory solitons strike the balance between the
energy fed by the driver and the energy lost to the dissipation.
This can be quantified, for example, by considering the integral
$N= \int_{-R}^R |\psi|^2 dx$. [When the nonlinear Schr\"odinger equation is employed to
describe planar stationary waveguides, this
integral measures the total power of light
captured by the $(-R,R)$ section of the waveguide.
There are energy-related interpretations of this integral in other contexts as well.]
Equation \eqref{NLS} yields
\begin{equation}
{\dot N}= 2h \int_{-R}^R \left. |\psi|^2 \sin (2 \theta) dx- 2 \gamma N + \Phi \right|^R_{-R},
\label{Ndot}
\end{equation}
where we have decomposed $\psi$ as $|\psi| e^{-i \theta}$
and defined
\begin{equation}
\Phi(x)= i ( \psi_x \psi^* -\psi_x^* \psi).
\label{Phi}
\end{equation}
The first term in the right-hand side of \eqref{Ndot} gives the rate at which the energy is pumped into the soliton
while
the second one quantifies the damping rate.
Assuming that the interval $(-R,R)$ is large enough to contain
the core of the soliton, the last term in \eqref{Phi} measures the radiation
flux through its endpoints.
The radiation through the points $x=\pm R$ outside the
soliton's core is governed by the linearisation of \eqref{NLS}
whose dispersion relation is
\begin{equation}
(\omega -i \gamma)^2= (k^2+1)^2-h^2.
\label{dr}
\end{equation}
The radiation waves emanate from the oscillating soliton
which plays the role of a pacemaker for these waves.
The pacemaker has a
a period
$T$ and will sustain waves with
frequencies
$\omega= 2 \pi n /T$ ($n=1,2,...$); hence $\omega$ should be taken real in \eqref{dr}.
For real $\omega$, equation \eqref{dr} has two pairs of complex roots
$k= \pm (p_{1,2} + iq_{1,2})$, with the imaginary components
\begin{equation}
q_{1,2}= \sqrt{
\frac{1}{2} \pm \frac{{\cal S}}{2}
+ \frac12 \sqrt{1+2({\cal Q} \pm {\cal S})}}, \label{q12}
\end{equation}
where
\[
{\cal S}=\sqrt{{\cal P}+{\cal Q}}, \quad {\cal Q}=\sqrt{{\cal P}^2+ \gamma^2 \omega^2},
\]
and $2{\cal P}=h^2-\gamma^2 +\omega^2$.
The decay rate $q_1$ is always greater than $q_2$
and bounded from below ($q_1 > 1$); it accounts for rapidly decaying core of the soliton.
Outside the core,
the decay rate crosses over
to $q_2$, and if $q_2$ is small the solution enters the long oscillatory wing
decaying in proportion to $e^ {-q_2 |x|}$.
Periodic solutions that we have reported in this paper
all have $\omega \sim 1$; for these, the exponent $q_2$ can be small only if
$\gamma$ is small.
Assuming that $\gamma$ is small, the expression \eqref{q12} simplifies. Defining the quantity
\begin{equation}
\sigma=\sigma(\omega)= \sqrt{ \omega^2+h^2-\gamma^2}-1,
\label{sigma}
\end{equation}
we have, in particular,
\begin{subequations} \label{sigamma}
\begin{align}
q_2 \to
\frac{\omega/2}{1+\sigma} \left(\frac{\gamma}{\sigma}\right)^{1/2} \gamma^{1/2}
\quad & \text{as} \ \frac{\gamma}{\sigma} \to 0
\ & (\sigma>0);
\label{q2} \\
q_2 \to \sqrt{\frac{\omega/2}{1+\sigma}} \, \gamma^{1/2} \quad & \text{as} \ \frac{\sigma}{\gamma} \to 0; \\
q_2 \to (-\sigma)^{1/2} \quad & \text{as} \ \frac{\gamma}{\sigma} \to 0 \ & (\sigma<0).
\end{align}
\end{subequations}
Thus $q_2$ is small if $\gamma$ is small
while $\sigma$ is positive. In particular, if $\sigma$ is much greater than $\gamma$, the exponent
$q_2 \sim \gamma/\sqrt{\sigma}$ while if $\sigma$ is much smaller than $\gamma$, we have
$q_2 \sim \sqrt{\gamma}$. The decay rate stays small ($\sim \sqrt{-\sigma}$) when $\sigma$ becomes
negative --- as long as it remains much smaller than $\gamma$
in absolute value. However, as $\sigma$ grows to larger negative values,
the exponent $q_2$ grows as well.
Hence the quantity $\sigma(\omega)$ serves as an indicator of whether the
long-range radiation with the frequency $\omega$ can be excited (i.e. whether $q_2$ is small)
or not.
\subsection{Radiation frequency selection}
The harmonic-wave solution of the linearised equation with the decay rate $q_2$ is
\begin{equation}
\psi = \eta \left( \cos \phi +
i \frac{\alpha \sin \phi- \beta \cos \phi}{({\cal S}+h)^2+ {\cal T}^2}
\right) e^{q_2x},
\label{hw}
\end{equation}
where the phase $\phi=\omega t- p_2x$, the wavenumber $p_2=-{\cal T} /(2q_2)$, and
the coefficients
\[
\alpha= \omega ({\cal S}+h) + \gamma {\cal T}, \quad
\beta= \gamma({\cal S}+h) - \omega {\cal T},
\]
and ${\cal T}=\sqrt{{\cal Q}-{\cal P}}$. The amplitude $\eta=\eta(\omega)$ is
arbitrary as far as the linearised equation is concerned but acquires a specific
value when \eqref{hw} is used to represent radiation from the soliton.
If $R$ is large enough, the solution $\psi(x,t)$ with $x$ near $\pm R$ is described
by \eqref{hw}.
Substituting \eqref{hw} in \eqref{Phi} we obtain the flux
through the points $x=\pm R$:
\begin{equation}
\Phi(\pm R)= \mp |\eta|^2 \frac{\cal T}{q_2} \, \frac{ \omega({\cal S}+h) +\gamma {\cal T} }
{({\cal S}+h)^2+{\cal T}^2} \, e^{-2q_2R}.
\label{flux}
\end{equation}
Assume, first, that $q_2$ is of order $1$; this occurs, in particular, when $\gamma \sim 1$.
Since $R$ is chosen to be large, the flux through the points $x=\pm R$ in this case
is exponentially small:
strongly damped radiation waves decay so quickly that they do not reach beyond the
core of the soliton (and essentially form part of it).
The exponential factor in \eqref{flux} is not exponentially small only if
$q_2$ is small; this, in turn, may only happen when the damping is weak.
These considerations have a simple physical interpretation.
When $\gamma$ is large, the soliton dissipates most of its excess energy within its core;
this is accounted for by the second term in \eqref{Ndot}. On the other hand,
when $\gamma$ is small, the dissipation within the core
is insufficient to balance the first term in \eqref{Ndot}.
The long-range radiation may provide an alternative energy-draining mechanism in this case,
but this is only available when $q_2$ is small which,
in turn, is indicated by $\sigma>0$.
If the period of the solution is small enough for the quantity $\sigma(2 \pi/T)$ to be positive
and much greater than $\gamma$
(which is assumed small), the decay rate $q_2$ is close to zero. Therefore, the flux
\eqref{flux} is not exponentially small and the first-harmonic radiation
is not blocked in this case.
As $T$ is increased so that $h^2-\gamma^2+ (2 \pi/T)^2$ drops below 1,
the decay rate $q_2$ grows to values of order $1$. According to eq.\eqref{flux},
the
first-harmonic radiation will be mostly suppressed in this case (i.e. damped outside a small neighbourhood of the soliton's core).
However, $\sigma(\omega)$ with $\omega= 4 \pi/T$ will be still positive; this means that the tails will be
dominated by the second-harmonic radiation now. Once the second harmonic is suppressed,
the third harmonic will take over, and so on.
We note that the second-harmonic radiation is present not only when the first-harmonic
radiation is suppressed. However, the amplitude $\eta(2 \pi n/T)$ typically scales
as $\epsilon^n$ and so Eq.\eqref{flux} implies that the second-harmonic
flux is much weaker than the first-harmonic flux in situations where both channels are open.
\subsection{First-harmonic radiation as a stabilising agent}
We routinely calculated the indicator $\sigma(2 \pi /T)$ as we
continued our periodic solitons. The frequency spectrum of the radiation
shows a remarkable correlation with the type of transition the soliton
is suffering.
The indicator \eqref{sigma} with $\omega= 2 \pi/T$ is
{\it positive\/} along each curve in Fig.\ref{ET_03} ---
moreover, $\sigma(\omega)$ is much greater than $\gamma$.
Therefore,
in the case $\gamma>0.275$, the long-range radiation from the
soliton is dominated by the first harmonic. On the other hand,
$\sigma(2\pi/T)$ is {\it negative\/} along each
$T1$ branch in Fig.\ref{weak_damping_fig} which implies that the
first-harmonic radiation is short-range when $\gamma$ is smaller than $0.25$.
(When $\gamma=0.10$, $\sigma$ decreases from values close to $-\gamma$ to about
$-6 \gamma$ along the curve;
when $\gamma=0.20$, $\sigma$ drops from zero to nearly $-2.5 \gamma$.)
In the latter case ($\gamma=0.10$ and $0.20$) the indicator $\sigma(\omega)$ becomes positive if
one lets $\omega= 4 \pi/T$; hence the radiation is second-harmonic
here (except for very large $T$ where it is third-harmonic).
This observation allows us to understand, in qualitative terms, why the
transformation of the soliton with $\gamma>0.275$ is different from
the scenario realised for very weak dampings ($\gamma<0.25$). When $\gamma$ is not very small, the soliton can
dispose the excess energy via two channels. Some energy will be lost
to the damping in the core region, where $|\psi| \sim 1$ and the term
$i \gamma \psi$ is $\sim \gamma$. The rest will be sent away in the
form of the strong first-harmonic radiation.
Due to the availability of this powerful back-up dissipation channel,
the soliton behaves as an overdamped system; hence its enhanced stability.
On the other hand, when $\gamma$ is very small, the dominant long-range radiation is second harmonic.
Since the amplitude of the $n$-th harmonic
radiation scales as $\epsilon^n$ where $\epsilon$ is the scale of the amplitude of
the oscillation in the core, the second-harmonic radiation is much weaker than the
first-harmonic one, and the radiation channel is effectively blocked. The soliton shows the volatility
of an underdamped system; hence its instability and period doublings.
The formation of three-soliton complexes observed
in the continuation of the oscillatory solitons with $\gamma=0.30$ and $\gamma=0.35$ is also made
possible due to the availability of strong radiation with the period equal to the
period of the central soliton. The lateral humps in the complex are seeded by the two maxima of the
radiation sinusoid which are closest to the central soliton.
On the other hand, the second-harmonic radiation available in the
$\gamma=0.10$ and $\gamma=0.20$ cases is unable to seed the lateral solitons
because it is weak and has the period different from the period of the central hump.
It is interesting to test these qualitative considerations using the case
of the intermediate damping, $\gamma=0.265$. In this case the first-harmonic indicator
$\sigma(2\pi/T)$ is positive but $\sigma/\gamma$ is {\it not\/} much greater than 1
over the most of the top $T1$ branch in
Fig.\ref{Fig_265}.
(As $h$ grows from $0.35$ to $0.71$ to $0.78$, the quotient
$\sigma/\gamma$ on the top branch changes from $0.5$ to $0$ to $1.$).
Accordingly, the first harmonic is strongly damped
($q_2 \sim \sqrt{\gamma}$) and the long-range radiation is second harmonic.
As a result, the soliton suffers period-doubling bifurcations, as
in the case of $\gamma=0.10$ and $0.20$. However, as we continue
further,
the quotient $\sigma/\gamma$ grows (to $2.$ at $h=0.82$)
and the first-harmonic radiation takes over
from the second-harmonic. Consequently, the solution transforms into
a three-soliton complex --- as in the case of $\gamma=0.30$ and $0.35$).
\section{Discussion and conclusions}
\label{DC}
\subsection{Homoclinic explosion}
\label{explosion}
The numerical continuation of the T1 solution with $\gamma< 0.25$
shows an unbounded growth of its period. The same is true for the T2
solution with $\gamma< 0.275$. These observations imply the presence of a
homoclinic bifurcation where the period becomes infinite. We now argue
that this bifurcation is a source of chaos.
We use results of Shilnikov \cite{Shilnikov_1}
who considered a three-dimensional dynamical system with a homoclinic orbit
connecting
a saddle-focus to itself. The saddle-focus is a fixed point with
a real unstable eigenvalue $\lambda_1>0$
and two complex-conjugate stable eigenvalues $\lambda_2=\lambda_3^*$,
${\rm Re\/} \, \lambda_{2,3}<0$.
Assuming that
\begin{equation}
\lambda_1 > | {\rm Re\/} \, \lambda_{2,3}|,
\label{Shil}
\end{equation}
Shilnikov proved that
at the bifurcation point, the system has infinitely many Smale's horseshoes
in a neighbourhood of the homoclinic orbit \cite{Shilnikov_1} (see also \cite{Shilnikov_2}).
Each of these horseshoes contains an invariant Cantor set with
a countable infinity of (unstable) periodic orbits and an uncountable infinity
of (unstable) bounded aperiodic orbits --- the strange invariant set
which manifests itself as attractive or transient chaos.
Later it was shown that finitely many of the horseshoes persist on one side of the homoclinic
bifurcation \cite{GG}.
The homoclinic bifurcation giving birth to this plethora of periodic and chaotic
orbits has been referred to as the
``homoclinic explosion" \cite{Sparrow}.
The homoclinic explosion may also occur in four- and higher-dimensional systems
with a homoclinic orbit. The necessary conditions for this are formulated
in terms of the eigenvalues of the linearisation about the fixed point connected to itself by the orbit: (a) out of all
eigenvalues in the right half of the complex plane, the closest to the imaginary axis
is a real eigenvalue $\lambda_1$; (b)
in the left half of the complex plane, a pair of complex-conjugate eigenvalues $\lambda_{2,3}$
are the closest eigenvalues to the imaginary axis. The homoclinic explosion occurs then if
the Shilnikov inequality \eqref{Shil} is satisfied \cite{Shilnikov_2}. A carefully studied and clearly explained
example of homoclinic explosion in partial differential equations is in
\cite{Knobloch}.
The role of the saddle-focus fixed point in our equation \eqref{NLS} is played by the
soliton $\psi_-$, Eq.\eqref{solmin}.
The linear spectrum of this fixed point consists of two
discrete eigenvalues, Eqs.\eqref{lambdas} of the Appendix, and a continuum of values
$\lambda= -\gamma+i \omega$,
where $|\omega| \ge \omega_0$, $\omega_0=\sqrt{1-h^2}$.
In the left half of the complex
plane, the eigenvalue $\lambda_2$
is further away from the imaginary axis than the continuous spectrum; hence
the dynamics near the saddle-focus will be dominated by the single positive eigenvalue $\lambda_1$
and the continuum of eigenvalues with negative real part. This is not exactly
the situation of Shilnikov; however it is similar
in the sense that the fixed point has a one-dimensional unstable manifold
and a more-than-one-dimensional stable manifold, formed by orbits
spiralling into the focus. It is natural to expect
Shilnikov's result to remain valid in this case, provided an
analog of the inequality \eqref{Shil}
is in place --- although we
do not have any rigorous proof of that. The analog of the Shilnikov inequality in our case
is
\begin{equation}
\lambda_1 > \gamma,
\label{inne}
\end{equation}
with $\lambda_1$ as in \eqref{lambdas}.
The inequality \eqref{inne} translates into $(A_-^2/ \sqrt{3}) \Lambda_0 > \gamma$,
where $\Lambda_0(\epsilon)$ is the positive eigenvalue of the problem
\eqref{La}.
One can easily demarcate the region where this inequality is valid.
From \eqref{epsie} we have
\begin{equation}
h=\sqrt{\left( \frac{\epsilon}{2+\epsilon} \right)^2+ \gamma^2 }.
\label{heps}
\end{equation}
Equation \eqref{heps} together with
\begin{equation}
\gamma = \frac{ A_-^2 }{\sqrt{3}} \Lambda_0(\epsilon)
\label{geps}
\end{equation}
define the parametric curve $h=h_{\rm Sh}(\gamma)$ on the $(\gamma,h)$ plane. The curve is
parabolic in shape (see the dotted curve in the main frame and inset in
fig.\ref{chart_b}); it connects the origin to the point
$\gamma=0$, $h=1$. The inequality \eqref{inne} is valid in the region bounded by this curve
and the $h$-axis.
This region obviously includes the loci of the homoclinic bifurcations for all $\gamma<0.275$
(see fig.\ref{chart_b}). Therefore the homoclinic bifurcation is indeed of the
homoclinic-explosion type in our case. It is this bifurcation that is responsible for the
appearance of temporally chaotic solitons in the parametrically
driven damped nonlinear Schr\"odinger equation.
\subsection{Attractor charts}
1. Our analysis answers a number of
questions raised by the attractor chart
\cite{Bondila} compiled via direct numerical simulations.
One such question concerns the existence of two types of the
transformation scenario of the parametrically driven soliton,
as the driving strength is raised for the fixed damping coefficient.
Is the transformation followed by the soliton with large $\gamma>0.275$ really
different from the transition to chaos suffered by its weakly damped counterpart
($\gamma<0.25$), or
is this difference caused merely by numerical approximations?
The answer to this question is
provided by diagrams in Figs. \ref{ET_03}, \ref{weak_damping_fig}, and \ref{Fig_265}.
In the region $\gamma<0.25$
(represented by the values $\gamma=0.10$ and $0.20$,
Fig.\ref{weak_damping_fig}) the increase of $h$ results in
a sequence of
period-doubling bifurcations of the periodic soliton. On
the contrary, no period-doubling
bifurcations occur as $h$ is increased for the fixed $\gamma>0.275$.
(See Fig.\ref{ET_03} showing the diagrams for $\gamma=0.30$ and $0.35$).
2. Another question is why no periodic solitons
are observed in numerical simulations with sufficiently large $h$.
Is there a well-defined
boundary of the domain of existence of stable
periodic solitons? We have shown that for
$\gamma>0.275$, the region of existence of periodic solitons
on the $h$ axis is bounded by the saddle-node bifurcation.
No such solutions, stable or
unstable, exist above this turning point.
For small $\gamma$, $\gamma<0.25$, the region of existence
of periodic solitons is also bounded by a turning point; however,
solitons cease to exist as attractors for lower values of $h$.
The boundary of the stability domain here is set by a
narrow strip of temporally-chaotic solitons.
These conclusions verify and explain the formation of the ``desert region" on the
$(\gamma,h)$ plane. The positions of the saddle-node bifurcation point
for $\gamma >0.275$ and the accumulation point of higher-periodic solutions
for $\gamma<0.25$, are in good agreement with the boundary of the
desert region detected in simulations \cite{Bondila}.
3. The third issue clarified by our analysis pertains to the
shape of the
region on the ($\gamma,h$)-plane where periodic solitons are observed in
numerical simulations. The question here is why does the $h(\gamma)$
curve bounding this region fold on itself for $\gamma$
between $0.25$ and $0.275$.
This phenomenon is accounted
for by the restabilisation of the periodic soliton as $h$ is increased
for $\gamma=0.265$ (Fig.\ref{Fig_265}).
We note that the case $\gamma=0.265$ is intermediate between
the small- and large-$\gamma$ transformation scenarios:
On the one hand, similar to the weakly-damped scenario,
the soliton undergoes a period-doubling cascade
as $h$ is raised for this $\gamma$ (more precisely, it undergoes
{\it two\/} period-doubling cascades). On the other, the domain occupied by
the periodic attractor is bounded by a saddle-node bifurcation in this case,
as in the strongly-damped situation.
\begin{figure}
\includegraphics[width = \linewidth]{fig7.pdf}
\vspace*{1mm}
\caption{\label{chart_b}
(Color online)
Single-soliton attractor
chart.
The Shilnikov inequality \eqref{inne} holds
inside the region bounded by the dotted curve.
The inset gives a wider perspective of this region.
}
\end{figure}
Fig.\ref{chart_b} summarises our conclusions on single-soliton periodic
attractors. This diagram is in a good agreement with the attractor chart produced using direct numerical
simulations, Fig.\ref{chart_from_PhysicaD}.
|
\section{Introduction}\label{sect1}
In phenomenological applications of QCD, the operator product expansion (OPE)
\cite{wil69,wit77} plays an essential role. It is for example employed in
applications of QCD sum rules \cite{svz79} or the analysis of the $\tau$
hadronic width \cite{bnp92}. Considering in particular two-point correlation
functions of mesonic currents, strictly speaking the OPE is only appropriate
in the Euclidean domain where $s\equiv q^2$ is sufficiently large and negative,
with $q$ being the four-momentum transfer between the two currents. In the
physical, Minkowskian domain with positive $s$, generally bound states or
resonances are present which are governed by long-distance physics and which
cannot be captured by a short-distance expansion.
In the Euclidean domain, a {\em duality} exists between descriptions of the
system in terms of the quark and gluon degrees of freedom of QCD, or the
physical hadronic degrees of freedom. Upon analytic continuation of correlation
functions towards the Minkowskian region, due to the presence of bound-state
poles, violations of the quark-hadron duality are expected to show up
\cite{shi94,shi95,cdsu96,bsz98}.\footnote{For a review the reader is also
referred to ref.~\cite{shi00}.} Since until now no analytic solution to QCD
has been found -- and it is unlikely that this will happen anytime soon -- the
precise functional form of {\em duality violations} (DVs) remains unknown. This
necessitates to resort to the study of models in order to investigate their
influence in phenomenological analyses of QCD. This is particularly pressing
when employing so-called QCD finite-energy sum rules \cite{ckt78,kpt83},
because here contour integrals in the complex energy plane have to be computed
down to the Minkowskian axis.
Based on the models presented in refs.~\cite{shi94,shi95,bsz98,shi00}, the
influence of duality violations on hadronic decays of the $\tau$ lepton has
been investigated in recent years \cite{cgp05,cgp08,cgp09,gpp10a,gpp10b},
because they may play a role in precision determinations of fundamental QCD
parameters, like the strong coupling $\alpha_s$
\cite{aleph05,bck08,bj08,ddhmz08,my08}. For example it has been observed that
the compatibility of values for the gluon condensate $\langle\alpha_s GG\rangle$, extracted from
fits to vector and axialvector $\tau$ decay spectra \cite{aleph05,ddhmz08},
is not very satisfactory. This poses the legitimate question if the inclusion
of DVs in a consistent analysis may improve the situation regarding the
compatibility of condensate parameters extracted from different channels,
and in how far this would affect the resulting value of $\alpha_s$.
Though the particular model for vector and axialvector correlation functions
employed in the analyses \cite{cgp05,cgp08} is based on features of large-$N_c$
QCD \cite{tho74} and Regge theory \cite{reg59,col71}, it is not directly rooted
in fundamental QCD. Thus, it appears important to detect other systems which
allow for an independent investigation of DVs, in particular their functional
form as well as their possible impact on phenomenological analyses. A system
that allows for an analytical treatment, and bears close relation to quarkonia
in QCD in the limit of a heavy quark mass, is the Coulomb system. In the
context of QCD moment sum rules, to some extent this example was already
studied in ref.~\cite{vol95,eid02}. Furthermore, in relation with hadronic
decays of $B$ mesons, duality-violating contributions resulting from charmonium
states were also considered in refs.~\cite{bbns09,bbf11}.
In the first half of this article, it will be explored in more detail what can
be learned from the Coulomb system regarding violations of duality. To this
end, the notion of duality will be adopted in a more general context. The point
of view taken in the following will be that a certain expansion is employed
which in general can only be considered to be of an asymptotic nature. In our
case this refers to the OPE, and {\em duality} is supposed to imply that the
asymptotic expansion provides an acceptable representation of the full function.
On the other hand, in some regions of the expansion, formally exponentially
suppressed contributions may become relevant, and those will be identified
with DVs.\footnote{This somewhat broader perspective could also be applied to
the perturbative expansion with the ``duality violations'' being the power
corrections due to the presence of QCD vacuum condensates.}
One finding of the study of the Coulomb system below will be that due to the
presence of zero-width bound-state poles on the real $s$-axis, the effects of
DVs turn out to be particularly strong. The same conclusion has been drawn on
the basis of the `t Hooft model \cite{mp09}. For this reason, the example of
the Coulomb system will mainly be used to gain further general insights into
the structure of DVs. For a more quantitative analysis, in the second half of
this work recourse will again be taken to a simplified version of the model
employed in refs.~\cite{cgp05,cgp08} in the case of a finite width for the
resonances. The finite width entails that the bound state poles move to the
unphysical region of the complex $s$-plane and consequently, DVs in the
physical region turn out to be less pronounced.
In refs.~\cite{cgp08,cgp09} a simplified Ansatz was advocated in order to
incorporate DV contributions into phenomenological QCD analyses. This Ansatz
is supposed to be admissible if the resonances are sufficiently broad and
thus lie sufficiently far away from the Minkowskian, physical axis. Towards
the end of this article, arguments will be given why a related Ansatz should
also be considered on an equal footing which is applicable in part of the
complex energy plane, and the corresponding Ansatz is provided.
\section{Sum rules for the Coulomb system}\label{sect2}
In the following, sum rules for the quantum mechanical Coulomb system will
be set up, which are analogous to the sum rules studied in QCD \cite{svz79}.
Quantum mechanical sum rules had already been investigated in the early years
of QCD sum rules \cite{vzns80,bb81,nsvz81,pt84}, as analytical examples in
which the techniques applied in QCD sum rule analysis could be studied and
tested. Here, this route shall be followed for the question of duality
violations.
Before going into the particular example of the Coulomb sum rule, let us
set up the framework of quantum mechanical sum rules in more general terms
\cite{pt84}. Consider a particle of mass $m$ under the influence of a potential
$V(\vec x)$. The Schr\"odinger equation for stationary states takes the form
\begin{equation}
\hat H \,\psi(\vec x) \,\equiv\, \biggl[\,-\,\frac{\Delta}{2m}
+ V(\vec x) \,\biggr] \psi(\vec x) \; = \; E\,\psi(\vec x) \,.
\end{equation}
Let $\psi_\alpha(\vec x)$ be the eigenfunction corresponding to the
eigenvalue $E_\alpha$. The resolvent operator $\hat G(z)$ of $\hat H$ is
defined as
\begin{equation}
\hat G(z) \,\equiv\, \Big[\,\hat H-z\hat 1\,\Big]^{-1} \,,
\end{equation}
where $z$ is an arbitrary complex number. Its matrix elements in the
position representation are given by
\begin{equation}
\label{eq:Gmat}
G(\vec x,\vec y;z) \,\equiv\, \langle\vec x|\hat G(z)|\vec y\rangle \; = \;
\sum\limits_\alpha\; \frac{\psi_\alpha(\vec x)\psi_\alpha^*(\vec y)}
{(E_\alpha-z-i0)} + \int \frac{\rho_{\rm cont}(\vec x,\vec y;E)}
{(E-z-i0)}\,{\rm d}E \,,
\end{equation}
where the sum runs over the discrete spectrum, the integral is taken over
the continuous spectrum and $\rho_{\rm cont}(\vec x,\vec y;E)$ is the spectral
density corresponding to the continuous spectrum.
Let us introduce the function $D(E)$ which plays the analogous role of
a physical correlation function in QCD sum rules, for example the Adler
function \cite{adl74}:
\begin{equation}
D(E) \,\equiv \, \biggl[\,\frac{{\rm d}}{{\rm d}E}\,G(\vec x,\vec y;E)\,
\biggr]_{\vec x=\vec y=0} \,.
\end{equation}
Using eq.~\eqn{eq:Gmat}, $D(E)$ can be expressed as
\begin{equation}
\label{DE}
D(E) \,=\, \sum\limits_\alpha\; \frac{|\psi_\alpha(0)|^2}{(E_\alpha-E-i0)^2} +
\int \frac{\rho_{\rm cont}(E')\,{\rm d}E'}{(E'-E-i0)^2} \,\equiv\,
\int \frac{\rho(E')\,{\rm d}E'}{(E'-E-i0)^2}\,,
\end{equation}
with appropriate limits of integration, and where we furthermore have
introduced the full spectral function
\begin{equation}
\label{rhoE}
\rho(E) \,=\, \rho_{\rm pole}(E) + \rho_{\rm cont}(E) \,=\,
\sum\limits_\alpha\, |\psi_\alpha(0)|^2\,\delta(E-E_\alpha) +
\rho_{\rm cont}(E) \,,
\end{equation}
containing both the discrete as well as the continuous spectrum.
Let us now particularise the general expressions to the Coulomb problem. The
ensuing sum rules turn out to be closely related to heavy-quark sum rules in
QCD. Consider two particles of mass $m$ coupled by a Coulomb potential of the
form
\begin{equation}
V(r) \,\equiv\, -\,\frac{\alpha}{r} \,.
\end{equation}
The Coulomb potential is written with a general coupling $\alpha$, though
one may well regard this as the leading term in the heavy-quark potential
containing the QCD coupling $\alpha_s$. Solving the relevant Schr\"odinger
equation, the radial Green function for the Coulomb potential is found to
be \cite{vol95}\footnote{As a matter of principle, also higher-order QCD
corrections could be included in a systematic way. See e.g. ref.~\cite{my98}.
However, for the present purposes, it suffices to stay at the leading order.}
\begin{equation}
G(r,0;E) \,=\, \frac{mk}{2\pi}\,e^{-kr}\,\Gamma(1-\lambda)\,
U(1-\lambda,2;2kr) \,,
\end{equation}
with
\begin{equation}
\lambda \,\equiv\, \frac{\alpha m}{2k} \qquad \mbox{and} \qquad
k \; \equiv \; \sqrt{-\,m(E+i0)} \,.
\end{equation}
$U(\alpha,\beta;z)$ is the confluent hypergeometric function. $G(r,0;E)$
is singular in the limit $r\rightarrow 0$, but the physical correlation
function $D(E)$ remains finite and is found to be
\begin{equation}
\label{MEcoul}
D(E) \,\equiv\, \frac{{\rm d}}{{\rm d}E}\,G(r,0;E) \Big|_{r=0} \,=\,
\frac{m\lambda}{4\pi\alpha}\Big[\, 1 + 2\lambda+2\lambda^2
\psi'(1-\lambda) \,\Big] \,,
\end{equation}
with $\psi(z)\equiv {\rm d}[\ln\Gamma(z)]/{\rm d}z$ being the logarithmic
derivative of the Gamma function. The function $D(E)$ has poles at
$\lambda_n=n$, $n=1,2,\ldots$ which translate to the bound-state energies
$E_n$ of the Coulomb system with
\begin{equation}
E_n \,=\, -\,\frac{\alpha^2m}{4n^2} \,.
\end{equation}
Expanding $D(E)$ around the position of the bound states, $E=E_n-\varepsilon$, from
the leading $1/\varepsilon^2$ singularity the square of the wave function at the
origin can be extracted,
\begin{equation}
|\psi_n(0)|^2 \,=\, \frac{\alpha^3m^3}{8\pi n^3} \,,
\end{equation}
which of course agrees with well known results in textbook quantum
mechanics \cite{gp90}.
Besides the discrete spectrum for $E<0$, in the case of the Coulomb problem
also a continuous spectrum is present for $E\geq0$. Taking the imaginary part
of eq.~\eqn{eq:Gmat}, one observes that the spectral density $\rho(E)$ is
related to the imaginary part of $G(r,0;E)$, namely
\begin{equation}
\rho(E) \,=\, \frac{1}{\pi}\,\mbox{\rm Im}\, G(0,0;E+i0) \,.
\end{equation}
Again, also $\mbox{\rm Im}\,G(r,0;E)$ is finite in the limit $r\rightarrow 0$ since
it is a physical quantity. Calculating $\rho_{cont}(E)$ from eq.~\eqn{MEcoul},
yields
\begin{equation}
\rho_{cont}(E) \,=\, \frac{\alpha m^2}{4\pi\Big[1-e^{-\pi\alpha\sqrt{m/E}}\Big]}
\,, \qquad E \, \geq \, 0 \,,
\end{equation}
which is known as the so-called Sommerfeld factor \cite{som31}. Putting
everything together, a finite-energy sum rule, analogous to the case of
hadronic $\tau$ decays, for the quantum mechanical Coulomb problem
can be written down:
\begin{eqnarray}
\label{coulsr}
R^w(E_0) \,\equiv\, \int\limits_{E_1}^{E_0} w(E)\,\rho(E)\, {\rm d}E
&\,=\,&
\frac{i}{2\pi} \oint\limits_{E_0} w(E)\, G(0,0;E)\, {\rm d}E \nonumber \\
\vbox{\vskip 8mm}
&\,=\,&
\frac{i}{2\pi} \oint\limits_{E_0} W(E)\, D(E)\, {\rm d}E \,,
\end{eqnarray}
which defines the moment $R^w(E_0)$ corresponding to an arbitrary analytic
weight function $w(E)$, and the weight function $W(E)$ is given by
\begin{equation}
\label{WE}
W(E) \,=\, \int\limits_E^{E_0} w(E')\, {\rm d}E' \,.
\end{equation}
The central idea of finite-energy sum rules is to employ a phenomenological
representation of the spectral function $\rho(E)$ on the left-hand side of
eq.~\eqn{coulsr} and a theoretically motivated, like the OPE, for $D(E)$ in
the contour integral on the right-hand side. Equating both sides then allows
to infer information on the physical spectrum from the theoretical description,
or to extract theoretical parameters from the phenomenological information. As
most probably in practical applications the theoretical expansion of $D(E)$ on
the right-hand side is only of an asymptotic nature, the consequences for the
finite-energy sum rule need to be investigated.
\section{Asymptotic expansion and numerical analysis}\label{sect3}
For sufficiently large positive or negative energy $E$, that is $|\lambda|<1$,
the $\psi'$-function appearing in $D(E)$ of eq.~\eqn{MEcoul} has a convergent
expansion \cite{as72}. Therefore, in this energy region no duality violations
are expected to occur. They may, however, arise in energy regions where
$\psi'(z)$ only has an asymptotic expansion, which is the case for
$|z|=|1-\lambda|$ going to infinity, while $|\arg z|<\pi$. This happens if
the energy is close to the continuum threshold $E=0$. In this region, the
asymptotic expansion of $\psi'(z)$ is given by \cite{as72}
\begin{equation}
\label{psipas}
\psi'(z) \,\sim\, \frac{1}{z} + \frac{1}{2z^2} + \sum\limits_{n=1}^\infty
\frac{B_{2n}}{z^{2n+1}} \qquad (z\to\infty\;{\rm in}\;|\arg z|<\pi) \,,
\end{equation}
where $B_{2n}$ are the Bernoulli numbers. Even though strictly speaking the
asymptotic expansion \eqn{psipas} should be valid in the full region
$|\arg z|<\pi$, due to the poles of $\psi(z)$ for $z=0,-1,-2,\ldots$, the
asymptotic expansion becomes very inefficient if $z$ approaches these poles.
Quite generally, in the left-half imaginary $z$-plane, that is
$\mbox{\rm Re}\,z<0$, the asymptotic expansion can be greatly improved
by applying the so-called {\em reflection relation} \cite{as72}
\begin{eqnarray}
\label{refrel}
\psi'(z) \,&=&\, -\;\psi'(-z) + \frac{1}{z^2} + \frac{\pi^2}{\sin^2(\pi z)}
\nonumber \\
\vbox{\vskip 8mm}
\,&\stackrel{z\to\infty}{\sim}&\, \frac{1}{z} + \frac{1}{2z^2} +
\sum\limits_{n=1}^\infty \frac{B_{2n}}{z^{2n+1}} + \frac{\pi^2}{\sin^2(\pi z)}
\qquad (\,\mbox{\rm Re}\,z < 0\,) \,.
\end{eqnarray}
A more quantitative account on the improvement achieved by including the
additional term in the expansion of \eqn{refrel} will be given in the second
example of the next section.
As the asymptotic expansions of $\psi'(z)$ and $-\,\psi'(-z)+1/z^2$ are
identical, it is clear that the additional term $\pi^2/\sin^2(\pi z)$ has to
be exponentially suppressed. This can easily be verified by rewriting the
$\sin$-function in terms of exponentials, which yields
\begin{equation}
\label{DVterm}
\frac{\pi^2}{\sin^2(\pi z)} \,=\, -\,4\pi^2 {\rm e}^{\pm 2\pi iz} +
{\cal O}\left({\rm e}^{\pm 4\pi iz}\right) \qquad {\rm for}\;\;
\mbox{\rm Im}\,z\!\!\phantom{x}^{>}_{<}\; 0 \,.
\end{equation}
On the other hand, close to the poles of $\psi'(z)$ the exponentially
suppressed term provides an essential contribution as it gets enhanced by
the poles contained in $1/\sin^2(\pi z)$ for $\mbox{\rm Im}\,z=0$. Thus, in the general
terminology employed in the present work, the additional term \eqn{DVterm}
can be considered a duality-violating contribution.\footnote{The appearance
of a DV-like term in eq.~\eqn{refrel} is related to the phenomenon of
Stokes discontinuities \cite{din73,pk01}. In the case of the $\psi$- and
$\psi'$-functions the imaginary axis is a Stokes ray, beyond which, that is
in the imaginary left-half plane, the formally exponentially suppressed
terms combine to yield a potentially non-negligible contribution.}
In the following numerical analysis, the FESR \eqn{coulsr} shall be
investigated in a way such that the asymptotic expansion can be performed in
a most transparent fashion. A convenient choice of the complex integration
contour to this end are paths with constant $|1-\lambda|=|z|\equiv z_0$.
The corresponding energy $E_z(\varphi)$ is then given by
\begin{equation}
\label{Ezphi}
E_z(\varphi) \,=\, -\,\frac{\alpha^2 m}{4(z_0{\rm e}^{i\varphi}-1)^2} \,,
\end{equation}
with the parametrisation $z=z_0\,{\rm e}^{i\varphi}$. Because of the square
in the denominator, in the full range of $0<\varphi<2\pi$, the contour
$E_z(\varphi)$ covers the complex $E$-plane twice. However, only the range
$\varphi_0<\varphi<2\pi-\varphi_0$, where the angle $\varphi_0=\arccos(1/z_0)$, and
which contains the section on which the DV term \eqn{DVterm} should be included
in the asymptotic expansion of $\psi'(z)$, provides the correct solution to
the corresponding transcendental equation. Of particular interest are the
two points
\begin{eqnarray}
\label{Ezzero}
E_0 \,\equiv\, E_z(\varphi_0) &\,=\,&
\phantom{-}\,\frac{\alpha^2 m}{4(z_0^2-1)} \,=\, \frac{-\,E_1}{(z_0^2-1)} \,, \\
\vbox{\vskip 8mm}
\label{Ezpi}
E_\pi \,\equiv\, E_z(\pi) &\,=\,&
-\,\frac{\alpha^2 m}{4(z_0+1)^2} \,=\, \frac{E_1}{(z_0+1)^2} \,,
\end{eqnarray}
where $E_z(\varphi)$ is real, and which have also been expressed in terms of
the lowest-lying bound-state energy $E_1$. Eq.~\eqn{Ezpi} shows that
in order for the contour not to hit a bound-state pole, $z_0$ should not
be integer, and that depending on $z_0$, poles with $n\geq{\rm int}(z_0)+2$,
where ${\rm int}(z_0)$ denotes the integer part of $z_0$, need to be included
on the phenomenological side of the sum rule, as they lie inside of the
integration region.
\begin{figure}[!thb]
\begin{center}
\includegraphics[angle=0, width=14cm]{Ez}
\end{center}
\vspace{-3mm}
\caption{Integration contour $E_z(\varphi)$ in the complex $E$-plane for the
FESR \eqn{coulsr}, employing $z_0=1.5$ and the further parameters provided in
the text. The part of the contour where the DV term \eqn{DVterm} should be
included has been depicted as the solid line, and the remaining one, where
the asymptotic expansion \eqn{psipas} is sufficient, as the dashed line. Some
bound state poles as well as the continuum cut are also shown as thick dots
and line, respectively.\label{fig1}}
\end{figure}
As a particular example, the sum rule \eqn{coulsr} will now be investigated
numerically for the trivial weight function $w(E)=1$ and the set of parameters
$\alpha=1/2$ and $m=10$. This leads to the lowest ground state being at
$E_1=-5/8$. The integration contour in the $E$-plane is displayed in
figure~\ref{fig1} for $z_0=1.5$, together with some bound state poles as
well as the continuum cut (thick dots and line). The section on which
$\psi'(z)$ admits an asymptotic expansion without DV term is depicted as the
dashed line, while the section where the DV term \eqn{DVterm} should be
included, is displayed as the solid line.
Next, figure~\ref{fig2} shows the moment $R^w(E_0)$ of eq.~\eqn{coulsr} for
the weight function $w(E)=1$, as a function of $z_0$. As expected, at the
locations where $E_z(\pi)$ crosses a bound-state pole, $R^w(E_0)$ is
discontinuous. Employing the full Adler-type function $D(E)$ of
eq.~\eqn{MEcoul}, it is a simple matter to verify that the equality between
the {\em phenomenological} left-hand side of \eqn{coulsr} in terms of $\rho(E)$
and the {\em theoretical} right-hand side is satisfied. On the other hand, when
simulating the OPE with the asymptotic expansions \eqn{psipas} and \eqn{refrel}
it is found that the DV term \eqn{DVterm} completely dominates the moment. At
the lowest considered $z_0=1.2$, the terms polynomial in $1/z$ only contribute
about $0.3\,$\% and at the highest plotted $z_0=5.5$, this contribution is
within the numerical uncertainties of the asymptotic expansion. Hence, in
the respective case, the moment $R^w(E_0)$ is completely saturated by the DV
contribution.
\begin{figure}[!thb]
\begin{center}
\includegraphics[angle=0, width=14cm]{Rcoul}
\end{center}
\vspace{-3mm}
\caption{The solid line represents the moment $R^w(E_0)$ of eq.~\eqn{coulsr}
for the weight function $w(E)=1$, as a function of $z_0$. The discontinuities
appear at the locations where the contour crosses a bound-state pole. The DV
contribution \eqn{DVterm} completely saturates this moment and is visually
indistinguishable from the solid line.\label{fig2}}
\end{figure}
There are two reasons that contribute to this behaviour. On the one hand, the
Coulomb bound states are zero-width resonances and thus the duality violations
are expected to be very strong. In such a case the DVs can be considered
``maximal''. On the other hand, the weight function $w(E)=1$ does not provide
any suppression of the resonance region which could soften the strong duality
violations. The second issue can be investigated further by employing weight
functions which provide some suppression of the resonance region, for example
power-like moments $W^{(k)}(E)=(E-E_\pi)^k (E-E_0)$. Studying these weights
for small $k=1,\,2,\,3$, it is found that indeed at low $z_0$ (close to the
breakdown of the asymptotic expansion) the contribution of the DVs is
suppressed. However, the dominance of DVs again quickly sets in for larger
$z_0$, and even at low $z_0$ the contribution of DVs is still sizeable. The
origin of the latter observation can be traced back to the fact that the DV
term \eqn{DVterm} penetrates some distance into the complex plane, before the
exponential decay becomes effective. Thus, even with the weight functions
$W^{(k)}(E)$, which nullify the contribution on the real axis, the residual
contribution from the full contour integration can remain sizeable.
The dependence of the DVs on the resonance structure could be studied further
by providing the Coulomb bound states with a finite width and varying this
width. However, in view of practical applications of DV models in analyses of
hadronic $\tau$ decays, the following discussion will be continued on the basis
of a second model for DVs, already investigated in
refs.~\cite{shi00,cgp05,cgp08}.
\section{A model for light-quark correlators}
Following refs.~\cite{shi00,cgp05,cgp08}, the structure of DVs shall be
investigated in a second model, which incorporates constraints from Regge
theory \cite{reg59,col71} on the light meson spectrum, and which can be chosen
to roughly resemble the physical spectrum of the light-quark vector current
correlator. The model employed below is a simplified version of the one
studied in refs.~\cite{cgp05,cgp08}, serving all purposes of the analysis
discussed in the following. Specifically, it is chosen to take the form
\begin{equation}
\label{PiV}
\Pi_V(s) \,=\, -\; \psi\Biggl(\frac{M_V^2+u(s)}{\Lambda^2}\Biggr) + {\rm const.} \,,
\end{equation}
where
\begin{equation}
\label{zzeta}
u(s) \,=\, \Lambda^2\biggl(\frac{-s}{\Lambda^2}\biggr)^{\!\zeta}
\qquad \mbox{and} \qquad \zeta \,=\, 1 - \frac{a}{\pi N_c} \,.
\end{equation}
As compared to refs.~\cite{cgp05,cgp08}, the lowest lying vector, that is
the would-be $\rho$ meson, is included in the $\psi$-function. The reason
for this will be explained below. Furthermore, in this work the global
normalisation is immaterial and has been dropped. For all remaining parameters,
in the numerical analysis numbers similar to the ones given in ref.~\cite{cgp08}
will be employed:
\begin{equation}
\label{nums}
M_V \,=\, 770\,\mbox{\rm MeV} \,, \qquad
\Lambda \,=\, 1.2\,\mbox{\rm GeV} \,, \qquad
a \,=\, 0.4 \,.
\end{equation}
In order to arrive at an OPE for the model, we require the expansion of the
digamma function at large $|s|$, or correspondingly, large $|u|$. The
expansion in question, which is however only asymptotic, takes the form
\cite{as72}
\begin{equation}
\label{psias}
\psi(z) \,\sim\, \ln z - \frac{1}{2z} - \sum\limits_{n=1}^{\infty}
\frac{B_{2n}}{2n\,z^{2n}} \qquad (z\to\infty\;{\rm in}\;|\arg z|<\pi) \,.
\end{equation}
In fact, the asymptotic expansion of $\psi'(z)$ of eq.~\eqn{psipas} is of
course just equal to the derivative of eq.~\eqn{psias}. Though in principle
one would have to further expand eq.~\eqn{psias} in terms of powers of $1/s$,
in order to arrive at an OPE-like expansion, this is not necessary for what
shall be discussed in the following, and thus we stick to the asymptotic
expansion in powers of $1/z$.
For finite-width resonances, the poles are on an unphysical sheet, and
$|\arg z|$ never reaches $\pi$. Still, like in the case of $\psi'(z)$,
in the region $\mbox{\rm Re}\,z<0$ the asymptotic expansion can be substantially
improved by making use of the reflection relation \cite{cgp08,as72}
\begin{eqnarray}
\label{refrel2}
\psi(z) \,&=&\, \psi(-z) - \frac{1}{z} - \pi\,\cot(\pi z) \\
\vbox{\vskip 8mm}
\,&\stackrel{z\to\infty}{\sim}&\, \ln z - \frac{1}{2z} -
\sum\limits_{n=1}^{\infty} \frac{B_{2n}}{2n\,z^{2n}}
- \pi\,[\,\cot(\pi z)\pm i\,] \quad
(\,\mbox{\rm Re}\,z < 0,\, \mbox{\rm Im}\,z^{\;>}_{\;<}\;0 \,)
\,. \nonumber
\end{eqnarray}
Again, the logarithm and rational parts of both asymptotic expansions
\eqn{psias} and \eqn{refrel2} agree, and it is a simple matter to convince
oneself that away from the real axis the additional term
$\pi\,[\,\cot(\pi z)\pm i\,]$ is exponentially suppressed, but takes once
more care of the nearby poles.
\begin{figure}[!htb]
\begin{center}
\includegraphics[angle=0, width=14cm]{psi}
\end{center}
\vspace{-3mm}
\caption{$\mbox{\rm Re}[\psi(z)]$ (left pane) and $\mbox{\rm Im}[\psi(z)]$ (right pane), for
$z=1.5\!\cdot\!\exp(i\varphi)$, as functions of $\varphi=\arg(z)$. The solid
lines correspond to the full result, while the dashed lines show the
asymptotic expansion of eq.~\eqn{psias}, including terms up to $n=5$.
\label{fig2b}}
\end{figure}
This is also demonstrated quantitatively in figure~\ref{fig2b}, where the real
(left pane) and imaginary (right pane) parts of $\psi(z)$ are displayed for
$z=1.5\!\cdot\!\exp(i\varphi)$ as functions of $\varphi=\arg(z)$. The solid lines
correspond to the full function, while the dashed lines show the asymptotic
expansion of eq.~\eqn{psias}, including terms up to $n=5$. For positive real
$z$ at this order the asymptotic series assumes its minimal term. As is
evident, the asymptotic expansion breaks down for $\varphi~{}_{\textstyle\sim}^{\textstyle >}~ 2.5$. The
difference between the total result and the asymptotic expansion is to a very
good approximation covered by the additional term in eq.~\eqn{refrel2}, such
that taking this contribution into account, the quality of the expansion is
analogous to the one for $\varphi<\pi/2$.
To continue, two finite-energy sum rules for the model \eqn{PiV} shall be
investigated. For an arbitrary, analytic weight function $w(s)$, they take
the generic form
\begin{equation}
\label{fesr}
R^w(s_0) \,\equiv\, \int\limits_0^{s_0} \frac{{\rm d}s}{s_0}\,w(s)\,\rho_V(s)
\,=\,
\frac{i}{2\pi}\!\oint\limits_{s_0} \frac{{\rm d}s}{s_0}\,w(s)\,\Pi_V(s) \,,
\end{equation}
where $\rho_V(s)\equiv\mbox{\rm Im}\,\Pi_V(s+i0)/\pi$, and the contour on the right-hand
side is chosen such that it starts and ends at a real $s_0>0$.
\begin{figure}[!htb]
\begin{center}
\includegraphics[angle=0, width=14cm]{sz}
\end{center}
\vspace{-3mm}
\caption{Integration contour $s_z(\varphi)$ of eq.~\eqn{szphi} in the complex
$s$-plane for the FESR \eqn{fesr}, employing $z_0=1.5$ and the further
parameters provided in the text. The part of the contour where the DV term
should be included has been depicted as the solid line, and the remaining one
as the dashed line. The cut of $\Pi_V(s)$ on the positive real $s$-axis is
also indicated as the thick line.\label{fig3}}
\end{figure}
The asymptotic expansion is again studied most cleanly for contours with
constant absolute value of the argument of the $\psi$-function, $|z|=z_0$.
The corresponding energy variable $s_z(\varphi)$ is then given by
\begin{equation}
\label{szphi}
s_z(\varphi) \,=\, -\,\Lambda^2\biggl( z_0\,{\rm e}^{i\varphi} -
\frac{M_V^2}{\Lambda^2} \,\biggr)^{\!1/\zeta} \,,
\end{equation}
and is displayed in figure~\ref{fig3} for the choice $z_0=1.5$. Again, the
portion of the contour on which the DV term should be included is depicted
as the solid line, while the remainder, on which the expansion \eqn{psias}
is admissible, is plotted as the dashed line. In order that the contour
closes below the cut of $\Pi_V(s)$, that is $s_z(0)<0$, one requires
$z_0>M_V^2/\Lambda^2\approx 0.41$. On the other hand this entails that
$s_0~{}_{\textstyle\sim}^{\textstyle >}~ 1.2\,\mbox{\rm GeV}^2$.\footnote{This is the reason why the $\rho$-meson
has been included into the $\psi$-function. In the original model of
ref.~\cite{cgp08}, the corresponding requirement would have been
$z_0>M_V^2/\Lambda^2\approx 1.6$, which would have necessitated values for $s_0$
at least as large as $s_0~{}_{\textstyle\sim}^{\textstyle >}~ 4.5\,\mbox{\rm GeV}^2$.} Because we have a model with
finite-width resonances and the poles are located on an unphysical sheet,
the angle $\varphi_0$ with $s_z(\varphi_0)=s_0$ is found smaller than $\pi$. For
example in the case of figure~\ref{fig3}, in which $z_0=1.5$, $\varphi_0=2.972$.
In the limit of zero-width resonances, corresponding to the large-$N_c$ limit,
and which in the model \eqn{PiV} is realised as $\zeta\to 1$, $\varphi_0$ would
go to $\pi$ and the poles would lie on the positive real $s$-axis. On the
contrary, for broader resonances which lie further away from the real $s$-axis,
the angle $\varphi_0$ is smaller, and the contribution of DVs gets reduced.
\begin{figure}[!htb]
\begin{center}
\includegraphics[angle=0, width=14cm]{R_w=1}
\end{center}
\vspace{-3mm}
\caption{The solid line shows the moment $R^w(s_0)$ of eq.~\eqn{fesr} for the
weight $w(s)=1$ and plotted as a function of $s_0$. The long- and short-dashed
lines correspond to the power-like asymptotic expansion up to fourth and sixth
order respectively.\label{fig4}}
\end{figure}
The first example of a moment $R^w(s_0)$ corresponding to eq.~\eqn{fesr}
is displayed in figure~\ref{fig4} for the trivial weight $w(s)=1$ and as a
function of $s_0$. The oscillatory behaviour of the moment results from the
presence of resonances, being damped for higher energies. As the long- and
short-dashed lines, the asymptotic expansion \eqn{psias} is plotted up to the
fourth and sixth order in $z$ respectively. As is apparent, the asymptotic
expansion breaks down at about $s_0\approx 1.5\,\mbox{\rm GeV}^2$. Furthermore, large
deviations of the asymptotic expansion to the exact moment are observed, which
are, however, perfectly described by the additional term in eq.~\eqn{refrel2}.
The largest deviation is found at the first minimum around
$s_0\approx 1.76\,\mbox{\rm GeV}^2$, where the DVs amount to about $-11\%$ of the
total contribution.
\begin{figure}[!htb]
\begin{center}
\includegraphics[angle=0, width=14cm]{R_w=1mx2}
\end{center}
\vspace{-3mm}
\caption{The solid line shows the moment $R^w(s_0)$ of eq.~\eqn{fesr} for the
weight $w(s)=(1-s/s_0)^2$ and plotted as a function of $s_0$. The long- and
short-dashed lines correspond to the power-like asymptotic expansion up to
fourth and sixth order respectively.\label{fig5}}
\end{figure}
To investigate the influence of weight functions which provide some pinch
suppression at $s=s_0$, in figure~\ref{fig5} the moment $R^w(s_0)$ is displayed
for the weight $w(s)=(1-s/s_0)^2$, which has a double zero at $s=s_0$, just
like the kinematical weight function in hadronic $\tau$ decays. Obviously,
now the influence of DVs is much smaller though still clearly visible. At
$s_0\approx 1.76\,\mbox{\rm GeV}^2$, the contribution of DVs amounts to only $1.6\%$
of the total moment, and a much stronger suppression is observed towards
higher energies.
\section{Conclusions}
In phenomenological analyses of QCD, the operator product expansion is a
widely used tool. While at sufficiently high energies the legitimacy of its
application is unquestionable in the Euclidian domain, the continuation of
the OPE towards the physical, Minkowskian region is inflicted with the
appearance of duality violations. The presence of DV contributions, besides
the OPE, is due to the physical bound states or resonances, whose physics
cannot be captured by the OPE alone.
In the present article, the structure and impact of DVs has been studied on
the basis of two models. Firstly, the quantum mechanical Coulomb system was
investigated, which can serve as a model for heavy quarkonia, and which has
not been considered in the context of finite-energy sum rules and DVs before.
Secondly, a simplified version of a model already discussed extensively in
refs.~\cite{shi00,cgp05,cgp08} was used for comparison with the features
found in the Coulomb model. The latter model incorporates constraints from
large-$N_c$ QCD, as well as Regge theory, and captures the central features
of the light-quark vector or axialvector correlators.
Qualitatively, the physical picture behind the appearance of duality violations
can be summarised as follows: in the Euclidean region, all effects of
resonances or bound-state poles have been sufficiently smoothed out,
quark-hadron duality is expected to work well, and the OPE provides a
satisfactory description of current correlation functions. When analytically
continuing through the complex energy plane towards the physical, Minkowskian
region, the presence of resonance poles becomes more and more prominent,
reflecting itself in a breakdown of the OPE. Mathematically, the DV
contributions show up as formally exponentially suppressed terms in an
asymptotic expansion, which are, however, enhanced by the nearby poles, and
thus can become numerically relevant.
The strength of effects resulting from DVs can be considered ``maximal'' if
the bound states have zero width and the poles lie on the Minkowskian axis.
This case is for example realised in the Coulomb model or when $N_c$ in
eq.~\eqn{zzeta} of the second model is taken to infinity \cite{cgp05}. On the
other hand, the influence of DVs is suppressed for resonances with a finite
width and becomes smaller when the resonances are broader and the poles lie
further away from the physical, Minkowskian axis. This is also seen from the
oscillations in the physical spectrum whose features cannot be described
adequately by the OPE. These oscillations are more prominent for narrower
resonances while they become weaker if the resonances get broader.
If the physical resonances are sufficiently broad, and thus lie sufficiently
far away from the physical axis, a simplified description of DVs as an
exponentially damped oscillatory function appears admissible. Assuming $s$ to
be large enough for the OPE to make sense, motivated by the $\psi$-function
model such a simplified Ansatz was advocated in refs.~\cite{cgp08,cgp09} for
the duality violating spectral function $\rho_{\rm DV}(s)$ in the form
\begin{equation}
\label{rhoDV}
\rho_{\rm DV}(s) \,=\, \kappa\,{\rm e}^{-\gamma s} \sin(\alpha+\beta s) \,,
\end{equation}
where $\kappa$, $\gamma$, $\alpha$ and $\beta$ are a priori free parameters
which in a specific channel have to be extracted from experiment, and which
bear no immediate QCD interpretation. The DV contribution to a moment is then
assumed to be given by
\begin{equation}
\label{RwDV}
R^w_{\rm DV}(s_0) \,=\, -\!\int\limits_{s_0}^\infty \frac{{\rm d}s}{s_0}
\,w(s)\,\rho_{\rm DV}(s) \,.
\end{equation}
In practical applications of QCD finite-energy sum rules, however, the DVs
naturally are required on the complex contour and thus a closely related
Ansatz, which corresponds to the continuation of eq.~\eqn{rhoDV} to the right
complex half-plane, may also be considered:
\begin{equation}
\label{PiDV}
\Pi_{\rm DV}(s) \,=\, \pi\,\kappa\,{\rm e}^{-\,\gamma s\,\pm\,i\,
(\alpha+\beta s)} \qquad
(\,\mbox{\rm Re}\,s > 0,\,\mbox{\rm Im}\,s^{\;>}_{\;<}\;0 \,) \,,
\end{equation}
with $\gamma$ and $\beta$ positive. Of course, the $\mbox{\rm Im}\,\Pi_{\rm DV}(s)$ on
the positive, real $s$-axis reproduces the spectrum of eq.~\eqn{rhoDV}, but
the Ansatz also reflects the form of the exponential suppression of the DV term
given in eq.~\eqn{DVterm}. Both Ans\"atze \eqn{rhoDV} and \eqn{PiDV} would
be totally equivalent, if the DV contribution would be active on all of the
complex contour. However, since the DV term should only be included on part of
the contour (the solid-line sections in figures \ref{fig1} and \ref{fig3}, or,
more generically, on the right-half $s$-plane $\mbox{\rm Re}\,s>0$), both choices differ
by an exponentially suppressed contribution. As a matter of principle this
should be numerically irrelevant, but it could play a certain role if in
practical applications the exponential suppression is not very pronounced.
This will be studied in the future in investigations like the ones of
refs.~\cite{cgp09,bcgjmop10}.
To conclude, in the presented article observations about duality violations
appearing in asymptotic expansions were summarised on the basis of two models.
One of them has already been discussed in the literature before, while the
other, the Coulomb system, has not been studied with respect to DVs previously.
The qualitative findings of this work should prove useful for phenomenological
QCD studies in the future.
\bigskip
\acknowledgments
I would like to thank Martin~Beneke, Diogo~Boito, Maarten Golterman,
Santi~Peris and Antonio~Pineda for interesting discussions. This work has been
supported in parts by the EU Contract No.~MRTN-CT-2006-035482 (FLAVIAnet),
by CICYT-FEDER FPA2008-01430, as well as by Spanish Consolider-Ingenio 2010
Programme CPAN (CSD2007-00042).
\bigskip
\providecommand{\href}[2]{#2}\begingroup\raggedright |
\section{Introduction and Literature Review}
Modeling communication channels with a state process fits well for many physical scenarios. For single-user channels, the characterization of the capacity with various degrees of channel state information at the transmitter (CSIT) and at the receiver (CSIR) is well understood. Among them, Shannon \cite{Sha_Side} determined the capacity formula when causal noiseless state information is available at the transmitter, where state is independent identically distributed (i.i.d.). The same problem with non-causal side information is considered in \cite{Gelfand}. In \cite{Salehi}, Shannon's result is extended to the case where noisy state observation is available at both the transmitter and the receiver. Later, in \cite{Caire} this result has been shown to be a special case of Shannon's model and the authors also determined that when CSIT is a deterministic function of CSIR optimal codes can be constructed directly on the input alphabet.
In the multi-user setting, \cite{Das} provides a multi-letter characterization of the capacity region of time-varying multiple access channels (MACs) with various degrees of CSIT and CSIR. In \cite{Jafar}, a general framework for the capacity region of MACs with causal and non-causal CSI is presented. In a related work, MACs where the encoders have degraded information on the channel state, which is coded to the encoders, is considered \cite{Cemal}. In \cite{lap-ste-1}, memoryless FS-MACs with two independent states (see also \cite{lap-ste-2} for the single state case), each known causally and strictly causally to one encoder, are considered and an achievable rate region, which is shown to contain an achievable region where each user applies Shannon strategies, is proposed. In \cite{lap-ste-1} and \cite{lap-ste-2} it is also shown that strictly casual state information does not increase the sum-rate capacity. More recently, in \cite{basher} finite-state Markovian MACs with asymmetric delayed state information at the transmitters are studied and their capacity region is determined.
The most relevant work to our paper is \cite{GiacomoSerdar}, which obtained a single letter characterization of the capacity region for memoryless FS-MAC in which transmitters have asymmetric partial quantized state observations and the receiver has full state information. In this work, the authors were inspired from team decision theory \cite{Yuksel}, \cite{Wits}. We herein mainly adopt the converse technique presented in \cite{GiacomoSerdar} and partially extend it to a noisy setup. The present paper, thus, studies the FS-MAC in which each of the transmitters have an asymmetric state information which is corrupted by an i.i.d. noise process and the receiver has complete state information. We provide a single letter inner bound to the capacity region, in terms of Shannon strategies \cite{Sha_Side}. By observing that this inner bound is tight for the sum-rate capacity, we also provide an outer bound to the channel's capacity region. We modify the approach in \cite{GiacomoSerdar} to account for the fact that the decoder does not have access to the state information at the encoders, and that the past state information does not lead to a tractable recursion.
The rest of the paper is organized as follows. In Section \ref{full} we formally state the problem, present inner and outer bounds to the capacity region with the achievability and converse proofs and in Section \ref{conc} we present concluding remarks.
Throughout the paper we will use the following notations. A random variable will be denoted by an upper case letter $X$ and its particular realization by a lower case letter $x$. For a vector $v$, and a positive integer $i$, $v_{i}$ will denote the $i$-th entry of $v$, while $v_{[i]} = (v_1, \cdots, v_i)$ will denote the vector of the first $i$ entries of $v$. For a finite set $\mathcal{A}$, $\mathcal{P}(\mathcal{A})$ will denote the simplex of probability distributions over $\mathcal{A}$. Probability distributions are denoted by $P(\cdot)$ and subscripted by the name of the random variables and conditioning, e.g., $P_{U,T|V,S}(u,t|v,s)$ is the conditional probability of $(U=u,T=t)$ given $(V=v,S=s)$. Finally, for a positive integer $n$, we shall denote by $\mathcal{A}^{(n)} := \bigcup_{0<s<n} \mathcal{A}^{s}$ the set of $\mathcal{A}$-strings of length smaller than $n$. We denote the indicator function of an event by $1_{\{E\}}$. All sets considered hereafter are finite.
\section{On the Capacity of FS-MAC with Noisy CSIT and Complete CSIR}\label{full}
Consider a two-user memoryless FS-MAC, with two encoders, $a, b$, and two independent message sources $W_a$ and $W_b$ which are uniformly distributed in the sets $W_a \in \{1,2,\cdots,M_a\}$ and $W_b \in \{1,2,\cdots,M_b\}$, respectively. The channel inputs of the encoders are $X^{a}$ and $X^{b}$, respectively. The channel state process is modeled as a sequence $\{S_t\}_{t=1}^{\infty}$ of i.i.d. random variables in some space $\cal S$. The two encoders have access to causal noisy version of the state information $S_t$ at each time $t\geq1$ modeled by $S_{t}^{a} \in {\cal S}^{a}$, $S_{t}^{b} \in {\cal S}^{b}$, respectively and as such the joint distribution of $(S_t,S_t^a,S_t^b)$ satisfies
\begin{eqnarray}
P_{S_{t}^{a}, S_{t}^{b},S_t}(s_{t}^{a}, s_{t}^{b},s_t)=P_{S_{t}^{a}|S_t}(s_{t}^{a}|s_t)P_{S_{t}^{b}|S_t}(s_{t}^{b}|s_t)P_{S_t}(s_t)\label{eq-sta-no}.
\end{eqnarray}
We also assume that $S_t$ is fully available at the receiver (see \ref{fig:macfb}) and that $(S_t,S_t^a,S_t^b)$ are independent of $(W_a,W_b)$ $\forall t\geq 1$. The channel inputs at time $t$, i.e., $X_t^{a}$ and $X_t^{b}$, are functions of the locally available information $(W_a,S_{[t]}^{a})$ and $(W_b,S_{[t]}^{b})$. Let $\mathbf{W}:=(W_a,W_b)$ and $\mathbf{X}:=(X^a,X^b)$. Then, the laws governing $n$-sequences of state, input and output letters are given by
\begin{eqnarray}
&&\hspace{-0.6in}P_{Y_{[n]}|\mathbf{W},\mathbf{X}_{[n]},S_{[n]},S_{[n]}^a,S_{[n]}^b}(y_{[n]}|\mathbf{w},\mathbf{x}_{[n]},s_{[n]},s_{[n]}^a,s_{[n]}^b)\nonumber\\
&&\quad\quad\quad=\prod _{t=1}^nP_{Y_t|X_{t}^{a}, X_{t}^{b}, S_t}(y_t|x_{t}^{a}, x_{t}^{b}, s_t), \label{eq-ch}
\end{eqnarray}
where the channel's transition distribution, $P_{Y_t|X_{t}^{a}, X_{t}^{b}, S_t}(y_t|x_{t}^{a}, x_{t}^{b}, s_t)$, is given a priori.
\begin{definition}\label{def:maccode}
An $(n, 2^{nR_a}, 2^{nR_b})$ code with block length $n$ and rates $(R_a, R_b)$ for an FS-MAC with noisy state feedback consists of
\begin{itemize}
\item [(1)] A sequence of mappings for each encoder
\begin{center}
$\phi_{t}^{(a)}: (\mathcal{S}^{a})^t \times \mathcal{W}_a \rightarrow {\cal X}_a, \enspace t=1,2,...n$;\\
\end{center}
\begin{center}
$\phi_{t}^{(b)}: (\mathcal{S}^{b})^t \times \mathcal{W}_b \rightarrow {\cal X}_b, \enspace t=1,2,...n$.\\
\end{center}
\item [2)] An associated decoding function
\begin{center}
$\psi: (\mathcal{S})^n\times{\cal Y}^n\rightarrow \mathcal{W}_a \times \mathcal{W}_b$.\\
\end{center}
\end{itemize}
\end{definition}
\begin{figure}
\hspace{-0.1in}
\setlength{\unitlength}{3947sp
\begingroup\makeatletter\ifx\SetFigFont\undefine
\gdef\SetFigFont#1#2#3#4#5
\reset@font\fontsize{#1}{#2pt
\fontfamily{#3}\fontseries{#4}\fontshape{#5
\selectfont
\fi\endgrou
\begin{picture}(4299,2635)(1281,-3202)
\thinlines
{\color[rgb]{0,0,0}\put(1346,-1379){\vector( 1, 0){300}}
{\color[rgb]{0,0,0}\put(1658,-1725){\framebox(846,709){}}
{\color[rgb]{0,0,0}\put(1656,-2709){\framebox(846,709){}}
{\color[rgb]{0,0,0}\put(1346,-2327){\vector( 1, 0){300}}
{\color[rgb]{0,0,0}\put(2510,-2328){\line( 1, 0){305}}
\put(2813,-2328){\line( 0, 1){344}}
\put(2813,-1984){\vector( 1, 0){162}}
{\color[rgb]{0,0,0}\put(2991,-2224){\framebox(1036,706){}}
{\color[rgb]{0,0,0}\put(2510,-1325){\line( 1, 0){308}}
\put(2817,-1325){\line( 0,-1){403}}
\put(2817,-1728){\vector( 1, 0){173}}
{\color[rgb]{0,0,0}\put(2075,-579){\vector( 0,-1){431}}
\put(2075,-579){\line( 1, 0){1379}}
\put(3454,-579){\line( 0,-1){931}}
{\color[rgb]{0,0,0}\put(2040,-3190){\vector( 0, 1){479}}
\put(2040,-3190){\line( 1, 0){1415}}
\put(3455,-3190){\line( 0, 1){970}}
{\color[rgb]{0,0,0}\put(4034,-1713){\vector( 1, 0){343}}
{\color[rgb]{0,0,0}\put(4034,-2026){\vector( 1, 0){343}}
{\color[rgb]{0,0,0}\put(4384,-2234){\framebox(846,709){}}
{\color[rgb]{0,0,0}\put(5237,-1696){\vector( 1, 0){300}}
{\color[rgb]{0,0,0}\put(5237,-2079){\vector( 1, 0){300}}
\put(1460,-1327){\makebox(0,0)[b]{\smash{{\SetFigFont{9}{10.8}{\rmdefault}{\mddefault}{\updefault}{\color[rgb]{0,0,0}$W_a$
}}}}
\put(2090,-1534){\makebox(0,0)[b]{\smash{{\SetFigFont{9}{10.8}{\rmdefault}{\mddefault}{\updefault}{\color[rgb]{0,0,0}$\phi_{t}^a(W_a,S_t^a)$
}}}}
\put(2088,-1292){\makebox(0,0)[b]{\smash{{\SetFigFont{9}{10.8}{\rmdefault}{\mddefault}{\updefault}{\color[rgb]{0,0,0}Encoder
}}}}
\put(1460,-2264){\makebox(0,0)[b]{\smash{{\SetFigFont{9}{10.8}{\rmdefault}{\mddefault}{\updefault}{\color[rgb]{0,0,0}$W_b$
}}}}
\put(2088,-2306){\makebox(0,0)[b]{\smash{{\SetFigFont{9}{10.8}{\rmdefault}{\mddefault}{\updefault}{\color[rgb]{0,0,0}Encoder
}}}}
\put(2090,-2521){\makebox(0,0)[b]{\smash{{\SetFigFont{9}{10.8}{\rmdefault}{\mddefault}{\updefault}{\color[rgb]{0,0,0}$\phi_{t}^b(W_b,S_t^b)$
}}}}
\put(3510,-2043){\makebox(0,0)[b]{\smash{{\SetFigFont{9}{10.8}{\rmdefault}{\mddefault}{\updefault}{\color[rgb]{0,0,0}$P(Y_t|\mathbf{X}_t,S_t)$
}}}}
\put(3502,-1796){\makebox(0,0)[b]{\smash{{\SetFigFont{9}{10.8}{\rmdefault}{\mddefault}{\updefault}{\color[rgb]{0,0,0}Channel
}}}}
\put(2682,-1250){\makebox(0,0)[b]{\smash{{\SetFigFont{8}{9.6}{\rmdefault}{\mddefault}{\updefault}{\color[rgb]{0,0,0}$X_t^a$
}}}}
\put(2682,-2476){\makebox(0,0)[b]{\smash{{\SetFigFont{9}{10.8}{\rmdefault}{\mddefault}{\updefault}{\color[rgb]{0,0,0}$X_t^b$
}}}}
\put(2164,-2969){\makebox(0,0)[b]{\smash{{\SetFigFont{9}{10.8}{\rmdefault}{\mddefault}{\updefault}{\color[rgb]{0,0,0}$S_{[t]}^b$
}}}}
\put(2222,-875){\makebox(0,0)[b]{\smash{{\SetFigFont{9}{10.8}{\rmdefault}{\mddefault}{\updefault}{\color[rgb]{0,0,0}$S_{[t]}^a$
}}}}
\put(4210,-1674){\makebox(0,0)[b]{\smash{{\SetFigFont{9}{10.8}{\rmdefault}{\mddefault}{\updefault}{\color[rgb]{0,0,0}$Y_t$
}}}}
\put(4210,-1945){\makebox(0,0)[b]{\smash{{\SetFigFont{9}{10.8}{\rmdefault}{\mddefault}{\updefault}{\color[rgb]{0,0,0}$S_{[t]}$
}}}}
\put(4764,-1797){\makebox(0,0)[b]{\smash{{\SetFigFont{9}{10.8}{\rmdefault}{\mddefault}{\updefault}{\color[rgb]{0,0,0}Decoder
}}}}
\put(4801,-2063){\makebox(0,0)[b]{\smash{{\SetFigFont{9}{10.8}{\rmdefault}{\mddefault}{\updefault}{\color[rgb]{0,0,0}$\psi(Y_{[n]},S_{[n]})$
}}}}
\put(5363,-1662){\makebox(0,0)[b]{\smash{{\SetFigFont{9}{10.8}{\rmdefault}{\mddefault}{\updefault}{\color[rgb]{0,0,0}$\hat{W}_a$
}}}}
\put(5363,-2054){\makebox(0,0)[b]{\smash{{\SetFigFont{9}{10.8}{\rmdefault}{\mddefault}{\updefault}{\color[rgb]{0,0,0}$\hat{W}_b$
}}}}
\end{picture
\caption{The multiple-access channel with noisy state feedback.}
\label{fig:macfb}
\end{figure}
The system's probability of error, $P_{e}^{(n)}$, is given by
\begin{eqnarray}
\frac{1}{2^{n(R_a+R_b)}}\sum_{w_a=1}^{2^{nR_a}}\sum_{w_b=1}^{2^{nR_b}}P\left(\psi(Y_{[n]},S_{[n]})\neq (w_a,w_b)| \mathbf{W}=\mathbf{w}\right).\nonumber
\end{eqnarray}
A rate pair $(R_a, R_b)$ is achievable if for any $\epsilon > 0$ there exists, for all $n$ sufficiently large, an $(n, 2^{nR_a}, 2^{nR_b})$ code such that $\frac{1}{N}\log M_a \geq R_a \geq 0$, $\frac{1}{N}\log M_b \geq R_b \geq 0$ and $P_{e}^{(n)} \leq \epsilon$. The capacity region of the FS-MAC, ${\cal C}_{FS}$, is the closure of the set of all achievable rate pairs $(R_a, R_b)$ and the sum-rate capacity is defined as ${\cal C}^{FS}_{\sum}:=\max_{(R_a,R_b) \in {\cal C}_{FS}}(R_a+R_b)$.
Before proceeding with the main result, we introduce \textit{memoryless stationary team policies} \cite{GiacomoSerdar} and their associated rate regions. We first define Shannon strategies.
\begin{definition}
Let the set of all possible functions from ${\cal S}^{a}$ to ${\cal X}^{a}$ and ${\cal S}^{b}$ to ${\cal X}^{b}$ be denoted by ${\cal T}^{a}$ and ${\cal T}^{b}$, respectively, where ${\cal T}^{a}={{\cal X}^{a}}^{|{\cal S}^a|}$ and ${\cal T}^{b}={{\cal X}^{b}}^{|{\cal S}^b|}$. Let $T^{a} \in {\cal T}^{a}$ and $T^b \in {\cal T}^{b}$ be two ${\cal T}^a$-valued and ${\cal T}^b$-valued random vectors, respectively, referred to as Shannon strategies.
\end{definition}
\begin{definition}\cite{GiacomoSerdar}\label{def:teampol}
A memoryless stationary (in time) team policy is a family
\begin{eqnarray}
\Pi=\left\{\pi=\left(\pi_{T^a}(\cdot),\pi_{T^b}(\cdot)\right)\in {\cal P}({\cal T}^a)\times{\cal P}({\cal T}^b)\right\}\label{eq-team-pol}
\end{eqnarray}
of probability distributions on the two sets of random functions. For every memoryless stationary team policy $\pi$, $\mathcal{R}(\pi)$ denotes the region of all rate pairs $R=(R_a,R_b)$ satisfying
\begin{eqnarray}
R_a &<& I(T^{a};Y|T^{b},S) \quad \label{eq-ra1-f}\\
R_b &<& I(T^{b};Y|T^{a},S) \quad \label{eq-ra2-f}\\
R_a+R_b &<& I(T^a,T^{b};Y|S) \label{eq-ra3-f}
\end{eqnarray}
where $S$, $T^a$, $T^b$ and $Y$ are random variables taking values in ${\cal S}$, ${\cal T}^{a}$, ${\cal T}^{b}$ and $\cal Y$, respectively and whose joint probability distribution factorizes as
\begin{eqnarray}
&&\hspace{-0.5in}P_{S,T^a,T^b,Y}(s,t^a,t^b,y) \nonumber\\
&&=P_S(s)P_{Y|T^a,T^b,S}(y|t^a,t^b,s)\pi_{T^a}(t^a)\pi_{T^b}(t^b)\label{eq-joidist-f}.
\end{eqnarray}
\end{definition}
We can now state the inner bound to the capacity region. Let ${\cal C}_{IN}:=\overline{co}\bigg(\bigcup_{\pi}\mathcal{R}(\pi)\bigg)$ denotes the closure of the convex hull of the rate regions $\mathcal{R}(\pi)$ given by (\ref{eq-ra1-f})-(\ref{eq-ra3-f}) associated to all possible memoryless stationary team polices as defined in (\ref{eq-team-pol}).
\begin{theorem}[Inner Bound]\label{the-main-achi-f}
${\cal C}_{IN}\subseteq {\cal C}_{FS}$.
\end{theorem}
The achievability proof follows the standard arguments of joint $\epsilon$-typical $n$-sequences \cite[Section 15.2]{cover}.
\begin{definition}\cite[Section 15.2]{cover}\label{def:jnt-typ}
The set ${\cal A}_{\epsilon}^{n}$ of $\epsilon$-typical $n$-sequences $\{(x^1_{[n]},\cdots,x^k_{[n]})\}$ with respect to the distribution $P_{X^1,\cdots,X^k}(x^1, \cdots, x^k)$ is defined by
\begin{eqnarray}
&&\hspace{-0.3in}{\cal A}_{\epsilon}^{n}=\left\{(x^1_{[n]},\cdots,x^k_{[n]})\in {\cal X}^1\times \cdots {\cal X}^k:\right.\nonumber\\
&&\hspace{-0.1in}\left.|-\frac{1}{n}\log\left(P(\textbf{s})\right)-H(S)|<\epsilon, \forall S \subseteq \{X^1,\cdots,X^k\}\right\} \nonumber
\end{eqnarray}
where $\textbf{s}$ denotes ordered set of sequences in $x^1_{[n]},\cdots,x^k_{[n]}$ corresponding to $S$.
\end{definition}
\begin{proof}[Proof of Theorem \ref{the-main-achi-f}]
Fix $(R_a,R_b)\in \mathcal{R}(\pi)$.
\textbf{\textit{Codebook Generation}}
Fix $\pi_{T^a}(t^a)$ and $\pi_{T^b}(t^b)$. For each $w_a \in \{1,\cdots, 2^{nR_a}\}$ randomly generate its corresponding $n$-tuple $t_{[n],w_a}^{a}$, each according to $\prod_{i=1}^n\pi_{T_{i}^{a}}(t_{i,w_a}^{a})$. Similarly, For each $w_b \in \{1,\cdots, 2^{nR_b}\}$ randomly generate its corresponding $n$-tuple $t_{[n],w_b}^{b}$, each according to $\prod_{i=1}^n\pi_{T_{i}^{b}}(t_{i,w_b}^{b})$. These codeword pairs form the codebook, which is revealed to the decoder while codewords $t_{i,w_l}^{l}$ is revealed to encoder $l$, $l=\{a,b\}$.
\textbf{\textit{Encoding}}
Define the encoding functions as follows: $x_{i}^{a}(w_a)=\phi_{i}^{a}(w_a,s_{[i]}^a)=t_{i,w_a}^a(s_i^a)$ and $x_{i}^{b}(w_b)=\phi_{i}^{b}(w_b,s_{[i]}^b)=t_{i,w_b}^b(s_i^b)$ where $t_{i,w_a}^a$ and $t_{i,w_b}^b$ denote the $i$th component of $t_{[n],w_a}^{a}$ and $t_{[n],w_b}^{b}$, respectively, and $s_i^a$ and $s_i^b$ denote the last component $s_{[i]}^a$ and $s_{[i]}^b$, respectively, $i=1,\cdots,n$.
Therefore, to send the messages $w_a$ and $w_b$, we simply transmit the corresponding $t_{[n],w_a}^{a}$ and $t_{[n],w_b}^{b}$, respectively.
\textbf{\textit{Decoding}}
After receiving $(y_{[n]},s_{[n]})$, the decoder looks for the only $(w_a,w_b)$ pair such that $(t_{[n],w_a}^{a}, t_{[n],w_b}^{b}$ $,y_{[n]},s_{[n]})$ are jointly $\epsilon-$typical and declares this pair as its estimate $(\hat{w}_a,\hat{w}_b)$.
\textbf{\textit{Error Analysis}}
Without loss of generality, we can assume that $(w_a,w_b)=(1,1)$ was sent. An error occurs, if the correct codewords are not typical with the received sequence or there is a pair of incorrect codewords that are typical with the received sequence. Define the events $E_{\alpha,\beta}\bydef\big\{(T_{[n],\alpha}^a,T_{[n],\beta}^b,Y_{[n]},S_{[n]})\in A_{\epsilon}^n\big\}$, $\alpha\in\{1,\cdots,2^{nR_a}\}$ and $\beta\in\{1,\cdots,2^{nR_b}\}$.
Then by the union bound we get
\begin{eqnarray}
&&\hspace{-0.3in}P_{e}^{n}=P\big(E_{1,1}^c\bigcup_{(\alpha,\beta)\neq(1,1)}E_{\alpha,\beta}\big)\leq P(E_{1,1}^c)\nonumber\\
&&\hspace{-0.3in}+\sum_{\alpha=1,\beta\neq1}P(E_{\alpha,\beta}) + \sum_{\alpha\neq1,\beta=1}P(E_{\alpha,\beta})+ \sum_{\alpha\neq1,\beta\neq1}P(E_{\alpha,\beta})\nonumber\\\label{eq-erbound}
\end{eqnarray}
where $P(E_{1,1}^c)$ denotes the probability that no message pair is jointly typical. It can easily be verified that $\{Y_i,S_i,T_i^a,T_i^b\}_{i=1}^{\infty}$ is an i.i.d. sequence and by \cite[Theorem 15.1.2]{cover}, $ P(E_{1,1}^c)\rightarrow0$ for sufficiently large $n$.
Next, let us consider the second term
\begin{eqnarray}
&&\hspace{-0.5in}\sum_{\alpha=1,\beta\neq1}P(E_{\alpha=1,\beta\neq1})\nonumber\\
&&\hspace{-0.2in}=\sum_{\alpha=1,\beta\neq1}P((T_{[n],1}^a,T_{[n],\beta}^b,Y_{[n]},S_{[n]})\in A_{\epsilon}^n)\nonumber\\
&&\hspace{-0.2in}\overset{(i)}{=}\sum_{\alpha=1,\beta\neq1}\sum_{(t_{[n]}^a,t_{[n]}^b,y_{[n]},s_{[n]})\in A_{\epsilon}^n}P_{T_{[n]}^b}(t_{[n]}^b)\nonumber\\
&&\quad \quad \quad P_{T_{[n]}^a, Y_{[n]},S_{[n]}}(t_{[n]}^a, y_{[n]},s_{[n]})\nonumber\\
&&\hspace{-0.2in}\overset{}{\leq}\sum_{\alpha=1,\beta\neq1}|A_{\epsilon}^n|2^{-n[H(T^b)-\epsilon]}2^{-n[H(T^a,Y,S)-\epsilon]}\nonumber\\
&&\hspace{-0.2in}\leq2^{nR_b}2^{-n[H(T^b)+ H(T^a,Y,S) - H(T^a,T^b,Y,S)-3\epsilon]}\nonumber\\
&&\hspace{-0.2in}\overset{(ii)}{=}2^{n[R_b-I(T^b;Y|S,T^a)-3\epsilon]}\label{eq-erbo1}
\end{eqnarray}
where $(i)$ holds since for $\beta\neq 1$, $T_{[n],\beta}^b$ is independent of $(T_{[n],1}^a,Y_{[n]},S_{[n]})$ and $(ii)$ follows since $T^b$ and $(T^a,S)$ are independent and $I(T^b;Y,T^a,S)=I(T^b;T^a,S)+I(T^b;Y|T^a,S)$
$=I(T^b;Y|T^a,S)$, where $I(T^b;T^a,S)=0$. Following the same steps for $(\alpha\neq1,\beta=1)$ and $(\alpha\neq1,\beta\neq1)$ we get
\begin{eqnarray}
&&\sum_{\alpha\neq1,\beta=1}P(E_{\alpha,\beta})\leq2^{n[R_a-I(T^a;Y|T^b,S)-3\epsilon]}, \nonumber\\ &&\sum_{\alpha\neq1,\beta\neq1}P(E_{\alpha,\beta})\leq2^{n[R_a+R_b-I(T^a,T^b;Y|S)-3\epsilon]} \label{eq-erbo2},
\end{eqnarray}
and the rate conditions of the $\mathcal{R}(\pi)$ imply that each term tends in (\ref{eq-erbound}) tends to zero as $n \rightarrow \infty$. This shows the achievability of a rate pair $(R_a, R_b) \in \mathcal{R}(\pi)$. Achievability of any rate pair in ${\cal C}_{IN}$ follows from a standard time-sharing argument.
\end{proof}
We now present an outer bound to ${\cal C}_{FS}$, which is obtained by providing a tight converse to the sum-rate capacity. Let \begin{eqnarray}
&&\hspace{-0.3in}{\cal C}_{OUT}:=\biggl\{(R_a,R_b)\in \mathds{R}^{+}\times\mathds{R}^{+}:\nonumber\\
&&\quad \quad \quad \quad \quad R_a+R_b\leq\sup_{\pi_{T^a}(t^a)\pi_{T^b}(t^b)}I(T^a,T^b;Y|S)\biggr\}\nonumber.
\end{eqnarray}
\begin{theorem}[Outer Bound]\label{the-main-outer-f}
${\cal C}_{FS}\subseteq {\cal C}_{OUT}$.
\end{theorem}
\begin{proof}
We need to show that all achievable rates satisfy
\begin{eqnarray}
R_a+R_b \leq \sup_{\pi_{T^a}(t^a)\pi_{T^b}(t^b)}I(T^a,T^b;Y|S),\nonumber
\end{eqnarray}
i.e., a converse for the sum-rate capacity. We use the converse technique of \cite{GiacomoSerdar} and extend it to a noisy setup. Therefore, following \cite{GiacomoSerdar} let $\alpha_{\mathbf{\sigma}}:= \frac{1}{n}P_{S_{[t-1]}}(\mathbf{\sigma})$, $\eta(\epsilon):=\frac{\epsilon}{1-\epsilon}\log|{\cal Y}|+\frac{H(\epsilon)}{1-\epsilon}$. Observe that $\lim_{\epsilon\rightarrow 0}\eta(\epsilon)=0$ and
\begin{eqnarray}
\sum_{\mathbf{\sigma} \in {{\cal S}}^{(n)}}\alpha_{\mathbf{\sigma}}=\frac{1}{n}\sum_{1\leq t \leq n}\enspace \sum_{\mathbf{\sigma} \in {{\cal S}}^{(t-1)}}P_{S_{[t-1]}}(\mathbf{\sigma})=1\nonumber,
\end{eqnarray}
where ${{\cal S}}^{(n)}$ and ${{\cal S}}^{(t-1)}$ are the sets of all ${\cal S}$-strings of length $n$ and $(t-1)$, respectively. First recall that, since $X_t^a=\phi_{t}^{(a)} \left(W_a,S_{[t-1]}^a,S_t^a\right)$ and $X_t^b=\phi_{t}^{(b)} \left(W_b,S_{[t-1]}^b, S_t^b\right)$, we have
\begin{eqnarray}
&&T_t^a=\phi_{t}^{(a)}\left(W_a,S_{[t-1]}^a\right) \in {{\cal X}^{a}}^{|{\cal S}^a|},\nonumber\\ &&T_t^b=\phi_{t}^{(b)}\left(W_b,S_{[t-1]}^b\right) \in {{\cal X}^{b}}^{|{\cal S}^b|}. \label{eq-t2}
\end{eqnarray}
We now show that the sum of any achievable rate pair can be written as the convex combination of conditional mutual information terms which are indexed by the realization of past complete state information.
\begin{lemma}\label{lem:conv-f}
Let $T_t^{a} \in {\cal T}^{a}$ and $T_t^b \in {\cal T}^{b}$ be the Shannon strategies induced by $\phi_{t}^a$ and $\phi_{t}^b$, respectively, as shown in (\ref{eq-t2}). Assume that a rate pair $R=(R_a,R_b)$, with block length $n\geq1$ and a constant $\epsilon \in (0,1/2)$, is achievable. Then,
\begin{eqnarray}
R_a+R_b\leq \sum_{\mathbf{\sigma} \in {\cal S}^{(n)}}\alpha_{\mathbf{\sigma}} I(T_t^a,T_t^b;Y_t|S_t,S_{[t-1]}=\mathbf{\sigma})+\eta(\epsilon) \label{eq-lf1}.
\end{eqnarray}
\end{lemma}
\begin{proof}
Let $\mathbf{T}_t:=(T_t^a,T_t^b)$. By Fano's inequality, we get
\begin{eqnarray}
H(\mathbf{W}|Y_{[n]},S_{[n]})\leq H(\epsilon) + \epsilon\log(|{\cal W}_a||{\cal W}_b|). \label{eq-c1}
\end{eqnarray}
Observing that
\begin{eqnarray}
\hspace{-0.2in}I(\mathbf{W};Y_{[n]}, S_{[n]})&=&H(\mathbf{W})-H(\mathbf{W}|Y_{[n]},S_{[n]})\nonumber\\
\hspace{-0.2in}&=&\log(|{\cal W}_a||{\cal W}_b|)-H(\mathbf{W}|Y_{[n]},S_{[n]}). \label{eq-c2}
\end{eqnarray}
Combining (\ref{eq-c1}) and (\ref{eq-c2}) gives
\begin{eqnarray}
(1-\epsilon)\log(|{\cal W}_a||{\cal W}_b|)\leq I(\mathbf{W};Y_{[n]}, S_{[n]})+H(\epsilon)\nonumber
\end{eqnarray}
and
\begin{eqnarray}
R_a+R_b&=& \frac{1}{n}\log(|{\cal W}_a||{\cal W}_b|)\nonumber\\
&\leq& \frac{1}{1-\epsilon}\frac{1}{n}\left(I(\mathbf{W};Y_{[n]}, S_{[n]})+H(\epsilon)\right)\label{eq-c3}.
\end{eqnarray}
Furthermore, $I(\mathbf{W};Y_{[n]}, S_{[n]})$ can be written as
\begin{eqnarray}
&&\hspace{-0.3in}\sum_{t=1}^{n}\left[H(Y_t,S_t|S_{[t-1]},Y_{[t-1]})-H(Y_t,S_t|\mathbf{W},S_{[t-1]},Y_{[t-1]})\right]\nonumber\\
&&\hspace{-0.3in}\overset{(i)}{=}\sum_{t=1}^{n}\left[H(Y_t|S_{[t]},Y_{[t-1]})-H(Y_t|\mathbf{W},S_{[t]},Y_{[t-1]})\right]\nonumber\\
&&\hspace{-0.3in}\overset{(ii)}{\leq}\sum_{t=1}^{n}\left[H(Y_t|S_{[t]})-H(Y_t|\mathbf{W},S_{[t]},Y_{[t-1]},\mathbf{T}_t)\right]\nonumber\\
&&\hspace{-0.3in}\overset{(iii)}{=}\sum_{t=1}^{n}\left[H(Y_t|S_{[t]})-H(Y_t|S_{[t]},\mathbf{T}_t)\right]\nonumber\\
&&\hspace{-0.262in}=\sum_{t=1}^{n}I(\mathbf{T}_t;Y_t|S_{[t]})\label{eq-c4}
\end{eqnarray}
where $(i)$ follows from the fact that $S_{t}$ is i.i.d. and independent of $\mathbf{W}$, in $(ii)$, $\mathbf{T}_t:=(T_t^a,T_t^b)$ are Shannon strategies whose realizations are mappings $t_t^i:S_t^{i}\rightarrow X_t^i$ for $i=\{a,b\}$ and thus $(ii)$ holds since conditioning reduces entropy. Finally, $(iii)$ follow since
\begin{eqnarray}
&&\hspace{-0.2in}P_{Y_t|\mathbf{W},S_t,S_{[t-1]},Y_{[t-1]},T_t^a,T_t^b}(y_t|\mathbf{w},s_t,s_{[t-1]},y_{[t-1]},t_t^a,t_t^b)\nonumber\\
&&=\sum_{s_t^a,s_t^b}P_{Y_t|S_t,S_t^a,S_t^b,T_t^a,T_t^b}(y_t|s_t,s_t^a,s_t^b,t_t^a,t_t^b)\nonumber\\ &&\hspace{0.9in}\times P_{S_t^a,S_t^b|S_t}(s_t^a,s_t^b|s_t)\nonumber\\
&&=P_{Y_t|S_t,T_t^a,T_t^b}(y_t|s_t,t_t^a,t_t^b)\label{eq-prob}
\end{eqnarray}
where the first equality is verified by (\ref{eq-ch}), where $x_t^i=t_t^i(s_t^i)$ for $i=\{a,b\}$, and by $\{S_t\}$ being i.i.d. and independent of $\mathbf{W}$. Now, let $\chi(\epsilon):=\frac{H(\epsilon)}{n(1-\epsilon)}$ and combining (\ref{eq-c3})-(\ref{eq-c4}) gives
\begin{eqnarray}
&&\hspace{-0.3in}R_a+R_b=\frac{1}{n}\log(|{\cal W}_a||{\cal W}_b|)\nonumber\\
&&\hspace{-0.15in}\leq\frac{1}{1-\epsilon}\frac{1}{n}\sum_{t=1}^nI(T_t^a,T_t^b;Y_t|S_{[t]})+\chi(\epsilon)+(n-1)\chi(\epsilon)\nonumber\\
&&\hspace{-0.15in}\overset{(a)}{\leq}\frac{1}{1-\epsilon}\frac{1}{n}\sum_{t=1}^nI(T_t^a,T_t^b;Y_t|S_{[t]})+\eta(\epsilon)\nonumber\\
&&\hspace{0.7in}-\frac{\epsilon}{1-\epsilon}\frac{1}{n}\sum_{t=1}^nI(T_t^a,T_t^b;Y_t|S_{[t]})\nonumber\\
&&\hspace{-0.15in}=\frac{1}{n}\sum_{t=1}^nI(T_t^a,T_t^b;Y_t|S_{[t]})+\eta(\epsilon)\label{eq-c5}
\end{eqnarray}
where $(a)$ is valid since $I(T_t^a,T_t^b;Y_t|S_{[t]})\leq \log|{\cal Y}|$. Furthermore,
\begin{eqnarray}
&&\hspace{-0.5in}I(T_t^a,T_t^b;Y_t|S_{[t]})\nonumber\\
&&=n\sum_{\mathbf{\sigma} \in {\cal S}^{(t-1)}}\alpha_{\mathbf{\sigma}}I(T_t^a,T_t^b;Y_t|S_t, S_{[t-1]}=\mathbf{\sigma}),
\end{eqnarray}
and substituting the above into (\ref{eq-c5}) yields (\ref{eq-lf1}).
\end{proof}
Observe now that, for any $t\geq1$, $I(T_t^a,T_t^b;Y_t|S_t,S_{[t-1]}=\mathbf{\sigma})$ is a function of the joint conditional distribution of channel state $S_t$, inputs $T_t^a, T_t^b$ and output $Y_t$ given the past realization $(S_{[t-1]}=\mathbf{\sigma})$. Hence, to complete the proof of the outer bound, we need to show that $P_{T_t^a,T_t^b,Y_t,S_t|S_{[t-1]}}(t_t^a,t_t^b,y_t,s_t|\mathbf{\sigma})$ factorizes as in (\ref{eq-joidist-f}). This is done in the lemma below. In particular, it is crucial to observe that the complete state observation at the decoder is enough to provide a product form on $T^a$ and $T^b$.
Before stating the lemma, let us introduce some more notations. Let $\mathbf{\sigma_a}$ and $\mathbf{\sigma_b}$ denote particular realizations of $S_{[t-1]}^a$ and $S_{[t-1]}^b$, respectively. Let
\begin{eqnarray}
&&\Upsilon_{\mathbf{\sigma_a}}^a(t^a):=\{w_a:\phi_{t}^{(a)}(w_a,s_{[t-1]}^a=\mathbf{\sigma_a})=t^a\}, \nonumber\\
&&\Upsilon_{\mathbf{\sigma_b}}^b(t^b):=\{w_b:\phi_{t}^{(b)}(w_b,s_{[t-1]}^b=\mathbf{\sigma_b})=t^b\}\label{eq-set1}
\end{eqnarray}
and
\begin{eqnarray}
\pi_{T^a}^{\mathbf{\sigma_a}}(t^a)&:=&\sum_{w_a\in \Upsilon_{\mathbf{\sigma_a}}^a(t^a)}\frac{1}{|{\cal W}_a|}, \nonumber\\
\pi_{T^b}^{\mathbf{\sigma_b}}(t^b)&:=&\sum_{w_b\in \Upsilon_{\mathbf{\sigma_b}}^b(t^b)}\frac{1}{|{\cal W}_b|}, \nonumber\\
\pi_{T^a}^{\mathbf{\sigma}}(t^a)&:=&\sum_{\mathbf{\sigma_a}}\pi_{T^a}^{\mathbf{\sigma_a}}(t^a)P_{S_{[t-1]}^a|S_{[t-1]}}(\mathbf{\sigma_a}|\mathbf{\sigma}),\nonumber\\ \pi_{T^b}^{\mathbf{\sigma}}(t^b)&:=&\sum_{\mathbf{\sigma_b}}\pi_{T^b}^{\mathbf{\sigma_b}}(t^b)P_{S_{[t-1]}^b|S_{[t-1]}}(\mathbf{\sigma_b}|\mathbf{\sigma}).\label{eq-pis}
\end{eqnarray}
\begin{lemma}\label{lem:fact-f}
For every $1\leq t\leq n$ and $\mathbf{\sigma} \in ({\cal S})^{t-1}$, the following holds
\begin{eqnarray}
&&\hspace{-0.45in}P_{T_t^a,T_t^b,Y_t,S_t|S_{[t-1]}}(t^a,t^b,y,s|\mathbf{\sigma})\nonumber\\
&&=P_S(s)P_{Y|S,T^a,T^b}(y|s,t^a,t^b)\pi_{T^a}^{\mathbf{\sigma}}(t^a)\pi_{T^b}^{\mathbf{\sigma}}(t^b).\label{eq-fact1-f}
\end{eqnarray}
\end{lemma}
\begin{proof}
Let $\textbf{s}:=(s,s_t^a,s_t^b)$ and observe that
\begin{eqnarray}
&&\hspace{-0.5in}P_{T_t^a,T_t^b,Y_t,S_t|S_{[t-1]}}(t^a,t^b,y,s|\mathbf{\sigma})\nonumber\\
&&=\sum_{s_t^a\in{\cal S}^a}\sum_{s_t^b\in{\cal S}^b}
P_{\textbf{S},T^a,T^b,Y|S_{[t-1]}}(\textbf{s},t^a,t^b,y|\mathbf{\sigma})\nonumber\\
&&=\sum_{s_t^a\in{\cal S}^a}\sum_{s_t^b\in{\cal S}^b}
P_{Y|\textbf{S},T^a,T^b}(y|\textbf{s},t^a,t^b)\nonumber\\
&&\hspace{0.8in} \times P_{\textbf{S},T^a,T^b|S_{[t-1]}}(\textbf{s},t^a,t^b|\mathbf{\sigma})\label{eq-fact2-f}
\end{eqnarray}
where the second equality verified by (\ref{eq-ch}) since $x_t^i=t_t^i(s_t^i)$ for $i=\{a,b\}$. Let us now consider the term $P_{\textbf{S},T^a,T^b|S_{[t-1]}}(\textbf{s},t^a,t^b|\mathbf{\sigma})$ above. We have the following
\begin{eqnarray}
&&\hspace{-0.5in}P_{\textbf{S},T^a,T^b|S_{[t-1]}}(\textbf{s},t^a,t^b|\mathbf{\sigma})\nonumber\\
&&\hspace{-0.4in}=\sum_{w_a\in {\cal W}_a}\sum_{w_b\in {\cal W}_b}\sum_{\mathbf{\sigma_a}}\sum_{\mathbf{\sigma_b}}\nonumber\\
&&\hspace{0.1in}P_{\mathbf{W},S_{[t-1]}^a,S_{[t-1]}^b,\textbf{S},T^a,T^b|S_{[t-1]}}(\mathbf{w},\mathbf{\sigma_a},\mathbf{\sigma_b},\textbf{s},t^a,t^b|\mathbf{\sigma})\nonumber\\
&&\hspace{-0.4in}\overset{(i)}{=}P_{\textbf{S}}(\textbf{s})\sum_{w_a\in {\cal W}_a}\sum_{w_b\in {\cal W}_b}\sum_{\mathbf{\sigma_a}}\sum_{\mathbf{\sigma_b}}\nonumber\\
&&\hspace{0.1in}P_{\mathbf{W},S_{[t-1]}^a,S_{[t-1]}^b,T^a,T^b|S_{[t-1]}}(\mathbf{w},\mathbf{\sigma_a},\mathbf{\sigma_b},t^a,t^b|\mathbf{\sigma})\nonumber\\
&&\hspace{-0.4in}\overset{(ii)}{=}P_{\textbf{S}}(\textbf{s})\sum_{w_a\in {\cal W}_a}\sum_{w_b\in {\cal W}_b}\sum_{\mathbf{\sigma_a}}\sum_{\mathbf{\sigma_b}}1_{\{t^l=\Phi_{t}^{(l)}(w_l,\mathbf{\sigma_l}),\enspace l=a,b\}}\nonumber\\
&&\hspace{0.1in}\times P_{\mathbf{W},S_{[t-1]}^a,S_{[t-1]}^b|S_{[t-1]}}(\mathbf{w},\mathbf{\sigma_a},\mathbf{\sigma_b}|\mathbf{\sigma})\nonumber\\
&&\hspace{-0.4in}\overset{(iii)}{=}P_{\textbf{S}}(\textbf{s})\sum_{w_a\in {\cal W}_a}\sum_{w_b\in {\cal W}_b}\sum_{\mathbf{\sigma_a}}\sum_{\mathbf{\sigma_b}}1_{\{t^l=\Phi_{t}^{(l)}(w_l,\mathbf{\sigma_l}),\enspace l=a,b\}}\nonumber\\
&&\hspace{0.1in}\times \frac{1}{|{\cal W}_a|}\frac{1}{|{\cal W}_b|}P_{S_{[t-1]}^a,S_{[t-1]}^b|S_{[t-1]}}(\mathbf{\sigma_a},\mathbf{\sigma_b}|\mathbf{\sigma})\nonumber\\
&&\hspace{-0.4in}\overset{(iv)}{=}P_{\textbf{S}}(\textbf{s})\sum_{\mathbf{\sigma_a}}P_{S_{[t-1]}^a|S_{[t-1]}}(\mathbf{\sigma_a}|\mathbf{\sigma})\sum_{\mathbf{\sigma_b}}P_{S_{[t-1]}^b|S_{[t-1]}}(\mathbf{\sigma_b}|\mathbf{\sigma})\nonumber\\
&&\hspace{-0.4in}\sum_{w_a\in {\cal W}_a} \frac{1}{|{\cal W}_a|}1_{\{t^a=\Phi_{t}^{(a)}(w_a,\mathbf{\sigma_a})\}}\sum_{w_b\in {\cal W}_b} \frac{1}{|{\cal W}_b|}1_{\{t^b=\Phi_{t}^{(b)}(w_b,\mathbf{\sigma_b})\}}\nonumber\\
&&\hspace{-0.4in}\overset{(v)}{=} P_{\textbf{S}}(\textbf{s})\sum_{\mathbf{\sigma_a}}P_{S_{[t-1]}^a|S_{[t-1]}}(\mathbf{\sigma_a}|\mathbf{\sigma}) \sum_{w_a\in \Upsilon_{\mathbf{\sigma_a}}^a(t^a)}\frac{1}{|{\cal W}_a|} \nonumber\\
&&\hspace{0.13in} \sum_{\mathbf{\sigma_b}}P_{S_{[t-1]}^b|S_{[t-1]}}(\mathbf{\sigma_b}|\mathbf{\sigma})\sum_{w_b\in \Upsilon_{\mathbf{\sigma_b}}^b(t^b)}\frac{1}{|{\cal W}_b|}\nonumber\\
&&\hspace{-0.4in}\overset{(vi)}{=} P_{\textbf{S}}(\textbf{s})\sum_{\mathbf{\sigma_a}}P_{S_{[t-1]}^a|S_{[t-1]}}(\mathbf{\sigma_a}|\mathbf{\sigma})\pi_{T^a}^{\mathbf{\sigma_a}}(t^a)\nonumber\\ &&\hspace{0.15in}\sum_{\mathbf{\sigma_b}}P_{S_{[t-1]}^b|S_{[t-1]}}(\mathbf{\sigma_b}|\mathbf{\sigma})\pi_{T^b}^{\mathbf{\sigma_b}}(t^b)\nonumber\\
&&\hspace{-0.4in}\overset{(vii)}{=}P_{\textbf{S}}(\textbf{s})\pi_{T^a}^{\mathbf{\sigma}}(t^a)\pi_{T^b}^{\mathbf{\sigma}}(t^b)\label{eq-pro1-f}
\end{eqnarray}
where $(i)$ is valid since the current state is independent of $\mathbf{W}$ and $(T^a,T^b)$, $(ii)$ is valid by (\ref{eq-t2}), $(iii)$ is valid since $\mathbf{W}$ is independent from the state processes, $(iv)$ is valid by (\ref{eq-sta-no}) and (\ref{eq-t2}), $(v)$ is valid due to (\ref{eq-set1}) and $(vi)-(vii)$ is valid due to (\ref{eq-pis}). Substituting (\ref{eq-pro1-f}) into (\ref{eq-fact2-f}) proves the lemma.
\end{proof}
We can now complete the proof of Theorem \ref{the-main-outer-f}. With Lemma \ref{lem:conv-f} it is shown that the sum of any achievable rate pair can be approximated by the convex combinations of rate conditions given in (\ref{eq-ra3-f}) which are indexed by $\mathbf{\sigma} \in {\cal S}^{(n)}$ and satisfy (\ref{eq-joidist-f}) for joint state-input-output distributions.
More explicitly, we have
\begin{eqnarray}
R_a+R_b&\leq& \sum_{\mathbf{\sigma} \in {\cal S}^{(n)}}\alpha_{\mathbf{\sigma}} I(T_t^a,T_t^b;Y_t|S_t,S_{[t-1]}=\mathbf{\sigma})+\eta(\epsilon)\nonumber\\
&=&\sum_{\mathbf{\sigma} \in {\cal S}^{(n)}}\alpha_{\mathbf{\sigma}} I(T_t^a,T_t^b;Y_t|S_t)_{\pi_{T^a}^{\mathbf{\sigma}}(t^a)\pi_{T^b}^{\mathbf{\sigma}}(t^b)}+\eta(\epsilon)\nonumber\\
&\leq&\sup_{\left(\pi_{T^a}^{\mathbf{\sigma}}(t^a)\pi_{T^b}^{\mathbf{\sigma}}(t^b),\enspace\mathbf{\sigma}\right)}I(T_t^a,T_t^b;Y_t|S_t)+\eta(\epsilon)\nonumber\\
&\leq&\sup_{\left(\pi_{T^a}(t^a)\pi_{T^b}(t^b)\in \Pi\right)}I(T_t^a,T_t^b;Y_t|S_t)+\eta(\epsilon)\nonumber
\end{eqnarray}
where the second step is valid since $I(T_t^a,T_t^b;Y_t|S_t,S_{[t-1]}=\mathbf{\sigma})$ is a function of the joint conditional distribution of channel state $S_t$, inputs $T_t^a, T_t^b$ and output $Y_t$ given the past realization $(S_{[t-1]}=\mathbf{\sigma})$. Hence, since $\lim_{n \rightarrow \infty}\eta(\epsilon)=0$, any achievable pair satisfies $R_a+R_b \leq \sup_{\pi_{T^a}(t^a)\pi_{T^b}(t^b)}I(T^a,T^b;Y|S)$.
\end{proof}
As a direct consequence of Theorem \ref{the-main-outer-f}, we have the following corollary.
\begin{corollary}\label{cor2}
\begin{eqnarray}
{\cal C}^{FS}_{\sum}=\sup_{\pi_{T^a}(t^a)\pi_{T^b}(t^b)}I(T^a,T^b;Y|S)\nonumber.
\end{eqnarray}
\end{corollary}
\begin{remark}
One main observation about the proof of Theorem \ref{the-main-outer-f} is the fact that, once we have the complete state information, conditioning on which allows a product form on $T^a$ and $T^b$, there is no loss of optimality (for the sum-rate capacity) in using associated memoryless team policies instead of using all the past information at the receiver. This fact is observed in \cite{GiacomoSerdar} when the information at the encoders are asymmetric quantized version of the information at the decoder.
\end{remark}
\begin{remark}
It should be noted that the main difference between the problem that we consider here and the one considered in \cite{GiacomoSerdar} is the information at the decoder about the information at the encoders. More explicitly, in \cite{GiacomoSerdar}, the information at the encoders are available at the decoder and as such, as the authors explicitly mention in their paper, the decoder does not need to estimate the coding policies used in a decentralized time-sharing. From this perspective, the main contribution of our work can be thought as showing that when this is not the case, by enlarging the input space, there is no loss of optimality (for the sum-rate capacity) if the optimization is performed by ignoring the past information at the encoders given that the decoder has complete CSI.
\end{remark}
\section{Conclusion and Remarks}\label{conc}
The present paper has investigated the memoryless FS-MAC with asymmetric noisy CSI at the encoders and complete CSI at the decoder. Single letter inner and outer bounds are presented when the channel state is a sequence of i.i.d. random variables. The main contribution of the paper, i.e., the tight converse for the sum-rate capacity and hence an outer bound to the capacity region, is realized by observing that the information available at the decoder is enough to attain a product form on the channel input functions and hence there is no loss of optimality if we ignore the past noisy state information at the encoders.
|
\section{Introduction}
The characterization of the over 500 detected exoplanets has now begun
to take place. Most of the studies are carried out at optical and
infrared wavelengths, because this is where the planetary reflected
light and thermal emission peak, respectively. The discovery of
transiting planets \citep{Charbonneau00, Henry00} allowed astronomers
to constrain new physical parameters such as the radii and masses of
the planets, which are not measurable by the radial velocity method
alone. It is on these systems that in the last years the planetary
atmosphere characterization has achieved the most exciting progress
through the use of spectroscopy and broadband photometry with space
telescopes. Examples are the identification of molecules such as water
absorption \citep[e.g.][]{Tinetti07} or methane \citep{Swain08}, or
the observation of the thermal emission variation with orbital phase
\citep{Knutson07}.
Although great improvements in characterizing the composition of
transiting Hot-Jupiters have been achieved, they only represent about
20\% of the known extrasolar planets\footnote{www.exoplanet.eu}. The
characterization of non transiting planets would require the direct
detection of their light, but the very low flux ratios between the
planets and their host stars makes a direct detection a very
challenging goal. Secondary eclipse observations from Spitzer show
that planet-to-star flux ratios can be as high as 2.5\tttt{-3} between
3.6 and 24 \microns\ \citep[e.g.][]{Knutson2008}. At 2.14 \microns\
the expected flux should be less than these values. Many authors have
attempted a direct detection of the Doppler-shifted signature in
high-resolution spectroscopy from ground-based telescopes. In the
optical \citet{Cameron99} tried to observe the starlight reflected
from the giant exoplanet Tau-Bo\"otis\,b, they found an upper limit to
the albedo and radius using a least-squares deconvolution method that
is well described in the appendices of \citet{Cameron02}, later the
author repeated the analysis on $\upsilon$ Andromeda\,b
\citep{Cameron02}. Recently \citet{Rodler08, Rodler10} searched in the
visible spectra of HD~75289Ab and Tau-Bo\"otis\,b and found upper
limits for their albedos using a model synthesis method. They
constructed a model of the observation composed by a stellar template
plus a shifted and scaled-down version of the stellar template to
simulate the starlight reflected from the planet, these models were
compared to the data by means of $\chi^2$. In the near-infrared,
several attempts have been made to detect Hot-Jupiters by trying to
distinguish the planetary thermal emission from the starlight
\citep{Wiedemann01, Lucas02, Barnes07, Barnes08, Barnes10}, they also
found upper limits for the emitted flux of the planets. All these
authors have used their own variation of a method based on the same
principle of separating the planetary and stellar spectra given their
relative Doppler shifts. Only recently, \citet{Snellen10} claimed the
detection of carbon monoxide from the transmission spectrum of
HD~209458\,b during a transit observation by using high-resolution
spectra; nonetheless, his technique required a transiting system.
In this work we present an effort to constrain new physical parameters
of the non-transiting Hot-Jupiter HD~217107\,b. We attempt to trace
its Doppler-shifted signature (estimated to be \sim 10$^{-4}$ times
dimmer than the star flux) with a correlation function between
high-resolution data and models of its atmospheric spectrum. With
positive detections this method would provide new information on its
characteristics, such as its temperature, chemical composition, and
the presence of chemical tracers associated with life. At the same
time, the method enables the calibration of high-resolution
spectroscopic models for a larger sample of planets that do not
necessarily transit their parent star.
In Section \ref{HD217} we review the planetary system HD~217107; in
Section \ref{data} we describe the observations, data reduction, and
calibration procedures; in Section \ref{analysis} we detail the
theoretical atmospheric spectrum of the planet and the method used to
extract and analyze the planetary signal and present the results of
our data; in Section \ref{simulations} we develop a strategy for the
ideal data acquisition situation and simulate observations of other
planetary systems; and in Section \ref{discussion} we give the
conclusions of our work.
\section{The planetary system HD~217107}
\label{HD217}
\subsection{HD~217107\,b discovery}
\label{HD217107b}
HD~217107 is a main-sequence star that is similar to the Sun in mass,
radius, and effective temperature; its spectral type, G8 IV, indicates
that it is starting to evolve into the red-giant phase
\citep{Wittenmyer07}. The presence of HD~217107\,b was first reported
by \citet{Fischer99} through radial velocity measurements of the star,
the detection was then confirmed by \citet{Naef01}. Later,
\citet{Fischer01} identified a trend in the residuals of the fit, and
\citet{Vogt05} postulated the existence of a third companion in an
external orbit with a period of 8.6 \pm\ 2.7 yr. The presence of this
third object promoted the study of this system in subsequent surveys
\citep{Butler06, Wittenmyer07, Wright09}, constraining more precisely
the companions' orbital parameters. Table \ref{tabpars} summarizes
the parameters used in this work.
\begin{table}[h]
\centering
\caption{Orbital parameters of HD 217107. \label{tabpars}}
\begin{tabular}{lr@{\,\pm\,}l@{\,}c}
\hline\hline
Parameter & \mctc{Value} & References\tablefootmark{a} \\
\hline
Star: \\
Spectral type & \mctc{G8 IV} & W07 \\
\math{T_\mathrm{eff}}(K) & 5\,646 & 26 & W07 \\
K (mag) & 4.536 & 0.021 & C03 \\
\math{d} (pc) & 19.72 & 0.30 & P97 \\
\math{M\sb{s}} (M\math{\sb{\odot}}) & 1.02 & 0.05 & S04 \\
\math{K\sb{s}} (m\,s\math{\sp{-1}}) & 140.6 & 0.7 & W07 \\
\math{v\sb{g}} (km\,s\math{\sp{-1}})& -14.0 & 0.6 & N04 \\
\hline
Planet: \\
\math{P} (days) & 7.12689 & 0.00005 & W07 \\
\math{T\sb{p}} (JD) & 2\,449\,998.50 & 0.04 & W07 \\
\math{e} & 0.132 & 0.005 & W07 \\
\math{m\sb{p} \sin i} (M\sb{\rm Jup}) & 1.33 & 0.05 & W07 \\
\math{a} (AU) & 0.074 & 0.001 & W07 \\
\math{\omega} (deg) & 22.7 & 2.0 & W07 \\
\hline
\end{tabular}
\tablefoottext{a}{W07: \citet{Wittenmyer07}, C03: \citet{Cutri03}, \\
P97: \citet{Perryman97}, S04: \citet{Santos04}.}
\end{table}
\subsection{ Radial velocity}
\label{subsecpc}
The radial velocity of the planet, \math{v\sb{p}\sin i}, around the
center of mass of the system is given by the reflex motion of the
star:
\begin{equation}
\label{eq:twobody}
v_p(t)\sin i\ =\ -\ v_s(t)\sin i\ \frac{m_s}{m_p\sin i}\times\sin i .
\end{equation}
It depends on the mass of the star, $m\sb{s}$; the minimum mass of the
planet, $m\sb{p}\sin i$; the projected radial velocity curve of the
host star, $v\sb{s}(t)\sin i$ (which in turn depends on the parameters
$T\sb{p},\ P,\ e,\ \omega,\ {\rm and}\ K\sb{s}$); and the inclination
of the orbit, $i$, and also on the velocity of the center of mass of
the system, $v\sb{g}$, when measured from Earth. Thus, the radial
velocity curve of the planet is a distinctive curve in time,
parameterized by the values summarized in Table \ref{tabpars}, where
the only unknown parameter is the inclination of the orbit. Figure
\ref{rvfig} shows the radial velocity curve of the star owing to the
interaction with HD~217107\,b, phased over one orbit, with the origin
in phase ($\phi = 0$) at the time of periastron. The radial velocity
of the planet is proportional to this radial velocity curve (Equation
\ref{eq:twobody}).
\begin{figure}[t]
\centering
\includegraphics[trim=15 0 50 0, width=\columnwidth, clip]{f1.eps}
\caption{ Radial velocity curve of HD 217107 \vs\ orbital phase. The
crosses mark the observations of \citet{Wittenmyer07}, which we used to
compute this orbital solution.
The boxes over the curve indicate the coverage of our observations, the
filled boxes represent the runs utilized in the analysis, while the
open boxes represent the discarded runs (details in Section
\ref{results}). }
\label{rvfig}
\end{figure}
\subsection{ Flux estimate}
\label{estimate}
By simulating the spectra of the planet and its host star as black
bodies, we can estimate the order of magnitude of the planet-to-star
flux ratio as a function of wavelength. The black body emission,
$F_\lambda(T)$, is determined by the surface temperature of the
object. While for the star the temperature is well known from models
(see Table \ref{tabpars}), for the planet our best approximation is
the equilibrium temperature
\begin{equation}
T_{eq} = \left( \frac{1-A}{4} \right)^{1/4}
\left( \frac{R_{s}}{a} \right)^{1/2} T_\mathrm{eff}.
\label{eq:teq}
\end{equation}
For a reference value of the bond albedo of $A=0$, we found an
equilibrium temperature for HD~217107\,b of $T_{eq} = 1040 \pm
19$~K. Figure \ref{fig:bbrad} shows the black body spectrum of the
star and the planet assuming a radius between one and two Jupiter
radii, which is the range of the radii for giant extrasolar planets
measured to date.
The planet-to-star flux ratio is given by
\begin{equation}
{\rm Flux\ ratio} \ =\ \frac{F_\lambda(T_{planet})}{F_\lambda(T_{star})}
=\ \frac{B_\lambda(T=1040\ \rm K)}{B_\lambda(T=5646\ \rm K)}
\left(\frac{R_p}{R_s}\right)^2 .
\end{equation}
\begin{figure}[h]
\centering
\includegraphics[trim=10 0 15 0, width=\columnwidth, clip]{f2.eps}
\caption{Black body emission of HD~217107 and HD~217107\,b
assuming a planet radius of 1.0 and 2.0 Jupiter
radii. The vertical dashed line marks the waveband of our data
(2.14 \microns). The directly reflected light component has
little contribution in the infrared and is thus omitted.}
\label{fig:bbrad}
\end{figure}
At 2.14 \microns, the flux ratio varies between 3\tttt{-5} and
1.5\tttt{-4} from one to two Jupiter radii of the planet's radii,
respectively. For shorter wavelengths the flux ratio decreases,
because the star light dominates the emission spectrum. For longer
wavelengths the net fluxes and thus the signal to noise ratio are
lower.
\section{Observations and data reduction}
\label{data}
\subsection{ Observations }
\label{observations}
We observed the planetary system HD~217107 in 11 nights between 2007
August 14 and November 28 using Phoenix \citep{Hinkle03}, a
high-resolution near-infrared spectrometer at the Gemini South
Observatory.
The spectrograph has a 256 x 1024 InSb Aladdin II array with a
resolving power of \math{\ttt{-5}} \microns\ per pixel, the slit
covers 14 arc seconds in length. Its gain is 9.2~e\math{\sp{-}}/ADU
and it has a readout noise of 40~electrons. An argon hollow cathode
wavelength calibration source is supplied with the instrument. Over
950 frames of the system were obtained in service mode, using the
standard ABBA nodding sequences to easily remove sky emission, they
cover a portion in the infrared spectral range from 2.136 to 2.145
\microns\ (see Table \ref{tab:obs}). We tuned the data acquisition
after receiving the data from the first runs since the instrument was
not fully characterized for use on the Gemini Telescope. For the
first two nights, the exposure time was set to 25 seconds, whereas for
the rest of the nights it was set to 80 seconds. We requested arc-lamp
calibration exposures as well.
\begin{table}[h]
\caption{Phoenix observations of HD 217107. \label{tab:obs}}
\centering
\begin{tabular}{ccccc}
\hline\hline
Date & Time on Target\tablefootmark{a} & Orbital Phase &
$\Delta v$\tablefootmark{b} & Status\tablefootmark{c} \\
UT & min & & ms\sp{-1} & \\
\hline
2007-08-14 & \phantom{0}45 & 0.30 & 15.72 & \\
2007-08-16 & \phantom{0}45 & 0.61 & 12.29 & \\
2007-08-22 & \phantom{0}22 & 0.43 & \phantom{0}2.77 & rejected \\
2007-08-26 & \phantom{0}45 & 0.99 & \phantom{0}3.57 & rejected \\
2007-10-02 & 192 & 0.16 & 25.51 & \\
2007-11-19 & \phantom{0}96 & 0.90 & \phantom{0}2.61 & rejected \\
2007-11-23 & \phantom{0}96 & 0.46 & \phantom{0}2.79 & rejected \\
2007-11-24 & \phantom{0}96 & 0.60 & 13.90 & \\
2007-11-25 & \phantom{0}96 & 0.74 & 12.33 & \\
2007-11-26 & \phantom{0}96 & 0.88 & \phantom{0}3.84 & rejected \\
2007-11-28 & \phantom{0}96 & 0.16 & 10.23 & \\
\hline
\end{tabular}
\tablefoottext{a}{Total exposure time of HD~217107.\\}
\tablefoottext{b}{Radial velocity span of the star during the observing time.\\}
\tablefoottext{c}{See Sections \ref{subseccc} and \ref{results} for details.\\}
\end{table}
\subsection{Reduction}
\label{reduction}
We wrote our own interactive data language
(IDL)\footnote{http://www.ittvis.com/ProductServices/IDL.aspx}
routines for the data reduction and analysis, processing each night
and slit position as an independent data set to minimize systematics
caused by different atmospheric conditions or instrumental set-up. We
used the flat-field images to identify hot pixels, marking a pixel as
bad if it had a value beyond 3.5 sigma from the median of the values
of the nine subsequent pixels in its neighborhood. Bad pixels were
masked in all further processing stages. Then, we divided the frames
by a per-night master flat-field and subtracted their corresponding
opposite A or B frame to remove bias and sky. Finally, we extracted
the spectra from the frames with an IDL
implementation\footnote{http://physics.ucf.edu/\math{\sim}jh/ast/software/optspecextr-0.3.1/doc/index.html}
of the optimal spectrum extraction algorithm described in
\citet{Horne86}, this algorithm identified cosmic ray hits, which were
also masked from subsequent processing.
\subsection{Wavelength calibration}
\label{subsecwc}
First, we calibrated the wavelength dispersion using the ThAr lamps,
identifying the line positions and strengths in a high-resolution ThAr
line atlas \citep{Hinkle01}. Because there was only one calibration
lamp for each night, this solution represented only a rough wavelength
calibration, because there are (sub pixel) offsets in wavelength in
the data. To reach the high precision needed for this work, we
fine-tuned the calibration with a high-resolution spectrum of the
Sun\footnote{http://bass2000.obspm.fr/solar\_spect.php} to identify
the telluric lines (identified as those present both in the solar
spectrum and in an average spectrum of our data set).
We constructed an average spectrum to increase the S/N ratio by
aligning and adding the spectra of each night. To determine the
relative shifts, we selected within each set the first spectrum as
reference, while the rest were shifted (using spline interpolations)
to calculate the shift that minimized the root-mean-square of the
correlation with the reference. The centers of fifteen common
absorption lines were identified in wavelength values for the solar
spectrum and in pixel position for our average spectrum. The
wavelength solution is obtained by fitting a second-order polynomial
($\lambda = c_0 + c_1\cdot p + c_2\cdot p^2$) to the solar wavelength
\vs\ the pixel position. Typical fitting coefficients are $c_0=
2.145407$, $c_1=-1.0305\tttt{-5}$, and $c_2=-1.8496\tttt{-10}$. The
dispersion of the residuals is $RMS=4.21\tttt{-6}$ \microns. No
pattern is seen in the residuals. Pixels at wavelengths dominated by
the identified telluric absorption lines were discarded from
subsequent processing owing to their highly variable nature. About
65\% of the pixels remained for the next analysis steps.
\section{ Data analysis and results}
\label{analysis}
\subsection{ Correlation }
\label{subseccc}
Because it is impossible to directly distinguish the planet's
signature from the stellar one in a single spectrum, following the
idea of \citet{Deming00} and \citet{Wiedemann01}, we searched for the
planetary Doppler-shift signature through a correlation method between
the (stellar-subtracted) residual data and a synthetic model of the
planet's spectrum. To remove the stellar flux, we aligned the spectra
for each set (Doppler-shifting them and using a spline interpolation)
in a reference system in which the star remains at rest, and
constructed a stellar template from the average of the set. Then, the
stellar templates and the spectra are normalized dividing by their
respective medians. Finally, the wavelengths of the stellar templates
are shifted according to the orbital phase of the star in each
individual spectrum, and then the stellar template is subtracted from
them. We avoided combining the different nights to obtain the stellar
template, because it is highly probable that other systematics would
be introduced.
Because the planet is approximately a thousand times less massive than
its host star, the planetary Doppler wobble is greater by the same
order of magnitude (see Eq. \ref{eq:twobody}), consequently the
planetary signature will not be added coherently in the stellar
template and thus appear blurred. The stellar template subtraction
leaves a residual spectrum that consists of the signature of the
planet, which is slightly attenuated in the averaging process and
immersed in Poisson noise. The blurring of the planet signature (see
Figure \ref{smoothing}) is determined by the planetary velocity span,
which in turn depends on the time span of an observation and the
orbital phase at the time of the observation. Observations near
inferior or superior conjunction provide the greatest radial velocity
spans, while observations close to the greater elongation of the
planet's orbit produce the smallest radial velocity spans, rendering
the data useless. The rejected data sets in Table \ref{tab:obs} were
observed near greater elongation.
\begin{figure}[t]
\centering
\includegraphics[trim=30 0 55 0, width=\columnwidth, clip]{f3.eps}
\caption{Planetary spectrum blurring in the stellar template. Using
our synthetic spectra of HD~217107\,b we simulated the smearing of
the planetary spectrum over one observing run. The dark-gray and
light-gray lines denote the first and last spectra of HD~217107\,b
during an observing night. The relative shift owing to the orbital
motion of HD 217107 b is 40 kms$^{-1}$ in this simulation. The
bottom black line shows averaged the planetary spectra.}
\label{smoothing}
\end{figure}
For the high-resolution synthetic planetary spectra of HD~217107\,b we
used customized theoretical thermal emission models of its atmosphere
\citep[model described in][]{Fortney05, Fortney06, Fortney08} at three
different distances from the star to account for the non-negligible
eccentricity of the planet. The models are cloud-free, with solar
metallicity, gravity $g=20$~m\,s$^{-2}$, and the molecular abundances
are those appropriate for chemical equilibrium. At these effective
temperatures, the main absorbing molecules are H$_2$O, CH$_4$, CO, and
CO$_2$. The chemistry is described in detail in \citet{Lodders02} and
\citet{Visscher06}. We empirically characterized the instrumental
resolution through the analysis of the emission lines in the
calibration lamps. We convolved the model spectra by the instrumental
resolution, which we determined to be \math{\lambda/\Delta\lambda
\approx 40\,000}. Then, for an assumed value of \math{\sin i}, the
synthetic spectra are Doppler-shifted to mimic the radial velocity of
the planet at the time that the data frame was obtained. The
correlation degree, $C(i)$, is calculated according to the formula
\begin{equation}
C(i) = \frac{1}{N}\sum\sb{k=1}\sp{N} \frac{ \sum\sb{j=1}\sp{N\sb{k}}
\left( r\sb{kj} - \bar r\sb{k} \right) \cdot
\left( \tau\sb{kj}(i) - \bar \tau_k(i) \right) }{ \sqrt{
\left\{ \sum\sb{j=1}\sp{N_k} \left( r\sb{kj} - \bar r_k \right)^2 \right\}
\left\{ \sum\sb{j=1}\sp{N_k} \left( \tau\sb{kj}(i) - \bar \tau_k(i) \right)^2
\right\} } } .
\label{cross}
\end{equation}
In this equation we used the notation $f\sb{kj}$ for the value of
the function at the pixel $j$ of the spectrum $k$, and $\bar f\sb{k}$
for the mean value of the function in the spectrum $k$. Here, ``$r$''
refers to the residual spectrum while ``$\tau(i)$'' to the shifted
planetary model spectrum, with $N_k$ the number of pixels in spectrum
$k$ and $N$ the total number of spectra. The denominator in the
expression normalizes the correlation, and thus a value of 1.0 would
indicate a perfect correlation.
We thus produce a curve of the correlation degree \vs\ the inclination
of the orbit, evaluated in the range $0 < i < \pi/2$. A positive
value of this function indicates that the data spectrum resembles that
of the model, while a negative one suggests anti-correlation. As
consequence of the random nature of the Poisson noise, the value of
the correlation between the residual spectra and the models should be
close to zero, except when the adopted $i$ matches that of the
planetary system. Therefore, an appreciable peak in the correlation
curve would represent a successful detection of the planetary
signature and immediately indicates the value of $i$. By constraining
the inclination with this method, the mass of the planet would be
immediately determined via Eq. 1.
\subsection{Data results}
\label{results}
For this analysis, we excluded the nights where the velocity span of
the star was less than 10 m\,s\sp{-1} (column 4 of Table
\ref{tab:obs}) since they do not represent any significant improvement
in the results, because the shift of the planet ($\sim8.3$
km\,s$^{-1}$) is not significantly higher than the instrumental
resolution ($\sim7.5$ km\,s$^{-1}$). Figure \ref{fig:results} (Top
panel) shows the correlation curve derived from our data as a function
of $\sin i$. The degree of correlation found was close to zero at all
inclinations, and we do not distinguish any identifiable positive peak
that could indicate an atmosphere with absorption features resembling
those of the models.
\subsection{Planet-to-star flux ratio fitting}
\label{detlim}
While the inclination determines the maximum of the correlation curve,
the planet-to-star flux ratio (\psfr) is the main physical parameter
bounded to the magnitude of the correlation. In this section we
determine the most probable values in the parameter space [\psfr\,,\
$\sin i$], which gave rise to our result, and estimate the statistical
significance of the value of the correlation reached. We searched for
the best fitting values comparing our data results
(Fig. \ref{fig:results} Top) with ``synthetic'' correlation curves.
We generated the synthetic correlation curves by recreating our
observations, adding a synthetic planetary spectrum, with known
inclination and planet-to-star flux ratio, according to the following
scheme:
\\
Step 1: We rearranged the order of the data set with random
permutations within each night, but kept the original order of the
dates of the observations. As a consequence, any real planet
signature disappeared, but the noise level of the data was conserved.
\\
Step 2: Using the atmospheric models of the planet, we injected a
synthetic spectrum in the scrambled data set, Doppler-shifted and with
a relative flux according to specific values of $\sin i$ and \psfr,
respectively. For simplicity, we adopted a constant \psfr\ along the
orbit.
\\
Step 3: We processed these synthetic data through the same routines as
in our original data (section \ref{subseccc}). We then iterated for a
grid of values in the ranges: $0 \le i \le \pi/2$ and $\ttt{-5} \le
\psfr \le \ttt{-2}$, obtaining a set of synthetic correlations for
$\sin i$ and \psfr.
Once we obtained these models, we searched for the best-fit parameters
through a $\chi^2$ minimization between the data correlation curve
and the synthetic correlation curves, generating a goodness-of-fit
map (Fig. \ref{fig:results}, bottom panel).
In addition, we used a bootstrap procedure to calculate
false-alarm-probability limits for this map. Following
\citet{Cameron02}, we determined the frequency with which the
correlation degree exceeds a given value as a result of noise in the
absence of a planet signal. The routine consists of performing a
random permutation of the data sets and the subsequent data analysis
(steps 1 and 3 of previous paragraphs) which we repeated a large
number of times (\sim 5\,000), recording the correlation curve after
each trial. This set of correlation curves represents the correlation
found in the absence of a planetary signal, and, because it is created
from the data themselves, defines an empirical probability
distribution that includes both the photon statistics and instrumental
systematics.
Then, at each inclination, we stacked and sorted the values of the
correlation in increasing order. We determined the 1, 2, 3, and
4--\math{\sigma} false-alarm confidence levels as the value of the
correlation degree at the 65, 90, 99 and 99.9 percentiles of the
trials. They represent the signal strengths at which spurious
detections occur with 35, 10, 1, and 0.1 percent false-alarm
probability respectively, at each value of the inclination. This
allows us to assess the probability of obtaining a certain correlation
degree in the absence of planetary emission.
Fig. \ref{fig:results} (Bottom panel) shows the probability map for
HD~217107\,b. The best fit occurs at $\sin i = 0.838$ and
$F\sb{p}/F\sb{s}=3.6\tttt{-3}$, although the relative improvement in
$\chi^2$ against the surrounding parameters is shallow. This value
disagrees with the maximum value of the correlation curve
(Fig. \ref{fig:results} Top), and furthermore, the bootstrap results
indicate that this value is below the 3--$\sigma$ confidence limit of
the signal not being a false positive. Also, this \psfr\ is much
higher than the predicted value from Sec. \ref{estimate} (between
3\tttt{-5} and 1.5\tttt{-4}). The disagreement of the results of the
top and bottom panel in Fig. \ref{fig:results} suggests that
systematics remain after the data reduction, while the strength of the
correlation value, two orders of magnitude above the expected flux
ratio, indicates that this result is not realistic. The 3--$\sigma$
confidence limit only allows us to establish an upper limit in the
flux ratio at 4--5 \tttt{-3} for inclinations greater than $\sin i =
0.6$. In conclusion, we cannot state the detection of \twoseventeenb.
\begin{figure}[t]
\centering
\includegraphics[trim = 5 5 15 5, width=\columnwidth, clip]{f4.eps}
\caption{ Top: Correlation result for our data set as a function of
\math{\sin i}. The correlation remains flat along every inclination
without any distinctive peak, the maximum value is reached at
\math{\sin i=0.71}. Bottom: Goodness of fit, \math{\chi\sp{2}}-map,
of the correlation models to the data. The horizontal and vertical
axes refer to the fitting parameters \math{\sin i} and
\math{F\sb{p}/F\sb{s}} respectively, at which the synthetic
planetary spectrum was added in the correlation models creation,
from which we calculated the minimum squares ($\chi^2\sb{i,{\rm
fr}}$). We plot $\chi^2$ relative to the best fit ($\Delta
\chi\sp{2}\sb{i,{\rm fr}} = \chi\sp{2}\sb{i,{\rm fr}} -
\chi\sp{2}\sb{{\rm min}}$) using the function $\exp(-\alpha\cdot
\Delta \chi\sp{2})$. The gray scale denotes the goodness of fit,
from black for the best fit (at $\chi\sp{2}\sb{{\rm min}}$), to
white for the poorest fit. The plotting parameter, \math{\alpha},
just enables a good contrast in the plots (the same value was used
for all plots). Additionally, we determined with bootstrap
procedures the solid lines (bottom to top) that mark the (1, 2, 3
and 4--\math{\sigma}) levels of false-alarm probability. The white
cross marks the best fit at \math{\sin i = 0.84} and
\math{F\sb{p}/F\sb{s} = 3.6 \tttt{-3}}, situated below the
3--\math{\sigma} confidence level.}
\label{fig:results}
\end{figure}
\section{Future prospects}
\label{simulations}
\subsection{Observational strategy}\label{obstrat}
Although our current data do not enable us to claim the detection of
HD~217107\,b, we identified a strategy to maximize the chances of a
successful detection. This involves selecting suitable candidate
systems and precisely choosing the phasing and span of the
observations. To exemplify the advantages, we simulated realistic
observations of other planetary systems.
\begin{figure*}[t]
\includegraphics[trim = 5 5 13 5, scale=0.425, clip]{f5a.eps}
\includegraphics[trim = 7 5 15 5, scale=0.425, clip]{f5b.eps}
\caption{ Correlation curves and \math{\chi\sp{2}}-maps of synthetic
data of HD 179949. A synthetic planetary signal was injected in the
spectra with
parameters: \math{\sin i=0.77} and \math{F\sb{p}/F\sb{s}=0.003}.
Left: Using our observing strategy. Right: random distributed
observing dates. The routine successfully recovers the signal at
\math{\sin i = 0.78} and \math{F\sb{p}/F\sb{s} = 2.8 \tttt{-3}} in
both cases, although, when using our strategy, the correlation
degree is stronger, and the parameters are better determined
compared with the right panel.}
\label{sim_179}
\end{figure*}
We limited our sample to the currently known extrasolar planets
without transits\footnote{http://exoplanet.eu}, observables from
Gemini South, for the first semester of the year. Even though we
constrained ourselves to the instrument we characterized and to a
fixed time span, the purpose of these simulations is to provide one
successful detection with our method. It is plausible that considering
the full extent of possibilities, stronger signals can be acquired.
The improvement in a detection, limited by purely photon noise, can be
quantified by the planet-to-star flux ratio and the stellar flux,
according to the expression (fluxes in number of photons)
\begin{equation}
\label{eq:snr}
F_p/{\rm Noise}\ =\ \frac{(\psfr) \cdot F_s}{\sqrt{|F_s|+|F_p|}}\
\approx\ (\psfr)\cdot\sqrt F_s .
\end{equation}
Then, for example, the spectrometer CRIRES with four times the
wavelength coverage of that of Phoenix, has twice the sensitivity of
Phoenix. We decided to simulate the Phoenix instrument, since it is
well characterized by our group, while other instruments should
present their own systematics, which are hard to quantify.
We simulated the planets as if the strength of the high-resolution
absorption features were the same as that of our models for
HD~217107\,b, but with the corresponding \psfr\ (estimated as in
Sec. \ref{estimate}). A caveat for this assumption is that the
strength of the lines is not very clear in planets that exhibit
thermal inversions. \citet{Burrows2008} and \citet{Fortney08} suggest
that the emission features should be weaker.
The target selection criteria are based first on the radial velocity
span of the planet, where we set a lower limit cutoff of 7.5
km\,s\sp{-1} (equivalent to the FWHM of the instrument spectral
resolution) for a three-hour observing run if the orbit was at \math{
i = 30\sp{\circ}}. Second, we look for higher apparent brightnesses
of the stars for better signal-to-noise ratios. Table \ref{tabplanets}
lists two of the better suited selected planetary systems (HD~217107
listed for comparison). A brighter K-band magnitude of the star
improves the signal-to-noise ratio, while a smaller semi-major axis
involves a higher radial velocity span, which enables a greater
Doppler shift of the planet spectra during the runs and at the same
time favors higher planet-to-star flux ratios.
\begin{table}[b]
\caption{ Favorable targets for Gemini South. \label{tabplanets} }
\centering
\begin{tabular}{lcccc}
\hline\hline
Target & \math{a} & M\sb{K} & \math{K\sb{p}} \\
& AU & & km\,s\sp{-1} \\
\hline
HD~179949 & 0.045 & 4.94 & 158.23 \\
Tau Boo & 0.046 & 3.51 & 150.62 \\
HD~217107 & 0.073 & 4.54 & 112.28 \\
\hline
\end{tabular}
\tablefoot{ Planet's radial velocity amplitude for an
orbit with \math{\sin i=1}.}
\end{table}
To simulate the observations, we recreated the same instrumental
settings of our data, but carefully selected the observing schedule.
For each one of the nights in the period and restricted to air masses
under 1.5, we selected the three-hour range that gives the maximum
velocity span. We recorded then, the radial velocity spans for each
night, and chose those with the biggest spans. We used the solar
spectrum to simulate the stellar spectrum, while for the planetary
component we used the atmospheric models of HD~217107\,b added with a
given planet-to-star flux ratio and inclination. Each component is
Doppler-shifted according to the orbital parameters. Finally we added
Poisson noise to the spectra, according to the signal-to-noise
corresponding to the magnitude of the target. The synthetic data were
processed in exactly the same way as our original data.
\subsection{ Simulations }
\label{simu}
In our first test, we present two simulations of an observing campaign
on a target with the physical parameters of HD~179949 to show the
improvements of our observing strategy in contrast with a regular
observation. The given parameters are \math{\sin i = 0.77} and
\math{F\sb{p}/F\sb{s} = 3 \tttt{-3}}. Figure \ref{sim_179} left shows
the simulation following our observing strategy, while Fig.
\ref{sim_179} right shows the simulation selecting random observing
dates. In both cases the correlation curves (top panels) mark the
inclination of the synthetic orbit with a increment in the correlation
degree near $\sin i = 0.77$, while the $\chi^2$-maps (bottom panels)
effectively indicate the best fit at $\sin i = 0.78$ and
$F\sb{p}/F\sb{s} = 2.8 \tttt{-3}$.
We identify the main differences between these two simulations: First,
given the larger radial velocity spans when implementing our strategy,
the planetary spectrum is more blurred in the stellar template and
consequently less reduced when the template is subtracted, the
planetary spectrum signal is thus stronger in the residual spectrum,
which increases the correlation degree. As consequence of these
greater correlation degrees, all confidence levels are generally
lower, because it is less probable to reach this correlation degree by
chance in the no-planet case, and lastly, the $\chi^2$-map peak is
much better determined. The improvement is reflected more in the
distinction of the best fit against other values of the parameter
space than in the distinction against the no-planet case.
In another simulation, we recreated the planetary system Tau Boo as
close as possible to its real physical characteristics
(Figure \ref{sim_others}). Tau~Boo\,b had the parameters $\sin i = 0.82$
and $F\sb{p}/F\sb{s}=4\tttt{-4}$, our routines returned the best fit:
\math{\sin i = 0.79} and \math{F\sb{p}/F\sb{s}=3.6\tttt{-4}},
slightly underestimating the values. Nevertheless, the
\math{\chi\sp{2}}-map shows an improvement in the region near the
injected inclination and flux ratio. The bootstrap results set the
3--$\sigma$ confidence limit near $F\sb{p}/F\sb{s}$ 1.5 \tttt{-4} (for
$\sin i > 0.5$), indicating a detection with 99\% confidence.
\section{Discussion and conclusions}
\label{discussion}
Because the instrument was not well characterized at the time and our
service-mode observational strategy had to be adapted after the first
few observing windows, the data for HD~217107 were not as sensitive as
expected. The correlation curve was featureless for all inclinations
and with values close to zero, with a maximum at $\sin i = 0.71$. By
fitting the sine of the inclination and the planet-to-star flux ratio
through least-squares, we found the best-fit parameters of $\sin i =
0.84$ and $F\sb{p}/F\sb{s} = 3.6 \tttt{-3}$ at a level below our
3--\math{\sigma} confident limit. As a consequence of the faint
features in the results, the disagreement between the peak in the
correlation (Fig. \ref{fig:results} Top panel) and the most probable
value of \math{\sin i} (Fig. \ref{fig:results} Bottom panel), and the
higher than predicted $F\sb{p}/F\sb{s}$, we could not claim a
detection of HD~217107\,b with our current data. Given the results of
the bootstrap procedure, we reject the flux ratio of HD~217107\,b to
its host star to be over 5 \tttt{-3} (3--\math{\sigma} confidence).
We attribute these results to the absence of an ideal strategy in the
data acquisition at the time of the observations and a needed further
treatment of the instrument systematics.
\begin{figure}[t]
\includegraphics[trim = 5 5 13 5, scale=0.425, clip]{f6.eps}
\caption{\label{sim_others} Similar to Fig. \ref{fig:results},
correlation curve (top panel) and \math{\chi\sp{2}}-map (bottom
panel) of a simulation of the planetary system Tau~Boo, with an
injected companion at \math{\sin i = 0.82} and \math{F\sb{p}/F\sb{s}
= 4 \tttt{-4}}.}
\end{figure}
We could not detect HD~217107\,b, but defined the outlines of future
campaigns by carefully defining a candidate selection criterion and an
observational strategy. We conclude that the best-suited candidates
for this technique are those in very close orbits, which allow the
planets to have high orbital velocities and higher planet-to-star flux
ratios. We propose an observing strategy where for the period of
observations we specifically select the nights with maximum radial
velocity spans. To explore the capabilities of our routines, we
simulated other planetary systems as observed by the Phoenix
spectrograph, with the same number of hours and an appropriate
schedule of observations. The system HD~179949 was recreated,
contrasting the use of our observing strategy with a regular
observation schedule. We recovered the planetary signature in both
cases, but showing an improvement in the correlation degree, precision
in the \math{\chi\sp{2}}-map, and lower \math{\sigma} limits when
using our observing schedule. Finally, we performed a realistic
simulation of the planetary system Tau Boo, and successfully detected
its signature.
In conclusion high-resolution instruments like Phoenix are capable of
detecting extrasolar planet Doppler-shifted signals with flux ratios
as low as \ttt{4} with this method if we perform a careful treatment
of the systematics (approaching the photon noise limit), if we count
with appropriate theoretical models, and if we follow an optimized
scheme in the data acquisition. Furthermore, using other instruments
like CRIRES or NIRSPEC could increase the confidence of the
results. Since our simulations exclude systematics effects specific to
the instrument, an adequate treatment to remove them would be
necessary. Our optimal observing strategy tends to select
observations at superior conjunctions of the planet's orbit, capturing
the highest amount of light possible from the planet and at the same
time covering the highest radial velocity span for a determined time
extent. Refinements of this technique will involve the optimization
of the distribution of time designated to the length of an observing
run \vs\ the number of nights of observation, while adding
phase-dependent functions of the planet's brightness to account for
the changing observed portion of the day/night side of the planet, and
for different amounts of irradiation in eccentric orbits, will
increase the accurateness of the fitted parameters.
\begin{acknowledgements}
Patricio Cubillos and Patricio Rojo are supported by the FONDAP
Center for Astrophysics 15010003, the center of excellence in
Astrophysics and Associated Technologies (PFB06) and the FONDECYT
project 11080271.
\end{acknowledgements}
|
\section{Acknowledgements}
The authors would like to acknowledge Hannie van den Broek, Cock Harteveld, L\'eon Woldering,
Willem Tjerkstra, Ivo Vellekoop, Christian Blum, and Vinod Subramaniam for their support
and insightful discussions.
This work is part of the research program
of the ``Stichting voor Fundamenteel Onderzoek der Materie (FOM)'', which is financially
supported by the ``Nederlandse Organisatie voor Wetenschappelijk Onderzoek (NWO)''.
JB is partially financed by the FIRB-MIUR "Futuro in Ricerca" project RBFR08UH60.
WLV thanks NWO-Vici and APM is supported by a Vidi grant from NWO.
\bibliographystyle{apsrev}
|
\section{Introduction}
Recently we have presented an algorithm~\cite{me-2010-PBCI} for the
approximation of the ground state of translationally invariant (TI) spin
chains with periodic boundary conditions (PBC) my means of TI MPS.
In this work we will use the ground states
obtained in~\cite{me-2010-PBCI} as the basis of an ansatz for
excited states with definite momentum.
We will consider only spin chain Hamiltonians that are translationally
invariant thereby fulfilling $[H,T]=0$ where $T$ is the translation
operator that shifts the lattice by one site.
Furthermore, as we will deal with finite chains in the following,
it means that there is no well defined momentum operator for our systems.
Nevertheless we can classify translationally invariant states
by their quasi-momentum which is defined in terms of the their
eigenvalue with respect to $T$. This definition is sensible
since in the thermodynamic limit, if we keep the chain length fixed,
the lattice spacing becomes infinitesimally small
and the quasi-momentum becomes identical to the momentum, which is
then well defined. For convenience, we will use
the term \emph{momentum} when we actually refer to the quasi-momentum.
This should not cause any confusion since we will only deal with
quasi-momenta throughout this work.
Since $H$ and $T$ commute, they can be diagonalized simultaneously.
This suggests that any variational ansatz based on eigenstates of
the translation operator will be well suited
to define families of states within which minimization with
respect to some variational parameters will yield \emph{momentum}
eigenstates with minimal energy.
Formulating this observation in terms of an MPS-based ansatz has
led in the past to some very interesting results about excitation spectra.
The first approach in this direction has been made in
\cite{rommer-ostlund-1995}
where the main result
is the celebrated insight that the fixed point of the density matrix
renormalization group (DMRG) can be written as an MPS.
In addition to this, based on
the MPS that is obtained for the ground state of the infinite Heisenberg
spin-$1$ chain, the authors suggest a variational ansatz for excitations
with definite \emph{momentum}.
Since the translationally invariant MPS they start
with is an approximation of the ground state in the thermodynamic limit,
their ansatz for excitations is only well suited in the limit $N\to\infty$.
For finite chains, the idea of using \emph{momentum} eigenstates for
the diagonalization of TI Hamlitonians has been used in~\cite{porras-2006}
in order to obtain a few of the lowest branches of excitations
for the bilinear-biquadratic (BB) spin-1 chain. The resulting state is a
TI superposition of a special class of tensor network states, which can
be viewed as an extension of MPS with PBC~\cite{frank-2004a} to
states that can accommodate multipartite entanglement.
Even though the multipartite entanglement is a nice feature which yields
a better variational ansatz in the cases when the approximated states have
that special entanglement structure (in~\cite{porras-2006} one has in
addition to the usual maximally entangled virtual bonds between nearest
neighbors a virtual GHZ state connecting all sites) we will not adopt it
in our present ansatz. Furthermore we would like to point out that the
individual MPS tensors produced by the minimization procedure
in~\cite{porras-2006} are not TI, only their superposition is.
Recent results~\cite{me-2010-PBCI} on the
approximation of ground states of TI PBC Hamiltonians opened up the
possibility of unifying the ideas
from~\cite{rommer-ostlund-1995} and~\cite{porras-2006} in order to obtain
an algorithm for excitations with definite \emph{momentum} in which only one
local tensor has to be determined, thereby avoiding the usual sweeping
procedure and the associated factor $N$ in the computational cost.
One of the main features of TI MPS is the fact that the
tensor network that has to be contracted for the computation of expectation
values contains big powers of a so-called transfer
matrix~\cite{fannes-1992}.
For non-critical systems the eigenvalues of this transfer matrix
usually decay rapidly enough s.t. big powers thereof can be
accurately approximated by considering only a few dominant eigenvectors.
In these cases the computational cost for the evaluation of expectation
values for systems with PBC
can be reduced significantly from $O(D^5)$ to $O(D^3)$, where $D$
denotes the virtual bond dimension of the MPS. For critical systems however
the eigenvalues of the transfer matrix decay much slower and the
algorithm that must be employed in order to obtain the optimal approximation
within the class of MPS with fixed $D$ has a scaling that depends in a
not yet fully understood way~\cite{me-2010-PBCI} on $D$, $N$ and on the
universality class of the simulated model.
The ansatz we present in this work is based on TI-MPS and thereby
all computed quantities will contain big powers of the transfer matrix.
We would like to emphasize that the computational cost can be reduced
by a factor of $D^2$ only in the case of non-critical systems. For critical
systems the full contraction of tensor networks (i.e. without using any
approximations of the transfer matrix) will turn out to have a more
favorable overall scaling of the computational cost.
Details on why this is the case and on the scaling of the computational cost
can be found in the next section.
\section{Overview}
\label{sec:overview}
Due to $T^N=1\hspace{-0.243em}\text{l}$, the translation operator $T$ that shifts a state on a
PBC lattice with $N$ sites by $1$ site is the generator of the cyclic group
of order $N$. Hence its eigenvalues $\tau_k$ must be
roots of the unity i.e. $\tau_k=e^{-i k \frac{2\pi}{N}}$ with
integer $k\in [0,N-1]$.
An ansatz for eigenstates of $T$ with eigenvalue $e^{-i k \frac{2\pi}{N}}$
is obviously given by
\begin{equation}\label{eq:ansatz}
\ket{\psi_k({\bvec{B}})}=\sum_{n=0}^{N-1} \frac{1}{\sqrt{N}}
e^{i\frac{2\pi k n}{N}} T^n \ket{\phi_{\bvec{A}}({\bvec{B}})}
\,\,\,.
\end{equation}
\noindent
Henceforth we will refer to states of the form (\ref{eq:ansatz}) as
Bloch states.
Note that we have used the convention that $T$ is the operator that realizes
a translation by one site to the right s.t.
$T\ket{\phi(i_1,i_2,\dots,i_N)}=\ket{\phi(i_N,i_1,i_2,\dots,i_{N-1})}$.
The state $\ket{\phi_{\bvec{A}}({\bvec{B}})}$ can in principle be any arbitrary state,
but in order to exploit the advantages of TI MPS, we choose
\begin{equation}\label{eq:ansatz_phi}
\ket{\phi_{\bvec{A}}({\bvec{B}})}=
\sum_{i_1,\dots i_N=1}^{d}
\mbox{Tr}\big(B_{i_1} A_{i_2} \dots A_{i_N}\big)
\ket{i_1 i_2 \dots i_N}
\end{equation}
\noindent
with identical matrices $A_i$ on all sites except the first one.
We will choose the $A_i$ to be the matrices corresponding
to the best TI MPS ground state approximation for a given model.
We emphasize that the $A_i$ remain fixed throughout the entire simulation.
This is the reason why we have omitted them from our labeling convention
for the Bloch states $\ket{\psi_k({\bvec{B}})}$.
We have used bold letters in order to denote objects that are
obtained if one rearranges the components of three indexed
MPS tensors into vectors, i.e.
$\bvec{A}:=\mbox{vec}( A_{i\phantom{\alpha}\beta}^{\phantom{i}\alpha} )$.
After fixing the \emph{momentum} $k$, the Bloch states $\ket{\psi_k({\bvec{B}})}$
will depend only on the tensors ${\bvec{B}}$ which will define the variational
manifold.
Our ansatz for Bloch eigenstates differs slightly from the ones
presented in Refs.~\cite{rommer-ostlund-1995,porras-2006,chung-2009}
although it is conceptually very similar. An important feature
of all these approaches is the reduction of the dimension of the
problem by a factor $N$. This is reached by effectively
projecting the original problem into the subspace with fixed
\emph{momentum} $k$ and minimizing the energy within the variational
manifold spanned by the free parameters in the ansatz. In our
case these free parameters are the components ${\bvec{B}}$ of an MPS tensor.
As it is always the case with MPS algorithms, one must eliminate
the ambiguities arising from the MPS representation by fixing the gauge.
Here this is done by starting with certain tensors ${\bvec{A}}$ in
(\ref{eq:ansatz_phi}) and not changing them throughout the entire
minimization procedure.
This automatically fixes the gauge of the tensors ${\bvec{B}}$ as they are
surrounded on both sides by ${\bvec{A}}$.
\section{The algorithm}
\label{sec:algorithm}
Ansatz (\ref{eq:ansatz}) defines a class of variational states
for the lowest energy states with fixed \emph{momentum}.
The energy is a quadratic expression in the tensor ${\bvec{B}}$ and thereby,
as it is usually the case in MPS based algorithms, minimizing
\begin{equation}\label{eq:variational_problem}
E_0(k)
=\min_{\bvec{B}\in\mathcal{C}^{dD^2}}
\frac{\bra{\psi_k(\bvec{B})} H \ket{\psi_k(\bvec{B})}}
{\braket{\psi_k(\bvec{B})|\psi_k(\bvec{B})}}
\,\,\,,
\end{equation}
\noindent
is equivalent to solving a generalized eigenvalue equation
\begin{equation}\label{eq:generalized_eigenvalue_problem}
H_{eff}(k) \bvec{B}_i(k) = E_i(k) N_{eff}(k) \bvec{B}_i(k)
\end{equation}
\noindent
where $H_{eff}(k)$ is defined by
\begin{equation}\label{eq:Heff_def}
{\bvec{B}}^{\dagger} H_{eff}(k) {\bvec{B}}
:= \bra{\psi_k(\bvec{B})} H \ket{\psi_k(\bvec{B})}
\end{equation}
\noindent
and $N_{eff}(k)$ by
\begin{equation}\label{eq:Neff_def}
{\bvec{B}}^{\dagger} N_{eff}(k) {\bvec{B}}
:= \braket{\psi_k(\bvec{B})|\psi_k(\bvec{B})}
\,\,\,.
\end{equation}
\noindent
The eigenvector corresponding to the smallest eigenvalue
$E_0(k)$ yields then the tensor ${\bvec{B}}_0(k)$ that when plugged into our
ansatz~(\ref{eq:ansatz}) gives the \emph{momentum}-$k$ state with
minimal energy.
Note that the variational principle guarantees that only the Bloch state
(\ref{eq:ansatz}) with lowest energy is the best approximation to
the exact eigenstate with that \emph{momentum}
within the subspace spanned by our ansatz states.
However if the lowest energy state is approximated accurately, due to the fact
that the other ${\bvec{B}}_i(k)$ are orthogonal to ${\bvec{B}}_0(k)$, the next solution
${\bvec{B}}_1(k)$ has a good chance to be close to the next higher energy state
with that momentum.
In fact it will turn out that quite a few
of the higher energy solutions of
(\ref{eq:generalized_eigenvalue_problem}) are
good approximations to low energy states with fixed \emph{momentum}.
Their precision is most of the time surprisingly good given the fact that
the variational principle does not hold for these states. The quality
of these solutions depends strongly on the bond dimension $D$, the
chain length $N$ and the model under consideration.
\begin{figure}[h]
\begin{center}
\includegraphics[width=1.0\columnwidth]{N_0m.pdf}
\end{center}
\caption{
(Color online).
Definition of $N_{0m}({\bvec{A}})$ as the norm of a TI MPS determined
by the tensor ${\bvec{A}}$.
}
\label{fig:N_0m}
\end{figure}
The bottleneck of our method is the computation of the effective
matrices $H_{eff}(k)$ and $N_{eff}(k)$. Let us first
consider $N_{eff}(k)$ since it is the slightly simpler one.
It reads
\begin{equation}\label{eq:Neff}
\begin{split}
{\bvec{B}}^{\dagger} &N_{eff}(k) {\bvec{B}} =\\
&=\frac{1}{N} \sum_{m,n=0}^{N-1} e^{i\frac{2\pi k (n-m)}{N}}
\bra{\phi_{\bvec{A}}({\bvec{B}})} T^{(n-m)} \ket{\phi_{\bvec{A}}({\bvec{B}})}\\
&=\sum_{\bar{n}=0}^{N-1} e^{-i\frac{2\pi k \bar{n}}{N}}
\bra{\phi_{\bvec{A}}({\bvec{B}})} T^{-\bar{n}} \ket{\phi_{\bvec{A}}({\bvec{B}})}\\
&= {\bvec{B}}^{\dagger} \bigg[ \sum_{m=0}^{N-1} e^{-i\frac{2\pi k m}{N}}
\cdot N_{0m}({\bvec{A}}) \bigg] {\bvec{B}}
\end{split}
\end{equation}
\noindent
where $N_{0m}({\bvec{A}})$ is a tensor network
resembling the norm of a TI MPS with empty slots $0$ and $m$
(see figure~\ref{fig:N_0m}).
To get from the second to the third line we have used the
fact that due to the PBC only the relative distance between $n$ and
$m$ plays a role. In the last line we have merely renamed the
summation index and introduced the quantity $N_{0m}({\bvec{A}})$.
Thus in order to obtain $N_{eff}(k)$ we have to compute the contraction of
the $N$ tensor networks $N_{0m}({\bvec{A}})$ and then take
the sum of these terms after weighting each one of them with the
corresponding phase factor. The computational cost for the contraction
of each tensor network is $O(D^6)$ s.t. the overall cost for computing
$N_{eff}(k)$ is $O(ND^6)$.
\begin{figure}[h]
\begin{center}
\includegraphics[width=1.0\columnwidth]{H_0nm.pdf}
\end{center}
\caption{
(Color online).
Definition of $H_{0nm}({\bvec{A}})$ as the expectation value of a
two-site operator with respect to a TI MPS determined
by the tensor ${\bvec{A}}$.
}
\label{fig:H_0nm}
\end{figure}
$H_{eff}(k)$ is constructed very much in the same spirit.
First, due to the translational invariance of the Hamiltonian we can write
\begin{equation}\label{eq:H_TI}
H=\sum_{l=0}^{N-1} h_{l,l+1} = \sum_{l=0}^{N-1} T^{l} h_{01} T^{-l}
\end{equation}
\noindent
where $h_{01}$ is the term acting between the first two sites of the
chain. Note that in~(\ref{eq:H_TI}) we have restricted ourselves
to nearest neighbor Hamiltonians since this is the type of models
we will treat numerically in this work. Generalizing the ideas developed
here to any local Hamiltonian is straightforward.
With~(\ref{eq:H_TI}), $H_{eff}(k)$ reads
\begin{equation}\label{eq:Heff_01}
\begin{split}
&{\bvec{B}}^{\dagger} H_{eff}(k) {\bvec{B}} =\\
&=\frac{1}{N} \sum_{l,m,n=0}^{N-1} e^{i\frac{2\pi k (n-m)}{N}}
\bra{\phi_{\bvec{A}}({\bvec{B}})} T^{l-m} h_{01} T^{n-l} \ket{\phi_{\bvec{A}}({\bvec{B}})}\\
&=\frac{1}{N} \sum_{l=0}^{N-1} \sum_{\bar{m},\bar{n}=-l}^{N-1-l}
e^{i\frac{2\pi k (\bar{n}-\bar{m})}{N}}
\bra{\phi_{\bvec{A}}({\bvec{B}})} T^{-\bar{m}} h_{01} T^{\bar{n}} \ket{\phi_{\bvec{A}}({\bvec{B}})}
\,\,\,.
\end{split}
\end{equation}
\noindent
Again, due to the fact that the $\bar{m}$ and $\bar{n}$ sums run over
all $N$ sites of a PBC chain, it is irrelevant where they begin s.t.
the $l$ sum merely yields a factor $N$. We rename the summation
indices for convenience and obtain
\begin{equation}\label{eq:Heff_02}
\begin{split}
&{\bvec{B}}^{\dagger} H_{eff}(k) {\bvec{B}} =\\
&=\sum_{m,n=0}^{N-1} e^{i\frac{2\pi k (n-m)}{N}}
\bra{\phi_{\bvec{A}}({\bvec{B}})} T^{n-m} T^{-n} h_{01} T^{n} \ket{\phi_{\bvec{A}}({\bvec{B}})}\\
&=\sum_{n=0}^{N-1} \sum_{\bar{m}=n}^{n-N+1}
e^{-i\frac{2\pi k \bar{m}}{N}}
\bra{\phi_{\bvec{A}}({\bvec{B}})} T^{-\bar{m}} h_{n,n+1} \ket{\phi_{\bvec{A}}({\bvec{B}})}\\
&={\bvec{B}}^{\dagger} \bigg[ \sum_{m,n=0}^{N-1}
e^{-i\frac{2\pi k m}{N}}
\cdot H_{0nm}({\bvec{A}}) \bigg] {\bvec{B}}
\end{split}
\end{equation}
\noindent
where $H_{0nm}({\bvec{A}})$ is a tensor network
resembling the expectation value of an operator acting on the sites
$n$ and $n+1$ with respect to a TI MPS where the slots $0$ and $m$
have been left open (see figure~\ref{fig:H_0nm}).
The computational cost for the contraction
of each tensor network is again $O(D^6)$ but now we have a total
of $N^2$ summands s.t. the overall cost for computing
$H_{eff}(k)$ is $O(N^2D^6)$. Note that to obtain $H_{eff}(k)$ is
computationally the most expensive part of our algorithm so we
can say that the overall computational cost scales like $O(N^2D^6)$.
\begin{figure*}[ht]
\begin{center}
\includegraphics[width=1.0\textwidth]{IS_Ei_g10_N50_D8D32.pdf}
\end{center}
\caption{
(Color online).
Lowest ten branches of the excitation spectrum for a critical
Ising chain with $N=50$. Left: $D=8$.
Right: $D=32$.
}
\label{fig:IS_Ei_g10_N50}
\end{figure*}
\subsection{Overall scaling of the computational cost}\label{sec:comp_cost}
At first sight the cost seems horrible for a 1D-algorithm.
Let us however have a closer look at what we get for this price.
First of all note that if we compute the sets of matrices
$\{N_{0m}({\bvec{A}})\}$ and $\{H_{0nm}({\bvec{A}})\}$ for $n,m\in[0,N-1]$ and store
these, we can obtain the $H_{eff}(k)$ and $N_{eff}(k)$ for
all $k$ trivially by just building the appropriately
weighted sums.
For each of these $k$ we then have to solve the generalized
eigenvalue equation~(\ref{eq:generalized_eigenvalue_problem}).
Since $H_{eff}(k)$ and $N_{eff}(k)$ are small $dD^2\times dD^2$ matrices
solving~(\ref{eq:generalized_eigenvalue_problem}) does not represent
any difficulty and can be done using any standard library eigenvalue solver.
Each eigenvalue problem leads to $dD^2$ orthonormal vectors ${\bvec{B}}_i(k)$
which plugged into the ansatz~(\ref{eq:ansatz}) yield
$dD^2$ states. Thus computing the sets $\{N_{0m}({\bvec{A}})\}$ and
$\{H_{0nm}({\bvec{A}})\}$ only once supplies immediately us with $NdD^2$ states!
By comparing our numerical results to exactly solvable
models we will show that the low energy states obtained
in this way are very accurate. This means that in terms of computational
time per state our algorithm performs quite well.
The computational bottleneck at the moment is that
we have to store $N^2$ $dD^2\times dD^2$ matrices in the memory.
With the present implementation, for a chain with $N=100$ sites,
we can go up to $D=32$. For larger $N$ simulations we have to settle
for smaller $D$. It is however straightforward how this boundary can
be pushed considerably towards larger $D$. First, instead of keeping
all matrices in the memory, one can write them to the hard disk after
computing each of them. Second, since the $\{N_{0m}({\bvec{A}})\}$ and
$\{H_{0nm}({\bvec{A}})\}$ are independent, one can parallelize their computation.
Thus the conceptual bottleneck becomes the contraction of the tensor
networks $\{N_{0m}({\bvec{A}})\}$ and $\{H_{0nm}({\bvec{A}})\}$. For non-critical
systems big powers of the transfer matrix can be well approximated by
a few of its dominant eigenvectors~\cite{me-2010-PBCI} and the contraction
of \emph{most} of the $\{N_{0m}({\bvec{A}})\}$ can be done with the computational
cost $O(n^2 D^3)$ while that of \emph{most} of the $\{H_{0nm}({\bvec{A}})\}$ with
the cost $O(n^3 D^3)$. Here $n$ represents the dimension of the subspace
within which we approximate the powers of the transfer
matrix~\cite{me-2010-PBCI}. This cannot be done however for critical
systems where in principle $n$ may grow as big as $D^2$ thereby
yielding a much worse scaling than the naive $O(D^6)$.
Note that since $\{N_{0m}({\bvec{A}})\}$ and $\{H_{0nm}({\bvec{A}})\}$ are
\emph{open} tensor networks the $O(D^5)$ contraction
scheme~\cite{frank-2004a} that works
for expectation values (i.e. \emph{closed} tensor networks) cannot be
applied here. Additionally, even if we restrict ourselves
to non-critical systems, not \emph{all} of the $\{N_{0m}({\bvec{A}})\}$
and $\{H_{0nm}({\bvec{A}})\}$ can be computed with the cost that scales
like $D^3$: if the distance between the \emph{open} slots is not
big enough, we cannot use the approximation for big powers of
the transfer matrix between the slots, and we are back to exact
contraction for this portion of the chain which in the case of
$\{N_{0m}({\bvec{A}})\}$ leads to the overall scaling $O(nD^5)$ and
in the case of $\{H_{0nm}({\bvec{A}})\}$ to the scaling $O(n^2D^5)$.
Thus the very naive exact contraction procedure that we use is
not so bad after all in this case even if it scales like $O(D^6)$.
There is one more subtlety we would like to point out here.
It turns out that the matrix $N_{eff}(k)$ is always singular which
presents a problem when we try to solve the generalized eigenvalue
equation~(\ref{eq:generalized_eigenvalue_problem}) since the solution
involves the inverse $N_{eff}^{-1}(k)$. We can circumvent this
problem by solving~(\ref{eq:generalized_eigenvalue_problem})
within the nonsingular subspace like it has been done
in~\cite{rommer-ostlund-1995}. Eigenvectors associated to the zero
eigenspace of $N_{eff}(k)$ will result in physical states
$\ket{\psi_k({\bvec{B}})}=0$, i.e. these are states of zero length in the Hilbert
space. Any physical operator will produce a zero when acting on these
states. In particular, the effective Hamiltonian $H_{eff}(k)$ will also
have zero eigenvalues for the same eigenvectors, and we do not loose any
information by restricting to the nonsingular subspace. The dimension of the
zero eigenspace can be shown to be $D^2(d-1)$ for $k\neq 0$ and
$D^2(d-1)+1$ for $k=0$ as we demonstrate in \cite{juthoEXC-2011}.
The tricky point is that for some models the strictly non-zero
eigenvalues of $N_{eff}(k)$ become so small that they yield the
generalized eigenvalue problem ill conditioned.
In general this behavior does not occur for small $D$.
For big $D$ or in certain regions of the phase diagram however
the nonsingular eigenvalues become so small that it is
hard to distinguish numerically between the singular subspace and
the nonsingular one. This issue might be the source of the mysterious
negative gap that appears in~\cite{rommer-ostlund-1995} in the
vicinity of the critical point.
We have employed a slightly different method for the regularization of
$N_{eff}(k)$. Instead of projecting the problem into the strictly
non-singular subspace, we restrict ourselves to the subspace
in which the eigenvalues of $N_{eff}(k)$ are larger than some $\epsilon$.
There is a tradeoff between loss in precision due to this projection
and loss in precision due to the bad conditioned generalized eigenvalue
problem. In the end we have settled for a seemingly optimal
$\epsilon=10^{-11}$.
\section{Numerical results}
\label{sec:results}
We have applied the algorithm presented above to two exactly solvable
nearest neighbour interaction spin models in order to benchmark its
accuracy: the quantum Ising model and the Heisenberg spin-$1/2$
antiferromagnetic chain.
Even though the Heisenberg model is exactly solvable by
means of Bethe ansatz, in practice it is much harder to obtain its
entire low-energy spectrum.
This is due to the fact that the elementary excitations
are two-spinon states and among these, the solution of the Bethe ansatz
equations for the two-spinon singlet states are computationally
very challenging~\cite{INTRO2BA-II}. Thus for long chains we have restricted
ourselves to check only the precision of the lowest two-spinon triplets.
For small chains that are accessible via exact diagonalization on the
other hand, we compare not only the entire low-energy spectrum
but also the fidelity of the states themselves.
\begin{figure*}[ht]
\begin{center}
\includegraphics[width=1.0\textwidth]{IS_reldiffEi_g10_N50_D8D32.pdf}
\end{center}
\caption{
(Color online).
Relative precision of the low excitation spectrum for a critical
Ising chain with $N=50$. Left: $D=8$.
Right: $D=32$.
}
\label{fig:IS_reldiffEi_g10_N50}
\end{figure*}
\begin{table*}[ht]
\begin{center}
\includegraphics[width=1.0\textwidth]{IS_g10_N50_spectrum_from_elementary_excitations_4branches.pdf}
\end{center}
\caption{
(Color online).
Quasi-particle structure of the lowest four branches for $g=1$.
The red/blue (grayscale: dark/light) boxes highlight states from the
odd-parity subspace respectively from the even parity subspace.
The ground state which is not shown in the table is the
fermionic vacuum in the even parity subspace
i.e. $\ket{GS}=\ket{\Omega}_{even}$.
}
\label{tab:IS_g10_N50_spectrum_from_elem_exc}
\end{table*}
\begin{figure}[ht]
\begin{center}
\includegraphics[width=1.0\columnwidth]{IS_LAMBDAk_g10_N50_EXACT.pdf}
\end{center}
\caption{
(Color online).
Exact solution for the dispersion relation of the Bogoliubov modes at
criticality (i.e., $g=1$).
}
\label{fig:IS_LAMBDAk_g10_N50_EXACT}
\end{figure}
\subsection{Quantum Ising model}
The Hamiltonian we have used in our simulations of the Quantum Ising
model is given by
\begin{equation}\label{eq:H_IS_sim}
H_{IS} = -\sum_{i} Z_i Z_{i+1} - g\sum_{i}X_i
\,\,\,.
\end{equation}
\noindent
We have used this version rather than
\begin{equation}\label{eq:H_IS_exact}
H'_{IS} = -\sum_{i} X_i X_{i+1} - g\sum_{i}Z_i
\,\,\,.
\end{equation}
\noindent
due to the fact that having a diagonal interaction term is more convenient
for the numerics.
Of course both versions are equivalent since they can be transformed into
each other by means of the unitary transformation
$U=\bigotimes_{i=1}^{N}H_i$ where the $H_i$ are 1-qubit Hadamard gates.
For the exact diagonalization however we have used (\ref{eq:H_IS_exact})
in order to stick closer to the procedure given in
\cite{lieb-schultz-mattis-1961} where the authors treat the
spin-$1/2$ XY chain.
The diagonalization of (\ref{eq:H_IS_exact}) for PBC in the limit of an
infinite number of sites is straightforward~\cite{pfeuty-1970}.
The first thing one has to do is to map the spin Hamiltonian to a fermionic
one via a Jordan-Wigner transformation.
Now the Jordan-Wigner transformation is non-local due to the fact
that it transforms local spin operators into fermionic ones
that anticommute if they act on different sites.
Luckily for almost all terms in the Hamiltonian the non-localities
cancel except for the term representing the interaction between
the last and the first site. This term ends up containing a global
parity operator acting on all sites and thus breaking the translational
invariance with respect to the fermionic modes.
Now, if we are interested in the thermodynamic limit, we will eventually
take the limes $N\to\infty$ at some point, and in this limit the
contribution of one interaction term to the energy can be neglected.
We thus have the freedom
to alter this term as we please in order to simplify things. One very
convenient choice are the so-called Jordan-Wigner boundary conditions
which are nothing more than simply neglecting the global parity operator
in the last term thereby yielding the fermionic Hamiltonian translationally
invariant. Note that the Jordan-Wigner boundary conditions cannot be
expressed in a trivial way in terms of spin operators.
The fermionic Hamiltonian obtained in this way is quadratic and
translationally invariant, but it is not particle conserving.
This can be fixed by a canonical transformation
\cite{lieb-schultz-mattis-1961} to non-interacting
Bogoliubov fermions. The ground state of the system is then given by
the new fermionic vacuum while excited states can be obtained by
sequentially filling the fermionic modes.
Ordering the eigenstates of the original spin model by momentum
and energy, it turns out that the lowest energy branch coincides
with the dispersion relation of the Bogoliubov fermions.
This happens because for a given
momentum, the lowest energy state is always a state where precisely
one fermionic mode is occupied.
Now, for finite systems with periodic boundary conditions, the Hamiltonian
after the Jordan-Wigner transformation presents a difficulty:
due to the fact that in this case we cannot choose the boundary conditions
freely, there is one term that contains a global parity operator as
prefactor (see Eq. 2.11' in \cite{lieb-schultz-mattis-1961}).
At first sight, this term makes the Bogoliubov transformation impossible.
However, if we project the Hamiltonian onto the subspaces with either
odd or even parity, we can replace the parity operator by its eigenvalue
in that subspace s.t. it becomes $\pm 1$, and
we can apply the Bogoliubov transformation in each subspace separately.
The spectrum of the original Hamiltonian can then be constructed by
picking from each subspace the states with the correct parity. It turns
out that we can arbitrarily choose the sign of the Bogoliubov
parity by shifting the Fermi surface. For example, if we choose the
fermionic vacuum to be the state with lowest energy, all excited states are
particle excitations
\footnote{
If the fermionic vacuum is not the lowest energy state there also exist
hole and particle-hole excitations. If we choose the vaccum in such a way
that there exists exactly one hole-mode we effecively switch the sign of the parity operator opposed to choosing no hole-modes at all.
}
and it turns out that the parity operator
changes its sign under the Bogoliubov transformation for fields
below the critical point i.e. $g < 1$. For $g \ge 1$ this choice of
the vacuum state leaves the parity operator invariant.
Thus for $g<1$, in principle we can switch the sign of the parity operator
by shifting the Fermi surface and thereby we could always
define the Bogoliubov modes such that the parity operator remains invariant.
We will however give numerical evidence for the fact that choosing the Fermi
surface to be the fermionic vacuum state is the physical choice.
Let us first present the results obtained at the critical point $g=1$.
In Figure~\ref{fig:IS_Ei_g10_N50} we have plotted the energy of the lowest
ten branches of excitations of a chain with $50$ spins obtained for MPS bond
dimensions $D=8$ and $D=32$. The results for $D=32$ are so close to the
exact spectrum that it makes much more sense to look at plots of the
relative energy precision rather than at plots of the energy itself.
This is shown in Figure~\ref{fig:IS_reldiffEi_g10_N50}.
\begin{figure*}[ht]
\begin{center}
\includegraphics[width=1.0\textwidth]{IS_reldiffEi_g10_N100D32_N500D20.pdf}
\end{center}
\caption{
(Color online).
Relative precision of the low excitation spectrum for the critical
Ising chain with different chain lengths. Left: $N=100$, $D=32$.
Right: $N=500$, $D=20$.
}
\label{fig:IS_reldiffEi_g10_N100_N500}
\end{figure*}
At first sight the crossing of the precision line for the first branch
of excitations with the
one for the second branch seems a little unusual. How can it be that
states with higher energy are approximated by roughly two orders of
magnitude better than states with lower energy? The answer to this question
is obvious if one looks at how the eigenstates emerge from the elementary
Bogoliubov modes. Table~\ref{tab:IS_g10_N50_spectrum_from_elem_exc}
shows which Bogoliubov modes contribute to each individual eigenstate
in the first four branches of excitations. Modes from the even parity
subspace have half-integer momentum while the ones from the odd parity
subspace have integer momentum.
Note that since only excitations with an even number of particles
are allowed in the even-parity subspace, the resulting states always
have integer momentum.
Henceforth $\ket{\Omega}_{even}$ shall denote the vacuum in the even parity
subspace and $\ket{\Omega}_{odd}$ the vacuum in the odd parity one.
The ground state of the critical chain is the Bogoliubov
vacuum in the even parity subspace i.e.
$\ket{GS}=\ket{\Omega}_{even}$.
The first excited state is the zero momentum state from the first branch
and is given by a Bogoliubov mode with zero momentum from the odd
parity subspace
\footnote{
Actually at $g=1$ the zero momentum mode has energy zero so the
states $\ket{\Omega}_{odd}$ and $\ket{0}$ have exactly
the same energy. However this happens only at the critical point $g=1$.
In general $\ket{\Omega}_{odd}$ and $\ket{0}$ have different energy.
}.
It is sufficient to show in
Table~\ref{tab:IS_g10_N50_spectrum_from_elem_exc}
how the spectrum emerges from elementary excitations for momenta
$0\leq k \leq N/2$ since the dispersion relation of the Bogoliubov
fermions is symmetric around $k=0$ as can be seen in
Figure~\ref{fig:IS_LAMBDAk_g10_N50_EXACT}.
The important thing to notice in
Table~\ref{tab:IS_g10_N50_spectrum_from_elem_exc} is that the lowest
branch of excitations does not contain solely one-particle excitations
as it does in the case of the infinite chain. Looking back at the right plot
in Figure~\ref{fig:IS_reldiffEi_g10_N50} we see immediately that
the one-particle excitations from the first two branches are approximated
with roughly the same accuracy between $10^{-11}$ and $10^{-9}$ with the
lower value for small momentum states.
One can easily check that the states with the same order of accuracy from
higher branches are precisely one-particle excitations.
On the other hand it is obvious that two-particle excitations from any
branch where one of the particles has fixed momentum $k=1/2$
can be found in the plateau with relative precision of roughly $10^{-7}$.
The other plateaus of similar precision in the $D=32$ plot of
Figure~\ref{fig:IS_reldiffEi_g10_N50} represent either two-particle
states where each particle has higher momentum or three and more particle
excitations.
\begin{figure*}[ht]
\begin{center}
\includegraphics[width=1.0\textwidth]{IS_reldiffEi_g11_N50_N100_D32.pdf}
\end{center}
\caption{
(Color online).
Relative precision of the low excitation spectrum for the
Ising chain at $g=1.1$ for different chain lengths. Left: $N=50$, $D=32$.
Right: $N=100$, $D=32$.
}
\label{fig:IS_reldiffEi_g11_N50_N100}
\end{figure*}
\begin{table*}[ht]
\begin{center}
\includegraphics[width=1.0\textwidth]{IS_g11_N50_spectrum_from_elementary_excitations_3branches.pdf}
\end{center}
\caption{
(Color online).
Quasi-particle structure of the lowest three branches for $g=1.1$.
The red/blue (grayscale: dark/light) boxes highlight states from the
odd-parity subspace respectively from the even parity subspace.
The ground state which is not shown in the table is the
fermionic vacuum in the even parity subspace
i.e. $\ket{GS}=\ket{\Omega}_{even}$.
}
\label{tab:IS_g11_N50_spectrum_from_elem_exc}
\end{table*}
This interpretation of the branch crossings in
Figure~\ref{fig:IS_reldiffEi_g10_N50} is strong evidence for the fact
that (\ref{eq:ansatz}) is a very good ansatz for one-particle excitations.
However it turns out that if $D$ is large enough, (\ref{eq:ansatz}) is
also a fairly good ansatz for many-particle excitations. The reason for
this is that the large bond dimension compensates for the localization
of excitations inherent in ansatz (\ref{eq:ansatz}) by spreading the
effect of the optimized tensor ${\bvec{B}}$ to a region around it whose size is
of the order of the induced correlation length of the MPS we start with.
This is exactly why for the Ising chain with $g=1$, $N=50$ and $D=32$
even states from tenth branch are approximated with an accuracy of
roughly $10^{-4}$.
Now let us have a closer look at the region of
the "level-crossing" between the lowest two-fermion branch from
the even parity subspace and the lowest one-fermion branch from the
odd parity subspace. In the case of $g=1$ this crossing turns out to be at
approximately $N/4$. In the immediate neighborhood of the crossing
the energy difference between states with identical momentum becomes very
small. Now if the bond dimension $D$ is chosen such that the precision
of the MPS is of the same order like the interlevel spacing, these two
levels cannot be discriminated properly by the MPS algorithm and thus
there is no clear interpretation we can give to these MPS states in terms
of one or two-particle states. As can be seen in the $D=8$ plot of
Figure~\ref{fig:IS_reldiffEi_g10_N50}, in this region the first and second
MPS branch interpolate between the one and the two-particle states which we
can safely discriminate. Note that this observation holds only on the
side of the level crossing where the one-particle state has higher energy
than the two-particle state (e.g. at $k\approx N/4$ on the left side of
the crossing). On the other side, the one-particle excitation has lower
energy and our one-particle MPS ansatz is perfectly suited to discriminate
between the first and the second branch even if the precision is smaller
than the actual gap between the levels.
The last thing we would like point out about
Figure~\ref{fig:IS_reldiffEi_g10_N50} is the gap in accuracy between the
states from the second and third branch at momentum $k=N/2$. It turns
out that this is a doubly degenerate state since it can be created
by two different superpositions of elementary excitations with
the same energy namely $\ket{\frac{49}{2},\frac{1}{2}}$ and
$\ket{-\frac{49}{2},-\frac{1}{2}}$. This is the reason why the precision
of the $k=N/2$ state in the second branch is better than that of the
surrounding two-particle states which are not degenerated:
variational algorithms are more precise if they try to approximate
the energy of an entire subspace of the Hilbert space rather than that
of a single state. However since all states generated by our algorithm
are orthogonal, the price we have to pay for the improved precision
in the second branch, is a slightly worse precision of the $k=N/2$
state in the third branch.
With this said, we can present the results we have obtained for different
chain lengths $N$ and different values of the magnetic field $g$.
Figure~\ref{fig:IS_reldiffEi_g10_N100_N500} shows the accuracy of the
algorithm for chains with $100$ and $500$ sites at $g=1$.
The plot for $N=100$ is very similar to the $D=32$ plot from
Figure~\ref{fig:IS_reldiffEi_g10_N50}. At small momenta $6\le k\le 11$
the one-particle excitations lie in the branches $4$ to $6$.
These states are not reliably reproduced by our algorithm within
the precision that is otherwise reached for one-particle excitations.
Presumably this would be fixed by increasing the bond dimension $D$
beyond $32$. However at the moment we cannot go to larger $D$ for $N=100$.
For $N=500$ the maximally accessible bond dimension is $D=20$.
The corresponding plot from Figure~\ref{fig:IS_reldiffEi_g10_N100_N500}
is very similar to the small $D$ plot for $N=50$.
Again in the region of the "level-crossing" between one-particle
and two-particle excitations around $k=N/4$ our algorithm has difficulties
obtaining the maximally reachable precision for the one-particle states.
\begin{figure*}[ht]
\begin{center}
\includegraphics[width=1.0\textwidth]{IS_reldiffEi_g9_N50_N100_D32.pdf}
\end{center}
\caption{
(Color online).
Relative precision of the low excitation spectrum for the
Ising chain at $g=0.9$ for different chain lengths. Left: $N=50$, $D=32$.
Right: $N=100$, $D=32$.
}
\label{fig:IS_reldiffEi_g9_N50_N100}
\end{figure*}
\begin{table*}[ht]
\begin{center}
\includegraphics[width=1.0\textwidth]{IS_g9_N50_spectrum_from_elementary_excitations_3branches.pdf}
\end{center}
\caption{
(Color online).
Quasi-particle structure of the lowest three branches for $g=0.9$.
The red/blue (grayscale: dark/light) boxes highlight states from the
odd-parity subspace respectively from the even parity subspace.
$\ket{\Omega}_{odd}$ denotes the fermionic vacuum in the odd-parity
subspace. The ground state which is not shown in the table is the
fermionic vacuum in the even parity subspace
i.e. $\ket{GS}=\ket{\Omega}_{even}$.
}
\label{tab:IS_g9_N50_spectrum_from_elem_exc}
\end{table*}
Now let us look at how the algorithm performs when we move away from
the critical point.
Figure~\ref{fig:IS_reldiffEi_g11_N50_N100} shows the relative energy
difference of the MPS approximation for $g=1.1$ i.e. above the critical
point. The most striking feature in this regime is clear separation
of the lowest branch of excitations from the higher ones.
This happens due to the fact that in this case the lowest branch
contains only one-particle states as can be seen in
Table~\ref{tab:IS_g11_N50_spectrum_from_elem_exc}.
Again if $D$ is large enough (e.g. $D=32$ for $N=50$), the different
plateaus of similar precision become clearly visible. The first one
at roughly $\Delta_{rel}E_i(k)\approx 10^{-8}$ contains two-particle
excitations from the second and third branch where one of the
fermionic modes has momentum $k=1/2$. The second one with precision
around $10^{-6}$ consists of states where one of the fermionic
modes has momentum $k=3/2$. Note that in the plot for $N=100$ the
lowest branch has slightly better precision than the one in the
$N=50$ plot even though the virtual bond dimension is the same.
This happens presumably because in this case the chain is long enough
such that the running particle cannot "feel its own tail" due to the PBC.
This is another piece of evidence that ansatz (\ref{eq:ansatz})
is very well suited to describe one-particle excitations. Whether
many-particle excitations are faithfully reproduced depends strongly
on the magnitude of $D$ with respect to $N$.
For $g<1$ the picture changes dramatically.
We can see in figure~\ref{fig:IS_reldiffEi_g9_N50_N100} that at $g=0.9$
the best precision for states from the lowest branch is five to seven
orders of magnitude worse than for $g=1.1$. Without any knowledge of the
quasi-particle structure of the spectrum this huge difference might look
a bit surprising. Even more surprising is the fact that the best precision
at $g=0.9$ is one order of magnitude worse than at the critical point $g=1$.
However looking at the quasi-particle structure in
Table~\ref{tab:IS_g9_N50_spectrum_from_elem_exc} clarifies the
situation. As already mentioned above the parity of the Bogoliubov
fermions in the odd-parity subspace can in principle be arbitrarily
chosen by shifting the Fermi surface. Throughout this work have made the
most natural choice of choosing all modes to have positive energy i.e.
none of the quasi-particle excitations are hole modes. For $g<1$ this choice
switches the sign of Bogoliubov parity operator such that we must
pick states with an even number of excitations from the odd-parity subspace.
One might argue against this convention and claim that it would be
much more natural to pick the Fermi surface s.t. the zero-momentum
mode is a hole excitation which yields the Bogoliubov parity operator
identical to the spin parity operator. In this case we would
have to construct all states from this subspace using an odd number
of quasi-particles. On the other hand
Table~\ref{tab:IS_g9_N50_spectrum_from_elem_exc}
clearly shows that our one-particle excitation ansatz (\ref{eq:ansatz})
is a poor approximation to all states in this regime thereby indicating
that indeed for $g<1$ there exist no one-particle excitations.
Thus our choice of the Fermi surface is justified and we have
to construct the spectrum by picking the even quasi-particle excitations
from the odd-parity subspace.
We can understand this behavior from another point of view if we consider
an infinite chain with open boundary conditions. It is well known that
in the region of the phase diagram where the ground state is
doubly degenerated, the elementary excitations are kink excitations.
If we would however impose
periodic boundary conditions on the infinite chain, the single kink states
would not be eigenstates any more since the existence of one domain
wall would automatically imply the existence of a second one.
In finite systems with PBC, the situation is a bit more complicated
since the ground state degeneracy is not exact
(the energy difference decays exponentially with $N$),
but we can still argue along similar lines that localized
perturbations, that interpolate between the states of the
almost degenerated ground state manifold, must always come in pairs.
\subsection{Heisenberg model}
The other model we have studied is the antiferromagnetic (AF)
Heisenberg spin-$1/2$ chain.
The Hamiltonian reads
\begin{equation}\label{eq:H_HB}
H_{HB} = \sum_{i=1}^{N} \vec{S}_i \vec{S}_{i+1}
= \frac{1}{4} \sum_{i=1}^{N}
(\sigma^x_i \sigma^x_{i+1} + \sigma^y_i \sigma^y_{i+1}
+\sigma^z_i \sigma^z_{i+1})
\end{equation}
\noindent
where $S^\alpha=\sigma^\alpha/2$ and $\sigma^\alpha$ denote as usually
the Pauli operators. As we already mentioned, the tensors
${\bvec{A}}$ that constitute the backbone of ansatz (\ref{eq:ansatz}) are
the results of the simulations presented in~\cite{me-2010-PBCI}.
In that work we have obtained a TI MPS approximation of ground states
for finite spin chains with PBC using matrices $A_i$ that were
real and symmetric. These results themselves were based
on previous work~\cite{me-2010-MPOR} where we have approximated the
ground state of infinite OBC chains by TI MPS with real symmetric matrices.
Thus the starting point in the entire procedure that leads ultimately
to the excited states presented here is the imaginary time evolution
for an infinite chain
with a set of real symmetric matrix product operators (MPO).
As we explained in~\cite{me-2010-MPOR} it is not possible to
construct these directly from the the Hamiltonian (\ref{eq:H_HB}).
However, by means of the unitary transformation
$U=U^{\dagger}=\prod_{j=1}^{N/2} \sigma^y_{2j-1}$
(i.e. acting with a $\sigma^y$-gate on every second site) we obtain
\begin{equation}\label{eq:H_HB_transf}
H'_{HB}=U^{\dagger} H_{HB} U
=\frac{1}{4} \sum_{i=1}^{N}
(-\sigma^x_i \sigma^x_{i+1} + \sigma^y_i \sigma^y_{i+1}
-\sigma^z_i \sigma^z_{i+1})
\end{equation}
\noindent
which allows us to express the imaginary time evolution in terms
of real symmetric MPO. Note that in order for this procedure to work
we have to restrict ourselves to chains with an even number of sites.
In this case it does not matter if we apply the $\sigma^y$-gates
on sites with an odd or an even index, so without loss of generality
we will apply them on the odd ones.
Now $H_{HB}$ and $H'_{HB}$ have the same spectrum and since their
eigenstates are simply related to eachother by
\begin{equation}\label{eq:eigenstate_transf}
\ket{\psi_i}=\prod_{j=1}^{N/2} \sigma^y_{2j-1} \ket{\psi_i'}
\end{equation}
\noindent
we can digonalize $H'_{HB}$ first and obtain the eigenstates of $H_{HB}$
subsequently with very little effort.
\begin{table}[ht]
\begin{center}
\includegraphics[width=0.8\columnwidth]{HB_N16_spectrum_totalspin_parity_10branches.pdf}
\end{center}
\caption{
(Color online).
Multiplet structure of the lowest ten branches of excitations
for a Heisenberg
$16$-site chain with Hamiltonian (\ref{eq:H_HB}).
The colors encode the multiplet information: yellow-singlet,
blue-triplet, red-quintuplet, dark red-septuplet
(grayscale: darker colors encode higher multiplets).
The states within each multiplet are ordered according to their
total spin projection quantum number.
The sign denotes the parity of a state.
}
\label{tab:HB_N16_spectrum_multiplets_parity}
\end{table}
\begin{table}[ht]
\begin{center}
\includegraphics[width=0.8\columnwidth]{HBtransf_N16_spectrum_totalspin_parity_10branches.pdf}
\end{center}
\caption{
(Color online).
Multiplet structure of the lowest ten branches of excitations
for a Heisenberg
$16$-site chain with Hamiltonian (\ref{eq:H_HB_transf}).
The colors encode the multiplet information: yellow-singlet,
blue-triplet, red-quintuplet
(grayscale: darker colors encode higher multiplets).
The sign denotes the parity of a state and the index
denotes the momentum $k$ if we apply the transformation
(\ref{eq:eigenstate_transf}) to an eigenstate with momentum $k'$.
}
\label{tab:HBtransf_N16_spectrum_multiplets_parity}
\end{table}
We will show below that the momentum of a state is not always invariant under
the transformation (\ref{eq:H_HB_transf}).
The easiest way to obtain the momentum for any given state is by
computing the expectation value of the translation operator $T$ with
respect to that state.
$H_{HB}$ and $H'_{HB}$ are both translationally invariant thus
all their eigenstates have well defined momentum so we can be
sure that the reverse transformation
$\ket{\psi_i'(k')} \rightarrow \ket{\psi_i(k)}$ will map momentum
eigenstates to momentum eigenstates albeit $k$ will generally differ
from $k'$.
The relation between the momenta follows easily from
\begin{equation}\label{eq:k_transf}
\begin{split}
&e^{-i\frac{2\pi k}{N}}=\braket{\psi_i(k)|T|\psi_i(k)} =\\
&=\braket{\psi_i'(k')|\bigg(\prod_{j=1}^{N/2} \sigma^y_{2j-1}\bigg)
T \bigg(\prod_{j=1}^{N/2} \sigma^y_{2j-1}\bigg) |\psi_i'(k')} \\
&=\braket{\psi_i'(k')|\bigg(\prod_{j=1}^{N/2} \sigma^y_{2j-1}\bigg)
\bigg(\prod_{j=1}^{N/2} \sigma^y_{2j}\bigg) T |\psi_i'(k')} \\
&=e^{-i\frac{2\pi k'}{N}}
\braket{\psi_i'(k')|\prod_{j=1}^{N} \sigma^y_{j}\,|\psi_i'(k')}
=e^{-i\frac{2\pi k'}{N}} \braket{P_y}_{i',k'}
\end{split}
\end{equation}
\noindent
where we have used $T \big(\prod_{j} O_{j}\big) T^{-1}=\prod_{j} O_{j+1}$
and $T\ket{\psi_i'(k')}=e^{-i\frac{2\pi k'}{N}}\ket{\psi_i'(k')}$.
Thus the change in momentum depends solely on the expectation value of
the operator $P_y=\prod_{j=1}^{N} \sigma^y_{j}$ which in the following
we will call the \emph{parity} operator. This naming convention
makes sense since $P_y=i^N\exp(i \pi S_T^y)$ where
$S_T^y=\sum_{j=1}^N S_j^y$ thus $P_y$ measures the parity of the
total spin along the $y$-direction. Note that due to the factor $i^N$
the meaning of positive and negative parity is interchanged for chains
with $N=0(\mod{4})$ and chains with $N=2(\mod{4})$.
The parity is a good quantum number for both $H_{HB}$ and $H_{HB}'$
so there exist eigenstates $\ket{\psi_i'(k')}$ that have well defined
parity plus or minus one. If $\braket{P_y}_{i',k'}=+1$ the momentum
remains unchanged i.e. $k=k'$, if
$\braket{P_y}_{i',k'}=-1=e^{\pm i\pi}$ we have
$k=k'\oplus_N N/2$
where $\oplus_N$ denotes addition modulo $N$.
Note that the parity itself is invariant under the mapping between
$H_{HB}$ and $H_{HB}'$ since $U^{\dagger}P_y U=P_y$.
\begin{figure*}[ht]
\begin{center}
\includegraphics[width=1.0\textwidth]{HB_Ei_reldiffEi_N16_D32.pdf}
\end{center}
\caption{
(Color online).
Results for the low excitation spectrum (left) and the corresponding
relative precision (right) for the Heisenberg spin-$1/2$ chain with
$N=16$ sites.
}
\label{fig:HB_Ei_reldiffEi_N16}
\end{figure*}
\begin{figure}[ht]
\begin{center}
\includegraphics[width=1.0\columnwidth]{HB_N16_canonicalANGLES.pdf}
\end{center}
\caption{
(Color online).
Distance between several degenerated subspaces obtained by our algorithm
to the corresponding degenerated subspaces obtained by exact
diagonalization. As a measure for the distance we have used the sine
of the canonical angle with the largest magnitude as defined
in~\cite{book-bhatia-1997}.
}
\label{fig:HB_N16_canonicalANGLES}
\end{figure}
Now the generators of the $SU(2)$ symmetry for $H_{HB}'$ do not
commute with the translation operator thus we cannot classify
the momentum eigenstates in terms of irreducible representations
of $SU(2)$. For $H_{HB}$ however we can do this, so we
know exactly the degeneracy structure of the spectrum in each
subspace with fixed momentum.
Thus if we encounter for instance a threefold degenerated eigenstate of
$H_{HB}'$, we know this is mapped to a spin triplet with
well defined momentum in the original Hamiltonian.
Accordingly it must contain a two-dimensional subspace with
negative parity corresponding to total spin along the $y$-direction
$\pm 1$ and a one-dimensional subspace corresponding to total spin $0$.
Since the spin triplet in the original Hamiltonian has well defined
momentum, according to the rules for the mapping $k\leftrightarrow k'$, we
will have one eigenstate of $H_{HB}'$ with momentum $k$ and a
two-dimensional subspace with the same energy but different momentum
$k'=k\ominus_N N/2$.
In this way
\footnote{
A singlet state would have parity $+1$ and thus there would be no
momentum shift in this case. A quintuplet would contain a three-dimensional
subspace with parity $+1$ and a two-dimensional subspace with parity
$-1$. Thus in this case we would observe three states with no momentum
shift and two states with a $\pi$-shift. The generalization
to higher multiplets is obvious.
},
after approximating the spectrum of $H_{HB}'$ and labeling
all energies with the corresponding momentum we can obtain
the spectrum of $H_{HB}$ by mere inspection of the degeneracy structure.
Table~\ref{tab:HB_N16_spectrum_multiplets_parity} and
Table~\ref{tab:HBtransf_N16_spectrum_multiplets_parity} illustrate
how the multiplets of $H_{HB}$ and $H_{HB}'$ are related to eachother.
\begin{figure*}[ht]
\begin{center}
\includegraphics[width=1.0\textwidth]{HB_Ei_reldiffEi_N50_N100_D32.pdf}
\end{center}
\caption{
(Color online).
Results for the low excitation spectrum and the corresponding
relative precision of the lowest triplet (insets) for the Heisenberg
spin-$1/2$ chain with $N=50$ (left) and $N=100$ (right) sites.
}
\label{fig:HB_Ei_reldiffEi_N50_N100_D32}
\end{figure*}
This procedure works very well for the lower branches of the
dispersion relation where the precision of our simulation is good
enough to discriminate unambigously between different multiplets.
For higher branches, on one hand the precision gets worse and
on the other hand the density of states increases such that
multiplets with similar energy become effectively undistiguishable
for our algorithm. In this case the eigenstates with well defined
momentum that we obtain for the Hamiltonian $H_{HB}'$ do not have
well defined parity i.e. they mix parity eigenstates with different
parity. Since states with same momentum and different parity are mapped by
(\ref{eq:eigenstate_transf}) to states with different momentum,
if we start with such a momentum eigenstate we obtain after the
transformation a superposition of states with different momenta which
is clearly not a momentum eigenstate.
There are however two ways to overcome this issue and obtain approximations
of the eigenstates of $H_{HB}$ that are at the same time exact momentum
eigenstates.
The \emph{first} one amounts to computing the matrix elements
of the translation operator $T$ in the subspace spanned by the transformed
states $\{M_{odd}^{y} \ket{\psi_i'(k)'}\}$ where
$M_{odd}^{y}:=\prod_{j=1}^{N/2} \sigma_{2j-1}^{y}$ and
then diagonalize this matrix.
It is not difficult to check that this can be done for each momentum $k'$
separately since
$M_{odd}^{y}\,T\,M_{odd}^{y}=M_{odd}^{y}\,M_{even}^{y}\,T=P_y T$ which
is a translationally invariant operator and thus it does not mix
states with different momentum.
Diagonalizing each of the
$T(k')_{ij}=U^{\dagger}(k')_{il}D(k')_{lm} U(k')_{mj}$
yields for each $k'$ a unitary $U(k')$ that is nothing more than
the transformation that we need to obtain the desired momentum eigenstates
via $\ket{\psi_i(k_i)}= U(k')_{ij} \ket{\psi_j'(k')}$.
The new momentum $k_i$ can be read off the diagonal matrix $D(k')$.
There are two drawbacks that come with this procedure. The first one is
that we must compute the matrix elements $T(k')_{ij}$ each of which is
done with the computational cost $O(N D^5)$. Since there are
$N b^2$ of these where $b$ is the number of branches, we obtain
the overall cost $O(N^2 b^2 D^5)$. Usually we compute enough
branches such that $b^2>D$ holds, thus the cost for this procedure
ends up being higher than the one for the diagonalization of $H_{HB}'$.
The second drawback is that the superpositions
$U(k')_{ij} \ket{\psi_j'(k')}$ mix the original approximations
of the energy levels thereby
slightly lowering the energy of higher excitations but increasing the
energy of lower excitations, which are usually the ones we are most
interested in.
The \emph{second} way to approximate the eigenstates of the
original Hamiltonian $H_{HB}$ such that they are at the same time
exact momentum eigenstates is to add to $H_{HB}'$ a perturbation
that splits degenerated levels with different parity.
This is easily achieved by taking $H_{HB}^{\pm}:=H_{HB}'\pm \lambda P_y$
where $\lambda$ must be chosen such that it is big enough for our
algorithm to deliver only states with a single parity, but as small
as possible in order to avoid numerical inaccuracies caused by
altering the Hamiltonian. In the case of the Heisenberg model, if we
choose to compute $b=10$ branches, $\lambda=0.1\cdot N$ fulfills these
requirements. In practice we first apply our algorithm to
$H_{HB}^{-}$ which yields for each momentum $k'$ $b$ states
with positive parity. These states do not change their momentum
under the transformation (\ref{eq:eigenstate_transf}).
Subsequently we apply the algorithm to $H_{HB}^{+}$ which yields
states with negative parity that change their momentum after
the transformation according to $k=k'\oplus_N N/2$.
In this way we end up with $2b$ branches of states that approximate
the spectrum of $H_{HB}$ and that are at the same time exact momentum
eigenstates. The computational cost per state is thus exactly the
same like diagonalizing only $H_{HB}'$.
Let us first look at the results we have obtained for a small chain with
$16$ sites. We have chosen to look at such a small system first for two
reasons: \emph{First}, even though the Heisenberg model is exactly
solvable via Bethe ansatz, obtaining \emph{all} energy levels can
be quite involved. Choosing $N$ as small as $16$ allows us to
compute the spectrum of this small chain by means of exact diagonalization.
\emph{Second}, even for the energy levels that are easily computable
with the Bethe ansatz (i.e., the triplet states in the subspace of
two-spinon excitations~\cite{INTRO2BA-II}) it is not possible to obtain
the eigenstates themselves. Exact diagonalization of a small chain
on the other hand allows us to compute and store the exact eigenstates
in order to check the fidelity of our MPS approximation.
Figure~\ref{fig:HB_Ei_reldiffEi_N16} shows the energy of the first ten
branches of excitations and the corresponding relative precision.
Note how states belonging to the same multiplet have very similar precision
even though they have different parity and thus correspond to eigenstates
of $H_{HB}'$ with different momentum. Since there are no one-particle
excitations in the low-energy spectrum of the AF Heisenberg model,
we do not obtain such a good precision like in the case of the
quantum Ising model. Nevertheless we get a very good approximation
of the first excited level, namely the triplet excitation at $k=N/2$.
We have also tested the accuracy of the states themselves:
for non-degenerated states, the absolute value of the overlap is a
perfect measure for this, and for reasons that will become clear
immediately, we have looked at the sine of the fidelity.
For degenerated states, in order to compare the
subspace spanned by our MPS to the one spanned by the exact eigenstates,
we have used as a measure for the distance the definition given in chapter
7 of~\cite{book-bhatia-1997}: the sine of the largest canonical angle
between the two subspaces. The canonical angles can be easily computed
from the matrix that has as its entries the overlaps between all states
of the subspaces that we want to compare. The results are plotted in
figure~\ref{fig:HB_N16_canonicalANGLES}. We see that only the MPS
with momentum $k=0$ and $k=N/2$ are extremely accurate. All other
states, especially those around $k=N/4$, are much further away from
the exact solutions, which is a bit surprising given the fact that
the energy precision for these states is comparable to the one obtained
for $k=0$.
The spectrum that we obtain for longer chains is plotted in
figure~\ref{fig:HB_Ei_reldiffEi_N50_N100_D32}. In this
regime we have only looked at the precision of the lowest two-spinon
triplet for which the exact results were obtained following
\cite{INTRO2BA-II}. Again we see that the states at momentum
$k=k_0\oplus N/2$ have the best accuracy. We would like to make
two more remarks concerning the chain with $N=50$. First, note
that the ground state has momentum $k_0=N/2$ in this case.
Second, unlike in the case
of $N=100$, where for all momenta $k\neq k_0$ the lowest excitation is
a triplet, we observe that for $N=50$ this does not happen.
Our simulations reveal that at $k\in\{2,3,47,48\}$ the quintuplet excitation
lies below the triplet while at $k\in\{23,27\}$ it is a singlet
that is the lowest lying excitation.
\section{Conclusions and Outlook}
\label{sec:conclusions}
Inspired by previous approaches~\cite{rommer-ostlund-1995,porras-2006}
we have introduced a new method for the simulation of translationally
invariant spin chains with periodic boundary conditions. We have used
an MPS based ansatz that corresponds to a particle-like excitation
with well defined momentum in order to obtain extremely accurate results
for models where the spectrum contains precisely one-particle states.
For states that can be expressed in terms of many quasi-particle
excitations, we still obtain feasible results if the MPS bond dimension
is chosen to be big enough. In the case of the quantum Ising model,
our results indicate that for $g<1$ the spectrum is built up entirely
out of excitations with an even number of quasi-particles.
Generalizations of our approach can go in two directions:
First, it is possible to adjust ansatz~(\ref{eq:ansatz}) in order to
treat infinite systems with open boundary condition, which
we are addressing in~\cite{juthoEXC-2011}.
Second, it seems feasible to generalize our approach to a many-particle
ansatz by using more than one MPS tensor in~(\ref{eq:ansatz}) in order
to define the variational manifold.
\section{Acknowledgements}
We thank G.Vidal, V. Murg, E. Rico and B. Nachtergaele for valuable
discussions. This work was supported by the
FWF doctoral program Complex Quantum Systems (W1210),
the Research Foundation Flanders, the ERC grant QUERG,
and the FWF SFB grants FoQuS and ViCoM.
\begin{mcitethebibliography}{16}
\expandafter\ifx\csname natexlab\endcsname\relax\def\natexlab#1{#1}\fi
\expandafter\ifx\csname bibnamefont\endcsname\relax
\def\bibnamefont#1{#1}\fi
\expandafter\ifx\csname bibfnamefont\endcsname\relax
\def\bibfnamefont#1{#1}\fi
\expandafter\ifx\csname citenamefont\endcsname\relax
\def\citenamefont#1{#1}\fi
\expandafter\ifx\csname url\endcsname\relax
\def\url#1{\texttt{#1}}\fi
\expandafter\ifx\csname urlprefix\endcsname\relax\def\urlprefix{URL }\fi
\providecommand{\bibinfo}[2]{#2}
\providecommand{\eprint}[2][]{\url{#2}}
\bibitem[{\citenamefont{\"Ostlund and Rommer}(1995)}]{rommer-ostlund-1995}
\bibinfo{author}{\bibfnamefont{S.}~\bibnamefont{\"Ostlund}} \bibnamefont{and}
\bibinfo{author}{\bibfnamefont{S.}~\bibnamefont{Rommer}},
\bibinfo{journal}{Phys. Rev. Lett.} \textbf{\bibinfo{volume}{75}},
\bibinfo{pages}{3537} (\bibinfo{year}{1995})\relax
\mciteBstWouldAddEndPuncttrue
\mciteSetBstMidEndSepPunct{\mcitedefaultmidpunct}
{\mcitedefaultendpunct}{\mcitedefaultseppunct}\relax
\EndOfBibitem
\bibitem[{\citenamefont{Rommer and \"Ostlund}(1997)}]{rommer-ostlund-1997}
\bibinfo{author}{\bibfnamefont{S.}~\bibnamefont{Rommer}} \bibnamefont{and}
\bibinfo{author}{\bibfnamefont{S.}~\bibnamefont{\"Ostlund}},
\bibinfo{journal}{Phys. Rev. B} \textbf{\bibinfo{volume}{55}},
\bibinfo{pages}{2164} (\bibinfo{year}{1997})\relax
\mciteBstWouldAddEndPuncttrue
\mciteSetBstMidEndSepPunct{\mcitedefaultmidpunct}
{\mcitedefaultendpunct}{\mcitedefaultseppunct}\relax
\EndOfBibitem
\bibitem[{\citenamefont{Pirvu et~al.}(2010{\natexlab{a}})\citenamefont{Pirvu,
Verstraete, and Vidal}}]{me-2010-PBCI}
\bibinfo{author}{\bibfnamefont{B.}~\bibnamefont{Pirvu}},
\bibinfo{author}{\bibfnamefont{F.}~\bibnamefont{Verstraete}},
\bibnamefont{and} \bibinfo{author}{\bibfnamefont{G.}~\bibnamefont{Vidal}}
(\bibinfo{year}{2010}{\natexlab{a}}), \eprint{arXiv:1005.5195}\relax
\mciteBstWouldAddEndPuncttrue
\mciteSetBstMidEndSepPunct{\mcitedefaultmidpunct}
{\mcitedefaultendpunct}{\mcitedefaultseppunct}\relax
\EndOfBibitem
\bibitem[{\citenamefont{Porras et~al.}(2006)\citenamefont{Porras, Verstraete,
and Cirac}}]{porras-2006}
\bibinfo{author}{\bibfnamefont{D.}~\bibnamefont{Porras}},
\bibinfo{author}{\bibfnamefont{F.}~\bibnamefont{Verstraete}},
\bibnamefont{and} \bibinfo{author}{\bibfnamefont{J.~I.} \bibnamefont{Cirac}},
\bibinfo{journal}{Phys. Rev. B} \textbf{\bibinfo{volume}{73}},
\bibinfo{pages}{014410} (\bibinfo{year}{2006})\relax
\mciteBstWouldAddEndPuncttrue
\mciteSetBstMidEndSepPunct{\mcitedefaultmidpunct}
{\mcitedefaultendpunct}{\mcitedefaultseppunct}\relax
\EndOfBibitem
\bibitem[{\citenamefont{Verstraete et~al.}(2004)\citenamefont{Verstraete,
Porras, and Cirac}}]{frank-2004a}
\bibinfo{author}{\bibfnamefont{F.}~\bibnamefont{Verstraete}},
\bibinfo{author}{\bibfnamefont{D.}~\bibnamefont{Porras}}, \bibnamefont{and}
\bibinfo{author}{\bibfnamefont{J.~I.} \bibnamefont{Cirac}},
\bibinfo{journal}{Phys. Rev. Lett.} \textbf{\bibinfo{volume}{93}},
\bibinfo{pages}{227205} (\bibinfo{year}{2004})\relax
\mciteBstWouldAddEndPuncttrue
\mciteSetBstMidEndSepPunct{\mcitedefaultmidpunct}
{\mcitedefaultendpunct}{\mcitedefaultseppunct}\relax
\EndOfBibitem
\bibitem[{\citenamefont{Fannes et~al.}(1992)\citenamefont{Fannes, Nachtergaele,
and Werner}}]{fannes-1992}
\bibinfo{author}{\bibfnamefont{M.}~\bibnamefont{Fannes}},
\bibinfo{author}{\bibfnamefont{B.}~\bibnamefont{Nachtergaele}},
\bibnamefont{and} \bibinfo{author}{\bibfnamefont{R.}~\bibnamefont{Werner}},
\bibinfo{journal}{Communications in Mathematical Physics}
\textbf{\bibinfo{volume}{144}}, \bibinfo{pages}{443} (\bibinfo{year}{1992}),
ISSN \bibinfo{issn}{0010-3616}, \bibinfo{note}{10.1007/BF02099178},
\urlprefix\url{http://dx.doi.org/10.1007/BF02099178}\relax
\mciteBstWouldAddEndPuncttrue
\mciteSetBstMidEndSepPunct{\mcitedefaultmidpunct}
{\mcitedefaultendpunct}{\mcitedefaultseppunct}\relax
\EndOfBibitem
\bibitem[{\citenamefont{Chung and Wang}(2009)}]{chung-2009}
\bibinfo{author}{\bibfnamefont{S.~G.} \bibnamefont{Chung}} \bibnamefont{and}
\bibinfo{author}{\bibfnamefont{L.}~\bibnamefont{Wang}},
\bibinfo{journal}{Physics Letters A} \textbf{\bibinfo{volume}{373}},
\bibinfo{pages}{2277 } (\bibinfo{year}{2009}), ISSN
\bibinfo{issn}{0375-9601}\relax
\mciteBstWouldAddEndPuncttrue
\mciteSetBstMidEndSepPunct{\mcitedefaultmidpunct}
{\mcitedefaultendpunct}{\mcitedefaultseppunct}\relax
\EndOfBibitem
\bibitem[{\citenamefont{Haegeman et~al.}(2011)\citenamefont{Haegeman, Pirvu,
Weir, Cirac, Osborne, Verschelde, and Verstraete}}]{juthoEXC-2011}
\bibinfo{author}{\bibfnamefont{J.}~\bibnamefont{Haegeman}},
\bibinfo{author}{\bibfnamefont{B.}~\bibnamefont{Pirvu}},
\bibinfo{author}{\bibfnamefont{D.}~\bibnamefont{Weir}},
\bibinfo{author}{\bibfnamefont{J.}~\bibnamefont{Cirac}},
\bibinfo{author}{\bibfnamefont{T.}~\bibnamefont{Osborne}},
\bibinfo{author}{\bibfnamefont{H.}~\bibnamefont{Verschelde}},
\bibnamefont{and}
\bibinfo{author}{\bibfnamefont{F.}~\bibnamefont{Verstraete}}
(\bibinfo{year}{2011}), \eprint{arXiv:1103.2286}\relax
\mciteBstWouldAddEndPuncttrue
\mciteSetBstMidEndSepPunct{\mcitedefaultmidpunct}
{\mcitedefaultendpunct}{\mcitedefaultseppunct}\relax
\EndOfBibitem
\bibitem[{\citenamefont{Karbach et~al.}(1998)\citenamefont{Karbach, Hu, and
Gerhard}}]{INTRO2BA-II}
\bibinfo{author}{\bibfnamefont{M.}~\bibnamefont{Karbach}},
\bibinfo{author}{\bibfnamefont{K.}~\bibnamefont{Hu}}, \bibnamefont{and}
\bibinfo{author}{\bibfnamefont{M.}~\bibnamefont{Gerhard}},
\bibinfo{journal}{Comp. in Phys.} \textbf{\bibinfo{volume}{12}},
\bibinfo{pages}{565} (\bibinfo{year}{1998}),
\eprint{arXiv:cond-mat/9809163}\relax
\mciteBstWouldAddEndPuncttrue
\mciteSetBstMidEndSepPunct{\mcitedefaultmidpunct}
{\mcitedefaultendpunct}{\mcitedefaultseppunct}\relax
\EndOfBibitem
\bibitem[{\citenamefont{Lieb et~al.}(1961)\citenamefont{Lieb, Schultz, and
Mattis}}]{lieb-schultz-mattis-1961}
\bibinfo{author}{\bibfnamefont{E.}~\bibnamefont{Lieb}},
\bibinfo{author}{\bibfnamefont{T.}~\bibnamefont{Schultz}}, \bibnamefont{and}
\bibinfo{author}{\bibfnamefont{D.}~\bibnamefont{Mattis}},
\bibinfo{journal}{Annals of Physics} \textbf{\bibinfo{volume}{16}},
\bibinfo{pages}{407 } (\bibinfo{year}{1961})\relax
\mciteBstWouldAddEndPuncttrue
\mciteSetBstMidEndSepPunct{\mcitedefaultmidpunct}
{\mcitedefaultendpunct}{\mcitedefaultseppunct}\relax
\EndOfBibitem
\bibitem[{\citenamefont{Pfeuty}(1970)}]{pfeuty-1970}
\bibinfo{author}{\bibfnamefont{P.}~\bibnamefont{Pfeuty}},
\bibinfo{journal}{Annals of Physics} \textbf{\bibinfo{volume}{57}},
\bibinfo{pages}{79 } (\bibinfo{year}{1970}), ISSN
\bibinfo{issn}{0003-4916}\relax
\mciteBstWouldAddEndPuncttrue
\mciteSetBstMidEndSepPunct{\mcitedefaultmidpunct}
{\mcitedefaultendpunct}{\mcitedefaultseppunct}\relax
\EndOfBibitem
\bibitem[{Not({\natexlab{a}})}]{Note1}
\bibinfo{note}{If the fermionic vacuum is not the lowest energy state there
also exist hole and particle-hole excitations. If we choose the vaccum in
such a way that there exists exactly one hole-mode we effecively switch the
sign of the parity operator opposed to choosing no hole-modes at all.}\relax
\mciteBstWouldAddEndPunctfalse
\mciteSetBstMidEndSepPunct{\mcitedefaultmidpunct}
{}{\mcitedefaultseppunct}\relax
\EndOfBibitem
\bibitem[{Not({\natexlab{b}})}]{Note2}
\bibinfo{note}{Actually at $g=1$ the zero momentum mode has energy zero so the
states $\mathinner {|{\Omega }\delimiter "526930B }_{odd}$ and $\mathinner
{|{0}\delimiter "526930B }$ have exactly the same energy. However this
happens only at the critical point $g=1$. In general $\mathinner {|{\Omega
}\delimiter "526930B }_{odd}$ and $\mathinner {|{0}\delimiter "526930B }$
have different energy.}\relax
\mciteBstWouldAddEndPunctfalse
\mciteSetBstMidEndSepPunct{\mcitedefaultmidpunct}
{}{\mcitedefaultseppunct}\relax
\EndOfBibitem
\bibitem[{\citenamefont{Pirvu et~al.}(2010{\natexlab{b}})\citenamefont{Pirvu,
Murg, Cirac, and Verstraete}}]{me-2010-MPOR}
\bibinfo{author}{\bibfnamefont{B.}~\bibnamefont{Pirvu}},
\bibinfo{author}{\bibfnamefont{V.}~\bibnamefont{Murg}},
\bibinfo{author}{\bibfnamefont{J.~I.} \bibnamefont{Cirac}}, \bibnamefont{and}
\bibinfo{author}{\bibfnamefont{F.}~\bibnamefont{Verstraete}},
\bibinfo{journal}{New J. Phys.} \textbf{\bibinfo{volume}{12}},
\bibinfo{pages}{025012} (\bibinfo{year}{2010}{\natexlab{b}})\relax
\mciteBstWouldAddEndPuncttrue
\mciteSetBstMidEndSepPunct{\mcitedefaultmidpunct}
{\mcitedefaultendpunct}{\mcitedefaultseppunct}\relax
\EndOfBibitem
\bibitem[{\citenamefont{Bhatia}(1997)}]{book-bhatia-1997}
\bibinfo{author}{\bibfnamefont{R.}~\bibnamefont{Bhatia}},
\emph{\bibinfo{title}{Matrix analysis}} (\bibinfo{publisher}{Springer-Verlag
New York}, \bibinfo{year}{1997})\relax
\mciteBstWouldAddEndPuncttrue
\mciteSetBstMidEndSepPunct{\mcitedefaultmidpunct}
{\mcitedefaultendpunct}{\mcitedefaultseppunct}\relax
\EndOfBibitem
\bibitem[{Not({\natexlab{c}})}]{Note3}
\bibinfo{note}{A singlet state would have parity $+1$ and thus there would be
no momentum shift in this case. A quintuplet would contain a
three-dimensional subspace with parity $+1$ and a two-dimensional subspace
with parity $-1$. Thus in this case we would observe three states with no
momentum shift and two states with a $\pi $-shift. The generalization to
higher multiplets is obvious.}\relax
\mciteBstWouldAddEndPunctfalse
\mciteSetBstMidEndSepPunct{\mcitedefaultmidpunct}
{}{\mcitedefaultseppunct}\relax
\EndOfBibitem
\end{mcitethebibliography}
\end{document}
|
\section{Introduction}
Matrix theory is expected to be a correct description of M-theory, and is given as
an U($N$) matrix quantum mechanics. The structure of the spectrum of states of this theory is
complicated, and we know little about it except that the spectrum is continuous with some discrete poles.
Instead, in this paper we investigate $D=2$ SU($N$) supersymmetric Yang-Mills matrix quantum mechanics,
whose action is given by the dimensional
reduction of $D=2=1+1$ super Yang-Mills theory to $0+1$ dimension. As a simplified model of Matrix theory,
which is from $D=10$ super Yang-Mills theory, we expect that it gives some hint to the spectrum of
Matrix theory. This quantum mechanics is almost free, but Gauss law associated with the gauge fixing procedure
makes it nontrivial.
The action of $D=2$ matrix quantum mechanics is given by
\bea
S & = & \frac{1}{g^2}\int dt\;\text{tr}\Big[
\frac{1}{2}DXDX+ig^2\theta^\dagger D\theta-g^4\theta^\dagger[X,\theta]
\Big],
\eea
where $g$ is the coupling constant, $DX=\p_tX-i[A_0,X]$, and $X, A_0, \theta$ are matrices
in {\it su}($N$) Lie algebra. $\theta$ is also a complex Grassmann odd operator.
This action is invariant under the gauge transformation, and the following supersymmetry transformation:
\beq
\delta X=i(\epsilon^\dagger\theta+\epsilon\theta^\dagger),\quad
\delta\theta=-\Pi\epsilon,\quad
\delta A_0=g^2\delta X.
\eeq
Quantization of this system gives the following commutation relations:
\beq
[X^a,\Pi^b]=i\delta_{ab},\quad
\{\theta^a,(\theta^b)^\dagger\}=\delta_{ab},
\eeq
where $\Pi=\frac{1}{g^2}DX$ and $a$ is an index of SU($N$) adjoint representation.
Physical states must be gauge invariant: $G^a\ket{\text{phys}}=0$, where $G$ is the generator of SU($N$) :
\beq
G= i[X,\Pi]+\{\theta^\dagger,\theta\}.
\eeq
Hamiltonian $H$ and supercharge $Q$ are
\bea
H & = & g^2\text{tr}\;\Big[\frac{1}{2}\Pi^2\Big] + (\text{terms proportional to 1st class constraints}), \\
Q & = & -\text{tr}\;[\Pi\theta].
\eea
On physical states $H$ acts as free Hamiltonian $h=g^2\text{tr}\;[\frac{1}{2}\Pi^2]$.
However, this system is not trivial, because of the physical condition $G\ket{\text{phys}}=0$.
Commutation relations of $H$ and $Q$ are usual ones, up to operators which vanish on physical states:
\bea
\{Q^\dagger,Q\} & = & \frac{2}{g^2}h+(\text{terms proportional to 1st class constraints}),
\\
\{Q,H\} & = & (\text{terms proportional to 1st class constraints}),
\\
\{Q^\dagger,H\} & = & (\text{terms proportional to 1st class constraints}).
\eea
The fermion part of the states is constructed by acting $(\theta^a)^\dagger$ on a vacuum
$\ket{0_F}$ satisfying $\theta^a\ket{0_F}=0$. Then the total states take the following form:
\beq
(\theta^{a_1})^\dagger\dots(\theta^{a_n})^\dagger\ket{0_F}
\int\prod_adX^a\left[\prod_a\ket{X^a}\right]\psi(X).
\eeq
A gauge invariant solution to Schr\"odinger equation in the sector of zero fermion number
has been found in \cite{ch85,s07}: When $X$ is diagonalized by the gauge transformation:
\beq
X=U
\begin{pmatrix}
x_1 & & & \\
& x_2 & & \\
& & \ddots & \\
& & & x_N
\end{pmatrix} U^{-1},
\eeq
then Claudson-Halpern-Samuel (CHS) solution, which has $N$ continuous parameters $k_i$, is given by
\beq
\psi(X)=\frac{1}{\cal M}\sum_{\sigma}\text{sgn}(\sigma)\exp\left[i\sum_{i=1}^Nk_{\sigma(i)}x_i\right],
\label{chss}
\eeq
where ${\cal M}=\prod_{i<j}(x_i-x_j)$ and $\sum_\sigma$ indicates summation over all the permutations.
This expression gives a solution in the case of U($N$) gauge group. To remove center of mass U(1) part
one just have to impose the condition $\sum_ik_i=0$. Then $\psi(X)$ depends on only traceless part
$w_i\equiv x_i-\frac{1}{N}\sum_jx_j$.
In SU(2) case this is the only energy eigenfunction with zero fermion number\cite{ch85,s07}.
In the case of higher groups there are more solutions \cite{trze06,k0910}.
Naively we can expect that the spectrum can be constructed by acting gauge invariant operators made of
$\Pi$ and $\theta^\dagger$ on states in the sector of zero fermion number\cite{s07}. Indeed in $N=2$ case
it has been proven that the whole spectrum is constructed in this way on CHS solution\cite{ch85,s07}.
To investigate the case of general $N$, in section 2 we compute the thermal partition function of this system,
and give an interpretation of the result. The form of the partition function gives support to
the naive expectation, and we give a conjecture that CHS solution is the unique eigenfunction of
simultaneously diagonalizable operators $\text{tr}\;[\Pi^n]~(n=2,3,\dots, N)$.
In section 3, we prove this conjecture in $N=3$ and $N=4$ cases.
Section 4 contains some discussion, and
Appendix contains necessary information on SU($N$) group theory.
\section{Thermal partition function and spectrum of $D=2$ matrix quantum mechanics}
In this section we compute the thermal partition function of $D=2$ quantum mechanics
$Z=\text{Tr}_{\text{phys}}\;[e^{-\beta H}]$, where the trace $\text{Tr}_{\text{phys}}$ is
taken over the physical states. Then from it we will obtain insight to the structure of the spectrum of this system.
First we compute $Z$ in the operator formalism.
$Z$ can be rewritten in terms of a trace without the condition of gauge invariance, by inserting
SU($N$) group element $e^{i\lambda^aG^a}$:
\beq
Z=\int d\lambda \text{Tr}\;[e^{-\beta h+i\lambda^aG^a}],
\eeq
where $\int d\lambda$ indicates integral over SU($N$) (See Appendix for notation about SU($N$) group theory used
here and the following),
and works as a projector onto the gauge invariant states.
We can set $\lambda^{(ij)}=0$ without loss of generality.
Since in $-\beta h+i\lambda^aG^a$ boson part and fermion part are decoupled,
$Z$ is decomposed into corresponding parts $Z_B$ and $Z_F$:
\beq
Z=\int d\lambda Z_B(\lambda)Z_F(\lambda),\quad
Z_B(\lambda)=\text{Tr}\;[e^{-\beta h+i\lambda^aG^a_B}],\quad Z_F(\lambda)=\text{Tr}\;[e^{i\lambda^a G^a_F}],
\eeq
where
\beq
G^a_B=f_{bac}X^b\Pi^c,\quad
G^a_F=-if_{bac}(\theta^b)^\dagger\theta^c.
\eeq
$Z_F$ can be computed by using the knowledge of group theory:
$e^{i\lambda^a G^a_F}$ is just the gauge group element for fermions:
\beq
e^{i\lambda^aG^a_F}(\theta_b)^\dagger e^{-i\lambda^a G^a_F}
=[\text{Ad}(\lambda)]_{bc}(\theta_c)^\dagger,
\eeq
and $(\theta^{a_1})^\dagger\dots(\theta^{a_n})^\dagger\ket{0_F}$ forms an antisymmetric tensor product representation
of adjoint representation. Therefore $Z_F$ is given by the sum of group theory characters of those representations:
\beq
Z_F(\lambda)=\sum_{n=0}^{N^2-1}\chi(\text{Alt}_n(\text{adj.}))=
2^{N-1}\prod_{\stackrel{\mbox{$\scriptstyle i,j=1$}}{\mbox{$\scriptstyle i\not\neq j$}}}^N(1+z_i/z_j),
\eeq
where we used \eqref{chaltadj}.
If we introduce an additional parameter $q$ for counting fermion number $F$:
$Z_F(\lambda,q)\equiv\text{Tr}\;[q^Fe^{i\lambda^a G^a_F}]$, then similarly,
\beq
Z_F(\lambda,q)=\sum_{n=0}^{N^2-1}q^n\chi(\text{Alt}_n(\text{adj.}))=(1+q)^{N-1}
\prod_{\stackrel{\mbox{$\scriptstyle i,j=1$}}{\mbox{$\scriptstyle i\not\neq j$}}}^N(1+qz_i/z_j).
\eeq
Next let us compute $Z_B(\lambda)$. By introducing eigenstates of $\Pi$: $\Pi^a\ket{p_b}=p^a\ket{p_b}$,
\bea
Z_B(\lambda) & = & \Big[\prod_a\int_{-\infty}^\infty \frac{dp_a}{2\pi}\Big]
\bra{p_b}e^{-\beta h}e^{i\lambda^c G^c_B}\ket{p_b}
\nn & = &
\Big[\prod_a\int_{-\infty}^\infty\frac{dp_a}{2\pi}\Big]e^{-\frac{\beta g^2}{2}p_d^2}
\bra{p_b}e^{i\lambda^c G^c_B}\ket{p_b}.
\eea
Noting that $e^{i\lambda^c G^c_B}$ generates gauge group elements,
\beq
\bra{p_b}e^{i\lambda^c G^c_B}\ket{p_b}=\vev{p_b|\text{Ad}(\lambda)_{bc}p_c}
=\prod_b2\pi\delta(p_b-\text{Ad}(\lambda)_{bc}p_c).
\eeq
Since $\text{Ad}(\lambda)_{bc}$ is diagonalized by decomposing adjoint indices into Cartan part $m$ and
ladder operator part $(ij)$,
\bea
\bra{p_b}e^{i\lambda^c G^c_B}\ket{p_b} & = &
(2\pi\delta(0))^{N-1}\prod_{\stackrel{\mbox{$\scriptstyle i,j=1$}}{\mbox{$\scriptstyle i\not\neq j$}}}^N
2\pi\delta([1-z_i/z_j]p_{(ij)})
\nn & = & V^{N-1}(2\pi)^{N(N-1)}
\prod_{\stackrel{\mbox{$\scriptstyle i,j=1$}}{\mbox{$\scriptstyle i\not\neq j$}}}^N(1-z_i/z_j)^{-1}\delta(p_{(ij)}),
\eea
where $V$ is the volume factor $V=2\pi\delta(p=0)$. Therefore
\bea
Z_B(\lambda) & = & \prod_{m=1}^{N-1}\Big[\int_{-\infty}^\infty \frac{dp_m}{2\pi}Ve^{-\frac{\beta g^2}{2}p_m^2}\Big]
\prod_{\stackrel{\mbox{$\scriptstyle i,j=1$}}{\mbox{$\scriptstyle i\not\neq j$}}}^N(1-z_i/z_j)^{-1}
\nn & = &
\left(\frac{V}{\sqrt{2\pi\beta g^2}}\right)^{N-1}
\prod_{\stackrel{\mbox{$\scriptstyle i,j=1$}}{\mbox{$\scriptstyle i\not\neq j$}}}^N(1-z_i/z_j)^{-1}.
\eea
Then $Z$ is given by
\bea
Z & = & \frac{1}{N!}\prod_{i=1}^{N-1}\int_0^1 dy_i
\prod_{\stackrel{\mbox{$\scriptstyle i,j=1$}}{\mbox{$\scriptstyle i\not\neq j$}}}^N(z_i-z_j)
\nn & &
\times 2^{N-1}\prod_{\stackrel{\mbox{$\scriptstyle i,j=1$}}{\mbox{$\scriptstyle i\not\neq j$}}}^N(1+z_i/z_j)
\left(\frac{V}{\sqrt{2\pi\beta g^2}}\right)^{N-1}
\prod_{\stackrel{\mbox{$\scriptstyle i,j=1$}}{\mbox{$\scriptstyle i\not\neq j$}}}^N(1-z_i/z_j)^{-1}
\nn & = &
\frac{1}{N!}\left(\frac{2V}{\sqrt{2\pi\beta g^2}}\right)^{N-1}
\prod_{i=1}^{N-1}\int_0^1 dy_i \Bigg[
\prod_{\stackrel{\mbox{$\scriptstyle i,j=1$}}{\mbox{$\scriptstyle i\not\neq j$}}}^N(1+z_i/z_j)\Bigg].
\eea
Note that the factor
$\prod_{i\not\neq j}(1-z_i/z_j)^{-1}$
from $Z_B$ is canceled with the measure factor from $\int d\lambda$.
To confirm the above result from different viewpoint, we will quickly explain the path integral calculation of $Z$
(For similar calculation in a different context, see e.g. \cite{yin03}).
In this calculation the overall normalization of $Z$ is somewhat ambiguous, and therefore we shall not keep track of
normalization factors.
In Euclidean path integral formulation $Z$ is given by
\beq
Z=\frac{1}{\text{Vol(SU($N$))}}\int\mathop{DA_0DXD\theta D\theta^\dagger}_{\text{periodicity cond.}}~
\exp\Bigg[-\frac{1}{g^2}\int_0^\beta dt~\text{tr}\,
\Big[\frac{1}{2}DXDX+ig^2\theta^\dagger D\theta-g^4\theta^\dagger[X,\theta]\Big]\Bigg],
\eeq
where $A_0(t)$ and $X(t)$ are periodic, and $\theta(t)$ is antiperiodic up to gauge transformation,
with periodicity $\beta$.
Vol(SU($N$)) is the infinite volume of the redundancy from the gauge symmetry.
By the shift $A_0\rightarrow A_0+g^2X$, we can eliminate the interaction term $\theta^\dagger[X,\theta]$:
\beq
Z=\frac{1}{\text{Vol(SU($N$))}}\int\mathop{DA_0DXD\theta D\theta^\dagger}_{\text{periodicity cond.}}~
\exp\Bigg[-\frac{1}{g^2}\int_0^\beta dt~\text{tr}\,
\Big[\frac{1}{2}DXDX+ig^2\theta^\dagger D\theta\Big]\Bigg].
\eeq
$A_0$ can be taken to be $\Delta$ given in the following, by taking Lorentz gauge $\p_tA_0=0$:
\beq
\Delta=\begin{pmatrix} \alpha_1 & & & \\ & \alpha_2 & & \\ & & \ddots & \\ & & & \alpha_N \end{pmatrix},
\quad
\begin{matrix}
0\leq\alpha_1\leq\alpha_2\leq\dots\leq\alpha_{N-1}<2\pi/\beta, \\
\alpha_N=-(\alpha_1+\alpha_2+\dots+\alpha_{N-1}), \\
\alpha_i: t\text{-independent},
\end{matrix}
\eeq
General $A_0$ is in the form of gauge transformation of $\Delta$:
$A_0(t)=U(t)\Delta U^{-1}(t)-i\p_t U(t)U^{-1}(t)$, and
the path integral measure $DA_0$ is decomposed as follows:
\beq
DA_0=DU\Big[\prod_{i=1}^{N-1}d\alpha_i\Big]\text{Det}'[D],
\eeq
where Det${}'$ is the functional determinant for periodic functions with the zero eigenvalues removed.
Although we will see that the Faddeev-Popov determinant Det${}'[D]$ cancels, let us see how this is computed.
By decomposing the adjoint index $a$ into Cartan subalgebra part $m$ and ladder operator part $(ij)$,
\beq
\text{Det}'[D]=\prod_{m=1}^{N-1}\text{Det}'[\p_t]
\prod_{\stackrel{\mbox{$\scriptstyle i,j=1$}}{\mbox{$\scriptstyle i\not\neq j$}}}^N
\text{Det}[\p_t-i(\alpha_i-\alpha_j)].
\eeq
For periodic functions eigenvalues of $\p_t$ are given by $i\frac{2\pi}{\beta}n~(n\in\mathbb{Z})$
and
\beq
\text{Det}'[\p_t]=\prod_{n\not\neq 0}i\frac{2\pi}{\beta}n=-i\beta,
\eeq
where we used the following zeta-function regularization:
\beq
\prod_{n=1}^\infty n^\alpha=(2\pi)^{\alpha/2},\quad \prod_{n=-\infty}^\infty a=1.
\eeq
Using the above and $\prod_{n=1}^\infty(1-\frac{x^2}{n^2})=\frac{\sin\pi x}{\pi x}$,
$\prod_{i\not\neq j}\text{Det}[\p_t-i(\alpha_i-\alpha_j)]$ can be computed similarly, and then
\beq
\text{Det}'[D]\sim
\prod_{\stackrel{\mbox{$\scriptstyle i,j=1$}}{\mbox{$\scriptstyle i\not\neq j$}}}^N(z_i-z_j),
\eeq
where $z_i=e^{2\pi iy_i}$, and $y_i=\frac{\beta}{2\pi}\alpha_i$. This is the measure factor
for the integration over SU($N$) group.
The path integral of $DU$ cancels the infinite volume factor: $\int DU=\text{Vol(SU($N$))}$.
Since $\alpha_i$ appear symmetrically, the integral of $\alpha_i$ can be done by regarding
all the $\alpha_i$ independent from each other, and dividing by the factor $1/N!$:
\beq
\frac{1}{N!}\prod_{i=1}^{N-1}\int_0^1dy_i.
\eeq
After the gauge fixing $X$ and $\theta$ can be taken as periodic and antiperiodic respectively.
The functional integral of them can be done as Gaussian integrals:
\bea
\int DX & \rightarrow & V^{N-1}\text{Det}'[D]^{-1},
\\
\int D\theta D\theta^\dagger & \rightarrow &
\prod_{m=1}^{N-1}\wt{\text{Det}}\Big[\p_t\Big]
\prod_{\stackrel{\mbox{$\scriptstyle i,j=1$}}{\mbox{$\scriptstyle i\not\neq j$}}}^N
\wt{\text{Det}}\Big[\p_t-i(\alpha_i-\alpha_j)\Big]
\sim
\prod_{\stackrel{\mbox{$\scriptstyle i,j=1$}}{\mbox{$\scriptstyle i\not\neq j$}}}^N(z_i+z_j),
\eea
where $\wt{\text{Det}}$ is the functional determinant for antiperiodic functions, and the volume factor $V^{N-1}$
comes from the integral of the zero modes of $X$. Then
\bea
Z & \sim & \frac{1}{N!}\prod_{i=1}^{N-1}\int_0^1dy_i~\text{Det}'[D]\cdot V^{N-1}\text{Det}'[D]^{-1}
\prod_{\stackrel{\mbox{$\scriptstyle i,j=1$}}{\mbox{$\scriptstyle i\not\neq j$}}}^N
(z_i+z_j)
\nn & = &
\frac{1}{N!}V^{N-1}\prod_{i=1}^{N-1}\int_0^1dy_i~\Bigg[
\prod_{\stackrel{\mbox{$\scriptstyle i,j=1$}}{\mbox{$\scriptstyle i\not\neq j$}}}^N(z_i+z_j)\Bigg].
\eea
Thus we can reproduce the result of operator formalism calculation, with the similar
cancellation of the measure factor.
To interpret the form of $Z$ and read off information on the spectrum,
let us consider mass deformation of this quantum mechanics given by the following action:
\bea
S & = & \frac{1}{g^2}\int dt\;\text{tr}\Big[
\frac{1}{2}DXDX-\frac{1}{2}g^4M^2X^2+ig^2\theta^\dagger D\theta-g^4M\theta^\dagger\theta
-g^4\theta^\dagger[X,\theta]
\Big].
\eea
The supersymmetry transformation is also deformed:
\beq
\delta X=i(\epsilon^\dagger\theta+\epsilon\theta^\dagger), \quad
\delta\theta=\left(iMX-\Pi\right)\epsilon, \quad
\delta A_0=g^2\delta X.
\eeq
Hamiltonian and supercharge are
\bea
H & = & g^2\text{tr}\;\Big[\frac{1}{2}\Pi^2+\frac{1}{2}M^2X^2+M\theta^\dagger\theta\Big]
\nn & &
+(\text{terms proportional to 1st class constraints}), \\
Q & = & -\text{tr}\;[(\Pi+iMX)\theta].
\eea
These satisfy the same commutation relations as those in the undeformed case.
Again $H$ acts as free Hamiltonian $h=g^2\text{tr}\;\big[
\frac{1}{2}\Pi^2+\frac{1}{2}M^2X^2+M\theta^\dagger\theta\big]$ on the physical states.
This system is easier than the undeformed case: Introducing annihilation operators
$A_a=\frac{1}{\sqrt{2M}}\Pi_a-i\sqrt{\frac{M}{2}}X_a$, $h$ is rewritten as
\beq
h=Mg^2(A^\dagger_aA_a+\theta^\dagger_a\theta_a).
\eeq
Therefore this system is equivalent to a supersymmetric harmonic oscillator system,
with the physical condition $G\ket{\text{phys}}=0$.
The spectrum of states is constructed by acting $A_a^\dagger$ and $\theta_a^\dagger$ on the vacuum
$\ket{0}$ defined by $A_a\ket{0}=0$ and $\theta_a\ket{0}=0$.
For states in the following form:
\beq
A_{a_1}^\dagger\dots A_{a_n}^\dagger\theta_{b_1}^\dagger\dots\theta_{b_m}^\dagger\ket{0},
\eeq
indices $a_1,\dots, a_n$ can be symmetrized and $b_1,\dots, b_m$ antisymmetrized, and
gauge invariant combinations of indices give physical states.
The number of such gauge invariant states can be computed by
\beq
\int d\text{SU}(N) \chi(\text{Sym}_n(\text{adj.}))\chi(\text{Alt}_m(\text{adj.})).
\eeq
Therefore the partition function of this system $Z_M=\text{Tr}_{\text{phys}}[e^{-\beta H}]$
can be computed by noting that the factor $e^{-\beta H}$ gives $q=e^{-\beta g^2 M}$ for each $A_a^\dagger$
and $\theta_a^\dagger$ on $\ket{0}$:
\bea
Z_M & = & \int d\text{SU}(N)\sum_{n=0}^\infty q^n\chi(\text{Sym}_n(\text{adj.}))
\sum_{n=0}^\infty q^n\chi(\text{Alt}_n(\text{adj.}))
\nn & = &
\frac{1}{N!}\int \prod_{i=1}^{N-1} dy_i
\Big[\prod_{\stackrel{\mbox{$\scriptstyle i,j=1$}}{\mbox{$\scriptstyle i\not\neq j$}}}^N(1-z_i/z_j)\Big]
\left(\frac{1+q}{1-q}\right)^{N-1}\prod_{\stackrel{\mbox{$\scriptstyle i,j=1$}}{\mbox{$\scriptstyle i\not\neq j$}}}^N
\frac{1+qz_i/z_j}{1-qz_i/z_j},
\label{udz}
\eea
where we used \eqref{chsymadj} and \eqref{chaltadj}. For some analysis of this expression see \cite{trze07-2}.
If we take massless limit $q\rightarrow 1$, we obtain
almost the same expression of the partition function as that in the undeformed case: the factor
$\prod_{i\not\neq j}(1-z_i/z_j)$ is canceled, and the factor $(1-q)^{1-N}$ corresponds to
the infinite volume factor $V^{N-1}$. Keeping this in mind, let us return to
the structure of the spectrum in the undeformed case and try to understand the structure of $Z$ more.
If we have an energy eigenstate $\ket{E}$ with zero fermion number,
we obtain new energy eigenstates by acting gauge invariant operators made of $\Pi$ and $\theta^\dagger$ on $\ket{E}$,
because those operators commute with the Hamiltonian\cite{s07}.
Those operators can be expressed as products of single trace operators such as
$\text{tr}\,[\Pi^n], \text{tr}\,[\Pi^n\theta^\dagger]$, or $\text{tr}\,[\Pi^n\theta^\dagger\Pi^m\theta^\dagger]$.
Note that at this stage we do not know if there are states which
cannot be constructed in this way from a solution with zero fermion number.
We can see a correspondence of states between mass deformed and undeformed systems:
\beq
A_{a_1}^\dagger\dots A_{a_n}^\dagger\theta_{b_1}^\dagger\dots\theta_{b_m}^\dagger\ket{0}
\quad \longleftrightarrow \quad
\Pi_{a_1}\dots \Pi_{a_n}\theta_{b_1}^\dagger\dots\theta_{b_m}^\dagger\ket{E}.
\eeq
Therefore number of these states in the undeformed case can be counted by setting $q=1$ in \eqref{udz}.
However, due to the factor $(1-q)^{-1}$ we cannot take the limit $q\rightarrow 1$. This is not surprising,
because there are infinitely many gauge invariant operators made of $\Pi$.
This problem can be avoided as follows:
Since $\text{tr}\,[\Pi^3], \text{tr}\,[\Pi^4], \dots$, and $\text{tr}\,[\Pi^N]$ are hermitian and commute
with each other, $h$ and $G$, these are simultaneously diagonalizable. (Note that in SU($N$) case $\text{tr}\,[\Pi^n]$
with $n>N$ can be expressed by products of those with lower $n$.) Therefore we consider an eigenstate
$\ket{p_2,p_3,\dots, p_N}$ defined by
\beq
\text{tr}\,[\Pi^n]\ket{p_2,p_3,\dots, p_N}=p_n\ket{p_2,p_3,\dots, p_N},\quad G^a\ket{p_2,p_3,\dots, p_N}=0.
\eeq
Since $\text{tr}\,[\Pi^n]$ do not produce new states when they act on $\ket{p_2,p_3,\dots, p_N}$,
we want to remove products of $\text{tr}\,[\Pi^n]$ from the list of the gauge invariant operators.
The list of those operators can be given by expanding the following expression:
\bea
& & [1+q^2\text{tr}\,[\Pi^2]+(q^2\text{tr}\,[\Pi^2])^2+(q^2\text{tr}\,[\Pi^2])^3+\dots] \nn
& & \times[1+q^3\text{tr}\,[\Pi^3]+(q^3\text{tr}\,[\Pi^3])^2+(q^3\text{tr}\,[\Pi^3])^3+\dots] \nn
& & \dots \nn
& & \times[1+q^N\text{tr}\,[\Pi^N]+(q^N\text{tr}\,[\Pi^N])^2+(q^N\text{tr}\,[\Pi^N])^3+\dots] \nn
& & = \prod_{n=2}^N\sum_{m=0}^\infty(q^n\text{tr}\,[\Pi^n])^m,
\eea
where $q$ is introduced to count the number of $\Pi$. If one just want to count the number of these
gauge invariant operators, $\text{tr}\,[\Pi^n]$ in the above expression can be replaced by 1:
\beq
\prod_{n=2}^N\sum_{m=0}^\infty(q^n)^m=\prod_{n=2}^N(1-q^n)^{-1}
=(1-q)^{1-N}\prod_{n=2}^N(1+q+q^2+\dots+q^{n-1})^{-1}.
\eeq
The coefficient of $q^n$ in the expansion of this expression gives the number of the operators made of $n$ $\Pi$s.
Therefore the number of gauge invariant operators which act nontrivially on $\ket{p_2,p_3,\dots, p_N}$
can be counted by
\bea
& & \Big[\int d\text{SU}(N)\sum_{n=0}^\infty q^n\chi(\text{Sym}_n(\text{adj.}))
\sum_{n=0}^\infty q^n\chi(\text{Alt}_n(\text{adj.}))\Big]\Big/
\nn & &
(1-q)^{1-N}\prod_{n=2}^N(1+q+q^2+\dots+q^{n-1})^{-1}
\nn
& = &
\frac{1}{N!}\int_0^1 \prod_{i=1}^{N-1} dy_i
\Big[\prod_{\stackrel{\mbox{$\scriptstyle i,j=1$}}{\mbox{$\scriptstyle i\not\neq j$}}}^N(1-z_i/z_j)\Big]
(1+q)^{N-1}\prod_{n=2}^N(1+q+q^2+\dots+q^{n-1})
\nn & & \times
\prod_{\stackrel{\mbox{$\scriptstyle i,j=1$}}{\mbox{$\scriptstyle i\not\neq j$}}}^N\frac{1+qz_i/z_j}{1-qz_i/z_j}.
\eea
We can safely take the limit $q\rightarrow 1$ in this expression:
\beq
\int_0^1 \prod_{i=1}^{N-1}dy_i~2^{N-1}
\prod_{\stackrel{\mbox{$\scriptstyle i,j=1$}}{\mbox{$\scriptstyle i\not\neq j$}}}^N(1+z_i/z_j).
\label{udnum}
\eeq
Note that this expression is very close to $Z$.
\eqref{udnum} times the contribution from integrals of parameters $p_2, p_3,\dots,p_N$ gives
the contribution to the partition function from states constructed by acting the gauge invariant operators
on $\ket{p_2,p_3,\dots, p_N}$. At this stage we do not know how many independent
solutions with the same eigenvalues of $\text{tr}\,[\Pi^n]$ exist in the sector of zero fermion number.
CHS solution gives a hint to the contribution from integrals of parameters $p_2, p_3,\dots, p_N$.
It has $N-1$ continuous parameters $k_1, k_2,\dots, k_{N-1}$ corresponding to free particle momenta,
and these appear symmetrically in the solution.
Therefore the contribution to the partition function is computed by regarding $k_i$ independent,
with the factor $1/N!$:
\beq
\frac{1}{N!}\left[\int\frac{dk}{2\pi}\bra{k}
e^{-\frac{1}{2}\beta g^2 k^2}\ket{k}\right]^{N-1}
=\frac{1}{N!}\left[V\int\frac{dk}{2\pi}e^{-\frac{1}{2}\beta g^2 k^2}\right]^{N-1}
=\frac{1}{N!}\left[\frac{V}{\sqrt{2\pi\beta g^2}}\right]^{N-1}.
\eeq
This times \eqref{udnum} gives exactly the same expression as $Z$:
\beq
\frac{1}{N!}\left[\frac{2V}{\sqrt{2\pi\beta g^2}}\right]^{N-1}\int_0^1 \prod_{i=1}^{N-1}dy_i~
\Big[\prod_{\stackrel{\mbox{$\scriptstyle i,j=1$}}{\mbox{$\scriptstyle i\not\neq j$}}}^N(1+z_i/z_j)\Big].
\eeq
Thus we have given evidence for the following naively conjectured structure of the spectrum of states:
\begin{quotation}
The whole spectrum of $D=2$ matrix quantum mechanics is constructed by acting
gauge invariant operators made of $\Pi$ and at least one $\theta^\dagger$
on eigenstates of $\text{tr}\,[\Pi^n]~(n=2,3,\dots,N)$ in the sector of zero fermion number.
\end{quotation}
and if the normalization of $Z$ inferred from CHS solution is correct,
the following conjecture follows:
\begin{quotation}
In the sector of zero fermion number,
CHS solution is the unique eigenfunction of $\text{tr}\,[\Pi^2], \text{tr}\,[\Pi^3],\dots$ and $\text{tr}\,[\Pi^N]$
for fixed eigenvalues.
\end{quotation}
Although we do not have proofs of this conjecture for general $N$,
for $N=2$ case this has been shown to be correct\cite{ch85,s07}.
In the next section we shall give proofs for $N=3$ and $N=4$ cases.
\section{Analysis of eigenfunctions of $\text{tr}\,[\Pi^n]$ in lower $N$ cases}
In this section we give proofs of the conjecture in the previous section
for $N=3$ and $N=4$ cases i.e.
we show that CHS solutions for any $N$ are eigenfunctions of $\text{tr}\,[\Pi^3]$ and $\text{tr}\,[\Pi^4]$,
and they are the unique eigenfunctions of those operators with zero fermion number
for fixed eigenvalues in $N=3$ and $N=4$ cases.
First we show that CHS solutions are eigenfunctions of $\text{tr}\,[\Pi^3]$ and $\text{tr}\,[\Pi^4]$.
Since the expression \eqref{chss} is regarded as the restriction of the gauge invariant solution to diagonal $X$,
we have to know how $\Pi$ acts on such restricted functions.
Note that on any function $f(X)$, $\Pi_{(ij)}$ can be replaced by $G_{(ij)}$ if $X$ is set
$Y\equiv\text{diag}(x_1,\dots,x_N)$\cite{wln89}.
\beq
\Pi_{(ij)}f(X)\Big|_{X=Y}=
\frac{i}{x_i-x_j}G_{(ij)}f(X)\Big|_{X=Y},
\eeq
where $\Pi_a$ is regarded as $-i\frac{\p}{\p X_a}$.
Therefore, action of two or more $\Pi_{(ij)}$s on a gauge invariant function can be reduced to that of less $\Pi_{(ij)}$s
by replacing the first $\Pi_{(ij)}$ by $G_{(ij)}$, and moving it rightward using the following commutation relation
until it annihilates the function:
\beq
[G_{(ij)},\Pi_{(kl)}]=i\delta_{il}\delta_{jk}(\p_i-\p_j)+\delta_{il}\Pi_{(jk)}-\delta_{jk}\Pi_{(il)},
\eeq
where $\p_i=\frac{\p}{\p x_i}$ and $\Pi_{(ii)}=0$. These can be shown from \eqref{suncom} and $w_i=(\nu^i)_mX_m$.
$\Pi_m$ is also replaced by $\p_i$ with the following relation:
\beq
(\nu^i)_m\Pi_m=-i(\nu^i)_m\frac{\p}{\p X_m}=-i\sum_j(\nu^i\cdot\nu^j)\p_j=-i(\p_i-P),
\eeq
where $P$ is the center of mass momentum operator $P=\frac{1}{N}\sum_{i=1}^N\p_i$, and summation over $m$ is understood.
Since we set the eigenvalue of $P$
zero and $P$ commutes with any operator dependent only on $\p_i$ and the differences $x_i-x_j$,
we ignore terms proportional to $P$ in the following.
Thus we can reduce action of $\Pi$s on gauge invariant functions to $\p_i$.
Therefore we obtain
\bea
\text{tr}\,[\Pi^2]f(X)\Big|_{X=Y} & = &
\Bigg[\sum_{i=1}^N(\nu^i)_{m_1}(\nu^i)_{m_2}\Pi_{m_1}\Pi_{m_2}
+\mathop{{\sum}'}_{i,j=1}^N\Pi_{(ij)}\Pi_{(ji)}\Bigg]f(X)\Big|_{X=Y}
\nn & = &
-\left[\sum_{i=1}^N\p_i\p_i+\mathop{{\sum}'}_{i,j=1}^N\frac{2}{x_i-x_j}\p_i\right]f(X)\Big|_{X=Y},
\\
\text{tr}\,[\Pi^3]f(X)\Big|_{X=Y} & = & \Bigg[
\sum_{i=1}^N(\nu^i)_{m_1}(\nu^i)_{m_2}(\nu^i)_{m_3}\Pi_{m_1}\Pi_{m_2}\Pi_{m_3}
\nn & &
+3\mathop{{\sum}'}_{i,j=1}^N(\nu^i)_m\Pi_m\Pi_{(ji)}\Pi_{(ij)}
+\mathop{{\sum}'}_{i,j,k=1}^N\Pi_{(ji)}\Pi_{(kj)}\Pi_{(ik)}\Bigg]f(X)\Big|_{X=Y}
\nn & = & i\Bigg[
\sum_{i=1}^N(\p_i)^3+3\mathop{{\sum}'}_{i,j=1}^N\frac{3}{x_i-x_j}(\p_i)^2
\nn & &
+3\mathop{{\sum}'}_{i,j,k=1}^N\frac{1}{x_j-x_i}\frac{1}{x_k-x_i}\p_i\Bigg]
f(X)\Big|_{X=Y},
\\
\text{tr}\,[\Pi^4]f(X)\Big|_{X=Y} & = & \Bigg[
\sum_{i=1}^N(\nu^i)_{m_1}(\nu^i)_{m_2}(\nu^i)_{m_3}(\nu^i)_{m_4}\Pi_{m_1}\Pi_{m_2}\Pi_{m_3}\Pi_{m_4}
\nn & &
+4\mathop{{\sum}'}_{i,j=1}^N(\nu^j)_{m_1}(\nu^j)_{m_2}\Pi_{m_1}\Pi_{m_2}\Pi_{(ji)}\Pi_{(ij)}
\nn & &
+2\mathop{{\sum}'}_{i,j=1}^N(\nu^j)_{m_1}(\nu^i)_{m_2}\Pi_{m_1}\Pi_{m_2}\Pi_{(ji)}\Pi_{(ij)}
\nn & &
+4\mathop{{\sum}'}_{i,j,k=1}^N(\nu^j)_m\Pi_m\Pi_{(ji)}\Pi_{(kj)}\Pi_{(ik)}
\nn & &
+\mathop{{\sum}'}_{i,j=1}^N\Pi_{(ji)}\Pi_{(ij)}\Pi_{(ji)}\Pi_{(ij)}
+2\mathop{{\sum}'}_{i,j,k=1}^N\Pi_{(ji)}\Pi_{(ij)}\Pi_{(kj)}\Pi_{(jk)}
\nn & &
+\mathop{{\sum}'}_{i,j,k,l=1}^N\Pi_{(ji)}\Pi_{(kj)}\Pi_{(lk)}\Pi_{(il)}
\Bigg]f(X)\Big|_{X=Y}
\nn & = & \Bigg[
\sum_{i=1}^N(\p_i)^4+\mathop{{\sum}'}_{i,j=1}^N\frac{4}{x_i-x_j}(\p_i)^3
\nn & &
+6\mathop{{\sum}'}_{i,j,k=1}^N\frac{1}{x_j-x_i}\frac{1}{x_k-x_i}(\p_i)^2
\nn & &
-4\mathop{{\sum}'}_{i,j,k,l=1}^N\frac{1}{x_j-x_i}\frac{1}{x_k-x_i}\frac{1}{x_l-x_j}\p_i
\Bigg]f(X)\Big|_{X=Y},
\eea
where $\sum'$ indicates summations with the condition that the dummy indices take different values from each other.
CHS solutions have the prefactor ${\cal M}$. This can be moved leftward using
\beq
\p_i\frac{1}{\cal M}=\frac{1}{\cal M}
\Big(\p_i+\sum_{\stackrel{\mbox{$\scriptstyle j=1$}}{\mbox{$\scriptstyle j\not\neq i$}}}^N\frac{1}{x_j-x_i}\Big),
\eeq
and with the following identities:
\bea
0 & = & \mathop{{\sum}'}_{i,j,k=1}^N\frac{1}{(x_j-x_i)(x_k-x_i)},
\nn
0 & = & \mathop{{\sum}'}_{i,j,k,l=1}^N\frac{1}{(x_j-x_i)(x_k-x_i)(x_l-x_i)},
\nn
0 & = & \mathop{{\sum}'}_{i,j,k,l,p=1}^N\frac{1}{(x_j-x_i)(x_k-x_i)(x_l-x_i)(x_p-x_i)},
\eea
we can show
\bea
\text{tr}\,[\Pi^2]\frac{1}{\cal M}f(X)\Big|_{X=Y} & = & -\frac{1}{\cal M}
\sum_{i=1}^N(\p_i)^2f(X)\Big|_{X=Y},
\label{mp2} \\
\text{tr}\,[\Pi^3]\frac{1}{\cal M}f(X)\Big|_{X=Y} & = & \frac{i}{\cal M}
\sum_{i=1}^N(\p_i)^3f(X)\Big|_{X=Y},
\\
\text{tr}\,[\Pi^4]\frac{1}{\cal M}f(X)\Big|_{X=Y} & = & \frac{1}{\cal M}
\sum_{i=1}^N(\p_i)^4f(X)\Big|_{X=Y}.
\label{mp4}
\eea
Obviously $\sum_{\sigma}\text{sgn}(\sigma)\exp\left[i\sum_{i=1}^Nk_{\sigma(i)}x_i\right]$
are eigenfunctions of $\sum_{i=1}^N(\p_i)^3$ and $\sum_{i=1}^N(\p_i)^4$, and therefore
the above equations mean that CHS solutions are eigenfunctions of $(\Pi^3)$ and $(\Pi^4)$.
Next, using the method in \cite{trze06}, we shall show that
eigenfunctions of $\text{tr}\,[\Pi^n]$ with zero fermion number for fixed eigenvalues
are unique in $N=3$ and $N=4$ cases.
Gauge invariant states in the sector of zero fermion number are in the following form:
\beq
\psi(X)=\sum_{n_2,n_3,\dots, n_N=0}^\infty a_{n_2,n_3,\dots,n_N}(X^2)^{n_2}(X^3)^{n_3}\dots(X^N)^{n_N},
\eeq
where $(A)=\text{tr}\,[A]$. The following expression in $(\Pi^n)\psi$,
\beq
(\Pi^n)[(X^2)^{n_2}(X^3)^{n_3}\dots(X^N)^{n_N}]
\eeq
can be rewritten as a sum of terms in the form of
$(X^2)^{m_2}(X^3)^{m_3}\dots(X^N)^{m_N}$. Therefore eigenvalue equations $(\Pi^n)\psi=p_n\psi$ give
recurrence relations for determining $a_{n_2,n_3,\dots,n_N}$. If these relations determine them uniquely
up to the overall scaling, the eigenfunction must be unique.
For $N=2$, the recurrence relation contains $a_n$ and $a_{n+1}$, which determines $a_n$
completely in terms of $a_0$ \cite{trze06}.
For $N=3$, case,
by explicit calculation using \eqref{tr1}, \eqref{tr2} and the following equations:
\bea
{}[(\Pi^2), (X^2)^n] & = & 2n(3-2n-N^2)(X^2)^{n-1}-4ni(X^2)^{n-1}(X\Pi), \\
{}[(\Pi^2), (X^3)^n] & = & 9n(n-1)\left\{\frac{1}{N}(X^2)^2-(X^4)\right\}(X^3)^{n-2}
-6ni(X^3)^{n-1}(X^2\Pi), \\
{}[(X\Pi), (X^3)^n] & = & -3ni(X^3)^n, \\
{}[(\Pi^3), (X^2)^n] & = & 8in(n-1)(n-2)(X^3)(X^2)^{n-3}
\nn & &
-6ni(X^2)^{n-1}(X\Pi^2)-12n(n-1)(X^2)^{n-2}(X^2\Pi), \\
{}[(\Pi^3), (X^3)^n] & = & i\left\{3n\left(\frac{4}{N}-5N+N^3\right)
-27n(n-1)\left(\frac{4}{N}-N\right)\right\}(X^3)^{n-1}
\nn & &
+27in(n-1)(n-2)\left\{\frac{2}{N^2}(X^2)^3-\frac{3}{N}(X^2)(X^4)+(X^6)\right\}(X^3)^{n-3}
\nn & &
-9ni(X^3)^{n-1}(X^2\Pi^2)+\frac{9n}{N}i(X^2)(X^3)^{n-1}(\Pi^2)
\nn & &
+9n\left(\frac{4}{N}-N\right)(X^3)^{n-1}(X\Pi)-27n(n-1)(X^3)^{n-2}(X^4\Pi)
\nn & &
+\frac{54}{N}n(n-1)(X^2)(X^3)^{n-2}(X^2\Pi),
\\
{}[(X\Pi^2), (X^3)^n] & = & \left\{3n\left(\frac{4}{N}-N\right)
+\frac{18}{N}n(n-1)\right\}(X^2)(X^3)^{n-1}-9n(n-1)(X^5)(X^3)^{n-2}
\nn & &
-6ni(X^3)^{n-1}(X^3\Pi)+\frac{6n}{N}i(X^2)(X^3)^{n-1}(X\Pi),
\\
{}[(X^2\Pi), (X^3)^n] & = & -3ni(X^4)(X^3)^{n-1}+\frac{3n}{N}i(X^2)^2(X^3)^{n-1},
\eea
and the following equations which hold in $N=3$ case,
\beq
(X^4)=\frac{1}{2}(X^2)^2,\quad (X^5)=\frac{5}{6}(X^2)(X^3),\quad (X^6)=\frac{1}{4}(X^2)^3+\frac{1}{3}(X^3)^2,
\eeq
we obtain two recurrence relations, respectively from $(\Pi^2)\psi=p_2\psi$ and $(\Pi^3)\psi=p_3\psi$:
\bea
p_2a_{n,m} & = & -4(n+1)(n+3m+4)a_{n+1,m}-\frac{3}{2}(m+1)(m+2)a_{n-2,m+2},
\label{rrsu3p2}\\
-ip_3a_{n,m} & = & (m+1)\{30n+9nm+6n(n-1)+9m(m+4)+40\}a_{n,m+1}
\nn & &
-\frac{3}{4}(m+1)(m+2)(m+3)a_{n-3,m+3}
\nn & &
+8(n+1)(n+2)(n+3)a_{n+3,m-1},
\label{rrsu3p3}
\eea
where $n, m\geq 0$ and terms proportional to $a_{n',m'}$ with negative $n'$ or $m'$ must be omitted.
From \eqref{rrsu3p2} for $n=0$ and 1,
\beq
a_{1,m}=-\frac{p_2}{4(3m+4)}a_{0,m},\quad
a_{2,m}=-\frac{p_2}{8(3m+5)}a_{1,m}=\frac{(p_2)^2}{32(3m+4)(3m+5)}a_{0,m},
\label{rrsu3p2-3}
\eeq
which means that $a_{1,m}$ and $a_{2,m}$ are determined by $a_{0,m}$.
Furthermore by the following relation from \eqref{rrsu3p2} for $n\geq 3$:
\beq
a_{n,m}=-\frac{1}{4n(n+3m+3)}\left\{p_2a_{n-1,m}+\frac{3}{2}(m+1)(m+2)a_{n-3,m+2}\right\},
\label{rrsu3p2-2}
\eeq
$a_{n,m}$ with $n\geq 3$ can also be determined recursively from $a_{0,m}$. Therefore there are
infinitely many solutions to $(\Pi^2)\psi=p_2\psi$ and are determined by giving $a_{0,m}$ as an initial condition.
Next we shall analyze \eqref{rrsu3p3}. For $n=0$ and $m\geq 1$,
\beq
-ip_3a_{0,m} = (m+1)\{9m(m+4)+40\}a_{0,m+1}+48a_{3,m-1}.
\eeq
Eliminating $a_{3,m-1}$ in this equation by using \eqref{rrsu3p2-2},
\beq
-ip_3a_{0,m} = (m+1)\{9m(m+4)+40\}a_{0,m+1}
-\frac{4}{3m+3}\left\{p_2a_{2,m-1}+\frac{3}{2}m(m+1)a_{0,m+1}\right\},
\eeq
and eliminating $a_{2,m-1}$ again in this equation by using \eqref{rrsu3p2-3},
\beq
0=\big[(m+1)\{9m(m+4)+40\}-2m\big] a_{0,m+1}+ip_3a_{0,m}-\frac{(p_2)^3}{8(3m+1)(3m+2)(3m+3)}a_{0,m-1}.
\label{su3frr}
\eeq
This means that $a_{0,m}$ with $m\geq 2$ are determined from $a_{0,0}$ and $a_{0,1}$.
Again from \eqref{rrsu3p3} for $m=0$ and $n=0,1,2$,
\beq
a_{0,1}=-\frac{ip_3}{40}a_{0,0},\quad
a_{1,1}=-\frac{ip_3}{70}a_{1,0},\quad
a_{2,1}=-\frac{ip_3}{112}a_{2,0}.
\eeq
The first equation of the above means that the entire solution is determined by $a_{0,0}$ and shows that the
solution for fixed $p_2$ and $p_3$ is unique. The second and third ones are consistent with \eqref{rrsu3p2-3}.
Since it seems difficult to give simple expression of the solution to \eqref{su3frr}, we only show some of $a_{n,m}$:
\beq
a_{0,2}=\frac{p_2^3 - 24 p_3^2}{161280}a_{0,0},\quad a_{0,3}=-\frac{ip_3(p_2^3-12p_3^2)}{35481600}a_{0,0},\quad
a_{0,4}=\frac{p_2^6-72p_2^3p_3^2+576p_3^4}{1549836288000}a_{0,0},\quad\dots
\eeq
This solution must coincide with CHS solution. Indeed by straightforward calculation
we have checked that lower terms in the expansion of CHS solution reproduce the above $a_{n,m}$.
Thus we see that for $N=3$ our conjecture given in the previous section is correct.
Next we shall analyze $N=4$ case. By the calculation similar to $N=3$ case we obtain three recurrence relations:
\bea
p_2a_{n,m,l} & = & 2(l+1)(l+2)a_{n-3,m,l+2}+
\frac{9}{4}(m+1)(m+2)a_{n-2,m+2,l}
\nn & &
-(l+1)(14 m+12 1+29)a_{n-1,m,l+1}-
\frac{4}{3}(l+1)(l+2)a_{n,m-2,l+2}
\nn & &
-9(m+1)(m+2)a_{n,m+2,l-1}-
2(n+1)(15+2n+6m+8l))a_{n+1,m,l},
\label{su4p2}
\\
-ip_3a_{n,m,l} & = &
8(n+1)(n+2)(n+3)a_{n+3,m-1,l}+
36(n+1)(n+2)(m+1)a_{n+2,m+1,l-1}
\nn & &
+4 (n+1)(l+1)(12 + 2 l + 3 m + 7 n)a_{n+1,m-1,l+1}
\nn & &
+9(m+1)(15 + 18 l + 4 l^2 + 8 m + 4 l m + m^2 + 7 n + 8 l n + 2 m n - n^2)a_{n,m+1,l}
\nn & &
-\frac{8}{9}(l+1)(l+2)(l+3)a_{n,m-3,l+3}
\nn & &
+(l+1)(l+2)(32 + 28 n + 10 m + 12 l)a_{n-1,m-1,l+2}
\nn & &
+9(m+1)(l+1)(3/2 + l - 2 n)a_{n-2,m+1,l+1}
\nn & &
+6(l+1)(l+2)(l+3)a_{n-3,m-1,l+3}-
\frac{9}{2}(m+1)(l+1)(l+2)a_{n-4,m+1,l+2},
\label{su4p3} \\
p_4 a_{n,m,l} & = &
\frac{27}{4}(m+1)(m+2)(l+1)(l+2)a_{n-5,m+2,l+2}
\nn & &
-(l+1)(l+2)(l+3)(37 + 12 l + 26 m + 8 n) a_{n-4,m,l+3}
\nn & &
+\frac{81}{64} (m+1)(m+2)(m+3)(m+4)a_{n-4,m+4,l}
\nn & &
+\frac{28}{3}(l+1)(l+2)(l+3)(l+4)a_{n-3,m-2,l+4}
\nn & &
-\frac{135}{8}(m+1)(m+2)(l+1)(9 + 4 l + 2 m) a_{n-3,m+2,l+1}
\nn & &
+\frac{1}{4}(l+1)(l+2)(1791 + 1240 l + 192 l^2 + 1878m + 688 lm
\nn & &
+ 318m^2 + 24n - 48mn - 96n^2) a_{n-2,m,l+2}
\nn & &
-\frac{81}{8} (m+1)(m+2)(m+3)(m+4)a_{n-2,m+4,l-1}
\nn & &
+\frac{2}{9}(l+1)(l+2)(l+3)(69 + 36 l + 50 m + 136 n) a_{n-1,m-2,l+3}
\nn & &
+\frac{9}{4} (m+1)(m+2)(117 + 222 l + 72 l^2 + 50m + 60 lm
\nn & &
+ 4m^2 - 51n - 24 ln - 14mn - 6n^2) a_{n-1,m+2,l}
\nn & &
+\frac{28}{27}(l+1)(l+2)(l+3)(l+4)a_{n,m-4,l+4}
\nn & &
+\frac{1}{2}(l+1)(1335 + 1384 l + 384 l^2 + 32 l^3 + 1284m + 764 lm + 64 l^2m
\nn & &
+ 360m^2 + 84 lm^2 + 24m^3 + 2102n + 1768 ln + 320 l^2n + 1944mn
\nn & &
+ 816 lmn + 336m^2n + 324n^2 + 192 ln^2 + 168mn^2 - 32n^3) a_{n,m,l+1}
\nn & &
+\frac{81}{4} (m+1)(m+2)(m+3)(m+4)a_{n,m+4,l-2}
\nn & &
+\frac{4}{9}(l+1)(l+2)(n+1)(135 + 32 l + 18m + 78n) a_{n+1,m-2,l+2}
\nn & &
+ 9 (m+1)(m+2)(n+1)(45 + 24 l + 14m + 6n) a_{n+1,m+2,l-1}
\nn & &
+(n+1)(n+2)(435 + 480 l + 96 l^2 + 396m + 192 lm
\nn & &
+ 72m^2 + 116n + 96 ln + 56mn) a_{n+2,m,l}
\nn & &
+\frac{32}{3}(l+1)(n+1)(n+2)(n+3)a_{n+3,m-2,l+1}
\nn & &
+16 (n+1)(n+2)(n+3)(n+4)a_{n+4,m,l-1},
\label{su4p4}
\eea
where $n, m,l\geq 0$ and terms proportional to $a_{n',m',l'}$ with negative $n'$, $m'$ or $l'$ must be omitted.
From \eqref{su4p2}, we obtain the following equation which determines $a_{n+1,m,l}$ with $n\geq 0$,
$m\geq 0$, and $l\geq 0$, in terms of $a_{n',m',l'}$ with $n'\leq n$:
\bea
a_{n+1,m,l} & = & \frac{1}{2(n+1)(2n+6m+8l+15)}\Big[2(l+1)(l+2)a_{n-3,m,l+2} \nn
& & +\frac{9}{4}(m+1)(m+2)a_{n-2,m+2,l} - (l+1)(14m+12l+29)a_{n-1,m,l+1} \nn
& & - \frac{4}{3}(l+1)(l+2)a_{n,m-2,l+2} - p_2a_{n,m,l} - 9(m+1)(m+2)a_{n,m+2,l-1}\Big].
\eea
Therefore we only have to determine $a_{0,m,l}$. By setting $n=0$ and $l\geq 3$ in \eqref{su4p3}
and using the above to rewrite $a_{n,m,l}$ with $n\geq 1$ in terms of $a_{0,m,l}$,
we obtain the following equation, which
determines $a_{0,m+5,l}$ with $m\geq 0$ and $l\geq 0$ in terms of $a_{0,m',l'}$ with $m'\leq 4$:
\bea
a_{0,m+5,l} & = & \frac{1}{(m+1)(m+2)(m+3)(m+4)(m+5)(37+5m+8 l)}\Big[
\nn & &
\frac{64}{19683}(l+4)(l+5)(l+6)(l+7)(l+8)(l+9)a_{0,m-7,l+9}
\nn & &
+\frac{16}{2187}\frac{(l+4)(l+5)(l+6)(l+7)(39+6m+8 l)}{35+6m+ 8 l}p_2a_{0,m-5,l+7}
\nn & &
+\frac{4}{729}\frac{(l+4)(l+5)(41+6m+8 l)}{33+6m+8 l} p_2^2a_{0,m-3,l+5}
\nn & &
+\frac{8}{6561}(l+4)(l+5)(l+6) (119064 + 69256 l + 13584 l^2 + 896 l^3
\nn & &
+ 53865 m + 21600 l m + 2160 l^2 m + 9450 m^2 + 1890 l m^2 + 540 m^3)a_{0,m-3,l+6}
\nn & &
+\frac{1}{729}\frac{(39+6m+8 l)(41+6m+8 l)}{(33+6m+8 l) (35+6m+8 l)}p_2^3a_{0,m-1,l+3}
\nn & &
+\frac{2}{729}\frac{(l+4)(39+6m+8 l)}{35+6m+8 l}(16569 + 9730 l + 1920 l^2 + 128 l^3
\nn & &
+ 7497 m + 2952 l m + 288 l^2 m +
1620 m^2 + 324 l m^2 + 108 m^3) p_2a_{0,m-1,l+4}
\nn & &
-\frac{i}{729}(37+6m+8 l)(39+6m+8 l)(41+6m+8 l)p_3a_{0,m,l+3}
\nn & &
-\frac{1}{81}\frac{(m+1)(33+3m+8 l)(41+6m+8 l)}{33+6m+8 l} p_2^2a_{0,m+1,l+2}
\nn & &
- \frac{1}{81}(m+1) (5618265 + 5767832 l + 2369580 l^2 + 486560 l^3 +
49920 l^4 + 2048 l^5
\nn & &
+ 3603906 m + 2956578 l m + 911180 l^2 m + 124800 l^3 m + 6400 l^4 m
\nn & &
+ 926103 m^2 + 567682 l m^2 + 116580 l^2 m^2 + 8000 l^3 m^2 + 117882 m^3
\nn & &
+ 47628 l m^3 + 4860 l^2 m^3 + 7668 m^4 + 1512 l m^4 + 216 m^5)a_{0,m+1,l+3}
\nn & &
-\frac{1}{9}\frac{(m+1)(m+2)(m+3)(39+6m+8 l)(70+9m+16 l)}{35+6m+8 l}p_2a_{0,m+3,l+1}
\Big].
\eea
For later use we give some more equations given by setting
$(n,m,l)=$
$(0,0,2)$,
$(0,0,1)$, $(0,3,0)$, $(0,2,0)$, $(0,1,0)$, $(0,0,0)$ in \eqref{su4p3}:
\bea
-ip_3a_{0, 0, 2} & = &
\frac{9}{899}p_2^2a_{0, 1, 1} + \frac{184275}{341}a_{0, 1, 2} + \frac{324}{319}p_2a_{0, 3, 0},
\label{eeq1} \\
-ip_3a_{0, 1, 1} & = & -\frac{1}{15525}p_2^3a_{0, 0, 1} - \frac{2024}{1395}p_2a_{0, 0, 2} +
\frac{40}{2001}p_2^2a_{0, 2, 0} + \frac{720720}{899}a_{0, 2, 1},
\\
-ip_3a_{0, 0, 1} & = & \frac{3}{161}p_2^2a_{0, 1, 0} + \frac{6885}{23}a_{0, 1, 1},
\\
-ip_3a_{0, 3, 0} & = & -\frac{8}{27621}p_2^2a_{0, 0, 2} -
\frac{4240}{341}a_{0, 0, 3} - \frac{1}{24273}p_2^3a_{0, 2, 0}
\nn & &
- \frac{780}{899}p_2a_{0, 2, 1} + \frac{555120}{341}a_{0, 4, 0},
\\
-ip_3a_{0, 2, 0} & = & -\frac{1}{12075}p_2^3a_{0, 1, 0} -
\frac{570}{667}p_2a_{0, 1, 1} + \frac{129789}{145}a_{0, 3, 0},
\\
-ip_3a_{0, 1, 0} & = & -\frac{1}{4845}p_2^3a_{0, 0, 0} -
\frac{336}{391}p_2a_{0, 0, 1} + \frac{181440}{437}a_{0, 2, 0},
\\
-ip_3a_{0, 0, 0} & = & 135a_{0, 1, 0}.
\label{eeq7}
\eea
Henceforth we only show schematic forms of equations using the following symbol:
\beq
[n,m,l]\equiv\text{a term proportional to}~a_{n,m,l},
\eeq
because explicit expressions are lengthy and it is not illuminating to show details of them.
However we have computed all the explicit expressions and have confirmed that there is no singular coefficient
which prevents us from solving linear equations.
Setting $n=0$ in \eqref{su4p4}, rewriting $a_{n,m,l}$ with $n\geq 1$ in terms of $a_{0,m,l}$, and
setting $m=0,1,2,3,4$, we obtain
\bea
[0,0,l] & = & [0,0,l-1]+[0,0,l]+[0,0,l+1]+[0,2,l-2]+[0,2,l-1]
\nn & &
+[0,4,l-3]+[0,4,l-2]+[0,6,l-4]+[0,8,l-5],
\label{a001} \\
{}[0,1,l] & = & [0,1,l-1]+[0,1,l]+[0,1,l+1]+[0,3,l-2]+[0,3,l-1]
\nn & &
+[0,5,l-3]+[0,5,l-2]+[0,7,l-4]+[0,9,l-5],
\label{a011} \\
{}[0,2,l] & = & [0,0,l+1]+[0,0,l+2]+[0,2,l-1]+[0,2,l]+[0,2,l+1]+[0,4,l-2]
\nn & &
+[0,4,l-1]+[0,6,l-3]+[0,6,l-2]+[0,8,l-4]+[0,10,l-5],
\label{a021} \\
{}[0,3,l] & = & [0,1,l+1]+[0,1,l+2]+[0,3,l-1]+[0,3,l]+[0,3,l+1]+[0,5,l-2]
\nn & &
+[0,5,l-1]+[0,7,l-3]+[0,7,l-2]+[0,9,l-4]+[0,11,l-5],
\label{a031} \\
{}[0,4, l] & = & [0,0,l+3]+[0,0,l+4]+[0,2,l+1]+[0,2,l+2]
\nn & &
+[0,4,l-1]+[0,4,l]+[0,4,l+1]+[0,6,l-2]+[0,6,l-1]
\nn & &
+[0,8,l-3]+[0,8,l-2]+[0,10,l-4]+[0,12,l-5].
\label{a041}
\eea
Rewriting $[0,m,l]$ with $m\geq 5$ further in terms of $[0,m',l']$ with $m'=0,1,2,3,4$,
we obtain the following: from \eqref{a001},
\bea
[0,0,0] & = & [0,0,0]+[0,0,1], \label{a000-1} \\
{}[0,0,1] & = & [0,0,0]+[0,0,1]+[0,0,2]+[0,2,0], \label{a000-2}\\
{}[0,0,2] & = & [0,0,1]+[0,0,2]+[0,0,3]+[0,2,0]+[0,2,1]+[0,4,0], \label{a000-3}\\
{}[0,0,3] & = & [0,0,2]+[0,0,3]+[0,0,4]+[0,2,1]+[0,2,2]+[0,4,0]+[0,4,1], \\
{}[0,0,4] & = & [0,0,3]+[0,0,4]+[0,0,5]+[0,1,3]+[0,2,2]+[0,2,3]
\nn & &
+[0,4,1]+[0,4,2], \\
{}[0,0,l] & = & [0,0,l-1]+[0,0,l]+[0,0,l+1]+[0,1,l-1]+[0,2,l-2]
\nn & &
+[0,2,l-1]+[0,3,l-2]+[0,4,l-3]+[0,4,l-2], \label{a000-5}
\eea
where $l\geq 5$. From \eqref{a011},
\bea
[0,1,0] & = & [0,1,0]+[0,1,1], \label{a010-1}\\
{}[0,1,1] & = & [0,1,0]+[0,1,1]+[0,1,2]+[0,3,0], \\
{}[0,1,2] & = & [0,0,3]+[0,1,1]+[0,1,2]+[0,1,3]+[0,3,0]+[0,3,1], \\
{}[0,1,3] & = & [0,0,3]+[0,0,4]+[0,1,2]+[0,1,3]+[0,1,4]+[0,3,1]+[0,3,2], \\
{}[0,1,4] & = & [0,0,4]+[0,0,5]+[0,1,3]+[0,1,4]+[0,1,5]+[0,2,3]
\nn & &
+[0,3,2]+[0,3,3], \label{a010-5} \\
{}[0,1,l] & = & [0,0,l]+[0,0,l+1]+[0,1,l]+[0,1,l+1]+[0,2,l-1]
\nn & &
+[0,3,l-1]+[0,4,l-2],
\eea
where $l\geq 5$. From \eqref{a021},
\bea
[0, 2, 0] & = & [0, 0, 1]+[0, 0, 2]+[0, 2, 0]+[0, 2, 1], \label{a020-1}\\
{}[0, 2, 1] & = & [0, 0, 2]+[0, 0, 3]+[0, 2, 0]+[0, 2, 1]+[0, 2, 2]+[0, 4, 0], \\
{}[0, 2, 2] & = & [0, 0, 3]+[0, 0, 4]+[0, 1, 3]+[0, 2, 1]+[0, 2, 2]+[0, 2, 3]
\nn & &
+[0, 4, 0]+[0, 4, 1], \\
{}[0, 2, 3] & = & [0, 0, 3]+[0, 0, 4]+[0, 0, 5]+[0, 1, 3]+[0, 1, 4]
\nn & &
+[0, 2, 2]+[0, 2, 3]+[0, 2, 4]+[0, 4, 1]+[0, 4, 2],
\\
{}[0, 2, 4] & = &[0, 0, 4]+[0, 0, 5]+[0, 0, 6]+[0, 1, 4]+[0, 1, 5]
\nn & &
+[0, 2, 3]+[0, 2, 4]+[0, 2, 5]+[0, 3, 3]+[0, 4, 2]+[0, 4, 3], \\
{}[0,2,l] & = & [0,0,l]+[0,0,l+1]+[0,0,l+2]+[0,1,l]+[0,2,l-1]
\nn & &
+[0,2,l]+[0,2,l+1]+[0,3,l-1]+[0,4,l-2]+[0,4,l-1], \label{a020-5}
\eea
where $l\geq 5$. From \eqref{a031},
\bea
[0, 3, 0] & = & [0, 1, 1]+[0, 1, 2]+[0, 3, 0]+[0, 3, 1], \label{a030-1}\\
{}[0, 3, 1] & = & [0, 0, 3]+[0, 1, 2]+[0, 1, 3]+[0, 3, 0]+[0, 3, 1]+[0, 3, 2], \\
{}[0, 3, 2] & = & [0, 0, 3]+[0, 0, 4]+[0, 1, 2]+[0, 1, 3]+[0, 1, 4]+[0, 2, 3]
\nn & &
+[0, 3, 1]+[0, 3, 2]+[0, 3, 3], \\
{}[0, 3, 3] & = & [0, 0, 4]+[0, 0, 5]+[0, 1, 3]+[0, 1, 4]+[0, 1, 5]
\nn & &
+[0, 2, 3]+[0, 2, 4]+[0, 3, 2]+[0, 3, 3]+[0, 3, 4], \\
{}[0, 3, 4] & = & [0, 0, 5]+[0, 0, 6]+[0, 1, 4]+[0, 1, 5]+[0, 1, 6]
\nn & &
+[0, 2, 4]+[0, 2, 5]+[0, 3, 3]+[0, 3, 4]+[0, 3, 5]+[0, 4, 3], \\
{}[0,3,l] & = & [0,0,l+1]+[0,0,l+2]+[0,1,l+1]
\nn & &
+[0,1,l+2]+[0,2,l]+[0,2,l+1]+[0,3,l]
\nn & &
+[0,3,l+1]+[0,4,l-1], \label{a030-5}
\eea
where $l\geq 5$. From \eqref{a041},
\bea
[0, 4, 0] & = & [0, 0, 3]+[0, 0, 4]+[0, 2, 1]+[0, 2, 2]+[0, 4, 0]+[0, 4, 1], \label{a040-1}\\
{}[0, 4, 1] & = & [0, 0, 3]+[0, 0, 4]+[0, 0, 5]+[0, 1, 3]+[0, 2, 2]+[0, 2, 3]
\nn & &
+[0, 4, 0]+[0, 4, 1]+[0, 4, 2], \\
{}[0, 4, 2] & = & [0, 0, 3]+[0, 0, 4]+[0, 0, 5]+[0, 0, 6]+[0, 1, 3]+[0, 1, 4]
\nn & &
+[0, 2, 2]+[0, 2, 3]+[0, 2, 4]+[0, 3, 3]+[0, 4, 1]+[0, 4, 2]+[0, 4, 3], \\
{}[0, 4, 3] & = & [0, 0, 4]+[0, 0, 5]+[0, 0, 6]+[0, 0, 7]+[0, 1, 4]+[0, 1, 5]+[0, 2, 3]
\nn & &
+[0, 2, 4]+[0, 2, 5]+[0, 3, 3]+[0, 3, 4]+[0, 4, 2]+[0, 4, 3]+[0, 4, 4], \\
{}[0, 4, 4] & = & [0, 0, 5]+[0, 0, 6]+[0, 0, 7]+[0, 0, 8]+[0, 1, 5]+[0, 1, 6]+[0, 2, 4]
\nn & &
+[0, 2, 5]+[0, 2, 6]+[0, 3, 4]+[0, 3, 5]+[0, 4, 3]+[0, 4, 4]+[0, 4, 5], \\
{}[0,4,l] & = & [0,0,l+1]+[0,0,l+2]+[0,0,l+3]+[0,0,l+4]+[0,1,l+1]
\nn & &
+[0,1,l+2]+[0,2,l]+[0,2,l+1]+[0,2,l+2]+[0,3,l]
\nn & &
+[0,3,l+1]+[0,4,l-1]+[0,4,l]+[0,4,l+1], \label{a040-5}
\eea
where $l\geq 5$.
From \eqref{a000-1}-\eqref{a000-5}, we see that $a_{0,0,l+1}$ is determined by $a_{0,0,l'}$ with $l'\leq l$,
$a_{0,1,l'}$ with $l'\leq l-1$, $a_{0,2,l'}$ with $l'\leq l-1$, $a_{0,3,l'}$ with $l'\leq l-2$,
and $a_{0,4,l'}$ with $l'\leq l-2$.
From \eqref{a010-1}-\eqref{a010-5}, we see that $a_{0,1,l+1}$ is determined by $a_{0,0,l'}$ with $l'\leq l+1$,
$a_{0,1,l'}$ with $l'\leq l$, $a_{0,2,l'}$ with $l'\leq l-1$, $a_{0,3,l'}$ with $l'\leq l-1$,
and $a_{0,4,l'}$ with $l'\leq l-2$.
From \eqref{a020-1}-\eqref{a020-5}, we see that $a_{0,2,l+1}$ is determined by $a_{0,0,l'}$ with $l'\leq l+2$,
$a_{0,1,l'}$ with $l'\leq l+1$, $a_{0,2,l'}$ with $l'\leq l$, $a_{0,3,l'}$ with $l'\leq l-1$,
and $a_{0,4,l'}$ with $l'\leq l-1$.
From \eqref{a030-1}-\eqref{a030-5}, we see that $a_{0,3,l+1}$ is determined by $a_{0,0,l'}$ with $l'\leq l+2$,
$a_{0,1,l'}$ with $l'\leq l+2$, $a_{0,2,l'}$ with $l'\leq l+1$, $a_{0,3,l'}$ with $l'\leq l$,
and $a_{0,4,l'}$ with $l'\leq l-1$.
We find it more useful to use the following equations obtained by eliminating $[0,0,l+4]$ in
\eqref{a040-1}-\eqref{a040-5} by using \eqref{a000-1}-\eqref{a000-5},
than to use \eqref{a040-1}-\eqref{a040-5}:
\bea
[0,4,0] & = & [0,0,2]+[0,0,3]+[0,2,1]+[0,2,2]+[0,4,0]+[0,4,1],
\\
{}[0,4,1] & = & [0, 0, 3]+[0, 0, 4]+[0, 1, 3]+[0, 2, 2]+[0, 2, 3]
\nn & &
+[0, 4, 0]+[0, 4, 1]+[0, 4, 2],
\\
{}[0, 4, 2] & = & [0, 0, 3]+[0, 0, 4]+[0, 0, 5]+[0, 1, 3]+[0, 1, 4]
\nn & &
+[0, 2, 2]+[0, 2, 3]+[0, 2, 4]+[0, 3, 3]
\nn & &
+[0, 4, 1]+[0, 4, 2]+[0, 4, 3],
\\
{}[0, 4, 3] & = & [0, 0, 4]+[0, 0, 5]+[0, 0, 6]+[0, 1, 4]+[0, 1, 5]
\nn & &
+[0, 2, 3]+[0, 2, 4]+[0, 2, 5]+[0, 3, 3]+[0, 3, 4]
\nn & &
+[0, 4, 2]+[0, 4, 3]+[0, 4, 4],
\\
{}[0, 4, 4] & = & [0, 0, 5]+[0, 0, 6]+[0, 0, 7]+[0, 1, 5]+[0, 1, 6]
\nn & &
+[0, 2, 4]+[0, 2, 5]+[0, 2, 6]+[0, 3, 4]+[0, 3, 5]
\nn & &
+[0, 4, 3]+[0, 4, 4]+[0, 4, 5],
\\
{}[0,4,l] & = & [0,0,l+1]+[0,0,l+2]+[0,0,l+3]+[0,1,l+1]+[0,1,l+2]
\nn & &
+[0,2,l]+[0,2,l+1]+[0,2,l+2]+[0,3,l+3]+[0,3,l+1]
\nn & &
+[0,4,l-1]+[0,4,l]+[0,4,l+1].
\eea
From these, we see that $a_{0,4,l+1}$ is determined by $a_{0,0,l'}$ with $l'\leq l+3$,
$a_{0,1,l'}$ with $l'\leq l+2$, $a_{0,2,l'}$ with $l'\leq l+2$, $a_{0,3,l'}$ with $l'\leq l+1$,
and $a_{0,4,l'}$ with $l'\leq l$.
In summary, $a_{0,0,l+1}$, $a_{0,1,l+1}$, $a_{0,2,l+1}$, $a_{0,3,l+1}$ and $a_{0,4,l+1}$ are determined
by $a_{0,0,l'}$, $a_{0,1,l'}$, $a_{0,2,l'}$, $a_{0,3,l'}$ and $a_{0,4,l'}$ with $\l'\leq l$.
(These can be determined in the following order:
$a_{0,0,l+1}$, $a_{0,0,l+2}$, $a_{0,1,l+1}$, $a_{0,1,l+2}$, $a_{0,2,l+1}$, $a_{0,0,l+3}$,
$a_{0,3,l+1}$, $a_{0,2,l+2}$, $a_{0,4,l+1}$.)
Therefore all the $a_{n,m,l}$ are deteremined in terms of $a_{0,0,0}$, $a_{0,1,0}$, $a_{0,2,0}$, $a_{0,3,0}$,
and $a_{0,4,0}$.
Finally we determine $a_{0,1,0}$, $a_{0,2,0}$, $a_{0,3,0}$, and $a_{0,4,0}$ in terms of $a_{0,0,0}$.
Explicit forms of \eqref{a000-1}, \eqref{a000-2}, \eqref{a000-3} and \eqref{a030-1} are
\bea
p_4a_{0,0,0} & = & \frac{29}{68}p_2^2a_{0,0,0} + \frac{5040}{17}a_{0,0,1},
\\
p_4a_{0,0,1} & = & \frac{1}{101745}p_2^4a_{0,0,0} + \frac{3159}{7820}p_2^2a_{0,0,1}
+ 1496a_{0,0,2} - \frac{4080}{437}p_2a_{0,2,0},
\\
p_4a_{0,0,2} & = & \frac{1}{450225}p_2^4a_{0,0,1} + \frac{24349}{61380}p_2^2a_{0,0,2}
+ \frac{45486360}{9889}a_{0,0,3}
\nn & &
+ \frac{8}{62031}p_2^3 a_{0,2,0} -
\frac{9360}{899}p_2a_{0,2,1} - \frac{13909320}{9889}a_{0,4,0},
\\
p_4a_{0,3,0} & = & -\frac{4}{89001}p_2^3a_{0,1,1} + \frac{36400}{37851}p_2a_{0,1,2}
+ \frac{619}{1276}p_2^2a_{0,3,0} + \frac{81900}{37}a_{0,3,1}.
\eea
Solving these and \eqref{eeq1}-\eqref{eeq7}, we obtain
\bea
a_{0, 1, 0} & = & -\frac{i}{135}p_3a_{0,0,0}, \\
a_{0, 2, 0} & = & \frac{-243 p_2^3 - 1748 p_3^2 + 684 p_2 p_4}{97977600}a_{0,0,0}, \\
a_{0, 3, 0} & = & \frac{i p_3 (207 p_2^3 + 580 p_3^2 - 540 p_2 p_4)}{29099347200}a_{0,0,0}, \\
a_{0, 4, 0} & = & [(37503 p_2^6 + 2834208 p_2^3 p_3^2 + 4492480 p_3^4 -
43092 p_2^4 p_4 - 7088256 p_2 p_3^2 p_4
\nn & &
- 688176 p_2^2 p_4^2 +
1373760 p_4^3)/366065131880448000]a_{0,0,0}.
\eea
This shows that $a_{n,m,l}$ are determined by $a_{0,0,0}$, which gives a proof of the conjecture
given in the previous section for $N=4$ case.
\section{Discussion}
We have computed the thermal partition function of SU($N$) $D=2$ matrix quantum mechanics, and from it we
have obtained some insight into the structure of the spectrum.
We have conjectured that CHS solution is the unique eigenstate of $\text{tr}\;[\Pi^2]$, $\text{tr}\;[\Pi^3], \dots$,
and $\text{tr}\;[\Pi^N]$ if we fix the eigenvalues, and have given proofs of it for $N=3$ and $N=4$ cases.
Our proofs of the uniqueness is too inefficient to apply to higher $N$ cases, and
it is desirable to invent more efficient method for investigating those cases.
From \eqref{mp2}-\eqref{mp4}, we can surmise that the following equation holds for any $n$:
\beq
\text{tr}\,[\Pi^n]\frac{1}{\cal M}f(X)\Big|_{X=Y} = \frac{(-i)^n}{\cal M}
\sum_{i=1}^N(\p_i)^nf(X)\Big|_{X=Y},
\eeq
and if this can be shown, CHS solution is proven to be an eigenfunction of
$\text{tr}\;[\Pi^2]$, $\text{tr}\;[\Pi^3], \dots$, and $\text{tr}\;[\Pi^N]$.
This system has been studied in the view of cut Fock space approach in \cite{k0910,cut},
which is a numerically tractable method, and some energy eigenfunctions
have been constructed. Relation between it and our analysis is not immediately clear to us.
It is important to make the relation clear and develop methods for numerical calculation more.
Although one may think that this system is too simplified for a model of Matrix theory,
it may give hint to the structure of asymptotic plane waves of Matrix theory, which is
important to consider scattering process\cite{petal}.
\renewcommand{\theequation}{\Alph{section}.\arabic{equation}}
|
\section{Introduction}
Spin-orbit coupling (SOC) has the potential to make novel
electronics applications possible,\cite{SOCapplication} as it
allows one to control the electron's spin degree of
freedom through its motion.
Most systems of interest for
such applications are nano- or mesoscopic in size, including
semiconductor quantum dots,\cite{rmtdotsreview} metallic
nanoparticles,\cite{npreview} and quantum
corrals defined on surfaces.\cite{qcreview} The energy spectrum,
and more generally, properties of these systems are (typically)
described by random matrix theory (RMT).\cite{rmtbook,chaosbook}
In RMT, the system's properties are a consequence of
its symmetries --- in the classic Wigner-Dyson
ensembles, the key symmetries are time-reversal ($T$)
and spin-rotation ($\sigma$) invariance.\cite{rmtbook,chaosbook}
With both $T$- and $\sigma$-invariance, the system
is described by the gaussian orthogonal ensemble (GOE);
SOC breaks the $\sigma$-invariance (while preserving
$T$-invariance), driving the system to the
gaussian symplectic ensemble (GSE).
[Systems with broken $T$-invariance are described by
the gaussian unitary ensemble (GUE).]
More specifically, systems with SOC are described by random
$N \times N$ matrices (with $N$$\rightarrow$$\infty$) having
quaternion components
\begin{equation}
H = S \otimes I_2 + i \frac{\lambda}{\sqrt{4N}}
\sum_{j=1}^3 A_j \otimes \sigma^j \, ,
\label{goegsematrix}
\end{equation}
where $S$ is an $N$$\times$$N$ symmetric matrix, and the
$\{ A_j \}$ are $N$$\times$$N$ antisymmetric matrices;
$\{ \sigma^j \}$ are the Pauli matrices, and $I_2$ is
the 2$\times$2 identity matrix.
$\lambda$ in Eq.\ \ref{goegsematrix} is related to the
SOC of the microscopic Hamiltonian --- $\lambda$=0 in the
GOE, while $\lambda$=$\sqrt{4N}$ in the GSE.
As most nanoscale systems of interest are described by
RMT,\cite{rmtdotsreview,npreview,qcreview,rmtbook,chaosbook}
it is important to understand the regimes/behaviors which
arise with SOC and the properties in these regimes.
In this work, we consider the spatial properties of wave
functions in (two-dimensional) chaotic nanoscale systems
with SOC.
The spatial properties of wave functions often determine
the system's response to experimental probes, and are
important for devices/applications.
\cite{conductanceexp,alhassid, conductance,annphys,nucleii,
rmttheoryexp,optics1,optics2}
While other works have discussed properties/consequences of
eigenvector statistics with SOC,\cite{grains,weakSO}
here we consider the spatial properties of eigenvectors
and, in particular, how these properties evolve with the
SOC.
In what follows, we consider the properties of the Hamiltonian
\begin{equation}
H = \frac{1}{2m} {\bf p}^2
+ \alpha~ \hat{z} \cdot \left(
{\bf p} \times \vec{\sigma} \right)
+ V({\bf r}) \ ,
\label{startinghammy}
\end{equation}
where $V({\bf r})$ is a confining and/or disorder potential.
Results obtained via RMT are compared with those obtained
by direct simulation of Eq.~\ref{startinghammy} for a
stadium billiard.\cite{bunimovich}
To characterize the system and understand its properties,
one- and two-point distribution functions were computed in
the crossover from the GOE to the GSE
(with increasing SOC).
In particular, it is found that excellent agreement between
RMT and microscopic simulations are obtained in a ``mean-field"
description of the (GOE-GSE) crossover (see below).
A key observation from our results is that correlations
of wave function amplitudes are suppressed with SOC.\@
Interestingly, however, these correlations play a more
important role in the two-point distribution function(s),
compared to the GOE (with vanishing SOC).
The rest of the paper is organized as follows.
The description of wave function statistics in RMT and, in
particular, the description of the GOE-GSE crossover is
discussed in Sec.~\ref{RMT}.
Details of our microscopic calculations --- namely the
stadium billiard considered as well as the numerical
approach employed --- are presented in Sec.~\ref{BWM}.
Our results are presented in Sec.~\ref{Res} --- one- and two-point
distribution functions obtained via RMT are compared with numerical
results from the stadium billiard.
Finally, Sec.\ \ref{concl} contains a summary of our results as
well as remarks on experimental consequences.
\section{Wave Function Statistics in RMT} \label{RMT}
In RMT, wave function correlations are governed by
the functional probability distribution\cite{srednicki,
localgaussian}
\begin{equation}
{\cal P}(\psi) = {\cal N} \exp\left[
-\frac{\beta}{2} \sum_{s,s'} \hspace{-0.046in}
\int \hspace{-0.046in} d{\bf r} d{\bf r}'
\psi^*_s({\bf r}) K_{s,s'} ({\bf r},{\bf r}')
\psi^{\phantom *}_{s'} ({\bf r}') \right] \, .
\label{fieldtheory}
\end{equation}
$K_{s,s'}({\bf r},{\bf r}')$
is the functional inverse of the two-point correlation
function $\langle \psi^*_s({\bf r})
\psi^{\phantom *}_{s'}({\bf r}') \rangle$,
where the angular brackets $\langle \cdots \rangle$
denote an average with respect to ${\cal P}(\psi)$;
the parameter $\beta$ depends on the system's symmetries
--- $\beta$=1 ($\beta$=2) in the GOE (GUE), while
$\beta$=4 in the GSE.
[${\cal N}$ is a normalization constant.]
${\cal P}(\psi)$ is the probability that a
particular energy eigenfunction with spin-$\sigma$
is equal to the specified function
$\psi_{\sigma}({\bf r})$.
A key property of
Eqs.\ \ref{goegsematrix} and \ref{startinghammy}
is their invariance
under time-reversal; as a result, the energy levels
are two-fold degenerate --- the eigenstates
$\{ \psi({\bf r}), {\cal T} \psi({\bf r}) \}$ are
degenerate, where ${\cal T}$ is the time-reversal operator.
Explicitly,
\begin{equation}
\psi({\bf r}) =
\left( \begin{array}{c}
\phi({\bf r}) \\ \chi({\bf r})
\end{array} \right)
, \ \
{\cal T} \psi({\bf r}) =
\left( \begin{array}{c}
-\chi^*({\bf r}) \\ \phi^*({\bf r})
\end{array} \right) .
\label{wavefunction}
\end{equation}
As a consequence of this two-fold degeneracy, the wave
function amplitude probed numerically and experimentally
is $|\psi_{\sigma}({\bf r})|^2$=$|\phi({\bf r})|^2
$+$|\chi({\bf r})|^2$.
As noted above, we are interested in the regimes/behaviors
which arise with SOC --- we will not only be interested in
the GSE, but also in the crossover from the GOE to the GSE.
As such, we decompose the complex wave functions
$\phi({\bf r})$ and $\chi({\bf r})$ in Eq.~\ref{wavefunction}
into their real and imaginary parts.
Then, the wave function amplitude is parameterized as
\begin{equation}
|\psi_{\sigma}({\bf r})|^2 = \gamma_1^2~ \phi^2_1({\bf r})
+ \gamma_2^2~ \phi^2_2({\bf r}) + \gamma_3^2~ \chi^2_1({\bf r})
+ \gamma_4^2~ \chi^2_2({\bf r}),
\label{wavefunctioncrossover}
\end{equation}
where the parameters $\{\gamma_i \}$,
which satisfy the constraint
$\gamma_1^2 + \gamma_2^2 + \gamma_3^2 + \gamma_4^2 = 1$,
characterize the crossover --- $\gamma_1$=1 with
$\gamma_i$=0 for $i \neq 1$ in the GOE, while
$\gamma_i$=$1/2$ ($i = 1 \cdots 4$) in the GSE;
in the crossover, the $\{ \gamma_i \}$ fluctuate and, hence,
physical quantities must be averaged over their distribution.
\begin{figure}[t]
\centering
\includegraphics[width=3.41in]{Fig1abcd}
\caption{(color online) Distribution of the $\{ \gamma_i \}$,
${\cal P}(\{ \gamma_i\})$ (from Eq.~\ref{wavefunctioncrossover}),
where $\lambda^*$=$\lambda \sqrt{4N}$.
(a) $\lambda^*=0.05$ (b) $\lambda^*=0.07$
(c) $\lambda^*=0.09$ (d) $\lambda^*=0.25$.
Notice all ${\cal P}$ change rapidly for $\lambda^* \lesssim 0.1$
and become sharply peaked at $\gamma \simeq 1/2$ for large $\lambda^*$.}
\label{fig:distribution}
\end{figure}
We obtained ${\cal P}(\{ \gamma_i \})$, the distribution of
the $\{ \gamma_i \}$, numerically from Eq.\ \ref{goegsematrix}
by considering the various orthogonal invariants\cite{goegse}
--- the results are shown in Fig.\ \ref{fig:distribution}.
We see that the ${\cal P}(\{ \gamma_i \})$ change rapidly
for $\lambda/\sqrt{4N} \lesssim 0.1$ --- in
particular, the ${\cal P}(\{ \gamma_i \})$ are broad for small
$\lambda$, but become sharply peaked gaussian-like for
larger values of $\lambda$, moving towards $\gamma_i$=1/2
($\forall~ i$) with increasing $\lambda$.
Figure \ref{fig:avgvariance} shows the variance of
the $\{ \gamma_i \}$, $\langle \gamma_i^2
\rangle$$-$$\langle \gamma_i \rangle^2$, as a function of
$\lambda$; the inset shows how the average values of the
$\{ \gamma_i \}$, $\langle \gamma_i \rangle$, evolve with
$\lambda$.
We see that the variance is extremely small for larger values
of $\lambda$; even for small values of $\lambda$ (where the
${\cal P}(\{ \gamma_i \})$ are broad and asymmetric), the
variance does not exceed 0.03.
As noted above, physical quantities must be averaged over the
${\cal P}(\{ \gamma_i \})$; however, as will be seen below,
rather good results are obtained in a ``mean-field" description
(due to the small variances), similar to what has been observed
in the GOE-GUE crossover\cite{annphys,whelan} --- rather good
results are obtained by approximating the $\{ \gamma_i \}$ by
their average values (rather than averaging over the
${\cal P}(\{ \gamma_i \}) )$.
\begin{figure}[t]
\centering
\includegraphics[width=3.44in]{Fig1ef}
\caption{(color online) Variance of $\{ \gamma_i \}$ vs.
$\lambda^*$=$\lambda \sqrt{4N}$.
Inset: average values of $\{ \gamma_i \}$ vs.\ $\lambda^*$.}
\label{fig:avgvariance}
\end{figure}
From Eq.\ \ref{fieldtheory}, all spatial correlations can be
obtained once the two-point correlation function
$\langle \psi^*_s({\bf r})
\psi^{\phantom *}_{s'}({\bf r}') \rangle$ is known.
To determine this, we expand the wave function as
\begin{equation}
\psi({\bf r}) = \sum_{\bf p}
\psi_{+,{\bf p}}({\bf r}) c_{+,{\bf p}}
+ \psi_{-,{\bf p}}({\bf r}) c_{-,{\bf p}} \ ,
\label{fourier}
\end{equation}
where the two-component spinors
$\psi_{+,{\bf p}}({\bf r})$ and $\psi_{+,{\bf p}}({\bf r})$
are eigenstates of Eq.~\ref{startinghammy} with
$V({\bf r})=0$. Explicitly, the eigenvalues
$\{ E_+,E_- \}$ and corresponding eigenstates
$\{ \psi_{+,{\bf p}}({\bf r}), \psi_{-,{\bf p}}({\bf r}) \}$
are ($\hbar = 1$)
\begin{equation}
E_{\pm} = \frac{|z|^2}{2m} \pm \alpha |z|; \
\psi_{\pm,{\bf p}}({\bf r}) = \frac{1}{\sqrt{2A}}
\left( \begin{array}{c}
1 \\ \pm i z/|z|
\end{array} \right)
e^{i {\bf p} \cdot {\bf r}}
\nonumber
\end{equation}
where $z$=$p_x$+$i$$p_y$.
The spectrum above describes
two spin-split chiral surfaces with energy $E$, shown
schematically in Fig.~\ref{fig:setup}a, where
$k_\pm = \sqrt{2mE+ m^2\alpha ^2} \mp m\alpha$.
To compute $\langle \psi^*_s({\bf r})
\psi^{\phantom *}_{s'}({\bf r}') \rangle$, the Fourier
coefficients (in Eq.~\ref{fourier}) are taken to be
gaussian random variables having zero mean and variance
given by\cite{alhassid} ($a$,$b$=+,$-$)
\begin{equation}
\langle c^*_{a, {\bf p}}
c^{\phantom *}_{b,{\bf k}} \rangle
= \delta_{a,b} \delta_{{\bf p},{\bf k}}
\frac{1}{N(\epsilon)} \delta(\epsilon({\bf p}) - \epsilon)
\, , \ \
\langle c_{a,{\bf p}} c_{b,{\bf k}} \rangle = 0 \ .
\label{conjecture}
\end{equation}
Writing the wave function as per Eq.~\ref{wavefunction}
and using the parameterization in Eq.~\ref{wavefunctioncrossover},
we obtain\cite{translation} ($i$,$j$=1,2)
\begin{subequations}
\begin{eqnarray}
& & \langle \phi_i({\bf r})
\phi^{\phantom *}_{j}({\bf r}') \rangle
= \langle \chi_i({\bf r})
\chi^{\phantom *}_{j}({\bf r}') \rangle
= \delta_{i,j} \, f \ ,
\label{correlator1} \\
& & \langle \phi_i({\bf r})
\chi^{\phantom *}_{j}({\bf r}') \rangle
= - \langle \chi_i({\bf r})
\phi^{\phantom *}_{j}({\bf r}') \rangle
= \delta_{i,j} \, g \ ,
\label{correlator2}
\end{eqnarray}
\end{subequations}
where
\begin{subequations}
\begin{eqnarray}
f & = & \frac{1}{2}
\left[ J_0(k_+ R) + J_0(k_- R) \right] ,
\label{correlatorf} \\
g & = & \frac{1}{2}
\left[ J_1(k_+ R) - J_1(k_- R) \right] .
\label{correlatorg}
\end{eqnarray}
\end{subequations}
In Eqs.~\ref{correlatorf} and \ref{correlatorg},
$J_0(x)$ ($J_1(x)$) is the Bessel function of order-0
(order-1),\cite{gradshteyn} $R$=$|{\bf r}-{\bf r}'|$, and
$k_{\pm}$ are the wave vectors associated with the
chiral branches at energy $E$.
The physics of Eq.~\ref{conjecture} (and Eqs.~\ref{correlator1}
and \ref{correlator2}) is that the system ergodically samples
the energy surfaces\cite{berry} (shown schematically in
Fig.\ \ref{fig:setup}a).
\begin{figure}[t]
\centering
\includegraphics[height=6.1cm,width=7cm]{Fig2ab}
\vspace{-0.35in} \hspace{-.11in}
\includegraphics[height=4.17cm]{Stadium_E405}
\caption{(color online) (a) Spin-split energy surfaces with
wave vectors $k_+$ and $k_-$.
(b) Stadium billiard considered in this work.
(c) Spatial scan of the LDOS of a typical chaotic
eigenfunction.}
\label{fig:setup}
\end{figure}
\section{Numerics} \label{BWM}
As described above, we are interested in comparing results
obtained via RMT with those obtained by direct simulation
of Eq.~\ref{startinghammy}. To this end, we have computed the
local density of states (LDOS) for a stadium
billiard,\cite{bunimovich} where the billiard's wall was
constructed with a unitary delta-function
potential\cite{boundarywall}
\begin{equation}
V({\bf r}) = V_0~ \delta\left( {\bf r} - {\bf R}(s) \right) \ ,
\end{equation}
with ${\bf R}(s)$ parameterizing the wall
(and $V_0 \rightarrow \infty$).
The retarded Green's function (GF) for the system,
\begin{equation}
G({\bf r}, {\bf r'}; \omega) = \langle {\bf r} |
\left( \omega - H + i0^{+} \right)^{-1} | {\bf r}' \rangle ,
\end{equation}
is computed from the Dyson equation
\begin{eqnarray}
G({\bf r}, {\bf r}'; \omega) & = &
G_0({\bf r}, {\bf r}'; \omega) \nonumber \\
& + & V_0~ \int_{\cal C} ds~ G_0({\bf r}, {\bf R}(s); \omega)~
G({\bf R}(s), {\bf r}'; \omega) \ .
\nonumber
\end{eqnarray}
In this equation, $G_0({\bf r}, {\bf r'}; \omega)$ is the
free-particle GF, i.e. the GF in the absence of the corral's
wall, but in the presence of SOC,\cite{spinorbitgreens,soccorral}
\begin{equation}
G_{0}({\bf r},{\bf r}';\omega) = G^{0}_0(R;\omega) ~I
+ G^{1}_0(R;\omega) \left( \begin{array}{c c}
0 & -i e^{-i\theta} \\
i e^{i\theta} & 0 \end{array} \right)
\nonumber
\end{equation}
where
\begin{eqnarray}
G^{0}_{0}(R;\omega) & = & -i \frac{m}{2k}
\left\{ k_{-} H^{(1)}_0 (R k_-)
+ k_{+} H^{(1)}_0 (R k_+) \right\} \ ,
\nonumber \\
G^{1}_{0}(R;\omega) & = & - \frac{m}{2k}
\left\{ k_{-} H^{(1)}_1 (Rk_-)
- k _{+} H^{(1)}_1 (Rk_+ ) \right\} \ ,
\nonumber \\
\end{eqnarray}
and
$\exp(i\theta) = [(x-x')+i(y-y')]/R$, with
$H^{(1)}_0(x)$ and $H^{(1)}_1(x)$ being Hankel
functions,\cite{gradshteyn}
with $R$=$|{\bf r}-{\bf r}'|$ and $k_\pm$ defined as before.
The LDOS is then obtained from the GF via
$A({\bf r},\omega) = -(1/\pi) {\rm I m} \, {\rm Tr}
\left[ G({\bf r}, {\bf r}; \omega) \right]$.
The stadium billiard we consider is shown schematically in
Fig.\ \ref{fig:setup}b. With energy in units of $E_0$=$
1/(2m R_0^2)$ and SOC in units of $\alpha_0$=1/$(mR_0)$,
where $R_0$ is the radius of the stadium's circular cap,
we have considered eigenstates with energy
$E$$\simeq$$405 E_0$, and have investigated SOCs in the
range $0$$\leq$$\alpha$$\leq 10\alpha_0$.
[Choosing $R_0$=70\AA, and $m$=$0.26m_e$ (with $m_e$ being the
electron's rest mass), one obtains
$\alpha_0$=$3.7\times10^{-11}$eVm, a value consistent with
e.g. electrons on an Au(111) surface.\cite{gold111,soccorral}]
A spatial scan of the LDOS for a typical eigenstate considered
is shown in Fig.\ \ref{fig:setup}c;
from the LDOS, one- and two-point distribution functions were
computed, going from the GOE to the GSE (with increasing $\alpha$).
\section{Results} \label{Res}
We now analyze the properties of the system, comparing
results from RMT with those obtained by direct simulation
of Eq.\ \ref{startinghammy} for a stadium billiard.
We begin by determining the regimes which arise as
function of the SOC strength.
To this end, we consider the one-point function
$ {\cal P}(\nu) = \langle \delta \left(
\nu - A |\psi_{\sigma} ({\bf r})|^2 \right) \rangle$,
which is obtained from Eq.~\ref{fieldtheory} by integrating
out the degrees of freedom except at ${\bf r}$.
Using Eq.\ \ref{wavefunctioncrossover}, we obtain
\begin{eqnarray}
& & {\cal P}(\nu) = \frac{\nu}
{4\gamma_1\gamma_2\gamma_3\gamma_4} \int_0^1 dz~
\label{goegsecrossover} \\ & & \hspace{0.087in} \times ~
\exp\left\{-\frac{\nu}{4} \left[ (1-z)
\left(\frac{1}{\gamma_1^2}+\frac{1}{\gamma_2^2} \right)
+ z \left(\frac{1}{\gamma_3^2}+\frac{1}{\gamma_3^2}
\right) \right] \right\}
\nonumber \label{1ptfunction}
\\ & & \hspace{0.087in} \times ~
I_0\left[\frac{\nu}{4} \left(\frac{1}{\gamma_1^2}
- \frac{1}{\gamma_2^2}\right) (1-z)\right]
I_0\left[\frac{\nu}{4} \left(\frac{1}{\gamma_3^2}
- \frac{1}{\gamma_4^2}\right) z \right] ,
\nonumber
\end{eqnarray}
where $I_0(x)$ is the modified Bessel function of
order-zero.\cite{gradshteyn}
This expression reduces to
${\cal P}_{\rm GOE}(\nu) = \exp(-\nu/2)/\sqrt{2\pi \nu}$ in
the GOE ($\gamma_1$=1 and $\{\gamma_i\}$=0 for $i$$\neq$1)
and ${\cal P}_{\rm GSE}(\nu) = 4\nu \exp (-2 \nu)$ in
the GSE ($\gamma_i$=1/2 $\forall i$).
\begin{figure}[b]
\includegraphics[width=3.41in]{Fig3}
\caption{(color online)
$ {\cal P}(\nu) = \langle \delta \left( \nu
- A |\psi_{\sigma} ({\bf r})|^2 \right) \rangle$.
From top to bottom:
$\alpha$=0 (GOE), $\alpha$=0.2,
$\alpha$=0.5, $\alpha$=1.5 (GSE).
Each curve has been vertically offset by one unit
for clarity.}
\label{fig:onept}
\end{figure}
Figure \ref{fig:onept} shows numerical results for
${\cal P}(\nu)$ for different values of the SOC; the
results are compared with Eq.~\ref{goegsecrossover}
in a ``mean-field" description, i.e.
with the $\{ \gamma_i \}$ evaluated at their average
values --- for $\alpha$=0.2$\alpha_0$ ($\alpha$=0.5$\alpha_0$),
we find $\lambda = 0.04\sqrt{4N}$
($\lambda = 0.08\sqrt{4N}$).\cite{fitting}
The physics of Eq.~\ref{startinghammy} is determined by
its two length scales --- the spin-flip length
$l_{\rm sf}$=$1/(m\alpha)$ and the linear dimension
of the system $L$ ($\simeq R_0$).
Figure \ref{fig:onept} shows how the system evolves toward
the GSE as the SOC is increased.
In particular, we find the system to be in the GSE for
$\alpha \gtrsim 1.5 \alpha_0$ i.e. $\l_{\rm sf} \lesssim 2R_0/3$;
once the system is in this GSE regime, the statistics
do not change further as the SOC is increased.
We now turn to spatial correlations of eigenfunctions.
We first consider the amplitude correlator
${\cal C}_{\sigma \sigma'}({\bf r},{\bf r}')
= \langle A |\psi_{\sigma}({\bf r})|^2
A |\psi_{\sigma '}({\bf r}')|^2 \rangle$.
Using the parameterization in Eq.~\ref{wavefunctioncrossover},
we obtain
\begin{eqnarray}
& & {\cal C}_{\sigma \sigma'}({\bf r},{\bf r}')
= 1 + 2 \left[ (\gamma_1^4 + \gamma_2^4
+ \gamma_3^4 +\gamma_4^4) f^2
\right. \nonumber \\ & & \left. \hspace{1in}
+ 2 \left( \gamma_1^2 \gamma_3^2 + \gamma_2^2 \gamma_4^2
\right) g^2 \right] \ .
\label{doscorrelator}
\end{eqnarray}
Notice that this reduces to
${\cal C}_{\sigma \sigma '}^{\rm GOE} ({\bf r}, {\bf r}')
= 1 + 2f^2$ in the GOE, and to
${\cal C}_{\sigma \sigma '}^{\rm GSE} ({\bf r}, {\bf r}')
= 1 + {\cal V}^2/2$ in the GSE
where ${\cal V}^2$=$f^2$+$g^2$.
Numerical results for
${\cal C}_{\sigma \sigma '} ({\bf r}, {\bf r}')$
are shown in Fig.\ \ref{fig:doscorrelate}
and are compared with Eq.~\ref{doscorrelator},
with the $\{ \gamma_i \}$ evaluated at their average
values (as before).
We see that the maximum is larger in the GOE; more generally,
the correlations decay more rapidly with SOC --- amplitude
correlations are suppressed as $\sigma$-invariance is
broken.
\begin{figure}[t]
\includegraphics[width=3.41in]{Fig4}
\caption{(color online) Amplitude correlator
${\cal C}_{\sigma \sigma'}({\bf r},{\bf r}')
= \langle A |\psi_{\sigma}({\bf r})|^2
A |\psi_{\sigma '}({\bf r}')|^2 \rangle$.
From top to bottom:
$\alpha$=0 (GOE), $\alpha$=0.2$\alpha_0$,
$\alpha$=0.5$\alpha_0$, $\alpha$=1.5$\alpha_0$ (GSE).
Each curve has been vertically offset by 1/4 unit for
clarity.}
\label{fig:doscorrelate}
\end{figure}
Having determined the parameter regimes and, in particular,
how large the SOC must be to be in the GSE, we now consider
in greater detail the properties of the system in the GSE.
To this end, we consider the joint distribution function
${\cal P}(\nu_1,\nu_2) = \left\langle
\delta \left( \nu_1 - A | \psi_{\sigma}({\bf r}) |^2 \right)
\delta \left( \nu_2 - A | \psi_{\sigma'}({\bf r}') |^2 \right)
\right\rangle$,
which is obtained from Eq.~\ref{fieldtheory} by integrating
out the degrees of freedom except those at
${\bf r}$ and ${\bf r}'$.
For the GSE we obtain
\begin{equation}
{\cal P}_{\rm GSE}(\nu_1, \nu_2) =
\frac{8\sqrt{\nu_1 \nu_2}}{{\cal V}(1-{\cal V}^2)}
\exp \left( -2 {\cal X}_S \right)
I_1\left( 4 {\cal X}_P \right) \ ,
\label{jointsymp}
\end{equation}
where $I_1(x)$ is the modified Bessel function of
order-$1$;\cite{gradshteyn}
for comparison, we also consider
${\cal P}(\nu_1,\nu_2)$ in the GOE\cite{srednicki}
\begin{equation}
{\cal P}_{\rm GOE}(\nu_1, \nu_2) =
\frac{\exp \left( - {\cal X}_S/2 \right)
\cosh \left( {\cal X}_P \right)}
{2\pi \sqrt{1-f^2} \sqrt{\nu_1 \nu_2} } \ .
\nonumber
\end{equation}
In the above equations,
${\cal X}_S$=$(\nu_1 + \nu_2)/(1-{\cal X}^2)$
and
${\cal X}_P$=${\cal X}\sqrt{\nu_1 \nu_2}/(1-{\cal X}^2)$
where ${\cal X}$=${\cal V}$ (${\cal X}$=$f$) for the GSE (GOE).
We now consider the properties and consequences of
${\cal P}({\nu_1},{\nu_2})$.
We begin by considering the conditional probability
\begin{equation}
{\cal P}_{\nu_1}({\nu_2})
= {\cal P}(\nu_1,\nu_2) / {\cal P}(\nu_1)
\label{conditional}
\end{equation}
which describes the wave function distribution at ${\bf
r}_2$, provided $A |\psi({\bf r}_1)|^2 = \nu_1$.
It follows from Eq.~\ref{conditional} that correlations
between fluctuations at different points depend on their
amplitudes\cite{prigodin} --- regions of high amplitude
(i.e. large $\nu_1$) are correlated over larger distances,
while regions of small amplitude are correlated over shorter
distances.
${\cal P}^{\rm GSE}_{\nu_1}({\nu_2})$ for the GSE is
shown in Fig.~\ref{fig:conditional}a for several values
of ${\cal V}=\sqrt{f^2+g^2}$;
${\cal P}^{\rm GOE}_{\nu_1}({\nu_2})$ for the GOE is
shown in Fig.~\ref{fig:conditional}b for comparison,
for several values of $f$.
\begin{figure}[t]
\centering
\includegraphics[width=3.5in]{Fig6}
\caption{(color online) Conditional probability
(a) ${\cal P}^{\rm GSE}_{\nu_1}({\nu_2})$, and
(b) ${\cal P}^{\rm GOE}_{\nu_1}({\nu_2})$, for
several values of ${\cal V}$ ($f$), for $\nu_1=10$.}
\label{fig:conditional}
\end{figure}
From Eq.~\ref{conditional}, one can obtain the average
$\langle \nu_2 \rangle_{\nu_1}$ and the mean squared
fluctuation $\langle (\delta\nu_2)^2 \rangle_{\nu_1}
= \langle \nu_2^2 \rangle^{\phantom 2}_{\nu_1}
- \langle \nu_2 \rangle^2_{\nu_1}$,
where $\langle \cdots \rangle_{\nu_1}$ denotes an average
with respect to $ {\cal P}_{\nu_1}({\nu_2})$:
\begin{eqnarray}
\langle \nu_2 \rangle_{\nu_1} & = & 1
+ {\cal X}^2(\nu_1 - 1) , \label{conditionalavg} \\
\langle (\delta \nu_2)^2 \rangle_{\nu_1}
& = & {\cal C} \left[ 1 + 2{\cal X}^2(\nu_1 - 1)
+ {\cal X}^4 (1 - 2\nu_1) \right] , \nonumber
\end{eqnarray}
where ${\cal C}$=2 for the GOE,\cite{rmttheoryexp}
while ${\cal C}$=1/2 for the GSE. [As before,
${\cal X}$=${\cal V}$ ($f$) for the GSE (GOE).]
From this, we see that fluctuations are suppressed in the
GSE compared to the GOE.
More generally, fluctuations are largest in the GOE
(compared with the GUE\cite{prigodin} and the
GSE, Eq.~\ref{conditionalavg}) and, hence,
correlations are the weakest.
We now consider the distribution of the product
$A |\psi_{\sigma}({\bf r})
\psi_{\sigma '}({\bf r}')|$,
${\cal P}(\Gamma) = \langle \delta \left( \Gamma
- A|\psi_{\sigma}({\bf r}) \psi_{\sigma'}({\bf r}')|
\right) \rangle$.
${\cal P}(\Gamma)$ determines a number of experimentally
relevant quantities, such as the form factor
in resonant scattering in
complex nucleii,\cite{nucleii} amplitudes in tunneling measurements,
and the conductance amplitude distribution through small quantum
dots.\cite{conductance}
From Eq.~\ref{jointsymp}, we obtain for the GSE
\begin{equation}
{\cal P}_{\rm GSE}(\Gamma) =
\frac{32~ \Gamma^{2}}{|{\cal V}|(1-{\cal V}^2)}
I_1\left( \frac{4|{\cal V}|~\Gamma}{1-{\cal V}^2}\right)
K_0 \left( \frac{4~\Gamma}{1-{\cal V}^2} \right) \ ,
\label{proddistribute}
\end{equation}
where $K_0(x)$ is a modified Bessel functions of
order-zero;\cite{gradshteyn}
in the GOE, we obtain
\begin{equation}
{\cal P}_{\rm GOE}(\Gamma) =
\frac{2}{\pi \sqrt{1-f^2}}
K_0 \left( \frac{\Gamma}{1-f^2} \right)
\cosh \left( \frac{f~\Gamma}{1-f^2} \right) \ .
\nonumber
\end{equation}
\begin{figure}[t]
\centering
\includegraphics[height=2.8in]{Fig7}
\caption{(color online) Product distribution
${\cal P}(\Gamma) = \langle \delta \left( \Gamma
- A|\psi_{\sigma}({\bf r}) \psi_{\sigma'}({\bf r}')|
\right) \rangle$
in the GSE for several values of ${\cal V}$.
For comparison, ${\cal P}_{\rm GOE}(\Gamma)$ is also shown for $f=0.5$.
Inset: Comparison of numerical and RMT results for $R=0.055R_0$.}
\label{fig:twopoint}
\end{figure}
Figure \ref{fig:twopoint} shows results for ${\cal P}(\Gamma)$
for several values of ${\cal V}$ ($f$) for the GSE (GOE).
We see that the maximum of ${\cal P}_{\rm GSE}(\Gamma)$
decreases with increasing ${\cal V}$ with the tail becoming
slightly longer.
For comparison, ${\cal P}_{\rm GOE}(\Gamma)$ is shown for
different values of $f$.
We see that correlations play a more significant role in the
GSE --- indeed, except for a very small region near $\Gamma$=0,
${\cal P}_{\rm GOE}(\Gamma)$ is essentially indistinguishable
from the result with $f$$\rightarrow$0.
This is a consequence of the fact that fluctuations are largest
in the GOE and correlations are the weakest.
Shown in the inset are numerical results for
${\cal P}_{\rm GSE}(\Gamma)$ for $R$=$0.055R_0$
in comparison with the RMT result, Eq.~\ref{proddistribute}.
\section{Concluding Remarks} \label{concl}
To summarize,
we have investigated the statistical properties of wave functions
in (two-dimensional) chaotic nanostructures with spin-orbit
interactions, focussing particularly on spatial correlations
of eigenfunctions.
Numerical results obtained for a chaotic stadium billiard
were compared with (analytic) results from RMT.
It was found that excellent agreement between RMT and microscopic
simulations are obtained in a ``mean-field" description of the
GOE-GSE crossover.
A key observation from our results is that correlations of
wave function amplitudes are suppressed with SOC.\@
Interestingly, however, these correlations with SOC play
a more significant role in the two-point distribution
function(s).
Our results have implications for a number of systems of
current interest.
Indeed, the effects of SOC have been observed in transport
through quantum dots.\cite{socdot}
These effects could also be observed in quantum corrals defined
on Au and Ag (111) surfaces,\cite{soccorral}
where large SOC has been observed recently,\cite{goldsurface}
especially as scanning tunneling microscopy techniques have
exquisite control of positioning and correlation measurements.
\section*{Acknowledgements}
We acknowledge helpful conversations with H.~U. Baranger
and P.~W. Brouwer.
EHK acknowledges the warm hospitality of the
Instituto de F\'isica T\'eorica (Madrid), where most of
this work was performed.
This work was supported in Ohio by NSF-DMR MWN/CIAM and
NSF-PIRE grants.
|
\section{Main result and discussion}\label{sec:i}
We consider the fractional Laplacian $\Delta^{\alpha/2}$
defined by
\begin{equation}\label{eq:fraclapl}
\Delta^{\alpha/2} u(x) = {\mathcal{A}_{d,-\alpha}} \lim_{\varepsilon\to 0^+}
\int_{\mathbb{R}^d \cap \{|y-x|>\varepsilon\}}\frac{u(y)-u(x)}{|x-y|^{d+\alpha}} \,dy.
\end{equation}
Here ${\mathcal{A}_{d,-\alpha}} = \frac{2^\alpha \Gamma(\frac{\alpha+d}{2})}{ \pi^{d/2} |\Gamma(-\frac{\alpha}{2})|}$.
This is an important operator for probability and analysis
\cite{MR2569321}, \cite{Landkof}.
The main results of this paper are the following formulae for the fractional Laplacian
of power functions.
\begin{thm}\label{thm:up}
Let $d\geq 1$, $0<\alpha<2$ and $p>-1$. Define
\begin{align}
u_p^{(d)}(x) &= (1-|x|^2)_+^p, \quad x\in \mathbb{R}^d,\label{upn}\\
v_p^{(d)}(x) &= (1-|x|^2)_+^p x_d, \quad x\in \mathbb{R}^d,\label{vpn}\\
\Phi^{(d)}_{p,\alpha}(x) &= \frac{{\mathcal{A}_{d,-\alpha}} B(-\frac{\alpha}{2}, p+1)\pi^{d/2}}{\Gamma(\frac{d}{2})} \;\;
{} _2F_1\Big(\frac{\alpha+d}{2}, -p+\frac{\alpha}{2}; \frac{d}{2}; x \Big).
\end{align}
If $x\in \mathbb{R}^d$ and $|x|<1$, then
\begin{align}
\Delta^{\alpha/2} u_p^{(d)}(x) &=\Phi^{(d)}_{p,\alpha}(|x|^2),\label{Deltaupn}\\
\Delta^{\alpha/2} v_p^{(d)}(x) &= x_d \Phi^{(d+2)}_{p,\alpha}(|x|^2).\label{Deltavpn}
\end{align}
\end{thm}
By $B$ and $\Gamma$ we denote Euler's beta and gamma functions,
respectively, and $_2F_1$ is Gauss' hypergeometric function \cite{Erdelyi}.
Some special cases of the above theorem were known before. A calculation
of $\Delta^{\alpha/2} u_{\frac{\alpha}{2}}^{(d)}$ was done by Getoor \cite{MR0137148}, and
explicit expression for
$\Delta^{\alpha/2} u_{\frac{\alpha}{2}-1}^{(d)}$ and $\Delta^{\alpha/2} v_{\frac{\alpha}{2}-1}^{(d)}$
may be deduced from \cite{Hmissi}, see also \cite{BogdanMartin}.
A formula for $\Delta^{\alpha/2} u_p^{(d)}$ with $p\in (0,1)$ and $\alpha\in (-1,1)$ is announced
in \cite{BilerImbertKarch}. Finally, (\ref{Deltaupn}) was given in \cite{DydaHardy1dim}
for $p=\frac{\alpha-1}{2}$ and $d=1$, where it was used to obtain
a fractional Hardy inequality with a remainder term.
For similar results on fractional derivatives we refer the reader to \cite{MR1347689}.
We note that if $p=n+\alpha/2$ and $n$ is a nonnegative integer,
then $\Delta^{\alpha/2} u_p^{(d)}$ and $ \Delta^{\alpha/2} v_p^{(d)}$ are
polynomials of degree $n$ and $n+1$, respectively.
For $d=1$, every polynomial can be expressed as a linear combination of polynomials
of the form $(1-x^2)^j$ and $x(1-x^2)^k$. Hence, if $P$ is \emph{any} polynomial of degree $n$, then
$\Delta^{\alpha/2}(P(x)(1-x^2)^{\alpha/2})$, for $x\in (-1,1)$, is a polynomial of degree $n$, too.
This also trivially holds for $\alpha=2$.
This may allow one for the development of an explicit integro--differential calculus for $\Delta^{\alpha/2}(\cdot 1_{|x|<1})$,
which is one of the motivations for this work.
It is noteworthy that $u_p^{(d)}(x) = {} _2F_1\Big(\frac{d}{2}, -p; \frac{d}{2}; |x|^2 \Big)$ for~\mbox{$|x|<1$},
therefore (\ref{Deltaupn}) may be rewritten in the following elegant form
\begin{equation}
\Delta^{\alpha/2}\left({} _2F_1\Big(\frac{d}{2}, -p; \frac{d}{2}; |x|^2 \Big) 1_{|x|<1}\right) =
C_{d,\alpha,p}\;\;
{} _2F_1\Big(\frac{d}{2} + \frac{\alpha}{2}, -p+\frac{\alpha}{2}; \frac{d}{2}; |x|^2 \Big),
\end{equation}
where $x\in\mathbb{R}^d$, $|x|<1$, $C_{d,\alpha,p}={{\mathcal{A}_{d,-\alpha}} B(-\frac{\alpha}{2}, p+1)\pi^{d/2}}/{\Gamma(\frac{d}{2})}$,
$d=1,2,\ldots$ and $\alpha\in (0,2)$.
Let $B=B_1^{(d)}=\{x\in \mathbb{R}^d:|x|<1\}$ be the unit ball in $\mathbb{R}^d$.
We consider the quadratic form
\[
\mathcal{E}(u)=\frac{{\mathcal{A}_{d,-\alpha}}}{2}
\int_{\mathbb{R}^d\times \mathbb{R}^d}
\frac{(u(x)-u(y))^2}{|x-y|^{d+\alpha}} \,dx\,dy,
\]
with the domain
\[
D(\mathcal{E})=\{u\in L^2(\mathbb{R}^d): \mathcal{E}(u)<\infty,\; \textrm{$u=0$ on $B^c$}\},
\]
and the corresponding symmetric bilinear form $\mathcal{E}(\cdot, \cdot)$.
Then $(\mathcal{E}(\cdot, \cdot), D(\mathcal{E}))$ is the Dirichlet form of the $\alpha$-stable, rotation
invariant L\'evy process killed upon leaving $B$.
To point out applications of Theorem~\ref{thm:up}, we consider the spectral problem
of finding $\phi\in D(\mathcal{E})$ such that
\begin{equation}\label{spprob}
\mathcal{E}(\phi,g) = \lambda \int \phi(x) g(x)\,dx, \quad g \in D(\mathcal{E}).
\end{equation}
It is known \cite{Getoor1959a} that there
exists an orthonormal basis of $L^2(B)\subset L^2(\mathbb{R}^d)$, consisting of eigenfunctions $\phi_1$, $\phi_2$, $\phi_3$, \ldots
with the corresponding eigenvalues
$0<\lambda_1<\lambda_2\leq \lambda_2 \leq \ldots$.
The latter means that $\phi=\phi_n$ and $\lambda=\lambda_n$ satisfy (\ref{spprob}).
We note that these eigenfunctions are in the domain of the generator, and we have that
$\Delta^{\alpha/2}\phi_n = -\lambda_n \phi_n$ on $B$ in the sense of definition (\ref{eq:fraclapl}),
see also \cite{MR1825645}. When there is a risk of confusion, we write the dimension of the underlying
space in superscripts, i.e., we write $\lambda_n^{(d)}$ for $\lambda_n$ and $\phi_n^{(d)}$ for $\phi_n$.
The eigenproblem is the main motivation for this research.
We should note that
the eigenvalues $\lambda_1$, $\lambda_2$, \ldots are not known explicitly even in the case of $d=1$ and $B=(-1,1)$.
A number of methods to study this one-dimensional case, and more general cases, were developed by
several authors
\cite{MR2056835}, \cite{MR2214145}, \cite{MR2273977}, \cite{MR2217951}, \cite{MR1840105},
\cite{Getoor1959b}, \cite{MR2158176}, \cite{MR2078980}, \cite{MR0147937}, \cite{MR0212621}, \cite{MR2679702},
\cite{KwasnickiInterval}, \cite{MR2445510}, \cite{MR2534231}, \cite{MR1119514}, \cite{MR2139213}.
The symmetry of eigenfunctions plays an important role in these investigations.
We also note that this spectral problem
may be formulated in terms of the isotropic $\alpha$-stable L\'evy process in $\mathbb{R}^d$.
For details we refer the reader to \cite{MR2322501}.
Let $\lambda_*$ be the smallest number such that there
exists an eigenfunction $\phi_*$ which is antisymmetric, $\phi_*(-x)=-\phi_*(x)$, and has eigenvalue $\lambda_*$.
It is conjectured, but not yet proved in the full range of $\alpha\in (0,2)$ and $d$, that $\lambda_*=\lambda_2$.
In the classical case ($\alpha=2$), and also in the $1$--dimensional case
for $\alpha\geq 1$ \cite{MR2679702}, we do have $\lambda_*=\lambda_2$.
It is natural to ask whether $\lambda_* = \lambda_2$.
While we do not answer this question here, the calculus of power functions given by
Theorem~\ref{thm:up} may be used to investigate the eigenfunctions of the fractional Laplacian
in the ball in this and related problems.
We should note that there always exists an antisymmetric eigenfunction.
Indeed, there exists a non-symmetric eigenfunction $\phi$, and we may
let $\tilde{\phi}(x)=\phi(x)-\phi(-x)$, which is an antisymmetric
eigenfunction with the same eigenvalue as $\phi$.
Similarly $\phi_*$ may and will be assumed antisymmetric
with respect to the last coordinate axis.
The similarity between \eqref{Deltaupn} and \eqref{Deltavpn} leads us to the following result.
\begin{prop}\label{prop:lambdaeq}
$\lambda_*^{(d)} = \lambda_1^{(d+2)}$.
\end{prop}
Noteworthy, for $p=n+\alpha/2$ the right-hand sides of (\ref{Deltaupn}) and (\ref{Deltavpn}) are
polynomials, which gives an efficient way to explicitly estimate
$\lambda_1$ and $\lambda_*$ using some versions of Barta inequality \cite{MR2139213}, see Sections~\ref{sec:lower} and \ref{sec:upper}.
In particular, we obtain the following corollary.
\begin{cor}\label{cor:est}
Let
\begin{equation}\label{eq:mu}
\mu_{d,\alpha} = \frac{2^\alpha \Gamma(\frac{\alpha}{2}+1) \Gamma(\frac{\alpha+d}{2}) (\alpha+2)(\alpha+d) (6-\alpha)}
{ \Gamma(\frac{d}{2})(12d+(16-2d)\alpha)}.
\end{equation}
We have
\begin{equation}
\lambda_1 \geq \mu_{d,\alpha}
\end{equation}
and
\begin{equation}
\lambda_* \geq \mu_{d+2,\alpha}.
\end{equation}
\end{cor}
We note that the above result improves estimates from
\cite{MR1119514} and \cite[Corollary 1]{MR1840105} by a~factor $(\alpha+2)(\alpha+1)(6-\alpha)/(12+14\alpha) > 1$,
and it also improves some of the estimates from \cite{KwasnickiInterval}.
The paper is structured as follows. In Section~\ref{sec:dim1} we prove Theorem~\ref{thm:up}
for $d=1$, then in Section~\ref{sec:dimd} we prove the $d$-dimensional case.
In Sections~\ref{sec:lower} and \ref{sec:upper} we prove Proposition~\ref{prop:lambdaeq} and
we derive lower and upper bounds for the eigenvalues
$\lambda_1$ and~$\lambda_*$.
\section{One--dimensional case}\label{sec:dim1}
The following technical result is the first step in the proof of Theorem~\ref{thm:up}.
\begin{lem}\label{lem:pv}
If $p>-1$, $0<\alpha<2$ and $x\in (-1,1)$, then
\begin{align}
I_m(p) &:= {p.v.}\int_{-1}^1 \frac{(1-tx)^{\alpha-m-2p}-1}{|t|^{1+\alpha}}\;
(1-t^2)^p\,dt \nonumber\\
&= B\Big(-\frac{\alpha}{2}, p+1\Big)
\left( {} _2F_1\Big(-\frac{\alpha}{2}, p+m-\frac{1}{2}-\frac{\alpha}{2}; \frac{1}{2}; x^2\Big)
-1 \right),\label{eq:pv}
\end{align}
where $m=1$ or $m=2$, and $p.v.$ means the Cauchy principal value.
\end{lem}
\begin{proof}
We assume that
$p\neq\frac{\alpha-2}{2}$, since otherwise the beta function on the right-hand side of (\ref{eq:pv}) is zero
and the result is obvious.
We have
\begin{align*}
I_m(p) &=
\, {p.v.}\int_{-1}^1 \frac{(1-tx)^{\alpha-m-2p}-1}{|t|^{1+\alpha}}\;
(1-t^2)^p\,dt \\
&=
\int_{-1}^1 \sum_{k=2}^\infty \frac{(2p+m-\alpha)_k (tx)^k}{ k! |t|^{1+\alpha}}\;
(1-t^2)^p\,dt \\
&=
2 \int_{0}^1 \sum_{k=1}^\infty \frac{ (2p+m-\alpha)_{2k} (tx)^{2k}}{ (2k)! |t|^{1+\alpha}}\;
(1-t^2)^p\,dt \\
&=\bigg(\sum_{k=0}^\infty \frac{ (2p+m-\alpha)_{2k} B\Big(k-\frac{\alpha}{2}, p+1\Big)}{ (2k)!} x^{2k}\bigg) -
B\Big(-\frac{\alpha}{2}, p+1\Big)\\
&=:S_m - B\Big(-\frac{\alpha}{2}, p+1\Big).
\end{align*}
Here $(a)_n$ is the Pochhammer symbol, that is, $(a)_0=1$ and $(a)_n = a(a+1)\ldots (a+n-1)$ for $n=1,2,\ldots$.
We observe that
\begin{equation}\label{eq:wz1}
(2p+m-\alpha)_{2k} = 2^{2k} \left(p+\frac{m}{2}-\frac{\alpha}{2}\right)_k \left(p+\frac{m+1}{2}-\frac{\alpha}{2}\right)_k,
\end{equation}
and, by the doubling formula,
\begin{equation}\label{doubling}
\Gamma(2x)=\Gamma(x)\Gamma(x+1/2)2^{2x-1}/\Gamma(1/2),
\end{equation}
applied to $2x=2k+1$, we have
\begin{equation}\label{eq:wz2}
(2k)! = 2^{2k} {\textstyle \left(\frac{1}{2}\right)\!_k }k!.
\end{equation}
We also have,
\begin{equation}\label{eq:wz3}
B\Big(k-\frac{\alpha}{2}, p+1\Big) = \frac{\left(-\frac{\alpha}{2}\right)_k }{ \left(p+1-\frac{\alpha}{2}\right)_k}B\Big(-\frac{\alpha}{2}, p+1\Big).
\end{equation}
Thus, by (\ref{eq:wz1}), (\ref{eq:wz2}) and (\ref{eq:wz3}) we obtain
\begin{align*}
S_m&= B\Big(-\frac{\alpha}{2}, p+1\Big)
\sum_{k=0}^\infty \frac{\left(-\frac{\alpha}{2}\right)_k \left(p+\frac{m}{2}-\frac{\alpha}{2}\right)_k \left(p+\frac{m+1}{2}-\frac{\alpha}{2}\right)_k
}{ \left(p+1-\frac{\alpha}{2}\right)_k \left(\frac{1}{2}\right)_k k! }\, x^{2k}.
\end{align*}
For $m=1$ or $m=2$ the factor $\left(p+1-\frac{\alpha}{2}\right)_k$ in the denominator cancels with one of the terms
in the numerator, and the result follows.
\end{proof}
In the following lemma we prove (\ref{Deltaupn}) for $d=1$.
Recall that
\begin{equation}\label{up}
u_p(x) = (1-x^2)_+^p,\quad x\in\mathbb{R}.
\end{equation}
\begin{lem}\label{lem:Deltaup}
If $0<\alpha<2$, $p>-1$ and $x\in (-1,1)$ then
\begin{align}\label{Deltaup}
\Delta^{\alpha/2} u_p(x) &={\mathcal{A}_{1,-\alpha}} B\Big(-\frac{\alpha}{2}, p+1\Big)
{} _2F_1\Big(-\frac{\alpha}{2}, p+\frac{1}{2}-\frac{\alpha}{2}; \frac{1}{2}; x^2\Big)
(1-x^2)^{p-\alpha}\\
&={\mathcal{A}_{1,-\alpha}} B\Big(-\frac{\alpha}{2}, p+1\Big)
{} _2F_1\Big(\frac{\alpha+1}{2}, -p+\frac{\alpha}{2}; \frac{1}{2}; x^2\Big).\label{Deltaup2}
\end{align}
\end{lem}
\begin{proof}
We recall from \cite[Lemma 2.1]{DydaHardy1dim} the following formula
\begin{align}
L u_p(x) &=
\frac{(1-x^2)^{p-\alpha}}{\alpha} \Bigg(
(1 - x)^\alpha + (1 + x)^\alpha
\nonumber\\
&
- (2p+2-\alpha)B(p+1,1-\alpha/2)
+\alpha I_1(p)
\Bigg), \label{Lup}
\end{align}
where $I_1(p)$ is given by (\ref{eq:pv}).
By $(2p+2-\alpha)B(p+1,1-\alpha/2) = - \alpha B(p+1,-\alpha/2)$ and Lemma~\ref{lem:pv},
\begin{align*}
\alpha I_1(p) - (2p+2-\alpha)&B(p+1,1-\alpha/2)\\
&=\alpha B\Big(-\frac{\alpha}{2}, p+1\Big) {} _2F_1\Big(-\frac{\alpha}{2}, p+\frac{1}{2}-\frac{\alpha}{2}; \frac{1}{2}; x^2\Big).
\end{align*}
This proves (\ref{Deltaup}). The formula (\ref{Deltaup2}) follows from \cite[formula 2.9(2), page 105]{Erdelyi}.
\end{proof}
\begin{table}[ht]
\caption{The fractional Laplacian for some functions vanishing outside of $(-1,1)$.}
\centering
\begin{tabular}{c c}
\hline\hline
$u(x)$ on $(-1,1)$ & $\Delta^{\alpha/2} u(x)$ for $x\in(-1,1)$ \\ [0.5ex]
\hline
$(1-x^2)^{-1+\alpha/2}$ & $0$ \\
$(1-x^2)^{\alpha/2}$ & $-\Gamma(\alpha+1)$ \\
$(1-x^2)^{1+\alpha/2}$ & $-\Gamma(\alpha+1) \frac{\alpha+2}{2}(1-(1+\alpha)x^2)$ \\
$(1-x^2)^{2+\alpha/2}$ & $-\Gamma(\alpha+1)\frac{\alpha+2}{2}\frac{\alpha+4}{4}
\Big( 1-(2\alpha+2)x^2 + (\frac{\alpha}{3}+1)(\alpha+1)x^4 \Big)$\\
\hline\hline
\end{tabular}
\label{table:values}
\end{table}
In the following lemma we prove (\ref{Deltavpn}) for $d=1$.
\begin{lem}\label{lem:Deltavp}
Let $p>-1$ and $v_p(x) = x(1-x^2)_+^p$ for $x\in \mathbb{R}$.
For $0<\alpha<2$ and $x\in (-1,1)$ we have
\begin{align}
\Delta^{\alpha/2} v_p(x) &={\mathcal{A}_{1,-\alpha}} B\Big(-\frac{\alpha}{2}, p+1\Big) (\alpha+1)\label{Deltavp}\\
&\quad \times {} _2F_1\Big(-\frac{\alpha}{2}, p+\frac{3}{2}-\frac{\alpha}{2}; \frac{3}{2}; x^2\Big)
x(1-x^2)^{p-\alpha}\nonumber\\
&={\mathcal{A}_{1,-\alpha}} B\Big(-\frac{\alpha}{2}, p+1\Big) (\alpha+1) \;\;
{} _2F_1\Big(\frac{\alpha+3}{2}, -p+\frac{\alpha}{2}; \frac{3}{2}; x^2\Big)x.\label{Deltavp2}
\end{align}
\end{lem}
\begin{proof}
We write
\begin{align*}
\mathcal{A}^{-1}_{1,-\alpha}&\Delta^{\alpha/2} v_p(x) \\
&=p.v. \int_{-1}^1 \frac{y(1-y^2)^p-x(1-x^2)^p}{|y-x|^{1+\alpha}}\,dy
- v_p(x) \int_{\mathbb{R}\setminus(-1,1)} \frac{dy}{|y-x|^{1+\alpha}}\\
&=p.v. \int_{-1}^1 \frac{y(1-y^2)^p-x(1-x^2)^p}{|y-x|^{1+\alpha}}\,dy
- \frac{v_p(x)}{\alpha} \left(\frac{1}{(x+1)^\alpha} + \frac{1}{(1-x)^\alpha} \right) \\
&=: I - \frac{v_p(x)}{\alpha} \left(\frac{1}{(x+1)^\alpha} + \frac{1}{(1-x)^\alpha} \right).
\end{align*}
To evaluate $I$, we change the variable to $t=\frac{x-y}{1-xy}$, see \cite[the proof of Lemma~2.1]{DydaHardy1dim}
for more details. We obtain
\begin{align*}
I &=
(1-x^2)^{p-\alpha}
p.v.\int_{-1}^1 \frac{(1-t^2)^p (x-t) - x(1-tx)^{2p+1}}{|t|^{1+\alpha}}\;
(1-tx)^{\alpha-2-2p}\,dt\\
&=
(1-x^2)^{p-\alpha} \Bigg[
x p.v.\int_{-1}^1 \frac{(1-tx)^{\alpha-2p-2} - 1}{|t|^{1+\alpha}}(1-t^2)^p\,dt\\
& + x p.v.\int_{-1}^1 \frac{(1-t^2)^p - 1}{|t|^{1+\alpha}}\,dt
+ x p.v.\int_{-1}^1 \frac{1 - (1-tx)^{\alpha-1}}{|t|^{1+\alpha}}\,dt\\
& - p.v.\int_{-1}^1 \frac{(1-t^2)^p t(1-tx)^{\alpha-2p-2}}{|t|^{1+\alpha}}\,dt \Bigg].
\end{align*}
We have by \cite[the proof of Lemma~2.1]{DydaHardy1dim}
\begin{align*}
p.v.\int_{-1}^1 \frac{(1-t^2)^p - 1}{|t|^{1+\alpha}}\,dt &= \frac{2}{\alpha}\left[1-(p+1-\alpha/2)B(p+1,1-\alpha/2)\right],\\
p.v.\int_{-1}^1 \frac{1 - (1-tx)^{\alpha-1}}{|t|^{1+\alpha}}\,dt &=
\frac{1}{\alpha}\Big( 1 - x \Big)^\alpha + \frac{1}{\alpha}\Big( 1 + x \Big)^\alpha - \frac{2}{\alpha}.
\end{align*}
By Lemma~\ref{lem:pv} we obtain
\begin{align*}
p.v.\int_{-1}^1 &\frac{(1-tx)^{\alpha-2p-2} - 1}{|t|^{1+\alpha}}(1-t^2)^p\,dt
=I_2(p)\\
& = B\Big(-\frac{\alpha}{2}, p+1\Big)
\left( {} _2F_1\Big(-\frac{\alpha}{2}, p+\frac{3}{2}-\frac{\alpha}{2}; \frac{1}{2}; x^2\Big) - 1\right).
\end{align*}
For $p\neq\frac{\alpha-2}{2}$ we have
\begin{align*}
K&:= p.v.\int_{-1}^1 \frac{(1-t^2)^p t(1-tx)^{\alpha-2p-2}}{|t|^{1+\alpha}}\,dt \\
&=p.v.\int_{-1}^1 \frac{(1-t^2)^p t}{|t|^{1+\alpha}}\sum_{k=0}^\infty \frac{(2p+2-\alpha)_k}{k!} (tx)^k \,dt \\
&=\sum_{k=0}^\infty 2 \int_0^1 \frac{(1-t^2)^p t}{|t|^{1+\alpha}} \frac{(2p+2-\alpha)_{2k+1}}{ (2k+1)!} (tx)^{2k+1} \,dt \\
&=\sum_{k=0}^\infty B(p+1,\frac{2k+2-\alpha}{2}) \frac{(2p+2-\alpha)_{2k+1}}{ (2k+1)!} x^{2k+1}.
\end{align*}
Using the doubling formula (\ref{doubling}) for $x=k+1$, we obtain
\[
(2k+1)! = 2^{2k+1} {\textstyle \left(\frac{1}{2}\right)\!_{k+1} }k!.
\]
We also have
\[
(2p+2-\alpha)_{2k+1} = 2^{2k+1} (p+1-\frac{\alpha}{2})_{k+1} (p+\frac{3}{2}-\frac{\alpha}{2})_{k}\;,
\]
and
\[
B(p+1,\frac{2k+2-\alpha}{2}) = B(-\frac{\alpha}{2}, p+1) \frac{ (-\frac{\alpha}{2})_{k+1}}{ (p+1-\frac{\alpha}{2})_{k+1}}.
\]
Hence
\begin{align*}
K&= B(-\frac{\alpha}{2}, p+1) \sum_{k=0}^\infty
\frac{ (-\frac{\alpha}{2})_{k+1} (p+\frac{3}{2}-\frac{\alpha}{2})_{k}}
{ \left(\frac{1}{2}\right)\!_{k+1} k! } x^{2k+1}.\\
&= -\alpha B(-\frac{\alpha}{2}, p+1) \sum_{k=0}^\infty
\frac{ (1-\frac{\alpha}{2})_{k} (p+\frac{3}{2}-\frac{\alpha}{2})_{k}}
{ \left(\frac{3}{2}\right)\!_{k} k! } x^{2k+1}\\
&=-\alpha B\Big(-\frac{\alpha}{2}, p+1\Big) x \cdot {}_2F_1\Big(1-\frac{\alpha}{2}, p+\frac{3}{2}-\frac{\alpha}{2}; \frac{3}{2}; x^2\Big).
\end{align*}
This holds also for $p=\frac{\alpha-2}{2}$, since in this case we have $K=0$.
Thus,
\begin{align*}
\Delta^{\alpha/2} v_p(x)& =\mathcal{A}_{1,-\alpha}
B\Big(-\frac{\alpha}{2}, p+1\Big) (1-x^2)^{p-\alpha} x \\
&\quad \times
\left( {} _2F_1\Big(-\frac{\alpha}{2}, p+\frac{3}{2}-\frac{\alpha}{2}; \frac{1}{2}; x^2\Big)
+ \alpha \;\; {}_2F_1\Big(1-\frac{\alpha}{2}, p+\frac{3}{2}-\frac{\alpha}{2}; \frac{3}{2}; x^2\Big) \right).
\end{align*}
Formula (\ref{Deltavp}) follows by \cite[formula 2.8(35), page 103]{Erdelyi},
and (\ref{Deltavp2}) is then a consequence of \cite[formula 2.9(2), page 105]{Erdelyi}.
\end{proof}
In Table~\ref{table:valuesvp} we list the fractional Laplacian of a few functions $v_p$ .
We note that the first example in Table~\ref{table:valuesvp} may be considered
a~linear combination of Martin kernels of the interval, see \cite[(89)]{MR2365478} and the references given there.
\begin{table}[ht]
\caption{The fractional Laplacian for some functions, when defined as zero outside of $(-1,1)$.}
\centering
\begin{tabular}{c c}
\hline\hline
$v(x)$ on $(-1,1)$ & $\Delta^{\alpha/2} v(x)$ for $x\in(-1,1)$ \\ [0.5ex]
\hline
$x(1-x^2)^{-1+\alpha/2}$ & $0$ \\
$x(1-x^2)^{\alpha/2}$ & $-\Gamma(\alpha+2)x$ \\
$x(1-x^2)^{1+\alpha/2}$ & $- \frac{\Gamma(\alpha+3)}{6}(3-(3+\alpha)x^2)x$ \\
$x(1-x^2)^{2+\alpha/2}$ & $- \frac{\Gamma(\alpha+3)(\alpha+4)}{120}
\Big( 15-(10\alpha+30)x^2 + (\alpha +3)(\alpha+5)x^4 \Big)x$\\
\hline\hline
\end{tabular}
\label{table:valuesvp}
\end{table}
\section{Multi-dimensional case}\label{sec:dimd}
We recall the notation (\ref{upn}) and note that $u_p^{(d)}(x) = u_p(|x|)$ for $x\in \mathbb{R}^d$,
with $u_p$ given by (\ref{up}).
We let $S^{(d-1)} = \{x\in \mathbb{R}^d:|x|=1\}$, the unit sphere in $\mathbb{R}^d$.
\begin{lem}\label{lem:Deltaradial}
Let $d\geq 2$, $0<\alpha<2$ and $p>-1$. If $x\in \mathbb{R}^d$ and $|x|<1$, then
\begin{equation}\label{eq:Deltaradial}
\Delta^{\alpha/2} u_p^{(d)}(x) =
\frac{{\mathcal{A}_{d,-\alpha}}}{2{\mathcal{A}_{1,-\alpha}}} \int_{S^{d-1}} (1- |x|^2 + |h_dx|^2)^{p-\alpha/2}
\Delta^{\alpha/2} u_p\bigg( \frac{|x|h_d}{\sqrt{1- |x|^2 + |h_dx|^2}} \bigg) \, dh.
\end{equation}
\end{lem}
\begin{proof}
Without loss of generality, we may assume that $x = (0,0,\ldots,0,|x|)$.
For $|x|<1$ we have,
\begin{align*}
\Delta^{\alpha/2}u_p^{(d)}(x) &= {\mathcal{A}_{d,-\alpha}} p.v. \int_{\mathbb{R}^d} \frac{u_p^{(d)}(y) - (1-|x|^2)^p}
{|x-y|^{d+\alpha}}\,dy\\
&=\frac{{\mathcal{A}_{d,-\alpha}}}{2} \int_{S^{d-1}} dh
\; p.v.\int_{\mathbb{R}}
\frac{ u_p^{(d)}(x+h t) - (1-|x|^2)^p}{|t|^{1+\alpha}} \,dt.
\end{align*}
We calculate the (inner) principal value integral by changing the variable
$t=-|x|h_d + s \sqrt{|h_dx|^2-|x|^2+1}$. We obtain
\begin{align*}
g&(x,h):=
p.v.\int_{\mathbb{R}}
\frac{ u_p^{(d)}(x+h t)- (1-|x|^2)^p}{|t|^{1+\alpha}} \,dt\\
&=p.v. \int_{\mathbb{R}}
\frac{ (1-s^2)^p(1-|x|^2+|h_dx|^2)^p - (1-|x|^2)^p }
{|-|x|h_d + s \sqrt{|h_dx|^2-|x|^2+1}|^{1+\alpha}}\,
\sqrt{|h_dx|^2-|x|^2+1}\,ds\\
&=
(1-|x|^2+|h_dx|^2)^{p-\alpha/2}
p.v.\int_{\mathbb{R}} \frac{ u_p(s) - (1- \frac{|h_dx|^2}{1-|x|^2+|h_dx|^2})^p }
{ |s - \frac{|x|h_d}{\sqrt{1-|x|^2+|h_dx|^2}}|^{1+\alpha}}\,ds\\
&=(1-|x|^2+|h_dx|^2)^{p-\alpha/2} {\mathcal{A}_{1,-\alpha}^{-1}}
\Delta^{\alpha/2} u_p\Big(\frac{|x|h_d}{\sqrt{1-|x|^2+|h_dx|^2}}\Big)
\end{align*}
\end{proof}
\begin{proof}[Proof of formula (\ref{Deltaupn}) of Theorem~\ref{thm:up} for $d>1$]
We have by Lemmata~\ref{lem:Deltaup} and~\ref{lem:Deltaradial},
\begin{align*}
\Delta^{\alpha/2} u_p^{(d)}(x) &=
\frac{{\mathcal{A}_{d,-\alpha}} B(-\frac{\alpha}{2}, p+1)}{2} \int_{S^{d-1}}
\frac{
{} _2F_1\Big(\frac{\alpha+1}{2}, -p+\frac{\alpha}{2}; \frac{1}{2}; \frac{|x|^2h_d^2}{1- |x|^2 + |h_dx|^2} \Big)
}{
(1- |x|^2 + |h_dx|^2)^{-p+\alpha/2}
}
\,dh\\
&=: \frac{{\mathcal{A}_{d,-\alpha}} B(-\frac{\alpha}{2}, p+1)}{2} I_{S^{d-1}}.
\end{align*}
We transform the integrand function using \cite[formula 2.9(4), page 105]{Erdelyi},
\begin{align*}
\frac{
{} _2F_1\Big(\frac{\alpha+1}{2}, -p+\frac{\alpha}{2}; \frac{1}{2}; \frac{|x|^2h_d^2}{1- |x|^2 + |h_dx|^2} \Big)
}{
(1- |x|^2 + |h_dx|^2)^{-p+\alpha/2}
} &= \frac{
{} _2F_1\Big(-\frac{\alpha}{2}, -p+\frac{\alpha}{2}; \frac{1}{2}; \frac{|x|^2h_d^2}{|x|^2-1} \Big)
}{
(1- |x|^2)^{-p+\alpha/2}
}.
\end{align*}
Therefore,
\begin{align*}
I_{S^{d-1}} &= \int_{S^{d-1}}
\frac{
{} _2F_1\Big(-\frac{\alpha}{2}, -p+\frac{\alpha}{2}; \frac{1}{2}; \frac{|x|^2h_d^2}{|x|^2-1} \Big)
}{
(1- |x|^2)^{-p+\alpha/2}
}\,dh \\
&= \frac{ 2\pi^{\frac{d-1}{2}}(1- |x|^2)^{p-\alpha/2}}{\Gamma(\frac{d-1}{2})}
\int_{-1}^1
{} _2F_1\Big(-\frac{\alpha}{2}, -p+\frac{\alpha}{2}; \frac{1}{2}; \frac{|x|^2h^2}{|x|^2-1} \Big)
(1-h^2)^{\frac{d-3}{2}} \,dh.
\end{align*}
Let
\[
\phi(z)=\int_{-1}^1
{} _2F_1\Big(-\frac{\alpha}{2}, -p+\frac{\alpha}{2}; \frac{1}{2}; \frac{z h^2}{z-1} \Big)
(1-h^2)^{\frac{d-3}{2}} \,dh, \quad z\in \mathbb{C}, |z|<1.
\]
Since $\re \frac{z h^2}{z-1} < \frac{1}{2}$ for $|z|<1$, the function $\phi$ is analytic
in the unit disc $\{z:|z|<1\}$.
For $|z|<\frac{1}{2}$ we calculate the integral defining $\phi$ by using Taylor's expansion,
\begin{align*}
\phi(z)&=
\sum_{k=0}^\infty \frac{\Gamma(-\frac{\alpha}{2}+k)}{\Gamma(-\frac{\alpha}{2})}
\frac{\Gamma(-p+\frac{\alpha}{2}+k)}{\Gamma(-p+\frac{\alpha}{2})}
\frac{\Gamma(\frac{1}{2})}{\Gamma(\frac{1}{2}+k) k!}
\left( \frac{z}{z-1} \right)^k
\int_{-1}^1 h^{2k}(1-h^2)^{\frac{d-3}{2}}\,dh\\
&=
\frac{\Gamma(\frac{d-1}{2}) \Gamma(\frac{1}{2}) }{ \Gamma(\frac{d}{2}) }
\sum_{k=0}^\infty \frac{\Gamma(-\frac{\alpha}{2}+k)}{\Gamma(-\frac{\alpha}{2})}
\frac{\Gamma(-p+\frac{\alpha}{2}+k)}{\Gamma(-p+\frac{\alpha}{2})}
\frac{\Gamma(\frac{d}{2})}{\Gamma(\frac{d}{2}+k) k!}
\left( \frac{z}{z-1} \right)^k\\
&=
\frac{\Gamma(\frac{d-1}{2}) \Gamma(\frac{1}{2}) }{ \Gamma(\frac{d}{2}) }
{} _2F_1\Big(-\frac{\alpha}{2}, -p+\frac{\alpha}{2}; \frac{d}{2}; \frac{z}{z-1} \Big)\\
&=\
\frac{\Gamma(\frac{d-1}{2}) \Gamma(\frac{1}{2}) }{ \Gamma(\frac{d}{2}) }
{} _2F_1\Big(\frac{\alpha+d}{2}, -p+\frac{\alpha}{2}; \frac{d}{2}; z \Big) (1-z)^{-p+\alpha/2}
=:\psi(z).
\end{align*}
In the last line we have used \cite[formula 2.9(4), page 105]{Erdelyi} again.
The functions $\phi$ and $\psi$ are both analytic in the unit disc, hence $\phi(z)=\psi(z)$
for all $|z|<1$. We put $z=|x|^2$ and the proof is finished.
\end{proof}
\begin{lem}\label{lem:DeltaradialV}
Let $d\geq 2$, $0<\alpha<2$ and $p>-1$. If $x\in \mathbb{R}^d$ and $|x|<1$, then
\begin{align}\label{eq:DeltaradialV}
\Delta^{\alpha/2} v_p^{(d)}(x) &= x_d \Delta^{\alpha/2} u_p^{(d)}(x)\\
&+
\frac{{\mathcal{A}_{d,-\alpha}}}{2{\mathcal{A}_{1,-\alpha}}} \int_{S^{d-1}} T^{p-\alpha/2} h_d
\left(
T^{1/2} \Delta^{\alpha/2} v_p
-\langle h,x\rangle \Delta^{\alpha/2} u_p \right)
\bigg( \frac{\langle h,x\rangle}{\sqrt{T}} \bigg) \, dh, \nonumber
\end{align}
where $T = T(x,h)=1- |x|^2 + \langle h,x\rangle^2$.
\end{lem}
\begin{proof}
We have for $|x|<1$,
\begin{align*}
\Delta^{\alpha/2}v_p^{(d)}(x) &= {\mathcal{A}_{d,-\alpha}} p.v. \int_{\mathbb{R}^d} \frac{v_p^{(d)}(y) - v_p^{(d)}(x)}
{|x-y|^{d+\alpha}}\,dy\\
&=\frac{{\mathcal{A}_{d,-\alpha}}}{2} \int_{S^{d-1}} dh
\; p.v.\int_{\mathbb{R}}
\frac{ v_p^{(d)}(x+h t) - v_p^{(d)}(x) }{|t|^{1+\alpha}} \,dt.
\end{align*}
We calculate the inner principal value integral by changing the variable
$t=-\langle h,x\rangle + s \sqrt{T}$. We obtain
\begin{align*}
g(x,h)&:=
p.v.\int_{\mathbb{R}}
\frac{ v_p^{(d)}(x+h t)- v_p^{(d)}(x)}{|t|^{1+\alpha}} \,dt\\
&=p.v. \int_{\mathbb{R}}
\frac{ (1-s^2)_+^p T^p (x_d-h_d \langle h,x\rangle+h_ds\sqrt{T}) - (1-|x|^2)^px_d }
{|-\langle h,x\rangle + s \sqrt{T}|^{1+\alpha}}\,
\sqrt{T}\,ds\\
&=
T^{p-\alpha/2}
p.v.\int_{\mathbb{R}} \frac{ (1-s^2)_+^p (x_d-h_d\langle h,x\rangle+h_ds\sqrt{T}) - (1- \frac{\langle h,x\rangle^2}{T})^px_d }
{|s - \frac{\langle h,x\rangle}{\sqrt{T}}|^{1+\alpha}}\,ds\\
&=
T^{p-\alpha/2} (x_d-h_d\langle h,x\rangle)
p.v.\int_{\mathbb{R}} \frac{ (1-s^2)_+^p - (1- \frac{\langle h,x\rangle^2}{T})^p }
{|s - \frac{\langle h,x\rangle}{\sqrt{T}}|^{1+\alpha}}\,ds\\
&\quad +T^{p+1/2-\alpha/2} h_d
p.v.\int_{\mathbb{R}} \frac{ (1-s^2)_+^ps - (1- \frac{\langle h,x\rangle^2}{T})^p \frac{\langle h,x\rangle}{\sqrt{T}} }
{|s - \frac{\langle h,x\rangle}{\sqrt{T}}|^{1+\alpha}}\,ds\\
&=T^{p-\alpha/2} \frac{x_d-h_d\langle h,x\rangle}{{\mathcal{A}_{1,-\alpha}}} \Delta^{\alpha/2} u_p\Big(\frac{\langle h,x\rangle}{\sqrt{T}}\Big)\\
&\quad +
T^{p+1/2-\alpha/2} \frac{h_d}{{\mathcal{A}_{1,-\alpha}}} \Delta^{\alpha/2} v_p\Big(\frac{\langle h,x\rangle}{\sqrt{T}}\Big).
\end{align*}
The result follows from Lemma~\ref{lem:Deltaradial}.
\end{proof}
\begin{proof}[Proof of formula (\ref{Deltavpn}) of Theorem~\ref{thm:up} for $d>1$]
We may assume that $x\neq 0$, since for $x=0$ the formula is obvious.
We denote $T = T(x,h)=1- |x|^2 + \langle h,x\rangle^2$.
By Lemmata~\ref{lem:Deltavp} and~\ref{lem:DeltaradialV},
\begin{align*}
\Delta^{\alpha/2} v_p^{(d)}(x) &= x_d \Delta^{\alpha/2} u_p^{(d)}(x)
+
\frac{{\mathcal{A}_{d,-\alpha}} B(-\frac{\alpha}{2}, p+1) }{2} \times \\
&\qquad \times \int_{S^{d-1}}
h_d\langle h,x\rangle T^{p-\alpha/2}
F(x,h)
\,dh,
\end{align*}
where
\[
F(x,h) =
(\alpha+1) {} _2F_1\Big(\frac{\alpha+3}{2}, -p+\frac{\alpha}{2}; \frac{3}{2}; \frac{\langle h,x\rangle^2}{T} \Big)
- _2F_1\Big(\frac{\alpha+1}{2}, -p+\frac{\alpha}{2}; \frac{1}{2}; \frac{\langle h,x\rangle^2}{T} \Big).
\]
We transform $F(x,h)$ using \cite[formula 2.8(35), page 103]{Erdelyi} and
\cite[formula 2.9(4), page 105]{Erdelyi},
\begin{align*}
F(x,h) &=
\alpha \cdot {} _2F_1\Big(\frac{\alpha+1}{2}, -p+\frac{\alpha}{2}; \frac{3}{2}; \frac{\langle h,x\rangle^2}{T} \Big)\\
&=
\alpha
\bigg(\frac{1-|x|^2}{T}\bigg)^{p-\alpha/2}
{} _2F_1\Big(1 - \frac{\alpha}{2}, -p+\frac{\alpha}{2}; \frac{3}{2}; \frac{\langle h,x\rangle^2}{|x|^2-1} \Big).
\end{align*}
Hence,
\begin{align*}
\int_{S^{d-1}}&
h_d\langle h,x\rangle T^{p-\alpha/2} F(x,h)\,dh\\
&=
\alpha
(1-|x|^2)^{p-\alpha/2}
\int_{S^{d-1}}
{} _2F_1&\Big(1-\frac{\alpha}{2}, -p+\frac{\alpha}{2}; \frac{3}{2}; \frac{\langle h,x\rangle^2}{|x|^2-1} \Big) h_d\langle h,x\rangle
\,dh\\
&=:\alpha
(1-|x|^2)^{p-\alpha/2} I_{S^{d-1}}.
\end{align*}
By symmetry, for any $e_1$, $e_2\in \mathbb{R}^d$ with $|e_1|=|e_2|=1$,
\[
\int_{S^{d-1}} f_1(\langle h,e_1\rangle) f_2(\langle h,e_2\rangle)\,dh =
\int_{S^{d-1}} f_1(\langle h,e_2\rangle) f_2(\langle h,e_1\rangle)\,dh,
\]
where $f_1$ and $f_2$ are any functions for which the integrals make sense.
Using this observation for $e_1=\frac{x}{|x|}$ and $e_2=(0,\ldots,0,1)$ we obtain
\begin{align*}
I_{S^{d-1}} &=
\int_{S^{d-1}}
{} _2F_1\Big(1-\frac{\alpha}{2}, -p+\frac{\alpha}{2}; \frac{3}{2}; \frac{ h_d^2 |x|^2}{|x|^2-1} \Big) \langle h,x\rangle h_d
\,dh\\
&=
\int_{S^{d-1}}
{} _2F_1\Big(1-\frac{\alpha}{2}, -p+\frac{\alpha}{2}; \frac{3}{2}; \frac{ h_d^2 |x|^2}{|x|^2-1} \Big) h_d^2 x_d
\,dh\\
&= \frac{2\pi^{\frac{d-1}{2}} x_d}{\Gamma(\frac{d-1}{2})}
\int_{-1}^1
{} _2F_1\Big(1-\frac{\alpha}{2}, -p+\frac{\alpha}{2}; \frac{3}{2}; \frac{ h^2 |x|^2}{|x|^2-1} \Big) h^2 (1-h^2)^{\frac{d-3}{2}}\,dh.
\end{align*}
Let
\[
\phi(z) = \int_{-1}^1
{} _2F_1\Big(1-\frac{\alpha}{2}, -p+\frac{\alpha}{2}; \frac{3}{2}; \frac{ h^2 z}{z-1} \Big) h^2 (1-h^2)^{\frac{d-3}{2}}\,dh,
\quad z\in \mathbb{C}, |z|<1.
\]
Similarly as in the proof of formula (\ref{Deltaupn}) we observe that $\phi$ is analytic in the unit disc,
and calculate $\phi(z)$ for $|z|<\frac{1}{2}$ by using Taylor's expansion,
\begin{align*}
\phi(z) &=\int_{-1}^1 \sum_{k=0}^\infty \frac{ (1-\frac{\alpha}{2})_k (-p+\frac{\alpha}{2})_k \Gamma(\frac{3}{2}) }
{\Gamma(k+\frac{3}{2}) k!} \bigg(\frac{z}{z-1}\bigg)^k h^{2k+2}(1-h^2)^{\frac{d-3}{2}}\,dh\\
&=\frac{ \Gamma(\frac{3}{2}) \Gamma(\frac{d-1}{2})}{\Gamma(\frac{d}{2}+1)}
\sum_{k=0}^\infty \frac{ (1-\frac{\alpha}{2})_k (-p+\frac{\alpha}{2})_k \Gamma(\frac{d}{2}+1) }{ \Gamma(k+\frac{d}{2}+1) k!}
\bigg(\frac{z}{z-1}\bigg)^k\,dh\\
&=\frac{ \Gamma(\frac{3}{2}) \Gamma(\frac{d-1}{2})}{\Gamma(\frac{d}{2}+1)}
{} _2F_1\Big(1-\frac{\alpha}{2}, -p+\frac{\alpha}{2}; 1+\frac{d}{2}; \frac{z}{z-1} \Big).
\end{align*}
Since the function in the last line is analytic in the unit disc (note that $\re \frac{z}{z-1}<\frac{1}{2}$ if $|z|<1$),
we conclude that
\begin{align*}
I_{S^{d-1}} &= \frac{\pi^{\frac{d}{2}} x_d}{\Gamma(\frac{d}{2}+1)}\;
{} _2F_1\Big(1-\frac{\alpha}{2}, -p+\frac{\alpha}{2}; 1+\frac{d}{2}; \frac{|x|^2}{|x|^2-1} \Big)\\
&= \frac{\pi^{d/2}x_d}{\Gamma(1+\frac{d}{2})}
\; {} _2F_1\Big(\frac{\alpha+d}{2}, -p+\frac{\alpha}{2}; 1+\frac{d}{2}; |x|^2 \Big)
(1-|x|^2)^{-p+\alpha/2}.
\end{align*}
In the last line we have used \cite[formula 2.9(4), page 105]{Erdelyi}.
By \cite[formula 2.8(35), page 103]{Erdelyi} we get
\begin{align*}
\Delta^{\alpha/2} v_p^{(d)}(x) &=
x_d \frac{{\mathcal{A}_{d,-\alpha}}\pi^{d/2} B(-\frac{\alpha}{2}, p+1) }{\Gamma(d/2)}
\bigg(
\; {} _2F_1\Big(\frac{\alpha+d}{2}, -p+\frac{\alpha}{2}; \frac{d}{2}; |x|^2 \Big)\\
&\qquad\qquad\qquad\qquad\qquad\quad
+ \frac{\alpha}{d} \; {} _2F_1\Big(\frac{\alpha+d}{2}, -p+\frac{\alpha}{2}; 1+\frac{d}{2}; |x|^2 \Big)
\bigg)\\
&=x_d \frac{{\mathcal{A}_{d,-\alpha}}\pi^{d/2} B(-\frac{\alpha}{2}, p+1) (\alpha+d) }{d\Gamma(d/2)} \times\\
&\qquad\times
\; {} _2F_1\Big(\frac{\alpha+d+2}{2}, -p+\frac{\alpha}{2}; 1+\frac{d}{2}; |x|^2 \Big),
\end{align*}
and the proof is finished.
\end{proof}
In Table~\ref{table:valuesd} we list the fractional Laplacian for some power functions in $\mathbb{R}^d$.
\begin{table}[ht]
\caption{Values of fractional Laplacian for some functions vanishing outside the unit ball in $\mathbb{R}^d$.}
\centering
\begin{tabular}{c c}
\hline\hline
$f(x)$ in the ball & $\Delta^{\alpha/2} f(x)$ in the ball \\ [0.5ex]
\hline
$(1-|x|^2)^{\alpha/2}$ & $-2^\alpha\Gamma(\frac{\alpha}{2}+1) \Gamma(\frac{d+\alpha}{2}) \Gamma(\frac{d}{2})^{-1}$ \\
$(1-|x|^2)^{1+\alpha/2}$ & $-2^\alpha\Gamma(\frac{\alpha}{2}+2) \Gamma(\frac{d+\alpha}{2})\Gamma(\frac{d}{2})^{-1}
\left(1 - (1+\frac{\alpha}{d})|x|^2\right)$\\
\hline
$(1-|x|^2)^{\alpha/2}x_d$ & $-2^\alpha\Gamma(\frac{\alpha}{2}+1) \Gamma(\frac{d+\alpha}{2}+1) \Gamma(\frac{d}{2}+1)^{-1}x_d$ \\
$(1-|x|^2)^{1+\alpha/2}x_d$ & $-2^\alpha\Gamma(\frac{\alpha}{2}+2) \Gamma(\frac{d+\alpha}{2}+1)\Gamma(\frac{d}{2}+1)^{-1}
\left(1 - (1+\frac{\alpha}{d+2})|x|^2\right)$\\
\hline\hline
\end{tabular}
\label{table:valuesd}
\end{table}
\section{Lower bounds for eigenvalues}\label{sec:lower}
Our approach to lower bounds is similar to that of \cite{KBBD-bc}.
The method is to use a suitable superharmonic function $\nu$ and
the resulting Hardy inequality with the weight given by the Fitzsimmons' ratio
${-\Delta^{\alpha/2} \nu }/{\nu}$ (\cite{Fitz}).
To estimate $\lambda_1$, the following calculation will be used.
\begin{lem}\label{lem:updest}
Let $\eta_{d,\alpha}=(6d-4+(4-d)\alpha-\alpha^2)/(\alpha+2)^2$ and
\[
\psi(x) = (1-|x|^2)_+^{\alpha/2} + \eta_{d,\alpha} (1-|x|^2)_+^{1+\alpha/2},\quad x\in \mathbb{R}^d.
\]
Then
\[
\frac{-\Delta^{\alpha/2} \psi(x)}{\psi(x)} \geq \mu_{d,\alpha},
\quad |x|<1,
\]
where $\mu_{d,\alpha}$ is defined in (\ref{eq:mu}).
\end{lem}
\begin{proof}
We have by Table~\ref{table:valuesd}
\[
\frac{-\Delta^{\alpha/2} \psi(x)}{\psi(x)} =
\frac{2^\alpha \Gamma(\frac{\alpha}{2}+1) \Gamma(\frac{\alpha+d}{2}) }{ \Gamma(\frac{d}{2})(1-|x|^2)^{\alpha/2}}
\frac{1 + \eta \frac{\alpha+2}{2}(1-(1+\frac{\alpha}{d})|x|^2)}{1+\eta(1-|x|^2)} =: f(|x|^2).
\]
After elementary but tedious calculations we obtain
\[
f'(x)=\frac{2^\alpha \Gamma(\frac{\alpha}{2}+1) \Gamma(\frac{\alpha+d}{2}) \alpha(\alpha+d)(\alpha+2) \eta^2}
{ 4d\Gamma(\frac{d}{2}) (1-x)^{1+\alpha/2}(1+\eta(1-x))^2 }
\left(x-\frac{(4-d)\alpha + 6d-8}{6d-4 + (4-d)\alpha-\alpha^2}\right)^2
\geq 0
\]
for $x\in [0,1)$. Hence
\[
\frac{-\Delta^{\alpha/2} \psi(x)}{\psi(x)} \geq \frac{-\Delta^{\alpha/2} \psi(0)}{\psi(0)}
=
\frac{2^\alpha \Gamma(\frac{\alpha}{2}+1) \Gamma(\frac{\alpha+d}{2}) (\alpha+2)(\alpha+d) (6-\alpha)}
{ \Gamma(\frac{d}{2})(12d+(16-2d)\alpha)}.
\]
\end{proof}
We will also consider the ratio ${-\Delta^{\alpha/2}\nu}/{\nu}$ for a~suitable antisymmetric function $\nu$.
We should note that $\nu$ changes sign and so it cannot be called superharmonic.
However, the Fitzsimmons ratio ${-\Delta^{\alpha/2}\nu}/{\nu}$ will prove nonnegative, and we will
obtain a Hardy inequality with the weight ${-\Delta^{\alpha/2}\nu}/{\nu}$.
To this end we start with the following simple lemma.
\begin{lem}\label{lem:utov}
We have
\begin{equation}\label{eq:utov}
\int_{B_1^{(d)}} x_d^2 \phi(|x|)\,dx = \frac{1}{\pi} \int_{B_1^{(d+2)}} \phi(|x|)\,dx,
\end{equation}
for any function $\phi$ for which the integrals are absolutely convergent.
\end{lem}
\begin{proof}
We recall the notation $S^{(d-1)} = \{x\in \mathbb{R}^d:|x|=1\}$ for the unit sphere in $\mathbb{R}^d$, and the formula
for its area, $|S^{(d-1)}|=\frac{2\pi^{d/2}}{\Gamma(d/2)}$. We have
\begin{align*}
\int_{B_1^{(d)}} x_d^2 \phi(|x|)\,dx &=
\frac{1}{d} \int_{B_1^{(d)}} |x|^2 \phi(|x|)\,dx
= \frac{2\pi^{d/2}}{d\Gamma(d/2)} \int_0^1 r^{d+1} \phi(r)\,dr\\
&= \frac{|S^{(d+1)}|}{\pi} \int_0^1 r^{d+1} \phi(r)\,dr
= \frac{1}{\pi} \int_{B_1^{(d+2)}} \phi(|x|)\,dx.
\end{align*}
\end{proof}
\begin{proof}[Proof of Proposition~\ref{prop:lambdaeq}]
We consider a linear subspace $R_{d+2} \subset L^2(B_{d+2})$ consisting of all
radial functions, and a linear subspace $A_d = \{ f: f(x)=x_d g(x) \text{ for some radial } g\in L^2(B_d)\}$
of $L^2(B_d)$. Let $T$ be an operator defined by
\[
(Tf)(x_1,\ldots, x_d) = \sqrt{\pi} x_d f(x_1,\ldots, x_d,0,0), \quad f\in R_{d+2}\cap C(B_{d+2}).
\]
By Lemma~\ref{lem:utov} we obtain that $\|Tf\|_{L^2(B_d)} = \|f\|_{L^2(B_{d+2})}$,
therefore $T$ may be extended to an isometry from $R_{d+2}$ onto $A_d$.
Let $G$ be the Green operator, i.e., a bounded operator on $\{f\in L^2(\mathbb{R}^d): f=0 \text{ on } B_d^c \}$
defined by $G_d \phi_n = \frac{1}{\lambda_n} \phi_n$. This operator is the inverse of $-\Delta^{\alpha/2}$
(understood as a~generator).
We observe that the following diagram commutes.
\[
\begin{tikzpicture}[]
\node (a) {$R_{d+2}\ni f$};
\node (b) [right=of a, xshift=1.0em] {$Tf\in A_d$};
\node (c) [below=of a, xshift=1.0em, yshift=-1.2em] {$G_{d+2}f$};
\node (d) [right=of c] {$G_d Tf$};
\draw ($(a.south) + (+1.5em,0)$)
edge[->] node[auto,swap,font=\scriptsize]{$G_{d+2}$}
($(c.north) + (+0.5em,0)$);
\draw ($(a.east) + (0,0.2em)$)
edge[->] node[auto,font=\scriptsize]{$T$}
($(b.west) + (0,0.2em)$);
\draw ($(a.east) + (0,-0.2em)$)
edge[<-] node[auto,swap,font=\scriptsize]{$T^{-1}$}
($(b.west) + (0,-0.2em)$);
\draw ($(c.east) + (0,0.2em)$)
edge[->] node[auto,font=\scriptsize]{$T$}
($(d.west) + (0,0.2em)$);
\draw ($(c.east) + (0,-0.2em)$)
edge[<-] node[auto,swap,font=\scriptsize]{$T^{-1}$}
($(d.west) + (0,-0.2em)$);
\draw ($(b.south) + (-1.2em,0)$)
edge[->] node[auto,font=\scriptsize]{$G_{d}$}
($(d.north) + (0,0)$);
\end{tikzpicture}
\]
Indeed, by \eqref{Deltaupn} and \eqref{Deltavpn} it commutes for functions
$f(x)=(1-|x|^2)_+^n$, where $x\in B_{d+2}$ and $n=0,1,2,\ldots$.
The linear span of the set of those functions is dense in $R_{d+2}$, hence
by boundedness of $G_d$, $G_{d+2}$, $T$ and $T^{-1}$ the diagram commutes
for all $f\in R_{d+2}$.
Therefore we obtain a one to one correspondence between the radial eigenfunctions
of $G_{d+2}$ (or $\Delta^{\alpha/2}$) in $B_{d+2}$ and the $x_d$-antisymmetric eigenfunctions of $G_d$,
moreover, the corresponding eigenvalues are the same. In particular, $\lambda_1^{(d+2)} = \lambda_*^{(d)}$.
\end{proof}
We note that for functions $u$ in the domain of the generator $\Delta^{\alpha/2}$ we have
\begin{equation}\label{eq:Egen}
\mathcal{E}(u) = -\int_B u \Delta^{\alpha/2}u\, dx,
\end{equation}
in particular (\ref{eq:Egen}) holds for $u(x)=(1-|x|^2)_+^p$ with $p\geq \alpha/2$.
We are now ready to prove our estimates of $\lambda_1$ and $\lambda_*$.
\begin{proof}[Proof of Corollary~\ref{cor:est}]
Let $\psi$ be as in Lemma~\ref{lem:updest}. From that lemma, (\ref{eq:Egen})
and \cite[Lemma~2.2]{DydaHardy1dim} (or \cite[Proposition~2.3]{MR2469027})
we obtain
\[
\mathcal{E}(u) \geq \int_B u^2(x) \frac{-\Delta^{\alpha/2} \psi(x)}{\psi(x)}\,dx
\geq \mu_{d,\alpha} \int_B u^2(x)\,dx,
\]
and hence $\lambda_1 \geq \mu_{d,\alpha}$.
The second part follows now from Proposition~\ref{prop:lambdaeq}.
\end{proof}
\begin{table}[!htb]
\caption{Lower and upper bounds for $\lambda_1$ and $\lambda_*$.
Lower bounds of Corollary~\ref{cor:est} are in the top row.
In the middle row, in italic, there are the upper bounds obtained from (\ref{eq:var}), by plugging in
a certain linear combination of functions
$u_{j+\alpha/2}$ with $j=0,1,\ldots,12$.
The (near optimal) linear combinations were found numerically.
In the bottom row, we give upper bounds obtained by considering
functions $u$ given by (\ref{eq:ff}) and $\eta$ given by (\ref{eq:eta}).
}
\centering
\begin{tabular}{c|c|c|c|c|c}
\hline\hline
\multirow{2}{*}{$\alpha$} &
\multirow{2}{*}{$\lambda_1$ for $d=1$} &
\multirow{2}{*}{$\lambda_1$ for $d=2$} &
$\lambda_1$ for $d=3$ & $\lambda_1$ for $d=4$ &$\lambda_1$ for $d=5$ \\
&&& $\lambda_*$ for $d=1$ & $\lambda_*$ for $d=2$ & $\lambda_*$ for $d=3$ \\
\hline
\multirow{3}{*}{0.1} &
0.9676 & 1.04874 & 1.08633 & 1.1102 & 1.12756 \\
& \emph{0.97261} & \emph{1.05096} & \emph{1.09221} & \emph{1.12082} & \emph{1.14301} \\
& 0.97273 & 1.05103 & 1.09225 & 1.12093 & 1.14327 \\
\hline
\multirow{3}{*}{0.2} &
0.94993 & 1.10549 & 1.18391 & 1.23565 & 1.27419 \\
& \emph{0.95747} & \emph{1.10993} & \emph{1.19655} & \emph{1.25903} & \emph{1.30877} \\
& 0.95764 & 1.11001 & 1.19663 & 1.25927 & 1.30934 \\
\hline
\multirow{3}{*}{0.5} &
0.96202 & 1.3313 & 1.56035 & 1.72814 & 1.86169 \\
& \emph{0.97017} & \emph{1.34374} & \emph{1.60155} & \emph{1.80843} & \emph{1.98572} \\
& 0.97029 & 1.3438 & 1.60173 & 1.8092 & 1.98766 \\
\hline
\multirow{3}{*}{1} &
1.15384 & 1.96349 & 2.60869 & 3.15561 & 3.63636 \\
& \emph{1.15778} & \emph{2.00612} & \emph{2.75476} & \emph{3.45334} & \emph{4.12131} \\
& 1.1578 & 2.00618 & 2.75548 & 3.45616 & 4.12824 \\
\hline
\multirow{3}{*}{1.5} &
1.58614 & 3.13569 & 4.61848 & 6.03622 & 7.39626 \\
& \emph{1.59751} & \emph{3.27594} & \emph{5.05977} & \emph{6.94732} & \emph{8.93319} \\
& 1.59751 & 3.27624 & 5.06201 & 6.95522 & 8.95256 \\
\hline
\multirow{3}{*}{1.8} &
2.01395 & 4.28394 & 6.65946 & 9.07867 & 11.51297 \\
& \emph{2.04874} & \emph{4.56719} & \emph{7.50312} & \emph{10.82218} & \emph{14.49989} \\
& 2.04876 & 4.56781 & 7.50715 & 10.83601 & 14.53414 \\
\hline
\multirow{3}{*}{1.9} &
2.19524 & 4.77496 & 7.54923 & 10.43088 & 13.37504 \\
& \emph{2.24406} & \emph{5.13213} & \emph{8.59576} & \emph{12.5934} & \emph{17.09653} \\
& 2.24409 & 5.1329 & 8.60059 & 12.60997 & 17.13776 \\
\hline\hline
\end{tabular}
\label{table:upper}
\end{table}
\section{Upper bounds for eigenvalues}\label{sec:upper}
For functions $u\in D(\mathcal{E})$, by the variational formula, we have
\begin{equation}\label{eq:var}
\lambda_1 \leq \frac{\mathcal{E}(u)}{\int_B u^2\,dx}.
\end{equation}
For $u$ being a linear combinations of functions $u_{j+\alpha/2}^{(d)}$,
it is easy to compute the right hand side of (\ref{eq:var})
by using (\ref{eq:Egen}), Theorem~\ref{thm:up} and the formula
\[
\int_B |x|^s (1-|x|^2)^t \,dx = \frac{\pi^{d/2} \Gamma( \frac{s+d}{2}) \Gamma( t+1)}
{\Gamma( \frac{d}{2}) \Gamma( \frac{s+d}{2} + t+1)},\quad s>-d,\;\; t>-1.
\]
In particular, for functions of the form
\begin{equation}\label{eq:ff}
u(x) = (1-|x|^2)_+^{\alpha/2} + \eta (1-|x|^2)_+^{1+\alpha/2},
\end{equation}
we can explicitly find $\eta$ minimising the right hand side of (\ref{eq:var}).
A calculation yields,
\begin{equation}\label{eq:eta}
\eta_{min} = \frac{\sqrt{w} + d^2+2d-2a^2-6a-4 }{ 4a^2+12a+8},
\end{equation}
where
\begin{align*}
w &= d^4+4ad^3+8d^3 +8a^2d^2 + 32ad^2 + 28d^2 + 8a^3d+48a^2d\\
& + 88ad + 48d + 4a^4 + 24a^3 + 52a^2 + 48a+16.
\end{align*}
It is possible to further improve the estimates by taking more functions
$u_{j+\alpha/2}^{(d)}$ to define~$u$.
Then, however, the ratio (\ref{eq:var}) should be minimised numerically.
In Table~\ref{table:upper}, we give bounds for $\lambda_1$ (and hence also for $\lambda_*$)
obtained by (\ref{eq:var}) and Corollary~\ref{cor:est}.
\subsection*{Acknowledgements}
The author wishes to thank Tadeusz Kulczycki for helpful discussions, Krzysztof Bogdan for
careful reading of the manuscript and numerous comments, and Mateusz Kwa\'snicki for
fruitful discussions and, in particular, suggesting the proof of Proposition~\ref{prop:lambdaeq}.
This research was partially supported by MNiSW grant N N201 397137,
and by the DFG through SFB-701 'Spectral Structures and Topological Methods in Mathematics'.
\def$'${$'$}
|
\section{Introduction}\label{Sec:Intro}
Non-relativistic superfluids provide a high-precision laboratory in which to probe many-body physics in the extreme quantum regime \cite{zwerger}. In an effort to bring the tools of holography \cite{Maldacena:1997re,Gubser:1998bc,Witten:1998qj} to bear on these systems, considerable effort has been devoted to studying non-relativistic deformations of relativistic examples\footnote{Since the non-relativistic conformal group is a subgroup of the relativistic group in one higher dimension, we can construct a non-relativistic conformal field theory (NRCFT) by turning on an operator in a relativistic conformal field theory (CFT) which breaks the relativistic group to its non-relativistic subgroup. Taking the operator to be marginal in the NRCFT \cite{Mehen:1999nd,Nishida:2007pj} requires it to be irrelevant in the CFT. Holographically, this corresponds to a 1-parameter deformation of the geometry which alters the asymptotic geometry from Anti de Sitter (AdS), whose isometries form the relativistic conformal group, to Schr\"odinger \cite{Son:2008ye, Balasubramanian:2008dm}, whose isometries fill out the non-relativistic conformal group.} which enjoy $z=2$ scaling \cite{Son:2008ye, Balasubramanian:2008dm,Adams:2008wt,Herzog:2008wg,Maldacena:2008wh}. Unfortunately, such deformations generate highly atypical states in the resulting NRCFT whose thermodynamic and other properties are tightly constrained by their relativistic births. In particular, they are in general far from the superfluid groundstates of the corresponding systems, for which we currently have no description.
In this paper we examine certain superfluid states in a holographic NRCFT in a probe approximation.
Our strategy is essentially the same as in the AdS case (see {\eg e.g.}\ \cite{Gubser:2008px, Hartnoll:2008vx}): we study an Abelian-Higgs theory in the background of a neutral asymptotically-Schr\"odinger black hole \cite{Adams:2008wt, Herzog:2008wg, Maldacena:2008wh}\ in the probe approximation. Several features of the geometry, however, make the resulting analysis qualitatively different. For example, we are now forced to turn on two components of the bulk gauge field: $A_{t}$, dual to a boundary charge current, and $A_{\xi}$, dual to a boundary Mass\footnote{%
In an NRCFT, each primary operator is characterized by not only a Dimension $\Delta$ but also by its Mass, $M$, where
Mass is the name of a central extension $\hat{M}$ in the NR conformal group. In the case of free fermions, $\hat{M}=M\psi^{\dagger}\psi$, hence it is often called the ``Number'' operator -- we prefer ``Mass'' to disambiguate the various meanings of ``number''.} current. By itself, this is not a big deal. What's surprising given intuition from the relativistic case is that the boundary value of this second vector component, $M_{o}=A_{\xi}|_{\partial}$, weasels its way into the {\em dimension} of the boundary order parameter as $\Delta = 2\pm \sqrt{4+m^2+q^{2}M_{o}^{2}}$. Specifying the boundary NRCFT thus requires not just specifying the bulk matter fields and their interactions, but also the asymptotic fall-offs of some of the bulk fields. Similar effects arise in the holographic renormalization of the theory, which as usual requires introducing counterterms which depend on the boundary operator dimensions; here, these counterterms will explicitly depend on the boundary values of some bulk fields, too (see {\eg e.g.}\ \cite{Guica:2010sw}\ for a discussion of such effects).
To build a truly NR superfluid, then, we must generate a condensate for a boundary operator with non-zero Mass eigenvalue, $M\neq0$. This is the role of the second component of the gauge field -- in a gauge where the phase of the condensate is constant, the Mass eigenvalue is simply $M=-q\,M_{o}$. The boundary value of the second component of the gauge field thus controls the breaking of the Mass symmetry in the superfluid phase.
While the background about which we perturb is a 1-parameter deformation of a relativistic example, the superfluid state we find is not, and indeed enjoys quite distinct phenomenology from its AdS cousins. Fundamentally, the non-relativistic condensate is characterized by one more quantum number than in the relativistic case -- the Mass eigenvalue, $M$, of the order parameter -- with the NR condensate breaking the symmetry generated by the Mass operator, a key signature of a non-relativistic superfluid. As we shall see, this leads to a host of interesting effects in the strongly NR regime, including the appearance of a thermodynamically unstable high-temperature condensed phase which drives the superconducting transition from 2$^{nd}$ order to 1$^{st}$ at a multicritical point, the persistence of a condensate even in the absence of a chemical potential for the charge density, and reentrance of the normal phase at low temperatures for sufficiently large background density.
It is tempting to interpret this re-entrance as signaling a zero temperature quantum phase transition as the background mass density is tuned. However, the re-entrant normal state is again the simple 1-parameter deformation which we do not expect to be the true equilibrium groundstate, so we do not expect this probe analysis to be the end of the story. Meanwhile, it remains possible that the system is in fact reentrant for all values of the background Mass density as $T\to0$, where our probe approximation becomes unreliable. Resolving these puzzles, however, requires going beyond the truncated probe approximation discussed in this paper; we leave them to future study.
The plan of the paper is as follows.
In Section 2 we quickly describe the basic strategy and computational setup, with various details elaborated in Appendices.
In Section 3 we explore the phenomenology and phase stucture of holographic superfluids outside a Schr\"odinger black hole (an analogous study in the background of a Schr\"odinger soliton \cite{Mann:2009xk}\ is performed in Appendix A -- while this is not in the same ensemble as the black hole, it provides an alternate example with surprising physics of its own).
We close in Section 5 with a summary and list of next steps.
\section{The Setup}\label{Sec:Probe}
Our basic strategy involves studying an Abelian Higgs system,
\begin{equation}\label{EQ:ProbeAction}
{\cal L}_{probe} ={1\over e^{2}}\( -{1\over4} F^{2} -|{\cal D}\Phi|^{2} -m^{2}|\Phi|^{2} \) \,,
\end{equation}
as a perturbation around the planar Schr\"odinger black hole background,
\begin{equation}\label{EQ:pappalardoBH}
ds^2 =
\(-f + {(f-1)^{2}\over 4(K-1)}\){dt^{2}\over K r^{4}}
+
{1+f \over r^{2} K} dt \, d\xi
+{K-1\over K}d\xi^{2}
+ {d\vec{x}^{2} \over r^{2}}
+ {dr^{2} \over f \, r^{2}}\,.
\end{equation}
in the probe limit, $e^{2}\to\infty$. Here,
$f=1-r^{4}(\pi T\Omega)^{4/3}$,
$K=1+r^2 \Omega^2$
and the metric is given in string frame. One can think of this as a rather extreme truncation of the charged Schr\"odinger black hole system \cite{Adams:2009dm,Imeroni:2009cs}\ where we drop the coupling of the vector to the scalar and massive vector of the black hole background, or simply as a holographic toy model. The geometry is controlled by two physical parameters, the background mass density, $\Omega$, and the temperature, $T$
with the horizon located at the radial coordinate $r_{H}=(\pi T\Omega)^{-1/3}$.
For spatially homogeneous solutions, we can without loss of generality set $\vec{A}=0$ and take
$\Phi = \phi(r)$
and $A = A_{t}(r)dt + A_{\xi}(r) d\xi$.
In Einstein frame, the equations of motion take the form,
\begin{eqnarray}\label{EQ:EOM1}
&& \hspace{-.65cm}
f^{2}r^{2}\phi'' - f(4-f)r\phi'
-\[\vphantom{\frac{A^{2}}{B_{H}}^{2}}
f\(q^{2}A_{\xi}^{2}
+2q^{2} r^{2}A_{\xi}A_{t}
+m^{2}K^{1/3}\)\right. \nonumber\\
&& \hspace{4cm}\left. -{q^{2}(f-1)^{2}\over 4(K-1)}\(A_{\xi} - 2r_{H}^{4}\Omega^{2}A_{t}\)^{2}
\vphantom{\frac{A^{2}}{B_{H}}^{2}}\]\phi =0~~~\\
\label{EQ:EOM2}
&& \hspace{-.65cm}f r^{2}A_{t}'' - \(2-{f\over3}(7K-4)\){r\over K}A_{t}'
- \(2+f(f-1)+{(f-1)^{2}\over K-1}\){1\over Kr}A_{\xi}' -2q^{2}K^{1/3} \phi^{2} A_{t}
= 0~~~~~~~~~~~\\
\label{EQ:EOM3}
&& \hspace{-.65cm}f r^{2}A_{\xi}'' - \(4K-2-{2+K\over3}f\){r\over K}A_{\xi}' -4(K-1){r^{3}\over K}A_{t}' -2q^{2} K^{1/3}\phi^{2} A_{\xi}
= 0 ~~~
\end{eqnarray}
Note that $A_{\xi}\neq A_{t}$.
\subsection{Asymptotic Behavior and the Holographic Dictionary}
Near the boundary at $r=0$, the vector components behave as,
\begin{equation}
A_t = \mu_{Q} +\r_{Q}\, r^2+\dots\,, ~~~~~~~~~~
A_\xi = M_{o} + \r_{M}\, r^2 + \dots
\end{equation}
where the various $\dots$ represent various (possibly non-normalizable) terms whose coefficients are entirely fixed by the equations of motion and the values of these integration constants, $\mu_{Q}$, $\r_{Q}$, $M_{o}$ and $\r_{M}$.\footnote{In particular, the leading term for $A_{t}$ runs as $-2\rho_{M}\log(r)$. While formally the dominant term, it is determined by the equations of motion and $\r_{M}$ and thus does not represent an independent mode of the system. Importantly, due to factors of the inverse metric, this log running does not lead any components of the bulk stress tensor to diverge. A complete holographic renormalization of this system would settle the dictionary, but is beyond the scope of the present paper; for the moment we simply take the above dictionary as a provisional interpretation which is supported by the consistency of the results below. Interestingly, while $A_{t}$ has no log in fully backreacted charged-black-hole solutions \cite{Adams:2009dm,Imeroni:2009cs}, linearizing the Maxwell equation around these solutions does generate a $\log$ without changing any other of the asymptotics of the vector, so this log is likely a simple consequence of an extreme truncation of the full charged black hole system. It would be interesting to study the full system and see what, if anything, changes.}\
As usual, $\mu_{Q}$ represents the chemical potential per unit charge, which effectively sets the zero of energy in the boundary theory -- the gauge-invariant bulk quantity that becomes the boundary hamiltonian acting on the operator dual to the bulk matter field of charge $q$ is $(i\partial_{t}+q\,A_{t})$; at the boundary, for plane waves $e^{-i\omega t}$, this becomes $(\omega+q\,\mu_{Q})$.
Thus, one insertion of the charged operator ${\cal O} \,e^{-i\omega t}$ costs $\delta E = (\omega+q\,\mu_{Q})$.
As usual, $\rho_{Q}$ computes the induced charge density.
It might be tempting to think of $M_{o}$ as a chemical potential for the Mass operator, $\hat{M}$. However, this is not quite right -- it is a superselection parameter. Recall that, holographically,
\begin{equation}
\hat{M} \equiv \hat{P}_{\xi}|_{\partial}= -i(\partial_{\xi}-iqA_{\xi})|_{\partial}\,,
\end{equation}
{\it i.e.}\ $\hat{M}$ is the boundary value of the gauge-invariant $\xi$-momentum in the bulk.
The mass eigenvalue of a boundary operator dual to a bulk field with $\xi$-momentum $\ell$ and charge $q$ is thus $M=(\ell-q\,M_{o})$, where $M_{o}=A_{\xi}|_{\partial}$.
Like a chemical potential, $M_{o}$ sets a bias for the mass $M$, shifting it away from its $\xi$-momentum, $\ell$. But the mass in an NRCFT is not a parameter, it is part of the definition of the theory. Thus, once we fix gauge in the bulk, {different values of $A_{\xi}|_{\partial}$ correspond to distinct NRCFTs}, not to a fixed theory with different background fields turned on. In particular, as we will see momentarily, the dimensions of various boundary operators depend on $A_{\xi}|_{\partial}$, an unfamiliar effect. $\rho_{M}$ computes the Mass density coupled to $A_{\mu}$. Henceforth we fix gauge in the bulk such that $\ell=0$ and $M = -qA_{\xi}|_{\partial}$.
As for our charged scalar, near the boundary at $r\to0$ it behaves as
\begin{equation}
\phi \sim \phi_1r^{\Delta_-} +\phi_2r^{\Delta_+}+\dots\,,
\end{equation}
where
\begin{equation}
\Delta_\pm
=2\pm \sqrt{4+m^2+q^{2}M_{o}^2} \, .
\end{equation}
(Note that we will occasionally write $\Delta_{1}$ and $\Delta_{2}$ for $\Delta_{-}$ and $\Delta_{+}$, respectively.) In the window $1\!<\!\Delta_{-}\!<\!2$, both components $\phi_{1,2}$ are normalizable, so we may interpret either of $\phi_{1,2}$ as the vev $\vev{{\cal O}}$, with the other representing the source ${\cal J}$. These two choices correspond to alternate quantizations of the boundary NRCFT \cite{Son:2008ye,Klebanov:1999tb}.
We will focus on the choice $\vev{{\cal O}}\propto\phi_{1}$ and ${\cal J}\propto\phi_{2}$ for reasons which will become clear in the next section.
Importantly, the dimensions, $\Delta_{\pm}$, depend not only on the mass of the bulk scalar, as in AdS,
but also on the boundary value of a bulk field, $M_{o}=A_{\xi}|_{\partial}$. As discussed above, by the holographic relation $\hat{M} \equiv -i(\partial_{\xi}-iqA_{\xi})|_{\partial}$, this quantity is nothing but the Mass eigenvalue of the dual operator,
$$
M=-qA_{\xi}|_{\partial} = -q\,M_{o}.
$$
Our expression for the dimensions above then becomes,
$
\Delta_{\pm} = 2\pm \sqrt{4+m^2+M^2}\,,
$
which is the expected form \cite{Son:2008ye,Balasubramanian:2008dm}, including the quadratic dependence on $M$ inside the radical.
Now, as discussed in \cite{Adams:2008wt,Herzog:2008wg,Maldacena:2008wh}, the free energy of the full system takes the form, $F \equiv E+\mu_{M}\hat{M}$, where $\mu_{M}={-1\over2\Omega^{2}r_{H}^{4}}$ is determined by the background spacetime. The total free energy per insertion of an operator dual to a bulk field with $\xi$-momentum $\ell$, frequency $\omega$, and coupled with charge $q$ to our gauge field is thus $\delta F = (\omega+q\,\mu_{Q})+\mu_{M} M$, where $M=(\ell-q\,M_{o})$
\subsection{Near-Horizon Behavior and Setting Up the Calculation}
In the bulk, we are thus left with a six-parameter family of solutions labeled by sources $\(\mu_{Q},\,M_{o},\,{\cal J}\)$ and responses $\(\r_{Q},\,\r_{M},\,\vev{{\cal O}}\)$. Holographically, we expect boundary conditions at the horizon, where the radial equations of motion degenerate, to impose three additional constraints. Together with the two parameters $T$ and $\Omega$ of the background geometry, this should leave us with a five-parameter phase space. To verify this, we need to study the behaviour of our solutions near the horizon.
The equations of motion degenerate at the black hole horizon, so we must impose boundary conditions to pick the appropriate solutions. As usual, it suffices to impose regularity at the horizon, which is in any case necessary for the validity of the probe approximation. Assuming regularity, the equations of motion as presented in (\ref{EQ:EOM1}) -- (\ref{EQ:EOM3}) degenerate into three algebraic equations relating the six horizon values of the fields and their derivatives, as expected,
\begin{eqnarray}
\(A_{\xi}(r_{H})- 2r_{H}^{4}\Omega^{2}A_{t}(r_{H})\)^{2} \phi(r_{H}) &=& 0 \\
- 2r_{H}^{2}A_{t}'(r_{H}) + \({2K_{H}-1\over K_{H}-1}\)A_{\xi}'(r_{H}) + 2q^{2}r_{H}K_{H}^{4/3}\phi^{2}(r_{H})A_{t}(r_{H}) &=& 0 \\
4r_{H}\phi'(r_{H}) +\( 4K_{H}(K_{H}-1)r_{H}^{4}q^{2}A_{t}^{2}(r_{H}) +{m^{2}K^{1/3}}\)\phi(r_{H}) &=&0\,.
\end{eqnarray}
This suggests a simple numerical strategy for constructing superfluid states of our holographic NRCFT. To specify a solution to the full equations of motion, we fix any three of $\mu_{Q}$, $M_{o}$, ${\cal J}$, $\r_{Q}$, $\r_{M}$ and $\vev{{\cal O}}$ at the boundary and impose the above regularity conditions at the horizon. Since we are interested in spontaneously generated condensates, we will generally set ${\cal J}=0$. The resulting two-point boundary value problem can be solved numerically in various ways. The most straightforward is a brute-force shooting method, as typically employed in the relativistic case.
In sweeping out parameter space, however, we must be careful to vary the {\em parameters} of the NRCFT while holding the NRCFT itself fixed -- {\it i.e.}, while holding the spectrum of quantum numbers fixed. This is straightforward in AdS, where fixing the set of dimensions reduces to fixing the bulk mass $m^{2}$ of the bulk scalar. Here, however, the dimension $\Delta$ and Mass $M$ of the boundary scalar operator depend on the asymptotic value of $A_{\xi}$ as $M = -q\,A_{\xi}|_{\partial}$ and $\Delta_{\pm} = 2\pm \sqrt{4+m^2+M^2}$. Before sweeping out parameter space, then, we must fix $A_{\xi}|_{\partial}$=$M_{o}$.
As we've already set ${\cal J}=0$, fixing the system thus leaves us with a three-parameter phase space labeled by $\mu_{Q}$, $\Omega$ and $T$.
This peculiar behavior -- that the definition of the boundary CFT depends on the boundary behavior of the bulk fields -- is a very general phenomenon in Schr\"odinger holography. Indeed, renormalizing the boundary stress tensor, say, or other operators in the boundary NRCFT, requires counterterms which are local in time and space, but which depend explicitly on dimensions, $\Delta$, and thus on the asymptotic values of $A_{\xi}$, in a mildly non-local fashion. Such field-dependent counterterms have appeared previously in attempts to renormalize holographic NRCFTs, most recently in \cite{Guica:2010sw}. A complete understanding of the holographic renormalization of these theories is clearly of considerable interest.
\subsection{Conductivity}
We can compute the conductivity in our superconducting background by studying linear response to a time-dependent vector potential $A_{x}\sim e^{-i\omega t}$. As usual, this boils down to solving the equation of motion for the bulk gauge component $A_{x}$ linearized about the superfluid background and subject to infalling boundary conditions at the horizon. Setting $A_{x}=a(x)e^{-i\omega t}$, we have
\begin{equation}\label{EQ:Cond}
f r^{2}a_{x}'' - \(4-f{2+7K\over3K}\)r a_{x}' +\({\omega^{2}r^{4}(K-1)\over f} - 2q^{2} K^{1/3}\phi^{2}\)a_{x} = 0\,.
\end{equation}
Near the horizon, this reduces to,
\begin{equation}
(\omega^2+(\frac{4 \epsilon}{r_H^3 \Omega} \frac{d}{d \epsilon})^2 ) a_x(\epsilon) \simeq 0
\end{equation}
where $r=r_{H}-\epsilon$. The infalling solutions thus takes the form,
\begin{equation}
A_{x} = a_{0}\,e^{-i\omega t}(r-r_{H})^{-i\omega\over4\pi T}(1+a_{1}(r-r_{H})+\dots)
\end{equation}
Near the boundary,
$$
A_{x} = A_{0} + A_{2} {r^2\over2}+\dot
$$
A short computation then verifies that the conductivity is given by,
$$
\sigma(\omega) = {\vev{J_{x}}\over\vev{E_{x}}} = -i{\vev{J_{x}}\over\omega\vev{A_{x}}} = -i{A_{2}\over \omega A_{0}}
$$
Note that, since we are solving a linear equation but only care about this ratio, the overall scale of $A_{x}$ is immaterial. We can use this freedom to set $a_{0}=1$, which simplifies the numerical problem.
Notably, we can analytically determine the $\omega$-dependence of $\sigma$ for large and small $\omega$ via standard power-series analysis. Importantly, the scaling in the superfluid phase to be independent of $\Delta$ and $M$. At small frequency, we find,
\begin{equation}
Im[\sigma(\omega \ll 1)] \propto \omega^{-1}
\end{equation}
while for large $\omega$ we have,
\begin{equation}
Re[\sigma(\omega \gg 1)]\propto \omega^{-1/3}
~~~~~~~~
Im[\sigma(\omega \gg 1)] \propto \omega^{-1/3}
\end{equation}
This last result unsurprisingly differs from the AdS case, where $ Re[\sigma(\omega \gg 1)]=1$. Reassuringly, they both match numerical results presented below, a nice sanity check.
\section{Phases of a Schr\"odinger Superfluid}\label{Sec:BHPhases}
Thus armed, we now get down to the business of finding a superfluid state in our non-relativistic holographic CFT and exploring its phase diagram. A priori, the phase space is fairly high-dimensional -- specifying a point involves fixing $\Delta$ and $M$ to fix the theory, then tuning $\mu_{Q}$, $T$ and $\Omega$ to sweep out the phase diagram. For simplicity, we will begin by picking convenient values $\Delta=6/5$, $M=1/2$ and $\mu_{Q}=1/8$, then dial the background Mass density, $\Omega$. This will reveal a zero-temperature quantum phase transition at a critical value $\Omega_{*}$. We will then fix $\Omega$ and vary $\mu_{Q}$, which will drive the superconducting phase transition from 2$^{nd}$ to $1^{st}$ order.
\subsection{Varying $\Omega$ and a Quantum Phase Transition?}
We begin by fixing $\Delta=6/5$ and $M=1/2$, then set $\mu_{Q}=1/8$ and vary $\Omega$ between 0 and 1. The basic results are presented in Figures 1, 2 and 3, which plot the condensate $\vev{{\cal O}(T)}$ as a function of temperature, as well as the AC conductivities $Re[\sigma(\omega)]$ and $Im[\sigma(\omega)]$, for $\Omega$ = ${1\over16}$, ${3\over8}$ and 1, respectively. These results are discussed in detail below.
\vspace{0.2cm}
{~~~~ $\bf\bullet ~~ \Omega\ll\Omega_{*}$}\\
For very small $\Omega$, the geometry remains essentially AdS until very close to the boundary, so we expect most low-energy physics -- such as superfluid condensation -- to very closely track familiar AdS results. This turns out to be almost correct, modulo a surprise we'll explore shortly.
\begin{figure}[!h]
\centerline{ \epsfig{figure=type1_condensate.pdf, height=1.4 in} \epsfig{figure=type1_REcondu.pdf, height=1.4 in} \epsfig{figure=type1_IMcondu.pdf, height=1.4 in} }
\caption{\em At small $\Omega$, the behavior of the superfluid is essentially the same as in AdS, with a $2^{nd}$ order mean-field phase transition at the onset of superconductivity at $T_{c}$, including the familiar gap-and-pole form in the AC conductivity, leaving us in a happy superfluid state at $T=0$.
Here, $(\mu_Q,\Omega)=(1/8,1/16)$, with $T_C=0.505$.
}
\label{FIG:TYPE1}
\end{figure}
Figure \ref{FIG:TYPE1}a shows the condensate as a function of temperature for the first conformal family, for $\Omega=1/16$. As is clear by eye and can be checked precisely from the numerics, the resulting condensate turns on at $T=T_{c}$ with classic mean-field behavior ($\b_{c}=\frac{1}{2}$) and grows as the temperature is lowered. Figures \ref{FIG:TYPE1}b and \ref{FIG:TYPE1}c then show the real and imaginary parts of the AC conductivity for various temperatures indicated by color, from high (violet) to low (red). These demonstrate the appearance of a superconducting state at $T_{c}$, with the gap growing as the temperature is lowered. Note, too, that the conductivity in the superfluid phase has $Im[\sigma(\omega \rightarrow 0)] \sim 1/\omega$, while $Re[\sigma(\omega \rightarrow \infty)]\sim Im[\sigma(\omega \rightarrow \infty)] \sim \omega^{-1/3}$. This scaling is expected on general grounds, so gives us confidence in our numerical results.
\newpag
{~~~~ $\bf\bullet ~~ \Omega\gg\Omega_{*}$}\\
As we increase the background number density, the story changes dramatically.
Figure \ref{FIG:TYPE3} shows the same plots as Figure \ref{FIG:TYPE1} but with $\Omega = 1$ rather than $\Omega=1/16$. The most obvious difference is that the order parameter {\em vanishes} at sufficiently low temperature, $T\le T_{L}$, doing so again with mean-field behavior. As is clear form the finite value of $Re[\sigma(0)]$, the extreme low-temperature phase is again metallic.
\begin{figure}[!h]
\centerline{ \epsfig{figure=type3_condensate.pdf, height=1.4 in} \epsfig{figure=type3_REcondu.pdf, height=1.4 in} \epsfig{figure=type3_IMcondu.pdf, height=1.4 in} }
\caption{\em At large $\Omega$, in addition to the original transition to a superconducting state at $T_{c}$, the system now exhibits reentrance of the normal phase at a new low-temperature $2^{nd}$ order transition at $T_{L}$, again with mean-field exponents. Below $T_{L}$, $\sigma$ behaves like the normal gas.
Here $(\mu_Q,\Omega)=(1/8,1)$ with $T_C=0.149$ and $T_L=0.009$
}
\label{FIG:TYPE3}
\end{figure}
Consider now the behavior of the system at zero temperature as a function of the background number density, $\Omega$. As $\Omega\to0$, the system is superconducting. As $\Omega\to1$, the metallic phase is reentrant. At some critical $\Omega_{*}$, then, the zero-temperature system appears to undergo a superconductor-metal quantum phase transition.
\vspace{0.2cm}
{~~~~ $\bf\bullet ~~ \Omega\to\Omega_{*}$}\\
It is tempting to try to determine what happens as we tune $\Omega$ towards this critical $\Omega_{*}$. Figure \ref{FIG:TYPE2} shows the same system at $\Omega$ slightly above $\Omega_{*}$.
\begin{figure}[!h]
\centerline{ \epsfig{figure=type2_condensate.pdf, height=1.4 in} \epsfig{figure=type2_REcondu.pdf, height=1.4 in} \epsfig{figure=type2_IMcondu.pdf, height=1.4 in} }
\caption{\em At intermediate $\Omega$, we again have a $2^{nd}$ order mean-field transition into a superfluid state at $T_{c}$. At low temperatures, however, the system undergoes a non-mean-field transition to an apparently insulating state.
Here $(\mu_Q,\Omega)=(1/8,3/8)$ with $T_C=0.123$.}
\label{FIG:TYPE2}
\end{figure}
As before, there is a phase transition at $T_{c}$ with standard mean-field behavior. The zero temperature behavior, however, differs dramatically from mean-field expectations; rather, at low temperature, the condensate decays exponentially, as does the superfluid density, while the normal density remains vanishing and the conductivity heavily suppressed at small but non-vanishing $\omega$, suggesting that the $T=0$ state is not metallic. It is tempting to read this as indicating a translationally-invariant insulating phase.
However, numerical results in this region should be taken with a sizeable grain of salt. Indeed, at sufficiently low temperature, the numerics simply fail to converge. More physically, in this regime, the probe approximation is becoming dangerously unreliable -- the matter field profiles which generate the required boundary values grow rapidly deep in the bulk (and in particular near the horizon) as we approach $T=0$ or $\Omega_{*}$. Backreaction may thus qualitatively alter the low-temperature physics, either near the transition at $\Omega\sim\Omega_{*}$ or for sufficiently low $T$ at any $\Omega$.
Indeed, it is entirely possible that the backreacted solution is re-entrant at any value of $\Omega$; our analysis is only reliable sufficiently far away from $T=0$.
To unambiguously exclude re-entrance at small $\Omega$ as $T\to0$ requires including backreaction, which is beyond the scope of this paper.
Note, however, that the probe approximation shows no signs of inconsistency for $\Omega>\Omega_{*}$, so we can be quite confident that the system is definitely re-entrant at sufficiently large $\Omega$.
\subsection{High Temperature Condensates and the Free Energy}
The surprise alluded to above involves the {\em high}-temperature limit. Figure \ref{FIG:HighT} shows the same system but now extending to higher temperatures. The surprise is the appearance of a {\em high temperature} condensate at $T\ge T_{H}$. Troublingly, the condensate appears to grow without bound as the temperature increases.\footnote{Such a high-temperature instability was predicted
by Cremonesi {\em et al} \cite{Cremonesi:2009gy}\ whenever $\Delta\le4$.}
\begin{figure}[!h]
\centerline{\epsfig{figure=type1_condensat_lowhighT.pdf, height=2 in} }
\caption{\em Surprisingly, there is another condensed phase at {\em high} temperatures. Here $(\mu_Q,\Omega)=(1/8,1/16)$.\label{FIG:HighT}}
\label{type i}
\end{figure}
Before we panic, however, we should verify that this high-temperature condensate is in fact thermodynamically favored over the trivial vacuum. Holographically, this means we computing the holographically renormalized on-shell action. Unfortunately, in asymptotically Schr\"odinger spacetimes, holographically renormalizing the action is exceedingly complicated. Happily, a simple strategy allows us to compute the difference in free energy between condensed and vacuum states without performing a full renormalization of the action.\!\footnote{We thank Nabil Iqbal for illuminating discussions on this topic, and refer the reader to \cite{Nabil}, in which this approach is further developed.}
The basic idea goes as follows. Generally, specifying the non-normalizable (source) mode $\phi_{1}$ of the bulk scalar determines the normalizable (response) mode, $\phi_{2}$. Smoothly varying the source thus traces out a curve $\phi_{2}(\phi_{1})$ in the ($\phi_{1}$, $\phi_{2}$) plane. Along this flow, we can ask how the free energy -- aka, the Euclidean action -- varies. Given the properly renormalized action, the variation of the full bulk action takes the form,
$
\delta S_{eff}=\dots \delta \phi_i+\dots \delta A_i
$
where $\delta \phi_i$ and $\delta A_i$ are the variations of the bulk fields and the $\dots$ correspond to the bulk equations of motion. So long as we satisfy the bulk equations of motion, this reduces to a simple boundary term,
$
\delta S=(\Delta_1-\Delta_2) \int_{\partial_M} \phi_2 \; \delta \phi_1 -2\int_{\partial_M} (\rho_M \; \delta \mu_Q+\rho_Q \; \delta M_{o})
$
Moreover, if we hold fixed the asymptotic values of $A_{i}$ (corresponding to fixing the values of the chemical potential $\mu_{Q}$=$A_{t}|_{\partial}$ and the mass $M$=$A_{\xi}|_{\partial}$), this further simplifies to,
$
\delta S=(\Delta_1-\Delta_2) \int_{\partial_M} \phi_2 \; \delta \phi_1
$
We can thus compute the relative free energy density (${\cal F}_{A}$$-{\cal F}_{B}$) between any two states $A$ and $B$ connected by such a flow by integrating $\delta S$ along the flow,
\begin{equation}
{\cal F}_B-{\cal F}_A =-T\int^B_A \frac{\delta S_E}{V_{D}}=-T(\Delta_1-\Delta_2) \int^B_A \phi_2 \; d \phi_1
\label{eq:F0}
\end{equation}
where $V_{D}$ is the volume of the boundary theory and the integral is performed along the flow specified above. By construction, this agrees with what we would get by evaluating the fully holographically renormalized free energy for each solution and subtracting. Happily, this allows us to compute the correct free energy without having to worry about the full holographic renormalization of the theory (for further comparison between holographic renormalization and our method, see \cite{Guica:2010sw},\cite{Ross:2009ar},\cite{Marolf:2006nd}).
Now consider the case of our holographic superfluid in alternate quantization, where $\phi_{1}$=$\vev{{\cal O}}$ is the response and $\phi_{2}$=${\cal J}$ is the source. In this case, the curve $\phi_{1}(\phi_{2})$ is multi-valued over $\phi_{2}$=0, with one solution corresponding to the trivial vacuum, $\vev{{\cal O}}$=0, and one to the nontrivial condensate, $\vev{{\cal O}}$$\neq$0. As outlined above, these two solutions are connected by a very specific flow in the ($\phi_{1},\,\phi_{2}$) plane. To compute the properly renormalized relative free energy, then, all we must do is find this flow and integrate along it,
\begin{equation}
{\cal F}_C-{\cal F}_N=-T(\Delta_1-\Delta_2) \int^C_N \phi_2 \; d \phi_1
\label{eq:F}
\end{equation}
where the integration is again along the flow defined above. If this difference is negative, the condensate is thermodynamically favored.
\begin{figure}[!t]
\centerline{ \epsfig{figure=lowT_area.pdf, height=2in} ~~~~~~~~ \epsfig{figure=highT_area.pdf, height=2in}}
\caption{\em (a) Low T condensate has smaller free energy than non-condensed phase, ${\cal F}_C-{\cal F}_N <0$. (b) High $T$ condensate has larger free energy than non-condensed phase, ${\cal F}_C-{\cal F}_N >0$. One can determine the sign of ${\cal F}_C-{\cal F}_N $ from the orientation of the curve. Here $(\mu_Q,\Omega)=(3/8,1/16)$.}
\label{lowhighTarea}
\end{figure}
Figure \ref{lowhighTarea} plots two such flows. On the left we have a flow connecting the trivial vacuum ($\phi_{1}$=$\phi_{2}$=0) and a non-trivial vacuum ($\phi_{1}$$\neq$$0$, $\phi_{2}$$=$$0$) of the first conformal family in the low-temperature regime, with the flow indicated by the solid line and the direction of flow defining the direction of integration. The area under the curve, corresponding to the free energy of the condensed state, is negative. On the right is the analogous flow in the high-temperature regime -- here the free energy is positive. We thus deduce that the low-temperature condensate is thermodynamically stable, while the high-temperature condensate is unstable, at least for this first conformal family.
\begin{figure}[!t]
\centerline{\epsfig{figure=SchBH_O2_T_Delta13-5.pdf, height=2in} ~~~~~~~~ \epsfig{figure=SchBH_O1_O2_Delta13-5.pdf, height=2in} }
\caption{\em (a) Condensate as a function of T for the second conformal family with $\Delta_2=13/5$. (b) Typical flow at generic temperature, indicating a thermodynamic instability at every temperature.}
\label{FIG:2ndFam}\end{figure}
What about the second conformal family? Figure \ref{FIG:2ndFam}a plots the condensate of the second family as a function of temperature. Note that there is no separate low vs high temperature condensate, just a single continuous instability whose profile grows with temperature.
Figure \ref{FIG:2ndFam}b then shows a typical flow at typical temperature. Importantly, the enclosed area is negative for every temperature, indicating a thermodynamic instability even at arbitrarily high temperature. This is why we quietly chose the first conformal family in Section 2. It would be interesting to understand the meaning of this instability in detail.
\subsection{Varying $\mu_{Q}$ and a Multicritical Point}
The thermodynamic instability of the high-temperature condensate leads to an important physical effect as we vary $\mu_{Q}$. Figure \ref{FIG:FirstOrder}a plots $T_{c}(\mu_{Q})$ and $T_{H}(\mu_{Q})$, the critical temperatures for the low- and high-temperature condensates as a function of the chemical potential $\mu_{Q}$. As we crank up $\mu_{t}$, holding all other parameters fixed, $T_{c}$ increases while $T_{H}$ decreases. At a critical value, $\mu_{*}$, the two critical points merge; above $\mu_{*}$, the condensate is non-zero for all temperatures. This is clear from Figure \ref{FIG:FirstOrder}b, where we plot the order parameter as a function of temperature for values of the chemical potential above and below this critical $\mu_{*}$.
\begin{figure}[!h]
\centerline{\epsfig{figure=mean_field_T_mut.pdf, height=2.0 in} ~~ \epsfig{figure=mean_field_2nd_1st_O_T.pdf, height=2.0in} }
\caption{\em At $\mu_{Q}>\mu_{*}\sim0.192$, the transition goes 1$^{st}$ order. Here, $\Omega$=${1 / 16}$.\label{FIG:FirstOrder}}
\label{type i}
\end{figure}
However, we have already checked that the condensed phase is thermodynamically disfavored at high temperatures. For $\mu_{Q}>\mu_{*}$, then, there must be a critical temperature, $T_{*}$, above which the smoothly varying, non-vanishing condensate becomes thermodynamically disfavored. This temperature is indicated by the red dashed curves in Figures \ref{FIG:FirstOrder}a and \ref{FIG:FirstOrder}b.
\begin{figure}[!h]
\centerline{ \epsfig{figure=O1_O2_area.pdf, height=2in} ~~~~~~~~ \epsfig{figure=flow.pdf, height=2in}}
\caption{\em (a) Flow lines in the neighborhood of the critical temperature, for $\mu_{t}=3$. (b) The flow switches direction discontinuously (lower solid to upper dashed curves) at a critical temperature $T_{*}$ indicated by the red dashed curve in
Figure \ref{type i}, leading to a first order phase transition at $T_{*}$. Here, $\mu_{Q}$=$3/8$ and $\Omega$=$1/16$.
\label{O1_O2_area}
} \end{figure}
We can verify this by computing the relative free energy of the condensed phase as we vary the temperature. Figure \ref{O1_O2_area}a shows the flows associated to points to the left and right of the critical temperature where the low- and high-temperature instabilities meet (indicated by the red dashed curve in Figure \ref{FIG:FirstOrder}b). Figure \ref{O1_O2_area}b focuses in on the immediate neighborhood of the transition temperature for $\mu_{Q}>\mu_{*}$. For all temperatures below the critical temperature, the flows go below the horizontal axis, corresponding to a negative free energy and a thermodynamically stable condensate. For all temperatures above the critical temperature, the flows go above the horizontal axis, so the condensate is thermodynamically disfavored at high temperatures.
Indeed, while the value of the condensate is non-vanishing and in fact completely smooth as we flow through $T_{*}$, the path which carries us from the trivial state to the condensate changes discontinuously as we pass through $T_{*}$. As a result, the integrated area -- and thus the free energy -- also changes discontinuously at $T_{*}$. Moreover, as we take $\mu_{Q}\to\mu_{*}$, the value of the condensate at the transition goes to zero, $\vev{{\cal O}(T_{*})}\to0$; this ensures that the latent heat of the transition goes to zero at the multicritical point where the transition switches from 1$^{st}$ to 2$^{nd}$ order, as expected on general grounds.
The upshot of all of the above is that as we raise $\mu_{Q}$, the phase transition from high-temperature metal to low-temperature superfluid switches from 2$^{nd}$ order to 1$^{st}$, with the transition occurring at a multicritical point where the low- and high-temperature superfluid phases collide.
Near the phase transition boundaries, including the multicritical point, the order parameter scales with simple mean-field exponents. More precisely, near the finite temperature 2$^{nd}$ order phase transition, $\langle {\cal O} \rangle \sim (T_{c}-T)^{1/2}$, while near the 1$^{st} $ order phase transition boundary when $\mu_Q>\mu_*$ the condensate jumps discontinuously at $T_*$, with $\langle {\cal O}(T_{*}) \rangle \sim (\mu_Q-\mu_*)^{1/2}$. This can be succinctly encoded in a simple mean-field free energy, $F(\varphi)= \frac{1}{2}c_2 (T-T_c(\mu_Q))\varphi^2+\frac{1}{4}c_4(\mu_*-\mu_Q) \varphi^4+ \frac{1}{6}c_6 \varphi^6$, with $\varphi\sim \langle {\cal O} \rangle$ and with coefficients $c_{2},c_{4},c_{6}>0$.
\subsection{Setting $\mu_{Q}=0$ and the Persistence of Condensates}
Playing with $\mu_{Q}$ raises another interesting point. Fundamental to our construction is that the scalar operator carries a charge $q$ under a global symmetry of the boundary theory. $\mu_{Q}$ tells us the energy cost for adding a unit of this charge to the system. In AdS, superfluid condensation is often induced by tuning $\mu_{Q}$ beyond a threshold. Is this also necessary in the non-relativistic case?
\begin{figure}[h]
\centerline{\epsfig{figure=zeromut_condensate_varymut.pdf, height=2 in} ~~~~~~~~ \epsfig{figure=zeromut_condensate, height=2 in} }
\caption{\em (a) Fixing $T\ll1$ and varying $\mu_{t}$. (b) Fixing $\mu_{t}=0$ and varying $T$. Here, $\Omega$=$1/16$.\label{FIG:Pairing}}
\end{figure}
By varying $\mu_{Q}$ to zero while holding all other parameters fixed (see Figure \ref{FIG:Pairing}a), we see that condensation persists even at vanishing $\mu_{Q}$. Indeed, by plotting $\vev{{\cal O}(T)}$ for $\mu_{Q}=0$ (see Figure \ref{FIG:Pairing}b), we see that the form of this curve is quite similar to the large-$\Omega$ case studied above with $\mu_{Q}\neq0$, modulo an overall scaling of the condensate. It is tempting to speculate that this indicates two distinct pairing mechanisms, one involving the charge and one involving the Mass eigenvalue alone. It would be interesting to explore this point further.
\section{Conclusions and Open Questions}\label{Sec:Conclusions}
In this paper we have constructed a toy model of superfluid states in holographic NRCFTs and studied the resulting phase diagram, finding several unanticipated features. First, as we lower the temperature in the disordered metallic state, the system generally undergoes a phase transition to a superfluid state. At small (and even vanishing) chemical potential, this transition is $2^{nd}$ order with mean-field exponents; at large chemical potential, however, the transition runs strongly $1^{st}$ order. Secondly, for large background mass density, the superfluid state only appears in a finite temperature window, with the metallic state reentering at sufficiently low temperature. Finally, at zero temperature, the reentrance of the metallic phase leads to an apparent quantum phase transition from superconducting to metallic as the background mass density is varied.
Several features of our results deserve further scrutiny. First, our low temperature results derive from a probe analysis which is not valid at zero temperature -- indeed, as we push the temperature to zero near or below the putative quantum phase transition at $\Omega_{*}$, the bulk profiles of various fields grow rapidly, diverging as we approach zero temperature. To be sure, we checked the consistency of the probe approximation in each calculation presented above.
However, it is entirely possible that various of our results could change qualitatively when we include backreaction. To nail down the $T=0$ physics, we must incorporate backreaction.
Relatedly, we have tacitly assumed that the neutral black hole geometry is the dominant saddle at $T=0$. However, for a variety of reasons including the strange thermodynamics of this black hole, this seems unlikely to be the case. It is tempting to speculate that the low-temperature phase is dominated by a Schr\"odinger soliton analogous to the AdS soliton which dominates the relativistic case a la Hawking-Page. Indeed, a simple such Schr\"odinger soliton solution is known, and we repeat the above analysis for this geometry in an Appendix. However, the black hole and the soliton enjoy incommensurate asymptotic periodicity conditions, so cannot contribute to the same ensembles. Understanding the true low-temperature ground state of the neutral black hole, even in the absence of any charge density in the system, is of considerable interest.
To this end, it is worth emphasizing that the basic trouble with the thermodynamics of this -- and indeed all known -- asymptotically Schr\"odinger black holes is the light-cone relation between the near-horizon killing vector $\partial_{\tau}$ and the asymptotic timelike killing vector, $\partial_{t}$. In most constructions, this follows from the structure of the salient solution generating technique. The challenge, then, is to build solutions which do not flow to $AdS$ black holes near the horizon.
Meanwhile, it's important to keep in mind that the simple Abelian-Higgs theory we study is an extremely stripped down toy model for which we do not have an explicit charged black hole solution. In the few examples where such a solution is known \cite{Adams:2009dm, Imeroni:2009cs}, the matter sectors are considerably more complicated, which is why we worked with the toy model at hand as a first step. It would be interesting to repeat our analysis in one of these more elaborate systems to disentangle the peculiarities of our toy model form general features of non-relativistic holographic superfluids.
Finally, several intriguing features of this system still need interpretation. Why does the second conformal family have a thermodynamically dominant high-temperature instability, what does that instability signal, and is there a simple low-energy way to see that this family is disfavored? Our conductivity calculations reveal a number of quasiparticle peaks with rather peculiar behavior, particularly near the critical point at $\Omega_{*}$ -- are these artifacts of the probe approximation, or do they signal real physics, and if so, what are they telling us? Does a proper holographic renormalization of the full system (which remains an open problem) alter any of our results, and if so how? What about the pairing mechanism -- what is the fermion spectral function in these systems, and can we correlate pairing of probe fermions with the condensation seen above? We hope to return to various of these questions in the future.
\section*{Acknowledgments}
We would like to thank
K.~Balasubramanian,
O.~DeWolfe,
N.~Iqbal,
S.~Kachru,
H.~Liu,
R.~Mahajan,
J.~McGreevy,
Y.~Nishida,
K.~Rajagopal,
and
D.~Vegh
for valuable discussions.
AA thanks the Aspen Center for Physics and the Stanford Institute for Theoretical Physics for hospitality during the completion of this work.
The work of AA is supported in part by funds provided by the U.S. Department of Energy (D.O.E.) under cooperative research agreement DE-FG0205ER41360.
|
\section{Introduction}
In this paper, we investigate the isomorphism types of combinatorial geometries arising from Hrushovski's flat strongly minimal structures and answer some questions from Hrushovski's original paper \cite{EH}. It is a sequel to \cite{MFDE1}, but can be read independently of it. In order to describe the main results it will be convenient to summarise some of the results from the previous paper.
Suppose $L$ is a relational language with, for convenience, all relation symbols of arity at least 3 and at most one relation symbol of each arity. Denote by $k(L)$ the maximum of the arities of the relation symbols in $L$ (allowing $k(L)$ to be $\infty$ if this is unbounded). The basic Hrushovski construction defines the \textit{predimension} of a finite $L$-structure to be its size minus the number of atomic relations on the structure. The class $\mathcal{C}_0(L)$ consists of the finite $L$-structures in which this is non-negative on all substructures. There is then an associated notion of \textit{dimension} $d$ and the notion of \textit{self-sufficiency} (denoted by $\leq$) of a substructure. All of this is reviewed in detail in Section 2 below. The class $(\mathcal{C}_0, \leq)$ has an associated \textit{generic structure} $\mathcal{M}_0(L)$ which also carries a dimension function $d$ giving it the structure of an infinite-dimensional pregeometry. The associated (combinatorial) geometry is denoted by $G(\mathcal{M}_0(L))$.
In \cite{MFDE1} we showed that:
\begin{enumerate}
\item[(1)] The collection of finite subgeometries of $G(\mathcal{M}_0(L))$ does not depend on $L$ (Theorem 3.8 of \cite{MFDE1}).
\item[(2)] For languages $L, L'$, the geometries $G(\mathcal{M}_0(L))$ and $G(\mathcal{M}_0(L'))$ are isomorphic iff the maximum arities $k(L)$ and $k(L')$ are equal. (Theorem 3.1 of \cite{MFDE1} for $\Leftarrow$ and see also Section 4.2 here; Theorem 4.3 of \cite{MFDE1} gives $\Rightarrow$.)
\item[(3)] The localization of $G(\mathcal{M}_0(L))$ over any finite set is isomorphic to $G(\mathcal{M}_0(L))$ (Theorem 5.5 of \cite{MFDE1}).
\end{enumerate}
For the strongly minimal set construction of \cite{EH}, one takes a certain function $\mu$ (see section 2 here) and considers a subclass $\mathcal{C}_\mu(L)$ of $\mathcal{C}_0(L)$. For appropriate $\mu$ there is a generic structure $\mathcal{M}_\mu(L)$ for the class $(\mathcal{C}_\mu(L), \leq)$ which is strongly minimal. The dimension function given by the predimension is the same as the dimension in the strongly minimal set and we are interested in the geometry of this. Our main result here is that this process of `collapse' is irrelevant to the geometry: under rather general conditions on $\mu$ we prove:
\begin{enumerate}
\item[(4)] The geometry $G(\mathcal{M}_\mu(L))$ of the strongly minimal set is isomorphic to the geometry $G(\mathcal{M}_0(L))$ (Theorem \ref{main}).
\end{enumerate}
Sections 5.1 and 5.2 of Hrushovski's paper \cite{EH} give variations on the construction which produce strongly minimal sets with geometries different from the $G(\mathcal{M}_0(L))$. However, we show, answering a question from \cite{EH} (see also Section 3 of \cite{Hasson}):
\begin{enumerate}
\item[(5)] the geometries of the strongly minimal sets in Sections 5.1 and 5.2 of \cite{EH} have localizations (over a finite set) which are isomorphic to one of the geometries $G(\mathcal{M}_0(L))$ (for appropriate $L$) (see Section 4.1 here).
\end{enumerate}
The first version of the result in (4) was proved by the second Author in his thesis \cite{MFThesis}: this was for the case where $L$ has a single 3-ary relation symbol (as in the original paper \cite{EH}). The somewhat different method of proof used in Sections 3 and 4 here was found later. It has the advantage of being simpler and more readily adaptable to generalization and proving the result in (5), however, the class of $\mu$-functions to which it is applicable is slightly more restricted than the result from \cite{MFThesis}: Theorem 6.2.1 of \cite{MFThesis} assumes only that $\mu \geq 1$.
\medskip
In summary, for each $k = 3, 4, \ldots, \infty$ we have a countably-infinite dimensional geometry $\mathcal{G}_k$ isomorphic to $G(\mathcal{M}_0(L))$ where $L$ has maximum arity $k$, and these are pairwise non-isomorphic. The geometry of each of the new (countable, saturated) strongly minimal sets in \cite{EH} has a localization isomorphic to one of these $\mathcal{G}_k$. Thus, whilst there is some diversity amongst the strongly minimal structures which can be produced by these constructions, the range of geometries which can be produced appears to be rather limited. It would therefore be very interesting to have a characterization of the geometries $\mathcal{G}_k$ in terms of a `geometric' condition (such as flatness, as in 4.2 of \cite{EH}, for example) and a condition on the automorphism group (such as homogeneity, but possibly with a stronger assumption).
\medskip
\noindent\textit{Acknowledgement:\/} Some of the results of this paper were produced whilst the second Author was supported as an Early Stage Researcher by the Marie Curie Research Training Network MODNET, funded by grant MRTN-CT-2004-512234 MODNET from the CEC. We thank the Referee for drawing our attention to the need to consider local isomorphisms in Theorem \ref{main}.
\section{Hrushovski constructions}\label{sec2}
We give a brief description of Hrushovski's constructions from \cite{EH}. Other presentations can be found in \cite{FW} and \cite{B&S}. The book \cite{AP2} of Pillay contains all necessary background material on pregeometries and model theory. The notation, terminology and level of generality is mostly consistent with that used in \cite{MFDE1}.
\subsection{Predimension and pregeometries}
Let $L$ be a relational language consisting of relation symbols $(R_i : i \in I)$ with $R_i$ of arity $n_i \geq 3$ (and $\vert I \vert \geq 1$). We suppose there are only finitely many relations of each arity here.
We work with $L$-structures $A$ where each each $R_i$ is symmetric: so we regard the interpretation $R_i^A$ of $R_i$ in $A$ as a set of $n_i$-sets. (By modifying the language, the arguments we give below can be adapted to deal with the case of $n_i$-tuples of not-necessarily-distinct elements: see Section 4.3 here.)
For finite $A$ we let the predimension of $A$ be $\delta(A) = \vert A \vert - \sum_{i \in I} \vert {R_i}^A\vert$ (of course this depends on $L$ but this will be clear from the context).
We let
$\mathcal{C}_0(L)$ be the set of finite $L$-structures $A$ such that $\delta(A') \geq 0$ for all $A' \subseteq A$.
Suppose $A \subseteq B \in \mathcal{C}_0(L)$. We write $A \leq B$ and say that $A$ is \textit{self-sufficient} in $B$ if for all $B'$ with $A \subseteq B' \subseteq B$ we have $\delta(A) \leq \delta(B')$. We will assume that the reader is familiar with the basic properties (such as transitivity) of this notion.
Let $\bar{\mathcal{C}}_0(L)$ be the class of $L$-structures all of whose finite substructures lie in $\mathcal{C}_0(L)$. We can extend the notion of self-sufficiency to this class in a natural way.
Note that if $A \subseteq B \in \bar{\mathcal{C}}_0(L)$ is finite then there is a finite $A'$ with $A \subseteq A' \subseteq B$ and $\delta(A')$ as small as possible. In this case $A' \leq B$ and it can be shown that there is a smallest finite set $C \leq B$ with $A \subseteq C$. We define the \textit{dimension} $d_B(A)$ of $A$ (in $B$) to be the minimum value of $\delta(A')$ for all finite subsets $A'$ of $B$ which contain $A$.
We define the $d$-\textit{closure} of $A$ in $B$ to be:
$${\rm cl}_B(A)=\{c\in B:d_{B}(Ac)=d_{B}(A)\}$$
where, as usual, $Ac$ is shorthand for $A\cup \{c\}$.
These notions can be relativized. If $A, C \subseteq B \in \mathcal{C}_0(L)$ define $\delta(A/C)$ to be $\delta(A\cup C)- \delta(C)$ and $d_B(A/C) = d_B(A\cup C) - d_B(C)$: the (pre)dimension \textit{over} $C$. In all of this notation, we will suppress the subscript for the ambient structure $B$ if it is clear from the context.
We can coherently extend the definition of $d$-closure to infinite subsets $A$ of $B$ by saying that the $d$-closure of $A$ is the union of the $d$-closures of finite subsets of $A$.
It can be shown that $(B,{\rm cl}_B)$ is a pregeometry and the dimension function (as cardinality of a basis) equals $d_{B}$ on finite subsets of $B$. We use the notation $PG(B)$ instead of $(B,{\rm cl}_B)$, and denote by $G(B)$ the associated (combinatorial) geometry: so the elements of $G(B)$ are the sets ${\rm cl}_B(x)\setminus{\rm cl}_B(\emptyset)$ for $x \in B \setminus {\rm cl}_B(\emptyset)$ and the closure on $G(B)$ is that induced by ${\rm cl}_B$.
Note that if $A \leq B \in \bar{\mathcal{C}}_0(L)$ then for $X \subseteq A$ we have $d_A(X) = d_B(X)$. Thus $G$ can be regarded as a functor from $(\bar{\mathcal{C}}_0(L), \leq)$ to the category of geometries (with embeddings of geometries as morphisms).
If $Y \subseteq B \in \bar{\mathcal{C}}_0(L)$ then the \textit{localization} of $PG(B)$ over $Y$ is the pregeometry with closure operation ${\rm cl}_B^Y(Z) = {\rm cl}_B(Y\cup Z)$. The corresponding geometry is denoted by $G_Y(B)$. Note that the dimension function here is given by the relative dimension $d_B(./Y)$.
It will be convenient to fix a first order language for the class of pregeometries. A reasonable choice for this is the language $LPI=\{I_n:n\geq 1\}$ where each $I_n$ is an $n$-ary relational symbol. A pregeometry $(P,{\rm cl})$ will be seen as a structure in this language by taking $I_n^P$ to be the set of independent $n$-tuples in $P$. Notice that we can recover a pregeometry just by knowing its finite independent sets. Note also that the isomorphism type of a pregeometry is determined by the isomorphism type of its associated geometry and the size of the equivalence classes of interdependence. In the case where these are all countably infinite, it therefore makes no difference whether we consider the geometry or the pregeometry.
\subsection{Self-sufficient amalgamation classes}
If $B_1, B_2 \in \bar{\mathcal{C}}_0(L)$ have a common substructure $A$ then the \textit{free amalgam} $E$ of $B_1$ and $B_2$ over $A$ consists of the disjoint union of $B_1$ and $B_2$ over $A$ and $R_i^E = R_i^{B_1} \cup R_i^{B_2}$ for each $i \in I$. It is well known that if $A \leq B_1$ then $B_2 \leq E$, so $E \in \bar{\mathcal{C}}_0(L)$, and $(\mathcal{C}_0, \leq)$ is an \textit{amalgamation class}. It can also be shown that if $A$ is $d$-closed in $B_1$ then $B_2$ is $d$-closed in $E$.
\smallskip
Suppose $Y \leq Z \in \mathcal{C}_0(L)$ and $Y \neq Z$. Following \cite{EH}, we say that this is an \textit{algebraic extension} if $\delta(Y) = \delta(Z)$. It is a simply algebraic extension if also $\delta(Z') > \delta(Y)$ whenever $Y \subset Z' \subset Z$. It is a minimally simply algebraic (msa) extension if additionally $Y' \subseteq Y'\cup (Z\setminus Y)$ is not simply algebraic whenever $Y' \subset Y$.
It can be shown that for each $n \geq 0$ there are arbitrarily large msa extensions $Y \leq Z$ in $\mathcal{C}_0(L)$ with $\delta(Y) = n$ (this make use of the fact that at least one of the relations $R_i$ has arity at least 3, which is certainly covered by our assumptions on $L$).
The following is trivial, but crucial for us:
\begin{lem}\label{21}
Suppose $Y \leq Z$ is a msa extension. Then for every $y \in Y$ there is some $w \in \bigcup_{i \in I} R_i^Z$ and $z \in Z\setminus Y$ such that $y, z \in w$. Moreover, if $Z\setminus Y$ is not a singleton and $z \in Z\setminus Y$, then there are at least two elements of $\bigcup_{i \in I} R_i^Z$ which contain $z$.
\end{lem}
\begin{proof}
Suppose this does not hold for some $y \in Y$. Let $Y' = Y\setminus \{y\}$. Then for every $U \subseteq Z\setminus Y$ we have $\delta(Y'\cup U) - \delta(Y') = \delta(Y \cup U) - \delta(Y)$. So $Y' \subseteq Y'\cup (Z\setminus Y)$ is simply algebraic: contradiction. Similarly, for the `moreover' part, if $z$ is in at most one relation in $\bigcup_{i \in I} R_i^Z$, then $\delta(Z\setminus\{z\}) \leq \delta(Z)$, which contradicts the simple algebraicity.
\end{proof}
We let $\mu$ be a function from the set of isomorphism types of minimally simply algebraic extensions in $\mathcal{C}_0(L)$ to the non-negative integers. The subclass $\mathcal{C}_\mu(L)$ consists of structures in $\mathcal{C}_0(L)$ which, for each msa $Y\leq Z$ in $\mathcal{C}_0(L)$, omit the atomic type consisting of $\mu(Y,Z)+1$ disjoint copies of $Z$ over $Y$.
We will work with $\mu$ where the following holds:
\begin{ass}[Assumed Amalgamation Lemma]\label{aalemma}
\begin{enumerate}
\item[(i)] If $A \leq B_1, B_2 \in \mathcal{C}_\mu(L)$ and the free amalgam of $B_1$ and $B_2$ over $A$ is not in $\mathcal{C}_\mu(L)$, then there exists $Y \subseteq A$ and minimally simply algebraic extensions $Y \leq Z_i \in B_i$ (for $i = 1,2$) which are isomorphic over $Y$ and $Z_i\setminus Y \subseteq B_i\setminus A$.
\item[(ii)] The class $(\mathcal{C}_\mu(L), \leq)$ is an amalgamation class (see below).
\end{enumerate}
\end{ass}
Note that (ii) here follows from (i) (cf. the proof of Lemma 4 in \cite{EH}), and by Section 2 of \cite{EH}, (i) holds if $\mu(Y,Z) \geq \delta(Y)$ for all msa $Y\leq Z$ in $\mathcal{C}_0(L)$.
\subsection{Generic structures and their geometries}
Suppose $\mathcal{A}$ is a subclass of $\mathcal{C}_0(L)$ such that $(\mathcal{A}, \leq)$ is an amalgamation class: meaning that if $B \in \mathcal{A}$ and $A \leq B$ then $A \in \mathcal{A}$, and if $A \leq B_1, B_2 \in \mathcal{A}$ then there is $C \in \mathcal{A}$ and embeddings $f_i : B_i \to C$ with $f_i(B_i) \leq C$ and $f_1 \vert A = f_2 \vert A$. Then there is a countable structure $\mathcal{M} \in \bar{\mathcal{C}}_0(L)$ satisfying the following conditions:\begin{enumerate}
\item [(G1)] $\mathcal{M}$ is the union of a chain $A_0 \leq A_1 \leq A_2 \leq \cdots$ of structures in $\mathcal{A}$.
\item [(G2)] (extension property) If $A\leq\mathcal M$ and $A\leq B\in\mathcal A$ then there exists an embedding $g:B\to\mathcal M$ such that $g(B)\leq\mathcal M$ and $g(a) = a$ for all $a \in A$.
\end{enumerate}
We refer to $\mathcal{M}$ as the \textit{generic structure} of the amalgamation class $(\mathcal{A}, \leq)$: it is determined up to isomorphism by the properties G1 and G2 (and G1 is automatic for countable structures in $\bar{\mathcal{C}}_0(L)$). Of course, Hrushovski's strongly minimal sets are the generic structures $\mathcal{M}_\mu(L)$ for the amalgamation classes $(\mathcal{C}_\mu(L), \leq)$. We will compare the geometries of these with that of the generic structure $\mathcal{M}_0(L)$ for the amalgamation class $(\mathcal{C}_0(L), \leq)$.
Suppose $(\mathcal{A}, \leq)$ and $(\mathcal{A}', \leq)$ are amalgamation classes, as above. We refer to the following as the Isomorphism Extension Property, and denote it by $\mathcal{A} \rightsquigarrow \mathcal{A}'$.
\begin{enumerate}
\item[(*)] Suppose $A \in \mathcal{A}$, $A' \in \mathcal{A}'$ and $f : G(A) \to G(A')$ is an isomorphism of geometries, and $A \leq B \in \mathcal{A}$. Then there is $B' \in \mathcal{A}'$ with $A' \leq B'$ and an isomorphism $f' : G(B) \to G(B')$ which extends $f$.
\end{enumerate}
\begin{lem}\label{isolemma}
Suppose $(\mathcal{A}, \leq)$ and $(\mathcal{A}', \leq)$ are amalgamation classes with generic structures $\mathcal{M}$, $\mathcal{M}'$ respectively. Suppose that both extension properties $\mathcal{A} \rightsquigarrow \mathcal{A}'$ and $\mathcal{A}' \rightsquigarrow \mathcal{A}$ hold. Then the geometries $G(\mathcal{M})$ and $G(\mathcal{M}')$ are isomorphic.
\end{lem}
\begin{proof}
We have already remarked that if $A \leq \mathcal{M}$, then the dimension of a subset of $A$ is the same whether computed in $A$ or in $\mathcal{M}$. Thus $G(A)$ is naturally a substructure of $G(\mathcal{M})$. We claim that the set $\mathcal{S}$ of geometry-isomorphisms
\[ f : G(A) \to G(A') \]
where $A \leq \mathcal{M}, A' \leq \mathcal{M}'$ are finite
is a back-and-forth system between $G(\mathcal{M})$ and $G(\mathcal{M}')$. Indeed (for the `forth'), given such an $f : G(A) \to G(A')$ and $A \leq B \leq \mathcal{M}$, there is $A' \leq B' \in \mathcal{A}'$ and an isomorphism $f' : G(B) \to G(B')$ extending $f$, by our assumption $\mathcal{A} \rightsquigarrow \mathcal{A}'$. The extension property G2 in $\mathcal{M}'$ means that we can take $B' \leq \mathcal{M}'$, as required. Similarly we obtain the `back' part from $\mathcal{A}' \rightsquigarrow \mathcal{A}$ and G2 in $\mathcal{M}$.
It follows that $G(\mathcal{M})$ and $G(\mathcal{M}')$ are isomorphic.
\end{proof}
\begin{rem} \label{isorem}\rm
We can adapt this slightly to give a criterion for local isomorphism of $G(\mathcal{M})$ and $G(\mathcal{M}')$. This is really just about adding parameters to the language, but we make it explicit. Suppose $X \in \mathcal{A}$ and $X' \in \mathcal{A}'$ are fixed finite structures with $X \leq \mathcal{M}$ and $X' \leq \mathcal{M}'$. We write $\mathcal{A}(X) \rightsquigarrow \mathcal{A}'(X')$ for the statement:
\begin{enumerate}
\item[]
Suppose $X \leq A \in \mathcal{A}$, $X' \leq A' \in \mathcal{A}'$ and $f : G_X(A) \to G_X'(A')$ is an isomorphism of geometries, and $A \leq B \in \mathcal{A}$. Then there is $B' \in \mathcal{A}'$ with $A' \leq B'$ and an isomorphism $f' : G_X(B) \to G_{X'}(B')$ which extends $f$.
\end{enumerate}
Here, recall that $G_X(A)$ is the localization of $G(A)$ over $X$, as defined in Section 2.1. It follows as in Lemma \ref{isolemma} that if $\mathcal{A}(X) \rightsquigarrow \mathcal{A}'(X')$ and $\mathcal{A}'(X') \rightsquigarrow \mathcal{A}(X)$ hold, then $G_X(\mathcal{M})$ and $G_{X'}(\mathcal{M}')$ are isomorphic.
\end{rem}
\section{Isomorphism of the strongly minimal set geometries}
Throughout, $(\mathcal{C}_0(L), \leq)$ and $(\mathcal{C}_\mu(L), \leq)$ are the amalgamation classes from the previous section. Note that $(\mathcal{C}_0(L), \leq)$ is an amalgamation class and we are \textit{assuming} that the amalgamation lemma \ref{aalemma} holds for $\mathcal{C}_\mu(L)$. We denote the generic structures by $\mathcal{M}_0(L)$ and $\mathcal{M}_\mu(L)$ respectively: so the latter is Hrushovski's strongly minimal set $D(L,\mu)$. The geometries are denoted by $G(\mathcal{M}_0(L))$ and $G(\mathcal{M}_\mu(L))$. We have already remarked that there are arbitrarily large msa extensions $\emptyset \leq Z$ in $\mathcal{C}_0(L)$, so if $\mu(\emptyset, Z) > 0$ for infinitely many of these, then ${\rm cl}_{\mathcal{M}_\mu(L)}(\emptyset)$ is infinite as it contains a copy of each such $Z$. Similarly, there are infinitely many msa extensions $Y \leq Z \in \mathcal{C}_0(L)$ with $\delta(Y) = 1$, so if $\mu(Y, Z) > 0$ for all of these, and $X \leq \mathcal{M}_\mu(L)$ is a singleton, then ${\rm cl}_{\mathcal{M}_\mu(L)}(X)$ is infinite. Our main result is:
\begin{thm}\label{main} Suppose \ref{aalemma} holds and $\mu(Y,Z) \geq 2$ for all msa $Y \leq Z \in \mathcal{C}_0(L)$ with $\delta(Y) \geq 2$ and $\mu(Y,Z) \geq 1$ when $\delta(Y) = 1$. Suppose that $X \leq \mathcal{M}_\mu(L)$ is finite and ${\rm cl}_{\mathcal{M}_\mu(L)}(X)$ is infinite. Then $G_X(\mathcal{M}_\mu(L))$ and $G(\mathcal{M}_0(L))$ are isomorphic geometries. In particular, $G(\mathcal{M}_\mu(L))$ and $G(\mathcal{M}_0(L))$ are locally isomorphic.
\end{thm}
\begin{proof}
Note that $X$ can be taken as $\emptyset$ or a singleton, by the remarks preceding the theorem.
We need to verify that the isomorphism extension property of Lemma \ref{isolemma} and Remarks \ref{isorem} holds in both directions. The main part will be to show that
$\mathcal{C}_0(L) \rightsquigarrow \mathcal{C}_\mu(L)(X)$. In fact, because of the symmetry of the argument, it will be convenient to take an arbitrary finite $W \leq \mathcal{M}_0(L)$ and show that $\mathcal{C}_0(L)(W) \rightsquigarrow \mathcal{C}_\mu(L)(X)$.
\medskip
So suppose we are given $W \leq A \leq B \in \mathcal{C}_0(L)$ and $X \leq A' \in \mathcal{C}_\mu(L)$ with an isomorphism $f : G_W(A) \to G_X(A')$. We want to find $B' \in \mathcal{C}_\mu(L)$ with $A' \leq B'$ and an isomorphism $f' : G_W(B) \to G_X(B')$ extending $f$. The main point will be to ensure that each point of $B'\setminus (A' \cup {\rm cl}_{B'}(X))$ is involved in only a small number of relations, and this gives us control over the msa extensions in $B'$.
Let $A_0 = {\rm cl}_A(W)$ and let $A_1,\ldots, A_r$ be the $d$-dependence classes (over $W$) on $A\setminus A_0$: the latter are the points of $G_W(A)$. Similarly let $B_0 = {\rm cl}_B(W)$ and $B_1, \ldots, B_s$ the $d$-dependence classes (over $W$) on $B\setminus B_0$, with $A_i \subseteq B_i$ for $i = 1,\ldots, r$. List the relations on $B$ which are not contained in $A$ or some $B_0\cup B_j$ as $\rho_1,\ldots, \rho_t$. So these are finite sets. Let $A_0' = {\rm cl}_{A'}(X)$ and $A_1',\ldots, A_r'$ be the classes of $d$-dependence (over $X$) on $A'\setminus A_0'$, labelled so that $f(A_i) = A_i'$. We construct $B_0', B_1', \ldots, B_s'$ with $A_i' \subseteq B_i'$ for $i = 0, \ldots, r$, and $B' = \bigcup_{i = 0}^s B_i'$ in the steps below.
\medskip
\noindent\textit{Terminology:\/} If $u, v \in E \in \mathcal{C}_0(L)$, say that $u,v$ are \textit{adjacent in $E$} if there exists $w \in \bigcup_i R_i^E$ such that $u,v \in w$.
\medskip
\noindent\textit{Step 1:\/} Construction of $A'' = A' \cup B_0' \in \mathcal{C}_\mu(L)$. \newline
Take $A_0' \leq V \in \mathcal{C}_\mu(L)$ with $\delta(V) = \delta(X)$ and $\vert V \setminus A_0' \vert $ sufficiently large. For example, by our assumption on $X$, $\mathcal{C}_\mu(L)$ contains arbitrarily large finite algebraic extensions of $X$, so we can take $V \in \mathcal{C}_\mu(L)$ to be an amalgamation of $A_0'$ and one of these over $X$. Let $A''$ be the free amalgam of $A'$ and $V$ over $A_0'$ and let $B_0'$ be the copy of $V$ inside this. As $A_0'$ is $d$-closed in $A'$ and $A_0' \leq V$ it follows from \ref{aalemma} that $A'' \in \mathcal{C}_\mu(L)$, $B_0'$ is $d$-closed in $A''$ and $A' \leq A''$. Note that $\delta(A'') = \delta(A')$.
\medskip
\noindent\textit{Step 2:\/} Construction of $B_0' \cup B_i' \in \mathcal{C}_\mu(L)$.\newline
We do this so that $B_0'$ is $d$-closed in $B_0' \cup B_i'$ and $\delta(B_i'\cup B_0'/B_0') = 1$. (As $\delta(B_0' \cup A_i'/B_0') = 1$, it then follows that $B_0' \cup A_i' \leq B_0'\cup B_i'$ when $1 \leq i \leq r$.) Let $m$ be sufficiently large. Choose some $R_i$: for example $R_1$ of arity $n \geq 3$. \newline
\smallskip
\textit{Case 1:\/} Suppose $i \leq r$. Pick $b_{i0} \in A_i'$ and let $s_{i1},\ldots, s_{im}$ be disjoint $(n-2)$-subsets of $B_0'\setminus A_0'$ (we adjust the choice of $V$ in step 1 to accommodate this). Let $B_i' = A_i' \cup \{b_{i1},\ldots, b_{im}\}$ and include as new $R_1$-relations on $B_0' \cup B_i'$ the $n$-sets $\{b_{i0}, b_{ij}\}\cup s_{ij}$ for $1\leq j\leq m$. We need to show that this has the required properties.
First, note that $B_0'\cup A_i' \leq B_0' \cup A_i' \cup \{b_{ij}\}$, so $B_0' \cup A_i' \cup \{b_{ij}\} \in \mathcal{C}_0(L)$.
Suppose $Y \leq Z$ is a msa extension in $B_0' \cup A_i' \cup \{b_{ij}\}$ not contained in $B_0'\cup A_i'$. So $s_{ij} \cup \{b_{i0},b_{ij}\} \subseteq Z$. If $b_{ij} \not\in Y$ then $Y \leq Z\setminus\{b_{ij}\} < Z$ is algebraic, so $Z\setminus \{b_{ij}\} = Y$ and $Y = s_{ij}\cup \{b_{i0}\}.$ As the elements of $s_{ij}$ are non-adjacent to $b_{i0}$ in $B_0'\cup A_i'$, it follows that there is only one copy of $Z$ over $Y$ in $B_0' \cup A_i' \cup \{b_{ij}\}$. If $b_{ij} \in Y$, then by Lemma \ref{21} $s_{ij}\cup \{b_{i0}\} \not\subseteq Y$. But then there is at most one copy of $Z$ over $Y$ in $B_0' \cup A_i' \cup \{b_{ij}\}$. Note that $\delta(Y) = \delta(Z) \geq \delta(Z\cap (A_i' \cup B_0')) \geq 1$. So in both cases we meet the requirements for $B_0' \cup A_i' \cup \{b_{ij}\} \in \mathcal{C}_\mu(L)$.
Now note that $B_0' \cup B_i'$ is the free amalgam over $B'_0\cup A_i'$ of the structures $B_0'\cup A_i' \cup \{b_{ij}\}$ (for $j = 1, \ldots, m$). Each $B_0'\cup A_i' \subseteq B_0'\cup A_i' \cup \{b_{ij}\}$ is an algebraic extension and the only msa extension in this with base in $B_0' \cup A_i'$ and which is not contained in $B_0'\cup A_i'$ is $s_{ij}\cup \{b_{i0}\} \leq s_{ij}\cup \{b_{i0}, b_{ij}\}$. So the amalgamation lemma \ref{aalemma} implies that $B_0' \cup A_i' \leq B_0'\cup B_i' \in \mathcal{C}_\mu(L)$. It is clear that $\delta(B_i'/B_0' \cup A_i') = 0$ so $\delta(B_0' \cup B_i') = \delta(B_0' \cup A_i') = \delta(B_0') + 1$.
Finally, note that as $B_0'$ is $d$-closed in $B_0' \cup A_i'$ and $B_0' \cup A_i' \leq B_0' \cup B_i'$, the $d$-closure of $B_0'$ in $B_0'\cup B_i'$ does not contain $b_{i0}$. It then follows easily that $B_0'$ is $d$-closed in $B_0'\cup B_i'$.
\smallskip
\noindent\textit{Case 2:\/} $i > r$. As in Case 1, let $s_{i1},\ldots, s_{im}$ be $(n-2)$-subsets of $B_0'\setminus A_0'$ with no relations on them. Let $B_i' = \{b_{i1},\ldots, b_{im}\}$ and include as new $R_1$-relations on $B_0' \cup B_i'$ the $n$-sets $\{b_{ij}, b_{i(j+1)}\}\cup s_{ij}$ for $1\leq j\leq m-1$. In this version of the construction we take the $s_{ij}$ to be $s_i$, independent of $j$.
It is clear that $\delta(B_0'\cup B_i'/B_0') = 1$ and if $\emptyset\neq Y \subseteq B_i'$ then $\delta(B_0' \cup Y/B_0') \geq 1$. So $B_0' \leq B_0'\cup B_i'$ and therefore $B_0'\cup B_i' \in \mathcal{C}_0(L)$, and $B_0'$ is $d$-closed in $B_0' \cup B_i'$. It remains to show that $B_0'\cup B_i' \in \mathcal{C}_\mu(L)$, so suppose $Y \leq Z_1$ is a msa extension in $B_0'\cup B_i'$. If $\delta(Y) = 0$ then $Z_1\subseteq B_0'$ and the same is true of any copy of $Z_1$ over $Y$. Similarly, if $Y \subseteq B_0'$ then all copies of $Z_1$ over $Y$ are contained in $B_0'$ as this is $d$-closed in $B_0'\cup B_i'$. So we can assume that $\delta(Y) \geq 1$ and $b_{ij} \in Y$ for some $j$. By Lemma \ref{21}, we can assume that one of the relations $s_i \cup \{b_{ij},b_{i(j\pm1)}\}$ is a subset of $Z_1$.
If $Z_1 \setminus Y$ is a singleton then there is at most one other copy $Z_2$ of $Z_1$ over $Y$, and in this case $Y = Z_1 \cap Z_2 = s_i\cup \{b_{ij}\}$. Note that $\delta(Y) \geq 2$ here, so $\mu(Y, Z_1) \geq 2$, by hypothesis.
Now suppose that $Z_1 \setminus Y$ has at least two elements. It will suffice to prove that there is no other copy $Z_2$ of $Z_1$ over $Y$ in $B_0'\cup B_i'$, so suppose there is such a $Z_2$. Take $j$ maximal such that $b_{ij} \in Z_1\cup Z_2$. By Lemma \ref{21}, $b_{ij}$ is in at least two relations in $Z_1\cup Z_2$; but $b_{ij}$ is only in two relations in $B_0'\cup B_i'$ and one of these also involves $b_{i(j+1)}$, so this is in $Z_1\cup Z_2$. This contradicts the choice of $j$.
\medskip
\noindent\textit{Step 3:\/} Other relations on $B'$. \newline
The relations on $B'$ not contained in $A'$ or some $B_0' \cup B_i'$ are $\rho_1', \ldots \rho_t'$. We can choose these to be subsets of $B'\setminus A''$ with $\rho_i' \cap \rho_j' = \emptyset$ if $i\neq j$, and $\rho_i' \cap B_j' \neq \emptyset$ iff $\rho_i \cap B_j \neq \emptyset$ (for $j \geq 1$). Note that this is possible if $m$ is sufficiently large. We make $\rho_i'$ of the same type as $\rho_i$ (that is, in the same $R_j$).
This completes the construction of $B'$. We now make a series of claims about it.
\medskip
\noindent\textit{Claim 1:\/} Let $U \subseteq \{1,\ldots, s\}$, and $Y = \bigcup_{i\in U} (B_i\cup B_0)$, $Y' = \bigcup_{i\in U} (B_i' \cup B_0')$. Then $Y \cap A$ is $d$-closed in $A$ iff $Y' \cap A'$ is $d$-closed in $A'$, and in this case we have $\delta(Y/B_0) = \delta(Y'/B_0').$
\smallskip
Let $U_0 = \{i\in U : i \leq r\}$ and $U_1 = \{i \in U : i > r\}$. Then $Y\cap A = \bigcup_{i \in U_0} (A_0\cup A_i)$ and $Y' \cap A' = \bigcup_{i \in U_0} (A_0'\cup A_i')$. Because $f$ is an isomorphism of geometries, one of these is $d$-closed (in $A$ or $A'$) iff the other is (remembering that a subset of a geometry is $d$-closed iff any set properly containing it has bigger dimension). So suppose this is the case. We compute that:
\[\delta(Y'/B_0') = \delta(A''\cap Y'/B_0') + \delta(Y'/A''\cap Y') = \delta(A''\cap Y'/B_0') + \vert U_1\vert - \vert J\vert\]
where $J = \{ j : \rho_j' \subseteq Y'\}$. This follows from the fact that $Y'$ consists of $\vert U_0\vert$ sets of $\delta$-value $0$ over $A''\cap Y'$ and $\vert U_1\vert$ sets of $\delta$-value $1$ over $A''\cap Y'$, and an extra $\vert J\vert$ relations $\rho'_j$ between them. Moreover
\[\delta(A'' \cap Y'/B_0') = \delta ((A'\cap Y')\cup B_0'/B_0') = \delta(A' \cap Y'/A_0'),\]
using, for example, the construction of $A''$ as a free amalgam in step 1. Thus we have
\[\delta(Y'/B_0') = \delta(A'\cap Y'/A_0') + \vert U_1\vert - \vert J\vert.\]
Now, by construction (step 3) we have $\rho_j \subseteq Y$ iff $\rho'_j \subseteq Y'$. So an identical calculation shows that
\[\delta(Y/B_0) = \delta(A\cap Y/A_0) + \vert U_1\vert - \vert J\vert.\]
(This uses the fact that $A, B_0$ are freely amalgamated over $A_0$, which follows from the definitions of $A_0$, $B_0$ and the assumption that $A\leq B$.)
By the isomorphism $f$, we have $d(A'\cap Y'/X) = d(A\cap Y/W)$. If $A'\cap Y'$, $A\cap Y$ are $d$-closed (in $A'$, $A$ respectively) then $d(A'\cap Y') = \delta(A'\cap Y')$ and $d(A\cap Y) = \delta(A \cap Y)$. So in this case $\delta(A\cap Y/ A_0) = \delta(A' \cap Y' / A_0')$, as $d(W) = \delta(A_0)$ and $d(X) =\delta(A_0')$. Thus we have $\delta(Y/B_0) = \delta(Y'/B_0')$, as required. \hfill$\Box_{Claim}$
\medskip
\noindent\textit{Claim 2:\/} $B' \in \mathcal{C}_0(L)$, $B_0'$ and $B_0'\cup B_i'$ are $d$-closed in $B'$ and $A'' \leq B'$. The map $f' : G_W(B) \to G_X(B')$ given by $f'(B_i) = B_i'$ is an isomorphism of geometries which extends $f$.
\smallskip
First, note that (by construction step 2) if $B_0'\subseteq C \subseteq B'$ and $Y' = \bigcup\{ B_0'\cup B_i' : C \cap B_i' \neq \emptyset\}$, then $\delta(C) \geq \delta(Y')$. If $A'' \subseteq C$ then by Claim 1, $\delta(Y'/B_0') = \delta(Y/B_0)$, where $Y = \bigcup\{ B_0\cup B_i : C \cap B_i' \neq \emptyset\}$. This contains $A$, so as $A \leq B$ we have $\delta(Y) \geq \delta(A)$. By the isomorphism, $\delta(A/A_0) = \delta(A'/A_0') = \delta(A''/B_0')$ (by step 1). Thus $\delta(C/B_0') \geq \delta(Y'/B_0') = \delta(Y/B_0) \geq \delta(A/A_0) = \delta(A''/B_0')$ (using that $\delta(A_0) = \delta(B_0)$). This shows $A''\leq B'$ and therefore (as $\emptyset \leq A''$) we also have $\emptyset \leq B'$.
Now suppose $Y' \geq X$ is $d$-closed in $B'$. Then $B_0' \subseteq Y'$ and as above, $Y'$ is of the form $\bigcup_{i \in U} (B_0'\cup B_i')$ for some $U$. Moreover $Y'\cap A'$ is $d$-closed in $A'$ and so we can apply Claim 1. It follows from this that $B_0'$ is $d$-closed in $B'$ and the $d$-closed sets of dimension 1 over $B_0'$ are the $B_0'\cup B_i'$, by using the fact that the corresponding statements hold in $B$.
It remains to show that $f'$ is an isomorphism of geometries. Let $Y$, $Y'$ be as in Claim 1. We need to show that $Y$ is $d$-closed in $B$ iff $Y'$ is $d$-closed in $B'$. So suppose $Y$ is $d$-closed in $B$. Then $Y\cap A$ is $d$-closed in $A$ and so we can apply Claim 1 to get that $\delta(Y/B_0) = \delta(Y'/B_0')$. Suppose for a contradiction that $Y'$ is not $d$-closed in $B'$. Let $Z'$ be its $d$-closure in $B'$. So $Z' = \bigcup_{i \in Q} (B_0'\cup B_i')$ for some $Q$ with $U \subset Q \subseteq \{1,\ldots, s\}$ and $\delta(Z') \leq \delta(Y')$. Let $Z = \bigcup_{i \in Q} (B_0\cup B_i)$. So $Y \subset Z$. Because $Z'$ is $d$-closed in $B'$ and therefore $Z'\cap A'$ is $d$-closed in $A'$, we can apply Claim 1 to get that $\delta(Z/B_0) = \delta(Z'/B_0')$. So we have
\[\delta(Z/B_0) = \delta(Z'/B_0') \leq \delta(Y'/B_0') = \delta(Y/B_0)\]
and this contradicts the fact that $Y$ is $d$-closed and $Y \subset Z$. Thus $Y'$ is $d$-closed in $B'$. The argument for the converse implication is the same.\hfill$\Box_{Claim}$
\medskip
\noindent\textit{Claim 3:\/} $B' \in \mathcal{C}_\mu(L)$.
\smallskip
Suppose that $Y \leq Z$ is a minimally simply algebraic extension in $B'$. First suppose $\delta(Y) = \delta(Z) \leq 1$. Then $d(Y) \leq 1$, so $Y \subseteq B_0'\cup B_i'$ for some $i$ and as $B_0'\cup B_i'$ is $d$-closed in $B'$, any copies of $Z$ over $Y$ in $B'$ are contained in $B_0'\cup B_i'$. So there are at most $\mu(Y,Z)$ of these as $B_0'\cup B_i' \in \mathcal{C}_\mu(L)$.
Now suppose that $\delta(Y) \geq 2$ and suppose for a contradiction that $Z_i$ (for $i = 1,\ldots, \mu(Y,Z)+1$) are disjoint copies in $B'$ of $Z$ over $Y \subseteq B'$ (meaning that the sets $Z_i \setminus Y$ are disjoint, of course).
If $y \in Y$ then $y$ is in some relation in $R_k^Z\setminus R_k^Y$ (for some $k \in I$) by Lemma \ref{21}. Thus $y$ is in at least three relations in $R_k^{B'}$ (one in each $R_k^{Z_i}\setminus R_k^Y$). By inspection of the construction one therefore sees that $y \in A''$ or $y = b_{ij}$ for some $i > r$. In the latter case, two of the (at most) three relations in $B'$ which involve $b_{ij}$ are $s_{i}\cup \{b_{i(j-1)}, b_{ij}\}$ and $s_{i}\cup \{b_{i(j+1)}, b_{ij}\}$. So we can assume that the first is a subset of $Z_1$ (and not a subset of $Y$) and the second is a subset of $Z_2$. But this implies that $s_{i} \subseteq Y$. However, there is no other relation which contains $\{b_{ij}\} \cup s_{i}$: contradicting the fact that $Z_3$ is a copy of $Z_1$ over $Y$.
Thus $Y \subseteq A''$. As $A'' \in \mathcal{C}_\mu(L)$ not all of the $Z_i$ are subsets of $A''$, so we can assume that $Z_1 \not\subseteq A''$. As $A'' \leq B'$ we have $Y \subseteq A'' \cap Z_1 \leq Z_1$ so (by the simplicity of the extension) $Z_1 \cap A'' = Y$.
Note that $Z_1$ is in the $d$-closure of $Y$ so we cannot have $Y \subseteq B_0'$. Let $y \in Y \setminus B_0'$. This is adjacent in $Z_1$ to some $z \in Z_1\setminus Y$. So $y \in A'' \setminus B_0'$ is adjacent in $B'$ to $z \in B'\setminus A''$. Inspection of the construction shows that $y = b_{i0}$ (for some $i \leq r$) and $z = b_{ij}$. Then the adjacency of $y$ and $z$ in $Z_1$ forces $s_{ij} \subseteq Z_1$, and so $s_{ij} \subseteq Y$ (as $A''\cap Z_1 = Y$). But then $Y \leq Y \cup \{b_{ij}\}$ is a simply algebraic extension in $Z_1$. As $Y \leq Z_1$ is a minimally simply algebraic extension, this implies $Y = \{b_{i0}\}\cup s_{ij}$ and $Z_1 = \{b_{i0}, b_{ij}\}\cup s_{ij}$. However, there is no other relation in $B'$ which contains this $Y$ (by construction), so we have a contradiction. \hfill$\Box_{Claim}$
\medskip
Claims 2 and 3 finish the proof of the isomorphism extension property $\mathcal{C}_0(L) (W)\rightsquigarrow \mathcal{C}_\mu(L)(X)$.
\medskip
For the other direction, we can use the same construction (it is a special case of the the above as $\mathcal{C}_\mu(L) \subseteq \mathcal{C}_0(L)$). Of course, in this case we do not need Claim 3.
\end{proof}
\section{Further isomorphisms}
\subsection{Localization of non-isomorphic geometries}
In this subsection the language $L$ has just a single $3$-ary relation $R$. We often suppress $L$ in the notation.
In 5.2 of \cite{EH} Hrushovski varies his strongly minimal set construction to produce examples where the model-theoretic structure of the strongly minimal set can be read off from the geometry: lines of the geometry have three points, and colinear points correspond to instances of the ternary relation. He thereby produces continuum-many non-isomorphic geometries of (countable, saturated) strongly minimal structures, but asks whether these examples are \textit{locally} isomorphic. We show that this is the case: in fact, localizing any of them over a 3-dimensional set gives a geometry isomorphic to $G(\mathcal{M}_0(L))$, the geometry of the generic structure for $(\mathcal{C}_0(L), \leq)$.
In 5.2 of \cite{EH}, Hrushovski considers
\[\mathcal{K}_0 = \{A \in \mathcal{C}_0(L) : B \leq A \mbox{ for all $B \subseteq A$ with $\vert B \vert \leq 3$}\}.\]
The class $(\mathcal{K}_0, \leq)$ is an amalgamation class: one shows that if $A \leq B_1, B_2 \in \mathcal{K}_0$ and the free amalgam of $B_1$ and $B_2$ over $A$ is not in $\mathcal{K}_0$, then there exist $a, a' \in A$ and $b_i \in B_i\setminus A$ with $R(a,a', b_i)$ holding in $B_i$ (for $i=1,2$).
More generally, given a function $\mu$ as before, we can consider $\mathcal{K}_\mu = \mathcal{K}_0 \cap \mathcal{C}_\mu(L)$ and for appropriate $\mu$, the class $(\mathcal{K}_\mu, \leq)$ will satisfy Assumption \ref{aalemma}. In fact, we only need to define $\mu(Y,Z)$ for $\delta(Y) \geq 3$. For suppose $Y \leq Z$ is a minimally simply algebraic extension in $\mathcal{K}_0$ and $\delta(Y) = \delta(Z) \leq 2$. Then $Z$ has at most 3 elements: otherwise there is a subset $W\subseteq Z$ of size 3 with $W \not\in R^Z$, and then $W$ is not self-sufficient in $Z$, contradicting the definition of $\mathcal{K}_0$. It follows that the value of $\mu(Y,Z)$ is irrelevant for such $Y \leq Z$: the multiplicity is already controlled by the definition of $\mathcal{K}_0$.
So we shall assume:
\begin{ass}\label{HAmalgLemma} With the above notation:
\begin{enumerate}
\item[(i)] If $A \leq B_1, B_2 \in \mathcal{K}_\mu$ and the free amalgam of $B_1$ and $B_2$ over $A$ is not in $\mathcal{K}_\mu$, then either there exist $a, a' \in A$ and $b_i \in B_i\setminus A$ with $R(a,a', b_i)$ holding in $B_i$ (for $i=1,2$), or there exists $Y \subseteq A$ with $\delta(Y) \geq 3$ and msa extensions $Y \leq Z_i \in B_i$ (for $i= 1,2$) which are isomorphic over $Y$ and $Z_i\setminus Y \subseteq B_i\setminus A$.
\item[(ii)] The class $(\mathcal{K}_\mu, \leq)$ is an amalgamation class.
\end{enumerate}
\end{ass}
Again, (ii) follows from (i) here and the condition $\mu(Y,Z) \geq \delta(Y)$ (for $\delta(Y) \geq 3$) guarantees that (i) holds.
Denote the generic structure of $(\mathcal{K}_\mu, \leq)$ by $\mathcal{N}_\mu$. The $d$-closure of two points in $\mathcal{N}_\mu$ has size $3$ (as above), so certainly $G(\mathcal{N}_\mu)$ and $G(\mathcal{M}_0(L))$ are non-isomorphic. In fact, we can recover the relation $R$ from the geometry $G(\mathcal{N}_\mu)$ as the $3$-sets with dimension 2. Thus different $\mu$ give different geometries. It can be shown that there are infinitely many msa extensions $Y \leq Z \in \mathcal{K}_0$ with $\delta(Y) = 3$, so if $\mu(Y, Z) \geq 1$ for infinitely many of these, then ${\rm cl}_{\mathcal{N}_\mu}(X)$ is infinite whenever $X \leq \mathcal{N}_\mu$ consists of 3 independent points.
We show:
\begin{thm}\label{surprise}
Suppose $\mu(Y,Z) \geq 3$ for all msa $Y \leq Z$ in $\mathcal{K}_0$ with $\delta(Y) \geq 3$ and Assumption \ref{HAmalgLemma} holds. Let $X \leq \mathcal{N}_\mu$ and $d(X) = 3$. Then the localization $G_X(\mathcal{N}_\mu)$ is isomorphic to $G(\mathcal{M}_0(L))$.
\end{thm}
\begin{proof}
We may assume that $X$ consists of 3 points (and no relations). By the remarks preceding the theorem, $X$ has arbitrarily large finite algebraic extensions in $\mathcal{K}_\mu$. We show that the isomorphism extension property $\mathcal{C}_0(L) \rightsquigarrow \mathcal{K}_\mu(X)$ holds.
Suppose we are given $A \leq B \in \mathcal{C}_0(L)$ and $X \leq A' \in \mathcal{K}_\mu$ and an isomorphism $f : G(A) \to G_X(A')$. We want to find $B' \in \mathcal{K}_\mu$ with $A' \leq B'$ and an isomorphism $f' : G(B) \to G_X(B')$ extending $f$. This is very similar to the the construction of $B'$ in the proof of Theorem \ref{main} and we will only indicate what needs to be modified and provide extra argument as required.
Let $A_0 = {\rm cl}_A(\emptyset)$ and let $A_1,\ldots, A_r$ be the $d$-dependence classes on $A\setminus A_0$: the latter are the points of $G(A)$. Similarly let $B_0 = {\rm cl}_B(\emptyset)$ and $B_1, \ldots, B_s$ the $d$-dependence classes on $B\setminus B_0$, with $A_i \subseteq B_i$ for $i = 1,\ldots, r$. List the relations on $B$ which are not contained in $A$ or some $B_0\cup B_j$ as $\rho_1,\ldots, \rho_t$. So these are 3-sets and note that each of them intersects three different $B_i$. Let $A_0' = {\rm cl}_{A'}(X)$ and $A_1',\ldots, A_r'$ be the classes of $d$-dependence over $X$ on $A'\setminus A_0'$, labelled so that $f(A_i) = A_i'$. We construct $B_0', B_1', \ldots, B_s'$ with $A_i' \subseteq B_i'$ for $i = 0, \ldots, r$, and $B' = \bigcup_{i = 0}^s B_i'$ in the following way.
\medskip
\noindent\textit{Step 1:\/} Construction of $A'' = A' \cup B_0' \in \mathcal{K}_\mu$. \newline
This is as before, but we need to take $V \in \mathcal{K}_\mu$: we can do this because algebraic extensions of $X$ can be arbitrarily large.
\medskip
\noindent\textit{Step 2:\/} Construction of $B_0' \cup B_i' \in \mathcal{K}_\mu$.\newline
The construction is as in Theorem \ref{main} for $i \leq r$. In the case $i > r$ we vary the construction by taking the $s_{ij}$ to be distinct. The proofs that $B_0'$ is $d$-closed in $B_0'\cup B_i'$ are as before; as are the arguments which show that if $Y \leq Z$ is a msa extension in $\mathcal{K}_0$ with $\delta(Y) \geq 3$ then there are at most $\mu(Y,Z)$ copies of $Z$ over $Y$ in $B_0'\cup B_i'$. So it remains to show that $B_0' \cup B_i' \in \mathcal{K}_0$.
If $i \leq r$, then using the amalgmation lemma \ref{HAmalgLemma} as in Step 2, Case 1 of Theorem \ref{main}, it will suffice to show that $B_0'\cup A_i' \cup \{b_{ij}\} \in \mathcal{K}_0$. This is the free amalgam of $\{s_{ij}, b_{i0}, b_{ij}\}$ and $B_0'\cup A_i'$ over $\{s_{ij},b_{i0}\}$. So we can apply \ref{HAmalgLemma} (because $\{s_{ij},b_{i0}\}$ is in no relation in $B_0'\cup A_i'$).
Now suppose $i > r$. We analyse the possibilities for $\delta(Y)$ when $Y \subseteq B_0'\cup B_i'$, $Y \not\subseteq B_0'$ and $\vert Y \vert > 1$. As $Y \cap B_0'$ is $d$-closed in $Y$ we have $\delta(Y) > \delta(Y \cap B_0')$. If $Y \cap B_0' = \emptyset$ then by the construction, $\delta(Y) = \vert Y \vert$. If $Y \cap B_0'$ is a singleton then $Y$ has at most one relation (because the $s_{ij}$ are distinct), so $\delta(Y) > 1$ and $\delta(Y) \geq \vert Y \vert -1$. In the remaining case, $\delta(Y \cap B_0') \geq 2$ (as $B_0' \in \mathcal{K}_0$), so $\delta(Y) \geq 3$. Thus, $Y$ consists of 2 points, or is 3 points in a relation, or has $\delta(Y) \geq 3$. It follows that $B_0'\cup B_i' \in \mathcal{K}_0$.
\medskip
\noindent\textit{Step 3:\/} Other relations on $B'$. \newline
As before.
\medskip
\noindent\textit{Claim 1:\/} Let $U \subseteq \{1,\ldots, s\}$, and $Y = \bigcup_{i\in U} (B_i\cup B_0)$, $Y' = \bigcup_{i\in U} (B_i' \cup B_0')$. Then $Y \cap A$ is $d$-closed in $A$ iff $Y' \cap A'$ is $d$-closed in $A'$, and in this case we have $\delta(Y) = \delta(Y'/B_0').$
As before.
\medskip
\noindent\textit{Claim 2:\/}
$B' \in \mathcal{C}_0(L)$, $B_0'$ and $B_0'\cup B_i'$ are $d$-closed in $B'$ and $A'' \leq B'$. The map $f' : G(B) \to G_X(B')$ given by $f'(B_i) = B_i'$ is an isomorphism of geometries which extends $f$.
As before, using Claim 1.
\medskip
\noindent\textit{Claim 3:\/} If $i\neq j$ then $B_0'\cup B_i' \cup B_j' \in \mathcal{K}_\mu$ and $B_0' \cup B_i'\cup B_j' \leq B'$.
\smallskip
By construction $B_0'\cup B_i' \cup B_j'$ is the free amalgam of $B_0'\cup B_i'$ and $B_0' \cup B_j'$ over $B_0'$, and $B_0'$ is $d$-closed in each. So the first statement follows from the assumed amalgamation lemma \ref{HAmalgLemma}. We have $\delta(B_0' \cup B_i'\cup B_j'/ B_0') = 2$. Moreover, as $B_0'\cup B_i'$ is $d$-closed in $B'$ (Claim 2), if $B_0' \cup B_i'\cup B_j'\subseteq Z$ then $\delta(Z/B_0') \geq 2$. This gives the second statement.\hfill$\Box_{Claim}$
\medskip
\noindent\textit{Claim 4:\/} $B' \in \mathcal{K}_0$.
\smallskip
We need to show that if $D \subseteq B'$ has size at most 3 then $D \leq B'$. If $\vert D \vert \leq 2$ then $D \subseteq B_0' \cup B_i' \cup B_j'$ for some $i, j$ and it follows from Claim 3 that $D \leq B_0'\cup B_i'\cup B_j' \leq B'$. So suppose $D$ has size 3 and $D \subseteq C$ with $\delta(C) < \delta(D)$. We must have $\delta(C) = 2$ (as any two points of $D$ are self-sufficient in $C$ and have $\delta$-value $2$). As $A'' \in \mathcal{K}_0$ there is an $i$ such that $C \cap (B_i'\setminus A'') \neq \emptyset$. Note that $C$ is not contained in $B_0'\cup B_i'$ (because this is in $\mathcal{K}_\mu$), so as $B_0'$ is $d$-closed in $B_0' \cup B_i'$ and the latter is $d$-closed in $B'$, we have
\[0 \leq \delta(C\cap B_0') < \delta(C \cap (B_0'\cup B_i')) < \delta(C) = 2.\]
Thus $\delta(C\cap B_0') = 0$, so $C \cap B_0' = \emptyset$.
It then follows from Step 2 of the construction that there is no adjacency in $C$ between points of $C\cap A''$ and points of $C \setminus A''$. Let $q = \vert\{j : \rho_j' \subseteq C\}\vert$. Then (using $A'' \leq B'$; so $C \cap A'' \leq C$)
\[2 \geq \delta(C/C\cap A'') = \vert C \setminus A''\vert - q \geq 2q\]
as the $\rho_j'$ are disjoint. If $q = 0$ then $C$ is $C\cap A''$ together with some isolated points, and this is in $\mathcal{K}_\mu$ (so not possible in this situation). If $q = 1$ then $C$ consists of 3 points in a single relation and this has no subset of the form required for $D$. \hfill$\Box_{Claim}$
\medskip
\noindent\textit{Claim 5:\/} $B' \in \mathcal{K}_\mu$\newline
We already know that $B' \in \mathcal{K}_0$, so we need to show that $B' \in \mathcal{C}_\mu(L)$, at least as far as msa extensions $Y \leq Z$ with $\delta(Y) \geq 3$ are concerned. So suppose $Z_1,\ldots, Z_4$ are disjoint copies of $Z$ over $Y$ in $B'$. Then each element $y \in Y$ is in at least $4$ relations (using \ref{21}, as before), so by construction, $y \in A''$. Thus $Y \subseteq A''$ and the rest of the proof is just as in Claim 3 of Theorem \ref{main}. \hfill$\Box_{Claim}$
\medskip
Claims 2 and 5 finish the proof of one direction of the isomorphism extension property. The direction $\mathcal{K}_\mu(X) \rightsquigarrow \mathcal{C}_0(L)$ follows from the property $\mathcal{C}_\mu(L)(X) \rightsquigarrow \mathcal{C}_0(L)$ proved in Theorem \ref{main}.
\end{proof}
\begin{rem}\rm Note that $\mathcal{K}_0$ can be considered as $\mathcal{K}_\mu$ where $\mu(Y,Z)$ is formally given the value $\infty$ for all msa $Y \leq Z \in \mathcal{K}_0$. Thus the above argument also shows that the geometry of $\mathcal{N}_0$, the generic for $(\mathcal{K}_0, \leq)$, is locally isomorphic to $G(\mathcal{M}_0(L))$.
\end{rem}
\begin{rem} \rm Another variation is given in 5.1 of \cite{EH}. Let $k \geq 2$ and consider the language $L$ with a single $(k+1)$-ary relation symbol $R$. Let
\[\mathcal{C}_0'(L) = \{A \in \mathcal{C}_0(L): \delta(B) \geq\min(\vert B\vert , k)\,\, \forall B \subseteq A\}.\]
So if $C \subseteq A \in \mathcal{C}_0'(L)$ and $\vert C \vert \leq k$ then $C \leq A$. Hrushovski observes that $(\mathcal{C}_0'(L), \leq)$ is a free amalgamation class and that the assumed amalgamation lemma \ref{aalemma} holds for $(\mathcal{C}_\mu', \leq)$, for suitable $\mu \geq 2$. The generic structures here are strongly minimal and any $k$ points are independent. So the geometries are again different from that of $\mathcal{M}_0(L)$. However, they are again locally isomorphic. To see this we proceed as in Theorem \ref{surprise}, but take $X$ to be a set of size $k$. The construction and proof are essentially the same as before, except for in Claim 4 where to show that $B' \in \mathcal{C}_0'(L)$ we modify the argument as follows.
Suppose $C \subseteq B'$ has $\delta(C) < k$ and $\vert C \vert \geq k+1$. Then for some $i$ we have:
\[0 \leq \delta(C\cap B_0') < \delta(C\cap (B_0' \cup B_i')) <\delta(C) \leq k-1.\]
So $\delta(C\cap B_0') \leq k-3$ and therefore $\vert C \cap B_0'\vert \leq k-3$. Then by construction there is no adjacency in $C$ between points of $C \cap A''$ and points of $C\setminus A''$. So (with $q$ as before):
\[k-1 \geq \delta(C/C\cap A'') = \vert C \setminus A''\vert - q\geq kq.\]
Thus $q = 0$ and we have a contradiction.
\end{rem}
\subsection{Changing the language and predimension}
Recall that the language $L$ consists of relation symbols $\{R_i: i \in I\}$ with $R_i$ of arity $n_i$ (and only finitely many symbols of each arity). Suppose that $L_0 = \{ R_i : i \in I_0\}$ is a sublanguage with the property that for every $i \in I$ there is $j \in I_0$ such that $n_i \leq n_j$. For example, if $I$ is finite we can take $L_0$ to consist of a relation symbol of maximal arity in $L$. The following is essentially Theorem 3.1 of \cite{MFDE1}, but working with sets rather than tuples: we omit most of the details of the proof.
\begin{thm}
The geometries $G(\mathcal{M}_0(L))$ and $G(\mathcal{M}_0(L_0))$ are isomorphic.
\end{thm}
\begin{proof}
We can use the construction in Theorem \ref{main} to show that $\mathcal{C}_0(L) \rightsquigarrow \mathcal{C}_0(L_0)$ holds. In step 3 of the construction, if $\rho_i$ is a $k$-set then we take $\rho_i'$ to be a $k'$-set with $k' \geq k$: the condition on $L_0$ allows us to do this. Claims 1 and 2 of Theorem \ref{main} then go through exactly as previously. The direction $\mathcal{C}_0(L_0) \rightsquigarrow \mathcal{C}_0(L)$ follows as in Theorem \ref{main}.
\end{proof}
\begin{rem}\rm
Theorem 3.1 of \cite{MFDE1} works with a predimension of the form:
\[\delta_\alpha(A) = \vert A \vert - \sum_{i \in I} \alpha_i\vert R_i^A\vert,\]
where the $\alpha_i$ are natural numbers. We can adapt the argument here to deal with such predimensions. For example, suppose $I$ is finite and $R_1$ is of maximal arity and $\alpha_1 = 1$. Let $L_0$ consist of $R_1$. Then, as in Theorem 3.1 of \cite{MFDE1}, $G(\mathcal{M}_0(L))$ is isomorphic to $G(\mathcal{M}_0(L_0))$. To show that $\mathcal{C}_0^\alpha(L) \rightsquigarrow \mathcal{C}_0(L_0)$ (where $\mathcal{C}_0^\alpha(L)$ is defined using the predimension $\delta_\alpha$) we perform the same construction except that in step 3, if $\rho_j$ is of type $R_i$ then we add $\alpha_i$ corresponding $\rho_j'$ (but still disjoint etc).
\end{rem}
\subsection{Sets versus tuples} We have chosen to work with structures $A$ where the relations $R_i^A$ are sets of $n_i$-sets. As was done in \cite{MFDE1} we could also have worked more generally with structures $A$ where the $R_i^A$ are sets of $n_i$-tuples and the predimension is still given by $\vert A \vert - \sum _i \vert R_i^A\vert$. Let $\hat{\mathcal{C}}_0(L)$ denote the class of these finite structures with $\emptyset \leq A$.
\begin{thm}
The geometries of the generic structures of the amalgamation classes $(\mathcal{C}_0, \leq)$ and $(\hat{\mathcal{C}}_0(L), \leq)$ are isomorphic.
\end{thm}
\begin{proof}
This is the usual sort of proof using the construction. For example, to show $\hat{\mathcal{C}}_0(L) \rightsquigarrow \mathcal{C}_0(L)$ we replace an $n_i$-tuple $\rho_j$ (in $R_i^B\setminus R_i^{A}$) by an $n_i$-set, using the new $d$-dependent points to eliminate repetitions of points in the tuple or different enumerations of the same set.
\end{proof}
\medskip
|
\section{\label{sec:level1}First-level heading:\protect\\ The line
\section{introduction}
Quantum computation (QC) is one of the most fascinating and fruitful area of research. It strives to utilize the principles of
quantum mechanics to improve the efficiency of computation. In recent years, with the discoveries of many quantum algorithms \cite{Ref1,PhysRevLett.79.325}, QC has been shown to work faster than any known classic computation. Though originally based upon discrete variables (DV), QC over continuous variables (CV) has also been addressed by Lloyd and Braunstein \cite{PhysRevLett.82.1784}.
In analogy to DV QC based on gate operations, universal CV QC can be carried out by executing a finite set of CV quantum logic gates, including displacement gate, shearing gate, cubic phase gate, and controlled-$Z$ ($C_Z$) gate \cite{PhysRevA.79.062318}.
CV $C_Z$ gate [also called as quantum nondemolition (QND) gate] is a canonical two-mode gate, which is a CV analog of a two-qubit CPHASE gate. Like the DV CPHASE gate, CV $C_Z$ gate is considered as one of the most fundamental CV
quantum gates.
Nowadays, much effort has been devoted
to the realization of such gate in many physical systems, especially in optical systems \cite{NATRUE10,PhysRevLett.57.2473,OC1,PhysRevA.42.2995,PhysRevLett.66.1418,PhysRevLett.62.28,PhysRevA.71.055801,PhysRevA.71.042308,PhysRevLett.101.250501}. However, by now optical CV CZ gate is still experimentally challenging.
Initial approaches to realize optical QND interactions are based on nonlinear optical systems \cite{PhysRevLett.57.2473,OC1,PhysRevA.42.2995,PhysRevLett.66.1418,PhysRevLett.62.28}.
$C_Z$ gate is performed by simply sending the beams through an optical crystal or an optical fiber.
These approaches, however, are hampered by the weak nonlinearities in the nonlinear media. Although the nonlinearities can be enhanced by embedding the crystal inside a cavity or using a long fiber, such enhancement techniques, on the other hand, make it difficult to inject the light states into the system and cause large loss rates. Recently, an alternative approach was proposed to circumvent cumbersome nonlinear interactions by using only linear optical elements and off-line squeezed light beams \cite{PhysRevA.71.042308}. This approach, however, requires efficient homodyne detection techniques and accurate feed forward control, which poses a challenge to the experimental realization \cite{PhysRevLett.101.250501}.
In recent years, with the emergence of one-way QC \cite{PhysRevLett.97.110501}, the interest in theoretical and experimental realization of optical $C_Z$ gate is further fuelled.
One-way QC is a new form of QC which eliminates unitary evolution and relies solely on adaptive measurements on a suitably prepared multiparty entangled state. This model is quite attractive because local projective measurements are often easier to implement than unitary evolution. Most of the challenging work of QC are then converted into the problem of creating the multiparty entangled state---the so-called \emph{cluster state}.
Cluster state in the CV regime is a multimode squeezed Gaussian state and has been proved universal resources for CV one-way QC. To date, several methods have been proposed to construct such state. One of them is called the \emph{canonical method} \cite{PhysRevA.73.032318}, which involves offline single-mode squeezers and $C_Z$ gates. Although this method is conceptually simple, it is not very practical because of the experimental challenges associated with $C_Z$ gates. Shortly after this method, another method, \emph{linear-optics method} \cite{PhysRevA.76.032321}, has also been proposed to eliminate the need for $C_Z$ gates. Any desired CV cluster state can be created by using only offline-squeezing and linear optics. This method, however, suffers from drawback that effects its scalability. Recently, Menicucci \emph{et al.} proposed the \emph{single-QND-gate method} \cite{PhysRevLett.104.250503} which reintroduces the $C_Z$ gates. Although this method has many distinct advantages, i.e. it saves the resources needed greatly and eliminates the need for long-time coherence of a large cluster state, still the inefficiency of the experimental realization of $C_Z$ gates casts it into the shade. Apparently, the only way to repolish the $C_Z$-based methods (and thus the CV one-way QC) is to devise many new schemes which enable us to implement the $C_Z$ gate more efficiently.
In this paper, we propose a simple and practical scheme to realize optical $C_Z$ gate using a free-space macroscopic atomic ensemble. We show that $C_Z$ gate between two optical beams can be performed by simply sending them perpendicularly to each other twice through a spin-squeezed atomic ensemble, see Fig. 1(a). The fidelities obtained in the scheme depend solely on the degree of the atomic squeezing. The more the amount of squeezing input, the higher the fidelity obtain.
Near-unity fidelity can be achieved under the condition that the atomic state is infinitely squeezed. Unlike previous measurement-based schemes \cite{PhysRevLett.101.250501,PhysRevA.71.042308,*PhysRevA.80.050303}, our scheme requires neither homodyne detection nor feedforward control during the gate operation, which greatly simplifies its experimental implementation. Within the presently experimentally available parameters, we find that the observed fidelities are quite high even with room-temperature atomic vapors.
\makeatletter
\newcommand{\rmnum}[1]{\romannumeral #1}
\newcommand{\Rmnum}[1]{\expandafter\@slowromancap\romannumeral #1@}
\makeatother
The remainder of this paper is organized as follows. In sec. \Rmnum{2}
we first review some basic theories, and then give details of $C_Z$ operation based on an atomic ensemble.
In sec. \Rmnum{3} we will consider the noise effects. After that, the experimental feasibility of the scheme is also discussed.
Finally, sec. \Rmnum{4} contains brief conclusions.
\begin{figure}[tp]
\centering
\includegraphics[scale=0.8]{fig_1.eps}
\caption{(Color online) (a) Scheme setup for realization of optical $C_Z$ gate in an atomic ensemble. Two light beams $L$ and $M$ simultaneously enter a spin-squeezed atomic ensembles. The outgoing beams then pass through a delay line. After the whole of the beams run through the atomic sample, they enter the ensemble again. When the beams exit, a $C_Z$ gate between $L$ and $M$ is performed. (b) Schematic chart of the $C_Z$ gate. The two interactions do not overlap in the time line. After them, the gate is performed.}
\end{figure}
\section{BASIC THEORY AND CONTROLLED-Z Gate}
\subsection{Basic theory}
For a two-mode two-party system described by the quadrature operators $\hat x_{1,2}$ and $\hat p_{1,2}$ satisfying the commutation relation $[\hat x_j,\hat p_k]=i\delta_{jk}$, the QND-gate coupling Hamiltonian can be written as: $\hat H_{QND}=\hbar \chi
\hat p_1\hat p_2$, where $\chi$ is the coupling coefficient. In the Heisenberg picture, one may deduce the ideal QND input-output relations for both position and momentum operators
\begin{eqnarray}
\hat x_1^{out} = \hat x_1^{in} + G\hat p_2^{in},~~~\hat p_1^{out} = \hat p_1^{in}, \hfill \nonumber\\
\hat x_2^{out} = \hat x_2^{in} + G\hat p_1^{in},~~~\hat p_2^{out} = \hat p_2^{in}, \label{equ:0}
\end{eqnarray}
where $G=\chi t$ is the gain of the interaction, and $t$ represents the interaction time. For nonzero $G$, these equations imply that the two sub-systems become Gaussian entangled \cite{RevModPhys.77.513}. Such entanglement has been widely used in quantum information processing \cite{PhysRevLett.85.5639,PhysRevLett.85.5643,PhysRevLett.91.060401}. Specifically, if we put the gain of the interaction $G=1$, then we obtain the CV $C_Z$ gate as desired.
To realize optical $C_Z$ gate in an atomic ensemble, let us first investigate the interaction between light and atoms.
Consider a cell filled with a large number of atoms interacting with a light pulse traveling along the $z$ direction. The atoms in the cell are initially prepared in a coherent spin state, i.e. a fully polarized state along the $x$ axis. As a result, we may treat the $x$ component of the collective spin as a $c$ number, that is $\hat J_x$ by $J_x$. In this case, we can map the transverse spin components into dimensionless canonical variables $(\hat x_a,\hat p_a)=(\hat J_y,\hat J_z)/\sqrt{J_x}$ obeying $[\hat x_a,\hat p_a]=i$. The light pulse interacting with atoms is also linearly polarized along the $x$ axis. Similarly, we may define the optical canonical operators as $(\hat x_{ph}(t),\hat p_{ph}(t))=(\hat S_y(t),\hat S_z(t))/\sqrt{S_x}$, which satisfy the commutation relation $[\hat x_{ph}(t),\hat p_{ph}(t)]=i\delta(t-t')$ and have the dimensions of $1/\sqrt{time}$, where $\hat S_i$ (with $i\in\{x,y,z\}$) denotes the time dependent Stokes vector component. Under the condition that the frequency of the beam was tuned far off resonance with atomic transition \cite{EPL1}, the interaction of light with atoms can be described by the effective Hamiltonian $\hat H=\tilde\kappa_0\hat p_{ph}\hat p_{a}$, with $\tilde\kappa_0=a\sqrt{J_xS_x}$, where $a$ is the effective coupling strength \cite{PhysRevLett.85.5639,EPL1}. Obviously, it is a QND-type. An important, immediate application of this Hamiltonian is quantum memory \cite{NATRUE3}. However, initially, such interactions were extensively investigated to produce spin squeezing \cite{EPL1,PhysRevA.60.4974,*PhysRevA.60.2346}.
We here briefly review the process of spin squeezing based on QND detection. Following Eq. (\ref{equ:0}) the input-output relations for light and atoms can be directly derived as: $\hat x_{ph}^{out} = \hat x_{ph}^{in}+ \kappa_0\hat p_a^{in} ,\hat p_{ph}^{out} = \hat p_{ph}^{in},\hat x_a^{out} = \hat x_a^{in}+ \kappa_0\hat p_{ph}^{in} ,\hat p_a^{out} = \hat p_a^{in} $, with $\kappa_0=\tilde{\kappa}_0\sqrt{T}$, where $T$ is the duration of the pulse. Next, a measurement of $\hat x_{ph}^{out}$ is performed, giving a random measurement outcome $\xi$. The momentum operator $\hat p_a^{out}$ is then displaced by an amount $g\xi$, where $g$ is a gain factor, resulting in $
\hat p_a^{out} = \hat p_a^{in} - gx_{ph}^{out} = ( {1 - g\kappa_0 } )\hat p_a^{in} + g\hat x_{ph}^{in}$. If the light pulse is initially in coherent state such that $\Delta {x_{ph}^{in}}^2=\Delta {p_{ph}^{in}}^2=1/2$, the variance of $\hat p_a^{out}$ can be easily calculated, giving
$(\Delta {\hat p_a^{out} })^2 = \frac{1}
{2}[ {( {1 - g\kappa_0 })^2 + g^2 }].$
Optimizing it, we get $(\Delta {\hat p_a^{out} })^2 =\frac{1}{2}\frac{1}{1+\kappa_0^2}$ for $g=\kappa_0/{(1+\kappa_0^2)}$. Obviously, for nonzero $\kappa_0$, the atomic momentum operator is then squeezed. Finally, we obtain the squeezed spin state (SSS) as:
\begin{eqnarray}
\hat x_a^{out}=\sqrt{1+\kappa_0^2}\hat x_a^{in},~~~\hat p_a^{out}=\frac{1}{\sqrt{1+\kappa_0^2}}\hat p_a^{in}.\label{equ:0-1}
\end{eqnarray}
\subsection{Controlled-Z gate}
Let us now consider the implementation of optical $C_Z$ gate in an atomic ensemble. Suppose that we have an atomic ensemble prepared in the SSS described above, through which two $x$-polarized light beams are transmitted simultaneously from two perpendicular directions, see Fig. 1(a). For the beam $L$ propagating along the $z$ direction, its interaction with atoms can then be described by $\hat H_L=\tilde{\kappa}_L\hat p_L\hat p_a$. The second beam $M$ propagates along the $y$ direction, leading to the second Hamiltonian $\hat H_M=\tilde\kappa_M\hat p_M\hat x_a$ \cite{PhysRevA.73.022331,PhysRevA.73.062329}. Thus, the complete Hamiltonian for this process can be written as
\begin{equation}
\hat H_1=\tilde\kappa\hat p_L\hat p_a+\tilde\kappa\hat p_M\hat x_a,\label{equ:1}
\end{equation}
where we have assumed $\tilde\kappa_L=\tilde\kappa_M=\tilde\kappa$.
Corresponding to this
Hamiltonian, one may straightly derive the Heisenberg equations for atoms and the Maxwell-Bloch equations (neglecting the effects of light retardation) for light as \cite{PhysRevA.73.022331,PhysRevA.73.062329}:
\begin{equation}
\partial_t\hat x_a \left( t \right) =\tilde\kappa\hat p_L^{in} \left( t \right),\label{equ:2a}
\end{equation}
\begin{equation}
\partial_t\hat p_a \left( t \right) = - \tilde\kappa\hat p_M^{in} \left( t \right),\label{equ:2b}
\end{equation}
\begin{equation}
\hat x_L^{out1} \left( t \right) = \hat x_L^{in} \left( t \right) + \tilde\kappa \hat p_a \left( t \right),\label{equ:3a}
\end{equation}
\begin{equation}
\hat p_L^{out1} \left( t \right) = \hat p_L^{in} \left( t \right),\label{equ:3b}
\end{equation}
\begin{equation}
\hat x_M^{out1} \left( t \right) =\hat x_M^{in} \left( t \right) + \tilde\kappa \hat x_a \left( t \right),\label{equ:3c}
\end{equation}
\begin{equation}
\hat p_M^{out1} \left( t \right) = \hat p_M^{in} \left( t \right),\label{equ:3d}
\end{equation}
where $\partial_t$ stands for the partial derivative with respect to $t$. Equations (\ref{equ:2a}) and (\ref{equ:2b}) can be readily solved by integrating over $t$ on both sides, giving
\begin{eqnarray}
\hat x_a \left( t \right) = \hat x_a \left( 0 \right) + \tilde\kappa\int_0^t {\hat p_L^{in} \left( \tau \right)} d\tau , \nonumber\\
\hat p_a \left( t \right) = \hat p_a \left( 0 \right) - \tilde\kappa \int_0^t {\hat p_M^{in} \left( \tau \right)d\tau } . \label{equ:4}
\end{eqnarray}
Inserting this set of equations into Eqs. (\ref{equ:3a}) and (\ref{equ:3c}), one will obtain the expressions for $\hat x_{L,M}^{out1}$,
\begin{eqnarray}
\hat x_L^{out1} \left( t \right) &=& \hat x_L^{in} \left( t \right) +\tilde\kappa \hat p_a \left( 0 \right) \nonumber- \tilde\kappa^2\int_0^t {\hat p_M^{in} \left( \tau \right)d\tau }, \nonumber\\
\hat x_M^{out1} \left( t \right) &=&\hat x_M^{in} \left( t \right) + \tilde\kappa \hat x_a \left( 0 \right)+\tilde\kappa^2\int_0^t {\hat p_L^{in} \left( \tau \right)} d\tau.\label{equ:5}
\end{eqnarray}
Next, we define the dimensionless collective light modes $\hat x_{j}=\int_0^T{f(t)\hat x_{j}(t)dt}$ ($j\in\{L,M\}$) with $[\hat x_{j},\hat p_{j}]=i$, where $f(t)$ is a temporal mode function specifying the mode in question. For the symmetric mode with $f(t)=1/\sqrt{T}$, Eqs. (\ref{equ:5}) are then changed into
\begin{eqnarray}
\hat {x}_L^{out1} &=& \hat {x}_L^{in} + \kappa\hat p_a\left(0\right)- \frac{\kappa^2}{T^{3/2}}\int_0^T {\left(T-t\right)\hat p_M^{in} \left( t \right)dt}, \nonumber\\
\hat {x}_M^{out1} &=& \hat {x}_M^{in} + \kappa\hat x_a\left(0\right)+ \frac{\kappa^2}{T^{3/2}}\int_0^T {\left(T-t\right)\hat p_L^{in}\left(t\right)dt},\nonumber\\
\label{equ:6}
\end{eqnarray}
where we have interchanged the order of the double integral and defined the dimensionless coupling strength $\kappa=\tilde\kappa\sqrt{T}$. Eqs. (\ref{equ:6}) indicate that, besides the momentum operators $\hat p_M^{in}$ and $\hat p_L^{in}$, the position operators $\hat x_L^{out1}$ and $\hat x_M^{out1}$ also pick up the information of the input atomic operators $\hat p_a(0)$ and $\hat x_a(0)$. Such information bring unfavorable influence on our
scheme, since, for an ideal optical $C_Z$ gate, only the information of $\hat p_M^{in}$ and $\hat p_L^{in}$ are allowed to be admixed into $\hat x_L^{out1}$ and $\hat x_M^{out1}$, respectively.
In the next step, to eliminate the influence of the atomic operators contained in Eqs. (\ref{equ:6}), we propose reflecting the two beams back into the cell after they have completely passed through the atomic ensemble, see Fig. 1(a). Before the second interaction, two delay lines are used to make sure that the two interactions do not overlap in time line, as shown in Fig. 2(b). For the second transmission, beam $M$ still propagates along $y$, leading to the same Hamiltonian as $\hat H_M$, while beam $L$ runs along $-z$ and sees therefore $-\hat x_A$. As a result, its interaction with atoms is changed into $-\hat H_L$. Consequently, the second complete Hamiltonian reads as
\begin{equation}
\hat H_2=-\tilde\kappa\hat p_L\hat p_a+\tilde\kappa\hat p_M\hat x_a.\label{equ:6-2}
\end{equation}
From this Hamiltonian, one can easily derive the evolution equations for both light and atoms along the same line outlined above. Using the output
state of the first interaction as the input for the second interaction, we can get the final input-output relations for light,
\begin{eqnarray}
\hat {x}_L^{out2} &=& \hat {x}_L^{out1} - \kappa\hat p_a\left(T\right)+\frac{\kappa^2}{T^{3/2}}\int_0^T {\left(T-t\right)\hat p_M \left( t \right)dt}, \nonumber\\
\hat {x}_M^{out2} &=& \hat {x}_M^{out1} + \kappa\hat x_a\left(T\right)- \frac{\kappa^2}{T^{3/2}}\int_0^T {\left(T-t\right)\hat p_L\left(t\right)dt},\nonumber\\\label{equ:7}
\end{eqnarray}
Inserting Eqs. (\ref{equ:4}) and (\ref{equ:6}) into Eq. (\ref{equ:7}), we finally obtain
\begin{equation}
\hat {x}_L^{out2} = \hat {x}_L^{in} +\kappa^2\hat p_M^{in},\label{equ:8a}
\end{equation}
\begin{equation}
\hat {p}_L^{out2} = \hat {p}_L^{in},\label{equ:8b}
\end{equation}
\begin{equation}
\hat {x}_M^{out2} = \hat {x}_M^{in}+\kappa^2\hat p_L^{in} +\frac{2\kappa}{\sqrt{1+\kappa^2_0}}\hat x_a^{in},\label{equ:8c}
\end{equation}
\begin{equation}
\hat {p}_M^{out2} = \hat {p}_M^{in},\label{equ:8d}
\end{equation}
where we have substituted the squeezed input atomic state (\ref{equ:0-1}) into Eq. (\ref{equ:8c}) and have assumed that the position quadrature (but not the momentum quadrature) is initially squeezed.
From above one can clearly see that, if the input atomic state is infinitely squeezed (such that $\kappa_0\rightarrow \infty$), then we can neglect the last two terms of Eq. (\ref{equ:8c}). In this case, if $\kappa$ is equal to $1$, then an ideal $C_Z$ gate between light $L$ and $M$ is performed.
However, in reality, the strength of coupling of light to atoms is always finite. As a result, the noise terms contained in Eq. (\ref{equ:8c}) can not be completely eliminated. In this case, we need to quantify the performance of the $C_Z$ operation. Often, it can be done by calculating the fidelity $F=\langle\psi|\hat\rho_{out}|\psi\rangle$, which is a measure of how well the output
state $\hat\rho_{out}$ compare to the original input state $|\psi\rangle$. Besides, we note that both the atomic state and the light states involved here are all Gaussian. For an $N$-modes Gaussian state, the Wigner function can be conveniently expressed in the form \cite{RevModPhys.77.513,arxiv1}: $W=1/(\pi^N\sqrt{\text{det}\gamma})\text{exp}[-(\xi-\textbf{m})\gamma^{-1}(\xi-\textbf{m})^T]$, where $\gamma$ stands for the covariance matrix, and $\xi$ denotes the $N$-dimensional vector having the quadrature
pairs of all $N$ modes as its components, while $\textbf{m}$ represents the mean values. Mathematically, the fidelity $F$
can be calculated by the overlap of the pure input state $W_{in}$ with the mixed output state $W_{out}$
\begin{eqnarray}
F &=&{\left(2\pi\right)}^N\int_{ - \infty }^{ + \infty } {dx} \int_{ - \infty }^{ + \infty } {dp} {W_{in}}\left( {x,p} \right){W_{out}}\left( {x,p} \right) \nonumber\\
&=& \frac{2^N}{{{\pi ^N}\sqrt {\det \gamma \det \gamma '} }}\int_{ - \infty }^{ + \infty } {dx} \int_{ - \infty }^{ + \infty } {dp} {e^{ - \left( {\xi - \textbf{m}} \right){\gamma ^{ - 1}}{{\left( {\xi - \textbf{m}} \right)}^T}}} \nonumber\\
&&\times {e^{ - \left( {\xi - \textbf{m}'} \right){{\gamma '}^{ - 1}}{{\left( {\xi - \textbf{m}'} \right)}^T}}} \nonumber\\
&=& 2^N\sqrt {\det{{{\left( {\gamma + \gamma '} \right)}^{ - 1}}}} {e^{ - \left( {\textbf{m} - \textbf{m}'} \right){{\left( {\gamma + \gamma '} \right)}^{ - 1}}{{\left( {\textbf{m} - \textbf{m}'} \right)}^T}}},\label{equ:9}
\end{eqnarray}
where the output state is characterized by the mean value $\textbf{m}'$ and the covariance matrix $\gamma'$. In the present case, we have $N=2$, and, if we assume the input variables $\hat x_a^{in}$ and $\hat p_{ph}^{in}$ are centered around zero such that $\textbf{m}=\textbf{m}'=0$, then the average values of the output states in Eqs. (\ref{equ:8a})$-$(\ref{equ:8d}) are conserved. Consequently, the fidelity of Eq. (\ref{equ:9}) can be further simplified as $F=4\sqrt{\text{det}(\gamma+\gamma')^{-1}}$. For the case of cluster state creation, the two beams before entering the $C_Z$ are always prepared in a squeezed vacuum state \cite{PhysRevA.73.032318,PhysRevLett.104.250503}. Hence, without loss of generality, we can assume that the two input light modes have a normalized variance $(\Delta\hat x_{L,M}^{in})^2=1/(\Delta\hat p_{L,M}^{in})^2=\frac{1}{2}e^{-2s}$ with $s$ the squeezing parameter. With this assumption, we finally obtain the fidelity as:
\begin{equation}
F = \frac{1}{{\sqrt {1 + \frac{2e^{2s}}{{1 + \kappa _0^2}}} }}\label{equ:10},
\end{equation}
\begin{figure}[tp]
\centering
\includegraphics[scale=0.65]{fig_2.eps}
\caption{(Color online) The fidelity of the gate operation vs the coupling strength $\kappa_0$ for different squeezing of the input light. The solid line denotes the coherent input.}
\end{figure}
where we have put $\kappa=1$. Corresponding to this expression, in Fig. 2, we are
able to show the fidelity of the gate operation in its dependence
on the coupling strength $\kappa_0$ for different squeezing of the input light (in dB). As can be seen from the figure that large coupling strength are required for the achievement of high fidelities.
\section{noise effects and experimental feasibility}
\subsection{Noise effects}
So far, we have neglected the noise effects. As in reality,
atoms are usually contained in glass cells. Therefore, light reflections by the cell walls become inevitable. Such reflections, however, can be modeled by a beam splitter type admixture of vacuum components \cite{PhysRevA.73.022331,PhysRevA.72.052313}, which transforms
the input light quadrature as $\hat \vartheta^{in}\to\hat \vartheta^{in'}=\sqrt{1-r}\hat \vartheta^{in}+\sqrt{r}\hat \vartheta^v$ with $\hat\vartheta\in\{\hat x_{L,M},\hat p_{L,M}\}$, where $r$ is the reflection coefficient and $\hat\vartheta^v$ is the vacuum noise quadrature. On the other hand, due to the weak excitation by the light
beams, the atoms also undergo dissipation. We assume that
the spontaneous emission happens at a rate of $\eta/T$ \cite{PhysRevA.73.062329}.
With these modeling, the evolution equations (\ref{equ:2a})$-$(\ref{equ:3d}) are then changed into
\begin{equation}
{\partial _t}{{\hat x}_a}\left( t \right) =\tilde \kappa' \hat p_L^{in}\left( t \right) - \frac{\eta }{{2T}}{{\hat x}_a}\left( t \right) + \sqrt {\frac{\eta }{T}} \hat x_a^v\left( t \right), \nonumber\\
\end{equation}
\begin{equation}
{\partial _t}{{\hat p}_a}\left( t \right) = - \tilde \kappa ' \hat p_M^{in}\left( t \right) - \frac{\eta }{{2T}}{{\hat p}_a}\left( t \right) + \sqrt {\frac{\eta }{T}} \hat p_a^v\left( t \right), \nonumber \\
\end{equation}
\begin{equation}
\hat x_L^{out1}\left( t \right) =\hat x_L^{in'}\left( t \right) + \tilde \kappa' {{\hat p}_a}\left( t \right), \nonumber \\
\end{equation}
\begin{equation}
\hat p_L^{out1}\left( t \right) =\hat p_L^{in'}\left( t \right), \nonumber\\
\end{equation}
\begin{equation}
\hat x_M^{out1}\left( t \right) = \hat x_M^{in'}\left( t \right) +\tilde \kappa' {{\hat x}_a}\left( t \right), \nonumber\\
\end{equation}
\begin{equation}
\hat p_M^{out1}\left( t \right) = \hat p_M^{in'}\left( t \right),
\end{equation}
where $\hat x_a^v$ and $\hat p_a^v$ are Langevin noise operators with zero
mean, having $\langle \hat x_a^v(t)\hat p_a^v(t')\rangle=\delta(t-t')/2$. Here, we have defined the reduced coupling strength $\tilde\kappa'=\sqrt{1-r}\tilde\kappa$. Corresponding to this set of equations, one may derive the modified input-output relations for light
\begin{eqnarray}
\hat x_L^{out1} &=& \hat x_L^{in'} + \kappa '{{\hat p}_a}\left( 0 \right)
- \frac{{{{\kappa '}^2}}}{{{T^{3/2}}}}\int_0^T {dt} \nonumber\\
&&\times{e^{ - \frac{{\eta t}}{{2T}}}}\left( {T - t} \right)\left[ {\hat p_M^{in'}\left( t \right) - \frac{\sqrt {\eta}}{{\kappa '}}\hat p_a^v\left( t \right)} \right],\nonumber\\
\hat x_M^{out1} &=& \hat x_M^{in'} + \kappa '{{\hat x}_a}\left( 0 \right)
+ \frac{{{{\kappa '}^2}}}{{{T^{3/2}}}}\int_0^T {dt}\nonumber\\
&&\times{e^{ - \frac{{\eta t}}{{2T}}}}\left( {T - t} \right)\left[ {\hat p_L^{in'}\left( t \right) + \frac{\sqrt {\eta}}{{\kappa '}}\hat x_a^v\left( t \right)} \right]\label{equ:11},
\end{eqnarray}
and for atoms
\begin{eqnarray}
{{\hat x}_a}\left( T \right) &=& {e^{ - \frac{\eta }{2}}}{{\hat x}_a}\left( 0 \right) + \frac{1}{\sqrt T}\int_0^T {dt} \nonumber\\
&&\times{e^{ - \frac{{\eta t}}{{2T}}}}\left[ {\kappa '\hat p_L^{in'}\left( t \right) + \sqrt {\eta} \hat x_a^v\left( t \right)} \right],\nonumber\\
{{\hat p}_a}\left( T \right) &=& {e^{ - \frac{\eta }{2}}}{{\hat p}_a}\left( 0 \right) - \frac{1}{\sqrt T}\int_0^T {dt}\label{equ:12}\nonumber\\
&&{e^{ - \frac{{\eta t}}{{2T}}}}\left[ {\kappa '\hat p_M^{in'}\left( t \right) - \sqrt {\eta} \hat p_a^v\left( t \right)} \right].
\end{eqnarray}
Before reflecting back into the vapor, the two light pulses will experience another two crossing of the cell wall [see Fig. 1(a)], which transfers the light states of Eqs. (\ref{equ:11}) into $\hat x_i^{out1}\to\hat x_i^{out1'}=\sqrt{1-2r}\hat x_i^{out1}+\sqrt{2r}\hat x_i^v$ with $i\in\{L,M\}$. Using the light state $\hat x_i^{out1'}$ and the atomic state of Eqs. (\ref{equ:12}) as the input states, the second interaction occurs, resulting in the final output states
\begin{eqnarray}
\hat x_L^{out2} &=& \sqrt {1 - 3r} \left[ {\hat x_L^{in} + \epsilon _1^ - {\kappa ^2}\hat p_M^{in}} \right.\nonumber\\
&&\left. { + \epsilon _2^ - {\kappa ^2}\hat{\tilde{p}}_M^{in} + \frac{\eta }{2}\kappa {{\hat p}_a}\left( 0 \right)} \right] + {{\hat f}_{v1}},\label{equ:13a}\nonumber\\
\hat p_L^{out2} &=& \sqrt {1 - 3r} \hat p_L^{in} + \sqrt {3r} \hat p_L^v,\label{equ:13b} \nonumber\\
x_M^{out2} &=& \sqrt {1 - 3r} \left[ {\hat x_M^{in} + \epsilon _1^ + {\kappa ^2}\hat p_L^{in}} \right. \nonumber\\
&&\left. { + \epsilon _2^ + {\kappa ^2}\hat {\tilde p}_L^{in} + 2\kappa \left( {1 - \frac{\eta }{4}} \right){{\hat x}_a}\left( 0 \right)} \right] + {{\hat f}_{v2}}, \label{equ:13c} \nonumber\\
\hat p_M^{out2} &=& \sqrt {1 - 3r} \hat p_M^{in} + \sqrt {3r} \hat p_M^v,\label{equ:13d}
\end{eqnarray}
\begin{figure}[tp]
\centering
\includegraphics[scale=0.7]{fig_3.eps}
\caption{(Color online) (a) Gate operation fidelity $F$ vs coupling $\kappa_0$ in the
presence of atomic decay and light reflection. The input light states
are assumed to be polarization squeezed with the degree of $5$ dB. (b) Optimized fidelity vs decay rate
$\eta$ for different reflection coefficient. The inset shows how the
optimal coupling $\kappa_0$ varies with $\eta$.}
\end{figure}
where $\epsilon_1^{\pm}=(1-r)(1 \pm 2r)( {1 - \frac{\eta }{2} \mp \frac{{2r}}{{1 \pm 2r}}})$ and
$\epsilon_2^{\pm}=(1-r)\frac{{(1 \pm 2r)}}{{\sqrt 3 }}( {\frac{\eta }{2} \pm \frac{{2r}}{{1 \pm 2r}}} )$, and we defined the total vacuum noise $\hat f_{v1}$ and $\hat f_{v2}$, which are a function of the vacuum operators $\hat x_i^v$ and $\hat p_{i}^v$, with $i\in\{a,L,M\}$. Here, $\hat{\tilde p}_{L,M}$ are new defined collective light modes with a temporal mode function $f(t)=(1-2t/T)$, satisfying $[ \hat{\tilde x}_j,\hat{\tilde p}_j ]=i$ with $j\in\{L,M\}$. It is easily checked that they are independent from all other modes \cite{PhysRevA.72.052313}.
Unlike the ideal case, Eqs. (\ref{equ:13a}) show that, besides the atomic position quadrature, the momentum quadrature $\hat p_a(0)$ now also appears. This term arising because of the decoherence of atoms, as we shall see, will lead to the optimal implementation of the scheme. Finally, after taking the end reflection losses into account (because of the fourth crossing of a cell wall) by
damping Eqs. (\ref{equ:13a}) with a factor
$\sqrt{1-r}$ and adding appropriate noise terms, the covariance matrix of the final output state (and, thus, the fidelity) can
then be calculated directly. Putting $\kappa=1$, the fidelity versus coupling strength $\kappa_0$ for squeezed input light with $s=\frac{1}{4}\text{ln}~10$ (corresponding to $5$ dB) in the presence of noises
is depicted in Fig. 3(a) for parameters $r=\eta=0.01$, $r=\eta=0.05$, and $r=\eta=0.1$. In each case, there exist an optimal fidelity, $F=0.89$, $F=0.66$, and $F=0.54$, for $\kappa_0=19.90$, $\kappa_0=8.78$, and $\kappa_0=6.12$, respectively. As illustrated by the
graph, decay and reflections losses have a significant effect on the quality of the gate operation. Losses of the
latter kind, however, can be greatly reduced down (to about $0.5\%$) with improved antireflection coating \cite{PhysRevA.74.064301}. Figure 3(b) shows the $\kappa_0$-optimized fidelity
versus $\eta$ for different (small) values of the reflection parameter $r$. For $\eta=0.1$ (corresponding to the experimental conditions of \cite{NATRUE8}) and $r=0.005$, a fidelity $F=0.71$ would be possible corresponding to $\kappa_0=6.05$.
\subsection{Experimental feasibility}
To successfully and efficiently implement the
$C_Z$ gate, it is required that (i) $T<T_{DL}$, where $T_{DL}$ represents a time in which the two beams
pass through the delay lines, (ii) the coupling strength $\kappa=1$, and (iii) large interaction strength $\kappa_0$ is achievable. Conditions (i), as analyzed in Ref. \cite{PhysRevA.78.010307}, is feasible within presently established techniques, i.e. by using a 1 $\mu s$ delay line \cite{phd2} and a sub-$\mu s$ pulse \cite{ph3}. Condition (ii) has been realized in many physical systems, e.g. room temperature atomic vapors \cite{NATRUE3}. Condition (iii), however, is by now still experimental challenge. Although, theoretically, one can always enhance the coupling strength by increasing the intensity of the light beams, or, the density of atoms, such enhancement, on the other hand, will cause high decay rate, and thus will lower the efficiency of current scheme. To overcome this limitation, we propose to inject squeezed light instead of coherent light during the spin squeezing process, which transfers the input quadratures $(\hat x_{ph}^{in},\hat p_{ph}^{in})$ into $(e^{-s}x_{ph}^{in},e^{s}\hat p_{ph}^{in})$. With this setting, the spin state of Eqs. (\ref{equ:0-1}) is then changed into
\begin{eqnarray}
\hat x_a^{out}=\sqrt{1+e^{2s}\kappa_0^2}\hat x_a^{in},~\hat p_a^{out}=\frac{1}{\sqrt{1+e^{2s}\kappa_0^2}}\hat p_a^{in}.
\end{eqnarray}
Note that the coupling strength is now enhanced by a factor $e^s$. We can define the effective coupling strength $\kappa^{eff}_0=e^s\kappa_0$. As a result, the higher the degree of squeezing is, the larger the effective coupling strength becomes. For room temperature atomic vapors, the feasible value of coupling strength is around $\kappa_0\approx1.4$ \cite{phd1}. In this case, if we inject a light with $8.5$ dB of squeezing, a large coupling strength $\kappa^{eff}_0\approx10$ can then be achieved.
\section{conclusions}
In conclusion, we have presented a simple and realistic scheme for realizing CV $C_Z$ gate in an atomic ensemble. The process is based on off-resonant interaction between light and spin-polarized
atomic ensembles. By sending two off-resonant pulses propagating in two orthogonal directions twice through an atomic ensemble which is initially prepared in spin squeezed state, we find that a $C_Z$ operation between the two pulses is performed. The more the amount of spin squeezing input, the higher the fidelity we will obtain. We also considered the influences of the noise effects including the atomic decay and photon reflections by the cell walls, showing that they have a strong effect on the fidelity. Noises of later kind, however, can be greatly suppressed by adding antireflection coating to cell walls. Such suppressions enable us to achieve quite high fidelities with current experimental parameters.
It is well known that offline squeezing and $C_Z$ gate together enable the construction of arbitrarily large CV cluster states \cite{PhysRevA.79.062318,PhysRevLett.104.250503}. Recently, offline squeezing based on atomic ensembles has been proposed by Sherson \emph{et al.} \cite{PhysRevLett.97.143602}. Therefore, our proposal paves the way for the implementation of CV one-way QC based only on atomic ensembles.
\begin{acknowledgments}
This work was supported by the Natural Science Foundation
of China (Grants No. 11074190 and No. 10947017), and the Natural Science Foundation of Zhejiang province, China (Grant No. Y6090529).
\end{acknowledgments}
\bibliographystyle{apsrev4-1}
|
\section{Introduction}
We address the problem of preparing a desired target quantum state
into the memory of a quantum computer. It is of course unreasonable to try to find an efficient
quantum algorithm to achieve this for general quantum states. Indeed, if any state could be prepared efficiently such
difficult tasks as preparing witnesses for QMA-complete problems~\cite{KKR:2006}
could be solved efficiently, a task believed to be
impossible even if classical side-information about the quantum state
is provided~\cite{AD:2010}. On the other hand, many interesting
computational problems can be related to quantum state generation
problems that carry some additional {\em structure} which might be
exploited by an efficient algorithm.
Among the most tantalizing examples of problems that are
reducible to quantum state generation is the {\sc Graph
Isomorphism} problem~\cite{KST:93} which could be solved by
preparing the quantum state $\ket{\Gamma} = \frac{1}{\sqrt{n!}}
\sum_{\pi\in S_n} \ket{\Gamma^\pi}$, \textit{i.\,e.}, the uniform superposition
of all the permutations of a given graph $\Gamma$. By generating such
states for two given graphs, one could then use the standard SWAP-test~\cite{BCWW:2001}
to check whether the two states are equal or
orthogonal and therefore decide whether the graphs are isomorphic or
not. Furthermore, it is known that all problems in statistical zero
knowledge (SZK) can be reduced to instances of quantum state
generation~\cite{aharonov03}, along with gap-versions of closest
lattice vector problems~\cite{Regev:2004,AR:2005} and subgroup
membership problems for arbitrary subgroups of finite groups~\cite{Watrous:2000,Watrous2001,Friedl2003}.
Aside from brute-force attempts that try to solve quantum state preparation
by applying sequences of controlled
rotations (typically of exponential length) to fix the amplitudes of the target state one qubit at a
time, not much is known regarding approaches to tackle the quantum
state generation problem while exploiting inherent structure.
In this regard, the only examples we are aware of are
(i)~a direct approach to generate states described by efficiently computable amplitudes~\cite{Grover2002},
(ii)~an approach via adiabatic computing~\cite{aharonov03} in which a sequence of local Hamiltonians has to be found such that the desired target state is the ground state of a final Hamiltonian and the overlap between intermediate ground states is large, and
(iii)~recent work on quantum analogues of classical annealing processes~\cite{BKS:2009,SBBK:2008} and of the Metropolis sampling procedure~\cite{TOV+:2009,YA:2010}.
Conversely, for some problems a {\em lower} bound on the complexity of
solving a corresponding quantum state generation problem would be
desirable, for instance to provide further evidence for the security
of quantum money schemes, see \textit{e.\,g.}~\cite{Aaronson:2009,FGH+:2010}.
Unfortunately, except for a recent result that generalizes the
adversary method to a particular case of quantum state generation problems (see~\cite{LMRSS11} and~\cite{AMRR:2011}), very little is known about lower bounds for quantum state generation problems in general.
\partitle{Rejection sampling}
The classical rejection sampling
method\footnote{It is also known as the accept/reject method or
``hit-and-miss'' sampling.} was introduced by
von~Neumann~\cite{vonNeumann51} to solve the following
\emph{resampling} problem: given the ability to sample according to some
probability distribution $P$, one is asked to produce samples from
some other distribution $S$. Conceptually, the method is extremely
simple and works as follows: let $\gamma \leq 1$ be the largest
scaling factor such that $\gammaS$ lies under $P$, formally,
$\gamma=\min_k (p_k/s_k)$. We accept a sample $k$ from $P$ with
probability $\gamma s_k/p_k$, otherwise we reject it and repeat. The
expected number $T$ of samples from $P$ to produce one sample from $S$
is then given by $T = 1/\gamma = \max_k(s_k/p_k)$. See also~\cite{devroye}
for further details and analysis of the method for
various special cases of $P$ and $S$. In a setting where access to the
distribution $P$ is given by a black box, this has been proved to be
optimal by Letac~\cite{Letac75}. The rejection sampling technique is
at the core of many randomized algorithms and has a wide range of
applications, ranging from
computer science to statistical physics, where it is used for
Monte Carlo simulations, the most well-known example being the
Metropolis algorithm~\cite{metropolis53}.
In the same way that quantum state preparation can be considered a
quantum analogue of classical sampling, it is natural to study a
quantum analogue of the classical resampling problem, \textit{i.\,e.}, the
problem of sampling from a distribution $S$ given the ability to
sample from another distribution $P$. We call this problem {\em
quantum resampling} and define it to be the following analogue of
the classical resampling problem: given an oracle generating a
quantum state $\ket{\vc{\pi}^\xi}= \sum_{k=1}^n \pi_k \ket{\xi_k} \ket{k}$,
where the amplitudes $\pi_k$ are known but the states $\ket{\xi_k}$
are unknown, the task is to prepare a target state
$\ket{\vc{\sigma}^\xi}=\sum_{k=1}^n \sigma_k \ket{\xi_k} \ket{k}$ with (potentially)
different amplitudes $\sigma_k$ but the same states $\ket{\xi_k}$. Note that while both the initial amplitudes $\pi_k$ and the final amplitudes $\sigma_k$ are fixed and known, the fact that the states $\ket{\xi_k}$ are unknown makes the problem non-trivial.
\partitle{Related work}
The query complexity of quantum state generation problems was studied in~\cite{AMRR:2011}, where the adversary method was extended to this model and used to prove a tight lower bound on a specific quantum state generation problem called $\IE$. The adversary method was later extended to quantum state \emph{conversion} problems---where the goal is to convert an initial state into a target state--- and shown to be nearly tight in the bounded error case for any problem in this class, which includes as special cases state generation and the usual model of function evaluation~\cite{LMRSS11}. In all these cases, the input to the problem is classical, as the oracle encodes some hidden classical data. This is where the quantum resampling problem differs from those models, as in that case the oracle encodes unknown quantum states.
Grover~\cite{Grover00} considered a special case of the quantum resampling problem, where the initial state
$\ket{\vc{\pi}}=\frac{1}{\sqrt{2^n}}\sum_x \ket{x}$
is a uniform superposition and one is given access to an oracle that for given input $x$ provides a classical description of $\sigma_x$, the amplitude in the target state
$\ket{\vc{\sigma}}=\sum_x \sigma_x \ket{x}$.
We significantly extend the scope of Grover's technique by considering any initial superposition and improving the efficiency of the algorithm when only an approximate preparation of the target state is required.
Techniques related to quantum resampling have already been used implicitly as a useful primitive for building quantum algorithms. For instance, it was used in a paper by Harrow, Hassidim, and Lloyd~\cite{HHL:09} for the problem of solving a system of linear equations $A x = b$, where $A$ is an invertible matrix over the real or complex numbers whose entries are efficiently computable, and $b$ is a vector. The quantum algorithm in~\cite{HHL:09} solves the problem of preparing the state $\ket{x}=A^{-1}\ket{b}$ by applying the following three basic steps. First, use phase estimation on the state
$\ket{b}=\sum_k b_k\ket{\psi_k}$
to prepare the state
$\sum_k b_k\ket{\lambda_k}\ket{\psi_k}$,
where $\ket{\psi_k}$ and $\lambda_k$ denote the eigenstates and eigenvalues of $A$. Next, convert this state to
$\sum_k b_k\lambda_k^{-1}\ket{\lambda_k}\ket{\psi_k}$.
Finally, undo the phase estimation to obtain the target state
$A^{-1}\ket{b}=\sum_k b_k\lambda_k^{-1}\ket{\psi_k}$.
The second step of this procedure performs transformation
$\sum_k b_k\ket{\lambda_k}\ket{\psi_k} \mapsto\sum_k b_k\lambda_k^{-1}\ket{\lambda_k}\ket{\psi_k}$
which can be seen as an instance of quantum resampling. Note that other works, such as~\cite{Childs2009a,Sheridan2009}, have used a similar technique---i.e., using phase estimation to simulate some function of an operator---to apply a unitary on an unknown quantum state, rather than preparing one particular sate.
The quantum Metropolis sampling algorithm has been proposed by Temme~\textit{et al.}~\cite{TOV+:2009} to solve the problem of preparing the thermal state of a quantum Hamiltonian. As it is heavily inspired by the classical Metropolis algorithm, the main step uses an accept/reject rule on random moves between eigenstates of the Hamiltonian. The main complication comes from reverting rejected moves, as the no-cloning principle prevents from keeping a copy of the previous eigenstate. We will show that this step actually reduces to a quantum resampling problem, and that quantum rejection sampling leads to an alternative solution which also provides a speed-up over the technique proposed in~\cite{TOV+:2009}.
Finally, another type of quantum resampling problem has been considered in a paper by Kitaev and Webb~\cite{KW:2009} in which the task is to prepare a superposition of basis states with Gaussian-distributed weights along a low-dimensional strip inside a high-dimensional space. Authors solve this problem by applying a sequence of lattice transformation and phase estimation steps.
For us, another important case in point are hidden shift problems over an abelian group $A$. Here it is easy to prepare a quantum state of the form $\ket{\vc{\pi}^\xi} = \sum_{\bv{w}\in A} \hat{f}(\bv{w}) \chi_{\bv{w}}(s) \ket{\bv{w}}$, where $\chi_{\bv{w}}$ denotes the characters of $A$ and $\hat{f}$ denotes the Fourier transform of $f$ (see e.\,g., \cite{vDHI:2003,Ivanyos:2008,Roetteler:2010}). If we could eliminate the Fourier coefficients $\hat{f}(\bv{w})$ from state $\ket{\vc{\pi}^\xi}$, we would obtain a state $\ket{\vc{\sigma}^\xi} = |A|^{-1/2} \sum_{\bv{w}\in A} \chi_{\bv{w}}(s) \ket{\bv{w}}$ from which the hidden shift $\bv{s}$ can be easily recovered by applying another Fourier transform. Note that in this case the states $\ket{\xi_k}$ are actually just the complex phases $\chi_{\bv{w}}(s)$. We will discuss an application of our general framework to the special case of hidden shift problems in Sect.~\ref{sect:hiddenshift}.
\partitle{Our results}
We denote the classical resampling problem mentioned above by $\sampling_{P \to S}$, where $P$ and $S$ are probability distributions on the set $[n]$. Note that in its simplest form, this problem is not meaningful in the context of query complexity (indeed, if distribution $S$ is known to begin with, there is no need to use the ability to sample from $P$). However, there is a natural modification of the problem, that actually models realistic applications, which does not suffer from this limitation. In this version of the problem, there is a function $\xi$ that deterministically associates some unknown data with each sample, and the problem is to sample pairs $(k,\xi(k))$, where $k$ follows the target distribution. Formally, the problem is therefore defined as follows: given oracle access to a black box producing pairs $(k,\xi(k))\in [n]\times[d]$ such that $k$ is distributed according to $P$, where $\xi:[n]\to[d]$ is an unknown function, the problem is to produce a sample $(k,\xi(k))$ such that $k$ is distributed according to $S$. Note that in this model it is not possible to produce the required samples without using the access to the oracle that produces the samples from $P$, and the algorithm is therefore restricted to act as in Fig.~\ref{fig:CRS}.
\begin{figure}[th]
\getpicture{0.35}{fig-model}
\caption{Classical rejection sampling: A black box produces samples $k$ according to a known probability distribution $P$ and accompanied by some unknown classical data $\xi(k)$. The algorithm $\mathcal{A}$ either accepts or rejects these samples, so that accepted samples are distributed according to a target distribution $S$.}
\label{fig:CRS}
\end{figure}
The problem studied in this article is a quantum analogue of
$\sampling_{P \to S}$, where probability distributions are replaced by
quantum superpositions. More precisely, let $\vc{\pi},\vc{\sigma}\in\mathbb{R}^n$ be such that
$\norm{\vc{\pi}}_2=\norm{\vc{\sigma}}_2=1$ and
$\pi_k,\sigma_k\geq 0$ for all $k\in [n]$. Let $O$ be a unitary
that acts on a default state $\0{dn} \in \mathbb{C}^d \otimes \mathbb{C}^n$ as
$O:\0{dn}\mapsto\ket{\vc{\pi}^\xi} := \sum_{k=1}^n \pi_k \ket{\xi_k}
\ket{k},$ where $\ket{\xi_k} \in \mathbb{C}^d$ are some fixed unknown
normalized quantum states. Given oracle access to unitary black
boxes $O,O^\dagger$, the $\qsamplingab$ problem is
to prepare the state
$\ket{\vc{\sigma}^\xi} := \sum_{k=1}^n \sigma_k \ket{\xi_k} \ket{k}$.
Note that a special case of this problem is the scenario $d = 1$, when
$\xi_k \in \mathbb{C}$ are just unknown phases (complex numbers of absolute
value $1$).
The main result of this article is a tight characterization of the
query complexity of $\qsamplingab$ for any success probability $p$ (the vector $\ve_{\vc{\quantump}\to\vc{\quantums}}^p$, as well as the probabilities $p_{\min},p_{\max}$, will be defined in Sect.~\ref{sect:problem}, intuitively, the vector $\ve_{\vc{\quantump}\to\vc{\quantums}}^p$ characterizes the amplitudes of the final state prepared by the best algorithm having success probability $p$):
\begin{restatable}{theorem}{thmqsampling}
\label{thm:qsampling}
For $p \in [p_{\min}, p_{\max}]$, the quantum query complexity of $\qsamplingab$ with success probability $p$ is $Q_{1-p}(\qsamplingab)=\Theta({1}/{\norm{\ve_{\vc{\quantump}\to\vc{\quantums}}^p}_2})$.
For $p\leq p_{\min}$, the query complexity is 1, and for $p> p_{\max}$, it is infinite.
\end{restatable}
\noindent Let us note that when $p=p_{\max}=1$, the query complexity reduces to $\max_k(\sigma_k/\pi_k)$ in analogy with the classical case, except that amplitudes replace probabilities. The lower bound comes from an extension of the automorphism principle (originally introduced in the context of the adversary method~\cite{AMRR:2011,NegativeWeights}) to our framework of quantum state generation problems with quantum oracles. The upper bounds follows from an algorithm based on amplitude amplification that we call quantum rejection sampling. We also show that a modification of this algorithm can also solve a quantum state conversion problem, which we call strong quantum resampling ($\sqsampling$).
Next, we illustrate the technique by providing different applications. We first show that the main steps in two recent algorithms, namely the quantum algorithm for solving linear systems of equations~\cite{HHL:09} and the quantum Metropolis sampling algorithm~\cite{TOV+:2009}, consists in solving quantum state conversion problems which we call $\qle$ and $\qmm$. We then observe that these problems reduce to $\sqsampling$, and can therefore be solved using quantum rejection sampling.
\begin{restatable}{theorem}{thmlinearequations}
\label{thm:linear-equations}
For any $\tilde{\kappa} \in [1,\kappa]$, there is a quantum algorithm that solves $\qle$ with success probability $p = (\vc{w}^{\mathsf{T}} \cdot \tilde{\vc{w}}) / (\norm{\vc{w}}_2 \cdot \norm{\tilde{\vc{w}}}_2)$ using an expected number of queries $O(\tilde{\kappa}/\norm{\tilde{\vc{w}}}_2)$, where $w_j := b_j / \lambda_j$, $\tilde{w}_j := b_j / \tilde{\lambda}_j$, and $\tilde{\lambda}_j := \max \set{\tilde{\kappa}^{-1}, \lambda_j}$.
\end{restatable}
\begin{restatable}{theorem}{thmmetropolis}
\label{thm:metropolis}
There is a quantum algorithm that solves $\qmm$ with success probability 1 using an expected number of queries $O(1/\|\vc{w}^{(i)}\|_2)$.
\end{restatable}
Let us note that while the quantum algorithm for linear systems of equations was indeed using this technique implicitly, this was not the case for quantum Metropolis sampling, where quantum rejection sampling provides some speed-up over the original algorithm.
Our final result is an application of quantum rejection sampling to the Boolean
hidden shift problem $\bhsp_f$, defined as follows. Let $f:
\bb{n} \to \b$ be a Boolean function, which is assumed to be
completely known. Furthermore, we are given oracle access to
\emph{shifted function} $f_{\bv{s}}(\bv{x}) := f(\bv{x} + \bv{s})$ for some
unknown bit string $\bv{s} \in \bb{n}$, with the promise that there
exists $\bv{x}$ such that $f(\bv{x} + \bv{s})\neq f(\bv{x})$. The task
is to find the bit string $\bv{s}$.
We show that we can solve this problem by solving the
$\qsamplingab$ problem for $\vc{\pi}$
corresponding to the Fourier spectrum of $f$, and $\vc{\sigma}$ being a
uniformly flat vector. This leads to the following upper bound which
expresses the complexity of our quantum algorithm for the Boolean
hidden shift problem in terms of a vector $\ve_{\vc{\quantump}\to\vc{\quantums}}^p$ (defined in
Sect.~\ref{sect:SDP}) that can be thought of as a ``water-filling'' of
the Fourier spectrum of $f$:
\begin{restatable}{theorem}{thmBHSP}
Let $f$ be a Boolean function and $\hat{f}$ be its Fourier transform. For any $p$, we have $Q_{1-p}(\bhsp_f) = O(1/\norm{\ve_{\vc{\quantump}\to\vc{\quantums}}^p}_2)$, where components of $\vc{\pi}$ and $\vc{\sigma}$ are given by $\pi_{\bv{w}}=\abs{\hat{f}_{\bv{w}}}$ and $\sigma_{\bv{w}}=1/\sqrt{2^n}$ for $\bv{w} \in \bb{n}$.
\label{thm:BHSP}
\end{restatable}
As special cases of this theorem we obtain the quantum algorithms for
hidden shift problem for delta functions, which leads to the Grover
search algorithm~\cite{Grover:96}, and for bent functions, which are
functions that have perfectly flat absolute values in their Fourier
spectrum~\cite{Roetteler:2010}. In general, the complexity of the algorithm is limited by the smallest Fourier coefficient of the function. By ignoring small Fourier coefficients, one can decrease the complexity of the algorithm, at the cost of a lower success probability. The final success probability can nevertheless be amplified using repetitions and by constructing a procedure to check a candidate shift, which leads to the following theorem.
\begin{restatable}{theorem}{thmBHSPboost}
\label{thm:BHSP2}
Let $f$ be a Boolean function and $\hat{f}$ be its Fourier transform. Moreover, let $p,\gamma\in[0,1]$ be such that $\norm{\hat{\vc{\varepsilon}}}_1 = \sqrt{2^n p}$, where $\varepsilon_w=\min \set{\abs{\hat{f}_w}, \gamma/\sqrt{2^n}}$. Then, for any $\delta > 0$, we have $Q_{\delta}(\bhsp_f) = O\bigl( \frac{1}{\sqrt{p}} (1/\norm{\vc{\varepsilon}}_2 + 1/\sqrt{\mathrm{I}_{f}}) \bigr)$.
\end{restatable}
\section{Definition of the problem} \label{sect:problem}
In this section, we define different notions related to the quantum resampling problem. It is important to note that this problem goes beyond the usual model of quantum query complexity, where the goal is to compute a function depending on some unknown classical data that can be accessed via an oracle (see~\cite{BuhrmanDeWolf02querysurvey} for a complete survey). In the usual model, the algorithm is therefore quantum but both its input and output are classical. A first extension of this model is the case were the output is also quantum, that is, the goal is to generate a target quantum state depending on some unknown classical data hidden by the oracle. The quantum adversary method has recently been extended to this model by~\cite{AMRR:2011}, where is was used to characterize the query complexity if a quantum state generation problem called $\textsc{IndexErasure}$. In both the usual model and this extension, the oracle acts as $O_x:\ket{b}\ket{i}\mapsto\ket{b+x_i}\ket{i}$, where $x$ is the hidden classical data. However, the quantum resampling problem corresponds to another extension of these models, where the input is also quantum, in the sense that the oracle hides unknown quantum states instead of classical data. Let us now define this extended model more precisely.
\begin{definition}\deftitle{Quantum state generation problem}
Let $\O := \set{O_x : x \in \mathcal{X}}$ and $\Psi := \set{ \ket{\psi_x} : x \in \mathcal{X}}$, respectively, be sets of quantum oracles (\textit{i.\,e.}, unitaries) and target quantum states labeled by elements of some set $\mathcal{X}$. Then $\P:=(\O, \Psi, \mathcal{X})$ describes the following \emph{quantum state generation problem:} given an oracle $O_x$ for some unknown value of $x \in \mathcal{X}$, prepare a state
\begin{equation*}
\ket{\psi}=\sqrt{p}\ket{\psi_x}\0{}+\ket{\mathrm{error}_x},
\end{equation*}
where $p$ is the desired success probability, $\0{}$ is a normalized standard state for some workspace and $\ket{\mathrm{error}_x}$ is an arbitrary error state with norm at most $\sqrt{1-p}$. The quantum query complexity of $\P$ is the minimum number of queries to $O_x$ or $O_x^\dagger$ necessary to solve $\P$ with success probability $p$ and will be denoted by $Q_{1-p}(\P)$.
\end{definition}
Intuitively, we want the final state $\ket{\psi}$ to have a component of length at least $\sqrt{p}$ in the direction of $\ket{\psi_x}\0{}$. We can restate the condition $\norm{\ket{\mathrm{error}_x}}_2 \leq \sqrt{1-p}$ as follows:
\begin{equation*}
1 - p \geq \norm{\ket{\psi} - \sqrt{p} \ket{\psi_x} \0{}}_2^2 = 1 + p - 2 \Re \bigl[ \bra{\psi} \cdot \sqrt{p} \ket{\psi_x} \0{} \bigr],
\end{equation*}
or equivalently:
\begin{equation}
\Re \bigl[ \bra{\psi} \cdot \ket{\psi_x} \0{} \bigr] \geq \sqrt{p}.
\label{eq:Closeness}
\end{equation}
Note that the main difference of the above definition with the usual model of quantum query complexity, and its extension to quantum state generation in~\cite{AMRR:2011}, is that the oracle is not restricted to act as $O_x:\ket{b}\ket{i}\mapsto\ket{b+x_i}\ket{i}$.
We now formally define $\qsamplingab$ as a special case of quantum state generation problem. Throughout this article, we fix positive integers $d,n$ and we assume that $\vc{\pi},\vc{\sigma}\in\mathbb{R}^n$ are vectors such that $\norm{\vc{\pi}}_2=\norm{\vc{\sigma}}_2=1$ and
$\pi_k,\sigma_k\geq 0$ for all $k\in [n]$. We also use the notation $\ket{\vc{\pi}}:=\sum_{k=1}^n \pi_k \ket{k}$ and $\ket{\vc{\sigma}}:=\sum_{k=1}^n \sigma_k \ket{k}$. For simplicity, we assume that $\sigma_k>0$ for all $k\in[n]$, but the general case can easily be obtained by taking the limit $\sigma_k\to 0$.
Let $\xi = (\ket{\xi_k} \in \mathbb{C}^d : k \in [n])$ be an ordered list of normalized quantum states. Then any unitary that maps a default state $\0{dn}$ to $\ket{\vc{\pi}^\xi} := \sum_{k=1}^n \pi_k \ket{\xi_k} \ket{k}$ is a valid oracle for $\qsamplingab$. Therefore, we will label valid oracles by a pair $(\xi,u)$, where $\xi$ denotes the states hidden by the oracle, and $u$ defines how the oracle acts on states that are orthogonal to $\0{dn}$.
More explicitly, we fix a default oracle $O \in \U{dn}$ as a unitary that acts on $\0{dn}$ as $O \0{dn} = \0{d} \ket{\vc{\pi}}$, and arbitrarily on the orthogonal complement. We then use $O$ as a reference point to define the remaining oracles:
\begin{equation}
O_{\xi,u} := V_{\xi} \, O \, u,
\label{eq:oracles}
\end{equation}
where $u \in \U{dn}$ is a unitary such that $u\0{dn}=\0{dn}$ and $V_\xi$ is a unitary that acts on $\0{d}\ket{k}$ as $V_\xi\0{d}\ket{k}=\ket{\xi_k}\ket{k}$ for any $k\in [n]$, and arbitrarily on the orthogonal complement of these states, so that $O_{\xi,u} \0{dn} = V_{\xi} \, O \0{dn} = V_{\xi} \sum_{k=1}^n \pi_k \0{d} \ket{k} = \sum_{k=1}^n \pi_k \ket{\xi_k} \ket{k} = \ket{\vc{\pi}^\xi}$ (note that how $O$ and $V_\xi$ are defined on the orthogonal complements is irrelevant as it only affects the exact labeling, but not the set of valid oracles).
\begin{definition}\deftitle{Quantum resampling problem}
The \emph{quantum resampling problem}, denoted by $\qsamplingab$, is an instance of quantum state generation problem $(\O, \Psi, \mathcal{X})$ with
\begin{align*}
\mathcal{X} &:= \set{(\xi,u) : \xi = (\ket{\xi_1}, \dotsc, \ket{\xi_n}) \in (\mathbb{C}^d)^n, u \in S}, \\
S &:= \set{u\in\U{dn}:u\0{dn}=\0{dn}}\cong \U{dn-1}.
\end{align*}
Oracles in $\O$ that are hiding the states $\ket{\vc{\pi}^\xi}$ are defined according to Eq.~(\ref{eq:oracles}) and the corresponding target states are defined by $\ket{\vc{\sigma}^\xi} := V_{\xi}\0{d} \ket{\vc{\sigma}} = \sum_{k=1}^n \sigma_k\ket{\xi_k} \ket{k}$.
\end{definition}
Let us note that the target states only depend on the index $\xi$, and not $u$. Moreover, note that amplitudes $\pi_k$ and $\sigma_k$ can be assumed to be real and positive without loss of generality, as any phase can be corrected using a controlled-phase gate, which does not require any oracle call since $\vc{\pi}$ and $\vc{\sigma}$ are fixed and known.
In~\cite{LMRSS11}, Lee~\textit{et al.} have proposed another extension of the query complexity model for quantum state generation of~\cite{AMRR:2011} by considering quantum state conversion, where the problem is now to convert a given quantum state into another quantum state, instead of generating a target quantum state from scratch. They have extended the adversary method to this model and showed that it is approximately tight in the bounded-error case for any quantum state conversion problem with a classical oracle. Here, we define a model that combines both extensions (from classical to quantum oracles as well as from state generation to state conversion), hence subsuming all previous models (see Fig.~\ref{fig:problems}).
\begin{definition}\deftitle{Quantum state conversion problem}
Let $\O := \set{O_x : x \in \mathcal{X}}$, $\Phi := \set{ \ket{\phi_x} : x \in \mathcal{X}}$ and $\Psi := \set{ \ket{\psi_x} : x \in \mathcal{X}}$, respectively, be sets of quantum oracles (\textit{i.\,e.}, unitaries), initial quantum states and target quantum states labeled by elements of some set $\mathcal{X}$. Then $\P:=(\O, \Phi, \Psi, \mathcal{X})$ describes the following \emph{quantum state conversion problem:} given an oracle $O_x$ for some unknown value of $x \in \mathcal{X}$ and a copy of the corresponding initial state $\ket{\phi_x}$, prepare a state
\begin{equation}
\ket{\psi}=\sqrt{p}\ket{\psi_x}\0{}+\ket{\mathrm{error}_x},
\end{equation}
where $p$ is the desired success probability, $\0{}$ is a normalized standard state for some workspace and $\ket{\mathrm{error}_x}$ is an arbitrary error state with norm at most $\sqrt{1-p}$. Again, $Q_{1-p}(\P)$ will denote the quantum query complexity of $\P$.
\end{definition}
We also define a strong version of the quantum resampling problem, which is a special case of the state conversion problem. Compared to the original resampling problem, it is made harder due to two modifications. First, instead of being given access to an oracle that maps $\0{dn}$ to $\ket{\vc{\pi}^\xi}$, we are only provided with one copy of $\ket{\vc{\pi}^\xi}$ and an oracle that reflects through it, making this a quantum state conversion problem. The second extension assumes that we only know the ratios between the amplitudes $\pi_k$ and $\sigma_k$ for each $k$, instead of the amplitudes themselves. More precisely, instead of vectors $\vc{\pi},\vc{\sigma}\in\mathbb{R}^n$ specifying the initial and target amplitudes, we fix a single vector $\vc{\tau}\in\mathbb{R}^n$ such that $\tau_k \geq 0$ and $\max_k \tau_k = 1$, specifying the ratios between those amplitudes. Let us now formally define the stronger version of the quantum resampling problem (this definition is motivated by the applications that will be presented in Sect.~\ref{sect:linear} and~\ref{sect:metropolis}).
\begin{definition}\deftitle{Strong quantum resampling problem}
Let $P := \set{\vc{\pi} \in \mathbb{R}^n : \norm{\vc{\pi}}_2 = 1, \forall k: \pi_k > 0}$. The \emph{strong quantum resampling problem $\sqsampling_{\vc{\tau}}$} is a quantum state conversion problem $(\O, \Phi, \Psi, \mathcal{X})$, where $\mathcal{X} := \set{(\xi,\vc{\pi}) : \xi = (\ket{\xi_1}, \dotsc, \ket{\xi_n}) \in (\mathbb{C}^d)^n, \vc{\pi} \in P}$, oracles in $\O$ are defined by $O_{\xi,\vc{\pi}}:=\reflection{\ket{\vc{\pi}^\xi}} = I - 2 \ket{\vc{\pi}^\xi} \bra{\vc{\pi}^\xi}$ with the corresponding initial and target states being $\ket{\vc{\pi}^\xi}$ and $\ket{\vc{\sigma}^\xi} = \sum_{k=1}^n \sigma_k\ket{\xi_k} \ket{k}$, respectively, where $\vc{\sigma} := \vc{\pi} \circ \vc{\tau} / \norm{\vc{\pi} \circ \vc{\tau}}_2$ so that $\sigma_k / \pi_k \propto \tau_k$.
\label{def:SQSampling}
\end{definition}
\begin{figure}[th]
\getpicture{0.45}{fig-problems}
\caption{Comparison of different classes of problems in quantum query complexity. The case of function evaluation has been extensively studied in the literature. The extension to quantum state generation with classical oracles, as well as the problem \textnormal{$\IE$} which belongs to that class, have been studied in~\cite{AMRR:2011}. The adversary method has been extended to the case of quantum state conversion with classical oracles in~\cite{LMRSS11}. The problems \textnormal{$\qsamplingab$} and \textnormal{$\sqsampling_{\vc{\tau}}$} studied in this article use quantum oracles and therefore belong to yet another extension of the quantum query complexity model.}
\label{fig:problems}
\end{figure}
The relationship between different classes of query complexity problems introduced above, and strong and regular quantum rejection sampling as special instances of them are summarized in Fig.~\ref{fig:problems}. Our main result is that the quantum query complexities of $\qsamplingab$ and $\sqsampling_{\vc{\tau}}$ for any success probability $p$ depend on a vector $\ve_{\vc{\quantump}\to\vc{\quantums}}^p$ defined as follows.
\begin{definition}\label{def:water-filling}
For any $\vc{\pi},\vc{\sigma}$, let us define the following quantities
\begin{align*}
p_{\min} &:= (\vc{\sigma}^{\mathsf{T}} \cdot \vc{\pi})^2, &
\gamma_{\min} &:= \min_{k:\pi_k>0} (\pi_k/\sigma_k), \\
p_{\max} &:= \sum_{k:\pi_k>0} \sigma_k^2, &
\gamma_{\max} &:= \max_{k} (\pi_k/\sigma_k).
\end{align*}
For any $\gamma\in[\gamma_{\min},\gamma_{\max}]$, let us define a vector $\vc{\varepsilon}(\gamma)$ and a scalar $p(\gamma)$ by
\begin{align*}
\varepsilon_k(\gamma)&:=\min \set{\pi_k, \gamma \sigma_k}, &
p(\gamma)&:=\biggl(\frac{\vc{\sigma}^{\mathsf{T}}\cdot\vc{\varepsilon}(\gamma)}{\norm{\vc{\varepsilon}(\gamma)}_2}\biggr)^2.
\end{align*}
For $p \in [p_{\min}, p_{\max}]$, let $\bar{\gamma}\in[\gamma_{\min},\gamma_{\max}]$ be such that $p(\bar{\gamma})=p$ and define a vector $\ve_{\vc{\quantump}\to\vc{\quantums}}^p := \vc{\varepsilon}(\bar{\gamma})$.
\end{definition}
To see that $\ve_{\vc{\quantump}\to\vc{\quantums}}^p$ is well-defined, note that $\norm{\vc{\varepsilon}(\gamma)}_2$ is monotonically increasing with $\gamma$, while $p(\gamma)$ is monotonically decreasing with $\gamma$. More precisely, for $\gamma=\gamma_{\min}$, the vector $\vc{\varepsilon}(\gamma)$ has components $\varepsilon_k(\gamma) = \gamma \sigma_k$ if $\pi_k \neq 0$ or zero otherwise, and $p(\gamma)=p_{\max}$. For $\gamma=\gamma_{\max}$, we have $\vc{\varepsilon}(\gamma)=\vc{\pi}$ and $p(\gamma)=p_{\min}$. Between these extreme cases, $p(\gamma)$ interpolates from $p_{\max}$ to $p_{\min}$, which means that for any $p\in[p_{\min}, p_{\max}]$, there exists a value $\bar{\gamma}$ such that $p(\bar{\gamma})=p$, which uniquely defines $\ve_{\vc{\quantump}\to\vc{\quantums}}^p$.
Intuitively, $\vc{\varepsilon}(\gamma)$ may be interpreted as a ``water-filling'' vector, where $\gamma$ defines the water level, and $\pi_k$ defines the height of ``tank'' number $k$. As $\gamma$ increases from $0$ to $\gamma_{\min}$, we have $\varepsilon_k(\gamma) = \gamma \sigma_k$, meaning that all tanks are progressively filled proportionally to $\gamma$. When $\gamma>\gamma_{\min}$, some tanks are filled ($\gamma \sigma_k>\pi_k$) and cannot hold more water, while others continue to get filled.
\section{Query complexity of quantum resampling} \label{sect:SDP}
Let us first show that $\ve_{\vc{\quantump}\to\vc{\quantums}}^p$ defines an optimal feasible point of a certain semidefinite program (SDP). Afterwards we will show that the same SDP characterizes the quantum query complexity of $\qsamplingab$.
\begin{restatable}{lemma}{lemSDPsolution}\label{lem:sdp-solution}
Let $p\in[p_{\min}, p_{\max}]$, and $\vc{\varepsilon}=\ve_{\vc{\quantump}\to\vc{\quantums}}^p$. Then, the following SDP
\begin{equation}
\begin{array}{rl}
\max_{M \succeq 0} \Tr M
\quad \text{s.t.} \quad
& \forall k: \pi_k^2 \geq M_{kk}, \\
& \Tr \bigl[ (\vc{\sigma} \cdot \vc{\sigma}^{\mathsf{T}} - p I) M \bigr] \geq 0.
\end{array}
\label{eq:Primal}
\end{equation}
has optimal value $\norm{\vc{\varepsilon}}_2^2$, which is achieved by the rank-1 matrix $M=\vc{\varepsilon}\cdot\vc{\varepsilon}^{\mathsf{T}}$.
\end{restatable}
\begin{proof}[sketch]
We first show that $M=\vc{\varepsilon}\cdot\vc{\varepsilon}^{\mathsf{T}}$, where $\vc{\varepsilon}=\ve_{\vc{\quantump}\to\vc{\quantums}}^p$, satisfies the constraints in SDP~(\ref{lem:sdp-solution}) and therefore constitutes a feasible point. Therefore, the optimal value of~(\ref{lem:sdp-solution}) is at least $\Tr M=\norm{\vc{\varepsilon}}_2^2$. Secondly, we dualize the SDP, and provide a dual-feasible point achieving the same objective value. The fact that this objective value is feasible for both the primal and the dual then implies that this is the optimal value.
The details of the proof are given in Appendix~\ref{apx:SDP}.
\qed
\end{proof}
Next, let us prove that SDP~(\ref{eq:Primal}) provides a lower bound for the $\qsamplingab^p$ problem. In Sect.~\ref{sect:Algorithm}, we will also show that this lower bound is tight by providing an explicit algorithm.
Let us emphasize the fact that the lower bound cannot be obtained from standard methods such as the adversary method~\cite{Amb00,NegativeWeights} (which has recently been proved to be tight for evaluating functions~\cite{Reichardt:2009,ReichardtReflections,LMRSS11}), nor from its extension to quantum state generation problems~\cite{AMRR:2011,LMRSS11}, because in this case the oracle is also quantum, in the sense that it encodes some unknown quantum state rather than some unknown classical data. To prove lower bounds it is useful to exploit possible symmetries of the problem. We extend the notion of automorphism group~\cite{AMRR:2011,NegativeWeights} to our framework of quantum state generation problems:
\begin{definition}\deftitle{Automorphism group}
We say that $G$ is the \emph{automorphism group} of problem $(\O, \Psi, \mathcal{X})$ if:
\begin{enumerate}
\item $G$ acts on $\mathcal{X}$ (and thus implicitly also on $\O$ as $g: O_x \mapsto O_{g(x)}$).
\item For any $g \in G$ there is a unitary $U_g$ such that $U_g \ket{\psi_x} = \ket{\psi_{g(x)}}$ for all $x \in \mathcal{X}$.
\item For any given $g \in G$ it is possible to simulate the oracles $O_{g(x)}$ for all $x \in \mathcal{X}$, using only a constant number of queries to the black box $O_x$.
\end{enumerate}
\end{definition}
While for the standard model of quantum query complexity, where the oracle encodes some unknown classical data, the automorphism group is restricted to be a permutation group and is therefore always finite, in this more general framework the automorphism group can be continuous. For example, in the case of $\qsamplingab^p$ it will involve the unitary group.
Taking such symmetries into account might significantly simplify the analysis of algorithms for the corresponding problem and is the key to prove our lower bound.
\newcommand{\g }[3]{ \gamma _{#1}(#2,#3)}
\newcommand{\gb}[1]{\bar{\gamma}_{#1}}
\newcommand{\dm}[1]{\;d\mu(#1)}
\newcommand{\twirling }[1]{\twirlingwithcoefficient{\frac{1}{\sqrt{2^n}}}{#1}}
\newcommand{\twirlingsquare}[1]{\twirlingwithcoefficient{\frac{1}{ 2^n }}{#1}}
\newcommand{\twirlingwithcoefficient}[2]{#1 \sum_{\bv{y} \in \bb{n}} \int_{v \in S} #2 \dm{v}}
\begin{restatable}{lemma}{lemlowerbound}
\label{lem:lower-bound}
Any quantum algorithm for $\qsamplingab^p$ with $p \in [p_{\min}, p_{\max}]$ requires at least $\Omega(1/\norm{\ve_{\vc{\quantump}\to\vc{\quantums}}^p}_2)$ queries to $O$ and $O^{\dagger}$.
\end{restatable}
\begin{proof}
The proof proceeds as follows: we first define a subset of oracles that are sufficiently hard to distinguish to characterize the query complexity of the problem. Exploiting the automorphism group of this subset of oracles, we then show that any algorithm may be symmetrized in such a way that the real part of the amplitudes of the final state prepared by the algorithm does not depend on the specific oracle it was given. These amplitudes define a vector $\bar{\vc{\gamma}}$ that should satisfy some constraints for the algorithm to have success probability $p$. Moreover, we can use the hybrid argument to show that the components of $\bar{\vc{\gamma}}$ should also satisfy some constraints for the algorithm to be able to generate the corresponding state in at most $T$ queries. Putting all these constraints together in an optimization program, we then show that such a vector $\bar{\vc{\gamma}}$ cannot exist unless $T$ is large enough. This optimization program is then shown to be equivalent to the semidefinite program in Lemma~\ref{lem:sdp-solution}, which proves the theorem.
Let us now give the details of the proof. For given $\vc{\pi}, \vc{\sigma} \in \mathbb{R}^n$, let us choose a subset of oracles $\O'_{\vc{\pi},\vc{\sigma}}\subset\O_{\vc{\pi},\vc{\sigma}}$ that are hard to distinguish. We choose the states hidden inside oracles to be of the form $\ket{\xi_k} = (-1)^{x_k} \0{d}$, where phases are given by some unknown string $\bv{x} \in \bb{n}$. We also choose $u$ so that any oracle in the subset acts trivially on the $d$-dimensional register holding the unknown states. In that case, this register is effectively one-dimensional, so we will omit it and write $(-1)^{x_k}$ as a relative phase.
More precisely, we consider a set of oracles $\O'_{\vc{\pi},\vc{\sigma}} := \set{O_{\bv{x},u} : \bv{x} \in \bb{n}, u \in S}$, where
\begin{equation}
S := \set{u \in \mathrm{U}(n) : u \0{n} = \0{n}} \cong \U{n-1}.
\label{eq:S}
\end{equation}
As in the general case, we define the first oracle $O_{\bv{0},I}$ as a unitary that acts on $\0{n}$ as $O_{\bv{0},I} \0{n} = \ket{\vc{\pi}}$, and arbitrarily on the orthogonal complement, and we use $O_{\bv{0},I}$ as a reference point to define the remaining oracles:
\begin{align*}
O_{\bv{x},u} &:= V_{\bv{x}} \, O_{\bv{0},I} u, &
\text{where} &&
V_{\bv{x}} &:= \textstyle \sum_{k=1}^n\ (-1)^{x_k} \ket{k} \bra{k}.
\end{align*}
The set of target states is $\Psi'_{\vc{\pi},\vc{\sigma}} := \set{\ket{\vc{\sigma}(\bv{x})} : \bv{x} \in \bb{n}, u \in S}$ where $\ket{\vc{\sigma}(\bv{x})} := V_{\bv{x}} \ket{\vc{\sigma}} = \sum_{k=1}^n (-1)^{x_k} \sigma_k \ket{k}$. For the quantum state generation problem corresponding to the restricted set of oracles $\O'_{\vc{\pi},\vc{\sigma}}$, the automorphism group is $G = \bb{n} \times \U{n-1}$ and it acts on itself, \textit{i.\,e.}, $\mathcal{X} = G$. Note that the target states depend only on $\bv{x}$, but $u$ is used for parameterizing the oracles. Intuitively, the reason we need this parameter is that the algorithm should not depend on how the black box acts on states other than $\0{n}$ (or how its inverse acts on states other than $\ket{\vc{\pi}^\xi}$). To make this intuition formal, we will later choose the parameter $u$ for different oracles adversarially.
Let us consider an algorithm that uses $T$ queries to the black box $O_{\bv{x},u}$ and its inverse, and let us denote the final state of this algorithm by $\ket{\psi_T(\bv{x},u)}$. If we expand the first register in the standard basis, we can express this state as
\begin{equation*}
\ket{\psi_T(\bv{x},u)} = \textstyle \sum_{k=1}^n (-1)^{x_k} \ket{k} \ket{\g{k}{\bv{x}}{u}}.
\end{equation*}
Here the workspace vectors $\ket{\g{k}{\bv{x}}{u}}$ can have arbitrary dimension and are not necessarily unit vectors, but instead satisfy the normalization constraint $\sum_{k=1}^n \|\ket{\g{k}{\bv{x}}{u}}\|_2^2 = 1$. If the algorithm succeeds with probability $p$, then according to Eq.~(\ref{eq:Closeness}) for any $\bv{x}$ and $u$ we have
\begin{align*}
\sqrt{p}
&\leq \Re \bigl[\bra{\vc{\sigma}(\bv{x})} \bra{\bar{0}} \cdot \ket{\psi_T(\bv{x},u)}\bigr]\\
&= \Re \bigl[\textstyle \sum_{k=1}^n \sigma_k\cdot
\braket{\bar{0}}{\g{k}{\bv{x}}{u}}\bigr]\\
&= \vc{\sigma}^{\mathsf{T}} \cdot \vc{\gamma}(\bv{x},u),
\end{align*}
where $\vc{\gamma}(\bv{x},u)$ is a real vector whose components are given by
\begin{equation*}
\g{k}{\bv{x}}{u} := \Re \bigl[ \braket{\bar{0}}{\g{k}{\bv{x}}{u}} \bigr ].
\end{equation*}
Note that $\norm{\vc{\gamma}(\bv{x},u)}_2 \leq 1$.
\begin{figure*
\getpicture{0.95}{fig-symmetrization}
\caption{Symmetrized algorithm. We symmetrize the algorithm by introducing unitaries $V_{\bv{y}}$ and $v$, controlled by an extra register prepared in the uniform superposition over all $\ket{\bv{y}}$ and $\ket{v}$.}
\label{fig:Symmetrization}
\end{figure*}
Next, let us show that we can symmetrize the algorithm without decreasing its success probability. We do this by replacing each oracle call by $O_{\bv{x}+\bv{y},uv} = V_{\bv{y}} O_{\bv{x},u} v$ and correcting the final state by applying $V_{\bv{y}}^{\dagger}$ (see Fig.~\ref{fig:Symmetrization}). Let $\mu$ be the Haar measure on the set $S$ defined in Eq.~(\ref{eq:S}). We define an operation that symmetrizes a set of states:
\begin{equation*}
\overline{\ket{\phi(\bv{x},u)}}\!
:= \!\!\twirling{
\!\Bigl[ (V_{\bv{y}}^{\dagger} \otimes I) \ket{\phi(\bv{x} + \bv{y},uv)} \Bigr]
\ket{\bv{y}} \ket{v}\!
}.
\end{equation*}
If we symmetrize the final state $\ket{\psi_T(\bv{x},u)}$, we get
\begin{multline*}
\overline{\ket{\psi_T(\bv{x},u)}} = \\
\twirling{
\sum_{k=1}^n (-1)^{x_k}
\ket{k} \ket{\g{k}{\bv{x} + \bv{y}}{uv}}
\ket{\bv{y}} \ket{v}
}.
\end{multline*}
Note that the target state $\ket{\vc{\sigma}(\bv{x})} \0{}$ is already symmetric, so symmetrization only introduces an additional workspace register in a default state (uniform superposition):
\begin{equation*}
\overline{\ket{\vc{\sigma}(\bv{x})} \0{}}
= \ket{\vc{\sigma}(\bv{x})}\0{}\twirling{ \ket{\bv{y}} \ket{v}
}.
\end{equation*}
The success probability of the symmetrized algorithm is
\begin{align*}
\sqrt{\bar{p}}
&:= \Re \Bigl[\overline{\bra{\vc{\sigma}(\bv{x})} \bra{\bar{0}}}
\cdot \overline{\ket{\psi_T(\bv{x},u)}} \Bigr] \\
& = \sum_{k=1}^n \sigma_k \cdot
\twirlingsquare{\Re \bigl[\braket{\bar{0}}{\g{k}{\bv{x} + \bv{y}}{uv}}}\bigr] \\
& = \vc{\sigma}^{\mathsf{T}} \cdot \bar{\vc{\gamma}},
\end{align*}
where, by changing variables, we get that $\bar{\vc{\gamma}}$ is the average of vectors $\vc{\gamma}(\bv{y},v)$ and thus does not depend on $\bv{x}$ and $u$:
\begin{equation*}
\bar{\vc{\gamma}} := \twirlingsquare{\vc{\gamma}(\bv{y},v)}.
\end{equation*}
Note that $\norm{\bar{\vc{\gamma}}}_2 \leq 1$ by triangle inequality. Also, note that $\bar{p} \geq p$, since the mean is at least as large as the minimum. Thus without loss of generality we can consider only symmetrized algorithms.
Let $\bv{x}, \bv{x}' \in \bb{n}$ and $u, u' \in S$. The difference of final states of the symmetrized algorithm that uses oracles $O_{\bv{x},u}$ and $O_{\bv{x}',u'}$ is given in Eq.~(\ref{eq:Distance lower bound}) on the next page.
\begin{figure*}
\normalsize
\setcounter{MYtempeqncnt}{\value{equation}}
\setcounter{equation}{9}
\begin{align}
\norm{\overline{\ket{\psi_T(\bv{x},u)}} - \overline{\ket{\psi_T(\bv{x}',u')}}}_2^2
&= \norm{
\sum_{k=1}^n
\twirling{
\ket{k}
\bigl(
(-1)^{x _k} \ket{\g{k}{\bv{x} + \bv{y}}{u v}} -
(-1)^{x'_k} \ket{\g{k}{\bv{x}'+ \bv{y}}{u'v}}
\bigr)
\ket{\bv{y}} \ket{v}
}
}_2^2 \nonumber \\
&= \sum_{k=1}^n
\twirlingsquare{
\norm{
(-1)^{x _k} \ket{\g{k}{\bv{x} + \bv{y}}{u v}} -
(-1)^{x'_k} \ket{\g{k}{\bv{x}'+ \bv{y}}{u'v}}
}_2^2
} \nonumber \\
&\geq
\sum_{k=1}^n
\twirlingsquare{
\bigl(
(-1)^{x _k} \g{k}{\bv{x} + \bv{y}}{u v} -
(-1)^{x'_k} \g{k}{\bv{x}'+ \bv{y}}{u'v}
\bigr)^2
} \nonumber \\
&\geq
\sum_{k=1}^n
\bigl(
(-1)^{x _k} \gb{k} -
(-1)^{x'_k} \gb{k}
\bigr)^2
= \sum_{k: x_k \neq x'_k} \bigl( 2 \gb{k} \bigr)^2.
\label{eq:Distance lower bound}
\end{align}
\setcounter{equation}{\value{MYtempeqncnt}}
Here the two inequalities were obtained from the following facts:
\begin{enumerate}
\item If $\0{}$ is a unit vector then for any $\ket{\gamma}$ we have:
$
\norm{\ket{\gamma}}_2^2 \geq
\norm{\0{} \braket{\bar{0}}{\gamma}}_2^2 =
\abs {\braket{\bar{0}}{\gamma}}^2 \geq
\bigl(\Re[\braket{\bar{0}}{\gamma}]\bigl)^2.
$
\item For any function $\gamma(\bv{y},v)$ by Cauchy--Schwarz inequality we have:
\begin{equation*}
\twirlingsquare{\gamma(\bv{y},v)^2}
\geq \Biggl( \twirlingsquare{\gamma(\bv{y},v)} \Biggr)^2.
\end{equation*}
\end{enumerate}
\hrulefill
\vspace*{4pt}
\end{figure*}%
\newcommand{\hybrid}{It follows by induction from the following fact. If $O$ and $O'$ are unitary matrices, then for any vectors $\ket{\psi}$ and $\ket{\psi'}$ it holds that
\begin{align*}
\norm{O \ket{\psi} - O' \ket{\psi'}}_2
&= \norm{O \ket{\psi} - O' \ket{\psi} + O' \ket{\psi} - O' \ket{\psi'}}_2 \\
&\leq \norm{(O - O') \ket{\psi}}_2 + \norm{O' (\ket{\psi} - \ket{\psi'})}_2 \\
&\leq \norm{O - O'}_\infty + \norm{\ket{\psi} - \ket{\psi'}}_2.
\end{align*}
}%
By the hybrid argument~\cite{bbbv97,Vazirani:98}, we get the following upper bound:\footnote{\hybrid}
\begin{equation}
\norm{\overline{\ket{\psi_T(\bv{x},u)}} - \overline{\ket{\psi_T(\bv{x}',u')}}}_2
\leq T \cdot \norm{O_{\bv{x},u}-O_{\bv{x}',u'}}_\infty,
\label{eq:Distance upper bound}
\end{equation}
where $\norm{\cdot}_\infty$ denotes the usual operator norm.
Bounds~(\ref{eq:Distance lower bound}) and~(\ref{eq:Distance upper bound}) together imply that for any $\bv{x}, \bv{x}' \in \bb{n}$ and $u, u' \in S$:
\begin{equation}
T \geq \frac{\sqrt{\sum_{k: x_k \neq x'_k} \bigl( 2 \gb{k} \bigr)^2}}
{\norm{O_{\bv{x},u}-O_{\bv{x}',u'}}_\infty}.
\label{eq:Lower bound}
\end{equation}
To obtain a good lower bound, we want to choose oracles $O_{\bv{x},u}$ and $O_{\bv{x}',u'}$ to be as similar as possible. First, let us fix $u := I$, $\bv{x} := \bv{0}$ and $\bv{x}' := \bv{e}_l$, where $\bv{e}_l$ is the $l$-th standard basis vector. Then, the numerator in Eq.~(\ref{eq:Lower bound}) is simply $2\gb{l}$. Let us choose $u'$ in order to minimize the denominator. Recall that $O_{\bv{0},I} \0{n} = \ket{\vc{\pi}}$ and note that any unitary matrix that fixes $\ket{\vc{\pi}}$ can be written as $O_{\bv{0},I} u' (O_{\bv{0},I})^{\dagger}$ for some choice of $u'$ fixing $\0{n}$. Since $O_{\bv{x}',u'} \0{n} = V_{\bv{e}_l} O_{\bv{0},I} u' \0{n} = V_{\bv{e}_l} \ket{\vc{\pi}}$, we also have $O_{\bv{x}',u'} (O_{\bv{0},I})^{\dagger} \ket{\vc{\pi}} = V_{\bv{e}_l} \ket{\vc{\pi}}$, and any unitary matrix that sends $\ket{\vc{\pi}}$ to $V_{\bv{e}_l} \ket{\vc{\pi}}$ can be expressed as $O_{\bv{x}',u'} (O_{\bv{0},I})^{\dagger}$ for some choice of $u'$. Let us choose $u'$ so that $O_{\bv{x}',u'} (O_{\bv{0},I})^{\dagger}$ acts as a rotation in the two-dimensional subspace $\spn \set{\ket{\vc{\pi}}, V_{\bv{e}_l} \ket{\vc{\pi}}}$ and as identity on the orthogonal complement. If $\theta$ denotes the angle of this rotation, then $\cos \theta = \bra{\vc{\pi}} V_{\bv{e}_l} \ket{\vc{\pi}} = \sum_{k=1}^n \pi_k^2 - 2 \pi_l^2 = 1 - 2 \pi_l^2$ and $\sin \theta = \sqrt{1 - (1 - 2 \pi_l^2)^2} = 2 \pi_l \sqrt{1- \pi_l^2}$. Then
\begin{align*}
\norm{O_{\bv{0},I} - O_{\bv{x}',u'}}_\infty
&= \norm{I - O_{\bv{x}',u'} (O_{\bv{0},I})^{\dagger}}_\infty \\
&= \norm{I -
\mx{
\cos \theta & -\sin \theta \\
\sin \theta & \cos \theta
}
}_{\infty}\\
&= 2 \pi_l
\norm{
\mx{
\pi_l & \sqrt{1 - \pi_l^2} \\
-\sqrt{1 - \pi_l^2} & \pi_l
}
}_{\infty} \\
&= 2 \pi_l,
\end{align*}
where the singular values of the last matrix are equal to $1$, since it is a rotation. By plugging this back in Eq.~(\ref{eq:Lower bound}), we get that for any $l \in [n]$:
\begin{equation*}
T \geq \frac{\abs{\gb{l}}}{\pi_l}.
\end{equation*}
Thus any quantum algorithm that solves $\qsamplingab^p$ with $T$ queries and success probability $p$ gives us some vector $\bar{\vc{\gamma}}$ such that
\begin{align}
\norm{\bar{\vc{\gamma}}}_2 &\leq 1, &
\vc{\sigma}^{\mathsf{T}} \cdot \bar{\vc{\gamma}} &\geq \sqrt{p}, &
\forall l: \abs{\gb{l}} \leq T \pi_l.
\label{eq:Feasibility}
\end{align}
To obtain a lower bound on $T$, we have to find the smallest possible $t$ such that there is still a feasible value of $\bar{\vc{\gamma}}$ that satisfies Eqs.~(\ref{eq:Feasibility}) (with $T$ replaced by $t$).
We can state this as an optimization problem:
\begin{equation}
\begin{array}{rl}
T \geq \min_{\bar{\vc{\gamma}}} t
\quad \text{s.t.} \quad
& \norm{\bar{\vc{\gamma}}}_2 \leq 1, \\
& \forall l: \abs{\gb{l}} \leq t \pi_l, \\
& \vc{\sigma}^{\mathsf{T}} \cdot \bar{\vc{\gamma}} \geq \sqrt{p}.
\end{array}
\label{eq:Subnormal gamma}
\end{equation}
Finally, let us show that we can start with a feasible solution $\bar{\vc{\gamma}}$ of problem~(\ref{eq:Subnormal gamma}) and modify its components, without increasing the objective value or violating any of the constraints, so that $\forall l: \gb{l} \geq 0$ and $\norm{\bar{\vc{\gamma}}}_2 = 1$. Clearly, making all components of $\bar{\vc{\gamma}}$ non-negative does not affect the objective value and makes the last constraint only easier to satisfy since $\sigma_k\geq 0$ for all $k$. To turn $\bar{\vc{\gamma}}$ into a unit vector, first observe that not all of the constraints $\gb{l} \leq t \pi_l$ can be saturated (indeed, in that case we would have $\bar{\vc{\gamma}} = t \vc{\pi}$ with $t < 1$ since $\norm{\bar{\vc{\gamma}}}_2 < \norm{\vc{\pi}}_2 = 1$, but the last constraint then implies $\vc{\sigma}^{\mathsf{T}} \cdot \vc{\pi} > \sqrt{p}$, which contradicts the assumption $p \geq p_{\min}$). If $\norm{\bar{\vc{\gamma}}}_2 < 1$, let $j$ be such that $\gb{j} < t \pi_j$. We increase $\gb{j}$ until either $\norm{\bar{\vc{\gamma}}}_2 = 1$ or $\gb{j} = t \pi_j$. We then repeat with another $j$ such that $\gb{j} < t \pi_j$, until we reach $\norm{\bar{\vc{\gamma}}}_2 = 1$. Note that while doing so, the other constraints remain satisfied. Therefore, the program reduces to
\begin{equation*}
\begin{array}{rl}
T \geq \min_{\bar{\vc{\gamma}}} t
\quad \text{s.t.} \quad
& \norm{\bar{\vc{\gamma}}}_2 = 1, \\
& \forall l: 0 \leq \gb{l} \leq t \pi_l, \\
& \vc{\sigma}^{\mathsf{T}} \cdot \bar{\vc{\gamma}} \geq \sqrt{p}.
\end{array}
\end{equation*}
Note that we need $p\leq p_{\max}$, otherwise this program has no feasible point. Setting $\vc{\varepsilon} = \bar{\vc{\gamma}}/t$, we may rewrite this program as in Eq.~(\ref{eq:Primal1}) in Appendix~\ref{apx:SDP}:
\begin{equation*}
\begin{array}{rl}
\frac{1}{T^2}\leq\max_{\varepsilon_k \geq 0} \norm{\vc{\varepsilon}}_2^2
\quad \text{s.t.} \quad
& \forall k: \pi_k \geq \varepsilon_k \geq 0, \\
& \vc{\sigma}^{\mathsf{T}} \cdot \vc{\varepsilon} \geq \sqrt{p} \norm{\vc{\varepsilon}}_2,
\end{array}
\end{equation*}
Finally, setting $M=\vc{\varepsilon}\cdot\vc{\varepsilon}^{\mathsf{T}}$, this program becomes the same as the SDP in Eq.~(\ref{eq:Primal}), with the additional constraint that $M$ is rank-1. However, we know from Lemma~\ref{lem:sdp-solution} that the SDP in Eq.~(\ref{eq:Primal}) admits a rank-1 optimal point, therefore adding this constraint does not modify the objective value, which is also $\norm{\ve_{\vc{\quantump}\to\vc{\quantums}}^p}_2^2$ by Lemma~\ref{lem:sdp-solution}.
\qed
\end{proof}
\section{Quantum rejection sampling\\algorithm} \label{sect:Algorithm}
In this section, we describe quantum rejection sampling algorithms for $\qsamplingab$ and $\sqsampling_{\vc{\tau}}$ problems. We also explain the intuition behind our method and its relation to the classical rejection sampling. Our algorithms are based on amplitude amplification~\cite{AmplitudeAmplification} and can be seen as an extension of the algorithm in~\cite{Grover00}.
\subsection{Intuitive description of the algorithm}
The workspace of our algorithm is $\mathbb{C}^d \otimes \mathbb{C}^n \otimes \mathbb{C}^2$, where the last register can be interpreted as a quantum coin that determines whether a sample will be rejected or accepted (this corresponds to basis states $\ket{0}$ and $\ket{1}$, respectively). Our algorithm is parametrized by a vector $\vc{\varepsilon} \in \mathbb{R}^n$ ($0 \leq \varepsilon_k \leq \pi_k$ for all $k$) that characterizes how much of the amplitude from the initial state will be used for creating the final state (in classical rejection sampling $\varepsilon_k^2$ corresponds to the probability that a specific value of $k$ is drawn from the initial distribution and accepted).
We start in the initial state $\0{d}\0{n}\ket{0}$ and apply the oracle $O$ to prepare $\ket{\vc{\pi}^\xi}$ in the first two registers:
\addtocounter{equation}{1}
\begin{equation}
O \0{dn} \otimes \ket{0}
= \ket{\vc{\pi}^\xi} \ket{0}
= \sum_{k=1}^n \pi_k \ket{\xi_k} \ket{k} \ket{0}.
\label{eq:a0}
\end{equation}
Next, for each $k$ let $\Rot(k)$ be a single-qubit unitary operation defined\footnote{For those $k$, for which $\pi_k = 0$, operation $\Rot(k)$ can be defined arbitrarily.} as follows (this is a rotation by an angle whose sine is equal to $\varepsilon_k/\pi_k$):
\begin{equation}
\Rot(k) := \frac{1}{\pi_k}
\mx{
\sqrt{\abs{\pi_k}^2 - \varepsilon_k^2} &
-\varepsilon_k \\
\varepsilon_k &
\sqrt{\abs{\pi_k}^2 - \varepsilon_k^2}
}.
\label{eq:Rotk}
\end{equation}
Let $R_{\vc{\varepsilon}} := \sum_{k=1}^n \ket{k} \bra{k} \otimes \Rot(k)$ be a block-diagonal matrix that performs rotations by different angles in mutually orthogonal subspaces. Then $I_d \otimes R_{\vc{\varepsilon}}$ corresponds to applying $\Rot(k)$ on the last qubit, controlled on the value of the second register being equal to $k$. This operation transforms state~(\ref{eq:a0}) into
\begin{align}
\ket{\Psi_{\vc{\varepsilon}}}
&:= (I_d \otimes R_{\vc{\varepsilon}}) \cdot \ket{\vc{\pi}^\xi} \ket{0} \nonumber \\
& = \sum_{k=1}^n \ket{\xi_k} \ket{k}
\Bigl( \sqrt{\abs{\pi_k}^2 - \varepsilon_k^2} \, \ket{0} + \varepsilon_k \ket{1} \Bigr).
\label{eq:Psie}
\end{align}
If we would measure the coin register of state $\ket{\Psi_{\vc{\varepsilon}}}$ in the basis $\set{\ket{0}, \ket{1}}$, the probability of outcome $\ket{1}$ (``accept'') and the corresponding post-measurement state would be
\begin{align}
q_{\ve} &:= \norm{\bigl(I_d \otimes I_n \otimes \ket{1} \bra{1}\bigr) \ket{\Psi_{\vc{\varepsilon}}}}_2^2
= \sum_{k=1}^n \varepsilon_k^2
= \norm{\vc{\varepsilon}}_2^2, \label{eq:qe} \\
\ket{\Psi_{\Pi,\vc{\varepsilon}}}
&:= \frac{1}{\norm{\vc{\varepsilon}}_2} \sum_{k=1}^n \varepsilon_k \ket{\xi_k} \ket{k} \ket{1}.\nonumber
\end{align}
Note that if the vector of amplitudes $\vc{\varepsilon}$ is chosen to be close to $\vc{\sigma}$, then the reduced state on the first two registers of $\ket{\Psi_{\Pi,\vc{\varepsilon}}}$ has a large overlap on the target state $\ket{\vc{\sigma}^\xi}$, more precisely,
\begin{equation}
\sqrt{p_{\ve}} := \bigl(\bra{\vc{\sigma}^\xi} \otimes \bra{1} \bigr) \ket{\Psi_{\Pi,\vc{\varepsilon}}}
= \vc{\sigma}^{\mathsf{T}} \cdot \frac{\vc{\varepsilon}}{\norm{\vc{\varepsilon}}_2},
\label{eq:pe}
\end{equation}
Depending on the choice of $\vc{\varepsilon}$, this can be a reasonably good approximation, so our strategy will be to prepare a state close to $\ket{\Psi_{\Pi,\vc{\varepsilon}}}$.
In principle, we could obtain $\ket{\Psi_{\Pi,\vc{\varepsilon}}}$ by repeatedly preparing $\ket{\Psi_{\vc{\varepsilon}}}$ and measuring its coin register until we get the outcome ``accept'' (we would succeed with high probability after $O(1/q_{\ve})$ steps). To speed up this process, we can use amplitude amplification~\cite{AmplitudeAmplification} to amplify the amplitude of the ``accept'' state $\ket{1}$ in the coin register of the state in Eq.~(\ref{eq:Psie}). This allows us to increase the probability of outcome ``accept'' arbitrarily close to $1$ in $O(1/\sqrt{q_{\ve}})$ steps.
\subsection{Amplitude amplification subroutine and\\quantum rejection sampling algorithm}
\begin{figure}[th]
\getpicture{0.20}{fig-Ue}
\caption{Quantum circuit for implementing $U_{\vc{\varepsilon}}$.}
\label{fig:Ue}
\end{figure}
We will use the following amplitude amplification subroutine extensively in all algorithms presented in this paper:
\begin{equation*}
\mc{S}_{\text{QRS}}(\reflection{\ket{\vc{\pi}^\xi}\ket{0}},\vc{\varepsilon},t) :=
\bigl(
\reflection{\ket{\Psi_{\vc{\varepsilon}}}} \cdot
\reflection{I_d \otimes I_n \otimes \ket{1} \bra{1}}
\bigl)^t,
\end{equation*}
where reflections through the two subspaces are defined as follows:
\begin{align*}
\reflection{I_d \otimes I_n \otimes \ket{1} \bra{1}}
&:= I_d \otimes I_n \otimes \bigl( I_2 - 2 \ket{1} \bra{1} \bigr)
= I_d \otimes I_n \otimes Z, \\
\reflection{\ket{\Psi_{\vc{\varepsilon}}}}
&:= (I_d \otimes R_{\vc{\varepsilon}}) \; \reflection{\ket{\vc{\pi}^\xi}\ket{0}} \; (I_d \otimes R_{\vc{\varepsilon}}) ^{\dagger}.
\end{align*}
Depending on the application, we will either be given an oracle $O$ for preparing $\ket{\vc{\pi}^\xi} \ket{0}$ or an oracle $\reflection{\ket{\vc{\pi}^\xi}\ket{0}}$ for reflecting through this state. Note that we can always use the preparation oracle to implement the reflection $\reflection{\ket{\vc{\pi}^\xi}\ket{0}}$ as
\begin{equation}
(O \otimes I_2) \;
\bigl( I_d \otimes I_n \otimes I_2 - 2 \ket{\bar{0},\bar{0},0} \bra{\bar{0},\bar{0},0} \bigl) \;
(O \otimes I_2)^{\dagger}.
\label{eq:refpxi}
\end{equation}
\algorithm{
\textbf{Amplitude amplification subroutine}
$\mc{S}_{\text{QRS}}(\reflection{\ket{\vc{\pi}^\xi}\ket{0}}, \vc{\varepsilon}, t)$
\textbf{for quantum rejection sampling} \\
Perform the following steps $t$ times:
\begin{enumerate}
\item Perform $\reflection{I_d \otimes I_n \otimes \ket{1} \bra{1}}$
by applying Pauli $Z$ on the coin register.
\item Perform $\reflection{\ket{\Psi_{\vc{\varepsilon}}}}$ by applying $R_{\vc{\varepsilon}}^{\dagger}$
on the last two registers, applying $\reflection{\ket{\vc{\pi}^\xi}\ket{0}}$,
and then undoing $R_{\vc{\varepsilon}}$.
\end{enumerate}
}
The quantum rejection sampling algorithm $\mc{A}_{\text{QRS}}(O,\vc{\pi},\vc{\varepsilon})$ starts with the initial state $\0{d} \0{n} \ket{0}$. First, we transform it into $\ket{\Psi_{\vc{\varepsilon}}}$ defined in Eq.~(\ref{eq:Psie}), by applying $U_{\vc{\varepsilon}} := (I_d \otimes R_{\vc{\varepsilon}}) \cdot (O \otimes I_2)$ (see Fig.~\ref{fig:Ue}). Then we apply the amplitude amplification subroutine $\mc{S}_{\text{QRS}}(\reflection{\ket{\vc{\pi}^\xi}\ket{0}}, \vc{\varepsilon}, t)$ with $t = O\bigl(1/\norm{\vc{\varepsilon}}_2\bigr)$. Finally, we measure the last register: if the outcome is $\ket{1}$, we output the first two registers, otherwise we output ``Fail''. To prevent the outcome ``Fail'' we can slightly adjust the angle of rotation in amplitude amplification so that the target state is reached exactly. More precisely, we prove that one can choose $\vc{\varepsilon} = r \cdot \ve_{\vc{\quantump}\to\vc{\quantums}}^p$ for some bounded constant $r$, so that amplitude amplification succeeds with probability $1$ (\textit{i.\,e.}, the outcome of the final measurement is always $\ket{1}$). Moreover, such algorithm is optimal as its cost matches the lower bound in Lemma~\ref{lem:lower-bound}.
\algorithm{
\textbf{Quantum rejection sampling algorithm}
$\mc{A}_{\text{QRS}}(O, \vc{\pi}, \vc{\varepsilon})$
\begin{enumerate}
\item Start in initial state $\0{d} \0{n} \ket{0}$.
\item Apply $U_{\vc{\varepsilon}}$.
\item On the current state apply the amplitude amplification subroutine
$\mc{S}_{\text{QRS}}(\reflection{\ket{\vc{\pi}^\xi}\ket{0}}, \vc{\varepsilon}, t)$
where $\reflection{\ket{\vc{\pi}^\xi}\ket{0}}$
is implemented according to Eq.~(\ref{eq:refpxi}) and
$t = O(1/\norm{\vc{\varepsilon}}_2)$.
\item Measure the last register. If the outcome is $\ket{1}$,
output the first two registers, otherwise output ``Fail''.
\end{enumerate}
}
\begin{lemma}\label{lem:algo-sdp}
For any $p_{\min}\leq p\leq p_{\max}$, there is a constant $r \in [\frac{1}{2},1]$, so that the algorithm $\mc{A}_{\text{QRS}}(O, \vc{\pi}, \vc{\varepsilon})$ with $\vc{\varepsilon} = r \cdot \ve_{\vc{\quantump}\to\vc{\quantums}}^p$ solves $\qsamplingab$ with success probability $p$ using $O \bigl( 1 / \norm{\ve_{\vc{\quantump}\to\vc{\quantums}}^p}_2 \bigr)$ queries to $O$ and $O^{\dagger}$.
\end{lemma}
\begin{proof}
By Def.~\ref{def:water-filling}, we have $0 \leq \varepsilon_k \leq \pi_k$ for all $k$, therefore $\ve_{\vc{\quantump}\to\vc{\quantums}}^p$ is a valid choice of vector $\vc{\varepsilon}$ for the algorithm. Instead of using $\ve_{\vc{\quantump}\to\vc{\quantums}}^p$ itself, we slightly scale it down by a factor $r$ so that the amplitude amplification never fails. Note that if we were to use $\vc{\varepsilon}=\ve_{\vc{\quantump}\to\vc{\quantums}}^p$, the probability that the amplitude amplification succeeds after $t$ steps would be $\sin^2 \bigl( (2t+1) \theta \bigr)$, where $\theta := \arcsin\norm{\ve_{\vc{\quantump}\to\vc{\quantums}}^p}_2$ (see \textit{e.\,g.}~\cite{BoyerBHT98,AmplitudeAmplification} for details). Note that this probability would be equal to one at $t = \frac{\pi}{4\theta}-\frac{1}{2}$, which in general might not be an integer. However, following an idea from~\cite[p.10]{AmplitudeAmplification}, we can ensure this by slightly decreasing $\theta$ to $\tilde{\theta} := \frac{\pi}{2(2\tilde{t}+1)}$ where $\tilde{t} := \lceil\frac{\pi}{4\theta}-\frac{1}{2}\rceil$. This can be done by setting $\vc{\varepsilon} := r \cdot \ve_{\vc{\quantump}\to\vc{\quantums}}^p$ with the scaling-down factor $r := \frac{\sin\tilde{\theta}}{\sin\theta}$. One can check that $r$ satisfies $\frac{1}{2} \leq r \leq 1$ (this follows from $0 \leq \theta \leq \frac{\pi}{2}$).
Together with Eq.~(\ref{eq:pe}), Def.~\ref{def:water-filling} also implies that for this choice, the algorithm solves $\qsamplingab$ with success probability $\frac{\vc{\sigma}^{\mathsf{T}} \cdot \vc{\varepsilon}}{\norm{\vc{\varepsilon}}_2} = \frac{\vc{\sigma}^{\mathsf{T}} \cdot \ve_{\vc{\quantump}\to\vc{\quantums}}^p}{\norm{\ve_{\vc{\quantump}\to\vc{\quantums}}^p}_2} = \sqrt{p}$. It therefore remains to prove that the cost of the algorithm is $O(1/\norm{\vc{\varepsilon}}_2)$, which follows immediately from its description: we need one query to implement the operation $U_{\vc{\varepsilon}}$ and two queries to implement $\reflection{\ket{\vc{\pi}^\xi}\ket{0}}$, thus in total we need $2t+1 = O\bigl(1/\sqrt{q_{\ve}}\bigr) = O(1/\norm{\vc{\varepsilon}}_2)$ calls to oracles $O$ and $O^{\dagger}$.
\qed
\end{proof}
We now have all the elements to prove Theorem~\ref{thm:qsampling}.
\thmqsampling*
\begin{proof
When $p\leq p_{\min}$, the state $\ket{\vc{\pi}^\xi}$ is already close enough to $\ket{\vc{\sigma}^\xi}$ to satisfy the constraint on the success probability, therefore one call to $O$ is sufficient, which is clearly optimal. When $p> p_{\max}$, the oracle gives no information about some of the unknown states $\ket{\xi_k}$ (when $\pi_k=0$), but the target state should have some overlap on $\ket{\xi_k}\ket{k}$ to satisfy the constraint on the success probability, therefore the problem is not solvable.
For the general case $p_{\min}\leq p\leq p_{\max}$, the upper bound comes from the algorithm in Lemma~\ref{lem:algo-sdp}, and the matching lower bound is given in Lemma~\ref{lem:lower-bound}.
\qed
\end{proof}
\subsection{Strong quantum rejection sampling algorithm}
Let us now describe how the algorithm can be modified to solve the stronger problem $\sqsampling_{\vc{\tau}}$. The first modification follows from the observation that in the previous algorithm, the oracle is only used in two different ways: it is used once to create the state $\ket{\vc{\pi}^\xi}$, and then only to reflect through that state. This means that we can still solve the problem if, instead of being given access to an oracle that maps $\0{dn}$ to $\ket{\vc{\pi}^\xi}$, we are provided with one copy of $\ket{\vc{\pi}^\xi}$ and an oracle that reflects through it.
In order to solve $\sqsampling_{\vc{\tau}}$, we should also be able to deal with the case where we only know the ratios between the amplitudes $\pi_k$ and $\sigma_k$ for each $k$, instead of the amplitudes themselves. As we will see, in that case we cannot use the algorithm given above anymore, as we do not know in advance how many steps of amplitude amplification are required. There are different approaches to solve this issue, one of them being to estimate $q_{\ve}$, and therefore the required number of steps, by performing a phase estimation on the amplitude amplification operator (this is sometimes referred to as amplitude estimation or quantum counting, see~\cite{BoyerBHT98,AmplitudeAmplification}). Another option, also proposed by~\cite{BoyerBHT98,AmplitudeAmplification}, is to repeat the algorithm successively with an increasing number of steps until it succeeds. One advantage of the first option would be that it provides an estimation of the initial acceptance probability $q_{\ve}$, which might be useful for some applications. Since this is not required for $\sqsampling_{\vc{\tau}}$, we will rather describe an algorithm based on the second option, as it is more direct. Note that for both options, we need to adapt the algorithms in~\cite{BoyerBHT98,AmplitudeAmplification} as they require to use a fresh copy of the initial state after each failed attempt, whereas for $\sqsampling_{\vc{\tau}}$ we only have one copy of that state. This issue can be solved by using the state left over from the previous unsuccessful measurement instead of a fresh copy of the state. More precisely, we can use the following algorithm.
\algorithm{
\textbf{Strong quantum rejection sampling algorithm}\\
$\mc{A}_{\text{SQRS}}(\ket{\vc{\pi}^\xi},\vc{\varepsilon},c)$
\begin{enumerate}
\item\label{item:sqrs-initialization}
Append an extra qubit prepared in the state $\ket{0}$ to the
input state $\ket{\vc{\pi}^\xi}$, and apply $I_d \otimes R_{\vc{\varepsilon}}$ on the resulting state.
\item\label{item:sqrs-measurement-1}
Measure the last register of the current state. If the outcome
is $\ket{1}$, output the first two registers and stop. Otherwise,
set $l=0$ and continue.
\item\label{item:sqrs-new-attempt}
Let $T_l:=\lceil c^l\rceil$ and pick an integer $t \in [T_l]$
uniformly at random.
\item\label{item:sqrs-qrs}
On the current state apply the amplitude amplification subroutine
$\mc{S}_{\text{QRS}}(\reflection{\ket{\vc{\pi}^\xi}\ket{0}}, \vc{\varepsilon}, t)$.
\item\label{item:sqrs-measurement-2}
Measure the last register of the current state. If the outcome
is $\ket{1}$, output the first two registers and stop. Otherwise,
increase $l$ by one and go back to Step~\ref{item:sqrs-new-attempt}.
\end{enumerate}
}
\begin{lemma}\label{lem:algo-sqsqmpling}
For any $\alpha \geq 1$, there is a quantum algorithm that solves $\sqsampling_{\vc{\tau}}$ with success probability $p(\gamma)$ using an expected number of queries $O(1/\norm{\vc{\varepsilon}(\gamma)}_2)$, where $\gamma = \alpha \norm{\vc{\pi} \circ \vc{\tau}}_2$. In particular, for $\alpha = 1$ the expected number of queries is $O(1/\norm{\vc{\pi} \circ \vc{\tau}}_2)$ and the success probability is equal to one.
\end{lemma}
Here the parameter $\alpha$ allows us to control the trade-off between the success probability and the required number of queries. However, we cannot predict the actual values of both quantities, because they depend on $\vc{\pi}$ and $\vc{\tau}$, but only $\vc{\tau}$ is known to us. The only exception is $\alpha=1$, when the success probability is equal to one. Also, increasing $\alpha$ above $1/(\min_{k:\tau_k > 0} \tau_k)$ will no longer affect the query complexity and success probability of the algorithm.
\begin{proof}
We show that for some choice of $c > 1$ and $\vc{\varepsilon}$ the algorithm $\mc{A}_{\text{SQRS}}(\ket{\vc{\pi}^\xi},\vc{\varepsilon},c)$ described above solves the problem.
Let us first verify that we can actually perform all steps required in the algorithm. We need one copy of the state $\ket{\vc{\pi}^\xi}$, which is indeed provided as an input for the $\sqsampling_{\vc{\tau}}$ problem. Note that for applying $R_{\vc{\varepsilon}}$ in Step~\ref{item:sqrs-initialization} it suffices to know only the ratio $\varepsilon_k/\pi_k$ (see Eq.~(\ref{eq:Rotk})), which can be deduced from $\tau_k$ as follows. Let $\vc{\varepsilon} := r \cdot \vc{\varepsilon}(\gamma)$ for some $r < 1$, and recall from Def.~\ref{def:water-filling} and~\ref{def:SQSampling} that $\varepsilon_k(\gamma) = \min \set{\pi_k, \gamma \sigma_k}$ and $\sigma_k = \pi_k \tau_k / \norm{\vc{\pi} \circ \vc{\tau}}_2$, respectively. Then
\begin{align}
\frac{\varepsilon_k}{\pi_k}
&= r \min \set{1, \gamma \frac{\sigma_k}{\pi_k}} \nonumber \\
&= r \min \set{1, \gamma \frac{\tau_k}{\norm{\vc{\pi} \circ \vc{\tau}}_2}} \nonumber \\
&= r \min \set{1, \alpha \tau_k},
\label{eq:r}
\end{align}
where we substituted $\gamma = \alpha \norm{\vc{\pi} \circ \vc{\tau}}_2$ from the statement of the Lemma. Note that once $r$ is chosen, the final expression in Eq.~(\ref{eq:r}) is completely known. Finally, applying $\mc{S}_{\text{QRS}}(\reflection{\ket{\vc{\pi}^\xi}\ket{0}}, \vc{\varepsilon}, t)$ in Step~\ref{item:sqrs-qrs} also requires the ability to apply $R_{\vc{\varepsilon}}$, as well as $\reflection{\ket{\vc{\pi}^\xi}\ket{0}}$, which can be done by using one oracle query. Therefore, we have all we need to implement the algorithm.
We now show that the algorithm has success probability $p(\gamma)$. Recall from Eq.~(\ref{eq:Psie}) that Step~\ref{item:sqrs-initialization} of the algorithm prepares the state
\begin{align*}
\ket{\Psi_{\vc{\varepsilon}}}
&= \sum_{k=1}^n \ket{\xi_k} \ket{k}
\Bigl( \sqrt{\abs{\pi_k}^2 - \varepsilon_k^2} \, \ket{0} + \varepsilon_k \ket{1} \Bigr) \\
&= \sin \theta \ket{\Psi_{\Pi,\vc{\varepsilon}}} + \cos \theta \ket{\Psi_{\Pi,\vc{\varepsilon}}^\perp},
\end{align*}
where $\theta := \arcsin{\norm{\vc{\varepsilon}}_2}$ and unit vectors
\begin{align*}
\ket{\Psi_{\Pi,\vc{\varepsilon}}}
&:= \frac{1}{\sin \theta}
\sum_{k=1}^n \varepsilon_k \ket{\xi_k} \ket{k} \ket{1}, \\
\ket{\Psi_{\Pi,\vc{\varepsilon}}^\perp}
&:= \frac{1}{\cos \theta}
\sum_{k=1}^n \sqrt{\abs{\pi_k}^2 - \varepsilon_k^2} \ket{\xi_k} \ket{k} \ket{0}
\end{align*}
are orthogonal and span a $2$-dimensional subspace. In this subspace $\reflection{I_d \otimes I_n \otimes \ket{1} \bra{1}}$ and $\reflection{\ket{\Psi_{\Pi,\vc{\varepsilon}}}}$ act in the same way, so each iteration of the amplitude amplification subroutine consists of a product of two reflections that preserve this subspace, and $\mc{S}_{\text{QRS}}(\reflection{\ket{\vc{\pi}^\xi}\ket{0}}, \vc{\varepsilon}, t)$ corresponds to a rotation by angle $2 t \theta$. Measurements in Steps~\ref{item:sqrs-measurement-1} and~\ref{item:sqrs-measurement-2} either project on $\ket{\Psi_{\Pi,\vc{\varepsilon}}}$, when the outcome is $\ket{1}$, or on $\ket{\Psi_{\Pi,\vc{\varepsilon}}^\perp}$, when the outcome is $\ket{0}$. Therefore, the algorithm always outputs the first two registers of the state $\ket{\Psi_{\Pi,\vc{\varepsilon}}}$, and by Eq.~(\ref{eq:pe}), the success probability is $p_{\ve} = p(\gamma)$, as claimed. In particular, for $\alpha = 1$ from Eq.~(\ref{eq:r}) we get $\varepsilon_k = r \pi_k \tau_k$ as $\tau_k \leq 1$. Thus, $\vc{\varepsilon} = r (\vc{\pi} \circ \vc{\tau})$ and since $\vc{\sigma} = (\vc{\pi} \circ \vc{\tau}) / \norm{\vc{\pi} \circ \vc{\tau}}_2$, we get $p_{\ve} = (\vc{\sigma}^{\mathsf{T}} \cdot \vc{\varepsilon}/\norm{\vc{\varepsilon}}_2)^2 = 1$.
Let us now bound the expected number of oracle queries. We follow the proof of Theorem 3 in~\cite{AmplitudeAmplification}, but there is an important difference: a direct analogue of the algorithm in~\cite[Theorem 3]{AmplitudeAmplification} would use a fresh copy of $\ket{\vc{\pi}^\xi}$ each time the measurement fails to give a successful outcome, whereas in this algorithm we start from the state left over from the previous measurement, since we only have one copy of $\ket{\vc{\pi}^\xi}$. Note that $\mc{S}_{\text{QRS}}(\reflection{\ket{\vc{\pi}^\xi}\ket{0}}, \vc{\varepsilon}, t)$ in Step~\ref{item:sqrs-qrs} is always applied on $\ket{\Psi_{\Pi,\vc{\varepsilon}}^\perp}$, since it is the post-measurement state corresponding to the unsuccessful outcome. Therefore, the state created by Step~\ref{item:sqrs-qrs} is $\sin (2t\theta) \ket{\Psi_{\Pi,\vc{\varepsilon}}} + \cos(2t\theta) \ket{\Psi_{\Pi,\vc{\varepsilon}}^\perp}$, and the next measurement will succeed with probability $\sin^2(2t\theta)$. Since $t$ is picked uniformly at random between $1$ and $T_l$, the probability that the $l$-th measurement fails is
\begin{align}
p_l &= \frac{1}{T_l} \sum_{t=1}^{T_l} \cos^2 (2 t \theta) \nonumber \\
&= \frac{1}{2} + \frac{1}{2 T_l} \sum_{t=1}^{T_l} \cos (4 t \theta) \nonumber \\
&\leq \frac{1}{2} + \frac{1}{2 T_l \norm{\vc{\varepsilon}}_2},
\label{eq:pl}
\end{align}
where the upper bound is obtained as follows:
\begin{align*}
\sum_{t=1}^T \cos (4t\theta)
&= \Re \biggl( e^{i 4 \theta} \sum_{t=0}^{T - 1} e^{i 4 t \theta} \biggr) \\
&= \Re \biggl( e^{i 4 \theta} \cdot \frac{1 - e^{i 4 T \theta}}{1 - e^{i 4 \theta}} \biggr) \\
&= \Re \biggl( e^{i 2(T+1)\theta} \cdot
\frac{e^{-i 2 T \theta} - e^{i 2 T \theta}}
{e^{-i 2 \theta} - e^{i 2 \theta}} \biggr) \\
&= \cos \bigl( 2(T+1)\theta \bigr)
\frac{\sin(2T\theta)}{\sin(2\theta)} \\
&\leq \frac{1}{\sin(2\theta)}
\leq \frac{1}{\sin \theta}
= \frac{1}{\norm{\vc{\varepsilon}}_2},
\end{align*}
where we forced the last inequality by picking $r := \sqrt{3}/2$, so that $\sin \theta = \norm{\vc{\varepsilon}}_2 \leq \sqrt{3}/2$ and thus $0 \leq \theta \leq \pi/3$. Recall from the algorithm that $T_l = \lceil c^l \rceil$ for some $c > 1$, so it is increasing and goes to infinity as $l$ increases. Let $\bar{T} := 1/(2 \Delta \norm{\vc{\varepsilon}}_2)$ for some $\Delta > 0$ and let $\bar{l}$ be the smallest integer such that $T_l \geq \bar{T}$ for all $l \geq \bar{l}$. Then according to Eq.~(\ref{eq:pl}) we get that $p_l \leq 1/2 + 1/(2 \bar{T} \norm{\vc{\varepsilon}}_2) = 1/2 + \Delta =: \bar{p}$ for all $l \geq \bar{l}$. Note that the $l$-th execution of the subroutine uses at most $2T_l$ oracle queries, so the expected number of oracle calls is at most $2T_0 + p_0 (2T_1 + p_1 (2T_2 + \ldots))$. This can be upper bounded by
\begin{align}
\sum_{l=0}^{\bar{l}} 2T_l + \sum_{d=1}^{\infty} 2T_{\bar{l}+d} \bar{p}^d
&= \sum_{l=0}^{\bar{l}} 2 \lceil c^l \rceil
+ \sum_{d=1}^{\infty} 2 \lceil c^{\bar{l}+d} \rceil \bar{p}^d \nonumber \\
&\leq 4 \Biggl(
\sum_{l=0}^{\bar{l}} c^l
+ c^{\bar{l}} \sum_{d=1}^{\infty} (c \bar{p})^d
\Biggr) \label{eq:expqueries} \\
& = 4 \Biggl(
\frac{c^{\bar{l}+1} - 1}{c - 1}
+ c^{\bar{l}} \frac{c \bar{p}}{1 - c \bar{p}}
\Biggr) \nonumber \\
&\leq 4 c^{\bar{l}+1}
\Biggl(
\frac{1}{c - 1}
+ \frac{\bar{p}}{1 - c \bar{p}}
\Biggr) \nonumber \\
&\leq \frac{2 c^2}{\Delta \norm{\vc{\varepsilon}}_2}
\Biggl(
\frac{1}{c - 1}
+ \frac{\bar{p}}{1 - c \bar{p}}
\Biggr), \label{eq:expqueriesbound}
\end{align}
where the first and last inequality is obtained from the following two observations, respectively:
\begin{enumerate}
\item $\lceil c^l \rceil = c^l + \delta$ for some $0 \leq \delta < 1$, so $T_l = \lceil c^l \rceil < c^l + 1 < 2c^l$ as $c > 1$.
\item $c^{\bar{l}+1} \leq c^2 \lceil c^{\bar{l}-1} \rceil = c^2 T_{\bar{l}-1} < c^2 \bar{T} = c^2 / (2 \Delta \norm{\vc{\varepsilon}}_2)$ by the choice of $\bar{l}$.
\end{enumerate}
Finally, we have to make a choice of $c > 1$ and $\Delta > 0$, so that the geometric series in Eq.~(\ref{eq:expqueries}) converges, \textit{i.\,e.}, $c \bar{p} < 1$ or equivalently $c < 2/(1+2\Delta)$. By choosing $c := 8/7$ and $\Delta := 1/4$ we minimize the upper bound in Eq.~(\ref{eq:expqueriesbound}) and obtain $128/\norm{\vc{\varepsilon}}_2 = O(1/\norm{\vc{\varepsilon}(\gamma)}_2)$. In particular, for $\alpha = 1$ this becomes $O(1/\norm{\vc{\pi} \circ \vc{\tau}}_2)$.
\qed
\end{proof}
\section{Applications} \label{sect:applications}
\subsection{Linear systems of equations} \label{sect:linear}
As a first example of application, we show that quantum rejection sampling was implicitly used in the quantum algorithm for linear systems of equations proposed by Harrow, Hassidim, and Lloyd~\cite{HHL:09}. This algorithm solves the following quantum state generation problem: given the classical description of an invertible $d \times d$ matrix $A$ and a unit vector $\ket{b} \in \mathbb{C}^d$, prepare the quantum state $\ket{x}/\norm{\ket{x}}_2$, where $\ket{x}$ is the solution of the linear system of equations $A \ket{x} = \ket{b}$. As shown in~\cite{HHL:09}, we can assume without loss of generality that $A$ is Hermitian. Similarly to classical matrix inversion algorithms, an important factor of the performance of the algorithm is the condition number $\kappa$ of $A$, which is the ratio between the largest and smallest eigenvalue of $A$. We will assume that all eigenvalues of $A$ are between $\kappa^{-1}$ and $1$, and we denote by $\ket{\psi_j}$ and $\lambda_j$ the eigenvectors and eigenvalues of $A$, respectively. We also define\footnote{We choose the global phase of each eigenvector $\ket{\psi_j}$ so that $b_j$ is real and non-negative.} the amplitudes $b_j := \braket{\psi_j}{b}$, so that $\ket{b} = \sum_{j=1}^d b_j \ket{\psi_j}$. Then, the problem is to prepare the state $\ket{x} := A^{-1}\ket{b} = \sum_{j=1}^d b_j \lambda_j^{-1} \ket{\psi_j}$ (up to normalization).
We now show how this problem reduces to the quantum state conversion problem $\sqsampling_{\vc{\tau}}$. Since $A$ is Hermitian, we can use Hamiltonian simulation techniques~\cite{Berry2006,Childs2009a,Childs2011} to simulate the unitary operator $e^{iAt}$ on any state. Using quantum phase estimation~\cite{Kitaev1995,CEMM98} on the operator $e^{iAt}$, we can implement an operator $E_A$ that acts in the eigenbasis of $A$ as $E_A: \ket{\psi_j} \0{} \mapsto \ket{\psi_j} \ket{\lambda_j}$, where $\ket{\lambda_j}$ is a quantum state encoding an approximation of the eigenvalue $\lambda_j$. Here, we will assume that this can be done exactly, that is, we assume that $\ket{\lambda_j}$ is a computational basis state that exactly encodes $\lambda_j$ (we refer the reader to~\cite{HHL:09} for a detailed analysis of the error introduced by this approximation). Under this assumption, the problem reduces to a quantum state conversion problem that we will call $\qle$. Its definition requires fixing a set of possible eigenvalues $\Lambda_{\kappa} \subset [\kappa^{-1},1]$ of finite cardinality $n := \abs{\Lambda_{\kappa}}$. Let us denote the set of $d \times d$ Hermitian matrices by $\Herm(d)$ and the eigenvalues of $A$ by $\spec(A)$.
\begin{definition}\deftitle{Quantum linear system of equations}
$\qle$, the \emph{quantum linear system of equations problem} is a quantum state conversion problem $(\O, \Phi, \Psi, \mathcal{X})$ with $\mathcal{X} := \set{(\ket{b},A) \in \mathbb{C}^d \times \Herm(d): \spec(A) \in \Lambda_{\kappa}^d}$, oracles in $\O$ being pairs $(O_{\ket{b},A}, E_A)$ where $O_{\ket{b},A} := \reflection{\ket{b}}$ and $E_A$ acts as $E_A: \ket{\psi_j} \0{n} \mapsto \ket{\psi_j} \ket{\lambda_j}$ where $\ket{\psi_j}$ are the eigenvectors of $A$, and the corresponding initial and target states being $\ket{b}$ and $A^{-1}\ket{b} / \norm{A^{-1}\ket{b}}_2$.
\label{def:qlse}
\end{definition}
Using Lemma~\ref{lem:algo-sqsqmpling} we can prove the following result.
\thmlinearequations*
\begin{proof}
Following~\cite{HHL:09}, the algorithm for this problem consists of three steps:
\begin{enumerate}
\item Apply the phase estimation operation $E_A$ on $\ket{b}$
to obtain the state $\sum_{j=1}^d b_j \ket{\psi_j} \ket{\lambda_j}$.
\item\label{step:lse-qrs} Convert this state to
$\sum_{j=1}^d w_j \ket{\psi_j} \ket{\lambda_j} / \norm{\vc{w}}_2$.
\item Undo the phase estimation operation $E_A$ to obtain the target state
$\sum_{j=1}^d w_j \ket{\psi_j} / \norm{\vc{w}}_2$.
\end{enumerate}
We see that Step~\ref{step:lse-qrs} is an instance of $\sqsampling_{\vc{\tau}}$, where the basis states $\set{\ket{\lambda} : \lambda \in \Lambda_{\kappa}}$ of the phase estimation register play the role of the states $\ket{k}$ and the vector $\vc{\tau}$ of the ratios between the initial and final amplitudes is given by $\tau_{\lambda} := (\kappa \lambda)^{-1}$ (here the normalization factor $\kappa$ is to make sure that $\max_{\lambda} \tau_{\lambda} = 1$). The rest of the reduction is summarized in Table~\ref{table:reduction-lse}. Therefore, we can use the algorithm from Lemma~\ref{lem:algo-sqsqmpling} to perform Step~\ref{step:lse-qrs}.
\begin{table}
\begin{center}
\begin{tabular}{|c|c|}
\hline
$\sqsampling_{\vc{\tau}}$ & $\qle$ \\
\hline
$\ket{k}$ & $\ket{\lambda}$ \\
$\pi_k$ & $\pi_\lambda :=
\begin{cases}
b_j \textrm{ if } \lambda = \lambda_j \in \spec(A) \\
0 \textrm{ if } \lambda \notin \spec(A)
\end{cases}$ \\
$\ket{\xi_k}$ & $\ket{\xi_\lambda} :=
\begin{cases}
\ket{\psi_j} \textrm{ if } \lambda = \lambda_j \in \spec(A) \\
\textrm{n/a if } \lambda \notin \spec(A)
\end{cases}$ \\
$\tau_k$ & $\tau_\lambda := (\kappa\lambda)^{-1}$ \\
\hline
\end{tabular}
\end{center}
\caption{Reduction from Step~\ref{step:lse-qrs} in the linear system of equations algorithm to the quantum resampling problem \textnormal{$\sqsampling_{\vc{\tau}}$}.}
\label{table:reduction-lse}
\end{table}
If we set $\alpha := \kappa / \tilde{\kappa}$ then from Table~\ref{table:reduction-lse} we get
\begin{align*}
\varepsilon_{\lambda_j}(\gamma)
&= \pi_{\lambda_j} \min \set{1, \alpha \tau_{\lambda_j}} \\
&= b_j \min \set{1, (\tilde{\kappa} \lambda_j)^{-1}} \\
&= \frac{b_j}{\tilde{\kappa} \max \set{\tilde{\kappa}^{-1}, \lambda_j}} \\
&= \frac{\tilde{w}_j}{\tilde{\kappa}},
\end{align*}
thus $\vc{\varepsilon}(\gamma) = \tilde{\vc{w}} / \tilde{\kappa}$ and the expected number of queries is $O(1/\norm{\vc{\varepsilon}(\gamma)}_2) = O(\tilde{\kappa} / \norm{\tilde{\vc{w}}}_2)$. Recall that the amplitudes of the target state are given by $\sigma_j = w_j / \norm{\vc{w}}_2$, so $\sigma = \vc{w} / \norm{\vc{w}}_2$ and the success probability is
\begin{equation*}
p = \biggl( \frac{\vc{\sigma}^{\mathsf{T}} \cdot \vc{\varepsilon}(\gamma)}{\norm{\vc{\varepsilon}(\gamma)}_2} \biggr)^2
= \frac{\vc{w}^{\mathsf{T}}}{\norm{\vc{w}}_2} \cdot
\frac{\tilde{\vc{w}}\;}{\norm{\tilde{\vc{w}}}_2}
\end{equation*}
as claimed.
\qed
\end{proof}
Note that even though we have a freedom to choose $\tilde{\kappa}$, we cannot predict the query complexity in advance, since it depends on $\tilde{w}_j = \braket{\psi_j}{b} / \tilde{\lambda}_j$, which in turn is determined by the lengths of projections of $\ket{b}$ in the eigenspaces of $A$, weighted by the corresponding truncated eigenvalues $\tilde{\lambda}_j$. Similarly, we cannot predict the success probability~$p$. However, by choosing $\tilde{\kappa} = \kappa$ we can at least make sure that $p = 1$ (since $\tilde{\lambda}_j = \lambda_j$ and thus $\tilde{\vc{w}} = \vc{w}$). In this case Step~\ref{step:lse-qrs} is performed exactly (assuming an ideal phase estimation black box) and the expected number of queries is $O(\kappa / \norm{\vc{w}}_2)$. By noting that $\lambda_j \leq 1$ for all $j$, we see that $\norm{\vc{w}}_2^2 = \sum_{j=1}^d b_j^2 \lambda_j^{-2} \geq 1$ and thus we can put a cruder upper bound of $O(\kappa)$, which coincides with the bound given in~\cite{HHL:09} for that step of the algorithm. For ill-conditioned matrices, \textit{i.\,e.}, matrices with a high condition number $\kappa$, the approach taken by~\cite{HHL:09} is to ignore small eigenvalues $\lambda_j\leq\tilde{\kappa}^{-1}$, for some cut-off $\tilde{\kappa}^{-1} \geq \kappa^{-1}$, which reduces the cost of the algorithm to $O(\tilde{\kappa})$, but introduces some extra error. In our case, by choosing $\alpha = \kappa / \tilde{\kappa}$ we obtained bound $O(\tilde{\kappa} / \norm{\tilde{\vc{w}}}_2)$, where $\norm{\tilde{\vc{w}}}_2 \geq 1$. Again, here $\tilde{\vc{w}}$ depends on additional structure of the problem and cannot be predicted beforehand.
In practical applications, we will of course not be given access to the ideal phase estimation operator $E_A$, but we can still approximate it by using the phase estimation algorithm~\cite{Kitaev1995,CEMM98} on the operator $A$. It is shown in~\cite{HHL:09} that if $A$ is $s$-sparse, this approximation can be implemented with sufficient accuracy at a cost $\tilde{O} \bigl( \log(d) s^2 \tilde{\kappa} / \epsilon \bigr)$, where $\epsilon$ is the overall additive error introduced by this approximation throughout the algorithm. Therefore, the total cost of the algorithm is at most $\tilde{O} \bigl( \log(d) s^2 \tilde{\kappa}^2 / \epsilon \bigr)$ (see~\cite{HHL:09} for details).
\subsection{Quantum Metropolis sampling} \label{sect:metropolis}
Since rejection sampling lies at the core of the (classical) Metropolis algorithm, it seems natural to use quantum rejection sampling to solve the corresponding problem in the quantum case. The quantum Metropolis sampling algorithm presented in~\cite{TOV+:2009} follows the same lines as the classical algorithm by setting up a (classical) random walk between eigenstates of the Hamiltonian, where each move is either accepted or rejected depending on the value of some random coin. The main complication compared to the classical version comes from the case where the move has to be rejected, since we cannot keep a copy of the previous eigenstate due to the no-cloning theorem. The solution proposed by Temme~\textit{et al.}~\cite{TOV+:2009} is to use an unwinding technique based on successive measurements to revert to the original state. Here, we show that quantum rejection sampling can be used to avoid this step, as it allows to amplify the amplitude of the ``accept'' state of the coin register, effectively eliminating rejected moves. This yields a more efficient algorithm as it eliminates the cost of reverting rejected moves and provides a quadratic speed-up on the overall cost of obtaining an accepted move.\footnote{Martin Schwarz has pointed out to us that this is similar to how~\cite{NWZ09} provides a speed-up over~\cite{MW05}, and that our technique can also be used to speed-up the quantum algorithm in~\cite{STV11} for preparing PEPS.}
Before describing in more details how quantum rejection sampling can be used to design a new quantum Metropolis algorithm, let us recall how the standard (classical) Metropolis algorithm works~\cite{metropolis53}. The goal is to solve the following problem: given a classical Hamiltonian associating energies $E_{j}$ to a set of possible configurations $j$, sample from the Gibbs distribution $p(j)=\exp(-\beta E_j)/Z(\beta)$, where $\beta$ is the inverse temperature and $Z(\beta)=\sum_j\exp(-\beta E_j)$ is the partition function. Since the size of the configuration space is exponential in the number of particles, estimating the Gibbs distribution itself is not an option, therefore the Metropolis algorithm proposes to solve this problem by setting up a random walk on the set of configurations that converges to the Gibbs distribution. More precisely, the random walk works as follows:
\begin{enumerate}
\item If $i$ is the current configuration with energy $E_i$, choose a random move to another configuration $j$ (\textit{e.\,g.}, for a system of spins, a random move could consist in flipping a random spin), and compute the associated energy $E_{j}$.
\item The random move is then accepted or rejected according to the following rule:
\begin{itemize}
\item if $E_j \leq E_i$, then the move is always accepted;
\item if $E_j > E_i$, then the move is only accepted
with probability $\exp \bigl( \beta(E_i - E_j) \bigr)$.
\end{itemize}
\end{enumerate}
It can be shown that this random walk converges to the Gibbs distribution.
The quantum Metropolis sampling algorithm by Temme~\textit{et al.}~\cite{TOV+:2009} follows the same general lines as the classical algorithm. It aims at solving the equivalent problem in the quantum case, where we need to generate the thermal state of a Hamiltonian $H$, that is, we need to generate a random eigenstate $\ket{\psi_j}$ where $j$ is sampled according to the Gibbs distribution. The fact that the Hamiltonian is quantum, however, adds a few obstacles, since the set of eigenstates $\ket{\psi_j}$ is not known to start with. The main tool to overcome this difficulty is to use quantum phase estimation~\cite{Kitaev1995,CEMM98} which, applied on the Hamiltonian $H$, allows to project any state on an eigenstate $\ket{\psi_j}$, while obtaining an estimate of the corresponding eigenenergy $E_j$. Similarly to the previous section, we will assume for simplicity that this can be done exactly, that is, we have access to a quantum circuit that acts in the eigenbasis of $H$ as $E_H: \ket{\psi_j} \0{} \mapsto \ket{\psi_j} \ket{E_j}$, where $\ket{E_j}$ exactly encodes the eigenenergy $E_j$. We will also assume that the eigenenergies of $H$ are nondegenerate, so that each eigenenergy $E_j$ corresponds to a single eigenstate $\ket{\psi_j}$, instead of a higher dimensional eigenspace. The quantum Metropolis sampling algorithm also requires to choose a set of quantum gates $\mathcal{C}$ that will play the role of the possible random moves between eigenstates. In this case, a given quantum gate $C_l\in\mathcal{C}$ will not simply move an initial eigenstate $\ket{\psi_i}$ to another eigenstate $\ket{\psi_j}$, but rather to a superposition $C_l \ket{\psi_i} = \sum_j c_{ij}^{(l)} \ket{\psi_j}$ where $c_{ij}^{(l)} := \bra{\psi_j} C_l \ket{\psi_i}$.
We can now give a high-level description of the quantum Metropolis sampling algorithm by Temme~\textit{et al.}~\cite{TOV+:2009}. Let $\ket{\psi_i}\ket{E_i}$ be an initial state, that can be prepared by applying the phase estimation operator $E_H$ on an arbitrary state, and measuring the energy register. The algorithm implements each random move by performing the following steps:
\begin{enumerate}
\item\label{step:qms-random-move}
Apply a random gate $C_l \in \mathcal{C}$ on the first register to prepare the state
$\bigl( C_l \ket{\psi_i} \bigr) \ket{E_i} = \sum_j c_{ij}^{(l)} \ket{\psi_j} \ket{E_i}$.
\item Apply the phase estimation operator $E_H$ on the $\ket{\psi_j}$ register
and an ancilla register initialized in the default state $\0{}$ to prepare the state
$\sum_j c_{ij}^{(l)} \ket{\psi_j} \ket{E_i} \ket{E_j}$.
\item\label{step:qms-coin}
Add another ancilla qubit prepared in the state $\ket{0}$ and apply a
controlled-rotation on this register to create the state
$\sum_j c_{ij}^{(l)} \ket{\psi_j} \ket{E_i} \ket{E_j}
\left[ \sqrt{f_{ij}} \ket{1} + \sqrt{1-f_{ij}} \ket{0} \right]$,
where $f_{ij} := \min \set{1, \exp(\beta(E_i-E_j))}$.
\item\label{step:qms-reject}
Measure the last qubit. If the outcome is $0$, reject the move by reverting the
state to $\ket{\psi_i} \ket{E_i}$ (see~\cite{TOV+:2009} for details) and go
back to Step~\ref{step:qms-random-move}. Otherwise, continue.
\item\label{step:qms-end}
Discard the $\ket{E_i}$ register and measure the $\ket{E_j}$ register
to project the state onto a new eigenstate $\ket{\psi_j} \ket{E_j}$.
\end{enumerate}
It is shown in~\cite{TOV+:2009} that by choosing a universal set of quantum gates for the set of moves $\mathcal{C}$, the algorithm simulates random walk on the set of eigenstates of $H$ that satisfies a quantum detailed balanced condition, which ensures that the walk converges to the Gibbs distribution, as in the classical case.
For a given initial state $\ket{\psi_i}\ket{E_i}$, the probability (over all choices of the randomly chosen gate $C_l$) that the measurement in Step~\ref{step:qms-reject} succeeds is $\frac{1}{\abs{\mathcal{C}}} \sum_{j,l} f_{ij} \abs{c_{ij}^{(l)}}^2$. If we define a vector $\vc{w}^{(i)}$ whose components are $w_{jl}^{(i)} := \sqrt{f_{ij}/\abs{\mathcal{C}}} c_{ij}^{(l)}$, then this probability is simply $\|\vc{w}^{(i)}\|_2^2$. Hence, after one execution of the algorithm (Steps~\ref{step:qms-random-move}-\ref{step:qms-end}) the initial state $\ket{\psi_i} \ket{E_i}$ gets mapped to $\ket{\psi_j}\ket{E_j}$ with probability $\sum_l \abs{w_{jl}^{(i)}}^2 / \|\vc{w}^{(i)}\|_2^2$. We could achieve the same random move by converting the initial state $\ket{\psi_i}\ket{E_i}$ to
\begin{multline}
\sum_{j,l} \frac{w_{jl}^{(i)}}{\|\vc{w}^{(i)}\|_2} \ket{l} \ket{\psi_j} \ket{E_i} \ket{E_j} \\
= \frac{1}{\|\vc{w}^{(i)}\|_2} \sum_j \sqrt{\frac{f_{ij}}{\abs{\mathcal{C}}}}
\biggl[ \sum_l c_{ij}^{(l)} \ket{l} \biggr] \ket{\psi_j} \ket{E_i} \ket{E_j}
\label{eq:target state}
\end{multline}
and discarding the $\ket{l}$ and $\ket{E_i}$ registers and measuring the $\ket{E_j}$ register to project on the state $\ket{\psi_j}\ket{E_j}$ with the correct probability. This implies that one random move reduces to a quantum state conversion problem that we will call $\qmm$. This problem assumes that we are able to perform a perfect phase estimation on the Hamiltonian $H$. Therefore, similarly to the previous section, we fix a set of possible eigenenergies $\mathcal{E}$ of finite cardinality $n := \abs{\mathcal{E}}$.
\begin{definition}\deftitle{Quantum Metropolis move}
The \emph{quantum Metropolis move problem}, denoted by $\qmm$, is a quantum state conversion problem $(\O, \Phi, \Psi, \mathcal{X})$, where
\begin{equation*}
\mathcal{X} := \set{(H,i) \in \Herm(d) \times [d]: \spec(H) \in \mathcal{E}^d}.
\end{equation*}
Oracles in $\O$ act as $E_H: \ket{\psi_j} \0{n} \mapsto \ket{\psi_j} \ket{E_j}$, with the corresponding initial states $\ket{\psi_i}$, the eigenvectors of $H$, and target states $\sum_{j,l} w_{jl}^{(i)} / \|\vc{w}^{(i)}\|_2 \ket{l} \ket{\psi_j}$ where
$w_{jl}^{(i)} := \sqrt{f_{ij}/\abs{\mathcal{C}}} c_{ij}^{(l)}$,
$f_{ij} := \min \set{1, \exp(\beta(E_i-E_j))}$, and
$c_{ij}^{(l)} := \bra{\psi_j} C_l \ket{\psi_i}$.
\label{def:qmm}
\end{definition}
A critical part of the algorithm from~\cite{TOV+:2009} described above is how to revert a rejected move in Step~\ref{step:qms-reject}. Temme~\textit{et al.} show how this can be done by using an unwinding technique based on successive measurements, but we will not describe this technique in detail, as we now show how this step can be avoided by using quantum rejection sampling. Intuitively, this can be done by using amplitude amplification to ensure that the measurement in Step~\ref{step:qms-reject} always projects on the ``accept'' state $\ket{1}$. This also avoids having to repeatedly attempt random moves until one is accepted, and the number of steps of amplitude amplification will be quadratically smaller than the number of random moves that have to be attempted until one is accepted. This leads to the following statement:
\thmmetropolis*
\begin{proof}
The modified algorithm follows the same lines as the original algorithm, except that Steps~\ref{step:qms-coin}-\ref{step:qms-reject} are replaced by a quantum rejection sampling step. We use a quantum coin to choose the random gate in Step~\ref{step:qms-random-move} in order to make it coherent. The algorithm starts by applying the phase estimation oracle $E_H$ on the initial state to prepare the state $\ket{\psi_i} \ket{E_i}$, and then proceeds with the following steps:
\begin{enumerate}
\item\label{step:qms2-random-move}
Prepare an extra register in the state $\frac{1}{\sqrt{\abs{\mathcal{C}}}}\sum_l\ket{l}$.
Conditionally on this register, apply the gate $C_l \in \mathcal{C}$
on the eigenstate register to prepare the state
\begin{equation*}
\frac{1}{\sqrt{\abs{\mathcal{C}}}} \sum_j
\left[ \sum_l c_{ij}^{(l)} \ket{l} \right]
\ket{\psi_j} \ket{E_i}.
\end{equation*}
\item\label{step:qms2-pe}
Apply the phase estimation operator $E_H$ on the second register and an ancilla
register initialized in the default state $\0{n}$ to prepare the state
\begin{equation*}
\frac{1}{\sqrt{\abs{\mathcal{C}}}} \sum_j
\left[ \sum_l c_{ij}^{(l)} \ket{l} \right]
\ket{\psi_j} \ket{E_i} \ket{E_j}.
\end{equation*}
\item\label{step:qms2-sqrs}
Convert this state to the state given in Eq.~(\ref{eq:target state}):
\begin{equation*}
\frac{1}{\|\vc{w}^{(i)}\|_2} \sum_j
\sqrt{\frac{f_{ij}}{\abs{\mathcal{C}}}}
\left[ \sum_l c_{ij}^{(l)} \ket{l} \right]
\ket{\psi_j} \ket{E_i} \ket{E_j}.
\end{equation*}
\item
Discard $\ket{E_i}$ and uncompute $\ket{E_j}$
by using one call to the phase estimation oracle $E_H^\dagger$.
\end{enumerate}
Note that Step~\ref{step:qms2-sqrs} is an instance of $\sqsampling_{\vc{\tau}}$, where the pair of basis states $\ket{E}\ket{E'}$ of the phase estimation registers plays the role of the states $\ket{k}$, the initial amplitudes $\pi_{E,E'}$ are given by $\frac{1}{\sqrt{\abs{\mathcal{C}}}}\sqrt{\sum_l\abs{c_{ij}^{(l)}}^2}$ for $(E,E')=(E_i,E_j)$ or $0$ for values $(E,E')$ that do not correspond to a pair of eigenvalues of $H$, the states $\left[ \sum_l c_{ij}^{(l)} \ket{l} \right] \ket{\psi_j} / \sqrt{\sum_l\abs{c_{ij}^{(l)}}^2}$ play the role of the unknown states $\ket{\xi_k}$, and the ratio between the initial and target amplitudes is given by $\tau_{E,E'} = \sqrt{\min \set{1, \exp(\beta(E-E'))}}$ (the reduction is summarized in Table~\ref{table:reduction-qms}). Therefore, this step may be performed using the algorithm in Lemma~\ref{lem:algo-sqsqmpling}. Here, we choose $\alpha=1$ since the full Quantum Metropolis Sampling algorithm requires to apply a large number of successive random moves, therefore each instance of $\qmm$ should be solved with high success probability. Choosing $\alpha=1$ ensures that each random move will have success probability 1 (under our assumption that the phase estimation oracle is perfect), using an expected number of phase estimation oracles $O(1/\|\vc{w}^{(i)}\|_2)$.
\qed
\end{proof}
Note that in this case it is critical that the algorithm only requires one copy of the initial state, hence solving the quantum state conversion problem $\sqsampling_{\vc{\tau}}$ (in contrast, the quantum algorithm for linear systems of equations used a unitary to create multiple copies of the initial state, which is allowed only in the weaker quantum state generation problem $\qsamplingab$). Indeed, creating the initial state requires one copy of the previous eigenstate $\ket{\psi_i}$, which cannot be cloned as it is unknown. Here, the algorithm only requires to reflect through the initial state, which can be done by inverting Steps~\ref{step:qms2-random-move}-\ref{step:qms2-pe}, applying a phase $-1$ conditionally on the eigenenergy register being in the state $\ket{E_i}$ (which is possible since $E_i$ is known), and applying Steps~\ref{step:qms2-random-move}-\ref{step:qms2-pe} again.
Repeating the algorithm for $\qmm$ a large number of times will simulate the same random walk on the eigenstates of $H$ as the original quantum Metropolis sampling algorithm in~\cite{TOV+:2009}, except that we have a quadratic speed-up over the number of attempted moves necessary to obtain an accepted move. In order to converge to the Gibbs distribution, we need to take into account this quadratic speed-up in order to decide when to stop the algorithm, effectively assuming that each move takes quadratically longer than it actually does. Another option would be to modify the algorithm for $\qsamplingab$ so that it also estimates $\|\vc{w}^{(i)}\|_2$ by using amplitude estimation or quantum counting~\cite{BoyerBHT98,AmplitudeAmplification}. We leave the full analysis of these technical issues for future work.
\begin{table*}
\begin{center}
\begin{tabular}{|c|c|}
\hline
$\sqsampling_{\vc{\tau}}$ & $\qmm$ \\
\hline
$\ket{k}$ & $\ket{E}\ket{E'}$\\
$\pi_k$ & $\pi_{E,E'}:=
\begin{cases}
\frac{1}{\sqrt{\abs{\mathcal{C}}}}\sqrt{\sum_l\abs{c_{ij}^{(l)}}^2}
\textrm{ if } (E,E') = (E_i,E_j) \textrm{ where } E_i, E_j \in \spec(H) \\
0 \textrm{ if } E \notin \spec(H) \textrm{ or } E' \notin \spec(H)
\end{cases}
$\\
$\ket{\xi_k}$ & $\ket{\xi_{E,E'}}:=
\begin{cases}
\sum_l c_{ij}^{(l)}\ket{l}\ket{\psi_j}/\sqrt{\sum_l\abs{c_{ij}^{(l)}}^2}
\textrm{ if } (E,E') = (E_i,E_j) \textrm{ where } E_i, E_j \in \spec(H) \\
\textrm{n/a if } E \notin \spec(H) \textrm{ or } E' \notin \spec(H)
\end{cases}
$\\
$\tau_k$ & $\tau_{E,E'} := \sqrt{\min \set{1, \exp(\beta(E-E'))}}$ \\
\hline
\end{tabular}
\end{center}
\caption{Reduction from Step~\ref{step:qms2-sqrs} in the new quantum Metropolis sampling algorithm to the quantum resampling problem \textnormal{$\sqsampling_{\vc{\tau}}$}.}
\label{table:reduction-qms}
\end{table*}
\subsection{Boolean hidden shift problem} \label{sect:hiddenshift}
Our final application of the quantum algorithm for the $\qsamplingab$ problem is a new quantum algorithm for the Boolean hidden shift problem $\bhsp_f$.
We refer to the recent review~\cite{DeWolf:2008} for a good overview
of basic properties of the Fourier transform of Boolean functions.
Fourier analysis on the Boolean cube studies the $2^n$-dimensional
vector space of all \emph{real}-valued functions defined on the
$n$-dimensional Boolean cube $\bb{n}$. Thus, in the following definition $f$
denotes a function of the form $\bb{n} \to \mathbb{R}$ (\emph{not} a
Boolean function). The Boolean case is discussed later.
\begin{definition}\deftitle{Fourier transform}
The \emph{Fourier basis} of $\bb{n}$ consists of functions $\set{\chi_{\bv{w}} : \bv{w} \in \bb{n}}$, where each $\chi_{\bv{w}} : \bb{n} \to \set{1,-1}$ is defined as
$
\chi_{\bv{w}}(\bv{x}) := (-1)^{\bv{w} \cdot \bv{x}},
$
where $\bv{w} \cdot \bv{x} := \sum_{i=1}^n w_i x_i$ is the inner product in $\b$. The \emph{Fourier transform} of a function $f : \bb{n} \to \mathbb{R}$ is the function $\hat{f}: \bb{n} \to \mathbb{R}$ defined as
$
\hat{f}(\bv{w}) :=
\frac{1}{2^n} \sum_{\bv{x} \in \bb{n}} (-1)^{\bv{w} \cdot \bv{x}} f(\bv{x}).
$
Note that $\hat{f}(\bv{w}) := \mathbb{E}_{\bv{x}}(\chi_{\bv{w}} f) = \frac{1}{2^n}
\sum_{\bv{x} \in \bb{n}} \chi_{\bv{w}}(\bv{x}) f(\bv{x})$ which is another way to write the
Fourier coefficients.
The set $\set{\hat{f}(\bv{w}) : \bv{w} \in
\bb{n}}$ of all values of $\hat{f}$ is called the \emph{spectrum} of
$f$ and each of its elements is called a \emph{Fourier coefficient} of
$f$.
\end{definition}
Let us consider a \emph{Boolean function} \mbox{$f: \bb{n} \to \b$}. To find its Fourier transform, it is
required to associate $f$ with a real-valued function \mbox{$F : \bb{n} \to \mathbb{R}$} in some way. Instead of the obvious correspondence (treating $0,1 \in \b$ as real numbers) for the purposes of this work it is more natural to let $F$ be the $(\pm 1)$-valued function defined by $F(\bv{x}) := (-1)^{f(\bv{x})}$.
\begin{definition}\deftitle{Fourier transform of a Boolean function}
By slight abuse of notation, the \emph{Fourier transform of a
Boolean function} \mbox{$f: \bb{n} \to \b$} is the function
$\hat{f}: \bb{n} \to \mathbb{R}$ defined as
\begin{equation*}
\hat{f}(\bv{w})
:= \mathbb{E}_{\bv{x}}(\chi_{\bv{w}} F)
= \frac{1}{2^n} \sum_{\bv{x} \in \bb{n}} (-1)^{\bv{w} \cdot \bv{x} + f(\bv{x})}.
\end{equation*}
\end{definition}
Based on a reduction to $\qsamplingab$, we can now prove the following upper bound on the query complexity of $\bhsp_f$.
\thmBHSP*
\begin{proof
We will use $\oracle{\bv{s}}$ to denote the \emph{phase oracle} for function $f_{\bv{s}}$, \textit{i.\,e.}, a diagonal matrix that acts on the standard basis vectors $\bv{x} \in \bb{n}$ as follows:
$\oracle{\bv{s}} \ket{\bv{x}} := (-1)^{f(\bv{x}+\bv{s})} \ket{\bv{x}}$.
Let us consider the quantum oracle $\oracle{\bv{s}}$ conjugated by the Hadamard transform. The resulting operation
\begin{equation*}
V(\bv{s}) := \H{n} \, \oracle{\bv{s}} \, \H{n}
\end{equation*}
is very useful, since, when acting on a register initialized in all-zeros state, it can be used to prepare the following quantum superposition:
\begin{equation*}
\ket{\psi_{\smash{\hat{f}}}(\bv{s})}
:= V(\bv{s}) \ket{0}\xp{n}
= \sum_{\bv{w} \in \bb{n}} (-1)^{\bv{w} \cdot \bv{s}} \hat{f}(\bv{w}) \ket{\bv{w}}.
\end{equation*}
If we could eliminate the Fourier coefficients $\hat{f}(\bv{w})$ from state $\ket{\psi_{\smash{\hat{f}}}(\bv{s})}$, we would obtain a state
\begin{equation*}
\ket{\psis} := \frac{1}{\sqrt{2^n}} \sum_{\bv{w} \in \bb{n}} (-1)^{\bv{w} \cdot \bv{s}} \ket{\bv{w}}
\end{equation*}
from which the hidden shift $\bv{s}$ can be easily recovered by applying the Hadamard transform $\H{n}$.
Luckily, the problem of transforming the state $\ket{\psi_{\smash{\hat{f}}}(\bv{s})}$ to $\ket{\psis}$ is a special case of $\qsamplingab$ with
\begin{align*}
\pi_{\bv{w}} &:= \abs{\hat{f}(\bv{w})}, &
\sigma_{\bv{w}} &:= 1/\sqrt{2^n}, &
\ket{\xi_{\bv{w}}} &:= (-1)^{\bv{w} \cdot \bv{s}} \ket{0}.
\end{align*}
(More precisely, the initial amplitudes are $\hat{f}(\bv{w})$ instead of $\abs{\hat{f}(\bv{w})}$. However, the function $f$ and therefore its Fourier transform $\hat{f}$ is completely known, so we can easily correct the phases using a controlled-phase gate.) As a consequence, Theorem~\ref{thm:qsampling} immediately gives us a quantum algorithm for solving this problem.
\qed
\end{proof}
The complexity of the algorithm is limited by the smallest Fourier coefficient of the function. By ignoring small Fourier coefficients, one can decrease the complexity of the algorithm, at the cost of a lower success probability. However, the success probability of this algorithm can be boosted using repetitions, which requires to construct a procedure to check a candidate shift. We propose such a checking procedure based on a controlled-SWAP test. The number of necessary repetitions may then be decreased quadratically using the amplitude amplification technique of~\cite{HMW03} (note that we cannot use the usual amplitude amplification algorithm since the checking procedure is imperfect). This leads to the following theorem (proved in Appendix~\ref{sect:boosting}):
\thmBHSPboost*
\section*{Conclusion and open problems}
We provide an algorithm for solving the quantum resampling problem. Our algorithm can be viewed as a quantum version of the classical rejection sampling technique.
It relies on amplitude amplification~\cite{AmplitudeAmplification} to increase the amplitude of some target ``accept'' state, and its query complexity is given by a semidefinite program.
The solution of this SDP and hence the cost of the algorithm depends on the ratio between the amplitudes of the initial and target states, similarly to the case of the classical rejection sampling where the cost is given by the ratio of probabilities. Using the automorphism principle over a unitary group, we derive an SDP for the lower bound that is identical to the one for the upper bound, showing that our algorithm has optimal query complexity.
While the original adversary method cannot be applied as is for this quantum state generation problem because the oracle encodes an unknown quantum state instead of some unknown classical data, it is interesting to note that the query complexity of this problem is also characterized by an SDP. Therefore, an interesting open question is whether the adversary method~\cite{Amb00,NegativeWeights}, which has been shown to be tight for evaluating functions~\cite{Reichardt:2009,ReichardtReflections,LMRSS11} and nearly tight for quantum state generation or conversion problems with classical oracles~\cite{LMRSS11}, can always be extended and shown to be tight for this more general framework of problems with quantum oracles.
In Sect.~\ref{sect:applications}, we illustrate how quantum rejection sampling may be used as a primitive in algorithm design by providing three different applications. We first show that it was used implicitly in the quantum algorithm for linear systems of equations~\cite{HHL:09}. By assuming a perfect phase estimation operator on the matrix of the system, we show that this problem reduces to a quantum state conversion problem which we call $\qle$, which itself reduces to $\sqsampling_{\vc{\tau}}$. An open question is how to combine the quantum rejection sampling approach with the variable time amplitude amplification technique that was proposed by Ambainis~\cite{Ambainis2010} to improve on the original algorithm by Harrow \textit{et al.}~\cite{HHL:09}. In order to do so, we should ``open'' the phase estimation black box since Ambainis's main idea is to stop some branches of the phase estimation earlier than others.
As a second application, we show that quantum rejection sampling can be used to speed up the main step in the original quantum Metropolis sampling algorithm~\cite{TOV+:2009}. The general idea is to use amplitude amplification to increase the acceptance probability of a move, and therefore quadratically reduce the number of moves that have to be attempted before one is accepted. While this approach also provides some type of quadratic speed-up, it is rather different from the ``quantum-quantum'' Metropolis algorithm proposed by Yung and Aspuru-Guzik~\cite{YA:2010}. The main difference is that the approach based on quantum rejection sampling still simulates the same classical random walk on the eigenstates of the Hamiltonian, whereas the quantum-quantum Metropolis algorithm replaces it by a quantum walk. Note that while random walks converge towards their stationary distribution from any initial state, this is not the case for quantum walks as they are reversible by definition. Therefore, while both the original quantum Metropolis sampling algorithm and our variation can start from any initial state and run at a fixed inverse temperature $\beta$ to converge to the corresponding Gibbs distribution, the quantum-quantum Metropolis sampling algorithm works differently: it starts from a uniform superposition, which corresponds to the Gibbs distribution at $\beta=0$, and uses a series of measurements to project this state onto superpositions corresponding to Gibbs distributions with increasingly large $\beta$, until the desired value is reached.
Finally. as shown in Sect.~\ref{sect:hiddenshift}, we can apply the quantum rejection sampling technique to solve the hidden shift problem for any Boolean function $f$. In the limiting cases of flat or highly peaked Fourier spectra we recover the quantum algorithm for bent functions~\cite{Roetteler:2010} or Grover's algorithm for delta functions~\cite{Grover:96}, respectively. For a general Boolean function the hidden shift problem can be seen as lying somewhere between these two extreme cases. While the algorithm is known to be optimal for the extreme cases of bent and delta functions, its optimality for more general cases remains an open problem. A related question is the optimality of the checking procedure that leads to Theorem~\ref{thm:BHSP2}.
\section*{Acknowledgements}
The authors acknowledge support by ARO/NSA under grant W911NF-09-1-0569
and also wish to thank Andrew Childs, Dmitry Gavinsky, Sean Hallgren and Guosen Yue for fruitful discussions. M.O. acknowledges support from QuantumWorks and the US ARO/DTO and would like to thank Martin Schwarz and Kristan Temme for fruitful discussions.
\bibliographystyle{abbrvurl}
|
\section{Introduction}\label{sec:introduction}
\label{s:introduction}
The fact that their peak absolute magnitude is correlated with the
width of the light curve has allowed Type Ia supernovae (SNe~Ia ) to
be used as standard candles in determining cosmological
distances. Despite this widespread use of SNe~Ia\ as standard
candles, many problems in describing how the explosion of the star
happens still remain.
A number of models have been proposed to explain SNe~Ia\ \citep[for a review, see][]{hillebrandt.niemeyer:type,podsiadlowski2008}.
Two promising candidates for an explosion mechanism are the
sub-Chandrasekhar double detonation model \citep{woosley1994,fink2010}
and the gravitational confined detonation (GCD) model
\citep{plewa.calder.ea:type}. In both models, most if not all of the
nuclear burning occurs in a detonation wave. In the sub-Chandrasekhar
model a layer of helium is deposited on the surface of a white
dwarf. The helium layer detonates resulting in a shock wave traveling
around the surface of the white dwarf. When the shock wave converges
at the antipode, a detonation is thought to be triggered off center in the carbon-oxygen core
\citep{woosley1994,fink2010,sim2010}. The nucleosynthetic
yield is set predominantly by the mass of the carbon-oxygen white dwarf
core and the mass of the helium layer \citep{woosley1994,fink2010}.
In the GCD model, the carbon burning runaway within the convective
core of a near Chandrasekhar-mass white dwarf is postulated to occur in a small
region displaced from the stellar center that becomes a highly
buoyant flame bubble and quickly rises to the stellar surface after
burning only a few percent of the star during its ascent
\citep[e.g.][]{plewa.calder.ea:type, Livne2005On-the-Sensitiv}. When the buoyant ash, as
well as unburned material pushed ahead of the rising flame bubble,
erupts forth from the stellar core it is largely confined to the
surface of the white dwarf by gravity. It then becomes a strong surface
flow that sweeps completely over the star, eventually converging in a
region opposite to the breakout location. Although the details of the detonation initiation process is still a
topic of active research \citep[see e.g.,][]{Ropke:2007fu,seitenzahl2009b,seitenzahl2009c} the
high temperatures and densities reached within the converging surface
flow are strong indicators that a detonation is likely to be
triggered.
The resulting GCD nucleosynthetic yield consists almost entirely of
detonation burning products and depends on how much the star has
expanded by the time the detonation initiates. More highly expanded
(hence lower density) cores at detonation result in a smaller fraction
of Fe peak nuclei, less \nickel[56], and consequently a larger
fraction of intermediate mass elements (IMEs) due to incomplete
relaxation to nuclear statistical equilibrium (NSE). Therefore, lower
luminosity (less \nickel[56] producing) explosions are accompanied by
a larger yield of IMEs, as has been observed \citep[see e.g.,][]{mazzali2007}.
The expansion of the star prior to detonation, in a GCD scenario, results from the work
done by the rising flame bubble. Deflagrations that burn
more mass prior to reaching the stellar surface excite higher
amplitude pulsations and hence more-expanded stars at the time of
detonation. It has been found that the expansion of the star due to
the deflagration is very well represented by the fundamental radial
pulsation mode of the underlying white dwarf \citep[see Fig. 2 and
Fig. 12 in][]{meakin2009}. Therefore, while it is essential to
understand the nature of the deflagration so as to better understand
the mapping between initial conditions and final outcomes, the range
of outcomes due to the deflagration can be parameterized by the
pulsation amplitude, resulting in a one parameter family of
models. While the phase of the pulsation at the time of detonation is
an additional parameter, it plays a lesser role since the pulsation
period is longer then the detonation crossing time.
In this Paper we explore the nucleosynthetic yields that result from
an edge-lit detonation of a pre-expanded near Chandrasekhar-mass white
dwarf core out of hydrostatic equilibrium. For our purposes we define edge-lit
detonation to be an off-center detonation where very little of the star has burned
before hand.
This is a toy-model which neglects the effect the deflagration has on the nucleosynthesis.
However, due to the small amount of mass burned in the deflagration it
captures most of the essential features of the GCD model. The physical
mechanism that gives rise to the abundance asymmetries in the ejecta
will furthermore manifest itself in other off-center detonation
models, such as the sub-Chandrasekhar double detonation models.
In Section~\ref{s:numerical-model} we describe the explosion model and
our computational method. Section~\ref{s:results} is comprised of
three parts. First, we discuss how the dynamics of an asymmetric
detonation affect the hydrodynamic profile of the SN~Ia\ remnant. We
then consider how variations in the thermodynamic trajectories of
detonated material affect the resulting nucleosynthesis. Finally, we
present the elemental yields found for our explosion model and
quantify the asymmetric distribution thereof in terms of center of
mass offsets, which we provide in tabulated form. We also show that
there is an asymmetric distribution in velocity space for elemental nickel. We then
conclude in Section~\ref{s:discussion}, with a discussion of how our
results relate to previous work and how our work can be improved.
\section{The Explosion Model}\label{s:numerical-model}
The explosion model discussed in this paper involves the
detonation of a cold (T=$4\ee{7}\numberspace\K$) white
dwarf of mass $1.365\ensuremath{M_\odot}$, comprised of 50\% \carbon\ and 50\% \oxygen\
by mass, which has been expanded according to its fundamental
radial-pulsation mode by such an amplitude that it has a central
density of $10^8\numberspace\grampercc$. A detonation is initiated in this expanded white dwarf at a radial location
of $r=2\times 10^7$cm, where the density of the white dwarf is
$\rho= 10^7\grampercc$, by heating a small spherical volume with
radius $r_{\rm ini}\sim 4$ km to high temperature $T_{\rm ini}\sim10^9$K. This
heating immediately triggered a detonation which propagates away from the
point of initiation.
The subsequent reactive-hydrodynamic evolution of the detonated white dwarf was conducted using
the FLASH code \citep{fryxell.ea:flash}. The code framework and the included physics
is identical to that described in \citet{meakin2009}. The star was set up on a
Cartesian grid in 2D with cylindrical symmetry. An effective adaptive mesh refinement
resolution of 1\numberspace\km\ is used. The detonation was initiated along the symmetry
axis, which is the only natural way to capture its evolution in this geometry.
Due to the cylindrical symmetry of the physical scenario and the lack of large scale instability in the detonation front
this setup gives the same results as a full 3D calculation of the same problem
(see \citet[][]{meakin2009} which provides an in depth description of the
detonation phase, and a comparison between 2D and 3D).
Energy release and bulk compositional evolution due to nuclear burning is coupled to the hydrodynamic flow by advancing a
system of progress variables which represent three stages of burning: (1) the burning
of carbon and oxygen to silicon group elements, (2) the relaxation of the silicon group elements
to NSE, and (3) the evolution of material which has already relaxed entirely to NSE, including its
adjustment to changing density and temperature and neutronization due to weak interactions
among the Fe peak elements. Additional details can be found in \citet{Calder2007Capturing-the-F,meakin2009}.
The pre-expansion of the white dwarf prior to detonation initiation results in $\sim
1\ensuremath{M_\odot}$ of high density material ($\rho>10^7\numberspace\grampercc$), which
burns to NSE in the detonation (primarily as \nickel[56]). The
remaining $\sim0.37\ensuremath{M_\odot}$ of material burns to IMEs, e.g. Si, S, Ca,
and Ar, resulting in only a small amount ($<0.04\ensuremath{M_\odot}$) of unburned
C/O in the outermost layers of the remnant.
Detailed yields are calculated by post processing Lagrangian tracer particles included
in the explosion calculation and are the primary focus of this
paper. We recorded the time history of 10$^4$ particles which were
initialized to evenly sample the initial mass of the white dwarf. Our reaction network incorporates
493 nuclides from n to \krypton[86] (Table~\ref{t:network}). We use the reaction rates from the Joint
Institute for Nuclear Astrophysics \code{reaclib}
Database\footnote{\url{http://groups.nscl.msu.edu/jina/reaclib/db/}}
\citep[][and references therein]{cyburt2010}; the light-element rates
are mostly experimental and are from compilations such as
\citet{caughlan88:_therm} and \citet{Iliadis2001Proton-induced-}. Weak
reaction rates are taken from \citet{Fuller1982Stellar-weak-in} and
\citet{langanke.martinez-pinedo:weak}. Screening is incorporated using
the formalism of \citet{graboske.dewitt.ea:screening}.
\begin{deluxetable}{crcrcrcr}
\tablecaption{493-Nuclide Reaction Network} \tablehead{\colhead{El.}
& \colhead{$A$} & \colhead{El.} & \colhead{$A$} & \colhead{El.} &
\colhead{$A$} & \colhead{El.} & \colhead{$A$}} \startdata
n & & & & & & & \\
H & 1--3 & Ne & 17--28 & K & 35--46 & Ni & 50--73 \\
He & 3--4 & Na & 20--31 & Ca & 35--53 & Cu & 54--70 \\
Li & 6--8 & Mg & 20--33 & Sc & 40--53 & Zn & 55--72 \\
Be & 7, 9--11 & Al & 22--35 & Ti & 39--55 & Ga & 58--73 \\
B & 8, 10-14 & Si & 22--38 & V & 43--57 & Ge & 59--76 \\
C & 9--16 & P & 26--40 & Cr & 43--60 & As & 62--76 \\
N & 12--20 & S & 27--42 & Mn & 46--63 & Se & 62--82 \\
O & 13--20 & Cl & 31--44 & Fe & 46--66 & Br & 71--81 \\
F & 15--24 & Ar & 31--47 & Co & 50--67 & Kr & 71--86 \\
\enddata
\label{t:network}
\end{deluxetable}
\section{Results}
\label{s:results}
\subsection{Explosion Dynamics and Remnant Asymmetry}
\label{s:explosionDynamics}
The detonation propagates from the point of initiation at nearly
the Chapman--Jouguet (CJ) speed, $D_{CJ}\sim 1.2\ee{9}\numberspace\cmpersec$
at a Mach number of $M_{CJ}\sim D_{CJ}/c_s \sim 3.4$. Because of the
weak upstream density dependence of the detonation speed under these
conditions, the detonation front remains very nearly spherical in
shape as it engulfs the star. The total time required for its passage
across the expanded white dwarf core is $t_{cross}\sim 2 r/D_{CJ}\sim
0.5\numberspace\unitstyle{s}$. This is followed by a period of $\sim
0.5\numberspace\unitstyle{s}$ in which the pressure forces drive the completely
incinerated remnant into a homologous expansion, characterized by a
purely radial expansion velocity profile with an expansion rate
proportional to the radial position $v \propto r$. After only a few
seconds, the remnant is expanding ballistically, and the total energy
budget is dominated by the kinetic energy. The homologous velocity
profile results in a self-similar density profile which persists until
the remnant begins to interact with the interstellar medium.
The expanding remnant resulting from the detonation is marked by
significant asymmetry. The late time density profiles along the
symmetry axis and the equator are shown in Fig~\ref{f:detprofile},
scaled by the peak density in the remnant $\rho_c$. The initial,
spherically symmetric white dwarf density profile is also shown for
comparison. It can be seen that the density peak is shifted into the
hemisphere in which the detonation was initiated, $y>0$ in this case,
resulting in a steeper density gradient in this hemisphere. This is
in agreement with the series of GCD simulations described in
\citet{meakin2009}(see their Fig. 9 - 11) which show that the density
isocontours in the remnant are well described by concentric circles
that have centers offset from the initial stellar center. The density
isocontours were found to have larger offsets at higher densities,
with the largest offset centered on the peak density in the remnant.
{\em In all cases the density peaks on the side of the
remnant where the detonation originated and has a steeper density gradient in that region.}
\begin{figure}[htbp]
\includegraphics[width= 3.5in]{f1}
\caption{Late time (t $>$ 3 s) density and velocity profiles for the
post-detonation state. The density is scaled by the peak value
and the position is scaled by the density e-folding distance in
the equatorial direction. The thick gray line shows the scaled
density profile of the initial white dwarf model, while the post-detonation
state model is shown by thick black lines for lines through the
equator (solid) and through the poles (dashed).}
\label{f:detprofile}
\end{figure}
Unlike the density profile, the velocity profile (also shown in
Fig.~\ref{f:detprofile}) does not show an asymmetry, but is everywhere
radially directed and spherically symmetric. This leads to an
asymmetry in the density as a function of expansion velocity, which is
likely to result in a viewing angle dependence for the light curve and
the spectral signature. Related composition asymmetries, discussed in
\S\S\ref{s:NucDepend} and \ref{s:Phenom} below, also contribute
to observable asymmetries and viewing angle dependencies.
A revealing format for presenting the dynamics of the detonation and
the subsequent expansion is the space-time diagram. In
Fig.~\ref{f:SpaceTimeTrajs} we present the space-time trajectories for
all of the tracer particles that were initialized near the symmetry
axis of the white dwarf. The left panel shows the time period over
which the detonation traverses the stellar core, while in the right
panel we show the later time evolution that ends in a radially
expanding, ballistic trajectory for each of the particles. The bold
dashed line in the left panel shows the path taken by a theoretical
detonation having a constant speed, which matches the kinks in the
particle trajectories very well.
\begin{figure*}[htbp]
\includegraphics[width= 3.5in]{f2a}
\includegraphics[width= 3.5in]{f2b}
\caption{(left) On-axis Lagrangian tracer particle positions are
shown as a function of time in this space-time diagram. The thick
dashed line shows the theoretical position of a constant speed
detonation originating from the location $y=0.2\ee{9}\numberspace\centi\meter$,
coincident with the detonation initiated in the simulation. A
theoretical detonation speed taken to be $v_{\rm det} =
1.1\ee{9}\numberspace\cmpersec$ for the detonation as it moves inward
toward the high density core and a speed that is 10\% lower as it
moves outward into the low density surface material provides a
very good match to the simulation data. (right) At times greater
than $1\unitstyle{s}$ the Lagrangian tracer particles exhibit homologous
expansion.}
\label{f:SpaceTimeTrajs}
\end{figure*}
The following features in this figure merit further discussion. (1)
The trajectories are slowly converging prior to detonation. This is
the signature of the stellar core undergoing mild contraction as a
result of having been expanded by a radial pulsation mode prior to
detonation. (2) The detonation accelerates material in the direction
it is propagating. This is the primary source of the asymmetry
imprinted on the remnant at late times. A large number of Lagrangian tracer
particles in the detonated hemisphere are first accelerated towards
the stellar center by the detonation before they are turned around by
pressure forces and accelerated to their final, outwardly directed
expansion velocities. The exact number of tracer particles accelerated towards
the stellar center by the detonation is dependent on distance from the center of the star to the point where the detonation was initiated, $a$.
A given tracer particle with a central angle $\theta$ and distance from the center of the star $r$ will be accelerated towards the center of the star if
$sin(\theta) > r/a$.
On the other hand, tracers in the opposite
hemisphere are accelerated by the detonation in the same direction as
their final expansion velocity, reaching their final velocity on a
shorter timescale. (3) The material in the detonated hemisphere is
accelerated to lower velocities overall compared to material in the
opposite hemisphere (see Fig.~\ref{f:SpaceTimeTrajs} (right)). This
mapping between initial position (and therefore initial density) and
resultant expansion velocity explains the density profile: the
material lines in the more rapidly expanding hemisphere are stretched
out over a larger region of space, and hence to a lower relative
density, than the more slowly expanding regions. (4) A natural
consequence of the explosion dynamics is an asymmetry in the expansion
timescale $t_{exp}$, defined as the time required for the detonated
material to drop from its post-detonation temperature maximum
$T_{max}$ to $e^{-1}T_{max}$, resulting from the differential rate at
which material cools (nearly adiabatically) due to the post-detonation
expansion. This follows directly from point (2) above. The expansion
timescale asymmetry is shown in Fig.~\ref{f:SpaceDist} where we have
plotted the tracer particles at their initial position in the stellar
core, color coded by their post-detonation expansion timescale. It is
obvious from this figure that the material in the detonated hemisphere
$(y>0)$ has overall a larger expansion timescale than in the opposite
hemisphere for a given initial upstream density. As will be discussed
in \S\S\ref{s:NucDepend} and \ref{s:Phenom} this expansion timescale
distribution imparts an asymmetry in the resultant nucleosynthetic
yield.
\begin{figure}[htbp]
\centering{\includegraphics[width=3.4in]{f3}}
\caption[Expansion time scale as a function of initial
position]{Initial spatial position of the Lagrangian tracer particles. Color
represents each particle's expansion time scale. The center of the
star is at X=0 Y=0 The detonation was initiated at X=0
Y=2. Particles on the side of the star where the detonation starts
have higher expansion time scales than the particles on the
opposite side of the star. The expansion time scale is calculated from the temperature profile of each tracer.}
\label{f:SpaceDist}
\end{figure}
\subsection{Nucleosynthesis Dependence}\label{s:NucDepend}
We find that nuclear burning in SNe~Ia\ progresses in three distinct
stages \citep{Khokhlov1983Deflagration-Fr,khokhlov1991}. The first
stage is carbon burning. During carbon burning \carbon+\carbon\ is
the primary reaction taking place. We find that carbon burning never
reaches an equilibrium state in a small region of the star where the
final carbon mass fraction is above $10^{-4}$. The carbon burning
reactions are sensitive to temperature with
\begin{equation}\label{eq:dYdt}
\frac{dY(\carbon)}{dt} \propto f(T_9) T_9^{-2/3} e^{-84.165 T_9^{-1/3}}.
\end{equation}
Here $dY(\carbon)/dt$ is change in \carbon\ abundance over change in
time, $T_{9}$ is temperature in units $10^{9}\numberspace\K$, and $f(T_9)$ is
a function defining the temperature effect on the branching ratio
between the $ \carbon ( \carbon , \alpha ) \neon$ and $ \carbon (
\carbon , p ) \sodium[23]$ reaction \citep{caughlan88:_therm}. Any
change in the thermal profile will result in a different abundance
pattern for material that does not proceed to the next phase of
burning.
The next stage is oxygen burning. Here \oxygen\ and the products of
carbon burning proceed to silicon group elements, like S, Ar, and Si.
As in carbon burning, oxygen burning never reaches an equilibrium state
and therefore also shows a dependence on the thermal history. We find
$X(\oxygen) > 10^{-5}$ if the next burning stage did not start. Very little
mass of the star ($<0.04\ensuremath{M_\odot}$) is in a region that does not complete either carbon
or oxygen burning, so most of the star proceeds to the next burning
stage.
At higher temperatures and densities, silicon burning is the dominant
form of nucleosynthesis. In silicon burning groups of nuclides enter
into equilibrium; a state known as quasi-statistical equilibrium
\citep[QSE; e.g.,][]{woosley1973}. There are two ways in which
changing the expansion time scale can affect the abundances in
QSE. First, while equilibrium holds within a group of nuclides, it
does not hold between groups. Second, within a group of nuclides in
equilibrium reactions freeze out at different temperatures, resulting
in an abundance pattern that depends on the expansion time scale. This
will be further explored below.
Fig.~\ref{f:Yvstexp} shows how the abundances of silicon and nickel
vary as functions of expansion time scale over a small range of
temperatures for tracer particles that never finish silicon
burning. Even with the scatter from plotting particles with different
peak temperatures, there is a clear dependence of the abundances on
expansion timescale. The directions of these trends are
counterintuitive, but Fig.~\ref{f:TComp} shows their origin. Both
tracer particles shown reach a peak temperature of $\sim
5\ee{9}\numberspace\K$, and their nucleosynthesis is nearly identical up to
that point. The particle with the longer $(0.6016\numberspace\unitstyle{s})$ thermal
expansion time scale, however, has a temperature that falls faster
over the first $0.1\numberspace\unitstyle{s}$ than the particle with the shorter
$(0.4428\numberspace\unitstyle{s})$ thermal expansion time scale. This rapid decrease
in temperature results in less \silicon\ burned to \nickel[56] even
though the thermal expansion time scale is longer.
\begin{figure}[htbp]
\centering{\includegraphics[width=3.4in]{f4}}
\caption[Abundance as a function of expansion time scale]{Mass
fraction of \silicon (red) and \nickel (blue) as a function of
expansion time scale. This plot was made from tracer particles
that had a maximum temperature between $4.99\ee{9}\numberspace\K$ and
$5.01\ee{9}\numberspace\K$ and a density of approximately
$1.5\ee{7}\numberspace\grampercc$.The lines shown are least squares fits
to the data.}
\label{f:Yvstexp}
\end{figure}
\begin{figure*}[htbp]
\includegraphics[width= 3.5in]{f5a}
\includegraphics[width= 3.5in]{f5b}
\caption{(left) Temperature profile for two particles with similar
peak temperatures but different expansion time scales. The
expansion time scale for the solid line is $0.4428\numberspace\unitstyle{s}$ and
the expansion time scale for the dashed line is
$0.6016\numberspace\unitstyle{s}$.
The times for both particles have been offset
such that the peak temperature is reached at
$1\ee{-3}\numberspace\unitstyle{s}$. (right) The mass fraction of \carbon,
\silicon, and \nickel\ as a function of time for the same
time-adjusted particles.}
\label{f:TComp}
\end{figure*}
Material exposed to high enough density and temperature conditions for
long enough will arrive in a state of nuclear statistical equilibrium
(NSE). In this state all nuclear reactions enter equilibrium and lose
all history of the thermal evolution up to that point. For material
that has reached this state, the only dependence that the final yield
has on expansion timescale occurs during the process of freeze
out. Freeze out occurs for a nuclide when the temperature drops low
enough that all strong reactions become too slow to change the
nuclides abundance again. Because this condition occurs at different
temperatures for different nuclides, the final yield depends on the
rate at which material cools. Like in QSE, different reactions freeze out
at different temperatures, leading to yields that depend on the
expansion timescale. Therefore, all three stages of burning and NSE in
SNe~Ia\ are affected by different expansion time scales, resulting
in a clear compositional asymmetry that will be discussed in the
following section.
\subsection{Phenomenology}\label{s:Phenom}
We now present the results of our reaction network calculations for a
near edge-lit detonation in a SN~Ia\ model. We find that a number of
nuclides exhibit pronounced asymmetries across the stellar remnant. We
quantify this effect by calculating the center of mass for a given
element. Suppose a tracer particle $i$ has a mass fraction $X_i(Z)$
of element $Z$, and its position is given by vector
$\mathbf{r}_i$. Then the center of mass for a given element,
$\mathbf{r}_{cm}(Z)$, is given by the equation
\begin{equation}\label{eq:cm}
\mathbf{r}_{cm}(Z) = \left(\sum_{i} X_i(Z) \mathbf{r}_i\right) / \left(\sum_{i} X_i (Z)\right).
\end{equation}
Due to the cylindrical symmetry of our explosion model the
displacement of the center of mass for any element lies along the y
axis. A velocity for the center of mass $\mathbf{v}_{cm}$ can be
calculated by replacing $\mathbf{r}_i$ by the velocity $\mathbf{v}_i$
in equation \ref{eq:cm}. Table \ref{t:gradtab} shows the total mass,
displacement of the center of mass, and velocity of the center of mass for elements between carbon and
germanium. For reference, the detonation was initiated at $\sim
2\ee{8}\numberspace\centi\meter$. These numbers are correct for the end of our
simulation $(t \sim 3\numberspace\unitstyle{s})$, where strong reactions have frozen out and homologous
expansion has been reached. Some of the isotopes making these
elements, \nickel[56] for example, decay so Table \ref{t:gradtab}
evolves with time. The last three columns show the total mass, displacement of the center of mass, and velocity of the center of mass assuming all radioactive elements
instantly decayed to their stable isotopes.
Note that nickel shows no change in center of mass or velocity since most of the mass of the star ends up as radioactive nickel in our model.
If the complete star was burned to nickel then by definition the change in center of mass would be zero since our model conserves mass and momentum.
Elements lighter than silicon have their masses
distributed more in the direction where the detonation was
initiated. Elements heavier than silicon are, for the most part,
distributed away from where the detonation was initiated. These
elements also display an odd-even pattern where odd Z nuclei, like cobalt and copper, are
predominantly distributed farther away from the start of the
detonation than their even Z counterparts, like iron and zinc. This asymmetry is due to the different thermal histories of the two sides of the star affecting the nucleosynthesis as outlined in \S\S\ref{s:NucDepend}.
\begin{deluxetable*}{ccccccc}
\tablecaption{Centers of mass and velocities for various elements}
\tablehead{ & \colhead{$\Delta_{cm}(\sim3\numberspace\unitstyle{s})$} & \colhead{$V_{cm}(\sim3\numberspace\unitstyle{s})$} & \colhead{mass$(\sim3\numberspace\unitstyle{s})$} & \colhead{$\Delta_{cm}(decayed)$} & \colhead{$V_{cm}(decayed)$} & \colhead{mass$(decayed)$}\\
& $(10^8\numberspace\centi\meter)$ & $(10^8\cmpersec)$ & $(\unitstyle{g})$ & $(10^8\numberspace\centi\meter)$ & $(10^8\cmpersec)$ & $(\unitstyle{g})$}
\startdata
C & 29.1 & 9.9 & 9.00\ee{29} & 29.1 & 9.9 & 9.00\ee{29} \\
N & 18.2 & 6.2 & 2.91\ee{25} & 11.6 & 4.1 & 8.41\ee{26} \\
O & 7.30 & 2.4 & 7.28\ee{31} & 7.30 & 2.4 & 7.28\ee{31} \\
F & 27.9 & 9.4 & 2.19\ee{22} & 35.2 & 12. & 3.34\ee{23} \\
Ne & 28.7 & 9.9 & 1.03\ee{30} & 28.7 & 9.9 & 1.03\ee{30} \\
Na & 25.6 & 8.8 & 1.12\ee{27} & 22.3 & 7.6 & 5.69\ee{27} \\
Mg & 13.8 & 4.9 & 1.37\ee{31} & 13.8 & 4.9 & 1.37\ee{31} \\
Al & 21.6 & 7.5 & 1.74\ee{28} & 18.5 & 6.4 & 3.13\ee{28} \\
Si & -1.09 & -0.6 & 2.18\ee{32} & -1.09 & -0.6 & 2.18\ee{32} \\
P & -4.44 & -1.9 & 8.34\ee{28} & 3.77 & 1.2 & 4.87\ee{28} \\
S & -2.68 & -1.2 & 1.18\ee{32} & -2.68 & -1.2 & 1.18\ee{32} \\
Cl & -9.31 & -3.7 & 1.82\ee{28} & -4.28 & -1.8 & 1.21\ee{28} \\
Ar & -2.95 & -1.3 & 2.60\ee{31} & -2.94 & -1.3 & 2.60\ee{31} \\
K & -4.53 & -1.7 & 5.12\ee{27} & -4.20 & -1.7 & 3.66\ee{27} \\
Ca & -2.55 & -1.1 & 2.80\ee{31} & -2.56 & -1.1 & 2.80\ee{31} \\
Sc & -3.36 & -1.3 & 2.02\ee{26} & -9.21 & -3.7 & 1.57\ee{25} \\
Ti & -2.78 & -1.1 & 3.93\ee{28} & -2.35 & -1.0 & 7.54\ee{29} \\
V & -2.04 & -0.7 & 3.68\ee{27} & -3.91 & -1.5 & 2.45\ee{28} \\
Cr & -2.37 & -1.0 & 7.60\ee{29} & -1.81 & -0.8 & 1.71\ee{31} \\
Mn & -3.68 & -1.4 & 2.58\ee{28} & -6.01 & -2.4 & 2.91\ee{29} \\
Fe & -1.80 & -0.8 & 1.71\ee{31} & 0.00 & 0.0 & 2.16\ee{33} \\
Co & -5.59 & -2.2 & 3.06\ee{29} & -3.49 & -1.3 & 2.57\ee{30} \\
Ni & 0.00 & 0.0 & 2.16\ee{33} & 0.13 & 0.1 & 3.52\ee{31} \\
Cu & -3.06 & -1.1 & 4.68\ee{30} & -2.08 & -0.7 & 2.46\ee{28} \\
Zn & 0.17 & 0.1 & 2.93\ee{31} & -1.49 & -0.5 & 6.46\ee{29} \\
Ga & -2.09 & -0.7 & 3.09\ee{28} & -0.83 & -0.3 & 3.54\ee{25} \\
Ge & -1.48 & -0.5 & 6.31\ee{29} & -0.47 & -0.1 & 3.11\ee{26} \\
\enddata
\label{t:gradtab}
\end{deluxetable*}
Our simulation was not run sufficiently past freeze out to allow all
beta unstable nuclei to decay. However, if we decay the unstable
nuclei and group them in elemental abundances, we find elemental nickel to
have a clear gradient over the star. In material not burned to
NSE we find the mass fraction of elemental nickel between $+90^\circ$
and $-90^\circ$ central angle to each tracer particle to increase by a factor of 2--3. Figure
\ref{f:sample-abun} shows the dependence of elemental nickel on central angle to each tracer particle.
\begin{figure}[htbp]
\centering{\includegraphics[width= 3.0in]{f6}}
\caption[Final mass fraction of elemental Ni as a function of central to
angle]{Final mass fraction of elemental(nonradioactive) Ni as a function of the ejection
angle relative to the center of the star. The detonation started
in the surface layer of the star in the theta = $90^{\circ}$ direction. The
particles with a Ni mass fraction above $10^{-3}$ are particles that
have burned to NSE.}
\label{f:sample-abun}
\end{figure}
An interesting side effect of the different expansion times is that
material on opposite sides of the star expands at different
velocities. This leads to a gradient in velocity space. Figure
\ref{f:sample-abun2} shows how elemental nickel, iron, manganese, and chromium vary with radial
velocity for different central angles. In material not burned to NSE, the part of the
remnant with the most nickel is also the part with the highest radial
velocity. This is self-consistent since the nickel mass fraction is greater on
the side of the remnant with the shortest expansion time. Therefore, it
follows it should have the highest radial velocity.
\begin{figure}[htbp]
\centering{\includegraphics[width=3.4in]{f7}}
\caption[Final mass fraction of elemental iron group elements as a function of radial
velocity]{Final mass fraction of elemental iron group elements as a function of the final
radial velocity. The particles with a Ni mass fraction above $10^{-3}$
are particles that have burned to NSE. Particles ejected between $-80^{\circ}<\theta<-70^{\circ}$ reach the highest radial velocities.}
\label{f:sample-abun2}
\end{figure}
\section{Discussion}
\label{s:discussion}
We have computed the abundances and spatial distributions of nuclides
in an explosion of an expanded near Chandrasekhar-mass white dwarf resulting
from an off-center initiated detonation, a toy model that captures the
thermodynamic profile of some SN~Ia\ explosion models. We find a compositional
asymmetry in the ejecta produced by the detonation. This compositional
asymmetry is connected with the thermal expansion time scale. The
different expansion timescales also result in a compositional
asymmetry in velocity space.
It is difficult to establish the observable features of our model since we have not conducted radiative transfer calculations to generate light curves and spectra. It is currently unclear how much of an observational effect this asymmetry will have. A series of synthetic spectra generated over a range of time allows for direct comparison with observed supernovae. We conjecture, that even if compositional effects are obscured the spectra will show some dependence on observing angle. This is because the side of the remnant that expands at higher velocities will also be at a lower density, making it more transparent at earlier times.
Even though it is difficult to determine observational features of our model, it is instructive to compare and contrast our model with other recent results. The `toy model' in \citet{Hillebrandt2007} and
\citet{Sim2007} is constructed similarly to our calculated yield distribution. They found that off-center distributions of burned material are likely to leave detectable imprints on observed light curves. An angular dependence of the light-curve peak brightness is introduced that might explain some over luminous SNe~Ia .
Our model has no deflagration ash so comparisons with deflagration to detonation models, like \citet{Livne1999}, is problematic.
We can compare to \citet{maeda2010}, who showed that expansion velocity gradients as
inferred from the Si II $\lambda6355$ absorption feature could be
explained by a velocity shift of $3500\numberspace\km\unitspace\unitstyle{s}^{-1}$ in Si
happening in an `opening angle' $105-110^\circ$ away from the ignition
points caused by deflagration ash. Even though our model has no deflagration ash
a similar effect occurs with \silicon\ having ejection
velocities $\sim10000\numberspace\km\unitspace\unitstyle{s}^{-1}$ faster on the side of
the star opposite of the detonation. The ejecta in our model, however,
do not have a sharp transition in velocity but a gradual increase in
velocity starting at a point $\sim90^\circ$ away from the ignition
point. It is difficult to determine if our model can reproduce the
observed velocity gradients. It is worth noting that the deflagration
to detonation transition model used in \citep{maeda2010} did not
produce the observed velocity shift or `opening angle' either.
\citet{kasen2005} attempted to calculate the spectral signatures of GCD by considering ejecta interacting with an extended atmosphere. The ejecta in their model were calculated from a 1-D simulation, and therefore lack the asymmetry in the nucleosynthesis that the reaction network calculations of our 2-D simulation find. In \citep{kasen2007} nucleosynthesis was done approximately with a 13 element reaction network. The surface flow, which consists partly of deflagration ash, which was excluded from our present model, needs to be considered. The surface flow might also have a spectral signature itself such as the presence of an high-velocity calcium absorber \citep{kasen2005} and should be compared with the underlying compositional asymmetry. In the case of
sub-Chandrasekhar models, detonation of a pure helium shell leads to a layer containing iron-group elements such as titanium and chromium around the core ejecta \citep{fink2010,sim2010,kromer2010}.
Another item to consider is that we do not include the effects of metallicity on the
nucleosynthesis. Our model is initially composed of \carbon\ and
\oxygen . However, it has been shown that prior to the explosion of a
carbon-oxygen white dwarf in a SN~Ia\ there is a long period during
which some \carbon\ is converted into \carbon[13] as well as heavier
elements \citep{Chamulak2007The-Reduction-o}. This process makes even
the most metal poor SNe~Ia\ have a composition of more diverse than pure \carbon\
and \oxygen. It is worth mentioning that for deflagration--detonation
transition (DDT) models where the detonation density was allowed to
vary in relation to the flame speed as a function of metallicity
\citep{Chamulak2007The-Laminar-Fla} the yield of \nickel[56] produced
also varied with metallicity \citep{bravo2010,jackson2010}. DDT
models with varying metallicity and fixed detonation density, however,
showed little variation in outcome \citep{Townsley:2009jl}.
\acknowledgements
The authors wish to acknowledge that this work was supported by the US
Department of Energy, Office of Nuclear Physics, under contract
DE-AC02-06CH11357. This work was also supported by the NSF, grant
AST-0507456, by the \textbf{J}oint \textbf{I}nstitute for
\textbf{N}uclear \textbf{A}strophysics at MSU under NSF-PFC grant
PHY~0822648, by the DOE through grant 08ER41570, and by the Deutsche
Forschungsgemeinschaft via the Emmy Noether Program (RO 3676/1-1). The
software used in this work was in part developed by the DOE-supported
ASC / Alliance Center for Astrophysical Thermonuclear Flashes at the
University of Chicago. Simulations presented in this work were run at
the High Performance Computing Center at Michigan State University.
\bibliographystyle{apj}
|
\section{Introduction}
Based on results obtained by Ford and collaborators~\cite{ford11,fn}
and Hu and Shiokawa~\cite{hu1}, recently three of us~\cite{krein1}
proposed an analog model for fluctuations of light cones
induced by quantum gravity effects. The model is based on two general
features of waves propagating in random fluids. First, acoustic
perturbations in a fluid define discontinuity surfaces that provide
a causal structure with sound cones. Second, propagation of
acoustic excitations in a random medium are generally described by
wave equations with a random speed of sound~\cite{ishimaru,book3,book2,book}.
Taken together, these features lead to fluctuations of sound cones,
analogous to the fluctuations of light cones. Phonons propagating in
such a random fluid are then modeling photon propagation in a gravitational
field with a fluctuating metric.
In the present paper we depart from the quest for analog models of
quantum gravity effects and study quantum fields in a disordered
environment in a more general context. Our study belongs to a wide
program devoted to propagation of quantum matter fields in a
classical background spacetime ~\cite{birrell} but with metric
fluctuations. Specifically, we consider a scalar quantum field
described by a Klein-Gordon-like equation, in which the parameters
of the equation, the mass and the coefficient of the second-order
time derivative, become random functions of the spatial coordinates.
The randomness of the parameters is due to {\em static} noise
sources that couple to the scalar field. In the limit of a vanishing
mass, one recovers the random wave equation considered in
Ref.~\cite{krein1}. While not addressing the issue of the origin of
the noise sources, we mention that one has in mind that they can be
induced by a variety of phenomena, like metric fluctuations due to
quantum creation of gravitons in a squeezed coherent state in the
presence of a black hole or interactions with background topological
defects, among others. Like in the case of quantum fields in the
presence of an external heat bath, the noise sources break Lorentz
symmetry since they define a preferred reference frame.
A scalar quantum field associated with acoustic waves in a
disordered medium can define a situation where sound cones fluctuate
randomly. In the study of such a
situation, it is important to observe that systems with disorder can
be divided into two wide groups, namely, systems with quenched or annealed
disorder ~\cite{debashish,glass,parisi1}. While in annealed systems
the random field is in thermal equilibrium with the others degrees
of freedom of the system, in quenched systems they are not.
The differentiation between the two types of disorder
is an important issue in the studies of the influence of impurities
on phase transition phenomena. Many authors have used field theory
and the renormalization group to investigate systems with quenched
disorders~\cite{lub}, with particular interest in the role played by
the disorder on critical exponents. In the case of quenched random
fields, by means of the replica trick it is possible to define a
quenched generating functional of connected $n$-point functions.
With this generating functional in hand, the issue of the
influence of the disorder on phase transitions can be studied. In
the present paper we consider annealed disorder. We consider weak
noise fields and implement a perturbative expansion controlled by
small parameters that characterize the noise correlation functions.
We obtain causal two- and four-point Green's functions of the scalar
field. Averaging these Green's functions over the noise
fluctuations, one obtains two- and four-point functions
qualitatively similar to a self-interacting $\lambda\varphi^{4}$
theory. As will be shown, we obtain a frequency-dependent coupling
constant.
Since Unruh's original paper~\cite{unruh1}, the possibility of
simulating aspects of general relativity and quantum fields in
curved space-time through analog models has been widely discussed in
the literature~\cite{mario,lectures}. One interesting proposal is
the generation of an acoustic metric in Bose-Einstein condensates
and superfluids~\cite{am7}. On the other hand, analog models with
sonic black-hole could find a sort of generalization within the
random fluid scenario. Since Bose-Einstein condensates are natural
candidates to produce an acoustic black-hole, one way to go beyond
the semi-classical approximation is to inquire how these systems
behave in the presence of disorder. The calculations presented in
the present paper are the first steps in the implementation of such
a program.
The organization of the paper is as follows: In Section~\ref{sec:PT} we
discuss the perturbative approach in a free scalar field in the
presence of annealed disorder described by two random functions.
Section III contains our conclusions. In the appendix some lengthy
calculations are presented.
\section{Perturbation Theory in Annealed-like Disordered Media}
\label{sec:PT}
\vspace{0.5cm}
Let us consider a scalar field $\varphi(t, \mathbf{r})$, defined in a $(d+1)$ dimensional
space-time, that
satisfies the random Klein-Gordon equation:
\begin{eqnarray}
\left[\left(1 + \mu \right)\frac{1}{u_{0}^{2}}
\frac{\partial^{2}}{\partial\,t^{2}}-\nabla^{2}+
\left(1 + \xi \right) m^{2}_{0} \right]
\varphi(t,{\mathbf{r}})=0.
\label{KG-eq}
\end{eqnarray}
The quantities $\mu = \mu(\mathbf{r})$ and $\xi = \xi(\mathbf{r})$ are dimensionless random variables,
functions of the spatial coordinates only, whose statistical properties will
be defined shortly. Here, the random coefficient of the time derivative
characterizes the possible discontinuity surfaces of the theory and provides a
kind of random causal structure to it. It is well known that the
principal part of a differential equation, i.e. the terms with
the higher-order derivatives, determines entirely the \textit{loci}
of points in space-time where a solution may possess non-null
discontinuities. This property is at the root of all analogue models
of classical gravitation, namely, an effective causality can be
obtained from the kinematical properties of a physical system.
The region of influence of excitations is given by an envelope which
characterizes the maximum speed of propagation, that is, the light cone
of the theory. From a physical point of view the region of influence of the
theory may be obtained by an \textit{eikonal} approximation, where
the mass term may be disregarded. On the other hand, when the wave
frequency is of the same magnitude of the mass term, excitations of
the system propagate inside the characteristic cone.
In the present paper we consider zero-mean random functions
$\mu(\mathbf{r})$ and $\xi(\mathbf{r})$:
\begin{equation}
\langle \mu(\mathbf{r}) \rangle_{\mu} = 0, \hspace{0.5cm}
\langle \xi(\mathbf{r}) \rangle_{\xi} = 0,
\label{means}
\end{equation}
and, for simplicity, we suppose white-noise correlations:
\begin{eqnarray}
\langle \mu(\mathbf{r}) \mu(\mathbf{r}') \rangle_{\mu} &=& \sigma^2_\mu \,
\delta(\mathbf{r} - \mathbf{r}') ,
\label{corr-mu}
\\
\langle \xi(\mathbf{r}) \xi(\mathbf{r}') \rangle_{\xi} &=& \sigma^2_\xi \,
\delta(\mathbf{r} - \mathbf{r}') ,
\label{corr-xi}
\end{eqnarray}
where $\langle\, \cdots \,\rangle_{\mu}$ and $\langle\, \cdots \,\rangle_{\xi}$
denote ensemble average of noise realizations and $\sigma_\mu$ and
$\sigma_\xi$ characterize the strengths of the noises. We also suppose
that the noises are Gaussian distributed (we use the notation $\mu(\mathbf{r}_i)
= \mu_i $):
\begin{eqnarray}
\langle\mu_{i_1} \cdots \mu_{i_{2n}} \rangle_\mu &=&
\langle \mu_{i_1} \mu_{i_2} \rangle_\mu \langle \mu_{i_3} \mu_{i_4} \rangle_\mu
\cdots \langle \mu_{i_{2n-1}} \mu_{i_{2n}} \rangle_\mu \nonumber \\
&& + \, {\rm permutations} ,
\label{wick}
\end{eqnarray}
and correlations of an odd number of noises are zero. The reader may
be aware, but it is worth remembering that these random functions
$\mu$ and $\xi$ are statistically independent.
Since Eq.~(\ref{KG-eq}) is linear in $\varphi(t,\mathbf{r})$, it is
useful to resort to Fourier transforms in order to find its
solutions. Therefore, we define the Fourier transform of
$\varphi(t,\mathbf{r})$ on the time variable $t$:
\begin{equation}
\varphi(t,\mathbf{r}\,) = \int \frac{d\omega}{2\pi} \, e^{-i \omega t}\,
\varphi(\omega,\mathbf{r}), \label{FTt-phi}
\end{equation}
and on the space variable $\mathbf{r}$:
\begin{equation}
\varphi(\cdot ,\mathbf{r} ) = \int \frac{d\mathbf{k}}{(2\pi)^d} \,
e^{i \mathbf{k} \cdot \mathbf{r} } \, \varphi(\cdot,\mathbf{k}).
\label{FTr-phi}
\end{equation}
In addition, we define Fourier transforms of the stationary noise
functions:
\begin{equation}
\mu(\mathbf{r} ) = \int \frac{d\mathbf{k}}{(2\pi)^d} \,
e^{i \mathbf{k}\cdot\mathbf{r} } \,
\mu(\mathbf{k} ) ,
\label{FT-noise}
\end{equation}
and similarly for $\xi(\mathbf{r})$. Using these in Eq.~(\ref{KG-eq}),
one obtains that the Fourier components of the field
$\varphi(\omega,\mathbf{k})$ satisfy an algebraic equation that can
be written as
\begin{equation}
\int d\mathbf{k}' \, \left[L_0(\mathbf{k},\mathbf{k}') + L_1(\mathbf{k},\mathbf{k}')\right] \, \varphi(\omega,\mathbf{k}') = 0,
\label{KG-mom}
\end{equation}
where $L_0$ is a non-stochastic matrix with elements:
\begin{eqnarray}
L_{0}(\mathbf{k},\mathbf{k}') = \left(\frac{\omega^{2}}{u_{0}^{2}}
- \mathbf{k}^{2} - m^{2}_{0}\right) \delta(\mathbf{k} -\mathbf{k}'),
\label{L0-mom}
\end{eqnarray}
and $L_1(\mathbf{k},\mathbf{k}')$ is a stochastic matrix, with elements:
\begin{eqnarray}
L_{1}(\mathbf{k},\mathbf{k}') = \frac{1}{(2\pi)^d}\left( \mu(\mathbf{k} -\mathbf{k}') \, \frac{\omega^2}{u^2_0} -
\xi(\mathbf{k} -\mathbf{k}') \, m^{2}_{0}\right).
\label{L1-mom}
\end{eqnarray}
In $\mathbf{r}$-space, $L_0$ and $L_1$ read:
\begin{eqnarray}
L_{0}(\mathbf{r}) &=& \frac{\omega^{2}}{u_{0}^{2}} + \nabla^{2} - m^{2}_{0},
\label{nami32} \\
L_{1}(\mathbf{r}) &=& \mu(\mathbf{r})
\frac{\omega^{2}}{u_{0}^{2}} - \xi(\mathbf{r})\,m^{2}_{0} .
\label{nami33}
\end{eqnarray}
As in $\mathbf{k}$-space, they act as integral convolution operators. Note
that while $L_{1}$ is diagonal, $L_{0}$ is non-diagonal in
$\mathbf{r}$-space; yet, in $\mathbf{k}$-space the situation is the opposite. In
terms of these operators, the random Klein-Gordon equation can be
written in matrix form:
\begin{equation}
(L_{0}+\,L_{1}) \, \varphi(\omega,\cdot) = 0 .
\label{nami31}
\end{equation}
From this, one can define the full (operator valued) Green's function
$G$:
\begin{equation}
G=( L_{0} + L_{1} )^{-1}.
\label{green}
\end{equation}
Now, if one assumes that the noises are ``weak",
a natural perturbative
expansion for~$G$ in the form of a Dyson formula can be defined:
\begin{eqnarray}
G &=& G^{(0)} - G^{(0)}\, L_1 \,G^{(0)} + G^{(0)}\, L_1 \,G^{(0)}\, L_1 \,G^{(0)}
+ \, \cdots \nonumber \\
&=& G^{(0)} - G^{(0)} \, \Sigma \, G^{(0)} ,
\label{green1}
\end{eqnarray}
with the self-energy $\Sigma$ given by
\begin{equation}
\Sigma = L_1 - L_1 G_0 L_1 + \cdots \, ,
\label{self}
\end{equation}
where $G^{(0)}=\,L_{0}^{-1}$ is the unperturbed (operator valued) Green's
function which, in $\mathbf{k}$-space, can be written as
\begin{equation}
G^{(0)}(\omega,\mathbf{k}) = \frac{1}{\omega^2 - (\mathbf{k}^2 + m^2_0) + i\epsilon} .
\label{nami34}
\end{equation}
(Hereafter we take $u_0$ equal to unity.) A schematic representation
of the expansion in Eq.~(\ref{green1}) is shown in
Fig.~\ref{expansion}.
\begin{figure}[htb]
\centering
\includegraphics[scale=0.17]{01.jpg}
\caption{Perturbative expansion of $G$ in terms of the
disorder. The wavy lines represent generically the random functions
$\mu$ and $\xi$.}
\label{expansion}
\end{figure}
We note that $i G^{(0)} = \Delta_0$, where $\Delta_0$ is the noninteracting Feynman
propagator, which is the vacuum expectation value of the time ordered product
of the quantum field operators $\varphi$ at space-time points
$x=(t,\mathbf{x})$ and $x'=(t',\mathbf{x}')$:
\begin{equation}
\Delta_0(x,x') = \langle 0|T[\varphi(t,\mathbf{x}) \varphi(t',\mathbf{x}')]|0\rangle .
\label{defDelta}
\end{equation}
Let us write down explicitly a few terms of the perturbative
series in $\mathbf{r}$-space, up to second order in the random fields:
\begin{widetext}
\begin{eqnarray}
G(\omega,\mathbf{r},\mathbf{r}') &=& G^{(0)}(\omega,\mathbf{r},\mathbf{r}') -
\int d\mathbf{r}_1 \, G^{(0)}(\omega,\mathbf{r},\mathbf{r}_1)
\left[\mu(\mathbf{r}_1) \,\omega^2 - \xi(\mathbf{r}_1)\,m^2_0 \right]
G^{(0)}(\omega,\mathbf{r}_1,\mathbf{r}') \nonumber \\
&+& \int d\mathbf{r}_1 \int d\mathbf{r}_2 \, G^{(0)}(\omega,\mathbf{r},\mathbf{r}_2)
\left[\mu(\mathbf{r}_{2}) \, \omega^{2} - \xi(\mathbf{r}_{2})\,m^{2}_{0} \right]
\, G^{(0)}(\omega,\mathbf{r}_{2},\mathbf{r}_{1}) \nonumber\\
&\times& \left[\mu(\mathbf{r}_{1}) \, \omega^{2} - \xi(\mathbf{r}_{1}) \, m^{2}_{0} \right]
\, G^{(0)}(\omega,\mathbf{r}_{1},\mathbf{r}') + \cdots \,.
\label{expans}
\end{eqnarray}
\end{widetext}
A pictorial representation in $\mathbf{r}$-space of this perturbative expression
in terms of diagrams can be done. They correspond to diagrams of multiple-scattering of the
Klein-Gordon field on random inhomogeneity scatterers located at positions
$\mathbf{r}_{1}, \mathbf{r}_{2},\,\cdots \,$. In $\mathbf{k}$-space, similar diagrams correspond to
multiple interactions of Fourier components of the Klein-Gordon field and
of the random inhomogeneities.
Let us now perform the averages over the random processes that appear in the
definition of the propagator $G$ given in Eq.~(\ref{expans}). Note that, because of
the Gaussian nature of noise correlations, terms with an odd-number
of noise fields do not contribute to the two-point function, and the
first nonzero correction to the two-point function comes from the
averages over the third term in Eq.~(\ref{expans}). Therefore,
up to second order in the noise fields, we get
\begin{eqnarray}
{G}^{(1)}(\omega,\mathbf{r},\mathbf{r}') & \equiv & \langle \, G(\omega,\mathbf{r},\mathbf{r}') \rangle_{\mu\xi}
\nonumber \\
&=& G^{(0)}(\omega,\mathbf{r},\mathbf{r}') + \bar G^{(1)}(\omega,\mathbf{r},\mathbf{r}'),
\label{onelooppropagator}
\end{eqnarray}
with the one-loop correction $\bar G^{(1)}(\omega,\mathbf{r},\mathbf{r}')$, represented pictorially
in Fig.~\ref{selfenergy} and given in terms of its Fourier transform
$\bar G^{(1)}(\omega,\mathbf{k})$ as
\begin{equation}
\bar G^{(1)}(\omega,\mathbf{r},\mathbf{r}') = \int \frac{d\mathbf{k}}{(2\pi)^d} \,
\bar G^{(1)}(\omega,\mathbf{k}) \, e^{i\mathbf{k}\cdot (\mathbf{r} -\mathbf{r}')}.
\end{equation}
This quantity can be written in terms of a self-energy $\Sigma(\omega,\mathbf{k})$ --
defined consistently with Eq.~(\ref{green1}) -- as:
\begin{equation}
\bar G^{(1)}(\omega,\mathbf{k}) = - G^{(0)}(\omega,\mathbf{k}) \, \Sigma(\omega,\mathbf{k}) \,
G^{(0)}(\omega,\mathbf{k}) ,
\end{equation}
with
\begin{equation}
\Sigma(\omega, \mathbf{k}) = - \left(\sigma^2_\mu \omega^4
+ \sigma^2_\xi m_0^4\right) \, \alpha(\omega) ,
\label{Sigma}
\end{equation}
where
\begin{equation}
\alpha(\omega) = \int \frac{d\mathbf{k}}{(2\pi)^d} \, G^{(0)}(\omega,\mathbf{k}) .
\label{alpha}
\end{equation}
Note that at one-loop order, the self-energy is actually independent
of $\mathbf{k}$. From Eq.~(\ref{green1}), one has $G^{-1} = G^{(0)-1} +
\Sigma$ and, therefore, in the $\mathbf{k}$-representation
\begin{eqnarray}
[G^{(1)}(\omega,\mathbf{k})]^{-1} &=& [G^{(0)}(\omega,\mathbf{k})]^{-1} + \Sigma(\omega,\mathbf{k})
\nonumber \\
&=& \omega^ 2 - \mathbf{k}^2 - m^2,
\end{eqnarray}
where
\begin{eqnarray}
m^2 &=& m^2_0 - \Sigma(\omega,\mathbf{k}) \nonumber \\
&=& m^2_0 + \left(\sigma^2_\mu \omega^4
+ \sigma^2_\xi m_0^4\right) \, \alpha(\omega) .
\label{mass}
\end{eqnarray}
\vspace{0.25cm}
\begin{figure}[h]
\centering
\includegraphics[scale=0.3]{02.jpg}
\caption{The one-loop correction to the two-point causal function
due to the random inhomogeneities.}
\label{selfenergy}
\end{figure}
From this result, it is clear that the effect of randomness is to
turn a free, conventional scalar quantum field theory into a
self-interacting theory, with a self-interaction qualitatively
similar to $\lambda \varphi^{4}$. The induced $\lambda \varphi^{4}$
theory is a consequence of the Gaussian nature of the functions
$\mu(\vec{r}\,)$ and $\xi(\vec{r}\,)$; a more general polynomial
self-interacting model is obtained if non-Gaussian noises are used.
The integral is divergent because of the white-noise nature of the
noise; a colored noise correlation function would lead to a finite
integral.
As mentioned earlier in the text, the induced coupling generated is
frequency-dependent, $\lambda(\omega) \sim \sigma^2_\mu \omega^4 +
\sigma^2_\xi m_0^4$. Also, for $d=3$ one can isolate the finite
part from the integral making use of the identity
\begin{equation} \frac{|\mathbf{k}|^2}{|{\mathbf{k}}|^2 - (\omega^2 - m^2_0)} = 1 +
\,\frac{\omega^2 - m^2_0}{|\mathbf{k}|^2 - (\omega^2 - m^2_0)},
\label{separat} \end{equation}
so that the finite part of the self-energy, $\Sigma_f(\omega,\mathbf{k})$,
is given by
\begin{eqnarray} \Sigma_{f}(\omega,\mathbf{k}) &=& \left(\sigma^2_\mu \omega^4 +
\sigma^2_\xi m_0^4\right) \int^{\infty}_0 \frac{d|\mathbf{k}|}{2\pi^2}\,
\frac{\omega^2 - m^2_0}{\omega^2 - |{\mathbf{k}}|^2 - m^2_0} \nonumber \\
&=& \frac{1}{4\pi}\left(\sigma^2_\mu \omega^4
+ \sigma^2_\xi m_0^4\right) |\omega^2 - m^2_0|^{1/2} {\cal A},
\end{eqnarray}
with
\begin{equation}
{\cal A} = \Biggl\{
\begin{array}{ll}
+1\,, & \hspace{0.25cm} m^2_0 > \omega^2 ,\\
-i \,,
& \hspace{0.25cm} \omega^2 > m^2_0 .
\end{array}
\label{f}
\end{equation}
The conclusion from this calculation is that the random fluctuations
induce a self-energy with a width, which gives a finite life-time
for the excitations.
Since randomness, as mentioned previously, led to an induced
coupling, one may ask on the form of the perturbative corrections to
the four-point function $G_4$. The reason for this particular
interest is the following. As well known, given the four-point
function, one can define the one-particle-irreducible ($1$PI)
four-point connected Green's function without the external legs
$\Gamma_4$:
\begin{equation}
\Gamma_{4}(k_1,k_2,k_3,k_4) =
G_{4}(k_1,k_2,k_3,k_4)\prod^4_{i=1}\Gamma_{k_i}, \label{vertex}
\end{equation}
where
\begin{equation} \Gamma_k = \omega^2-(\mathbf{k}^2 + m_0^2). \label{Gamma-k} \end{equation}
In a covariant quantum field theory, one may compute the
renormalized coupling constant from the proper vertex function,
under certain conditions. So, even though we are not dealing with a
covariant field theory, it is natural to expect that such function
can bring us some information on the nature of the induced coupling.
Since the random field equation is linear in the field variable, the
corresponding action of the system is quadratic in the field,
resulting in a Gaussian generating functional. Then, the quantum
$n$-point Green's functions of the system are products of two-point
Green's functions. In particular the four-point Green's function
will be of the form:
\begin{eqnarray}
G_4(x_1,x_2,x_3,x_4) &=& \langle G(x_1,x_2)G(x_3,x_4)
\rangle_{\mu\xi} \nonumber\\
&+& \langle G(x_1,x_3)G(x_2,x_4) \rangle_{\mu\xi} \nonumber\\
&+& \langle G(x_1,x_4)G(x_2,x_3) \rangle_{\mu\xi}. \label{g4a}
\end{eqnarray}
As before, $x = (t,\mathbf{x})$. Fig.~\ref{g12g34} presents a graphical
representation of this expression.
\begin{figure}[h]
\includegraphics[scale=0.19]{03.jpeg}
\caption{The four-point Green function up to one-loop, after noise
averaging. `1PR graphs' means one-particle reducible
graphs.}
\label{g12g34}
\end{figure}
In Appendix~\ref{app:G2-G4}, Eq.~(\ref{g4a}) is computed up to one
loop; the result can be written as $G_4(x_1,x_2,x_3,x_4) =
G_4^{(0)}(x_1,x_2,x_3,x_4) + G_4^{(1)}(x_1,x_2,x_3,x_4)$, where
$G_{4}^{(0)}$ and $G_4^{(1)}$ are the tree-level and one-loop
contributions, respectively. The noise averaging leading to the
tree-level four-point function $G_{4}^{(0)}$ is illustrated in
Fig.~\ref{treelevel4point1} and the one-loop
corrections $G_4^{(1)}$ are pictorially represented in Figs.~\ref{oneloop4point} and \ref{gra_apend}.
\begin{figure}[h]
\centering
\includegraphics[scale=0.25]{04.jpg}
\caption{The disorder-induced, tree-level four-point Green's
function.}
\label{treelevel4point1}
\end{figure}
From Eq. (\ref{vertex}) and the results derived in the Appendix, one
can express the tree-level proper vertex function as:
\begin{eqnarray} \Gamma^{(0)}_4(k_1,k_2,k_3,k_4) &=&
(2\pi) \Bigl[ \delta(\omega_3 + \omega_4)
\left(\sigma^2_\mu \omega^2_2 \omega^2_4 +
\sigma^2_\xi m^4_0\right) \nonumber \\
&+& \delta(\omega_2 + \omega_4) \left(\sigma^2_\mu \omega^2_3
\omega^2_4 + \sigma^2_\xi m^4_0\right)\nonumber \\ &+& \delta(\omega_2 +
\omega_3) \left(\sigma^2_\mu \omega^2_3 \omega^2_4 + \sigma^2_\xi
m^4_0\right) \Bigr]. \label{Gamma4-0}
\end{eqnarray}
As mentioned earlier, the proper understanding of the nature of the induced interacting
theory requires the study of the four-point Green's function.
We will not dwell into such a study here. Besides being involved (see the
Appendix for the one-loop contribution), such a study is out of the
scope of the present paper and therefore we reserve it for a future
publication.
\begin{figure}[h]
\centering
\includegraphics[scale=0.25]{05.jpg}
\caption{One-loop corrections of first kind to the four-point Green's function.}
\label{oneloop4point}
\end{figure}
The most striking feature that comes out from these calculations is
that the induce coupling is qualitatively similar to a $\lambda
\phi^4$ theory, with a frequency-dependent coupling. The important
meaning of this is that if one starts with an interacting model, the
induced coupling due to the random scatterings in the medium changes
the value of the renormalized coupling constant of the model. As a
result, a relevant question to be answered is to know what would be
the sign of this effective coupling. For instance, if it turns out
to be negative, it is clear that we would have a decrease in the
value of the renormalized coupling constant. Furthermore, as well
known, a negative coupling constant can be a source of instability
in field theory, which means that one should look for alternative
solutions to circumvent the above mentioned problem. One possibility
is to discuss the ground state properties of the system in a theory
with metastable vacua. This can be obtained in non-simply connected
manifolds. For simplicity one might assume periodic boundary
conditions in all $d+1$ directions, with compactified lengths $L_1$,
$L_2$,..,$L_{d+1}$. At this point we remark that in theories defined
on a non-simply connected Euclidean space, one of the compactified
length is related to the temperature, remembering the
Kubo-Martin-Schwinger (KMS) condition \cite{matsu}. Therefore in a
Euclidean theory describing bosons, we have to impose periodic
boundary condition in one compactified direction. In the other $d$
directions, we are free to choose any boundary conditions. However,
imposing periodic boundary conditions in all compactified dimensions
enables us to maintain translational invariance~\cite{ford}.
Non-translational invariant systems were studied in Refs. \cite{sy}.
We take this opportunity to call attention for the results obtained
by Arias and co-workers \cite{arias}, where the thermodynamic of the
massless self-interacting scalar field model with negative coupling
constant was analyzed. We believe that such results from this
reference can be useful to proceed with the investigations of the
consequences of impurities in a relativistic model.
\begin{figure}[h]
\includegraphics[scale=0.29]{06.jpg}
\caption{One-loop corrections of second kind to the four-point Green's function.}
\label{gra_apend}
\end{figure}
\section{Conclusions}
\label{sec:concl}
Recently it was proposed that an analog model for quantum
gravity effects can be implemented in condensed matter physics when
the fluid is a random medium \cite{krein1}. In order to study the
wave propagation of the elementary excitations in the fluid when the
acoustic wave propagates in random medium, it was considered that
there are no random fluctuation in the density, but only in the
reciprocal of the bulk modulus. In this analog model of quantum
gravity with a colored noise a free quantum field theory describing
phonons becomes a self-interacting model.
In this paper we studied a massive scalar field theory in a
disordered medium. We determined the perturbation theory in
annealed-like disordered medium, with two distinct random functions.
After performing the random averages over the noise function a free
scalar quantum field theory became the $\lambda\varphi^{4}$
self-interacting model.
The one-loop two and four-point functions were presented. We obtained
that the induced coupling constant is frequency dependent. Note
that although we employed Gaussian
functions, it is not difficult to extend the method to non-Gaussian
functions by means of a cluster expansion \cite{cluster}.
There exist several possibilities to use the above results in analog
models. For example, a random fluid with a supersonic acoustic flow
can be an analog model that allows one to discuss the effect of the
fluctuation of the geometry in the Hawking radiation. Since
Bose-Einstein condensate is a natural candidate to produce acoustic
black-holes, in order to introduce randomness, i.e., the effects of
the metric fluctuations in the acoustic black-hole, we have to
investigate how superfluid and Bose-Einstein condensate behave in
the presence of disorder. Sound waves in Bose-Einstein condensate,
described by random wave equation must reproduce physics beyond the
semi-classical approximation in an analog model.
An interesting course of action would be the search of a possible
localization of the Hawking thermal flux in such analog model, since
it is well known that the effect of the impurities is the
localization of classical waves and elementary excitations
\cite{anderson}. In conclusion, the study of quantum fields in
disorder medium in the analog models scenario introduces new
experimental and theoretical challenges. The study of the
relativistic Bose-Einstein condensation in the presence of a random
potential \cite{new}, and also the consequences of introducing
metric fluctuations in the analog model proposed in Ref.~\cite{Fagnocchi:2010sn}
is under investigation by the authors.
\acknowledgments
Work partially financed by CAPES, CNPq and FAPESP (Brazilian agencies).
|
\section{Introduction}
Studies of trilinear couplings between the gauge bosons ($W, Z, \gamma$) test the standard model (SM) description of gauge sector interactions and provide sensitivity to physics beyond the SM through examination of production rates and kinematics~\cite{CDFZllAGC,CDFWWAGC,D0ZllAGC,D0ZnunuAGC,D0WZAGC, LEPEWWG}. In the case of neutral couplings, $ZZ\gamma$ and $Z\gamma\gamma$ vertex interactions vanish at tree level and, while allowed via internal particle loops, are highly suppressed in the SM. However, these trilinear gauge couplings can be non-negligible if loop contributions occur via non-SM particles. Models such as those incorporating compositeness or supersymmetry can alter the predicted cross section and production kinematics of \ensuremath{Z\gamma}\ events~\cite{newZgPhys,tgc,prd41_1990_1476,prd47_1993_4889}.
In the SM, given the suppression of $ZZ\gamma$ and $Z\gamma\gamma$ couplings, the production of \ensuremath{Z\gamma}\ events is dominated by production of a $Z$ boson along with the radiation of a photon off either an incoming parton or a $Z$ decay product. These production mechanisms are interesting in their own right, serving as an important background
to searches for new physics (e.g.~in gauge-mediated supersymmetry breaking models~\cite{gmsb}) and Higgs boson searches.
In this Letter, the production properties of \ensuremath{Z\gamma}\ events are compared
to SM predictions, and limits are set on anomalous trilinear gauge couplings.
The measurements of $Z \gamma$ couplings are performed with
\ensuremath{p\bar{p}}\ collision data at $\sqrt{s}$ = 1.96 TeV from the Tevatron Collider using the Collider Detector at Fermilab (CDF).
We seek two types of \ensuremath{Z\gamma}\ events: those where the $Z$ decays to charged leptons (by identifying
lepton candidate pairs and a prompt photon~\cite{photonID} with large transverse energy \ensuremath{E_{T}}~\cite{coords}), and those where the $Z$ decays to
neutrinos (by identifying an event with only a solitary, prompt, high-\ensuremath{E_{T}}\ photon). In the former case, data corresponding to an integrated luminosity of 5.1 \ensuremath{\mathrm{fb^{-1}}}\ are used; in the latter, 4.9 \ensuremath{\mathrm{fb^{-1}}} . These measurements use over twice as much data as the previous published CDF result~\cite{CDFZllAGC} and incorporate looser muon selection requirements. As no significant disagreement is found between the SM prediction and the data, we set limits that are not only far more restrictive than those measured in~\cite{CDFZllAGC}, but are approximately half the magnitude of the previous best published limits~\cite{D0ZnunuAGC}.
In beyond-the-SM scenarios with enhanced \ensuremath{Z\gamma}\
couplings, not only does the \ensuremath{Z\gamma}\ production cross section increase, but the photon \ensuremath{E_{T}}\ spectrum is modified due to an enhancement in the production of high-\ensuremath{E_{T}}\ photons~\cite{prd47_1993_4889}. We take advantage of this enhancement by
comparing the photon \ensuremath{E_{T}}\ distribution in data to both SM and beyond-the-SM predictions. Binned maximum likelihood measurements of the coupling parameters that describe \ensuremath{Z\gamma}\ interactions in the Lagrangian are performed. We calculate separate likelihoods for the \ensuremath{Z \rightarrow l^+l^-}\ and \ensuremath{Z \rightarrow \nu \bar{\nu}}\ samples and combine the likelihoods to produce the final result.
The CDF detector is covered in detail elsewhere~\cite{CDF,CDF_run2}. The transverse momenta (\ensuremath {p_{{T}}}) of charged particles
are measured by an eight-layer silicon strip detector~\cite{SVX}
and a 96-layer drift chamber (COT)~\cite{COT} inside a 1.4\,T magnetic field.
The COT provides tracking coverage with high efficiency for the pseudorapidity range $|\eta|<1$~\cite{coords}.
Electromagnetic and hadronic calorimeters
surround the tracking system. They are segmented in a projective tower geometry and measure
the energies of charged and neutral particles in the central ($|\eta|<1.1$) and forward ($1.1<|\eta|<3.6$) regions.
Each calorimeter has an electromagnetic shower profile detector
positioned at the shower maximum~\cite{showerMax}.
The calorimeters are surrounded by drift chambers that detect muons.
The measurements of anomalous trilinear gauge coupling parameters in the \ensuremath{Z \rightarrow l^+l^-}\ and $Z \rightarrow \nu\bar{\nu}$ decay channels differ both in event selection and background estimation. For the \ensuremath{Z \gamma \rightarrow l^+l^- \gamma}\ decay channel we identify events containing \ensuremath{Z \rightarrow \mu^+ \mu^-}\ and \ensuremath{Z \rightarrow e^+e^-}\ candidates along with prompt photon candidates with $\ensuremath{E_{T}^{\gamma}} > 50$ GeV. According to experiments performed on simulated events, this choice of \ensuremath{E_{T}^{\gamma}}\ requirement maximizes the ability of the analysis to exclude anomalous couplings assuming SM physics, although a serious loss in sensitivity only occurs if the \ensuremath{E_{T}^{\gamma}}\ requirement is placed at 100 GeV or higher. The previous CDF analysis used a much less restrictive requirement of $\ensuremath{E_{T}^{\gamma}} > 7$ GeV, as the \ensuremath{Z\gamma}\ cross section was being measured in addition to trilinear gauge coupling parameters~\cite{CDFZllAGC}; additionally, placing the cut at 50 GeV allows for a control region to be based off of lower-$E_T$ photons. Event selection starts with inclusive muon (electron) triggers that require muon $\ensuremath {p_{{T}}} > 18$ \ensuremath{\GeV\!/c}\ (electron $\ensuremath{E_{T}} > 18 $ GeV). For electrons, a track must be reconstructed in the COT or in the silicon detector; additionally, the energy deposited by the candidate in the calorimeter must be isolated. For muons, a track must be reconstructed in the COT; additionally, no more than a few GeV of energy may be deposited in the calorimeters so that the candidate is compatible with a minimum ionizing particle. The two lepton candidates must correspond to the same flavor, with a requirement of $\ensuremath {p_{{T}}} > 20$ \ensuremath{\GeV\!/c}\ ($\ensuremath{E_{T}} > 20$ GeV) on one muon (electron) candidate and $\ensuremath {p_{{T}}} > 10$ \ensuremath{\GeV\!/c}\ ($\ensuremath{E_{T}} > 10$ GeV) on the other; furthermore, if the charges of both leptons are well-measured, the signs of these charges must be opposite. Studies of the invariant mass distributions of the two lepton candidates indicate that we retain a very high purity of $Z$ bosons (over $99\%$) despite the loose selection requirements.
Once we have selected events with \ensuremath{Z \rightarrow l^+l^-}\ candidates, we look for isolated photons that pass standard CDF requirements~\cite{photonIsolation} in the central region $(|\eta| < 1.1)$ with $\ensuremath{E_{T}^{\gamma}} > 50$ GeV and are well separated from the $Z$ decay leptons ($\Delta R_{\ell\gamma} > 0.7$, with $\Delta R = \sqrt{ (\phi_l - \phi_{\gamma})^{2} + (\eta_l - \eta_{\gamma})^{2}}$). Additionally, we require that the two lepton candidates and the photon candidate form a three-body invariant mass greater than 100 \ensuremath{\GeV\!/c^2}\ in order to discriminate against events where the photon is radiated from one of the leptons from the $Z$ boson decay. The estimated contribution of SM $Z\gamma$ events is derived from Monte Carlo (MC) simulations that use the Baur-Berger package at the generator level~\cite{prd47_1993_4889} and \textsc{pythia}~\cite{pythia} for particle showering. This method yields a prediction of 87.2 $\pm$ 7.8 \ensuremath{Z\gamma}\ events that pass our selection requirements, where the uncertainty is dominated by the uncertainty on the luminosity and the predicted cross section. The non-\ensuremath{Z\gamma}\ events that pass these selection requirements result from hadronic jets being reconstructed as prompt photons and leptons (more commonly electrons). This background is estimated by calculating separate probabilities for a jet to mimic a photon or lepton as a function of jet \ensuremath{E_{T}}, and applying them to jets in events to which all our requirements have been applied except those pertaining to the mimicked particle. For photons and electrons, these probabilities are calculated by taking the ratio of the number of individual photon or electron candidates to the number of jets in a sample of data events where only the presence of at least one jet is required. The number of photon and electron candidates is corrected for the expected contribution of true photons or electrons in this sample. We estimate the probability for a false muon candidate from the number of dimuon $Z$ decay candidates in which both muon candidates have the same charge. Overall, the non-\ensuremath{Z\gamma}\ background contribution is very low: of the 91 events that pass our requirements, less than one event involving a mimicked photon or lepton is expected.
In order to identify $Z\gamma$ candidate events in the $Z \rightarrow \nu\bar{\nu}$ decay channel, we require solitary high-\ensuremath{E_{T}}\ photons and a transverse energy imbalance~\cite{MET_defn} in the detector. These events must pass a trigger requirement of an electromagnetic cluster with $\ensuremath{E_{T}} > 25$ GeV and $|\eta| < 1.1$ as well as missing transverse energy in excess of 25 GeV. For our signal region we require $\ensuremath{E_{T}^{\gamma}} > 100$ GeV, a threshold optimized in the same manner as the \ensuremath{Z \rightarrow l^+l^-}\ case. To account for the neutrinos we require a transverse energy imbalance of at least 50 GeV. In order to discriminate against $W$ boson contamination in our sample, we reject events containing any tracks with $\ensuremath {p_{{T}}} > 10$ GeV, any electron candidates with $\ensuremath{E_{T}} > 15$ GeV, or any muon candidates with $\ensuremath {p_{{T}}} > 10$ \ensuremath{\GeV\!/c}. Additionally, we reject events that have any jets with $\ensuremath{E_{T}} > 15$ GeV in order to reduce the mismeasurement of missing transverse energy. The primary SM source for photons passing these requirements is $Z\gamma$ events in which the $Z$ has decayed to a pair of neutrinos, as shown in Table~\ref{tab:Znunu_bkgnds}. The method of estimating the expected number of $Z\gamma$ events is the same as that used for the \ensuremath{Z \rightarrow l^+l^-}\ candidate sample.
The primary source of non-\ensuremath{Z\gamma}\ events in the final $Z\gamma \rightarrow \nu\bar{\nu}\gamma$ candidate sample is cosmic ray interactions. High-\ensuremath{E_{T}}\ photons from cosmic rays leave large transverse energy imbalances in our detector, mimicking the presence of neutrinos. Therefore, additional event requirements are applied to reduce the contributions from cosmic ray events. First, we require that the energy deposited in the electromagnetic calorimeter appear within a timing window centered on the \ensuremath{p\bar{p}}\ interaction. Second, we use a relevance vector machine (RVM) multivariate discriminator~\cite{RVM} to distinguish whether a photon came from a collision or a non-collision source; the three inputs used for the RVM discriminator are the $\phi$-angle between the photon candidate and the closest muon candidate (if any), the ratio of energies from the photon candidate in the electromagnetic and hadronic calorimeters, and the ratio of energies from the electromagnetic shower profile detector and the electromagnetic calorimeter. We use photons outside the timing window to train the RVM for non-collision sources, and photons recoiling against jets to train for collision sources. The RVM discriminator reduces the contribution from cosmic ray events by an additional 90\%. Finally, we require the event to have a reconstructed vertex of at least three tracks from a \ensuremath{p\bar{p}}\ interaction. After applying these selection requirements, we have 85 candidate events in our sample. Despite the anti-cosmic ray requirements, cosmic ray events remain the second largest contributor to our sample, after \ensuremath{Z\gamma}\ events.
We model two other major categories of non-\ensuremath{Z\gamma}\ events: one in which a charged lepton from $W \rightarrow e\nu$, $W \rightarrow \mu\nu$, or $W \rightarrow \tau\nu$ decay is reconstructed as a photon, and the other in which a true photon is produced but another object (e.g.~a lepton) is lost or only partially reconstructed, creating a large transverse energy imbalance. For the former case, the rate at which electrons are reconstructed as photons in the detector has been calculated using events with an electron and photon pair candidate that has an invariant mass near the mass of the $Z$, i.e., events in which the photon candidates are almost entirely electrons in actuality. The rate at which $\mu$'s and $\tau$'s are reconstructed as photons is taken from MC. For the latter case, which encompasses $W\gamma \rightarrow l \nu \gamma$ events in which a lepton is lost and $\gamma\gamma$ events in which a photon is lost, a two-step process is used to calculate the expected number of events. First, events in data are selected such that we obtain a very pure sample of one of the aforementioned event types in which there is no lost object. Then, we calculate the fraction of the corresponding events in MC in which an object is not detected, and this fraction is used to scale the photon \ensuremath{E_{T}}\ distribution of the data events so as to provide an estimate of this background's photon \ensuremath{E_{T}}\ distribution in the signal sample. An exception to this method is the case in which a $\tau$ is lost; due to the difficulty of reliably identifying $\tau$ candidates, this background is estimated purely from MC simulations. Further details on these methods of background prediction can be found in~\cite{LED}, a CDF analysis which used very similar event requirements. We see excellent agreement between the SM predictions and the data in the control regions of $15 < \ensuremath{E_{T}^{\gamma}}\ < 40$ GeV (\ensuremath{Z \rightarrow l^+l^-}\ case) and $70 < \ensuremath{E_{T}^{\gamma}}\ < 100$ GeV (\ensuremath{Z \rightarrow \nu \bar{\nu}}\ case).
\begin{table}[!hbtp]
\begin{center}
\begin{tabular}{lc}
\hline
\hline
Process & Events \\
\hline
$Z\gamma \rightarrow \nu \bar{\nu}\gamma$ & $52.8 \pm 4.6$ \\
cosmics & $14.9 \pm 1.4$ \\
$W \rightarrow e\nu$ & $3.9 \pm 0.8$ \\
$W \rightarrow \mu/\tau \nu \gamma$ & $1.6 \pm 0.3$ \\
$W\gamma \rightarrow e\nu\gamma $ & $1.1 \pm 1.1$ \\
$W\gamma \rightarrow \mu\nu\gamma $ & $1.8 \pm 1.3$ \\
$W\gamma \rightarrow \tau\nu\gamma $ & $4.5 \pm 1.3$ \\
$\gamma \gamma$ & $5.3 \pm 1.9$ \\
\hline
SM Total & $85.9 \pm 5.6$ \\
\hline
Data & 85 \\
\hline
\hline
\end{tabular}
\caption{SM expected contributions to the $Z\gamma \rightarrow \nu\bar{\nu}\gamma$ candidate sample. Uncertainties shown are systematic only and thus exclude the statistical uncertainties. }
\label{tab:Znunu_bkgnds}
\end{center}
\end{table}
Assuming gauge and Lorentz invariance, eight parameters are needed to describe $Z\gamma$ couplings, denoted by $h^V_{i0}$ where $V$ is either a $Z$ or a $\gamma$ and the index $i$ runs from 1 to 4; these parameters are all zero at tree level~\cite{prd47_1993_4889}. Interaction amplitudes are linear in these parameters. Indices 1 and 2 represent CP-violating terms while indices 3 and 4 represent CP-conserving terms. We assume CP conservation in these interactions by setting $h^V_{10} = h^V_{20} = 0$ and we investigate the possibility of non-zero values for $h^V_{30}$ and $h^V_{40}$, corresponding to electric dipole and magnetic quadrupole transition moments~\cite{dipole_quadrupole}. In order to preserve unitarity at large incoming parton center-of-mass energy $\sqrt{\hat{s}}$, an $\hat{s}$-dependent form factor is used to suppress the coupling, constructed as $h^V_i(\hat{s}) = {{h^V_{i0}}\over{(1+\hat{s} / \Lambda^2)^n}}$, where $n = i$ for $h^V_{30}$ and $h^V_{40}$~\cite{prd47_1993_4889}. The parameter $\Lambda$ describes the predicted energy scale of the new physics that creates anomalous \ensuremath{Z\gamma}\ couplings.
For a given set of anomalous coupling parameter values, we compute a likelihood for the \ensuremath{E_{T}^{\gamma}}\ distribution. Hence, we have $\prod_{j=1}^{N} L(x_j | h^V_{i})$, where $x_j$ represents the number of entries in the $j$th of $N$ bins in our \ensuremath{E_{T}^{\gamma}}\ distribution and $h^V_{i}$ denotes the coupling parameter being measured (the other three being held fixed at zero). The bin-by-bin likelihood $L$ is simply the Poisson probability of the number of observed entries given the expected number of entries for the value of $h^V_{i}$. This limit method requires a predicted \ensuremath{E_{T}}\ distribution for each combination of the four coupling parameters. To create these distributions, we produce \ensuremath{Z\gamma}\ MC events at the generator level using the Baur-Berger package~\cite{prd47_1993_4889}. Modeling the particle showering process and detector response in MC separately for every parameter value is computationally impractical. To mimic fully-simulated MC events we first determine the efficiency for a generated event to pass all of the event requirements as a function of generator-level \ensuremath{E_{T}^{\gamma}}\ and $|\eta^{\gamma}|$; these functions are derived from a SM MC sample which has used the full simulation of the detector. Due to the correlation between \ensuremath{E_{T}^{\gamma}}\ and the $Z$ kinematics, we create and combine separate templates for the cases of central-central, central-forward, and forward-forward lepton pairs, ``central'' denoting $0 < |\eta| < 1.1$ and ``forward'' denoting $1.1 < |\eta| < 2.8$. We then apply this efficiency function to generator-level MC samples to get the expected \ensuremath{E_{T}^{\gamma}}\ distributions. The final prediction is the sum of this \ensuremath{Z\gamma}\ prediction with the predictions of the non-\ensuremath{Z\gamma}\ backgrounds.
In Fig.~\ref{fig:et_comps}, for both the \ensuremath{Z \rightarrow l^+l^-}\ and \ensuremath{Z \rightarrow \nu \bar{\nu}}\ cases, the \ensuremath{E_{T}^{\gamma}}\ distributions in data are
compared to the SM prediction and beyond-the-SM predictions; it can be seen that the production of high-\ensuremath{E_{T}}\ photons is far more likely in the beyond-the-SM cases compared to the SM case.
The uncertainty bands shown for the SM predictions illustrate the systematic uncertainties on
those predictions.
These uncertainties are dominated by the 7\% uncertainty on the theoretical \ensuremath{Z\gamma}\
cross section~\cite{theoryerror} and the 6\% uncertainty on the luminosity~\cite{lumerror}; the
other sources are the reconstructed photon's energy scale and efficiency, as well as uncertainties
on the number of non-\ensuremath{Z\gamma}\ background events. The effect of these systematic uncertainties on the
limits is negligible --- of the order of a couple of percent of the limit values.
\begin{figure}
\begin{center}
\includegraphics[width=1.0 \columnwidth]{./dat_vs_pred_Zll_signal_PRL_v9}
\includegraphics[width=1.0 \columnwidth]{./dat_vs_pred_Znunu_signal_PRL_v9}
\caption{Comparison of the measured \ensuremath{E_{T}^{\gamma}}\ distribution with the predicted distributions from both the SM and beyond-the-SM scenarios for $Z \to l^+l^-$ (top) and $Z \to \nu\bar{\nu}$ (bottom) candidate samples, at $\Lambda = 1.5$\ TeV. The beyond-the-SM scenarios chosen here can be excluded at 95\% Bayesian credibility level in each sample. Note the greatest difference in the SM and beyond-the-SM cases is found offscale at \ensuremath{E_{T}^{\gamma}} $> 200$ GeV; the lack of data events in this region indicates good agreement with the SM.
\label{fig:et_comps}}
\end{center}
\end{figure}
With the likelihood distribution for a given $h^V_i$, taking a flat Bayesian prior in $h^V_i$ allows us to set Bayesian credibility limits on the parameter. These limits are defined as the values of $h^V_i$ which demarcate the central 95\% of the integral of the likelihood distribution. The resulting allowed ranges for the strength of anomalous couplings are shown in Table~\ref{tab:AGClimits}. The values $\Lambda$ = 1.2 TeV and $\Lambda = 1.5$ TeV have been chosen to allow direct comparisons with earlier CDF~\cite{CDFZllAGC} and D0~\cite{D0ZnunuAGC} results, respectively. We see no evidence for anomalous couplings.
In conclusion, we find that the \ensuremath{E_{T}^{\gamma}}\ distribution of photons produced in association with $Z$ bosons in both the \ensuremath{Z \rightarrow \nu \bar{\nu}}\ and \ensuremath{Z \rightarrow l^+l^-}\ decay channels in a data sample corresponding to an integrated luminosity of approximately 5 \ensuremath{\mathrm{fb^{-1}}}\ is consistent with SM couplings. We place 95\% Bayesian credibility limits of $|h_3^{\gamma,Z}| < 0.027 $ and $|h_4^{\gamma,Z}| < 0.0013$ on the CP-conserving \ensuremath{Z\gamma}\ couplings at $\Lambda = 1.2$\ TeV and $|h_3^{\gamma,Z}| < 0.022 $ and $|h_4^{\gamma,Z}| < 0.0009$ at $\Lambda = 1.5$ TeV; these are significantly tighter constraints on beyond-the-SM contributions than those provided by previously measured limits.
\begin{table}
\begin{center}
\begin{tabular}{ccc}
\hline
\hline
Parameter & ($\Lambda = 1.2$ TeV) & ($\Lambda = 1.5$ TeV)\\
\hline
$h_3^Z$ & $-0.024$, 0.027 & $-0.020$, 0.021 \\
$h_4^Z$ & $ -0.0013$, 0.0013 & $-0.0009$, 0.0009 \\
\hline
$h_3^{\gamma}$ & $-0.026$, 0.026 & $-0.022$, 0.020 \\
$h_4^{\gamma}$ & $ -0.0012$, 0.0013 & $-0.0008$, 0.0008 \\
\hline
\hline
\end{tabular}
\caption{Allowed ranges (95\% Bayesian credibility limits) of anomalous $Z\gamma$ couplings for $\Lambda=1.2$ and 1.5 TeV using notation from Ref.~\cite{prd47_1993_4889}. Each parameter's limits are set assuming the other three parameters have values fixed at 0.
\label{tab:AGClimits}}
\end{center}
\end{table}
We thank the Fermilab staff and the technical staffs of the participating institutions for their vital contributions. This work was supported by the U.S. Department of Energy and National Science Foundation; the Italian Istituto Nazionale di Fisica Nucleare; the Ministry of Education, Culture, Sports, Science and Technology of Japan; the Natural Sciences and Engineering Research Council of Canada; the National Science Council of the Republic of China; the Swiss National Science Foundation; the A.P. Sloan Foundation; the Bundesministerium f\"ur Bildung und Forschung, Germany; the World Class University Program, the National Research Foundation of Korea; the Science and Technology Facilities Council and the Royal Society, UK; the Institut National de Physique Nucleaire et Physique des Particules/CNRS; the Russian Foundation for Basic Research; the Ministerio de Ciencia e Innovaci\'{o}n, and Programa Consolider-Ingenio 2010, Spain; the Slovak R\&D Agency; and the Academy of Finland.
|
\section{Introduction}
Galaxies are the primary tracers by which astronomers map and measure
the Universe, so understanding how galaxies form and evolve is a
longstanding central goal of astronomy. The physics of galaxy
formation involves a diverse set of processes spanning an enormous
dynamic range, from black holes and stellar physics on sub-parsec
scales to cosmology and large-scale structure on gigaparsec scales.
As a consequence, despite rapidly advancing observations of galaxies
across cosmic time, a comprehensive theory for the formation and
evolution of galaxies that can explain all their observed properties
remains elusive.
As galaxies are essentially collections of stars, understanding the
growth of the stellar component of galaxies is of fundamental
importance. Recent observations have probed the star formation
rates and stellar masses of galaxies out to redshift $z\sim 7$.
Key trends have emerged. For instance, fairly massive galaxies are
in place at early epochs, and galaxies exhibit ``downsizing" in the
sense that more massive galaxies form their stars earlier and have
lower birthrates today. While some claim that these trends are in
conflict with expectations from hierarchical structure formation,
they are actually not. Since the most massive galaxies form within
the largest mass density peaks, they collapse first and start forming
stars earlier than lower-mass systems~\citep{dav06b,nei06}. The
collapse is driven by gravitational instability and strong radiative
cooling that (left unchecked) grow galaxies very rapidly at early
epochs, so that models without strong feedback yield galaxies that
are {\it too large} at high redshifts~\citep[e.g.][]{dav06}. It
is important to note that no confirmed high-$z$ galaxy yet observed
is too large to be produced in a hierarchical model, and in fact
strong feedback appears to be necessary to suppress galaxy formation
from the earliest epochs. The evolution of galaxy stellar masses
and star formation rates therefore provide valuable insights into
the processes, particularly related to feedback, that drive galaxy
formation from the highest redshift until today.
The canonical scenario for galaxy formation posits that galaxies
assemble first as rotationally-supported disks cooling out of
virial-temperature gas within a dark matter
halo~\citep[e.g.][]{ree77,whi78,whi91}, and then such objects merge
to form larger and earlier-type systems~\citep[e.g.][]{mih96}.
Naively, one expects then that star-forming disks would be surrounded
by hot gaseous halos~\citep{ben00}, and that early low-mass objects
would tend to be disky while more massive objects would be ellipticals.
Recent observations, however, do not clearly support this paradigm.
For instance, star-forming disks and hot (X-ray emitting) gaseous
halos are rarely if ever found together~\citep[e.g.][though see
\citealt{cra10} and references therein]{ras09}. Observed dynamics
of $z\sim 1-2$ galaxies indicate that larger objects tend to be
disky, while smaller ones are dispersion-dominated~\citep{for09}.
Furthermore, galaxies appear to be consuming their gas rapidly
enough that fuel must be constantly replenished~\citep{tac10,pap10},
but merger-induced starbursts do not appears to be driving global
star formation~\citep{bri04,noe07,jog09}. Hence there is some doubt
that the canonical halo/merger-centric view provides the appropriate
framework for galaxy formation, despite the wide-ranging successes
that models based on this have enjoyed~\citep[e.g.][]{ben10}.
Improvements in cosmological hydrodynamic simulations over the past
decade have made them a competitive approach towards understanding
the physics of galaxy formation in a hierarchical structure formation
context. The complex three-dimensional geometry of large-scale
structure is self-consistently evolved together with the baryonic
components that make up observable galaxies and the co-evolving
intergalactic medium (IGM). These simulations have long suggested
that accretion onto galaxies is powered by cold, filamentary streams
that connect to the cosmic web on larger scales~\citep{ker05,dek09},
and that most accreted gas fueling star formation never heats to
the virial temperature~\citep{kat91,dek06}. Gas replenishment from
the IGM is a natural outcome of this model, and generally occurs
smoothly rather than being driven by major
mergers~\citep[e.g.][]{mur02,ker05,vdvoort10}. These simulations
provide many insights into how galaxies form and grow, some of which
are in common with the traditional scenario, while some differ.
Recently, simulations incorporating strong and ubiquitous outflows
have had success matching a wide range of data on
galaxies~\citep[e.g.][]{dav06,fin08,opp10} and the
IGM~\citep[e.g.][]{opp06,opp08,opp09,opp09b}. The outflow models
in these simulations were originally tuned to enrich the IGM in
accord with observations of $z\sim 2-4$ \ion{C}{iv}
absorbers~\citep{opp06}. The favoured scalings turn out to be those
expected for momentum-conserved winds~\citep{mur05,zha10}, namely
that the wind speed scales with circular (or escape) velocity, while
the amount of mass ejected per unit star formation (i.e. the mass
loading factor) scales inversely with it. This provides sufficiently
early IGM enrichment by expelling the majority of metals formed in
small high-redshift galaxies, while having moderate wind speeds
that does not to over-heat the IGM~\citep{opp06}. In subsequent work
we showed that an outflow model with similar scalings consistently
does as well or better than any other outflow model that we have
tried at matching a wide range of IGM properties. We note that the
actual implementation of this wind model in our simulations has
evolved owing to advances in modeling techniques and changes in
cosmology~\citep{opp08,opp10}. Successes include very early ($z\ga
5$) enrichment~\citep{opp09b}, \ion{O}{vi} absorbers at low-$z$~\citep[with
the introduction of a physically-motivated IGM turbulence
model][]{opp09}, the metallicities in intragroup and intracluster
gas~\citep{dav08b}, and the overall cosmic metal
distribution~\citep[including the so-called ``missing metals
problem";][]{dav07}.
Examining galaxy properties, we again found that an outflow model
following similar scalings was the most successful amongst the ones
we tried. Successes included matching the mass-metallicity relation
at $z=2$~\citep{fin08}, the shape of the present-day stellar mass
function~\citep{opp10} below $M^\star$, damped Ly$\alpha$ absorber
kinematics~\citep{hon10}, and observations of very high-$z$ ($z\ga
4$) galaxies~\citep{fin06,dav06,bou07,fin10}. This is not to say
that this particular outflow model matches all the data; in fact,
it does not, it simply does better than others attempted, including
the oft-used \citet{spr03b} model. Moreover, the implementation
remains highly heuristic, and may be impacted by numerical
effects~\citep[e.g. see discussion in][]{opp10}. Nevertheless, the
consistent (comparative) success of an outflow model with momentum-driven
wind scalings suggests that it captures some essential ingredients
required for a successful model of galactic outflows, and motivates
a more detailed investigation of the implications of such outflows.
Two key predictions from models incorporating outflows with such
scalings are that (i) the total amount of material ejected from
galaxies significantly exceeds that going into stars~\citep[hereafter
OD08]{opp08}; and (ii) the ejected material often re-accretes onto
galaxies~\citep{opp10}. The strong and ubiquitous ejection and
re-accretion of material suggest that galaxies are best viewed as
evolving within an ecosystem of surrounding intergalactic gas with
which they continually exchange mass, energy, and metals.
While the above results are encouraging, these simulations have yet
to be comprehensively tested against observations of galaxy evolution
from the present epoch back into the early universe. Such a
comparison would provide the most stringent test to date of how
well these simulations can reproduce the observed Universe, and
would better highlight the current successes and failures of galaxy
evolution simulations. This series of papers aims to understand
how basic galaxy statistics and scaling relations between their
stellar, gas, and metal content constrain the physics of galaxy
formation. In Paper I (this paper) we focus on the statistics of
and scaling relations between stellar mass and star formation rate.
In Paper II (Dav\'e, Finlator, \& Oppenheimer, in prep.) we will
additionally investigate the gas and metal content of galaxies.
The primary goal of these works is to understand and characterise
the key physical processes that drive the observable properties of
galaxies across cosmic time.
In these papers, we will demonstrate that for a particular choice
of outflow model, namely outflows following scalings as expected
for momentum-driven winds, the simulations can reproduce key
observations of scaling relations between stars, gas, and metals
for large star-forming galaxies that dominate cosmic star formation.
This is a first for cosmological hydrodynamic simulations. However,
all simulations we examine here fail to match selected galaxy
properties at high and low masses, indicating that additional or
different physics is required even for our most successful model.
Using insights gained from the simulations, we show that relations
between basic galaxy constituents can be broadly understood through
a simple scenario in which inflows are balanced by outflows and
star formation. This contrasts with the canonical halo/merger-centric
model of galaxy formation; in our scenario, the dark matter halo's
virial radius and galaxy mergers play a sub-dominant role in the
evolution of star-forming systems. We argue that galactic outflows
are the key moderator of the stellar content, star formation rate,
and gas and metal evolution in galaxies at all epochs. In turn,
observations of these properties provide valuable constraints on
the cosmic ecosystem within which galaxies form and grow.
This paper is organised as follows. In \S\ref{sec:sims} we describe
our hydrodynamic simulations including our galactic outflow model.
In \S\ref{sec:msfrfcn} we investigate the galaxy stellar mass and
star formation rate functions, and explain the three-tier behavior
generically seen in wind simulations by invoking differential wind
recycling. In \S\ref{sec:ssfr} we look at the scaling relation
between SFR and $M_*$ and its evolution to high-$z$, emphasizing
its connection with the matter accretion rate into halos. In
\S\ref{sec:sat} we further break down these trends in terms of
centrals vs. satellite galaxies, and raise some issues regarding
observed dwarf galaxies that none of the models reproduce. Finally,
we summarise and discuss the broader implications of our work in
\S\ref{sec:summary}.
\section{Simulations}\label{sec:sims}
\subsection{Code}
Our simulations are run with an extended version of the {\sc Gadget-2}~N-body
+ Smoothed Particle Hydrodynamic (SPH) code \citep[OD08]{spr05}.
We assume a $\Lambda$CDM cosmology using the cosmological parameters
based on recent WMAP results~\citep{hin09}: $\Omega_{\rm M}=0.28$,
$\Omega_{\rm \Lambda}=0.72$, $h\equiv H_0/(100 \;{\rm km}\,{\rm s}^{-1}\;{\rm Mpc}^{-1})=0.7$, a
primordial power spectrum index $n=0.95$, an amplitude of the mass
fluctuations scaled to $\sigma_8=0.82$, and $\Omega_b=0.046$. We
refer to these runs as the r-series, where our general naming
convention for a simulation run is r[{\it boxsize}]n[{\it
particles/side}][{\it wind model}]. Our primary runs use the boxsize
of $48h^{-1}{\rm Mpc}$ on a side with $384^3$ dark matter and $384^3$ gas
particles. To expand our dynamic range we also run two additional
sets of simulations that are identical except that they have box
sizes of $24h^{-1}{\rm Mpc}$ and $48h^{-1}{\rm Mpc}$, and have only $256^3$ particles
of each species. The particle masses and gravitational softening
lengths are listed in Table~\ref{tab:sims}.
An overview of the {\sc Gadget-2}~code can be found in \S2 of K09a. Additions
to the public version of the code include cooling processes using
the primordial abundances as described in \citet{kat96} and metal-line
cooling as described in \citet{opp06}. Star formation is modeled
using a subgrid recipe introduced in \citet{spr03a} where a gas
particle above a critical density is modeled as a fraction of cold
clouds embedded in a warm ionised medium following \citet{mck77}.
Star formation (SF) follows a Schmidt law \citep{sch59} with the
SF timescale scaled to match the $z=0$ Kennicutt law \citep{ken98a}.
The density threshold for SF is $n_{\rm H}=0.13 \;{\rm cm}^{-3}$.
We use a \citet{cha03} initial mass function (IMF) throughout. The
fraction of mass of the IMF going into massive stars, defined here
as $\ge 10 \;{\rm M}_{\odot}$ and assumed to result in Type II supernovae
(SNe), is $f_{\rm SN}=0.18$; for reference, a Salpeter IMF
has $f_{\rm SN}=0.10$.
Our simulations directly account for metal enrichment from sources
including Type II supernovae (SNe), Type Ia SNe, and asymptotic
giant branch (AGB) stars. Gas particles eligible for SF undergo
self-enrichment from Type II SNe using the instantaneous recycling
approximation, where mass, energy, and metallicity are assumed to
immediately return to the ISM. Type II SN metal enrichment uses
the metallicity-dependent yields calculated from the \citet{chi04}
SNe models. Our prescriptions for feedback, described below, also
assume the instantaneous input of energy from O and B stars (i.e.
stars that are part of $f_{\rm SN}$). We input the Type Ia SNe
rates of \citet{sca05}, where an instantaneous component is tied
to the SFR, and a delayed component is proportional to the stellar
mass as described in OD08. Each Type Ia SN adds $10^{51}$ ergs of
thermal energy and the calculated metal yields of \citet{thi86} to
surrounding gas; the mass returned is small compared to that from
Type II SNe. AGB stellar enrichment occurs on delayed timescales
from 30 Myrs to 14 Gyrs, using a star particle's age and metallicity
to calculate mass loss rates and metallicity yields as described
in OD08. AGB mass and metal loss is returned to the three nearest
surrounding gas particles. OD08 showed that the largest effect of
AGB stars is to replenish the ISM, because a star particle can lose
over 30\% of its mass over $t_{\rm H}$, nearly double the $1-f_{\rm
SN}=18\%$ that is recycled instantaneously via Type II SNe.
Galactic outflows are implemented using a Monte Carlo approach
analogous to star formation. Outflows are directly tied to the
SFR, using the relation $\dot M_{\rm wind}= \eta \dot M_{\rm SF}$,
where $\eta$ is the {\it mass loading factor}. The probabilities
for a gas particle to spawn a star particle is calculated as described
above, and the probability to be launched in a wind is $\eta$ times
that. If the particle is selected to be launched, it is given an
additional velocity of $v_w$ in the direction of {\bf v}$\times${\bf
a}, where {\bf v} and {\bf a} are the particle's instantaneous
velocity and acceleration, respectively; this would create a bipolar
outflow for a thin disk, but in practice the outflow has a much
larger opening angle. This subgrid model circumvents our inability
to resolve the detailed physics of the ISM that generates winds,
at the cost of introducing additional parameters. Choices of the
parameters $\eta$ and $v_w$ define the ``wind model".
Once a gas particle is launched, its hydrodynamic (not gravitational)
forces are turned off until either $1.95\times10^{10}/(v_{\rm w} (\;{\rm km}\,{\rm s}^{-1}))$
years have passed or, more often, the gas particle has reached 10\%
of the SF critical density. This is intended to simulate the
formation of hot chimneys extending out of a disk galaxy providing
a low resistance avenue for outflows to escape into the IGM, which
the spherically-averaged SPH algorithm at our current resolution
is incapable of resolving. However, explicit decoupling neglects
energy losses owing to the initial creation of such channels as well as
any drag along their interfaces, and as such cannot drive
turbulence in the ISM. We note that decoupling has a significant
impact on wind propagation; simulations without such decoupling
explored e.g. by \citet{dal08} show that decoupling greatly enhances
the ability of material to escape a galaxy's ISM. Hence decoupling
must be regarded as an integral part of the wind models described
in this work, having a significant impact on the results, particularly
those associated with recycling of wind material (which would be
much more prevalent if the hydrodynamic forces were never turned
off). \citet{spr03b} demonstrated that this decoupling achieves
resolution convergence in the cosmic star formation history. We
will later explicitly demonstrate resolution convergence for many
of the galaxy properties of interest here.
\begin{table*}
\caption{Simulation runs}
\begin{tabular}{lccccccccccc}
\hline
Name &
$L^a$ &
$\epsilon^b$ &
$m_{\rm SPH}$ &
$m_{\rm dark}$ &
$m_{\rm gal}^c$ &
$v_{\rm w}$ ($\;{\rm km}\,{\rm s}^{-1}$) &
$\eta^{d}$ &
$E_{\rm wind}/E_{\rm SN}^{e}$ &
$p_{\rm wind}/p_{\rm *}^{f}$ &
$f_*^g$ &
$f_{\rm wind}^h$
\\
\hline
\multicolumn {4}{c}{} \\
r48n384nw & 48 & 2.5 & $3.6\times 10^7$ & $1.8\times 10^8$ & $\sim 10^{10}$ & 0 & 0 & 0 & 0 & 0.220 & 0 \\
r48n384vzw & 48 & 2.5 & $3.6\times 10^7$ & $1.8\times 10^8$ & $1.1\times 10^9$ & $\propto\sigma$ & $\propto\sigma^{-1}$ & 0.49$^i$ & 5.2 & 0.097 & 0.345\\
r48n384cw & 48 & 2.5 & $3.6\times 10^7$ & $1.8\times 10^8$ & $1.1\times 10^9$ & 680 & 2 & 0.95 & 8.7 & 0.045 & 0.174\\
r48n384sw & 48 & 2.5 & $3.6\times 10^7$ & $1.8\times 10^8$ & $1.1\times 10^9$ & 340 & 2 & 0.24 & 4.4 & 0.124 & 0.476\\
\hline
r24n256nw$^j$ & 24 & 1.875 & $1.5\times 10^7$ & $7.6\times 10^7$ & $\sim 4\times 10^9$ & 0 & 0 & 0 & 0 & - & 0 \\
r24n256vzw$^j$ & 24 & 1.875 & $1.5\times 10^7$ & $7.6\times 10^7$ & $4.7\times 10^8$ & $\propto\sigma$ & $\propto\sigma^{-1}$ & - & - & - & -\\
r24n256cw$^j$ & 24 & 1.875 & $1.5\times 10^7$ & $7.6\times 10^7$ & $4.7\times 10^8$ & 680 & 2 & 0.95 & 8.7 & - & -\\
r24n256sw$^j$ & 24 & 1.875 & $1.5\times 10^7$ & $7.6\times 10^7$ & $4.7\times 10^8$ & 340 & 2 & 0.24 & 4.4 & - & -\\
\hline
r48n256nw & 48 & 3.75 & $1.2\times 10^8$ & $6.1\times 10^8$ & $\sim 3\times 10^{10}$ & 0 & 0 & 0 & 0 & 0.182 & 0 \\
r48n256vzw & 48 & 3.75 & $1.2\times 10^8$ & $6.1\times 10^8$ & $3.7\times 10^9$ & $\propto\sigma$ & $\propto\sigma^{-1}$ & 0.52$^i$ & 5.1 & 0.081 & 0.257 \\
r48n256cw & 48 & 3.75 & $1.2\times 10^8$ & $6.1\times 10^8$ & $3.7\times 10^9$ & 680 & 2 & 0.95 & 8.7 & 0.046 & 0.179 \\
r48n256sw & 48 & 3.75 & $1.2\times 10^8$ & $6.1\times 10^8$ & $3.7\times 10^9$ & 340 & 2 & 0.24 & 4.4 & 0.116 & 0.448 \\
\hline
\end{tabular}
\\
$^a$ Box size in comoving $h^{-1}{\rm Mpc}$.\\
$^b$ Gravitational softening length in comoving $h^{-1}{\rm kpc}$ (equivalent Plummer).\\
$^c$ Galaxy mass resolution limit in $M_\odot$; see discussion in \S\ref{sec:massfcn} of no-wind case.\\
$^d$ $\eta\equiv\dot{M}_{\rm wind}/$SFR.\\
$^e$ Total feedback energy divided by SN energy (Chabrier IMF).\\
$^f$ Total feedback momentum divided by momentum output by $10^7$~yr old stellar population (Chabrier IMF).\\
$^g$ Global mass fraction of baryons in stars at $z=0$.\\
$^h$ Global mass fraction of baryons that have been ejected in winds by $z=0$.\\
$^i$ Volume-averaged to $z=0$; in this model it is $\propto\sigma$.\\
$^j$ Owing to a hardware failure, we only have information at z=0,1,3 for these runs.\\
\label{tab:sims}
\end{table*}
\subsection{Wind Models}\label{sec:windmodels}
For this paper we run four wind models, described as follows:
{\bf No winds (nw):} We turn off our galactic winds. Note that,
as with all simulations, energy is imparted thermally to ISM SPH
particles using the \citet{spr03a} subgrid two-phase recipe where
all SN energy is instantaneously coupled to the ISM. Owing to the
explicit coupling of the hot and cold ISM phases, such energy cannot
drive a wind.
{\bf Constant winds (cw):} Feedback energy is added kinetically to
gas particles at a constant rate relative to the SFR, with $\eta=2$
and a constant velocity $v_{\rm w}=680 \;{\rm km}\,{\rm s}^{-1}$. This is intended to mimic
the constant-wind model of \citet{spr03a}, but with some notable
differences. Here $\eta$ is defined relative to the SFR of the
entire Chabrier IMF, and not just the long-lived stars in a Salpeter
IMF as it is defined by \citet{spr03a}. Second, $v_{\rm w}=680 \;{\rm km}\,{\rm s}^{-1}$ is
used rather than $484 \;{\rm km}\,{\rm s}^{-1}$, because there is more SN energy per
mass formed in a Chabrier IMF. The kinetic energy imparted to wind
particles per unit SF is $9.25\times 10^{48} {\rm erg}/\;{\rm M}_{\odot}$,
which is 95\% of the SN energy in this IMF assuming all stars $\ge
10 \;{\rm M}_{\odot}$ add $10^{51}$ ergs/SN. The hydrodynamic decoupling
time in this model has a maximum of $2.9\times 10^7$ years. We
note that this model is similar to the ``REFERENCE" outflow model
of the OverWhelmingly Large Simulations~\citep[OWLS][]{sch10}, except
that we include hydrodynamic decoupling while they do not; in
fact, they have also considered models likw nw and vzw (below)
except without decoupling.
{\bf Slow winds (sw):} We run an alternative constant-wind model
with the only difference being the outflow velocities are half as
high as the cw winds (i.e. $v_{\rm w}=340 \;{\rm km}\,{\rm s}^{-1}$ and $\eta=2$). Therefore
only roughly a quarter of the SN energy couples kinetically into
these outflows.
{\bf Momentum-conserved winds (vzw):} This model uses the analytic
derivations by \citet{mur05} based on the galaxy velocity dispersion
($\sigma$). Its implementation is explained in \citet{opp06} and
updated in OD08 with the addition of an {\it in situ} group finder
to calculate $\sigma$ using galaxy mass following \citet{mo98}.
The wind parameters vary between galaxies using the following
relations:
\begin{eqnarray}
v_{\rm w} &=& 3\sigma \sqrt{f_L-1}, \label{eqn: windspeed} \\
\eta &=& \frac{\sigma_0}{\sigma} \label{eqn: massload},
\end{eqnarray}
where $f_L$ is the luminosity factor in units of the galactic
Eddington luminosity (i.e. the critical luminosity necessary to expel
gas from the galaxy potential), and $\sigma_0$ is the normalization of
the mass loading factor. Here we randomly select $f_L=[1.05,2]$ for
each SPH particle, following observations by \citet{rup05}, and include a
metallicity dependence for $f_L$ owing to more UV photons output by
lower metallicity stellar populations
\begin{equation}
f_L = f_{\rm{L},\odot} \times 10^{-0.0029*(\log{Z}+9)^{2.5} + 0.417694}.
\label{eqn: zfact}
\end{equation}
We further add an extra kick to get out of the potential of the
galaxy in order to simulate continuous pumping of winds until it
is far from the galaxy disk (as argued by MQT05). These parameters
$v_{\rm w}$ and $f_L$ are well-constrained by observations, leaving the
only unconstrained free parameter as $\sigma_0$. We set $\sigma_0=150
\;{\rm km}\,{\rm s}^{-1}$, which is the same efficiency factor of the mass loading used
in OD08. The energy budget for winds scales as $\propto \eta
v_w^2\propto \sigma$, and exceeds the supernova energy budget only
at quite large halo masses of $\ga 10^{14} M_\odot$ (at $z=0$).
Since we do not include any type of feedback that quenches star
formation in massive galaxies~\citep{gab10}, our simulations do not
accurately model these systems in any case. In general, the wind
energetics in this model are quite modest, and despite the large
mass outflow rates and speeds around the escape velocity, are
generally well below the supernova energy budget (globally, about half
in the r48n384vzw run). This does not even account for radiative
energy input from massive stars, which can exceed the energy output from
supernovae.
These wind models allow us to illustrate how outflows impact the
scaling relations of galaxies. An important difference comes from
the wind speed scaling: In the momentum-conserved outflows case,
the wind speed is typically around the galaxy escape velocity, and
scales with it. Hence in this model, the potential well of the
galaxy does not play a governing role in wind dynamics as a function
of mass. This stands in contrast to the canonical model for
outflows~\cite[e.g.][]{dek86} in which outflows escape from small
galaxies owing to their small potential wells while being recaptured
in large galaxies. If gravity was the only relevant force, this
model should show exactly the same escape fraction of outflowing
material from galaxies of all masses. In actuality, hydrodynamic
slowing plays an important role~\citep{opp08}, and because more
massive galaxies live in denser gaseous environments, this retards
winds more around more massive galaxies, causing greater return of
wind material to these systems~\citep{opp10}.
The constant-$\eta$ models (sw and cw) behave more like the canonical
model for winds, in the sense that at low galaxy masses their
(constant-velocity) winds can escape the potential well, while at
high masses they cannot. One can envision this as an ``effective"
mass loading factor that is roughly the input value of 2 at small
masses, but drops to zero at large masses~\citep{fin08}. Alternatively,
one can view this as rapid recycling of wind material back into
galaxies at high masses~\citep{opp10}. In any case, the constant
wind speed imparts a feature at a mass scale corresponding to where
the assumed wind velocity is comparable to the galaxy escape velocity.
Because of the different wind speeds in cw and sw, this feature
should occur at different masses (roughly a factor of eight apart
in halo mass, given the factor of two difference in wind speeds).
Hence comparing these two models enables us to identify features
in scaling relations that are related to how the effective mass
loading factor scales with mass; this is discussed extensively in
\citet[see e.g. their Figure~4]{opp10}. As we will argue later,
it is the scaling of the (effective) mass loading factor with $M_*$
that primarily governs most galaxy scaling relations.
Our twelve runs are listed in Table \ref{tab:sims}. They were run
at the University of Arizona's SGI cluster, ICE, the National Center
for Supercomputing Applications' Intel cluster, Abe, and the
University of Massachusetts' Xeon cluster, Eagle. Each run took
between 100,000 and 150,000 CPU hours, although nearly 1 million
CPU hours were used when including smaller box sizes and various
test runs. We note that due to technical difficulties, we only
have galaxy catalogs for the r24 runs at $z=0,1,3$, and do not have
any information regarding satellite galaxies in these runs. However,
all the important trends can be seen from the other runs at the
redshifts where the r24 runs are not available.
We emphasize that all these models are heuristic and intended only
to explore how these wind scalings impact galaxy properties. We are
not directly advocating any particular physical process, though we
are attempting to use physically-motivated recipes. In particular,
the momentum-driven wind scalings can arise from models that only employ
supernova energy to drive winds, owing to the interactions of super-heated
ISM material with surrounding gas~\citep[e.g.][]{dal08}. Hence the name
``momentum-driven" is only intended to indicate the scalings rather than
a detailed physical model.
Nevertheless, owing to its prominence in this paper and others, we briefly
discuss some issues related to the physics of the momentum-driven scalings
wind model. The basic physical motivation is a model in which radiation
pressure drives dust outwards, which couples to the gas to generate
an outflow~\citep{mur05,zha10}. First, it is important to note that
the momentum input required to drive winds at hundreds of km/s with
mass loading factors as assumed above significantly exceeds the photon
momentum emitted by the stellar populations~\citep[M. Haas \& J. Schaye,
priv. comm.][]{haa10}. We show this ratio in Table~\ref{tab:sims}, and
is a factor of 5.2 for our r48n384vzw simulation\footnote{The momentum
requirements are similarly large for the constant-$v_w$ models, but in
this case the physical motivation is supernova energy, so the momentum
input is not relevant.} Following \citet{haa10}, this is calculated from
the luminosity of stars for a solar-metallicity $10^7$~yr old stellar
population assuming a \citet{cha03} IMF, divided by $c$. The large
value indicates that momentum alone, under the assumption of single
photon scattering, cannot power the outflows in the vzw model.
It is possible that subsequent re-radiation of photons could cause
significantly more momentum deposition than what is emitted by the
stars if the gas is optically thick to infrared radiation. For an
infrared optical depth of $\tau_{IR}$, momentum deposition is increased
by $1+\tau_{IR}$. This factor can only be significantly greater than
unity in highly dense gas where young stars form. \citet{hop11} computed
$\tau_{IR}$ in simulations of radiation-driven winds from high-resolution
disk galaxies, and indeed found that in some cases $\tau_{IR}$ can
approach an order of magnitude, driving an outflow velocity of $\sim
200$~km/s leaving the disk with mass loading factors of order unity.
While this is encouraging, the issue of whether enough momentum is
available from stars, and whether it can effectively couple to the dust
and in turn to the gas, remains an open question.
Another possibility is that supernovae and photons from young stars work
together to drive winds~\citep[e.g.][]{nat09,mur10}. Invariably, this
must be happening at some level, since both processes are energetically
comparable. \citet[Figure 2.19]{haa10} showed that the momentum from
SNe can be an order of magnitude larger than that from stellar radiation
at $10^7$~yr after the burst, although it drops rapidly with time.
\citet{mur10} showed that for a 300~km/s wind, radiation pressure can
yield a mass loading factor up to $\eta=0.4$ in the single-scattering
regime, which is factors of many less than what our momentum-driven wind
scalings require. Hence supernova momentum input would have to dominate
overall, particularly at large radii where the optical depth to absorb
photons becomes small. Despite being driven primarily by supernovae and
not radiation pressure, the scalings predicted in this scenario continue
to follow those expected for momentum-driven winds~\citep{mur10}.
In short, the momentum-driven wind scalings that we have assumed
challenge the available momentum budget, but could plausibly arise from
a combination of physical processes. Much work remains to be done to
marry large-scale constraints from cosmological simulations with detailed
studies of wind expulsion from star-forming regions.
\subsection{Galaxy and halo identification}
We use SKID\footnote{http://www-hpcc.astro.washington.edu/tools/skid.html}
(Spline Kernel Interpolative Denmax) to identify galaxies as bound
groups of star-forming gas and stars~\citep{ker05,opp10}. Since
we specifically include just star-forming gas, our simulated galaxies
only account for gas within the galactic ISM and not from the
extended halo; this defines the galaxy's gas mass. Our galaxy
stellar mass limit is set to be $\ge 64$ star particles~\citep{fin06},
resulting in the masses listed under $m_{\rm gal}$ in Table~\ref{tab:sims}.
We will argue that for the no-wind case the resolution convergence
is poorer, and we list approximate masses that show convergence.
Unless otherwise noted, we will only consider galaxies with stellar
mass $M_*\geq m_{\rm gal}$ in our analysis.
We separate galaxies into central and satellite galaxies by associating
each galaxy with a halo. We identify halos via a spherical overdensity
algorithm~\citep{ker05}. In brief, we begin at the center of each
galaxy and expand out spherically until the mean density enclosed
equals that of a virialized halo~\citep{kit96}. We then subsume
all galaxies whose centers are within that of a larger galaxy's
halo into that larger halo. In this way, each galaxy is a member
of one and only one halo. The central galaxy is the largest one
that is at the center of the halo, and any other galaxies within
that halo are satellites.
\section{Stellar Mass and Star Formation Rate Functions}\label{sec:msfrfcn}
\subsection{Stellar Mass Function}\label{sec:massfcn}
\begin{figure*}
\vskip -1.0in
\setlength{\epsfxsize}{0.85\textwidth}
\centerline{\epsfbox{fig.mfcomp.ps}}
\vskip -1.0in
\caption{Stellar mass functions at $z=0,1,2,3$ in cosmological
hydrodynamic simulations employing our four galactic outflow
scalings: momentum-conserved winds (green), constant winds (red),
no winds (blue), and slow winds (magenta). Solid lines show results
from the r48n384 series of runs, dotted lines show r24n256 series,
and dashed lines show r48n256 series; these illustrate resolution
convergence, although when interpreting convergence one should keep
in mind that these runs are not equally (logarithmically) spaced in
mass resolution: r48n256 is $2.4\times$ r48n384, which is $3.3\times$
r24n256. The momentum-conserved wind case shows Poisson error bars as
a representative example. Cyan dashed curve at $z=3$ shows the result
of adding a Gaussian random scatter of $\sigma=0.3$~dex to each galaxy's
stellar mass in the momentum-conserved wind case. Data points at $z=0$
from \citet{bel03}, curve at $z=0$ from \citet{bal08}, curves at $z=1,2,3$
from \citet{kaj09}, and data points at $z=2,3$ from \citet{mar09}.
}
\label{fig:massfcn}
\end{figure*}
We begin by considering the stellar mass function and its evolution
in our simulations. Stellar mass appears to be a governing parameter
for galaxy properties, having fairly tight scaling relations with
many other fundamental galaxy properties, relatively independently
of environement~\citep[e.g.][]{li10}. Hence the stellar mass
function provides a key barometer for simulated galaxy populations.
Figure~\ref{fig:massfcn} shows the galaxy stellar mass function (GSMF) out
to $z=3$ in our four outflow simulations. Data is shown for comparison
from \citet[$z\sim 0$, points]{bel03}, \citet[$z\sim 0$, curve]{bal08},
\citet[$z=1,2,3$, curves]{kaj09}, and \citet[$z=2,3$, curves]{mar09}.
These were selected as fairly recent determinations of the stellar mass
function from some of the deepest available near-infrared observations,
though it is far from a complete sample of available data. The curves
represent Schechter function best fits, while the data points (from
different samples) give an indication of the uncertainties. Note that
the \citet{kaj09} data plotted at $z=1$ is actually centered at $z\approx
0.75$, but the evolution is not rapid during this period. Also, we have
scaled all data to a \citet{cha03} IMF (as assumed in the simulations),
by dividing the stellar masses by 1.25 for those quoted using a diet
Salpeter IMF~\citep[namely,][]{bel03,bal08}, and by 1.8 for those quoted
using a Salpeter IMF~\citep{kaj09}. \citet{mar09} used a \citet{kro01}
IMF which is very similar to Chabrier, so we apply no IMF correction.
Finally, the \citet{bel03} data has been corrected to $h=0.7$; all other
data shown assumed this value for the Hubble constant.
A test of numerical resolution convergence is provided by comparing
the solid, dotted, and dashed lines for each model, which are from
the r48n384, r24n256, and r48n256 series of runs, respectively.
The difference in mass resolution between r48n384 and r48n256 is a factor
of 2.4, while between r48n384 and r24n256 it is 3.3; without accounting
for this, convergence will appear slightly better at low resolution.
In actuality, for the wind models at every redshift the GSMFs show very
good convergence (modulo some stochastic fluctuations at the massive end),
and the faint end of the higher-resolution run joins smoothly onto that
of the larger-volume run. For the no-wind simulations, this is not the
case: The higher-resolution run deviates from the lower-resolution run
at a $M_*$ that is much larger than the nominal 64 star particle galaxy
mass resolution. The reason can be traced back to resolution convergence
for halos, as opposed to galaxies: Halos also require approximately
64 particles to robustly host galaxies, and in the no-wind case this
corresponds to a much higher $M_*$ than in the wind models where star
formation at a given halo mass is much suppressed~\citep{opp10}.
Hence convergence for the no-wind case differs from the wind runs,
and its resolved galaxy mass is higher by roughly an order of magnitude
(as indicated in Table~\ref{tab:sims}). The low-mass peak of the GSMF
is shifted over by roughly the difference in mass resolution between
the two models, further indicating that the GSMF here is impacted by
resolution effects. Hence the turn-down in the no-wind GSMF at low
masses is a numerical artifact, not a prediction. For the wind models,
this turn-down happens below the nominal stellar mass resolution.
The no-wind case (nw; blue), where resolution-converged, shows the
classic behavior that the GSMF well exceeds observations at all
masses~\citep[see e.g.][for a more detailed discussion]{ker09b}.
This is symptomatic of the overcooling problem~\citep{whi91,bal01,dav01},
in which there are too many baryons locked in stars at all epochs
without strong feedback. We note that this occurs at {\it all}
redshifts and masses back to $z\sim 4$~\citep[and even farther back;
see][]{dav06}, indicating that the overcooling problem is not just
a late-time phenomenon and must be mitigated in the early Universe.
It has long been suggested that galactic outflows could suppress
star formation to solve the overcooling problem. The magnitude of
suppression, factors of several even at maximum galaxy formation
efficiency around $M^\star$ increasing to much larger factors at
larger and smaller masses, indicates that outflows must eject a
substantial amount of gas from galaxies and/or suppress accretion
by adding energy to the surrounding IGM. In star-forming galaxies
(typically $M_*\la M^\star$), the dominant energy source is star
formation, either from supernovae or young stars. This motivates
all of our wind models having mass loading factors exceeding unity.
For concreteness, Table~\ref{tab:sims} shows as its final two columns
the global mass fraction of baryons in stars at $z=0$ and ejected
in winds by $z=0$, respectively. For the momentum-conserved run,
the ratio between the two is roughly 3.5, i.e. $3.5\times$ more
baryons have been ejected from galaxies than have formed into stars
by $z=0$. The constant-$\eta$ models have a ratio of approximately
3.8. That this ratio signficantly exceeds unity is a generic feature
of all models that solve overcooling via an ejective feedback
mechanism such as galactic outflows.
The exact shape of the $z=0$ GSMF depends sensitively on outflows,
albeit in a subtle way. As discussed in \citet{opp10}, it is not
the overall suppression by ejection that governs the shape, but
rather the effectiveness of subsequent re-accretion of ejected
material, i.e. recycled wind mode accretion. As shown in \citet{opp10},
without wind recycling, all GSMFs look fairly similar, and show a
steep power law behavior with a faint-end slope comparable to that
of the halo mass function. Because wind material re-accretes into
larger galaxies faster (``differential recycling"), it boosts
high-mass galaxy masses relative to low-mass ones, and therefore
flattens the faint end of the stellar mass function. At sufficiently
small masses, the timescale for wind re-accretion exceeds the Hubble
time, at which point the mass function returns towards tracking the
dark matter halo mass function as there is no differential effect
from recycling. At sufficiently large masses, recycling is so fast
that it effectively is like having no winds at all. This creates
a three-tier behavior for the GSMF in simulations: It follows the
steep dark matter mass function both above and below the mass range
where differential recycling operates, and it is flatter within
that mass range.
All our wind models display this three-tiered behavior, though it
is less evident in some cases. At high masses, they have a slope
similar to that of the no-wind model. In the intermediate mass
regime, the slope is flatter or even inverted, and finally at low
masses once again the GSMF is reverts to being steep. We note that
all models grossly overpredict the massive end of the GSMF compared
to observations. This is because we do not have any feedback
mechanism to truncate star formation in massive galaxies, such as
feedback from active galactic nuclei~\citep[AGN; e.g.][]{dim05}.
Hence our first key point is that our current simulations can only
reliably probe the regime of star-forming galaxies, which dominate
at masses $\la M^\star\approx 10^{11}M_\odot$~\citep{sal07}. In
\citet{gab10} we show that it is possible to include empirical
feedback mechanisms to quench star formation in massive galaxies,
without significantly affecting the population of galaxies below
$M^*$. This is consistent with the observation that strong AGN
activity is not seen in lower-mass galaxies today~\citep[$M_*\la
10^{10.5}M_\odot$;][]{kau04}.
The exact transitions between the various GSMF tiers depend strongly
on wind model. In the constant-$\eta$ cases, the three-tiered
behavior is manifested as a bump in the GSMF at high masses at
$z=0$. In the constant wind case (cw; red), the middle tier covers
a small mass range, and differential recycling is so strong that
it actually produces an inverted slope of the GSMF within that small
mass range. Outside that, the GSMF is quite steep, roughly following
the slope of the no-wind case. The slow-wind case (sw; magenta)
follows a similar pattern, but the mass scale of the differential
recycling is lower because wind speeds are slower, and hence wind
recycling becomes effective at a smaller scale. In neither case
does the GSMF resemble the observed one at $z=0$ even in the regime
where quenching feedback is unimportant, although the slow-wind
case is not as far off.
The momentum-driven scalings case (vzw; green) produces a more gradual
differential recycling curve than the constant-$\eta$ cases~\citep{opp10}.
Hence the middle tier is not inverted, but merely shallower than
the high and low mass ends of the GSMF. This mimics the behavior
exhibited by the observed GSMF. Recent observations by \citet{bal08}
have conclusively detected an upturn to a steep faint-end slope.
In detail, the vzw simulation is still mildly steeper than observed
and the upturn occurs at slightly too high a mass. If our physical
interpretation for the origin of the GSMF is correct, this would
imply that the momentum-conserved wind model has a differential
recycling curve that is close to, but not quite, correct. It is
worth mentioning that modeling the dynamics of wind recycling is
subject to significant numerical difficulties as discussed in
\citet[][though more so at in higher-mass halos where hot gas is
present]{opp10}, so the level of discrepancy with data is probably
within modeling uncertainties, but more careful simulations are
needed to determine this.
The evolution of the GSMF can now be probed to high redshifts thanks
to advancing deep near-IR surveys. At the massive end, the observed
GSMF has a less pronounced cutoff at high-$z$. Since this truncation
is associated with the presence of ``red and dead" galaxies, the
relative dearth of such galaxies at $z\ga 2$ explains why the
truncation is much less sharp. At $z\sim 2-3$, the vzw run yields
a GSMF at $M_*\ga 10^{10.5}M_\odot$ that is in fair agreement with
data, although it appears somewhat too steep. Meanwhile, sw
overproduces the number of galaxies at these stellar masses similar
to the no-wind case. The constant wind case generally matches well
here, although it still shows a bump at $M_*\ga 10^{11} M_\odot$
that is not seen in the data. Considering all redshift from $z=0-3$
in toto, the momentum-driven scalings model yields the best match
from amongst these models to the population of star-forming galaxies
at $10^{10}\la M_*\la 10^{11}M_\odot$.
The faint-end slope of the GSMF has been observed to evolve towards
being steeper at high redshifts. Our simulations follow this trend
qualitatively, but only within the intermediate mass regime where
differential recycling flattens (or inverts) the GSMF slope at later
times. Outside of this regime, the slope always approaches the
faint-end halo mass slope of $\alpha\approx 2$. For instance, in
the vzw run, the intermediate mass regime occurs at $10^{9.7}\la
M_*\la 10^{10.7} M_\odot$, and the slope at $z=0$ is $\alpha\approx
1.4$, which is close to the observed slope in that mass range. In
the sw case, the intermediate mass regime is narrower, $10^{10}\la
M_*\la 10^{10.5} M_\odot$, while in the cw case it is equally narrow
but occurs at a mass $\sim 8\times$ higher. In both constant-$\eta$
simulations, the GSMF slope at $z=0$ is inverted in this regime,
i.e. $\alpha<1$, but at higher redshifts $\alpha$ increases.
The general reason for the increase in $\alpha$ within the intermediate
mass regime is that wind recycling becomes less important at higher
redshifts. This is primarily because there is less time for ejected
material to return to galaxies~\citep{opp10}. Our simulations,
particularly the momentum-conserved case, tend to find median
recycling times that are fairly constant at $\sim 1-2$~Gyr, independent
of cosmic epoch~\citep{opp08}; this is partly because wind speeds
at a given $M_*$ are higher~\citep{opp08} at high-$z$. Hence the
features associated with recycling become less prominent, and in
fact they are essentially invisible at $z\ga 2$ in the vzw case.
Observationally, $\alpha$ is seen to evolve as $\alpha\approx
\alpha_0+(0.082\pm0.033)z$~\citep{fon06}, where $\alpha_0\approx
1.2-1.3$. Effectively, this corresponds to the slope in the mass
regime of $10^{10}\la M_*\la 10^{11}M_\odot$, which is what can be
probed in practice out to high-$z$ at this time; at lower masses,
the present-day GSMF becomes steeper, but this cannot be reliably
traced at high-$z$. The momentum-driven scalings model matches
well at low-$z$ (below $M^\star$), but at high-$z$ the faint end
is quite a bit steeper: e.g. at $z=2$, $\alpha\approx 2$ in the
regime where it is observed to be more like 1.5 or so. Hence while
the evolutionary trend is qualitatively as observed, there are
quantitative discrepancies in the sense that this model has too
many small galaxies (as is the case with all models). This has
also been noted at $z\sim 4-7$ in a comparison of various simulations
to mass function observations~\citep{gon11}; all models have too
steep a mass function at $M_*\la 10^9 M_\odot$. This may indicate
that feedback processes in very small systems are not being correctly
represented in these models, a point that will be reiterated
throughout this paper.
One issue that may explain part of the difference between the simulations
and data are observational uncertainties in determinations of stellar
masses, which can become significant particularly at high-$z$.
To explore this, at $z=3$ we added a Gaussian random scatter with
$\sigma=0.3$~dex to each galaxy's stellar mass in the momentum-conserved
wind simulation~\citep[which is a typical observational uncertainty;
see][]{mar09}, and recomputed the stellar mass function. The result is
shown as the dashed cyan line in Figure~\ref{fig:massfcn}. Because the
mass function is steep, more small galaxies get scattered to larger masses
than vice versa, and the mass function does becomes slightly shallower
(and in this case, agrees slightly better with observations at the massive
end). But this does not go far towards reducing the overproduction of
small galaxies in the simulations. There are, additionally, substantial
systematic uncertainties in determining stellar masses from photometric
data~\citep[e.g.][]{mar09}; we have not considered such effects here,
but it's conceivable that they could dominate over the statistical errors.
Hence while this effect should be accounted for when comparing carefully
to data, it does not qualitatively alter the faint end discrepancy.
Overall, the momentum-driven scalings model does the best job of
the models considered in reproducing the observed GSMFs around
$M^\star$ from $z=3\rightarrow 0$. There are still significant
discrepancies at higher and lower masses, so the agreement is
only good in the range of $10^{10}\la M_*\la 10^{11}M_\odot$.
Nevertheless, this is a significant success, a first for cosmological
hydrodynamic simulations (that we are aware of). While the discrepancy
at high masses is well-known and strongly suggestive of quenching
feedback absent in these simulations, the discrepancy at small
masses is less certain owing to still substantial observational
uncertainties at high-$z$. Upcoming deeper observations should
quantify this discrepancy to greater precision, possibly providing
insights into the nature of feedback processes in high-redshift
dwarf galaxies.
\subsection{Star Formation Rate Function}\label{sec:sfrfcn}
\begin{figure*}
\vskip -1.0in
\setlength{\epsfxsize}{0.85\textwidth}
\centerline{\epsfbox{fig.sfrfcn.ps}}
\vskip -1.0in
\caption{Star formation rate functions $\Phi_{\rm SFR}$ at $z=0,1,2,3$
in our r48n384 series of simulations employing our four galactic outflow
scalings: momentum-conserved winds (green), constant winds (red), no
winds (blue), and slow winds (magenta). Solid lines show results from
the r48n384 series of runs, dotted lines show r24n256 series ($3.3\times$
better mass resolution), and dashed lines show r48n256 series ($2.4\times$
worse mass resolution). The momentum-conserved case shows Poisson error
bars as a representative example. Dotted lines indicate the SFR at
galaxy stellar mass limit (see Table~\ref{tab:sims}) at each redshift,
from a linear fit to the $M_*-$SFR relation for each model; below this
SFR, the results are less likely to be robust. The thick line at $z=0$
shows $\Phi_{\rm SFR}$ derived from ultraviolet plus infrared data
from \citet{dmar05}, while points at $z=1$ and thick line at $z=2$ show
$\Phi_{\rm SFR}$ derived from observed H$\alpha$ luminosity functions at
$z\approx 0.81$ by \citet{ly10} and at $z\approx 2.2$ by \citet{hay10},
respectively.
}
\label{fig:sfrfcn}
\end{figure*}
Figure~\ref{fig:sfrfcn} shows the star formation rate function
($\Phi_{\rm SFR}$) for galaxies in our four wind models at four
redshifts. The dotted line shows the SFR at the galaxy stellar
mass limit listed in Table~\ref{tab:sims}, obtained by fitting the
relationship between SFR and $M_*$ as displayed in Figure~\ref{fig:ssfr}.
Since we do not have a clean stellar mass threshold in this plot,
we show all galaxies, but $\Phi_{\rm SFR}$ at SFRs below the dotted
line may not be robust. Because the relationship between SFR and
$M_*$ varies slightly with wind model, each model has a different
SFR threshold. Note that the no-wind case has a significantly
higher resolved galaxy mass threshold in Table~\ref{tab:sims} for
reasons described in the previous section, and hence its dotted
line is at a much higher mass. Observations at $z=0$ from a
GALEX+IRAS compilation by \citet{dmar05} are shown as the solid
black line, at $z\sim 1$ from \citet{ly10}, and at $z=2$ from an
H$\alpha$ survey by \citet{hay10}.
At $z=0$, $\Phi_{\rm SFR}$ shows different behaviors amongst the various
wind models that mimic many features seen in the GSMF. The effect
of differential recycling is seen in the three-tiered behavior of the
constant-$\eta$ models, but the intermediate regime is not obviously
seen in the vzw case. Nevertheless, relative to no-winds, vzw shifts
star formation from smaller to larger systems, owing both to its mass
loading dependence as well as differential recycling. The large amount of
recycled wind accretion at late times~\citep{opp10} produces a $\Phi_{\rm
SFR}$ at high masses that is actually higher for all the wind models
than for the no-wind case. The winds are metal-enriched and so are
particularly effective at cooling back onto the galaxy (often failing
to shock heat to the temperature of surrounding halo gas despite their
large outflow velocities), and they delay star formation towards later
epochs making the effect at high masses stronger towards lower redshifts.
Moving to higher redshifts, the strength of the differential recycling
features in the cw and sw simulations lessen as with the GSMF, but
they are still prominent even at the highest redshifts. For vzw,
$\Phi_{\rm SFR}$ becomes quite a bit steeper at high-$z$. The shape
of the no-wind $\Phi_{\rm SFR}$ does not vary, but the amplitude
evolution of $\Phi_{\rm SFR}$ is fairly constant from $z=3\rightarrow
2$, and then drops with time reflecting the overall drop in cosmic
SFR. For example, $\Phi_{\rm SFR}(10 M_\odot/$yr) at $z\sim 2-3$
is about $\sim 10^{-1.5}$, dropping to $\sim 10^{-2}$ at $z=1$ and
$10^{-3}$ at $z=0$. In contrast, the momentum-conserved wind model
shows a less rapidly evolving $\Phi_{\rm SFR}$, as $\Phi_{\rm SFR}(10
M_\odot/$yr)$\approx -2.3$ at all redshifts, although $\Phi_{\rm
SFR}$ shows a flatter faint-end slope at low redshifts.
Compared to observations at $z=0$~\citep{dmar05}, none of the models
fare well, as they all overproduce the number of star-forming
galaxies at higher SFR's. This is not surprising, and can be
straightforwardly traced to the lack of any feedback mechanism to
quench star formation in massive galaxies~\citep{gab10} as observed
in the real Universe~\citep[e.g.][]{kau04}. Note that the number
of star-forming galaxies must be suppressed down to quite moderately
star-forming systems. For instance, in the momentum-conserved case,
$\Phi_{\rm SFR}$ is overproduced at SFR$\ga 1 M_\odot$/yr. Therefore
galaxies larger than the Milky Way must be increasingly quenched,
which coincides with the discrepancy in the GSMF above $M^\star$.
The star formation rate function therefore provides a strong
constraint on models for quenching massive galaxies.
At higher redshifts, we compare to the H$\alpha$ luminosity function
determined at $z=0.81$ by \citet{ly10} and at $z\approx 2.2$ by
\citet{hay10}, both of which probe down to $\sim 1 M_\odot/$yr. We
obtain SFR from $L_{H\alpha}$ using the \citet{ken98b} conversion
divided by $1.8$ to correct from Salpeter to Chabrier IMF, namely
SFR$=4.4\times 10^{-42} L_{H\alpha} M_\odot/$yr. For \citet{ly10}
we show their extinction- and incompleteness-corrected 2.5$\sigma$
sample. For \citet{hay10} we show their ``combined" best-fit
Schechter function which is supplemented by bright-end data from
\citet{gea08}, and correct for 0.977 magnitudes of extinction
($A_V=1.19$ at 6563\AA) independent of $L_{H\alpha}$. At these
epochs, quenched massive galaxies are increasingly rare, so the
comparison should be more meaningful than at $z=0$.
At both epochs, the no-winds run overproduces the number of galaxies
with SFR$\la 100 M_\odot$/yr, sw overproduces galaxies with SFR$\ga
1 M_\odot/$yr, and cw underproduces galaxies with $1\la$SFR$\la 100
M_\odot/$yr. The momentum-driven scalings model fares reasonably
well at both redshifts. At $z\sim 1$ it still slightly overproduces
high-SFR galaxies, which is the onset of a trend that becomes much
more prominent by $z=0$. At $z\sim 2$ the bright end matches well,
while the faint end is modestly steeper than observed. Note however
that \citet{hay10} derive a steeper H$\alpha$ luminosity function
from their faint-end data alone ($\alpha=1.72$ instead of $\alpha=1.60$
for the fit shown here). Furthermore, convolving in the uncertainties
in $M_*$ determinations will tend to make the predicted $\Phi_{\rm
SFR}$ shallower, as we showed with the GSMF in Figure~\ref{fig:massfcn}.
Hence the discrepancies may not be significant.
Overall, the simulated star formation rate functions of the
constant-$\eta$ models display a three-tiered trend arising from
differential wind recycling as seen in the GSMF. The form and
evolution of $\Phi_{\rm SFR}$ provides independent constraints on
how outflows govern star formation across cosmic time. At low-$z$,
all current models produce too many rapidly star-forming galaxies,
owing to a lack of quenching feedback in massive galaxies. At $z\ga
1$, the momentum-driven scalings simulation predicts $\Phi_{\rm
SFR}$ in good agreement with available H$\alpha$ data, while other
wind models do not.
\section{Specific Star Formation Rate}\label{sec:ssfr}
The specific star formation rate (sSFR) measures the intensity of
ongoing versus past-averaged star formation, so it provides an
important barometer for how a given galaxy has assembled. It also
most directly governs the observed colors of galaxies. The main
sequence of galaxies reflects a relation between sSFR and $M_*$,
and yields important insights into galaxy growth. In this section
we study the specific star formation rate in our simulations. We
will particularly focus on how the slowly-evolving equilibrium
between inflows, star formation, and outflows governs the behavior
of sSFR as a function of mass and epoch.
\subsection{Specific Star Formation Rate vs. $M_*$}\label{sec:sfrmstar}
\begin{figure*}
\vskip -1.0in
\setlength{\epsfxsize}{0.85\textwidth}
\centerline{\epsfbox{fig.ssfr.ps}}
\vskip -1.0in
\caption{Specific star formation rate (sSFR$\equiv$SFR$/M_*$) as a function of
stellar mass
at $z=0$ in our r48n384 series of simulations employing our
four galactic outflow scalings: momentum-conserved winds (upper left),
constant winds (upper right), no winds (lower left), and slow winds
(lower right). Points are color-coded by gas fraction within
each stellar mass bin: Galaxies with $f_{\rm gas}$ exceeding $0.5\sigma$ above
the median are blue, less than $0.5\sigma$ below the median are red,
and in between are green. Galaxies with sSFR$=0$ are shown as red
points along the bottom. A running median is shown in magenta,
excluding non-starforming galaxies. Thick lines show
observed GALEX+SDSS best-fit relations~\citep{sal07} for non-AGN
star-forming galaxies
(solid) and blue galaxies with $NUV-r<4$ (dashed, broken power law).
}
\label{fig:ssfr}
\end{figure*}
Figure~\ref{fig:ssfr} shows the sSFR of galaxies at $z=0$ in our
four wind models. A running median within mass bins is shown in
magenta with $1\sigma$ variance; this does not include the
non-starforming galaxies shown along the bottom of the plot, which
do not substantially change the median except at the lowest masses
in the constant-$\eta$ models (we will return to this in \S\ref{sec:sat}).
The points are color-coded by gas fraction within each mass bin,
blue meaning high $f_{\rm gas}$ ($>0.5\sigma$ above median), red for low,
and green in between. Here we define gas fraction as $f_{\rm
gas}\equiv M_{\rm gas}/(M_{\rm gas}+M_*)$, with $M_{\rm gas}$ including
all gaseous phases in the star-forming ISM. Observations of galaxies
are shown as the thick lines from a compilation of GALEX and SDSS
data by \citet{sal07}: Solid line represents the median sSFR for
all galaxies classified as having no AGN content in a \citet{bal81}
diagram, while the dashed double power-law represents those classified
as red (NUV$-r>4$). The observed median for all galaxies drops
sharply above $\ga 10^{11}M_\odot$ as red and dead galaxies dominate
in that mass range.
Simulations generically produce a fairly tight relation between SFR
and $M_*$~\citep[e.g.][]{dav00,fin06}. In all cases, the relatively
flat sSFR($M_*$) implies that star formation rate scales roughly
linearly with stellar mass, and at the high-mass end in all cases
the scaling is sub-linear. The basic similarity indicates that the
sSFR is not highly sensitive to feedback, as has been noted
previously~\citep{dav08,dut10}. The basic reason is that, to first
order, feedback processes reduce SFR and $M_*$ in conjunction,
thereby keeping sSFR roughly similar. All models produce massive
galaxies that are forming stars vigorously, in contrast with
observations that indicate most galaxies at high masses are passive,
indicating the need for some quenching mechanism~\citep{gab11}.
All models also produce a population of low-mass galaxies that are
not forming stars; these are predominantly satellite galaxies, and
will be discussed in \S\ref{sec:sat}.
The basic trends in this plot can be understood in the context of
a balance between accretion, outflow, and star formation. As shown
in \citet{fin08}, galaxies live in a slowly-evolving quasi-equilibrium
between these quantities, such that
\begin{equation}\label{eqn:equil}
\dot{M}_{\rm inflow} = \dot{M}_{\rm outflow} + {\rm SFR},
\end{equation}
where the first two terms are the mass inflow and outflow rates,
respectively. We can rewrite this using the definition of the mass
loading factor $\eta\equiv \dot{M}_{\rm outflow}/$SFR, so that
\begin{equation}\label{eqn:sfr}
{\rm SFR} = \dot{M}_{\rm inflow}/(1+\eta).
\end{equation}
This equation, while simple, implicitly makes some non-trivial
assumptions. First, it assumes that the variations in the inflow rate
are slow compared to the rate of processing gas into stars, so that
galaxies are never strongly out of equilibrium (e.g. this is not valid
during major merger events). Second, it assumes that galaxies do not
collect large reservoirs of gas and then process it, but rather process
gas as it is made available to the ISM (from inflows and outflows); we
call this ``supply-regulated" star formation, to distinguish it from
other scenarios where gas aggregates in or around galaxies before being
consumed rapidly~\citep[e.g.][]{egg62,mar10}. Third, there is no explicit
dependence on cosmic epoch or environment, as these are assumed to be of
secondary importance in galaxy growth. Fourth, it neglects additional
reservoirs of gas such as stellar mass loss~\citep{sch10,lei10} and wind
recycling~\citep{opp10} that are particularly important at late times,
and hence this equation is expected to be more valid at earlier epochs.
We will show in this paper and Paper~II~\citep[see also][]{fin08} that
simulated galaxy properties generally follow those expected from the
equilibrium relation (eq.~\ref{eqn:equil}), although moreso at early
epochs, suggesting that at least in these models, the above assumptions
are broadly valid.
Within the context of this equilibrium model, let us return to
examining sSFRs. In the no-wind case, $\eta=0$, so the trend of
sSFR($M_*$) reflects the inflow rate into galaxies of a given $M_*$.
Owing to the growth of hot gaseous halos that retard inflow in
larger halos~\citep[e.g.][]{ker05,dek06}, the sSFR is smaller at
larger masses. This is one manifestation of downsizing, namely
star formation rate downsizing as defined in \citet{fon09}. This
model also produces archaeological downsizing, which means that
larger galaxies have older stellage populations~\citep[also called
``natural downsizing"][]{dav06b,nei06}, as we will show in
\S\ref{sec:age}. Note that these downsizing trends, while qualitatively
in agreement with data, do not by themselves produce massive red
and dead galaxies~\citep{gab10}. In the no-wind case, the slope
of sSFR($M_*$) agrees well with observations, but the amplitude is
too low by $\sim\times 2$. Hence paradoxically, while no winds
produce far too much global star formation (and stellar mass), they
produce too little star formation at a given stellar mass today.
They also produce $\sim\times 10$ more stellar mass for a given
halo mass~\citep{ker09a,opp10}, so that the excess global star
formation arises from too much star formation per halo, particularly
at high and low masses.
The wind models deviate from the no wind case based on $\eta(M_*)$
and wind recycling. In the cw and sw models, since $\eta$ is constant,
in the simplest scenario both SFR and $M_*$ are reduced equally. This
shifts a given galaxy down in $M_*$ from the no-wind case by a factor
of $(1+\eta)=3$, when the mass is small enough that winds don't recycle
(i.e. the effective mass loading factor is identical to the input mass
loading factor). Since the overall slope of the no-wind case is negative,
this results in a lower sSFR at a given mass. An additional suppression
is provided by preventive feedback from heat added to surrounding gas by
winds~\citep{vdvoort10}, which lowers the accretion rate and hence sSFR at
a given $M_*$. This is particularly noticeable in the cw case which adds
substantial energy to surrounding gas and the even the IGM~\citep{opp06}.
At the highest masses, all winds are gravitationally and hydrodynamically
trapped, resulting in an effective mass loading factor of $\eta\approx 0$
as discussed in \S\ref{sec:windmodels}, and so also follows the no-wind
case trend with a slightly higher amplitude. The reason for the amplitude
difference is that wind recycling removes mass from small systems and
adds it back into larger systems at later epochs. Hence wind recycling
shifts star formation from low to high masses, an effect that is stronger
with the higher wind speeds owing to the increase in the recycling time
and increasing preventive feedback from added heat. Both these models
are in poor agreement with observations, as they have too low sSFR by
up to an order of magnitude at the low-mass end, and the observations
do not indicate any positive-slope feature in the predicted mass range.
In the case of momentum-conserved winds, $\eta$ varies with mass.
At the lowest masses where there is no wind recycling, $\eta\propto
\sigma^{-1}\propto M_{\rm gal}^{-1/3}\propto M_*^{-1/3}$ (roughly,
although in detail smaller galaxies tend to have lower stellar fractions,
making the slope slightly shallower), since our wind model ties $\sigma$
directly to the galaxy baryonic mass $M_{\rm gal}$ as identified by
our group finder~\citep[see eq.~6 of][]{opp08}. Hence the sSFR shows
a slope that exceeds the no-wind case by roughly $1/3$. This by itself
would results in a slope of around zero. But differential wind recycling
produces an additional effect, which in this model is more gradual and
occurs over a wider range in mass than in the constant-$\eta$ cases. At
intermediate masses, the additional fuel from recycling causes a positive
slope for sSFR($M_*$), for the same reason as in the constant-$\eta$
case except the effect is more gradual. \citet{fir10} also pointed
out that wind recycling tends to increase the slope of sSFR($M_*$).
At the largest masses, the slope again reverts to the no-wind case as
recycling times are very short~\citep{opp10}, so winds have little effect.
The gradual rolling of the slope of sSFR($M_*$) is also seen in
observations, but it less pronounced and really only evident at
masses below what are plotted here~\citep[see Figure~16 of][]{sal07}.
Even at $M_*\la 10^9 M_\odot$, the observed sSFR($M_*$) has a
negative slope~\citep{sal07}, whereas this model predicts a positive
slope. We will see in Paper~II that this is also reflected in the
relation between gas fraction and $M_*$, which has a turnover to
low masses that is likewise absent in observations. This could
signify that wind recycling should be more effective in small
systems, or perhaps less material should be ejected from these
systems in the first place. Just as with the faint end of the
high-redshift GSMF, this highlights that even our most successful
model has difficulty correctly reproducing observations of dwarf
galaxies. We will return to this point in \S\ref{sec:age}.
An interesting second-parameter trend is indicated by the color of
the simulated data points. Galaxies with high gas fractions at a
given $M_*$ lie above the mean sSFR relation, and those with low
gas fractions lie below. This is not surprising since the presence
of gas drives star formation based on our assumed star formation
law, but it provides an interesting generic prediction (independent
of outflows) of this class of models where star formation is driven
by accretion. Such a trend would not be expected, for example, if
mergers drove elevated SFRs at a given $M*$; in that case, it is
the gas configuration rather than overall gas content that drives
elevated SFRs, as gas is much more concentrated in mergers owing
to tidal dissipation~\citep[e.g.][]{mih96}. In Paper~II we will
discuss second-parameter dependences of the relationships between
stellar mass, metallicity, and star formation rate in much greater
depth, all within the context of the aforementioned equilibrium
model (eq.~\ref{eqn:equil}) for galaxy growth.
\subsection{Specific Star Formation Rate Evolution}\label{sec:ssfrevol}
\begin{figure}
\setlength{\epsfxsize}{0.65\textwidth}
\centerline{\epsfbox{fig.ssfrevol.ps}}
\vskip -3.2in
\caption{Evolution of the specific star formation rate
at a stellar mass of $M_*=10^{10}M_\odot$
from $z=6\rightarrow 0$ in our r48n384 series of simulations
employing our four galactic outflow scalings. Observations are
shown at $z=0,1$ from \citet{elb07} and at $z\geq 2$ from \citet{gon10}.
The downward arrow on the data point at $z\sim 2$ indicates the
$\sim\times 2$ lowering of the UV-derived SFR based on {\it Herschel}
data. Dashed lines show the typical observed $1\sigma$ variance
about the $M_*-$SFR relation~\citep{noe07,dad07}. The black dotted
line shows a power-law scaling of $(1+z)^{2.25}$ as predicted
from cold accretion-driven galaxy growth.
}
\label{fig:ssfrevol}
\end{figure}
Figure~\ref{fig:ssfrevol} shows the evolution from $z=6\rightarrow
0$ of the specific star formation rate at a particular stellar mass
of $M_*=10^{10}M_\odot$ in our four wind models. This mass is
chosen because observations can directly measure sSFR at this mass
at all epochs (with small extrapolations), and it is well-resolved
by simulations but is expected to be negligibly affected by quenching
feedback. In the simulations, this sSFR is computed as the median
sSFR within a stellar mass bin of $\pm$0.25~dex around $10^{10}M_\odot$.
Most models show a characteristic evolution of the sSFR which is
essentially a power law in $(1+z)$. The fundamental physics giving
rise to this is gas inflow driven by growth of structure. In the
cold accretion paradigm that captures the basic behavior in
hydrodynamic simulations~\citep[e.g.][]{ker05}, the amount of gas
inflowing into the star forming region is correlated with the gas
inflowing at the halo virial radius, since cold streams efficiently
channel material to the center of the halo~\citep{dek09}. Feedback
processes that strongly impact the surrounding gas to retard infall
can cause significant deviations from this~\citep{vdvoort10}, but
let us set that aside for now to examine the implications of this
simplest scenario. The amount of gas entering the virial radius
can be estimated by the total mass accretion rate times the baryon
fraction. The total mass accretion rate onto halos has been well
measured in simulations~\citep{dek09}, and scales with redshift as
$(1+z)^{2.25}$.
The scaling of $(1+z)^{2.25}$ is indicated by the dotted line,
arbitrarily normalised to the observed $z=0$ sSFR. The no-wind and
slow-wind cases follow this power law essentially at all redshifts.
The no-wind case is easy to understand, since there is no additional
physics regulating star formation beyond accretion. The slow-wind
case also follows this because it simply ejects a constant two-thirds
of the accreted material. Hence the SFR reduction translates into
a similar reduction in $M_*$, so the sSFR evolution is nearly
identical to the no-wind case. In the sw case, two additional
processes are at work: Feedback heats surrounding gas to retard
accretion into the ISM, which is countered by additional accretion
provided by recycled winds. Note that at this stellar mass, about
half the $z=0$ accretion is recycled winds~\citep{opp10}, and both
recycling and heating increase to low redshifts. The impact of
these processes does not appear to be large (barring a remarkable
cancellation), as the net effect is only a slight increase in sSFR
in the sw model over the no-wind case.
The constant wind case shows a substantial departure from no-winds,
with sSFR lower by $\sim\times 2-5$, increasing at late times. As
discussed previously, this model produces significant heating of
surrounding gas, substantially retarding inflow into the ISM from
the virial radius~\citep{opp10,vdvoort10}. This effect becomes
stronger at low-$z$ as feedback energy accumulates in the IGM, as
evidenced e.g. by a much higher fraction of cosmic gas in the
Warm-Hot Intergalactic Medium in this model~\citep{dav10}. At
$M_*=10^{10}M_\odot$, recycling is rare in this wind model~\citep{opp10},
so the net effect is to strongly lower the sSFR relative to the
no winds case.
In contrast, the momentum-driven wind scalings model increases the
sSFR relative to no winds. In this case, the relationship to
no-winds is not so straightforward to understand, because the mass
loading factor depends on galaxy mass and epoch. Overall, the
amount of energy input in vzw is smaller than in cw. Recyling is
dominant at late times at $M_*\sim 10^{10}M_\odot$, but at early
epochs it is not infrequent. The net effect is larger suppressions
of the sSFR at early epochs, while at late times the wind recycling
at this mass causes the sSFR to be higher relative to a simple
inflow model extrapolated from high-$z$. Hence the vzw model yields
a shallower evolution of sSFR($z$) because it suppresses star
formation at early epochs and later recycles that material into
galaxies.
Observations show a very rapid rise from $z=0\rightarrow 2$, and
then essentially no evolution out to very high
redshifts~\citep{dad07,sta10,gon10}. \citet{bou10} showed that
this is difficult to reproduce in a simple accretion-driven scenario,
and they argued for a rather radical modification to the galaxy
formation paradigm in which no stars formed in halos with virial
masses less than $10^{11}M_\odot$ (as well as above $10^{12}M_\odot$).
This strongly suppresses early star formation until the characteristic
mass of star-forming halos exceeds this threshold mass, at which
point it reverts to the accretion-driven scaling. Clearly this
hypothesis is incorrect today; there is no evident feature in the
galaxy population around halo masses of $\sim 10^{11} M_\odot$,
unlike at $\sim 10^{12} M_\odot$ where many transitions occur. It
is also difficult to accomodate at high-$z$, as smaller halos need
to form stars to reionise the Universe and form observed high-$z$
galaxy populations~\citep{mun11}. Nevertheless, while implausible,
this scenario correctly emphasizes that some process must suppress
high-redshift star formation significantly in order to match current
sSFR evolution observations.
The observations are themselves uncertain to some extent, particularly
at high redshifts. For instance, the original $z\sim 2$ data point
(black circle) was taken from dust-corrected UV-derived star formation
rates of star-forming BzK galaxies~\citep{dad07}. Recent {\it
Herschel} data characterizing the infrared continuum of such sources
suggests that the extinction corrections were overestimated by up
to $\sim\times 2$~\citep{nor10}; this modification is shown by the
red arrow and circle in Figure~\ref{fig:ssfrevol}. Prior to this
correction, the data exceeded the accretion rate evolution, making
it difficult to reconcile this evolution within the cold accretion
paradigm. This prompted \citet{dav08} to make the radical suggestion
that the IMF may evolve with redshift, so that the true SFR's were
lower by $\sim\times 3$. If the {\it Herschel}-derived correction
is right, the $z\sim 2$ data point is now consistent with being
powered by ongoing accretion, making it viable within the cold
accretion paradigm. \citet{dav10b} pointed out that there still
may be a problem with respect to the momentum-driven scalings model;
this may be particularly true if the corrections at these masses
(which are not directly constrained by {\it Herschel} data) are
less than a factor of two~\citep[see also][]{nor10}. But this is
no longer a fundamental difficulty, and this level of discrepancy
could in principle be alleviated by modifications of the feedback
prescription. Indeed, \citet{genel11} find that doing high-resolution
disk simulations using a momentum-conserved feedback model with
somewhat different parameters results in galaxies that are in very
good agreement with (corrected) sSFR data at $z\approx 2$.
The slope $\beta$ of the main sequence (defined by SFR$\propto
M_*^{\beta}$) is also observed to evolve downwards from being unity
at high-$z$~\citep{sta10,lab10}, to $\beta\approx 0.9$ at $z\sim
1-2$~\citep{dad07,elb07}, to $\beta\approx 0.7$ at $z\la 1$~\citep{noe07}
down to $z\sim 0$~\citep{bri04}. There is some sensitivity to
selection effects, particularly in how one chooses a sample of
star-forming galaxies at the massive end; the preponderance of
low-sSFR massive galaxies particularly at low-$z$ pulls the slope
down depending on how many such systems are included in the fit. Hence
it is not straightforward to compare to our models where we do not
reproduce large passive galaxies. In general, our momentum-conserved
winds case produces slopes of $\beta\approx 0.9-1$, roughly
independently of redshift, and the no-wind case produces a shallower
slope. For the constant-$\eta$ models $\beta$ is ill-defined as
they show a strong three-tiered behavior. The trend of the slope
becoming more shallow with time is therefore not obviously evident
even in our favored model. While this may be a consequence of the lack
of quenching feedback in massive galaxies, we note that the
semi-analytic model of \citet{som08} includes AGN feedback but still
produces a slope around unity for the present-day main sequence.
In summary, the evolution of the specific star formation rate encodes
information about the feedback processes that have regulated star
formation up to that epoch. Hence this is a critical quantity to
measure with upcoming surveys. The fundamental driver is the gas
accretion rate onto halos, which evolves as $(1+z)^{2.25}$, but
feedback can modify this by either retarding accretion onto the
galaxy or adding recycled wind accretion. Matching observations
that suggest a significant departure from this scaling at $z\ga 2$
requires suppressing star formation at early epochs. Of the feedback
models examined here, only the momentum-conserved scalings case works in
this sense, though it is still well away from the unevolving sSFR
seen from $z\sim 4-7$. Observations at these epochs remain fairly
uncertain~\citep[see e.g.][]{schaerer10}, so we await more robust
constraints from upcoming surveys.
\section{Satellite vs. Central Galaxies}\label{sec:sat}
The equilibrium paradigm for galaxy evolution introduced earlier
is appropriate for central galaxies, because cold accretion channels
material to the centers of halos. Satellite galaxies, in contrast,
must scavenge their fuel from ambient halo gas, and therefore can
be impacted by a range of processes related to their surrounding
environment. Satellite galaxies could thus display a broadly
different class of properties than central galaxies, for which
environment is not a major driver except through wind recycling.
Here we investigate the properties of satellite vs. central galaxies
in our simulations, contrasting the statistics we have examined
earlier between the satellite and central galaxy populations. We
particularly focus on dwarf galaxies where many environmental
processes could play a large role.
\subsection{Satellite Fraction}
\begin{figure*}
\vskip -0.2in
\setlength{\epsfxsize}{0.85\textwidth}
\centerline{\epsfbox{fig.mfsat.ps}}
\vskip -0.5in
\caption{Four larger panels show stellar mass functions at $z=0$
(solid lines) and $z=2$ (dashed lines) for our four r48n384 simulations.
Black lines are the total mass functions (reproduced from
Figure~\ref{fig:massfcn}), which are subdivided into into centrals
(magenta) versus satellites (cyan). Accompanying smaller panels below
each large panel show the fraction of satellite galaxies as a function
of $M_*$ at $z=0$ (solid) and $z=2$ (dashed).
}
\label{fig:mfsat}
\end{figure*}
Figure~\ref{fig:mfsat} shows stellar mass functions as in
Figure~\ref{fig:massfcn} (black curves), subdivided into central
galaxies (magenta curves) and satellite galaxies (cyan curves).
Solid curves show $z=0$ results, while dashed curves show $z=2$.
Smaller panel below each main panel show the fraction of satellite
galaxies as a function of stellar mass, also at $z=0$ (solid) and
$z=2$ (dashed). Only the r48n384 series of runs are shown in this
plot.
Satellite galaxy mass functions are, for the most part, simply
scaled-down versions of the central galaxy mass functions. They
display the same features arising from differential wind recycling.
In general, the properties of satellites appear to be established
when they are central galaxies in their own halo, and are not
dramatically affected by falling into a larger halo. This is broadly
consistent with trends inferred from halo occupation distribution
models of galaxy clustering within dark matter halos~\citep[e.g.][]{con06}.
For all models at $z=0$, central galaxies dominate by number at all
resolved masses. This may be surprising, as it is often assumed
that most small galaxies are satellites, and that the fraction of
satellites increases strongly to small masses. The no-wind model
does show an increase in satellite fraction to lower masses, but
only mildly so. In contrast, the wind models show a satellite
fraction that is approximately constant at one-third of all galaxies
for $M_*\la M^\star$. For $M_*>M^\star$ the satellite fraction
decreases, although it is high again at very large masses -- these
are satellites in galaxy groups where the halo occupation distribution
is steeply rising~\citep[e.g.][]{ber03}. These results agree with
those obtained by \citet[see their Figure~8]{vbosch08} based on the
conditional luminosity function constrained by the 2dF Galaxy
Redshift Survey, and recent SAMs also yield similar trends~\citep{guo10}.
At $z=2$, there are even fewer satellite galaxies at a given stellar
mass, typically only around 20\% of all galaxies for $M_*\la
10^{11}M_\odot$. Hence dwarf galaxies are not typically satellites,
but rather central galaxies in small halos.
The relatively constant (and low) satellite fraction has a significant
implication for galaxy formation. It means that that whatever
physics is responsible for establishing the mass function and other
such ensemble-averaged observables must predominantly govern central
galaxies, not satellites~\citep[see also][]{fon09}. It is not
possible, for instance, to alter satellite galaxy physics to reconcile
the steep faint-end GSMF slope in the constant-$\eta$ or no-wind
simulations with observations. This is even more true at higher
redshifts. Hence the evolution of the ensemble population of
galaxies is governed by central galaxies, and understanding galaxy
evolution is, to first order, tantamount to understanding how central
galaxies form and grow.
\subsection{Quenched Satellites}\label{sec:satquench}
\begin{figure*}
\vskip -0.2in
\setlength{\epsfxsize}{0.85\textwidth}
\centerline{\epsfbox{fig.birthrate.ps}}
\vskip -0.5in
\caption{Larger panels show the birthrate (i.e. $t_{\rm SFR}$SFR/$M_*$,
where $t_{\rm SFR}=t_{\rm Hubble}-1$Gyr) for galaxies from
our r48n384 runs at $z=0$
color-coded by central (blue) and satellite (red) in our four
wind models. Galaxies
with SFR$=0$ (``quenched") are shown along the bottom of the plot.
Running median (solid curves), including quenched galaxies, are
shown for central galaxies (magenta) and satellites (cyan). Dashed
above and below show running 75\% and 25\% percentiles, respectively.
Small panels below show the fraction of quenched galaxies
(having birthrate $b=0$) in a given mass
bin for centrals (magenta) and satellites (cyan).
}
\label{fig:ssfrsat}
\end{figure*}
Figure~\ref{fig:ssfrsat} shows birthrates for galaxies in our four wind
models, separated into central (blue) and satellite (red) galaxies.
The birthrate is defined as the specific star formation rate multiplied
by the time over which the galaxy has been forming stars; the sSFR values
are noted on the right axis, for comparison with Figure~\ref{fig:ssfr}.
\citet{dav08} showed that the timescale for star formation in most
galaxies is reasonably well approximated by the Hubble time less 1~Gyr,
since none but the largest galaxies (today) form stars in the first Gyr.
A running median birthrate and lines enclosing 25\% and 75\% of galaxies
in a given mass bin are given by the solid and dashed lines, for centrals
(magenta) and satellites (cyan).
Birthrates are typically around unity for all models. At the massive
end all models show a negative slope in birthrates with $M_*$,
indicative of archaeological downsizing. At small masses, this can
invert owing to differential wind recycling as discussed in the
previous section. Overall, however, star-forming galaxies show
generically smooth and fairly constant SFHs over much of cosmic
time to $z\sim 0$~\citep{dav08}, similar to that inferred for the
Milky Way.
There are a significant number of galaxies with no ongoing star formation
(i.e. ``quenched"), predominantly at low masses. The smaller panels show
the fraction of quenched central (magenta) and satellite (cyan) galaxies
as a function of $M_*$. There are a few small central galaxies with zero
instantaneous SFR's. It is not clear what these are; one possibility
is that they are actually quenched satellites whose orbits carry them
outside the virial radius calculated by our spherical overdensity halo
finder, and therefore are identified as centrals. Another possibility
is that quenching is partly numerical, since this may impact a satellite
galaxy's ability to cool gas from a hot halo. Since they are very few
in number, we will leave a detailed investigation of these possibilities
for the future.
The satellite quenched fractions become large towards small masses.
Below a few$\times 10^9 M_\odot$, the median sSFR drops to zero for
satellites in all models. Even above this mass, satellites show
lower typical sSFR's than central galaxies, hence environmental
processes can affect even large satellite galaxies. The different
wind models (including no winds) show only minor differences in
terms of quenched satellite fraction, despite significant differences
in sSFR. This indicates that it is not outflows that are responsible
for quenching satellites, and instead is some process(es) related
to environment that depends more on the growth of structure common
to all models.
Possible environmental processes that quench satellites are ram
pressure stripping, tidal harrassment, and strangulation (i.e.
cutting off of the gas supply). \citet{sim09} showed that, in SPH
simulations similar to ours, quenching of a galaxy occurs on a $\sim
1$~Gyr timescale after entering another galaxy's halo. The process
is gradual because subhalos retain their identity for quite some
time after entering a larger halo, so the satellite does not
immediately see the full effect of the hot gas in the larger halo.
In the cold accretion paradigm, the gas cooling along filaments
tends to fall to the center of the halo, and hence galaxies that
turn into satellites become disconnected from their feeding filaments.
But strangulation cannot be the entire story, because the gas
consumption timescales for these systems are of order a few Gyr (as
we will show in Paper~II), whereas quenching occurs on a timescale
of $\la 1$~Gyr. Hence harrassment and stripping must be playing a
role. Unfortunately, while the exact processes that drive satellite
quenching are interesting, they are a great numerical challenge to
model properly~\citep[e.g.][]{age07}, and the numerical resolution
and methodology used here is probably insufficient to make robust
quantitative statements.
Although there are seen to be a significant number of dwarf galaxies with
little or no ongoing star formation, SDSS data~\citep{wei06,vbosch08}
actually indicate a trend opposite to that predicted here, namely a
smaller fraction of red satellites (and centrals) at smaller masses.
One difference may be that these analyses defined red based on a color cut
between the red sequence and blue cloud, which does not straightforwardly
correspond to being fully quenched or not; the observed galaxies may
simply have lower specific SFRs. Interestingly, a recent semi-analytic
model that is more successful than our simulations at matching a wider
range of data still suffers from the same discrepancy, showing an upturn
in red fraction at low masses~\citep{kim09}. Hence the qualitative
discrepancy may hint at generic difficulties in modeling environmental
effects in satellites within hierarchical models, and bears further
investigation.
In summary, the fraction of satellite galaxies in our wind models
is roughly constant with mass for $M_*\la M^\star$ at around 20-30\%,
declining slightly with redshift. Their mass functions show trends
similar to centrals, likely reflecting the fact that these trends
are set when satellites were centrals. Satellites are increasingly
quenched to smaller masses, reflecting environmental processes that
gradually disconnect these galaxies from their feeding cold streams.
High-mass quenching must arise from another source that is currently
not included such as AGN feedback, but low-mass quenching can be
understood as a result of environmental effects that are already
included (at least in principle) in these types of models.
\subsection{Stellar Ages and Dwarf Galaxy Evolution}\label{sec:age}
\begin{figure*}
\vskip -0.2in
\setlength{\epsfxsize}{0.85\textwidth}
\centerline{\epsfbox{fig.agesat.ps}}
\vskip -0.5in
\caption{Larger panels show mean mass-weighted age of stars in galaxies
in our r48n384 runs
as a function of $M_*$, color-coded by central (blue) and satellite
(red) in our four wind models at $z=0$. Running median (solid
curves), including quenched galaxies, are shown for central galaxies
(magenta) and satellites (cyan). The smaller panels show the fraction
of satellites identified as ``bursting", i.e. with birthrate $b>2.5$,
as a function of mass, for centrals (magenta) and satellites (cyan).
}
\label{fig:agesat}
\end{figure*}
We have seen that below masses of $M_*\la 10^{10} M_\odot$, the
wind models show a turn-down in sSFR and birthrate (and in Paper
II we will see a similar turn-down in gas fraction), for both
satellites and centrals. This is in contradiction with
observations that show no such turn-down.
To investigate this further, we show in Figure~\ref{fig:agesat} the
mean (mass-weighted) stellar ages of galaxies as a function of
$M_*$, for satellites (red) and centrals (blue), in our four wind
models at $z=0$. This is essentially an archaeological downsizing
plot, showing whether more massive galaxies have older stellar
populations as generally observed. Running medians are shown for
centrals (magenta) and satellites (cyan).
All models exhibit archaeological downsizing for galaxies at the massive
end. But for the wind models, below some mass the trend reverses,
and smaller galaxies begin to be progressively older. This is opposite
to the trend observed by \citet{pas10} in a group catalog derived from
SDSS (see their Figure~4). For the vzw and sw models, this happens at
$M_*\sim 10^{10} M_\odot$; above this mass, the ages are in reasonable
agreement with data. They also show that satellites are slightly older
than centrals at a given mass, which is also qualitatively as observed,
although \citet{pas10} find the median mass-weighted age difference at
$M_*=10^{10} M_\odot$ is $\sim 0.1$~Gyr (becoming smaller to higher masses)
while our wind models predict a smaller difference. The no-wind case
shows an archaeological downsizing trend at all masses.
The mass where the turn-down happens is exactly the mass where the
specific star formation rate turns over (Figure~\ref{fig:ssfrsat}).
We argued previously that this owes to wind recycling, in particular
the lack of it in the smallest systems. Hence we begin to build a
consistent story for the failure of wind models in the dwarf galaxy
regime: The smallest galaxies in these models begin forming stars
early, and owing to ejection of mass they lose their gas and cannot
reacquire it as fast as larger galaxies. This is not because small
galaxies are satellites: central galaxies show exactly the same
trend, and most small galaxies are centrals anyway.
One possibility is that the simulations do not resolve the (supposedly)
bursty star formation histories of dwarf galaxies, which could give
rise to younger ages. It is true that our simulations lack the
resolution to model internal resonances that can drive gas towards
the center of galaxies during interactions, fueling a burst~\citep{mih96}.
However, recent data has strongly questioned whether bursts are
common in dwarfs. \citet{lee09} used the volume-limited 11HUGS
sample to show that the burst fraction, defined as having
birthrate$>2.5$, is only $6^{+4}_{-2}\%$ in dwarfs (typically $\la
0.1L^\star$), comprising $23^{+14}_{-9}\%$ of that population's
star formation. These numbers are similar to what is seen for
more massive galaxies~\citep[e.g.][]{bri04}
In the small panels in Figure~\ref{fig:agesat} we show the burst
($b>2.5$) fraction as a function of mass, for centrals (magenta)
and satellites (cyan). It is generally quite low, and in our favored
vzw model it is mostly independent of mass at $\sim 5\%$ for $M_*\la
10^{10.5}M_\odot$, in both satellites and centrals. In the no wind
case, the burst fraction rises dramatically at small masses, but
mostly in central galaxies not satellites. The three-tiered shapes
of the birthrate curves in the cw and sw model result in odd behaviors
in the burst fraction, and result in very low burst fractions for
small systems. The favoured momentum-conserved wind case agrees
well with observations of burst fractions, although the fraction
of star formation in bursts is $\sim 15$\% which is $\approx 1\sigma$
below that observed. This model's concordance with data suggests
that, contrary to popular wisdom, small dwarfs are not significantly
burstier compared to larger star-forming systems.
The higher ages, steep high-$z$ GSMF, and lower sSFR's for dwarfs
suggest that wind models should suppress early star formation in
smaller systems even more than is currently done in the momentum-conserved
wind case, and produce a larger reservoir of gas for consumption
at later epochs. This could arise if mass loading factors are
larger at early epochs than assumed here, and/or wind speeds are
reduced at late times such that these galaxies can reacquire their
ejected gas. This could also arise if the star formation law is
different in these systems owing to a less efficient conversion of
atomic gas into molecular~\citep[e.g.][]{rob08} -- this would tend
to delay star formation, providing a younger, more vigorously
star-forming dwarf population today. Investigating this discrepancy
further is likely to reveal new physical processes that are important
for the evolution of small galaxies at all cosmic epochs.
\section{Summary and Discussion}\label{sec:summary}
Cosmological hydrodynamic simulations are now reaching a maturity
level such that they can be informatively compared to observations
of galaxies across cosmic time. This is a new era in this type of
modeling, as past such efforts have generally produced quite poor
agreement with data, and have been severely limited by dynamic range
particularly when evolved to low redshifts. With improvement in
both computing power and input physics, simulations can now plausibly
reproduce a wider suite of observed relations (though still far
from all), which in return can yield insights into the governing
physics. A particularly important new physical aspect has been the
incoporation of galactic outflows, originally motivated to explain
IGM enrichment, and simultaneously having a wide-ranging impact on
galaxies' stellar, gas, and metal content.
In this paper we present a study of how the stellar content and
star formation rates of galaxies are impacted by galactic outflows.
We concentrate on understanding the underlying physical drivers of these
properties, and how observations can enlighten us on the way in which
these drivers operate in the real Universe. To this end we present
some comparisons with observations, showing that including outflows
with momentum-conserved scalings provide the best overall match to
the ensemble of observations considered in this work, although notable
discrepancies remain. Given the broader success of this class of models
in matching IGM (and other galaxy) properties, and that these scalings
are consistent with those observed for outflows, this adds to the growing
body of evidence that strong and ubiquitous outflows with these scalings
are an important piece in understanding the overall evolution of galaxies.
The evolution of the stellar component can be broadly understood within
the context of a cycle of gas inflow and outflow between galaxies and
the IGM. This differs somewhat from the traditional view of galaxy
formation in which halos and their mergers drive galaxy evolution.
The cycle of baryons can at its most basic level be analytically
represented by a slowly-evolving equilibrium between inflow, outflow,
and star formation (eq.~\ref{eqn:equil}), where the outflows govern
how much of the inflowing material turns into stars. The galaxy
mass dependence of the outflow rate thus directly governs e.g. the
GSMF, the specific star formation rate of galaxies, and galaxy ages.
A particularly key aspect is the idea of differential wind recycling, in
which material ejected returns to galaxies in a mass-dependent manner.
This drives noticeable features in observable properties of galaxies
as a function of mass, providing a way to constrain how outflows move
material within the cosmic ecosystem surrounding galaxies. We reiterate
that this scenario does not invoke the halo virial radius, environment,
or mergers as playing a governing role in star-forming galaxy evolution.
The cold streams feeding star-forming galaxies generally take no notice of
the virial radius, are broadly unaffected by environment, and while such
streams also carry in galaxies that merge, the mass growth rate in such
mergers is sub-dominant~\citep{dek09}. In contrast, the virial radius,
environment, and mergers seem to play a pivotal role in quenching star
formation, and more generally in the evolution of massive halos hosting
quiescent galaxies.
With that framework in mind, we summarize the key conclusions of this
paper:
\begin{itemize}
\item Galactic outflows are required to suppress stellar mass growth at
all epochs. The stellar mass function is quite steep and nearly a power
law at high redshifts. At lower redshifts it develops a more pronounced
three-tier behavior, where the middle tier reflects the steepness of
differential (i.e. mass-dependent) wind recycling~\citep{opp10}, and the
higher and lower tiers tend towards the slope of the halo mass function.
The middle tier (typically just below $M^\star$) becomes shallower with
time owing primarily to wind recycling, which returns ejected material
faster to more massive galaxies. Observations of the $z\approx 0$
GSMF also show a three-tier behavior, and the momentum-conserved wind
simulation comes closest to matching its behavior (although the three
tiers are not obviously evident).
\item The star formation rate functions show many of the same general
features as the GSMF, but the detailed evolution, mass dependence,
and dependence on outflows are somewhat different. Comparisons of
these models to present-day data are hampered by their lack of
quenching feedback in massive galaxies in our models. At $z\ga 1$
where massive quenched galaxies are less common, momentum-conserved
winds produces a good match to the observed star formation rate
function derived from H$\alpha$ data.
\item The specific star formation rate's dependence on mass is
governed by how the effective mass loading factor of outflows varies
with stellar mass. The no-wind case shows too little star formation
at a given $M_*$, despite having far too much star formation globally.
The momentum-conserved wind model matches observations well at
$\sim 0.1-1\;M^\star$, but all wind models show too little star
formation at the smallest masses. This may indicate that there
should be less mass loading or more wind recycling at the $M_*\la
0.1 M^\star$ compared to what is currently in these models, or
else the conversion of gas into stars is not being modeled properly
in low-mass systems.
\item The evolution of the sSFR is generally well-described as
being driven by cosmic mass accretion into halos, which scales as
$(1+z)^{2.25}$. The no-wind and slow-wind cases follow this trend at
all redshifts. Observations do so at $z\la 2$, but at $z\ga 2$ they
depart from this scaling, indicating that small high-redshift galaxies
must be significantly suppressed. The momentum-conserved wind model
qualitatively yields this, but not enough to match current high-$z$
observations. Pinning down the sSFR evolution at $z\ga 2$ will have a
major impact on understanding early galaxy evolution and feedback
processes.
\item Satellite galaxies are the minority of galaxies at all masses
probed by these simulations (down to $M_*\sim 10^9M_\odot$), with
a fraction of around one-third at $z=0$ independent of $M_*$. Hence
understanding the growth of galaxies must focus on developing a
model for central galaxies. The satellites' GSMF follows the same
trends as that of centrals. However, there are many more non-star
forming satellites than centrals, particularly at lower masses,
owing to environmental effects wherein the satellites no longer
effectively receive inflow from the IGM; this trend may be contrary
to that observed. Small galaxies are at most only mildly more
bursty than larger ones, and the burst fraction is small in accord
with observations. Our results challenge conventional notions that
most small galaxies are bursty satellites.
\end{itemize}
Taking a broader view, it has long been recognised that the galaxy
formation process must provide some feedback mechanism that preferentially
suppresses mass growth at both high and low masses. In this paper
we have advocated that it is galactic outflows that are primarily
responsible for regulating the low-mass end. The advantages of invoking
this particular mechanism are that (i) outflows are observed and seem
to follow scalings that yield many desirable properties in models; (ii)
they concurrently enrich the IGM at all cosmic epochs as observed; (iii)
they lower the baryon content of halos preferentially to lower masses as
observed~\citep[e.g.][]{dav10d}; and (iv) they are an ejective feedback
mechanism that works in conjunction with the cold accretion paradigm,
in which it is difficult to prevent filaments from channeling gas
efficiently into galaxies. There are other ideas to suppress low-mass
star formation by regulating it as a function of mass internally within
the ISM (i.e. by varying the star formation efficiency as a function
of $M_*$), or by preventing material entering the halo from accreting
into small galaxies. But these scenarios must explain IGM enrichment
separately, and must hide most of the gravitationally-accreted baryons
in halos in some undetected form which is becoming increasingly difficult
to accomodate~\citep[e.g.][]{mcg10,dai10}.
While attractive, the idea of invoking outflows is not without
its difficulties. For instance, the amount of mass being driven
out of these galaxies is large, globally many times the amount of
mass forming into stars. This is a generic result, because some
process must suppress stellar masses by such large factors at the
low mass end~\citep{ker09a,opp10}; in our models, this process is
galactic outflows, as opposed to ISM physics or preventive feedback.
The physical mechanism(s) that would drive such a sizeable amount of
material out of galaxies is unclear. The canonical idea of driving
outflows from overpressurized bubbles in the ISM tends to (when
modeled) create holes by which energy and metals, but little mass,
are ejected~\citep[e.g.][]{maclow99}, and there are difficulties
with entraining cold clouds as observed without disrupting them.
The scalings preferred by our current comparisons to data are those
expected for radiation-driven winds, but it is not well known how far out
the photons' momentum can drive dust, and how far out the dust remains
coupled to the gas. For instance, \citet{yos10} use spectropolarimetry
to determine that the dust in M82 is moving outwards much more slowly
than the gas, suggesting that the gas and dust are not strongly coupled.
Furthermore, the momentum budget of typical stellar populations is, under
the assumption of single photon scattering, insufficent to drive outflows
as appear to be necessary, indicating that supernovae or other sources of
energy are still required. Without a solid understanding of the dynamics
driving outflows, simulations such as the ones presented here cannot be
considered a complete description of the physics of galaxy formation.
Fortunately, observations of outflows across cosmic time are gaining
substantial traction~\citep{mar05,wei09,ste10}, providing empirical
constraints for models while the theory of outflow propagation develops.
Our momentum-conserved winds provides the best match to data
on star-forming galaxies of the four models considered here.
But the agreement is only good in the rather narrow mass range from
$10^{10}\la M_*\la 10^{11} M_\odot$. At larger masses, the current
models require some additional quenching mechanism perhaps associated
with AGN~\citep[e.g.][]{dim05,cro06,gab11}. At smaller masses,
this model produces too vigorous early star formation in dwarfs,
resulting in too-old stellar populations, too steep mass functions,
and insufficient star formation today~\citep[a result also seen in
many SAMs, e.g.][]{fon09,guo10}. This may indicate that feedback in
these dwarf systems evolves in some way not captured by this particular
outflow prescription, or else that there is some additional physical
process that is important. It is possible that the star formation
law is different in smaller systems to make gas consumption less
efficient~\citep[e.g.][]{rob08}, or else that these simulations have
too low a density threshold for star formation (owing to resolution
limitations) such that they do not properly represent star formation
in low surface brightness, gas-rich systems~\citep[e.g.][]{gov10}.
More broadly, none of our simulations can simultaneously reproduce
the shallow stellar mass function, archaeological downsizing, and
star formation rate downsizing as observed; this remains a fundamental
challenge for galaxy formation models. Even though the concordant mass
range of the momentum-conserved wind model is fairly narrow, it is still
a significant step forward as it provides physical insights into the
processes that govern galaxy evolution at all masses, pinpointing both
successes and failures. And it is worth recognizing that this particular
dex in stellar mass happens to be where most of cosmic star formation
occurs, so it is a critical mass range for understanding global stellar
mass growth.
Overall, the general simulation-inspired framework for how inflows
and outflows govern galaxy formation is compelling from both an
observational and theoretical perspective, and is well-situated
within the current hierarchical structure formation paradigm. This
makes it an attractive scenario to pursue both analytically and
numerically. In Paper~II we will examine how this framework provides
intuition about the gaseous and metal content of galaxies, thereby
together covering the primary constituents that determine the
observable multi-wavelength properties of galaxies at all cosmic
epochs.
\section*{Acknowledgements}
The authors acknowledge A. Dekel, M. Fardal, M. Haas, N. Katz, D. Kere\v{s},
J. Kollmeier, C. Papovich, and D. Weinberg for helpful discussions,
S. Weinmann for useful comments on an early draft, J. Schaye for helpful
refereeing, and V. Springel for making {\sc Gadget-2}\ publicly available.
RD thanks North West University in Mafikeng, South Africa and the
University of Cape Town, South Africa for their hospitality during
much of the writing of this paper. The simulations used here were run
on University of Arizona's SGI cluster, ice. Support for this work
was provided by NASA through grant number HST-AR-11751 from the Space
Telescope Science Institute, which is operated by AURA,Inc. under NASA
contract NAS5-26555. This work was also supported by the National
Science Foundation under grant numbers AST-0847667 and AST-0907998.
Computing resources were obtained through grant number DMS-0619881 from
the National Science Foundation.
|
\section{Introduction}
\label{introduction}
The coupling between magnetism and ferroelectricity is one of the intriguing phenomena in solid-state physics.\cite{wang2009} Apart from the ongoing studies of the underlying microscopic mechanisms,\cite{katsura2005,moskvin2008} the effect itself is relevant for applications,\cite{wang2009} and stimulates extensive experimental work on diverse systems varying from bulk transition-metal compounds \cite{kimura2003,ikeda2005,kagawa2010} to heterostructures.\cite{wu2010} Magnetoelectric effects in bulk systems typically conform to one of the two following scenarios: (i) the magnetism arises from transition-metal cations with a partially filled $d$ shell, and the ferroelectricity is driven by lone-pair cations, such as Bi$^{3+}$ or Se$^{4+}$ (Refs.~\onlinecite{khomskii2006,hill2002}), or (ii) electronic effects [a spiral (helicoidal) magnetic structure or a charge ordering] break the symmetry and cause ferroelectricity.\cite{khomskii2006,cheong2007,brink2008} The former mechanism ensures large electric polarization, which, however, is weakly coupled to the magnetism. The latter scenario provides a strong coupling, but a small electric polarization. The combination of the two approaches is clearly advantageous, but difficult to achieve. The best known example is BiFeO$_3$,\cite{catalan2009} which exhibits a plethora of interesting effects related to the coupling between ferroelectricity and magnetism.\cite{zhao2006,rovillain2010}
\begin{figure}
\includegraphics{Fig1}
\caption{\label{structure}
(Color online) The crystal structure of BiMnFe$_2$O$_6$. The Fe and Mn cations are situated in the oxygen octahedra (in the color version, blue and violet polyhedra denote the FeMn1 and FeMn2 positions, respectively).
}
\end{figure}
The recently discovered complex oxide BiMnFe$_2$O$_6$ represents another system combining the two scenarios plausible for multiferroicity: the lone-pair Bi$^{3+}$ cation and the spiral magnetic ground state. Neither of the two, however, lead to ferroelectricity. The polar displacements of Bi$^{3+}$ are ordered in an antiferroelectric manner,\cite{yang2010} whereas the spiral magnetic structure is non-polar due to a strong antiferromagnetic coupling along the crystallographic $c$ direction.
The unique crystal structure of BiMnFe$_2$O$_6$ features fragments of the hypothetical $\textit{hcp}$ oxygen-based MO building blocks that are related by a mirror operation into a polysynthetically twinned structure.\cite{yang2010} The Fe$^{3+}$ ($d^{5}$) and Mn$^{3+}$ ($d^{4}$) cations are nearly randomly distributed over two crystallographically distinct positions, both octahedrally coordinated by oxygen atoms (Fig.~\ref{structure}). The octahedra are interconnected into a framework through corner-, edge-, and face-sharing, providing complex paths for the magnetic exchange. In spite of the random distribution of the Fe and Mn cations, a long-range magnetic order was reported in BiMnFe$_2$O$_6$ below $T_N\simeq 220$~K, according to $^{57}$Fe M\"ossbauer spectroscopy, magnetic susceptibility, and heat capacity measurements. The low-temperature neutron powder diffraction revealed the incommensurate propagation vector of the magnetic structure, yet no details of the ground-state spin arrangement have been reported.\cite{yang2010}
Spiral magnetic structures are capable of inducing ferroelectricity in a number of transition-metal oxides, such as TbMnO$_3$,\cite{kenzelmann2005} Ni$_3$V$_2$O$_8$,\cite{lawes2004,lawes2005} MnWO$_4$,\cite{heyer2006} and Ba$_{0.5}$Sr$_{0.5}$Zn$_2$Fe$_{12}$O$_{22}$.\cite{kimura2005} In contrast to the aforementioned compounds, BiMnFe$_2$O$_6$ remains paraelectric below $T_N$.\cite{yang2010} To understand the lack of ferroelectricity, we studied the magnetic structure and explored the underlying frustrated spin lattice. Although the strong frustration leads to the formation of magnetic helices propagating along $b$, the unfrustrated antiferromagnetic (AFM) coupling along $c$ induces the 222 point symmetry of the magnetically ordered state and cancels the possible electric polarization.
\section{Methods}
\label{methods}
The powder sample of BiMnFe$_2$O$_6$ was synthesized by high-temperature solid-state reaction in air. Stoichiometric amounts of raw materials (Bi$_2$O$_3$, Mn$_2$O$_3$, Fe$_2$O$_3$) were ground thoroughly and heated up to 800~$^{\circ}$C\ in 5~h. After annealing for 10~h, the powder was reground, pressed into a pellet, and heated at 1000~$^{\circ}$C\ for 100~h with several intermediate grindings.
\begin{figure}
\includegraphics{Fig2}
\caption{\label{NPD}
(Color online) Neutron diffraction patterns for BiMnFe$_2$O$_6$ at different temperatures. The arrows show the magnetic reflections.
}
\end{figure}
Neutron powder diffraction (NPD) data were collected on a $\,\simeq\!12$~g sample of BiMnFe$_2$O$_6$, contained in a 9.5~mm diameter vanadium can. A closed cycle He refrigerator was used for temperature control. Patterns were collected with the BT-1 32-detector high-resolution neutron powder diffractometer at the National Institute of Standards and Technology Center for Neutron Research, Gaithersburg, MD. A Cu(311) monochromator with a $90^{\circ}$ takeoff angle and 15 min in-pile collimation was used. The neutron wavelength was 1.5402(1)~\r A. Data from the 32 detectors were combined to give pseudo one-detector data over a total scan range of $3^{\circ}\leq 2\theta\leq 167.75^{\circ}$ with a step size of $0.05^{\circ}$ ($2\theta$). The magnetic structure was analyzed with the JANA2006 program.\cite{jana2006}
Density functional theory (DFT) band structure calculations were performed in the full-potential local-orbital (\texttt{FPLO}) code.\cite{fplo} We used local density approximation (LDA)\cite{pw92} supplied with a mean-field (DFT+$U$) correction for correlation effects in the Fe/Mn 3$d$ shell. The $k$-mesh comprised 192 points for the 40-atom unit cell and 64 points for the 80-atom supercell. The correlated shell was parameterized with an effective on-site Coulomb repulsion $U_d=5$~eV and exchange $J_d=1$~eV,\cite{leonov2004,abakumov2010} whereas the double-counting was corrected in the fully-localized-limit (atomic limit) fashion. To evaluate individual exchange couplings, total energies for a number of collinear spin configurations were mapped onto a classical Heisenberg model. The validity of the computational results was checked by calculations within generalized gradient approximation (GGA)\cite{pbe} and by choosing $U_d$ values of 4 and 6~eV. Similar to Cu$^{2+}$ oxides,\cite{tsirlin2010} the exchange-correlation potential (LDA $vs.$ GGA) has marginal effect on the spin model, whereas a change in $U_d$ only shifts the exchange couplings in a systematic way and keeps their ratios nearly constant. The DFT-based spin model was further studied by classical Monte-Carlo simulations with the \texttt{spinmc} algorithm of the ALPS package.\cite{alb2007}
\begin{table*}
\begin{minipage}{14cm}
\caption{\label{characters}
Characters of the irreducible representations of the little group of the propagation vector $\mathbf k=[0,\beta,0]$ for the space group $Pbcm$ ($a=e^{i\pi\beta}$) and the corresponding magnetic superspace groups. The decomposition of the magnetic representation $\Gamma_{\text{mag}}$ for all four-fold magnetic sites is: $\Gamma_{\text{mag}} = 3\Gamma_1+3\Gamma_2+3\Gamma_3+3\Gamma_4$.}
\begin{ruledtabular}
\begin{tabular}{cccccc}
& $(E\,|\,0\,0\,0)$ & $(m_x\,|\,0\,\frac12\,0)$ & $(2_y\,|\,0\,\frac12\,0)$ & $(m_z\,|\,0\,0\,0)$ & Superspace group \\\hline
$\Gamma_1$ & 1 & $a$ & $a$ & 1 & $Pbcm1'(0\beta0)000s$ \\
$\Gamma_2$ & 1 & $-a$ & $a$ & $-1$ & $Pbcm1'(0\beta0)s0ss$ \\
$\Gamma_3$ & 1 & $-a$ & $-a$ & 1 & $Pbcm1'(0\beta0)s00s$ \\
$\Gamma_4$ & 1 & $a$ & $-a$ & $-1$ & $Pbcm1'(0\beta0)00ss$ \\
\end{tabular}
\end{ruledtabular}
\end{minipage}
\end{table*}
\begin{table*}
\caption{\label{operators}
Symmetry operators for the $Pbcm1'(0\beta0)s00s$ and $Pbcm1'(0\beta0)00ss$ magnetic superspace groups for origin shift $\delta=0,\frac14$ along $x_4$. The operators resulting from combination with the $(1'\,|\,0\,0\,0\,\frac12)$ generator, always present, are not shown ($-m$ means a ``time inversion'' operation, while $m$ is an operation without time inversion). The operators forming the common subgroups are printed in boldface.}
\begin{ruledtabular}
\begin{tabular}{rccc}
& $Pbcm1'(0\beta0)s00s$, & $Pbcm1'(0\beta0)00ss$, & $Pbcm1'(0\beta0)00ss$,\\
& origin at 0,0,0,0 & origin at 0,0,0,0 & origin at $0,0,0,\frac14$ \\\hline
$E$ & $x_1,x_2,x_3,x_4,m$ & $\mathbf{x_1,x_2,x_3,x_4,m}$ & $\mathbf{x_1,x_2,x_3,x_4,m}$ \\
$2_{1,z}$ & $-x_1,-x_2,x_3+\frac12,-x_4,m$ & $-x_1,-x_2,x_3+\frac12,-x_4+\frac12,m$ & $\mathbf{-x_1,-x_2,x_3+\frac12,-x_4,m}$ \\
$2_{1,y}$ & $-x_1,x_2+\frac12,-x_3+\frac12,x_4+\frac12,m$ & $\mathbf{-x_1,x_2+\frac12,-x_3+\frac12,x_4+\frac12,m}$ & $\mathbf{-x_1,x_2+\frac12,-x_3+\frac12,x_4+\frac12,m}$ \\
$2_x$ & $x_1,-x_2+\frac12,-x_3,-x_4+\frac12,m$ & $x_1,-x_2+\frac12,-x_3,-x_4,m$ & $\mathbf{x_1,-x_2+\frac12,-x_3,-x_4+\frac12,m}$ \\
$\bar{1}$ & $-x_1,-x_2,-x_3,-x_4,m$ & $\mathbf{-x_1,-x_2,-x_3,-x_4,m}$ & $-x_1,-x_2,-x_3,-x_4+\frac12,m$ \\
$m$ & $x_1,x_2,-x_3+\frac12,x_4,m$ & $x_1,x_2,-x_3+\frac12,x_4+\frac12,m$ & $x_1,x_2,-x_3+\frac12,x_4+\frac12,m$ \\
$c$ & $x_1,-x_2+\frac12,x_3+\frac12,-x_4+\frac12,m$ & $\mathbf{x_1,-x_2+\frac12,x_3+\frac12,-x_4+\frac12,m}$ & $x_1,-x_2+\frac12,x_3+\frac12,-x_4,m$ \\\medskip
$b$ & $-x_1,x_2+\frac12,x_3,x_4+\frac12,m$ & $-x_1,x_2+\frac12,x_3,x_4,m$ & $-x_1,x_2+\frac12,x_3,x_4,m$ \\
& Subgroup & $P12_1/c11'(0\beta0)s0s$, & $P22_12_11'(0\beta0)0s0s$,
\cr & & origin at $0,0,0,0$ & origin at $0,\frac14,0,0$ \\
\end{tabular}
\end{ruledtabular}
\end{table*}
\section{Results}
\label{results}
\subsection{Magnetic structure}
At room temperature (RT), BiMnFe$_2$O$_6$ crystallizes in an orthorhombic unit cell with $a=5.03590(3)$~\r A, $b=7.07342(4)$~\r A, $c=12.65425(6)$~\r A, and $Pbcm$ space symmetry.\cite{foot11} Below $T_N\simeq 220$~K, extra reflections appear on the NPD patterns (Fig.~\ref{NPD}). These reflections are of magnetic origin and can not be attributed to a structural phase transition because the x-ray diffraction experiment does not show any change down to $T=120$~K.\cite{yang2010} The magnetic reflections on the $T=7$~K pattern are indexed with a propagation vector $\mathbf k=[0,\beta,0]$ with $\beta=0.1379(1)$. The propagation vector is inside of the Brillouin zone and has a star with two arms {$\mathbf k,-\mathbf k$}. The little group of the propagation vector $G_k$ is $Pb2_1m$. In BiMnFe$_2$O$_6$, there are two symmetrically independent magnetic species in the nuclear $Pbcm$ structure: FeMn1 ($8e$: $0.4891,-0.1597,-0.6036$) and FeMn2 ($4a$: $0,0,\frac12$) (Fig.~\ref{structure}). In the little group $Pb2_1m$, this corresponds to two four-fold magnetic sites for the FeMn1 position and one four-fold magnetic site for the FeMn2 position (origin at $0,0,\frac14$). There are four one-dimensional irreducible representations (irreps) for the propagation vector $\mathbf k=[0,\beta,0]$ in the space group $Pbcm$; their characters are given in Table~\ref{characters}. The magnetic structure is transformed according to one of the irreps or their combination. Alternatively, we can describe the symmetry of the incommensurately modulated magnetic structure by embedding it into a higher-dimensional space and applying magnetic superspace groups defined in (3+1)-dimensional superspace.\cite{janner1980,schobinger2006,schobinger2007,petricek2010} The magnetic moment on atom $i$ is expressed as a vector function:
$$\mathbf M_i(x_4)=\mathbf M_{i0}+\sum_{n=1}^{N}[\mathbf M_{ins}\sin(2\pi nx_4)+\mathbf M_{inc}\cos(2\pi nx_4)],$$
where $n$ denotes terms of the Fourier series, \mbox{$x_4=\mathbf k(\mathbf T+\mathbf r_i)$} is an internal coordinate, $\mathbf T$ is the lattice translation of the nuclear structure, and $\mathbf r_i$ is the position of the atom $i$ in the unit cell of the nuclear structure. Monitoring of the intensity of the magnetic and nuclear reflections upon varying temperature revealed that there is no magnetic impact into the intensity of the nuclear reflections, and therefore $\mathbf M_{i0}=0$ for all magnetic sites. Since only the first-order satellites were observed, the Fourier series are reduced to the $n=1$ terms. A magnetic (Shubnikov) superspace group describing the transformations of the magnetic modulation waves can be set in correspondence with each irrep. The magnetic superspace groups are based on generators of the little group $G_k$, but the symmetry elements transforming the propagation vector $\mathbf k$ to $-\mathbf k$ also should be taken into account. This yields four possible magnetic superspace groups listed in Table~\ref{characters}. The explanation of the magnetic superspace group symbols is provided in Ref.~\onlinecite{petricek2010}
All four magnetic superspace groups were tested in the refinement. Acceptable solutions were found in magnetic superspace groups $Pbcm1'(0\beta0)s00s$ and $Pbcm1'(0\beta0)00ss$, both with the same reliability factor for magnetic reflections $R_I^{\text{mag}}=0.057$. In spite of the relatively low reliability factor, the correspondence between the experimental and calculated NPD profile was far from ideal for both models. This indicates that the actual solution requires a combination of the irreps $\Gamma_3$ and $\Gamma_4$, and can be realized in a common subgroup of $Pbcm1'(0\beta0)s00s$ and $Pbcm1'(0\beta0)00ss$. This subgroup depends on a relative shift $\delta$ along the internal space of the conventional origins of two superspace groups. The list of operators of $Pbcm1'(0\beta0)s00s$ and their intersections with the operators of $Pbcm1'(0\beta0)00ss$ for \mbox{$\delta=0,\frac14$} are provided in Table~\ref{operators}. The resulting common subgroups are $P12_1/c11'(0\beta0)s0s$ and $P22_12_11'(0\beta0)0s0s$. For any $\delta$ values inequivalent to the cases mentioned above the common subgroup is $P12_111'(0\beta0)ss$.
The solutions in the $P12_1/c11'(0\beta0)s0s$ and $P22_12_11'(0\beta0)0s0s$ groups provide the same quality of the Rietveld fit and can not be distinguished on this basis. The $P12_1/c11'(0\beta0)s0s$ magnetic superspace symmetry results in a collinear magnetic structure with an antiferromagnetic transverse amplitude modulated wave with the magnetic moments confined to the $ac$ plane. The amplitude of the magnetic moment modulation varies from almost zero to 5.6~$\mu_B$, which is unrealistically high for the Mn$^{3+}$ and Fe$^{3+}$ cations. Thus, the $P12_1/c11'(0\beta0)s0s$ solution was ruled out. In the $P22_12_11'(0\beta0)0s0s$ model, there are three magnetic symmetrically unequivalent atoms: FeMn1a, FeMn1b, and FeMn2 (see Table \ref{coordinates}), all at the general four-fold sites of the $P22_12_1$ space group. The $(1'\,|\,0\,0\,0\,\frac12)$ operator of the magnetic superspace group requires the $\mathbf M_{i0}$ term to be zero and constrains the magnetic moment modulation functions to odd harmonics only, thereby resulting in the absence of the magnetic impact into intensity of the nuclear reflections and the absence of even-order magnetic satellites.\cite{petricek2010} The components of the magnetic moment modulation function for the general four-fold site are related by symmetry elements of the $P22_12_11'(0\beta0)0s0s$ magnetic superspace group, as shown in Table~\ref{functions}.
\begin{table*}
\caption{\label{coordinates}
Crystallographic parameters and atomic coordinates in BiMnFe$_2$O$_6$ at $T=7$~K and room temperature (RT). Atomic displacement parameters $U_{\text{iso}}$ are given in $10^{-2}$~\r A$^2$. The symmetry operators of the $P22_12_1$ space group are listed in Table~\ref{operators}.}
\begin{ruledtabular}
\begin{tabular}{c@{\hspace{3em}}rrrr@{\hspace{3em}}rrrr}
& \multicolumn{4}{c}{$T=7$~K} & \multicolumn{4}{c}{RT} \\\hline
Space group & \multicolumn{4}{c}{$P22_12_1$} & \multicolumn{4}{c}{$Pbcm$} \\
$a$ (\r A) & \multicolumn{4}{c}{5.02305(7)} & \multicolumn{4}{c}{5.03589(3)} \\
$b$ (\r A) & \multicolumn{4}{c}{7.06232(9)} & \multicolumn{4}{c}{7.07341(4)} \\
$c$ (\r A) & \multicolumn{4}{c}{12.6424(2)} & \multicolumn{4}{c}{12.65425(6)} \\\hline
Atom & \multicolumn{1}{c}{$x$} & \multicolumn{1}{c}{$y$} & \multicolumn{1}{c}{$z$} & $U_{\text{iso}}$ & \multicolumn{1}{c}{$x$} & \multicolumn{1}{c}{$y$} & \multicolumn{1}{c}{$z$} & \multicolumn{1}{c}{$U_{\text{iso}}$} \\\hline
Bi & 0.9711(3) & $-0.1301(2)$ & 0.751(1) & 0.27(4) & 0.9705(1) & $-0.1305(1)$ & $\frac34$ & 0.68(2) \\
FeMn1a\footnote{FeMn1a and FeMn1b: 0.7Fe+0.3Mn;\quad FeMn2: 0.6Fe+0.4Mn} & 0.486(3) & $-0.160(2)$ & $-0.607(1)$ & 0.32(8) & 0.4891(3) & $-0.1597(2)$ & $-0.60355(7)$ & 0.80(3) \\
FeMn1b$^\text{a}$ & $-0.488(3)$ & 0.160(3) & 0.607(1) & 0.32(8) & \\
FeMn2$^\text{a}$ & 0.000(6) & 0.000(6) & 0.500(2) & 0.32(8) & 0 & 0 & $\frac12$ & 0.37(4) \\
O1a & 0.157(3) & 0.439(3) & $-0.6349(9)$ & 0.47(2) & 0.1629(4) & 0.4372(3) & $-0.6366(2)$ & 0.65(5) \\
O1b & $-0.162(3)$ & $-0.441(3)$ & 0.6386(9) & 0.47(2) & \\
O2a & 0.672(2) & $-0.408(2)$ & $-0.5783(9)$ & 0.47(2) & 0.6641(4) & $-0.4077(4)$ & $-0.5765(2)$ & 0.96(5) \\
O2b & $-0.660(2)$ & $-0.593(2)$ & 0.5746(9) & 0.47(2) & \\
O3a & 0.787(3) & $\frac14$ & $\frac12$ & 0.47(2) & 0.7913(6) & $\frac14$ & $\frac12$ & 0.86(7) \\
O3b & $-0.799(3)$ & $\frac34$ & $\frac12$ & 0.47(2) & \\
O4 & 0.3444(6) & $-0.3044(4)$ & 0.753(1) & 0.47(2) & 0.3457(7) & $-0.3022(4)$ & $\frac34$ & 0.69(7) \\
\end{tabular}
\end{ruledtabular}
\end{table*}
Since in the nuclear structure the atoms FeMn1a and FeMn1b are crystallographically equivalent, it is reasonable to assume that the modulation of the magnetic moment at these positions follows the same type of modulation waves. The solution was found with the following restrictions (not imposed by the magnetic superspace group) on the coefficients of the magnetic moment modulation functions:
\begin{widetext}
\begin{center}
\begin{minipage}{12cm}
\begin{align*}
M_{s,z}\text{(FeMn1a)}&=& M_{c,x}\text{(FeMn1a)}&=& M_{s,z}\text{(FeMn1b)}&=& M_{c,x}\text{(FeMn1b)}&=& M_1 \\
M_{s,x}\text{(FeMn1a)}&=& -M_{c,z}\text{(FeMn1a)}&=& -M_{s,x}\text{(FeMn1b)}&=& M_{c,z}\text{(FeMn1b)}&=& M_2 \\
M_{s,z}\text{(FeMn2)}&=& M_{c,x}\text{(FeMn2)} && && &=& M_3 \\
M_{c,z}\text{(FeMn2)}&=& M_{s,x}\text{(FeMn2)} && && &=& 0 \\
\end{align*}
\end{minipage}
\end{center}
\end{widetext}
The refined magnetic moment components along the $b$ axis were smaller than their standard deviations and were fixed to zero. The refined magnetic parameters at different temperatures are provided in Table~\ref{moments}. The experimental, calculated, and difference NPD profiles at $T=7$~K are shown in Fig.~\ref{Rietveld}. The temperature dependencies of the ordered magnetic moment for the FeMn1 and FeMn2 positions are given in Fig.~\ref{momtemp}. Extrapolations of these dependencies with the $M=M_0[(1-T/T_N)^{\alpha}]^{\beta}$ function give coinciding $T_N$ values of 221(2)~K and 219(4)~K for the FeMn1 and FeMn2 positions, respectively, and the magnetic moments of $M_0\text{(FeMn1)}=3.76(4)$~$\mu_B$ and $M_0\text{(FeMn2)}=4.0(1)$~$\mu_B$.
\begin{table*}
\begin{minipage}{14cm}
\caption{\label{functions}
Symmetry-imposed relations between the components of the magnetic moment modulation functions for the general four-fold site of the $P22_12_11'(0\beta0)0s0s$ magnetic superspace group with the origin at 0,$\frac14$,0,0.}
\begin{ruledtabular}
\begin{tabular}{cccc}
Operator & \multicolumn{3}{c}{Components of $\mathbf M_i(x_4)$} \\\hline
$(E\,|\,0\,0\,0,0)$ & $M_x(x_4)$ & $M_y(x_4)$ & $M_z(x_4)$ \\
$(2_x\,|\,0\,\frac12\,0, \frac12)$ & $M_x(-x_4+\frac12)$ & $M_y(-x_4)$ & $M_z(-x_4)$ \\
$(2_y\,|\,0\,\frac12\,\frac12, \frac12)$ & $M_x(x_4)$ & $M_y(x_4+\frac12)$ & $M_z(x_4)$ \\
$(2_z\,|\,0\,0\,\frac12, 0)$ & $M_x(-x_4+\frac12)$ & $M_y(-x_4+\frac12)$ & $M_z(-x_4)$ \\
\end{tabular}
\end{ruledtabular}
\end{minipage}
\end{table*}
\begin{table*}
\begin{minipage}{14cm}
\caption{\label{moments}
Refined magnetic parameters for BiMnFe$_2$O$_6$ at different temperatures (see text for notations). Reliability factors are listed in the order of $R_I$ (overall), $R_I$ (nuclear reflections), $R_I$ (magnetic satellites), $R_P$.}
\begin{ruledtabular}
\begin{tabular}{ccccccc}
$T$ & $M_1$ & $M_2$ & $M_3=M$(FeMn2) & $M$(FeMn1) & $\beta$ & $R$-factors \\
(K) & ($\mu_B$) & ($\mu_B$) & ($\mu_B$) & ($\mu_B$) & & \\\hline
7 & 3.25(2) & 1.94(2) & 3.98(3) & 3.80(3) & 0.13801(8) & 0.017, 0.014, 0.028, 0.036 \\
57 & 3.13(2) & 1.87(2) & 3.71(2) & 3.65(3) & 0.13740(4) & 0.016, 0.013, 0.025, 0.030 \\
100 & 3.09(4) & 1.75(3) & 3.50(4) & 3.55(5) & 0.1316(1) & 0.025, 0.022, 0.040, 0.048 \\
150 & 2.74(6) & 1.40(5) & 2.57(6) & 3.08(8) & 0.1232(2) & 0.030, 0.029, 0.036, 0.067 \\
200 & 1.69(6) & 0.88(5) & 1.56(6) & 1.91(8) & 0.1175(3) & 0.029, 0.027, 0.044, 0.051 \\
215 & 0.95(9) & 0.54(9) & 0.68(9) & 1.1(1) & 0.1155(9) & 0.028, 0.028, 0.036, 0.049 \\
\end{tabular}
\end{ruledtabular}
\end{minipage}
\end{table*}
\begin{figure}
\includegraphics{Fig3}
\caption{\label{Rietveld}
(Color online) Experimental, calculated, and difference NPD pattern for BiMnFe$_2$O$_6$ at $T=7$~K. The black (dark) and green (light) bars mark the reflection positions for the nuclear and magnetic structures, respectively.
}
\end{figure}
\begin{figure}
\includegraphics{Fig4}
\caption{\label{momtemp}
(Color online) Temperature dependence of the magnetic moment for the FeMn1 and FeMn2 positions.
}
\end{figure}
The spin arrangement in the BiMnFe$_2$O$_6$ magnetic structure is shown in Fig.~\ref{spinstr}. BiMnFe$_2$O$_6$ adopts a spiral magnetic structure consisting of antiferromagnetic helixes propagating along the $b$-axis with a period of $\,\simeq\!3.5b$. The helixes are associated with the FeMn1 and FeMn2 atomic chains running along the $b$-axis, where the magnetic atoms are separated by $b/2$. The magnetic moments rotate for $\pi(1+\beta)\simeq 204.8$~deg about the $b$-axis on going between the adjacent magnetic atoms in the chains. The refined parameters $M_1$, $M_2$, and $M_3$ represent the magnetic moment of the FeMn1 atom $\left[M\text{(FeMn1)}=\sqrt{M^2_1+M^2_2}\right]$, the magnetic moment of the FeMn2 atom $\left[ M\text{(FeMn2)}=|M_3|\,\right]$, and the phase shift $\varphi$ between the FeMn1 and FeMn2 helixes \mbox{($\tan\varphi=-{\frac{M_2}{M_1}}$)}.
The refined structural parameters and interatomic distances at $T=7$~K (space group $P22_12_1$) and RT (space group $Pbcm$) are listed in Tables \ref{coordinates} and \ref{distances}, respectively. The crystal structure at $T=7$~K was refined with fixed parameters of the magnetic structure. The crystal structures at both temperatures are virtually identical indicating that the magnetic ordering does not influence the nuclear structure. With the $P22_12_11'(0\beta0)0s0s$ magnetic symmetry, the space group of the average nuclear structure should be $P22_12_1$ (point group 222). Although the spiral magnetic ordering eliminates the inversion center, it does not create a polar direction. Indeed, no indication of ferroelectricity below $T_N$ was found in BiMnFe$_2$O$_6$ by dielectric permittivity measurements.\cite{yang2010}
\begin{table*}
\begin{minipage}{16cm}
\caption{\label{distances}
Selected interatomic distances (in~\r A) in BiMnFe$_2$O$_6$ at $T=7$~K and room temperature (RT).}
\begin{ruledtabular}
\begin{tabular}{cc@{\hspace{4em}}cc@{\hspace{4em}}cc}
Bond & Length & Bond & Length & Bond & Length \\\hline
$T=7$~K & & & & & \\
Bi--O1a & 2.20(2), 2.69(2) & FeMn1a--O1b & 1.95(2) & FeMn2--O1a & 1.93(3) \\
Bi--O1b & 2.16(2), 2.70(2) & FeMn1a--O2a & 2.02(2), 2.56(2) & FeMn2--O1b & 1.98(3) \\
Bi--O4 & 2.243(3) & FeMn1a-O2b & 1.97(2) & FeMn2--O2a & 2.03(3)\
\cr & & FeMn1a--O3b & 2.07(2) & FeMn2--O2b & 2.06(3) \
\cr & & FeMn1a--O4 & 2.00(2) & FeMn2--O3a & 2.07(4) \
\cr & & FeMn1b--O1a & 1.95(2) & FeMn2--O3b & 2.03(4) \
\cr & & FeMn1b--O2a & 1.99(2) & & \
\cr & & FeMn1b--O2b & 1.99(2), 2.51(2) & & \
\cr & & FeMn1b--O3a & 2.03(2) & & \\
\cr & & FeMn1b--O4 & 1.92(2) & & \\\hline
RT & & & & & \\
Bi--O1($\times$2) & 2.207(2) & FeMn1--O1 & 1.928(3) & FeMn2--O1($\times$2) & 1.964(2) \\
Bi--O1($\times$2) & 2.686(2) & FeMn1--O2 & 1.973(3) & FeMn2--O2($\times$2) & 2.056(2) \\
Bi--O4 & 2.246(3) & FeMn1--O2 & 1.992(3) & FeMn2-O3($\times$2) & 2.057(2)\\
Bi--O2($\times$2) & 2.696(2) & FeMn1--O4 & 2.007(2) & & \\
Bi--O4 & 2.816(3) & FeMn1--O3 & 2.029(3) & & \
\cr & & FeMn1--O2 & 2.489(2) & & \\
\end{tabular}
\end{ruledtabular}
\end{minipage}
\end{table*}
\subsection{Electronic structure}
Owing to the complex crystal structure, an empirical assignment of individual exchange couplings in BiMnFe$_2$O$_6$ is a formidable challenge. The problem can be solved by electronic structure calculations that evaluate individual couplings and, therefore, establish a reliable microscopic magnetic model. Recent studies prove the remarkable accuracy of DFT for diverse correlated systems, including frustrated magnets with highly intricate spin lattices.\cite{gorelov2010,mazurenko2008,tsirlin2010a} Prior to computing exchange integrals, we discuss the electronic structure of BiMnFe$_2$O$_6$, which is a key to understanding the magnetic behavior.
\begin{figure}
\includegraphics{Fig5}
\caption{\label{spinstr}
(Color online) The arrangement of magnetic moments in BiMnFe$_2$O$_6$. The unit cell for the nuclear structure is outlined.
}
\end{figure}
Computational analysis of BiMnFe$_2$O$_6$ is complicated by the intrinsic disorder of Fe and Mn atoms. Unfortunately, state-of-the-art computational tools, such as virtual crystal approximation (VCA) or coherent potential approximation (CPA), cannot be applied to this case, since they provide an averaged description of a disordered system, while we are targeting magnetic properties that depend on local interactions. The insulating nature of BiMnFe$_2$O$_6$ \cite{yang2010} suggests localized 3$d$ electrons of transition metals; therefore, the distinct spin-$\frac52$ Fe$^{3+}$ ($d^{5}$) and \mbox{spin-2} Mn$^{3+}$ ($d^{4}$) sites are randomly distributed on the spin lattice. Further, orbital ordering for Mn$^{3+}$ results in dramatic differences among the Fe--O--Fe, Mn--O--Mn, and Fe--O--Mn superexchanges (see below). An averaged VCA or CPA picture would not reproduce any of these features. To capture effects arising from the localized moments of Fe$^{3+}$ and Mn$^{3+}$, we calculated exchange integrals for several systems with ordered Fe and Mn atoms, and accessed all possible scenarios of the superexchange. In these calculations, we kept the experimental atomic positions, but imposed different arrangements of Fe and Mn, spanning purely Fe (BiFe$_3$O$_6$) and purely Mn (BiMn$_3$O$_6$) cases as well as intermediate (BiMnFe$_2$O$_6$ and BiMn$_2$FeO$_6$) configurations.
The LDA band structure depends only marginally on the Fe/Mn ordering. The density of states (DOS) features Bi 6$s$ bands below $-10$~eV, O $2p$ states between $-7$~eV and $-2$~eV, transition-metal 3$d$ states at the Fermi level, and Bi $6p$ states above 3~eV (Fig.~\ref{ldados}). The effect of substituting Mn for Fe is a change in the electron count and the ensuing shift of the Fermi level within the $3d$ bands. Irrespective of the Fe/Mn ratio, LDA band structures are metallic due to the heavy underestimation of strong electronic correlations. The insulating spectrum is correctly reproduced by DFT+$U$ (see upper panel of Fig.~\ref{lda+udos}). In particular, the band gaps of about 1.5~eV for BiFe$_3$O$_6$ and 1.1~eV for BiMn$_3$O$_6$ at $U_d=5$~eV are consistent with the black color of the compound.
\begin{figure}
\includegraphics{Fig6}
\caption{\label{ldados}
(Color online) LDA density of states for BiFe$_3$O$_6$. The Fermi level is at zero energy. The gapless energy spectrum is caused by underestimation of the electronic correlations in LDA.}
\end{figure}
Superexchange couplings in insulating compounds intimately depend on the orbital state of the transition metal. Five unpaired electrons of Fe$^{3+}$ fill five $d$ orbitals and leave no orbital degrees of freedom. By contrast, Mn$^{3+}$ has four unpaired electrons only, and hence one of the $d$ orbitals is unoccupied in the Mott-insulating state. Since there is a sizable crystal-field splitting (about 1.5~eV) driven by the octahedral coordination of the FeMn1 and FeMn2 sites, the three $t_{2g}$ states are half-occupied, while the two $e_g$ states have to share the remaining electron. The half-filled orbital is picked out by a weak distortion of the octahedral local environment. In the following, we analyze such distortions in more detail to determine which of the two $e_g$ orbitals is half-filled in the insulating state.
\begin{figure}
\includegraphics{Fig7}
\caption{\label{lda+udos}
(Color online) Top: LSDA+$U$ DOS for one spin channel in the ground-state AFM configuration of BiMn$_3$O$_6$ ($U_d$ = 5~eV). Middle and bottom: orbital-resolved DOS for 3$d$ states of Mn$^{3+}$. Four out of five $d$ orbitals show filled spin-up and empty spin-down states, whereas the states of the remaining $d$ orbital ($d_{x^{2}-y^{2}}$ for Mn1 and $d_{3z^{2}-r^{2}}$ for Mn2) are mostly above the Fermi level for both spin directions. The Fermi level is at zero energy.}
\end{figure}
The FeMn1 position reveals one long bond of about 2.5~\r A (FeMn1--O2, Table~\ref{distances}) that we choose as the local $z$ axis. According to simple electrostatic arguments of crystal-field theory, the long bond along $z$ shifts the $d_{3z^{2}-r^{2}}$ orbital down in energy with respect to the $d_{x^{2}-y^{2}}$ orbital. The opposite scenario is found for FeMn2, where the local environment is a squeezed octahedron. The short FeMn2--O1 bond (about 1.96~\r A, Table~\ref{distances}), which we take as the local $z$-axis, shifts the $d_{3z^{2}-r^{2}}$ orbital up in energy. Therefore, the unpaired electron of Mn$^{3+}$ should occupy the $d_{3z^{2}-r^{2}}$ orbital for FeMn1 and the $d_{x^{2}-y^{2}}$ orbital for FeMn2. This conclusion is verified by DFT+$U$ calculations that place unpaired electrons on the respective orbitals, while shifting both spin-up and spin-down states for the remaining $e_g$ orbital ($d_{x^{2}-y^{2}}$ and $d_{3z^{2}-r^{2}}$ for FeMn1 and FeMn2, respectively) above the Fermi level (see Fig.~\ref{lda+udos}).\cite{foot12}
The orbital order in BiMn$_3$O$_6$ is formally of the ferro-type within the FeMn1 and FeMn2 sublattices, yet of the antiferro-type between the two sublattices. However, this notation is deceptive because of the different local axes on neighboring atoms. The magnetic model largely deviates from the conventional scenario\cite{kugel1982} of AFM superexchange for ferro-type orbital order and ferromagnetic (FM) superexchange for antiferro-type orbital order. In the following, we show that the peculiar arrangement of empty $e_g$ orbitals in BiMn$_3$O$_6$ induces sizable FM superexchange for nearly all couplings and alters the spiral ground state that arises from the purely AFM spin lattice of BiFe$_3$O$_6$.
\subsection{Microscopic magnetic model}
Exchange couplings calculated for different superexchange scenarios (Fe--O--Fe, \mbox{Mn--O--Mn}, Fe--O--Mn, and Mn--O--Fe) are listed in Table \ref{couplings}. The Fe--O--Fe, Mn--O--Mn, and Fe--O--Mn superexchanges show sharp differences for most of the couplings, while the Fe--O--Mn and Mn--O--Fe cases are only different for interactions between the FeMn1 and FeMn2 sublattices (i.e., when the two metal sites are not related by symmetry). In Table~\ref{couplings}, we restrict ourselves to short-range couplings matching direct connections between the FeMn octahedra. Long-range interactions are expected to be weak, as confirmed by the following qualitative argument. Long-range superexchange requires suitable overlap of atomic orbitals along a M--O--O--M (or even more complex) pathway, and can be achieved for one orbital channel only. By contrast, short-range superexchange is possible for most of the orbital channels, and should therefore dominate in systems with several magnetic orbitals. To verify this conclusion for BiMnFe$_2$O$_6$, we calculated the exchange couplings within the crystallographic unit cell and within a supercell doubled along $a$. The resulting values of $J_i$ agreed within $10\,$\% and indicated weak long-range couplings in BiMnFe$_2$O$_6$. Below, we demonstrate that our minimum microscopic model, restricted to short-range couplings, is sufficient to explain the spiral magnetic structure of BiMnFe$_2$O$_6$ and the lack of ferroelectricity in this compound.
\begin{table*}
\begin{minipage}{14cm}
\caption{\label{couplings}
Interatomic distances (in~\r A) and leading exchange couplings (in~K) in BiMnFe$_2$O$_6$ calculated with the supercell procedure (LSDA+$U$, $U_d$ = 5~eV) for different scenarios of superexchange. Negative $J_i$ denotes FM coupling. The intralayer couplings $J_1-J_7$ and $J_9$ are depicted in Fig.~\ref{spinlattice}, whereas $J_8$ connects the layers along $c$.}
\begin{ruledtabular}
\begin{tabular}{cc@{\hspace{3em}}ccc}
& Distance & \multicolumn{3}{c}{Exchange couplings $J_i$} \\\hline
& & Fe--O--Fe & Mn--O--Mn & Fe--O--Mn/Mn--O--Fe \\
$J_1$ & 2.916 & 1 & $-1$ & $-50$ \\
$J_2$ & 3.010 & 19 & $-9$ & 29/10 \\
$J_3$ & 3.101 & 25 & $-38$ & 2/$-23$ \\
$J_4$ & 3.462 & 26 & 7 & $-38$ \\
$J_5$ & 3.537 (FeMn2) & 30 & 13 & 43 \\
$J_6$ & 3.538 (FeMn1) & 61 & $-8$ & 48 \\
$J_7$ & 3.685 & 37 & $-1$ & 23/30 \\
$J_8$ & 3.706 & 74 & 79 & 127 \\
$J_9$ & 3.759 & 66 & $-12$ & 20/70 \\
\end{tabular}
\end{ruledtabular}
\end{minipage}
\end{table*}
The spin lattice of BiMnFe$_2$O$_6$ incorporates nine inequivalent exchanges (Fig.~\ref{spinlattice} and Table~\ref{couplings}). Eight of these couplings are found in the $ab$ plane, whereas $J_8$ connects the layers along~$c$. The interlayer coupling is weakly influenced by the Fe/Mn substitution, and remains one of the leading AFM interactions for all superexchange scenarios (Table \ref{couplings}). The robustness of $J_8$ should be traced back to the orbital order for Mn$^{3+}$. The local $z$-axis of the FeMn1 octahedra roughly matches the crystallographic $c$ direction. Therefore, the replacement of Fe by Mn results in an empty $d_{x^{2}-y^{2}}$ orbital (Fig.~\ref{lda+udos}) that contributes weakly to the superexchange along $c$. By contrast, most of the intralayer couplings are heavily affected by changing Mn for Fe.
In BiFe$_3$O$_6$, eight intralayer couplings are AFM. Excluding the apparently weak $J_1$, we arrive at seven AFM couplings ranging from 19 K to 74 K (Table~\ref{couplings}). Triangular loops abound (Fig.~\ref{spinlattice}) and lead to the strong frustration of the spin lattice. This frustration is largely released by Mn$^{3+}$ that renders most of the intralayer couplings FM, and reduces the remaining AFM couplings below 13 K (Table \ref{couplings}). The Fe--O--Mn and Mn--O--Fe cases are intermediate, with partially reduced AFM couplings. Large FM contributions to the superexchange can be traced back to empty $d$ orbitals of Mn$^{3+}$. These orbitals provide the strong $\sigma$-overlap with oxygen orbitals and a leading contribution to the superexchange.
\begin{figure}
\includegraphics{Fig8}
\caption{\label{spinlattice}
(Color online) Spin lattice (left) and the respective part of the crystal structure (right) showing frustrated exchange couplings in the $ab$ plane. Open and filled circles denote the FeMn1 and FeMn2 positions, respectively.}
\end{figure}
To evaluate the ground state of the proposed model, we performed classical Monte-Carlo simulations for a \mbox{$16\times16\times16$} finite lattice with periodic boundary conditions. The unit cell of the spin lattice incorporated six magnetic atoms within one layer (half of the crystallographic unit cell). The temperature was set to 10 K, well below the ordering temperature $T_N$ (see further in this section). Spin-spin correlation functions for purely Fe$^{3+}$ ($S=\frac52$) and purely Mn$^{3+}$ ($S=2$) lattices are given in Table~\ref{correlations}. The correlation functions normalized for $S^{2}$ are $-1$ for antiparallel spins and $+1$ for parallel spins. Intermediate values indicate non-collinear configurations.
\begin{table}
\caption{\label{correlations}
Normalized spin-spin correlations $\langle S_iS_j\rangle/S^2$ for BiFe$_3$O$_6$ (Fe$^{3+}$, $S=\frac52$) and BiMn$_3$O$_6$ (Mn$^{3+}$, $S=2$).}
\begin{minipage}{6cm}
\begin{ruledtabular}
\begin{tabular}{crr}
& BiFe$_3$O$_6$ & BiMn$_3$O$_6$ \\\hline
$J_1$ & 0.288 & 0.957 \\
$J_2$ & 0.240 & 0.954 \\
$J_3$ & 0.234 & 0.974 \\
$J_4$ & $-0.835$ & 0.945 \\
$J_5$ & $-0.693$ & 0.940 \\
$J_6$ & $-0.706$ & 0.951 \\
$J_7$ & $-0.827$ & 0.945 \\
$J_8$ & $-0.992$ & $-0.993$ \\
$J_9$ & $-0.835$ & 0.956 \\
\end{tabular}
\end{ruledtabular}
\end{minipage}
\end{table}
We consider the Fe$^{3+}$ case first. The normalized spin-spin correlation for $J_8$ is close to $-1$; therefore, the interlayer ordering is collinear and AFM. Similar correlations on the $J_5$ and $J_6$ bonds indicate the same propagation vector along $b$ for the FeMn1 and FeMn2 sublattices. Further on, similar correlations on the $J_2$ and $J_3$ as well as on $J_7$ and $J_9$ bonds signify the same magnetic order along the respective bonds and, consequently, a twice shorter periodicity along $b$ for the magnetic unit cell compared to the crystallographic unit cell. To find out the ordering pattern along $a$, we note that the respective FeMn2 atoms are connected via $J_2$ and $J_3$ or $J_7$ and $J_9$ bonds. The correlations along these paths are different; therefore, the spins on the FeMn2 atoms should be parallel (i.e., the magnetic moment rotates for a certain angle $\varphi_1$ on the $J_2$ bond and for the opposite angle $-\varphi_1$ on the $J_3$ bond). Thus, the propagation vector of the magnetic structure is $[0,\beta,0]$, in agreement with the experiment. The principal spin arrangement is described by three parameters, which are the angles between the spins on the $J_5$ ($J_6$), $J_2$, and $J_7$ bonds. According to our simulations, these angles are $\varphi=226$~deg, $\varphi_1=76$~deg, and $\varphi_2=146$~deg, respectively, in remarkable agreement with the experimental values of $\varphi=206$~deg, $\varphi_1=39$~deg, and $\varphi_2=167$~deg. Our microscopic model for BiFe$_3$O$_6$ reproduces the experimental magnetic structure of BiMnFe$_2$O$_6$ quite well. The remaining discrepancies are likely related to the partial replacement of Fe by Mn.
The complete substitution of Mn for Fe changes the magnetic ground state. According to Table \ref{correlations}, most of the normalized spin-spin correlations in BiMn$_3$O$_6$ are close to +1. The only negative correlation refers to the $J_8$ bond and indicates AFM interlayer coupling. The intralayer ordering is now collinear FM due to predominantly FM exchange couplings ($J_1$, $J_2$, $J_3$, $J_6$, $J_7$, and $J_9$, see Table \ref{couplings}). The spin lattice is still frustrated by AFM couplings $J_4$ and $J_5$, which, however, are not strong enough to induce the spiral order. While BiFe$_3$O$_6$ and BiMn$_3$O$_6$ present two opposite scenarios of strong and weak frustration, respectively, BiMnFe$_2$O$_6$ lies between these distinct regimes. The spiral ground state of this compound is driven by the frustration of intralayer exchange couplings that remain predominantly AFM for mixed Fe--O--Mn superexchange pathways (see Table \ref{couplings}).
For an additional test of our microscopic model, we calculated magnetic susceptibility ($\chi$) and estimated the N\'eel temperature $T_N$ as the position of the kink in the temperature dependence of $\chi$. Transition temperatures for BiFe$_3$O$_6$ (420~K) and BiMn$_3$O$_6$ (150~K) reasonably agree with the experimental $T_N$ of 220~K that lies between the two calculated values.
\section{Discussion and summary}
\label{discussion}
Our experimental and computational study provides microscopic insight into the physics of BiMnFe$_2$O$_6$. This compound features a strongly frustrated spin lattice with predominantly AFM exchange couplings that induce the spiral magnetic order. The comparison between BiFe$_3$O$_6$ and BiMn$_3$O$_6$ suggests that AFM exchange couplings are an essential prerequisite for the spiral magnetic order. The crystal structure itself may still allow for different ground states; therefore, it is the transition-metal cation that determines the type of the long-range magnetic order. In this respect, first reports on cation substitution\cite{yang2010} look promising, because the change in the Fe/Mn ratio or an incorporation of other magnetic cations could alter the magnetic structure and other physical properties.
BiMnFe$_2$O$_6$ is a magnetic compound that conforms to two mechanisms of ferroelectricity. First, lone pairs of Bi$^{3+}$ induce polar displacements. Second, the non-collinear magnetic structure breaks the inversion symmetry and allows for magnetic-field-induced electric polarization. However, none of the two mechanisms succeed in rendering the compound ferroelectric. Bi$^{3+}$-related polar displacements form an antiferroelectric pattern that leads to zero net polarization. The electronic mechanism meets a similar obstacle of the AFM interlayer coupling and the ensuing non-polar (albeit non-centrosymmetric) magnetic structure. The lack of polarity in both atomic and magnetic structures naturally explains the absence of the ferroelectric response in BiMnFe$_2$O$_6$ below $T_N$.\cite{yang2010}
Our results demonstrate that the simple criteria of ferroelectricity and multiferroicity are not universal, since the polarization induced by any kind of polar distortion (atomic displacement or spin spiral) can be wiped out by an overall antiferroelectric/antiferromagnetic order. The combination of lone-pair and transition-metal cations does not necessarily lead to a magnetic ferroelectric, whereas an arbitrary incommensurate magnetic structure may not allow for the electronic mechanism of ferroelectricity. These simple observations make the search for multiferroics a challenge, and put forward the charge-ordering mechanism as a more robust approach to the design of magnetoelectric materials.
The cancellation of polarity in an incommensurate magnetic structure has been proposed for the spin-chain cuprate NaCu$_2$O$_2$\cite{capogna2010} and for a number of layered compounds, such as $\alpha$-CaCr$_2$O$_4$\cite{chapon2011,toth2011} and \mbox{$\alpha$-SrCr$_2$O$_4$},\cite{dutton2011} that were considered as potential multiferroics. The non-frustrated AFM interlayer coupling seems to be a general obstacle for ferroelectricity in layered systems, like $\alpha$-CaCr$_2$O$_4$,\cite{singh} or in three-dimensional systems with two-dimensional frustrated units, as in BiMnFe$_2$O$_6$. To overcome this problem, one has to design materials with FM or frustrated interlayer couplings. In the case of BiMnFe$_2$O$_6$, cation substitutions preserve the long-range order \cite{yang2010} and can be promising for tuning this compound toward ferroelectric and possibly multiferroic behavior.
In summary, we have solved the spiral magnetic structure of BiMnFe$_2$O$_6$ and proposed a microscopic magnetic model for this compound. The ground state features spins lying in the $ac$ plane and propagating along $b$ by a $\,\simeq\!204.8$ deg rotation about the $b$ axis. The two inequivalent FeMn positions reveal the same propagation vector. The spiral magnetic structure is driven by the strong frustration of antiferromagnetic exchange couplings on a complex spin lattice in the $ab$ plane. However, the coupling along the $c$ direction is non-frustrated, leading to antiparallel spin arrangement in neighboring layers. The resulting magnetic structure is non-polar and, alike the centrosymmetric (antiferroelectric) atomic structure, precludes the ferroelectric behavior of BiMnFe$_2$O$_6$.
\acknowledgments
We are grateful to Oleg Janson for granting access to the \texttt{flyswatter} program and fruitful discussions. A.T. was funded by Alexander von Humboldt Foundation. Contributions by Dr. Judith Stalick of NIST are gratefully acknowledged. Use of the National Synchrotron Light Source, Brookhaven National Laboratory, was supported by the U.S. Department of Energy, Office of Basic Energy Sciences, under Contract No. DE-AC02-98CH10886.
|
\section{Introduction}
Magnetic helicity is an important quantity in dynamo theory
\cite{FrischPouquet1975,BrandenbSubramanianReview2005},
astrophysics \cite{RustKumar1996ApJ,Low1996SoPh}
and plasma physics
\cite{Taylor1974,Taylor1982,JensenChu1984PhFl,BergerField1984JFM}.
In the limit of high magnetic Reynolds numbers
it is a conserved quantity \cite{L.Woltjer061958}. This conservation is responsible
for an inverse cascade which can be the
cause for large-scale magnetic fields as we observe them in astrophysical
objects.
The small-scale component of magnetic helicity is responsible for the
quenching of dynamo action \cite{Gruzinov1994}
and has to be shed in order to obtain magnetic
fields of equipartition strength and sizes larger then the underlying turbulent
eddies \cite{BlackmanBrandenburg2003ApJ}.
Helical magnetic fields are observed on the Sun's
surface \cite{Seehafer1990,Pevtsov1995ApJ}.
Such fields are also produced in tokamak
experiments for nuclear fusion to contain the plasma \cite{Nelson1995PhPl}.
It could be shown that the helical
structures on the Sun's surface are more likely to erupt in coronal
mass ejections \cite{Canfield1999}, which could imply that the Sun sheds
magnetic helicity \cite{Zhang05}.
In \cite{Zhang06} it was shown that, for a force-free magnetic field
configuration, there exists an upper limit of the magnetic helicity below
which the system is in equilibrium.
Exceeding this limit leads to coronal mass ejections which drag magnetic
helicity from the Sun.
Magnetic helicity is connected with the linking of magnetic field lines.
For two separate
magnetic flux rings with magnetic flux $\phi_{1}$ and $\phi_{2}$
it can be shown that magnetic helicity is equal to twice the number of
mutual linking $n$ times the product of the two fluxes
\cite{MoffattKnottedness1969}:
\begin{equation} \label{eq: helicity linking}
\mbox{$H_{\rm M}$} = \int_{V} \BoldVec{A} {}\cdot\BoldVec{B} {} d V = 2n\phi_{1}\phi_{2},
\end{equation}
where $\BoldVec{B} {}$ is the magnetic flux density, expressed in terms of the
magnetic vector potential $\BoldVec{A} {}$ via $\BoldVec{B} {} = \BoldVec{\nabla} {}\times\BoldVec{A} {}$
and the integral is taken over the whole volume.
As we emphasize in this paper, however, that this formula does not
apply to the case of a single interlocked flux tube.
The presence of magnetic helicity constrains the decay of magnetic
energy \cite{L.Woltjer061958,Taylor1974} due to the
the realizability condition \cite{MoffattBook1978} which imposes a lower
bound on the spectral magnetic energy if magnetic helicity is finite; that is,
\begin{equation} \label{eq: realizability}
M(k) \ge k|H(k)|/2\mu_{0},
\end{equation}
where $M(k)$ and $H(k)$ are magnetic energy and helicity at
wave number $k$ and $\mu_{0}$ is the vacuum permeability.
These spectra are normalized such that
$\int M(k) d k = \langle \BoldVec{B} {}^{2}\rangle/2\mu_{0}$ and
$\int H(k) d k = \langle \BoldVec{A} {}\cdot\BoldVec{B} {}\rangle$,
where angular brackets denote volume averages.
Note that the energy at each scale is bound separately, which constrains
conversions from large to small scales and vice versa.
For most of our calculations we assume a periodic domain
with zero net flux.
Otherwise, in the presence of a net flux, magnetic helicity would not
be conserved \cite{SMG94,Ber97}, but it would be produced at a constant
rate by the $\alpha$ effect \cite{BM04}.
The connection with the topology of the magnetic field makes the magnetic
helicity a particularly interesting quantity for studying relaxation processes.
One could imagine that the topological structure imposes limits on how magnetic
field lines can evolve during magnetic relaxation.
To test this it has been studied whether
the field topology alone can have an effect on the decay process
or if the presence of magnetic helicity is needed \cite{fluxRings10}.
The outcome was that, even for topologically nontrivial configurations, the decay
is only effected by the magnetic helicity content.
This was, however, questioned \cite{Yeates_Topology_2010}
and a topological invariant was introduced
via field line mapping which adds another constraint even in absence of magnetic
helicity.
Further evidence for the importance of extra constraints came from
numerical simulations of braided magnetic field with zero magnetic
helicity \cite{Pontin_etal11} where, at the end of a complex cascade-like
process, the system relaxed into an approximately force-free field state
consisting of two flux tubes of oppositely signed twist.
Since the net magnetic helicity is zero, the evolution of the field
would not be governed by Taylor relaxation \cite{Taylor1974} but by
extra constraints.
A serious shortcoming of some of the earlier work is that
the nonhelical field configurations considered so far were
still too simple.
For example, in the triple ring of \cite{fluxRings10} it would
have been possible to rearrange freely one of the outer rings on top
of the other one without crossing any other field lines.
The magnetic flux of these rings would annihilate to zero, making this
configuration trivially nonhelical.
Therefore, we construct in the present paper
more complex nonhelical magnetic field configurations
and study the decay of the magnetic field in a similar fashion as in
our earlier work.
Candidates for suitable field configurations
are the IUCAA logo \footnote{IUCAA = The Inter-University Centre
for Astronomy and Astrophysics in Pune, India.}
(which is a single nonhelically interlocked flux rope that will be
referred to below as the IUCAA knot)
and the Borromean rings for which $\mbox{$H_{\rm M}$} = 0$.
The IUCAA knot is commonly named $8_{18}$ in knot theory.
Furthermore, we test if Eq.\ \eqref{eq: helicity linking} is applicable for
configurations where there are no separated flux tubes while magnetic
helicity is finite. Therefore we
investigate setups where the magnetic field has the shape of a particular knot
which we call $n$--foil knot.
\section{Model}
\subsection{Representation of $n$--foil knots}
In topology a knot or link can be described via the braid notation
\cite{ArtinBraids1947}, where the crossings are plotted sequentially,
which results in a diagram that resembles a braid.
Some convenient starting points have
to be chosen from where the lines are drawn in the direction
according to the sense of the knot (Figs.\ \ref{fig: 4_foil_braid}
and \ref{fig: 4_foil_plane_color})
\begin{figure}[t!]\begin{center}
\includegraphics[width=0.7\columnwidth]{4_foil_braid}
\end{center}\caption[]{(Color online) Braid representation of the 4--foil knot.
The letters denote the starting position and the numbers denote the
crossings.}
\label{fig: 4_foil_braid}
\end{figure}
For each crossing either a capital or small letter is assigned
depending on whether it is a positive or negative crossing.
\begin{figure}[t!]\begin{center}
\includegraphics[width=0.7\columnwidth]{4_foil_plane_color}
\end{center}\caption[]{(Color online) $xy$ projection of the 4--foil knot. The
numbers denote the crossings while the colors (line styles)
separate different parts of the curve.
The letters denote the different starting positions
for the braid representation in Fig.\ \ref{fig: 4_foil_braid}. The
arrow shows the sense of the knot.}
\label{fig: 4_foil_plane_color}
\end{figure}
For the trefoil knot the braid representation is simply AAA. For each new
foil a new starting point is needed; at the same time
the number of crossings for each line increases by one. This means
that, for the 4--foil knot, the braid representation is ABABABAB, for the
5--foil ABCABCABCABCABC, etc.
We construct an initial magnetic field configuration in the form of an $n$--foil
knot with $\mbox{$n_{\rm f}$}$ foils or leaves. First, we construct its spine or backbone as
a parametrized curve in three-dimensional space.
In analogy to \cite{PP_IdealTrefoil_01} we apply the convenient parametrization
\begin{equation}
\BoldVec{x}{}(s) = \left( \begin{array}{c}
(C+\sin{s\mbox{$n_{\rm f}$}})\sin[{s(\mbox{$n_{\rm f}$}-1)}] \\
(C+\sin{s\mbox{$n_{\rm f}$}})\cos[{s(\mbox{$n_{\rm f}$}-1)}] \\
D\cos{s\mbox{$n_{\rm f}$}} \\
\end{array} \right),
\end{equation}
where $(C-1)$ is some minimum distance from the origin,
$D$ is a stretch factor in the $z$ direction
and $s$ is the curve parameter (see Fig.\ \ref{fig: 5foil_schematic}).
\begin{figure}[t!]\begin{center}
\includegraphics[width=0.7\columnwidth]{5foil_schematic}
\end{center}\caption[]{(Color online)
Projection of the 5--foil on the $xy$ plane. The lines
show the meaning of the distance $C$,
which has to be larger than $1$ to make sense.}
\label{fig: 5foil_schematic}
\end{figure}
The strength of the magnetic field across the tube's cross section
is constant and equal to $B_0$.
In the following we shall use $B_0$ as the unit of the magnetic field.
Since we do not want the knot to touch
itself we set $C = 1.6$ and $D = 2$.
The full three-dimensional magnetic field is constructed radially around
this curve (Fig.\ \ref{fig: 4foil iso initial}), where the thickness of the
cross section is set to $0.48$.
\begin{figure}[t!]\begin{center}
\includegraphics[width=0.5\columnwidth]{4foil}
\end{center}\caption[]{(Color online)
Isosurface of the initial magnetic field energy
for the 4--foil configuration.}
\label{fig: 4foil iso initial}
\end{figure}
\subsection{The IUCAA knot}
A prominent example of a single nonhelically interlocked flux rope
is the IUCAA knot.
For the IUCAA knot we apply a very similar parametrization as for the $n$--foil
knots. We have to consider the faster variation in $z$ direction, which yields
\begin{equation}
\label{eq: param iucaa}
\BoldVec{x}{}(s) = \left( \begin{array}{c}
(C+\sin{4s})\sin{3s} \\
(C+\sin{4s})\cos{3s} \\
D\cos{(8s-\varphi)} \\
\end{array} \right),
\end{equation}
where $C$ and $D$ have the same meaning as for the $n$--foil knots and
$\varphi$ is a phase shift of the $z$ variation.
The full three-dimensional magnetic field is constructed radially around
this curve (Fig.\ \ref{fig: iucaa_iso}), where the thickness of the
cross section is set to $0.48$.
\begin{figure}[t!]\begin{center}
\includegraphics[width=0.45\columnwidth]{iucaa_iso}
\includegraphics[width=0.45\columnwidth]{iucaa_iso_side}
\end{center}\caption[]{(Color online)
Isosurface of the initial magnetic field energy
for the IUCAA knot seen from the top (left panel) and slightly from the side
(right panel).}
\label{fig: iucaa_iso}
\end{figure}
\subsection{Borromean rings}
The Borromean rings are constructed with three ellipses whose surface
normals point in the direction of the unit vectors
(Fig.\ \ref{fig: borromean iso initial}).
\begin{figure}[t!]\begin{center}
\includegraphics[width=0.5\columnwidth]{Borromean_iso_t0}
\end{center}\caption[]{(Color online)
Isosurface of the initial magnetic field energy
for the Borromean rings configuration.}
\label{fig: borromean iso initial}
\end{figure}
The major and minor axes are set to $2.5$ and $1$, respectively,
and the thickness of the cross section is set to $0.6$.
If any one of the three rings were removed, the remaining 2 rings
would no longer be interlocked.
This means that there is no mutual
linking and hence no magnetic helicity. One should, however, not consider this
configuration as topologically trivial, since the rings cannot be separated,
which is reflected in a nonvanishing third-order topological invariant
\cite{ruzmaikin:331}.
\subsection{Numerical setup}
We solve the resistive magnetohydrodynamical (MHD) equations
for an isothermal compressible gas,
where the gas pressure is given by $p=\rho c_{S}^{2}$, with the density
$\rho$ and isothermal sound speed $c_{S}$. Instead of solving for
the magnetic field $\BoldVec{B} {}$ we solve for its vector potential $\BoldVec{A} {}$
and choose the resistive gauge, since it is numerically well behaved
\cite{helicityAdvective11}.
The equations we solve are
\begin{equation}
\frac{\partial \BoldVec{A} {}}{\partial t} = \BoldVec{U} {} \times \BoldVec{B} {} + \eta\nabla^{2}\BoldVec{A} {},
\label{dAdt}
\end{equation}
\begin{equation}
\frac{{\rm D} {} \BoldVec{U} {}}{{\rm D} {} t} = -c_{\rm S}^{2} \BoldVec{\nabla} {} \ln{\rho} +
\BoldVec{J} {}\times\BoldVec{B} {}/\rho + \BoldVec{F} {}_{\rm visc},
\label{dUdt}
\end{equation}
\begin{equation}
\frac{{\rm D} {} \ln{\rho}}{{\rm D} {} t} = -\BoldVec{\nabla} {} \cdot \BoldVec{U} {},
\label{drhodt}
\end{equation}
where $\BoldVec{U} {}$ is the velocity field, $\eta$ is the magnetic diffusivity,
$\BoldVec{J} {} = \BoldVec{\nabla} {}\times\BoldVec{B} {}/\mu_{0}$ is the current density,
$\BoldVec{F} {}_{\rm visc} = \rho^{-1}\BoldVec{\nabla} {}\cdot2\nu\rho\bm{\mathsf{S}}$ is the viscous
force with the traceless rate of strain tensor $\bm{\mathsf{S}}$ with components
${\sf S}_{ij}=\frac{1}{2}(u_{i,j}+u_{j,i})-\frac{1}{3}\delta_{ij}\BoldVec{\nabla} {}\cdot\BoldVec{U} {}$,
$\nu$ is the kinematic viscosity, and
${\rm D} {}/{\rm D} {} t=\partial/\partial t + \BoldVec{U} {}\cdot\nabla$ is the advective
time derivative.
We perform simulations in a box of size $(2\pi)^{3}$ with fully periodic
boundary conditions for all quantities. To test how boundary effects play a role
we also perform simulations with perfect conductor boundary conditions (i.e.,
the component of the magnetic field perpendicular to the surface vanishes).
In both choices of boundary conditions, magnetic helicity is gauge invariant
and is a conserved quantity in ideal MHD (i.e., $\eta=0$).
As a convenient parameter we use the Lundquist number
$\mbox{\rm Lu} = U_{\rm A}L/\eta$, where $U_{\rm A}$ is the Alfv\'en velocity
and $L$ is a typical length scale of the system.
The value of the viscosity is characterized by the magnetic Prandtl
number $\mbox{\rm Pr}_{\rm M}=\nu/\eta$.
However, in all cases discussed below we use $\mbox{\rm Pr}_{\rm M}=1$.
To facilitate comparison of different setups it is convenient to
normalize time by
the resistive time $t_{\rm res} = r^{2}\pi/\eta$, where $r$ is the
radius of the cross section of the flux tube.
We solve \Eqss{dAdt}{drhodt} with the {\textsc Pencil Code} \cite{BD02PC},
which employs
sixth-order finite differences in space and a third-order time
stepping scheme.
As in our earlier work \cite{fluxRings10}, we use $256^3$ meshpoints
for all our calculations.
We recall that we use explicit viscosity and magnetic diffusivity.
Their values are dominant over numerical contributions associated
with discretization errors of the scheme \footnote{The
discretization error of the temporal scheme scheme implies a
small diffusive contribution proportional to $\nabla^4$, but even
at the Nyquist frequency this is subdominant.}.
\section{RESULTS}
\subsection{Helicity of $n$--foil knots}
We test equation \eqref{eq: helicity linking} for the $n$--foil knots in order
to see how the number of foils $\mbox{$n_{\rm f}$}$ relates to the number of mutual linking
$n$ for the separated flux tubes.
From our simulations we know the
magnetic helicity $\mbox{$H_{\rm M}$}$ and the magnetic flux $\phi$ through the tube.
Solving \eqref{eq: helicity linking} for $n$ will lead to an apparent
self-linking number which we call $\mbox{$n_{\rm app}$}$. It turns out that
$\mbox{$n_{\rm app}$}$ is much larger then $\mbox{$n_{\rm f}$}$ and increases faster
(Fig.\ \ref{fig: n_apparent}).
We note that \eqref{eq: helicity linking} does not apply to this setup
of flux tubes and propose therefore a different formula for the magnetic
helicity:
\begin{eqnarray} \label{eq: H_nf}
\mbox{$H_{\rm M}$} & = & (\mbox{$n_{\rm f}$}-2)\mbox{$n_{\rm f}$} \phi^{2}/2.
\end{eqnarray}
In \Fig{fig: n_apparent} we plot the apparent linking number together
with a fit which uses equation \eqref{eq: H_nf}.
\begin{figure}[t!]\begin{center}
\includegraphics[width=\columnwidth]{n_apparent}
\includegraphics[width=\columnwidth]{length_nfoil}
\end{center}\caption[]{(Color online) The apparent self-linking number
for $n$--foil knots with respect to $\mbox{$n_{\rm f}$}$ (upper panel).
The fit is obtained by equating
\eqref{eq: helicity linking} and \eqref{eq: H_nf}.
The length of a $n$--foil knot is plotted with respect to $\mbox{$n_{\rm f}$}$
(lower panel), which can be fit almost perfectly by a linear function.
}
\label{fig: n_apparent}
\end{figure}
Equation \eqref{eq: H_nf} can be motivated via the number of crossings.
The flux tube is projected onto the $xy$ plane such that the number of crossings
is minimal. The linking number can be determined by adding all
positive crossings and subtracting all negative crossings according
to \Fig{fig: crossings schematically}.
\begin{figure}[t!]\begin{center}
\includegraphics[width=\columnwidth]{crossings_schematic}
\end{center}\caption[]{(Color online) Schematic representation
illustrating the sign of a crossing.
Each crossing has a handedness which can be either positive or negative. The
sum of the crossings gives the number of linking and eventually the magnetic
helicity content via equation \eqref{eq: H_nf}.}
\label{fig: crossings schematically}
\end{figure}
The linking number is then simply given as \cite{SchareinPhD}
\begin{equation}
n_{\rm linking} = (n_{+}-n_{-})/2,
\end{equation}
where $n_{+}$ and $n_{-}$ correspond to positive and negative crossings,
respectively. If we set $n_{\rm linking} = n_{\rm app}$ then we
easily see the validation of \eqref{eq: H_nf}.
Each new foil creates a new ring of crossings and adds up one crossing
in each ring (see Fig. \ref{fig: 4foil_crossings}), which explains the
quadratic increase.
\begin{figure}[t!]\begin{center}
\includegraphics[width=0.7\columnwidth]{4foil_crossings}
\end{center}\caption[]{(Color online) The isosurface for the 4--foil knot field
configuration. The sign of the crossing is always negative. The
rings show the different areas where crossings occur.}
\label{fig: 4foil_crossings}
\end{figure}
\subsection{Magnetic energy decay for $n$--foil knots}
Next, we plot in \Fig{fig: magnetic energy nfoil}
the magnetic energy decay for $n$--foil knots with $\mbox{$n_{\rm f}$}=3$ up to
$\mbox{$n_{\rm f}$}=7$ for periodic boundary conditions.
It turns out that, at later times, the decay slows down as $\mbox{$n_{\rm f}$}$ increases.
The decay of the magnetic energy obeys an approximate $t^{-2/3}$ law
for $\mbox{$n_{\rm f}$}=3$ and a $t^{-1/3}$ law for $\mbox{$n_{\rm f}$}=7$.
The rather slow decay is surprising in view of earlier results that,
for turbulent magnetic fields, the magnetic energy decays like $t^{-1}$
in the absence of magnetic helicity and like $t^{-1/2}$ with magnetic
helicity \cite{CHB05}.
Whether or not the decay seen in \Fig{fig: magnetic energy nfoil} really
does follow a power law with such an exponent remains therefore open.
\begin{figure}[t!]\begin{center}
\includegraphics[width=\columnwidth]{Emag}
\end{center}\caption[]{(Color online) Time dependence of the normalized
magnetic energy for a given number of foils with
periodic boundary conditions.
The power law for the energy decay varies between $-2/3$
for $\mbox{$n_{\rm f}$}=3$ (solid/blue line) and $-1/3$ for $\mbox{$n_{\rm f}$}=7$
(solid/black).}
\label{fig: magnetic energy nfoil}
\end{figure}
\begin{figure}[t!]\begin{center}
\includegraphics[width=\columnwidth]{realizability_k2}
\end{center}\caption[]{
(Color online) Time dependence of the quotient from the realizability
condition \eqref{eq: realizability} for $k=2$.
It is clear that, for larger $\mbox{$n_{\rm f}$}$, the energy approaches its minimum faster.
}
\label{fig: realizability_k2}
\end{figure}
The different power laws for a given number of foils $\mbox{$n_{\rm f}$}$ are unexpected
because the setups differ only in their magnetic helicity and
magnetic energy content and not in the qualitative nature of the knot.
Indeed, one might have speculated that the faster $t^{-2/3}$ decay applies
to the case with larger $\mbox{$n_{\rm f}$}$, because this structure is more complex
and involves sharper gradients.
On the other hand, a larger value of $\mbox{$n_{\rm f}$}$ increases the total helicity,
making the resulting knot more strongly packed.
This can be verified by noting that the magnetic
helicity increases quadratically with $\mbox{$n_{\rm f}$}$ while the magnetic energy
increases only linearly.
This is because the energy is proportional to the length
of the tube which, in turn, is proportional to $\mbox{$n_{\rm f}$}$ (\Fig{fig: n_apparent}).
Therefore we expect that, for the higher $\mbox{$n_{\rm f}$}$ cases,
the realizability condition should play a more significant role
at early times.
This can be seen in \Fig{fig: realizability_k2},
where we plot the ratio $2M(k)/(k|H(k)|)$
for $\mbox{$n_{\rm f}$} = 3$ to $\mbox{$n_{\rm f}$} = 7$ for $k = 2$. Since the magnetic helicity
relative to the magnetic energy is higher for larger values of $\mbox{$n_{\rm f}$}$, it plays
a more significant role for high $\mbox{$n_{\rm f}$}$.
This would explain a different onset of the power law
decay, although it would not explain a change in the exponent.
Indeed the decay of $\mbox{$H_{\rm M}$}$ shows approximately the same behavior for
all $\mbox{$n_{\rm f}$}$ (\Fig{fig: mag hel decay}).
We must therefore expect that the different decay laws
are described only approximately by power laws.
\begin{figure}[t!]\begin{center}
\includegraphics[width=\columnwidth]{Hmag}
\end{center}\caption[]{(Color online) Time dependence of the normalized
magnetic helicity for a given number of foils with
periodic boundary conditions.}
\label{fig: mag hel decay}
\end{figure}
For periodic boundary conditions it is possible that the flux tube reconnects
over the domain boundaries which could lead to additional magnetic field
destruction.
To exclude such complications we compare simulations with perfectly
conducting or closed boundaries with periodic boundary conditions
(Fig.\ref{fig: magnetic energy nfoil PC}).
Since there is no difference in the two cases we can exclude the
significance of boundary effects for the magnetic energy decay.
\begin{figure}[t!]\begin{center}
\includegraphics[width=\columnwidth]{Emag_PC}
\end{center}\caption[]{(Color online) Time dependence of the normalized
magnetic energy
for the trefoil and 4--foil knot with periodic and perfect
conductor (PC) boundary conditions. There is no significant difference
in the energy decay for the different boundary conditions.}
\label{fig: magnetic energy nfoil PC}
\end{figure}
In all cases the magnetic helicity can only decay on a resistive time
scale (Fig.\ \ref{fig: mag hel decay}). This means that, during
faster dynamical processes like magnetic reconnection, magnetic helicity
is approximately conserved. To show this we plot the magnetic field lines
for the trefoil knot at different times
(Fig.\ \ref{fig: trefoil0}). Since magnetic helicity
does not change significantly, the self-linking is transformed into
a twisting of the flux tube which is topologically equivalent to linking.
Such a process has also been mentioned in connection with Fig.~1 of
Ref.~\cite{PhysRevLett.83.1155}, while the opposite process of the conversion
of twist into linkage has been seen in Ref.~\cite{Linton_etal01}.
We can also see that the reconnection process, which transforms the trefoil
knot into a twisted ring, does not aid the decay of magnetic helicity.
\begin{figure}[t!]\begin{center}
\includegraphics[width=0.8\columnwidth]{field_lines_0}
\includegraphics[width=0.8\columnwidth]{field_lines_39}
\end{center}\caption[]{(Color online) Magnetic field lines for the trefoil knot
at time $t = 0$ (upper panel) and
$t=7.76\times10^{-2}\,t_{\rm res}$ (lower panel).
Both images were taken from the same viewing position to make
comparisons easier. The Lundquist number was chosen to be $1000$.
The colors indicate the field strength.
}
\label{fig: trefoil0}
\end{figure}
\subsection{Decay of the IUCAA knot}
For the nonhelical triple-ring configuration of Ref.~\cite{fluxRings10}
it was found that
the topological structure gets destroyed after only
$10$ Alfv\'en times.
The destruction was attributed to the absence of magnetic
helicity whose conservation would pose constraints on the relaxation process.
Looking at the magnetic field lines of the IUCAA knot at different times
(\Fig{fig: iucaa_256a1_t39}), we see that the field remains structured
and that some helical features emerge above and below the $z=0$ plane.
These localized helical patches could then locally impose constraints
on the magnetic field decay.
\begin{figure}[t!]\begin{center}
\includegraphics[width=0.9\columnwidth]{iucaa_256a1_t39}
\includegraphics[width=0.9\columnwidth]{iucaa_256a1_t78}
\end{center}\caption[]{(Color online) Magnetic field lines for the IUCAA knot at
$t=0.108\,t_{\rm res}$
(upper panel) and at $t=0.216\,t_{\rm res}$ (lower panel)
for $\mbox{\rm Lu} = 1000$ and $\varphi = (4/3)\pi$.}
\label{fig: iucaa_256a1_t39}
\end{figure}
The asymmetry of the IUCAA knot in the $z$ direction leads to different
signs of magnetic helicity above and below the $z=0$ plane.
This is shown in \Figs{fig: helicity iucaa 256c}{fig: helicity iucaa 256c2}
where we plot the magnetic helicity for the upper and lower
parts for two different values of $\varphi$; see \Eq{eq: param iucaa}.
In the plot, we refer to the upper and lower parts as north and south,
respectively.
These plots show that
there is a tendency of magnetic helicity of opposite sign to emerge
above and below the $z=0$ plane.
Given that the magnetic helicity was initially zero, one
may speculate that higher order topological invariants could
provide an appropriate tool to characterize the emergence of
such a ``bi-helical'' structure from an initially nonhelical one.
\begin{figure}[t!]\begin{center}
\includegraphics[width=\columnwidth]{helicity_t_c}
\end{center}\caption[]{(Color online)
Normalized magnetic helicity in the northern (green/dashed line) and southern
(red/dotted line) domain half together with the total magnetic helicity
(blue/solid line) for the IUCAA knot with $\mbox{\rm Lu} = 2000$ and $\varphi = (4/3)\pi$.}
\label{fig: helicity iucaa 256c}
\end{figure}
\begin{figure}[t!]\begin{center}
\includegraphics[width=\columnwidth]{helicity_t_c2}
\end{center}\caption[]{(Color online)
Normalized magnetic helicity in the northern (green/dashed line) and southern
(red/dotted line) domain half together with the total magnetic helicity
(blue/solid line) for the IUCAA knot with $\mbox{\rm Lu} = 2000$ and $\varphi = (4/3+0.2)\pi$.}
\label{fig: helicity iucaa 256c2}
\end{figure}
Note that there is a net increase of magnetic helicity over the
full volume.
Furthermore, the initial magnetic helicity is not exactly zero either,
but this is probably a consequence of discretization errors
associated with the initialization.
The subsequent increase of magnetic helicity can only occur on the longer
resistive time scales, since magnetic helicity is conserved on dynamical
time scales.
Note, however, that the increase of magnetic helicity is exaggerated
because we divide by the mean magnetic energy density which is decreasing
with time.
\begin{figure}[t!]\begin{center}
\includegraphics[width=\columnwidth]{helicity_iucaa_xy_c}
\end{center}\caption[]{(Color online)
$xy$-averaged magnetic helicity density profile in
$z$ direction for the IUCAA knot with $\mbox{\rm Lu} = 2000$ and $\varphi = (4/3)\pi$.
There is an apparent asymmetry in the distribution amongst the hemispheres.}
\label{fig: helicity iucaa xyaveraged 256c2}
\end{figure}
In \Fig{fig: helicity iucaa xyaveraged 256c2} we plot the $xy$-averaged
magnetic helicity as a function of $z$ and $t$.
This shows that the asymmetry between upper and lower parts
increases with time, which we attribute
to the Lorentz forces through which the knot shrinks and compresses its interior.
This is followed by the ejection of magnetic field.
To clarify this we plot slices of the magnetic
energy density in the $xz$ plane for different
times (\Fig{fig: energy_slice_c}).
The slices are set in the center of the domain.
\begin{figure}[t!]\begin{center}
\includegraphics[width=\columnwidth]{energy_slice_xz_c_0} \\
\includegraphics[width=\columnwidth]{energy_slice_xz_c_100}
\end{center}\caption[]{
(Color online) Magnetic energy density in the $xz$ plane for $y=0$ at $t=0$
(upper panel) and $t=5.58\times10^{-2}\,t_{\rm res}$ (lower panel) for
the IUCAA knot with
$\mbox{\rm Lu} = 2000$ and $\varphi = (4/3)\pi$.
}
\label{fig: energy_slice_c}
\end{figure}
Due to the rose-like shape, our representation of the IUCAA knot is
not quite symmetric and turns out to be narrower in the lower half
(negative $z$) than in the upper half (positive $z$), which
is shown in \Fig{fig: iucaa_iso} (right panel).
When the knot contracts due to the Lorentz force, it begins to touch the
inner parts which creates motions in the positive $z$ direction which,
in turn, drag the magnetic field away from the center (\Fig{fig: energy_slice_c}).
The pushing of material can, however, be decreased when the phase $\varphi$
is changed.
For $\varphi = (4/3+0.2)\pi$ there is no such upward motion visible
and the configuration stays nearly symmetric (\Fig{fig: energy_slice_c2}).
\begin{figure}[t!]\begin{center}
\includegraphics[width=\columnwidth]{energy_slice_xz_c2_0} \\
\includegraphics[width=\columnwidth]{energy_slice_xz_c2_100}
\end{center}\caption[]{(Color online)
Magnetic energy density in the $xz$-plane for $y=0$ at $t=0$
(upper panel) and $t=5.58\times10^{-2}\,t_{\rm res}$ (lower panel)
for the IUCAA knot with
$\mbox{\rm Lu} = 2000$ and $\varphi = (4/3+0.2)\pi$. The magnetic field stays centered.}
\label{fig: energy_slice_c2}
\end{figure}
In \Fig{fig: energy decay compare} the decay behavior of the magnetic energy
is compared with previous work \cite{fluxRings10}.
We note in passing that
the power law of $t^{-1}$ is expected for nonhelical turbulence \cite{CHB05},
but it is different from the
helical ($t^{-1/2}$) and nonhelical ($t^{-3/2}$) triple-ring configurations
studied earlier.
A possible explanation is the conservation of magnetic structures for
the IUCAA knot, whereas the nonhelical triple-ring configuration loses its
structure.
\subsection{Borromean rings}
Previous calculations showed a significant difference in the decay process
of three interlocked flux rings in the helical and nonhelical case
\cite{fluxRings10}.
In \Fig{fig: energy decay compare} we compare
the magnetic energy decay found from previous calculations
using triple-ring configurations
with the IUCAA knot and the Borromean rings.
\begin{figure}[t!]\begin{center}
\includegraphics[width=\columnwidth]{energy_decay_compare}
\end{center}\caption[]{(Color online) Magnetic energy versus time for the
different initial field configurations together with power
laws which serve as a guide.
The decay speed of the IUCAA knot and Borromean rings lies
well in between the helical and nonhelical triple-ring configuration.
}
\label{fig: energy decay compare}
\end{figure}
The Borromean rings show a similar behavior as the IUCAA knot where the
magnetic energy decays like $t^{-1}$. Similarly to the IUCAA knot we
expect some structure, which is conserved during the relaxation process
and causes the relatively slow energy decay compared to other nonhelical
configurations.
We plot the magnetic field lines at times $t = 0.248\,t_{\rm res}$
and $t = 0.276\,t_{\rm res}$; see
\Figs{fig: Borromean_256a1_t70}{fig: Borromean_256a1_t78}, respectively.
At $t = 0.248\,t_{\rm res}$ there are two interlocked
flux rings in the lower left corner, while in the opposite half of
the simulation domain a clearly twisted flux ring becomes visible.
The interlocked rings reconnect at $t = 0.276\,t_{\rm res}$
and merge into one flux tube with a twist
opposite to the other flux ring. The magnetic helicity stays zero during the
reconnection, but changes locally, which then
imposes a constraint on the magnetic energy decay and could explain
the power law that we see in \Fig{fig: energy decay compare}.
This finding is similar to that of Ruzmaikin and Akhmetiev
\cite{ruzmaikin:331} who propose that, after
reconnection, the Borromean ring configuration transforms first into a trefoil
knot and three 8-form flux tubes and after subsequent reconnection into two
untwisted flux rings (so-called unknots) and six 8-form flux tubes.
We can partly reproduce this behavior, but
instead of a trefoil knot we obtain two interlocked flux rings and, instead
of the 8-form flux tubes, we obtain internal twist in the flux rings.
\begin{figure}[t!]\begin{center}
\includegraphics[width=\columnwidth]{Borromean_256a1_t70}
\end{center}\caption[]{(Color online)
Magnetic field lines at $t=0.248\,t_{\rm res}$ for the
Borromean rings configuration for $\mbox{\rm Lu} = 1000$.
In the lower-left corner the interlocked flux rings are clearly visible
which differs from the proposed trefoil knot \cite{ruzmaikin:331}.
The flux ring in the opposite corner has an internal twist which
makes it helical.
The colors denote the strength of the field,
where the scale goes from red over green to blue.}
\label{fig: Borromean_256a1_t70}
\end{figure}
\begin{figure}[t!]\begin{center}
\includegraphics[width=\columnwidth]{Borromean_256a1_t78}
\end{center}\caption[]{(Color online)
Magnetic field lines at $t=0.276\,t_{\rm res}$ for the
Borromean rings configuration for $\mbox{\rm Lu} = 1000$.
The two flux rings in the corners both have an internal twist which makes them
helical. The twist is, however, of opposite sign which means that the whole
configuration does not contain magnetic helicity.
The colors denote the strength of the field,
where the scale goes from red over green to blue.}
\label{fig: Borromean_256a1_t78}
\end{figure}
\section{Conclusions}
In this paper we have analyzed for the first time the decay of complex
helical and nonhelical magnetic flux configurations.
A particularly remarkable one is the IUCAA knot for which the
linking number is zero, and nevertheless, some finite magnetic helicity
is gradually emerging from the system on a resistive time scale.
It turns out that both the IUCAA knot and the Borromean rings
develop regions of opposite magnetic
helicity above and below the midplane, so the net magnetic helicity
remains approximately zero.
In that process, any slight imbalance can then lead to the amplification
of the ratio of magnetic helicity to magnetic energy---even though the
magnetic field on the whole is decaying.
This clearly illustrates the potential of nonhelical configurations to
exhibit nontrivial behavior, and thus the need for studying the evolution
of higher order invariants that might capture such processes.
The role of resistivity in producing magnetic helicity from a nonhelical
initial state has recently been emphasized \cite{Low11}, but it remained
puzzling how a resistive decay can increase the topological complexity
of the field, as measured by the magnetic helicity.
Our results now shed some light on this.
Indeed, the initial field in our examples has topological complexity
that is not captured by the magnetic helicity as a quadratic invariant.
This is because of mutual cancellations that can gradually undo themselves
during the resistive decay process, leading thus to finite magnetic
helicity of opposite sign in spatially separated locations.
We recall in this context that the magnetic helicity over the periodic
domains considered here is gauge invariant and should thus agree with
any other definition, including the absolute helicity defined
in Ref.~\cite{Low11}.
Contrary to our own work on a nonhelical interlocked flux
configuration \cite{fluxRings10}, which was reducible to a
single flux ring after mutual annihilation of two rings,
the configurations studied here are non-reducible even
when mutual annihilation is taken into account.
For the helical $n$--foil knot, we have shown that the magnetic
helicity increases quadratically with $n$.
Furthermore, their decay exhibits different power laws
of magnetic energy which lie between $t^{-2/3}$ for the
3--foil knot and $t^{-1/3}$ for the 7--foil knot.
The latter case corresponds well with the
previously discussed case of three interlocked flux rings that
are interlocked in a helical fashion.
The appearance of different power
laws seems surprising since we first expected a uniform power law
in all helical cases in the regime where the magnetic helicity is so large that
the realizability condition plays a role.
This makes us speculate whether there are other quantities that are different
for the various knots and constrain magnetic energy decay.
Such quantities would be higher order topological invariants
\cite{ruzmaikin:331}, which are so far only defined for spatially separated
flux tubes. In order to investigate their role they need to be generalized
such that they can be computed for any magnetic field configuration,
similar to the integral for the magnetic helicity.
The power law of $t^{-1}$ in the decay of the magnetic energy for the IUCAA
knot and the Borromean rings is different from the $t^{-3/2}$ behavior
found earlier for the nonhelical triple-ring configuration.
The observed decay rate can be attributed
to the creation of local helical structures that
constrain the decay of the local magnetic field.
But we cannot exclude higher order invariants \cite{ruzmaikin:331} whose
conservation would then constrain the energy decay.
The Borromean rings showed clearly that local helical structures can
be generated without forcing the system. These can then impose constraints
on the field decay. We suggest that spatial variations should be taken into
account to reformulate the realizability condition \eqref{eq: realizability},
which would increase the lower bound for the magnetic energy.
For astrophysical systems local magnetic helicity variations have to
be considered to give a more precise description of both relaxation and
reconnection processes.
Both the IUCAA logo and the Borromean rings do not stay stable during
the simulation time and split up into two separated helical
magnetic structures.
On the other hand we see that the helical $n$--foil knots stay stable.
A similar behavior was seen in \cite{Braithwaite2010MNRAS}, where
magnetic fields in bubbles inside galaxy clusters were simulated.
In the case of a helical initial magnetic field the field decays into a
confined structure, while for sufficiently low initial magnetic helicity,
separated structures of opposite magnetic helicity seem more preferable.
\acknowledgements
We thank the anonymous referee for making useful suggestions
and the
Swedish National Allocations Committee for providing computing resources at the
National Supercomputer Centre in Link\"oping and the Center for
Parallel Computers at the Royal Institute of Technology in Sweden.
This work was supported in part by the Swedish Research Council,
Grant 621-2007-4064, the European Research Council under the
AstroDyn Research Project 227952,
and the National Science Foundation under Grant No.\ NSF PHY05-51164.
\defAnn. Rev. Astron. Astrophys.{Ann. Rev. Astron. Astrophys.}
\defAstrophys. J. Lett.{Astrophys. J. Lett.}
\defSol. Phys.{Sol. Phys.}
\defAstrophys. J.{Astrophys. J.}
\defJ. Geophys. Res.{J. Geophys. Res.}
\bibliographystyle{ieeetr}
|
\section{Introduction}
\subsection{Problem statement}
Let $\R^{n\times p}$ denote the set of all $n\times p$ real matrices. For any $M\in\mathbb R^{n\times p}$, we denote by $M^{t}$ its transpose and by $\|\cdot\|$ its operator norm:
\bean
\|M\|&:=&\max_{x\in\R^p, \|x\|_2=1} \|Mx\|_2, \quad \|x\|_2^2=x^t x.
\eean
Let $X\in\R^{n\times p}$ and $T$ be a random index subset of size $s$ of $\{1,\ldots,p\}$ drawn from the
uniform distribution. Let $X_T$ denote the submatrix obtained by extracting the columns $X_j$'s of $X$ indexed by $j\in T$. We say that
$X_T$ is an $r_0$-quasi-isometry if $\|X_T^tX_T-\Id\|\le r_0$ (quasi-isometry property).
The goal of this paper is to propose a new upper bound for the probability that
the submatrix $X_T$ fails to be an $r_0$-quasi-isometry. In the sequel, we assume that the
columns of $X$ have unit norm.
Proving that the quasi-isometry property holds with high probability has applications in Compressed Sensing and high-dimensional statistics based on sparsity. The uniform version of the quasi-isometry property,
i.e., satisfied for all possible $T$'s, is called the Restricted Isometry Property (RIP) and has been widely studied for independent, identically distributed (i.i.d.) sub-Gaussian matrices \cite{Mendelson:ConstApp08}. Recent works
such as \cite{CandesPlan:AnnStat09} proved that the quasi-isometry property holds with high probability
for matrices with sufficiently small coherence $\mu(X):=\max_{j\neq j'} |X_j^tX_{j'}|$. Unlike checking the RIP, computing $\mu(X)$ can be achieved in polynomial time. Such types of result are therefore of great potential
interest for a wide class of problems involving high-dimensional linear or nonlinear regression models.
Let $\{\delta_j \}$ denote a sequence of i.i.d. Bernoulli 0--1 random variables with expectation $\delta$. Let $R$ denote the square diagonal "selector matrix" whose $j^{th}$ diagonal entry is $\delta_j$.
Following the landmark papers of Bourgain and Tzafriri \cite{Bourgain:IsrJM87} (see also \cite{ProbBSp91})
and Rudelson \cite{Rudelson:JFA99}, Tropp \cite{Tropp:CRAS08} established, in particular, a bound for $(\pE \|R(X^tX-\Id)R\|^\rho )^{1/\rho}$, $\rho\in [2,\infty)$. As in \cite{Rudelson:JACM07}, the proof heavily relies on the Non-Commutative Khintchine inequality.
Using Tropp's result, Cand\`es and Plan proved in \cite[Theorem 3.2]{CandesPlan:AnnStat09} that
$X_T$ is a $1/2$-quasi-isometry with probability greater than $1-p^{-2\log(2)}$ when $s\le p/(4\|X\|^2)$ and the coherence $\mu(X)$ is sufficiently small. The quasi-isometry property for $r_0=\frac12$ then holds with high probability under
easily-checked assumptions on $X$.
\subsection{Our contribution} The present paper aims at giving a more precise and self-contained version of
Theorem 3.2 in \cite{CandesPlan:AnnStat09}.
Our result yields explicit constants, which improve the previously known values by a large factor. The analysis relies on a tail-decoupling argument, of independent interest, and a recent version of a Non-Commutative Chernoff inequality (NCCI) \cite{Tropp:ArXiv10}.
\subsection{Additional notations}
\label{not}
For $S\subset \left\{1,\cdots,p\right\}$, we denote by $|S|$ the cardinality of $S$.
Given a vector $x\in \R^{p}$, we set $x_{T}=(x_{j})_{j\in T}\in\R^{|T|}$.
We denote by $\|M\| _{ 1\rightarrow 2}$ the
maximum $l_2$-norm of a column of $M\in\R^{n\times p}$ and $\|M\|_{\max}$ is the maximum absolute entry of $M$.
In the present paper, we consider the 'hollow Gram' matrix $H$:
\bea
H &=& X^tX-\Id \label{holo}.
\eea
In the sequel, $R^\prime$ will always denote an independent copy of the selector matrix $R$.
Let $R_s$ be a diagonal matrix whose diagonal is a random vector $\delta^{(s)}$ of length $p$, uniformly distributed on the set of all vectors with $s$ components equal to 1 and $p-s$ components equal to 0. Notice that when $\delta=s/p$, the support of the diagonal of $R$ has cardinality close to $s$ with high probability, by a standard
concentration argument.
\section{Preliminary results}
\subsection{On Rademacher chaos of order $2$ }
Let $\{\eta_i\}$ be a sequence of i.i.d. Rademacher random variables.
Theorem 3.2.2 in \cite[p.113]{Gine:Decoupling99} gives the following general result:
a Banach-valued homogeneous chaos $X$ of order $d$
\bean
X &=& \sum_{1\le i_1<\cdots<i_d\le p}X_{ i_1\cdots i_d}\eta_{i_1}\cdots \eta_{i_d}
\eean
verifies
$\left(\pE \|X\|^q \right)^{\frac{1}{q}} \le \left(\frac{q-1}{p-1}\right)^{d/2} \left(\pE \|X\|^p \right)^{\frac{1}{p}}$, $1<p<q<\infty$.
We give an elementary proof in the real case with $d=2$ and $q=2p=4$, which yields a better constant.
\begin{lemm} \label{chaos}
Let $x_{ij}\in\R$, $1\le i,j\le p$. The homogeneous Rademacher chaos of order $2$:
$\xi = \sum_{i<j} x_{ij} \eta_i \eta_j$
verifies
\bea
\pE \ \xi^4 & \le & 9\ \left(\pE \ \xi^2\right)^2.
\eea
\end{lemm}
\proof
The multinomial formula applied to $\xi$ raised to the positive power $q$, gives
\bea \label{multi}
\xi^q & = & \sum_{} \frac{q!}{\prod \alpha_{ij}!} \prod x_{ij}^{\alpha_{ij}} (\eta_i\eta_j)^{\alpha_{ij}},
\eea
where the sum is over all integers $\alpha_{ij}$'s, $i< j$, such that $\sum \alpha_{ij}=q$, and the products are over all the indices $(i,j)$, $i<j$, ordered via the lexicographical order, still denoted by '$<$'. As from now, let these conventions hold.
Case $q=2$ --- The partitions of $2$ are $2+0's$ and $1+1+0's$. Consider the partition $1+1+0's$, say $\alpha_{kl}=\alpha_{k'l'}=1$ for some $4$-uple $(k,l,k',l')$ with $k\le k'$. We have $(k,l)\neq(k',l')$, $k<l$ and $k'<l'$. Thus,
\bean
\pE [\eta_{k}\eta_{l}\eta_{k'}\eta_{l'}]= \left\{
\begin{array}{l}
\pE [\eta_{k}] \ \pE[\eta_{l}\eta_{k'}\eta_{l'}] \ (=0) \ \text{ if } k<k' \\
\\
\pE[\eta^2_{k}] \ \pE [\eta_{l}] \ \pE[\eta_{l'}]\ (=0) \ \text{ if } k=k'.
\end{array}
\right.
\eean
Therefore, $\pE \ \xi^2$ only depends on the partition $2+0's$, and one has
\bea \label{xi2}
\pE \ \xi^2 & = & \sum_{i<j} x_{ij}^2.
\eea
Case $q=4$ --- The partitions of $4$ are $4$, $2+2$, $3+1$, $2+1+1$ and $1+1+1+1$ (we now omit the zeros).
First, using the same arguments as in the case $q=2$, we show that the terms in $\pE\ \xi^4$ corresponding to the partitions $3+1$ and $2+1+1$ vanish.
Second, the partitions $1+1+1+1$ involve four different couples $(i,i')$, $(j,j')$, $(k,k')$ and $(l,l')$ (recall that $i<i'$, etc., and that the couples are lexicographically ordered). The only terms corresponding to the partitions $1+1+1+1$ whose expectation does not vanish are of the form
\bean
x_{i_1i_1'}x_{i_1i_2'}x_{i_2i_1'}x_{i_2i_2'}\ \eta_{i_1}^2 \eta_{i_1'}^2 \eta_{i_2}^2 \eta_{i_2'}^2=x_{i_1i_1'}x_{i_1i_2'}x_{i_2i_1'}x_{i_2i_2'},
\eean
i.e., the four couples $(i_1,i_1')<(i_1,i_2')<(i_2,i_1')<(i_2,i_2')$ are the vertices of a rectangle into the upper off diagonal part of the matrix $(x_{ij})$. We denote by $\cal R$ the set of all these rectangles whose vertices are lexicographically ordered.
\begin{figure}[!h]
\begin{center} \label{fig}
\begin{tikzpicture}[scale=.6]
\draw (1.5,7) -- (8.5,0);
\draw[fill=gray!10] (8,5.5) rectangle (5,4.5);
\draw[dashed] (8,5.5) -- (5,4.5);
\draw[dashed] (8,4.5) -- (5,5.5);
\draw[dashed] (8,5.5) -- (8,7.3);
\draw[dashed] (5,5.5) -- (5,7.3);
\draw[dashed] (5,4.5) -- (.4,4.5);
\draw[dashed] (5,5.5) -- (.4,5.5);
\draw (8,5.5) node {$\bullet$};
\draw (5,4.5) node {$\bullet$};
\draw (8,4.5) node {$\bullet$};
\draw (5,5.5) node {$\bullet$};
\draw[very thick] (1,0) .. controls (.5,0) and (.5,7) .. (1,7);
\draw[very thick] (9,0) .. controls (9.5,0) and (9.5,7) .. (9,7);
\draw (.2,5.5) node {$i_1$};
\draw (.2,4.5) node {$i_2$};
\draw (8,7.5) node {$i'_2$};
\draw (5,7.5) node {$i'_1$};
\end{tikzpicture}
\end{center}
\caption{The matrix $(x_{ij})$ where a 'rectangle' of $\cal R$ is drawn.}
\end{figure}
Finally, the $\alpha_{ij}$'s corresponding to the partitions $4$ and $2+2$ are even: $\alpha_{ij}=2\beta_{ij}$, with $\sum \beta_{ij}=2$. Therefore
\bean
\pE\ \xi^4 & = & \sum_{} \frac{4!}{\prod (2\beta_{ij})!} \prod x_{ij}^{2\beta_{ij}}+ \sum_{\cal R} 4!\ x_{i_1i_1'}x_{i_1i_2'}x_{i_2i_2'}x_{i_2i_1'} \ := A+B.
\eean
But
\bean
A\le 3\sum_{} \frac{2!}{\prod \beta_{ij}!} \prod (x_{ij}^{2})^{\beta_{ij}} = 3\left(\sum_{i<i'} x_{ii'}^2 \right)^2,
\eean
and
\bean
B\le \frac{4!}{2}\sum_{\cal R} \left(x_{i_1i_1'}^2x_{i_2i_2'}^2 + x_{i_1i_2'}^2 x_{i_2i_1'}^2 \right) \le 6\sum_{\stackrel{i<i', j<j'}{(i,i')<(j,j')}} 2!\ x_{ii'}^2x_{jj'}^2 = 6\left(\sum_{i<i'} x_{ii'}^2 \right)^2.
\eean
The second inequality for $B$ stems from relaxing the constraints induced by $\cal R$ and illustrated in Fig. \ref{fig}.
Using (\ref{xi2}), we obtain the desired result.
\qed
\begin{rema}
The ratio $\pE \ \xi^4 /(\pE \ \xi^2)^2$ will be used in the proof of Prop. \ref{condctrl}. We gain a factor $9$ compared to the constant $(\frac{4-1}{2-1})^{\frac22\cdot 4}=81$.
\end{rema}
\smallskip
\subsection{A Non-Commutative Chernoff inequality}
We will also need a corollary of a Matrix Chernoff's inequality recently established in \cite{Tropp:ArXiv10}.
\begin{theo}{\bf (Matrix Chernoff Inequality \cite{Tropp:ArXiv10})}
\label{bern}
Let $X_1$,\ldots,$X_p$ be independent random positive semi-definite matrices taking values in $\mathbb R^{d\times d}$. Set $S_p=\sum_{j=1}^p X_j$. Assume that for all $j\in\{1,\cdots,p\}$,
$\|X_j\|\le B$ a.s. and
\bean
\left\| \pE\ S_p \right\| & \le & \mu_{\max}.
\eean
Then, for all $r\ge e\ \mu_{\max}$,
\bean
\bP\left( \left\| S_p \right\|\ge r \right) & \le & d \left(\frac{e\ \mu_{\max}}{r} \right)^{r/B}.
\eean
\end{theo}
(Set $r=(1+\delta)\mu_{\max}$ and use $e^\delta\le e^{1+\delta}$ in Theorem 1.1 \cite{Tropp:ArXiv10}.)
\section{Main results}
\subsection{Singular-value concentration theorem}
\begin{theo}\label{main1}
Let $r\in(0,1)$, $\alpha \ge 1$. Let us be given a full-rank matrix $X\in \mathbb R^{n\times p}$ and a
positive integer $s$, such that
\bea
\mu(X) & \le & \frac{r}{2(1+\alpha)\log p}\\
s & \le & \frac{r^2}{4(1+\alpha) e^2}\ \frac{p}{ \|X\|^{2} \log p }.
\eea
Let $T\subset \left\{1,\ldots,p\right\}$
be a set with cardinality $s$, chosen randomly from the uniform distribution.
Then the following bound holds:
\bea \label{sing}
\bP \left(\|X_T^tX_T-\Id_s \|\ge r \right) & \le & \frac{216}{p^{\alpha}}.
\eea
\end{theo}
\subsection{Remarks on the various constants}\
The constant $216$ stems from the following decomposition:
$2$ (poissonization) $\times 36$ (decoupling)
$\times 3$ (union bound). This constant might look large. However, in many statistical applications as in sparse models, $p$ is often assumed to be very large.
Let us now compare the constants $C_s$ and $C_\mu$ in the inequalities
\bea
\mu(X) & \le & \frac{C_\mu}{\log p} \label{form-cmu}\\
s & \le &C_s\ \frac{p}{ \|X\|^{2} \log p },
\eea
to the one of \cite{CandesPlan:AnnStat09}. The larger $C_s$ and $C_\mu$ are, the better the result is.
One of the various constraints on the rate $\alpha$ in \cite{CandesPlan:AnnStat09} is given by the theorem of Tropp in \cite{Tropp:CRAS08}. In this setting, $\alpha = 2\log 2$ and $r_{0} = 1/2$,
the author's choice of $1/2$ being unessential.
To obtain such a rate $\alpha$, they need to impose the r.h.s. of (3.15) in \cite{CandesPlan:AnnStat09} to be less than $1/4$, that is
$ 30 C_{\mu} + 13 \sqrt{2C_{s}} \le \frac{1}{4}$.
This yields $C_{s} < 1.19 \times 10^{-4}$. Choosing $C_{s}$ close to $1.19 \times 10^{-4}$, e.g.
$C_{s} \simeq 1.18\ 10^{-4}$, we obtain:
\bean
C_{s} \simeq 1.18\ 10^{-4},\quad
C_{\mu} \simeq 1.7\ 10^{-3}.
\eean
Our theorem allows to choose any rate $\alpha>0$. To make a fair comparison, let us choose $\alpha = 2\log 2$ and $r = 1/2$.
We obtain:
\bean
C_{s} \simeq 3.5\ 10^{-3}, \quad
C_{\mu} \simeq 0.1.
\eean
\section{Proof of Theorem \ref{main1}}
\label{Proofmain1}
In order to study the invertibility condition, we want to obtain bounds for the distribution tail of random sub-matrices of $H=X^tX-\Id$.
Let $R^{\prime}$ be an independent copy of $R$.
Let us recall two basic estimates:
\bean
\|H\|_{1\rightarrow 2}^2 \le \|X\|^2,\quad
\|H\|^2 \le \|X\|^4.
\eean
As a preliminary, let us notice that
\beq \label{poisse}
\bP \left(\|R_sHR_s\|\ge r \right) \ \le \ 2\ \bP \left(\|RHR\|\ge r\right),
\eeq
which can be actually proven using the same kind of 'Poissonization argument' as in Claim $(3.29)$ p. 2173 in \cite{CandesPlan:AnnStat09}.
To study the tail-distribution of $\|RHR\|$, we use a decoupling technique which consists of replacing $\|RHR\|$ with $\|RHR'\|$.
\begin{prop}
\label{condctrl}
The operator norm of $RHR$ satisfies
\bea \label{dec}
\bP\left(\|RHR\|\ge r\right) & \le & 36\ \bP\left(\|RHR^\prime\|\ge r/2\right).
\eea
\end{prop}
The main feature of this inequality is that the numerical constants are improved by a great factor when compared to the general result \cite[Theorem 1 p.224]{DeLaPena:BAMS94} (cf. Remark \ref{cste}). In addition to this decoupling argument, we need the following technical concentration result.
\begin{prop}\label{exp_bound}Let $X\in \mathbb R^{n\times p}$ be a full-rank matrix. For all parameters $s,r,u,v$ such that $\frac{p}{s} \frac{r^2}{e} \ge u^{2} \ge \frac{s}{p}\|X\|^{4}$ and
$v^2 \ge \frac{s}{p}\|X\|^{2}$, the following bound holds:
\bea\label{inv_bound}
\bP \left(\|RHR'\|\ge r\right) & \le & 3 \ p\ \mathcal V(s,[r,u,v]),
\eea
with
\bean
\mathcal V(s,[r,u,v]) & = & \left(e\frac{s}{p} \frac{u^2}{r^2} \right)^{\frac{r^2}{v^2}} + \left(e \frac{s}{p}\frac{\|X\|^4}{u^2} \right)^{u^2/\|X\|^2} + \left(e \frac{s}{p}\frac{\|X\|^2}{v^2} \right)^{v^2/\mu(X)^2}.
\eean
\end{prop}
\smallskip
We now have to analyze carefully the various quantities in Proposition \ref{exp_bound} in order to obtain for $P\left(\|RHR^\prime\|\ge r/2 \right)$ a bound of the order $e^{-\alpha \log p}$.
Set $\alpha'=\alpha+1$ and $r'=r/2$. We tune the parameters so that
\bea
\frac{u^2}{\|X\|^2} & = & \alpha' \log p \label{u} \\
\frac{v^2}{\mu(X)^2} & = & \alpha' \log p \label{v} \\
\frac{r'^2}{v^2} & \ge & \alpha' \log p, \label{tv}
\eea
and
\bea
e \frac{s}{p}\frac{\|X\|^4}{u^2} & \le & e^{-1} \label{u'} \\
e \frac{s}{p}\frac{\|X\|^2}{v^2} & \le & e^{-1} \label{v'} \\
e \frac{s}{p}\frac{u^2}{r'^2} & \le & e^{-1}. \label{ut}
\eea
A crucial quantity turns out to be $\frac{s}{p}\|X\|^{2}$. Keeping in mind that the hypothesis on the coherence reads
\bea \label{cmu}
\mu(X) & \le & \frac{C_{\mu}}{\log p},
\eea
it is necessary to impose that $s$ satisfies
\bea \label{sm}
\frac{s}{p} \|X\|^{2} & = & \frac{C_s}{\log p},
\eea
The constants $C_{\mu}$ and $C_{s}$ will be tuned according to several constraints.
The equalities (\ref{u}-\ref{v}) determine the values of $u$ and $v$. It remains to show that the previous inequalities are satisfied for a suitable choice of $C_{\mu}$ and $C_{s}$.
First, substituting (\ref{u}) into (\ref{ut}), we obtain:
\bean
\alpha' \frac{s}{p}\|X\|^2 \log p & \le e^{-2}r'^2.
\eean
Using (\ref{sm}), it follows that
\bea \label{csr}
C_s & \le & \frac{r'^2}{\alpha' e^2}.
\eea
Now, the bound (\ref{u'}) is satisfied if
\bean
\frac{e^2 C_s}{\log p} & \le & \alpha' \log p.
\eean
Based on (\ref{csr}), it suffices to have $\frac{r'^2}{\alpha'^2} \le \log^2 p$, that is
$p\ge e > e^{r'/\alpha'}$.
Second, substituting (\ref{v}) into (\ref{v'}), we obtain:
\bean
e^2 \frac{s}{p}\|X\|^2 & \le & \alpha' \mu(X)^2 \log p.
\eean
Using (\ref{cmu}) and (\ref{sm}), it follows that
\bean \label{cs}
e\sqrt{\frac{C_s}{\alpha'}} & \le & C_\mu.
\eean
Finally, (\ref{v}-\ref{tv}) yields
$r'^2 \ge \alpha'^2 \mu(X)^2 \log^2 p$.
In view of (\ref{cmu}), it thus suffices to have $r' \ge \alpha' \ C_\mu$.
To reach the desired conclusion, in order to ensure the six previous constraints, it suffices to choose $C_s$ and $C_\mu$ such that:
\bean
C_\mu \le \frac{r'}{1+\alpha} \quad {\rm and} \quad
C_s \le \min\left( \frac{r'^2}{(1+\alpha) e^2}, (1+\alpha) \frac{C_\mu^2}{e^2} \right).
\eean
This completes the proof of Theorem \ref{main1}.
\section{Proof of the tail-decoupling and the concentration result}
\subsection{Proof of Proposition \ref{condctrl}}
Let us write
\bean
RHR & = & \sum_{i\neq j} \delta_{i}\delta_{j} H_{ij}.
\eean
Let $\{\eta_i\}$ be a sequence of i.i.d. independent Rademacher random variables, mutually independent of
$\mathcal D := \{\delta_i, 1\le i\le p\}$.
Following Bourgain and Tzafriri \cite{Bourgain:IsrJM87}, and de la Pe\~{n}a and Gin\'e \cite{Gine:Decoupling99}, we construct an auxiliary random variable:
\bean
Z=Z(\eta,\delta):=\sum_{i\neq j} (1-\eta_{i}\eta_{i}) \delta_{i}\delta_{j} H_{ij}.
\eean
Setting $Y = \sum_{i\neq j} \delta_{i}\delta_{j} H_{ij} \eta_i \eta_j$, we can write
\bea
Z & =& RHR + Y. \label{expand}
\eea
For the sake of completeness, we recall basic arguments from Corollary 3.3.8 p.12 in de la Pe\~{n}a and Gin\'e \cite{Gine:Decoupling99} (applied to (\ref{expand})) to obtain a lower bound for $\bP(\|Z\| \ge \|RHR\|)$. (We henceforth work conditionally on $\mathcal D$.)
Hahn-Banach's theorem gives a linear form $x^*$ on $\cal \R^{p\times p}$ such that
\bea \label{x*}
\bP(\|Z\| \ge \|RHR\| \ |\cal D) & \ge & \bP(x^*(Z) \ge x^*(RHR) \ |\cal D) \nonumber\\
& \ge & \bP(x^*(Y) \ge 0\ |\cal D).
\eea
For any centered real random variable $\xi$, one obtains using H\"{o}lder's inequality twice (first with $\pE |\xi|=2\pE \ \xi \1_{\xi>0}$, second with $\pE\ \xi^2=\pE\ \xi^{2/3}\xi^{4/3}$):
\bea\label{xi}
\bP(\xi \ge 0) \ge \frac14 \frac{(\pE |\xi|)^2}{\pE\ \xi^2} \ge \frac14 \frac{(\pE \ \xi^2)^2}{\pE\ \xi^4}.
\eea
Noticing that $x^*(Y)$ is a centered homogeneous real chaos of order $2$, we deduce from (\ref{x*}), (\ref{xi}) and Lemma \ref{chaos},
\bea
\bP \left( \left\| Z\right\| \ge \|RHR\| \ | \mathcal D \right) &\ge & \frac{1}{4\times 9}=\frac{1}{36}.
\eea
Multiplying both sides by $\1_{\{ \|RHR\|\ge r \}}$ and taking the expectation, one has
\bea \label{324}
\frac{1}{36}\ \bP(\|RHR\|\ge r ) & \le & \bP \left( \left\| Z\right\| \ge r \right).
\eea
As from now, we can use similar arguments to \cite[Prop. 2.1]{Tropp:CRAS08}. There is a $\eta^{*}\in\{-1,1\}^{p}$ for which
\bean
\bP \left( \left\| Z\right\| \ge r \right) = \pE\ \pE \left[ \1_{ \{\| Z\| \ge r\} } |(\eta_{i})\right] & \le & \pE\ \1_{ \{\| Z(\eta^*,\delta)\| \ge r\} }=\bP(\| Z(\eta^*,\delta)\| \ge r).
\eean
Hence, setting $T=\{i, \eta^{*}_{i}=1\}$, we can write
\bean
Z(\eta^*,\delta) & = & 2 \sum_{j\in T, \ k\in T^{c}} \delta_{j}\delta_{k} H_{jk} + 2 \sum_{j\in T^c, \ k\in T} \delta_{j}\delta_{k} H_{jk}.
\eean
Since $H$ is hermitian, we have
\bean
\left\| \sum_{j\in T, \ k\in T^{c}} \delta_{j}\delta_{k} H_{jk} + \sum_{j\in T^c, \ k\in T} \delta_{j}\delta_{k} H_{jk}\right\| = \left\| \sum_{j\in T, \ k\in T^{c}} \delta_{j}\delta_{k} H_{jk} \right\|.
\eean
Now, let $(\delta_i')$ be an independent copy of $(\delta_i)$. Set $\wdt \delta_i=\delta_i$ if $i\in T$ and $\wdt \delta_i=\delta'_i$ if $i\in T^c$. Since the vectors $(\delta_i)$ and $(\wdt \delta_i)$ have the same law, we then obtain:
\bean
\bP\left(\|Z\|\ge r\right) & \le & \bP\left( 2\ \left\| \sum_{j\in T, \ k\in T^{c}} \delta_{j}\delta'_{k} H_{jk}\right\|\ge r\right).
\eean
Re-introducing the missing entries in $H$ yields
\bean
\bP \left( \left\| Z\right\| \ge r \right) & \le & \bP(\|RHR'\|\ge r/2 ),
\eean
which concludes the proof of the lemma due to (\ref{324}).
\begin{rem} \label{cste}
The previous result can be seen as a special case of Theorem 1 p.224 of the seminal paper \cite{DeLaPena:BAMS94}. Tracing the various constants involved in this theorem, we obtained the inequality
\bea \label{decbad}
\bP\left(\|RHR\|\ge r\right) & \le & 10^3\ \bP\left(\|RHR'\|\ge \frac{r}{18}\right).
\eea
\end{rem}
\subsection{Proof of Proposition \ref{exp_bound}}
We first apply the NCCI to $\|RHR'\|$ by conditioning on $R$.
\begin{lem}
The following bound holds:
\bea
P\left(\|RHR^\prime\|\ge r\right) & \le & \bP\left(\|RH\| \ge u\right) + \bP\left( \|RH\|_{1\rightarrow 2} \ge v\right) \nonumber\\
& & +\ p \left(e\frac{s}{p}\ \frac{u^2}{r^2} \right)^{\frac{r^2}{v^2}}.
\eea
\end{lem}
\begin{proof}
We have $\|RHR^\prime\|^{2}=\|RH R^{\prime 2} HR\|$. But ${R^\prime}^2=R^\prime$, so
\bea
\bP\left(\|RHR^\prime\|\ge r\right) & = & P\left(\|RH R^\prime HR\|\ge r^2\right).
\eea
We will first compute the conditional probability
\bea
\bP\left(\|RHR^\prime HR\|\ge r^2\mid R\right) & := & \pE\left[\1_{\{\|RHR^\prime HR\|\ge r^2\}}|\ R\right].
\eea
Let $Z_j$ be the $j^{th}$ column of $RH$, $j\in\{1,\cdots,p\}$. Notice that
\bean
RHR^\prime HR \ = \ \sum_{j=1}^p \delta_j^\prime Z_jZ_j^t \ := \ \sum_{j=1}^p A_j.
\eean
Since $\sum_{j=1}^p Z_jZ_j^t =RH^2R$ and $\|Z_jZ_j^t\|=\|Z_j\|_2^2$ ,
we then obtain
\bea
\|A_j\| & \le & \|RH\|_{1\rightarrow 2}^2\\
\left\| \sum_{j=1}^p \pE A_j \right\|& \le & \frac{s}{p} \|RH\|^2.
\eea
The NCCI then yields
\bea
\bP\left( \|RHR^\prime HR\|\ge r^2 \mid R \right) & \le & \ p \left(e \frac{s}{p}\frac{\|RH\|^2}{r^2} \right)^{r^2/\|RH\|_{1\rightarrow 2}^2},
\eea
provided that
\bea
e \frac{s}{p}\frac{\|RH\|^2}{r^2} & \le & 1.
\eea
Let us now introduce the events
\bean
\mathcal A = \left\{ \|RHR^\prime HR\|\ge r^2 \right\};\quad
\mathcal B = \left\{\|RH\|\ge u \right\}; \quad
\mathcal C = \left\{\|RH\|_{1\rightarrow 2} \ge v \right\}.
\eean
We have
\bean
\bP(\mathcal A) & = & \bP(\mathcal A\mid \mathcal B\cup \mathcal C)\bP(\mathcal B\cup \mathcal C)+\bP(\mathcal A\cap \mathcal B^c\cap \mathcal C^c)\\
& \le & \bP(\mathcal B)+ \bP(\mathcal C) +\bP(\mathcal A\cap \mathcal B^c\cap \mathcal C^c).
\eean
The identity
$
\bP(\mathcal A\cap \mathcal B^c \cap \mathcal C^c) = \pE \left[\1_{\mathcal A\cap \mathcal B^c \cap \mathcal C^c} \right]
=\pE \left[\bP\left(\mathcal A \mid R\right) \: \1_{\mathcal B^c \cap \mathcal C^c} \right]
$
concludes the lemma.
\hfill
\end{proof}
We now have to control the norm of $\frac{s}{p}RH^2R$, the norm of $RH$ and the column norm of $RH$.
Let us begin with $\|RH\|=\|HR\|$.
\begin{lem}
The following bounds hold:
\bean
P\left( \|HR\| >u \right) &\le & p \left(e \frac{s}{p}\frac{\|X\|^4}{u^2} \right)^{u^2/\|X\|^2}\\
\bP\left( \|RH\|_{1\rightarrow 2}\ge v\right) &\le & p \left(e \frac{s}{p}\frac{\|X\|^2}{v^2} \right)^{v^2/\mu(X)^2},
\eean
provided that $e \frac{s}{p}\frac{\|X\|^4}{u^2}$ and $e \frac{s}{p}\frac{\|X\|^2}{v^2}$ are less than $1$.
\end{lem}
\begin{proof}
The steps are of course the same as what we have just done in the proof of Lemma \ref{condctrl}. Notice that
\bean
\bP\left( \|RH\| >u \right) = \bP\left( \|HR\|^2 >u^2 \right) = \bP\left( \|HRH\| >u^2 \right).
\eean
The $j^{th}$ column of $H$ is $H_{j}=X^tX_j-e_j$. Moreover,
\bea
HRH=\sum_{j=1}^p \delta_j H_j H_j^t.
\eea
We have
$\ \|H_j H_j^t \| \ = \ \|H_j \|_2^2 \ \le\ \|H \|_{1\rightarrow 2}^2 \ \le \ \|X\|^2$,
and
\bea
\left\| \sum_{j=1}^p \pE [\delta_j H_j H_j^t] \right\|& \le & \frac{s}{p} \|H\|^2 \ \le\ \frac{s}{p} \|X\|^4.
\eea
We finally deduce from the NCCI that
\bea
\bP\left(\|HRH\| \ge u^2 \right) & \le & \ p \left(e \frac{s}{p}\frac{\|X\|^4}{u^2} \right)^{u^2/\|X\|^2}.
\eea
Let us now control the supremum $\ell_2$-norm of the columns of $RH$. Set
\bea
M & = & \sum_{k=1}^{p} \delta_{k}\ {\rm diag}(H_{k}H^{t}_{k}).
\eea
Notice that
\bean
\|RH\|^{2}_{1\rightarrow 2} & = \max_{k=1}^p \|(RH)_k\|_2^2= \left\|{\rm diag} \left((RH)^tRH \right) \right\| = & \left\|{\rm diag} \left(H^tRH \right) \right\|.
\eean
Thus,
\bean
\|RH\|^{2}_{1\rightarrow 2} & = & \left\|{\rm diag} \left(\sum_{k=1}^p \delta_k (H^t)_kH_k^t \right) \right\|.
\eean
Using symmetry of $H$ and interchanging the summation and the "diag" operation,
we obtain that $\|RH\|^{2}_{1\rightarrow 2} = \|M\|$. Moreover, we have for all $k\in \{1,\cdots,p\}$,
\bea
\| {\rm diag}(H_{k}H^{t}_{k})\| & = \max_{j=1}^p (X_jX_k)^2 \le & \mu(X)^{2},
\eea
and
\bean
\| \pE M \| = \frac{s}{p} \| {\rm diag}(HH^{t})\|^2 = \frac{s}{p} \|H\|^{2}_{1\rightarrow 2} \le \frac{s}{p} \|X\|^2.
\eean
Applying the NCCI completes the lemma.
\end{proof}
\subsection*{Ackowledgment} The authors thank the referee for valuable comments that improved the paper. They thank Max H\"ugel for pointing out a mistake in a constant involved in a previous version of the tail-decoupling inequality.
\bibliographystyle{amsplain}
|
\section{Introduction}
In this paper, we explore $\mathcal{N}=1$ supersymmetric vacua of $E_8
\times E_8$ heterotic string~\cite{Gross:1984dd,Candelas:1985en} and
$M$-theory~\cite{Horava:1995qa, Horava:1996ma, Witten:1996mz,
Lukas:1997fg, Lukas:1998yy, Lukas:1998tt}. The four-dimensional
effective theory is specified by a Calabi-Yau threefold $X$ and a
slope-stable holomorphic vector bundle, $\ensuremath{\mathscr{V}}$. The detailed
structure of the low energy theory is determined~\cite{Green:1987sp}
by the choice of a Ricci-flat metric, $g$, on the threefold and an
$\mathcal{N}=1$ supersymmetry gauge connection, $A$, on the vector
bundle. Existence proofs, such as Yau's theorem~\cite{MR480350} for
the Ricci-flat metric and the Donaldson-Uhlenbeck-Yau
theorem~\cite{duy1,duy2} for the Hermitian Yang-Mills connection,
provide us with numerous examples of such geometries. However, the
explicit metric and gauge connection are not known analytically,
except in very special cases~\cite{Candelas:1987rx, Candelas:1990rm,
Greene:1993vm, Donagi:2006yf, Braun:2006me,Anderson:2009ge}. The difficulty of
determining these quantities has presented an obstacle to the
systematic search for realistic heterotic vacua. Even for a known
vacuum, this has precluded the computation of physically relevant
parameters in the effective theory, such as the Yukawa couplings.
In recent years, the development of sophisticated numerical
approximation schemes have provided a new approach to these
problems~\cite{MR2161248,MR1916953, DonaldsonNumerical, MR1064867,
MR2154820,Douglas:2006hz,
MR2194329,Doran:2007zn,Headrick:2009jz,Douglas:2008es}. With the
development of powerful new algorithms and modern computer speed, it
is now possible to numerically approximate Ricci-flat metrics and
Hermitian Yang-Mills connections to a high degree of accuracy. We will
refer to these tools collectively as the ``generalized Donaldson
algorithm". Using them, the structure of the four-dimensional
effective theory can be explored in remarkable new ways. The goal of
this program is to determine all coefficients in the superpotential,
the explicit form of the K\"ahler potential and, ultimately, to
perform first-principle calculations of physical quantities such as
the relative quark and lepton masses.
In this paper, we make substantial progress towards this goal by
extending previous work~\cite{Anderson:2010ke} to include vector
bundles defined over manifolds with higher-dimensional K\"ahler cones;
that is, for which $h^{1,1}>1$. Importantly, our results allow one to
study arbitrary vector bundles arising in heterotic string
compactifications and to determine whether such geometries admit
$\mathcal{N}=1$ supersymmetric vacua. The problem of finding the
K\"ahler cone substructure, that is, the regions in K\"ahler moduli
space where a given holomorphic vector bundle is or is not
slope-stable, is a notoriously difficult one. In particular, the
difficulty of a direct stability analysis generally increases rapidly
with the dimension $h^{1,1}$ of the K\"ahler cone. One of the great
advantages of the algorithm presented in this paper is that, unlike a
standard analytic analysis, our numerical calculations can be
performed with essentially equal ease in arbitrary $h^{1,1}$. This
provides an important new tool in the study of supersymmetric
heterotic vacua.
The structure of the paper is as follows. To begin, in
\autoref{review_section} we provide a brief review of the numerical
algorithm for computing the Ricci-flat metric, $g$, and the Hermitian
Yang-Mills connection, $A$. Starting in \autoref{metric_review}, an
overview of Donaldson's algorithm for computing the Ricci flat
metric~\cite{MR2161248, MR1916953, DonaldsonNumerical} on a Calabi-Yau
manifold is given. In particular, the numerical implementations
developed in~\cite{Braun:2007sn, Braun:2008jp}
and~\cite{Douglas:2006rr, Douglas:2006hz} are discussed. Next, in
\autoref{connection_alg}, we outline the recent generalizations of
Donaldson's algorithm presented
in~\cite{MR2154820, Douglas:2006hz, Anderson:2010ke}. These make it
possible to compute Hermite-Einstein metrics on holomorphic vector
bundles over a Calabi-Yau manifold and, hence, to solve for the unique
gauge connection satisfying the conditions for $\mathcal{N}=1$
supersymmetry.
Both the original Donaldson algorithm and its generalizations to
connections rest on finding a particularly ``nice'' projective
embedding. In the
case of the Ricci-flat metric on a Calabi-Yau manifold, the embedding
is defined from $X$ into some higher-dimensional projective space via
the global sections of some ample line bundle, ${\cal L}^{\otimes k_g}$, on $X$. In
the case of the connection on a rank $n$ bundle $\ensuremath{\mathscr{V}}$, a map into
the Grassmannian $G(n,N_{k_{H}}-1)$ is constructed out of the
$N_{k_{H}}$ sections of $\ensuremath{\mathscr{V}} \otimes {\cal L}^{\otimes k_{H}}$, where $
{\cal L}^{\otimes k_{H}}$ is some ample line bundle on $X$. Using either
one of these embeddings\footnote{Technically, we only require a map that is an immersion rather than an embedding. However, we need not make the distinction in this paper.}, a metric can be pulled-back to the Calabi-Yau
manifold and vector bundle, respectively. One obtains a
$k_g$-dependent sequence of K\"ahler metrics $X$ and a $k_H$-dependent
sequence of Hermitian fiber metrics on $\ensuremath{\mathscr{V}}$. The degrees of
freedom of every embedding parametrize a family of pulled-back
metrics. By tuning the embedding to the so-called ``balanced embedding'' for
each degree $k_g$ and $k_H$, the K\"ahler metrics converge to the Ricci-flat
metric on $X$ and the Hermitian fiber metrics converge to a
Hermite-Einstein metric on $\ensuremath{\mathscr{V}}$. Finding the balanced embedding
is solved by Donaldson's T-operator and its generalization due to Wang
and others~\cite{MR2161248, 2008arXiv0804.4005S, MR1916953,
DonaldsonNumerical, MR1064867, MR2154820, 2007arXiv0709.1490K,
2002math......3254P}. Roughly, the T-operator acts on embeddings of
fixed degree and has the balanced embedding as a fixed point. For a
Calabi-Yau manifold, iterating the T-operator will always converge to
a balanced embedding, and, therefore, to the Ricci-flat metric on $X$
in the limit that $k_g \to \infty$. In the case of the connection, the
iteration of the T-operator for fixed $k_H$ is not guaranteed to
converge. In fact, it converges to a balanced embedding if and only if
the bundle $\ensuremath{\mathscr{V}}$ is Gieseker-stable. Furthermore, when $\ensuremath{\mathscr{V}}$
is slope-stable then the sequence of balanced embeddings define a
fiber metric converging to the Hermite-Einstein fiber metric in the
limit that $k_H \to \infty$.
In \autoref{substruc_sec}, we modify a number of the numerical tools
developed in previous work~\cite{Anderson:2010ke} to enable us to
compare the convergence of the generalized Donaldson algorithm for
different rays (or ``polarizations") in K\"ahler moduli space. These
results are illustrated in \autoref{eg_sec} by an indecomposable, rank
$2$ vector bundle defined over the $K3$ surface via the monad
construction~\cite{okonek, Anderson:2007nc,
Anderson:2008uw}. Furthermore, in \autoref{volker_int}, we present
the technical details of a newly developed, rapid numerical scheme for
integrating over a Calabi-Yau manifold.
\autoref{t_iter_error} provides a criterion to decide whether or not a
vector bundle is slope-stable for a given polarization, without the
need to explicitly compute the connection. Hence, this check can be
rapidly applied to decide slope stability. In particular, we use a
result of Wang~\cite{MR2154820} which states that the iteration of the
T-operator will reach a fixed point if and only if the defining vector
bundle is Gieseker-stable. While Gieseker stability is not sufficient
to guarantee a solution to the Hermitian Yang-Mills equations (and,
hence, a supersymmetric heterotic vacuum), it still provides valuable
information. In particular, while a slope-stable bundle is
automatically Gieseker stable, the converse does not follow. A
Gieseker-stable bundle need only be slope semi-stable. Despite these
subtleties, we extract results from the T-operator convergence which
can be used to determine the K\"ahler cone substructure. Related to
the question of semi-stability, we consider the dependence of
slope-stability on the vector bundle moduli $H^1(\ensuremath{\mathscr{V}} \otimes
\ensuremath{\mathscr{V}}^{\vee})$ in \autoref{stab_wall_sec}. In particular, the
numerical algorithm is tested on a ``stability
wall''~\cite{Anderson:2009sw,Anderson:2009nt} in K\"ahler moduli
space, the boundary between slope-stable and unstable regions. We find
that the generalized Donaldson algorithm is sensitive to the bundle
moduli dependence and, hence, our results also distinguish the
marginal cases of slope poly-stable bundles from strictly semi-stable
ones.
In \autoref{cy3_sec}, we extend our study to higher-dimensional spaces by
presenting an example of K\"ahler cone substructure of a rank $3$
monad bundle defined over a Calabi-Yau threefold constructed as a
complete intersection in a product of projective spaces. We conclude
and discuss future work in \autoref{conclusion_sec}.
\section{Hermitian Yang-Mills Connections and Fiber Metrics}
\label{review_section}
A supersymmetric $E_8 \times E_8$ heterotic string compactification is
specified by 1) a complex $d$-dimensional Calabi-Yau manifold, $X$,
and 2) a holomorphic vector bundle, $\ensuremath{\mathscr{V}}$, with structure group $K
\subset E_{8}$ defined over $X$. The gauge connection, $A$, on
$\ensuremath{\mathscr{V}}$ with associated field strength, $F$, must satisfy the
well-known Hermitian Yang-Mills (HYM)
equations~\cite{Green:1987sp}. For general $U(n)$ structure groups,
these equations are given by
\begin{equation}
\label{hym_genr}
F_{ij}=F_{\bar{i}\bar{j}}=0
,\qquad
g^{i\bar{j}}F_{i\bar{j}}
=
\mu(\ensuremath{\mathscr{V}})\cdot \mathbf{1}_{n \times n} ,
\end{equation}
where $ g^{i\bar{j}}$ is the Calabi-Yau metric, $n$ is the rank of
$\ensuremath{\mathscr{V}}$, the scalar $\mu(\ensuremath{\mathscr{V}})$ is a real number associated with
$\ensuremath{\mathscr{V}}$ and $i,j=1,\ldots d$, run over the holomorphic indices of
the Calabi-Yau $d$-fold. Our primary interest is in Calabi-Yau
threefolds, since compactification on these give rise to $\mathcal{N}=1$
supersymmetric theories in four dimensions. However, in order to
present simple illustrations of the techniques introduced in this
paper, we will discuss Calabi-Yau twofold ($K3$) as well as threefold
examples. It is not strictly necessary for the first Chern class of
the bundle to vanish~\cite{Blumenhagen:2006ux}, and the methods used
in this paper would work just as well in that setting. However, most
realistic compactifications are based on structure groups $K=SU(n)
\subset U(n)$, and these will be our main focus. When $K=SU(n)$, the
parameter $\mu(\ensuremath{\mathscr{V}})=0$ and eq.~\eqref{hym_genr} reduces to
\begin{equation}
\label{the_hym}
F_{ij}=F_{\bar{i}\bar{j}}=0
,\qquad
g^{i\bar{j}}F_{i\bar{j}}=0 .
\end{equation}
While eq.~\eqref{the_hym} are the relevant equations for realistic
heterotic compactifications, mathematically it will often be useful to
discuss the Hermitian Yang-Mills equations in full generality.
A solution to \eqref{hym_genr} is equivalent to the
bundle $\ensuremath{\mathscr{V}}$ carrying a particular Hermitian structure. An
Hermitian structure (or Hermitian fiber metric), $G$, on $\ensuremath{\mathscr{V}}$ is
an Hermitian scalar product $G_{x}$ on each fiber $\ensuremath{\mathscr{V}}(x)$ which
depends differentiably on $x$. The pair $(\ensuremath{\mathscr{V}},G)$ is often
referred to as an Hermitian vector bundle.
For a given frame, $e_{a}(x)$,
the Hermitian structure specifies an inner product as
\begin{equation}
\label{Gdef}
(e_a, e_b)=G_{{\bar a}b}
,\qquad
G=G^{\dagger}
.
\end{equation}
A choice of frame provides the necessary coordinates to express the
covariant derivative in terms of the connection,
\begin{equation}
\label{covd}
D(v^a e_a) =(dv^a)e_a + v^a A^{b}_{a}e_{b}.
\end{equation}
Imposing compatibility of the connection with the holomorphic
structure of the bundle and the fiber metric determines the
connection uniquely up to gauge transformations. Written in the most
useful gauge choice for our purposes, the connection is
\begin{equation}
\label{Amath}
\bar{A}=0
,\qquad
A=G^{-1}\partial G
.
\end{equation}
One can then rephrase the Hermitian Yang-Mills equation for
$F^{(1,1)}$ in \eqref{hym_genr} as a condition on the bundle metric,
\begin{equation}
\label{herm_met}
\mu(\ensuremath{\mathscr{V}})\cdot \mathbf{1}_{n \times n}
=
g^{i\bar{j}}F_{i\bar{j}}
=
g^{i\bar{j}}\bar{\partial}_{\bar{j}}A_i
=
g^{\bar{j}i}\bar{\partial}_{\bar{j}}(G^{-1}\partial_i G)
.
\end{equation}
A metric $G$ on the fiber of $\ensuremath{\mathscr{V}}$ satisfying this equation is
called an ``Hermite-Einstein metric''. By integration, this metric can
be used to define an inner product on the space of global sections of
$\ensuremath{\mathscr{V}}$, $s^{a}_\alpha$ where $\alpha=1,\ldots h^0(X,\ensuremath{\mathscr{V}})$,
\begin{equation}
\label{sec_prod}
\big\langle s_{\alpha} \big| s_{\beta} \big\rangle
=\int_X
s^{b}_{\beta}G_{b\bar{a}}\bar{s}^{\bar{a}}_{\bar{\alpha}}
~\dVol
.
\end{equation}
The above notions in differential geometry can be related to seemingly very different concepts in the algebraic geometry of holomorphic vector bundles. Relating the two approaches has made it possible to better understand both. For K\"ahler manifolds, the relationship
can be summarized as follows:
\begin{theorem}[Donaldson-Uhlenbeck-Yau~\cite{duy1,duy2}]
On each slope \emph{poly-stable} holomorphic vector bundle,
$\ensuremath{\mathscr{V}}$, there exists a unique connection satisfying the general
Hermitian Yang-Mills equations eq.~\eqref{hym_genr}. Moreover, such
a connection exists if and only if $\ensuremath{\mathscr{V}}$ is slope poly-stable.
\end{theorem}
Thus, in the heterotic string context, to verify that a gauge vector
bundle is consistent with supersymmetry one need only verify that it
is slope poly-stable. The notion of slope-stability of a bundle ${\cal F}$
over a K\"ahler manifold $X$ is defined by means of a real number (the
same which appeared in eq.~\eqref{hym_genr}), called the \emph{slope}:
\begin{equation}
\label{slope}
\mu ({\cal F})
\equiv
\frac{1}{\mathop{{\rm rk}}({\cal F})}\int_{X}c_{1}({\cal F})\wedge \omega^{d-1}
,
\end{equation}
where $d$ is the complex dimension of the K\"ahler manifold. Here,
$\omega$ is the K\"ahler form on $X$, while $\mathop{{\rm rk}}({\cal F})$ and
$c_1({\cal F})$ are the rank and the first Chern class of ${\cal F}$
respectively. A bundle $\ensuremath{\mathscr{V}}$ is called \emph{stable}
(\emph{semi-stable}) if, for all sub-sheaves ${\cal F}\subset \ensuremath{\mathscr{V}}$ with
$0<\mathop{{\rm rk}}({\cal F})<\mathop{{\rm rk}}(\ensuremath{\mathscr{V}})$, the slope satisfies
\begin{equation}
\label{slope_req}
\mu ({\cal F})
~
\begin{smallmatrix}
\displaystyle <
\\[1mm]
(\leq)
\end{smallmatrix}
~
\mu(\ensuremath{\mathscr{V}})
.
\end{equation}
A bundle is \emph{poly-stable} if it can be decomposed into a direct sum of
stable bundles
which all have the same slope. That is,
\begin{equation}
\label{slope_themovie}
\ensuremath{\mathscr{V}}=\bigoplus_n \ensuremath{\mathscr{V}}_n , \qquad \mu(\ensuremath{\mathscr{V}}_i)=\mu(\ensuremath{\mathscr{V}}) .
\end{equation}
From the above definitions, it is clear that the condition of
slope-stability on a Calabi-Yau manifold depends on all moduli of the
heterotic compactification. To be specific, consider a Calabi-Yau
threefold. Here, the moduli are the $h^{1,1}(X)$ K\"ahler moduli, the
$h^{2,1}(X)$ complex structure moduli, and the $h^1(\End(\ensuremath{\mathscr{V}}))$
vector bundle moduli. The dependence on K\"ahler moduli is explicit in
eqns.~\eqref{slope} and~\eqref{hym_genr}. Since slope stability is an
open property~\cite{huybrechts}, it depends only on a K\"ahler form,
$\omega$, defined up to an overall scale. We refer to this
one-parameter family of K\"ahler forms (which define a \emph{ray} in
K\"ahler moduli space) as a choice of ``polarization'' and frequently
make no distinction between a particular $\omega$ and its associated
polarization. It is possible to expand the K\"ahler form $\omega$ in
\eqref{slope} as $\omega=t^{r}\omega_{r}$, where $\omega_{r}$ are a
basis of $(1,1)$-forms and $t^r$ are the real parts of the K\"ahler
moduli. Written in terms of the triple intersection numbers $d_{rst}$
of the threefold, the slope is simply
\begin{equation}
\label{slope2the revenge}
\mu(\ensuremath{\mathscr{V}})=\frac{1}{\mathop{{\rm rk}}(\ensuremath{\mathscr{V}})}
\sum_{r,s,t=1}^{h^{1,1}(X)}
d_{rst}
c_1(\ensuremath{\mathscr{V}})^r t^s t^t
.
\end{equation}
The complex structure moduli of the Calabi-Yau manifold and
the vector bundle moduli enter through the notion of a subsheaf ${\cal F} \subset \ensuremath{\mathscr{V}}$. Thus,
finding a solution to the Hermitian Yang-Mills equations, or
determining whether the bundle is slope-stable, is a question that
must be asked after selecting a particular point in moduli space.
\section{The Generalized Donaldson Algorithm}
\label{metric_review}
Many of the challenges associated with string compactifications on a
Calabi-Yau $d$-fold $X$ arise from the difficulty in determining the
explicit geometry. The simplest $\mathcal{N}=1$ supersymmetric vacuum
solutions require a Ricci-flat metric, $g_{i\bar{j}}$, on $X$ and a
Hermite-Einstein bundle metric, $G_{\bar{a}b}$, satisfying \eqref{herm_met} as
discussed above. While Yau's theorem~\cite{MR480350} ensures that a
Ricci flat metric exists on a Calabi-Yau manifold, and the
Donaldson-Uhlenbeck-Yau theorem~\cite{duy1,duy2} provides for the
existence of a Hermite-Einstein metric on a slope-stable bundle, no
analytic solutions for either the metric or connection have yet been
found.
However, recent work has made it possible to find accurate numerical
solutions for both metrics and connections. An algorithm was initially
proposed by Donaldson for the computation of Ricci-flat
metrics~\cite{MR2161248, MR1916953, DonaldsonNumerical}, and was
implemented numerically and extended in~\cite{MR2283416, Braun:2007sn,
Braun:2008jp, Douglas:2006rr, Douglas:2006hz,
Headrick:2009jz,Anderson:2010ke}. What we refer to as the
``generalized Donaldson algorithm'' is an extension of Donaldson's
approximation scheme which numerically approaches an Hermite-Einstein
bundle metric, solving \eqref{herm_met}. This was developed
mathematically in~\cite{MR1064867,MR2154820} and implemented
numerically in~\cite{Douglas:2006hz,Anderson:2010ke}. A thorough
review of the Donaldson algorithm and its extensions is beyond the
scope of this paper. We refer the reader to~\cite{Anderson:2010ke} for
more details. However, in order to proceed with our present
investigation of K\"ahler cone substructure, we provide here a brief
review of the central ingredients of the (generalized) Donaldson's
algorithm and set the notation that will be used throughout this work.
\subsection{Donaldson's Algorithm}\label{metric_alg}
We begin with an overview of Donaldson's algorithm for approximating
the Ricci flat metric on a Calabi-Yau manifold. The first ingredient
we need is one of the simplest K\"ahler metrics, the Fubini-Study
metric on $\mathbb{P}^n$. This is given by $g_{FS
i\bar{j}}=\frac{i}{2}\partial_{i} \bar{\partial}_{\bar{j}}K_{FS}$,
where
\begin{equation}
\label{FS}
K_{FS} = \frac{1}{\pi} \ln \sum_{i\bar{j}}
h^{i\bar{j}}z_{i}\bar{z}_{\bar{j}}
\end{equation}
and $h^{i\bar{j}}$ is any Hermitian, positive, non-singular matrix.
Since it is always possible to embed $X\subset \mathbb{P}^n$ for some
large enough $n$, the Fubini-Study metric can be used to induce some
metric on any Calabi-Yau manifold $X$. Such a metric will not be
Ricci-flat, for otherwise one could easily write down an analytic
expression for the Calabi-Yau metric. It is tempting to wonder whether
there exists a generalized version of eq.~\eqref{FS} with enough free
parameters to provide a more versatile induced metric on $X$? The
central idea of Donaldson's algorithm is to find such a generalization
and a procedure for successively tuning its free parameters to
approximate the Ricci-flat metric. The obvious generalization of
eq.~\eqref{FS} is to replace the degree one polynomials with polynomials
of higher degree. That is,
\begin{equation}
\label{FS_gen}
K = \frac{1}{k\pi} ~\ln\!\! \sum_{i_{1}\ldots
i_{k}\bar{j}_{1}\ldots\bar{j}_{k}} h^{i_{1}\ldots
i_{k}\bar{j}_{1}\ldots\bar{j}_{k}}z_{i_{1}}\ldots
z_{i_{k}}\bar{z}_{\bar{j}_{1}}\ldots \bar{z}_{\bar{j}_{k}}
\end{equation}
where $h^{i_{1}\ldots i_{k}\bar{j}_{1}\ldots\bar{j}_{k}}$ is
Hermitian. This new K\"ahler potential now has $(n+1)^{2k}$ real
parameters. This generalization can, in fact, be seen in a more
systematic way by using holomorphic line bundles over $X$. The Kodaira
Embedding Theorem~\cite{MR507725} tells us that given an ample
holomorphic line bundle $\ensuremath{\mathscr{L}}$ over $X$ with $n_{1}=h^0(X,
\ensuremath{\mathscr{L}})$ global sections, one can define an embedding of $X$ into
projective space via the sections of $\mathcal{L}^{k}
=\ensuremath{\mathscr{L}}^{\otimes k}$ for some $k$. That is, choosing a basis for the
space of sections, $s_{\alpha} \in H^0(X,\ensuremath{\mathscr{L}}^k)$ where $0 \leq
\alpha \leq n_{k}-1$, allows one to define a map from $X$ to
$\mathbb{P}^{n_{k}-1}$ given by
\begin{equation}
\label{embed}
i_{k}:~ X \to \mathbb{P}^{n_{k}-1}
,\quad
(x_{0},\ldots,x_{d-1}) \mapsto \big[ s_0(x): \ldots : s_{n_{k}-1}(x) \big]
,
\end{equation}
where $x_i$ are holomorphic coordinates on the Calabi-Yau manifold. If
$\mathcal{L}$ is sufficiently ample, eq.~\eqref{embed} will define an
embedding of $ X \subset \mathbb{P}^{n_{k}-1}$ for all $\mathcal{L}^k$
with $k \geq k_0$ for some $k_0$.
In terms of this embedding via a line bundle ${\cal L}$, one can view the
generalized K\"ahler potential in eq.~\eqref{FS_gen}, restricted to
$X$, as simply
\begin{equation}
\label{Lmetric}
K_{h,k}
=\frac{1}{k\pi} \ln \sum_{\alpha,\bar{\beta}=0}^{n_{k}-1} h^{\alpha\bar{\beta}}
s_{\alpha}{\bar{s}}_{\bar{\beta}}=\ln ||s||^{2}_{h,k} .
\end{equation}
Geometrically, \eqref{Lmetric} defines an Hermitian
fiber metric on the line bundle $\ensuremath{\mathscr{L}}^{k}$ itself. It provides a natural
inner product on the space of global sections
\begin{equation}
\label{sec_metric}
M_{\alpha\bar{\beta}}=\left<s_{\beta}|s_{\alpha}\right>
=
\frac{n_{k}}{\Vol_{CY}(X)} \int_{X}
\frac{s_{\alpha}{\bar{s}}_{\bar{\beta}}}{||s||^{2}_{h}}\dVol_{CY} ,
\end{equation}
where
\begin{equation}
\label{metric_themovie}
\dVol_{CY}=\Omega \wedge {\bar{\Omega}}
\end{equation}
and $\Omega$ is the holomorphic (3,0) volume form on $X$.
With the initial K\"ahler metric eq.~\eqref{Lmetric} in hand, we must
now proceed to systematically adjust it towards Ricci flatness. To
accomplish this, the notion of a balanced metric is required. Note
that, in general, the matrices $h^{\alpha\bar{\beta}}$ and
$M_{\alpha\bar{\beta}}$ in \eqref{sec_metric} are completely
unrelated. However, for special metrics, they may coincide. The
metric $h$ on the line bundle $\ensuremath{\mathscr{L}}$ is called \emph{balanced} if
\begin{equation}
(M_{\alpha\bar{\beta}})^{-1}
=
h^{\alpha\bar{\beta}}
.
\end{equation}
Donaldson first recognized that balanced metrics lead to special
curvature properties. These can be summarized as
follows~\cite{MR2161248, MR1916953, DonaldsonNumerical,
2007arXiv0709.1490K}:
\begin{theorem}[Donaldson, Keller]
For each $k \geq 1$, the balanced metric $h$ on $\ensuremath{\mathscr{L}}^k$ exists
and is unique. As $k \to \infty$, the sequence of metrics
\begin{equation}
\label{g_approx}
g_{i\bar{j}}^{(k)}
=
\frac{1}{k\pi}\partial_{i} \bar{\partial}_{\bar{j}}
\ln \sum_{\alpha,\bar{\beta}=0}^{n_{k}-1}
h^{\alpha\bar{\beta}}s_{\alpha}\bar{s}_{\bar{\beta}}
\end{equation}
on $X$ converges to the unique Ricci-flat metric for the given
K\"ahler class and complex structure.
\end{theorem}
The central task of Donaldson's algorithm is thus to find the balanced
metric for each $k$. To this end, Donaldson defined the T-operator as
\begin{equation}
\label{t_oper}
T(h)_{\alpha\bar{\beta}}
=
\frac{n_{k}}{\Vol_{CY}(X)}
\int_{X}
\frac{s_{\alpha}{\bar{s}}_{\bar{\beta}}}{\sum_{\gamma\bar{\delta}}
h^{\gamma\bar{\delta}}s_{\gamma}{\bar{s}}_{\bar{\delta}}}\dVol_{CY}
.
\end{equation}
For a given metric $h$, it computes a matrix $T(h)$. If this matrix
equals $M_{\alpha\bar{\beta}}$, we have a balanced embedding. To find
this fixed point, simply iterate \eqref{t_oper} as follows.
\begin{theorem}
For any initial metric $h_0$ (and basis $s_{\alpha}$ of
global sections of $\ensuremath{\mathscr{L}}^k$), the sequence
\begin{equation}
h_{m+1}=\big(T(h_m)\big)^{-1}
\end{equation}
converges to the balanced metric as $m \to \infty$.
\end{theorem}
\noindent Happily, in practice, very few ($\approx 10$) iterations are
needed to approach the fixed point. Henceforth, we will also refer to
$g^{(k)}_{i\bar{j}}$ in eq.~\eqref{g_approx}, the approximating metric
for fixed $k$, as the balanced metric. It should be noted that, to
find the balanced metric at each step $k$, one must be able to
integrate over the Calabi-Yau threefold. In \autoref{volker_int}, we
will discuss the new adaptive mesh numerical integration scheme used
throughout this work.
As one final ingredient in the algorithm, one must be able to quantify
how closely the numerical metric approximates the Ricci-flat metric. A
variety of such error measures were given
in~\cite{Anderson:2010ke}. Recall that, given an sufficiently ample
line bundle $\mathcal{L}$, one can find a K\"ahler form
\begin{equation}
\label{kform}
\omega_k
=
\frac{i}{2}g^{(k)}_{i\bar{j}}dz_{i}\wedge d\bar{z}_{\bar j}
\end{equation}
corresponding to the balanced metric associated with
$\mathcal{L}^k$. Note that the K\"ahler class of this K\"ahler form is
$[\omega_k]=2\pi c_{1}(\mathcal{L}^k)$ and the associated volume is
\begin{equation}\label{voldef}
\Vol_k = \frac{1}{d!}\int_X \omega_k^d
,
\end{equation}
where $\omega_{k}^{d}$ denotes the $(d,d)$ volume form $\omega \wedge \dots \wedge \omega$.
In this paper, we will measure convergence of the Donaldson algorithm via the Ricci scalar in
\begin{equation}
||EH||_k =
\Vol_k^{(1-d)/d}
\int |R_{k}| \; \sqrt{\det g_{k}} \; \text{d}^{2d}\! x
.
\label{again2}
\end{equation}
On a Calabi-Yau manifold, $||EH||_k =O(k^{-1})$ as $k\to \infty$ and,
hence, this error measure should approach zero. As a final note, we
will henceforth denote the degree of twisting, given by the
integer $k$ in ${\cal L}^{k}$, as $k_g$ to make it clear that this integer
is associated with the computation of the metric.
A summary of Donaldson's algorithm for Ricci flat metrics is provided in \autoref{algorithm_review}. We now turn to the generalized Donaldson algorithm for computing Hermite-Einstein fiber metrics on holomorphic vector bundles.
\subsection{Hermite-Einstein Bundle Metrics}\label{connection_alg}
As we saw in the previous section, Donaldson's algorithm is a powerful
tool for numerically approximating the Ricci-flat metric on a
Calabi-Yau manifold. In this section, we investigate a generalization
of these techniques which can be used to approximate the field
strength $F^{(1,1)}$ of a holomorphic connection which satisfies
\eqref{hym_genr}. As discussed in \autoref{metric_alg}, Donaldson's
algorithm for Calabi-Yau metrics can be viewed as a method for
numerically obtaining a particular Hermitian structure on the ample
line bundle ${\cal L}^k$. This balanced fiber metric on ${\cal L}^k$ allows one
to define a balanced embedding of the Calabi-Yau space $X$ into
$\mathbb{P}^{n_{k}-1}$. By mapping the coordinates $x \in X$ into the
global sections $s_{\alpha} \in H^0(X,{\cal L}^{k})$, that is,
\begin{equation}
\xymatrix@R=5mm@C=2cm{
x \ar@{|->}[r] &
[ s_0(x) : \cdots : s_{n_k-1}(x)] ,
}
\end{equation}
we produced a map $i_{k}: X \to \mathbb{P}^{n_{k}-1}$ where
$n_k=h^0(X,{\cal L}^k)$. The pull-back of the associated Fubini-Study
metric was shown in \autoref {metric_alg} to converge to the
Ricci-flat metric on $X$ in the limit that $k \to \infty$. Viewed in
terms of Hermitian fiber metrics on line bundles, it is a natural
question to ask whether Donaldson's algorithm could be extended to
develop an analogous approximation to Hermitian metrics on higher rank
vector bundles. In particular, could one find an approximation scheme
to produce an Hermitian metric on an arbitrary stable bundle $\ensuremath{\mathscr{V}}$
of rank $n$ such that it satisfies condition \eqref{herm_met}?
Fortunately, precisely this question has been addressed in the
mathematics literature~\cite{MR2154820} and implemented for physics
in~\cite{Douglas:2006hz, Anderson:2010ke}.
To generalize Donaldson's algorithm, consider defining an embedding
via the global sections of a twist of some holomorphic vector bundle
$\ensuremath{\mathscr{V}}$ with non-Abelian structure group. That is, consider a map
\begin{equation}\label{bundle_embed}
\xymatrix@R=5mm@C=2cm{
x \ar@{|->}[r] &
{
\left[
\begin{pmatrix}
S_0^1(x) \\ \vdots \\ S_0^n(x)
\end{pmatrix}
: \cdots :
\begin{pmatrix}
S_{N_k-1}^1(x) \\ \vdots \\ S_{N_k-1}^n(x)
\end{pmatrix}
\right]
.
}}
\end{equation}
from $x \in X$ into the global sections $S^{a}_{\alpha} \in
H^0(X,\ensuremath{\mathscr{V}}\otimes{\cal L}^k)$, where $\alpha=0 \ldots N_k-1$ indexes the
$h^0(X,\ensuremath{\mathscr{V}}\otimes {\cal L}^k)$ global sections and the index $a=1,\ldots
n$ is valued in the fundamental representation of structure group $K
\subseteq U(n)$ of the rank $n$ bundle $\ensuremath{\mathscr{V}}$. We hope then to
define the embedding
\begin{equation}
\label{V_embed}
X \longrightarrow G(n,N_{k}-1) ,
\end{equation}
where $G(n,N_{k}-1)$ denotes the Grassmannian of the relevant
dimension\footnote{In this language, the Abelian case in \eqref{embed}
is simply an embedding $X \rightarrow G(1,n_{k}-1)$.}. By the
Kodaira embedding theorem~\cite{MR507725}, given a holomorphic vector
bundle, $\ensuremath{\mathscr{V}}$, and an ample line bundle, ${\cal L}$, there must exist a
finite integer $k_0$ such that, for any $k>k_0$, the twisted bundle
$\ensuremath{\mathscr{V}} (k)=\ensuremath{\mathscr{V}}\otimes {\cal L}^k$ defines an embedding, $i_k: X \to
G(n,N_k-1)$.
As in the Abelian case in the previous section, one can attempt to use
this embedding to define a Hermite-Einstein bundle metric on $\ensuremath{\mathscr{V}}
\otimes {\cal L}^k$ and, hence, an Hermitian Yang-Mills connection as in
\eqref{Amath} and \eqref{herm_met}. If ${\cal L}$ is ample then, for some
sufficiently large $k$, $\ensuremath{\mathscr{V}} \otimes {\cal L}^k$ will be generated by
its global sections. That is, it will define an embedding as in
\eqref{V_embed}. In our search for a solution to the Hermitian
Yang-Mills equation \eqref{hym_genr}, the connection on the twisted
bundle $\ensuremath{\mathscr{V}} \otimes {\cal L}^{k}$ will be closely related to the
original connection, since such a twist only modifies the trace part
of the field strength. Stated in terms of algebraic geometry, the
process of twisting will not modify the slope-stability properties of
$\ensuremath{\mathscr{V}}$ since $\ensuremath{\mathscr{V}}\otimes {\cal L}^k$ is stable if and only if
$\ensuremath{\mathscr{V}}$ is.
As at the beginning of \autoref{metric_alg}, where we chose the trial
form of the K\"ahler potential in \eqref{Lmetric}, here we begin with
another simple anzatz for the Hermitian structure $G$ in
eq.~\eqref{Gdef}. Consider the matrix
\begin{equation}
\label{G_anzatz}
(G^{-1})^{a\bar{b}}
=
\sum_{\alpha,\beta=0}^{N_k-1}H^{\alpha\bar{\beta}}S_{\alpha}^{a}
(\bar{S})_{\bar{\beta}}^{\bar{b}}
,
\end{equation}
where $H^{\alpha\bar{\beta}}$ is a Hermitian matrix of constants and
$S^{a}_{\alpha}$ are the global sections of $\ensuremath{\mathscr{V}}\otimes {\cal L}^k$. As
in \eqref{sec_prod}, this fiber metric induces an inner product on the
space of sections $H^0(X, \ensuremath{\mathscr{V}}\otimes {\cal L}^k) =
\Span\{S_{\alpha}\}$ via
\begin{equation}
\big< S_{\beta} \big| S_{\alpha} \big>=\frac{N_k}{\Vol_{CY}}
\int_{X}
S_{\alpha}^{a}(G^{a\bar{b}})^{-1}\bar{S}^{\bar{b}}_{\bar{\beta}}\dVol_{CY}
=\frac{N_k}{\Vol_{CY}}\int_{X}
S_{\alpha}^{a}
(S^{a}_{\gamma}H^{\gamma\bar{\delta}}{\bar
S}^{\bar{b}}_{\bar{\delta}})^{-1}
\bar{S}^{\bar{b}}_{\bar{\beta}}\dVol_{CY}
.
\label{coffee2}
\end{equation}
With this definition of the inner product on sections, one can give a
natural generalization of the T-operator eq.~\eqref{t_oper}. This
generalization,
\begin{equation}\label{t_gen}
T(H)_{\alpha\bar\beta} =\frac{N_k}{\Vol_{CY}}
\int_X
S_\alpha
\Big( S^\dagger H S \Big)^{-1}
\bar S_{\bar \beta}
~\dVol_{CY} ,
\end{equation}
was introduced in~\cite{MR2154820} and studied numerically
in~\cite{Douglas:2006hz,Anderson:2010ke}. Note that if $\ensuremath{\mathscr{V}}$ is a
line bundle then eq.~\eqref{t_gen} reduces to \eqref{t_oper} and one
recovers the case of a balanced embedding into
$\mathbb{P}^{N_{k}-1}$. As in the previous section, we will now
describe how the iteration of the generalized T-operator can produce a
fixed point which describes an Hermite-Einstein bundle metric.
To do this however, we must introduce one additional notion of
stability, namely that of ``Gieseker
stability''~\cite{huybrechts}. Let ${\cal L}$ be an ample line bundle and
${\cal F}$ be a torsion-free sheaf. The Hilbert polynomial of ${\cal F}$ with
respect to ${\cal L}$ is defined as
\begin{equation}\label{gieseker}
p_{{\cal L}}({\cal F})(n)=\frac{\chi({\cal F}\otimes {\cal L}^{n})}{\mathop{{\rm rk}}({\cal F})}
\end{equation}
where $\chi({\cal F}\otimes {\cal L}^{n})$ is the index of ${\cal F}\otimes
{\cal L}^{n}$. Given two polynomials $f$ and $g$, we will write $f \prec g$
if $f(n) < g(n)$ for all $n \gg 0$. Then a bundle $\ensuremath{\mathscr{V}}$ is said to
be \emph{Gieseker stable} if, for every non-zero torsion free subsheaf
${\cal F} \subset \ensuremath{\mathscr{V}}$,
\begin{equation}\label{gieseker_ineq}
p_{{\cal L}}({\cal F}) \prec
p_{{\cal L}}(\ensuremath{\mathscr{V}})
.
\end{equation}
With this definition in hand, it was shown in~\cite{MR2154820} that
\begin{theorem}[Wang]\label{wang2}
A bundle $\ensuremath{\mathscr{V}}$ is Gieseker stable if and only if the $k$-th
embedding, defined by $\ensuremath{\mathscr{V}}\otimes {\cal L}^{k}$ as in
eq.~\eqref{V_embed}, can be moved to a ``balanced'' place. That is,
if there exists an orthonormal section-wise metric on the twisted
bundle such that
\begin{equation}
\label{T_gen_bal}
\big( T(H)_{\alpha\bar\beta} \big)^{-1}
=
H^{\alpha\bar{\beta}}
\end{equation}
is a fixed point of the generalized T-operator .
\end{theorem}
We can use this special metric on $\ensuremath{\mathscr{V}} \otimes {\cal L}^k$ to define an
Hermitian metric on $\ensuremath{\mathscr{V}}$ itself. Let $G_{{\cal L}}$ denote the
balanced metric on ${\cal L}$, and $G^{(k)}$ the balanced metric on $\ensuremath{\mathscr{V}}
\otimes {\cal L}^{k}$. Then
\begin{equation}
\label{tensor_met}
G_{k}=G^{(k)}\otimes G_{{\cal L}}^{-k}
\end{equation}
is an Hermitian metric on $\ensuremath{\mathscr{V}}$. This appears in the following
important theorem~\cite{2008arXiv0804.4005S, MR2154820, MR2180559}.
\begin{theorem}[Seyyedali, Wang]
\label{wang1}
Suppose $\ensuremath{\mathscr{V}}$ is a Gieseker stable bundle of rank $n$. If $G_k
\to G_{\infty}$ as $k \to \infty$, then the metric $G_\infty$ solves
the ``weak Hermite-Einstein equation''
\begin{equation}
\label{weak_hym}
g^{i \bar j} F_{i \bar j}
=
\left(
\mu + \frac{\overline{R}-R}{2}
\right)
\mathbf{1}_{n\times n}
\end{equation}
where
\begin{itemize}
\item $R$ is the scalar curvature.
\item $\overline{R} = \int R \sqrt{\det g} \; \text{d}^{2d}\!x$ is
the averaged scalar curvature, which is zero for any K\"ahler
metric on a manifold of vanishing first Chern class.
\end{itemize}
\end{theorem}
We will, henceforth, denote the degree $k$ of the embedding defined
above as $k_{H}$, to make clear its association with the Hermitian
matrix in eq.~\eqref{G_anzatz} and distinguish it from
$k_g$. Procedurally, the process of obtaining the Hermite-Einstein
fiber metric on a slope-stable bundle $\ensuremath{\mathscr{V}}$ is very similar to
that outlined for the Ricci-flat connection in \autoref{metric_alg}:
for each value $k_{H}$ of the twisting, we iterate the T-operator
associated with the embedding defined by $H^0(X,\ensuremath{\mathscr{V}} \otimes
{\cal L}^{k_{H}})$ until a fixed point is reached. Then, by
\autoref{wang1}, the induced connection approximates solutions to
eq.~\eqref{weak_hym} as $k_{H} \to \infty$. However, there is an
immediate and important difference between this generalized algorithm
and Donaldson's algorithm for Ricci-flat metrics. While all Calabi-Yau
manifolds admit a Ricci-flat metric, not all holomorphic vector
bundles will admit an Hermite-Einstein metric satisfying
\eqref{herm_met}. That is, if one applies the algorithm to a bundle
that is not slope-stable, it will not converge to a solution of the
Hermitian Yang-Mills equations, \eqref{hym_genr}. Moreover, it should
be noted that while all slope-stable bundles are Gieseker
stable~\cite{huybrechts}, the converse does not hold: not all
Gieseker-stable bundles are slope-stable. That is, there exist cases
where the iteration of the T-operator does converge for fixed $k_{H}$,
but the sequence of metrics does not converge towards a solution of
the Hermite-Einstein fiber metric. However, if $\ensuremath{\mathscr{V}}$ is a
slope-stable holomorphic bundle, then the iteration $H_{m+1} =
T(H_m)^{-1}$ \emph{will} converge at each $k_{H}$, and in the limit
that $k_{H} \to \infty$, produce the Hermitian bundle metric
$G_{\infty}$ satisfying \eqref{weak_hym} via its associated field
strength defined in \eqref{Amath} and \eqref{herm_met}. Moreover, in
the case where the Calabi-Yau metric $g^{i \bar j}$ is Ricci-flat,
\eqref{weak_hym} simply reduces to \eqref{hym_genr}. Thus, we have
found a solution to the Hermitian Yang-Mills equations.
However, one must be careful. Despite having found a Hermite-Einstein
bundle metric (and, hence, HYM connection) associated with the twisted
bundle $\ensuremath{\mathscr{V}}\otimes {\cal L}^{k_H}$, and an Hermitian metric
$G_{\infty}$ satisfying \eqref{weak_hym}, our task is not yet
complete. We still need to explicitly determine the connection on the
bundle $\ensuremath{\mathscr{V}}$ itself satisfying the Hermitian Yang-Mills equations,
\eqref{hym_genr}. Since the process of twisting $\ensuremath{\mathscr{V}}$ by a line
bundle ${\cal L}^{k_{H}}$ in the above construction clearly modifies the
trace-part of the connection, one must subtract this line bundle
contribution to get the connection on $\ensuremath{\mathscr{V}}$ \emph{only}. To do
this, we have to separately find a suitable metric $G_{{\cal L}}$ on
$\ensuremath{\mathscr{L}}$. For example, one could compute the balanced metric
$G_{\cal L}^{(k_h)} = s^\dagger h s$ on $\ensuremath{\mathscr{L}}^{k_h}$ for some
sufficiently large $k_h$. Then $G_{{\cal L}}=(s^\dagger h s)^{1/k_{h}}$
would approximate the constant curvature Hermitian fiber metric on
$\ensuremath{\mathscr{L}}$ and, as in eq.~\eqref{tensor_met}, we find that
\begin{equation}
G
=
G^{(k_H)}\times G_{{\cal L}}^{-k_H}
=
\Big( S^\dagger H S\Big) \Big( s^\dagger h s\Big)^{-k_H/k_h}
\end{equation}
is the fiber metric \eqref{Gdef} on $\ensuremath{\mathscr{V}}$. As before, $S\in
H^0(X,\ensuremath{\mathscr{V}} \otimes {\cal L}^{k_H})$ and $s \in H^0(X, {\cal L}^{k_h})$ are
the relevant global sections. Using eqns.~\eqref{Amath} and
\eqref{herm_met}, in terms of the Hermitian metric, the connection on
$\ensuremath{\mathscr{V}}$ is then given by
\begin{equation}\label{Auntwist}
\begin{split}
A(\ensuremath{\mathscr{V}})
=&\;
\partial
\left[
\Big( S^\dagger H S\Big)
\Big( s^\dagger h s\Big)^{-k_H/k_h}
\right]
\Big( S^\dagger H S\Big)^{-1}
\Big( s^\dagger h s\Big)^{k_H/k_h}
\\=&\;
A(\ensuremath{\mathscr{V}} \otimes \ensuremath{\mathscr{L}}^{k_H})
- \frac{k_H}{k_h}
A(\ensuremath{\mathscr{L}}^{k_h}) .
\end{split}
\end{equation}
That is, one can ``untwist'' the connection simply by subtracting the
trace of the Abelian connection on
${\cal L}^{k_{H}}$ to produce the $U(n)$ connection on $\ensuremath{\mathscr{V}}$. The curvature is given by
\begin{equation}
\label{eq:Funtwisted}
F^{(0,2)} = F^{(2,0)} = 0
,\quad
g^{i \bar j} F_{i \bar j} = g^{i \bar j} \partial_{\bar j} A_i
=
g^{i \bar j} \partial_{\bar j} \partial_i \ln
\Big( S^\dagger H S\Big) \Big( s^\dagger h s\Big)^{-k_H/k_h} .
\end{equation}
As shown in~\cite{Anderson:2010ke}, when $\ensuremath{\mathscr{V}}$ is a $U(n)$ bundle,
the most efficient way to perform this untwisting is not by computing
an independent balanced metric $(G_{{\cal L}})^{k_h}$, but directly using
the induced Hermitian fiber metric on the determinant line bundle
$\wedge^{n}(\ensuremath{\mathscr{V}}\otimes\ensuremath{\mathscr{L}}^{k_H})$ of $\ensuremath{\mathscr{V}}\otimes
{\cal L}^{k_{H}}$. In particular, we choose $k_h = \rank(\ensuremath{\mathscr{V}}) k_H$. It
follows that the Hermitian metric on ${\cal L}^{k_{h}}$ is
\begin{equation}
( G_{{\cal L}})^{k_h}
=
\det(G^{(k_H)})
=
\det \big( S^\dagger H S\big) .
\end{equation}
Let $\lambda^{(k_H)}$ be the eigenvalues of $g^{i\bar
j}F^{(k_H)}_{i\bar j}$ on $\ensuremath{\mathscr{V}}\otimes \ensuremath{\mathscr{L}}^{k_{H}}$, and let
$\lambda$ be the corresponding eigenvalues of $ g^{i\bar j}F_{i\bar
j}$ on $\ensuremath{\mathscr{V}}$ after untwisting. Using eq.~\eqref{eq:Funtwisted},
we obtain that
\begin{equation}\label{lambdai}
\lambda_i
= \lambda^{(k_H)}_i -
\frac{1}{\rank \ensuremath{\mathscr{V}}} g^{i\bar j} \tr F^{(k_H)}_{i\bar j}
= \lambda^{(k_H)}_i - \frac{\sum_j
\lambda^{(k_H)}_j}{\rank \ensuremath{\mathscr{V}}} ,
\end{equation}
where $i=1,\ldots n$ where $n$ is the dimension of the fundamental
representation of the structure group of $\ensuremath{\mathscr{V}}$. Therefore, the
effect of this untwisting is precisely to subtract, at each point, the
average of the eigenvalues. Hence, in~\cite{Anderson:2010ke} we
referred to this untwisting as subtracting the trace.
The eigenvalues in \eqref{lambdai} are a pointwise measure of the
error in the numerically derived connection. For a slope-stable
bundle, $\lambda_i \to \mu$ as one increases $k_{H} \to \infty$. To
properly define an error measure for the approximation to the
Hermitian Yang-Mills connection, we must test the approximation at all
points and, hence, integrate \eqref{lambdai} over $X$. As
in~\cite{Anderson:2010ke}, we define the $L^1$ error measure
\begin{equation}
\label{eq:taudef}
\tau(A_\ensuremath{\mathscr{V}}) =
\frac{1}{2\pi}
\frac{
k_g
}{
\Vol_{k_{g}}
\rank(\ensuremath{\mathscr{V}})
}
\int_X \Big( \sum |\lambda_i| \Big)
\sqrt{g} \mathop{d}\nolimits^{2d}\!x ,
\end{equation}
where $\Vol_{k_{g}}$ is the volume computed in \eqref{voldef}. For a
slope-stable bundle, $ \tau(A_\ensuremath{\mathscr{V}}) \to \rank(\ensuremath{\mathscr{V}}) \mu$ as
$k_{H} \to \infty$. This is simply a global check of the eigenvalues
in \eqref{lambdai}. To summarize the results of this section, the
generalized Donaldson algorithm for numerically approximating a
Hermitian Yang-Mills connection is presented in
\autoref{algorithm_review}.
\begin{sidewaystable}[htbp]
\begin{center}
\renewcommand{\arraystretch}{2}
\begin{tabular}{|c|p{0.4\linewidth}|p{0.4\linewidth}|}
\hline
\textbf{Step} &
\textbf{Ricci-flat metric on \boldmath$X$} &
\textbf{Hermite-Einstein metric on \boldmath$\ensuremath{\mathscr{V}}$}
\\ \hline\hline
1
&
Choose an ample line bundle ${\cal L}$ and a degree $k_g$. &
Choose an ample line bundle ${\cal L}$, a degree $k_{H}$ and form the
twisted bundle $\ensuremath{\mathscr{V}} \otimes {\cal L}^{k_H}$.
\\ \hline
2
&
Find a basis $\{s_{\alpha}\}_{\alpha=0}^{n_{k}-1}$ for $H^0(X,\mathcal{L}^{k_g})$
at the chosen $k_g$.
&
Find a basis $\{S_{\alpha}\}_{\alpha=0}^{N_{k_{H}}-1}$ for $H^0(X,\ensuremath{\mathscr{V}} \otimes {\cal L}^{k_{H}})$
at the chosen $k_{H}$.
\\ \hline
3 &
Choose an initial positive, Hermitian matrix
$h^{\gamma\bar{\delta}}$ for the ansatz eq.~\eqref{Lmetric}. Numerically
integrate to compute the T-operator in eq.~\eqref{t_oper}.
&
Choose an initial positive, Hermitian matrix,
$H^{\gamma\bar{\delta}}$ for the ansatz \eqref{G_anzatz}. Numerically
integrate to compute the T-operator in \eqref{t_gen}.
\\ \hline
4
&
Set the new $h^{\alpha\bar{\beta}}$ to be
$h^{\alpha\bar{\beta}}=(T_{\alpha\bar{\beta}})^{-1}$.
&
Set the new $H^{\alpha\bar{\beta}}$ to be
$H^{\alpha\bar{\beta}}=(T_{\alpha\bar{\beta}})^{-1}$.
\\ \hline
5 &
Return to item 3 and repeat until
$h^{\alpha\bar{\beta}}$ approaches its fixed point ($\approx$ $10$ iterations).
&
Return to item 3 and repeat until
$H^{\alpha\bar{\beta}}$ approaches its fixed point ($\approx$ $10$ iterations).
\\ \hline
6 & &
Compute the ``untwisted'' connection and
field strength via \eqref{Auntwist} and \eqref{eq:Funtwisted}.
\\ \hline
7 &
Measure the error $||EH||_{k_{g}}$.
&
Measure the error $\tau(A_\ensuremath{\mathscr{V}})_{k_{H}}$.
\\ \hline
\end{tabular}
\caption{The Donaldson algorithm and the generalized Donaldson
algorithm. In the first column is an outline of the original
algorithm for the computation of the Ricci-flat metric on a
Calabi-Yau manifold, $X$. In the second column is an outline of
the generalized Donaldson algorithm for numerically approximating
a Hermite-Einstein bundle metric on a slope-stable bundle,
$\ensuremath{\mathscr{V}}$, over $X$.}
\label{algorithm_review}
\end{center}
\end{sidewaystable}
\section{K\"ahler Cone Substructure}\label{substruc_sec}
\subsection{Modifications For Higher Dimensional K\"ahler Cones}
\label{modific}
One of our central motivations in this work is to understand the
generalized Donaldson algorithm on manifolds with higher dimensional
K\"ahler cones, $\ensuremath{\mathcal{K}}$, that is, $\dim (\Pic(X))>1$. In particular, we
will compare the behavior of bundles under the algorithm for
\emph{different choices of polarization}. In general, holomorphic
vector bundles can display different slope-stability properties for
different choices of polarization, that is, along different rays in
the K\"ahler cone.\footnote{The radial direction along a fixed ray
only parametrizes the overall volume and does not change the
stability properties.} That is, a given bundle may be slope-stable
in some sub-cone $\ensuremath{\mathcal{K}}_\text{stable} \subset \ensuremath{\mathcal{K}}$, but be
slope-unstable (and, hence, break supersymmetry) in other sub-regions
$\ensuremath{\mathcal{K}}_\text{unstable} \subset \ensuremath{\mathcal{K}}$. This substructure is of
interest both mathematically and physically, with applications ranging
from supersymmetry breaking in heterotic $\mathcal{N}=1$ supersymmetric
vacua \cite{Sharpe:1998zu, Anderson:2009sw, Anderson:2009nt,
Anderson:2010tc, Anderson:2010mh, Anderson:2011cz} to the
computation of Donaldson-Thomas invariants on Calabi-Yau threefolds
\cite{Thomas:math9806111, 2010arXiv1002.4080L, Anderson:2010ty}. In
general, it is a difficult task to determine the global
slope-stability properties of a vector bundle, $\ensuremath{\mathscr{V}}$, throughout the
K\"ahler cone. In particular, this analysis scales badly with the
dimension of K\"ahler moduli space, $h^{1,1}(X)$. Already for
$h^{1,1}>4$ it becomes prohibitively difficult to analytically analyze
the stability of a bundle except in special cases. As a result, it is
of considerable interest to ask the question: \emph{Can the
generalized Donaldson algorithm provide an efficient probe of
K\"ahler cone substructure and vector bundle stability for higher
dimensional K\"ahler cones?} In principle, the connection algorithm
reviewed in \autoref{connection_alg} shows no difference in
computational difficulty for \emph{any dimension $h^{1,1}$}. That is,
it depends only on a one-dimensional ray (defined by the line bundle
${\cal L}$ in Step $1$ of \autoref{algorithm_review} and the embedding
\eqref{bundle_embed}) and not on the dimension of the K\"ahler cone
containing that ray. As we will see in the following sections, the
generalized Donaldson algorithm does indeed provide a powerful new
tool for analyzing K\"ahler cone substructure.
In order to pursue this goal, however, one will need a way to compare
the convergence of the algorithms (for both metric and connection) for
different rays in K\"ahler moduli space. A number of properties change
in the case that $h^{1,1}>1$ and, in particular, a few of the
definitions introduced in \autoref{review_section}, and in the
previous literature~\cite{Douglas:2006hz,Anderson:2010ke}, need some
modifications in order to make sensible comparisons for different
polarizations. One of the first of these is the way in which we
measure the complexity of the embeddings in \eqref{embed} and
\eqref{bundle_embed}. Recall that the algorithms described in
\autoref{review_section} rely on defining an embedding into some
high-dimensional Grassmanian, \eqref{bundle_embed}. For example, to
compute the Ricci-flat metric of $X$, we define the embedding $X \to
\mathbb{P}^n$ via the global sections $H^0(X,{\cal L}^{k_{g}})$. For a
manifold with $h^{1,1}=1$, it is clear that as we increase the degree,
$k_g$, of twisting, we increase the number of global sections and,
hence, as described in \autoref{metric_alg}, the accuracy of the
metric approximation. For example, in~\cite{Anderson:2010ke} we
computed Ricci-flat metrics on the Quintic hypersurface in
$\mathbb{P}^4$, where the global sections of the embedding line
bundle, ${\cal L}={\cal O}(1)$, increase with $k_g$ as $H^0\big(X,
{\cal O}(k_g)\big)=\frac{5}{6}(5k_g +{k_{g}^3)}$.
However, for manifolds with $h^{1,1}>1$ the situation becomes more
subtle if one wants to compare results for two different
polarizations, defined by line bundles ${\cal L}_1$ and ${\cal L}_2$. As an
example, consider the Calabi-Yau $3$-fold $X$ defined as a $(2,4)$
hypersurface in $\mathbb{P}^1 \times \mathbb{P}^3$. This manifold has
$h^{1,1}=2$ and its Picard group is spanned by the restriction of the
respective hyperplanes of $\mathbb{P}^1$ and $\mathbb{P}^4$ to $X$
(respectively the line bundles ${\cal O}(1,0)$ and ${\cal O}(0,1)$). Now,
consider two distinct polarizations defined by ${\cal L}_1={\cal O}(2,1)$ and
${\cal L}_2={\cal O}(1,2)$. We can define an embedding of $X$ into some
projective space using either of these ample line bundles. However,
the sections of each grow very differently in
$H^0(X,{{\cal L}_{i}}^{k_{g}})$ where $i=1,2$. These sections grow with
$k_{g}$ as
\begin{equation}
H^0\big(X,{\cal O}(2,1)^{\otimes k_{g}}\big)
=
\frac{1}{3}k_{g}(23+13{k_{g}}^2)
,\quad
H^0\big(X,{\cal O}(1,2)^{\otimes k_{g}}\big)
=
\frac{1}{3}k_{g}(28+32{k_{g}}^2)
.
\end{equation}
Hence, if we computed the metric for each of these polarizations to the same degree, say $k_{g}=10$, we would have very different results. From ${\cal L}_1$ we would have defined an embedding with $4,410$ sections, while with ${\cal L}_2$ we would have $107,600$ sections -- and, via the algorithm of \autoref{metric_alg}, a far more accurate approximation to the Ricci-flat metric.
In order to sensibly compare results for different polarizations then, instead of the degree of twisting $k_{g}$ (or $k_{H}$ in the case of the connection), we specify \emph{the number of sections} in $H^0(X,{\cal L}^{k_{g}})$. When we compare the results for different polarizations, we compare them at orders chosen so that there is an approximately equal number of sections. This procedure was first introduced for Ricci flat metrics in~\cite{Braun:2007sn, Braun:2008jp}.
There is one further modification one must make to the definitions of
\autoref{review_section}. This arises in error measures used to test
the accuracy of the approximation to an Hermitian Yang-Mills
connection. In order to determine whether or not a bundle is
slope-stable for different polarizations, we must charge the
normalization of our error measure in eq.~\eqref{eq:taudef}. For
example, take the case of a general rank $n$ stable vector bundle with
structure group $U(n)$ satisfying the general Hermitian Yang-Mills
equations \eqref{hym_genr} on a $d$-dimensional K\"ahler manifold
$X$. Since $g^{i{\bar j}}F_{i{\bar j}}=\mu(\ensuremath{\mathscr{V}})\cdot \mathbf{1}_{n
\times n}$, it is straightforward to see that in this case
\begin{equation}
\tau(A_{\ensuremath{\mathscr{V}}})=\int_X c_1(\ensuremath{\mathscr{V}} \otimes {\cal L}^{k_{H}}) \wedge \omega^{d-1}
~\in \mathbb{Z}
.
\end{equation}
As a result, $\tau$ manifestly depends not only on the first Chern
class of $\ensuremath{\mathscr{V}}$, but also on the choice of polarization
$\omega$. In fact, it jumps by an integral amount if one changes the
polarization. In order to compare the $\tau$ error measure for
different choices of polarization, we will introduce a new
normalization that will remove the polarization dependence and make
the initial values of $\tau$ more uniform for different twists
$\ensuremath{\mathscr{V}} \otimes {\cal L}^{k_{H}}$.
\begin{figure}[tb]
\centering
\input{StabilityWall-EllipticK3-Sum.tex}
\caption{A plot of the normalized error measure,
\eqref{tau_norm}. The results shown are for the sum of line
bundles, $\ensuremath{\mathscr{O}}(0,1)\oplus \ensuremath{\mathscr{O}}(0,-1)$, on the $K3$ defined
as a $(2,3)$ hypersurface in $\mathbb{P}^1\times
\mathbb{P}^3$. Using the error measure introduced in
\eqref{eq:taudef}, the results would vary for different rays in
K\"ahler moduli space. However, with the new normalization in
\eqref{tau_norm} the results are uniform.}
\label{tab:Sum}
\end{figure}
As an example, take a sum of line bundles $\bigoplus_a {\cal L}_a$ of
different first Chern class and let $\mu_{sum}=\bigoplus_a
\mu({\cal L}_a)$. For any given polarization
${\cal L}={\cal O}(t_r)={\cal O}(t_1,t_2,\ldots)$ with $r=1,\ldots, h^{1,1}$, define
$\mu_{pol}=\mu({\cal L})$. We then introduce the new normalized error
measure
\begin{equation}
\label{tau_norm}
\tau(t_r)= \frac{\mu_{pol}}{\mu_{sum}}\tau(A_{\ensuremath{\mathscr{V}}})_{k_{H}}
.
\end{equation}
This choice is made so that $\tau(t_r)$ will be independent of the
polarization and equal to $1$ for any sum of line bundles. This
normalization is shown in \autoref{tab:Sum} for the sum of line
bundles
\begin{equation}
{\cal O}(0,1)\oplus {\cal O}(0,-1)
\end{equation}
on the $K3$ surface defined via a degree $(2,3)$ hypersurface in
$\mathbb{P}^1\times \mathbb{P}^3$. Without this normalization the
error measure of the sum of line bundles would vary according to the
polarization, making it hard to compare the different directions in
the K\"ahler moduli space. However, as is illustrated by the figure,
the error measure in \eqref{tau_norm} produces uniform results. With
these new definitions in hand, we turn to our first systematic study
of K\"ahler cone substructure.
\subsection{A Bundle On An Elliptic K3 Surface}
\label{eg_sec}
In this section, we explore the K\"ahler cone substructure described
above in an explicit example. In particular, we consider the elliptic
$K3$ surface $X$ defined as a degree $(2,3)$ hypersurface in
$\mathbb{P}^1\times \mathbb{P}^2$. This representation of the $K3$ has
$\dim (\Pic(X))=2$, that is, a $2$-dimensional algebraic K\"ahler cone,
$\ensuremath{\mathcal{K}}$. Expanding the K\"ahler form on $X$ in a harmonic basis of
$(1,1)$-forms as $\omega=t_1 \omega_1 + t_2 \omega_2$, the K\"ahler
cone is the positive quadrant, $t_1,t_2 >0$ in the coefficients
$t_r$. Over this space, we will denote line bundles by ${\cal O}(t_1,t_2)$,
where ${\cal O}(1,0)$ and ${\cal O}(0,1)$ are the pull-back of the hyperplane
bundles of $\mathbb{P}^1$ and $\mathbb{P}^2$, respectively. We will
consider below the stability properties of a sample $SU(2)$ bundle
over different polarizations in the K\"ahler cone, $\ensuremath{\mathcal{K}}$.
We compute both the Ricci-flat metric on the base manifold and the
Hermitian Yang-Mills connection on a bundle for a number of different
polarizations. In the case of the metric, this amounts to computing
the fiber metric on the line bundle ${\cal L}={\cal O}(t_1,t_2)$ for different
positive choices of $t_1,t_2$. As discussed above, in order to compare
the accuracy of these approximations to a Ricci-flat metric for
different polarizations, $t_{2}/t_{1}$, one can no longer simply
specify a twisting ${\cal L}^{k_{g}}$. Instead, as we vary the coefficients
$t_r$, we will compare the number of global sections that are
generated by ${\cal L}^{k_{g}}$ for different choices of ${\cal L}$. That is,
for each choice of ${\cal L}$, we will compute the metric up to the largest
degree $k_g$ in \autoref{metric_alg} such that there are $\leq 500$
sections in $H^0(X,{\cal L}^{k_{g}})$. In the calculation, the metric
algorithm utilized $202,800$ points in the adaptive numeric
integration (as will be described further in \autoref{volker_int}) and
the metric T-operator was iterated $30$ times.
On this Calabi-Yau twofold we now define a rank $2$, holomorphic
vector bundle with structure group $SU(2)$. This sample bundle is
defined through the so-called monad construction~\cite{okonek,
Anderson:2007nc, Anderson:2008uw, Anderson:2009mh},
\begin{equation}
\label{k3_eg1}
0
\longrightarrow
{\cal O}(-2,-1)
\stackrel{f}{\longrightarrow}
{\cal O}(-2,0)^{\oplus 2}\oplus {\cal O}(2,-1)
\longrightarrow
\ensuremath{\mathscr{V}}
\longrightarrow
0 .
\end{equation}
Here $\ensuremath{\mathscr{V}}$ is defined as the cokernel of a generic map $f$ with
bi-degrees $((0,1), (0,1), (4,0))$ between the direct sums of line
bundles. Using the techniques of~\cite{huybrechts,Anderson:2008ex}, it
is straightforward to prove that $\ensuremath{\mathscr{V}}$ in \eqref{k3_eg1} is
destabilized in part of the K\"ahler cone by the rank $1$ sheaf
\begin{equation}
\label{destabilizing}
0
\longrightarrow
{\cal F}
\longrightarrow
{\cal O}(2,0)^{\oplus 2}
\stackrel{f}{\longrightarrow}
{\cal O}(2,1)
\longrightarrow
0
\end{equation}
with $c_1({\cal F})=(2,-1)$. The intersection numbers $d_{rs}$ of the two
hyperplane classes in the $(2,3)$ $K3$ surface are $d_{12}=d_{21}=3$,
$d_{22}=2$, and $d_{11}=0$. Using the definition of slope in
\eqref{slope}, it can be verified that $\ensuremath{\mathscr{V}}$ is slope-stable when
$t_1/t_2 >4/3$ and unstable when $t_1/t_2<4/3$. That is, the K\"ahler
cone exhibits the substructure
\begin{equation}\label{eg1substruc}
\begin{split}
\mathcal{K}_\text{stable}
=&\;
\big\{
\ensuremath{\mathscr{O}}(t_1, t_2)
\big|~
\tfrac{t_1}{t_2} > \tfrac{4}{3}
\big\}
, \\
\mathcal{K}_\text{unstable}
=&\;
\big\{
\ensuremath{\mathscr{O}}(t_1, t_2)
\big|~
\tfrac{t_1}{t_2} < \tfrac{4}{3}
\big\} .
\end{split}
\end{equation}
Hence, in the stable region we expect the T-operator to converge,
whereas in the unstable region we do not expect to find a fixed point.
As discussed in \autoref{connection_alg}, to apply the generalized
Donaldson algorithm one must define the embedding $i_k: X \to
G(n,N_{k_{H}})$. To do this, one must compute the global sections of
the twisted line bundle $\ensuremath{\mathscr{V}} \otimes {\cal L}^{k_H}$ for some ample
line bundle ${\cal L}$. For the bundle defined in \eqref{k3_eg1}, the
global sections $H^0(X,\ensuremath{\mathscr{V}} \otimes {\cal L}^{k_H})$ can be computed for
any choice of twisting. Multiplying eq.~\eqref{k3_eg1} by
${\cal L}^{k_H}={\cal O}(t_1,t_2)^{k_{H}}$, we obtain the short exact sequence
\begin{multline}
{\tiny 0
\to
{\cal O}(k_{H}t_1-2,k_{H}t_2-1)
\stackrel{f}{\longrightarrow}
{\cal O}(k_{H}t_1-2,k_{H}t_2)^{\oplus 2} \oplus
{\cal O}(k_{H}t_1+2,k_{H}t_2-1)} \\
{\tiny \to
\ensuremath{\mathscr{V}} \otimes {\cal O}(t_1,t_2)^{\otimes k_{H}}
\to
0}.
\label{aaa}
\end{multline}
\begin{figure}[tb]
\centering
\input{StabilityWall-EllipticK3-Substructure.tex}
\caption{The K\"ahler cone substructure associated with the $SU(2)$ bundle in \eqref{k3_eg1}. The normalized error measure, \eqref{tau_norm}, is shown for different choices of polarization in the K\"ahler cone. The presence of $\ensuremath{\mathscr{V}}$ clearly divides the K\"ahler cone, $\ensuremath{\mathcal{K}}$, into two chambers, $\ensuremath{\mathcal{K}}_\text{stable}$ and $\ensuremath{\mathcal{K}}_\text{unstable}$, corresponding to the stable/unstable regions described in \eqref{eg1substruc}.}
\label{tab:Substructure}
\end{figure}
Then the global sections are given simply as the cokernel
\begin{equation}\label{h0_def}
H^0\big(X,\ensuremath{\mathscr{V}} \otimes {\cal O}(t_1,t_2)^{k_{H}} \big)
=
\frac{
H^0\big(X,
{\cal O}(k_{H}t_1-2,k_{H}t_2)^{\oplus 2}\oplus
{\cal O}(k_{H}t_1+2,k_{H}t_2-1)
\big)
}{
f\big(H^0(X, {\cal O}(k_{H}t_1-2,k_{H}t_2-1))\big)
}
,
\end{equation}
where both parts of this quotient are the global sections of
sums of ample line bundles when $k_{H}t_1> 2$ and $k_{H}t_2>1$.
For $\ensuremath{\mathscr{V}}$ in \eqref{k3_eg1}, we will apply the generalized
Donaldson algorithm as reviewed in \autoref{connection_alg} and
\autoref{algorithm_review}. In particular, we will compute Hermitian
bundle metrics on $\ensuremath{\mathscr{V}}$ and determine whether they converge to the
Hermitian Yang-Mills connection for different choices of
polarization. Specifically, we will compare the results for six
different choices of polarization: three lying within
$\ensuremath{\mathcal{K}}_\text{stable}$, two in $\ensuremath{\mathcal{K}}_\text{unstable}$ and one on the
boundary defined by $t_1/t_2=4/3$. The results are shown in
\autoref{tab:Substructure}, where we plot the normalized $L^1$ error
measure, eq.~\eqref{tau_norm}, for the Yang-Mills connection in
various directions in the K\"ahler cone. Since $\ensuremath{\mathscr{V}}$ is a cokernel
associated with the direct sum $\bigoplus_i {\cal L}_i={\cal O}(-2,0)^{\oplus
2}\oplus {\cal O}(2,-1)$, the normalization $\mu_{sum}$ in
eq.~\eqref{tau_norm} was chosen to be
$\mu_{sum}=2\mu\big({\cal O}(-2,0))+\mu({\cal O}(2,-1)\big)$ in order to
meaningfully compare the different rays in the K\"ahler cone. In the
stable sub-cone, the eigenvalues of $g^{i{\bar j}}F_{i{\bar j}}$ can
clearly be seen to be approaching zero as one increases the number of
sections in $H^0(X,\ensuremath{\mathscr{V}}\otimes {\cal L}^{k_H})$, as expected. In the
unstable region there is no such convergence. Both observations are in
perfect agreement with eq.~\eqref{eg1substruc}. The connection on
$\ensuremath{\mathscr{V}}$ was computed with $N_G = 74,892$ points (adaptive) and the
connection T-operator was computed with $100$ iterations at each graph
point in \autoref{tab:Substructure}.
The results of this section clearly indicate that the generalized
Donaldson algorithm can be used to investigate K\"ahler cone
sub-structure. We will explore this in more detail in the following
sections, but first we will provide a description of the novel
integration method implemented and used throughout this work. This
integration scheme provides a significant increase in computation
speed and makes it possible for us to analyze a wider range of
examples.
\section{Adaptive Integration}\label{volker_int}
\subsection{Rectangle Method vs. Monte-Carlo}
All approaches to numerical geometry of Calabi-Yau threefolds, be it
Donaldson's algorithm~\cite{MR2161248, MR1916953, DonaldsonNumerical,MR2283416, Braun:2007sn, Braun:2008jp, Douglas:2006rr,
Douglas:2006hz, Headrick:2009jz,Anderson:2010ke}, its generalization to Hermitian
Yang-Mills bundles~\cite{MR2154820,Douglas:2006hz,Anderson:2010ke}, or direct
minimization~\cite{Headrick:2009jz} all use a spectral representation
of the geometric data. That is, the tensors describing the geometry
are eventually expanded in a suitable basis of functions, and the
problem reduces to finding the ``best fit'' coefficients. This is in
contrast to the traditional finite elements methods, where one
directly discretizes spacetime. As a general rule, finite elements
work well in low dimensions, but spectral representations are
necessary in high-dimensional problems.
The basic numerical step that every spectral algorithm relies on at
its core is to integrate over the base manifold. This is necessarily
so because only by evaluating the spectral basis \emph{everywhere} on
the manifold can one draw conclusions about the global behavior of the
geometric object of interest. The most straightforward integration
scheme is to split the integration domain into equal-sized pieces and
use the multidimensional generalization of the rectangle rule. In
practice, the volume elements can only be chosen of equal size with
respect to an auxiliary metric. But this is then easily corrected for
by weighting the individual points accordingly,
\begin{equation}
\sqrt{g} \; d^n x =
w(x) \; \sqrt{g_\text{aux}} \; d^n x
\quad \Rightarrow \quad
w(x)
= \frac{\sqrt{g}}{\sqrt{g_\text{aux}}}
.
\end{equation}
Here, the scalar weight function $w(x)$ is simply determined by the
auxiliary measure (given by the point distribution) and the desired
measure $\sqrt{g}\; d^n x$. The Calabi-Yau case is particularly
simple, since the Calabi-Yau volume form $\sqrt{g_\text{CY}}\; d^n x =
\Omega\wedge \overline \Omega$ is known analytically.
The disadvantage of this direct approach is that it requires extensive
knowledge about the geometry and topology of the manifold to construct
a constant (auxiliary) volume cell decomposition. This is why it has
only been used for complex surfaces~\cite{2005math.....12625D,
2008arXiv0803.0987B}, that is, real 4-dimensional manifolds. A
solution to this problem was devised in~\cite{Douglas:2006rr}, where a
Monte-Carlo integration scheme was proposed that needs as input only
the defining equation of a Calabi-Yau hypersurface (or complete
intersection). The key to this approach is that one can generate
random points with known distributions using zeroes of random sections
of line bundles~\cite{Zelditch:Shif, MR1616718}. As an example,
consider the quintic $i:Q\to\ensuremath{\mathop{\null {\mathbb{P}}}}\nolimits^4$ embedded in projective space. A
line $\ensuremath{\mathop{\null {\mathbb{P}}}}\nolimits^1\subset \ensuremath{\mathop{\null {\mathbb{P}}}}\nolimits^4$ is defined by three linear equations in the
homogeneous coordinates, that is, three sections of
$\ensuremath{\mathscr{O}}_{\ensuremath{\mathop{\null {\mathbb{P}}}}\nolimits^4}(1)$. Any one linear equation is defined by its $5$
coefficients, so one can talk about random sections with an
$SU(5)$-invariant distribution. Three linear equations intersect the
quintic hypersurface in $5$ points, and by the general theory the
probability distribution has measure $i^*(\omega_\text{FS})^3$. The
Calabi-Yau volume form is given by the residue integral
\begin{equation}
\Omega = \oint \frac{d^4 \rho}{Q(\rho)}
\end{equation}
in a local (holomorphic) coordinate patch $(\rho_1,\rho_2,\rho_3,\rho_4)$, thus
determining the weight function $w(x)$ for the random points. The
disadvantage of this Monte-Carlo integration scheme is that one is
forced to use the random point set without modification. In
particular, it tuns out that the point weights $w(x)$ fluctuate over a
large range with the error accumulating in the badly-sampled
regions. This was mitigated in\cite{Keller:2009vj} using stratified
sampling, at the cost of having to work with higher-dimensional
Kodaira embeddings.
\subsection{An Adaptive Integration Algorithm}
For the purposes of this paper, we developed a combination of the best
features of the naive higher-dimensional rectangle rule and the
Monte-Carlo sampling. In a nutshell, the idea is to first parametrize
point(s) of the Calabi-Yau threefold similarly to the parametrization
by sections of three line bundles. Then integrate using the standard
higher-dimensional rectangle rule by constructing a suitable cell
decomposition of the \emph{parameter space}, not the Calabi-Yau
manifold.
As the simplest example, consider the $(2,3)$-hypersurface in
$X\subset \ensuremath{\mathop{\null {\mathbb{P}}}}\nolimits^1\times\ensuremath{\mathop{\null {\mathbb{P}}}}\nolimits^2$, which is a $K3$ surface. The projection
$\pi:X\to\ensuremath{\mathop{\null {\mathbb{P}}}}\nolimits^2$ onto the $\ensuremath{\mathop{\null {\mathbb{P}}}}\nolimits^2$ factor is, generically,
two-to-one. Locally, there are two inverses $\pi^{-1}_1$, $\pi^{-1}_2$
for the two-sheeted cover $\pi$. With it, one can rewrite the
integration as
\begin{equation}
\int_X f(x) \sqrt{g} \; d^4x
=
\int_{\ensuremath{\mathop{\null {\mathbb{P}}}}\nolimits^2}
\sum_{i=1}^2
f\big(\pi^{-1}_i(z)\big)
\underbrace{
\left|\frac{\partial\pi^{-1}_i}{\partial z}\right|
\;
\frac{\sqrt{g}}{\sqrt{g_\text{aux}}}
}_{w_i(z)}
\;
\sqrt{g_\text{aux}} \; d^4z ,
\end{equation}
where the sum runs over the different sheets. The weights $w_i(z)$ are
the product of the Jacobian of the coordinate transformation and, as
before, the scalar factor required to transform the auxiliary measure
into the desired Calabi-Yau measure. Finally, it is easy to integrate
over projective space. For our purposes, we will use an adaptive
integration scheme where we start with a decomposition of $\ensuremath{\mathop{\null {\mathbb{P}}}}\nolimits^2$ into
cells of equal volume with respect to some convenient auxiliary
measure. Then, if any of the weights $w_i(z)$ is significantly larger
than the average weight, we recursively subdivide the cell until the
weight is acceptably small. Finally, we use the usual rectangle rule
and sum over all cells to compute the integral.
It is not really necessary for the map $\pi:X\to \ensuremath{\mathop{\null {\mathbb{P}}}}\nolimits^2$ to be a
multisheeted cover or even to be defined everywhere. In particular, a
dominant rational map would be perfectly fine. To summarize, our
integration algorithm
\begin{itemize}
\item does not require a cell decomposition of the Calabi-Yau manifold,
\item converges as $O(\tfrac{1}{N})$ with the number of
points, just like the standard rectangle rule, and
\item makes it easy to adapt each integration step to reduce numerical
errors.
\end{itemize}
\subsection{Integrating Over Projective Space}
Thus far, we reduced the integration to one over projective space
$\ensuremath{\mathop{\null {\mathbb{P}}}}\nolimits^n$ with some convenient auxiliary metric. We now describe a way
of decomposing $\ensuremath{\mathop{\null {\mathbb{P}}}}\nolimits^n$ into equal-sized cells that is suitable for
adaptive subdivision of the cells. First, let us decompose $\ensuremath{\mathop{\null {\mathbb{P}}}}\nolimits^n$
into $n+1$ polydiscs
\begin{equation}
\ensuremath{\mathop{\null {\mathbb{P}}}}\nolimits^n = \bigcup_{i=0}^n D_i
,\qquad
D_i =
\Big\{
[z_0:
\cdots:z_{i-1}:
\underbrace{1}_{\mathpalette\mathclapinternal{\text{$i$-th entry}}}:
z_{i+1}:\cdots:
z_n]
~
\Big|
~
|z_j|\leq 1
\Big\}
\simeq
D^n
.
\end{equation}
\begin{figure}[p]
\centering
\begin{tabular}{c@{\ }c}
& \parbox{12cm}{\hfill
coarse ($\approx 20$ points)\hfill\hfill
medium ($\approx 80$ points)\hfill\hfill
fine ($\approx 300$ points)\hfill \strut}\\
\begin{sideways}
\parbox{4.28cm}{\centering Euclidean: $dz\wedge d\bar{z}$}
\end{sideways}
& \includegraphics{Mesh-regular.pdf} \\
\begin{sideways}
\parbox{4.28cm}{\centering Fubini-Study: $\omega_{FS}$}
\end{sideways}
& \includegraphics{Mesh-FS.pdf} \\
\begin{sideways}
\parbox{4.28cm}{\centering Calabi-Yau: $\Omega\wedge\overline{\Omega}$}
\end{sideways}
& \includegraphics{Mesh-CY.pdf}
\end{tabular}
\caption{The adaptive mesh generation for four different
integrations over the disk $D=\{|z|\leq 1\}$.
In each case, the points are chosen such that their weight is
approximately constant. The point distributions in the three rows
are further explained in items~\ref{item:adapt1},
\ref{item:adapt2}, and \ref{item:adapt3} on page~\pageref{item:adapt1}.}
\label{fig:adapt}
\end{figure}
Note that, by dividing with the homogeneous coordinate of largest
magnitude, the homogeneous coordinates of any point can be rescaled
such that
\begin{itemize}
\item one homogeneous coordinate equals unity, say, $z_i=1$, and
\item all other homogeneous coordinates have equal or smaller
magnitude $|z_j|\leq 1$, $j\not=i$.
\end{itemize}
Hence, a generic point of $\ensuremath{\mathop{\null {\mathbb{P}}}}\nolimits^n$ is contained in precisely one of the
polydiscs $D_i$. The polydiscs overlap in the measure-zero set where
two or more homogeneous coordinates attain the maximum magnitude.
It remains to decompose each polydisc $D_i\simeq D^n$. For this
purpose, we use that the polydisc is the Cartesian product of $n$
individual disks
\begin{equation}
D = \Big\{
r e^{i\varphi}
~\Big|~
r\leq 1
,\;
0\leq \varphi <2\pi
\Big\} \subset \ensuremath{{\mathbb{C}}}
.
\end{equation}
Any cell decomposition of $D\subset \ensuremath{{\mathbb{C}}}$ induces one of the
polydisc. For our implementation, we chose to decompose $D$ into
annuli of width $\delta r$, and each annulus into segments of angles
$\delta \varphi \approx \delta r /(2\pi r)$. The annulus segments are
approximately quadratic and of area $2\pi r \delta \varphi \cdot
\delta r$ in the Euclidean measure.
To summarize, we decompose the polydiscs $D_i$ into hypercubes of
constant volume in the Euclidean metric, which we use as the auxiliary
measure. If the weight of the cell is too large, we subdivide it by
splitting the annulus segment in one of the $n$ discs\footnote{We
chose one of the $n$ discs randomly, but this could clearly be
improved by separating the cell in the direction of the biggest
gradient for the weights.} that constitute $D_i$. Summing over each
of the $n+1$ polydiscs according to the rectangle rule then computes
the integral over $\ensuremath{\mathop{\null {\mathbb{P}}}}\nolimits^n$. To illustrate the adaptive integration
algorithm, we plot the point distribution for three different volume
forms in \autoref{fig:adapt}. The first, second, and third column
shows successive refinements with growing number of points. Each row
corresponds to a different volume form that the point distribution is
adapted to:
\begin{enumerate}
\item\label{item:adapt1} In the top row, points are distributed
regularly on the disc. This distribution, with constant weight
attached to each point, approximates integration with respect to the
Euclidean measure.
\item\label{item:adapt2} In the second row, we try to approximate
integration with respect to the Fubini-Study measure. Here, the
points are chosen adaptively and are denser towards the center where
the Fubini-Study volume form is denser.
\item\label{item:adapt3} In the third row, we illustrate integration
over the $(2,2)$-hypersurface
\begin{multline}
w_0^2 z_0^2
- \tfrac{9+i}{10} w_0^2 z_0 z_1
- \tfrac{3-9i}{10} w_0^2 z_1^2
- \tfrac{1-7i}{10} w_0 w_1 z_0^2
- \tfrac{5-10i}{10} w_0 w_1 z_0 z_1
\\
- \tfrac{4-i}{10} w_0 w_1 z_1^2
- \tfrac{10+9i}{10} w_1^2 z_0^2
- \tfrac{2+2i}{10} w_1^2 z_0 z_1
+ \tfrac{5-i}{10} w_1^2 z_1^2
= 0
\end{multline}
in $\ensuremath{\mathop{\null {\mathbb{P}}}}\nolimits^1_{[w_0:w_1]}\times\ensuremath{\mathop{\null {\mathbb{P}}}}\nolimits^1_{[z_0:z_1]}$ with respect to the
Calabi-Yau volume form. Using the projection
$\pi:X\to\ensuremath{\mathop{\null {\mathbb{P}}}}\nolimits^1_{[z_0:z_1]}$ we rewrite this integration as the
integration over a single $\ensuremath{\mathop{\null {\mathbb{P}}}}\nolimits^1$. The disc in \autoref{fig:adapt}
shows the first patch $[z_0:z_1] = [z:1]$, $|z|\leq 1$. The points
are adaptively chosen to have approximately constant weight.
\end{enumerate}
\section{Diagnosing Stability}\label{t_iter_error}
One of the goals of this paper to use the generalized Donaldson
algorithm to probe K\"ahler cone substructure in higher-dimensional
K\"ahler cones. That is, we would like to efficiently be able to scan
a K\"ahler cone for regions where a given bundle is stable. To this
end, we present in this section a new and efficient numerical measure
of the stability properties of $\ensuremath{\mathscr{V}}$ at a fixed polarization.
The central idea is as follows. If one just wants to numerically
determine whether a bundle is stable or not, it is not necessary to
explicitly compute the Hermitian Yang-Mills connection. Instead, one
can simply use the first step of the generalized Donaldson algorithm,
namely the convergence properties of the T-operator. As we saw in
\autoref{connection_alg}, according to Wang's theorem, for a fixed
embedding defined by $H^0(X, \ensuremath{\mathscr{V}}\otimes {\cal L}^{k_{H}})$, the
T-operator can be moved to a balanced place if and only if it is
Gieseker stable as defined in eq.~\eqref{gieseker}. Thus, for an
embedding defined by even a relatively small number of sections (that
is, a small degree of twisting $k_H$), it should be straightforward to
check whether or not the iterations of the T-operator in \eqref{t_gen}
converge to a fixed point. If such a fixed point exists, then we know
that the bundle is Giesker stable. In general, this computation is
easier and faster than a complete computation of the connection and
its integrated error measure.
\begin{figure}[htb]
\centering
\input{StabilityWall-EllipticK3-Substructure-r.tex}
\caption{The T-operator convergence measured for the same bundle
and polarizations as in \autoref{tab:Substructure}.}
\label{tab:r}
\end{figure}
As part of our program of probing the stability of general bundles
$\ensuremath{\mathscr{V}}$ in higher dimensional K\"ahler cones, it is intriguing to
see that one can gain important information regarding Gieseker
stability from the T-operator and only a minimal number of twists
$\ensuremath{\mathscr{V}} \otimes {\cal L}^{k_{H}}$. However, one must be careful. While it
is certainly true that all slope-stable bundles are Gieseker stable,
properly Gieseker stable bundles need only be slope semi-stable. We
can only conclude that if the T-operator fails to converge, then
$\ensuremath{\mathscr{V}}$ is \emph{not} slope-stable. However, if it does converge, we
only know that $\ensuremath{\mathscr{V}}$ is semi-stable, but not necessarily a
solution to \eqref{hym_genr}. In order to be certain that $\ensuremath{\mathscr{V}}$ is properly slope-stable, one would then need to augment the analysis of the T-iteration with a full computation of the $\tau$ error measure in \eqref{tau_norm}, as discussed in the previous sections. In the next section, we
will investigate this subtle semi-stable behavior in detail. For now, however, we
only explore how much can be gained from a simple check of T-operator
convergence.
With this idea in mind, we develop a new error measure based on the convergence of the T-operator. Considering the matrix $H_{\alpha\bar\beta}$ in \eqref{T_gen_bal}, we would like to know how much the matrix changes as one moves from the $m$-th to the $(m+1)$-th iteration of the T-operator in \eqref{t_gen}. To answer this question, consider the eigenvalues of the $H_{\alpha\bar\beta}$-matrix. At each step of the
iteration, the largest eigenvalues are the relevant features; small
eigenvalues only give small corrections to the connection on the
bundle. Let
\begin{equation}
v_{t_1,t_2}^\text{max}(m)
\end{equation}
be the largest eigenvalue of the $H_{\alpha\bar\beta}$ after $m$
iterations of the T-operator. Except for the overall numerical scale, one expects
that the details of the initial values are washed out by the
iteration. However, the overall scale is preserved by the iteration
and, moreover, does not enter the connection. Therefore, a good
quantity to measure the convergence is
\begin{equation}\label{r_def}
r_{(t_1,t_2)}^\text{max}(m) =
\frac{
v_{(t_1,t_2)}^\text{max}(m)
}{
v_{(t_1,t_2)}^\text{max}(m-1)
}
- 1
.
\end{equation}
By \autoref{wang2} and the above considerations, we know that
\begin{equation}
\lim_{m\to \infty} r_{(t_1,t_2)}^\text{max}(m) =0
\end{equation}
if $\ensuremath{\mathscr{V}}$ is Gieseker stable as in eq.~\eqref{gieseker}. For the
purposes of this paper, we are interested in slope-stability and
solutions to the Hermitian Yang-Mills equations. Since slope-stability
implies Gieseker stability, it follows that
$r_{(t_1,t_2)}^\text{max}(m)$ should have the following behavior in
the presence of K\"ahler cone substructure:
\begin{equation}
\lim_{m\to \infty}
r_{(t_1,t_2)}^\text{max}(m)
\quad
\begin{cases}
\quad
= 0 & \text{if}~
(t_1,t_2) \in \mathcal{K}_\text{stable}
\\
\quad
> 0 & \text{if}~
(t_1,t_2) \in \mathcal{K}_\text{unstable}
\end{cases}
\end{equation}
We will put this to the test in the example introduced in \autoref{eg_sec}. For the $SU(2)$
bundle $\ensuremath{\mathscr{V}}$ in \eqref{k3_eg1}, one would expect
\begin{equation}
\lim_{m\to \infty}
r_{(t_1,t_2)}^\text{max}(m)
\quad
\begin{cases}
\quad
= 0 & \text{if}~
\frac{t_1}{t_2}> \frac{4}{3}
\\
\quad
> 0 & \text{if}~
\frac{t_1}{t_2} < \frac{4}{3}
\end{cases}
\end{equation}
The results are shown in \autoref{tab:r} for the same six polarizations chosen in \autoref{tab:Substructure}. The connection T-operator was computed at $30$ iterations and the comparison of \eqref{r_def} made between the $29$th and $30$th iterations. As expected, $r_{(t_1,t_2)}^\text{max}(30)$ is approximately zero in the slope-stable region of K\"ahler moduli space, but non-zero in the unstable region.
\begin{figure}[htb]
\centering
\input{StabilityWall-EllipticK3-Angular.tex}
\caption{Convergence of the T-iteration on the $SU(2)$ bundle in \eqref{k3_eg1} with
substructure for different K\"ahler moduli. The
radial direction $\gcd(t_1,t_2)$ is always chosen to be as large
as possible subject to the constraint that there are $\dim
H^0\big( \ensuremath{\mathscr{O}}(t_1,t_2) \big) \leq 200$ sections.}
\label{tab:Angular}
\end{figure}
On the boundary between $\ensuremath{\mathcal{K}}_\text{stable}$ and
$\ensuremath{\mathcal{K}}_\text{unstable}$, however, we find that the
$r_{(4,3)}^\text{max}(30)$ is also close to zero, despite the fact
that $\ensuremath{\mathscr{V}}$ in \eqref{k3_eg1} is only slope semi-stable for this
polarization. This is to be expected however, since it can be shown
via direct computation (and \eqref{gieseker}) that $\ensuremath{\mathscr{V}}$ is still
properly Gieseker stable for this line in K\"ahler moduli space and
Gieseker stability implies only slope semi-stability. It follows that
to accurately determine the behavior on this boundary, one would need
to also investigate the $\tau$ error measure of the previous section,
which can distinguish between slope-stable and semi-stable
behavior. We return to the boundary behavior in the next section.
For now, it should be noted that one need not have checked the
T-operator for all $500$ sections in \autoref{tab:r}. In fact, the
convergence of the T-operator is already evident at a much smaller
projective embedding. While we must have enough sections to make sure
that \eqref{bundle_embed} is a proper embedding, we can see the
stability properties of $\ensuremath{\mathscr{V}}$ from the first embedding which makes
$H^0(X,\ensuremath{\mathscr{V}} \otimes {\cal L}^{k_{H}})$ non-vanishing. In
Figures~\ref{tab:Angular} and~\ref{tab:Substructure2D} we present the
same results from a purely ``angular'' point of view; that is,
computing the T-operator at only a single point along each of the rays
plotted in \autoref{tab:r}. These points (that is, the twistings
${\cal L}^{k_{H}}$) were chosen so that there are $ \leq 200$ sections at
each point. Note that the oscillation in the height of the points in
\autoref{tab:Angular} in the unstable region is not significant since
we are comparing data from different polarizations, which leads to
different normalizations. The only meaningful comparison is between
zero and non-zero values of $r_{(t_1,t_2)}^{max}(m)$. We also present
a $3$-dimensional plot of the same K\"ahler cone substructure in
\autoref{tab:Substructure2D}. For all calculations, the connection was
computed with $N_G = 74,892$ points (adaptive) and $30$ iterations of
the T-operator.
\begin{figure}[tbp]
\centering
\input{StabilityWall-EllipticK3-2D.pspdftex}
\caption{The K\"ahler cone substructure associated with the $SU(2)$ bundle in \eqref{k3_eg1} for different
K\"ahler moduli. Green points are stable, blue points are on the
line of stability, and red points are unstable polarizations.}
\label{tab:Substructure2D}
\end{figure}
Finally, in \autoref{tab:ev-k} we consider the rate at which the
T-operator approaches its fixed point as the number of sections
defining the embedding is increased, see eq.~\eqref{bundle_embed}. We
observe that convergence of the T-iteration is generally slower for
more sections. In particular, in \autoref{tab:ev-k} we present the
convergence (and divergence) of the T-iteration for polarizations
$\ensuremath{\mathscr{O}}(t,t)$, which is in the unstable region
$\ensuremath{\mathcal{K}}_\text{unstable}$ of the K\"ahler moduli space.
\begin{figure}[htb]
\centering
\input{StabilityWall-EllipticK3-EigenvalueRatio-k.tex}
\caption{Convergence (for the polarization $\ensuremath{\mathscr{O}}(2t,t)$) and
divergence (for $\ensuremath{\mathscr{O}}(t,t)$) of the T-iteration on the bundle
$\ensuremath{\mathscr{V}}$ in \eqref{k3_eg1}. Substructure is shown for different K\"ahler moduli
along a stable and an unstable ray, respectively. The orange and gray lines are associated with two different values of the bundle moduli for the unstable polarization.}
\label{tab:ev-k}
\end{figure}
\section{On The Line Of Semi-Stability}\label{stab_wall_sec}
In the previous sections, we demonstrated that the generalized
Donaldson algorithm is capable of broadly probing K\"ahler cone
substructure. In this section, we take a detailed look at the boundary
between stable/unstable regions, the so-called ``stability
wall"~\cite{Anderson:2009sw,Anderson:2009nt}.
\begin{figure}[htb]
\centering
\input{Line-EllipticK3-tau.tex}
\caption{The $\tau$ error measure associated with the $SU(2)$ bundle in \eqref{k3_eg1} at the stability wall in K\"ahler moduli space. The bundle modulus $\varphi$ defined in \eqref{f_varphi} determines whether $\ensuremath{\mathscr{V}}$ is strictly semi-stable (when $\varphi \neq 0$) or poly-stable (when $\varphi=0$). In the latter case, the reducible connection satisfies the Hermitian Yang-Mills equations \eqref{the_hym}.
}
\label{tab:Wall}
\end{figure}
Let us revisit the $SU(2)$ bundle $\ensuremath{\mathscr{V}}$ on the $K3$ surface
defined by \eqref{k3_eg1}. As discussed in \autoref{eg_sec}, $\ensuremath{\mathscr{V}}$
is destabilized in part of the K\"ahler cone by a rank $1$ subsheaf
${\cal F} \subset \ensuremath{\mathscr{V}}$ defined in \eqref{destabilizing}. In the region
$\ensuremath{\mathcal{K}}_\text{stable}$ in \eqref{eg1substruc}, $\mu({\cal F})<0$ and in
$\ensuremath{\mathcal{K}}_\text{unstable}$, $\mu({\cal F})>0$. What happens on the boundary
line between these two regions, where $\mu({\cal F})=\mu(\ensuremath{\mathscr{V}})=0$? By
definition, on a line with $\mu({\cal F})=\mu(\ensuremath{\mathscr{V}})$ the bundle
$\ensuremath{\mathscr{V}}$ is semi-stable. Hence, for generic values of the bundle
moduli its connection will not solve the Hermitian Yang-Mills
equations. However, semi-stable bundles are distinguished from
unstable bundles in that they can provide supersymmetric solutions for
special loci in their moduli space. Looking at the definitions in
\eqref{slope_req} and \eqref{slope_themovie} in
\autoref{review_section}, we see that the only way that $\ensuremath{\mathscr{V}}$ can
satisfy the Hermitian Yang-Mills equations when
$\mu({\cal F})=\mu(\ensuremath{\mathscr{V}})$ is for it to be \emph{poly-stable} rather than
strictly semi-stable. That is, if the connection on $\ensuremath{\mathscr{V}}$ is
decomposable, it is possible that a connection may exist which
satisfies \eqref{hym_genr}. Mathematically, this property of slope
semi-stable bundles is characterized by the notion of S-equivalence
classes and the Harder-Narasimhan filtration \cite{huybrechts,
MR2180559}, which states that every semi-stable bundle has a unique
poly-stable representative in its moduli space.
For the bundle in \eqref{k3_eg1}, we find that the poly-stable representative arises when $\ensuremath{\mathscr{V}}$ decomposes on the semi-stable wall as the direct sum
\begin{equation}\label{split}
\ensuremath{\mathscr{V}} \longrightarrow {\cal F} \oplus {\cal O}(-2,1)
.
\end{equation}
\begin{figure}[htb]
\centering
\input{Line-EllipticK3-iterations.tex}
\caption{The convergence/divergence of T-iteration associated with the $SU(2)$ bundle in \eqref{k3_eg1} at the stability wall in K\"ahler moduli space. The bundle modulus $\varphi$ defined in \eqref{f_varphi} determines whether $\ensuremath{\mathscr{V}}$ is strictly semi-stable (when $\varphi \neq 0$) or poly-stable (when $\varphi=0$). In the latter case, the reducible connection satisfies the Hermitian Yang-Mills equations \eqref{the_hym}.
}
\label{tab:r_iterations}
\end{figure}
One can quantify this decomposable locus in the bundle moduli space as
follows. The bundle moduli space of $\ensuremath{\mathscr{V}}$ is described by the
parameters in the map $f$ in \eqref{k3_eg1}. We can describe the
decomposable locus in \eqref{split} in terms of this map by
parametrizing one particular vector bundle modulus as $\varphi$ in the
monad map
\begin{equation}\label{f_varphi}
f_\varphi
=
\begin{pmatrix}
(9-i) y_0 + (-6-6i) y_1 + (-8-8i) y_2 \\
(-9-9i) y_0 + (-2+9i) y_1 + ( 4-4i) y_2 \\
\varphi
\big(
{ \scriptstyle
(-10+7i) x_0^4 +
( 6-9i) x_0^3 x_1+
( 3-8i) x_0^2 x_1^2+
( -9-4i) x_0 x_1^3+
( 8+4i) x_1^4
}
\big)
\end{pmatrix}
\end{equation}
which determines $\ensuremath{\mathscr{V}}_{\varphi}$,
\begin{equation}
\label{eq:Vphi}
\vcenter{\xymatrix{
0
\ar[r] &
\ensuremath{\mathscr{O}}(-2,-1)
\ar[r]^-{f_\varphi} &
\ensuremath{\mathscr{O}}(-2,0)^{\oplus 2} + \ensuremath{\mathscr{O}}(2,-1)
\ar[r] &
\ensuremath{\mathscr{V}}_\varphi
\ar[r] &
0~.
}}
\end{equation}
When $\varphi=0$, the bundle splits as in \eqref{split} and the
resulting direct sum is poly-stable and a solution to the Hermitian
Yang-Mills equations. When $\varphi \neq 0$, $\ensuremath{\mathscr{V}}$ is only
semi-stable, one cannot solve \eqref{the_hym}, and supersymmetry is
broken.
Let us now revisit the numerical results of the past few sections in light of this structure. Can the error measures $\tau(A_{\ensuremath{\mathscr{V}}})$ and $r^{max}(m)$, introduced in \autoref{modific} and \autoref{t_iter_error} respectively, accurately reveal this $\varphi$-dependent structure? To answer this question, we have run the algorithm for the same bundle again, this time keeping the K\"ahler moduli fixed directly on the line $(t_1,t_2) \sim (4,3)$, but allowing the bundle moduli to vary through the variable $\varphi$. The metric was computed for this polarization with $N_g = 202,800$ points (adaptive), $30$ iterations of the metric T-operator and with a K\"ahler form determined by ${\cal O}(4,3)^{\otimes 3}$. Meanwhile, the connection utilized $N_G = 74,892$ points (adaptive) and $100$ iterations of the connection T-operator.
The results for the $\tau(t_1,t_2)$ error measure of equation \eqref{tau_norm} are shown in \autoref{tab:Wall}. Significantly, the algorithm clearly distinguishes between the semi-stable ($\varphi \neq 0$) and poly-stable ($\varphi=0$) behavior of $\ensuremath{\mathscr{V}}$. Moreover, one can repeat the same analysis with the T-operator convergence measure, $r^{max}(12,9)(m)$, of \eqref{r_def}. For this analysis, we hold ourselves fixed at one point in K\"ahler moduli space ($t_1=12, t_2=9$) and allowing the number of T-iterations $m$ to increase. Once again, we find that this error measure definitively distinguishes between the supersymmetric and non-supersymmetric configurations. The convergence of the T-iteration is shown in \autoref{tab:r_iterations}.
The graph \autoref{tab:r_iterations} shows a high degree of
sensitivity to the bundle moduli. The fact that the T-iterations can
distinguish the $\varphi=0$ and $\varphi \neq 0$ cases is significant,
since it means that for a general value of the bundle moduli $\ensuremath{\mathscr{V}}$
is \emph{not} Gieseker stable for the boundary wall polarization. This
can be verified directly using the destabilizing subsheaf ${\cal F}$ in
\eqref{destabilizing} and the definitions of Gieseker stability,
\eqref{gieseker}, in \autoref{review_section}. In fact, for
${\cal L}={\cal O}(4,3)$ we have
\begin{equation}
p_{{\cal L}}({\cal F})(n)=45n^2-3 ,
\end{equation}
while
\begin{equation}
p_{{\cal L}}(\ensuremath{\mathscr{V}})(n)=\frac{1}{2}(90n^2-8)~.
\end{equation}
Therefore,
\begin{equation}
p_{{\cal L}}(\ensuremath{\mathscr{V}})(n)\prec p_{{\cal L}}({\cal F})(n)
\end{equation}
and, by \eqref{gieseker_ineq}, $\ensuremath{\mathscr{V}}$ is not Gieseker stable. The results of \autoref{tab:r_iterations} are in complete agreement with this; showing the $\varphi=0$ reducible connection T-operator converging to its fixed point while the Gieseker unstable configurations with $\varphi \neq 0$ fail to converge. It follows that the T-operator is sensitive enough to the bundle moduli that it can distinguish between these important cases.
\section{A Calabi-Yau Threefold Example}
\label{cy3_sec}
In the previous sections, we investigated K\"ahler cone substructure with a number of new tools and from a variety of perspectives. Here, we present a final example involving geometry directly relevant to realistic $\mathcal{N}=1$ heterotic compactifications; namely, a Calabi-Yau threefold with a higher-rank vector bundle defined over it.
As a base manifold, we choose a Calabi-Yau threefold determined by a
generic degree $(4,2)$-hypersurface in $\ensuremath{\mathop{\null {\mathbb{P}}}}\nolimits^3\times\ensuremath{\mathop{\null {\mathbb{P}}}}\nolimits^1$. For this
threefold, $h^{1,1}=2$ and all line bundles can be written in the form
${\cal O}(a_1,a_2)$, where ${\cal O}(1,0)$ arises from the hyperplane descending
from $\mathbb{P}^1$ and ${\cal O}(0,1)$ descends from the hyperplane of
$\mathbb{P}^3$.
\begin{figure}[htb]
\centering
\input{StabilityWall-Threefold24-Angular.tex}
\caption{Convergence of the T-iteration on the $SU(3)$ bundle defined by \eqref{cy3_eg_dual} for different
K\"ahler moduli. The radial direction $\gcd(t_1,t_2)$ is always
chosen to be as large as possible subject to the constraint that
there are $\dim H^0\big( \ensuremath{\mathscr{O}}(t_1,t_2) \big) \leq 500$
sections. The results agree with the K\"ahler cone substructure in \eqref{cy3substruc}.}
\label{tab:AngularThreefold}
\end{figure}
Over this threefold, define the $SU(3)$ monad bundle
\begin{equation}\label{cy3_eg}
0 \longrightarrow \ensuremath{\mathscr{V}} \longrightarrow {\cal O}(-2,1) \oplus {\cal O}(3,-1) \oplus {\cal O}(2,0) \oplus {\cal O}(2,1)^{\oplus 2} \stackrel{f}{\longrightarrow} {\cal O}(4,1) \oplus {\cal O}(3,1) \longrightarrow 0
\end{equation}
and its associated dual bundle
\begin{multline}
\label{cy3_eg_dual}
0 \longrightarrow {\cal O}(-4,-1) \oplus
{\cal O}(-3,-1) \longrightarrow
\\
{\cal O}(-2,-1)^{\oplus 2} \oplus {\cal O}(-2,0)
\oplus {\cal O}(-3,1) \oplus {\cal O}(2,-1) \longrightarrow \ensuremath{\mathscr{V}}^{\vee}
\longrightarrow 0 .
\end{multline}
This example was chosen because it once again the stability depends on
the precise value of the K\"ahler moduli. It can be verified, using
the definitions of \autoref{review_section}, that the bundle $\ensuremath{\mathscr{V}}$
in eq.~\eqref{cy3_eg} has two potentially destabilizing
sub-sheaves. These are ${\cal F}_1$,${\cal F}_2 \subset \ensuremath{\mathscr{V}}$ given by
\begin{equation}
\begin{split}
&0 \longrightarrow {\cal F}_1 \longrightarrow
{\cal O}(3,-1) \oplus {\cal O}(2,0) \oplus {\cal O}(2,1)^{\oplus 2}
\stackrel{f}{\longrightarrow}
{\cal O}(4,1) \oplus {\cal O}(3,1) \longrightarrow 0 ,\\
&0 \longrightarrow {\cal F}_2 \longrightarrow
{\cal O}(-2,1) \oplus {\cal O}(2,0) \oplus {\cal O}(2,1)^{\oplus 2}
\stackrel{f}{\longrightarrow} {\cal O}(4,1)
\oplus {\cal O}(3,1) \longrightarrow 0
.
\end{split}
\end{equation}
Both are of rank $2$ with $c_1({\cal F}_1)=(2,-1)$ and
$c_1({\cal F}_2)=(-3,1)$. Due to these two destabilizing sub-sheaves, the
K\"ahler cone divides into three regions
\begin{equation}\label{cy3substruc}
\begin{split}
\mathcal{K}_\text{stable}
=&\;
\big\{
\ensuremath{\mathscr{O}}(t_1, t_2)
\;\big|\;
\tfrac{4}{5} < \tfrac{t_2}{t_1} < \tfrac{4}{3}
\big\}
, \\
\mathcal{K}_\text{unstable}
=&\;
\big\{
\ensuremath{\mathscr{O}}(t_1, t_2)
\;\big|\;
\tfrac{t_2}{t_1} < \tfrac{4}{5}
\big\}
~\cup~
\big\{
\ensuremath{\mathscr{O}}(t_1, t_2)
\;\big|\;
\tfrac{t_2}{t_1} > \tfrac{4}{3}
\big\}
.
\end{split}
\end{equation}
Moreover, a direct calculation using \eqref{gieseker} shows that
$\ensuremath{\mathscr{V}}$ is \emph{still} Gieseker stable at each of the stability
walls defined by $t_2/t_1=4/3$ and $t_2/t_1=4/5$. Thus, while the
bundle is strictly semi-stable along these rays (and, hence, does not
solve the Hermitian Yang-Mills equations eq.~\eqref{the_hym}), one would
still expect the T-iteration to converge for points along these
boundaries.
For ease of embedding, the T-operator is computed for
$\ensuremath{\mathscr{V}}^{\vee}$. However, since $\ensuremath{\mathscr{V}}$ and its dual must be
slope-stable/unstable in the same regions of K\"ahler moduli space,
this does not affect our results. Since we have demonstrated in the
previous sections that the T-iteration is a fast and efficient check
of stability, we will consider here only the error measure
$r_{(t_1,t_2)}^{max}(m)$. It is interesting to observe that, even in
this more complex example, the data shown in
\autoref{tab:AngularThreefold} clearly reproduces the K\"ahler cone
substructure described in \eqref{cy3substruc}. Moreover, at the points
on the two stability walls we find, as expected, that
$r_{(t_1,t_2)}^{max}(30)=0$, since $\ensuremath{\mathscr{V}}$ is still Gieseker stable
for these lines of slope semi-stability. The connection integration
was performed with $N_G = 119,164$ points (adaptive) and $30$
iterations of the T-operator. The plot shown in
\autoref{tab:AngularThreefold} are analogous to the results in
\autoref{tab:Angular} for the $K3$ surface.
\section{Conclusions and Future Work}\label{conclusion_sec}
In this paper, we extended the generalized Donaldson algorithm to manifolds with
higher-dimensional K\"ahler cones and presented a method for approximating the connection on slope-stable holomorphic vector bundles in this context. We also introduced a numerical criterion for determining the existence of supersymmetric heterotic vacua, without having to compute the connection.
These techniques clearly can be used to search for new classes of smooth $\mathcal{N}=1$ supersymmetric vacua in heterotic string theory. However, the explicit knowledge of the gauge connection that they provide allows us to go far beyond a simple categorization of vacua.
It is a long-standing goal of string theory to produce low-energy
theories whose effective actions reproduce the symmetries, particle
spectrum, and properties of elementary particle physics. Within the
context of smooth heterotic string or M-theory
compactifications~\cite{Candelas:1985en, Lukas:1998yy, Lukas:1998ew},
there has been considerable progress in recent
years~\cite{Braun:2004xv, Braun:2005nv, Braun:2005ux, Braun:2005zv,
Buchbinder:2007ad, Bouchard:2005ag, Anderson:2009mh}. Despite these
successes, certain observable quantities of particle physics, such as
the gauge and Yukawa couplings, are difficult to compute
directly. Normalized Yukawa couplings, for example, depend on both the
coefficients of cubic terms in the superpotential and the explicit
form of the K\"ahler potential. In turn, these quantities depend on
the detailed structure of the underlying geometry -- that is, the
metric and gauge connection on the Calabi-Yau threefold and
holomorphic vector bundle respectively. Hence, the Yukawa couplings in
the four-dimensional effective theory are not known, except in very
special cases. One can rarely do better than the qualitative statement
that such coupling coefficents either vanish or are of order
one. However, the methods introduced
in~\cite{Douglas:2006rr,Anderson:2010ke} and extended in this paper
allow one, for the first time, to explicitly compute both the metric
and the gauge connection and, hence, the Yukawa coefficients. In
future work~\cite{yukawas_numeric}, we will use these techniques to
calculate these physical couplings.
\section*{Acknowledgments}
The work of L. Anderson and B. A. Ovrut is supported in part by the DOE under contract No. DE-AC02-76-ER-$03071$ and by NSF RTG Grant DMS-$0636606$ and NSF-$1001296$.
\bibliographystyle{utcaps}
\renewcommand{\refname}{Bibliography}
\addcontentsline{toc}{section}{Bibliography}
|
\section{Introduction}
The measurement of the $t \bar t$ forward-backward (FB) asymmetry at Tevatron~\cite{:2007qb,Aaltonen:2008hc,Aaltonen:2011kc} has motivated a plethora of models which attempt to accommodate the experimental values, up to $3.4\sigma$ larger than the prediction of the Standard Model (SM). This task is not straightforward because the measured $t \bar t$ cross section is in good agreement with the SM: any ``generic'' addition to the $t \bar t$ production amplitude, large enough to produce the observed FB asymmetry, will easily give rise to too large a departure in the total rate. Many of the proposed models circumvent this problem at the expense of a cancellation between (linear) interference and (quadratic) new physics terms in the total cross section,
\begin{equation}
\sigma(t\bar t) = \sigma_\text{SM} + \delta \sigma_\text{int} + \delta \sigma_\text{quad} \,,
\end{equation}
where $\sigma_\text{SM}$ is the SM cross section, $\delta \sigma_\text{quad}$ the one corresponding to the new physics and $\delta \sigma_\text{int}$ the interference term.
The cancellation $\delta \sigma_\text{int} + \delta \sigma_\text{quad} \simeq 0$ requires a new large amplitude $A_\text{new} \sim - 2 A_\text{SM}$ which, obviously, should have observable effects elsewhere. The ideal candidate to search for these effects is the Large Hadron Collider (LHC).
New physics in $u \bar u,d \bar d \to t \bar t$ which produces such a large cancellation in the Tevatron cross section will likely produce an observable enhancement in the $t \bar t$ tail at LHC, even if at this collider top pair production is dominated by gluon fusion. (The tail is also enhanced at Tevatron energies but in the majority of the proposed models the deviations are compatible with present measurements~\cite{Aaltonen:2011kc}.) In some cases, this effect should be visible already with the data collected in 2010. Conversely, if large deviations are not reported, a number of candidates to explain the Tevatron $t \bar t$ asymmetry will be excluded from the list.
In this paper we make these arguments quantitative for a wide class of SM extensions. We consider general new vector bosons and scalars, classified by their transformation properties under the SM gauge group $\text{SU}(3) \times \text{SU}(2)_L \times \text{U}(1)_Y$, and study their possible effects in $t \bar t$ production. We use effective field theory to consistently (i) integrate out the new heavy states and obtain their contribution to $u \bar u,d \bar d \to t \bar t$ in terms of four-fermion operators; (ii) obtain the cross sections in terms of effective operator coefficients. This allows us to find, for each vector boson and scalar representation, a relation between the possible values of the asymmetry $A_\text{FB}$ at Tevatron and the excess in the $t \bar t$ tail at LHC. The discussion of all possible vector boson and scalar representations within the model-independent effective operator framework allows us to make stronger statements than in other model-independent studies~\cite{Zhang:2010dr,Degrande:2010kt,Blum:2011up,Delaunay:2011gv}. At the same time, any model with several vector bosons and scalars can be considered in our framework by simply summing the effective operator coefficients corresponding to the integration of each new particle. Previous studies of LHC signals associated to the FB asymmetry within particular models have been presented in
Refs.~\cite{Dorsner:2009mq,Jung:2010yn,Choudhury:2010cd,Cao:2011ew,
Bai:2011ed,Berger:2011ua,Bhattacherjee:2011nr,Patel:2011eh,Gresham:2011dg}.
After this analysis, we explore an alternative way to produce a large asymmetry with moderate effects in the $t \bar t$ tail at LHC. The key for this mechanism is the observation that one can also obtain a large $A_\text{FB}$ without significant changes in the total $t \bar t$ cross section by introducing new physics which only contributes to the latter at quadratic order. This has been studied before, in the effective formalism, in Ref.~\cite{Degrande:2010kt}. If we write
\begin{eqnarray}
\sigma^F (t \bar t) & = & \sigma_\text{SM} ^F + \sigma_\text{int}^F + \sigma_\text{quad}^F \,, \notag \\
\sigma^B (t \bar t) & = & \sigma_\text{SM} ^B + \sigma_\text{int}^B + \sigma_\text{quad}^B \,,
\end{eqnarray}
for the forward (F) and backward (B) cross sections, the total rate is maintained at first order provided $\sigma_\text{int}^F + \sigma_\text{int}^B \simeq 0$, which can be achieved, for example, with a new vector boson and a scalar. For models fulfilling this cancellation the size of the new physics contributions required to accommodate the experimental value of $A_\text{FB}$ are smaller. Therefore, these models provide a better agreement of the $t \bar t$ tail with Tevatron measurements, and predict a much smaller tail at LHC, which is still potentially observable with forthcoming measurements.
We remark that the use of effective field theory (with the assumption that the new physics is too heavy to be directly produced at LHC) does not limit much the generality of our conclusions. For $t$-channel exchange of new vector bosons or scalars, integrating out the new particles gives a good estimate for masses $M \gtrsim 1$ TeV. For $s$-channel exchange this approximation is worse, but the cross section enhancement produced by the new particle(s) is always larger than the one from the corresponding four-fermion operator(s), and in this sense our predictions are conservative. On the other hand, if new physics in the $t \bar t$ tail is not seen at LHC, the new resonances are heavy and the effective operator framework
can be safely used. It is also worth pointing out that we include $1/\Lambda^4$ corrections arising from the quadratic terms in new physics in the cross sections. The contributions from the interference of $1/\Lambda^4$ operators with the SM can be neglected, as they are suppressed with respect to the former for the values of parameters that are required to explain the $t \bar t$ asymmetry.
\section{Extra bosons, operators and $t \bar t$ production}
\label{sec:2}
There are ten possible $\text{SU}(3) \times \text{SU}(2)_L \times \text{U}(1)_Y$ representations~\cite{delAguila:2010mx} for new vector bosons contributing to $u \bar u,d \bar d \to t \bar t$, while for scalars eight representations contribute. They are collected in Table~\ref{tab:lagr}, where the first column indicates the label used to refer to them.\footnote{We note that for $\mathcal{W}_\mu$ and $\mathcal{H}_\mu$ the normalisation in the Lagrangian differs from Ref.~\cite{delAguila:2010mx} by a factor of two, to simplify the presentation of the limits.}
The relevant interaction Lagrangian is included as well, indicating the symmetry properties, if any, of the coupling matrices $g_{ij}$. We use standard notation with left-handed doublets $q_{Li}$ and right-handed singlets $u_{Ri}$, $d_{Ri}$; $\tau^I$ are the Pauli matrices, $\lambda^a$ the Gell-Mann matrices normalised to $\text{tr}(\lambda^a \lambda^b) = 2 \delta_{ab}$ and $\tilde \phi = \epsilon \phi$, $\psi^c = C \bar \psi^T$, with $\epsilon=i\tau^2$ and $C$ the charge conjugation matrix. The indices $a,b,c$ denote colour, and $\varepsilon_{abc}$ is the totally antisymmetric tensor.
\begin{table}[p]
\begin{center}
\begin{tabular}{c|clc}
Label & Rep. & \multicolumn{1}{c}{Interaction Lagrangian} & Sym. \\
\hline
$\mathcal{B}_\mu$ & $(1,1)_0$
& $-\left( g_{ij}^q \bar q_{Li} \gamma^\mu q_{Lj}
+ g_{ij}^u \bar u_{Ri} \gamma^\mu u_{Rj}
+ g_{ij}^d \bar d_{Ri} \gamma^\mu d_{Rj} \right) \mathcal{B}_\mu $ & $g=g^\dagger$ \\[1mm]
$\mathcal{W}_\mu$ & $(1,3)_0$
& $- g_{ij} \bar q_{Li} \gamma^\mu \tau^I q_{Lj} \, \mathcal{W}_\mu^I$
& $g=g^\dagger$ \\[1mm]
$\mathcal{B}_\mu^1$ & $(1,1)_1$
& $- g_{ij} \bar d_{Ri} \gamma^\mu u_{Rj} \, \mathcal{B}_\mu^{1\dagger} + \text{h.c.}$ & -- \\[1mm]
$\mathcal{G}_\mu$ & $(8,1)_0$
& $- \left( g_{ij}^q \bar q_{Li} \gamma^\mu \frac{\lambda^a}{2} q_{Lj}
+ g_{ij}^u \bar u_{Ri} \gamma^\mu \frac{\lambda^a}{2} u_{Rj}
+ g_{ij}^d \bar d_{Ri} \gamma^\mu \frac{\lambda^a}{2} d_{Rj} \right) \mathcal{G}_\mu^a$ & $g=g^\dagger$ \\[1mm]
$\mathcal{H}_\mu$ & $(8,3)_0$
& $- g_{ij} \bar q_{Li} \gamma^\mu \tau^I \frac{\lambda^a}{2} q_{Lj} \, \mathcal{H}_\mu^{aI}$ & $g=g^\dagger$ \\[1mm]
$\mathcal{G}_\mu^1$ & $(8,1)_1$
& $- g_{ij} \bar d_{Ri} \gamma^\mu \frac{\lambda^a}{2} u_{Rj} \, \mathcal{G}_\mu^{1a\dagger} + \text{h.c.}$ & -- \\[1mm]
$\mathcal{Q}_\mu^1$ & $(3,2)_{\frac{1}{6}}$
& $-g_{ij} \varepsilon_{abc} \bar d_{Rib} \gamma^\mu \epsilon q_{Ljc}^c \, \mathcal{Q}_\mu^{1a\dagger} + \text{h.c.}$ & -- \\[1mm]
$\mathcal{Q}_\mu^5$ & $(3,2)_{-\frac{5}{6}}$
& $-g_{ij} \varepsilon_{abc} \bar u_{Rib} \gamma^\mu \epsilon q_{Ljc}^c \, \mathcal{Q}_\mu^{5a\dagger} + \text{h.c.}$ & -- \\[1mm]
$\mathcal{Y}_\mu^1$ & $(\bar 6,2)_{\frac{1}{6}}$
& $-g_{ij} \textstyle \frac{1}{2} \left[ \bar d_{Ria} \gamma^\mu \epsilon q_{Ljb}^c +
\bar d_{Rib} \gamma^\mu \epsilon q_{Lja}^c \right] \mathcal{Y}_\mu^{1ab\dagger} + \text{h.c.}$ & -- \\[1mm]
$\mathcal{Y}_\mu^5$ & $(\bar 6,2)_{-\frac{5}{6}}$
& $-g_{ij} \textstyle \frac{1}{2} \left[ \bar u_{Ria} \gamma^\mu \epsilon q_{Ljb}^c +
\bar u_{Rib} \gamma^\mu \epsilon q_{Lja}^c \right] \mathcal{Y}_\mu^{5ab\dagger} + \text{h.c.}$ & -- \\[1mm]
$\phi$ & $(1,2)_{-\frac{1}{2}}$
& $- g_{ij}^u \bar q_{Li} u_{Rj} \, \phi - g_{ij}^d \bar q_{Li} d_{Rj} \, \tilde \phi + \text{h.c.}$ & -- \\[1mm]
$\Phi$ & $(8,2)_{-\frac{1}{2}}$
& $- g_{ij}^u \bar q_{Li} \frac{\lambda^a}{2} u_{Rj} \, \Phi^a - g_{ij}^d \bar q_{Li} \frac{\lambda^a}{2} d_{Rj} \, \tilde \Phi^a + \text{h.c.}$ & -- \\[1mm]
$\omega^1$ & $(3,1)_{-\frac{1}{3}}$
& $- g_{ij} \varepsilon_{abc} \bar d_{Rib} u_{Rjc}^c \, \omega^{1a\dagger} + \text{h.c.}$ & -- \\[1mm]
$\Omega^1$ & $(\bar 6,1)_{-\frac{1}{3}}$
& $-g_{ij} \textstyle \frac{1}{2} \left[ \bar d_{Ria} u_{Rjb}^c +
\bar d_{Rib} u_{Rja}^c \right] \Omega^{1ab\dagger} + \text{h.c.}$ & -- \\[1mm]
$\omega^4$ & $(3,1)_{-\frac{4}{3}}$
& $- g_{ij} \varepsilon_{abc} \bar u_{Rib} u_{Rjc}^c \, \omega^{4a\dagger} + \text{h.c.}$ & $g=-g^T$ \\[1mm]
$\Omega^4$ & $(\bar 6,1)_{-\frac{4}{3}}$
& $-g_{ij} \textstyle \frac{1}{2} \left[ \bar u_{Ria} u_{Rjb}^c +
\bar u_{Rib} u_{Rja}^c \right] \Omega^{4ab\dagger} + \text{h.c.}$ & $g=g^T$ \\[1mm]
$\sigma$ & $(3,3)_{-\frac{1}{3}}$
& $- g_{ij} \varepsilon_{abc} \bar q_{Lib} \tau^I \epsilon q_{Ljc}^c \, \sigma^{a\dagger} + \text{h.c.}$ & $g=-g^T$ \\[1mm]
$\Sigma$ & $(\bar 6,3)_{-\frac{1}{3}}$
& $-g_{ij} \textstyle \frac{1}{2} \left[ \bar q_{Lia} \tau^I \epsilon q_{Ljb}^c +
\bar q_{Lib} \tau^I \epsilon q_{Lja}^c \right] \Sigma^{Iab\dagger} + \text{h.c.}$ & $g=g^T$
\end{tabular}
\end{center}
\caption{Vector bosons and scalar representations mediating $u \bar u,d \bar d \to t \bar t$.}
\label{tab:lagr}
\end{table}
\begin{table}[p]
\begin{center}
\begin{small}
\begin{tabular}{c|ccccccc}
& $C_{qq}^{3113}$ & $C_{qq'}^{1133}$ & $C_{uu}^{3113}$
& $C_{ud'}^{3311}$ & $C_{qu}^{1331}$ & $C_{qu}^{3113}$
& $C_{qd}^{3113}$ \\[1mm]
\hline \\[-4mm]
$\mathcal{B}_\mu$ & $-|g_{13}^q|^2$ & -- & $-|g_{13}^u|^2$ & -- & -- & -- & -- \\[1mm]
$\mathcal{W}_\mu$ & $|g_{13}|^2$ & $-2 |g_{13}|^2$ & -- & -- & -- & -- & -- \\[1mm]
$\mathcal{G}_\mu$ & $\textstyle \frac{1}{6} |g_{13}^q|^2$ & $-\textstyle \frac{1}{2} g_{11}^q g_{33}^q$
& \!\!\begin{tabular}{c}$\textstyle \frac{1}{6} |g_{13}^u|^2$ \\ $-\textstyle \frac{1}{2} g_{11}^u g_{33}^u$
\end{tabular}\!\!
& $-\textstyle \frac{1}{4} g_{33}^u g_{11}^d$ & $\textstyle \frac{1}{2} g_{11}^q g_{33}^u$ & $\textstyle \frac{1}{2} g_{33}^q g_{11}^u$ & $\textstyle \frac{1}{2} g_{33}^q g_{11}^d$ \\[1mm]
$\mathcal{H}_\mu$
& \!\!\begin{tabular}{c}$-\textstyle \frac{1}{6} |g_{13}|^2$ \\ $- g_{11} g_{33}$
\end{tabular}\!\!
& \!\!\begin{tabular}{c}$\textstyle \frac{1}{3} |g_{13}|^2$ \\ $+\textstyle \frac{1}{2} g_{11} g_{33}$
\end{tabular}\!\!
& -- & -- & -- & -- & -- \\[1mm]
$\mathcal{B}_\mu^1$ & -- & -- & -- & $-\textstyle \frac{1}{2}|g_{13}|^2$ & -- & -- & -- \\[1mm]
$\mathcal{G}_\mu^1$ & -- & -- & -- & $\textstyle \frac{1}{12} |g_{13}|^2$ & -- & -- & -- \\[1mm]
$\mathcal{Q}_\mu^1$ & -- & -- & -- & -- & -- & -- & $|g_{13}|^2$ \\[1mm]
$\mathcal{Q}_\mu^5$ & -- & -- & -- & -- & $|g_{31}|^2$ & $|g_{13}|^2$ & -- \\[1mm]
$\mathcal{Y}_\mu^1$ & -- & -- & -- & -- & -- & -- & $-\textstyle \frac{1}{2} |g_{13}|^2$ \\[1mm]
$\mathcal{Y}_\mu^5$ & -- & -- & -- & -- & $-\textstyle \frac{1}{2}|g_{31}|^2$ & $-\textstyle \frac{1}{2}|g_{13}|^2$\\[1mm]
$\phi$ & -- & -- & -- & -- & $\textstyle \frac{1}{2} |g_{13}^u|^2$ & $\textstyle \frac{1}{2} |g_{31}^u|^2$ & $\textstyle \frac{1}{2} |g_{31}^d|^2$\\[1mm]
$\Phi$ & -- & -- & -- & -- & $-\textstyle \frac{1}{12} |g_{13}^u|^2$ & $-\textstyle \frac{1}{12} |g_{31}^u|^2$ & $-\textstyle \frac{1}{12} |g_{31}^d|^2$\\[1mm]
$\omega^1$ & -- & -- & -- & $-\textstyle \frac{1}{4} |g_{13}|^2$ & -- & -- & -- \\[1mm]
$\Omega^1$ & -- & -- & -- & $\textstyle \frac{1}{8} |g_{13}|^2$ & -- & -- & -- \\[1mm]
$\omega^4$ & -- & -- & $-2|g_{13}|^2$ & -- & -- & -- & -- \\[1mm]
$\Omega^4$ & -- & -- & $|g_{13}|^2$ & -- & -- & -- & -- \\[1mm]
$\sigma$ & $-2|g_{13}|^2$ & $-2|g_{13}|^2$ & -- & -- & -- & -- & -- \\[1mm]
$\Sigma$ & $|g_{13}|^2$ & $|g_{13}|^2$ & -- & -- & -- & -- & --
\end{tabular}
\end{small}
\end{center}
\caption{Coefficients of effective operators interfering with the SM amplitudes for $u \bar u,d\bar d \to t \bar t$. The new physics scale $\Lambda$ equals the mass of the new particle or multiplet.}
\label{tab:CAint}
\end{table}
\begin{table}[p]
\begin{center}
\begin{tabular}{c|ccccccc}
& $C_{qq}^{1133}$ & $C_{qq'}^{3113}$ & $C_{uu}^{1133}$
& $C_{ud}^{3311}$ \\[1mm]
\hline \\[-4mm]
$\mathcal{B}_\mu$ & $-g_{11}^q g_{33}^q$ & -- & $-g_{11}^u g_{33}^u$ & $-\textstyle \frac{1}{2} g_{33}^u g_{11}^d$ \\[1mm]
$\mathcal{W}_\mu$ & $g_{11}^q g_{33}^q$ & $-2 g_{11}^q g_{33}^q$ & -- & -- \\[1mm]
$\mathcal{G}_\mu$ & $\textstyle \frac{1}{6} g_{11}^q g_{33}^q$ & $-\textstyle \frac{1}{2} |g_{13}^q|^2$
& \!\!\begin{tabular}{c}$\textstyle \frac{1}{6} g_{11}^u g_{33}^u$ \\ $-\textstyle \frac{1}{2} |g_{13}^u|^2$
\end{tabular}
& $\textstyle \frac{1}{12} g_{33}^u g_{11}^d$ \\[1mm]
$\mathcal{H}_\mu$
& \!\!\begin{tabular}{c}$-\textstyle \frac{1}{6} g_{11} g_{33}$ \\ $- |g_{13}|^2$
\end{tabular}
& \!\!\begin{tabular}{c}$\textstyle \frac{1}{2} |g_{13}|^2$ \\ $+\textstyle \frac{1}{3} g_{11} g_{33}$
\end{tabular}
& -- & -- \\[1mm]
$\mathcal{G}_\mu^1$ & -- & -- & -- & $-\textstyle \frac{1}{4} |g_{13}|^2$ \\[1mm]
$\omega^1$ & -- & -- & -- & $\textstyle \frac{1}{4} |g_{13}|^2$ \\[1mm]
$\Omega^1$ & -- & -- & -- & $\textstyle \frac{1}{8} |g_{13}|^2$ \\[1mm]
$\omega^4$ & -- & -- & $2|g_{13}|^2$ & -- \\[1mm]
$\Omega^4$ & -- & -- & $|g_{13}|^2$ & -- \\[1mm]
$\sigma$ & $2|g_{13}|^2$ & $2|g_{13}|^2$ & -- & -- \\[1mm]
$\Sigma$ & $|g_{13}|^2$ & $|g_{13}|^2$ & -- & --
\end{tabular}
\end{center}
\caption{Coefficients of $\bar L L \bar L L$ and $\bar R R \bar R R$ effective operators contributing to $u \bar u,d\bar d \to t \bar t$ only at quadratic level. For the representations not listed all these coefficients vanish. The new physics scale $\Lambda$ equals the mass of the new particle or multiplet.}
\label{tab:CA1}
\end{table}
\begin{table}[p]
\begin{center}
\begin{tabular}{c|ccccccc}
& $C_{qu}^{3311}$ & $C_{qu'}^{1331}$ & $C_{qu'}^{3113}$ & $C_{qu'}^{3311}$
& $C_{qd'}^{3113}$ \\[1mm]
\hline \\[-4mm]
$\mathcal{B}_\mu$ & -- & $g_{11}^q g_{33}^u$ & $g_{33}^q g_{11}^u$ & $2 g_{13}^{q*} g_{13}^u$ & $g_{33}^q g_{11}^d$ \\[1mm]
$\mathcal{G}_\mu$ & $g_{13}^{q*} g_{13}^u$ & $-\textstyle \frac{1}{6} g_{11}^q g_{33}^u$ & $-\textstyle \frac{1}{6} g_{33}^q g_{11}^u$ & $-\textstyle \frac{1}{3} g_{13}^{q*} g_{13}^u$ & $-\textstyle \frac{1}{6} g_{33}^q g_{11}^d$ \\[1mm]
$\mathcal{Q}_\mu^1$ & -- & -- & -- & -- & $-|g_{13}|^2$ \\[1mm]
$\mathcal{Q}_\mu^5$ & $2 g_{13} g_{31}^*$ & $-|g_{31}|^2$ & $-|g_{13}|^2$ &
$-2 g_{13} g_{31}^*$ & -- \\[1mm]
$\mathcal{Y}_\mu^1$ & -- & -- & -- & -- & $-\textstyle \frac{1}{2} |g_{13}|^2$ \\[1mm]
$\mathcal{Y}_\mu^5$ & $- g_{13} g_{31}^*$ & $-\textstyle \frac{1}{2} |g_{31}|^2$ & $-\textstyle \frac{1}{2} |g_{13}|^2$ &
$-g_{13} g_{31}^*$ & -- \\[1mm]
$\phi$ & $g_{11}^{u*} g_{33}^u$ & -- & -- & -- & -- \\[1mm]
$\Phi$ & $-\textstyle \frac{1}{6} g_{11}^{u*} g_{33}^u$ & $\textstyle \frac{1}{4} |g_{13}^u|^2$ & $\textstyle \frac{1}{4} |g_{31}^u|^2$ & $\textstyle \frac{1}{2} g_{11}^{u*} g_{33}^u$ & $\textstyle \frac{1}{4} |g_{31}^d|^2$ \\[1mm]
\end{tabular}
\end{center}
\caption{Coefficients of $\bar L R \bar R L$ effective operators contributing to $u \bar u,d\bar d \to t \bar t$ only at quadratic level. For the representations not listed all these coefficients vanish. The new physics scale $\Lambda$ equals the mass of the new particle or multiplet.}
\label{tab:CA2}
\end{table}
\begin{table}[htb]
\begin{center}
\begin{tabular}{c|ccccccc}
& $C_{qq\epsilon}^{1331}$ & $C_{qq\epsilon}^{3311}$
& $C_{qq\epsilon'}^{1331}$ & $C_{qq\epsilon'}^{3311}$ \\
\hline \\[-4mm]
$\phi$ & $g_{13}^u g_{31}^d$ & $g_{33}^u g_{11}^d$ & -- & -- \\[1mm]
$\Phi$ & $-\textstyle \frac{1}{6} g_{13}^u g_{31}^d$ & $-\textstyle \frac{1}{6} g_{33}^u g_{11}^d$ & $\textstyle \frac{1}{2} g_{13}^u g_{31}^d$ & $\textstyle \frac{1}{2} g_{33}^u g_{11}^d$ \\[1mm]
\end{tabular}
\end{center}
\caption{Coefficients of $\bar L R \bar L R$ effective operators contributing to $u \bar u,d\bar d \to t \bar t$ only at quadratic level. For the representations not listed all these coefficients vanish. The new physics scale $\Lambda$ equals the mass of the new particle or multiplet.}
\label{tab:CA3}
\end{table}
The four-fermion operators contributing to $t\bar t$ production, including those that do not interfere with the SM QCD amplitude and only appear at quadratic level, have been given in Ref.~\cite{AguilarSaavedra:2010zi}. (The operators which interfere with the SM were given in Ref.~\cite{Jung:2009pi}.)
We collect in Tables~\ref{tab:CAint}--\ref{tab:CA3} the values of the corresponding coefficients for all the vector and scalar irreducible representations that can induce these operators. For vector bosons the coefficients have previously been obtained in Ref.~\cite{delAguila:2010mx}. The notation for four-fermion operators is given in appendix~\ref{sec:a}. We stress again that including
both interference $1/\Lambda^2$ and quadratic $1/\Lambda^4$ terms in our calculations is not inconsistent, despite the fact that we are not considering dimension-eight operators. When quadratic terms are relevant for $t \bar t$ production (large couplings) the missing dimension-eight terms are sub-leading in the classes of SM extensions we consider. A related discussion about the importance of $1/\Lambda^2$ and $1/\Lambda^4$ contributions from dimension-six operators has been presented in Ref.~\cite{AguilarSaavedra:2010sq}.
We evaluate the FB asymmetry
\begin{equation}
A_\text{FB} = \frac{\sigma^F-\sigma^B}{\sigma^F+\sigma^B} =
\frac{\sigma^F_\text{SM}+ \delta \sigma^F -\sigma_\text{SM}^B- \delta \sigma^B}{\sigma^F_\text{SM}+\delta \sigma^F +\sigma_\text{SM}^B +\delta \sigma^B} \,,
\end{equation}
using the SM predictions~\cite{Campbell:1999ah}
\begin{align}
& A_\text{FB}^\text{SM} = 0.058 \pm 0.009 && \text{(inclusive)} \,, \notag \\
& A_\text{FB}^\text{SM} = 0.088 \pm 0.013 && (m_{t\bar t} > 450~\text{GeV}) \,.
\label{ec:afbSM}
\end{align}
and the new contributions from four-fermion operators $\delta \sigma^{F,B}$, parameterised in terms of effective operator coefficients and numerical constants. The explicit expressions are collected in appendix~\ref{sec:a}. It is important to point out that positive operator coefficients always increase $A_\text{FB}$ at first order ($1/\Lambda^2$ interference with the SM), as it follows from Eqs.~(\ref{ec:int}).
We choose, among the recently reported measurements of the FB asymmetry~\cite{Aaltonen:2011kc},
\begin{align}
& A_\text{FB}^\text{exp} = 0.158 \pm 0.075 && \text{(inclusive)} \,, \notag \\
& A_\text{FB}^\text{exp} = 0.475 \pm 0.114 && (m_{t\bar t} > 450~\text{GeV}) \,.
\label{ec:afbX}
\end{align}
the one for high $t \bar t$ invariant masses which exhibits the largest deviation ($3.4\sigma$) with respect to the SM prediction.
The total cross section at LHC is evaluated including four-fermion operators in a similar way. In order to display the effect of new contributions on the $t \bar t$ tail we evaluate the cross section for $t \bar t$ invariant masses larger than 1 TeV.
We note that our calculation in terms of effective operators gives a larger (smaller) tail than the exact calculation for $t$-channel ($s$-channel) resonances. In the former case the differences are not dramatic but in the latter our results can be quite conservative, depending on the mass and width of the new resonance. A detailed comparison is presented in appendix~\ref{sec:b}.
\begin{figure}[htb]
\begin{center}
\begin{tabular}{ccc}
\epsfig{file=Figs/bmu,height=4.8cm,clip=} & \quad &
\epsfig{file=Figs/gmu,height=4.8cm,clip=} \\
\epsfig{file=Figs/bg1,height=4.8cm,clip=} & \quad &
\epsfig{file=Figs/QY1,height=4.8cm,clip=} \\
\multicolumn{3}{c}{\epsfig{file=Figs/QY5,height=4.8cm,clip=}}
\end{tabular}
\caption{Allowed regions for the Tevatron $t \bar t$ asymmetry and the $t \bar t$ tail at LHC for a single vector boson in each representation.}
\label{fig:VB}
\end{center}
\end{figure}
The relation between the predictions for the Tevatron $t \bar t$ asymmetry and the $t \bar t$ tail at LHC is tested by performing a random scan over the relevant couplings $g_{ij}$ corresponding to each new particle or multiplet. The results for the ten vector boson representations are presented in Fig.~\ref{fig:VB}.
(For $\mathcal{B}_\mu^1$, $\mathcal{G}_\mu^1$, $\mathcal{Q}_\mu^1$ and $\mathcal{Y}_\mu^1$ the regions are one-dimensional because there is only one coupling involved.) There are several interesting conclusions which can be drawn from these plots:
\begin{enumerate}
\item For $\mathcal{W}_\mu$ and $\mathcal{H}_\mu$ the allowed regions are inside the corresponding ones for $\mathcal{B}_\mu$, $\mathcal{G}_\mu$, respectively. This is expected because the interactions of the former correspond to a particular case of the latter, with only left-handed couplings.
\item For new colour-singlet neutral bosons $\mathcal{B}_\mu$, $\mathcal{W}_\mu$ the linear terms have negative coefficients and decrease $A_\text{FB}$, which can only reach the experimental value for large couplings when quadratic terms dominate. Hence,
accommodating a large asymmetry automatically implies a large $t \bar t$ tail.
For example, for $A_\text{FB} \gtrsim 0.3$ the enhancement is more than a factor of five. This implies that these possible explanations for $A_\text{FB}$ can be probed, and eventually excluded, with the luminosity collected in the 2010 run of LHC. The same conclusion applies to the vector bosons $\mathcal{G}_\mu^1$ and $\mathcal{B}_\mu^1$, as they only contribute in $d \bar d$ initial states and require a huge coupling to produce a large asymmetry.
\item For colour-octet isosinglet bosons $\mathcal{G}_\mu$ it is possible to have a large $A_\text{FB}$ and still a moderate tail at LHC. This model provides an example of cancellation of linear $1/\Lambda^2$ terms in the cross section, provided that $g_{ii}^q = - g_{ii}^u = - g_{ii}^d$, {\em i.e.} the vector boson couples as an axigluon. However, in order to have positive coefficients in the interference terms the couplings for the third and first generation must have opposite sign. This does not happen for the isotriplet boson $\mathcal{H}_\mu$ which only has left-handed couplings, and for these SM extensions the predicted $t \bar t$ tail is large.
\item Another interesting candidate is a colour triplet $\mathcal{Q}_\mu^5$, which has positive coefficients in interference terms as well. Its inclusion gives some enhancement to the asymmetry, which can reach the experimental value with the further addition of a scalar. Note that quadratic terms from operators $O_{qu^{(')}}$ decrease the asymmetry, as it can be derived from Eqs.~(\ref{ec:quad}) and is clearly seen in Fig.~\ref{fig:VB}. Hence, this model is interesting only for moderate couplings.
\item The rest of vector bosons, $\mathcal{Q}_\mu^1$, $\mathcal{Y}_\mu^1$ and $\mathcal{Y}_\mu^5$ have little interest for the $t \bar t$ asymmetry because they do not allow for a value appreciably larger than the SM prediction.
\end{enumerate}
\begin{figure}[t]
\begin{center}
\begin{tabular}{ccc}
\epsfig{file=Figs/phi,height=4.8cm,clip=} & \quad &
\epsfig{file=Figs/sigma,height=4.8cm,clip=} \\
\epsfig{file=Figs/Om1,height=4.8cm,clip=} & \quad &
\epsfig{file=Figs/Om4,height=4.8cm,clip=}
\end{tabular}
\caption{Allowed regions for the Tevatron $t \bar t$ asymmetry and the $t \bar t$ tail at LHC for a single scalar in each representation.}
\label{fig:sc}
\end{center}
\end{figure}
Aside from these remarks, we also note that for (i) $\mathcal{B}_\mu^1$ and $\mathcal{G}_\mu^1$; (ii) $\mathcal{Q}_\mu^1$ and $\mathcal{Y}_\mu^1$; (iii) $\mathcal{Q}_\mu^5$ and $\mathcal{Y}_\mu^5$, the linear $1/\Lambda^2$ terms have opposite sign, which explains the behaviour observed in the plots for these vector bosons.
The results for the eight scalar representations are presented in Fig.~\ref{fig:sc}.
Except for the isodoublets $\phi$, $\Phi$, the regions are one-dimensional because there is only one coupling involved. We point out that:
\begin{enumerate}
\item A colour-singlet isodoublet $\phi$ (with the same quantum numbers as the Higgs boson) can give an asymmetry compatible with the experimental value, and still produce a moderate tail at LHC. (Note that quadratic terms involving the operators $O_{qu^{(')}}$, $O_{qd^{(')}}$ decrease the asymmetry.) On the other hand, a colour octet $\Phi$ produces an asymmetry smaller than the SM value because the interference terms have opposite sign.
\item For colour-sextets $\Omega^4$ and $\Sigma$ the interference terms increase the asymmetry because the operator coefficients are positive.\footnote{In Ref.~\cite{Shu:2009xf} the SM and colour-sextet contributions have a wrong relative sign, resulting in a decrease of the cross section at first order. The sign has been corrected in Refs.~\cite{Arhrib:2009hu,Ligeti:2011vt}.}
However, producing an asymmetry $A_\text{FB} \gtrsim 0.3$ requires large couplings $g_{13}$ and implies a large $t \bar t$ tail at LHC, which might be already excluded. For the colour triplets $\omega^4$, $\sigma$ the situation is worse because the interference terms have negative operator coefficients, and even larger couplings are required to produce a large asymmetry.
\item For the two scalars $\Omega^1$, $\omega^1$ which only contribute in $d \bar d \to t \bar t$, the deviations in $A_\text{FB}$ are always very small.
\end{enumerate}
To conclude this survey, we remark again that this correlation between $A_\text{FB}$ and the $t \bar t$ tail at LHC applies to SM extensions with a single vector boson or scalar (as many of the ones proposed in the literature) but when more than one particle is present the contributions can add up or cancel, making it easier to fit the experimental data
and predict moderate effects in the $t \bar t$ tail, as we discuss in detail in the next section.
\section{A large $t \bar t$ asymmetry with a small $t \bar t$ tail}
\label{sec:3}
By inspection of Eqs.~(\ref{ec:int}) it is clear that the cancellation of the linear $1/\Lambda^2$ contributions to the cross section takes place provided
\begin{align}
& C_{qq'}^{1133} + C_{qq}^{3113} + C_{uu}^{3113} = C_{qu}^{1331} + C_{qu}^{3113} \equiv c_1 \,, \notag \\
& C_{qq'}^{1133} + 2 C_{ud'}^{3311} = C_{qu}^{1331} + C_{qd}^{3113} \equiv c_2
\label{ec:cancel}
\end{align}
(see also Ref.~\cite{Degrande:2010kt}).
Notice that in the left-hand side of both equations we have $LL$ and $RR$ couplings, whereas on the right-hand side we have $LR$ and $RL$ ones. As we have mentioned, one simple example where both equalities are fulfilled is an axigluon with flavour-diagonal couplings $g_{ii}^q = -g_{ii}^u = -g_{ii}^d$, with the additional requirement that first and third generation couplings have opposite sign, to have positive coefficients. (This model may be excluded by low-energy measurements, however~\cite{Chivukula:2010fk}.) But there are many other possibilities which can be constructed combining particles in Table~\ref{tab:CAint}, for instance, a colour triplet $\mathcal{Q}_\mu^5$ together with a colour sextet $\Omega^4$ or $\Sigma$.
In these SM extensions with vanishing (or very small) contributions to the total cross section at first order, the FB asymmetry is
\begin{equation}
A_\text{FB} = A_\text{FB}^\text{SM} + \frac{2 c_1 (D_\text{int}^F-D_\text{int}^B)_{u \bar u}
+ 2 c_2 (D_\text{int}^F-D_\text{int}^B)_{d \bar d}}{\sigma_\text{SM}}
\end{equation}
plus smaller corrections from quadratic terms, which depend on the specific operators which yield $c_1$ and $c_2$.
Remarkably, one can obtain a good fit to both asymmetry measurements in Eqs.~(\ref{ec:afbX}) with values of $c_1$, $c_2$ of order unity.\footnote{With two parameters at hand we can fit the exact central values of the two measurements, but this requires huge values of the constants $c_1$, $c_2$.} For instance, assuming $c_1=c_2$ and equal $LL$, $RR$, $LR$ and $RL$ terms, the best fit to both measurements is $c_{1,2}=2$, for which
\begin{align}
& A_\text{FB}^{4F} = 0.225 && \text{(inclusive)} \,, \notag \\
& A_\text{FB}^{4F} = 0.366 && (m_{t\bar t} > 450~\text{GeV})
\label{ec:afb4F}
\end{align}
including linear and quadratic terms, with a $\chi^2$ of 1.72. For $c_2=0$ the best fit is found for $c_1 = 2.34$, giving similar predictions for the asymmetry.
We have investigated the effects on the $t \bar t$ tails by implementing four-fermion operators in the generator {\tt Protos}~\cite{AguilarSaavedra:2008gt}. We have first checked that this class of SM extensions does not produce a too large tail at Tevatron. Figure~\ref{fig:mtt-tev} shows the $t \bar t$ invariant mass distributions for the SM and with $LL+RR$, $RL+LR$ four-fermion contributions corresponding to $c_1=c_2=2$, which yield the asymmetries in Eqs.~(\ref{ec:afb4F}). Above $m_{t \bar t} = 700$ GeV the cross section is enhanced by $+56\%$, which is consistent with data~\cite{Aaltonen:2011kc}.
\begin{figure}[t]
\begin{center}
\begin{tabular}{ccc}
\epsfig{file=Figs/mtt-tev,height=4.8cm,clip=} & \quad &
\epsfig{file=Figs/mttL-tev,height=4.8cm,clip=}
\end{tabular}
\caption{Invariant mass distribution for $t \bar t$ pairs at Tevatron, for the SM and with four-fermion contributions predicting the FB asymmetries in Eqs.~(\ref{ec:afb4F}). The plot on the left panel has linear scale whereas for the one in the right panel it is logarithmic.}
\label{fig:mtt-tev}
\end{center}
\end{figure}
\begin{figure}[t]
\begin{center}
\begin{tabular}{ccc}
\epsfig{file=Figs/mtt-lhc,height=4.8cm,clip=} & \quad &
\epsfig{file=Figs/mttL-lhc,height=4.8cm,clip=}
\end{tabular}
\caption{Invariant mass distribution for $t \bar t$ pairs at LHC, for the SM and with four-fermion contributions predicting the FB asymmetries in Eqs.~(\ref{ec:afb4F}). The plot on the left panel has linear scale whereas for the one in the right panel it is logarithmic.}
\label{fig:mtt-lhc}
\end{center}
\end{figure}
For LHC the invariant mass distributions are presented in Fig.~\ref{fig:mtt-lhc}. The cross section above 1 TeV is a factor of $2.3$ above the SM one. This deviation could be visible with the luminosity to be collected in 2011, provided that the systematic uncertainties (jet energy scale, jet energy resolution, $b$ jet energy scale, etc.) are low enough.
Here it is necessary to mention that for Tevatron a smaller efficiency for $t$-channel new physics at high $m_{t \bar t}$ has been recently claimed~\cite{Gresham:2011pa,Jung:2011zv}, which could help maintain the cross section at the high $m_{t \bar t}$ bins in agreement with measurements while reproducing the FB asymmetry. For LHC the efficiency decrease at $m_{t \bar t}>1$ TeV is not significant because of the larger detector coverage up to $|\eta| = 2.5$ for charged leptons and $|\eta| = 5$ for jets (see appendix~\ref{sec:b}).
\section{Conclusions}
\label{sec:4}
If the $t \bar t$ asymmetry measured at Tevatron corresponds to new physics, this new physics should also manifest at the $t \bar t$ tail at LHC. The size of the effect of course depends on the new physics itself which gives rise to the FB asymmetry, and it can serve to discriminate among different explanations. These issues have been investigated here using an effective operator framework and classifying all possible new vector bosons and scalars by their transformation properties under the SM gauge group. Particular models in the literature attempting to explain the observed asymmetry often fall into one of these classes.
For models which reproduce $A_\text{FB}$ with $t$- and/or $s$-channel $Z'$ exchange~\cite{Jung:2009jz,Cao:2009uz,Choudhury:2010cd,Cao:2011ew,
Berger:2011ua,Gresham:2011pa}
we have found that the tail above 1 TeV should be enhanced by a factor of five at LHC, at least for $Z'$ bosons heavier than 1 TeV. With $\sigma(m_{t \bar t} > 1~\text{TeV}) = 1.22$ pb at the tree level and a luminosity of approximately 35 pb$^{-1}$ collected in the 2010 run, the SM predicts 6.3 events in the semileptonic decay channel, not including the detector acceptance nor efficiencies. Then, it seems likely that an enhancement by a factor of five in the tail could be excluded just by analysing 2010 data. The same argument applies to $t$-channel $W'$ exchange~\cite{Cheung:2009ch,Cao:2010zb,Cheung:2011qa,Gresham:2011pa} or a mixture of both~\cite{Barger:2011ih}.
More exotic models explain the observed asymmetry by the exchange of a colour-triplet isosinglet scalar $\omega^4$ (see for example Refs.~\cite{Shu:2009xf,Arhrib:2009hu,Choudhury:2010cd,Ligeti:2011vt,Gresham:2011pa}) or its colour-sextet counterpart $\Omega^4$~\cite{Shu:2009xf,Arhrib:2009hu,
Grinstein:2011yv,Patel:2011eh,Ligeti:2011vt,Gresham:2011pa}.
These models could also be tested, and eventually excluded, with data already analysed by CMS and ATLAS in the search for $t \bar t$ resonances.
On the other hand, models with axigluons~\cite{Ferrario:2008wm,Frampton:2009rk,Cao:2010zb,Bai:2011ed,Choudhury:2010cd} or other types of colour-octet bosons~\cite{Djouadi:2009nb,Chen:2010hm,Burdman:2010gr,
Alvarez:2010js,Grinstein:2011yv,Gresham:2011pa} can in principle accommodate the measured $t \bar t$ asymmetry predicting a moderate $t \bar t$ tail at LHC. (Our discussion obviously does not directly apply to models where the
new physics produces $t \bar t$ plus other particles in the final state, see for example Ref.~\cite{Isidori:2011dp}.)
Besides, we note that Ref.~\cite{Delaunay:2011gv} has recently found that SM extensions explaining the FB asymmetry without predicting too large $t \bar t$ tails must have interference with the SM amplitudes.\footnote{This reference has appeared in the arXiv one day before the present paper, and our findings are consistent with theirs, where they overlap.} As we have shown, our conclusions are stronger because in many extensions with interference this is not possible either. Moreover, for all vector bosons and scalars in Table~\ref{tab:lagr} there is interference unless the involved couplings vanish.
Finally, in this paper we have investigated the conditions under which the first order $1/\Lambda^2$ contributions to the total $t \bar t$ cross section cancel while still producing a FB asymmetry compatible with experimental data. (A previous study at this order has been presented in Ref.~\cite{Degrande:2010kt}.)
Clearly, in this situation the tails of the $t \bar t$ invariant mass distribution are much smaller, both at Tevatron and LHC, as it has been shown explictly. A popular example of a model fulfilling these conditions is an axigluon with opposite couplings to the first and third generation, but there are many other possibilities which can be worked out from Table~\ref{tab:CAint}. All these SM extensions can accommodate the Tevatron $t \bar t$ asymmetry and cross section with small couplings, and predict a moderate enhancement of the $t \bar t$ tail at Tevatron and LHC, which is not in contradiction with experiment.
Interestingly, a small excess in the $t \bar t$ tail with boosted tops has been already observed at Tevatron~\cite{cdf10234}. In any case, these possible departures will soon be tested with forthcoming LHC data.
\section*{Acknowledgements}
This work has been partially supported by projects FPA2010-17915 (MICINN), FQM 101 and FQM 437 (Junta de Andaluc\'{\i}a) and CERN/FP/116397/2010 (FCT).
|
\section*{Introduction}
Microtubules (MT) are involved in
key processes of cell functions such as mitosis, cell morphogenesis and motility.
The building blocks of microtubules are $\alpha \beta$-tubulin heterodimers
which can associate either laterally or longitudinally \cite{Desai_Mitchison_MT:97}.
In biological systems, microtubules display unusual non-equilibrium dynamic behaviors, which are relevant for cell functioning. One such behavior, termed treadmilling involves a flux of subunits from one polymer end to the other, and is created by a difference of critical concentrations of the two ends \cite{wilson:1998}. In another behavior, termed dynamic instability, microtubules undergo alternating phases of elongation and rapid shortening \cite{mitchison:1984}. The two behaviors, treadmilling and dynamic instability result from an interplay between the polymerization and the GTP hydrolysis.
The cap model provides a simple explanation for the dynamic instability: a growing microtubule is stabilized by a cap of unhydrolyzed units at its extremity, and when this cap is lost, the microtubule undergoes a sudden change to the shrinkage state, a so called catastrophe.
The transitions between growth and shrinking can be described
by a two state model with prescribed stochastic transitions \cite{Bayley:89,hill:84}.
This model has lead to a number of theoretical and experimental studies \cite{Verde:1992,Leibler:93,Leibler-cap:96}, which have shown in particular the existence of a phase boundary between a bounded growth and an unbounded growth regimes.
Although many features of microtubules dynamics can be captured in this way, this model remains phenomenological, because of the unknown dependence of the transition rates as function of external factors, such as tubulin concentration or temperature.
To go beyond phenomenological models, one needs to account for the main chemical
reactions occurring at the level of a single monomer \cite{margolin:2006,Wolynes:06}.
These reactions can be assumed to occur between discrete states, and the corresponding
transition rates can be observed experimentally. In this way, discrete models can be constructed,
which capture remarkably well the main dynamical features of
single actin or single microtubule filaments \cite{kolomeisky:06,Antal-etal-PRE:07,Ranjith2009,Ranjith2010}.
These discrete models have the additional advantage of being free from some of the
limitations inherent to continuous models.
The question of the precise mechanism of hydrolysis in microtubules or actin has been
controversial for many years despite decades of experimental work.
In the vectorial model, hydrolysis occurs only at the unique interface between units bound to GTP/ATP and units bound to GDP/ADP, while
in the random model, hydrolysis can occur on any unhydrolyzed unit of the filament leading to a multiplicity of interfaces at a given time.
Between these two limits, models with an arbitrary level of cooperativity in the hydrolysis have been considered (see for instance \cite{wegner-1996,kierfeld-2010} for actin and \cite{Leibler-cap:96} for microtubules). The idea that the filament dynamics depends on the mechanism of hydrolysis in its interior or more generally on the internal
structure of the filament has been recently emphasized and it has been given the name of structural plasticity \cite{mitchison:2009}. As a practical recent illustration of that idea, the dynamical properties of microtubules can be tuned by incorporating in them GDP-tubulin in a controlled way \cite{valiron:2010}.
In microtubules, many experimental facts point towards a
mechanism of hydrolysis which is non-vectorial but random or cooperative.
Studies of the statistics of catastrophes \cite{jason-dogterom:03,Walker-1988,voter:1991} already provided
hints about this, but there are now more direct evidences.
The observation of GTP-tubulin remnants inside a microtubule using a specific antibody \cite{perez:2008} is probably one
of the most compelling evidences.
With the development of microfluidic devices for biochemical applications,
similar experiments probing the internal structure and the dynamics of single bio-filaments are becoming more and more accessible. Furthermore, it is now possible to record the dynamics of microtubule plus-ends at nanometer resolution \cite{schek:2007,kerssemakers:2006},
thus allowing essentially to detect the addition and departure of single tubulin dimers from microtubule ends.
In view of all these recent developments,
there is a clear need to organize all this information on
microtubules dynamics with a theoretical model.
Here, we propose a simple one dimensional non-equilibrium model,
accounting for the hydrolysis occurring within the filament.
We show that this model successfully explains
for many known experimental observations with microtubules such as:
the cap size, the dependence of the catastrophe time versus monomer concentration and the delay before a catastrophe following a dilution \cite{jason-dogterom:03,Walker-1988,voter:1991}.
Our interpretation of this data confirms and goes beyond results obtained in a recent numerical and theoretical study of the dynamic instability of MT \cite{Brun-2009}.
In vivo, the dynamics of microtubules is controlled by a variety of binding proteins,
which typically modify the polymerization process. Here we
focus on the physical principles which control the dynamic instability
of microtubules in vitro in the absence of any microtubule associated proteins.
Our model differs from previous attempts to address this problem, in that it is sufficiently simple to
be analytically solvable to a large extend, while still capturing the main features of MT dynamics.
\noindent
\section*{Model}
GTP hydrolysis is a two steps process: the first step,
the GTP cleavage produces GDP-Pi and is rapid, while the
second step, the release of the phosphate (Pi), leads to
GDP-tubulin and is by comparison much slower. This suggests that
many kinetic features of tubulin polymerization can be explained
by a simplified model of hydrolysis, which takes into account only the second
step of hydrolysis and treats tubulin subunits bound to GTP and tubulin subunits
bound to GDP-Pi as a single specie \cite{kolomeisky:06,Ranjith2009,Ranjith2010}.
This is the assumption which we make here. Therefore what we mean by random hydrolysis here
is the random process of phosphate release, which as we argue,
controls the dynamic instability of microtubules.
Our second main assumption has to do with the neglect of the protofilament structure of microtubules. Protofilaments are likely to be strongly interacting and should experience mechanical stresses in the MT lattice. We agree that modeling these effects is important to provide a complete microscopic picture of the transition from the growing phase to the shrinking phase, since this transition should involve protofilament curling near the MT ends \cite{van-buren-2005,kulic-2010}.
Here, we do not account for such effects, because as in Ref.~\cite{Leibler-cap:96}, we are interested in constructing a minimal dynamic model for microtubules, which would describe in a coarse-grained way the main aspects of the dynamics of this polymer.
We also assume that the filament contains a single active end and is
in contact with a reservoir of subunits bound to GTP.
The parameters of the model are as in Refs.~\cite{kolomeisky:06,Ranjith2009,Ranjith2010}:
the rate of addition of subunits $U$,
the rate of loss of subunits bound to GTP, $W_T$, the rate of loss of
subunits bound to GDP, $W_D$, and finally the rate of GTP hydrolysis
$r$ assumed to occur randomly on any unhydrolyzed subunits within the filament.
In Fig. \ref{fig-sketch},
all these possible transitions have been depicted. We
have assumed that all the rates are independent of the
concentration of free GTP subunits $C$ except for the on-rate \cite{jason-dogterom:03},
which is $U=k_0 c$.
All the rates of this model have
been determined precisely experimentally except for $r$. The
values of these rates are given in table \ref{table-rates}.
\begin{figure}
\begin{center}
\includegraphics[scale=0.3]{fig1-pnas}
\caption{(a): Representation of the various elementary transitions considered in the model with their corresponding rates, $U$ the on-rate of GTP-subunits, $W_T$ the off-rate of GTP-subunits, $W_D$ the off-rate of GDP-subunits and $r$ the hydrolysis rate for each unhydrolyzed unit within the filament. (b) Pattern for a catastrophe with $N$ terminal units in the GDP state.
\label{fig-sketch} }
\end{center}
\end{figure}
\begin{table}
\caption{Various rates used in the model and corresponding references. The conditions are that of a low ionic strength buffer. The value used for the rate of hydrolysis result from the analysis of the present paper. \label{table-rates}}
\begin{tabular}{lcrr}
On-rate of T subunits at + end &$k_0$ ($\mu$M$^{-1}s^{-1}$)& 3.2 & \cite{howard-book} \cr
\hline
Off-rate of T subunits from + end & $W_T $($s^{-1}$) & 24 & \cite{jason-dogterom:03} \cr
\hline
Off-rate of D subunits from + end & $W_D$($s^{-1}$) & 290 & \cite{howard-book} \cr
\hline
Hydrolysis rate (random model) & $r$ (s$^{-1}$)& 0.2 & \cr
\hline
\end{tabular}
\end{table}
As a result of the random hydrolysis, a typical filament configuration contains
many islands of unhydrolyzed subunits within the filament.
The last island containing the terminal unit is called the cap.
\noindent
\section*{Results and discussion}
In this section, we obtain the nucleotide content of the filament within
a mean-field approximation (for earlier references on this model, see \cite{Ranjith2010,kolomeisky:06,wegner:86}).
We denote by $i$ the position of a monomer within the filament, from the terminal unit at $i=1$.
For a given configuration, we introduce for each subunit
$i$ an occupation number $\tau_i$, such that $\tau_i=1$ if the subunit is bound to GTP
and $\tau_i=0$ otherwise. In the reference frame associated with the end of the filament, the equations
for the average occupation number are for $i=1$,
\begin{equation} \frac{d \langle \tau_1 \rangle}{dt}=U (1- \langle \tau_1 \rangle ) - W_T \langle \tau_1 (1-\tau_2)
\rangle + W_D \langle \tau_2 (1-\tau_1) \rangle - r \langle \tau_1 \rangle, \label{recursion1} \end{equation}
and for $i >1$,
\begin{eqnarray} \frac{d \langle \tau_i \rangle}{dt} & = & U \langle \tau_{i-1} -\tau_i \rangle + W_T \langle \tau_1 (\tau_{i+1} - \tau_i ) \rangle \nonumber \\
& + & W_D \langle (1-\tau_1) ( \tau_{i+1} - \tau_i ) \rangle - r \langle \tau_i \rangle. \label{recursioni} \end{eqnarray}
In a mean-field approach, correlations are neglected, which means that for any $i,j$, $\langle \tau_i \tau_j \rangle$ is replaced by $\langle \tau_i \rangle \langle \tau_j \rangle$.
At steady state, the left-hand sides of Eqs.~\ref{recursion1}-\ref{recursioni} are both zero,
which leads to recursion relations for the $\langle \tau_i \rangle$.
Let us denote $\langle \tau_1 \rangle=q$ as the probability that the terminal
unit is bound to GTP. The recursion relations have a solution of the form for $i\geq1$,
\begin{equation} \frac{ \langle \tau_{i+1} \rangle}{\langle \tau_i \rangle}=b, \label{recursion} \end{equation}
where $b=(U-q(W_T+r))/(U-q W_T)$.
Combining Eqs.~\ref{recursion1}-\ref{recursion}, one obtains $q$ explicitly as function of all the rates
as the solution of a cubic equation which is given in the appendix of Ref.~\cite{Ranjith2010}.
The mean filament velocity (namely the average rate of change of the total filament length) is given by
\begin{equation}
v= \left( U - W_T q - W_D (1-q) \right) d,
\label{velocity}
\end{equation}
in terms of the monomer size $d$. At the critical concentration $c_c$, the mean velocity vanishes, which corresponds to the boundary between a phase of bounded growth for $c<c_c$ and a phase of unbounded growth for $c>c_c$ \cite{Ranjith2010}.
The plot of this velocity versus concentration exhibits a kink shape near the critical
concentration, which is not particularly sensitive to the mechanism of hydrolysis since it is present both in the vectorial and random model \cite{kolomeisky:06,Ranjith2010}. This kink is well known from studies with actin \cite{hill:85} but has not been studied experimentally with microtubules except in Ref.~\cite{carlier-hill-1984} in a specific medium containing glycerol.
The distribution of the nucleotide along the filament length has a well defined steady-state in the tip reference frame at arbitrary value of the monomer concentration $c$. Using Eq.~\ref{recursion}, it follows that $\langle \tau_i \rangle= b^{i-1} q$, and therefore, the steady-state probability that the cap has exactly a length $l$, $P_l$,
is $P_l=(\prod_{i=1}^l \langle \tau_i \rangle) (1-\langle \tau_{l+1} \rangle)$. This leads to the following expression:
\begin{equation}
P_l=b^{l(l-1)/2} q^l \left( 1 - b^l q \right),
\label{SS proba}
\end{equation}
and, the corresponding average cap size is :
\begin{equation}
\langle l \rangle = \sum_{l \ge 1} l P_l = \sum_{l \ge 1} b^{l(l-1)/2} q^l.
\end{equation}
In figure \ref{fig-cap}, we show how this average cap size varies as function of the free
tubulin concentration. The average cap becomes longer than approximatively one subunit above the critical concentration, $c_c$ defined above, and which is about 7$\mu$M for the parameters of table \ref{table-rates} used here. At concentrations significantly larger than this value, the cap grows more slowly, as $\sqrt{\pi U/2r}$ as $U \rightarrow \infty$ \cite{Antal-etal-PRE:07,Leibler-cap:96}. In the range of concentration [0:100 $\mu$M], the cap stays smaller than about 47 subunits, which represents only 3.6 layers (or 28 nm). This estimate indicates that the cap is below optical resolution in the range of tubulin concentration generally used, which could explain the difficulty for observing it experimentally.
\begin{figure}
\begin{center}
\includegraphics[scale=1]{fig2-pnas}
\caption{Average cap size in number of subunits as function of the free tubulin concentration $c$ in $\mu$M. The line is the mean-field analytical solution and the
filled squares are simulation points.
\label{fig-cap} }
\end{center}
\end{figure}
A long standing view in the literature is that the cap could be as small as a single layer, as shown by experiments based on a chemical detection of the phosphate release \cite{wilson:2002}.
This view has been recently challenged by two experiments, in which the length fluctuations of microtubules were probed at the nanoscale, \cite{schek:2007,kerssemakers:2006}. The interpretation of these experiments still generate debates \cite{howard:2009,odde:2008}. In any case, taken together these two experimental studies reported a highly variable MT plus-end growth behavior, which suggests that the cap size is a fluctuating quantity, larger than one layer but smaller than about 5 layers. We note that such a range is compatible with our prediction and agrees with the estimation obtained from dilution experiments \cite{voter:1991}. Furthermore, our stochastic model naturally incorporates a fluctuating cap size. Even if the cap is indeed below optical resolution, we note that this does not rule out the possibility that it could be observed with the technique of Ref.~\cite{perez:2008}.
In figure \ref{fig-cap}, we also compare the predictions of the mean-field approximation with an exact simulation of the dynamics. We find that mean-field theory provides an excellent approximation of the exact solution when the free tubulin concentration is above the critical concentration, which corresponds to the conditions of most experiments \cite{jason-dogterom:03,Walker-1988}. Deviations can be seen between the exact solution and its mean-field approximation in figure \ref{fig-cap} but only below the critical concentration.
Many other quantities of interest follow from the determination of the nucleotide content of a given subunit, namely $\langle \tau_i \rangle$, such as the length fluctuations of the filament \cite{Ranjith2010} or the islands distribution of hydrolyzed or non-hydrolyzed subunits \cite{Antal-etal-PRE:07,kierfeld-2010}. These predictions should prove particularly useful in testing this model against experiments, since the island distribution of unhydrolyzed units or "remnants" will become accessible in future experiments similar to that of \cite{perez:2008} but carried out in in vitro conditions.
\noindent
\subsection*{Frequency of catastrophes and rescues vs. concentration}
One difficulty in bridging the gap between a model of the dynamic instability
and experiments, lies in a proper definition of the event which is called a catastrophe, since
the number of reported catastrophes is affected by several factors depending on the experimental conditions, such as for instance the experimental resolution of the observation \cite{schek:2007}.
Although a catastrophe manifests itself experimentally as an abrupt reduction of the total filament length, we choose to define it from the nucleotide content of the terminal region.
Following closely Ref.~\cite{Brun-2009}, we define a shrinking configuration as one in which the last $N$ units of the filament are all in the GDP state (irrespective of the state of the other units) as shown in figure \ref{fig-sketch}.
The remaining configurations (with an unhydrolyzed cap of any size or when the number of hydrolyzed subunits at the end is less than $N$) are assumed to belong to the growing phase.
In such a two states description of the dynamics (with a growing and a shrinking phase),
which is implicitly assumed in the analysis of most experiments, the catastrophe frequency $f_c(N)$ is the inverse of the average time spent in the growing phase, while the rescue frequency $f_r(N)$ is the inverse of the average time spent in the shrinking phase. It follows from this that the catastrophe frequency $f_c(N)$ can be obtained as the probability flux out of the growing state divided by the probability to be in the growing state. For instance for $N=1$, this flux condition is
\begin{equation}
\label{catastrophe frequencyN=1}
f_c(1) q = (W_T+r) P_1 + r \sum_{j \geq 2} P_j,
\end{equation}
where the terms on the right proportional to $P_1$ correspond to a transition of the terminal unit from the GTP to the GDP state, which can occur either through hydrolysis or depolymerization of that unit, while the last term corresponds to hydrolysis of the terminal unit from cap states of length larger or equal than 2.
We have derived the general expression of $f_c(N)$ in the case of an arbitrary $N$ as shown in Supporting Information (SI) Methods, and we have checked these results by comparing them with stochastic simulations using the Gillespie algorithm \cite{gillespie:77}.
In the case of the vectorial model, the last term in Eq.~\ref{catastrophe frequencyN=1} is absent and the catastrophe frequency is non-zero only below the critical concentration. The fact that catastrophes are observed in \cite{jason-dogterom:03} significantly above the critical concentration indicates that this data is incompatible with a vectorial mechanism. For this reason, we only discuss here the predictions of the random model.
The catastrophe time $T_c(N)=1/f_c(N)$ is shown as function of growth velocity for $N=2$ in figure \ref{fig-N}a, and as a function of the concentration of free subunits, $c$, for $N=1$ in figure \ref{fig-N}b. The growth velocity is simply proportional to the concentration of free subunits. For both plots, one sees that below the critical concentration which is in the range of 5-10 $\mu$M, the catastrophe time is zero as expected since there is no stable filament in that region of concentration. Note that $T_c(N)$ behaves linearly as function of $c$ for $N=2$ but it behaves non-linearly for $N=1$. Since the experimental data of \cite{jason-dogterom:03} shows a linear dependence, this comparison indicates that the data can be explained with the model for $N=2$ but not for $N=1$. The same observation has been made in Ref.~\cite{Brun-2009}, where the same data has been analyzed. Note however, in comparing this work with this reference the following differences: first, the model of Ref.~\cite{Brun-2009} neglects rescues and assumes that the duration of a catastrophe once started is zero while the present model includes rescues, and takes into account the finite rate of loss of GDP units. Secondly the results of Ref.~\cite{Brun-2009} corresponds to the regime of high concentration of free subunits while the present model holds at any concentration even in the proximity or below the critical concentration. Thirdly the present approach leads to analytical results with the assumption that the filament has no protofilament structure while the results of Ref.~\cite{Brun-2009} are numerical but that model includes a protofilament structure.
Our analytical derivation of the catastrophe time confirms that the case $N=1$ differs in an essential way from the $N \geq 2$ case at high concentration. Indeed, the catastrophe time reaches a plateau when the concentration goes to infinity for $N=1$, while it goes to infinity for $N \geq 2$. This trend is already apparent in the figure \ref{fig-N}.
\begin{figure}
\begin{center}
\rotatebox{-90}{\includegraphics[scale=0.4]{fig3-pnas}}
\caption{(a): catastrophe time $T_c$ vs growth velocity for the $N=2$ case, with the theoretical prediction (solid line) together with simulation points ($\blacktriangle$) and experimental data points taken from \cite{jason-dogterom:03} for constrained growth ($\circ$) and free growth ($\bullet$). (b): catastrophe time $T_c$ vs free subunit concentration for the $N=1$ case, with the theoretical prediction (black solid line) and simulations (filled diamonds). In addition, the dash-dotted line and the dashed line represent respectively $\langle T(1) \rangle$ and $\langle T(20) \rangle$, which have been calculated using Eq.~\ref{full Tk}.
\label{fig-N} }
\end{center}
\end{figure}
In figure \ref{fig-N}, we have used a value for the rate of hydrolysis $r=0.2$, which is higher than that estimated in Ref.~\cite{Leibler-cap:96} (there the estimate was 0.002). The reason is that the hydrolysis rate is a global factor which controls the amplitude of the catastrophe time, basically $T_c(N)$ scales for an arbitrary $N$ as $1/r^N$. The value $r=0.002$ leads to a reasonable estimate for $T_c(N)$ for $N=1$ (albeit with the wrong dependence on concentration), but if we take seriously as we do here, the observation that only the definition with $N=2$ is compatible with the measured concentration dependence of the catastrophe time, then $r$ must have a significantly larger value than expected, and
$0.2$ is the value that is needed for $T_c$ in order to match the experimental data. Finally, we also note that the scaling of $T_c(N)$ as a power law of $r$ means that large values of $N$ (such as $N>2$) can be excluded given the observed range of catastrophe times.
We also show the distribution of catastrophe times calculated with the parameters given in table \ref{table-rates}, for $N=1$ and $N=2$ in figure \ref{fig:distrib-catastrophe2}. These distributions in both cases are essentially exponential (except at a very short time which is probably inaccessible in practice in the experiments), in agreement with the observations reported in Ref.~\cite{jason-dogterom:03} with free filaments.
\begin{figure}
\begin{center}
\includegraphics[scale=1.4]{fig4-pnas}
\caption{(left)The distribution of catastrophe time (N=1) for different concentration values. $C=9\mu M$(filled squares) and $C=12\mu M$(open circles). (right) The distribution of catastrophe time (N=2) for different concentration values $C=9\mu M$(filled squares) and $C=12\mu M$(open circles). The distributions are normalized.
\label{fig:distrib-catastrophe2}}
\end{center}
\end{figure}
One advantage of our microscopic model is that it can explain and predict different related aspects of the dynamic instability of microtubules. Specifically, it also allows to predict the statistics of rescue events when the polymer switches from the shrinking phase back into the growing phase. Assuming that the system reached a steady-state behavior, the frequency of rescues $f_{r}(N)$ can be calculated using flux conditions similar to the ones used to obtain $f_c(N)$ (see SI Methods for more details). The corresponding expression is rather simple and it can be written as
\begin{equation} \label{fr}
f_{r}(N)=U+W_{D} b^{N} q.
\end{equation}
We have carried out a complete numerical test of this frequency of rescues using stochastic simulations, which is shown in SI Fig.~1.
Our model predicts that rescue events should be observable under typical cellular conditions and in experiments. However, surprisingly there is a very limited experimental information on rescues. The analysis of Eq. (\ref{fr}) might shed some light on this issue. At low concentrations of GTP monomers in the solution, when the rate $U$ is small, the average time before the rescue event, $T_{r} \simeq 1/U$, might be very large. As a result, it might not be observable in experiments since the polymer with $L$ monomers could collapse faster ($T_{collapse} \simeq L/W_{D}$) before any rescue event could take place. At large $U$, rescues are more frequent given that the polymer is in the shrinking state. But the frequency of the catastrophes is very small under these conditions, the microtubule is almost always in the growing phase. Therefore in these conditions, rescues are not observed \cite{jason-dogterom:03}.
\noindent
\subsection*{First passage time of the cap and dilution experiments}
In dilution experiments, the concentration of free tubulin is abruptly reduced to a small value, resulting in catastrophes within seconds, independent of the initial concentration \cite{Walker-1991,voter:1991}. This observation is an evidence that the cap is short and independent of the initial concentration. The idea that the cap is short is also supported by the observation that cutting the end of a microtubule typically with a laser results in catastrophe. As we shall see below, all these well-known experimental facts about microtubules can be explained by the present model.
Here, we are interested in the time until the first catastrophe appears following the dilution. For simplicity, we take the definition of catastrophe introduced in the previous section for $N=1$, which means that a catastrophe starts as soon as the cap has disappeared (as shown in the previous section, one could extend this result to the more general case of an arbitrary $N$). Let us then introduce $F_k(t)$ the distribution of the first passage time $T_k$ for an initial condition corresponding to a cap of length $k$, and a filament in contact with a medium of arbitrary concentration.
As explained in SI Methods, it is possible to calculate analytically $F_k(t)$, by a method recently used in the context of polymer translocation \cite{PK-KironeJStatMech-2010}.
After numerically inverting the Laplace transform of $F_k(t)$,
one obtains the distribution $F_k(t)$ which is shown as solid lines in figure \ref{fig:DFPT} for the particular case of $k=2$. As can be seen in this figure, the predicted distributions agree very well with the results obtained from the stochastic simulation in this case.
\begin{figure}
\begin{center}
\rotatebox{-90}{\includegraphics[scale=0.4]{fig5-pnas}}
\caption{Distributions of the first passage time of the cap for an initial cap of $k=2$ units, $F_2(t)$ as function of the time $t$, for various initial concentration of free monomers. The solid lines are the theoretical predictions deduced from Eq.~(13) of the SI Methods after numerically inverting the Laplace transform, while the symbols are simulations. The circles correspond to a dilution into a medium with no free monomers, squares correspond similarly to a dilution into a medium with a concentration of free monomers of 2$\mu$M, diamonds to 5$\mu$M, and triangles to 9$\mu$M.
\label{fig:DFPT}}
\end{center}
\end{figure}
From the distribution $F_k(t)$ we obtain its first moment,
the mean first passage time of the cap $\langle T(k) \rangle$. As shown in SI Methods, we find that
\begin{equation}
\langle T(k) \rangle= \sum_{j=0}^{k-1} y^j \frac{J_{n+j+1}(\bar{y})}{\sqrt{U W_T} J_n(\bar{y})
- U J_{n+1}(\bar{y})},
\label{full Tk}
\end{equation}
where $y=\sqrt{W_T/U}$, $\bar{y}=2\sqrt{U W_T}/r$, $n=(U+W_T)/r$, and the functions $J_n(y)$ are Bessel functions.
The dependance of $\langle T(k) \rangle$ as a function of the initial size of the cap $k$ is shown in figure \ref{fig:Tk}: at small $k$, $\langle T(k) \rangle$ is essentially linear in $k$ as would be expected at all $k$ in the vectorial model of hydrolysis \cite{Ranjith2009}, while here it saturates at large values of $k$ (the value of this plateau can be calculated analytically but only for $U=0$ see SI Methods). To understand this saturation, consider a cap which is initially infinitively large, then after a time of order $1/r$, the cap abruptly becomes of a finite much smaller size as a result of the hydrolysis of one unit at a random position within the filament. This feature will always happen irrespective of the monomer concentration, and indeed in figure \ref{fig:Tk}, $\langle T(k) \rangle$ has a plateau for $k \rightarrow \infty$ for all values of the monomer concentration. We note that such a behavior of
$\langle T(k) \rangle$ as function of $k$ has similarities with the case of non-compact exploration investigated in \cite{condamin:2007}, while the vectorial model of hydrolysis would correspond in the language of this reference to the case of compact exploration.
\begin{figure}
\begin{center}
\rotatebox{-90}{\includegraphics[scale=0.4]{fig6-pnas}}
\caption{Mean first passage time $\langle T(k) \rangle$ as function of $k$ for three values of the monomer concentration from bottom to top 0, 2 and 4$\mu$M. The presence of steps in these curves is due to the fact that $\langle T(k) \rangle$ is only defined on integer values of $k$. Note the existence of a plateau for all values of the monomer concentration.
\label{fig:Tk}}
\end{center}
\end{figure}
Let us now turn to a practical use of this quantity for characterizing the dynamic instability.
In the previous section, we calculated the catastrophe time $T_c$. We expect that this quantity is an average of $\langle T(k) \rangle$, and indeed we find for the case of $N=1$ that $T_c$ is bounded by $\langle T(1) \rangle$ and $\langle T(20) \rangle$ (the choice of $20$ is purely illustrative) as shown in figure \ref{fig-N}.
The characteristic time observed in dilution experiments is another average of $\langle T(k) \rangle$. More precisely, let us denote $\langle T(k) \rangle_{post}$ as the first passage time in post-dilution conditions given that the initial length of the cap is $k$. The dilution time $T_{dilution}$ is then the average of $\langle T(k) \rangle_{post}$ with respect to the steady-state probability distribution of the initial conditions before the dilution occurs. In other words,
\begin{equation}
T_{dilution}=\sum_{k} \langle T(k) \rangle_{post} P_k({\rm predilution}),
\label{dilution}
\end{equation}
where $P_k(pre-dilution)$ is the stationary probability given in Eq.~\ref{SS proba} in pre-dilution conditions.
In the case that the final medium after dilution is very dilute, one can assume that the final free tubulin concentration is zero, which allows to simplify the general expression given in Eq.~\ref{full Tk} as explained in SI Methods.
Using Eq.~\ref{dilution}, one obtains the dilution time for the parameters of the table \ref{table-rates} which is shown in figure \ref{fig-dilution}.
\begin{figure}
\begin{center}
\includegraphics[scale=1]{fig7-pnas}
\caption{Dilution time (s) as function of free tubulin concentration (in $\mu$M) before dilution in the case that the post-dilution tubulin concentration is zero. Solid line is the mean-field prediction based on Eq.~\ref{dilution} and the symbols are simulation points. As found experimentally, the dilution time is essentially independent of the concentration of tubulin in the pre-dilution state, and the time to observe the first catastrophe is of the order of seconds or less.
\label{fig-dilution} }
\end{center}
\end{figure}
The figure confirms that the dilution time can be as short as a fraction of seconds in this case.
It is straightforward to extend this calculation to the case of an arbitrary value of the post-dilution medium ({\it i.e} for the case of a dilution of arbitrary strength) using the general expression derived in Eq.~\ref{full Tk}.
As the amplitude of the dilution is reduced (by increasing the post-dilution concentration), the dilution time increases as well but the general sigmoidal shape remains, with in particular a plateau at concentrations above the critical concentration. The presence of these plateaux means that the dilution time is essentially independent of the concentration of the monomers in pre-dilutions conditions as observed experimentally.
Note that the height of these plateaux scale with the hydrolysis rate. For instance, to explain the dilution times reported in \cite{Walker-1991}, one needs to use a smaller value of $r$ as given in the table because of the use of the $N=1$ definition of catastrophe. Alternatively, just as in the calculation of the catastrophe frequencies, it is possible to keep the expected large value of $r$ provided the $N=2$ definition of catastrophe is chosen. Thus, complementary information can be obtained from the catastrophe frequencies and the dilution times.
\noindent
\section*{Conclusion}
In this work, we have explained several important features about microtubules dynamics using a model for the random release of phosphate within the filament. The results of our mean-field approach are analytical to a large extend. With this approach we could recover some well known features of MT dynamics such as the mean catastrophe time and its distribution or the delays following a dilution, but we have also investigated much less studied aspects concerning the cap size, the role of the definition of catastrophes (via the parameter $N$) and the first passage time of the cap.
The theoretical model and ideas presented in this paper for the case of microtubules could also apply to other biofilaments such as actin or Par-M, for which the random hydrolysis model may be relevant as well. Furthermore, although the model describes a priori only single free filaments dynamics, it is also potentially useful for understanding constrained filaments,
in the broader context of force generation and force regulation by ensembles of biofilaments.
For this reason, it would be interesting to study extensions of the model to account for the various effects of MAPs on microtubules, which should shed light on the behavior of microtubules in more realistic biological conditions.
We hope that this theoretical work will stimulate further experimental and theoretical studies of these questions.
\begin{acknowledgments}
We thank F. Perez, F. Nedelec and M. F. Carlier for inspiring discussions. We also would like to thank K. Mallick for pointing to us Ref.~\cite{PK-KironeJStatMech-2010}, and M. Dogterom for providing us with the data of Ref.~\cite{jason-dogterom:03}. RP acknowledges
support through IYBA, from Department of Biotechnolgy, India.
\end{acknowledgments}
|
\section{Introduction}
In its most general sense, a (smooth) dynamical system, from the
geometrical point of view, is simply a vector field on some manifold.
But many dynamical systems of interest in physics and engineering
applications are more specialized than that:\ they are of second-order
type. By a dynamical system, or vector field, of second-order type,
or a second-order differential equation field, we mean a vector field
$\Gamma$ on the tangent bundle $\tau:T(Q)\to Q$ of some configuration
manifold $Q$ with the property that $\tau_{*y}\Gamma=y$ for all $y\in
T(Q)$, so that in terms of coordinates $(x^i,y^i)$ where the $y^i$ are
the canonical fibre coordinates corresponding to coordinates $x^i$ on
$Q$,
\[
\Gamma=y^i\vf{x^i}+\Gamma^i(x,y)\vf{y^i}.
\]
It is of interest therefore, to find criteria for determining whether
a given dynamical system, which may be represented in some arbitrary
coordinates, is actually of second-order type, in that coordinates may
be found with respect to which it takes the form above. This is a
problem which has both a local aspect, just described, and a global
one, which includes the question of whether the manifold on which the
dynamical system resides is in fact a tangent bundle.
In the recent article \cite{RR}, Ricardo and Respondek deal with a
version of the problem in the context of control theory. Assume $M$
to be an even dimensional manifold, not known to be the tangent
manifold of some configuration manifold. Under which conditions does
a coordinate change exist such that a given nonlinear control system
\[
\dot z^\alpha = F^\alpha(z) + u_r G^\alpha_r(z), \qquad
\alpha = 1,\ldots, 2n,
\]
is transformed into the form of a so-called `mechanical control
system', meaning a dynamical system of the form
\[
{\ddot q}^a = \conn abc(q) {\dot q}^b {\dot q}^c
+ P^a_b(q){\dot q}^b +Q^a(q) + u_r g^a_r(z), \qquad a=1, \ldots, n?
\] The solution of the problem in \cite{RR} is cast in
terms of a certain vector space $\mathcal{V}$, which is a subspace of the
infinite dimensional vector space of vector fields on $M$, with
dimension exactly the half of the dimension of the manifold, and
which, among other properties, contains the control forces $G_r$ and
satisfies $[\mathcal{V},\mathcal{V}]=0$.
In this paper, we will address a more general problem. First of all,
we will not assume that the dimension of the manifold is even. This
is mainly motivated by the observation that even for a system of
time-dependent second-order differential equations the manifold on
which the dynamics is described is odd-dimensional:\ it is the first
jet manifold of a spacetime manifold, or event space, fibred
over the real numbers (see e.g.\ \cite{CMS,MaPa}).
Further, we will allow some of the new coordinates to play simply the
role of parameters. That is to say, we will not require the number
$n$ of second-order equations to be exactly the half of the dimension
of the manifold $M$.
Next to extending the results of \cite{RR} to a broader class of
manifolds, we will also make some conceptual modifications. In a
nutshell, the results of \cite{RR} claim that if $\mathcal{V}$ (which is a
vector space constructed from the given control forces $G_r$ and from
$F$) satisfies certain conditions the vector field $F=F^\alpha
\partial/\partial z^\alpha$ (the so-called drift vector field)
transforms into an appropriate coordinate form, and, as a side-effect,
so does also the controlled dynamical field $F+u_rG_r$. We will take
the space $\mathcal{V}$ to be the primary given object of our study, and ignore
that it was constructed from some given control forces. Consequently,
we shift the attention to specific coordinate expressions for $F$, and
leave the control system given by the vector field $F+u_rG_r$ out of
the picture all together. A second deviation is that for us $\mathcal{V}$ will
not be a vector space of vector fields, but rather the distribution it
generates.
In Section~2 we investigate under what conditions a given vector field
$F$ can be transformed into the coordinate expression of a
second-order differential equation field with possible parameters, in
the presence of an arbitrary involutive distribution $\mathcal{V}$ (of
arbitrary dimension). Our framework has the advantage that it leaves
open the possibility that the transformed dynamics become either
autonomous or time-dependent. We show in Section~3 how one can
associate various
connections to $F$, and we argue that these connections provide a
coordinate-independent method to express that the dynamics of $F$ is
of quadratic type (or of mechanical type, in the sense as above) in a
yet unknown set of coordinates.
Working with a distribution $\mathcal{V}$ has the further advantage that it
brings an associated almost tangent structure (and almost jet
structure) to the foreground. These geometric structures find their
equivalence in standard tangent bundle and jet bundle geometry, but
they went unnoticed in \cite{RR}. Based on results in
\cite{CT,DeF,MP2}
we further address in Section~4 the global issues that arise in this
context, such as e.g.\ the affine fibre bundle structure of $M$ and
the relation of $F$ to second-order differential equation fields on a
certain tangent or jet manifold.
In the last section we illustrate the theory in the context of a
Lagrangian system with an Abelian symmetry group, where a second-order
differential equation field with multiple parameters naturally shows
up.
\section{Local coordinate transformations}
Let $M$ be a manifold of dimension $m$ and $\mathcal{V}$ an involutive
distribution on $M$ of dimension $n$ such that $2n\leq m$.
We must be a little careful here about the meaning of the term
distribution and related terms. A distribution on $M$ is of course a
choice of subspace of $T_z(M)$ at each $z\in M$, of constant
dimension, depending smoothly on $z$ in the sense that it admits local
smooth bases. There is a related, but distinct, concept, which we may
call a vector field system. A vector field system $\mathcal{S}$ on $M$ is a
collection of (smooth) vector fields on $M$ which is a $C^\infty(M)$
submodule of $\vectorfields{M}$, the module of vector fields on $M$.
For each $z\in M$, we denote by $\dim_z(\mathcal{S})$ the dimension of the
subspace of $T_z(M)$ spanned by the values at $z$ of the vector fields
in $\mathcal{S}$. Now $\dim_z(\mathcal{S})$ need not be constant. However, if vector
fields $X_a$ are linearly independent at $z$ they are linearly
independent in a neighbourhood of $z$, which means that
$\dim_{z'}(\mathcal{S})\geq\dim_z(\mathcal{S})$ for all $z'$ in a neighbourhood of $z$.
Moreover, $\dim_z(\mathcal{S})$ has a maximal value on $M$, which we call the
maximal dimension of $\mathcal{S}$, and the set of points at which the maximal
dimension is attained is an open subset of $M$.
A vector field system $\mathcal{S}$ on $M$ restricts to a vector field system
$\mathcal{S}|_U$ on any open subset $U$ of $M$, considered as a submanifold.
On the other hand, any vector field in $\mathcal{S}_U$ may be extended to a
vector field in $\mathcal{S}$ by multiplying it by a bump function whose
support is contained in $U$ (taking advantage of the fact that we are
working in the $C^\infty$ category). So it is permissible to discuss
local aspects of vector field systems in coordinates.
An alternative definition of the term distribution is that a
distribution is a vector field system $\mathcal{S}$ for which $\dim_z(\mathcal{S})$ is
constant, or for which the maximal dimension is attained everywhere.
We assume that $\mathcal{V}$, mentioned in the opening sentence of the section,
is a distribution in the strict sense. Now suppose that we have a
vector field $F$ on $M$ not belonging to $\mathcal{V}$. We denote by
$\mathcal{V}+[F,\mathcal{V}]$ the collection of vector fields on $M$ which may be
written in the form $V_1+[F,V_2]$ with $V_1,V_2 \in \mathcal{V}$. This is a vector
field system, essentially because for any $f \in C^\infty(M)$ and $V
\in \mathcal{V}$, $f[F,V]=[F,fV] \pmod\mathcal{V}$. We will be concerned with this
vector field system, for a given involutive distribution $\mathcal{V}$ and
vector field $F$, throughout this paper.
\begin{prop}\label{propo1}
Suppose that the vector field $F$ is such that $[F,\mathcal{V}]\cap\mathcal{V}=\{0\}$,
that is, if $V\in\mathcal{V}$ and $[F,V]\in\mathcal{V}$ then $V=0$. Then the maximal
dimension of $\mathcal{V}+[F,\mathcal{V}]$ is $2n$, and the open subset of $M$ on which
it is attained is dense, that is, its closure is $M$.
\end{prop}
\begin{proof}
We denote the vector field system $\mathcal{V}+[F,\mathcal{V}]$ by $\mathcal{W}$ for convenience.
Clearly the maximal dimension of $\mathcal{W}$ is at most $2n$, and the set of
points $z$ where $\dim_z(\mathcal{W})=2n$ is open, though it may be empty.
Suppose that $z$ is a point of $M$ with $\dim_z(\mathcal{W})<2n$. We show that
there can be no open neighbourhood of $z$ such that $\dim_{z'}(\mathcal{W})<2n$
for all $z'$ in the neighbourhood.
Since $\mathcal{V}$ is involutive there is a coordinate neighbourhood $U$ of
$z$ and coordinates $(q^a,y^i)$ with $a=1,2,\ldots,m-n$ and
$i=1,2,\ldots,n$ such that the coordinate fields $\partial/\partial
y^i$ span $\mathcal{V}|_U$. We may set
\[
F|_U=f^a\vf{q^a}+f^i\vf{y^i}
\]
for some smooth functions $f^a,f^i$ on $U$, so that
\[
\left[F|_U,\vf{y^i}\right]=-\fpd{f^a}{y^i}\vf{q^a} \pmod{\mathcal{V}},
\]
and for any $V\in\mathcal{V}_U$, with $V=V^i\partial/\partial y^i$,
\[
[F|_U,V]=-V^i\fpd{f^a}{y^i}\vf{q^a} \pmod{\mathcal{V}}.
\]
Now since $m-n\geq n$ the rank of the matrix $(\partial f^a/\partial
y^i)$ at any point is at most $n$. If it is $n$ at $z$ then
\[
V^i(z)\fpd{f^a}{y^i}(z)=0\quad\Longrightarrow\quad V^i(z)=0,
\]
the vector fields $[F|_U,\partial/\partial y^i]$ are linearly
independent at $z$, and $\dim_z(\mathcal{W})=2n$. So if $\dim_{z'}(\mathcal{W})<2n$ for
all $z'$ in a neighbourhood of $z$, which we can take to be a
coordinate neighbourhood as above, then the rank of the matrix
$(\partial f^a/\partial y^i)(z')$ is less than $n$, and we can find
functions $V^i$ on a neighbourhood $U$ of $z$, not all vanishing, such
that
\[
V^i\fpd{f^a}{y^i}=0.
\]
Then the vector field $V=V^i\partial/\partial y^i$ on $U$ satisfies
$[F|_U,V]\in\mathcal{V}$. So (by multiplying by a suitable bump function) we
can find a vector field $V'$ on $M$, not identically zero, with
$V'\in\mathcal{V}$, such that $[F,V']\in\mathcal{V}$, which is a contradiction. So
every neighbourhood of a point $z$ where $\dim_z(\mathcal{W})<2n$ must contain a
point $z'$ where $\dim_{z'}(\mathcal{W})=2n$. Thus the set of points $z$ where
$\dim_z(\mathcal{W})=2n$ is nonempty and open, and its closure is $M$; that is,
the set of points where $\dim_z(\mathcal{W})=2n$ is an open dense subset of
$M$.
\end{proof}
From now on we will assume that the vector field $F$ does indeed
satisfy the condition of Proposition~\ref{propo1}, and we will
restrict our attention to the open subset where $\dim_z(\mathcal{W})=2n$, that
is to say, we will effectively assume that $\mathcal{W}$ is a distribution. We
will make the further assumption that the distribution $\mathcal{W}$ is also
involutive.
We will work locally for the rest of this section, and drop explicit
notational reference to the neighbourhood on which we are working.
In the proof of Proposition~\ref{propo1} we showed that if $\{V_i\}$
is a local basis of $\mathcal{V}$ consisting of coordinate fields
$\partial/\partial y^i$ of a local coordinate system $(q^a,y^i)$ and
we set $W_i=[F,V_i]$ then $\{V_i,W_i\}$ is a local basis for $\mathcal{W}$.
Indeed, this will be true for any local basis $\{V_i\}$ of $\mathcal{V}$. If
we change basis to $\tilde{V}_i=A_i^jV_j$ (where the $A_i^j$ are
locally defined smooth functions and $(A_i^j)$ is nonsingular) then
$W_i$ changes to $\tilde{W}_i=A_i^jW_j+F(A_i^j)V_j$. If the basis
$\{V_i\}$ is such that $[V_i,V_j]=0$, so that the $V_i$ are coordinate
fields, the necessary and sufficient condition for $\{\tilde{V}_i\}$
also to satisfy $[\tilde{V}_i,\tilde{V}_j]=0$ is that
$A_i^lV_l(A^k_j)=A_j^lV_l(A^k_i)$.
Our next aim is to show that, under the assumptions stated earlier,
one can choose a commuting basis $\{V_i\}$ for $\mathcal{V}$ such that it and
the corresponding $W_i$ satisfy $[V_i,W_j]\in\mathcal{V}$. For any basis
$\{V_i\}$ we can write
\[
[V_i,W_j]=\alpha^k_{ij}V_k+\beta^k_{ij}W_k.
\]
Notice that if $[V_i,V_j]=0$ then both coefficients are symmetric in
their lower indices:
\[
0=[F,[V_i,V_j]]=[W_i,V_j]+[V_i,W_j]=[V_i,W_j]-[V_j,W_i].
\]
If we change the basis of $\mathcal{V}$ to $\tilde{V}_i=A_i^jV_j$ we have
\begin{align*}
[\tilde{V}_i,\tilde{W}_j]&=[A_i^kV_k,A_j^lW_l+F(A^l_j)V_l]\\
&=(A_i^lV_l(A_j^k)+A_i^lA_j^m\beta^k_{lm})W_k \pmod\mathcal{V}.
\end{align*}
So to make $[\tilde{V}_i,\tilde{W}_j]\in\mathcal{V}$ we want to choose $A^j_i$
such that $V_l(A_j^k)+A_j^m\beta^k_{lm}=0$. Note that since
$\beta^k_{ij}$ is symmetric, we will then have
$A_i^lV_l(A_j^k)=A_j^lV_l(A_i^k)$, and if the $V_i$ pairwise commute
then the $\tilde{V}_i$ will also pairwise commute.
The equations
\[
\fpd{A_j^k}{y^l}+A_j^m\beta^k_{lm}=0
\]
are linear first-order partial differential equations for the
unknowns $A^i_j$, and admit solutions if and only if their
integrability conditions, which are
\[
\fpd{\beta_{jk}^l}{y^i}-\fpd{\beta_{ik}^l}{y^j}
+\beta_{im}^l\beta_{jk}^m-\beta_{jm}^l\beta_{ik}^m=0,
\]
are satisfied. Now
\[
0=[[V_i,V_j],W_k]=[[V_i,W_k],V_j]+[V_i,[V_j,W_k]],
\]
and
\[
[V_i,[V_j,W_k]]=(V_i(\beta_{jk}^l)+\beta_{im}^l\beta_{jk}^m)W_l \pmod\mathcal{V}.
\]
So it follows from the identity $[V_i,[V_j,W_k]]-[V_j,[V_i,W_k]]=0$
that the integrability conditions are indeed satisfied. If we take a
solution $A^i_j$ for which the matrix $(A^i_j)$ is nonsingular on a
local cross-section of the $\mathcal{V}$ foliation, for example by taking
$A^i_j=\delta^i_j$ there, then $(A^i_j)$ will be nonsingular on an
open subset containing the cross-section. We have shown the
following.
\begin{prop}\label{beta}
If both $\mathcal{V}$ and $\mathcal{W}$ are involutive, there is a commuting basis
$\{V_i\}$ of $\mathcal{V}$ such that for all $i,j$, $[V_i,W_j]\in\mathcal{V}$ (where
$W_i=[F,V_i]$).
\end{prop}
The remaining freedom in the choice of commuting basis (such that
$[V_i,W_j]\in\mathcal{V}$ still holds) is to take $A^j_i$ to satisfy
$V_k(A_i^j)=0$.
The condition $[V_i,W_j]\in\mathcal{V}$ says that $W_j$ is invariant under the
action of $\mathcal{V}$, modulo $\mathcal{V}$.
Let us take a coordinate neighbourhood $U$ in $M$, with coordinates
$(q^a,y^i)$ such that $V_i=\partial/\partial y^i$; we may suppose
without essential loss of generality that $U$ is the image of a
product of open subsets $O\subset\mathbb{R}^{m-n}$ and $P\subset\mathbb{R}^n$, where
$0\in P$. Then $y^i=0$ is a submanifold of $U$ of codimension $n$,
say $N$, and $U$ is fibered over $N$ with fibres the integral
submanifolds of $\mathcal{V}$. Denote by $\pi:U\to N$ the corresponding
projection. Then the restriction of $W_j$ to $U$ is projectable to
$N$:\ that is to say, there is a well-defined vector field $\bar{W}_j$
on $N$ which is $\pi$-related to $W_j$. More generally, a vector
field $X\in\mathcal{W}$, say $X=X^iW_i\pmod{\mathcal{V}}$, is projectable if, and only
if, the coefficients $X^i$ satisfy $V(X^i)=0$ for all $V\in\mathcal{V}$, or
indeed if $V_j(X^i)=0$. Let us denote by $\bar{\mathcal{W}}$ the distribution
on $N$ spanned by the $\bar{W}_i$, in other words, the distribution
consisting of the projections of projectable vector fields in $\mathcal{W}$.
Then $\bar{\mathcal{W}}$ is involutive, since it $\pi$-related to the
involutive distribution $\mathcal{W}$. We may therefore choose coordinates
$(t^p,x^i)$ on $N$, where $p=1,2,\ldots,m-2n$, such that the integral
submanifolds of $\bar{\mathcal{W}}$ are given by $t^p=\mbox{constant}$. Then
with respect to the coordinates $(t^p,x^i,y^i)$ on $U$ we have
\[
V_i=\vf{y^i},\quad
W_i=W^j_i(x)\vf{x^j}\pmod{\mathcal{V}},
\]
where the coefficients $W^j_i$ are everywhere the components of a
nonsingular matrix. We still have at our disposal the freedom to
change the original basis to $\tilde{V}_i=A_i^j(t,x)V_j$. If we do so
with $A_i^kW_k^j=\delta_i^j$ then
\[
\tilde{V}_i=A_i^j\vf{y^j},\quad
\tilde{W}_i=\vf{x^i}\pmod{\mathcal{V}}.
\]
If we make a further change of coordinates to
\[
\tilde{t}^p=t^p,\quad\tilde{x}^i=-x^i,\quad
\tilde{y}^i=W^i_j(x)y^j,
\]
then
\[
\vf{\tilde{t}^p}=\vf{t^p}\pmod{\mathcal{V}},\quad
\vf{\tilde{x}^i}=-\vf{x^i}\pmod{\mathcal{V}},\quad
\vf{\tilde{y}^i}=A_i^j\vf{y^j}
\]
We have proved the following result.
\begin{prop}\label{coords}
With $\mathcal{V}$, $\mathcal{W}$, $V_i$, $W_i$ as above, we can find local coordinates
$(t^p,x^i,y^i)$ on $M$, $p=1,2,\ldots,m-2n$, $i=1,2,\ldots,n$ such
that
\[
V_i=\vf{y^i},\quad
W_i=-\vf{x^i}\pmod{\mathcal{V}}.
\]
\end{prop}
We move on now to investigate the form of $F$. There are two cases to
consider, depending on whether $F$ does or does not belong to $\mathcal{W}$.
\begin{thm} \label{sode} Assume both $\mathcal{V}$ and $\mathcal{W}$ are involutive.
\begin{enumerate}
\item Suppose that $F\in\mathcal{W}$, and assume that the set $N\subset
M=\{z\in M:F(z)\in\mathcal{V}\}$ is nonempty. Then we may choose coordinates
with respect to which
\[
F=y^i\vf{x^i}+F^i(t,x,y)\vf{y^i};
\]
that is to say, $F$ takes the form of a second-order differential
equation field in terms of the coordinates $(x^i,y^i)$, with the $t^p$
merely behaving as parameters.
\item Suppose that $F$ is everywhere independent of $\mathcal{W}$ (so that in
particular $M>2n$) and that $[F,\mathcal{W}]\subset\mathcal{W}$. Then we may choose
coordinates with respect to which
\[
F=\vf{t^1}+y^i\vf{x^i}+F^i(t,x,y)\vf{y^i};
\]
that is to say, $F$ takes the form of a time-dependent second-order
differential equation field in terms of the coordinates
$(t^1,x^i,y^i)$, with the $t^p$ with $p>1$ merely behaving as
parameters.
\end{enumerate}
\end{thm}
\begin{proof}
1.\ For the first case, set $F=a^iV_i+b^iW_i$ with respect to a frame
with $[V_i,V_j]=0$ and $[V_i,W_j]\in\mathcal{V}$. Then
\[
W_i=[F,V_i]=-V_i(a^j)V_j+b^j[W_j,V_i]-V_i(b^j)W_j,
\]
so we must have $V_i(b^j)=-\delta^j_i$. Now $N$ is the zero level set
of $(b^i)$, and the rank of the Jacobian of the map $M\to\mathbb{R}^n:z\mapsto
(b^i(z))$ is $n$, or in other words the 1-forms $db^i$ are
independent, since $(V_i(b^j))$ is nonsingular. So $N$ is an immersed
submanifold of $M$ of codimension $n$, and $\mathcal{V}$ is transverse to it.
We may choose coordinates $(t^p,x^i,y^i)$ as in
Proposition~\ref{coords}, such that $V_i=\partial/\partial y^i$, $N$
is given by $y^i=0$, $(t^p,x^i)$ are coordinates on $N$, and
\[
W_i=-\vf{x^i}\pmod{\mathcal{V}}.
\]
With respect to such coordinates set
\[
F=f^i(t,x,y)\vf{x^i}+F^i(t,x,y)\vf{y^i}.
\]
Then from its definition
\[
W_i=-\fpd{f^j}{y^i}\vf{x^j}\pmod{\mathcal{V}},
\]
and therefore
\[
\fpd{f^j}{y^i}=\delta^j_i.
\]
Taking into account the fact that $f^i(t,x,0)=0$ we have
$f^i(t,x,y)=y^i$.
2.\ For the second case, take coordinates as in
Proposition~\ref{coords}, and suppose that
\[
F=\varphi^p(t,x,y)\vf{t^p}\pmod{\mathcal{W}}.
\]
From the assumption that $[F,\mathcal{W}]\subset\mathcal{W}$ it follows that in fact
$\varphi^p$ depends only on the $t^q$. By assumption the $\varphi^p$
cannot vanish simultaneously, and so by a transformation of the
coordinates $t^p$ we may take
\[
F=\vf{t^1}+f^i(t,x,y)\vf{x^i}+F^i(t,x,y)\vf{y^i}.
\]
Arguing as above we see that
\[
\fpd{f^j}{y^i}=\delta^j_i.
\]
We may only conclude now that $f^i(t,x,y)=y^i+k^i(t,x)$. However, a
further coordinate transformation $y^i\mapsto y^i+k^i(t,x)$, with
$t^p$ and $x^i$ unchanged, leads to
\[
\vf{t^p}\mapsto\vf{t^p}\pmod{\mathcal{V}},\quad \vf{x^i}\mapsto\vf{x^i}\pmod{\mathcal{V}},
\]
with $\partial/\partial y^i$ unchanged, and so leads to the required
form for $F$.
\end{proof}
The remaining freedom in transforming the coordinates $(x^i,y^i)$ in
the first case, so as to preserve the form of $F$, is
\[
\tilde{x}^i=\tilde{x}^i(x),\quad \tilde{y}^i=\fpd{\tilde{x}^i}{x^j}y^j.
\]
That is to say, the $y^i$ transform like canonical fibre coordinates
on a tangent bundle. From this point of view it is natural to think
of the coordinates $y^i$ in use before the final transformation
leading to Proposition~\ref{coords} as quasi-velocities.
The remaining freedom in the second case is
\[
\tilde{x}^i=\tilde{x}^i(t^1,x),\quad
\tilde{y}^i=\fpd{\tilde{x}^i}{x^j}y^j+\fpd{x^i}{t^1}.
\]
Here the $y^i$ transform like the jet coordinates of the 1-jet bundle
of a manifold fibred over $\mathbb{R}$.
\section{Induced connections}
The coefficients $\beta^k_{ij}$ used in the proof of
Proposition~\ref{beta} have the appearance of the components of a
connection, and the integrability conditions quoted in the proof have
the form of the vanishing of the curvature of this connection. We
begin this section by explaining in what sense the $\beta^k_{ij}$ are
indeed the components of a flat symmetric connection.
Let $\mathcal{V}$ be an involutive distribution on any manifold $M$. For any
vector field $X$ on $M$ denote by $X+\mathcal{V}$ the equivalence class of $X$
modulo $\mathcal{V}$, that is, the collection of vector fields differing from
$X$ by an element of $\mathcal{V}$. For any $V\in\mathcal{V}$, set
\[
D_V(X+\mathcal{V})=[V,X]+\mathcal{V}.
\]
This is a well-defined operator on equivalence classes, which is
$\mathbb{R}$-linear in both arguments, and for $f\in C^\infty(M)$ satisfies
\[
D_{fV}(X+\mathcal{V})=fD_V(X+\mathcal{V}),\quad
D_V(fX+\mathcal{V})=fD_V(X+\mathcal{V})+V(f)(X+\mathcal{V}).
\]
That is to say, $D$ has connection-like properties. By the Jacobi
identity, for any $V_1,V_2\in\mathcal{V}$
\[
D_{V_1}D_{V_2}(X+\mathcal{V})-D_{V_2}D_{V_1}(X+\mathcal{V})-D_{[V_1,V_2]}(X+\mathcal{V})=0;
\]
that is to say, if $D$ were a connection it would have zero curvature.
More particularly, let $\mathcal{W}$ be another involutive distribution on $M$,
with $\mathcal{V}\subset\mathcal{W}$; then we may restrict $X$ in the construction above
to lie in $\mathcal{W}$. The same conclusions hold, mutatis mutandis. We may
think of $\mathcal{V}$ and $\mathcal{W}$ as vector sub-bundles of $T(M)$, and vector
fields in the distributions as sections of the corresponding bundles
$\mathcal{V}\to M$, $\mathcal{W}\to M$. If $W$ is a section of $\mathcal{W}\to M$ then we may
think of $W+\mathcal{V}$ as a section of the vector bundle $\mathcal{W}/\mathcal{V}\to M$. Then
(using the terminology of Lie algebroid theory) $D$
is a $\mathcal{V}$-connection on $\mathcal{W}/\mathcal{V}$.
Now take $\dim\mathcal{W}=2n$, $\dim\mathcal{V}=n$, and suppose there is a type $(1,1)$
tensor field $S$ on $\mathcal{W}$ (that is, a section of the bundle
$\mathcal{W}\otimes\mathcal{W}^*\to M$) with the algebraic properties of an almost
tangent structure (so that\ $\mathop{\mathrm{im}} S=\mathop{\mathrm{ker}} S$), with kernel $\mathcal{V}$. Then $S$
defines an isomorphism between sections of $\mathcal{W}/\mathcal{V}$ and sections of
$\mathcal{V}$. So we may define a $\mathcal{V}$-connection on $\mathcal{V}$, say $\nabla$, by
\[
\nabla_{V_1}V_2=S(D_{V_1}(W+\mathcal{V}))\quad
\mbox{for any $W\in\mathcal{W}$ such that $S(W)=V_2$.}
\]
That is, $\nabla_V S(W)=S(D_V(W+\mathcal{V}))=S([V,W])$. This is well-defined
as a $\mathcal{V}$-connection, and has vanishing curvature. Since $\nabla$ is
a $\mathcal{V}$-connection on $\mathcal{V}$, it makes sense to talk about its torsion.
But for any $W_1,W_2\in\mathcal{W}$,
\begin{eqnarray*}
\lefteqn{\nabla_{S(W_1)}S(W_2)-\nabla_{S(W_2)}S(W_1)-[S(W_1),S(W_2)]}\\
&=&S[S(W_1),W_2]-S[S(W_2),W_1]-[S(W_1),S(W_2)]\\
&=&-([S(W_1),S(W_2)]-S[S(W_1),W_2]-S[W_1,S(W_2)])\\
&=&-N_S(W_1,W_2).
\end{eqnarray*}
That is to say, the torsion vanishes if and only if the formal
Nijenhuis torsion $N_S$ of $S$ (a type $(2,1)$ $\mathcal{W}$-tensor) vanishes.
If $S$ has vanishing Nijenhuis torsion in this sense, and we restrict
attention to any leaf of the involutive distribution $\mathcal{V}$, we obtain a
flat symmetric connection there.
We now show how to construct such a $\mathcal{W}$-tensor $S$ in the case of
interest.
\begin{prop}\label{prop1} Assume both $\mathcal{V}$ and $\mathcal{W}$ are involutive.
There is a unique type $(1,1)$ $\mathcal{W}$-tensor field $S$ for which
\[
S(V) = 0 \qquad\mbox{and}\qquad S([F,V]) = -V, \qquad V\in\mathcal{V};
\]
it satisfies $\mathop{\mathrm{ker}} S=\mathop{\mathrm{im}} S =\mathcal{V}$ and $N_S=0$.
\end{prop}
\begin{proof}
Let $\{V_i\}$ be a basis of $\mathcal{V}$, and set $W_i=[F,V_i]$:\ then
$\{V_i,W_i\}$ is a basis for $\mathcal{W}$. So it is enough to know how $S$
acts on elements of the form $V$ and $[F,V]$. The definition above is
consistent:\ if $V \in\mathcal{V}$, then also $fV\in\mathcal{V}$ with $f$ a function on
${M}$, and $S([F,fV]) = S(F(f)V) + S(f[F,V]) = fS([F,V]) = -fV$. We
have $S^2=0$, $\mathop{\mathrm{ker}} S=\mathop{\mathrm{im}} S =\mathcal{V}$. The formal Nijenhuis torsion $N_S$
obviously vanishes for two elements in $\mathcal{V}$. Moreover, for any
$V_1,V_2\in\mathcal{V}$,
\[
N_S(V_1,[F,V_2]) = - S[V_1,V_2] = 0
\]
because of the assumed integrability of $\mathcal{V}$. Likewise, by making use
of the Jacobi identity (and because $[V_1,V_2]\in\mathcal{V}$),
\begin{align*}
N_S([F,V_1],[F,V_2]) &= [V_1,V_2] +S[V_1,[F,V_2]] + S [[F,V_1],V_2]\\
&= [V_1,V_2] + S [F,[V_1,V_2]] = 0.\qedhere
\end{align*}
\end{proof}
In the case of interest, where $S$ is as defined above, the
$\beta^k_{ij}$ are the connection coefficients of this connection with
respect to the basis $\{V_i\}$.
Suppose we have a further distribution $\mathcal{H}$ on $M$, of dimension $n$,
contained in $\mathcal{W}$, and everywhere transverse to $\mathcal{V}$; in other words a
complement to $\mathcal{V}$ in $\mathcal{W}$. We call such a distribution horizontal.
Then the restriction of $S$ to $\mathcal{H}$ is a $C^\infty(M)$-isomorphism
$\mathcal{H}\to\mathcal{V}$. For any $V\in\mathcal{V}$, denote by $\hlift{V}$ the unique element
of $\mathcal{H}$ such that $S(\hlift{V})=V$.
We can extend the $\mathcal{V}$-connection $\nabla$ on $\mathcal{V}$ to a
$\mathcal{W}$-connection on $\mathcal{V}$ as follows:\ for any $W\in\mathcal{W}$ and any $V\in\mathcal{V}$
set
\[
\nabla_WV=P_\mathcal{V}([P_\mathcal{H}(W),V])+S([P_\mathcal{V}(W),\hlift{V}]),
\]
where $P_\mathcal{H}$ and $P_\mathcal{V}$ are the projectors on $\mathcal{H}$ and $\mathcal{V}$,
respectively. The right-hand side belongs to $\mathcal{V}$ and depends
$\mathbb{R}$-linearly on the arguments. For $f\in C^\infty(M)$
\begin{align*}
\nabla_{fW}V&=P_\mathcal{V}([fP_\mathcal{H}(W),V])+S([fP_\mathcal{V}(W),\hlift{V}])\\
&=f\nabla_WV-V(f)P_V(P_\mathcal{H}(W))-\hlift{V}(f)S(P_\mathcal{V}(W))\\
&=f\nabla_WV
\end{align*}
while
\begin{align*}
\nabla_W(fV)&=P_\mathcal{V}([P_\mathcal{H}(W),fV])+S([P_\mathcal{V}(W),f\hlift{V}])\\
&=f\nabla_WV+P_\mathcal{H}(W)(f)P_\mathcal{V}(V)+P_\mathcal{V}(W)(f)S(\hlift{V})\\
&=f\nabla_WV+(P_\mathcal{H}(W)+P_\mathcal{V}(W))(f)V\\
&=f\nabla_WV+W(f)V.
\end{align*}
So $\nabla$ is a covariant derivative.
If $W\in\mathcal{V}$, say $W=V_1$, then the new definition gives
$\nabla_{V_1}V_2=S([V_1,\hlift{V_2}])$. According to the old definition,
$\nabla_{V_1}V_2=S([V_1,W])$ for any $W$ such that $S(W)=V_2$. But
$W=\hlift{V_2}$ is such that $S(W)=V_2$; so the two definitions agree in
this case. On the other hand, suppose that $W\in\mathcal{H}$ and that $W$ is
projectable in the sense that $[W,\mathcal{V}]\subset\mathcal{V}$ (the horizontal
projection of any projectable vector field is projectable, and the
$W_i$ are projectable as we pointed out before). Then
$\nabla_WV=[W,V]$.
Assuming as before that $[F,\mathcal{W}]\subset\mathcal{W}$ (which is automatically the
case if $F\in\mathcal{W}$, and is an assumption in Part 2 of Theorem~\ref{sode}
if not), it is possible to define a Lie derivative by $F$ of
$\mathcal{W}$-tensors:\ for example, in the case of a type $(1,1)$ $\mathcal{W}$-tensor
$T$ as the commutator of operators $\mathop{\mathrm{ad}}\nolimits F$ and $T$:
\[
(\lie{F}T)(W)=[F,T(W)]-T[F,W].
\]
\begin{prop} If $[F,\mathcal{W}]\subset\mathcal{W}$, the vector field $F$ defines a
complement $\mathcal{H}$ of $\mathcal{V}$ in $\mathcal{W}$.
\end{prop}
\begin{proof}
We show now that, with the above definition, $\lie{F} S$ defines two
projection operators on $\mathcal{W}$. We first show that $(\lie{F}S)^2 =
\mathop{\mathrm{id}}$. We have, for $V\in \mathcal{V}$,
\[
(\lie{F} S) (V) = [F,S(V)] - S[F,V] = V
\]
and thus $(\lie F S)^2(V)=V$. Also,
\[
(\lie F S) ([F,V]) = [F,S[F,V]] - S[F,[F,V]] = - [F,V] -S[F,[F,V]].
\]
Since $S[F,[F,V]]\in\mathcal{V}$, we have $((\lie{F} S) (S[F,[F,V]]) =
S[F,[F,V]] $, and therefore
\[
(\lie{F} S)^2 ([F,V]) = - (\lie{F} S) ([F,V]) - S[F,[F,V]] = [F,V].
\]
The conclusion is that $P_\mathcal{H} = {\textstyle\frac12} (\mathop{\mathrm{id}} - \lie{F} S )$ and $P_\mathcal{V} =
{\textstyle\frac12} (\mathop{\mathrm{id}} + \lie{F} S )$ are complementary projection operators,
with e.g.\ $P_\mathcal{V}(V)=V$ and $P_\mathcal{H}(V)=0$; $\mathcal{H}=\mathop{\mathrm{im}} P_\mathcal{H}$ is therefore a
complement to $\mathcal{V}$ in $\mathcal{W}$.
\end{proof}
We will use this complement from now on.
In the case where $M$ is a tangent manifold $T(Q)$, we can take $\mathcal{V}$
to be the canonical vertical distribution, and in particular
$\mathcal{W}=\vectorfields{T(Q)}$. The connection with covariant derivative
$\nabla$ is then the Berwald connection associated to a system of
autonomous second-order differential equations, see e.g.\ \cite{Berw}
(taking into account the fact that the current connection is expressed
in terms of vertical vector fields rather than vector fields along the
tangent bundle projection). A similar construction exists for the
case where $M$ is the first jet manifold of a bundle $E\to\mathbb{R}$, and
where the second-order dynamics are time-dependent, see e.g.\
\cite{CMS,MS}.
The Berwald connection can be used to describe special classes of
second-order differential equation fields, such as the ones of
quadratic type we had encountered in the introduction.
The `mixed curvature' of the $\mathcal{V}$-connection $\nabla$ is the (1,2)
$\mathcal{V}$-tensor field $\theta$ given by
\[
\theta(V_1,V_2)V_3=
\nabla_{\hlift{V}_1}\nabla_{V_2}V_3 - \nabla_{V_2}\nabla_{\hlift{V}_1} V_3
- \nabla_{[\hlift{V}_1,V_2]}V_3.
\]
\begin{prop}
Let $\mathcal{V}$ and $\mathcal{W}$ both be involutive and $[F,\mathcal{W}]\subset\mathcal{W}$. The
necessary and sufficient condition for the existence of coordinates
in which $F$ takes the form of a quadratic second-order differential
equation field is that $\theta=0$.
\end{prop}
\begin{proof}
Let $\hat\mathcal{V}$ denote the set of $V\in \mathcal{V}$ for which the corresponding
$\hlift V$ is projectable, i.e.\ $V$ satisfies $[\hlift V, V_2] \in
\mathcal{V}$ for all $V_2\in\mathcal{V}$. This set defines a module over the
projectable functions on $M$ (those functions $f$ for which
$V_1(f)=0$, for all $V_1\in\mathcal{V}$). Alternatively, $V_2\in\hat\mathcal{V}$ if and
only if $\nabla_{V_1} V_2 = 0$, for all $V_1\in\mathcal{V}$.
Let $V_1,V_2\in\hat\mathcal{V}$. Then
\[
\nabla_{V_3} \nabla_{\hlift{V}_1}V_2=
-\theta(V_1,V_3)V_2 - \nabla_{[\hlift{V}_1,V_3]}V_2
+ \nabla_{\hlift{V}_1}\nabla_{V_3}V_2= - \theta (V_1,V_3)V_2,
\]
meaning that $\nabla_{\hlift{V}_1}V_2$ is again projectable if and
only if $\theta=0$.
Let $\theta=0$. If we set,
for $V_1,V_2 \in \hat V$,
\[
D_{V_1} V_2 = \nabla_{\hlift{V}_1} V_2,
\]
one easily verifies that the operator $D$ satisfies connection-like
properties with respect to the multiplication of elements of the
module $\hat\mathcal{V}$ with projectable functions $f$:
\[
D_{fV_1} V_2 = f D_{V_1} V_2 \qquad \mbox{and} \qquad
D_{V_1} f V_2 = f D_{V_1} V_2 + \hlift{V}_1(f) V_2.
\]
Remark that $\hlift{V}_1(f)$ is again projectable, since for any
$V_2\in\mathcal{V}$, $V_2 \hlift{V}_1(f) = \hlift{V}_1V_2(f) + [V_2,
\hlift{V}_1](f)=0$.
In the coordinates as defined in Theorem~\ref{sode} (regardless of
whether $F$ lies in $\mathcal{W}$ or not) the connection coefficients are
given by
\[
D_{ \frac{\partial}{\partial y^i} } \fpd{}{y^j} =
\conn kij \fpd{}{y^k}, \qquad
\mbox{where\,\,} \Gamma^i_j = -{\textstyle\frac12} \fpd{f^i}{y^j} \,
\mbox{\,and\,} \, \conn ijk = \fpd{\Gamma^i_j}{y^k}.
\]
It is clear from this expression that $\conn kij = \conn kji$, or,
equivalently, that the connection $D$ is symmetric, in the sense that
the torsion
\[
D_{V_1} V_2 - D_{V_2} V_1 - S[\hlift{V}_1,\hlift{V}_2]
\]
vanishes.
We can therefore conclude that the functions $\Gamma^k_{ij}$ are
projectable if and only if $\theta=0$. For that to be the case, $f^k$
must be of the form $f^k = \conn kij y^iy^j + P^k_i y^i + Q^k$, for
some projectable functions $P^k_i(t,x)$ and $Q^k(t,x)$.
\end{proof}
The advantage of the current description is that the criterion
$\theta=0$ can be verified in any given set of coordinates on $M$.
\section{Global properties}\label{globalsection}
In this section we will address the global bundle structure of a
manifold $M$ in the context of a given involutive distribution and a
vector field $F$, assuming from the start that the set where the
dimension of $\mathcal{V}+[F,\mathcal{V}]$ is maximal is the whole of $M$, or in other
words that $\mathcal{V}+[F,\mathcal{V}]$ is actually a distribution.
Let $M$ be an $m$-dimensional manifold with an involutive (and thus
integrable) $n$-dimensional distribution $\mathcal{V}$. The foliation of the
distribution defines an equivalence relation on $M$ by declaring two
elements of $M$ to be equivalent if they lie in the same leaf of $\mathcal{V}$.
The quotient of $M$ by means of this equivalence relation, say $Q$,
will have the structure of a differentiable manifold if for every leaf
one can find a smooth embedded local submanifold $N$ through a
point of the leaf, of dimension $m-n$, which has the property that
each other leaf it meets is intersected in only one point. Then
$\pi_1: M\to Q$ defines a fibration, for which the fibres have dimension $n$,
and for which the distribution $\mathcal{V}$ coincides with the tangents to the
fibres.
We will assume that this condition is satisfied.
In the case of interest, $Q$ comes also equipped with an integrable
distribution. Indeed, a projectable vector field on $M$ is
$\pi_1$-related to a vector field on $Q$. Those projectable vector
fields that happen to lie in $\mathcal{W}$ define therefore a distribution on
$Q$, say $\bar \mathcal{W}$, which is involutive by construction. As above,
the corresponding equivalence relation therefore defines a new
quotient, $T$, again assumed to be a manifold. We will denote the
corresponding fibration by $\pi_2: Q\to T$. Alternatively, we could
have defined a fibration by quotienting out the distribution $\mathcal{W}$
from the beginning. This structure will coincide with the composition
projection $\pi_2 \circ\pi_1: M\to T$.
An almost tangent structure on an even dimensional manifold is a
(1,1)-tensor field $S$ on that manifold, for which the kernel of $S$
at each point coincides with its image. The almost tangent structure
is said to be integrable if its Nijenhuis torsion vanishes. If that
is the case then the kernel of $S$ is an involutive distribution. We
now recall a result from \cite{CT}. Suppose that the kernel of an
integrable almost tangent structure on a manifold defines a fibration
over some base manifold (as above, by taking the quotient of that
distribution). Suppose that each fibre is connected and simply
connected, and that there exists a flat connection on each fibre, for
which the fibre is geodesically complete. Then the manifold is the
total manifold of an affine bundle modelled over the tangent bundle of
the base manifold.
This theorem can be applied to the current setting, if we take a
particular leaf $L_{\mathcal{W}}$ of $\mathcal{W}$ to be the even dimensional manifold
of interest. The projection $\pi_1$ will project this leaf $L_\mathcal{W}$ of
$\mathcal{W}$ onto a corresponding leaf $L_{\bar\mathcal{W}}$ of $\bar\mathcal{W}$. Therefore we
may consider the fibration given by the restriction
$\pi_1|_{L_\mathcal{W}}: L_\mathcal{W} \to L_{\bar\mathcal{W}} $. Vectors that are tangent to
its fibres can be identified with vectors in $\mathcal{V}$, and the fibres
themselves can be identified with leaves of $\mathcal{V}$. We have defined a
$\mathcal{W}$-tensor field $S$ on $\mathcal{W}$ in Proposition~\ref{prop1}. It
restricts naturally to an almost tangent structure on $L_\mathcal{W}$ (i.e.\
$S^2=0$, $N_S=0$, and $\mathop{\mathrm{ker}} S=\mathop{\mathrm{im}} S= \mathcal{V}_{L_\mathcal{W}}$).
Moreover, we have seen that the restriction of the $\mathcal{V}$-connection
$\nabla_V$ ($V\in\mathcal{V}$) to a leaf of $\mathcal{V}$ gives a flat connection on
that leaf. We can conclude therefore:
\begin{thm} \label{thm2}
Suppose that each leaf of $\mathcal{V}$ is connected and simply connected, and
assume that each leaf of $\mathcal{V}$ is geodesically complete with respect to
the restriction of $\nabla_V$ to that leaf. Then, for each $L_\mathcal{W}$ of
$\mathcal{W}$, $\pi_1|_{L_\mathcal{W}}: L_\mathcal{W} \to L_{\bar\mathcal{W}}$ is an affine bundle,
modelled over the tangent bundle $T(L_{\bar\mathcal{W}})\to L_{\bar\mathcal{W}}$.
Suppose further that the set $N=\{z\in L_\mathcal{W}:F(z)\in\mathcal{V}\}$ is a global
cross-section of $\pi_1|_{L_\mathcal{W}}$:\ then $L_\mathcal{W}$ may be identified with
$T(L_{\bar\mathcal{W}})$ and $N$ with the zero section.
\end{thm}
\begin{cor} \label{cor1}
In case $F\in\mathcal{W}$, and under the assumptions of the previous theorem,
the restriction of $F$ to a certain leaf $L_\mathcal{W}\equiv T(L_{\bar\mathcal{W}})$ will
be a second-order differential equation field on $T(L_{\bar\mathcal{W}})$.
\end{cor}
\begin{proof} This follows easily from the coordinate expression of
$F$. Restricting $F$ to a leaf is the same as fixing the parameters
$t^p$ to some constant values.
\end{proof}
For completeness, we mention that one may find an alternative
formulation of the theorem of \cite{CT} in \cite{DeF}, where the
global conditions on an (assumed given) symmetric connection are
replaced by global conditions on an (assumed given) vector field. In
our current framework, the restriction of the vector field $S(F)$ to
$L_\mathcal{W}$ plays the role of that vector field.
In case $F$ does not belong to $\mathcal{W}$, but leaves it invariant, $F$
defines a vector field $\tilde F$ on $T$. This vector field defines a
1-dimensional involutive distribution on $T$, leading as before to a
fibration $T\to T_0$. If we assume that the vector field $\tilde
F$ is complete, an integral curve of $\tilde F$ will define a
1-dimensional submanifold $T_1$ of $T$. In turn, the preimage of
$T_1$ under $\pi_2$ is a collection $E$ of leaves of $\bar\mathcal{W}$ lying
over that integral curve. We can think of the restriction of $\pi_2$
to $E$ as defining a fibration $\pi_2|_E: E\to T_1$. Let's denote
its 1-jet bundle by $J^1(E)\to E$.
The distribution $\mathcal{W}_F = \langle F \rangle \oplus \mathcal{W}$ is also
involutive. Its leaves $L_{\mathcal{W}_F}$ are $(2n+1)$-dimensional manifolds
that are projected by means of $\pi_1$ onto one of the above described
manifolds $E$, corresponding to a certain integral curve (with image
$T_1$) of $\tilde F$. The fibres of $\pi_1|_{L_{\mathcal{W}_F}}: L_{\mathcal{W}_F} \to
E$ can again be identified with $\mathcal{V}$. Recall that we had defined a
symmetric flat connection on each leaf of $\mathcal{V}$.
By setting $S(F)=0$ we can extend $S$ to a $\mathcal{W}_F$-tensor field, which
has the property that its restriction to a leaf of $\mathcal{W}_F$ satisfies
$S^2=0$, $N_S = 0$ and $\mathop{\mathrm{rank}} S = n$. These properties are exactly
those that define, in the terminology of \cite{MP2}, an `almost jet
structure'. The above mentioned theorem in \cite{CT} has been
generalized to $(2n+1)$-dimensional manifolds with almost jet
structures in \cite{MP2}, where a 1-jet bundle replaces the role
played by a tangent bundle (see the Theorem on page
90 of \cite{MP2}). We are in the situation that we can apply this
theorem, since each $L_{\mathcal{W}_F}$ is a $(2n+1)$-dimensional manifold with
a $2n$-dimensional distribution $\mathcal{W}|_{L_{\mathcal{W}_F}}$ which is completely
integrable and which is such that $\mathop{\mathrm{im}} S = \mathcal{V}|_{L_{\mathcal{W}_F}} \subset
\mathcal{W}|_{L_{\mathcal{W}_F}}$. We may therefore conclude:
\begin{thm}
Suppose that each leaf of $\mathcal{V}$ is connected and simply connected, and
assume that each leaf of $\mathcal{V}$ is geodesically complete with respect to
the restriction of $\nabla_V$ to that leaf. In case $F\notin\mathcal{W}$ and
$[F,\mathcal{W}]\subset\mathcal{W}$, each leaf $L_{\mathcal{W}_F}$ of $\mathcal{W}_F$ is diffeomorphic to the
1-jet bundle $J^1(E)$ of $E\to T_1$.
\end{thm}
\begin{cor}
Under the assumptions of the previous theorem, the restriction of $F$
to a certain leaf $L_{\mathcal{W}_F}\equiv J^1(E)$ will be a time-dependent
second-order differential equation field on $J^1(E)$.
\end{cor}
\begin{proof}
This follows again from the coordinate expression of $F$. Restricting
$F$ to a leaf is the same as fixing all parameters $t^p$ to some
constant values, except for $t^1$.
\end{proof}
The cases of most obvious interest are those in which the dimension of
$M$ is either $2n$ or $2n+1$ ($n$ being the dimension of $\mathcal{V}$). We
end this section with a statement of our global results in these
cases, in a form which collects together the assumptions we have made.
\begin{thm}
Let $\mathcal{V}$ be an involutive distribution of dimension $n$ on a manifold
$M$ of dimension $m$, $m=2n$ or $2n+1$; and $F$ a vector field such
that $\mathcal{V}\cap[F,\mathcal{V}]=\{0\}$. Assume that
\begin{itemize}
\item $M$ is fibred over a manifold $Q$ and the leaves of $\mathcal{V}$ are
the fibres of this fibration;
\item each leaf of $\mathcal{V}$ is connected and simply connected;
\item each leaf of $\mathcal{V}$ is geodesically complete with respect to the
flat symmetric connection induced on it (as described in Section~3);
\item $\mathcal{V}+[F,\mathcal{V}]$ is a distribution (necessarily of dimension $2n$).
\end{itemize}
In the case $m=2n$, assume further that
\begin{itemize}
\item the set $\{z\in M:F(z)\in\mathcal{V}\}$ is a global cross-section of
$M\to Q$.
\end{itemize}
Then $M$ may be identified with $T(Q)$ and $F$ with a second-order
differential equation field on $T(Q)$.
In the case $m=2n+1$ assume further that
\begin{itemize}
\item $\mathcal{W}=\mathcal{V}+[F,\mathcal{V}]$ is involutive;
\item $F\notin\mathcal{W}$, $[F,\mathcal{W}]\subset\mathcal{W}$;
\item $F$ is complete;
\item $Q$ is fibred over $\mathbb{R}$, and the leaves of $\bar{\mathcal{W}}$ (the
projection of $\mathcal{W}$ to $Q$) are the fibres of this fibration.
\end{itemize}
Then $M$ may be identified with $J^1(Q)$ and $F$ with a time-dependent
second-order differential equation field on $J^1(Q)$.
\end{thm}
\section{An illustrative example}
The cases in which $\dim M$ is $2n$ or $2n+1$ may be of most obvious
interest, but they are by no means the only cases of interest, as we
now show by an example.
Let ${\sf Q}$ be the configuration space of a Lagrangian system with
regular Lagrangian $L$ and assume that $L$ is invariant under the
(free and proper) action of a symmetry Lie group $G$. In that case,
the Euler-Lagrange field $\Gamma\in\vectorfields{T({\sf Q})}$ is
$G$-invariant and it can be reduced to a vector field $\check\Gamma$
on $T({\sf Q})/G$. The corresponding equations for finding the integral
curves of $\check\Gamma$ are known in the literature as the
`Lagrange-Poincar\'e equations', see e.g.\ \cite{Cendra}. The
equations determining the reduced vector field $\check\Gamma
\in\vectorfields{T({\sf Q})/G}$ can be cast in terms of the reduced
Lagrangian $l$ on $T({\sf Q})/G$.
We will follow here closely the description we have given in
\cite{invlag}. With the aid of a principal connection on ${\sf Q}\to{\sf Q}/G$
one may decompose $T({\sf Q})/G$ into $T({\sf Q}/G) \oplus ({\sf Q}\times\mathfrak{g})/G$,
where the action of $G$ on $\mathfrak{g}$ is the adjoint action. In what
follows $(x^i,v^i,w^p)$ are local coordinates on $T({\sf Q})/G$, where the
$(x^i)$ are coordinates on ${\sf Q}/G$, and $(v^i,w^p)$ are fibre
coordinates corresponding to the decomposition.
We will assume that the symmetry group is Abelian. This has the
advantage that the adjoint action is trivial. The vector field
$\check\Gamma$ can then be determined from:
\begin{eqnarray*}
&& \check\Gamma({x}^i) = v^i\\
&&\check{\Gamma}\left( \fpd{l}{v^i}\right) -\fpd{l}{x^i} =
K^p_{ik}v^k \fpd{l}{w^p}\nonumber\\
&&\check{\Gamma}\left( \fpd{l}{w^p}\right) =0.
\end{eqnarray*}
Here $K^p_{ik}$ are the components of the curvature of the principal
connection with respect to an invariant basis. The coordinate
expression of the reduced field is therefore of the form
$\check\Gamma=v^i\partial/\partial x^i + \Gamma^i \partial/\partial
v^i + \Gamma^p \partial/\partial w^p$, where $\Gamma^i$ and $\Gamma^p$
are functions on $T({\sf Q})/G$.
In the assumption that the matrix $(\partial^2 l/\partial w^p\partial
w^q)$ is everywhere non-singular, and that the relation $\partial
l/\partial w^p=\mu_p$ can therefore be rewritten in the form
$w^p=\rho^p(x,v,\mu)$, we can perform a coordinate transformation
$(x^i,v^i,w^p) \to ({\bar x}^i = x^i,{\bar v}^i = v^i,\mu_p = \partial
l/\partial w^p)$. The last equation is then simply
$\check{\Gamma}\left( \mu_p\right) =0$, that is to say: the
coordinates $\mu_p$ can be regarded as parameters. In the new
coordinates the reduced vector field becomes $\check\Gamma={\bar
v}^i\partial/\partial {\bar x}^i + \Gamma^i \partial/\partial {\bar
v}^i + 0 \partial/\partial \mu_p $.
The first two equations determine a system of second-order ordinary
differential equations in the variables $x^i$ with parameters $\mu_p$.
By introducing Routh's (reduced) function
\[
{\mathcal R}^\mu (x,v)= l(x,v, \rho(x,v,\mu)) - \mu_p \rho^p(x,v,\mu)
\]
these equations, restricted to fixed values for the $\mu$\,s, can
equivalently be rewritten as
\[
\check\Gamma\left( \fpd{{\mathcal R}^\mu}{v^i}\right)
-\fpd{{\mathcal R}^\mu}{x^i} = K^p_{ik}v^k\mu_p.
\]
This equation is known as Routh's (reduced) equation for an Abelian
symmetry group, see e.g.\ \cite{MaSch}.
We show that the situation described above is in agreement with the
statements of Theorem~\ref{sode} and Corollary~\ref{cor1}. Recall
first the definition of the momentum map $J: T({\sf Q}) \to {\mathfrak{g}}^*$, where
$\langle J(v),\xi \rangle = (\vlift{\tilde\xi}L)(v)$ (for each
$\xi\in\mathfrak{g}$, $\tilde\xi\in\vectorfields{{\sf Q}}$ is the corresponding
fundamental vector field). It is well-known that the map $ T({\sf Q}) \to
{\sf Q}\times {\mathfrak{g}}^*, (q,v) \mapsto (q,J(v))$ is $G$-equivariant, where the
action of $G$ on $\mathfrak{g}^*$ is the coadjoint action. Therefore, it
reduces to a map ${\check J}: T({\sf Q})/G \to ({\sf Q} \times {\mathfrak{g}}^*)/G$. But
for an Abelian group the adjoint action is trivial, so the coadjoint
action is also trivial. It follows that $J$ is invariant, and that
the image of the reduced momentum $\check J$ is ${\sf Q}/G\times \mathfrak{g}^*$. In
the current coordinates $\check J$ is simply $(\bar x,\bar v,\mu)
\mapsto (\bar x,\mu)$.
Let $M=T({\sf Q})/G$ and $F=\check\Gamma$. The distribution $\mathcal{V} = \mathop{\mathrm{ker}}
T\check J$ is clearly involutive. It has the commuting basis given by
the vector fields $\partial / \partial {\bar v}^i$. It is easy to see
that $[F,\mathcal{V}]\cap \mathcal{V} = \{0\}$. The distribution $\mathcal{W} = \mathcal{V} + [F,\mathcal{V}]$ is
spanned by $\{ \partial /\partial {\bar x}^i, \partial /\partial {\bar
v}^i \}$ and is involutive as well. It is the distribution formed by
the vector fields on $T({\sf Q})/G$ which are tangent to the level sets of
momentum. A leaf of $\mathcal{W}$ is thus a particular level set, $\mu_p
=\mu_p^0$. The corresponding $N$ in the statement of
Theorem~\ref{sode} can be identified with $\mathop{\mathrm{im}} \check J =
{\sf Q}/G\times\mathfrak{g}^*$, and is non-empty. So Theorem~\ref{sode} applies.
The quotient space $Q$ of Section~\ref{globalsection} can be identified with ${\sf Q}/G \times
\mathfrak{g}^*$. It is trivially fibred over $T=\mathfrak{g}^*$; $\bar \mathcal{W}$ is the
distribution formed by the (projections of the) $\partial/\partial
{\bar x}^i$, and any leaf of $\bar\mathcal{W}$ can therefore be identified with
${\sf Q}/G$. According to Theorem~\ref{thm2}, therefore,
the restriction of $F$ to a level set of momentum $\mu=\mu^0$ is a
second-order differential equation field on ${\sf Q}/G$ (a leaf of
$\bar\mathcal{W}$); it is of course the one which satisfies Routh's reduced
equation with $\mu_p=\mu^0_p$.
\subsubsection*{Acknowledgements}
The first author is a Postdoctoral Fellow of the Research Foundation
-- Flanders (FWO). The second author is a Guest Professor at Ghent
University:\ he is grateful to the Department of Mathematics for its
hospitality. This work is part of the {\sc irses} project {\sc
geomech} (nr.\ 246981) within the 7th European Community Framework
Programme. We are indebted to W.\ Sarlet for many useful discussions.
|
\section{Introduction}
\label{S.intro}
Smoluchowski's coagulation equation provides a mean-field description of
binary coalescence of clusters.
If $\xi$ denotes the size of a cluster and $f(\xi,t)$ the corresponding number density
at time $t$ then the equation is
\begin{equation}\label{coag1}
\frac{\partial}{\partial t} f (\xi,t) = \tfrac 1 2 \int_0^\xi \,d\eta\, K(\xi-\eta,\eta)
f(\eta,t) f(\xi{-}\eta,t)
\,-
\, f(\xi,t) \int_0^{\infty}\,d\eta\, K(\xi,\eta) f(\eta,t)\,,
\end{equation}
where $K(\xi,\eta)$ is a kernel that describes the rate of the coalescence process.
Here we consider a specific diagonal kernel of homogeneity $\gamma < 1$,
given by $K(\xi,\eta) = \delta_{(\xi-\eta)} \xi^{1+\gamma}$,
that reduces (\ref{coag1}) to
\begin{equation}\label{coag2}
\frac{\partial}{\partial t} f (\xi,t) = \frac 1 4 \Big( \frac{\xi}{2},t\Big)^{1+\gamma} f^2\Big(\frac{\xi}{2},t
\Big) - \xi^{1+\gamma} f^2(\xi,t)\,.
\end{equation}
In the following we study self-similar solutions of (\ref{coag2}).
Such solutions are of the form
\begin{equation}\label{sss1}
f(\xi,t) = t^{-(1+(1+\gamma)\beta)} g\Big( \frac{\xi}{t^{\beta}}\Big)
\end{equation}
for some positive $\beta$, where
$g$ satisfies, with $x=\xi/t^{\beta}$, that
\begin{equation}\label{sss2}
-\big( 1 + (1+\gamma)\beta \big)g - \beta x g'(x)
= \frac 1 4 \Big( \frac{x}{2}\Big)^{1+\gamma} g^2\left( \frac{x}{2}\right)
-x^{1+\gamma} g^2(x)\,.
\end{equation}
If one looks for solutions with conserved mass, then $\beta$
is uniquely determined by $\beta=\beta_*:= 1/(1{-}\gamma)$.
For further reference we also note that we can integrate the equation in (\ref{sss2})
to obtain
\begin{equation}\label{sss2b}
\beta x^2 g(x) = \int_{x/2}^x s^{2+\gamma} g^2(s)\,ds + (1{-}\gamma)(\beta - \beta_*)
\int_0^x s g(s)\,ds\,.
\end{equation}
Here we assumed implicitly that $x g(x)$ and $x^{2+\gamma} g^2(x)$ are
integrable at zero and that
$\lim_{x\to 0} x^2 g(x)=0$. As we will see below (cf. (\ref{sol1})), these properties
will be satisfied by the solutions we are going to consider.
Notice also that we have the well-known power-law solution
\begin{equation}\label{sss5}
g = x^{-(1+\gamma)} \frac{1}{1{-}\theta} \qquad \mbox{ with } \theta:=2^{\gamma-1}<1\,.
\end{equation}
In \cite{Le06} a
mass-conserving solution of (\ref{sss2b}), that is a solution for $\beta = \beta_*$,
is constructed that is decaying exponentially fast and satisfies
\begin{equation}\label{le1}
g(x) = x^{-(1+\gamma)} \Big( \frac{1}{1{-}\theta} - c x^{\mu/(1-\gamma)} + o(x^{\mu/(1-\gamma)})
\Big) \qquad \mbox{ as } x \to 0\,,
\end{equation}
where $\mu>0$ satisfies a certain transcendental equation. The constant $c>0$
is not determined due to an invariance of (\ref{sss2}) under the rescaling
$g(x) \mapsto a^{1+\gamma} g(ax)$ for any $a>0$. In the case of mass-preserving solutions
the constant can be fixed by normalizing the mass of the solution.
As is pointed out in \cite{Le06}, the solution is unique in the class of
functions satisfying (\ref{le1}), but uniqueness in general is not known.
In \cite{Le06} the question is raised whether solutions with algebraic decay, others from the
one in (\ref{sss5}), exist in analogy to the ones that have been
found in \cite{MP1} for the constant and additive kernel. More precisely, for example for the constant
kernel, it is established in \cite{MP1} that there exists a family of self-similar solutions with infinite
mass and the decay behavior $x^{-(1+\rho)}$ for all $\rho \in (0,1)$. Furthermore, it is
shown that a solution
of the coagulation equation converges to the self-similar solution with decay behavior
$x^{-(1+\rho)}$ if and only if the
mass-distribution of the initial data is regularly varying with exponent $\rho$. In this note we prove
for the diagonal kernel
the existence of a corresponding family of self-similar solutions with infinite mass and asymptotic behavior $x^{-(1+\rho)}$ as $x \to \infty$
with $\rho \in (\gamma,1)$. Notice, that this includes solutions that are
increasing as $x \to \infty$ if $\gamma <-1$.
Our proof is simple and exploits strong monotonicity properties
of a suitably rescaled version
of the equation for the self-similar solution. We presently do not know, however, how
to characterize the domains of attraction of these self-similar solutions. The analysis in \cite{MP1}
relies on the fact that the Laplace transform of the equation satisfies a simple ODE, a method
that is not applicable in the present situation.
Our main result is the following.
\begin{theorem}
\label{T.1}
Let $\gamma <1$ and $\mu$ be the unique positive solution of
\begin{equation}\label{mueq}
\frac{1+\beta \mu}{2} = \frac{1-2^{\gamma-1-\mu}}{1-2^{\gamma-1}}\,.
\end{equation}
Then there exists for any $\beta>\beta_*$ a solution $g$ of
(\ref{sss2b}) such that
\begin{equation}\label{sol1}
g(x) = x^{-(1+\gamma)} \Big( \frac{1}{1{-}\theta} - c x^{\mu/(1{-}\gamma)} +
o\big (x^{\mu/(1{-}\gamma)} \big)\Big)
\end{equation}
as $x \to 0$ with a positive constant $c$ Furthermore, $x^{-(1+\gamma)} g(x)$ is
monotonically decreasing and satisfies
\begin{equation}\label{sol2}
g(x) \sim \frac{d}{x^{1+\gamma+1/\beta}} \qquad \mbox{ as } x \to \infty
\end{equation}
for some positive constant $d$.
\end{theorem}
As explained above, the constants $c$ and $d$ in Theorem \ref{T.1} are not determined
due to the invariance of the equation under appropriate rescaling.
\section{Proof}
Our proof proceeds similarly to the one in \cite{Le06} for the mass-conserving solutions.
First, to scale out the singular behavior as $x \to 0$, we introduce
$h(x)=g(x) x^{1+\gamma}
$
such that
$h$ solves
\begin{equation}\label{sss4}
-\beta x h'(x) - h(x) = \theta
h^2 \left ( \frac{x}{2} \right) - h^2(x)\,
\end{equation}
or, due to (\ref{sss2b}),
\begin{equation}\label{sss4b}
\beta x^{1-\gamma} h(x) = \int_{x/2}^x s^{-\gamma} h^2(s)\,ds + (1{-}\gamma)(\beta -\beta_*) \int_0^x s^{-\gamma}
h(s)\,ds\,.
\end{equation}
Notice, that the power-law solution (\ref{sss5}) corresponds to the constant solution
$h \equiv 1/(1{-}\theta)$. It is also clear that any solution of $(\ref{sss4})$ for which
$\lim_{x \to 0} h(x) $ exists, that this limit must equal $1/(1{-}\theta)$. We are now looking
for solutions that bifurcate from this constant at $x \to 0$.
In order to identify the next order behavior, we make the ansatz $h(x) = 1/(1{-}\theta) + x^{\mu} +o(x^{\mu})$
as $x \to 0$. Plugging this into (\ref{sss4b}),
recalling that $\beta_*=1/(1{-}\gamma)$ and rearranging we find that $\mu$ must indeed satisfy (\ref{mueq}).
If we denote by $F(\mu)=(1-2^{\gamma-1-\mu})/(1{-}\theta)$ we see that $F(0)=1>1/2$. On the other hand,
$F$ is increasing and $\lim_{\mu \to \infty} F(\mu) = 1/(1{-}\theta)$. Hence, there must be a unique
positive solution of (\ref{mueq}).
Next, we introduce the function $j(x)$ via
\begin{equation}\label{jdef}
h(x)=\frac{1}{1{-}\theta} + x^{\mu}\big( -c + j(x) \big)\,,
\end{equation}
where $c \in \mathbb{R}$ is a constant.
Using equations (\ref{mueq}) and (\ref{sss4b}) we obtain that $j$ satisfies
\begin{equation}\label{jeq}
\begin{split}
j(x)&= \frac{1}{\beta} x^{-(1-\gamma+\mu)} \Big( \int_{x/2}^x s^{-\gamma+\mu} \frac{2}{1{-}\theta}
j(s)\,ds + \int_{x/2}^x s^{-\gamma+ 2\mu} (-c+j(s))^2\,ds\\
&\quad + (1{-}\gamma)(\beta{-}\beta_*) \int_0^x s^{-\gamma+\mu} j(s)\,ds \Big) =: T[j]\,.
\end{split}
\end{equation}
In order to prove that a local solution of (\ref{jeq}) exists, we can proceed analogously to \cite{Le06}.
We only indicate the main steps here.
We define for some $\eps \in (0,\mu)$ and $z>0$ the space
\[
C_{\eps}(z):=\{ f \in C[0,z]\,;\, f(0)=0\,;\, \|f\|:=\sup_{x \in [0,z]} x^{-\eps} |f(x)| <\infty\}\,.
\]
It is clear that the operator $T$ maps $C_{\eps}(z)$ into itself. Next, we are going to show that
$T$ maps a ball in $C_{\eps}(z)$ of a sufficiently small
radius $R$ into itself if $z$ is sufficiently small.
This follows from
\[
\begin{split}
\|T[j]\|& \leq \frac{1}{\beta}\|j\| \Big( \frac{2}{1{-}\theta}
\frac{1}{1{-}\gamma{+}\mu{+}\eps}
\big( 1 {-} 2^{\gamma-1-\mu-\eps}\big)
+ \|j\| \frac{2 z^{\mu}}{1{-}\gamma{+}2\mu{+}2\eps} \big( 1 - 2^{\gamma-1-\mu-2 \eps}\big)
\\
& \qquad + c^2 \frac{ 2z^{\mu}}{1{-}\gamma{+}2\mu} + \frac{(1{-}\gamma)
(\beta{-}\beta_*) }{1{-}\gamma{+}\mu{+}\eps} \Big)
\end{split}
\]
that implies
\[
\|T[j]\| \leq \|j\| \Big(\frac{1}{\beta(1{-}\gamma{+}\mu{+}\eps)} \big( 2 F(\mu+\eps) + (1{-}\gamma) \beta -1\Big)
+ C z^{\mu} \big( \|j\|^2 +1\big)\,.
\]
Now we know by the definition of $\mu$ that $2 F(\mu+\eps) < 1+\beta(\mu+\eps)$
and hence
\[
\frac{1}{\beta(1{-}\gamma{+}\mu{+}\eps)} (2 F(\mu+\eps) + (1{-}\gamma) \beta -1)
< \frac{1}{1{-}\gamma{+}\mu{+}\eps }(\mu{+}\eps {+}1{-}\gamma)=1\,.
\]
Thus, there exists a constant $\kappa=\kappa(\eps)<1$ such that
if $\|j\| \leq R$ we find
$\|T[j]\| \leq \kappa R + Cz^{\mu}(R^2+1)$.
For sufficiently small $z$ and an appropriately small $R$ the right hand side
is bounded by $R$. Similarly one can show that $T$ is a contraction, we omit
the details here.
Hence, a local solution to (\ref{jeq}) exists, and thus also to (\ref{sss4}).
Next, we choose $c>0$, and claim that $h$ is decreasing in a neighbourhood of zero.
To see this, notice that it follows from (\ref{jeq}) that $j'(x)$ exists for $x>0$ and that we have
the estimate
$|j'(x)| \leq C \frac{|j(x)|}{x} + C x^{\mu-1}$ for $x \in (0,z)$.
This in turn implies that
\[
h'(x) = \mu x^{\mu-1} \big( -c+j(x)) + x^{\mu} j'(x) \leq x^{\mu-1} \big( - c \mu + C \mu|j(x)| + C x^{\mu}
\big)\,.
\]
If $z$ is sufficiently small, we find that $h'(x)<0$ for $x \in (0,z)$.
We are going to show that as long as $h$ exists and is positive this property is
conserved. Indeed, assume that there exists
$x_0>0$ such that $h'(x_0)=0$. Then (\ref{sss4}) and the fact
that $h$ is decreasing for $x<x_0$ imply
that
\[
0= h(x_0)^2 - h(x_0)- \theta h^2\Big( \frac{x_0}{2}\Big)^2
< \big(1{-}\theta\big) h^2(x_0) - h(x_0)
=h(x_0) \big( (1{-}\theta) h(x_0) -1\big) \,.
\]
As long as $h$ is positive, the right hand side is strictly negative, since
$h(x_0)<1/(1{-}\theta)$ and we obtain the desired contradiction.
Moreover, equation (\ref{sss4b}) implies for $\beta \geq \beta_*$
that $h$ is positive whenever it exists. Hence, using standard results on ordinary differential equations,
we obtain global existence of a solution $h$
to (\ref{sss4}) which is strictly decreasing.
Since equation (\ref{sss4}) has the only stationary points $1/(1{-}\theta)$ and $0$, it also follows
that $h(x) \to 0$ as $x \to \infty$.
It remains to show that $h(x) \sim d x^{-1/\beta}$ as $x \to \infty$ from which (\ref{sol2}) follows. First, due to the invariance of equation (\ref{sss4})
under the transformation $x \to ax$ for $a>0$, we can assume without
loss of generality that $h(1)=1/2$. Since
$h$ satisfies
$\beta x h'(x) + h(x) \leq h^2(x)$
we have by simple comparison that
\begin{equation}
\label{supersol}
h(x) \leq
\frac{1}{1+ x^{1/\beta}} \qquad \mbox{ for } x\geq 1\,.
\end{equation}
We now introduce $p(x)=x^{1/\beta}h(x)$ that solves
\begin{equation}\label{peq}
\beta p'(x) = x^{-(1+1/\beta)} \Big( p^2(x) - \theta 2^{2/\beta} p^2
\Big (\frac{x}{2}\Big) \Big )\,.
\end{equation}
The estimate (\ref{supersol}) in particular implies that $p(x) \leq 1$
for all $x \geq 1$ and thus (\ref{peq}) implies that $\beta |p'(x)|
\leq 2 x^{-(1+1/\beta)} $ for all $x \geq 2$. Hence
$|p(x)-p(x_0)| \leq 2 x_0^{-1/\beta}
$
for any $x_0\geq 2$ which implies that $\lim_{x\to \infty} p(x)$ exists. In order
to complete the proof of Theorem \ref{T.1} it remains
to establish that this limit is strictly positive. To this end we note that
(\ref{sss4b}) implies
\begin{equation}\label{hineq}
\beta x^{1-\gamma }h\left( x\right) >\left( 1{-}\gamma \right) \left( \beta
-\beta _{\ast }\right) \int_{0}^{x}s^{-\gamma }h\left( s\right) ds\,.
\end{equation}
If we define
$\Phi \left( x\right): =\int_{0}^{x}s^{-\gamma }h\left( s\right) ds$ then (\ref{hineq})
implies that $
\beta x\Phi ^{\prime }\left( x\right)$\linebreak $ -\left( 1-\gamma \right) \left( \beta
-\beta _{\ast }\right) \Phi \left( x\right) >0$. Integrating this last inequality we obtain \linebreak
$\left( x^{-\frac{\left( 1-\gamma \right) \left( \beta -\beta _{\ast }\right)
}{\beta }}\Phi \left( x\right) \right) ^{\prime }>0
$ and thus
\[
x^{-\frac{\left( 1-\gamma \right) \left( \beta -\beta _{\ast }\right) }{%
\beta }}\Phi \left( x\right) \geq \Phi(1) = \int_0^1 s^{-\gamma} h(s)\,dx=:c_0 >0\,
\]%
for all $x\geq 1$.
Thus
\[
\Phi \left( x\right) \geq c_0 x^{\frac{\left( 1-\gamma \right) \left( \beta
-\beta _{\ast }\right) }{\beta }}=c_0 x^{1-\gamma }x^{-\frac{1}{\beta }}\ \ \
\]%
for $x \geq 1$ and plugging this into (\ref{hineq}) we find
$h\left( x\right) \geq \frac{c_0}{\beta }x^{-\frac{1}{\beta }}$
for all $x \geq 1$, that finishes the proof.
\bigskip
{\bf Acknowledgment.} This work was supported by the EPSRC Science and Innovation award to the
Oxford Centre for Nonlinear PDE (EP/E035027/1),
the DGES Grant MTM2007-61755, the Proyecto Intramural 200850I248
and the Isaac Newton Institute.
\bibliographystyle{amsplain}
|
\section{Introduction}
The history of fractional calculus can be traced all the way back to
Leibniz. In a letter to L'Hopital (1695), Leibniz mentions that he has an
expression that looks like a fractional derivative [1, 2]. Later, Euler
notices that using his gamma function fractional integrals may be possible.
However, a systematic development of the subject does not come until
nineteenth century, where Riemann, Liouville, Gr\"{u}nwald and Letnikov play
key roles [1-6]. On the application side, after it became clear that
anomalous diffusion can be investigated in terms of the fractional diffusion
equation, the situation has been changing rapidly [1-6]. The first
applications to quantum mechanics started with Laskin in 2000, where he
considered the path integral formulation of quantum mechanics over L\'{e}vy
paths and showed that the corresponding equation of motion is the space
fractional Schr\"{o}dinger equation [7-17]. In 2004, Naber [18] discussed
the time fractional Schr\"{o}dinger equation [18-24] and obtained the time
dependent part of the wave function for the separable solutions as the
Mittag-Leffler function with an imaginary argument, which he wrote as a
purely oscillatory term plus an integral that can not be evaluated
analytically by standard techniques [18]. The main arguments and the
conclusions of the paper were based on the \textit{critical} assumption that
this integral is uniformly decaying, hence could be neglected for
sufficiently large times. This affected the main conclusions of the paper by
giving probabilities greater than $1.$ Later, others have also followed this
\textit{critical} assumption [22-24].
We evaluate this integral exactly by using the elegant language of Fox's $H-
functions [25-27] and show that the decomposition of the Mittag-Leffler
function as the sum of its purely oscillating and purely decaying parts is
not possible. Since this point implies dramatic changes in the conclusions
of the previous works [18, 22-24], we reconsider the problem of the time
fractional Schr\"{o}dinger equation.
After introducing the time and space fractional Schr\"{o}dinger equation in
Section II, in Section III we concentrate on the time fractional Schr\"{o
dinger equation and discuss its separable solutions. We show that the time
dependence is given as the Mittag-Leffler function with an imaginary
argument by two different methods. After separating the Mittag-Leffler
function into its real and imaginary parts, we show that the total
probability is always $\leq 1$ and decays with time. We also discuss the
asymptotic forms of the wave function. In Section IV, we discuss the simple
box problem, and in Section V, we introduce the effective potential
approach. Using the effective potential, in Section VI we show that the
Mittag-Leffler function with an imaginary argument can be written, not as
the sum, but as the product of its purely decaying and purely oscillating
parts. In Section VII, we discuss our results. \ \ \ \
\section{Fractional Schr\"{o}dinger Equation}
To write the fractional Schr\"{o}dinger equation\textbf{,} we use the fact
that the Schr\"{o}dinger equation is analytic in the lower half complex $t-
plane, and perform a Wick rotation, $t\rightarrow -it,$ on the one
dimensional Schr\"{o}dinger equation,
\begin{equation}
i\hbar \frac{\partial \Psi }{\partial t}=-\frac{\hbar ^{2}}{2m}\frac
\partial ^{2}\Psi }{\partial x^{2}}+V(x)\Psi (x,t),
\end{equation
to write
\begin{equation}
-\hbar \frac{\partial \Psi }{\partial t}=-\frac{1}{2m}\left( \hbar \frac
\partial }{\partial x}\right) ^{2}\Psi +V(x)\Psi (x,t).
\end{equation
This is nothing but the Bloch equation [6]
\begin{equation}
\frac{\partial \Psi }{\partial t}=\text{ }\check{D}\frac{\partial ^{2}\Psi }
\partial x^{2}}-\frac{1}{\hbar }V(x)\Psi (x,t),\text{ }\check{D}>0,
\end{equation
where $\check{D}=\dfrac{\hbar }{2m}$ is the quantum diffusion constant and
V(x)$ is the potential. We can now write the time and space fractional
version of the Bloch equation as
\begin{equation}
_{0}^{C}D_{t}^{\alpha }\Psi (x,t)=\frac{1}{\hbar }\check{D}_{\alpha ,\beta
}\hbar ^{\beta }R_{x}^{\beta }\Psi (x,t)-\frac{1}{\hbar }V(x)\Psi (x,t)
\text{ }0<\alpha <1,\text{ }1<\beta <2,
\end{equation
where $_{0}^{C}D_{t}^{\alpha }$ is the Caputo derivative and $R_{x}^{\beta }$
is the Riesz derivative (Appendix A). When there is no room for confusion,
we also write $_{0}^{C}D_{t}^{\alpha }$ $\equiv \frac{\partial ^{\alpha }}
\partial t^{\alpha }}$ and $R_{x}^{\beta }$ $\equiv \nabla ^{\beta }\equiv
\frac{\partial ^{\beta }}{\partial x^{\beta }}$. We have also introduced a
new quantum diffusion constant, $\check{D}_{\alpha ,\beta },$ with the
appropriate units. Performing an inverse Wick rotation, $t\rightarrow it,$
we obtain the most general fractional version of the Schr\"{o}dinger
equation as
\begin{equation}
\frac{\partial ^{\alpha }}{\partial t^{\alpha }}\Psi (x,t)=\frac{i^{\alpha
}{\hbar }\check{D}_{\alpha ,\beta }\left( \hbar \frac{\partial }{\partial x
\right) ^{\beta }\Psi (x,t)-\frac{i^{\alpha }}{\hbar }V(x)\Psi (x,t).
\end{equation}
\subsection{Space Fractional Schr\"{o}dinger Equation}
When $\alpha=1,$ Equation (5) becomes the space fractional Schr\"{o}dinger
equation investigated by Laskin [7]:
\begin{equation}
\frac{\partial}{\partial t}\Psi(x,t)=\frac{i}{\hbar}\check{D
_{1,\beta}\left( \hbar\frac{\partial}{\partial x}\right) ^{\beta}\Psi(x,t)
\frac{i}{\hbar }V(x)\Psi(x,t),
\end{equation}
which he wrote as
\begin{equation}
i\hbar\frac{\partial}{\partial t}\Psi(x,t)=-D_{\beta}\left[ \hbar \nabl
\right] ^{\beta}\Psi(x,t)+V(x)\Psi(x,t).
\end{equation}
Laskin called $\left[ \hbar\nabla\right] ^{\beta}$ the quantum Riesz
derivative and $D_{\beta}$ in the above equation is the quantum diffusion
constant for the space fractional Schr\"{o}dinger equation, where as
\beta\rightarrow2,$ $D_{\beta}\rightarrow1/2m.$
\subsection{Time Fractional Schr\"{o}dinger Equation}
For $\beta =2,$ we get
\begin{equation}
\frac{\partial ^{\alpha }}{\partial t^{\alpha }}\Psi (x,t)=\frac{i^{\alpha
}{\hbar }\check{D}_{\alpha ,2}\left( \hbar \frac{\partial }{\partial x
\right) ^{2}\Psi (x,t)-\frac{i^{\alpha }}{\hbar }V(x)\Psi (x,t),
\end{equation
or
\begin{equation}
\frac{\partial ^{\alpha }}{\partial t^{\alpha }}\Psi (x,t)=i^{\alpha }\check
D}_{\alpha ,2}\hbar \frac{\partial ^{2}}{\partial x^{2}}\Psi (x,t)-\frac
i^{\alpha }}{\hbar }V(x)\Psi (x,t).
\end{equation
By defining a new quantum diffusion constant, $D_{\alpha }=\check{D}_{\alpha
,2}\hbar ,$ we can write the time fractional Schr\"{o}dinger equation as
\begin{equation}
\frac{\partial ^{\alpha }}{\partial t^{\alpha }}\Psi (x,t)=i^{\alpha
}D_{\alpha }\frac{\partial ^{2}}{\partial x^{2}}\Psi (x,t)-\frac{i^{\alpha
}{\hbar }V(x)\Psi (x,t),\text{ }0<\alpha <1,
\end{equation
where $D_{\alpha }\rightarrow \hbar /2m$ as $\alpha \rightarrow 1.$ Note
that there is a certain amount of ambiguity in writing the time fractional
version of the Schr\"{o}dinger equation [18]. We will address this point
again in Section VII.
\section{Separable Solutions of the Time Fractional Schr\"{o}dinger Equation}
We now reconsider the time fractional Schr\"{o}dinger equation. Assuming a
separable solution of the form $\Psi (x,t)=X(x)T(t),$ we obtain the
equations to be solved for $X(x)$ and $T(t),$ respectively, as
\begin{align}
D_{\alpha }\frac{d^{2}X(x)}{dx^{2}}-\frac{V(x)}{\hbar }X(x)& =\lambda
_{n}X(x), \\
\frac{\partial ^{\alpha }T(t)}{\partial t^{\alpha }}& =i^{\alpha }\lambda
_{n}T(t),\text{ }0<\alpha <1,
\end{align
where the spatial equation [Eq. (11)] has to be solved with the appropriate
boundary conditions, and $\lambda _{n}$ is the separation constant. At this
point, the index $n$ is redundant but we keep it for the cases where
\lambda $ is discrete. Since in the limit as $\alpha \rightarrow 1,$
D_{\alpha }\rightarrow \hbar /2m$ and $\lambda _{n}\rightarrow -E_{n}/\hbar
, for physically interesting cases we will take $\lambda _{n}<0.$
\subsection{Time Dependence $-$ Method I}
Using Equation (A7) in Appendix A, we take the Laplace transform of Equation
(12):
\begin{equation}
\pounds \left\{ \frac{\partial ^{\alpha }T(t)}{\partial t^{\alpha }}\right\}
=i^{\alpha }\lambda _{n}\pounds \left\{ T(t)\right\} ,
\end{equation
to write
\begin{equation}
s^{\alpha }\widetilde{T}(s)-s^{\alpha -1}T(0)=i^{\alpha }\lambda _{n
\widetilde{T}(s),
\end{equation
where $\widetilde{T}(s)=\pounds \left\{ T(t)\right\} .$ This gives the
Laplace transform of the time dependence of the wave function as
\begin{align}
\widetilde{T}(s)& =T(0)\frac{s^{\alpha -1}}{s^{\alpha }-i^{\alpha }\lambda
_{n}}\ \\
& =T(0)\frac{s^{-1}}{1-i^{\alpha }\lambda _{n}s^{-\alpha }}.
\end{align
Using geometric series, we can also write this as
\begin{align}
\widetilde{T}(s)& =T(0)\sum_{m=0}^{\infty }\left( i^{\alpha }\lambda
_{n}s^{-\alpha }\right) ^{m}s^{-1} \\
& =T(0)\sum_{m=0}^{\infty }i^{\alpha m}\lambda _{n}^{m}s^{-m\alpha -1},
\end{align
which converges for $\left\vert i^{\alpha }\lambda _{n}s^{-\alpha
}\right\vert <1.$ The inverse Laplace transform of this can be found easily
to yield the time dependent part of the wave function as
\begin{align}
T(t)& =T(0)\sum_{m=0}^{\infty }\frac{i^{\alpha m}\lambda _{n}^{m}t^{\alpha m
}{\Gamma (1+\alpha m)} \\
& =T(0)\sum_{m=0}^{\infty }\frac{\left( i^{\alpha }\lambda _{n}t^{\alpha
}\right) ^{m}}{\Gamma (1+\alpha m)} \\
& =T(0)E_{\alpha }(i^{\alpha }\lambda _{n}t^{\alpha }),
\end{align
where $E_{\alpha }(i^{\alpha }\lambda _{n}t^{\alpha })$ is the
Mittag-Leffler function with an imaginary argument.
\subsection{Time Dependence $-$ Method II}
To find the time dependence of the wave function, an alternate path can also
be taken by directly evaluating the integral [18]
\begin{equation}
T(t)=T(0)\left[ \frac{\ 1}{2\pi i}\int_{\gamma-i\infty}^{\gamma+i\infty
\frac{e^{st}s^{\alpha-1}ds}{s^{\alpha}-i^{\alpha}\lambda_{n}}\right] .
\end{equation}
The integrand has a pole at $s_{1}=0$ due to the numerator, and a pole at
s_{2}=i\lambda_{n}^{1/\alpha}$ due to the denominator, hence can be
rewritten as (Bayin [6] and supplements of Ch. 14, [18])
\begin{equation}
T(t)=T(0)\left[ \frac{e^{i\lambda_{n}^{1/\alpha}t}}{\alpha}-\frac{\sigma
\sin\alpha\pi}{\pi}\int_{0}^{\infty}\frac{e^{-xt}x^{\alpha-1}dx}{x^{2\alpha
}-2\sigma\cos\alpha\pi\text{ }x^{\alpha}+\sigma^{2}}\right] \ .
\end{equation}
For $\lambda_{n}>0,$ the above expression is valid for all $0<\alpha<1.$ For
$\lambda_{n}<0,$ the second pole, $s_{2}=\left\vert \lambda_{n}\right\vert
^{1/\alpha}e^{i(\pi/\alpha+\pi/2)},$ has both real and imaginary parts.
Since the branch cut is located along the negative real axis, we have to
exclude the $\alpha=\frac{2}{5},\frac{2}{9},\frac{2}{13},\cdots$ values,
that is, $\alpha=2/(5+4n),$ $n=0,1,2,\ldots$, so that the pole lies inside
the contour. We now write Equation (23) a
\begin{equation}
T(t)=T(0)\left[ \frac{e^{i\lambda_{n}^{1/\alpha}t}}{\alpha
-F_{\alpha}(\sigma;t)\right] ,\text{ }\sigma=\lambda_{n}i^{\alpha},
\end{equation}
wher
\begin{equation}
F_{\alpha}(\sigma;t)=\frac{\sigma\sin\alpha\pi}{\pi}\int_{0}^{\infty}\frac
e^{-xt}x^{\alpha-1}dx}{x^{2\alpha}-2\sigma\cos\alpha\pi\text{
x^{\alpha}+\sigma^{2}}.
\end{equation}
In previous works, the above integral was assumed to be a decaying
exponential, hence neglected in subsequent calculations [18, 22-24]. This
resulted in probabilities greater than $1.$ We now show that an exact
treatment of this integral, which can not be evaluated analytically by
standard techniques, proves otherwise. Using the elegant language of Fox's
H- functions, details of which are presented in Appendices B and C, we
evaluate the above integral exactly and show that it has an oscillatory part
along with a Mittag-Leffler function as
\begin{equation}
F_{\alpha}(\sigma;t)=\pm\frac{e^{i\lambda_{n}^{1/\alpha}t}}{\alpha
-E_{\alpha }(\lambda_{n}i^{\alpha}t^{\alpha}).
\end{equation}
Thus, the second method gives the time dependent part of the wave function
as
\begin{equation}
T(t)=T(0)\left[ \frac{e^{i\lambda_{n}^{1/\alpha}t}}{\alpha}\mp\frac
e^{i\lambda_{n}^{1/\alpha}t}}{\alpha}+E_{\alpha}(\lambda_{n}i^{\alpha
}t^{\alpha})\right] .
\end{equation}
To be consistent with the robust result of Method I [Eq. (21)], we pick the
minus sign, hence the time dependent part of the wave function is again
obtained as
\begin{equation}
T(t)=E_{\alpha}(\lambda_{n}i^{\alpha}t^{\alpha}),
\end{equation}
where without loss of any generality we have set $T(0)=1.$
\subsection{Probability and Asymptotic Behavior}
\subsubsection{Total Probability:}
Since the general time dependence of the separable solutions of the time
fractional Schr\"{o}dinger equation is established as the Mittag-Leffler
function with an imaginary argument [Eq. (28)], we can write the series
\begin{align}
T(t) & =E_{\alpha}(\lambda_{n}i^{\alpha}t^{\alpha})=\sum_{\nu=0}^{\infty
\frac{(\lambda_{n}i^{\alpha}t^{\alpha})^{\nu}}{\Gamma(1+\alpha\nu)},\text{
\left\vert \lambda_{n}i^{\alpha}t^{\alpha}\right\vert <1 \\
& =\left[ 1+\frac{\lambda_{n}i^{\alpha}t^{\alpha}}{\Gamma(1+\alpha)}+\frac
\left( \lambda_{n}i^{\alpha}t^{\alpha}\right) ^{2}}{\Gamma (1+2\alpha)}
\frac{\left( \lambda_{n}i^{\alpha}t^{\alpha}\right) ^{3}}{\Gamma(1+3\alpha)
+\cdots\right] ,
\end{align}
which can be separated into its real and imaginary parts as
\begin{align}
E_{\alpha}(\lambda_{n}i^{\alpha}t^{\alpha}) & =E_{\alpha}^{R}(t)+iE_{\alpha
}^{I}(t),\text{ }0<\alpha<1, \\
& =\sum_{\nu=0}^{\infty}\frac{\lambda_{n}^{\nu}\left[ \cos\frac{\nu\alpha \p
}{2}\right] t^{\alpha\nu}}{\Gamma(1+\alpha\nu)}+i\sum_{\nu=0}^{\infty }\frac
\lambda_{n}^{\nu}\left[ \sin\frac{\nu\alpha\pi}{2}\right] t^{\alpha\nu}}
\Gamma(1+\alpha\nu)}.
\end{align}
When $\alpha=1,$ naturally, $E_{\alpha}(\lambda_{n}i^{\alpha}t^{\alpha})$
becomes the Euler equation, that is,
\begin{align}
T_{\alpha=1}(t) & =\cos\lambda_{n}t+i\sin\lambda_{n}t \\
& =e^{i\lambda_{n}t}.
\end{align}
The eigenvalues, $\lambda_{n},$ come from the solution of the space part of
the Schr\"{o}dinger equation with the appropriate boundary conditions [Eq.
(11)]. Note that we can also write Equation (32) as
\begin{equation}
T(t)=E_{\alpha}\left( \lambda_{n}\cos^{1/\nu}\left( \frac{\nu\alpha\pi}{2
\right) \text{ }t^{\alpha}\right) +iE_{\alpha}\left(
\lambda_{n}\sin^{1/\nu}\left( \frac{\nu\alpha\pi}{2}\right) \text{
t^{\alpha}\right) ,
\end{equation}
which in terms of $H-$functions becomes (Appendix B, [25])
\begin{align}
T(t) & =\left[ H_{1,2}^{1,1}\left( \left. -\lambda_{n}\cos^{1/\nu}\left(
\frac{\nu\alpha\pi}{2}\right) \text{ }t^{\alpha}\right\vert
_{(0,1),(0,\alpha )}^{(0,1)}\right) \right. \notag \\
& \left. +iH_{1,2}^{1,1}\left( \left. -\lambda_{n}\sin^{1/\nu}\left( \frac
\nu\alpha\pi}{2}\right) \text{ }t^{\alpha}\right\vert _{(0,1),(0,\alpha
)}^{(0,1)}\right) \right] .
\end{align}
An important consequence of this result is that for the time fractional Sch
\"{o}dinger equation, the total probability, $\int_{-\infty}^{+\infty
}\left\vert \Psi(x,t)\right\vert dx,$ which is a function of time:
\begin{align}
\int_{-\infty}^{+\infty}\left\vert \Psi(x,t)\right\vert dx & =\left\vert
T(t)\right\vert ^{2}\int_{-\infty}^{+\infty}\left\vert X(x)\right\vert ^{2}dx
\notag \\
& =\left[ \left\vert E_{\alpha}^{R}(t)\right\vert ^{2}+\left\vert E_{\alpha
}^{I}(t)\right\vert ^{2}\right] \int_{-\infty}^{+\infty}\left\vert
X(x)\right\vert dx \notag \\
& =\left\vert E_{\alpha}^{R}(t)\right\vert ^{2}+\left\vert
E_{\alpha}^{I}(t)\right\vert ^{2},
\end{align}
is not conserved. The normalization constant is fixed by the total
probability at $t=0$:
\begin{equation}
\int_{-\infty}^{+\infty}\left\vert \Psi(x,0)\right\vert dx=\int_{-\infty
}^{+\infty}\left\vert X(x)\right\vert dx=1.
\end{equation}
\subsubsection{Small Time Behavior:}
For small times, we use the $H-$function representation of the
Mittag-Leffler function:
\begin{equation}
E_{\alpha }(z)=H_{1,2}^{1,1}\left( -z_{(0,1),(0,\alpha )}^{(0,1)}\right) ,
\end{equation
and write the time dependence as
\begin{equation}
T(t)=H_{1,2}^{1,1}\left( \left. -\lambda _{n}i^{\alpha }t^{\alpha
}\right\vert _{(0,1),(0,\alpha )}^{(0,1)}\right) .
\end{equation
Using Equation (C3) we again obtain the series [Eq. (32)]
\begin{align}
T(t)& =\left[ 1+\frac{\cos (\alpha \pi /2)}{\Gamma (1+\alpha )}(\lambda
_{n}t^{\alpha })+\frac{\cos (\alpha \pi )}{\Gamma (1+2\alpha )}(\lambda
_{n}t^{\alpha })^{2}+\frac{\cos (3\alpha \pi /2)}{\Gamma (1+3\alpha )
(\lambda _{n}t^{\alpha })^{3}+\cdots \right] \notag \\
& +i\left[ \frac{\sin (\alpha \pi /2)}{\Gamma (1+\alpha )}(\lambda
_{n}t^{\alpha })+\frac{\sin (\alpha \pi )}{\Gamma (1+2\alpha )}(\lambda
_{n}t^{\alpha })^{2}+\frac{\sin (3\alpha \pi /2)}{\Gamma (1+3\alpha )
(\lambda _{n}t^{\alpha })^{3}+\cdots \right] ,
\end{align
which for small times gives the total probability as
\begin{align}
\int_{-\infty }^{+\infty }\left\vert \Psi (x,t)\right\vert ^{2}dx&
=\left\vert T(t)\right\vert ^{2}\int_{-\infty }^{+\infty }\left\vert
X(x)\right\vert ^{2}dx \\
& \simeq \left( 1+\frac{2\cos (\alpha \pi /2)(\lambda _{n}t^{\alpha })}
\alpha \Gamma (\alpha )}+0(t^{2\alpha })\right) .
\end{align
For $\lambda _{n}<0,$ this makes the total probability $\leq 1$ for small
times.
\subsubsection{Large Time Behavior:}
For large times, we use the expansion in Equation (C7) to write
\begin{align}
T(t)& =\sum_{v=0}^{\infty }\frac{-1}{\Gamma (1-\nu \alpha )(\lambda
_{n}i^{\alpha }t^{\alpha })^{1+\nu }} \\
& =\left[ -\frac{1}{\lambda _{n}i^{\alpha }t^{\alpha }}-\frac{1}{\Gamma
(1-\alpha )(\lambda _{n}i^{\alpha }t^{\alpha })^{2}}-\frac{1}{\Gamma
(1-2\alpha )(\lambda _{n}i^{\alpha }t^{\alpha })^{3}}-\cdots \right] .
\end{align
We can also write this in terms of its real and imaginary parts as
\begin{align}
T(t)& =\left[ -\frac{\cos (\alpha \pi /2)}{\lambda _{n}t^{\alpha }}-\frac
\cos (\alpha \pi )}{\Gamma (1-\alpha )(\lambda _{n}t^{\alpha })^{2}}-\frac
\cos (3\alpha \pi /2)}{\Gamma (1-2\alpha )(\lambda _{n}t^{\alpha })^{3}
+\cdots \right] \notag \\
& +i\left[ \frac{\sin (\alpha \pi /2)}{\lambda _{n}t^{\alpha }}+\frac{\sin
(\alpha \pi )}{\Gamma (1-\alpha )(\lambda _{n}t^{\alpha })^{2}}+\frac{\sin
(3\alpha \pi /2)}{\Gamma (1-2\alpha )(\lambda _{n}t^{\alpha })^{3}}+\cdots
\right] ,
\end{align
In agreement with Diethelm et. al. [28], for large times, the total
probability decays as
\begin{equation}
\left\vert T(t)\right\vert ^{2}\varpropto \frac{1}{\lambda
_{n}^{2}t^{2\alpha }}.
\end{equation}
Since the wave function is zero on and outside the boundary, the energy does
not leak out. Due to the presence of the fractional time derivative, the
particle inside the infinite well is not free according to the ordinary Sch
\"{o}dinger equation. In Section V, we introduce the \textit{effective
potential} in terms of the ordinary Schr\"{o}dinger equation, where the
imaginary part of the complex effective potential describes the dissipation
process. Hence, the energy is basically exchanged within the box, between
the physical agent that generates the effective potential and the particle.
\section{The Box Problem}
\ \ We now consider a particle in an infinite potential well, where
\begin{equation}
V(x)=\left\{
\begin{tabular}{lll}
$0$ & , & $0<x<a$ \\
& & \\
$\infty $ & , & elsewher
\end{tabular
\ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \right. .
\end{equation
For the separable solutions of the time fractional schr\"{o}dinger equation
[Eq. (10)], we have already determined the time dependent part [Eq. (28)] of
the wave function, hence we can write $\Psi (x,t)=E_{\alpha }(\lambda
_{n}i^{\alpha }t^{\alpha })X(x),$ where the spatial part comes from the
solution of Equation (11) with the appropriate boundary conditions. For\ the
particle in a box with impenetrable walls, we solve
\begin{equation}
D_{\alpha }\frac{d^{2}X(x)}{dx^{2}}=\lambda _{n}X(x),\text{ \ }\Psi
(0,t)=\Psi (a,t)=0,
\end{equation
which yields the eigenvalues and the eigenfunctions as
\begin{equation}
\lambda _{n}=-D_{\alpha }\left( \frac{n\pi }{a}\right) ^{2},\text{ \
X_{n}(x)=C_{0}\sin \left( \frac{n\pi }{a}x\right) ,\text{ }n=1,2,\ldots
\text{ }D_{\alpha }>0.
\end{equation
Since we are considering the time fractional Schr\"{o}dinger equation, non
locality is in terms of time, which means that the system has memory. For
the separable solutions, the differential equation to be solved for the
space part of the wave equation is an ordinary differential equation, hence
the boundary conditions used in Eq. (49) are the natural boundary conditions
for impenetrable walls. Using the normalization condition $\int \left\vert
\Psi _{n}(x,0)\right\vert ^{2}dx=1,$ we write the complete wave function as
\begin{equation}
\Psi _{n}(x,t)=\sqrt{\frac{2}{a}}\left[ \sin \left( \frac{n\pi }{a}x\right)
\right] E_{\alpha }(\lambda _{n}i^{\alpha }t^{\alpha }).
\end{equation
The total probability is time dependent, which for small times behaves as
\begin{align}
\int \left\vert \Psi _{n}(x,t)\right\vert ^{2}dx& =\left\vert E_{\alpha
}^{R}(t)\right\vert ^{2}+\left\vert E_{\alpha }^{I}(t)\right\vert ^{2} \\
& \simeq 1-\frac{2\cos (\alpha \pi /2)(\left\vert \lambda _{n}\right\vert
t^{\alpha })}{\alpha \Gamma (\alpha )}+0(t^{2\alpha }),
\end{align
and for large times decays as
\begin{equation}
\lim_{t\rightarrow \infty }\int \left\vert \Psi _{n}(x,t)\right\vert
^{2}dx\thicksim \frac{1}{\lambda _{n}^{2}t^{2\alpha }}\rightarrow 0.
\end{equation
This is in contrast with the previous works [18, 22-24], where for large
times the total probability is constant and greater than 1:
\begin{equation}
\lim_{t\rightarrow \infty }\int \left\vert \Psi _{n}(x,t)\right\vert ^{2}dx
\frac{1}{\alpha ^{2}}\geq 1.
\end{equation}
\subsection{The Energy Operator and the Box Problem}
For the time fractional Schr\"{o}dinger equation, we take the new energy
operator as
\begin{equation}
E=-\frac{\hbar}{i^{\alpha}}\frac{\partial^{\alpha}}{\partial t^{\alpha}},
\end{equation}
where the Hamiltonian is
\begin{equation}
H_{\alpha}=-D_{\alpha}\hbar\frac{\partial^{2}}{\partial x^{2}}+V(x).
\end{equation}
With this operator, the energy values are real but time dependent
\begin{align}
E & =-\frac{\hbar}{i^{\alpha}}\int\Psi^{\ast}(x,t)\frac{\partial^{\alpha}}
\partial t^{\alpha}}\Psi(x,t)dx \notag \\
& =-\hbar\lambda_{n}\left\vert E_{\alpha}(\lambda_{n}i^{\alpha}t^{\alpha
})\right\vert ^{2}\int\ \left\vert X(x)\right\vert ^{2}dx \notag \\
& =-\hbar\lambda_{n}\left\vert E_{\alpha}(\lambda_{n}i^{\alpha}t^{\alpha
})\right\vert ^{2}
\end{align}
For the box problem, the energy eigenvalues are now obtained as
\begin{eqnarray}
E_{n} &=&-\frac{\hbar }{i^{\alpha }}\int \Psi _{n}^{\ast }(x,t)\frac
\partial ^{\alpha }}{\partial t^{\alpha }}\Psi _{n}(x,t)dx, \\
&=&\frac{\hbar \pi ^{2}n^{2}D_{\alpha }}{a^{2}}\ \left[ \left\vert E_{\alpha
}^{R}(t)\right\vert ^{2}+\left\vert E_{\alpha }^{I}(t)\right\vert ^{2}\right]
\end{eqnarray
which for small times are given as \
\begin{equation}
E_{n}\simeq \frac{\hbar \pi ^{2}n^{2}D_{\alpha }}{a^{2}}\left( 1-\frac{2\cos
(\alpha \pi /2)(\left\vert \lambda _{n}\right\vert t^{\alpha })}{\alpha
\Gamma (\alpha )}\right) . \notag
\end{equation
For large times $(\alpha \neq 1),$ the system dissipates all of its energy
as
\begin{equation}
\lim_{t\rightarrow \infty }E_{n}\thicksim \frac{\hbar \pi ^{2}n^{2}D_{\alpha
}}{a^{2}\lambda _{n}^{2}t^{2\alpha }}\rightarrow 0.
\end{equation
In the limit as $\alpha \rightarrow 1,$ $D_{\alpha }\rightarrow $ $\hbar /2m
, the energy eigenvalues approach to their usual values predicted by the Sch
\"{o}dinger equation:
\begin{equation}
E_{n}=\frac{\hbar ^{2}\pi ^{2}n^{2}}{2ma^{2}}.
\end{equation}
Note that the ordinary energy operator, $i\hbar \dfrac{\partial }{\partial t}
$, yields complex eigenvalues, which only in the limit as $t\rightarrow
\infty $ approach to real values that are larger than their usual values in
Equation (62) by a factor of $1/\nu ^{2}$.
\section{The Effective Potential\textbf{\ }}
To gain a better understanding of the effects of the fractional time
derivative, we look for an effective potential in the (ordinary) Schr\"{o
dinger equation that yields the same wave function. For simplicity we will
set $V(x)=0$ in Equation (11). We now use the following useful formula [18]:
\begin{equation}
\frac{dT(t)}{dt}=_{0}^{C}D_{t}^{1-\alpha }\left[ _{0}^{C}D_{t}^{\alpha }T(t
\right] +\frac{\left[ _{0}^{C}D_{t}^{\alpha }T(t)\right] _{t=0}}{\Gamma
(\alpha )t^{1-\alpha }},
\end{equation
which is particularly helpful in isolating and exploring the effects of
fractional time derivatives in terms of an effective potential. Using
separable solutions of the time fractional Schr\"{o}dinger equation [Eq.
(12)], we can write
\begin{equation}
\frac{dT(t)}{dt}=i^{\alpha }\lambda _{n}\left[ _{0}^{C}D_{t}^{1-\alpha }T(t)
\frac{T(0)}{\Gamma (\alpha )t^{1-\alpha }}\right] .
\end{equation
The quantity inside the square brackets is nothing but the Riemann-Liouville
derivative [Eq. (A5)], thus
\begin{equation}
\frac{dT(t)}{dt}=i^{\alpha }\lambda _{n}\left[ _{0}^{R-L}D_{t}^{1-\alpha
}T(t)\right] .
\end{equation}
To define an effective potential for the separable solutions, we also write
the (ordinary) Schr\"{o}dinger equation a
\begin{equation}
i\hbar\frac{dT(t)}{dt}X(x)=-\frac{\hbar^{2}}{2m}\frac{d^{2}X(x)}{dx^{2}
T(t)+V_{eff.}(t)X(x)T(t),
\end{equation}
and substitute the wave function found from the solution of the time
fractional Schr\"{o}dinger equation [Eq. (65) and Eq. (11) with $V(x)=0$]:
\begin{equation}
i\hbar i^{\alpha}\lambda_{n}\left[ _{0}^{R-L}D_{t}^{1-\alpha}T(t)\right]
X(x)=-\frac{\hbar^{2}}{2m}\frac{\lambda_{n}}{D_{\alpha}
X(x)T(t)+V_{eff.}(t)X(x)T(t),
\end{equation}
This yields the effective potential as
\begin{equation}
V_{eff.}(t)=i^{1+\alpha}\hbar\lambda_{n}\frac{_{0}^{R-L}D_{t}^{1-\alpha}T(t
}{T(t)}+\frac{\hbar^{2}}{2m}\frac{\lambda_{n}}{D_{\alpha}}.
\end{equation}
Since the time dependence is given as $T(t)=E_{\alpha}(\lambda_{n}i^{\alpha
}t^{\alpha}),$ we can also write
\begin{equation}
V_{eff.}(t)=i^{1+\alpha}\hbar\lambda_{n}\frac{_{0}^{R-L}D_{t}^{1-\alpha
}E_{\alpha}(\lambda_{n}i^{\alpha}t^{\alpha})}{E_{\alpha}(\lambda_{n}i^
\alpha }t^{\alpha})}+\frac{\hbar^{2}}{2m}\frac{\lambda_{n}}{D_{\alpha}}.
\end{equation}
In the limit as $\alpha\rightarrow1,$ $D_{\alpha}\rightarrow\dfrac{\hbar}{2m
,$ the Mittag-Leffler function becomes the exponential function,
T(t)=e^{i\lambda_{n}t}$, hence the effective potential vanishes as it
should. With this effective potential, the Schr\"{o}dinger equation yields
the same wave function as the time fractional Schr\"{o}dinger equation and
satisfies the same boundary conditions. In other words, the wave function
found from the time fractional Schr\"{o}dinger equation, satisfies the Sch
\"{o}dinger equation with the above effective potential.
Using the $H-$function [25] representation of the Mittag-Leffler function
(Appendix B):
\begin{equation}
E_{\alpha }(\lambda _{n}i^{\alpha }t^{\alpha })=H_{1,2}^{1,1}\left( \left.
-\lambda _{n}i^{\alpha }t^{\alpha }\right\vert _{(0,1),(0,\alpha
)}^{(0,1)}\right)
\end{equation
and Equation (B7) for the Riemann-Liouville derivative of the $H-$function,
we can also write the following closed expression for the effective
potential:
\begin{equation}
V_{eff.}(t)=\frac{\left( i^{1+\alpha }\hbar \lambda _{n}\right) }
t^{1-\alpha }}\frac{H_{2,3}^{1,2}\left( \left. -\lambda _{n}i^{\alpha
}t^{\alpha }\right\vert _{(0,1),(0,\alpha ),(1-\alpha ,\alpha )}^{(0,\alpha
),(0,1)}\right) }{H_{1,2}^{1,1}\left( \left. -\lambda _{n}i^{\alpha
}t^{\alpha }\right\vert _{(0,1),(0,\alpha )}^{(0,1)}\right) }+\frac{\hbar
^{2}}{2m}\frac{\lambda _{n}}{D_{\alpha }}.
\end{equation
Using the symmetries of the $H-$function [25], we can further simplify the
above equation as
\begin{equation}
V_{eff.}(t)=\frac{\left( i^{1+\alpha }\hbar \lambda _{n}\right) }
t^{1-\alpha }}\frac{H_{1,2}^{1,1}\left( \left. \left( (-\lambda
_{n})^{1/\alpha }it\right) ^{\alpha }\right\vert _{(0,1),(1-\alpha ,\alpha
)}^{(0,1)}\right) }{H_{1,2}^{1,1}\left( \left. -\lambda _{n}i^{\alpha
}t^{\alpha }\right\vert _{(0,1),(0,\alpha )}^{(0,1)}\right) }+\frac{\hbar
^{2}}{2m}\frac{\lambda _{n}}{D_{\alpha }},
\end{equation
or as
\begin{equation}
V_{eff.}(t)=\frac{\left( i^{1+\alpha }\hbar \lambda _{n}\right) }
t^{1-\alpha }}\frac{E_{\alpha ,\alpha }(\lambda _{n}i^{\alpha }t^{\alpha })}
E_{\alpha }(\lambda _{n}i^{\alpha }t^{\alpha })}+\frac{\hbar ^{2}}{2m}\frac
\lambda _{n}}{D_{\alpha }},
\end{equation
where
\begin{equation}
E_{\alpha ,\beta }(z)=\sum_{k=0}^{\infty }\frac{z^{k}}{\Gamma (\alpha
k+\beta )},\text{ }\alpha ,\beta \in \mathbb{C},\text{ }\func{Re}(a)>0,\text{
}\func{Re}(\beta )>0,\text{ }z\in \mathbb{C},
\end{equation
is the generalized Mittag-Leffler function, and $E_{\alpha ,1}(z)=E_{\alpha
}(z)$ [25].
\subsection{Hamiltonian}
Since the effective potential is time dependent and complex, the energies
corresponding to the operator $i\hbar\dfrac{\partial}{\partial t}$ are also
time dependent and complex, that is, the ordinary Hamiltonian operator:
\begin{equation}
H=-\frac{\hbar^{2}}{2m}\frac{d^{2}}{dx^{2}}+V_{eff.}(t)
\end{equation}
is not hermitian. However, complex potentials in quantum mechanics have
found interesting applications in the study of dissipative systems [29-32].
Let us now write the effective potential for the time fractional Schr\"{o
dinger equation explicitly. Using Equation (73) and the expression
\begin{equation}
E_{\alpha ,\alpha }(\lambda _{n}i^{\alpha }t^{\alpha })=\sum_{k=0}^{\infty
\frac{\lambda _{n}^{k}\cos (k\alpha \pi /2)t^{\alpha k}}{\Gamma (1+k\alpha )
+i\sum_{k=0}^{\infty }\frac{\lambda _{n}^{k}\sin (k\alpha \pi /2)t^{\alpha k
}{\Gamma (1+k\alpha )},
\end{equation
we can write the effective potential in terms of its real and imaginary
parts as
\begin{equation}
V_{eff.}(t)=V_{eff.}^{R}(t)+iV_{eff.}^{I}(t),
\end{equation
where
\begin{align}
V_{eff.}^{R}(t)& =\frac{\hbar ^{2}\lambda _{n}}{2mD_{\alpha }}-\frac{\hbar
\lambda _{n}\sin (\alpha \pi /2)t^{\alpha -1}}{\Gamma (\alpha )} \notag \\
& -\hbar \lambda _{n}^{2}\sin (\alpha \pi )\left( \frac{1}{\Gamma (2\alpha )
-\frac{1}{\Gamma (1+\alpha )\Gamma (\alpha )}\right) t^{2\alpha -1}+\cdots
\end{align
and
\begin{align}
V_{eff.}^{I}(t)& =\frac{\hbar ^{2}\lambda _{n}}{2mD_{\alpha }}-\frac{\hbar
\lambda _{n}\cos (\alpha \pi /2)t^{\alpha -1}}{\Gamma (\alpha )} \notag \\
& +\hbar \lambda _{n}^{2}\sin (\alpha \pi )\left( \frac{1}{\Gamma (1+\alpha
)\Gamma (\alpha )}+\frac{\cos (\alpha \pi )}{\Gamma (2\alpha )}-\frac{2\cos
^{2}(\alpha \pi /2)}{\Gamma (\alpha )\Gamma (1+\alpha )}\right) t^{2\alpha
-1}+\cdots \text{ }.
\end{align
The imaginary part of the complex effective potential plays an important
role in describing the decay of the energy of the stationary states of the
time-fractional Schr\"{o}dinger equation.
\section{The Mittag-Leffler Function, $E_{\protect\alpha }(\protect\lambda
_{n}i^{\protect\alpha }t^{\protect\alpha }),$ and it's Purely Decaying and
Oscillating Parts}
Solving\ the Schr\"{o}dinger equation [Eq. (66)] with the above effective
potential, we now obtain the following expression for the time dependent
part of the separable solutions of the time fractional Schr\"{o}dinger
equation:
\begin{equation}
T(t)=\left[ \exp \left( \dfrac{1}{\hbar }\dint_{0}^{t}V_{eff.}^{I}(t^{\prime
})dt^{\prime }\right) \right] \exp \left( \dfrac{i}{\hbar }\left[ \dfrac
\hbar ^{2}\lambda _{n}t}{2mD_{\alpha }}-\dint_{0}^{t}V_{eff.}^{R}(t^{\prime
})dt^{\prime }\right] \right) ,
\end{equation
where $\lambda _{n}$ is a separation constant. This is naturally equivalent
to the solution found before [Eq. (28)], that is, $T(t)=E_{\alpha }(\lambda
_{n}i^{\alpha }t^{\alpha })$. Note that when $\alpha =1,$ the effective
potential vanishes, hence the time dependence reduces t
\begin{equation}
T(t)=\exp \left( -\frac{iE_{n}t}{\hbar }\right) ,\text{ }E_{n}=\dfrac{\hbar
^{2}\lambda _{n}}{2mD_{1}}.
\end{equation}
In other words, the time dependent part of the separable solution of the
time fractional Schr\"{o}dinger equation, $E_{\alpha }(\lambda _{n}i^{\alpha
}t^{\alpha }),$ can be written, not as the sum, but as the product of its
purely oscillating and purely exponentially decaying/growing parts. Despite
the fact that the effective potential corresponding to the time fractional
Schr\"{o}dinger equation is complex, the advantage of the effective
potential approach is that one can use the mathematical structure of the Sch
\"{o}dinger theory.
Finally, using the expansion
\begin{align}
V_{eff.}(t)& \simeq \left( \frac{\hbar ^{2}\lambda _{n}}{2mD_{\alpha }}
\frac{\hbar \lambda _{n}\sin (\alpha \pi /2)}{\Gamma (\alpha )}t^{\alpha
-1}+0(t^{2\alpha -1})\right) \notag \\
& +i\left( \frac{\hbar \lambda _{n}\cos (\alpha \pi /2)}{\Gamma (\alpha )
t^{\alpha -1}+0(t^{2\alpha -1})\right) ,
\end{align
the small time behavior of $T(t),$ to lowest order, becomes
\begin{equation}
T(t)\simeq \left[ \exp \left( \frac{\lambda _{n}\cos (\alpha \pi /2)}{\alpha
\Gamma (\alpha )}t^{\alpha }\right) \right] .\left[ \exp \left( \frac{i}
\hbar }\left( \frac{\hbar \lambda _{n}\sin (\alpha \pi /2)}{\alpha \Gamma
(\alpha )}t^{\alpha }\right) \right) \right] .
\end{equation
For the box problem, where $\lambda _{n}=-D_{\alpha }\left( \frac{n\pi }{a
\right) ^{2},$ $n=1,2,\ldots ,$ $D_{\alpha }>0,$ the total probability
decays exponentially, which for small times becomes
\begin{equation}
\left\vert T(t)\right\vert ^{2}\simeq 1-\frac{2\cos (\alpha \pi /2)}{\alpha
\Gamma (\alpha )}\left\vert \lambda _{n}\right\vert t^{\alpha }+0(t^{2\alpha
}).
\end{equation
This is in accordance with our previous result [Eq. (43)] obtained from
T(t)=E_{\alpha }(\lambda _{n}i^{\alpha }t^{\alpha })$.
\section{Conclusions}
The fact that anomalous diffusion can be studied by fractional calculus has
attracted researchers from many different branches of science and
engineering into this intriguing branch of mathematics [1-6]. Since 60's,
successful examples are given in random walk, anomalous diffusion, economics
and finance. Introduction of fractional calculus into a certain branch of
science is usually initiated by replacing certain derivatives in the
evolution or the transport equations with their fractional counterparts. A
general feature of these applications is that replacing a time derivative
with its fractional counterpart basically introduces memory effects into the
system and makes the process non Markovian, while a replacement of a space
derivative introduces global or non local effects. However, it is important
to note that not all such effects can be covered by fractional
generalizations of the basic equations [20, 21].
In this paper, we first established the time dependent part of the separable
solutions of the time fractional Schr\"{o}dinger equation as the
Mittag-Leffler function with an imaginary argument by two different methods.
This showed that the total probability is $\leq 1$ and decays with time. We
also introduced the effective potential approach, where with the effective
potential, the (ordinary) Schr\"{o}dinger equation yields the same wave
function, with the same boundary conditions, as that the time fractional Sch
\"{o}dinger equation.
For the energy, the standard operator used in Schr\"{o}dinger theory,
E=i\hbar \dfrac{\partial }{\partial t},$ yields complex values since the
corresponding Hamiltonian is complex and non hermitian. \ On the other hand,
the operator $E=-\dfrac{\hbar }{i^{\alpha }}\dfrac{\partial ^{\alpha }}
\partial t^{\alpha }},$ which follows naturally from the time fractional Sch
\"{o}dinger equation, yields real but time dependent energies [Eq. (58)]. In
the light of these, we also discussed the particle in a box problem, which
is essentially the prototype of a device or a detector with internal degrees
of freedom. With the operator $E=-\dfrac{\hbar }{i^{\alpha }}\dfrac{\partial
^{\alpha }}{\partial t^{\alpha }},$ the energy eigenvalues are real and time
dependent, and decay with the imaginary part of the complex effective
potential [Eq. (79)] as
\begin{equation}
E_{n}=\frac{\hbar \pi ^{2}D_{\alpha }n^{2}}{a^{2}}\exp \left( \dfrac{2}
\hbar }\dint_{0}^{t}V_{eff.}^{I}(t^{\prime })dt^{\prime }\right) .
\end{equation
Since the wave function is zero on and outside the boundary, the energy is
exchanged within the box between the source of the effective potential and
the particle. The imaginary part of the complex potential describes the
dissipative processes involved.
Complex potentials have been very useful in the study of dissipative
processes in both classical and quantum physics. Sinha et. al. [29] have
extended the factorization technique of Kuru and Negro [30] to study complex
potentials in classical systems that are analogues of non hermitian quantum
mechanical systems. In the context of quantum device modeling, a general non
hermitian Hamiltonian operator has been investigated by Ferry et. al. [31].
They have also discussed the general behavior of complex potentials with
applications to ballistic quantum dots and its implications for trajectories
and histories in the dots. Barraf [32] have used uniform complex potentials
to model particle capture from an incident beam by quantum wells. We have
pointed out that there is a certain ambiguity in writing the time fractional
Schr\"{o}dinger equation. For separable solutions, the time fractional Sch
\"{o}dinger equation that Naber [18] and Dong and Xu [22] used gives the
time dependence of the wave function as $E_{\alpha }(\lambda
_{n}(-i)^{\alpha }t^{\alpha }).$ However, for the total probability [Eq.
(37)] and the energy [Eq. (58)], both conventions yield the same time
dependence since
\begin{equation}
\left\vert E_{\alpha }(\lambda _{n}(-i)^{\alpha }t^{\alpha })\right\vert
=\left\vert E_{\alpha }(\lambda _{n}i^{\alpha }t^{\alpha })\right\vert
\text{ }0<\alpha <1.
\end{equation}
|
\section{\label{section-introduction}Introduction}
The Jaynes-Cummings model (JCM) was originally
proposed for describing the spontaneous emission
in a semi-classical manner
by Jaynes and Cummings during the 1960s
\cite{Jaynes1963,Shore1993,Louisell1973,Schleich2001}.
The JCM consists of a single two-level atom
and a single cavity mode of the electromagnetic field.
The JCM interaction between the atom and the cavity mode is obtained
by the rotating wave approximation,
so that
each photon creation causes an atomic de-excitation
and
each photon annihilation causes an atomic excitation.
The JCM is a soluble quantum mechanical model.
Moreover, it is simple enough for expressing the basic and most important characteristics
of the matter-radiation interaction.
Because of these distinct advantages,
both theoretical and experimental researchers
in the field of quantum optics
have been studying the JCM eagerly for decades.
If we initially prepare the atom in the ground state
and the cavity mode in the coherent state,
the JCM shows the periodic spontaneous collapse and the revival of the Rabi oscillations
during its time-evolution
\cite{Cummings1965,Eberly1980,Narozhny1981,Yoo1981,Yoo1985}.
This phenomenon was experimentally demonstrated in the 1980s
\cite{Rempe1987}.
We can regard this phenomenon as a direct evidence for discreteness of energy states of photons.
Thus, the JCM has a fully quantum property,
which cannot be explained by semi-classical physics.
In this paper,
we investigate the non-dissipative JCM at finite low temperature.
We assume that the atom and the cavity mode are initially
in the thermal equilibrium state and the thermal coherent state, respectively,
at a certain finite low temperature
$\beta[=1/(k_{\mbox{\scriptsize B}}T)]$.
Thus, we can describe an initial probability distribution of quantum states
of the system with the canonical ensemble.
Moreover,
we assume the time-evolution of the system is governed
by a unitary operator generated with the JCM Hamiltonian.
This implies that the system does not suffer from dissipation and its time-evolution is reversible.
The Thermo Field Dynamics (TFD) formalism was developed by Takahashi and Umezawa
for dealing with phenomena in isolated systems during the 1970s
\cite{Takahashi1975,Ojima1981,Umezawa1982,Umezawa1992}.
The TFD formalism has a wide range of applications in equilibrium situations of closed systems, as well.
In this paper,
we think about applications of the TFD
to non-dissipative closed systems.
The TFD formalism helps us to be more successful than conventional formalisms do as follows:
calculating an expectation value of a pure state created by the TFD mechanics,
we can obtain a thermal average in statical mechanics.
In return for this benefit,
the TFD formalism requires us to let the original Hilbert space double in size for the tensor product.
Then,
the TFD formalism induces thermal-like noises in two mode squeezed states
generated by the thermal Bogoliubov transformation.
Describing the JCM at finite low temperature with the TFD formalism,
we can write down a thermal average (an expectation value) of an observable
as a series expansion
containing powers of $\theta(\beta)$ in a systematic manner.
[The explicit form of the function $\theta(\beta)$ is given by
$\theta(\beta)=\mbox{arctanh}[\exp(-\beta\hbar\omega/2)]$,
where $\omega$ is a frequency of the cavity mode.
We note that $\theta(\beta)\to +0$ as $\beta\to +\infty$.]
We call this series the low-temperature expansion.
This prescription is a new key point of this paper as compared with the other past works.
Strictly speaking,
because we introduce a finite temperature into both the atom and the cavity field,
the perturbation theory has to include two parameters of the temperature.
In this paper,
we let $\Theta(\beta)$ and $\theta(\beta)$ denote the parameters of the temperature
for the atom and the cavity field, respectively.
[The explicit form of the function $\Theta(\beta)$ is given by
$\Theta(\beta)=\arctan[\exp(-\beta\hbar\omega_{0}/2)]$,
where $\omega_{0}$ is a transition frequency of the two-level atom.
We note that $\Theta(\beta)\to +0$ as $\beta\to +\infty$.]
However,
the Hilbert space of the atom is two-dimensional,
so that we can solve the problem concerning the atom exactly.
On the other hand,
the dimension of the Hilbert space of the cavity field is infinite,
so that we cannot give a rigorous treatment for the problem
concerning the cavity field,
which interacts with the atom.
This is the reason why the low-temperature expansion contains
powers of $\theta(\beta)$, but not $\Theta(\beta)$.
As mentioned before,
in this paper,
we initially prepare the cavity mode in the thermal coherent state,
which is
proposed by Barnett, Knight, Mann and Revzen
as a natural extension of the zero-temperature ordinary coherent state according to the TFD
\cite{Barnett1985,Mann1989a}.
Then, we investigate the time-evolution of the atomic population inversion
using the low-temperature expansion up to the third order terms
in $\theta(\beta)$.
We regard this low-temperature expansion as the third order perturbation theory.
To take another approach,
we give an intuitive discussion and obtain the following expectation:
the period of the revival of the Rabi oscillations becomes longer as the temperature rises,
where the temperature is finite but low enough and varies in the neighborhood of the absolute zero.
Numerical results obtained by the third order perturbation theory
reproduce well this temperature dependence of the period.
Here, we write about related works.
The thermal JCM without dissipation is investigated
in Refs.~\cite{Arroyo-Correa1990,Liu1992,Chumakov1993,Klimov1999,Azuma2008,Azuma2010}.
In these references,
the cavity field is prepared initially in a thermal equilibrium state,
whose statistical behavior is given by the Bose-Einstein distribution.
Then,
the system consisting of
the atom and the cavity mode is assumed to evolve
with a unitary operator generated by the JCM Hamiltonian,
so that its time-evolution is reversible.
However,
in these works,
thermal coherent states are not considered to be the initial states of the cavity field.
The thermal JCM with dissipation is investigated
in Refs.~\cite{Eiselt1991,Daeubler1992,Murao1995,deFaria1999,Kuang1997}.
In Refs.~\cite{Eiselt1991,Daeubler1992,Murao1995,deFaria1999},
the cavity damping in the JCM is discussed.
In these works,
the equation of motion for the density operator $\rho$ includes
the term $\kappa L_{\mbox{\scriptsize ir}}(\rho)$,
which describes the irreversible motion caused by the cavity damping.
A typical form of $L_{\mbox{\scriptsize ir}}(\rho)$ is given by
$(2a\rho a^{\dagger}-\rho a^{\dagger}a-a^{\dagger}a\rho)$
in Refs.~\cite{Daeubler1992,deFaria1999},
where $a^{\dagger}$ and $a$ are the creation and annihilation operators of the cavity photons,
respectively.
On the other hand, in Ref.~\cite{Kuang1997},
the phase damping in the JCM is discussed.
In this work,
the equation of motion for the density operator $\rho$ includes the term
$(-\gamma[H,[H,\rho]])$,
which describes the phase damping.
To obtain this term,
we assume the system to interact with a heat-bath environment,
namely an infinite set of harmonic oscillators.
Considering the Liouville-von Neumann equation for the system and the heat bath,
and applying the Markovian approximation to it,
we can derive the term of the phase damping.
Some researchers attempt to extend the JCM Hamiltonian according to the TFD formalism
\cite{Barnett1985,Azuma2010,Fan2004}.
However,
in these works,
they introduce the temperature only into the cavity field
and let the atom maintain the temperature at absolute zero.
In this paper,
we explain that such extended JCM Hamiltonians cannot be genuine ones
based on the TFD formalism.
The organization of this paper is as follows:
In Sec.~\ref{section-review-JCM},
we give a brief review of the original JCM.
In Sec.~\ref{section-review-TFD},
we give a brief review of the TFD.
In Sec.~\ref{section-thermal-effects-period-qualitative-estimation},
we give an intuitive discussion about the thermal effects found
in the period of the revival of the Rabi oscillations.
In Sec.~\ref{section-formulation-perturbetion-theory},
we formulate the perturbation theory under the low-temperature limit.
In Sec.~\ref{section-comparison-of-TFD-and-Liouville-von-Neumann-equation},
we compare our TFD formalism applied to the JCM with the Liouville-von Neumann equation.
Although the equilibrium TFD is equivalent to the Liouville-von Neumann equation in principle,
we find that the TFD formalism is more effective and easier than the Liouville-von Neumann equation
for deriving the low-temperature expansion.
In Secs.~\ref{section-0th-order-correction-term},
\ref{section-1st-order-correction-term}
and
\ref{section-2nd-order-correction-term},
we calculate the zero-th, first and second order perturbation corrections, respectively.
In Sec.~\ref{section-numerical-calculations},
we give numerical results of calculations.
In Sec.~\ref{section-counter-rotating-terms},
we consider thermal effects of the counter-rotating terms.
In Sec.~\ref{section-discussion},
we give a brief discussion.
In Appendix~\ref{section-3rd-order-correction-term},
we give details of calculations of the third order perturbation correction.
The evaluation of the perturbation corrections described in
Secs.~\ref{section-0th-order-correction-term},
\ref{section-1st-order-correction-term},
\ref{section-2nd-order-correction-term}
and
Appendix~\ref{section-3rd-order-correction-term},
we make use of techniques for calculations
developed in Ref.~\cite{Azuma2010}.
Thus, we can regard this paper as a sequel of Ref.~\cite{Azuma2010}.
\section{\label{section-review-JCM}A brief review of the JCM}
In this section, we review the JCM briefly.
The Hamiltonian of the original
JCM is expressed in the form,
\begin{equation}
H
=
\frac{\hbar}{2}\omega_{0}\sigma_{z}
+
\hbar\omega a^{\dagger}a
+
\hbar\kappa(\sigma_{+}a+\sigma_{-}a^{\dagger}),
\label{JCM-Hamiltonian-0}
\end{equation}
where
$\hbar=h/2\pi$,
$\sigma_{\pm}=(1/2)(\sigma_{x}\pm i\sigma_{y})$
and
$[a,a^{\dagger}]=1$.
The Pauli matrices
($\sigma_{i}$ for $i=x,y,z$)
are operators acting on quantum states of the single atom in the cavity.
The creation and annihilation operators
$a^{\dagger}$
and
$a$
act on quantum states of the single cavity mode.
The transition frequency of the two-level atom is given by
$\omega_{0}$.
The frequency of the single cavity mode is given by
$\omega$.
From now on,
for simplicity,
we assume
$\kappa$
to be real.
We can divide the Hamiltonian $H$ given by Eq.~(\ref{JCM-Hamiltonian-0})
into two parts as follows:
\begin{eqnarray}
H&=&\hbar(C_{1}+C_{2}), \nonumber \\
C_{1}&=&\omega(\frac{1}{2}\sigma_{z}+a^{\dagger}a), \nonumber \\
C_{2}&=&\kappa(\sigma_{+}a+\sigma_{-}a^{\dagger})-\frac{\Delta\omega}{2}\sigma_{z},
\label{JCM-Hamiltonian-decomposition-0}
\end{eqnarray}
where
$\Delta\omega=\omega-\omega_{0}$.
We can confirm $[C_{1},C_{2}]=0$.
Moreover, we can diagonalize $C_{1}$ at ease.
Because of the above facts,
we take the following interaction picture for expressing the time-evolution of the quantum state:
First, we write down the state vector of the atom and the cavity photons in the Schr{\"o}dinger picture as
$|\psi_{\mbox{\scriptsize S}}(t)\rangle$.
Second,
assuming $|\psi_{\mbox{\scriptsize I}}(0)\rangle=|\psi_{\mbox{\scriptsize S}}(0)\rangle$,
we define the state vector in the interaction picture as
$|\psi_{\mbox{\scriptsize I}}(t)\rangle
=\exp(iC_{1}t)|\psi_{\mbox{\scriptsize S}}(t)\rangle$.
Thus,
we can describe the time-evolution
as
$|\psi_{\mbox{\scriptsize I}}(t)\rangle=U(t)|\psi_{\mbox{\scriptsize I}}(0)\rangle$,
where
$U(t)=\exp(-iC_{2}t)$.
We give eigenstates of the two-level atom and the cavity mode in the following forms:
First, we describe the ground state and the excited state of the atom as two-component vectors,
\begin{equation}
|g\rangle_{\mbox{\scriptsize A}}
=
|0\rangle_{\mbox{\scriptsize A}}
=
\left(
\begin{array}{c}
0 \\
1
\end{array}
\right),
\quad\quad
|e\rangle_{\mbox{\scriptsize A}}
=
|1\rangle_{\mbox{\scriptsize A}}
=
\left(
\begin{array}{c}
1 \\
0
\end{array}
\right),
\label{atom-basis-vectors}
\end{equation}
where
$|g\rangle_{\mbox{\scriptsize A}}(=|0\rangle_{\mbox{\scriptsize A}})$
and
$|e\rangle_{\mbox{\scriptsize A}}(=|1\rangle_{\mbox{\scriptsize A}})$
are eigenvectors of
$\sigma_{z}$,
and their corresponding eigenvalues are
$(-1)$
and
$1$,
respectively.
The index $\mbox{A}$ stands for the atom.
We can regard
$|i\rangle_{\mbox{\scriptsize A}}$ for $i\in\{0,1\}$
as the number states of the fermions.
Second,
we
write down the number states of the cavity photons
as
$|n\rangle_{\mbox{\scriptsize P}}
=(1/\sqrt{n!})(a^{\dagger})^{n}|0\rangle_{\mbox{\scriptsize P}}$ for $n=0,1,2,...$.
The index $\mbox{P}$ stands for the photons.
We describe the unitary operator for the time-evolution $U(t)$
as the $2\times 2$ matrix,
\begin{equation}
U(t)
=
\exp[-it
\left(
\begin{array}{cc}
-\Delta\omega/2 & \kappa a \\
\kappa a^{\dagger} & \Delta\omega/2
\end{array}
\right)
]
=
\left(
\begin{array}{cc}
u_{11} & u_{10} \\
u_{01} & u_{00}
\end{array}
\right),
\label{unitary-evolution-1}
\end{equation}
where
\begin{eqnarray}
u_{11}
&=&
\cos(\sqrt{a^{\dagger}a+c+1}|\kappa|t)
+
i
\frac{\Delta \omega}{|\Delta\omega|}
\sqrt{c}
\frac{\sin(\sqrt{a^{\dagger}a+c+1}|\kappa|t)}{\sqrt{a^{\dagger}a+c+1}}, \nonumber \\
u_{10}
&=&
-i
\frac{\kappa}{|\kappa|}
\frac{\sin(\sqrt{a^{\dagger}a+c+1}|\kappa|t)}{\sqrt{a^{\dagger}a+c+1}}
a, \nonumber \\
u_{01}
&=&
-i
\frac{\kappa}{|\kappa|}
\frac{\sin(\sqrt{a^{\dagger}a+c}|\kappa|t)}{\sqrt{a^{\dagger}a+c}}
a^{\dagger}, \nonumber \\
u_{00}
&=&
\cos(\sqrt{a^{\dagger}a+c}|\kappa|t)
-
i
\frac{\Delta \omega}{|\Delta\omega|}
\sqrt{c}
\frac{\sin(\sqrt{a^{\dagger}a+c}|\kappa|t)}{\sqrt{a^{\dagger}a+c}},
\label{unitary-evolution-2}
\end{eqnarray}
and
\begin{equation}
c
=
(\frac{\Delta\omega}{2\kappa})^{2}.
\label{unitary-evolution-3}
\end{equation}
Because we take the basis vectors $\{|1\rangle_{\mbox{\scriptsize A}},|0\rangle_{\mbox{\scriptsize A}}\}$,
the indices $i,j\in\{1,0\}$ for $u_{ij}$
are arranged in descending order.
That is to say,
we take the index `$1$' for representing the first row and the first column of the $2\times2$ matrix $U(t)$,
and
we take the index `$0$' for representing the second row and the second column of the $2\times2$ matrix $U(t)$.
After these preparations,
the probability for detecting the ground state of the atom $|g\rangle_{\mbox{\scriptsize A}}$
at the time $t$
is given by
\begin{equation}
P_{g}(t)
=
\|_{\mbox{\scriptsize A}}\langle g|\psi_{\mbox{\scriptsize I}}(t)\rangle\|^{2}.
\end{equation}
Moreover,
the atomic population inversion is given by
\begin{equation}
\langle\sigma_{z}(t)\rangle
=
\langle\psi_{\mbox{\scriptsize I}}(t)|\sigma_{z}|\psi_{\mbox{\scriptsize I}}(t)\rangle
=
1-2P_{g}(t).
\label{definition-atomic-population-inversion}
\end{equation}
As a particular case, we consider
the initial state
to be in $|\psi_{\mbox{\scriptsize I}}(0)\rangle
=|g\rangle_{\mbox{\scriptsize A}}|\alpha\rangle_{\mbox{\scriptsize P}}$,
where $|\alpha\rangle_{\mbox{\scriptsize P}}$ represents the coherent state,
\begin{eqnarray}
|\alpha\rangle_{\mbox{\scriptsize P}}
&=&
\exp(\alpha a^{\dagger}-\alpha^{*}a)|0\rangle_{\mbox{\scriptsize P}} \nonumber \\
&=&
e^{-|\alpha|^{2}/2}
\sum_{n=0}^{\infty}
\frac{\alpha^{n}}{\sqrt{n!}}|n\rangle_{\mbox{\scriptsize P}},
\label{zero-temperature-coherent-state}
\end{eqnarray}
and $\alpha$ is an arbitrary complex number.
From now on, for simplicity,
we always let the parameter $\alpha$
characterizing the coherent state $|\alpha\rangle_{\mbox{\scriptsize P}}$
be real.
Then,
we obtain $P_{g}(t)$ in the form,
\begin{eqnarray}
P_{g}(t)
&=&
\left\|
\begin{array}{c}
{
(
\begin{array}{cc}
0 & 1
\end{array}
)
} \\
\quad
\end{array}
\left(
\begin{array}{cc}
u_{11} & u_{10} \\
u_{01} & u_{00}
\end{array}
\right)
\left(
\begin{array}{c}
0 \\
|\alpha\rangle_{\mbox{\scriptsize P}}
\end{array}
\right)
\right\|^{2} \nonumber \\
&=&
{}_{\mbox{\scriptsize P}}
\langle\alpha|u_{00}^{\dagger}u_{00}|\alpha\rangle_{\mbox{\scriptsize P}} \nonumber \\
&=&
e^{-\alpha^{2}}
\sum_{n=0}^{\infty}
\frac{\alpha^{2n}}{n!}
[\cos^{2}(\sqrt{n+c}|\kappa|t)
+
c
\frac{\sin^{2}(\sqrt{n+c}|\kappa|t)}{n+c}].
\label{Probability-collapse-and-revival-Rabi-oscillations-0}
\end{eqnarray}
The atomic population inversion given by Eqs.~(\ref{definition-atomic-population-inversion})
and (\ref{Probability-collapse-and-revival-Rabi-oscillations-0})
shows
the collapse and the revival of the Rabi oscillations.
Here, we examine the time scale of the initial collapse
and the period of the revival of the Rabi oscillations
\cite{Barnett1997}.
Learning from experience,
we know this phenomenon becomes more distinct as $\alpha^{2}$ increases.
Thus,
we assume $\alpha^{2}\gg 1$.
We rewrite the index of the summation $n$ as $n=\alpha^{2}+\delta n$
in Eq.~(\ref{Probability-collapse-and-revival-Rabi-oscillations-0}).
Then,
because of the Poisson distribution,
the major contribution for the summation
in Eq.~(\ref{Probability-collapse-and-revival-Rabi-oscillations-0})
comes from $|\delta n|<\alpha^{2}$.
Moreover,
for simplicity,
we assume $c\ll \alpha^{2}$,
and we neglect the term
$[c/(n+c)]\sin^{2}(\sqrt{n+c}|\kappa|t)$
in the right-hand side of Eq.~(\ref{Probability-collapse-and-revival-Rabi-oscillations-0}).
From the above discussions and Eqs.~(\ref{definition-atomic-population-inversion})
and (\ref{Probability-collapse-and-revival-Rabi-oscillations-0}),
writing
$\sqrt{n+c}
\simeq
(\alpha^{2}+n)/(2|\alpha|)$,
we obtain the following approximation:
\begin{eqnarray}
\langle\sigma_{z}(t)\rangle
&\simeq&
-
e^{-\alpha^{2}}
\sum_{n=0}^{\infty}
\frac{\alpha^{2n}}{n!}
\frac{1}{2}
[
e^{i(\alpha^{2}+n)|\kappa|t/|\alpha|}
+
e^{-i(\alpha^{2}+n)|\kappa|t/|\alpha|}
] \nonumber \\
&=&
-
\exp[\alpha^{2}(\cos\frac{|\kappa|t}{|\alpha|}-1)]
\cos(|\alpha||\kappa|t+\alpha^{2}\sin\frac{|\kappa|t}{|\alpha|}).
\label{atomic-population-inversion-approximation}
\end{eqnarray}
In Eq.~(\ref{atomic-population-inversion-approximation}),
$\exp[\alpha^{2}(\cos(|\kappa|t/|\alpha|)-1)]$
represents
the amplitude envelope of the wave,
and
$\cos[|\alpha||\kappa|t+\alpha^{2}\sin(|\kappa|t/|\alpha|)]$
represents
the Rabi oscillations.
Therefore,
we can estimate
the time scale of the initial collapse
and the period of the revival of the Rabi oscillations at
$|\kappa|^{-1}$
and
$2\pi|\alpha|/|\kappa|$ around,
respectively.
Moreover,
paying attention to
$\alpha^{2}\sin(|\kappa|t/|\alpha|)\sim|\alpha||\kappa|t$ for $\alpha^{2}\gg 1$,
we can estimate the period of the Rabi oscillations at about
$\pi/(|\alpha||\kappa|)$.
\section{\label{section-review-TFD}A brief review of the TFD}
In this section,
we give a brief review of the TFD developed by Takahashi and Umezawa
\cite{Takahashi1975,Ojima1981,Umezawa1982,Umezawa1992}.
The TFD is a method for describing the quantum mechanics at finite temperature.
Using this formalism,
we can describe
the statistical average of an observable at finite temperature
as a pure state expectation value.
Thus,
if we take the TFD formalism,
we do not need to deal with a mixed state,
which is a statistical ensemble of pure states at finite temperature.
In return for the above advantage,
the TFD lets us introduce
the so-called tilde particles corresponding to the ordinary particles.
Then,
we understand that the ordinary particles and the tilde particles
represent the dynamical degree of freedom and the thermal degree of freedom,
respectively.
Thus, to construct the TFD formalism,
we introduce a fictitious Hilbert space $\tilde{\cal H}$
corresponding to an original Hilbert space ${\cal H}$
and handle quantum mechanics on ${\cal H}\otimes\tilde{\cal H}$.
In this section,
according to the TFD formalism,
we define the thermal vacua of bosons and fermions.
Moreover, we discuss the thermal coherent state proposed
by Barnett, Knight, Mann and Revzen
\cite{Barnett1985,Mann1989a}.
After these preparations,
we rewrite the Hamiltonian of the JCM according to the TFD.
First, we consider the TFD formalism for describing the system of the bosons.
We define
the ordinary Hilbert space
\begin{equation}
{\cal H}_{\mbox{\scriptsize B}}:
\{|0\rangle_{\mbox{\scriptsize B}},|1\rangle_{\mbox{\scriptsize B}},
|2\rangle_{\mbox{\scriptsize B}},...\},
\end{equation}
and
its corresponding tilde space,
\begin{equation}
\tilde{\cal H}_{\mbox{\scriptsize B}}:
\{|\tilde{0}\rangle_{\mbox{\scriptsize B}},|\tilde{1}\rangle_{\mbox{\scriptsize B}},
|\tilde{2}\rangle_{\mbox{\scriptsize B}},...\}.
\end{equation}
Then,
the TFD formalism for the bosons is defined on the following space:
\begin{equation}
\{|n\rangle_{\mbox{\scriptsize B}}\otimes |\tilde{m}\rangle_{\mbox{\scriptsize B}}
\in
{\cal H}_{\mbox{\scriptsize B}}\otimes\tilde{\cal H}_{\mbox{\scriptsize B}}:
n,m\in\{0,1,2,..\}\}.
\end{equation}
We write the creation and annihilation operators on the Hilbert space ${\cal H}_{\mbox{\scriptsize B}}$
as $a^{\dagger}$ and $a$, respectively.
Moreover,
we write the creation and annihilation operators on the Hilbert space $\tilde{\cal H}_{\mbox{\scriptsize B}}$
as $\tilde{a}^{\dagger}$ and $\tilde{a}$, respectively.
Then,
we assume the commutation relations,
\begin{equation}
[a,a^{\dagger}]=[\tilde{a},\tilde{a}^{\dagger}]=1,
\quad
[a,\tilde{a}]=[a,\tilde{a}^{\dagger}]=0.
\label{ordinary-tilde-commutation-relations}
\end{equation}
Next,
we introduce the temperature $\beta=1/(k_{\mbox{\scriptsize B}}T)$
as follows:
\begin{equation}
\hat{U}_{\mbox{\scriptsize B}}(\theta)=\exp[i\theta(\beta)\hat{G}_{\mbox{\scriptsize B}}],
\label{definition-hat-U-operator}
\end{equation}
\begin{equation}
\hat{G}_{\mbox{\scriptsize B}}=i(a\tilde{a}-\tilde{a}^{\dagger}a^{\dagger}),
\label{definition-hat-G-operator}
\end{equation}
\begin{eqnarray}
\cosh\theta(\beta)&=&[1-\exp(-\beta\epsilon)]^{-1/2}, \nonumber \\
\sinh\theta(\beta)&=&[\exp(\beta\epsilon)-1]^{-1/2},
\label{definition-theta-beta-1}
\end{eqnarray}
where $\epsilon=\hbar\omega$.
We note that the relations
$\hat{G}_{\mbox{\scriptsize B}}^{\dagger}=\hat{G}_{\mbox{\scriptsize B}}$,
$\hat{U}_{\mbox{\scriptsize B}}^{\dagger}(\theta)=\hat{U}_{\mbox{\scriptsize B}}^{-1}(\theta)$
and
$\hat{U}_{\mbox{\scriptsize B}}(-\theta)=\hat{U}_{\mbox{\scriptsize B}}^{\dagger}(\theta)$ hold.
In Eqs.~(\ref{definition-hat-U-operator}) and (\ref{definition-hat-G-operator}),
to emphasize that $\hat{G}_{\mbox{\scriptsize B}}$ and $\hat{U}_{\mbox{\scriptsize B}}(\theta)$
are operators acting on both
${\cal H}_{\mbox{\scriptsize B}}$
and
$\tilde{\cal H}_{\mbox{\scriptsize B}}$,
we put an accent (a hat) on them.
Moreover,
we pay attention to the fact that
$\theta(\beta)(\geq 0)$ is real
and
$\theta(\beta)\to +0$ as $\beta\to +\infty$
in Eq.~(\ref{definition-theta-beta-1}).
The index $\mbox{B}$ appearing
in $\hat{G}_{\mbox{\scriptsize B}}$ and $\hat{U}_{\mbox{\scriptsize B}}(\theta)$
stands for the boson.
Because of introducing the temperature,
the creation and annihilation operators
defined on ${\cal H}_{\mbox{\scriptsize B}}$ and $\tilde{\cal H}_{\mbox{\scriptsize B}}$
are transformed as follows:
\begin{eqnarray}
a
&\rightarrow&
a(\theta)
=\hat{U}_{\mbox{\scriptsize B}}(\theta)a\hat{U}_{\mbox{\scriptsize B}}^{\dagger}(\theta)
=
\cosh \theta(\beta)a
-
\sinh \theta(\beta)\tilde{a}^{\dagger}, \nonumber \\
\tilde{a}
&\rightarrow&
\tilde{a}(\theta)
=\hat{U}_{\mbox{\scriptsize B}}(\theta)\tilde{a}\hat{U}_{\mbox{\scriptsize B}}^{\dagger}(\theta)
=
\cosh \theta(\beta)\tilde{a}
-
\sinh \theta(\beta)a^{\dagger}.
\label{temperature-transformations-a-tilde-a}
\end{eqnarray}
The transformation given by Eq.~(\ref{temperature-transformations-a-tilde-a})
is called the Bogoliubov transformation for two mode squeezed states.
From Eqs.~(\ref{ordinary-tilde-commutation-relations})
and
(\ref{temperature-transformations-a-tilde-a}),
we can derive the commutation relations,
\begin{equation}
[a(\theta),a^{\dagger}(\theta)]=[\tilde{a}(\theta),\tilde{a}^{\dagger}(\theta)]=1,
\quad
[a(\theta),\tilde{a}(\theta)]=[a(\theta),\tilde{a}^{\dagger}(\theta)]=0.
\label{ordinary-tilde-commutation-relations-finite-temperature}
\end{equation}
Thus, we can regard $a(\theta)$ and $\tilde{a}(\theta)$
as quasi-particles at the temperature $\beta$.
Here, we define the zero-temperature vacuum as
\begin{eqnarray}
|0,\tilde{0}\rangle_{\mbox{\scriptsize B}}
=
|0\rangle_{\mbox{\scriptsize B}}
\otimes
|\tilde{0}\rangle_{\mbox{\scriptsize B}}
\in{\cal H}_{\mbox{\scriptsize B}}\otimes\tilde{\cal H}_{\mbox{\scriptsize B}}.
\end{eqnarray}
Furthermore,
referring to
Sec.~2.2.3 of Ref.~\cite{Umezawa1992},
we define the thermal vacuum in the form,
\begin{eqnarray}
|0(\theta)\rangle_{\mbox{\scriptsize B}}
&=&
\hat{U}_{\mbox{\scriptsize B}}(\theta)
|0,\tilde{0}\rangle_{\mbox{\scriptsize B}} \nonumber \\
&=&
\exp(-\ln \cosh\theta)\exp[(\tanh\theta)a^{\dagger}\tilde{a}^{\dagger}]
|0,\tilde{0}\rangle_{\mbox{\scriptsize B}}.
\label{definition-thermal-vacuum}
\end{eqnarray}
From Eqs.~(\ref{temperature-transformations-a-tilde-a}) and (\ref{definition-thermal-vacuum}),
we obtain
\begin{equation}
a(\theta)|0(\theta)\rangle_{\mbox{\scriptsize B}}
=
\tilde{a}(\theta)|0(\theta)\rangle_{\mbox{\scriptsize B}}
=
0.
\end{equation}
Thus,
we can consider the thermal vacuum
$|0(\theta)\rangle_{\mbox{\scriptsize B}}$
to be a vacuum for the quasi-particles, which are represented by $a(\theta)$ and $\tilde{a}(\theta)$.
Second, we consider the TFD formalism for describing the system of the fermions.
We define
the ordinary Hilbert space,
\begin{equation}
{\cal H}_{\mbox{\scriptsize F}}:
\{|0\rangle_{\mbox{\scriptsize F}},|1\rangle_{\mbox{\scriptsize F}}\},
\end{equation}
and its corresponding
tilde space,
\begin{equation}
\tilde{\cal H}_{\mbox{\scriptsize F}}:
\{|\tilde{0}\rangle_{\mbox{\scriptsize F}},|\tilde{1}\rangle_{\mbox{\scriptsize F}}\}.
\end{equation}
Then,
the TFD formalism for the fermions is defined on the following space:
\begin{equation}
\{|n\rangle_{\mbox{\scriptsize F}}\otimes |\tilde{m}\rangle_{\mbox{\scriptsize F}}
\in
{\cal H}_{\mbox{\scriptsize F}}\otimes\tilde{\cal H}_{\mbox{\scriptsize F}}:
n,m\in\{0,1\}\}.
\end{equation}
We write the creation and annihilation operators on the Hilbert space
${\cal H}_{\mbox{\scriptsize F}}$
as $c^{\dagger}$ and $c$, respectively.
Moreover, we write the creation and annihilation operators on the Hilbert space
$\tilde{\cal H}_{\mbox{\scriptsize F}}$ as
$\tilde{c}^{\dagger}$ and $\tilde{c}$,
respectively.
Then,
we assume the anti-commutation relations,
\begin{equation}
\{c,c^{\dagger}\}=\{\tilde{c},\tilde{c}^{\dagger}\}=1,
\quad
\{c,\tilde{c}\}=\{c,\tilde{c}^{\dagger}\}=0.
\label{ordinary-tilde-commutation-relations-fermion}
\end{equation}
Next,
we introduce the temperature $\beta$ as follows:
\begin{equation}
\hat{U}_{\mbox{\scriptsize F}}(\theta)=\exp[i\theta(\beta)\hat{G}_{\mbox{\scriptsize F}}],
\label{definition-hat-U-operator-fermion}
\end{equation}
\begin{equation}
\hat{G}_{\mbox{\scriptsize F}}=i(c\tilde{c}-\tilde{c}^{\dagger}c^{\dagger}),
\label{definition-hat-G-operator-fermion}
\end{equation}
\begin{eqnarray}
\cos\theta(\beta)&=&[1+\exp(-\beta\epsilon)]^{-1/2}, \nonumber \\
\sin\theta(\beta)&=&\exp(-\beta\epsilon/2)[1+\exp(-\beta\epsilon)]^{-1/2}.
\label{definition-theta-beta-2}
\end{eqnarray}
We note that the relations
$\hat{G}_{\mbox{\scriptsize F}}^{\dagger}=\hat{G}_{\mbox{\scriptsize F}}$,
$\hat{U}_{\mbox{\scriptsize F}}^{\dagger}(\theta)=\hat{U}_{\mbox{\scriptsize F}}^{-1}(\theta)$
and
$\hat{U}_{\mbox{\scriptsize F}}(-\theta)=\hat{U}_{\mbox{\scriptsize F}}^{\dagger}(\theta)$ hold.
Moreover, we pay attention to the fact that
$\theta(\beta)(\geq 0)$ is real and $\theta(\beta)\to +0$ as $\beta\to +\infty$
in Eq.~(\ref{definition-theta-beta-2}).
$\hat{U}_{\mbox{\scriptsize F}}(\theta)$ in Eq.~(\ref{definition-hat-U-operator-fermion})
and
$\hat{G}_{\mbox{\scriptsize F}}$ in Eq.~(\ref{definition-hat-G-operator-fermion})
are again of the same forms as
$\hat{U}_{\mbox{\scriptsize B}}(\theta)$ in Eq.~(\ref{definition-hat-U-operator})
and
$\hat{G}_{\mbox{\scriptsize B}}$ in Eq.~(\ref{definition-hat-G-operator}),
respectively.
The index $\mbox{F}$ appearing
in $\hat{U}_{\mbox{\scriptsize F}}(\theta)$ and $\hat{G}_{\mbox{\scriptsize F}}$
stands for the fermion.
Because of introducing the temperature,
the creation and annihilation operators defined on
${\cal H}_{\mbox{\scriptsize F}}$
and
$\tilde{\cal H}_{\mbox{\scriptsize F}}$
are transformed as follows:
\begin{eqnarray}
c
&\rightarrow&
c(\theta)
=\hat{U}_{\mbox{\scriptsize F}}(\theta)c\hat{U}_{\mbox{\scriptsize F}}^{\dagger}(\theta)
=
\cos \theta(\beta)c
+
\sin \theta(\beta)\tilde{c}^{\dagger}, \nonumber \\
\tilde{c}
&\rightarrow&
\tilde{c}(\theta)
=\hat{U}_{\mbox{\scriptsize F}}(\theta)\tilde{c}\hat{U}_{\mbox{\scriptsize F}}^{\dagger}(\theta)
=
\cos \theta(\beta)\tilde{c}
-
\sin \theta(\beta)c^{\dagger}.
\label{temperature-transformations-c-tilde-c}
\end{eqnarray}
From Eqs.~(\ref{ordinary-tilde-commutation-relations-fermion})
and (\ref{temperature-transformations-c-tilde-c}),
we can derive the anti-commutation relations,
\begin{equation}
\{c(\theta),c^{\dagger}(\theta)\}=\{\tilde{c}(\theta),\tilde{c}^{\dagger}(\theta)\}=1,
\quad
\{c(\theta),\tilde{c}(\theta)\}=\{c(\theta),\tilde{c}^{\dagger}(\theta)\}=0.
\label{ordinary-tilde-commutation-relations-finite-temperature-fermion}
\end{equation}
Thus,
we can regard $c(\theta)$ and $\tilde{c}(\theta)$
as quasi-particles at the temperature $\beta$.
Here,
we
define the zero-temperature vacuum as
\begin{eqnarray}
|0,\tilde{0}\rangle_{\mbox{\scriptsize F}}
=
|0\rangle_{\mbox{\scriptsize F}}
\otimes
|\tilde{0}\rangle_{\mbox{\scriptsize F}}
\in{\cal H}_{\mbox{\scriptsize F}}\otimes\tilde{\cal H}_{\mbox{\scriptsize F}}.
\end{eqnarray}
Furthermore,
referring to Sec.~2.4.2 of Ref.~\cite{Umezawa1992},
we define the thermal vacuum in the form,
\begin{eqnarray}
|0(\theta)\rangle_{\mbox{\scriptsize F}}
&=&
\hat{U}_{\mbox{\scriptsize F}}(\theta)
|0,\tilde{0}\rangle_{\mbox{\scriptsize F}} \nonumber \\
&=&
[\cos\theta+(\sin\theta)c^{\dagger}\tilde{c}^{\dagger}]
|0,\tilde{0}\rangle_{\mbox{\scriptsize F}}.
\label{definition-thermal-vacuum-fermion}
\end{eqnarray}
From Eqs.~(\ref{temperature-transformations-c-tilde-c})
and
(\ref{definition-thermal-vacuum-fermion}),
we obtain
\begin{equation}
c(\theta)|0(\theta)\rangle_{\mbox{\scriptsize F}}
=
\tilde{c}(\theta)|0(\theta)\rangle_{\mbox{\scriptsize F}}
=
0.
\end{equation}
Thus,
we can consider the thermal vacuum
$|0(\theta)\rangle_{\mbox{\scriptsize F}}$
to be a vacuum for the quasi-particles,
which are represented by
$c(\theta)$ and $\tilde{c}(\theta)$.
Third, we consider the thermal coherent state.
After the above preparations,
Barnett, Knight, Mann and Revzen
define the thermal coherent state as follows
\cite{Barnett1985,Mann1989a}:
\begin{equation}
|\alpha,\tilde{\gamma};\theta\rangle_{\mbox{\scriptsize B}}
=
\exp[\alpha a^{\dagger}(\theta)
+\tilde{\gamma}^{*}\tilde{a}^{\dagger}(\theta)
-\alpha^{*}a(\theta)
-\tilde{\gamma}\tilde{a}(\theta)]
|0(\theta)\rangle_{\mbox{\scriptsize B}},
\label{definition-thermal-coherent-state-1}
\end{equation}
where $\alpha(\theta)$ and $\tilde{\alpha}(\theta)$
are given by Eq.~(\ref{temperature-transformations-a-tilde-a}).
However,
the TFD formalism requires
all state vectors to be invariant under the tilde conjugation,
which is given by
\begin{eqnarray}
(XY)\tilde{\;\;}&=&\tilde{X}\tilde{Y}, \nonumber \\
(\xi_{1}X+\xi_{2}Y)\tilde{\;\;}&=&\xi_{1}^{*}\tilde{X}+\xi_{2}^{*}\tilde{Y}, \nonumber \\
(X^{\dagger})\tilde{\;\;}&=&\tilde{X}^{\dagger}, \nonumber \\
(\tilde{X})\tilde{\;\;}&=&\sigma X,
\end{eqnarray}
where
\begin{equation}
\sigma
=
\left\{
\begin{array}{ll}
1 & \mbox{(boson)}\\
-1 & \mbox{(fermion)}
\end{array}
\right.,
\end{equation}
$X$ and $Y$
are arbitrary operators defined on
${\cal H}_{\mbox{\scriptsize B}}$
and
${\cal H}_{\mbox{\scriptsize F}}$,
and
$\xi_{1}$ and $\xi_{2}$ are arbitrary complex numbers.
For example,
the thermal vacua are obviously invariant under the tilde conjugation,
that is to say,
$|0(\theta)\rangle_{\mbox{\scriptsize B}}\tilde{\;\;}
=|0(\theta)\rangle_{\mbox{\scriptsize B}}$,
${}_{\mbox{\scriptsize B}}\langle 0(\theta)|\tilde{\;\;}
={}_{\mbox{\scriptsize B}}\langle 0(\theta)|$,
$|0(\theta)\rangle_{\mbox{\scriptsize F}}\tilde{\;\;}
=|0(\theta)\rangle_{\mbox{\scriptsize F}}$
and
${}_{\mbox{\scriptsize F}}\langle 0(\theta)|\tilde{\;\;}
={}_{\mbox{\scriptsize F}}\langle 0(\theta)|$.
Moreover,
we can show
$(\hat{G}_{\mbox{\scriptsize B}})\tilde{\;\;}=-\hat{G}_{\mbox{\scriptsize B}}$,
$(\hat{U}_{\mbox{\scriptsize B}}(\theta))\tilde{\;\;}=\hat{U}_{\mbox{\scriptsize B}}(\theta)$,
$(\hat{G}_{\mbox{\scriptsize F}})\tilde{\;\;}=-\hat{G}_{\mbox{\scriptsize F}}$
and
$(\hat{U}_{\mbox{\scriptsize F}}(\theta))\tilde{\;\;}=\hat{U}_{\mbox{\scriptsize F}}(\theta)$,
at ease.
In the TFD formalism,
all state vectors realized actually in the physical system
have to be invariant under the tilde conjugation.
Thus, not only the time-evolution but also all possible transitions of state vectors
have to be invariant
under the tilde conjugation.
Requiring $|\alpha,\tilde{\gamma};\theta\rangle_{\mbox{\scriptsize B}}$
given by Eq.~(\ref{definition-thermal-coherent-state-1})
to be invariant under the tilde conjugation,
we obtain
\begin{eqnarray}
|\alpha;\theta\rangle_{\mbox{\scriptsize B}}
&=&
|\alpha,\alpha^{*};\theta\rangle_{\mbox{\scriptsize B}} \nonumber \\
&=&
\exp[\alpha a^{\dagger}(\theta)
+\alpha\tilde{a}^{\dagger}(\theta)
-\alpha^{*}a(\theta)
-\alpha^{*}\tilde{a}(\theta)]
|0(\theta)\rangle_{\mbox{\scriptsize B}}.
\label{definition-thermal-coherent-state-2}
\end{eqnarray}
From now on,
we call this state the thermal coherent state
\cite{Mann1989b,Kireev1989,Mann1989c}.
Moreover, for simplicity,
we assume $\alpha$ characterizing $|\alpha;\theta\rangle_{\mbox{\scriptsize B}}$
to be always real.
In the following paragraphs,
we examine the physical meanings of the thermal vacua,
$|0(\theta)\rangle_{\mbox{\scriptsize B}}$
and
$|0(\theta)\rangle_{\mbox{\scriptsize F}}$.
First,
we clarify the physical meanings of the thermal vacuum for the bosons
$|0(\theta)\rangle_{\mbox{\scriptsize B}}$.
We begin by considering the density operator defined on
${\cal H}_{\mbox{\scriptsize B}}$,
\begin{equation}
\rho_{\mbox{\scriptsize B}}(\theta)
=
\mbox{Tr}_{\tilde{\cal H}}|0(\theta)\rangle_{\mbox{\scriptsize B}}
{}_{\mbox{\scriptsize B}}\langle 0(\theta)|.
\label{density-operator-boson-thermal-vacuum-0}
\end{equation}
Then,
we derive an explicit representation of
$\rho_{\mbox{\scriptsize B}}(\theta)$
as follows:
At first,
from Eq.~(\ref{definition-thermal-vacuum}) and
(\ref{density-operator-boson-thermal-vacuum-0}),
we obtain
\begin{eqnarray}
\rho_{\mbox{\scriptsize B}}(\theta)
&=&
\frac{1}{\cosh^{2}\theta}
\sum_{n=0}^{\infty}
\frac{1}{n!}
{}_{\mbox{\scriptsize B}}\langle\tilde{0}|\tilde{a}^{n}
\exp[(\tanh\theta)a^{\dagger}\tilde{a}^{\dagger}]
|0,\tilde{0}\rangle_{\mbox{\scriptsize B}} \nonumber \\
&&\times
{}_{\mbox{\scriptsize B}}\langle 0,\tilde{0}|
\exp[(\tanh\theta)\tilde{a}a](\tilde{a}^{\dagger})^{n}
|\tilde{0}\rangle_{\mbox{\scriptsize B}}.
\label{density-operator-thermal-vacuum-boson-1}
\end{eqnarray}
Next, we apply the following relation
to Eq.~(\ref{density-operator-thermal-vacuum-boson-1}):
\begin{equation}
\exp[-(\tanh\theta)a^{\dagger}\tilde{a}^{\dagger}]
\tilde{a}^{n}
\exp[(\tanh\theta)a^{\dagger}\tilde{a}^{\dagger}]
=
[\tilde{a}+(\tanh\theta)a^{\dagger}]^{n}.
\label{formula-a-tilde-a-condensation-1}
\end{equation}
Then using Eq.~(\ref{definition-theta-beta-1}),
we obtain
\begin{eqnarray}
\rho_{\mbox{\scriptsize B}}(\theta)
&=&
\frac{1}{\cosh^{2}\theta}
\sum_{n=0}^{\infty}
\tanh^{2n}\theta
|n\rangle_{\mbox{\scriptsize B}}{}_{\mbox{\scriptsize B}}\langle n| \nonumber \\
&=&
(1-e^{-\beta\epsilon})\sum_{n=0}^{\infty}e^{-n\beta\epsilon}
|n\rangle_{\mbox{\scriptsize B}}{}_{\mbox{\scriptsize B}}\langle n|.
\label{explicit-form-density-operator-thermal-vacuum-boson}
\end{eqnarray}
Looking at Eq.~(\ref{explicit-form-density-operator-thermal-vacuum-boson}),
we notice that
$\rho_{\mbox{\scriptsize B}}(\theta)$
is an ensemble of quantum states
$\{|n\rangle_{\mbox{\scriptsize B}}:n=0,1,2,...\}$,
into each of which $n$ bosons ($a$-particles) are put.
Moreover,
the statistical probability of $|n\rangle_{\mbox{\scriptsize B}}$
is given by $(1-e^{-\beta\epsilon})e^{-n\beta\epsilon}
=\mbox{Const.}\times e^{-\beta(n\epsilon)}$,
so that
$\rho_{\mbox{\scriptsize B}}(\theta)$
represents a canonical ensemble of the Bose-Einstein distribution in thermal equilibrium.
From these considerations,
we understand that
the $a$-particle represents the dynamical degree of freedom
and the $\tilde{a}$-particle represents the thermal degree of freedom
in
$|0(\theta)\rangle_{\mbox{\scriptsize B}}
\in
{\cal H}_{\mbox{\scriptsize B}}
\otimes
\tilde{\cal H}_{\mbox{\scriptsize B}}$.
Second,
we clarify the physical meanings of the thermal vacuum for fermions
$|0(\theta)\rangle_{\mbox{\scriptsize F}}$.
We begin by considering
the density operator defined on ${\cal H}_{\mbox{\scriptsize F}}$,
\begin{equation}
\rho_{\mbox{\scriptsize F}}(\theta)
=
\mbox{Tr}_{\tilde{\cal H}}|0(\theta)\rangle_{\mbox{\scriptsize F}}
{}_{\mbox{\scriptsize F}}\langle 0(\theta)|.
\label{density-operator-fermion-thermal-vacuum-0}
\end{equation}
From Eqs.~(\ref{definition-theta-beta-2}),
(\ref{definition-thermal-vacuum-fermion})
and
(\ref{density-operator-fermion-thermal-vacuum-0}),
we derive an explicit representation of
$\rho_{\mbox{\scriptsize F}}(\theta)$
in the form,
\begin{eqnarray}
\rho_{\mbox{\scriptsize F}}(\theta)
&=&
\sum_{n\in\{0,1\}}
{}_{\mbox{\scriptsize F}}\langle\tilde{n}|
[\cos\theta+(\sin\theta)c^{\dagger}\tilde{c}^{\dagger}]
|0,\tilde{0}\rangle_{\mbox{\scriptsize F}}
{}_{\mbox{\scriptsize F}}\langle 0,\tilde{0}|
[\cos\theta+(\sin\theta)\tilde{c}c]
|\tilde{n}\rangle_{\mbox{\scriptsize F}} \nonumber \\
&=&
\cos^{2}\theta|0\rangle_{\mbox{\scriptsize F}}{}_{\mbox{\scriptsize F}}\langle 0|
+
\sin^{2}\theta|1\rangle_{\mbox{\scriptsize F}}{}_{\mbox{\scriptsize F}}\langle 1| \nonumber \\
&=&
\frac{1}{1+e^{-\beta\epsilon}}
|0\rangle_{\mbox{\scriptsize F}}{}_{\mbox{\scriptsize F}}\langle 0|
+
\frac{e^{-\beta\epsilon}}{1+e^{-\beta\epsilon}}
|1\rangle_{\mbox{\scriptsize F}}{}_{\mbox{\scriptsize F}}\langle 1|.
\label{density-operator-thermal-vacuum-fermion-1}
\end{eqnarray}
Looking at Eq.~(\ref{density-operator-thermal-vacuum-fermion-1}),
we notice that
$\rho(\theta)_{\mbox{\scriptsize F}}$
is an ensemble of quantum states
$\{|n\rangle_{\mbox{\scriptsize F}}:n\in\{0,1\}\}$,
into each of which $n$ fermions ($c$-particles) are put.
Moreover,
the statistical probability of $|n\rangle_{\mbox{\scriptsize F}}$
is given by
$(1+e^{-\beta\epsilon})^{-1}e^{-n\beta\epsilon}
=\mbox{Const.}\times e^{-\beta(n\epsilon)}$,
so that $\rho(\theta)_{\mbox{\scriptsize F}}$
represents a canonical ensemble of the Fermi-Dirac distribution in thermal equilibrium.
From these considerations,
we understand that
the $c$-particle represents the dynamical degree of freedom
and
the $\tilde{c}$-particle represents the thermal degree of freedom
in
$|0(\theta)\rangle_{\mbox{\scriptsize F}}
\in
{\cal H}_{\mbox{\scriptsize F}}
\otimes
\tilde{\cal H}_{\mbox{\scriptsize F}}$.
For convenience of calculations
that appear in the remains of this paper,
using Eqs.~(\ref{ordinary-tilde-commutation-relations})
and (\ref{temperature-transformations-a-tilde-a}),
we rewrite $|\alpha;\theta\rangle_{\mbox{\scriptsize B}}$
given by Eq.~(\ref{definition-thermal-coherent-state-2})
as
\begin{eqnarray}
|\alpha;\theta\rangle_{\mbox{\scriptsize B}}
&=&
\exp[\hat{U}_{\mbox{\scriptsize B}}(\theta)
(\alpha a^{\dagger}
+\alpha\tilde{a}^{\dagger}
-\alpha a
-\alpha\tilde{a})
\hat{U}_{\mbox{\scriptsize B}}^{\dagger}(\theta)]
\hat{U}_{\mbox{\scriptsize B}}(\theta)
|0,\tilde{0}\rangle_{\mbox{\scriptsize B}} \nonumber \\
&=&
\hat{U}_{\mbox{\scriptsize B}}(\theta)
\exp[\alpha(a^{\dagger}-a)]
\exp[\alpha(\tilde{a}^{\dagger}-\tilde{a})]
|0,\tilde{0}\rangle_{\mbox{\scriptsize B}} \nonumber \\
&=&
\hat{U}_{\mbox{\scriptsize B}}(\theta)
|\alpha\rangle_{\mbox{\scriptsize B}}
|\tilde{\alpha}\rangle_{\mbox{\scriptsize B}},
\label{another-form-of-thermal-coherent-state-1}
\end{eqnarray}
where
$|\alpha\rangle_{\mbox{\scriptsize B}}$
and
$|\tilde{\alpha}\rangle_{\mbox{\scriptsize B}}$
are the coherent states at zero temperature
defined on
${\cal H}_{\mbox{\scriptsize B}}$
and
$\tilde{\cal H}_{\mbox{\scriptsize B}}$,
respectively.
Next, after the above preparations,
we discuss how to construct the finite-temperature JCM according to the TFD formalism
from the original JCM Hamiltonian.
For example,
we consider a system,
which consists of bosons $a$ and fermions $c$.
We assume that
the time-evolution of the system is reversible and it never causes dissipation.
This implies that
the Hamiltonian $H$,
which is the Hermitian operator corresponding to the total energy of the system
(the $a$-particles and the $c$-particles),
is equivalent to the generator of the unitary operator for the time-evolution.
Obeying the TFD formalism,
we introduce the $\tilde{a}$-particle corresponding to the $a$-particle
and
the $\tilde{c}$-particle corresponding to the $c$-particle
into the system.
Because $\tilde{a}$ and $\tilde{c}$ are fictitious particles
representing the thermal degree of freedom,
the Schr{\"o}dinger equation governing the time-evolution of the particles $a$ and $c$
never suffers form the thermal effects.
This observation suggests
that the Schr{\"o}dinger equation for $a$ and $c$
never changes in spite of introducing $\tilde{a}$ and $\tilde{c}$.
From these considerations,
we can conclude that
the particles ($a$ and $c$) are never coupled to the tilde particles ($\tilde{a}$ and $\tilde{c}$)
direct in the Hamiltonian.
Hence,
the total Hamiltonian based on the TFD formalism
has to be a sum
of the interaction terms of the particles ($a$ and $c$)
and the interaction terms of the tilde particles ($\tilde{a}$ and $\tilde{c}$).
Here, we describe the total Hamiltonian for the TFD formalism as $\hat{H}$.
Then,
letting the Schr{\"o}dinger equation for the whole system consisting of the particles
of $a$, $\tilde{a}$, $c$ and $\tilde{c}$
be invariant under the tilde conjugation,
we can express $\hat{H}$ in the form,
\begin{equation}
\hat{H}=H-\tilde{H},
\label{hat-Hamiltonian-0}
\end{equation}
where $\tilde{H}=(H)\tilde{\;\;}$
\cite{Takahashi1975,Ojima1981,Umezawa1982,Umezawa1992}.
If we give the Hamiltonian of the system $\hat{H}$ in the form of Eq.~(\ref{hat-Hamiltonian-0}),
the time-evolution of the whole system consisting of the particles
and the tilde particles
is reversible and it never causes dissipation.
In Eq.~(\ref{hat-Hamiltonian-0}),
we emphasize the following:
The Hamiltonian $H$ is constructed from $a$ and $c$.
Then,
because of $\tilde{H}=(H)\tilde{\;\;}$,
the Hamiltonian $\tilde{H}$ has to include both $\tilde{a}$ and $\tilde{c}$.
This implies that we have to introduce the temperature into
both the $a$-particles and the $c$-particles.
If the Hamiltonian of the total system is given by Eq.~(\ref{hat-Hamiltonian-0}),
the thermal vacua $|0(\theta)\rangle_{\mbox{\scriptsize B}}$ given by Eq.~(\ref{definition-thermal-vacuum})
and $|0(\theta)\rangle_{\mbox{\scriptsize F}}$ given by Eq.~(\ref{definition-thermal-vacuum-fermion})
are not dependent on time,
so that they are stationary states.
We can explain this fact as follows:
Turning our eyes towards Eq.~(\ref{definition-thermal-vacuum}),
we notice that
the condensation of $(a\tilde{a})$-pairs into $|0(\theta)\rangle_{\mbox{\scriptsize B}}$ occurs.
Thus, in the vacuum $|0(\theta)\rangle_{\mbox{\scriptsize B}}$,
the $a$-particle receives a phase factor $\exp[-i(H/\hbar)t]$
and the $\tilde{a}$-particle receives a phase factor $\exp[i(\tilde{H}/\hbar)t]$.
Then,
these phase factors cancel out their effects with each other,
and the ($a\tilde{a}$)-pair acts like a zero-energy boson.
Hence, we understand that the thermal vacuum
$|0(\theta)\rangle_{\mbox{\scriptsize B}}$
is not dependent on time.
We can apply a similar discussion to the thermal vacuum of the fermions,
because we can observe the condensation of $(c\tilde{c})$-pairs
into $|0(\theta)\rangle_{\mbox{\scriptsize F}}$ in Eq.~(\ref{definition-thermal-vacuum-fermion}).
To obtain the JCM Hamiltonian
$\hat{H}$ which takes the form given by Eq.~(\ref{hat-Hamiltonian-0}),
we rewrite the original Hamiltonian $H$ of the JCM
given by Eq.~(\ref{JCM-Hamiltonian-0})
as
\begin{equation}
H
=
\frac{\hbar}{2}\omega_{0}(2c^{\dagger}c-1)
+
\hbar\omega a^{\dagger}a
+
\hbar\kappa(c^{\dagger}a+ca^{\dagger}).
\label{JCM-Hamiltonian-1}
\end{equation}
The Pauli matrices $\sigma_{z}$, $\sigma_{+}$ and $\sigma_{-}$
in the original Hamiltonian $H$ given by Eq.~(\ref{JCM-Hamiltonian-0})
are replaced with the creation and annihilation operators of the fermions
$(2c^{\dagger}c-1)$, $c^{\dagger}$ and $c$
in the rewritten Hamiltonian $H$ given by Eq.~(\ref{JCM-Hamiltonian-1}).
Extending the Hamiltonian $H$ given by Eq.~(\ref{JCM-Hamiltonian-1})
according to Eq.~(\ref{hat-Hamiltonian-0}),
we obtain
\begin{equation}
\hat{H}
=
(H_{0}+H_{\mbox{\scriptsize I}})
-
(\tilde{H}_{0}+\tilde{H}_{\mbox{\scriptsize I}}),
\label{extended-JCM-Hamiltonian-0}
\end{equation}
where
\begin{eqnarray}
H_{0}
&=&
\frac{\hbar}{2}\omega_{0}
(2c^{\dagger}c-1)
+
\hbar\omega a^{\dagger}a, \nonumber \\
H_{\mbox{\scriptsize I}}
&=&
\hbar\kappa(c^{\dagger}a+ca^{\dagger}), \nonumber \\
\tilde{H}_{0}
&=&
\frac{\hbar}{2}\omega_{0}
(2\tilde{c}^{\dagger}\tilde{c}-1)
+
\hbar\omega \tilde{a}^{\dagger}\tilde{a}, \nonumber \\
\tilde{H}_{\mbox{\scriptsize I}}
&=&
\hbar\kappa(\tilde{c}^{\dagger}\tilde{a}+\tilde{c}\tilde{a}^{\dagger}).
\label{extended-JCM-Hamiltonian-1}
\end{eqnarray}
In this paper,
we concentrate on examining the Hamiltonian $\hat{H}$
given by Eqs.~(\ref{extended-JCM-Hamiltonian-0}) and (\ref{extended-JCM-Hamiltonian-1}).
Here, we think about
some Hamiltonians,
which are proposed in the other works.
Especially, we examine whether or not we can regard them
as genuine JCM Hamiltonians based on the TFD formalism.
Barnett, Knight and Azuma consider the following Hamiltonian
in Refs.~\cite{Barnett1985,Azuma2010}:
\begin{equation}
\hat{H}_{\mbox{\scriptsize BKA}}
=
\frac{\hbar}{2}\omega_{0}\sigma_{z}
+
\hbar\omega(a^{\dagger}a-\tilde{a}^{\dagger}\tilde{a})
+
\hbar\kappa(\sigma_{+}a+\sigma_{-}a^{\dagger}).
\label{TFD-JCM-Hamiltonian-Azuma}
\end{equation}
However,
the Hamiltonian $\hat{H}_{\mbox{\scriptsize BKA}}$ given by Eq.~(\ref{TFD-JCM-Hamiltonian-Azuma})
does not include the tilde operators
corresponding to the atomic operators $\sigma_{z}$ and $\sigma_{\pm}$,
so that the states of the atom are always at zero temperature.
Because
the Hamiltonian $\hat{H}_{\mbox{\scriptsize BKA}}$ given by Eq.~(\ref{TFD-JCM-Hamiltonian-Azuma})
does not introduce the temperature into the atom,
we cannot regard it as a genuine Hamiltonian based on the TFD formalism.
Fan and Lu propose the following Hamiltonian in Ref.~\cite{Fan2004}:
\begin{eqnarray}
\hat{H}_{\mbox{\scriptsize FL}}
&=&
\frac{\hbar}{2}\omega_{0}\sigma_{z}
+
\hbar\omega(a^{\dagger}a-\tilde{a}^{\dagger}\tilde{a}) \nonumber \\
&&
+
\hbar\kappa
[
\sigma_{+}
\sqrt{
\frac{a-\tilde{a}^{\dagger}}{a^{\dagger}-\tilde{a}}
}
\sqrt{a^{\dagger}a-\tilde{a}^{\dagger}\tilde{a}}
+
\sqrt{a^{\dagger}a-\tilde{a}^{\dagger}\tilde{a}}
\sqrt{
\frac{a^{\dagger}-\tilde{a}}{a-\tilde{a}^{\dagger}}
}
\sigma_{-}
].
\label{Fan-Lu-Hamiltonian}
\end{eqnarray}
The reason why Fan and Lu construct the Hamiltonian $\hat{H}_{\mbox{\scriptsize FL}}$
given by Eq.~(\ref{Fan-Lu-Hamiltonian})
is as follows:
They find the commutation relations,
\begin{eqnarray}
{[}
a^{\dagger}a-\tilde{a}^{\dagger}\tilde{a},
\sqrt{\frac{a^{\dagger}-\tilde{a}}{a-\tilde{a}^{\dagger}}}
{]}
&=&
\sqrt{\frac{a^{\dagger}-\tilde{a}}{a-\tilde{a}^{\dagger}}}, \nonumber \\
{[}
a^{\dagger}a-\tilde{a}^{\dagger}\tilde{a},
\sqrt{\frac{a-\tilde{a}^{\dagger}}{a^{\dagger}-\tilde{a}}}
{]}
&=&
-
\sqrt{\frac{a-\tilde{a}^{\dagger}}{a^{\dagger}-\tilde{a}}}.
\end{eqnarray}
Thus,
if we regard
$(a^{\dagger}a-\tilde{a}^{\dagger}\tilde{a})$
as an extended number operator,
we can think
\\
$\sqrt{(a^{\dagger}-\tilde{a})/(a-\tilde{a}^{\dagger})}$
and
$\sqrt{(a-\tilde{a}^{\dagger})/(a^{\dagger}-\tilde{a})}$
to be extended creation and annihilation operators, respectively.
From these suggestions,
Fan and Lu propose the Hamiltonian $\hat{H}_{\mbox{\scriptsize FL}}$
given in Eq.~(\ref{Fan-Lu-Hamiltonian}).
However,
Fan and Lu's Hamiltonian $\hat{H}_{\mbox{\scriptsize FL}}$
does not include the tilde operators
corresponding to the atomic operators $\sigma_{z}$ and $\sigma_{\pm}$.
Thus, it can not be regarded as a genuine Hamiltonian based on the TFD formalism.
Moreover,
in the Hamiltonian $\hat{H}_{\mbox{\scriptsize FL}}$,
the operators $a$ and $a^{\dagger}$
are coupled direct to
the tilde operators $\tilde{a}$ and $\tilde{a}^{\dagger}$.
Thus, Fan and Lu's system suffers from dissipation
during the time-evolution.
To examine their system,
we have to deal with non-equilibrium states.
Because it is beyond the purpose of this paper,
we do not involve ourselves in it.
From now on,
we examine the Hamiltonian $\hat{H}$
given by Eqs.~(\ref{extended-JCM-Hamiltonian-0}) and (\ref{extended-JCM-Hamiltonian-1}).
We can divide the Hamiltonian $\hat{H}$ into two parts as follows:
\begin{eqnarray}
\hat{H}
&=&
\hbar(\hat{C}_{1}+\hat{C}_{2}), \nonumber \\
\hat{C}_{1}
&=&
\omega
[(c^{\dagger}c-\tilde{c}^{\dagger}\tilde{c})
+
(a^{\dagger}a-\tilde{a}^{\dagger}\tilde{a})], \nonumber \\
\hat{C}_{2}
&=&
\kappa(c^{\dagger}a+ca^{\dagger})
-
\kappa(\tilde{c}^{\dagger}\tilde{a}+\tilde{c}\tilde{a}^{\dagger})
-
\Delta\omega(c^{\dagger}c-\tilde{c}^{\dagger}\tilde{c}).
\label{decomposition-hat-Hamiltonian}
\end{eqnarray}
Then, we obtain a commutation relation
$[\hat{C}_{1},\hat{C}_{2}]=0$.
Moreover, we can diagonalize $\hat{C}_{1}$ at ease.
Thus,
we describe the time-evolution of the total system with the interaction picture
as follows:
First, we write
the state vector of the total system in the Schr{\"o}dinger picture
as $|\Psi_{\mbox{\scriptsize S}}(t)\rangle$.
Second,
assuming $|\Psi_{\mbox{\scriptsize I}}(0)\rangle
=|\Psi_{\mbox{\scriptsize S}}(0)\rangle$,
we define
the state vector of the total system in the interaction picture
as
$|\Psi_{\mbox{\scriptsize I}}(t)\rangle
=\exp(i\hat{C}_{1}t)|\Psi_{\mbox{\scriptsize S}}(t)\rangle$.
Hence,
we can describe the time-evolution
as
$|\Psi_{\mbox{\scriptsize I}}(t)\rangle
=\hat{U}(t)|\Psi_{\mbox{\scriptsize I}}(0)\rangle$,
where
$\hat{U}(t)=\exp(-i\hat{C}_{2}t)$.
The unitary operator for the time-evolution of $|\Psi_{\mbox{\scriptsize I}}(t)\rangle$
is given by
\begin{eqnarray}
\hat{U}(t)
&=&
\exp[-i\hat{C}_{2}t] \nonumber \\
&=&
U(t)\otimes \tilde{U}(t) \nonumber \\
&=&
\exp[-it
\left(
\begin{array}{cc}
-\Delta\omega/2 & \kappa a \\
\kappa a^{\dagger} & \Delta\omega/2
\end{array}
\right)
]
\otimes
\exp[it
\left(
\begin{array}{cc}
-\Delta\omega/2 & \kappa \tilde{a} \\
\kappa \tilde{a}^{\dagger} & \Delta\omega/2
\end{array}
\right)
].
\label{hat-time-evolution-unitary-operator-0}
\end{eqnarray}
In the right-hand side of Eq.~(\ref{hat-time-evolution-unitary-operator-0}),
the first $2\times 2$ matrix of the tensor product acts on ${\cal H}_{\mbox{\scriptsize F}}$
and
the second $2\times 2$ matrix of the tensor product acts on $\tilde{\cal H}_{\mbox{\scriptsize F}}$.
The $2\times 2$ matrix $U(t)$ appearing in Eq.~(\ref{hat-time-evolution-unitary-operator-0})
and
the unitary operator for the time-evolution defined in Eq.~(\ref{unitary-evolution-1})
are in the same form.
Thus,
the elements of the $2\times 2$ matrices $U(t)$ and $\tilde{U}(t)$
appearing in Eq.~(\ref{hat-time-evolution-unitary-operator-0})
are given by Eq.~(\ref{unitary-evolution-2}).
\section{\label{section-thermal-effects-period-qualitative-estimation}
Thermal effects of the period of the revival of the Rabi oscillations}
In this paper,
putting the single cavity mode and the atom
in the thermal coherent state $|\alpha;\theta\rangle_{\mbox{\scriptsize B}}$
and
the thermal vacuum $|0(\theta)\rangle_{\mbox{\scriptsize F}}$,
respectively,
at the time $t=0$,
we aim at examining the time-evolution of the JCM.
We assume that
the system consisting of
the single cavity mode and the atom
evolve in time without dissipation,
and it maintains the constant temperature $\beta$ all the time.
As shown in Sec.~\ref{section-review-JCM},
we can estimate that
the period of the revival of the Rabi oscillations at zero temperature
is around $2\pi|\alpha|/|\kappa|$.
In this section,
we discuss how the period changes at finite low temperature.
We evaluate the thermal effects of the period in an intuitive manner.
The parameter $|\alpha|$,
which characterizes the ordinary zero-temperature coherent state $|\alpha\rangle$,
is given by
\begin{equation}
|\alpha|
=
(\langle\alpha|a^{\dagger}a|\alpha\rangle)^{1/2}.
\end{equation}
Thus,
we can guess that the parameter $|\alpha|$ varies
with the thermal effects of the finite low temperature
as
\begin{equation}
|\alpha|
\rightarrow
(\langle\alpha;\theta|a^{\dagger}a|\alpha;\theta\rangle)^{1/2}.
\end{equation}
On the other hand,
using Eqs.~(\ref{ordinary-tilde-commutation-relations}),
(\ref{definition-hat-U-operator}),
(\ref{definition-hat-G-operator}),
(\ref{temperature-transformations-a-tilde-a})
and
(\ref{another-form-of-thermal-coherent-state-1}),
and paying attention to
$a|\alpha\rangle=\alpha|\alpha\rangle$
and
$\tilde{a}|\tilde{\alpha}\rangle=\alpha|\tilde{\alpha}\rangle$,
we obtain
\begin{eqnarray}
&&
\langle\alpha;\theta|a^{\dagger}a|\alpha;\theta\rangle \nonumber \\
&=&
\langle\alpha|\langle\tilde{\alpha}|
\hat{U}_{\mbox{\scriptsize B}}(-\theta)
a^{\dagger}a
\hat{U}_{\mbox{\scriptsize B}}^{\dagger}(-\theta)
|\alpha\rangle|\tilde{\alpha}\rangle \nonumber \\
&=&
\langle\alpha|\langle\tilde{\alpha}|
a^{\dagger}(-\theta)a(-\theta)
|\alpha\rangle|\tilde{\alpha}\rangle \nonumber \\
&=&
\alpha^{2}e^{2\theta}
+
(1/4)(e^{2\theta}+e^{-2\theta}-2).
\end{eqnarray}
From the above observations, we can expect the following:
assuming $\alpha^{2}\gg 1$ and $\theta\ll 1$,
the thermal effects let
the parameter characterizing the coherent state
change
under the low-temperature limit
as
\begin{equation}
|\alpha|
\rightarrow
(\langle\alpha;\theta|a^{\dagger}a|\alpha;\theta\rangle)^{1/2}
\simeq
|\alpha|e^{\theta}.
\label{thermal-effect-of-alpha-coherent-state-0}
\end{equation}
Thus,
we can expect that
the period of the revival of the Rabi oscillations
at finite low temperature varies as
\begin{equation}
2\pi|\alpha|/|\kappa|
\rightarrow
2\pi|\alpha|e^{\theta}/|\kappa|.
\label{thermal-effect-period-rivival-Rabi-oscillations}
\end{equation}
This phenomenon is confirmed by numerical calculations in Sec.~\ref{section-numerical-calculations}.
The intuitive discussions given in this section is effective,
when we can specify the whole system with the constant temperature $\beta$.
If the system is in a non-equilibrium state and the temperature $\beta$ varies during its time-evolution,
we cannot apply the above intuitive discussions to the system,
so that Eq.~(\ref{thermal-effect-period-rivival-Rabi-oscillations}) does not hold.
\section{\label{section-formulation-perturbetion-theory}The formulation of the perturbation theory}
In this section,
we initially
put the states of the atom and the cavity field
in the thermal vacuum of the fermions $|0(\theta)\rangle_{\mbox{\scriptsize F}}$
and
the thermal coherent state $|\alpha;\beta\rangle_{\mbox{\scriptsize B}}$,
respectively.
Then,
we formulate the time-evolution of the JCM based on the TFD
discussed in Sec.~\ref{section-review-TFD}
as the perturbation theory under the low-temperature limit.
After formulating the perturbation theory here,
we estimate the zero-th, first, second and third order corrections
in Secs.~\ref{section-0th-order-correction-term},
\ref{section-1st-order-correction-term},
\ref{section-2nd-order-correction-term}
and
Appendix~\ref{section-3rd-order-correction-term}.
To evaluate these correction terms,
we make use of techniques for calculations developed in Ref.~\cite{Azuma2010}.
At first, we express the state of the system for $t=0$ in the form,
\begin{equation}
|\Psi_{\mbox{\scriptsize I}}(0)\rangle
=
|0(\Theta)\rangle_{\mbox{\scriptsize F}}
|\alpha;\theta\rangle_{\mbox{\scriptsize B}}.
\label{extended-JCM-initial-state}
\end{equation}
During the time-evolution of the system,
we assume the atom and the cavity mode do not suffer dissipation
and maintain the constant temperature $\beta$.
Thus,
from Eqs.~(\ref{definition-theta-beta-1})
and
(\ref{definition-theta-beta-2}),
the parameters of the temperature for the fermionic atom $\Theta(\beta)$
and
the bosonic cavity mode $\theta(\beta)$
are given in the following forms, respectively:
\begin{eqnarray}
\cos\Theta(\beta)
&=&
[1+\exp(-\beta\hbar\omega_{0})]^{-1/2}, \nonumber \\
\sin\Theta(\beta)
&=&
\exp(-\beta\hbar\omega_{0}/2)
[1+\exp(-\beta\hbar\omega_{0})]^{-1/2}, \nonumber \\
\cosh\theta(\beta)
&=&
[1-\exp(-\beta\hbar\omega)]^{-1/2}, \nonumber \\
\sinh\theta(\beta)
&=&
[\exp(\beta\hbar\omega)-1]^{-1/2},
\label{definition-temperature-JCM-0}
\end{eqnarray}
where
$\omega_{0}$ represents the transition frequency of the two-level atom
and
$\omega$ represents the frequency of the single cavity mode,
as defined in Eq.~(\ref{JCM-Hamiltonian-0}).
From Eqs.~(\ref{another-form-of-thermal-coherent-state-1})
and
(\ref{hat-time-evolution-unitary-operator-0}),
we obtain $|\Psi_{\mbox{\scriptsize I}}(t)\rangle$
as
\begin{eqnarray}
|\Psi_{\mbox{\scriptsize I}}(t)\rangle
&=&
\hat{U}(t)|\Psi_{\mbox{\scriptsize I}}(0)\rangle \nonumber \\
&=&
[U(t)\otimes\tilde{U}(t)]
|0(\Theta)\rangle_{\mbox{\scriptsize F}}
|\alpha;\theta\rangle_{\mbox{\scriptsize B}} \nonumber \\
&=&
[U(t)\otimes\tilde{U}(t)]
[
|0(\Theta)\rangle_{\mbox{\scriptsize F}}
\hat{U}_{\mbox{\scriptsize B}}(\theta)
|\alpha\rangle_{\mbox{\scriptsize B}}|\tilde{\alpha}\rangle_{\mbox{\scriptsize B}}
].
\label{state-vector-Psi_I_t-0}
\end{eqnarray}
From Eq.~(\ref{definition-thermal-vacuum-fermion}),
we can rewrite $|0(\Theta)\rangle_{\mbox{\scriptsize F}}$
in the form,
\begin{equation}
|0(\Theta)\rangle_{\mbox{\scriptsize F}}
=
\cos\Theta|0,\tilde{0}\rangle_{\mbox{\scriptsize F}}
+
\sin\Theta|1,\tilde{1}\rangle_{\mbox{\scriptsize F}}.
\label{thermal-vacuum-fermion-2}
\end{equation}
Taking $\{|i,\tilde{j}\rangle_{\mbox{\scriptsize F}}:i,j\in\{1,0\}\}$
for the basis vectors of the four-dimensional Hilbert space
${\cal H}_{\mbox{\scriptsize F}}\otimes\tilde{\cal H}_{\mbox{\scriptsize F}}$,
we can write down $|0(\Theta)\rangle_{\mbox{\scriptsize F}}$
as a four-component vector,
\begin{equation}
|0(\Theta)\rangle_{\mbox{\scriptsize F}}
=
\left(
\begin{array}{c}
\sin\Theta \\
0 \\
0 \\
\cos\Theta
\end{array}
\right),
\label{fermion-thermal-vacuum-4-component-vector-0}
\end{equation}
where
the components of the above vector are arranged
in the order of
$|1,\tilde{1}\rangle_{\mbox{\scriptsize F}}$,
$|1,\tilde{0}\rangle_{\mbox{\scriptsize F}}$,
$|0,\tilde{1}\rangle_{\mbox{\scriptsize F}}$
and
$|0,\tilde{0}\rangle_{\mbox{\scriptsize F}}$.
Thus,
writing $|\Psi_{\mbox{\scriptsize I}}(t)\rangle$
as the four-component vector,
we obtain
\begin{equation}
|\Psi_{\mbox{\scriptsize I}}(t)\rangle
=
[U(t)\otimes\tilde{U}(t)]
\left(
\begin{array}{c}
\sin\Theta
\hat{U}_{\mbox{\scriptsize B}}(\theta)
|\alpha\rangle_{\mbox{\scriptsize B}}|\tilde{\alpha}\rangle_{\mbox{\scriptsize B}} \\
0 \\
0 \\
\cos\Theta
\hat{U}_{\mbox{\scriptsize B}}(\theta)
|\alpha\rangle_{\mbox{\scriptsize B}}|\tilde{\alpha}\rangle_{\mbox{\scriptsize B}}
\end{array}
\right).
\end{equation}
Moreover,
expressing $U(t)\otimes\tilde{U}(t)$
in the form of the $4\times 4$ matrix,
\begin{equation}
U(t)\otimes\tilde{U}(t)
=
\left(
\begin{array}{cc}
u_{11}\tilde{U}(t) & u_{10}\tilde{U}(t) \\
u_{01}\tilde{U}(t) & u_{00}\tilde{U}(t)
\end{array}
\right),
\end{equation}
\begin{equation}
\tilde{U}(t)
=
\left(
\begin{array}{cc}
\tilde{u}_{11} & \tilde{u}_{10} \\
\tilde{u}_{01} & \tilde{u}_{00}
\end{array}
\right),
\label{definition-tilde-U-0}
\end{equation}
where
$\{u_{ij}:i,j\in\{1,0\}\}$
and
$\{\tilde{u}_{ij}:i,j\in\{1,0\}\}$
are given by Eq.~(\ref{unitary-evolution-2}),
we can write down the four-component vector
$|\Psi_{\mbox{\scriptsize I}}(t)\rangle$
as the following explicit form:
\begin{equation}
|\Psi_{\mbox{\scriptsize I}}(t)\rangle
=
\left(
\begin{array}{c}
\psi_{1\tilde{1}} \\
\psi_{1\tilde{0}} \\
\psi_{0\tilde{1}} \\
\psi_{0\tilde{0}}
\end{array}
\right),
\end{equation}
where
\begin{eqnarray}
\psi_{1\tilde{1}}
&=&
(\sin\Theta u_{11}\tilde{u}_{11}+\cos\Theta u_{10}\tilde{u}_{10})
\hat{U}_{\mbox{\scriptsize B}}(\theta)
|\alpha\rangle_{\mbox{\scriptsize B}}|\tilde{\alpha}\rangle_{\mbox{\scriptsize B}}, \nonumber \\
\psi_{1\tilde{0}}
&=&
(\sin\Theta u_{11}\tilde{u}_{01}+\cos\Theta u_{10}\tilde{u}_{00})
\hat{U}_{\mbox{\scriptsize B}}(\theta)
|\alpha\rangle_{\mbox{\scriptsize B}}|\tilde{\alpha}\rangle_{\mbox{\scriptsize B}}, \nonumber \\
\psi_{0\tilde{1}}
&=&
(\sin\Theta u_{01}\tilde{u}_{11}+\cos\Theta u_{00}\tilde{u}_{10})
\hat{U}_{\mbox{\scriptsize B}}(\theta)
|\alpha\rangle_{\mbox{\scriptsize B}}|\tilde{\alpha}\rangle_{\mbox{\scriptsize B}}, \nonumber \\
\psi_{0\tilde{0}}
&=&
(\sin\Theta u_{01}\tilde{u}_{01}+\cos\Theta u_{00}\tilde{u}_{00})
\hat{U}_{\mbox{\scriptsize B}}(\theta)
|\alpha\rangle_{\mbox{\scriptsize B}}|\tilde{\alpha}\rangle_{\mbox{\scriptsize B}}.
\end{eqnarray}
Hence,
the probability that we detect the ground state of the atom at zero temperature
in the state of the total system $|\Psi_{\mbox{\scriptsize I}}(t)\rangle$
is given by
\begin{eqnarray}
P_{g}(\Theta,\theta;t)
&=&
\|{}_{\mbox{\scriptsize F}}\langle 0,\tilde{0}|\Psi_{\mbox{\scriptsize I}}(t)\rangle\|^{2}
+
\|{}_{\mbox{\scriptsize F}}\langle 0,\tilde{1}|\Psi_{\mbox{\scriptsize I}}(t)\rangle\|^{2} \nonumber \\
&=&
\|\psi_{0\tilde{0}}\|^{2}+\|\psi_{0\tilde{1}}\|^{2} \nonumber \\
&=&
{}_{\mbox{\scriptsize B}}\langle\alpha|
{}_{\mbox{\scriptsize B}}\langle\tilde{\alpha}|
\hat{U}_{\mbox{\scriptsize B}}^{\dagger}(\theta)
[
\sin^{2}\Theta u_{01}^{\dagger}u_{01}\tilde{u}_{01}^{\dagger}\tilde{u}_{01} \nonumber \\
&&
+
\sin\Theta\cos\Theta
(u_{01}^{\dagger}u_{00}\tilde{u}_{01}^{\dagger}\tilde{u}_{00}
+
u_{00}^{\dagger}u_{01}\tilde{u}_{00}^{\dagger}\tilde{u}_{01}) \nonumber \\
&&
+
\cos^{2}\Theta u_{00}^{\dagger}u_{00}\tilde{u}_{00}^{\dagger}\tilde{u}_{00} \nonumber \\
&&
+
\sin^{2}\Theta u_{01}^{\dagger}u_{01}\tilde{u}_{11}^{\dagger}\tilde{u}_{11} \nonumber \\
&&
+
\sin\Theta\cos\Theta
(u_{01}^{\dagger}u_{00}\tilde{u}_{11}^{\dagger}\tilde{u}_{10}
+
u_{00}^{\dagger}u_{01}\tilde{u}_{10}^{\dagger}\tilde{u}_{11}) \nonumber \\
&&
+
\cos^{2}\Theta u_{00}^{\dagger}u_{00}\tilde{u}_{10}^{\dagger}\tilde{u}_{10}
]
\hat{U}_{\mbox{\scriptsize B}}(\theta)
|\alpha\rangle_{\mbox{\scriptsize B}}
|\tilde{\alpha}\rangle_{\mbox{\scriptsize B}}.
\label{Probability-atom-ground-state-theta-0}
\end{eqnarray}
Here,
we pay attention to the following fact:
Because $\tilde{U}(t)$ given by Eq.~(\ref{definition-tilde-U-0})
is a unitary matrix,
we obtain
\begin{eqnarray}
\tilde{u}_{11}^{\dagger}\tilde{u}_{11}+\tilde{u}_{01}^{\dagger}\tilde{u}_{01}
&=&
1, \nonumber \\
\tilde{u}_{10}^{\dagger}\tilde{u}_{10}+\tilde{u}_{00}^{\dagger}\tilde{u}_{00}
&=&
1, \nonumber \\
\tilde{u}_{11}^{\dagger}\tilde{u}_{10}+\tilde{u}_{01}^{\dagger}\tilde{u}_{00}
&=&
0, \nonumber \\
\tilde{u}_{10}^{\dagger}\tilde{u}_{11}+\tilde{u}_{00}^{\dagger}\tilde{u}_{01}
&=&
0.
\label{relations-unitary-tilde-U-components}
\end{eqnarray}
Substitution of Eq.~(\ref{relations-unitary-tilde-U-components})
into Eq.~(\ref{Probability-atom-ground-state-theta-0})
yields
\begin{eqnarray}
P_{g}(\Theta,\theta;t)
&=&
{}_{\mbox{\scriptsize B}}\langle\alpha|
{}_{\mbox{\scriptsize B}}\langle\tilde{\alpha}|
\hat{U}_{\mbox{\scriptsize B}}^{\dagger}(\theta)
[
\sin^{2}\Theta u_{01}^{\dagger}u_{01}
+
\cos^{2}\Theta u_{00}^{\dagger}u_{00}
]
\hat{U}_{\mbox{\scriptsize B}}(\theta)
|\alpha\rangle_{\mbox{\scriptsize B}}
|\tilde{\alpha}\rangle_{\mbox{\scriptsize B}} \nonumber \\
&=&
\cos^{2}\Theta
{}_{\mbox{\scriptsize B}}\langle\alpha|
{}_{\mbox{\scriptsize B}}\langle\tilde{\alpha}|
\hat{U}_{\mbox{\scriptsize B}}^{\dagger}(\theta)
g_{1}(a^{\dagger}a+c)
\hat{U}_{\mbox{\scriptsize B}}(\theta)
|\alpha\rangle_{\mbox{\scriptsize B}}
|\tilde{\alpha}\rangle_{\mbox{\scriptsize B}} \nonumber \\
&&
+
\sin^{2}\Theta
{}_{\mbox{\scriptsize B}}\langle\alpha|
{}_{\mbox{\scriptsize B}}\langle\tilde{\alpha}|
\hat{U}_{\mbox{\scriptsize B}}^{\dagger}(\theta)
g_{2}(a^{\dagger}a+c+1)
\hat{U}_{\mbox{\scriptsize B}}(\theta)
|\alpha\rangle_{\mbox{\scriptsize B}}
|\tilde{\alpha}\rangle_{\mbox{\scriptsize B}},
\label{Probability-atom-ground-state-theta-1}
\end{eqnarray}
where
\begin{eqnarray}
g_{1}(a^{\dagger}a+c)
&=&
u_{00}^{\dagger}u_{00} \nonumber \\
&=&
\cos^{2}(\sqrt{a^{\dagger}a+c}|\kappa|t)
+
c
\frac{\sin^{2}(\sqrt{a^{\dagger}a+c}|\kappa|t)}{a^{\dagger}a+c}, \nonumber \\
g_{2}(a^{\dagger}a+c+1)
&=&
u_{01}^{\dagger}u_{01} \nonumber \\
&=&
a
\frac{\sin^{2}(\sqrt{a^{\dagger}a+c}|\kappa|t)}{a^{\dagger}a+c}a^{\dagger} \nonumber \\
&=&
\frac{\sin^{2}(\sqrt{a^{\dagger}a+c+1}|\kappa|t)}{a^{\dagger}a+c+1}[(a^{\dagger}a+c+1)-c].
\label{definition-functions-g-1&2}
\end{eqnarray}
Using Eqs.~(\ref{definition-hat-U-operator})
and
(\ref{definition-hat-G-operator}),
we can rewrite $P_{g}(\Theta,\theta;t)$
given by Eqs.~(\ref{Probability-atom-ground-state-theta-1})
and
(\ref{definition-functions-g-1&2})
as
\begin{eqnarray}
P_{g}(\Theta,\theta;t)
&=&
\cos^{2}\Theta
\langle\alpha|
\langle\tilde{\alpha}|
\exp[\theta(a\tilde{a}-\tilde{a}^{\dagger}a^{\dagger})]
g_{1}(a^{\dagger}a+c)
\exp[-\theta(a\tilde{a}-\tilde{a}^{\dagger}a^{\dagger})]
|\alpha\rangle
|\tilde{\alpha}\rangle \nonumber \\
&&
+
\sin^{2}\Theta
\langle\alpha|
\langle\tilde{\alpha}|
\exp[\theta(a\tilde{a}-\tilde{a}^{\dagger}a^{\dagger})]
g_{2}(a^{\dagger}a+c+1) \nonumber \\
&&\times
\exp[-\theta(a\tilde{a}-\tilde{a}^{\dagger}a^{\dagger})]
|\alpha\rangle
|\tilde{\alpha}\rangle.
\label{Probability-atom-ground-state-theta-2}
\end{eqnarray}
Thus,
we obtain the perturbative expansion of $P_{g}(\Theta,\theta;t)$
in the small parameter $\theta(\beta)$
as
\begin{equation}
P_{g}(\Theta,\theta;t)
=
\sum_{n=0}^{\infty}
\frac{\theta(\beta)^{n}}{n!}
P^{(n)}_{g}(\Theta,\theta;t),
\label{Perturbation-theory-formula-0}
\end{equation}
where
\begin{equation}
P^{(n)}_{g}(\Theta,\theta;t)
=
\cos^{2}\Theta P^{(n)}_{g,1}(t)
+
\sin^{2}\Theta P^{(n)}_{g,2}(t),
\label{Perturbation-theory-formula-1}
\end{equation}
\begin{eqnarray}
P^{(0)}_{g,1}(t)
&=&
\langle\alpha|
\langle\tilde{\alpha}|
g_{1}(a^{\dagger}a+c)
|\alpha\rangle
|\tilde{\alpha}\rangle, \nonumber \\
P^{(0)}_{g,2}(t)
&=&
\langle\alpha|
\langle\tilde{\alpha}|
g_{2}(a^{\dagger}a+c+1)
|\alpha\rangle
|\tilde{\alpha}\rangle, \nonumber \\
P^{(1)}_{g,1}(t)
&=&
\langle\alpha|
\langle\tilde{\alpha}|
[a\tilde{a}-\tilde{a}^{\dagger}a^{\dagger},g_{1}(a^{\dagger}a+c)]
|\alpha\rangle
|\tilde{\alpha}\rangle, \nonumber \\
P^{(1)}_{g,2}(t)
&=&
\langle\alpha|
\langle\tilde{\alpha}|
[a\tilde{a}-\tilde{a}^{\dagger}a^{\dagger},g_{2}(a^{\dagger}a+c+1)]
|\alpha\rangle
|\tilde{\alpha}\rangle, \nonumber \\
P^{(2)}_{g,1}(t)
&=&
\langle\alpha|
\langle\tilde{\alpha}|
[a\tilde{a}-\tilde{a}^{\dagger}a^{\dagger},
[a\tilde{a}-\tilde{a}^{\dagger}a^{\dagger},g_{1}(a^{\dagger}a+c)]]
|\alpha\rangle
|\tilde{\alpha}\rangle, \nonumber \\
P^{(2)}_{g,2}(t)
&=&
\langle\alpha|
\langle\tilde{\alpha}|
[a\tilde{a}-\tilde{a}^{\dagger}a^{\dagger},
[a\tilde{a}-\tilde{a}^{\dagger}a^{\dagger},g_{2}(a^{\dagger}a+c+1)]]
|\alpha\rangle
|\tilde{\alpha}\rangle, \nonumber \\
&&...,
\label{Perturbation-theory-formula-2}
\end{eqnarray}
\begin{eqnarray}
P^{(n)}_{g,1}(t)
&=&
\langle\alpha|
\langle\tilde{\alpha}|
\underbrace{
[a\tilde{a}-\tilde{a}^{\dagger}a^{\dagger},
...,
[a\tilde{a}-\tilde{a}^{\dagger}a^{\dagger},
g_{1}(a^{\dagger}a+c)]...]
}_{\mbox{\scriptsize $n$-fold bracket}}
|\alpha\rangle
|\tilde{\alpha}\rangle, \nonumber \\
P^{(n)}_{g,2}(t)
&=&
\langle\alpha|
\langle\tilde{\alpha}|
\underbrace{
[a\tilde{a}-\tilde{a}^{\dagger}a^{\dagger},
...,
[a\tilde{a}-\tilde{a}^{\dagger}a^{\dagger},
g_{2}(a^{\dagger}a+c+1)]...]
}_{\mbox{\scriptsize $n$-fold bracket}}
|\alpha\rangle
|\tilde{\alpha}\rangle \nonumber \\
&&
\mbox{for $n=1,2,3,...$}.
\label{Perturbation-theory-formula-3}
\end{eqnarray}
Here, we pay attention to the following fact:
The perturbative expansion given by
Eqs.~(\ref{Perturbation-theory-formula-0}),
(\ref{Perturbation-theory-formula-1}),
(\ref{Perturbation-theory-formula-2})
and
(\ref{Perturbation-theory-formula-3})
is a power series in the small parameter $\theta(\beta)$.
On the other hand,
all the terms of the parameter $\Theta(\beta)$ included
in the perturbative expansion,
namely
$\cos^{2}\Theta$ and $\sin^{2}\Theta$,
are expressed as explicit rigorous forms.
Thus,
we can strictly compute the functions of $\Theta(\beta)$ at ease
in Eqs.~(\ref{Perturbation-theory-formula-0})
and (\ref{Perturbation-theory-formula-1}),
so that
we do not need to worry about perturbative corrections of the parameter $\Theta(\beta)$.
In this paper,
we consider the power series in the small parameter $\theta(\beta)$
to be the perturbative expansion under the low-temperature limit.
In contrast,
we do not regard $\Theta(\beta)$ as the parameter for the perturbation.
Furthermore,
the following trick lets
actual computations of correction terms,
that is to say,
$P^{(n)}_{g}(\Theta,\theta;t)$ for $n=1,2,3,...$,
be tractable.
We can write down the functions $g_{1}(x)$ and $g_{2}(x)$
defined in Eq.~(\ref{definition-functions-g-1&2})
as
\begin{eqnarray}
g_{1}(x)
&=&
\cos^{2}(\sqrt{x}|\kappa|t)
+
c
\frac{\sin^{2}(\sqrt{x}|\kappa|t)}{x}, \nonumber \\
g_{2}(x)
&=&
\frac{\sin^{2}(\sqrt{x}|\kappa|t)}{x}(x-c),
\label{definition-function-g-1&2}
\end{eqnarray}
so that we can rewrite each of them as the Taylor series at $x=0$,
\begin{equation}
g_{i}(x)
=
\sum_{m=0}^{\infty}g^{(m)}_{i}x^{m}
\quad
\mbox{for $-\infty<x<+\infty$},
\label{g-function-Taylor-series-0}
\end{equation}
where
\begin{equation}
g^{(m)}_{i}
=
\frac{1}{m!}\frac{d^{m}}{dx^{m}}g_{i}(x)
\bigg|_{x=0}
\quad
\mbox{for $i\in\{1,2\}$}.
\label{g-function-Taylor-series-1}
\end{equation}
Thus,
we can rewrite Eq.~(\ref{Perturbation-theory-formula-3}) as
\begin{eqnarray}
P^{(n)}_{g,1}(t)
&=&
\sum_{m=0}^{\infty}
g^{(m)}_{1}
\langle\alpha|
\langle\tilde{\alpha}|
\underbrace{
[a\tilde{a}-\tilde{a}^{\dagger}a^{\dagger},
...,
[a\tilde{a}-\tilde{a}^{\dagger}a^{\dagger},
(a^{\dagger}a+c)^{m}]...]
}_{\mbox{\scriptsize $n$-fold bracket}}
|\alpha\rangle
|\tilde{\alpha}\rangle, \nonumber \\
P^{(n)}_{g,2}(t)
&=&
\sum_{m=0}^{\infty}
g^{(m)}_{2}
\langle\alpha|
\langle\tilde{\alpha}|
\underbrace{
[a\tilde{a}-\tilde{a}^{\dagger}a^{\dagger},
...,
[a\tilde{a}-\tilde{a}^{\dagger}a^{\dagger},
(a^{\dagger}a+c+1)^{m}]...]
}_{\mbox{\scriptsize $n$-fold bracket}}
|\alpha\rangle
|\tilde{\alpha}\rangle \nonumber \\
&&\mbox{for $n=1,2,3,...$.}
\label{Perturbation-theory-formula-4}
\end{eqnarray}
In Secs.~\ref{section-0th-order-correction-term},
\ref{section-1st-order-correction-term},
\ref{section-2nd-order-correction-term}
and
Appendix~\ref{section-3rd-order-correction-term},
using the perturbative expansion given by
Eqs.~(\ref{Perturbation-theory-formula-0}),
(\ref{Perturbation-theory-formula-1}),
(\ref{Perturbation-theory-formula-2}),
(\ref{Perturbation-theory-formula-3}),
(\ref{definition-function-g-1&2}),
(\ref{g-function-Taylor-series-0}),
(\ref{g-function-Taylor-series-1})
and
(\ref{Perturbation-theory-formula-4}),
we compute
$P_{g}(\Theta,\theta;t)$.
\section{\label{section-comparison-of-TFD-and-Liouville-von-Neumann-equation}
Comparison of the TFD formalism and the
\\
Liouville-von Neumann equation}
In the previous sections,
we discuss a method for examining the time-evolution
caused by the Hamiltonian $\hat{H}$
defined in Eqs.~(\ref{extended-JCM-Hamiltonian-0}) and (\ref{extended-JCM-Hamiltonian-1})
with the initial state
$|\Psi(0)\rangle
=|0(\Theta)\rangle_{\mbox{\scriptsize F}}|\alpha;\theta\rangle_{\mbox{\scriptsize B}}$
given by Eq.~(\ref{extended-JCM-initial-state})
according to the TFD formalism.
This method is equivalent to
solving the following Liouville-von Neumann equation:
\begin{equation}
\frac{\partial}{\partial t}
\rho(t)
=
-\frac{i}{\hbar}
[H,\rho(t)],
\label{Liouville-von-Neumann-equation-0}
\end{equation}
where
\begin{eqnarray}
\rho(0)
&=&
\mbox{Tr}_{\tilde{\cal H}}
[
|\Psi(0)\rangle\langle\Psi(0)|
] \nonumber \\
&=&
\mbox{Tr}_{\tilde{\cal H}}
[
|0(\Theta)\rangle_{\mbox{\scriptsize F}}{}_{\mbox{\scriptsize F}}\langle 0(\Theta)|
\otimes
|\alpha;\theta\rangle_{\mbox{\scriptsize B}}{}_{\mbox{\scriptsize B}}\langle\alpha;\theta|
],
\label{initial-density-operator-0}
\end{eqnarray}
and the Hamiltonian $H$ appearing in Eq.~(\ref{Liouville-von-Neumann-equation-0})
is given by Eqs.~(\ref{JCM-Hamiltonian-0}) and (\ref{JCM-Hamiltonian-1}).
Both the Hamiltonian $\hat{H}$ based on the TFD formalism
defined by Eqs.~(\ref{extended-JCM-Hamiltonian-0}) and (\ref{extended-JCM-Hamiltonian-1})
and
the Liouville-von Neumann equation
given by Eq.~(\ref{Liouville-von-Neumann-equation-0})
represent that
the total system evolves in time
with maintaining the constant temperature,
so that it never suffers from dissipation
and
its time-evolution is reversible.
Thus, we understand
that
the Hermitian operator corresponding to the energy of the total system
is equivalent
to the generator of the unitary operator for the time-evolution.
Here,
thinking about the Liouville-von Neumann equation given by Eq.~(\ref{Liouville-von-Neumann-equation-0}),
we divide $H$ into the two parts $C_{1}$ and $C_{2}$
as shown in Eq.~(\ref{JCM-Hamiltonian-decomposition-0})
and take the interaction picture.
Assuming $\rho_{\mbox{\scriptsize I}}(0)=\rho(0)$,
we introduce the density operator described in the interaction picture
as
\begin{equation}
\rho_{\mbox{\scriptsize I}}(t)
=
e^{iC_{1}t}\rho(t)e^{-iC_{1}t}.
\end{equation}
Then,
using the commutation relation $[C_{1},C_{2}]=0$,
we obtain
\begin{equation}
\frac{\partial}{\partial t}
\rho_{\mbox{\scriptsize I}}(t)
=
-i
[C_{2},\rho_{\mbox{\scriptsize I}}(t)].
\label{Liouville-von-Neumann-equation-1}
\end{equation}
From Eq.~(\ref{Liouville-von-Neumann-equation-1}),
we notice that we can rewrite
$\rho_{\mbox{\scriptsize I}}(t)$
as
\begin{eqnarray}
\rho_{\mbox{\scriptsize I}}(t)
&=&
e^{-iC_{2}t}\rho_{\mbox{\scriptsize I}}(0)e^{iC_{2}t} \nonumber \\
&=&
\mbox{Tr}_{\tilde{\cal H}}
\Bigl[
e^{-iC_{2}t}
[
|0(\Theta)\rangle_{\mbox{\scriptsize F}}{}_{\mbox{\scriptsize F}}\langle 0(\Theta)|
\otimes
|\alpha;\theta\rangle_{\mbox{\scriptsize B}}{}_{\mbox{\scriptsize B}}\langle\alpha;\theta|
]e^{iC_{2}t}
\Bigr].
\label{Liouville-von-Neumann-equation-2}
\end{eqnarray}
Moreover,
the probability that
we detect the ground state of the atom at zero temperature
is given by
\begin{equation}
P_{g}(\Theta,\theta;t)
=
{}_{\mbox{\scriptsize F}}\langle 0|
\mbox{Tr}_{\mbox{\scriptsize B}}[
\rho_{\mbox{\scriptsize I}}(t)
]
|0\rangle_{\mbox{\scriptsize F}}.
\label{Liouville-von-Neumann-equation-3}
\end{equation}
The physical meaning of Eqs.~(\ref{Liouville-von-Neumann-equation-2})
and (\ref{Liouville-von-Neumann-equation-3})
is equivalent to
the discussion developed
from Eq.~(\ref{extended-JCM-initial-state})
until Eq.~(\ref{Probability-atom-ground-state-theta-2})
in Sec.~\ref{section-formulation-perturbetion-theory}.
Thus,
comparing Eqs.~(\ref{Probability-atom-ground-state-theta-2})
and
(\ref{Liouville-von-Neumann-equation-3}),
we cannot find distinct differences
between the TFD formalism and the Liouville-von Neumann equation.
However,
if we take the TFD formalism,
we can express a physical quantity
as a power series in $\theta(\beta)$
such as Eqs.~(\ref{Perturbation-theory-formula-0}),
(\ref{Perturbation-theory-formula-1}),
(\ref{Perturbation-theory-formula-2})
and
(\ref{Perturbation-theory-formula-3}).
Because of this advantage,
the TFD formalism is superior than the Liouville-von Neumann equation
for computing physical quantities actually.
The reason why we take the TFD formalism in this paper
for describing the JCM at finite temperature is the fact mentioned above.
And this prescription is a new key point of this paper as compared with the other past works.
In fact,
if we rewrite Eqs.~(\ref{Liouville-von-Neumann-equation-2})
and
(\ref{Liouville-von-Neumann-equation-3})
as a low-temperature expansion without using the TFD formalism,
we have to carry out the following calculations:
\begin{eqnarray}
&&
\mbox{Tr}_{\tilde{\cal H}}
[|\alpha;\theta\rangle_{\mbox{\scriptsize B}}{}_{\mbox{\scriptsize B}}\langle\alpha;\theta|] \nonumber \\
&=&
\mbox{Tr}_{\tilde{\cal H}}
[
\hat{U}_{\mbox{\scriptsize B}}(\theta)
|\alpha\rangle_{\mbox{\scriptsize B}}
|\tilde{\alpha}\rangle_{\mbox{\scriptsize B}}
{}_{\mbox{\scriptsize B}}\langle\alpha|
{}_{\mbox{\scriptsize B}}\langle\tilde{\alpha}|
\hat{U}_{\mbox{\scriptsize B}}^{\dagger}(\theta)
] \nonumber \\
&=&
\mbox{Tr}_{\tilde{\cal H}}
\Bigl(
\exp[-\theta(a\tilde{a}-\tilde{a}^{\dagger}a^{\dagger})]
|\alpha\rangle_{\mbox{\scriptsize B}}
|\tilde{\alpha}\rangle_{\mbox{\scriptsize B}}
{}_{\mbox{\scriptsize B}}\langle\alpha|
{}_{\mbox{\scriptsize B}}\langle\tilde{\alpha}|
\exp[\theta(a\tilde{a}-\tilde{a}^{\dagger}a^{\dagger})]
\Bigr) \nonumber \\
&=&
\mbox{Tr}_{\tilde{\cal H}}
\Bigl(
|\alpha\rangle_{\mbox{\scriptsize B}}
|\tilde{\alpha}\rangle_{\mbox{\scriptsize B}}
{}_{\mbox{\scriptsize B}}\langle\alpha|
{}_{\mbox{\scriptsize B}}\langle\tilde{\alpha}|
-\theta
[a\tilde{a}-\tilde{a}^{\dagger}a^{\dagger},
|\alpha\rangle_{\mbox{\scriptsize B}}
|\tilde{\alpha}\rangle_{\mbox{\scriptsize B}}
{}_{\mbox{\scriptsize B}}\langle\alpha|
{}_{\mbox{\scriptsize B}}\langle\tilde{\alpha}|] \nonumber \\
&&
+\frac{\theta^{2}}{2!}
[a\tilde{a}-\tilde{a}^{\dagger}a^{\dagger},
[a\tilde{a}-\tilde{a}^{\dagger}a^{\dagger},
|\alpha\rangle_{\mbox{\scriptsize B}}
|\tilde{\alpha}\rangle_{\mbox{\scriptsize B}}
{}_{\mbox{\scriptsize B}}\langle\alpha|
{}_{\mbox{\scriptsize B}}\langle\tilde{\alpha}|]]
+...
\Bigr).
\end{eqnarray}
The above calculations are essentially equivalent to
Eqs.~(\ref{Perturbation-theory-formula-0}),
(\ref{Perturbation-theory-formula-1}),
(\ref{Perturbation-theory-formula-2})
and
(\ref{Perturbation-theory-formula-3}).
However,
the perturbation theory via
the TFD formalism provides us
a clearer insight and a more accurate understanding
than the Liouville-von Neumann equation does.
\section{\label{section-0th-order-correction-term}The zero-th order correction}
From Eqs.~(\ref{definition-functions-g-1&2}),
(\ref{Perturbation-theory-formula-1})
and
(\ref{Perturbation-theory-formula-2}),
we can write down the zero-th order correction as
\begin{equation}
P^{(0)}_{g}(\Theta,\theta;t)
=
\cos^{2}\Theta P^{(0)}_{g,1}(t)
+
\sin^{2}\Theta P^{(0)}_{g,2}(t),
\label{0th-order-perturbation-0}
\end{equation}
\begin{eqnarray}
P^{(0)}_{g,1}(t)
&=&
\langle\alpha|
g_{1}(a^{\dagger}a+c)
|\alpha\rangle \nonumber \\
&=&
e^{-\alpha^{2}}
\sum_{n=0}^{\infty}
\frac{\alpha^{2n}}{n!}
g_{1}(n+c) \nonumber \\
&=&
e^{-\alpha^{2}}
\sum_{n=0}^{\infty}
\frac{\alpha^{2n}}{n!}
[
\cos^{2}(\sqrt{n+c}|\kappa|t)
+
c
\frac{\sin^{2}(\sqrt{n+c}|\kappa|t)}{n+c}
], \nonumber \\
P^{(0)}_{g,2}(t)
&=&
\langle\alpha|
g_{2}(a^{\dagger}a+c+1)
|\alpha\rangle \nonumber \\
&=&
e^{-\alpha^{2}}
\sum_{n=0}^{\infty}
\frac{\alpha^{2n}}{n!}
g_{2}(n+c+1) \nonumber \\
&=&
e^{-\alpha^{2}}
\sum_{n=0}^{\infty}
\frac{\alpha^{2n}}{n!}
\frac{\sin^{2}(\sqrt{n+c+1}|\kappa|t)}{n+c+1}(n+1).
\label{0th-order-perturbation-term}
\end{eqnarray}
Referring to Eq.~(\ref{Probability-collapse-and-revival-Rabi-oscillations-0}),
we note that
$P^{(0)}_{g,1}(t)=P_{g}(t)$
and
$P^{(0)}_{g}(0,0;t)=P_{g}(t)$.
For the convenience of calculations
carried out in the remains of this paper,
we define the following functions,
each of which is represented as an infinite series:
\begin{eqnarray}
Q_{1}^{(l)}(t)
&=&
e^{-\alpha^{2}}
\sum_{n=0}^{\infty}
\frac{\alpha^{2n}}{n!}
g_{1}(n+c+l) \nonumber \\
&=&
e^{-\alpha^{2}}
\sum_{n=0}^{\infty}
\frac{\alpha^{2n}}{n!}
[
\cos^{2}(\sqrt{n+c+l}|\kappa|t)
+
c
\frac{\sin^{2}(\sqrt{n+c+l}|\kappa|t)}{n+c+l}
], \nonumber \\
Q_{2}^{(l)}(t)
&=&
e^{-\alpha^{2}}
\sum_{n=0}^{\infty}
\frac{\alpha^{2n}}{n!}
g_{2}(n+c+1+l) \nonumber \\
&=&
e^{-\alpha^{2}}
\sum_{n=0}^{\infty}
\frac{\alpha^{2n}}{n!}
\frac{\sin^{2}(\sqrt{n+c+1+l}|\kappa|t)}{n+c+1+l}(n+1+l).
\end{eqnarray}
From the above definitions,
we obtain the zero-th order correction terms as
\begin{eqnarray}
P^{(0)}_{g,1}(t)
&=&
Q_{1}^{(0)}(t), \nonumber \\
P^{(0)}_{g,2}(t)
&=&
Q_{2}^{(0)}(t).
\label{0th-order-perturbation-term-1}
\end{eqnarray}
\section{\label{section-1st-order-correction-term}The first order correction}
From Eqs.~(\ref{Perturbation-theory-formula-1})
and
(\ref{Perturbation-theory-formula-4}),
we can write down the first order correction as
\begin{equation}
P^{(1)}_{g}(\Theta,\theta;t)
=
\cos^{2}\Theta P^{(1)}_{g,1}(t)
+
\sin^{2}\Theta P^{(1)}_{g,2}(t),
\label{1st-order-perturbation-0}
\end{equation}
\begin{eqnarray}
P^{(1)}_{g,1}(t)
&=&
\sum_{n=0}^{\infty}
g^{(n)}_{1}
\langle\alpha|
\langle\tilde{\alpha}|
[a\tilde{a}-a^{\dagger}\tilde{a}^{\dagger},
(a^{\dagger}a+c)^{n}]
|\alpha\rangle
|\tilde{\alpha}\rangle, \nonumber \\
P^{(1)}_{g,2}(t)
&=&
\sum_{n=0}^{\infty}
g^{(n)}_{2}
\langle\alpha|
\langle\tilde{\alpha}|
[a\tilde{a}-a^{\dagger}\tilde{a}^{\dagger},
(a^{\dagger}a+c+1)^{n}]
|\alpha\rangle
|\tilde{\alpha}\rangle.
\label{first-order-perturbation-terms-0}
\end{eqnarray}
From Eq.~(\ref{first-order-perturbation-terms-0}),
we notice that we have to calculate the commutation relations,
\begin{equation}
[a\tilde{a}-a^{\dagger}\tilde{a}^{\dagger},
(a^{\dagger}a+c)^{n}]
\quad\mbox{for $n=0,1,2,...$}.
\end{equation}
At first,
we define the following three operators:
\begin{equation}
\hat{A}
=
a^{\dagger}\tilde{a}^{\dagger}-a\tilde{a},
\quad
\hat{B}
=
a^{\dagger}a+c,
\quad
\hat{C}
=
a^{\dagger}\tilde{a}^{\dagger}+a\tilde{a}.
\label{definition-hatA-hatB-hatC-1}
\end{equation}
[We pay attention to the fact that
the operator $\hat{C}$ defined in Eq.~(\ref{definition-hatA-hatB-hatC-1})
is different from
$C_{1}$, $C_{2}$, $\hat{C}_{1}$ and $\hat{C}_{2}$
given by Eqs.~(\ref{JCM-Hamiltonian-decomposition-0})
and
(\ref{decomposition-hat-Hamiltonian}).]
Then,
we obtain the commutation relations,
\begin{eqnarray}
[\hat{A},\hat{B}]
&=&
-\hat{C}, \nonumber \\
{[}\hat{A},\hat{B}^{2}{]}
&=&
\hat{A}-2\hat{B}\hat{C}, \nonumber \\
{[}\hat{A},\hat{B}^{3}{]}
&=&
-\hat{C}+3\hat{B}\hat{A}-3\hat{B}^{2}\hat{C}, \nonumber \\
{[}\hat{A},\hat{B}^{4}{]}
&=&
\hat{A}-4\hat{B}\hat{C}+6\hat{B}^{2}\hat{A}-4\hat{B}^{3}\hat{C}, \nonumber \\
&&....
\label{commutation-relations-A-Bn-1}
\end{eqnarray}
Next,
we define the following two operators:
\begin{equation}
\hat{\mu}=a^{\dagger}\tilde{a}^{\dagger},
\quad
\hat{\nu}=a\tilde{a},
\label{definition-operators-mu-nu-1}
\end{equation}
and
we obtain
\begin{equation}
\hat{A}=\hat{\mu}-\hat{\nu},
\quad
\hat{C}=\hat{\mu}+\hat{\nu}.
\label{definition-operators-mu-nu-2}
\end{equation}
Using the operators $\hat{\mu}$ and $\hat{\nu}$,
we can rewrite Eq.~(\ref{commutation-relations-A-Bn-1})
as the general form,
\begin{equation}
[\hat{A},\hat{B}^{n}]
=
(\hat{B}-1)^{n}\hat{\mu}-(\hat{B}+1)^{n}\hat{\nu}-\hat{B}^{n}\hat{A}
\quad\mbox{for $n=1,2,3,...$}.
\label{commutation-relations-A-Bn-3}
\end{equation}
We can prove Eq.~(\ref{commutation-relations-A-Bn-3})
with the mathematical induction as follows:
First, we can confirm that Eq.~(\ref{commutation-relations-A-Bn-3}) holds for $n=1$, at ease.
Second, we assume Eq.~(\ref{commutation-relations-A-Bn-3}) holds for some unspecified number $n(\geq 1)$.
Third, we compute the commutation relation,
\begin{eqnarray}
[\hat{A},\hat{B}^{n+1}]
&=&
[\hat{A},\hat{B}^{n}]\hat{B}+\hat{B}^{n}[\hat{A},\hat{B}] \nonumber \\
&=&
\Bigl((\hat{B}-1)^{n}\hat{\mu}-(\hat{B}+1)^{n}\hat{\nu}-\hat{B}^{n}\hat{A}\Bigr)\hat{B}
+\hat{B}^{n}[\hat{A},\hat{B}] \nonumber \\
&=&
(\hat{B}-1)^{n+1}\hat{\mu}-(\hat{B}+1)^{n+1}\hat{\nu}-\hat{B}^{n+1}\hat{A},
\end{eqnarray}
where we use $\hat{\mu}\hat{B}=(\hat{B}-1)\hat{\mu}$ and $\hat{\nu}\hat{B}=(\hat{B}+1)\hat{\nu}$.
Thus,
we obtain
\begin{eqnarray}
&&
{[}
a\tilde{a}-a^{\dagger}\tilde{a}^{\dagger}, (a^{\dagger}a+c)^{n}
{]} \nonumber \\
&=&
-(a^{\dagger}a+c-1)^{n}\tilde{a}^{\dagger}a^{\dagger}
+(a^{\dagger}a+c+1)^{n}a\tilde{a}
-(a^{\dagger}a+c)^{n}(a\tilde{a}-a^{\dagger}\tilde{a}^{\dagger}) \nonumber \\
&=&
-a^{\dagger}\tilde{a}^{\dagger}(a^{\dagger}a+c)^{n}
+(a^{\dagger}a+c+1)^{n}a\tilde{a}
-(a^{\dagger}a+c)^{n}a\tilde{a} \nonumber \\
&&
+a^{\dagger}\tilde{a}^{\dagger}(a^{\dagger}a+c+1)^{n}.
\label{first-order-perturbation-commutation-relations-a}
\end{eqnarray}
Moreover,
replacing $c$ in Eq.~(\ref{first-order-perturbation-commutation-relations-a})
with $(c+1)$,
we obtain the commutation relation,
\begin{eqnarray}
&&
{[}
a\tilde{a}-a^{\dagger}\tilde{a}^{\dagger}, (a^{\dagger}a+c+1)^{n}
{]} \nonumber \\
&=&
-a^{\dagger}\tilde{a}^{\dagger}(a^{\dagger}a+c+1)^{n}
+(a^{\dagger}a+c+2)^{n}a\tilde{a} \nonumber \\
&&
-(a^{\dagger}a+c+1)^{n}a\tilde{a}
+a^{\dagger}\tilde{a}^{\dagger}(a^{\dagger}a+c+2)^{n}.
\label{first-order-perturbation-commutation-relations-b}
\end{eqnarray}
Hence,
using the relations $a|\alpha\rangle=\alpha|\alpha\rangle$
and $\tilde{a}|\tilde{\alpha}\rangle=\alpha|\tilde{\alpha}\rangle$,
substitution of Eq.~(\ref{first-order-perturbation-commutation-relations-a})
into Eq.~(\ref{first-order-perturbation-terms-0})
yields
\begin{eqnarray}
P^{(1)}_{g,1}(t)
&=&
\sum_{n=0}^{\infty}
g^{(n)}_{1}
\langle\alpha|
\langle\tilde{\alpha}|
[
-a^{\dagger}\tilde{a}^{\dagger}(a^{\dagger}a+c)^{n}
+(a^{\dagger}a+c+1)^{n}a\tilde{a} \nonumber \\
&&
-(a^{\dagger}a+c)^{n}a\tilde{a}
+a^{\dagger}\tilde{a}^{\dagger}(a^{\dagger}a+c+1)^{n}
]
|\alpha\rangle
|\tilde{\alpha}\rangle, \nonumber \\
&=&
\langle\alpha|
\langle\tilde{\alpha}|
[
-a^{\dagger}\tilde{a}^{\dagger}g_{1}(a^{\dagger}a+c)
+g_{1}(a^{\dagger}a+c+1)a\tilde{a} \nonumber \\
&&
-g_{1}(a^{\dagger}a+c)a\tilde{a}
+a^{\dagger}\tilde{a}^{\dagger}g_{1}(a^{\dagger}a+c+1)
]
|\alpha\rangle
|\tilde{\alpha}\rangle, \nonumber \\
&=&
-2\alpha^{2}
\langle\alpha|
\langle\tilde{\alpha}|
[g_{1}(a^{\dagger}a+c)-g_{1}(a^{\dagger}a+c+1)]
|\alpha\rangle
|\tilde{\alpha}\rangle \nonumber \\
&=&
-2\alpha^{2}
e^{-\alpha^{2}}\sum_{n=0}^{\infty}
\frac{\alpha^{2n}}{n!}[g_{1}(n+c)-g_{1}(n+c+1)] \nonumber \\
&=&
-2\alpha^{2}[Q_{1}^{(0)}(t)-Q_{1}^{(1)}(t)].
\label{1st-order-perturbation-term-g1}
\end{eqnarray}
Similarly,
substitution of Eq.~(\ref{first-order-perturbation-commutation-relations-b})
into
Eq.~(\ref{first-order-perturbation-terms-0})
yields
\begin{eqnarray}
P^{(1)}_{g,2}(t)
&=&
-2\alpha^{2}
e^{-\alpha^{2}}\sum_{n=0}^{\infty}
\frac{\alpha^{2n}}{n!}[g_{2}(n+c+1)-g_{2}(n+c+2)] \nonumber \\
&=&
-2\alpha^{2}[Q_{2}^{(0)}(t)-Q_{2}^{(1)}(t)].
\label{1st-order-perturbation-term-g2}
\end{eqnarray}
\section{\label{section-2nd-order-correction-term}The second order correction}
From Eq.~(\ref{Perturbation-theory-formula-4}),
we can write down the second order terms as
\begin{eqnarray}
P^{(2)}_{g,1}(t)
&=&
\sum_{n=0}^{\infty}
g^{(n)}_{1}
\langle\alpha|
\langle\tilde{\alpha}|
[a\tilde{a}-a^{\dagger}\tilde{a}^{\dagger},
[a\tilde{a}-a^{\dagger}\tilde{a}^{\dagger},
(a^{\dagger}a+c)^{n}]]
|\alpha\rangle
|\tilde{\alpha}\rangle, \nonumber \\
P^{(2)}_{g,2}(t)
&=&
\sum_{n=0}^{\infty}
g^{(n)}_{2}
\langle\alpha|
\langle\tilde{\alpha}|
[a\tilde{a}-a^{\dagger}\tilde{a}^{\dagger},
[a\tilde{a}-a^{\dagger}\tilde{a}^{\dagger},
(a^{\dagger}a+c+1)^{n}]]
|\alpha\rangle
|\tilde{\alpha}\rangle.
\label{second-order-perturbation-terms-0}
\end{eqnarray}
Looking at Eqs.~(\ref{definition-hatA-hatB-hatC-1})
and
(\ref{second-order-perturbation-terms-0}),
we notice that we need to calculate the commutation relation
$[\hat{A},[\hat{A},\hat{B}^{n}]]$.
From now on,
referring to Eq.~(\ref{commutation-relations-A-Bn-3}),
we divide the commutation relation
$[\hat{A},[\hat{A},\hat{B}^{n}]]$
into the following two parts
and examine each of them:
\begin{equation}
[\hat{A},[\hat{A},\hat{B}^{n}]]
=
\hat{R}_{n}+\hat{S}_{n}
\quad\mbox{for $n=1,2,3,...$},
\label{commutation-relation-Rn-Sn}
\end{equation}
where
\begin{eqnarray}
\hat{R}_{n}
&=&
[\hat{A},(\hat{B}-1)^{n}]\hat{\mu}
-
[\hat{A},(\hat{B}+1)^{n}]\hat{\nu}
-
[\hat{A},\hat{B}^{n}]\hat{A}, \nonumber \\
\hat{S}_{n}
&=&
(\hat{B}-1)^{n}[\hat{A},\hat{\mu}]
-
(\hat{B}+1)^{n}[\hat{A},\hat{\nu}].
\label{defintion-Rn-Sn-i}
\end{eqnarray}
According to Eq.~(\ref{commutation-relation-Rn-Sn}),
we divide one of the second order terms given by Eq.~(\ref{second-order-perturbation-terms-0})
into two parts as
\begin{equation}
P^{(2)}_{g,1}(t)
=
\sum_{n=0}^{\infty}
g^{(n)}_{1}
\langle \alpha|\langle\tilde{\alpha}|
\hat{R}_{n}
|\alpha\rangle|\tilde{\alpha}\rangle
+
\sum_{n=0}^{\infty}
g^{(n)}_{1}
\langle \alpha|\langle\tilde{\alpha}|
\hat{S}_{n}
|\alpha\rangle|\tilde{\alpha}\rangle.
\label{second-order-correction-i}
\end{equation}
Here,
we show $P^{(2)}_{g,1}(t)$
as a concrete example in Eq.~(\ref{second-order-correction-i}).
We understand obviously
that we can compute $P^{(2)}_{g,2}(t)$
after the manner of $P^{(2)}_{g,1}(t)$.
Thus,
for simplicity,
we concentrate on evaluating $P^{(2)}_{g,1}(t)$ in the following paragraphs.
At first,
we examine the part which includes $\{\hat{R}_{n}\}$ in Eq.~(\ref{second-order-correction-i}).
From Eq.~(\ref{defintion-Rn-Sn-i}),
we obtain
\begin{eqnarray}
\hat{R}_{0}
&=&
0, \nonumber \\
\hat{R}_{1}
&=&
0, \nonumber \\
\hat{R}_{2}
&=&
-2[\hat{A},\hat{B}]\hat{C}, \nonumber \\
\hat{R}_{3}
&=&
-3[\hat{A},\hat{B}^{2}]\hat{C}
+3[\hat{A},\hat{B}]\hat{A}, \nonumber \\
\hat{R}_{4}
&=&
-4[\hat{A},\hat{B}^{3}]\hat{C}
+6[\hat{A},\hat{B}^{2}]\hat{A}
-4[\hat{A},\hat{B}]\hat{C}, \nonumber \\
\hat{R}_{5}
&=&
-5[\hat{A},\hat{B}^{4}]\hat{C}
+10[\hat{A},\hat{B}^{3}]\hat{A}
-10[\hat{A},\hat{B}^{2}]\hat{C}
+5[\hat{A},\hat{B}]\hat{A}, \nonumber \\
\hat{R}_{6}
&=&
-6[\hat{A},\hat{B}^{5}]\hat{C}
+15[\hat{A},\hat{B}^{4}]\hat{A}
-20[\hat{A},\hat{B}^{3}]\hat{C}
+15[\hat{A},\hat{B}^{2}]\hat{A}
-6[\hat{A},\hat{B}]\hat{C}, \nonumber \\
&&....
\label{defintion-Rn-Sn-ii}
\end{eqnarray}
Thus,
using Eq.~(\ref{commutation-relations-A-Bn-3}) and the following formula:
\begin{eqnarray}
\sum_{n=0}^{\infty}
\left(
\begin{array}{c}
n+m+1 \\
m
\end{array}
\right)
g^{(n+m+1)}_{1}
x^{n+1}
&=&
\frac{1}{m!}\frac{d^{m}}{dx^{m}}g_{1}(x)-g^{(m)}_{1} \nonumber \\
&&
\quad
\mbox{for $m=1,2,3,...$},
\label{formula-g-polynomial-1}
\end{eqnarray}
where $g^{(m)}_{1}$ is defined
in Eqs.~(\ref{g-function-Taylor-series-0}) and (\ref{g-function-Taylor-series-1}),
we can rewrite
the part including $\{\hat{R}_{n}\}$
in the second order term of Eq.~(\ref{second-order-correction-i})
as
\begin{eqnarray}
&&
\langle \alpha|\langle\tilde{\alpha}|
\Bigr(
-
\sum_{n=0}^{\infty}
\left(
\begin{array}{c}
n+2 \\
1
\end{array}
\right)
g^{(n+2)}_{1}[\hat{A},\hat{B}^{n+1}]\hat{C} \nonumber \\
&&
+
\sum_{n=0}^{\infty}
\left(
\begin{array}{c}
n+3 \\
2
\end{array}
\right)
g^{(n+3)}_{1}[\hat{A},\hat{B}^{n+1}]\hat{A} \nonumber \\
&&
-
\sum_{n=0}^{\infty}
\left(
\begin{array}{c}
n+4 \\
3
\end{array}
\right)
g^{(n+4)}_{1}[\hat{A},\hat{B}^{n+1}]\hat{C}
+
...
\Bigr)
|\alpha\rangle|\tilde{\alpha}\rangle \nonumber \\
&=&
\langle \alpha|\langle\tilde{\alpha}|
\Bigr(
-\hat{F}_{1}\hat{C}
+\frac{1}{2!}\hat{F}_{2}\hat{A}
-\frac{1}{3!}\hat{F}_{3}\hat{C}
+...
\Bigr)
|\alpha\rangle|\tilde{\alpha}\rangle \nonumber \\
&=&
\langle \alpha|\langle\tilde{\alpha}|
\Bigr(
\sum_{n=0}^{\infty}
\frac{(-1)^{n}}{n!}\hat{F}_{n}\hat{\mu}
-
\sum_{n=0}^{\infty}
\frac{1}{n!}\hat{F}_{n}\hat{\nu}
\Bigr)
|\alpha\rangle|\tilde{\alpha}\rangle \nonumber \\
&&
-
\langle \alpha|\langle\tilde{\alpha}|
\Bigr(
g_{1}(\hat{B}-1)\hat{\mu}-g_{1}(\hat{B}+1)\hat{\nu}-g_{1}(\hat{B})\hat{A}
\Bigr)\hat{A}
|\alpha\rangle|\tilde{\alpha}\rangle,
\label{second-order-correction-R-part-i}
\end{eqnarray}
where
\begin{equation}
\hat{F}_{n}
=
\frac{d^{n}}{dx^{n}}g_{1}(x)\bigg|_{x=\hat{B}-1}\hat{\mu}
-
\frac{d^{n}}{dx^{n}}g_{1}(x)\bigg|_{x=\hat{B}+1}\hat{\nu}
-
\frac{d^{n}}{dx^{n}}g_{1}(x)\bigg|_{x=\hat{B}}\hat{A}.
\label{definition-operator-Fn}
\end{equation}
In the derivation of Eq.~(\ref{second-order-correction-R-part-i}),
we use Eq.~(\ref{commutation-relations-A-Bn-3})
in an effective manner.
The form of $\hat{F}_{n}$ in Eq.~(\ref{definition-operator-Fn})
reflects Eq.~(\ref{commutation-relations-A-Bn-3}).
Using the operators $e^{\pm d/dx}$,
we can rewrite the first term of the right-hand side of Eq.~(\ref{second-order-correction-R-part-i})
as
\begin{eqnarray}
&&
\langle \alpha|\langle\tilde{\alpha}|
\Bigr(
[
e^{-d/dx}g_{1}(\hat{B}-1)\hat{\mu}
-
e^{-d/dx}g_{1}(\hat{B}+1)\hat{\nu}
-
e^{-d/dx}g_{1}(\hat{B})\hat{A}
]\hat{\mu} \nonumber \\
&&
-
[
e^{d/dx}g_{1}(\hat{B}-1)\hat{\mu}
-
e^{d/dx}g_{1}(\hat{B}+1)\hat{\nu}
-
e^{d/dx}g_{1}(\hat{B})\hat{A}
]\hat{\nu}
\Bigr)
|\alpha\rangle|\tilde{\alpha}\rangle.
\label{second-order-correction-R-part-ii}
\end{eqnarray}
Then,
we apply the following technique to Eq.~(\ref{second-order-correction-R-part-ii}):
\begin{equation}
e^{\pm d/dx}g_{1}(\hat{X})
=
\sum_{n=0}^{\infty}
\frac{(\pm 1)^{n}}{n!}
\frac{d^{n}}{dx^{n}}g_{1}(x)\bigg|_{x=\hat{X}}
=
g_{1}(\hat{X}\pm 1),
\end{equation}
where $\hat{X}$ is an arbitrary operator.
Thus,
we can rewrite Eq.~(\ref{second-order-correction-R-part-ii})
as
\begin{eqnarray}
&&
\langle \alpha|\langle\tilde{\alpha}|
\Bigr(
[
g_{1}(\hat{B}-2)\hat{\mu}
-
g_{1}(\hat{B})\hat{\nu}
-
g_{1}(\hat{B}-1)\hat{A}
]\hat{\mu} \nonumber \\
&&
-
[
g_{1}(\hat{B})\hat{\mu}
-
g_{1}(\hat{B}+2)\hat{\nu}
-
g_{1}(\hat{B}+1)\hat{A}
]\hat{\nu}
\Bigr)
|\alpha\rangle|\tilde{\alpha}\rangle.
\label{second-order-correction-R-part-iii}
\end{eqnarray}
Next,
we examine the part including $\{\hat{S}_{n}\}$
in Eq.~(\ref{second-order-correction-i}).
From Eqs.~(\ref{definition-operators-mu-nu-1})
and
(\ref{definition-operators-mu-nu-2}),
we obtain
\begin{equation}
[\hat{A},\hat{\mu}]=[\hat{A},\hat{\nu}]=[\hat{\mu},\hat{\nu}]=-\hat{D},
\label{commutation-relationd-related-D}
\end{equation}
where
\begin{equation}
\hat{D}=a^{\dagger}a+\tilde{a}^{\dagger}\tilde{a}+1.
\end{equation}
In the remains of this section and Appendix~\ref{section-3rd-order-correction-term},
we use Eq.~(\ref{commutation-relationd-related-D}) and the following commutation relations often
without notice:
\begin{equation}
[\hat{A},\hat{C}]=-2\hat{D},
\quad\quad
[\hat{A},\hat{D}]=-2\hat{C}.
\end{equation}
Then,
we can rewrite $\hat{S}_{n}$ given by Eq.~(\ref{defintion-Rn-Sn-i}) as
\begin{equation}
\hat{S}_{n}
=
-
\Bigl(
(\hat{B}-1)^{n}-(\hat{B}+1)^{n}
\Bigr)
\hat{D}.
\label{formula-Sn-1}
\end{equation}
[In Eq.~(57) of Ref.~\cite{Azuma2010},
a calculation concerning
$\hat{S}_{n}$
is wrong.]
Thus, we can write down the part including $\{\hat{S}_{n}\}$
in the second order term given by Eq.~(\ref{second-order-correction-i})
as
\begin{equation}
-
\langle \alpha|\langle\tilde{\alpha}|
\Bigr(
g_{1}(\hat{B}-1)-g_{1}(\hat{B}+1)
\Bigr)
\hat{D}
|\alpha\rangle|\tilde{\alpha}\rangle.
\label{second-order-correction-S-part-i}
\end{equation}
Putting together Eqs.~(\ref{second-order-correction-i}),
(\ref{second-order-correction-R-part-i}),
(\ref{second-order-correction-R-part-iii})
and
(\ref{second-order-correction-S-part-i}),
we can write down the whole of the second order term
as
\begin{eqnarray}
&&
P^{(2)}_{g,1}(t) \nonumber \\
&=&
\langle \alpha|\langle\tilde{\alpha}|
[
g_{1}(\hat{B}-2)\hat{\mu}^{2}
+
g_{1}(\hat{B}-1)(-\hat{D}-\hat{\mu}\hat{A}-\hat{A}\hat{\mu}) \nonumber \\
&&
+
g_{1}(\hat{B})(\hat{A}^{2}-\hat{\nu}\hat{\mu}-\hat{\mu}\hat{\nu})
+
g_{1}(\hat{B}+1)(\hat{D}+\hat{\nu}\hat{A}+\hat{A}\hat{\nu}) \nonumber \\
&&
+
g_{1}(\hat{B}+2)\hat{\nu}^{2}]
|\alpha\rangle|\tilde{\alpha}\rangle.
\label{second-order-correction-ii}
\end{eqnarray}
Here,
to compute $P^{(2)}_{g,1}(t)$ given by Eq.~(\ref{second-order-correction-ii}),
we arrange $\hat{\mu}$ in the left side of the product of operators
and $\hat{\nu}$ in the right side of the product of operators.
For the arrangement of operators,
we carry out the calculations,
\begin{eqnarray}
(\hat{B}-2)^{n}\hat{\mu}^{2}
&=&
\hat{\mu}^{2}
\hat{B}^{n}, \nonumber \\
(\hat{B}-1)^{n}(-\hat{D}-\hat{\mu}\hat{A}-\hat{A}\hat{\mu})
&=&
-2\hat{\mu}^{2}(\hat{B}+1)^{n}
+2\hat{\mu}\hat{B}^{n}\hat{\nu}, \nonumber \\
\hat{B}^{n}(\hat{A}^{2}-\hat{\nu}\hat{\mu}-\hat{\mu}\hat{\nu})
&=&
\hat{\mu}^{2}(\hat{B}+2)^{n}
-4\hat{\mu}(\hat{B}+1)^{n}\hat{\nu}
-2\hat{B}^{n}\hat{D}+\hat{B}^{n}\hat{\nu}^{2}, \nonumber \\
(\hat{B}+1)^{n}(\hat{D}+\hat{\nu}\hat{A}+\hat{A}\hat{\nu})
&=&
2\hat{\mu}(\hat{B}+2)^{n}\hat{\nu}
-2(\hat{B}+1)^{n}\hat{\nu}^{2}
+2(\hat{B}+1)^{n}\hat{D} \nonumber \\
&&
\mbox{for $n=1,2,3,...$.}
\end{eqnarray}
Substitution of the above relations into Eq.~(\ref{second-order-correction-ii})
yields
\begin{eqnarray}
P^{(2)}_{g,1}(t)
&=&
\langle\alpha|\langle\tilde{\alpha}|
[
\hat{\mu}^{2}g_{1}(\hat{B})
-2\hat{\mu}^{2}g_{1}(\hat{B}+1)
+2\hat{\mu}g_{1}(\hat{B})\hat{\nu}
+\hat{\mu}^{2}g_{1}(\hat{B}+2) \nonumber \\
&&
-4\hat{\mu}g_{1}(\hat{B}+1)\hat{\nu}
-2g_{1}(\hat{B})\hat{D}
+g_{1}(\hat{B})\hat{\nu}^{2}
+2\hat{\mu}g_{1}(\hat{B}+2)\hat{\nu} \nonumber \\
&&
-2g_{1}(\hat{B}+1)\hat{\nu}^{2}
+2g_{1}(\hat{B}+1)\hat{D}
+g_{1}(\hat{B}+2)\hat{\nu}^{2}
]
|\alpha\rangle|\tilde{\alpha}\rangle \nonumber \\
&=&
4\alpha^{4}
\langle\alpha|\langle\tilde{\alpha}|
[
g_{1}(\hat{B})
-2g_{1}(\hat{B}+1)
+g_{1}(\hat{B}+2)
]
|\alpha\rangle|\tilde{\alpha}\rangle \nonumber \\
&&
-2
\langle\alpha|\langle\tilde{\alpha}|
g_{1}(\hat{B})\hat{D}
|\alpha\rangle|\tilde{\alpha}\rangle
+2
\langle\alpha|\langle\tilde{\alpha}|
g_{1}(\hat{B}+1)\hat{D}
|\alpha\rangle|\tilde{\alpha}\rangle.
\end{eqnarray}
Moreover,
preparing the following formula:
\begin{eqnarray}
&&
\langle\alpha|\langle\tilde{\alpha}|
g_{1}(\hat{B})\hat{D}
|\alpha\rangle|\tilde{\alpha}\rangle \nonumber \\
&=&
\langle\alpha|\langle\tilde{\alpha}|
e^{-\alpha^{2}}
\sum_{n=0}^{\infty}
\sum_{m=0}^{\infty}
\frac{\alpha^{n+m}}{\sqrt{n!m!}}
(n+m+1)g_{1}(n+c)
|n\rangle|\tilde{m}\rangle \nonumber \\
&=&
e^{-\alpha^{2}}
\sum_{n=0}^{\infty}
\frac{\alpha^{2n}}{n!}
g_{1}(n+c)
+
e^{-\alpha^{2}}
\sum_{n=0}^{\infty}
\frac{\alpha^{2n}}{n!}
n
g_{1}(n+c) \nonumber \\
&&
+
e^{-2\alpha^{2}}
\sum_{n=0}^{\infty}
\sum_{m=0}^{\infty}
\frac{\alpha^{2(n+m)}}{n!m!}
m
g_{1}(n+c), \nonumber \\
&=&
(1+\alpha^{2})
e^{-\alpha^{2}}
\sum_{n=0}^{\infty}
\frac{\alpha^{2n}}{n!}
g_{1}(n+c)
+
\alpha^{2}
e^{-\alpha^{2}}
\sum_{n=0}^{\infty}
\frac{\alpha^{2n}}{n!}
g_{1}(n+c+1),
\label{g-B-D-foemula}
\end{eqnarray}
we arrive at the final representation of $P^{(2)}_{g,1}(t)$ as
\begin{eqnarray}
P^{(2)}_{g,1}(t)
&=&
4\alpha^{4}
e^{-\alpha^{2}}
\sum_{n=0}^{\infty}
\frac{\alpha^{2n}}{n!}
[g_{1}(n+c)-2g_{1}(n+c+1)+g_{1}(n+c+2)] \nonumber \\
&&
-2
(1+\alpha^{2})
e^{-\alpha^{2}}
\sum_{n=0}^{\infty}
\frac{\alpha^{2n}}{n!}
g_{1}(n+c)
-2
\alpha^{2}
e^{-\alpha^{2}}
\sum_{n=0}^{\infty}
\frac{\alpha^{2n}}{n!}
g_{1}(n+c+1) \nonumber \\
&&
+2
(1+\alpha^{2})
e^{-\alpha^{2}}
\sum_{n=0}^{\infty}
\frac{\alpha^{2n}}{n!}
g_{1}(n+c+1)
+2
\alpha^{2}
e^{-\alpha^{2}}
\sum_{n=0}^{\infty}
\frac{\alpha^{2n}}{n!}
g_{1}(n+c+2) \nonumber \\
&=&
2(2\alpha^{2}+1)(\alpha+1)(\alpha-1)
Q_{1}^{(0)}(t) \nonumber \\
&&
-2(2\alpha^{2}+1)(2\alpha^{2}-1)
Q_{1}^{(1)}(t) \nonumber \\
&&
+2\alpha^{2}(2\alpha^{2}+1)
Q_{1}^{(2)}(t).
\label{2nd-order-perturbation-term-g1}
\end{eqnarray}
Similarly,
we obtain
$P^{(2)}_{g,2}(t)$
as
\begin{eqnarray}
P^{(2)}_{g,2}(t)
&=&
2(2\alpha^{2}+1)(\alpha+1)(\alpha-1)
Q_{2}^{(0)}(t) \nonumber \\
&&
-2(2\alpha^{2}+1)(2\alpha^{2}-1)
Q_{2}^{(1)}(t) \nonumber \\
&&
+2\alpha^{2}(2\alpha^{2}+1)
Q_{2}^{(2)}(t).
\label{2nd-order-perturbation-term-g2}
\end{eqnarray}
\section{\label{section-numerical-calculations}The numerical calculations}
In this section,
we show numerical results for the atomic population inversion
obtained with the third order perturbation theory under the low-temperature limit.
In Secs.~\ref{section-0th-order-correction-term},
\ref{section-1st-order-correction-term},
\ref{section-2nd-order-correction-term}
and Appendix~\ref{section-3rd-order-correction-term},
we obtain
$\{P_{g,1}^{(n)}(t),P_{g,2}^{(n)}(t):n\in\{0,1,2,3\}\}$
in the form of Eqs.~(\ref{0th-order-perturbation-term-1}),
(\ref{1st-order-perturbation-term-g1}),
(\ref{1st-order-perturbation-term-g2}),
(\ref{2nd-order-perturbation-term-g1}),
(\ref{2nd-order-perturbation-term-g2}),
(\ref{3rd-order-perturbation-term-g1})
and
(\ref{3rd-order-perturbation-term-g2}).
Thus,
from Eqs.~(\ref{definition-atomic-population-inversion}),
(\ref{Perturbation-theory-formula-0})
and
(\ref{Perturbation-theory-formula-1}),
we can calculate
$\langle\sigma_{z}(t)\rangle$
as the third order perturbation theory,
\begin{equation}
\langle\sigma_{z}(t)\rangle
=
1-2
[
\cos^{2}\Theta(\beta)
\sum_{n=0}^{3}
\frac{\theta(\beta)^{n}}{n!}
P_{g,1}^{(n)}(t)
+
\sin^{2}\Theta(\beta)
\sum_{n=0}^{3}
\frac{\theta(\beta)^{n}}{n!}
P_{g,2}^{(n)}(t)
].
\label{sigma_z_t_upto-3rd-order-perturbation}
\end{equation}
\begin{figure}
\begin{center}
\mbox{\scalebox{1.0}[1.0]{\includegraphics{Figure01.eps}}}
\vspace*{8pt}
\caption{The atomic population inversion $\langle\sigma_{z}(t)\rangle$
as a function of the time $t$
obtained from numerical calculations of Eq.~(\ref{sigma_z_t_upto-3rd-order-perturbation})
with
$\alpha=4$, $c=1$, $\kappa=1$ and $\Theta(\beta)=\theta(\beta)=0$.
Looking at the graph,
we estimate the time scale of the initial collapse
and
the period of the revival
of the Rabi oscillations
at unity and $8\pi$ around, respectively.}
\label{Figure01}
\end{center}
\vspace*{8pt}
\begin{center}
\mbox{\scalebox{1.0}[1.0]{\includegraphics{Figure02.eps}}}
\vspace*{8pt}
\caption{The atomic population inversion $\langle\sigma_{z}(t)\rangle$
as a function of the time $t$
obtained from numerical calculations of Eq.~(\ref{sigma_z_t_upto-3rd-order-perturbation})
with
$\alpha=4$, $c=1$, $\kappa=1$,
$\omega_{0}=2$, $\omega=4$,
$\theta(\beta)=\pi/32$
and
$\Theta(\beta)=\arctan[\tanh^{1/2}(\pi/32)]$.
Looking at the graph,
we estimate the time scale of the initial collapse
and
the period of the revival
of the Rabi oscillations
at unity and $8\pi e^{\pi/32}\simeq (8.83)\pi$ around, respectively.}
\label{Figure02}
\end{center}
\end{figure}
\begin{figure}
\begin{center}
\mbox{\scalebox{1.0}[1.0]{\includegraphics{Figure03.eps}}}
\vspace*{8pt}
\caption{The atomic population inversion $\langle\sigma_{z}(t)\rangle$
as a function of the time $t$
obtained from numerical calculations of Eq.~(\ref{sigma_z_t_upto-3rd-order-perturbation})
with
$\alpha=8$, $c=1$, $\kappa=1$ and $\Theta(\beta)=\theta(\beta)=0$.
Looking at the graph,
we estimate the time scale of the initial collapse
and
the period of the revival
of the Rabi oscillations
at unity and $16\pi$ around, respectively.}
\label{Figure03}
\end{center}
\vspace*{8pt}
\begin{center}
\mbox{\scalebox{1.0}[1.0]{\includegraphics{Figure04.eps}}}
\vspace*{8pt}
\caption{The atomic population inversion $\langle\sigma_{z}(t)\rangle$
as a function of the time $t$
obtained from numerical calculations of Eq.~(\ref{sigma_z_t_upto-3rd-order-perturbation})
with
$\alpha=8$, $c=1$, $\kappa=1$,
$\omega_{0}=2$, $\omega=4$,
$\theta(\beta)=\pi/60$
and
$\Theta(\beta)=\arctan[\tanh^{1/2}(\pi/60)]$.
Looking at the graph,
we estimate the time scale of the initial collapse
and
the period of the revival
of the Rabi oscillations
at unity and $16\pi e^{\pi/60}\simeq (16.9)\pi$ around, respectively.}
\label{Figure04}
\end{center}
\end{figure}
Figure~\ref{Figure01} shows
the atomic population inversion $\langle\sigma_{z}(t)\rangle$
given by Eq.~(\ref{sigma_z_t_upto-3rd-order-perturbation})
as a function of the time $t$
with
$\alpha=4$, $c=1$, $\kappa=1$ and $\Theta(\beta)=\theta(\beta)=0$.
Figure~\ref{Figure03} shows
the atomic population inversion $\langle\sigma_{z}(t)\rangle$
given by Eq.~(\ref{sigma_z_t_upto-3rd-order-perturbation})
as a function of the time $t$
with
$\alpha=8$, $c=1$, $\kappa=1$ and $\Theta(\beta)=\theta(\beta)=0$.
Carrying out numerical calculations for Figs.~\ref{Figure01} and \ref{Figure03},
we replace the summation $\sum_{n=0}^{\infty}$ in $Q_{1}^{(0)}(t)$
with $\sum_{n=0}^{100}$,
so that we compute the sum of first one hundred and one terms in the series.
(In this section,
whenever we carry out numerical calculations of $Q_{1}^{(l)}(t)$ and $Q_{2}^{(l)}(t)$,
we replace their summation $\sum_{n=0}^{\infty}$ with $\sum_{n=0}^{100}$.)
In Figs.~\ref{Figure01} and \ref{Figure03},
we assume the system to be at zero temperature.
Thus, the graphs in Figs.~\ref{Figure01} and \ref{Figure03}
do not suffer from thermal effects.
We can observe the collapse and the revival of the Rabi oscillations
obviously in these graphs.
Figure~\ref{Figure02} shows
the atomic population inversion $\langle\sigma_{z}(t)\rangle$
given by Eq.~(\ref{sigma_z_t_upto-3rd-order-perturbation})
as a function of the time $t$
with
$\alpha=4$, $c=1$, $\kappa=1$, $\theta(\beta)=\pi/32$, $\omega_{0}=2$ and $\omega=4$.
From Eq.~(\ref{definition-temperature-JCM-0}),
we obtain
\begin{eqnarray}
\theta(\beta)
&=&
\mbox{arctanh}(e^{-\beta\hbar\omega/2}), \nonumber \\
\Theta(\beta)
&=&
\arctan(e^{-\beta\hbar\omega_{0}/2}),
\label{definitions-theta-Theta-1}
\end{eqnarray}
so that the relation
$\exp(-2\beta\hbar)=\tanh[\theta(\beta)]=\tanh(\pi/32)$
holds.
Thus,
we can derive the following relation:
\begin{equation}
\Theta(\beta)
=\arctan(e^{-\beta\hbar})
=\arctan[\tanh^{1/2}(\theta(\beta))]
=\arctan[\tanh^{1/2}(\pi/32)].
\label{definitions-theta-Theta-2}
\end{equation}
Because the system of Fig.~\ref{Figure02} evolves in time with maintaining constant low temperature,
its time-evolution is under the thermal effects.
Comparing the graphs shown in Figs.~\ref{Figure01} and \ref{Figure02},
we notice that the period of Fig.~\ref{Figure02} is longer than the period of Fig.~\ref{Figure01}.
Thus,
we can suppose that the thermal effects let the period of the revival of the Rabi oscillations
become longer.
Figure~\ref{Figure04} shows
the atomic population inversion $\langle\sigma_{z}(t)\rangle$
given by Eq.~(\ref{sigma_z_t_upto-3rd-order-perturbation})
as a function of the time $t$
with
$\alpha=8$, $c=1$, $\kappa=1$, $\theta(\beta)=\pi/60$, $\omega_{0}=2$ and $\omega=4$.
Then,
in a similar manner for obtaining Eqs.~(\ref{definitions-theta-Theta-1})
and
(\ref{definitions-theta-Theta-2}),
we achieve
$\Theta(\beta)
=\arctan[\tanh^{1/2}(\theta(\beta))]
=\arctan[\tanh^{1/2}(\pi/60)]$.
Comparing the graphs shown in Figs.~\ref{Figure03} and \ref{Figure04},
we notice that the period of Fig.~\ref{Figure04} is longer than the period of Fig.~\ref{Figure03},
so that
we can suppose that the thermal effects let the period of the revival of the Rabi oscillations
become longer.
\begin{table}
\caption{The ranges of numerical values of the perturbation corrections with
$\alpha=4$, $c=1$, $\kappa=1$, $0\leq t\leq 20\pi$ and $\theta(\beta)=\pi/32$.
The estimations of the minimum and the maximum in every row of the table are based on values
of each correction term,
which we obtain numerically at equally spaced intervals $\Delta t=20\pi\times 10^{-4}$
during $0\leq t\leq 20\pi$.}
\label{Table01}
\begin{center}
\begin{tabular}{|c|c|c|}
\hline
correction term & min & max \\
\hline
$\theta(\beta)P_{g,1}^{(1)}(t)$ & $-0.227$\hphantom{0} & $0.199$\hphantom{0} \\
$\theta(\beta)P_{g,2}^{(1)}(t)$ & $-0.208$\hphantom{0} & $0.225$\hphantom{0} \\
$(1/2)\theta(\beta)^{2}P_{g,1}^{(2)}(t)$ & $-0.108$\hphantom{0} & $0.102$\hphantom{0} \\
$(1/2)\theta(\beta)^{2}P_{g,2}^{(2)}(t)$ & $-0.0996$ & $0.109$\hphantom{0} \\
$(1/6)\theta(\beta)^{3}P_{g,1}^{(3)}(t)$ & $-0.0441$ & $0.0483$ \\
$(1/6)\theta(\beta)^{3}P_{g,2}^{(3)}(t)$ & $-0.0478$ & $0.0467$ \\
\hline
\end{tabular}
\end{center}
\end{table}
\begin{table}
\caption{The ranges of numerical values of the perturbation corrections with
$\alpha=8$, $c=1$, $\kappa=1$, $0\leq t\leq 40\pi$ and $\theta(\beta)=\pi/60$.
The estimations of the minimum and the maximum in every row of the table are based on values
of each correction term,
which we obtain numerically at equally spaced intervals $\Delta t=40\pi\times 10^{-4}$
during $0\leq t\leq 40\pi$.}
\label{Table02}
\begin{center}
\begin{tabular}{|c|c|c|}
\hline
correction term & min & max \\
\hline
$\theta(\beta)P_{g,1}^{(1)}(t)$ & $-0.246$\hphantom{0} & $0.247$\hphantom{0} \\
$\theta(\beta)P_{g,2}^{(1)}(t)$ & $-0.248$\hphantom{0} & $0.245$\hphantom{0} \\
$(1/2)\theta(\beta)^{2}P_{g,1}^{(2)}(t)$ & $-0.125$\hphantom{0} & $0.128$\hphantom{0} \\
$(1/2)\theta(\beta)^{2}P_{g,2}^{(2)}(t)$ & $-0.127$\hphantom{0} & $0.126$\hphantom{0} \\
$(1/6)\theta(\beta)^{3}P_{g,1}^{(3)}(t)$ & $-0.0566$ & $0.0570$ \\
$(1/6)\theta(\beta)^{3}P_{g,2}^{(3)}(t)$ & $-0.0565$ & $0.0567$ \\
\hline
\end{tabular}
\end{center}
\end{table}
When we take $\alpha=4$, $c=1$, $\kappa=1$, $0\leq t\leq 20\pi$ and $\theta(\beta)=\pi/32$,
a numerical value of each order perturbation correction varies as shown in Table~\ref{Table01}.
On the other hand,
when we take $\alpha=8$, $c=1$, $\kappa=1$, $0\leq t\leq 40\pi$ and $\theta(\beta)=\pi/60$,
a numerical value of each order perturbation correction varies as shown in Table~\ref{Table02}.
Turning our eyes towards Table~\ref{Table01},
we observe that the contribution of the third order correction is nearly equal to
a half of the contribution of the second order correction
in the perturbative expansion.
From Table~\ref{Table01},
we consider the perturbative expansion to be reliable for $\theta(\beta)=\pi/32$.
Thus,
taking $\alpha=4$, $c=1$ and $\kappa=1$,
we can conclude that the third order perturbation theory is effective
for the parameter $0\leq \theta(\beta) \leq \pi/32$.
We notice that a similar thing happens in Table~\ref{Table02}, as well.
Thus,
taking $\alpha=8$, $c=1$ and $\kappa=1$,
we can conclude that the third order perturbation theory is effective
for the parameter $0\leq \theta(\beta) \leq \pi/60$.
\begin{figure}
\begin{center}
\includegraphics[scale=1.0]{Figure05.eps}
\vspace*{8pt}
\caption{The period of the revival of the Rabi oscillations $T(\theta)$
plotted
as a function of the parameter of the temperature $\theta(\beta)$.
The points are
obtained from numerical calculations of the third order perturbation theory
with taking
$\alpha=4$, $c=1$, $\kappa=1$,
$\omega_{0}=2$, $\omega=4$
and
$0\leq \theta(\beta)\leq\pi/32$.
In the graph,
the vertical axis is scaled logarithmically as $\ln[T(\theta)]$
and the horizontal axis is scaled linearly as $\theta(\beta)$.
Fitting the points with the linear function according to the least-squares method,
we obtain $\ln[T(\theta)]=3.25+(0.988)\theta$,
which is drawn in the graph.}
\label{Figure05}
\end{center}
\end{figure}
\begin{figure}
\begin{center}
\includegraphics[scale=1.0]{Figure06.eps}
\vspace*{8pt}
\caption{The period of the revival of the Rabi oscillations $T(\theta)$
plotted
as a function of the parameter of the temperature $\theta(\beta)$.
The points are
obtained from numerical calculations of the third order perturbation theory
with taking
$\alpha=8$, $c=1$, $\kappa=1$,
$\omega_{0}=2$, $\omega=4$
and
$0\leq \theta(\beta)\leq\pi/60$.
In the graph,
the vertical axis is scaled logarithmically as $\ln[T(\theta)]$
and the horizontal axis is scaled linearly as $\theta(\beta)$.
Fitting the points with the linear function according to the least-squares method,
we obtain $\ln[T(\theta)]=3.92+(1.07)\theta$,
which is drawn in the graph.}
\label{Figure06}
\end{center}
\end{figure}
In Figs.~\ref{Figure05} and \ref{Figure06},
we plot
the period of the revival of the Rabi oscillations $T(\theta)$
as a function of the parameter of the temperature $\theta(\beta)$.
The points in Fig.~\ref{Figure05} are
obtained from numerical calculations of the third order perturbation theory
with taking
$\alpha=4$, $c=1$, $\kappa=1$,
$\omega_{0}=2$, $\omega=4$
and
$0\leq \theta(\beta)\leq\pi/32$.
The points in Fig.~\ref{Figure06} are obtained similarly with taking
$\alpha=8$, $c=1$, $\kappa=1$,
$\omega_{0}=2$, $\omega=4$
and
$0\leq \theta(\beta)\leq\pi/60$.
In Fig.~\ref{Figure05},
we compute the period $T(\theta)$ numerically as follows:
First,
we calculate $\langle\sigma_{z}(t)\rangle$
given by Eq.~(\ref{sigma_z_t_upto-3rd-order-perturbation})
for a certain $\theta(\beta)$.
For every point of Fig.~\ref{Figure05}, taking the interval of the time $\Delta t=5\pi\times 10^{-4}$,
we obtain $\langle\sigma_{z}(t)\rangle$ at each time step
during $15\pi/2\leq t\leq 10\pi$.
We write the time at which $\langle\sigma_{z}(t)\rangle$ takes the maximum value
as $t_{\mbox{\scriptsize max}}$
and
write the time at which $\langle\sigma_{z}(t)\rangle$ takes the minimum value
as $t_{\mbox{\scriptsize min}}$.
Second, we obtain the period $T(\theta)$
with taking $T=(t_{\mbox{\scriptsize max}}+t_{\mbox{\scriptsize min}})/2$.
[For example,
taking $\theta=\pi/32$,
we obtain
$t_{\mbox{\scriptsize max}}\approx 28.52$
for $\langle\sigma_{z}(t_{\mbox{\scriptsize max}})\rangle\approx 0.3923$
and
$t_{\mbox{\scriptsize min}}\approx 28.86$
for $\langle\sigma_{z}(t_{\mbox{\scriptsize min}})\rangle\approx -0.4763$.
Thus,
we obtain $T(\pi/32)\approx 28.69$.]
The points of $T(\theta)$ in Fig.~\ref{Figure06} are
obtained in a similar manner with taking the interval of the time
$\Delta t=10\pi\times 10^{-4}$
and carrying out calculations of $\langle\sigma_{z}(t)\rangle$
at each time step during $15\pi\leq t\leq 20\pi$.
In the graphs of Figs.~\ref{Figure05} and \ref{Figure06},
the vertical axes are scaled logarithmically as $\ln[T(\theta)]$
and the horizontal axes are scaled linearly as $\theta(\beta)$.
In the graph of Fig.~\ref{Figure05},
the points form groups consisting of twos, threes and fours,
so that they appear in the shape of the stairs
as $\ln[T(\theta)]$ increases gradually.
The reason why the points appear in the shape of the stairs is as follows:
The atomic population inversion $\langle\sigma_{z}(t)\rangle$
is a bunch of the Rabi oscillations whose period is $\pi/(|\alpha||\kappa|)\simeq\pi/4$ around.
(We obtain this approximation in Sec.~\ref{section-review-JCM}.)
At the same time,
it shows the revival of the amplitude envelope with the period
$T(\theta)\simeq 2\pi|\alpha|e^{\theta}/|\kappa|=8\pi e^{\theta}$ around.
Thus,
calculating $t_{\mbox{\scriptsize max}}$ and $t_{\mbox{\scriptsize min}}$
numerically,
the rapid Rabi oscillations give us the smallest interval measurable as about
$\pi/8$, which is the half of the period of the Rabi oscillations.
This resolution of the time lets the points in Fig.~\ref{Figure05}
form the shape of the stairs.
Contrastingly,
the points in Fig.~\ref{Figure06} do not appear in the distinct shape of the stairs.
This is because the resolution of Fig.~\ref{Figure06} is finer than that of Fig.~\ref{Figure05}.
Indeed, in Fig.~\ref{Figure06},
the period of the Rabi oscillations is given by $\pi/8$ around,
so that the resolution of $T(\theta)$ is nearly equal to $\pi/16$.
Fitting the points in Fig.~\ref{Figure05}
with the linear function according to the least-squares method,
we obtain
\begin{equation}
\ln{[}T(\theta){]}
=
3.25+(0.988)\theta.
\end{equation}
We can interpret the above result as
\begin{equation}
T(\theta)
\simeq
e^{3.25}\times e^{(0.988)\theta}
\simeq
8\pi e^{\theta},
\end{equation}
which reminds us of Eq.~(\ref{thermal-effect-period-rivival-Rabi-oscillations}).
On the other hand,
fitting the points in Fig.~\ref{Figure06}
with the linear function according to the least-squares method,
we obtain
\begin{equation}
\ln{[}T(\theta){]}
=
3.92+(1.07)\theta.
\end{equation}
We can interpret the above result as
\begin{equation}
T(\theta)
\simeq
e^{3.92}\times e^{(1.07)\theta}
\simeq
16\pi e^{\theta},
\end{equation}
which also reminds us of Eq.~(\ref{thermal-effect-period-rivival-Rabi-oscillations}).
\section{\label{section-counter-rotating-terms}Thermal effects of the counter-rotating terms}
In this section,
we address thermal effects of the counter-rotating terms.
Because this topic is difficult and includes subtle problems,
we treat it with an intuitive manner.
First of all,
we have to go back to the derivation of the JCM.
At the beginning,
we consider the Hamiltonian for a magnetic dipole in a magnetic field,
and we obtain
\begin{equation}
H
=
\frac{\hbar}{2}\omega_{0}\sigma_{z}
+
\hbar\omega a^{\dagger}a
+
\hbar\kappa(\sigma_{-}+\sigma_{+})
(a+a^{\dagger}).
\label{dipole-field-Hamiltonian-0}
\end{equation}
Assuming near resonance $\omega\simeq\omega_{0}$,
the interaction terms $\sigma_{+}a$ and $\sigma_{-}a^{\dagger}$ are
practically independent of the time $t$,
while the terms $\sigma_{-}a$ and $\sigma_{+}a^{\dagger}$ vary
rapidly at frequencies $\pm(\omega_{0}+\omega)$.
Then,
applying the rotating wave approximation to Eq.~(\ref{dipole-field-Hamiltonian-0})
and removing the term $\hbar\kappa(\sigma_{+}a^{\dagger}+\sigma_{-}a)$,
we obtain the Hamiltonian of the JCM written down as Eq.~(\ref{JCM-Hamiltonian-0}).
As mentioned above,
the rotating wave approximation is used often in the field of the quantum optics.
However,
it is shown that the rotating wave approximation cannot always be a good treatment,
and sometimes it causes serious defects.
Ford {\it et al}. examine the Hamiltonian for an oscillator of the frequency $\omega_{0}$
interacting with a reservoir and its rotating wave approximation
\cite{Ford1988,Ford1997}.
The Hamiltonian of the original model is given by
\begin{equation}
H
=
\hbar\omega_{0}a^{\dagger}a
+
\sum_{j}\hbar\omega_{j} b_{j}^{\dagger}b_{j}
+
(a+a^{\dagger})\sum_{j}\lambda_{j}(b_{j}+b_{j}^{\dagger}),
\label{oscillator-reservoir-Hamiltonian-0}
\end{equation}
where $[a,a^{\dagger}]=1$, $[b_{j},b_{j}^{\dagger}]=1$ $\forall j$,
and its rotating wave approximation is given by
\begin{equation}
H_{\mbox{\scriptsize RWA}}
=
\hbar\omega_{0}a^{\dagger}a
+
\sum_{j}\hbar\omega_{j} b_{j}^{\dagger}b_{j}
+
\sum_{j}\lambda_{j}(a^{\dagger}b_{j}+ab_{j}^{\dagger}).
\label{oscillator-reservoir-Hamiltonian-RWA-1}
\end{equation}
Then,
the Hamiltonian $H_{\mbox{\scriptsize RWA}}$ defined
in Eq.~(\ref{oscillator-reservoir-Hamiltonian-RWA-1})
causes the following problem:
The expectation value (the energy) of $H_{\mbox{\scriptsize RWA}}$ has no lower bound,
so that we cannot specify the ground state.
Thus, we have to think the system described with $H_{\mbox{\scriptsize RWA}}$ to be unphysical.
As explained above, the rotating wave approximation sometimes manifests anomalous aspects.
Someone might complain that the rotating wave approximation brings us the JCM
that is an exactly soluble quantum mechanical model
for arbitrary $\Delta \omega$ and $\kappa$.
However, the JCM also has a defect,
which we cannot neglect.
Here, we think around the eigenvalues and the eigenvectors of the Hamiltonian of the JCM
given by Eq.~(\ref{JCM-Hamiltonian-0}).
They are written down as follows
\cite{Louisell1973}:
\begin{eqnarray}
E_{n,1}
&=&
\hbar[\omega(n+\frac{1}{2})+\lambda_{n}], \nonumber \\
E_{n,2}
&=&
\hbar[\omega(n+\frac{1}{2})-\lambda_{n}]
\quad\quad
\mbox{for $n=0,1,2,...$}, \nonumber \\
E_{0,0}
&=&
-\frac{\hbar}{2}\omega_{0},
\label{eigenvalues-JCM}
\end{eqnarray}
\begin{eqnarray}
|\varphi(n,1)\rangle
&=&
\cos\theta_{n}|n+1\rangle_{\mbox{\scriptsize P}}|g\rangle_{\mbox{\scriptsize A}}
+
\sin\theta_{n}|n\rangle_{\mbox{\scriptsize P}}|e\rangle_{\mbox{\scriptsize A}}, \nonumber \\
|\varphi(n,2)\rangle
&=&
-\sin\theta_{n}|n+1\rangle_{\mbox{\scriptsize P}}|g\rangle_{\mbox{\scriptsize A}}
+
\cos\theta_{n}|n\rangle_{\mbox{\scriptsize P}}|e\rangle_{\mbox{\scriptsize A}}
\quad\quad
\mbox{for $n=0,1,2,...$}, \nonumber \\
|0,0\rangle
&=&
|0\rangle_{\mbox{\scriptsize P}}|g\rangle_{\mbox{\scriptsize A}},
\label{eigenvectors-JCM}
\end{eqnarray}
where
\begin{equation}
\lambda_{n}=\sqrt{(\frac{\Delta\omega}{2})^{2}+\kappa^{2}(n+1)},
\end{equation}
$\Delta\omega=\omega-\omega_{0}$,
and
\begin{equation}
\tan\theta_{n}=\frac{\kappa\sqrt{n+1}}{(\Delta\omega/2)+\lambda_{n}}.
\end{equation}
\begin{figure}
\begin{center}
\mbox{\scalebox{1.0}[1.0]{\includegraphics{Figure07.eps}}}
\vspace*{8pt}
\caption{$E_{0,0}$ and $\{E_{n,2}:n=0,1,2,...,24\}$
as functions of $\kappa$,
where $0\leq \kappa\leq 10$.
Plotting them as graphs,
we assume $\omega_{0}=\omega=1$ and $\hbar=1$.
Because $\{E_{n,1}\}$ never can be the ground-state energy,
we do not plot them in this figure.
Looking these graphs,
we notice the following facts.
When $\kappa=0$, the ground-state energy is equal to $E_{0,0}$.
On the other hand, when $\kappa=10$, the ground-state energy is given by $E_{24,2}$.
In fact, these graphs show that the ground-state energy changes from $E_{0,0}$
to $E_{n,2}$ for $n\gg 1$ gradually as $\kappa$ becomes larger.}
\label{Figure07}
\end{center}
\end{figure}
\begin{figure}
\begin{center}
\mbox{\scalebox{1.0}[1.0]{\includegraphics{Figure08.eps}}}
\vspace*{8pt}
\caption{An excitation energy $\Delta E$,
which is required to promote the JCM system from the ground state to the first excited state,
as an function of $\kappa$,
where $0\leq\kappa\leq 10$.
Looking at this graph,
we notice that $\Delta E$ oscillates and its amplitude decreases rapidly.
When $\Delta E=0$, the JCM system has the degenerate ground states.}
\label{Figure08}
\end{center}
\vspace*{8pt}
\begin{center}
\mbox{\scalebox{1.0}[1.0]{\includegraphics{Figure09.eps}}}
\vspace*{8pt}
\caption{An excitation energy $\Delta E$,
which is required to promote the JCM system from the ground state to the first excited state,
as an function of $\kappa$,
where $0\leq\kappa\leq 10$.
The graph uses the logarithmic scale on the vertical axis
and the linear scale on the horizontal axis.
In Fig.~\ref{Figure08}, we show that the system has the degenerate ground states
at certain values of the parameter $\kappa$.
Because $\ln\Delta E\to -\infty$ as $\Delta E\to 0$,
we cannot plot small $\Delta E$, which is nearly equal to zero, in the graph.
Looking at this graph,
we notice that the amplitude of $\Delta E$,
which oscillates in the parameter $\kappa$,
decreases exponentially.}
\label{Figure09}
\end{center}
\end{figure}
Looking at Eq.~(\ref{eigenvalues-JCM}),
we notice that the ground state changes from $|0,0\rangle$
to $|\varphi(n,2)\rangle$ for $n\gg 1$ gradually
as $|\kappa|$ becomes larger.
To confirm it numerically,
we plot $E_{0,0}$ and $\{E_{n,2}:n=0,1,2,...,24\}$
as functions of $\kappa$ in Fig.~\ref{Figure07}.
At the same time,
an excitation energy, which is required to promote the JCM system from the ground state
to the first excited state,
becomes smaller rapidly as $|\kappa|\to\infty$.
To confirm it numerically,
we plot the excitation energy as a function of $\kappa$ in Figs.~\ref{Figure08} and \ref{Figure09}.
From the analyses performed in Figs.~\ref{Figure07}, \ref{Figure08} and \ref{Figure09},
we can conclude as follows:
If we take a large value of $|\kappa|$, the ground state of the JCM contains many photons.
Then,
the excitation energy takes a small value.
These properties of the ground state of the JCM relate to the uncertainty principle
$\Delta N \Delta \phi\geq (1/2)$.
Because $\Delta E$ decreases exponentially
as $|\kappa|$ becomes larger as shown in Figs.~\ref{Figure08} and \ref{Figure09},
the system of the JCM is able to jump from the ground state to excited states
at ease for $|\kappa|\gg 1$.
Thus, the fluctuation of the number of photons $\Delta N$ becomes very larger.
Hence, according to the uncertainty principle
$\Delta N \Delta \phi\geq (1/2)$,
the system of the JCM around the ground state acquires
very small fluctuation of the phase of each photon,
so that $\Delta \phi\to 0$.
However, the ground state that contains a large number of photons with small fluctuation of the phase
seems not to be practical.
We may realize the ground state for $\Delta N\gg 1$ and $\Delta\phi\simeq 0$
in the laboratory
by using a two-level atom in the cavity field,
which is induced by a very strong laser beam.
From the viewpoint explained above,
we cannot regard the JCM derived with the rotating wave approximation as a proper model
in the field of the quantum optics.
Hence, the JCM is valid and has physical meanings
if and only if a near resonance $\omega\simeq\omega_{0}$ is assumed and $|\kappa|$ is small enough.
To overcome the defects of the rotating wave approximation in the JCM,
some researchers try to extend and generalize the JCM.
Ng {\it et al}. investigate the two-photon JCM
and the intensity-dependent JCM with the counter-rotating terms
\cite{Ng1999,Ng2000}.
In these models,
the nonlinearity of the interaction between the two-level atom and the cavity field is emphasized.
In general,
it is very difficult and complicated to evaluate the contributions of the counter-rotating terms
in the JCM.
Feranchuk {\it et al}. study the Schr{\"o}dinger equation,
whose Hamiltonian is given by Eq.~(\ref{dipole-field-Hamiltonian-0}),
numerically
\cite{Feranchuk1996}.
Especially,
Phoenix presents several perturbative approaches to investigate this problem.
Here, we review one of his perturbation methods,
which is called short time expansion of the inversion.
First,
we begin with the Heisenberg picture of $\sigma_{z}$,
\begin{equation}
\sigma_{z}(t)
=
\exp(\frac{i}{\hbar}Ht)\sigma_{z}\exp(-\frac{i}{\hbar}Ht),
\label{Heisenberg-picture-sigma-z-0}
\end{equation}
where the Hamiltonian is given by Eq.~(\ref{dipole-field-Hamiltonian-0}).
Moreover,
we assume $\omega=\omega_{0}$,
so that we consider the optical resonance.
Second,
we expand Eq.~(\ref{Heisenberg-picture-sigma-z-0}) in a power series in $t$ and neglect third-order terms
\cite{Phoenix1989}.
So that,
we obtain
\begin{eqnarray}
\sigma_{z}(t)
&=&
\sigma_{z}(0)
+
\frac{it}{\hbar}[H,\sigma_{z}(0)]
-
\frac{t^{2}}{2\hbar^{2}}(H^{2}\sigma_{z}(0)+\sigma_{z}(0)H^{2})
+
\frac{t^{2}}{\hbar^{2}}H\sigma_{z}(0)H \nonumber \\
&&
+
{\cal O}(t^{3}) \nonumber \\
&=&
1-2(\kappa t)^{2}(a+a^{\dagger})^{2}
+
{\cal O}(t^{3}).
\label{short-time-expansion-inversion-0}
\end{eqnarray}
Third,
we assume the initial state as $|\alpha\rangle_{\mbox{\scriptsize P}}|e\rangle_{\mbox{\scriptsize A}}$,
where $\alpha=\sqrt{\bar{n}}e^{i\phi}$,
and substitute it into Eq.~(\ref{short-time-expansion-inversion-0}).
Finally,
we obtain
\begin{eqnarray}
\langle \sigma_{z}(t)\rangle_{\alpha}
&=&
{}_{\mbox{\scriptsize A}}\langle e|_{\mbox{\scriptsize P}}\langle\alpha|
\sigma_{z}(t)
|\alpha\rangle_{\mbox{\scriptsize P}}|e\rangle_{\mbox{\scriptsize A}} \nonumber \\
&=&
1-2(\kappa t)^{2}(4\bar{n}\cos^{2}\phi+1)+{\cal O}(t^{3}).
\label{short-time-expansion-inversion-coherent-state-1}
\end{eqnarray}
Because Eq.~(\ref{short-time-expansion-inversion-coherent-state-1}) is valid
for $0\leq t\ll 1$,
we can expect it to describe the initial collapse of the Rabi oscillations.
We note that Eq.~(\ref{short-time-expansion-inversion-coherent-state-1})
depends on the phase $\phi$.
This characteristic can always be found
in any perturbative expansion of $\langle\sigma_{z}(t)\rangle_{\alpha}$.
(This fact is indicated by Phoenix first.)
In Sec.~\ref{section-thermal-effects-period-qualitative-estimation},
we give the intuitive discussions about the thermal effects of the JCM,
and we obtain Eq.~(\ref{thermal-effect-of-alpha-coherent-state-0})
under the low-temperature limit.
Hence, we can add the thermal effects to Eq.~(\ref{short-time-expansion-inversion-coherent-state-1})
as
\begin{equation}
\langle\sigma_{z}(t)\rangle_{\alpha}
=
1-2(\kappa t)^{2}(4\bar{n}e^{2\theta}\cos^{2}\phi+1)+{\cal O}(t^{3}),
\label{short-time-expansion-inversion-coherent-state-2}
\end{equation}
where $\theta$ is given by Eq.~(\ref{definition-theta-beta-1}).
To examine whether or not Eq.~(\ref{short-time-expansion-inversion-coherent-state-2}) holds
remains to be solved in the future.
In this section,
we argue only the short time expansion of $\langle\sigma_{z}(t)\rangle_{\alpha}$
with the counter-rotating terms.
We point out that to examine long time behaviour of the JCM with counter-rotating terms
is very difficult even if we use perturbative techniques.
\section{\label{section-discussion}Discussion}
Turning our eyes towards the graphs shown
in Figs.~\ref{Figure05} and \ref{Figure06},
we observe that the period of the revival of the Rabi oscillations becomes longer
as the temperature rises.
This phenomenon is predicted form an intuitive discussion
in Sec.~\ref{section-thermal-effects-period-qualitative-estimation}.
In Sec.~\ref{section-numerical-calculations},
we confirm this phenomenon (or this expectation)
with numerical calculations based on the third order low-temperature expansion.
Why is the low-temperature expansion in $\theta(\beta)$ given
by Sec.~\ref{section-formulation-perturbetion-theory}
effective for a perturbation theory?
The reason why is as follows:
The thermal coherent state is defined
in Eqs.~(\ref{definition-thermal-coherent-state-2})
and
(\ref{another-form-of-thermal-coherent-state-1}).
This definition is suitable for the low-temperature expansion
because of the Baker-Hausdorff theorem
\cite{Louisell1973}.
|
\section{introduction}
\medskip
In \cite{c55}, Dolfi, Jabara and Lucido determine the finite groups in which the centraliser of every element
of order $5$ is a $5$-group. However, the main result in \cite {c55} is flawed: when looking at
groups $G $ with $F^*(G/F(G)) \cong \Alt(5)$ and $F(G)$ of odd order the authors of \cite{c55} erroneously prove that $F(G)$ must be abelian. We
believe that the error occurs on page 1060 of their paper. Our main result is a corrected version of their theorem:
\begin{theorem}[Dolfi, Jabara and Lucido]\label{MainThm} Suppose that $G$ is a finite group in which the centraliser of every element of order $5$ is a $5$-group. Then one of the following holds:
\begin{enumerate}
\item $G$ is a $5$-group.
\item $G$ is a soluble group and one of the following hold:
\begin{enumerate}
\item $ G$ is a soluble Frobenius group such that either the Frobenius kernel or a Frobenius complement
is a 5-group;
\item $G$ is a 2-Frobenius group such that $F(G)$ is a $5'$-group and $G/ F(G)$ is a Frobenius group,
whose kernel is a cyclic $5$-group and whose complement is cyclic of order 2 or 4; or
\item $G$ is a 2-Frobenius group such that $F(G)$ is a 5-group and $G/ F(G)$ is a Frobenius group,
whose kernel is a cyclic $5'$-group and whose complement is a cyclic 5-group.
\end{enumerate}
\item $G/F(G) \cong \Alt(5)$ or $\Sym(5)$ and $F(G)= O_2(G)O_{2'}(G)$ where $O_2(G)$ is nilpotent of
class at most three and $O_{2'}(G)=O_{\{2,5\}'}(G)$ is nilpotent of class at most two.
\item $G/F(G) \cong \Alt(6)$, $\Sym(6)$ or $\M(9)$ and $F(G)= O_2(G)O_3(G)$ where $O_2(G)$ and $O_3(G)$
are elementary abelian.
\item $G/F(G) \cong \Alt(7)$ and $F(G)=O_2(G)$ is elementary abelian.
\item $G/F(G) \cong {}^2\mathrm B_2(8)$ or ${}^2\B_2(32)$ and $F(G)=O_2(G)$ is elementary abelian.
\item $G/F(G) \cong \PSU_4(2)$ or $\Aut(\PSU_4(2))$ and $F(G)=O_2(G)$ is elementary abelian.
\item $G/F(G) \cong \PSL_2(49)$, $\PGL_2(49) $ or $\M(49)$ and $F(G)=O_7(G)$ is elementary abelian.
\item $E(G) \cong \PSL_2(5^e)$ with $e \ge 2$ and either $G \cong \PSL_2(5^e)$ or $G \cong \PGL_2(5^e)$ or
$e$ is even and $G\cong \M(5^e)$.
\item $E(G) \cong \PSL_2(p)$ where $p$ is a prime which can be written as $p = 2\cdot5^e \pm 1$ for some non-negative integer $e$.
\item $G \cong \PSL_2(11)$, $\PSL_3(4)$, $\PSL_3(4){:}2_f$, $\PSL_3(4){:}2_i$, $\PSp_4(7)$, $\PSU_4(3)$, $\Mat(11)$ or $\Mat(22)$.
\end{enumerate}
\end{theorem}
We adopt
notation from \cite{c55}, so, for all odd primes $p$ and all $m \in \N$, we write $\M(p^{2m})$ to denote
the non-split extension of $\PSL_2(p^{2m})$ such that $|\M(p^{2m}):\PSL_2(p^{2m})|=2$.
In part (xi) of Theorem~\ref{MainThm}, we denote by $2_f$ a field automorphism of order $2$ and by $2_i$ the inverse transpose automorphism. Throughout the statement, elementary abelian groups may be trivial. We finally point out that the $G/F(G)$-modules involved in $F(G)$ are explicitly known in all cases.
The paper is organised as follows. In Section 2, we construct examples of $r$-groups of class two for all
odd primes $r$, $r \neq 5$, which admit an action of $\Alt(5)$ with an element of order $5$ acting without non-trivial fixed
points. This demonstrates that the statement in \cite{c55} is false. However, in Section 3, we show
that there are no groups $G$ with $O_{\{2,5\}'}(G)$ nilpotent of class three with
$G/O_{\{2,5\}'}(G) \cong \Alt(5)$ and with
an element of order $5$ acting fixed point
freely. This is the main contribution of this paper and our proof of this fact is guided by some of the arguments in \cite{Holt}.
We further note that the authors of \cite{c55} use the fact that if $\Alt(5)$ acts on a $7$-group, then the
$7$-group is abelian to show that $\PSL_2(49)$ cannot act on a non-abelian $7$-group. In Section 4, we provide an alternative
proof for this fact and prove, in addition, that such an abelian group must be elementary abelian. We also show that, in part (iv) of Theorem~\ref{MainThm}, the subgroup $O_3(G)$ is elementary abelian. This sharpens the result
stated in \cite{c55}.
\bigskip
\noindent {\bf Acknowledgement.} The second author is grateful to the DFG for their support and both the first and second author thank the mathematics department in Halle for their hospitality.
\bigskip
\section{Construction of a class two group admitting $\Alt(5)$ with an element of order 5 acting fixed point freely}
\medskip
In this section we demonstrate that Theorem \ref{MainThm}~(iii) cannot be strengthened to say that $O_{2'}(G)$ is abelian by constructing examples with $O_{2'}(G)$ of class two.
Suppose that $r$ is an odd prime with $r \neq 5$ and that $W$ is a
$5$-dimensional vector space over $\GF(r)$ with basis $\{a_1,\dots, a_5\}$. Let $X:=\Sym(5)$ naturally permute this basis. This permutation action gives rise to a faithful linear action of $X$ on $W$ which is in fact a reduction modulo $r$ of the corresponding integral representation of $X$. We
let $V$ be the $\GF(r) X$-submodule of $W$ with basis $\mathfrak V=
\{v_1, \dots, v_4\}$ where, for all $1\le i \le 4$, we set $v_i:= a_1-a_{i+1}$. It is easy to check that an element of order $5$ in $X$ acts fixed point freely on $V$ and that $W=V \oplus \langle a_1+a_2+a_3+a_4+a_5\rangle$.
Finally we let $Y:= X' \cong \Alt(5)$ and note that the restriction of $V$ to $Y$ is also an irreducible $\GF(r)Y$-module. We also denote this module by $V$.
\medskip
Recall that $V \wedge V$ is isomorphic to the submodule of $V \otimes V$ generated by the vectors $ v_i \otimes
v_j - v_j \otimes v_i$ for $1\le i < j \le 4$.
\begin{lemma} \label{modfacts} The $\GF(r)Y$-module $V \wedge V$ has no composition factor isomorphic to $V$.
\end{lemma}
\begin{proof}
Let $U:=V \wedge V$. Calculating most easily with $S = \langle (2,3)(4,5),(2,4)(3,5)\rangle$, we obtain that $C_U(S)=0$. Since $C_V(S) = \langle v_1+v_2+v_3+v_4\rangle$, the result follows.
\end{proof}
We need the following statement about self-extensions of $V$ by $V$.
\begin{lemma}\label{coho} Any $\GF(r)Y$-module with all composition
factors isomorphic to $V$ is completely reducible.
\end{lemma}
\begin{proof} We need to show that $\mathrm {Ext}^1_Y(V,V)= 0$.
(For notation see for example \cite{benson}.)
If $r > 3$, then, as $r \neq 5$, the result follows
from Maschke's Theorem. So suppose that $r=3$. Let $H$ denote a subgroup of $Y$ which is isomorphic to $\Alt(4)$. We recall that $V$ is a direct summand of the natural
permutation module $W$ for $Y$. Hence if we let $1_H$ denote the trivial $\GF(3)H$-module, then $W$ is the induced module $1_H\uparrow ^Y$.
It follows that
$$\mathrm {Ext}_Y^1(W,V)= \mathrm {Ext}_Y^1(1_H\uparrow ^Y,V)= \mathrm
{Ext}_H^1(1_H,V\downarrow_ H),$$
where the second equality comes from Shapiro's Lemma \cite[Corollary 3.3.2]{benson}.
Now $V\downarrow_ H$ is a direct sum of a faithful $3$-dimensional module and a trivial module. As $H$ contains a Klein fours subgroup and as this subgroup acts coprimely on the
$3$-dimensional module, this module only has trivial extensions with the trivial $H$-module.
Since $\dim \mathrm {Ext}^1_Y(1_H,1_H)= 1$ (see for example \cite[Corollary 3.5.2]{benson}), we have that $\dim \mathrm {Ext}^1_Y(1_H,V\downarrow H)= 1$.
On the other hand
$$\mathrm {Ext}^1_Y(W,V)= \mathrm
{Ext}^1_Y(1_Y\oplus V,V)= \mathrm {Ext}^1_Y(1_Y,V)\oplus \mathrm {Ext}^1_Y(V,V).$$ Now let $Y \ge D \cong \mathrm
{Dih}(10)$. Then $1_D\uparrow^Y$ is a uniserial module with socle and head of dimension $1$ and heart of dimension
$4$. Hence $\dim \mathrm {Ext}^1_Y(1_Y,V) \ge 1$. We infer that $\mathrm {Ext}^1_Y(V,V)=0$ as claimed.
\end{proof}
Constructions of $r$-groups admitting the action of a further group are intimately related to tensor products and homomorphisms between modules. We therefore study $\Hom(V,V)$ in the next lemma.
Let $\sigma = v_1+v_2+v_3+v_4$ and define $\theta \in \Hom(V,V)$ as follows: for all $1\le i \le 4$,
$$v_i \mapsto
v_i\theta :=\sigma - v_i.$$
\medskip
With respect to the basis $\mathfrak V$,
we calculate that $\theta$ has matrix $$\left(\begin{array}{cccc}
0&1&1&1\\1&0&1&1\\1&1&0&1\\1&1&1&0\end{array}\right).$$
\medskip
\begin{lemma}\label{homsubmod} The $\GF(r)X$-submodule of $\Hom(V,V)$
generated by $\theta$ has dimension $4$ and is isomorphic to $V$.
\end{lemma}
\begin{proof} Let $X_1 = \mathrm {Stab}_X(1) \cong \Sym(4)$. Then, for all $1\le i \le
4$ and for all $\pi \in X_1$, we have $$v_i\pi^{-1} \theta\pi=
(v_{(i+1)\pi^{-1}-1})\theta\pi = (\sigma - v_{((i+1)\pi^{-1}-1)\pi}) =
\sigma - v_i= v_i\theta.$$ Since $v_1(1,2)\theta
(1,2)\neq v_1\theta$, the orbit of $X$ containing $\theta$ has
exactly $5$ elements. Therefore the subspace $U$ of $\Hom(V,V)$
spanned by $\theta$ has dimension $4$ or $5$. If $\dim U = 5$, then
the sum $\tau$ of the translates of $\theta$ under $X$ is
centralised by $X$ and is non-trivial. By Schur's Lemma we then have
that $\tau$ is a scalar matrix. Since $r$ is odd, if $\tau$ is non-zero, it has
non-zero trace. However, $\theta$ and its translates have trace $0$
and therefore $\tau$ must also have trace $0$. Thus $\tau=0$ and
$\dim U=4$.
\end{proof}
\begin{theorem}\label{cls2}
For all odd primes $r$, $r \neq 5$, there are $r$-groups of class two
which admit $\Sym(5)$ with an element of order $5$ acting without
fixed points.
\end{theorem}
\begin{proof} Let $U$ be an $8$-dimensional vector space over
$\GF(r)$. We define a subgroup $J$ of $\GL_8(r)$ by
$$J=\langle \left(
\begin{array}{cc}I_4&0\\\theta&I_4\end{array}\right),\left(
\begin{array}{cc}x_\pi&0\\0&x_\pi\end{array}\right)\mid \pi \in
X\rangle$$ where $x_\pi$ is the matrix corresponding to the action of $\pi\in X$ on $V$ with respect to the basis
$\mathfrak V$ and $I_4$ is the $4\times 4$ identity matrix. Then $J \cong r^4{:}\Sym(5)$ by Lemma~\ref{homsubmod}. Set $K= U \rtimes J$. Then $O_r(K)$ has
order $r^{12}$ and class two. Furthermore, the elements of order $5$ in $K$ are self-centralising.
\end{proof}
Table~\ref{tab1} describes an element $\gamma\in \Hom(V\otimes V,V)$ by defining the images of the basic tensors.
\begin{table}[h]
$$\begin{array}{c|cccc}
\otimes&v_1&v_2&v_3&v_4\\
\hline
v_1&-5v_1+2\sigma&\sigma&\sigma&\sigma\\
v_2&\sigma&-5v_2+2\sigma&\sigma&\sigma\\
v_3&\sigma&\sigma&-5v_3+2\sigma&\sigma\\
v_4&\sigma&\sigma&\sigma&-5v_4+2\sigma\\
\end{array}.
$$
\caption{The module homomorphism $\gamma$}\label{tab1}
\end{table}
The next lemma will be used in Section 3.
\begin{lemma}\label{unique} Up to scalar multiplication $\gamma$ is the unique element of $\Hom_{Y}(V\otimes V,V)= \Hom_X(V\otimes V,V)$.
\end{lemma}
\begin{proof} We note that $X=\langle (1,2), (2,3,4,5)\rangle$ and that the second generator permutes
$\{v_1, v_2,v_3,v_4\}$ as a $4$-cycle. In particular, $\gamma$ commutes with this element. Hence we only need to
verify that $\gamma$ commutes with the transposition. As an illustration we show that $(v_1\otimes v_2)\gamma
(1,2)= (v_1\otimes v_2)(1,2)\gamma$. The left hand side is seen to be $\sigma(1,2)= -5 v_1+\sigma$, while the
right hand side equals $-(v_1\otimes v_2)\gamma + (v_1\otimes v_1)\gamma= \sigma -5v_1$. Thus $\gamma \in \Hom_X(V\otimes V,V)$.
Assume that $\mu \in \Hom_{Y}(V\otimes V,V)$ is not a scalar multiple of $\gamma$. We consider $(v_1\otimes v_1)\mu = \sum_{i=1}^5\lambda_ia_i$ where
$\l_i \in \GF(r)$ and $\sum_{i=1}^5\lambda_i= 0$. Using the fact that $\mu$ commutes with elements from $Y$, we
see, by applying $(3,4,5)$, that $\lambda_3=\lambda_4=\lambda_5$ which we define to be $\lambda$. Next, applying
$(1,2)(3,4)$, we deduce that $\lambda_1=\lambda_2$. So $$(v_1\otimes v_1)\mu = \l_1(a_1+a_2)+\l(a_3+a_4+a_5)$$ and,
therefore, $2\lambda_1 + 3\l =0$. Since $\mu$ is not the zero map, after adjusting $\mu$ by a scalar, we may suppose that $$(v_1\otimes v_1)\mu = -5v_1+2\sigma= (v_1\otimes
v_1)\gamma.$$ Replacing $\mu $ with $\mu - \gamma$, we only need to consider the case where $\mu$ maps
$v_1\otimes v_1$ to zero. The action of $Y$ now gives $(v_j\otimes v_j)\mu =0$ for all $1\le j \le 4$.
Let
$(1,j,k)$ be a $3$-cycle in $Y$. Since $(v_j\otimes v_j)\mu =0$, we have that $$0= (v_j\otimes v_j)(1,j,k)\mu =
((v_k-v_j)\otimes (v_k-v_j))\mu = -(v_k\otimes v_j)\mu- (v_j\otimes v_k)\mu.$$ It follows that $\mu$ is
alternating and hence that $V \wedge V$ has a quotient isomorphic to $V$. But then Lemma~\ref{modfacts} implies that $\mu$ is the zero map and this is our contradiction. The lemma is now proved.
\end{proof}
\bigskip
\section{A non-existence Theorem about class three groups admitting $\Alt(5)$}
\medskip
We have seen in the previous section that $\Alt(5)$ can act on a class two group with an element of order 5 acting fixed point freely. Our objective in this section is to show that if $E(G/F(G)) \cong \Alt(5)$, then Theorem~\ref{MainThm} (iii) holds. In particular, we show that $\Alt(5)$ cannot act on a class three group of odd order with an element of order 5 acting fixed point freely.
Hence suppose that $G$ is a finite group such that the centraliser of every element of order $5$ is a $5$-group and $E(G/F(G)) \cong \Alt(5)$. Assume that $O_5(G)\neq 1$. Then, as $F(G)$ is nilpotent, the fact that elements of order $5$ are $5$-groups implies that $F(G)= O_5(G)$. Let $B$ be a Sylow $2$-subgroup of $G$. Then $B$ is elementary abelian of order $4$. By coprime action we have that $O_5(G) = \langle C_{O_5(G)}(b)\mid b \in B^\#\rangle$. So, as the elements of order $5$ in $G$ do not commute with involutions, we have a contradiction. Thus $O_5(G)=1$. Therefore we only have to restrict the
structure of $O_r(G)$ for all primes $r\neq 5$ dividing $|F(G)|$. If $r=2$, then \cite[Lemma~4.1]{prince} implies that $G$ contains a subgroup isomorphic to $\Alt(5)$ and so we have that $O_2(G)$ admits
$\Alt(5)$ with an element of order $5$ acting fixed point freely. In this case \cite[Theorem
2]{Holt} yields that $O_2(G)$ has class at most three. So we may assume that $r$ is odd with $r \neq 5$. Our aim
will be achieved once we prove the following theorem:
\begin{theorem}\label{cl3} Suppose that $r$ is an odd prime with $r\neq 5$ and that $G$ is a group with $R:=F(G)=O_r(G)$ an $r$-group and $F^*(G/R)\cong \Alt(5)$. If
an element of order $5$ in $G$ acts fixed point freely on $R$, then $R$ has class at most two.
\end{theorem}
The remainder of this section is devoted to proving Theorem~\ref{cl3}, and
it suffices to consider the case where $G/O_r(G) \cong \Alt(5)$ and $O_r(G)$ has class three.
The most direct approach to the proof of Theorem~\ref{cl3} would choose $G$ of minimal order with $R:=O_r(G)$ of class three and then derive a contradiction. This approach works very smoothly when $r \neq 3$ because in these cases $G$ splits over $R$. However, when $r=3$ it is possible that $G$ does not split over $R$ (as there exist groups $X$ with $X/O_3(X) \cong \Alt(5)$ and $O_3(X)$ isomorphic to $V$ as an $X/O_3(X)$-module which do not contain subgroups isomorphic to $\Alt(5)$). Thus in the straightforward approach to the proof of Theorem~\ref{cl3}, when we choose a normal subgroup $U$ of $G$ such that $U$ is proper in $R$, there may not exist a group $G^*$ with $O_3(G^*)= U$. Consequently our inductive hypothesis is not strong enough to assert that $U$ has class at most two.
Instead, we assume that Theorem~\ref{cl3} is false and we choose a counter-example $G$ of minimal order such that $R:=O_r(G)$ has class three. Then we choose a normal subgroup $Q$ of $G$ that is contained in $R$ and that is minimal with respect to having class three. We set $\ov{G}:=G/R$ and fix all this notation.
\begin{lemma}\label{QProps}
The following hold:
\begin{enumerate}
\item $Q'$ is abelian.
\item Every normal subgroup of $G$ that is properly contained in $Q$ has class at most two.
\item $\Gamma_2(Q)$ is elementary abelian of order $r^4$ and is isomorphic to $V$ (as defined before Lemma \ref{unique}) as a $\GF(r)\ov{G}$-module.
\item $Q'/(Q' \cap Z(Q))$ is a $\GF(r)\ov{G}$-module.
\item $\Phi(Q) \le Z_2(Q) \le C_Q(Q') <Q$. In particular, $Q/Z_2(Q)$ is elementary abelian.
\item Every $G$-composition factor of $Q$ is isomorphic as a
$\GF(r)\ov{G}$-module to $V$. Every elementary abelian $G$-invariant
section of $Q$ which is centralised by $R$ is a direct sum of
modules isomorphic to $V$.
\end{enumerate}
\end{lemma}
\begin{proof} We have $Q' \le Z_2(Q)$ and $[Q',Z_2(Q)]=1 $ follows from the three subgroup lemma.
Thus $Q'\le Z_2(Q) \le C_Q(Q') <Q$ and, especially, (i) and some of the inclusions in (v) hold.
For part (ii) let $Q_0$ be a normal subgroup of $G$ that is properly contained in $Q$.
Then the minimal choice of $Q$ immediately gives that $Q_0$ has class at most two.
$\Gamma_2(Q)$ lies in $Z(Q)$ because $Q$ has class three, so it is abelian. Let $P:=\Phi(\Gamma_2(Q))$. Then $R$ centralises $P$ and hence, if $P \neq 1$, then $\wh{G}:=G/P$ is a counter-example of smaller order because $\wh{Q}$ is still of class three.
This contradiction shows that $P=1$ and so $\Gamma_2(Q)$ is elementary abelian and it follows similarly that $\Gamma_2(Q)$ is irreducible as a $\GF(r)\ov{G}$-module.
This proves (iii).
Let $a \in Q'$ and $q\in Q$. Then $[a^r,q]= [a,q]^r$. But $[a,q]\in
\Gamma_2(Q)$ which is elementary abelian by (iii). So $[a^r,q]=1$.
Hence $a^r \in Z(Q)$ and therefore $Q'/Q' \cap Z(Q)$ is elementary abelian.
For this factor group to be a $\GF(r)\ov{G}$-module
it remains to show that $[Q',R] \le Q' \cap Z(Q)$.
Of course $[Q',R] \le Q'$, and also
$[Q,Q',R] \le [\Gamma_2(Q),R]<\Gamma_2(Q)$. Therefore
$[Q,Q',R]=1$ by (iii).
As $R$ has class three, we know that
$[R,R'] \le Z(R)$ and hence $[Q,R] \le R' \le Z_2(R)$.
Therefore $[R,Q,Q']=1$ and hence the three subgroup lemma implies that $[Q',R,Q]=1$.
We deduce that $[Q',R] \le Z(Q)$ whence $[Q',R] \le Q' \cap Z(Q)$.
For the last statement of (v) suppose that $a, q_1, q_2 \in Q$. Then
$[a^r,q_1,q_2] = [a,q_1,q_2]^r =1$ which means that $a^r \in
Z_2(Q)$. Hence $Q/Z_2(Q)$ is elementary abelian.
As stated in \cite{c55}, $V$ is the unique $\GF(r)\ov{G}$-module which admits an element of
order $5$ from $\ov{G}$ acting fixed point freely. Thus Lemma~\ref{coho}
gives (vi).
\end{proof}
\begin{lemma}\label{3Ms}
Suppose that $M_1$, $M_2$ and $M_3$ are subgroups of $Q$ that are maximal subject to being normal in $G$ and contained in $Q$. Set $D= M_1\cap M_2\cap M_3$ and suppose that $|Q:D|=r^{12}$. Then $D \le Z_2(Q)$ and in particular $Q/Z_2(Q)$ is a direct product of at most three minimal
normal subgroups of $G/Z_2(Q)$ each of order $r^4$.
\end{lemma}
\begin{proof}
By Lemma~\ref{QProps} (ii), the subgroups $M_1$, $M_2$ and $M_3$ have class at most two. Thus for $1 \le i<j \le 3$, we have that
$[M_i\cap M_j ,M_j]\le M_j'\le Z(M_j)$.
As $Q=M_iM_j$, we obtain that $$[M_i \cap M_j,Q] \le [M_i \cap M_j,M_i][M_i \cap M_j,M_j]\le
Z(M_i)Z(M_j).$$ In particular, $$[D,Q]\le Z(M_1)Z(M_2)\cap Z(M_2)Z(M_3)\cap Z(M_1)Z(M_3).$$
Since $|Q:D|=r^{12}$, we have $M_1 \cap M_2 \not \le M_3$. Therefore $Q= (M_1\cap M_2)M_3$ and
$$M_1= M_1 \cap (M_1\cap M_2)M_3= (M_1 \cap M_2)(M_1\cap M_3).$$
Similarly, $M_2= (M_1 \cap M_2)(M_2\cap M_3)$ and so $$Q= M_1M_2=(M_1 \cap M_2)(M_2\cap M_3)(M_1\cap M_3).$$
Hence it follows that $[D,Q,Q]$ is contained in
$$[Z(M_1)Z(M_2)\cap Z(M_2)Z(M_3)\cap Z(M_1)Z(M_3),(M_1 \cap M_2)(M_2\cap M_3)(M_1\cap
M_3)]=1$$ and consequently $D\le Z_2(Q)$.
We know from Lemma~\ref{QProps} (v) that $Q/Z_2(Q)$ is elementary abelian. As $R$ has class three and $Q \unlhd G$, we see that $[Q,R] \le R' \cap Q \le Z_2(R) \cap Q \le Z_2(Q)$, so
$Q/Z_2(Q)$ is centralised by $R$ and hence Lemma~\ref{QProps} (vi) implies that $Q/Z_2(Q)$ is a direct product of minimal normal subgroups of $G/Z_2(Q)$. If there are at least
three minimal normal subgroups of $G/Z_2(G)$ involved in this product, then there are exactly three by the first part of the lemma. This completes the proof.
\end{proof}
The remainder of the proof is organised as a series of claims.
\medskip
\begin{claim} \label{M1} Suppose that $M$ is a normal subgroup of $Q$ such that $M/Z_2(Q)$ is a minimal normal subgroup of $G/Z_2(Q)$. Then $M' \le Z(Q)$. In particular
$|Q/Z_2(Q)| > r^4$.
\end{claim}
\medskip
We know that $|M/Z_2(Q)| = r^4$ by hypothesis. Since $Z_2(Q)' \le Q'
\cap Z(Q)$ and $Q'/(Z(Q)\cap Q')$ is elementary abelian by Lemma~\ref{QProps} (iv), the commutator map defines a
$\GF(r)\ov G$-module homomorphism from $M/Z_2(Q) \otimes M/Z_2(Q)$ to $Q'/(Q'\cap Z(Q))$ which is well-defined as
$[Q,Z_2(Q)]\le Q'\cap Z(Q)$. Since $[a,b]=[b,a]^{-1}$ for all $a,b\in Q$, this map factors through $(M/Z_2(Q))
\wedge (M/Z_2(Q))$. But $(M/Z_2(Q)) \wedge (M/Z_2(Q))$ has no $4$-dimensional quotients by
Lemma~\ref{modfacts} and so the commutator map is trivial. Therefore, $M'\le Z(Q)\cap Q'$ as claimed. If
$|Q/Z_2(Q)| =r^4$, then we may take $M= Q$ and conclude that $Q$ has class two which is absurd.
\hfill$\blacksquare$
\bigskip
We now simultaneously define bases for all the $G$-composition
factors of $Q$ as follows. Let $U$ be such a composition factor. Then there is an isomorphism $\psi_U$
from $U$ to $V$. For $1\le i \le 4$, we let $u_i = (v_i)\psi_U^{-1}$. We call this a \emph{standard basis} of $U$. Given a standard basis $u_1, \dots, u_4$, we define $\sigma_u = u_1u_2u_3u_4$.
Recall the homomorphism $\gamma$ from Table~\ref{tab1}. Suppose that $R,S$ and $T$ are $G$-composition factors with standard bases $r_1, \dots, r_4$ and $s_1, \dots , s_4$. Then define a map from $R\otimes S$ to $T$ by setting, for all $i,j \in \{1, \dots 4\}$:
$$r_i\otimes s_j \mapsto (r_i\psi_R\otimes s_j\psi_S)\gamma\psi_T^{-1}.$$
Now we can determine the image
of $r_i\otimes s_j$ in $T$ represented in the standard basis for $T$ directly from the table describing
$\gamma$. Thus, for example, $r_1\otimes s_2 $ maps to $\sigma_t$ and $r_1\otimes s_1$ maps to
$\sigma_t^2t_1^{-5}$. Replacing $\gamma$ by a scalar multiple $m\gamma$ we get that $r_1\otimes s_1$ maps to
$(\sigma_t^2t_1^{-5})^m$.
We are now going to exploit the Hall-Witt identity \cite[Lemma 5.6.1 (iv)]{Gorenstein} which in a class three
group such as $Q$ takes the form
$$[x,y,z][y,z,x][z,x,y]=1$$
for all $x,y,z \in Q$.
\medskip
\begin{claim}\label{M2} $|Q/Z_2(Q)| = r^{12}$.
\end{claim}
\medskip
As $|Q/Z_2(Q)| > r^4$ by (\ref{M1}), it is sufficient to exclude the case where $|Q/Z_2(Q)| = r^{8}$.
Hence assume that $|Q/Z_2(Q)| = r^{8}$.
We let $M_1$ and $M_2$ be normal subgroups of $G$ such that $Q= M_1M_2$ and such that $C:= M_1/Z_2(Q)$ and $D:=M_2/Z_2(Q)$ have
order $r^4$. We choose standard bases $c_1, \dots, c_4$ for $C$ and $d_1, \dots, d_4$ for $D$. Then, as in
\ref{M1}, the commutator map defines an $\GF(r)\ov G$-module homomorphism from $C \otimes D$ to $Q'/(Q'\cap Z(Q))$. If
this map is trivial, then $[M_1,M_2]\le Q' \cap Z(Q)$ and \ref{M1} implies that $Q'\le Z(Q)$ which is not
the case. By Lemmas~\ref{unique} and \ref{QProps}(vi), the image of this commutator map is $4$-dimensional and
isomorphic to $V$. We let the image be $E$ and take a standard basis $e_1, \dots, e_4$ such that the commutator
map is defined by $\gamma$.
Finally we may assume that $\Gamma_2(Q)=[M_1,M_2,M_2]$ and take a standard
basis $f_1, \dots, f_4$, again chosen so that $\gamma$ represents the commutator map from $E \otimes D$ to
$\Gamma_2(Q)$. We consider the Hall-Witt identity with the elements $c_1$, $d_1$ and $d_2$ (and here it is
critical to note that the identity is independent of the representatives for $c_1$, $d_1$ and $d_2$ that we
choose). We have $[d_1,d_2] \in Z(Q)$ from \ref{M1} and so $[d_1,d_2,c_1]=1$. Calculating further commutators,
using the map $\gamma$ from Table~\ref{tab1}, we get
$$[c_1,d_1,d_2] = [ (c_1\psi_C\otimes d_1\psi_D)\gamma \psi_E^{-1}, d_2]=[\sigma_e^2e_1^{-5} ,d_2] =
(\sigma_e^2e_1^{-5} \psi_E \otimes d_2\psi_D)\gamma\psi_F^{-1} = \sigma_f^{5}f_2^{-10}
$$ and, similarly,
$$[d_2,c_1,d_1] = [\sigma_e,d_1] = \sigma_f^5f_1^{-5}$$
which is a contradiction as their product is not the identity.
Thus $|Q/Z_2(Q)| = r^{12}$. \hfill$\blacksquare$
\bigskip
From \ref{M2} we have that $|Q/Z_2(Q)|= r^{12}$. Let $M_1$, $M_2$ and $M_3$ be normal
subgroups of $G$ such that $Q=
M_1M_2M_3$ and such that $C:= M_1/Z_2(Q)$, $D:=M_2/Z_2(Q)$ and $E:=M_3/Z_2(Q)$ have order $r^4$. We take standard bases $c_1, \dots, c_4$, $d_1, \dots, d_4$ and $e_1, \dots, e_4$ for
$C$, $D$ and $E$ as described. Since $M_i' \le Z(Q)$ by \ref{M1}, we may assume that $F:= [M_1,M_2](Q' \cap
Z(Q))/(Q' \cap Z(Q))$ is non-trivial.
We set
$$H:= [M_1,M_3](Q' \cap Z(Q))/(Q' \cap Z(Q))$$
and
$$J:= [M_2,M_3](Q'
\cap Z(Q))/(Q' \cap Z(Q)).$$
These may be trivial groups. We let $f_1,\dots, f_4$ be a standard basis for $F$,
and when $H$ or $J$ is non-trivial, we take standard bases $h_1,\dots, h_4$ and $j_1,\dots,j_4$ for $H$ and $J$
respectively, where all the bases are chosen so that the commutator map is represented by $\gamma$. Let $k_1,
\dots, k_4$ be a standard basis for $\Gamma_2(Q)$. Notice that the basis $k_1,\dots, k_4$ cannot necessarily be
chosen so that all the possible commutator maps are represented by $\gamma$ itself. However, they are represented
by scalar multiples of $\gamma$ and the powers appearing in the images below do not affect the failure of the
Hall-Witt identity which we now proceed to check for the elements $c_1$, $d_1$ and $e_2$. Taking all commutator
maps to be $\gamma$, we calculate
$$[c_1,d_1,e_2]= [\sigma_f^2f_1^{-5}, e_2] = \sigma_k^5k_2^{-10},$$
$$[d_1,e_2,c_1] = \begin{cases}1&J=1\cr [\sigma_j,c_1] =
\sigma_k^5k_1^{-5}&\text{ otherwise}\end{cases}$$ and
$$[e_2,c_1,d_1] = \begin{cases}1&H=1\cr [\sigma_h,c_1] =
\sigma_k^5k_1^{-5}&\text{ otherwise}\end{cases}$$ and, in full generality, the images lie in the cyclic groups
generated by the elements presented above. Thus it follows that the Hall-Witt identity does not hold in the
putative group $Q$. This concludes the proof of the theorem.
\section{Final remarks}
In this section we prove that if $E(G/F(G))\cong \PSL_2(49)$, then $F(G)$ is an elementary abelian $7$-group. We also take the opportunity to add more
detail to the statement of Theorem~\ref{MainThm} by showing that the abelian $3$-group in Theorem~\ref{MainThm}
(iv) is elementary abelian.
\begin{lemma}\label{notin} The group $\PSL_2(49)$ is not isomorphic to a subgroup of
$\GL_4(\mathbb Z/49\mathbb Z)$.
\end{lemma}
\begin{proof} We calculate using {\sc Magma} \cite{Magma}. Let $J \cong \PSL_2(49)$ and suppose that $J$ is a
subgroup of $K=\GL_4(\mathbb Z/49\mathbb Z)$. Then, as $J$ contains a subgroup isomorphic to $\Alt(5)$, we can again take the integral
version $A$ of $\Alt(5)$ from Section~{2} and suppose that it is a subgroup of $J$. Now we take a Sylow
$5$-subgroup $S$ of $A$ and determine $C_K(S)$ using {\sc Magma}. It has order $2^5\cdot 3\cdot 5^2\cdot 7^4$ and is
$5$-closed. We let $T \in \sy_5(C_K(S))$. Then $T$ is the unique subgroup of order $25$ in $K$ containing $S$. It
follows that $J = \langle A,T\rangle$ and a calculation shows that $J$ has order $2^4\cdot 3\cdot
5^2\cdot 7^8$. But this is a contradiction because $7^8$ does not divide the order of $J \cong \PSL_2(49)$. Hence $\PSL_2(49)$ is not isomorphic to a subgroup of $K$.
\end{proof}
\begin{lemma}\label{74 unique} Suppose that $k$ is an algebraically closed field of characteristic 7 and that $V$ is an irreducible $k\PSL_2(49)$-module which admits an
element of order $5$ without non-zero fixed points. Then $V$ is isomorphic to $N \otimes N^\sigma$ where $N$ is the
natural $k\SL_2(49)$-module and
$\sigma$ is the automorphism of $k$ obtained by raising every element to its seventh power. Furthermore, $V$ is not a composition factor of $V \otimes V$.
\end{lemma}
\begin{proof} By \cite[Section 30 (98)]{BN} we have $V = U \otimes W^\sigma$ where $U$ and $W$ are basic
$k\SL_2(49)$-modules. The basic $k\SL_2(49)$-modules can be identified with the seven modules $U_j$, $0\le j \le
6$, obtained as degree $j$ homogeneous polynomials in $k[x,y]$. Then $\dim U_j=j+1$. Now let $\phi$ in $\SL_2(49) \le \GL_2(k)$ be of
order $5$. Then $\phi$ diagonalises in $\GL_2(k)$ and so we may assume that $\phi$ acts as the diagonal matrix
$\diag(\l,\l^{-1})$. It is straightforward to check that, for $j \ge 3$, $\phi$ has all possible non-trivial
eigenvalues on $U_j$ and so if $U_j$ or $U_j^\sigma$ appears in the tensor product defining $V$, then as $\phi$
acts fixed point freely on $V$, one of the tensor factors in $V$ must be $U_0$. But then $\dim U_j$ must be even and we see
that $V$ is a module for $\SL_2(49)$, not for $\PSL_2(49)$. Similarly, we now deduce the only contenders for
$V$ are $U_1 \otimes U_1^\sigma$, $U_2 \otimes U_1^\sigma$ and $U_1\otimes U_2^\sigma$. In the latter two cases
we again get a representation of $\SL_2(49)$ rather than $\PSL_2(49)$ and this shows that the only possibility is
that $V= U_1 \otimes U_1^\sigma$ and it is easy to check that this module has the required properties.
Finally, we have \begin{eqnarray*}V \otimes V&=& (U_1 \otimes U_1^\sigma)\otimes (U_1 \otimes U_1^\sigma)= (U_1
\otimes U_1)\otimes (U_1^\sigma \otimes U_1^\sigma) \\&=& (U_0 \oplus U_2)\otimes (U_0^\sigma\oplus U_2^\sigma)=
U_0 \oplus U_2^\sigma \oplus U_2 \oplus ( U_2\otimes U_2^\sigma)\end{eqnarray*} and none of these irreducible
summands are isomorphic to $V$.
\end{proof}
\begin{lemma}\label{cr49} Any $\GF(7)\PSL_2(49)$-module which has all composition factors isomorphic to the module described in Lemma~\ref{74 unique} is completely reducible.
\end{lemma}
\begin{proof} A {\sc Magma} calculation has shown that $\mathrm
{Ext}^1_X(V,V)=\mathrm H^1(X,V\otimes V^*)= 0$. This proves the lemma. \end{proof}
\begin{lemma}\label{49}
Suppose that $r$ is a prime and that $G$ is a group such that $X:=G/O_r(G) \cong\PSL_2(49)$ and $G$ has an element of order $5$ acting fixed point freely on $O_r(G)$. Then $r=7$ and $O_7(G)$ is elementary abelian and completely reducible as a $\GF(7)X$-module.
\end{lemma}
\begin{proof}
Let $Q:=O_r(G)$.
By \cite{c55} we may suppose that $r=7$ and thus that $Q$ is a $7$-group.
Suppose first that $Q$ is abelian, but not elementary abelian, and that $G$ has minimal order with these properties. Every $G$-chief factor of $Q$ is isomorphic to the irreducible $4$-dimensional module
$V$ of $X$ described in Lemma~\ref{74 unique}. Let $N$
be a minimal $G$-invariant subgroup of $Q$. Then $N$ has order $7^4$ and $Q/N$ is elementary abelian. Since every maximal $G$-invariant subgroup is also elementary abelian, we infer that $Q$ has a unique
such subgroup. By Lemma~\ref{cr49}, $Q/N$ is completely reducible. Therefore $Q/N$ has order $7^4$ and so $Q $ is homocyclic of order
$49^4$ and Lemma~\ref{notin} provides the contradiction. Thus, if $Q$ is abelian, it is elementary abelian and completely reducible as a $\GF(7)X$-module.
Suppose now, aiming for a contradiction, that $G$ is chosen so that $Q$ is a class two group of minimal order. Thus $Q'$ is elementary abelian of order $7^4$ and $Q/Q'$ is elementary abelian by the previous paragraph. In addition we may assume that $Q'= Z(Q)$ as $Z(Q)$ is completely reducible as a $\GF(7)X$-module. Since the
commutator map from $Q/Z(Q)\times Q/Z(Q)$ determines an $\GF(r)X$-module epimorphism from $Q/Z(Q)\otimes Q/Z(Q)$ to
$Q'=Z(Q)$, and $V \otimes V$ has no quotients isomorphic to $V$ by Lemma~\ref{74 unique}, we have a contradiction. This proves the lemma.
\end{proof}
\begin{lemma}\label{notin2} The group $\Alt(6)$ is not isomorphic to a subgroup of
$\GL_4(\mathbb Z/9\mathbb Z)$.\end{lemma}
\begin{proof} Take the integral
copy of $A\cong \Alt(5)$ from Section~2 and determine the
normaliser $N$ of a Sylow $2$-subgroup using {\sc Magma}. This has order $7776$. None
of the groups $\langle A,x\rangle$, $x \in N$, is isomorphic to
$\Alt(6)$. This proves the claim.
\end{proof}
Because of Lemmas~\ref{coho} and \ref{notin2} we can now prove our version of Theorem~\ref{MainThm} (iv) using similar arguments to
those used towards the end of the proof of Lemma~\ref{49}.
\begin{lemma} If $\Alt(6)$ acts on an abelian $3$-group $Q$ with an element of order $5$ acting fixed point freely, then $Q$ is elementary
abelian and completely reducible as a $\GF(3)\Alt(6)$-module.\qed
\end{lemma}
|
\section{Introduction and notation}
Recently, $q$-analogs of classical congruences have been studied by several
authors including {\cite{clark-qbin95}}, {\cite{andrews-qcong99}},
{\cite{shipan-qwolst07}}, {\cite{pan-qlehmer07}}, {\cite{chapman-qwilson08}},
{\cite{dilcher-qharm08}}. Here, we consider the classical congruence
\begin{equation}\label{eq:classical}
\binom{a p}{b p} \equiv \binom{a}{b} \modulo{p^3}
\end{equation}
which holds true for primes $p \geqslant 5$. This also appears as Problem 1.6
(d) in {\cite{stanley-ec1}}. Congruence (\ref{eq:classical}) was proved in
1952 by Ljunggren, see {\cite{granville-bin97}}, and subsequently generalized
by Jacobsthal, see Remark \ref{rk-jacobsthal}.
Let $[n]_q := 1 + q + \ldots q^{n - 1}$, $[n]_q ! := [n]_q [n - 1]_q
\cdots [1]_q$ and
\[ \qbinom{n}{k}{q} := \frac{[n]_q !}{[k]_q ! [n - k]_q !} \]
denote the usual $q$-analogs of numbers, factorials and binomial coefficients
respectively. Observe that $\qn{n}{1}=n$ so that in the case $q=1$ we recover
the usual factorials and binomial coefficients as well. Also, recall that the
$q$-binomial coefficients are polynomials in $q$ with nonnegative integer
coefficients. An introduction to these $q$-analogs can be found in
\cite{stanley-ec1}.
We establish the following $q$-analog of \eqref{eq:classical}:
\begin{theorem}\label{thm:q}
For primes $p\ge5$ and nonnegative integers $a, b$,
\begin{equation}\label{eq:qclassical}
\qbinom{ap}{bp}{q} \equiv \qbinom{a}{b}{q^{p^2}} - \binom{a}{b + 1} \binom{b + 1}{2}
\frac{p^2 - 1}{12} (q^p - 1)^2 \modulo{\qn{p}{q}^3} .
\end{equation}
\end{theorem}
The congruence \eqref{eq:qclassical} and similar ones to follow are to be
understood over the ring of polynomials in $q$ with integer coefficients. We remark
that $p^2-1$ is divisible by $12$ for all primes $p\ge5$.
Observe that \eqref{eq:qclassical} is indeed a $q$-analog of
\eqref{eq:classical}: as $q\to1$ we recover \eqref{eq:classical}.
\begin{example}
Choosing $p=13$, $a=2$, and $b=1$, we have
\begin{align*}
\qbinom{26}{13}{q} &= 1+q^{169} - 14 (q^{13}-1)^2
+ (1+q+\ldots+q^{12})^3 f(q)
\end{align*}
where $f(q) = 14-41q+41q^2-\ldots+q^{132}$ is an irreducible polynomial with
integer coefficients. Upon setting $q=1$, we obtain $\binom{26}{13}\equiv2$
modulo $13^3$.
\end{example}
Since our treatment very much parallels the classical case, we give a brief
history of the congruence \eqref{eq:classical} in the next section before
turning to the proof of Theorem \ref{thm:q}.
\section{A bit of history}
A classical result of Wilson states that $(n-1)! + 1$ is divisible by $n$ if
and only if $n$ is a prime number. ``In attempting to discover some analogous
expression which should be divisible by $n^2$, whenever $n$ is a prime, but
not divisible if $n$ is a composite number'', \cite{babbage}, Babbage is led to
the congruence
\begin{equation}\label{eq:babbage}
\binom{2p-1}{p-1} \equiv 1 \modulo{p^2}
\end{equation}
for primes $p\ge3$. In 1862 Wolstenholme, \cite{wolstenholme}, discovered
\eqref{eq:babbage} to hold modulo $p^3$, ``for several cases, in testing
numerically a result of certain investigations, and after some trouble
succeeded in proving it to hold universally'' for $p\ge5$. To this end, he
proves the fractional congruences
\begin{align}
\sum_{i=1}^{p-1} \frac{1}{i} &\equiv 0 \modulo{p^2}, \label{eq:wol1}\\
\sum_{i=1}^{p-1} \frac{1}{i^2} &\equiv 0 \modulo{p} \label{eq:wol2}
\end{align}
for primes $p\ge5$. Using \eqref{eq:wol1} and \eqref{eq:wol2} he then extends
Babbage's congruence \eqref{eq:babbage} to hold modulo $p^3$:
\begin{equation}\label{eq:wolstenholme}
\binom{2p-1}{p-1} \equiv 1 \modulo{p^3}
\end{equation}
for all primes $p\ge5$. Note that \eqref{eq:wolstenholme} can be rewritten as
$\binom{2p}{p} \equiv 2$ modulo $p^3$. The further generalization of
\eqref{eq:wolstenholme} to \eqref{eq:classical}, according to
\cite{granville-bin97}, was found by Ljunggren in 1952. The case $b=1$ of
\eqref{eq:classical} was obtained by Glaisher, \cite{glaisher}, in 1900.
In fact, Wolstenholme's congruence \eqref{eq:wolstenholme} is central to the
further generalization \eqref{eq:classical}. This is just as true when
considering the $q$-analogs of these congruences as we will see here in Lemma
\ref{lem:2}.
A $q$-analog of the congruence of Babbage has been found by Clark
{\cite{clark-qbin95}} who proved that
\begin{equation}\label{eq:clark}
\qbinom{ap}{bp}{q} \equiv \qbinom{a}{b}{q^{p^2}} \modulo{\qn{p}{q}^2} .
\end{equation}
We generalize this congruence to obtain the $q$-analog \eqref{eq:qclassical} of
Ljunggren's congruence \eqref{eq:classical}. A result similar to
\eqref{eq:clark} has also been given by Andrews in \cite{andrews-qcong99}.
Our proof of the $q$-analog proceeds very closely to the history just outlined.
Besides the $q$-analog \eqref{eq:clark} of Babbage's congruence
\eqref{eq:babbage} we will employ $q$-analogs of Wolstenholme's harmonic
congruences \eqref{eq:wol1} and \eqref{eq:wol2} which were recently supplied by
Shi and Pan, \cite{shipan-qwolst07}:
\begin{theorem}\label{thm:shipan}
For primes $p\ge5$,
\begin{equation}\label{eq:qwol1}
\sum_{i = 1}^{p - 1} \frac{1}{[i]_q} \equiv - \frac{p - 1}{2} (q
- 1) + \frac{p^2 - 1}{24} (q - 1)^2 [p]_q \modulo{\qn{p}{q}^2}
\end{equation}
as well as
\begin{equation}\label{eq:qwol2}
\sum_{i = 1}^{p - 1} \frac{1}{[i]_q^2} \equiv - \frac{(p - 1) (p
- 5)}{12} (q - 1)^2 \modulo{\qn{p}{q}}.
\end{equation}
\end{theorem}
This generalizes an earlier result {\cite{andrews-qcong99}} of Andrews.
\section{A $q$-analog of Ljunggren's congruence}
In the classical case, the typical proof of Ljunggren's congruence
\eqref{eq:classical} starts with the Chu-Vandermonde identity which has the
following well-known $q$-analog:
\begin{theorem}\label{thm:qchu}
\begin{equation*}
\qbinom{m+n}{k}{q} = \sum_j \qbinom{m}{j}{q} \qbinom{n}{k-j}{q} q^{j(n-k+j)}.
\end{equation*}
\end{theorem}
We are now in a position to prove the $q$-analog of \eqref{eq:classical}.
\begin{proof}[of Theorem \ref{thm:q}]
As in {\cite{clark-qbin95}} we start with the identity
\begin{equation}
\qbinom{ap}{bp}{q} = \sum_{c_1 + \ldots + c_a = b p}
\qbinom{p}{c_1}{q} \qbinom{p}{c_2}{q} \cdots \qbinom{p}{c_a}{q}
q^{^{p \sum_{1 \leqslant i \leqslant a} (i - 1) c_i
- \sum_{1 \leqslant i < j \leqslant a} c_i c_j}}
\label{eq-qchu}
\end{equation}
which follows inductively from the $q$-analog of the Chu-Vandermonde identity
given in Theorem \ref{thm:qchu}. The summands which are not divisible by
$[p]_q^2$ correspond to the $c_i$ taking only the values $0$ and $p$. Since
each such summand is determined by the indices $1 \le j_1 < j_2 < \ldots <
j_b \le a$ for which $c_i = p$, the total contribution of these terms is
\[ \sum_{1 \leqslant j_1 < \ldots < j_b \leqslant a} q^{p^2 \sum_{k = 1}^b
(j_k - 1) - p^2 \binom{b}{2}} = \sum_{0 \leqslant i_1
\leqslant \ldots \leqslant i_b \leqslant a - b} q^{p^2 \sum_{k = 1}^b i_k}
= \qbinom{a}{b}{q^{p^2}} . \]
This completes the proof of \eqref{eq:clark} given in {\cite{clark-qbin95}}.
To obtain \eqref{eq:qclassical} we now consider those summands in
(\ref{eq-qchu}) which are divisible by $[p]_q^2$ but not divisible by
$[p]_q^3$. These correspond to all but two of the $c_i$ taking values $0$ or
$p$. More precisely, such a summand is determined by indices $1 \leqslant
j_1 < j_2 < \ldots < j_b < j_{b + 1} \leqslant a$, two subindices $1
\leqslant k < \ell \leqslant b + 1$, and $1 \leqslant d \leqslant p - 1$
such that
\[ c_i = \left\{ \begin{array}{l}
d \text{ for $i = j_k$},\\
p - d \text{ for $i = j_{\ell}$},\\
p \text{ for $i \in \{j_1, \ldots, j_{b + 1} \}\backslash\{j_k,
j_{\ell} \}$},\\
0 \text{ for $i \not\in \{j_1, \ldots, j_{b + 1} \}$} .
\end{array} \right. \]
For each fixed choice of the $j_i$ and $k, \ell$ the contribution of the
corresponding summands is
\[ \sum_{d = 1}^{p - 1} \qbinom{p}{d}{q} \qbinom{p}{p-d}{q}
q^{p \sum_{1 \leqslant i \leqslant a} (i - 1) c_i
- \sum_{1 \leqslant i < j \leqslant a} c_i c_j} \]
which, using that $q^p \equiv 1$ modulo $\qn{p}{q}$, reduces modulo $\qn{p}{q}^3$ to
\[ \sum_{d = 1}^{p - 1} \qbinom{p}{d}{q} \qbinom{p}{p-d}{q} q^{d^2}
= \qbinom{2p}{p}{q} - [2]_{q^{p^2}} . \]
We conclude that
\begin{equation}\label{eq:cong2}
\qbinom{ap}{bp}{q} \equiv \qbinom{a}{b}{q^{p^2}} + \binom{a}{b + 1}
\binom{b + 1}{2} \left( \qbinom{2p}{p}{q} - [2]_{q^{p^2}} \right)
\modulo{\qn{p}{q}^3}.
\end{equation}
The general result therefore follows from the special case $a = 2$, $b = 1$
which is separately proved next.
\end{proof}
\section{A $q$-analog of Wolstenholme's congruence}
We have thus shown that, as in the classical case, the congruence
\eqref{eq:qclassical} can be reduced, via \eqref{eq:cong2}, to the case $a=2$,
$b=1$. The next result therefore is a $q$-analog of Wolstenholme's congruence
\eqref{eq:wolstenholme}.
\begin{lemma}\label{lem:2}
For primes $p\ge5$,
\[ \qbinom{2p}{p}{q} \equiv [2]_{q^{p^2}} - \frac{p^2 - 1}{12} (q^p -
1)^2 \modulo{\qn{p}{q}^3} . \]
\end{lemma}
\begin{proof}
Using that $\qn{a n}{q} = \qn{a}{q^n} \qn{n}{q}$ and $\qn{n + m}{q} =
\qn{n}{q} + q^n \qn{m}{q}$ we compute
\[ \qbinom{2p}{p}{q} = \frac{\qn{2 p}{q} \qn{2 p - 1}{q} \cdots \qn{p + 1}{q}}
{\qn{p}{q} \qn{p - 1}{q} \cdots \qn{1}{q}} = \frac{\qn{2}{q^p}}{\qn{p - 1}{q} !} \prod_{k = 1}^{p
- 1} \left( \qn{p}{q} + q^p \qn{p - k}{q} \right) \]
which modulo $\qn{p}{q}^3$ reduces to (note that $\qn{p-1}{q} !$ is relatively
prime to $\qn{p}{q}^3$)
\begin{equation}
\qn{2}{q^p} \left( q^{(p - 1) p} + q^{(p - 2) p} \sum_{1 \leqslant i
\leqslant p - 1} \frac{\qn{p}{q}}{\qn{i}{q}} + q^{(p - 3) p} \sum_{1
\leqslant i < j \leqslant p - 1} \frac{\qn{p}{q} \qn{p}{q}}{\qn{i}{q} \qn{j}{q}}
\right) . \label{eq-2ppharmonic}
\end{equation}
Combining the results \eqref{eq:qwol1} and \eqref{eq:qwol2} of Shi and Pan,
\cite{shipan-qwolst07}, given in Theorem \ref{thm:shipan}, we deduce that for
primes $p\ge5$,
\begin{equation}
\sum_{1 \leqslant i < j \leqslant p - 1} \frac{1}{\qn{i}{q} \qn{j}{q}}
\equiv \frac{(p - 1) (p - 2)}{6} (q - 1)^2 \modulo{\qn{p}{q}} .
\end{equation}
Together with \eqref{eq:qwol1} this allows us to rewrite
(\ref{eq-2ppharmonic}) modulo $\qn{p}{q}^3$ as
\begin{align*}
&& \qn{2}{q^p} \left( q^{(p - 1) p} + q^{(p - 2) p} \left( - \frac{p -
1}{2} (q^p - 1) + \frac{p^2 - 1}{24} (q^p - 1)^2 \right) + \right.\\
&& \left. + q^{(p - 3) p} \frac{(p - 1) (p - 2)}{6} (q^p - 1)^2 \right).
\end{align*}
Using the binomial expansion
\[ q^{m p} = ((q^p - 1) + 1)^m = \sum_k \binom{m}{k} (q^p - 1)^k \]
to reduce the terms $q^{m p}$ as well as $\qn{2}{q^p} = 1 + q^p$ modulo the
appropriate power of $\qn{p}{q}$ we obtain
\[ \qbinom{2p}{p}{q} \equiv 2 + p (q^p - 1) + \frac{(p - 1) (5 p -
1)}{12} (q^p - 1)^2 \modulo{\qn{p}{q}^3} . \]
Since
\[ [2]_{q^{p^2}} \equiv 2 + p (q^p - 1) + \frac{(p - 1) p}{2} (q^p - 1)^2
\modulo{\qn{p}{q}^3} \]
the result follows.
\end{proof}
\begin{remark}
\label{rk-jacobsthal}Jacobsthal, see {\cite{granville-bin97}}, generalized
the congruence (\ref{eq:classical}) to hold modulo $p^{3 + r}$ where $r$ is
the $p$-adic valuation of
\[ a b (a - b) \binom{a}{b} = 2 a \binom{a}{b + 1} \binom{b + 1}{2} . \]
It would be interesting to see if this generalization has a nice analog in
the $q$-world.
\end{remark}
\acknowledgements
Most parts of this paper have been written during a visit of the author at
Grinnell College. The author wishes to thank Marc Chamberland for his
encouraging and helpful support. Partial support of grant NSF-DMS 0713836 is
also thankfully acknowledged.
|
\section{Introduction}
Although a large fraction of nearby stars occur in multiple systems, direct detection exoplanet searches have generally selected against binary targets because a secondary star complicates observations and limits the detection sensitivity. In addition to observational complications, binary systems were traditionally assumed to be a hostile environment for the formation and evolution of planetary systems. However, direct imaging of stars known to harbor planets detected by radial velocities has revealed several binary star systems \mycitep{Pat02, Egg09}. Protoplanetary disks, which form the material basis for planet formation, are observed in both circumstellar and circumbinary configurations in binary systems \mycitep{Rod96, Tri07}, and the growth and settling of dust grains is common in binary star systems \mycitep{Pas08}. Theoretical numerical simulations successfully model the evolution of protoplanets \mycitep{Pie07}, terrestrial planets \mycitep{Qui06}, giant planets \mycitep{Pie08} and brown dwarfs \mycitep{Jia04} in circumbinary disks. Furthermore, theoretical models also permit planet formation through gravitational instability in binary systems. Planets formed in this way have large separations from their stars and are therefore particularly important targets for direct imaging. The influence of a secondary star can in some cases prevent a collapse through tidal heating \mycitep{May05} or trigger the collapse for an otherwise stable disk \mycitep{Bos06}, leaving a characteristic imprint on the demographics of planets in binary systems, the observation of which would provide an invaluable test for planet formation theory. A secondary star can also act as a source of angular momentum for a circumbinary planet and either scatter or tidally push the planet into wider orbits, thus enriching the expected population of wide-orbit stars \mycitep{Nel03, Ver04, Hol99, Kle00}.
Direct detection techniques are critical for investigating extrasolar planets on large ($>$ 10 AU) orbital separations, and the first direct detections of extrasolar planets have confirmed that planets do exist in these orbits (e.g., \mycitealt{Mar08,Kal08,Lag09}). High contrast instrumentation and direct detection techniques are being developed to probe planet formation and the evolution of planetary systems. In order to achieve high contrast ($>$ 10$^{-6}$) at small angular separations ($<$ 1$\arcsec$), one must control the diffracted light from the host star in the image plane. However, the majority of coronagraphs are designed to remove light from single host stars. If the secondary star lies close to the target star and it is not suppressed, the secondary star's light will overwhelm the signal from any faint companions. The peak of the G dwarf companion distribution is near 30 AU \mycitep{Duq91}, and while also accounting for random orbital inclinations and phases on the sky, the vast majority of nearby binary stars ($<$ 100 pc) targeted by high contrast imaging surveys would benefit from having a specialized coronagraph.
Coronagraphs capable of simultaneously blocking the light from binary stars can be grouped into two categories: coronagraphs with linear masks, and coronagraphs with dual circular masks. Each type has advantages and disadvantages; in the following, we present examples of each and discuss their performance in the presence of a typical low-order wavefront noise following an adaptive optics system. In particular, we present a dual-mask design based on an APLC which allows for a small inner working angle and a large discovery space even for an obstructed aperture, and which minimizes cross-talk between the masks. We also address manufacturing and implementation concerns such as obstructed apertures, field rotation, manufacturing, and mask placement, and outline a control scheme for maintaining alignment of the system even in the presence of atmospheric errors.
\section{Types of Coronagraphs for High Contrast Imaging}
The general class of Lyot-type coronagraphs have a common structure to the optical design. The primary components of a Lyot coronagraph consist of an initial pupil equivalent to the pupil of the telescope, potentially with an apodization; a focal plane mask, which can be hard-edged or vary in amplitude and phase; and a Lyot stop at the reimaged pupil, which is hard-edged and often (but not always) undersized with respect to the telescope pupil. Lyot's original design \mycitep{Lyo39} was equipped with hard-edged masks in all planes; however, subsequent variants include apodized pupil Lyot coronagraphs (APLCs) \mycitep{Sou03}, band-limited coronagraphs (BLCs) \mycitep{Kuc02}, four-quadrant phase mask coronagraphs (4QPMs) \mycitep{Rou00}, and vortex coronagraphs (VCs) \mycitep{Foo05, Maw10}. \mycitet{Guy06} provides a detailed analysis of comparative coronagraphy.
For the specialized case of binary stars, these coronagraphs split into two general categories: (1) coronagraphs with a single focal plane mask which is oriented to block both stars simultaneously, and (2) coronagraphs with two masks, one placed to block each star.
\subsection{Single-mask case}
The main purpose behind using a single mask is simplicity---it avoids additional mechanisms and complexity by having a single fixed mask. In exchange, this mask will tend to block out more of the field of the view than just the target stars, obscuring edge-on systems in particular.
These masks work by using a separable apodization $A(x, y) = A(x)A(y)$, generally with $A(y) = 1$. The mask is then oriented so that $y$-axis is aligned with the two stars, and both will be suppressed identically. Coronagraphs that could be considered for this approach include the BLC---in fact, this has been demonstrated on-sky \mycitep{Cre10}---and an APLC designed for a linear mask, following \mycitet{Aim02}.
\subsection{Dual-mask case}
The rationale behind using two masks in the image plane is to block only the regions of the field of view containing the starlight, so the remainder of the field can be investigated as discovery space. However, the discovery space of the coronagraphic mask may not be symmetric (as is the case for the BLC), which could lead to significant preference for certain orbital inclinations. Observations of T-Tauri stars, for example, indicate circumstellar disks in binary systems tend to be parallel to each other \mycitep{Mon07}. Assuming that all spatial orientations are randomly distributed on the sky, the abundance of orbits with a specific inclination angle is proportional to the cosine of the inclination. This represents a significant bias towards encountering edge-on binary systems and suggests that the binary orientation direction is a favorable place to find a planetary companion, assuming the planet formed in the disk and was not scattered far from the invariable plane. A dual-mask coronagraph enables sensitivity along this orientation, while the BLC completely obscures planets along that axis. Therefore, the BLC has a strong selection preference for observing face-on systems, where this effect is minimized, but the dual-mask coronagraph has no such requirement. On the other hand, a dual-mask coronagraph system is significantly more complex, as the two masks must be moved in the image plane to coincide with the target stars.
A source located at an angular separation of $\psi_1$ with respect to the optical axis produces a tilted wavefront at the aperture of the telescope; without loss of generality, we can place this source along the $x_1$-axis with unit intensity to give the time-averaged electric field $\Psi_{1}(x_1,y_1; \psi_1)$ following the first pupil plane:
\begin{equation} \label{tilt1}
\Psi_{1}(x_1,y_1; \psi_1) = P(x_1,y_1) e^{[-2 \pi i x_1 \sin{\psi_1}/ \lambda]}
\end{equation}
with $P(x_1,y_1)$ the pupil function of the telescope and $\lambda$ the wavelength under consideration. The electric field at the subsequent image plane ($\Psi_{2}(x_2,y_2; \psi_1)$) can be written:
\begin{align}
\Psi_{2}(x_2,y_2; \psi_1) &= M(x_2, y_2) \times \frac{1}{i \lambda f} \int_{-\infty}^{\infty} \int_{-\infty}^{\infty} e^{-\frac{2 \pi i}{\lambda f} (x_1 x_2 + y_1 y_2)} P(x_1,y_1) e^{-\frac{2 \pi i x_1 \sin{\psi_1}}{\lambda}} \intd{x_1} \intd{y_1} \\
&= M(x_2, y_2) \times \frac{1}{i \lambda f} \int_{-\infty}^{\infty} \int_{-\infty}^{\infty} P(x_1,y_1) e^{-\frac{2 \pi i}{\lambda f} (x_1 [x_2 + f \sin{\psi_1}] + y_1 y_2)} \intd{x_1} \intd{y_1} \label{tilt2}
\end{align}
with $f$ the focal length of the telescope and $M(x_2, y_2)$ the focal plane mask. In the on-axis case,
\begin{align}
\Psi_{2}(x_2,y_2; 0) &= M(x_2, y_2) \times \frac{1}{i \lambda f} \int_{-\infty}^{\infty} \int_{-\infty}^{\infty} P(x_1,y_1) e^{-\frac{2 \pi i}{\lambda f} (x_1 x_2 + y_1 y_2)} \intd{x_1} \intd{y_1}, \label{tilt2a}
\end{align}
and so the shift by $f \sin{\psi_1}$ from the tilt of the incident wavefront creates a misalignment with the mask. We can compensate by moving the mask by $-f \sin{\psi_1}$:
\begin{align}
\Psi_{2s}(x_2,y_2; \psi_1) &= M(x_2 + f \sin{\psi_1}, y_2) \times \frac{1}{i \lambda f} \int_{-\infty}^{\infty} \int_{-\infty}^{\infty} P(x_1,y_1) e^{-\frac{2 \pi i}{\lambda f} (x_1 [x_2 + f \sin{\psi_1}] + y_1 y_2)} \intd{x_1} \intd{y_1} \label{tilt3a} \\
&= \Psi_{2}(x_2 + f \sin{\psi_1},y_2; 0) \label{tilt3b}
\end{align}
so the system with an off-axis source provides identical suppression at $x_2 = -f \sin{\psi_1}$ as does an on-axis mask and star. (The image plane masks behave achromatically in this configuration, where the wavelength does not change the magnitude of the shift of the PSF in the image plane.) If the two stars are located at $\pm \psi_1$ from the optical axis, then, the two masks would be placed at $\pm f \sin{\psi_1}$, and would suppress both stars simultaneously. A schematic of this system is shown in \myfig{fig:bscdiagram}, for a simple telescope with circular symmetry.
\begin{figure}
\begin{center}
\includegraphics[width=6.5in]{Slide3a}
\end{center}
\caption{APLC optical components for binary observations. Plane $1$ has an apodizer and the pupil, shown face-on at top; plane $2$ has the dual image plane masks, shown face-on above the plane; plane $3$ is the pupil again, shown face-on above; plane $4$ represents the science camera image plane. The red and green rays show the path of the starlight from the binary pair through the system, while the blue ray shows the path of light from the planet.} \label{fig:bscdiagram}
\end{figure}
Coronagraphs that could in theory be considered for this dual-mask approach include nearly every Lyot-type coronagraph with a radially-symmetric focal-plane mask, including Lyot coronagraphs, APLCs, BLCs with radially-symmetric masks \mycitep{Kuc03}, 4QPMs, and VCs. Unfortunately, this simple analysis ignores one large difficulty: not only does mask A block the core of star A, but mask B will suppress a portion of the sidelobes of star A, which then will scatter light across the image plane, limiting the contrast that can be achieved. Thus, only a coronagraph with the lowest sidelobes at the first image plane is likely to prove promising for a binary system. We investigate the APLC because it includes an apodization at the entrance pupil that reshapes the PSF. As a result, the APLC has optimally low sidelobes \mycitep{Sou03} and will minimize the interaction between the two masks.
\subsection{Additional design considerations}
Most existing telescopes have obstructed apertures, and the effects of secondaries and their support structures on the coronagraphs must be considered in the modeling of the optical performance. In many cases, the effects of these optical obstructions can be mitigated by carefully designing the coronagraph. For example, Lyot stops can be designed to block the diffraction from aperture obstructions \mycitep{Siv05}, spider removal plates can remove the spiders from the image \mycitep{Loz09}, and for APLCs, the apodization can be optimized for arbitrary apertures \mycitep{Sou09}. However, the reduced performance should be taken into account when designing these coronagraphs. \mycitet{Cre10} makes binary star observations with the unobstructed subaperture \mycitep{Ser07} on the Palomar Observatory's 200-inch Hale Telescope, neatly avoiding these coronagraphic complications, but only utilizing $\sim9\%$ of the collecting area of the telescope.
A further difficulty arises from the field rotation, which causes the binary system to rotate with respect to the telescope pupil for telescopes in altitude-azimuth mounts. In high contrast observations, this characteristic can be employed for Angular Differential Imaging (ADI) \mycitep{Mar06}. The major consequence of this is that the coronagraph must be able to maintain its performance regardless of the orientation of the aperture with respect to the image plane mask, including any obstructions. For an unobstructed circular aperture, the circular symmetry of the pupil renders this trivial. The performance of the system is also dependent on the aberrations introduced by the atmosphere and the ability of the adaptive optics (AO) system to correct these wavefront errors.
\section{Simulated performance}
To examine the performance of the coronagraphic options discussed above, we simulate their performance on a realistic system. Here, we use parameters of the Subaru Telescope and HiCIAO \mycitep{Hod08} for simulation purposes: the pupil is a $8.2$m diameter circular aperture with an IR secondary $1.265$m diameter circular obscuration concentric with the primary mirror \mycitep{Usu03}. The secondary is held in place with four spiders, as shown in \myfig{subaruPupil}. From a numerical perspective, we make extensive use of the Fast Lyot algorithm described in \mycitet{Sou07} for efficient propagation.
The simulated field contains a binary system and a faint companion. The system is simulated at $1.65\mu$m for a 10s exposure, using simulated residual wavefronts from the atmosphere following the AO188 adaptive optics system \mycitep{Min10} at 0.5ms intervals, with 200nm of residual static rms wavefront error. The two stars are placed at $\pm 500$\,mas along the $y$-axis, with intensity normalized to 1; the companion is placed at $-1000$\,mas along the $x$-axis and is attenuated to an intensity of $10^{-4}$ for clarity. We choose a 1'' separation of the binary companions because closer pairs interfere with the wavefront sensing capability of the AO188 system (S. Egner, private communication) and thus are unlikely to have a well-corrected field.
We examined the performance of four coronagraphic configurations, beginning with the case of no coronagraph in the system; the performance is shown in \myfig{m1}. The HiCIAO system currently uses a Lyot coronagraph for its observations; we looked at the performance with a standard round hard-edged mask doubled, with each mask centered at $\pm 500$\,mas. At $1.65\mu$m, the mask diameter is $4.8\lambda/D$. In addition, the Lyot stop is undersized to $80\%$ and the secondary is enlarged to $140\%$ \mycitep{Tam06}, to match the existing instrumentation. The performance for this configuration is shown in \myfig{m2}.
Using the techniques described in \mycitep{Sou09}, we designed an apodized pupil Lyot coronagraph for the Subaru pupil which explicitly incorporated the spiders; this pupil is shown in \myfig{m3a}. Note that the inclusion of the spiders leads to breaking of circular symmetry in the apodization; in this case, there is higher throughput in the center of the upper and lower regions than in the side regions. Again, the hard-edged image plane mask was doubled, and each mask centered at $\pm 500$\,mas in the image plane; the mask is identical to the one used for the Lyot coronagraph. The Lyot stop is identical to the pupil in \myfig{subaruPupil}, as the eigenfunction equation underlying the APLC requires the Lyot stop match the pupil. The performance of the coronagraph is shown in \myfig{m3b}.
For the band-limited coronagraph, we aligned a 4th order linear mask along the direction of the two stars. While higher-order designs exist \mycitep{Kuc05}, \mycitet{Cre07} shows that the 4th order mask generally outperforms them except in the presence of very high Strehl. We use a modified Lyot stop, following \mycitep{Siv05}, to compensate for the large spiders present in the Subaru pupil. The band-limited mask and Lyot stop used in our simulations are shown in \myfig{m4a}. The Lyot stop is required to be quite severe to prevent the spiders from leaking into the image plane. The performance of the mask is shown in \myfig{m4b}.
\begin{figure}
\begin{center}
\includegraphics[width=3.1in]{subaruPupil}
\caption{The pupil of Subaru telescope.}
\label{subaruPupil}
\end{center}
\end{figure}
\begin{figure}
\begin{center}
\subfigure[The binary system with no coronagraph.]{\label{m1} \includegraphics[width=3.1in]{NCwAOstat200at-4circ}}
\subfigure[The binary system with a Lyot coronagraph.]{\label{m2} \includegraphics[width=3.1in]{LCwAOstat200at-4circ}}
\subfigure[The binary system with an apodized pupil Lyot coronagraph.]{\label{m3b} \includegraphics[width=3.1in]{APLCwAOstat200at-4circ}}
\subfigure[The binary system with a band-limited coronagraph.]{\label{m4b} \includegraphics[width=3.1in]{BLCwAOstat200at-4circ}}
\caption{Performance of the four coronagraphs in the presence of atmospheric aberrations, wavefront control, and static error. The intensity at each point is shown on a log scale. The companion in each image is at (-1000mas, 0mas). The vertical band in the center of panel (d) is obscuration from the mask and no companions can be detected along that axis.}
\end{center}
\end{figure}
\begin{figure}
\begin{center}
\includegraphics[width=3.1in]{aplcApodization}
\caption{The pupil of the apodized pupil Lyot coronagraph.}
\label{m3a}
\end{center}
\end{figure}
\begin{figure}
\begin{center}
\subfigure{
\includegraphics[width=3.1in]{blcImagePlane}
\includegraphics[width=3.1in]{blcLyotStop}
}
\caption{\emph{Left.} The mask for the band-limited coronagraph. \emph{Right.} The Lyot stop for the band-limited coronagraph.}
\label{m4a}
\end{center}
\end{figure}
We find that the BLC and the APLC are both effective at suppressing the central cores of the PSFs and the spiders, although the throughput of the planet is reduced in both cases. This reduced throughput is due to modifications to the system to eliminate spiders: the APLC has a throughput of $39.0\%$, with the reduction driven primarily by the apodizer, and the BLC has throughput of $35.36\%$, with the reduction driven primarily by the Lyot stop. This result is not unexpected; \mycitet{Cre07} showed that in lower-Strehl cases, hard-edged masks will perform similarly to apodized image-plane masks, and the reduction in the opening of the Lyot stop between \myfig{subaruPupil} and \myfig{m4a} suggests a corresponding reduction in throughput. As the apodizer is designed to concentrate the light into a central core, and the Lyot stop is not, the planet PSF is sharper for the APLC.
We note that rotation of the field will cause the spikes from the spiders to rotate, potentially obscuring the planet multiple times during the observation. However, the BLC and the APLC prove effective at suppressing the diffraction from the spiders given their design considerations; this will be a major advantage to using either method on an obstructed aperture.
\section{Manufacturing considerations for the APLC}
We can conceptualize the masks as being two circles deposited on two discs of glass, each large enough such that the mask can be moved to any part of the first image plane without the edge of the glass appearing in the final image. These discs would then be placed in a pair of X-Y actuated mounts so that the metal sides face each other, are separated by a very small amount, and bracket the image plane on either side. For observations of single stars, the masks can be moved to coincide and perform as a standard APLC, or one moved aside completely if there is sufficient room.
Other physical arrangements should be possible, as well. If the use of transmissive optics is excluded due to unwanted physical by-products such as ghosting or chromaticity, the masks can instead be etched to be freestanding with small spiders to hold them in place. These spiders could then be included in the eigenfunction calculations, or simply included in the simulations as error sources.
The placement of the masks near the image plane will need to be precise to mm-scale, so one of the primary challenges in building this system is likely to be optomechanical---ensuring both masks remain in planes parallel to the image plane without hitting each other or becoming defocused. The finite thickness of the masks is not likely to be a consideration, however, as existing freestanding image-plane masks can be sub-mm in thickness (e.g., 400$\mu$m for the ``bowtie'' masks used on the high-contrast testbed at Princeton \mycitep{Bel07}) as can coatings on coronagraph masks placed on glass (e.g., 200$\mu$m for 8th-order band-limited masks \mycitep{Lay05}).
In practice, it is better \emph{not} to optimize the APLC apodization for a specific tilt angle by creating an eigenfunction with two masks in the image plane. An optimized APLC apodized mask will marginally improve the coronagraphic performance for a binary system with a specific angular separation, but it would significantly degrade performance for stars that do not match to the apodization. This configuration would require the apodizer to be replaced for every system. Therefore, we believe it is better to compute the APLC apodization for an on-axis star and let Fourier optics solve the off-axis situation. There are a number of methods by which the pupil apodization can be created. For example, HEBS glass \mycitep{Boc08} and microdots \mycitep{Tho08, Mar09} are options being considered for APLCs on SPHERE and GPI, respectively, and shaped pupil APLCs offer another alternative for apodization \mycitep{Cad09b}.
One well-known shortcoming of the APLC is its chromaticity. This could be compensated using standard linear optimization techniques to create an apodization that concentrates light and minimizes sidelobes across a designated bandpass. Alternatively, the mask could simply be sized for the longest wavelength in the band, while shorter wavelengths create narrower PSFs whose cores are still blocked by the mask. This may still scatter light in undesirable ways in the image plane. A third option is the use of eigenvalues less than $\Lambda_0$, which may suppress more light off the central wavelength while still maintaining high suppression. This method is employed in the APLC design for GPI \mycitep{Mac08}. Some investigation would be required to determine which method is preferable, depending on the telescope arrangement.
\section{Tracking}
The majority of large telescopes today have obstructed apertures and alt-az mounts. Field rotation will cause the binary pair to rotate in the image plane, with each star maintaining a copy of the telescope PSF. There are two approaches that can be taken to compensate for the field rotation: simultaneous counter-rotation of the pupil and Lyot stop with the image rotator activated, and rotation of the image-plane masks with the image rotator deactivated. Pupil rotation has been tested previously on sky: the pupil-tracking mode of the NACO instrument on the VLT rotates the entire instrument to keep the pupil fixed \mycitep{Kas09a, Tut10}. While either is possible, rotating the image-plane masks may be preferred, as it minimizes the number of actuators required; the image-plane masks must be actuated regardless, to match the separation of the target binary.
The simplest way to place the masks in the image plane is with open-loop control--identify the image plane locations for the masks without the masks in, then move them into position and make science observations. A fast steering mirror may be used to provide coarse correction, though as there are two sources being blocked independently, at least one mask will still have to be adjusted. If this can be aligned with sufficient precision, the alignment procedure ends here.
If necessary, we suggest an additional closed-loop control to maintain mask alignment. Closed-loop control will also be necessary if a system is run with the image rotator off, to ensure the masks rotate with the image plane. One possible method of performing closed-loop control is to use flux in the image plane as a feedback signal. Decentering the masks will tend to increase flux sharply, as this is equivalent to introducing tilt errors, which increases the flux leaking through an APLC \mycitep{Siv08}. We note that pointing control using science camera imagery has already been demonstrated on-sky with a vector vortex mask, in the imaging of HR8799 from Palomar \mycitep{Ser10}, although the technique used to close the loop was not the same.
The total energy in the pupil plane is shown in \myfig{f2} for a number of sky angles in the vicinity of a mask. As shown, the region of suppression surrounding the correct alignment is distinct, though smeared by atmospheric errors. Improved AO correction will improve the sensitivity. Moving downward into this region would be relatively straightforward, if one mask is held fixed. This has the added advantage that no additional optics would be required, as the control uses feedback from the the imaging camera. Furthermore, we can iterate the positioning of the two masks if necessary, so only a single variable is used at a time.
We can then outline an observing procedure as follows: center the telescope between the two target stars, and obtain an estimate of the angular separation between them, and their angle relative to horizontal on the image plane. This would be done with the masks displaced from the vicinity of the stars; the astrometric measurement creates an initial placement for the two masks in the image plane. Closed-loop control, as outlined above, may proceed from there.
\section{Conclusions}
\label{sec:conclusion}
A sizable fraction of the overall planet population is expected to reside in binary systems; yet most current high-contrast facilities are not suitably equipped to observe such systems with coronagraphy. We have presented a conceptual design for a APLC-based coronagraph which would allow faint companions to be seen around binary stars. This design provides comparable throughput to observations using existing methods with band-limited coronagraphs, suppresses the diffraction spikes introduced by central obstructions and spiders, and allows observations without blocking edge-on systems. We also outline a control scheme and observing procedure which would allow the coronagraph to lock onto the stars. In particular, this would be a prime use for multi-object adaptive optics systems.
As a next step, we will examine mask tolerancing in simulation; a full closed-loop control system should be designed and closed-loop performance under noisy conditions should be simulated. We would also hope to verify the performance experimentally with a small testbed model, and show the masks can track effectively.
\section*{Acknowledgments}
The authors would like to thank Sebastian Egner for graciously providing simulated wavefront data, Laurent Pueyo for useful discussions, and R{\'e}mi Soummer for providing code for creating apodizers for APLCs. M.W.M. acknowledges support from NSF Astronomy \& Astrophysics Postdoctoral Fellowship under award AST-0901967.
|
\section{Introduction and theoretical motivations}
Heavy spin-1 resonances are a generic prediction of many Beyond-the-Standard Model (BSM) theories. The most frequently discussed case is that of new gauge bosons associated with extensions of the SM gauge group. While neutral states, known in the literature as $Z'$ \cite{Langacker:2008yv,Salvioni:2009p068,Salvioni:2010p010,Accomando:2010fz}, can be introduced by simply adding an extra $U(1)$ factor to the SM gauge group $G_{\mathrm{SM}}=SU(3)_{C}\times SU(2)_{L}\times U(1)_{Y}$, electrically charged states need a non-Abelian extension of the SM gauge symmetry. Some well-known examples of such extensions are those appearing in grand unified theories, including Left-Right (LR) models, in Little Higgs models, and in models where the Higgs is a pseudo-Goldstone boson arising from the spontaneous breaking of an extended global symmetry. A $W'$ can also appear as a Kaluza-Klein excitation of the $W$ in theories with extra dimensions.
On the other hand, there is also the interesting possibility that such heavy spin-1 particles are composite states, bound by a new strong interaction responsible for ElectroWeak Symmetry Breaking (EWSB). The most convenient approach in the discussion of the LHC reach on such composite resonances is to write the most general effective Lagrangian describing interactions of the new state with the SM fields and invariant under $G_{\mathrm{SM}}$ (see, e.g., Refs.~\cite{Bauer:2010p280,Barbieri:2011p2759,Han:2010p2696}). Once the representation in which the extra state transforms is specified, the Lagrangian is fully determined by a set of free parameters, namely the mass of the heavy state and its couplings to the SM particles. A specific gauge model, in which the vector is the gauge boson associated with the gauging of some extra symmetry, can then be recovered by taking some special values of these free parameters.
We apply an effective approach to study the early LHC phenomenology of a $W'$ transforming in the representation
\begin{equation} \label{111}
(\mathbf{1},\mathbf{1})_{1}
\end{equation}
of $G_{\mathrm{SM}}$, where the notation $(SU(3)_{c},SU(2)_{L})_{Y}$ has been adopted. A similar approach has been employed by the authors of Ref.~\cite{Aguila:2010p1781}, where however the focus was on computing constraints from electroweak data. In Ref.~\cite{Aguila:2010p1781}, bounds from ElectroWeak Precision Tests (EWPT) were discussed for all the irreducible representations of the SM gauge group which can have linear and renormalizable couplings to SM fields. There it was shown that the only such representations containing a color-singlet $W'$ (for a study of colored resonances at the early LHC, see Ref.~\cite{Han:2010p2696}) coupled to the SM fermions, in addition to that in Eq.~\eqref{111}, are $(\mathbf{1},\mathbf{3})_{0}$ and $(\mathbf{1},\mathbf{2})_{-3/2}$. The $(\mathbf{1},\mathbf{2})_{-3/2}$ multiplet does not have any renormalizable coupling to quarks or gluons, and as a consequence its production at the LHC would be very suppressed: therefore, we do not discuss it any further in the present work. Our choice to discuss the representation $(\mathbf{1},\mathbf{1})_{1}$ is motivated by the fact that in this case we can add to the SM only a charged resonance, without any associated neutral state. This is in contrast with the other representation commonly obtained in specific models, namely the $SU(2)_{L}$ triplet $(\mathbf{1},\mathbf{3})_{0}$. In the latter case, the $W'$ and $Z'$ masses are degenerate, apart from electroweak scale corrections, and as a result the strong bounds from neutral currents (including LEP2 data on four-fermion operators) apply also to the $W'$, pushing its mass well into the TeV range (and thus out of the LHC reach in its first run) unless its couplings to leptons are very small. On the other hand, a $W'$ transforming as $(\mathbf{1},\mathbf{1})_{1}$, because its only couplings to leptons arise through $W$-$W'$ mixing and are therefore strongly suppressed\footnote{Since we only consider the SM field content, we do not include right-handed neutrinos; or, equivalently for our purposes, we assume them to be heavier than the $W'$, so that the decay $W'\rightarrow \ell_{R}\nu^{\ell}_{R}$ is forbidden.}, is only constrained by hadronic processes (except for the oblique $T$ parameter). As we will discuss later, if particular forms for the right-handed quark mixing matrix are chosen as to evade constraints from $\Delta F=2$ transitions, the coupling of the $W'$ to quarks is only constrained by Tevatron direct searches, and therefore it can be sizable, without violating any existing constraint, even for a $W'$ mass below one TeV, making a discovery of the resonance at the early LHC possible. Furthermore, as discussed in Refs.~\cite{Hsieh:2010p2579,Schmaltz:2010p2610}, in Left-Right (LR) models, which give a $(\mathbf{1},\mathbf{1})_{1}$ charged state after LR symmetry breaking, the splitting between the masses of the $W'$ and $Z'$ (with the latter being a singlet under $G_{\mathrm{SM}}$) can be large, without violating EWPT constraints, if one takes $g_{X}\gg g_{R}$, where $g_{X}$ and $g_{R}$ are the couplings of the Abelian factor and of $SU(2)_{R}$, respectively. Also assigning the Higgs responsible for $SU(2)_{R}\times U(1)_{X}\rightarrow U(1)_{Y}$ breaking to a higher dimensional representation (for example, introducing a $SU(2)_{R}$ triplet Higgs) can help in increasing the mass splitting between the $W'$ and $Z'$. If such splitting is large enough, constraints from the $Z'$ can be made negligible, and one can study the phenomenology of the $W'$ using an effective theory for a $(\mathbf{1},\mathbf{1})_{1}$ state. Another example of a construction where the $W'$ we consider arises is the Littlest Higgs with custodial symmetry \cite{Chang:2003p057} (incidentally, we remark that several Little Higgs models contain in the spectrum a spin-1 $SU(2)_{L}$ triplet). While these provide specific examples of $W'$ that are described by the effective theory we consider, the interest of our approach goes much further, as it encompasses any composite state, whose properties could depart significantly from those of the gauge boson of a minimal non-abelian extension of $G_{\mathrm{SM}}$. We also note that a $W'$ with flavor-violating couplings to quarks has been invoked as an explanation of the anomaly in the top pair forward-backward asymmetry observed by CDF: we briefly comment on how such a $W'$ is described by our framework in Section~\ref{indirect bounds}. Composite vectors are usually considered in Higgsless models or in models where the Higgs is a composite state, where they have been shown to play an important role in keeping perturbative unitarity in the longitudinal $WW$ scattering up to the cut-off \cite{Csaki:2003dt,Barbieri:2008p1580}. The LHC phenomenology of these composite states is discussed, e.g., in Refs.~\cite{He:2008p3365, Barbieri:2010p144, Barbieri:2010p1577, CarcamoHernandez:2010p1578, Cata:2009p713,Birkedal:2004au,Martin:2009gi}.
We discuss the prospects of the early LHC to discover the $W'$ in the dijet channel, which, together with the $tb$ final state \cite{Gopalakrishna:2010p2723}, is the main avenue to look for the `leptophobic' $W'$ we are considering. A particularly striking difference between gauge models and the effective theory we consider is the presence in the latter case of a sizable $W'W\gamma$ interaction, which is very suppressed if the $W'$ is a fundamental gauge boson. As a consequence, observation of the $W'\rightarrow W\gamma$ decay at the LHC would be a hint of the compositeness of the resonance. In this light, we discuss the LHC prospects for discovery of the $W'\rightarrow W\gamma$ decay. We also present the prospects for observing the $W'\rightarrow WZ$ decay at the early LHC, and compare the reach in this channel to that in the $W\gamma$ final state. For previous relevant work on the phenomenology of a $W'$ at the LHC, see Refs.~\cite{Schmaltz:2010p2610,Frank:2010p2250,Rizzo:2007p3440,Nemevvsek:2011hz}. In Ref.~\cite{Schmaltz:2010p2610}, the early LHC reach on two simple $W'$ models was discussed. Our work differs from the discussion of a right-handed $W'$ in Ref.~\cite{Schmaltz:2010p2610} in two ways: firstly, as already detailed above we adopt an effective approach, without relying on any specific model; secondly, we make the `pessimistic' assumption that the decay of $W'$ into right-handed neutrinos, which was studied in Ref.~\cite{Schmaltz:2010p2610} (see also Ref.~\cite{Nemevvsek:2011hz}), be kinematically closed, and discuss the reach in the dijet and diboson final states.
Our paper is organized as follows: after introducing the effective Lagrangian in Section~\ref{model-indep approach}, we discuss bounds on the parameter space of the model coming from electroweak and low-energy data in Section~\ref{indirect bounds}, and from Tevatron searches in Section~\ref{Tevatron}. Section~\ref{secLHC} is devoted to the study of the early LHC reach on the $W'$ we are discussing: in Section~\ref{LHC} we present results for the dijet final state, in Section~\ref{LHCWgamma} we study the $W'\rightarrow W\gamma$ channel and we discuss how it could be used to obtain information on the theoretical nature of the resonance; the complementary search for $W'\rightarrow WZ$ is discussed in Section~\ref{WZsearch}. Finally, we present our conclusions in Section~\ref{summary}. Appendix~\ref{Partialdecaywidths} contains the partial decay widths of the $W'$, whereas in App.~\ref{EffLagr130} the effective Lagrangian for a $W'$ transforming as an $SU(2)_{L}$ triplet is given for completeness. In App.~\ref{minimalgaugemodels} we set the notation for the most economic gauge extensions of the SM containing either an iso-singlet or an iso-triplet $W'$.
\section{`Model independent' approach} \label{model-indep approach}
We consider, in addition to the SM field content, a complex spin-1 state transforming as a singlet under color and weak isospin, and with hypercharge equal to unity, according to Eq.~\eqref{111}. The extra vector is therefore electrically charged, with unit charge (we adopt a normalization for the hypercharge such that the electric charge is $Q=T_{3L}+Y$, where $T_{3L}$ is the third component of the weak isospin). We do not make any assumption on the theoretical origin of the extra state, and in particular we do not assume it to be a gauge boson associated with an extended gauge symmetry. Taking a \mbox{model-independent} approach, we write down all the renormalizable interactions between the new vector and the SM fields which are allowed by the $SU(3)_{c}\times SU(2)_{L}\times U(1)_{Y}$ gauge symmetry. Higher-Dimensional Operators (HDO) would be suppressed with respect to renormalizable ones by the cut-off of the theory; we neglect them in our analysis. We expect HDO to give corrections roughly of order $M^{2}_{W'}/\Lambda^{2}$ to our results: in Section~\ref{LHCWgamma} we show that the cut-off always satisfies $\Lambda\gtrsim 5M_{W'}$, so we can conservatively estimate our results to hold up to 10 percent corrections due to HDO. Within this framework, we write down the Lagrangian
\begin{equation} \label{lagrangian}
\mathcal{L}=\mathcal{L}_{SM}+\mathcal{L}_{V}+\mathcal{L}_{V-SM}\,,
\end{equation}
where $\mathcal{L}_{SM}$ is the SM Lagrangian, and\footnote{To be general, we should also include the operators $V_{\mu}^{+}V^{+\mu}V_{\nu}^{-}V^{-\nu}$ and $V_{\mu}^{+}V^{-\mu}V_{\nu}^{+}V^{-\nu}$; however, these operators only contribute to quartic interactions of vectors and can thus be neglected for the scope of this study. On the other hand, a cubic self-interaction of $V_{\mu}$ is forbidden by gauge invariance.}
\begin{align}
\mathcal{L}_{V}\,=\,& \nonumber D_{\mu}V_{\nu}^{-}D^{\nu}V^{+\mu}-D_{\mu}V_{\nu}^{-}D^{\mu}V^{+\nu}+\tilde{M}^{2}V^{+\mu}V_{\mu}^{-}\\
&+\frac{g_{4}^{2}}{2}|H|^{2}V^{+\mu}V_{\mu}^{-}-ig_{B}B^{\mu\nu}V^{+}_{\mu}V^{-}_{\nu}\,,\\ \label{V-SM}
\mathcal{L}_{V-SM}\,=\,& V^{+\mu}\left(ig_{H}H^{\dagger}(D_{\mu}\tilde{H})+ \frac{g_{q}}{\sqrt{2}}(V_{R})_{ij}\overline{u_{R}^{i}}\gamma_{\mu}d_{R}^{j}\right) +\text{h.c.}\,,
\end{align}
where we have denoted the extra state with $V^{\pm}_{\mu}$, and have defined $\tilde{H}\equiv i\sigma_{2}H^{\ast}$. We remark that we have not introduced right-handed neutrinos, in order to avoid making any further assumptions about the underlying model. The coupling of $V_{\mu}$ to left-handed fermionic currents is forbidden by gauge invariance. The covariant derivative is referred to the SM gauge group: for a generic field $\mathcal{X}$, neglecting colour we have
\begin{equation}
D_{\mu}\mathcal{X}=\partial_{\mu}\mathcal{X}-igT^{a}\hat{W}^{a}_{\mu}\mathcal{X}-ig'YB_{\mu}\mathcal{X}\,,
\end{equation}
where $T^{a}$ are the generators of the $SU(2)_{L}$ representation where $\mathcal{X}$ lives, and we have denoted the $SU(2)_{L}$ gauge bosons with a hat, to make explicit that they are gauge (and not mass) eigenstates. In fact, upon electroweak symmetry breaking the coupling $g_{H}$ generates a mass mixing between $\hat{W}_{\mu}^{\pm}$ and $V_{\mu}^{\pm}$. This mixing is rotated away by introducing mass eigenstates
\begin{equation}
\begin{pmatrix} W^{+}_{\mu} \\
W^{\prime\,+}_{\mu} \end{pmatrix} = \begin{pmatrix}
\cos\hat\theta\,\,\, & \sin\hat\theta \\
-\sin\hat\theta\,\,\, & \cos\hat\theta
\end{pmatrix} \begin{pmatrix}
\hat{W}_{\mu}^{+} \\
V_{\mu}^{+}
\end{pmatrix}\,.
\end{equation}
The expression of the mixing angle is
\begin{equation}
\tan(2\hat{\theta})=\frac{2\Delta^{2}}{m^{2}_{\hat{W}}-M^{2}}\,,
\end{equation}
where
\begin{equation}
m_{\hat{W}}^{2}=\frac{g^{2}v^{2}}{4},\quad \Delta^{2}=\frac{g_{H}gv^{2}}{2\sqrt{2}},\quad M^{2}=\tilde{M}^{2}+\frac{g_{4}^{2}v^{2}}{4}\,.
\end{equation}
We denote with $v\approx 246$ GeV the SM Higgs \textsc{vev}. We assume that Eq.~\eqref{lagrangian} is written in the mass eigenstate basis for fermions. We have written the heavy vector mass explicitly: the details of the mass generation mechanism will not affect our phenomenological study, as long as additional degrees of freedom possibly associated with such mechanism are heavy enough. We assume that the standard redefinition of the phases of the quark fields has already been done in $\mathcal{L}_{SM}$, thus leaving only one CP-violating phase in the Cabibbo--Kobayashi--Maskawa (CKM) mixing matrix $V_{CKM}$. The \mbox{right-handed} mixing matrix $V_{R}$ does not need to be unitary in the framework we adopt here: it is in general a complex $3\times 3$ matrix. This is a relevant difference with respect to LR models, where $V_{R}$ must be unitary, as a consequence of the gauging of $SU(2)_{R}$. We normalize $g_{q}$ in such a way that $|\det (V_{R})|=1$ (a generalization of this condition can be applied if $V_{R}$ has determinant zero).
In the mass eigenstate basis both for spin-$1/2$ and spin-1 fields, the charged current interactions for quarks read:
\begin{equation}
\mathcal{L}^{q}_{cc}\,=\,W^{+}_{\mu}\,\overline{u}^{i}\left(\gamma^{\mu}v_{ij}+\gamma^{\mu}\gamma_{5}a_{ij}\right)d^{j}+
W^{\prime\,+}_{\mu}\,\overline{u}^{i}\left(\gamma^{\mu}v^{\prime}_{ij}+\gamma^{\mu}\gamma_{5}a^{\prime}_{ij}\right)d^{j}+\text{h.c.}\,,
\end{equation}
where $u^{i}, d^{j}$ are Dirac fermions, and the couplings have the expressions
\begin{align} \nonumber
v_{ij}=& \frac{1}{2\sqrt{2}}\left(g_{q}\sin\hat\theta(V_{R})_{ij}+g\cos\hat\theta(V_{CKM})_{ij}\right)\,, \\
a_{ij}=& \frac{1}{2\sqrt{2}}\left(g_{q}\sin\hat\theta(V_{R})_{ij}-g\cos\hat\theta(V_{CKM})_{ij}\right)\,, \nonumber\\
v^{\prime}_{ij}=& \frac{1}{2\sqrt{2}}\left(g_{q}\cos\hat\theta(V_{R})_{ij}-g\sin\hat\theta(V_{CKM})_{ij}\right)\,, \nonumber\\
a^{\prime}_{ij}=& \frac{1}{2\sqrt{2}}\left(g_{q}\cos\hat\theta(V_{R})_{ij}+g\sin\hat\theta(V_{CKM})_{ij}\right)\,. \nonumber
\end{align}
We note that in general, $g_{H}$ is a complex parameter: for example, it is complex in LR models, see Eq.~\eqref{LRmodelcorrespondence}. However, the transformation $g_{H}\to g_{H}e^{-i\alpha}$ (with $\alpha$ an arbitrary phase) on the Lagrangian \eqref{lagrangian} only results, after diagonalization of $W$-$W'$ mixing, in $V_{R}\to e^{i\alpha}V_{R}$, therefore its effects are negligible for our scopes. Thus for simplicity we take $g_{H}$ to be real. The charged current interactions for leptons have the form
\begin{equation}
\mathcal{L}^{\ell}_{cc}\,=\,W^{+}_{\mu}\cos\hat\theta\frac{g}{\sqrt{2}}\overline{\nu}^{i}_{L}\gamma^{\mu}e^{i}_{L}-W^{\prime\,+}_{\mu}\sin\hat\theta \frac{g}{\sqrt{2}}\overline{\nu}^{i}_{L}\gamma^{\mu}e^{i}_{L}\,.
\end{equation}
The trilinear couplings involving the $W'$, the $W$ and the Higgs and the $W'$ and two SM gauge bosons read
\begin{subequations}
\begin{align} \nonumber
\mathcal{L}_{W'Wh}&=\Big[-\frac{1}{2}g^{2}vh\sin\hat\theta\cos\hat\theta+\frac{g_{H}g}{\sqrt{2}}vh(\cos^{2}\hat\theta -\sin^{2}\hat\theta)+\frac{g_{4}^{2}}{2}hv\sin\hat\theta\cos\hat\theta\, \Big] \\
&\hspace{4mm}\times (W^{+\,\mu}W_{\mu}^{\prime\,-}+W^{-\,\mu}W_{\mu}^{\prime\,+})\,, \\
\label{W'Wgamma-111}
\mathcal{L}_{W'W\gamma}&=-i\,e(c_{B}+1)\sin\hat\theta \cos\hat\theta F_{\mu\nu}(W^{+\,\mu}W^{\prime\,-\,\nu}+W^{\prime\,+\,\mu}W^{-\,\nu})\,, \\
\mathcal{L}_{W'WZ}&=\, i\sin\hat\theta \cos\hat\theta\Big[(g\cos\theta_{w}+g'\sin\theta_{w})(W^{-\,\mu}W^{\prime\,+}_{\nu\mu}+W^{\prime\, -\, \mu}W^{+}_{\nu\mu}- W^{\prime\,+\, \mu}W^{-}_{\nu\mu} \nonumber\\ \label{W'WZvertex}
&\hspace{4mm} -W^{+\,\mu} W^{\prime\,-}_{\nu\mu})Z^{\nu} -(g\cos\theta_{w}-g'\sin\theta_{w}c_{B})\left(W^{+\,\mu} W^{\prime\,-\,\nu}+ W^{\prime\,+\,\mu} W^{-\,\nu}\right) Z_{\mu\nu}\Big]\,,
\end{align}
\end{subequations}
where $\theta_{w}$ is the weak mixing angle. Partial widths for decays into two-body final states are collected in App.~\ref{Partialdecaywidths}.
In summary, in addition to the $W'$ mass, 4 couplings appear in our phenomenological Lagrangian: $g_{q}$, $g_{H}$ (or equivalently the mixing angle $\hat\theta$), $g_{B}$ and $g_{4}$. We find it useful to normalize $g_{B}$ to the SM hypercharge coupling, so we will refer to $c_{B}\equiv g_{B}/g'$ in what follows. Our phenomenological Lagrangian describes the low energy limit of a LR model\footnote{Here we are assuming the $Z'$ to be sufficiently heavier than the $W'$, and we are neglecting effects coming from a different scalar spectrum.} for the following values of the parameters (see App.~\ref{minimalgaugemodels}):
\begin{equation}
\begin{array}{llll} \label{LRmodelcorrespondence}
g=g_{L}\,, \qquad &\displaystyle g^{\prime}=\frac{g_{X}g_{R}}{\sqrt{g_{X}^{2}+g_{R}^{2}}}\,, \qquad &g_{q}=g_{R}\,, \qquad &\displaystyle g_{H}=-2\sqrt{2}g_{R}\frac{kk'e^{-i\alpha_{1}}}{v^{2}}\,,\\\\
c_{B}=-1\,,\qquad &\displaystyle g_{4}^{2}=2g_{R}^{2}\frac{k^{2}+k^{\prime\,2}}{v^{2}}\,, \qquad &\displaystyle \tilde{M}^{2}=\frac{g_{R}^{2}v_{R}^{2}}{4}\,,\qquad &v^{2}=2(k^{2}+k^{\prime\,2})\,.
\end{array}
\end{equation}
\section{Indirect bounds} \label{indirect bounds}
In this section we discuss indirect bounds on the couplings: $g_{q}$ is mainly constrained by $K$ and $B$ meson mixings, i.e. $\Delta F=2$ transitions (as we discuss in the next paragraph, the bounds are however strongly dependent on the structure of the right-handed mixing matrix $V_{R}$), while $\hat\theta$ is constrained by EWPT and by $u\rightarrow d$ and $u\rightarrow s$ transitions. The coupling $g_{B}$ is weakly constrained by Trilinear Gauge Couplings (TGC) measured at LEP, while $g_{4}$ is essentially unconstrained and marginal in our analysis (it only affects, and in a subleading way, the partial width for the decay $W'\rightarrow Wh$).
\subsection{Bounds on the coupling to quarks $g_{q}$ from $\Delta F=2$ processes}
The heavy charged vector we are considering, being coupled to right-handed quark currents, contributes in general to the $K_{L}$-$K_{S}$ mass difference via box diagrams. The experimental determination of $\Delta m_{K}$ thus gives a constraint on the mass $M_{W'}$ and on the coupling of the $W'$ to quarks $g_{q}$; however, the bound has evidently a strong dependence on the assumed form for the right-handed quark mixing matrix $V_{R}$ (we remark that $V_{R}$ does not need to be unitary in our effective approach). It was shown in Ref.~\cite{Langacker:1989p2578} that for some special choices of $V_{R}$ the constraint is weakened significantly (notice that the discussion of Ref.~\cite{Langacker:1989p2578} was performed in the context of LR models, so unitarity of $V_{R}$ was assumed). We choose for our phenomenological analysis the least constrained of these special forms, namely
\begin{equation} \label{RHmixing}
\left|V_{R}\right|=\mathbb{1}\,,
\end{equation}
for which the bound reads at $90 \%$ CL \cite{Langacker:1989p2578}
\begin{equation} \label{Klimit}
M_{W'}> \frac{g_{q}}{g}\,300\,\mathrm{GeV}\,.
\end{equation}
We note that in specific models, the bound can be much stronger: for example, if a discrete symmetry (P or C) relating the left and right sectors is imposed in LR models, then the bound reads approximately $M_{W'}>(2\,\text{--}\,3)\,\textrm{TeV}$ (see, e.g., Refs.~\cite{Zhang:2007da,Maiezza:2010ic}). This happens because the discrete symmetry forces $V_{R}$ to be close to $V_{CKM}$, implying a mixing of the order of the Cabibbo angle between the first two generations.
While mixing among the first two families is strongly constrained by $K_{L}$-$K_{S}$ data, we could consider the case where significant mixing between the first and third, or between the second and third families is present; we should accordingly take into account the constraints from $B$ meson physics, as discussed in Ref.~\cite{Frank:2010p3284}, where constraints on the elements of the right-handed mixing matrix from $b\rightarrow s\gamma$ and from $B^{0}_{d,s}$-$\overline{B}^{0}_{d,s}$ mixing were analysed in the context of a LR model. However, this goes beyond the scope of our work, so we simply take the form \eqref{RHmixing}, which automatically satisfies the constraints from $B$ meson mixing. The corresponding upper bound from $K$ mixing, Eq.~\eqref{Klimit}, is negligible with respect to the constraints coming from Tevatron direct searches (see Section~\ref{Tevatron}). Also notice that, as discussed in Ref.~\cite{Langacker:1989p2578}, this bound still holds if each $(V_{R})_{ij}$ is varied of $\epsilon=0.01$ from its central value, so that extreme fine tuning is avoided. For a study of the LHC phenomenology of a LR model with large off-diagonal $V_{R}$ elements, see Ref.~\cite{Frank:2010p2250}.
\subsubsection{Flavor-violating $W'$ as an explanation of the top $A_{FB}$ puzzle}
The `anomaly' observed by CDF in the forward-backward asymmetry of top pairs has recently drawn a lot of attention. The most recent measurement of $A^{t\overline{t}}_{FB}$ found a discrepancy of around $2\sigma$ with respect to the SM prediction \cite{CDFCollaboration:2010p3438}; furthermore, the asymmetry is observed to be larger in the region of large invariant mass of the $t\overline{t}$ pair, and in the region of large rapidity difference $|y_{t}-y_{\overline{t}}|$. The $t$-channel exchange of a $W'$ that only couples to $t$ and $d$ quarks was suggested in Ref.~\cite{Jung:2010p015004} as a possible explanation of the anomaly, and in Refs.~\cite{Cheung:2009p3324,Cheung:2011qa} it was shown that the observed asymmetry can be reproduced with the introduction of a right-handed $W'$ with mass in the range $200\,\text{--}\, 600$ GeV, and coupling $W'$-$t$-$d$ of magnitude $0.85\,\text{--}\, 2.1$. Similar values were chosen in Refs.~\cite{Shelton:2011p3339,Barger:2011ih}. Such $W'$ is described by our framework, where the right-handed mixing matrix does not need to be unitary, and as a consequence can accommodate a large $W'$-$t$-$d$ coupling, while having the remaining entries tuned to evade, e.g., the strong bounds coming from meson mixing.
\subsection{Bounds on the $W$-$W'$ mixing angle $\hat{\theta}$}
The main constraints on the $W$-$W'$ mixing angle $\hat\theta$ come from EWPT and from semileptonic $u\rightarrow d$ and $u\rightarrow s$ transitions. The $W$-$W'$ mixing term in \eqref{V-SM} breaks custodial symmetry, and is therefore strongly constrained by EWPT. A recent electroweak fit (including LEP2 data) performed in Ref.~\cite{Aguila:2010p1781} gives at $95\%$ CL
\begin{equation} \label{LEPbound}
\left|\frac{g_{H}}{M}\right|< 0.11\, \mathrm{TeV}^{-1}\,.
\end{equation}
We have checked that, as already remarked in Ref.~\cite{Aguila:2010p1781}, this constraint is essentially due to the negative contribution the $W$-$W'$ mixing gives to the $T$ parameter: the leading term in the $v^{2}/M^{2}$ expansion reads
\begin{equation} \label{Tcontribution}
\hat{T}_{V}=-\frac{\Delta^{4}}{M^{2}m_{\hat{W}}^{2}}\,.
\end{equation}
The LEP2 lower limit on the Higgs mass thus forces such a contribution to be very small. The bound \eqref{LEPbound} was in fact computed in Ref.~\cite{Aguila:2010p1781} leaving the Higgs mass as a free fit parameter, and including data from direct Higgs searches at LEP2. The results of our study depend very weakly on the mass of the Higgs, as long as it is light. We can translate Eq.~\eqref{LEPbound} into an upper bound on $\hat\theta$: the resulting limit becomes stronger when the mass of the $W'$ is increased, and varies from $|\hat\theta|\lesssim 4\times 10^{-3}$ for $M_{W'}=300\,\text{GeV}$ to $|\hat\theta|\lesssim 5\times 10^{-4}$ for $M_{W'}=2$ TeV.
A bound on the mixing angle $\hat\theta$ of different origin comes from the precise \mbox{low-energy} measurement of $u\rightarrow d$ and $u\rightarrow s$ transitions (i.e. from the measurements of the corresponding entries of the CKM matrix). Integrating out both the $W$ and the $W'$, we obtain a four-fermion effective Lagrangian that can be used to compute constraints from such measurements. The operators relevant to semileptonic processes, which give the strongest bounds, are
\begin{equation}
\mathcal{L}_{\text{eff}}=-\frac{4G_{F}}{\sqrt{2}}\,\overline{u}\,\gamma^{\mu}\Big[(1+\epsilon_{L})V_{CKM}P_{L}+\epsilon_{R}V_{R}P_{R}\Big]d\,(\overline{\ell}_{L}\gamma_{\mu}\nu^{\ell}_{L})+\text{h.c.}\,,
\end{equation}
where, neglecting $O(v^{4}/M^{4}_{W'})$ terms, $\epsilon_{L}=0$ and $\epsilon_{R}=g_{q}\,\hat\theta/g$. In Ref.~\cite{Buras:2010p3239} the bound $\epsilon_{R}\,\mathrm{Re}(V_{R}^{ud})=(0.1\pm 1.3)\times 10^{-3}$ was obtained, which assuming small CP phases implies at 95$\%$ CL
\begin{equation}
-2\times 10^{-3} < \epsilon_{R}V_{R}^{ud} < 3\times 10^{-3}\,.
\end{equation}
On the other hand, such bound is strongly relaxed if CP phases in $V_{R}$ are large: in the limit of maximal CP phases, only a milder second-order constraint survives, leading (assuming $V_{R}^{ud}\approx 1$) roughly to $|\epsilon_{R}|<10^{-(2\,\text{--}\,1)}$, as discussed in Ref.~\cite{Langacker:1989p2578}.
By making use of soft-pion theorems, constraints from nonleptonic processes such as $K\to 2\pi$ and $K\to 3\pi$ were also computed \cite{Donoghue:1982mx,Bigi:1981rr}. However, such bounds were obtained neglecting long distance chiral loop effects, which are known to be important and can offset tree-level results. Therefore, we do not consider such constraints in the following.
\subsection{Bounds from trilinear gauge couplings}
The $WWV_{0}$ vertex ($V_{0}=\gamma, Z$) can be described, assuming C- and P- conservation, by an effective Lagrangian containing 6 parameters (see for example Ref.~\cite{TripleGaugeCouplongsWorkingGroup:1996p2576}):
\begin{equation*}
\mathcal{L}_{\text{eff}}^{WWV_{0}}=ig_{WWV_{0}}\Big[g_{1}^{V_{0}}V_{0}^{\mu}(W^{-}_{\mu\nu}W^{+\nu}-W^{+}_{\mu\nu}W^{-\nu})+k_{V_{0}}W^{+}_{\mu}W_{\nu}^{-}V_{0}^{\mu\nu}+\frac{\lambda_{V_{0}}}{m^{2}_{W}}V_{0}^{\mu\nu}W_{\nu}^{+\rho}W^{-}_{\rho\mu}\Big]
\end{equation*}
where $g_{WW\gamma}=e,\,g_{WWZ}=g\cos\theta_{w}$, and the SM values of the parameters are given by $g_{1}^{\gamma,Z}=\kappa_{\gamma,Z}=1$ and $\lambda_{\gamma,Z}=0$. Assuming $SU(2)_{L}\times U(1)_{Y}$ gauge invariance reduces the number of independent parameters to three, which can be taken to be $\Delta g_{1}^{Z}\equiv g_{1}^{Z}-1$, $\Delta k_{\gamma}\equiv k_{\gamma}-1$, and $\lambda_{\gamma}$. In the case under discussion, the expressions of these parameters read
\begin{equation}
\Delta g_{1}^{Z}=-\sin^{2}\hat\theta(1+\tan^{2}\theta_{w})\,,\qquad \Delta k_{\gamma}=-\sin^{2}\hat\theta(1+c_{B})\,,\qquad \lambda_{\gamma}=0\,.
\end{equation}
Thus we can use the fits to LEP2 data performed by the LEP experiments \cite{DELPHICollaboration:2010p2577,ALEPHCollaboration:2004p0726,L3Collaboration:2002p151,L3Collaboration:2004p151,OPALCollaboration:2004p463} letting $\Delta g_{1}^{Z},\Delta k_{\gamma}$ free to vary while keeping fixed $\lambda_{\gamma}=0$, to constrain the values of our model parameters $(c_{B},\hat\theta)$. By combining this limit with the upper bound on the mixing angle $\hat\theta$ presented in the previous subsection, we can in principle constrain $c_{B}$. However, since as discussed above the mixing angle is required to be very small, in practice TGC constrain only extremely weakly the value of $c_{B}$. For example, using the analysis performed by the DELPHI Collaboration \cite{DELPHICollaboration:2010p2577}, we find that even considering a very large mixing angle $|\hat\theta| \sim 10^{-1}$, the wide range $-11 < c_{B}< 20$ (i.e. $-3.9< g_{B}< 7.1$) is allowed by TGC measurements at $95\%$ CL.
\section{Bounds from Tevatron direct searches} \label{Tevatron}
Data collected by the CDF and D0 experiments at the Tevatron in the dijet and $tb$ final states\footnote{By $tb$ we will always mean the sum $t\overline{b}+\overline{t}b$.} can be used to set an upper limit on the coupling to quarks of the $W'$ we are discussing as a function of its mass. In this section, we assume negligible $W$-$W'$ mixing, $\hat\theta \approx 0$, so the only relevant parameters are the $W'$ mass and the coupling $g_{q}$, and we obtain an upper bound on $g_{q}$ as a function of $M_{W'}$. If $W$-$W'$ mixing happens to be sizable, then the branching ratio into quarks is reduced, and the upper bound on $g_{q}$ gets relaxed accordingly. For instance, taking the relatively large value $\hat\theta = 10^{-2}$, the upper bound on $g_{q}$ is relaxed by approximately $10\%$ for $M_{W'}\gtrsim 1\,\mathrm{TeV}$, and less for lighter $W'$. The dependence of the ratio $\Gamma_{W'}/M_{W'}$ on the coupling $g_{q}$ is plotted in the left panel of Fig.~\ref{fig:totalwidth}, while the branching ratios as functions of $M_{W'}$ are shown in the right panel of the same figure, for representative values of the parameters.
Unless explicitly noted, we take the latest average value of the top mass, namely $m_{t}=173.3$ GeV \cite{TevatronElectroweakWorkingGroup:2010p3432}, and make use of the CTEQ6L set of parton distribution functions \cite{Pumplin:2002p3433}. Cross sections are computed using the CalcHEP matrix element generator \cite{Pukhov:1999p1261,Calchepweb}.
\begin{figure}[t]
\includegraphics[scale=0.34]{FIGURES/totalwidth.pdf}
\hspace{1mm}\includegraphics[scale=0.34]{FIGURES/BRplot.pdf}
\caption{\textsl{Left panel.} $W'$ width over mass ratio as a function of $g_{q}/g$ for negligible mixing, $\hat\theta \approx 0$, for $M_{W'}=300\,\mathrm{GeV}$ (dashed, red) and 1.5 TeV (blue). \textsl{Right panel.} Branching ratios of the $W'$ as a function of its mass, for the following choice of the remaining parameters: $g_{q}=g$, $\hat\theta=10^{-3}$, $c_{B}=-3$, $g_{4}=g$. From top to bottom: $ud$, $tb$, $WZ$, $Wh$, $W\gamma$, $\ell\nu$ (the latter includes all the three lepton families).}
\label{fig:totalwidth}
\end{figure}
\subsection{Dijet final state}
Searches for resonances in the invariant mass spectrum of dijet events at CDF and D0 are sensitive to the $W'$ we are discussing, which decays into quarks with branching ratio close to unity. The most recent dijet search, based on 1.13 fb$^{-1}$ of data, has been performed by the CDF collaboration \cite{CDFCollaboration:2009p2710}. Since no discrepancy with the SM prediction was observed, upper limits on the product $\sigma(p\overline{p}\rightarrow W'\rightarrow jj)\times \mathcal{A}$, where $\mathcal{A}$ is the geometrical acceptance for having both jets with $|y|<1$, have been set in Ref.~\cite{CDFCollaboration:2009p2710} for several types of resonance, including a $W'$. Therefore, we can compute $\sigma(p\overline{p}\rightarrow W'\rightarrow jj)\times \mathcal{A}$ using our phenomenological Lagrangian, and extract an upper bound on $g_{q}$ for each value of $M_{W'}$, which is reported in the left panel of Fig.~\ref{fig:Tevatronlimits}. We use cross sections at Leading Order (LO).
\begin{figure}[t]
\includegraphics[scale=0.34]{FIGURES/limitsTevatron.pdf}
\hspace{2.5mm}\includegraphics[scale=0.34]{FIGURES/comparisonLO-NLO.pdf}
\caption{\textit{Left panel}. The region of the ($M_{W'},g_{q}/g$) plane excluded at $95\%$ CL by Tevatron searches in the dijet final state (red region extending up to $1.4\,\mathrm{TeV}$) and $tb$ final state (blue region extending up to $950\,\mathrm{GeV}$). The dashed lines correspond to exclusion limits computed assuming $\sigma\propto g_{q}^{2}$, see text for details. Also shown in grey is the region excluded at $95\%$ CL by CMS dijet searches, see Section~\ref{LHC}. \textit{Right panel}. Upper limit from $tb$ searches on the $W'$ coupling to quarks as a function of $M_{W'}$, obtained using cross sections at LO (dashed) and at NLO (continuous) as reported in Ref.~\cite{Sullivan:2002p2617}. The scaling behaviour $\sigma\propto g_{q}^{2}$ was assumed.}
\label{fig:Tevatronlimits}
\end{figure}
The acceptance $\mathcal{A}$ is $36\%$ at $M_{W'}=300\, \mathrm{GeV}$, reaches a maximum of $51\%$ for $M_{W'}\sim 800$ GeV, and decreases for larger masses, being $34\%$ at $1.4$ TeV. The decreasing behavior of the acceptance at high resonance masses is due to a threshold effect: for a $W'$ mass around $1\, \mathrm{TeV}$ and above (that is, close to the kinematic limit of the Tevatron, which has a center of mass energy of $1.96\,\mathrm{TeV}$), the probability that the on-shell production condition $x_{1}x_{2}\approx M_{W'}^{2}/s$ is satisfied is so small that the off-shell contribution to the $p\overline{p}\rightarrow W'\rightarrow jj$ cross section becomes relevant, making the acceptance behave differently from what we would naively expect for an on-shell production mechanism. To make the relevance of this threshold effect more manifest, we also plot the upper bound on the coupling obtained by rescaling, for each value of $M_{W'}$, the value of $\sigma\times \mathcal{A}$ computed for $g_{q}=g$ according to the relation $\sigma\times \mathcal{A}\propto g_{q}^{2}\,$, which holds exactly in the Narrow Width Approximation (NWA). It is evident that while for masses below approximately 800 GeV the NWA (pure \mbox{on-shell} production) provides an excellent description of the dijet resonant production, for a larger mass of the resonance the NWA is not reliable anymore. Because of these relevant off-shell effects, we prefer to avoid using the notation $\sigma(W')\times \mathrm{BR}(W'\rightarrow jj)$.
The method we use to compute limits is valid for a resonance width smaller than the dijet energy resolution, which for the CDF experiment is of the order of $10\%$ of the dijet mass. The $W'$ we are studying has a width of $\sim 10\%$ of its mass for $g_{q}\sim2g$, as can be read off the left panel in Fig.~\ref{fig:totalwidth}; for larger couplings, the resonance width cannot be neglected, and the analysis would need to be corrected for this effect.
\subsection{$tb$ final state}
Another final state which is relevant to our model is $tb$. The CDF and D0 collaborations have searched for narrow resonances decaying into $tb$, with the $W$ coming from the top decaying into a lepton and missing transverse energy. The most recent search from CDF is based on 1.9 fb$^{-1}$ of data \cite{CDFCollaboration:2009p3125}, whereas D0 has carried out a similar analysis with 0.9 fb$^{-1}$ \cite{D0Collaboration:2008p2709}\footnote{The latest D0 analysis is based on 2.3 fb${}^{-1}$ of data \cite{Abazov:2011xs}. However, unfolding of the cross section limits, which is necessary to interpret them in our framework, was performed by the D0 collaboration only after completion of this work. Therefore the bounds from Ref.~\cite{Abazov:2011xs} are not included here. We expect that they will be slightly more stringent than those presented in Fig.~\ref{fig:Tevatronlimits}.}. Both analyses give as result upper limits on $\sigma(p\overline{p}\rightarrow W'\rightarrow tb)$, so we can compute the latter quantity using our phenomenological Lagrangian to extract an upper bound on $g_{q}$ for each value of the $W'$ mass. The strongest constraints are given by the CDF analysis, and in the left panel of Fig.~\ref{fig:Tevatronlimits} we compare them with the dijet limits discussed in the previous subsection. Analogously to what happened for the dijet final state, for $M_{W'}\gtrsim 800$ GeV threshold effects become relevant, and correspondingly the upper limit on $g_{q}$ computed assuming the NWA relation $\sigma\propto g_{q}^{2}$ differs from the correct limit.
\subsubsection{Comparison of LO and NLO $tb$ limits}
We have carried out our analysis at LO. However, for illustration purposes, it is useful to compare in Fig.~\ref{fig:Tevatronlimits} the upper bound on $g_{q}$ from $tb$ searches computed using cross sections at NLO and at LO, both as given in Ref.~\cite{Sullivan:2002p2617}. Since in the latter paper all cross sections were computed for $g_{q}=g$, we assume the relation $\sigma(p\overline{p}\rightarrow W'\rightarrow tb)\propto g_{q}^{2}$, which is exact in the NWA\footnote{We stress, however, that deviations from the NWA, and as a consequence from the $\sigma\propto g_{q}^{2}$ behaviour, arise due to off-shell effects for $M_{W'} \gtrsim 800$ GeV, as already discussed in the previous paragraph.}, to extract the upper bound on the coupling. Notice that the dependence on $M_{W'}$ of the NLO upper bound on $g_{q}$ differs significantly from those reported by the CDF and D0 collaborations in Refs.~\cite{CDFCollaboration:2009p3125} and \cite{D0Collaboration:2008p2709} respectively: this is due to the wrong assumption made there, that the cross section is proportional to the fourth power of the coupling. We remark that the upper limits shown in the right panel of Fig.~\ref{fig:Tevatronlimits} are obtained assuming $m_{t}=175$ GeV and $V_{R}=V_{CKM}$ (see Ref.~\cite{Sullivan:2002p2617}); even though the difference is at the level of a few percent, for the sake of consistency in what follows we will use the limits computed with $m_{t}=173.3$ GeV and $V_{R}=\mathbb{1}$, and reported in the left panel of Fig.~\ref{fig:Tevatronlimits}.
\section{LHC phenomenology} \label{secLHC}
In this section we discuss the reach of the early LHC on the composite $W'$ we are studying. We analyse first the prospects for discovery of the resonance as an excess of events in the dijet invariant mass spectrum, and subsequently move on to discuss decays into two gauge bosons. We study first the $W'\rightarrow W\gamma$ decay, which is of special interest since it is strongly suppressed in gauge models. As a consequence, its observation would be a hint of the compositeness of the $W'$. Finally, we discuss the $W'\rightarrow WZ$ channel.
\subsection{Dijet searches} \label{LHC}
The search for resonances in the dijet mass spectrum is one of the first new physics analyses performed by the CMS \cite{CMSCollaboration:2010p2595} and ATLAS \cite{ATLASCollaboration:2010p1746,ATLAS:2010Upd} experiments at the LHC, with an integrated luminosity of 2.9 and 3.1 pb${}^{-1}$ respectively at 7 TeV. Due to the very small data sample analysed so far, such searches are not competitive yet with those performed at the Tevatron: from Fig.~\ref{fig:Tevatronlimits} we see that only in a very narrow interval around $M_{W'}\sim 500$ GeV does the CMS search place a meaningful (even if weaker than the Tevatron one) upper limit on the $W'$ coupling to quarks. For larger masses, the CMS upper bound on the $W'$ cross section is saturated for values of the coupling $g_{q}>2g$, which implies that the width of the resonance is larger than the dijet mass resolution, and as a consequence the experimental analysis would need to be modified to account for a broad resonance. We use the CMS results because their limits were computed also for resonances decaying into a $qq$ final state, while the ATLAS analysis only assumes a resonance decaying into the final state $qg$, which leads to more radiation and as a consequence to a broader resonance shape, which has an effect on the cross section limits.
Future LHC analyses, however, will soon overtake the Tevatron results, so it is interesting to discuss the reach of dijet searches on the $W'$ we are considering. We assume the CMS kinematic cuts, namely on the pseudorapidity $|\eta|<2.5$ of each jet, and on the pseudorapidity difference $|\Delta\eta|<1.3$ \cite{CMSCollaboration:2010p2595}. For values of $M_{W'}$ between \mbox{300 GeV} and 2.6 TeV, in intervals of 100 GeV, we compute as a function of the coupling $g_{q}$ the integral of the signal differential invariant mass distribution $d\sigma_{S}/dM_{jj}$ over the region $M_{jj}>M_{W'}(1-\epsilon/2)$, and compare the result with the integral of the background distribution over the same range, to obtain $5\sigma$ discovery and 95$\%\, \mathrm{CL}$ exclusion contours in the $(M_{W'},g_{q}/g)$ plane. Here $\epsilon$ is the dijet mass resolution, which following Ref.~\cite{CMSCollaboration:2010p2595} we assume to vary from 8$\%$ at $M_{W'}=500$ GeV to 5$\%$ at $2.5\,\mathrm{TeV}$. The results are shown in Fig.~\ref{fig:LHCcouplingVSmass} for three different integrated luminosities, namely \mbox{$L=\int\mathcal{L}=0.1,1,5$ fb${}^{-1}$}, and for two LHC center of mass energies, namely 7 and 8 TeV\footnote{It has recently been decided that the LHC will run at 7 TeV in 2011. However, a higher energy for 2012 cannot be excluded at the time of writing \cite{Carli:2010zz}.}. We find that \mbox{100 pb${}^{-1}$} are not sufficient for a discovery, even at 8 TeV (except perhaps for a very small region around $M_{W'}=1\,\mathrm{TeV}$). On the other hand, if we focus on the exclusion contours, we see that the LHC can do better than the Tevatron already with 100 pb${}^{-1}$ for \mbox{$M_{W'}\gtrsim 700\,\text{GeV}$}, and for essentially all $W'$ masses if the luminosity is increased to $1\,\text{fb}^{-1}$.
We also report in Fig.~\ref{fig:LHCluminosityVSmass}, as a function of $M_{W'}$, the integrated luminosity needed for discovery or exclusion of a $W'$ with coupling to quarks equal to that of the SM $W$($g_{q}=g$), both for the 7 and 8 TeV LHC.
We choose to compare the integrals over $M_{jj}>M_{W'}(1-\epsilon/2)$ of the signal and background differential dijet mass distributions rather than their integrals in a finite interval centered on the $W'$ mass, because the former method is less sensitive to smearing effects generated by hadronization and jet reconstruction, which we cannot take into account in our parton-level analysis. In this way, we expect our estimate of the reach of the early LHC to be closer to the actual experimental results than it would be if we compared signal and background in an interval centered around the $W'$ mass.
\begin{figure}[t]
\includegraphics[scale=0.34]{FIGURES/LHC_discovery_coupling.pdf}
\hspace{2mm}\includegraphics[scale=0.33]{FIGURES/LHC_exclusion_coupling.pdf}
\caption{Contours in the $(M_{W'},g_{q}/g)$ plane for $5\sigma$ discovery (left) and $95\%$ CL exclusion (right) at the 7 TeV and 8 TeV LHC, for an integrated luminosity of $L=0.1,\,1$ and $5$ fb${}^{-1}$, corresponding to the continuous, dashed and dotted lines, respectively (for each different dashing, the upper, purple line is for $7$ TeV and the lower, green line is for $8$ TeV). Also shown are the Tevatron dijet (red) and $tb$ (blue) exclusions, together with the CMS exclusion with $2.9$ pb${}^{-1}$ (grey).}
\label{fig:LHCcouplingVSmass}
\end{figure}
\begin{figure}[t]
\includegraphics[scale=0.335]{FIGURES/LHC_discovery_mass.pdf}
\hspace{3mm}\includegraphics[scale=0.33]{FIGURES/LHC_exclusion_mass.pdf}
\caption{Integrated luminosity needed for $5\sigma$ discovery (left) and $95\%$ CL exclusion (right) as a function of the $W'$ mass, for the 7 TeV (continuous) and 8 TeV (dashed) LHC. The region shaded in grey, corresponding to $M_{W'}<913$ GeV, is excluded at $95\%$ CL by Tevatron searches.}
\label{fig:LHCluminosityVSmass}
\end{figure}
In addition to those in the dijet final state, also LHC searches in the $tb$ channel will be of course relevant to the $W'$ we are studying. We do not discuss them here, and refer the reader to the recent, extensive analysis of Ref.~\cite{Gopalakrishna:2010p2723}.
\subsection{Search for the $W'\rightarrow W\gamma$ decay} \label{LHCWgamma}
We now move on to consider decay channels of the $W'$ which have partial widths proportional to the $W$-$W'$ mixing angle $\hat{\theta}$. These include $WZ,\,Wh$ and $W\gamma$ final states. We will focus first on the last channel, which is of special interest since it is very suppressed in the gauge models containing a $(\mathbf{1},\mathbf{1})_{1}$ $W'$, such as for instance LR models. Therefore, observation of $W'\to W\gamma$ would point to a composite nature of the $W'$. The partial width for decay into $W\gamma$ reads
\begin{equation} \label{widthWgamma}
\Gamma(W'\rightarrow W\gamma)=\frac{e^{2}}{96\pi}(c_{B}+1)^{2}\sin^{2}\hat\theta\cos^{2}\hat\theta \left(1-\frac{M^{2}_{W}}{M^{2}_{W'}}\right)^{3}\left(1+\frac{M^{2}_{W'}}{M^{2}_{W}}\right)M_{W'}\,.
\end{equation}
Since the width for decay into this channel is controlled by $\hat\theta$ and $c_{B}$, it is interesting to estimate which values of these parameters will be accessible to the LHC in its first run. To assess the discovery potential, we choose two benchmark values for the $W'$ mass, namely 800 and 1200 GeV, and we assume two representative values of the integrated luminosity, namely 1 and 5 fb${}^{-1}$, at a center of mass energy of 7 TeV. We set the coupling to quarks to $g_{q}=0.84\,(1.48)g$ for $M_{W'}=800\,(1200)$ GeV, that is, to the largest value allowed by Tevatron $jj$ and $tb$ searches (see Fig.~\ref{fig:Tevatronlimits}). Notice that the upper limit on $g_{q}$ from Tevatron searches in quark final states was computed for $\hat{\theta}=0$; when the mixing is introduced, the bound on the coupling weakens, due to the smaller branching ratio of the resonance into quarks.
A direct constraint on the mixing angle $\hat\theta$ comes from the non-observation of resonances decaying into $WZ$ in a search performed by the D0 collaboration \cite{D0Collaboration:2010p3231}: we take such constraint into account in our analysis for the $W\gamma$ final state. On the other hand, the CDF Collaboration has performed a search in the $\ell\gammaE\llap{/\kern1.5pt}_T$ ($\ell=e,\mu$) final state \cite{CDFCollaboration:2007p3228}, without observing any discrepancies with the SM prediction. Also the constraints coming from this channel were taken into account; however, they turn out to be less stringent than those obtained from the $WZ$ channel, because of the smaller dataset analyzed.
We select decays of the $W$ into an electron and a neutrino, and apply the following cuts on the $e\gammaE\llap{/\kern1.5pt}_T$ final state: $p_{T}^{\gamma}>250\,(400)$ GeV, $p_{T}^{e}>50$ GeV, $E\llap{/\kern1.5pt}_T >50\,\,\textrm{GeV}$, $|\eta_{e,\gamma}|<2.5$, and $|M(W\gamma)-M_{W'}|<0.05\,(0.10)M_{W'}\,$, for $M_{W'}=800\,(1200)$ GeV. We note that, even though the neutrino longitudinal momentum $p_{z}^{\,\nu}$ is not measured experimentally, it can be reconstructed by imposing that the lepton and neutrino come from an on-shell $W$: a quadratic equation for $p_{z}^{\,\nu}$ is thus obtained. It follows that a criterion must be chosen to unfold this ambiguity. The assessment of the effects of such choice on the cuts on $E\llap{/\kern1.5pt}_T$ and on the total invariant mass $M(W\gamma)$ goes beyond the scope of this work, and we leave it to the experimental collaborations\footnote{In this regard, we also note that, at the detector level, fluctuations in the measured $E\llap{/\kern1.5pt}_T$ can lead to events where no solution for $p_{z}^{\,\nu}$ can be found even though the lepton and neutrino come from the decay of a $W$ (see, e.g., the section on top quark mass measurements in Ref.~\cite{ATLASCollaboration:2008p3435}).}. We neglect the interference between $W$ and $W'$, which is due to the $O(\hat\theta)$ coupling of $W'$ to left-handed quark currents. The main background process is the SM $W\gamma$ production, which we include in our analysis, while we leave out the $W+j$ production with the jet misidentified as a photon. We have checked that applying the rejection factor for misidentification into a $\gamma$ of very high-$p_{T}$ jets, which is of the order of $5\times 10^{3}$ if photon identification and isolation cuts are applied (see, e.g., Ref.~\cite{ATLASCollaboration:2008p3435}), the $W+j$ background contribution is roughly one order of magnitude smaller than the irreducible $W\gamma$ process. This estimate suffers from the fact that we are not including NLO corrections to $W+j$, and from the fact that requiring photon identification and isolation has an efficiency of $\sim80\%$ on `real' photons \cite{ATLASCollaboration:2008p3435}, which would slightly reduce the number of signal events detected. Other possibly relevant instrumental backgrounds that we do not include in our exploratory study are $eeE\llap{/\kern1.5pt}_T$ with $e$ misidentified as a photon, and QCD jets faking $e+E\llap{/\kern1.5pt}_T$. We leave the proper treatment of such detector-dependent backgrounds to the experimental analyses; we just note that doubling the statistics by including also the $W\rightarrow \mu\nu$ channel would help in balancing the sensitivity loss, in case the sum of instrumental backgrounds -- such as those mentioned above -- happened to be of the same order of magnitude of the irreducible $W\gamma$ background (for example, in the D0 $\ell\gammaE\llap{/\kern1.5pt}_T$ search, the total background was estimated to be roughly twice as large as the irreducible $W\gamma$, see Ref.~\cite{CDFCollaboration:2007p3228}).
In Fig.~\ref{fig:LHCWgammaDISTRIB}, we show the distributions of the reconstructed invariant mass of the $W'$ and of the $p_{T}$ of the photon, compared to the SM $W\gamma$ background. We stress that experimentally, reconstruction of the longitudinal component of the neutrino momentum by imposing the on-shell condition for the $W$ will have an impact on the resolution of the $W'$ invariant mass. From the invariant mass distribution, it is also evident that for the values of the parameters chosen, the $W'$ of mass 1.2 TeV has a quite large width, which motivated the use of a broader cut around the peak, as discussed above. While the number of events predicted at the early LHC is clearly small, these distributions can be used as a guideline also for searches at higher integrated luminositites, after rescaling cross section to higher LHC center of mass energy.
\begin{figure}[t]
\centering
\begin{minipage}{20cm}
\includegraphics[scale=0.36]{FIGURES/wgamma_minv.pdf}\hspace{-3mm}
\includegraphics[scale=0.36]{FIGURES/wgamma_pt_new.pdf}
\end{minipage}
\caption{Invariant mass (left) and photon $p_{T}$ (right) distributions for the \mbox{$W'\rightarrow W\gamma\rightarrow e\nu\gamma$} signal and for the irreducible background. The values of the couplings are as follows: \mbox{$g_{q}=0.84g$} and $\hat\theta = 10^{-2}$ for $M_{W'}=800\,\mathrm{GeV}$, and $g_{q}=1.48g$ and $\hat\theta =4\times 10^{-2}$ for $M_{W'}=1.2\,\mathrm{TeV}$.}
\label{fig:LHCWgammaDISTRIB}
\end{figure}
Our main results are shown in Fig.~\ref{fig:LHCWgamma}. As can be read off the left side of the figure, for $M_{W'}=800$ GeV, assuming $c_{B}=5$ (which corresponds to $g_{B}=5g'\sim 1.8$), the interval $5\times 10^{-3}<\hat\theta<1.25\times 10^{-2}$ is accessible for a discovery with 5 fb${}^{-1}$. Such values of $\hat\theta$, while being excluded by EWPT if we assume the $W'$ is the only new physics contributing to precision data (see the discussion after Eq.~\eqref{Tcontribution}), are however allowed by $u\rightarrow d$ and $u\rightarrow s$ transitions if the CP phases are not small. It is conceivable that a positive contribution to the $T$ parameter coming from additional new physics (such as, for example, a heavy neutral spin-1 state) relaxes the bound from EWPT, allowing for such relatively large values of $\hat\theta$. On the other hand, from the right side of Fig.~\ref{fig:LHCWgamma} we see that setting the mixing angle to the value $\hat\theta = 10^{-2}$, discovery of a $W'$ with mass \mbox{$800$ GeV} is possible with 5 fb$^{-1}$ for $c_{B}\gtrsim 2$, which corresponds to a moderate value of the coupling $g_{B}\sim 0.7$. The prospects for a heavier $M_{W'}=1200$ GeV are similar, except that in this case there is no relevant bound from Tevatron searches.
\begin{figure}[h]
\centering
\begin{minipage}{20cm}
\includegraphics[scale=0.37]{FIGURES/LHC7Theta800Plot5.pdf}\hspace{-5mm}
\includegraphics[scale=0.37]{FIGURES/LHC7cB800Plot.pdf} \\
\end{minipage}
\begin{minipage}{20cm}
\includegraphics[scale=0.37]{FIGURES/LHC7Theta1200Plot5.pdf} \hspace{-7mm}
\includegraphics[scale=0.37]{FIGURES/LHC7cB1200Plot.pdf}
\end{minipage}
\caption{`$5\sigma$' discovery prospects of the 7 TeV LHC for the $W'\rightarrow W\gamma\rightarrow e\nu \gamma$ process, for $M_{W'}=800$ GeV, $g_{q}=0.84g$ (top row) and $M_{W'}=1200$ GeV, $g_{q}=1.48g$ (bottom). $N_{\mathrm{events}}$ is the number of signal events after applying the cuts described in the text. The red curves show the expected number of events as a function of the parameters of our phenomenological Lagrangian, whereas the blue flat lines represent the number of events needed for a $5\sigma$ discovery, taking into account the SM background. The signal cross sections, after all cuts, are simply given by $\sigma_{S}=N_{\mathrm{events}}/L$; the background cross sections after all cuts are $\sigma_{B}\,(M_{W'}=800,1200\, \mathrm{GeV})= (9.6,\,2.7)\times 10^{-2}\,\mathrm{fb}$.
The region shaded in grey is excluded at $95\%$ CL by Tevatron searches for resonances decaying into $WZ$.}
\label{fig:LHCWgamma}
\end{figure}
For illustrative purposes, we also give in Table \ref{14TeV800GeV} an estimate of the sensitivity on $c_{B}$ for the 14 TeV LHC with 300 fb${}^{-1}$ luminosity. Background events are due to the irreducible SM $W\gamma$ process only. Cuts on the final state kinematics are the same as for the early LHC case discussed above.
\begin{table}[t]
\centering
\begin{tabular} {| c | c | c | c || c | c | c | c |}
\hline
$c_{B}+1$ & $N_{\mathrm{s}}$ & $N_{\mathrm{bckgr}}$ & $N_{\sigma}$ & $c_{B}+1$ & $N_{\mathrm{s}}$ & $N_{\mathrm{bckgr}}$ & $N_{\sigma}$ \\
\hline
$0.6$ & $57$ & $102$ & $5.7$ & $0.4$ & $34$ & $45$ & $5.0$ \\
\hline
$0.5$ & $40$ & $102$ & $4.0$ & $0.3$ & $23$ & $45$ & $3.4$ \\
\hline
$0.4$ & $26$ & $102$ & $2.6$ & $0.2$ & $9$ & $45$ & $1.5$ \\
\hline
\end{tabular}
\caption{Sensitivity on $c_{B}$ at the 14 TeV LHC with 300 fb${}^{-1}$, for $M_{W'}=800$ GeV, $g_{q}=0.84g$ and $\hat\theta = 10^{-2}$ (left), and for $M_{W'}=1.2$ TeV, $g_{q}=1.48g$ and $\hat\theta = 4\times 10^{-2}$ (right).}
\label{14TeV800GeV}
\end{table}
Clearly, it is very interesting to understand what are the predictions for the strength of the $W'W\gamma$ coupling in extensions of the SM. In Ref.~\cite{Ferrara:1992yc} it was shown that the gyromagnetic ratio of any elementary particle of mass $M$ (of any spin) coupled to the photon has to take the value $g=2$, which can be equivalently written as $c_{B}=-1$ in our effective language, in order for perturbative unitarity to be preserved up to energies $E\gg M/e$. As a consequence, in any gauge extension of the SM, where the $W'$ is the fundamental gauge boson of some extra symmetry, $g=2$ has to be expected, since unitarity is preserved up to much larger scales. Indeed, in the `minimal' gauge model containing an isosinglet $W'$, namely a LR model (see App.~\ref{minimalgaugemodels} for the notation), we find that $c_{B}=-1$ at the renormalizable level. Including dimension-6 operators, we expect $c_{B}=-1+O(v_{R}^{2}/\Lambda^{2})$, where $\Lambda$ is the cut-off of the LR model. Therefore, $c_{B}\approx -1$ will still hold, and observation of $W'\to W\gamma$ is likely to be out of the reach of the LHC.
On the other hand, if the $W'$ is a composite state of some new strong interaction, then the requirement of preservation of perturbative unitarity is relaxed, and significant departures from $c_{B}=-1$ can be envisaged. The only condition that needs to be satisfied even in the composite case is that the scale of violation of perturbative unitarity be sufficiently larger than the $W'$ mass. To verify that this is indeed the case, and since $c_{B}$ only appears in the $BVV$ vertex (see Eq.~\eqref{V-SM}), where $B$ is the hypercharge gauge boson and $V$ is the extra vector, we compute the amplitude for $BB\to VV$ scattering. The two independent amplitudes that grow the most with energy are $B_{+}B_{\pm}\to V_{L}V_{L}$, where $B_{\pm}$ are the two transverse polarizations of the $B$, and $V_{L}$ is the longitudinally polarized $V$. The leading term of these amplitudes in the high-energy limit reads
\begin{equation} \label{UVamplitudes}
A_{++\to LL}\approx \frac{(1-c_{B}^{2})g^{\prime\,2}s}{2M^{2}}\,,\qquad A_{+-\to LL}\approx \frac{(1+c_{B})^{2}g^{\prime\,2}s}{4M^{2}}\,.
\end{equation}
Notice that for $c_{B}\to -1$, the dangerous high-energy behavior is removed, as it was anticipated above. Requiring the amplitudes in Eq.~\eqref{UVamplitudes} not to exceed $16\pi^{2}$, we find the cut-off $\Lambda$ at which perturbative unitarity is lost\footnote{Other definitions of the perturbative unitarity bound are possible, and have been used in the literature. A different choice would simply change the numerical factors appearing in the definition of the cut-off.}, as a function of $c_{B}$: taking the maximum value we used in the phenomenological analysis, namely $c_{B}=10$, we find $\Lambda \approx 5 M$; for smaller values of $c_{B}$, the cut-off is obviously larger. This result guarantees that we can safely study the phenomenology at scale $M$ with relatively large values of $c_{B}$, without encountering any perturbative unitarity violation issues.
We conclude that, since the size of the $W'W\gamma$ coupling is expected to be very small if the $W'$ is a fundamental gauge boson, observation of $W'\to W\gamma$ at the LHC would be a hint of the composite nature of the $W'$.
\subsubsection{$W'\rightarrow W\gamma$ for a $W'$ belonging to an $SU(2)_{L}$ triplet}
It would be interesting to understand how the prospects in the search for \mbox{$W'\to W\gamma$} change if we consider a $W'$ transforming in the $(\mathbf{1},\mathbf{3})_{0}$ representation, which appears for example in some Little Higgs models and in models with large extra dimensions (the effective Lagrangian for such representation can be found in App.~\ref{EffLagr130}). Even though the $W'W\gamma$ interaction has the same structure for both the $(\mathbf{1},\mathbf{1})_{1}$ and the $(\mathbf{1},\mathbf{3})_{0}$ representations, see Eqs.~\eqref{W'Wgamma-111} and \eqref{W'Wgamma-130}, the results of our LHC study do not straightforwardly apply to the latter representation by just making the substitution \mbox{$c_{B}+1\,\to (c'_{B}+1)/(1-\widetilde{g}^{2})$}, because of $W$-$W'$ interference effects, which are potentially relevant for the isotriplet $W'$ (since it couples significantly to left-handed currents), and because of the different width (the triplet $W'$ also decays into light leptons). Furthermore, present constraints on the triplet $W'$ are different (and stronger) than those for the isosinglet $W'$ we consider in this paper. A detailed analysis of the $(\mathbf{1},\mathbf{3})_{0}$ $W'$ goes beyond the scope of this work.
From a theoretical perspective, in analogy with the isosinglet case, if the $SU(2)_{L}$ triplet $W'$ is a fundamental gauge boson (see App.~\ref{minimalgaugemodels} for the minimal gauge extension of the SM that contains such state), then $c'_{B}\approx -1$, and observation of the $W'\rightarrow W\gamma$ decay is likely to be out the LHC reach. On the other hand, if the $W'$ is a composite state, significant departures from $c'_{B}=-1$ are possible.
\subsection{Search for $W'\rightarrow WZ$} \label{WZsearch}
We also discuss the $W'\rightarrow WZ$ decay, which is complementary to $W'\rightarrow W\gamma$ because, being the rate for resonant $WZ$ production almost independent of the parameter $c_{B}$, its measurement would allow one to estimate the size of the mixing angle $\hat\theta$. Since we consider the early LHC reach, where integrated luminosity will be $\lesssim 5$ fb$^{-1}$, the most promising final state is $WZ\rightarrow \ell E\llap{/\kern1.5pt}_T jj$, which has a larger rate with respect to the purely leptonic channel; on the contrary, selecting leptonic decays of the $Z$ together with a hadronic $W$ has been shown to be less promising \cite{Alves:2009aa}.
Therefore, we implement simple cuts on the $e\nu jj$ final state (we only consider $W$ decays into an electron, in analogy to what we did for the $W'\rightarrow W\gamma$ process) to enhance the ratio of signal over background, namely: $p_{T}^{e,j}>50$ GeV, $E\llap{/\kern1.5pt}_T > 50$ GeV, $|\eta_{e,j}|<2.5$, and in addition we require the invariant mass of the dijet system to reconstruct a $Z$, $|M(jj)-M_{Z}|<20\,\mathrm{GeV}$. Finally, we select events which have an invariant mass compatible with $M_{W'}$ as follows: $|M(e\nu jj)-M_{W'}|<0.10M_{W'}\,$. The background we consider is the SM $pp\rightarrow e\nu jj$, which includes a large contribution from $W+jj$. The $t\overline{t}$ background can be efficiently reduced to roughly one order of magnitude less than the QCD background by applying a central jet veto \cite{Alves:2009aa}, and we do not consider it here. The invariant mass distributions of signal and background for this channel are shown in Fig.~\ref{fig:WZdistrib}.
Our results are shown in Fig.~\ref{fig:WZ} for the same choices of the $W'$ mass and couplings that we already discussed when studying $W'\rightarrow W\gamma$, so that a direct comparison between the two searches can be made. We can see that with 5 fb${}^{-1}$, a mixing angle larger than $\hat\theta\approx 3\,\text{--}\, 4\times 10^{-3}$ is accessible for discovery; this result is to a good approximation independent of the size of $c_{B}$. We also notice that the number of signal events can be sizable, which is the main reason why this channel is more favorable than the purely leptonic one for limited LHC luminosity. We remark that the size of the cut on the total invariant mass of $e\nu jj$ agrees with Ref.~\cite{Alves:2009aa}, where it was chosen to retain most of the signal in the presence of jet energy smearing. In addition, the cut we set on the invariant mass of the $jj$ system is even looser than the one adopted in Ref.~\cite{Alves:2009aa}. Therefore we believe our results to be reasonably stable with respect to jet smearing, which was not included in our parton-level analysis.
Finally, we do not discuss $W'$ decays into $Wh$, because the choice of the most relevant final states is strongly dependent on the Higgs mass, and such a detailed study goes beyond the scope of this work. We refer the interested reader to Refs.~\cite{Azuelos:2004dm,Bao:2011nh} and to the references cited therein.
\begin{figure}[t]
\centering
\includegraphics[scale=0.37]{FIGURES/wz_decay_minv.pdf}
\caption{Invariant mass distribution of the $e\nu jj$ system for the $W'$ signal and for the $pp\rightarrow e\nu jj$ background. The values of the coupling $g_{q}$ are the same as in Fig.~\ref{fig:LHCWgammaDISTRIB}.}
\label{fig:WZdistrib}
\end{figure}
\begin{figure}[t]
\centering
\begin{minipage}{20cm}
\includegraphics[scale=0.37]{FIGURES/wz_LHC7Theta800Plot_hadr.pdf}\hspace{-5mm}
\includegraphics[scale=0.37]{FIGURES/wz_LHC7Theta1200Plot5_hadr.pdf}
\end{minipage}
\caption{`$5\sigma$' discovery prospects on the mixing angle $\hat\theta$ via the $W'\rightarrow WZ\rightarrow e\nu jj$ process at the 7 TeV LHC, for $M_{W'}=800\,\mathrm{GeV}$, $g_{q}=0.84g$ (left) and $M_{W'}=1200\,\mathrm{GeV}$, $g_{q}=1.48g$ (right). The results are almost independent of $c_{B}$. The interpretation of the curves is analogous to Fig.~\ref{fig:LHCWgamma}; after all cuts, the background cross sections are $\sigma_{B}\,(M_{W'}=800,1200\, \mathrm{GeV})= (3.5,\,0.73)\,\mathrm{fb}$. The region shaded in grey is excluded at $95\%$ CL by Tevatron searches for resonances decaying into $WZ$.}
\label{fig:WZ}
\end{figure}
\section{Summary and conclusions} \label{summary}
We have applied an effective approach to study the phenomenology of a heavy $W'$ transforming as a singlet under weak isospin. Such a $W'$ is very weakly coupled to light leptons, and is therefore only constrained by $\Delta F=2$ hadronic processes (mainly $K$ and $B$ meson mixing) and Tevatron direct searches, provided the $W$-$W'$ mixing angle $\hat\theta$ is small enough to evade the important bounds from the oblique $T$ parameter and from precision measurements of $u\rightarrow d$ and $u\rightarrow s$ semileptonic transitions.\footnote{We remark that in the effective approach, the coupling to quarks and the mixing angle are independent parameters.} Furthermore, for suitable choices of the right-handed quark mixing matrix $V_{R}$, the only constraints on the coupling of the $W'$ to quarks come from Tevatron direct searches. Therefore, a $W'$ with mass even below a TeV and sizable coupling to quarks is allowed by present data. We have estimated the early LHC reach in the dijet channel on such a resonance. We have also noted that, if different choices for $V_{R}$ are made, our effective approach encompasses the class of $W'$ with flavor-violating couplings that has been recently called for as an explanation of the anomaly observed by CDF in the top pair forward-backward asymmetry.
Subsequently we have discussed the possibility that the $W$-$W'$ mixing angle be large enough to allow observation of the decays $W'\rightarrow W\gamma$ and $W'\to WZ$ at the early LHC. We have shown that discovery of these decays is possible for values of $\hat\theta$ allowed by semileptonic processes, if the CP phases in $V_{R}$ are not small. Although such values of $\hat\theta$ are excluded by EWPT because of the too large negative contribution the $W'$ gives to the $T$ parameter, it is conceivable that some additional new physics, such as for example an additional heavy neutral vector, could relax such constraint.
We have shown that the $W'\to W\gamma$ channel is of significant relevance to gain insight on the nature of the $W'$ after a discovery in the dijet (or $tb$) final state. We have compared the experimentally accessible values of the parameters $(c_{B},\hat\theta)$ to the prediction for the strength of the $W'W\gamma$ coupling both in weakly coupled gauge extensions of the SM, and in strongly interacting theories where the $W'$ is a composite state. We have shown that observation of $W'\rightarrow W\gamma$ at the early LHC would be a hint of the composite nature of the resonance. We also briefly commented on the relevance of the decay into $W\gamma$ in case the $W'$ belongs to a triplet under weak isospin, the other representation which is commonly encountered in BSM constructions.
\acknowledgments
We thank G.~Isidori, D.~Pappadopulo, G.~Villadoro and F.~Zwirner for discussions. We are grateful to M.~Narain for clarifications about Ref.~\cite{Abazov:2011xs}. This research has been partly supported by the European Commission under the contract ERC Advanced Grant 226371 {\it MassTeV} and the contract PITN-GA-2009-237920 {\it UNILHC}. The work of R.~T. has been supported in part by the European Commission under the contract PITN-GA-2010-264564 {\it LHCPhenoNet}. E.~S. was partly supported by the Fondazione Cariparo Excellence Grant \textit{String-derived supergravities with branes and fluxes and their phenomenological implications}.
|
\section*{Introduction}
The discreteness of the spectrum of the permitted velocity $v = p/m $ states because of the Bohr's quantization $pr = n\hbar $ goes down with the radius $r$ and mass $m$ increase, in agreement with the Bohr's correspondence principle. Therefore the quantum phenomena, such as the persistent current of electrons, can be observed in nano-rings with $r > 300 \ nm$ only at low temperature \cite{Nikulov01,Nikulov02} even at microscopic mass $m = 9 \ 10^{-31} \ kg$. Superconductivity is the exception thanks to the impossibility for Cooper pairs to change their quantum state $n$ individually \cite{Nikulov03}. This exception gives us an opportunity for more detailed research of phenomena connected with the Bohr's quantization: a ring, in contrast to atom, can be made in different forms.
\section {Periodical dependencies in magnetic field}
We can observe the change of the quantum number $n$ in magnetic field $B$ thanks to a "huge" radius $r > 500 \ nm$ of superconducting ring in comparison with atom orbit radius $r_{B} \approx 0.05 \ nm $. This change is observed because of the Aharonov-Bohm effect \cite{Nikulov01,Nikulov04}, i.e. the influence of the magnetic vector potential $A$ of the canonical momentum $p = mv + qA$ of a particle with charge $q$. The momentum of a free particle is described with the gradient $p = \hbar \bigtriangledown \varphi $ of the phase $\varphi $ of the wave function $\Psi = |\Psi |\exp{i\varphi }$, according to the quantum formalism. Because of the requirement that the complex wave function, describing a particle state, must be single-valued $\Psi = |\Psi |\exp i\varphi = |\Psi |\exp i(\varphi + n2\pi)$ at any point, the phase $\varphi $ must change by integral $n$ multiples of $2\pi $ following a complete turn along the path of integration, yielding $\oint_{l}dl \bigtriangledown \varphi = \oint_{l}dl p/\hbar = \oint_{l}dl (mv + qA)/ \hbar = m\oint_{l}dl v/ \hbar + 2\pi \Phi/\Phi_{0}= n2\pi $. The integer number $n$ of the permitted state with minimum energy $\propto v^{2} \propto (n - \Phi/\Phi_{0})^{2}$ should change with magnetic flux $ \Phi = BS \approx B\pi r^{2}$ at $ \Phi = (n+0.5)\Phi_{0}$. The flux quantum $\Phi _{0} = 2\pi \hbar /q $ equals $\approx 4140 \ T \ nm^{2}$ for electron $q = e$ and $\approx 2070 \ T \ nm^{2}$ for Cooper pair $q = 2e$. Consequently the $n$ change can be observed in magnetic field $B > \Phi_{0}/ \pi r^{2} \approx 0.0026 \ T $ in a superconducting ring with $r \approx 500 \ nm$ and in the inaccessibly high magnetic fields $\Phi_{0}/ \pi r_{B}^{2} \approx 530000 \ T$ for atom orbit with $r_{B} \approx 0.05 \ nm $.
The magnetic periodicity in different parameters, such as: resistance \cite{Nikulov04,Nikulov05,Nikulov06,Nikulov07}, magnetic susceptibility \cite{Nikulov08}, critical current \cite{Nikulov09,Nikulov10} and dc voltage \cite{Nikulov06,Nikulov07,Nikulov09} demonstrate unambiguously the $n$ change with the magnetic flux $ \Phi = BS $ inside ring: the resistance oscillations have minimum value at $ \Phi = n\Phi_{0}$ and maximum value at $ \Phi = (n+0.5)\Phi_{0}$, in accordance with the theoretical relation $\Delta R \propto v^{2} \propto (n - \Phi/\Phi_{0})^{2} $; magnetic susceptibility and dc voltage change the sign at $ \Phi = n\Phi_{0}$ and $ \Phi = (n+0.5)\Phi_{0}$, in accordance with the theoretical relation $M \propto V_{p} \propto \overline{v} \propto \overline {n - \Phi/\Phi_{0}}$; the critical current of symmetric ring have minimum value at $ \Phi = (n+0.5)\Phi_{0}$ and maximum value at $ \Phi = (n+0.5)\Phi_{0}$, in accordance with the relation $I_{c} = I_{c0} - 2I_{p}$, where $I_{p} = s2en_{s}v = I_{p,A}2(n - \Phi /\Phi _{0})$ is the persistent current circulating in the ring clockwise or anticlockwise \cite{Nikulov09}. No jump, which could be connected with the jump of the permitted velocity $v \propto n - \Phi/\Phi_{0}$ and the persistent current, because of the $n$ change at $ \Phi = (n+0.5)\Phi_{0}$ is observed at these measurements in accordance with the theory: the squared velocity $v^{2} \propto (n - \Phi/\Phi_{0})^{2}$ and the absolute value of the persistent current $|I_{p}| = I_{p,A}2|n - \Phi /\Phi _{0}| $ should not change with $n$ change from $n$ to $n+1$ at $ \Phi = (n+0.5)\Phi_{0}$; the average velocity at $\overline{v} \propto \overline {n - \Phi/\Phi_{0}} = (-0.5 + 0.5)/2 = 0$ at $ \Phi = (n+0.5)\Phi_{0}$.
\begin{figure}[]
\includegraphics{Fig1.eps}
\caption{\label{fig:epsart} The aluminum ring $r \approx 1 \mu m$ with asymmetric link-up of current leads, $l_{sh} \approx 0.7\pi r$, $l_{long} \approx 1.3\pi r$. }
\end{figure}
\begin{figure}[b]
\includegraphics{Fig2.eps}
\caption{\label{fig:epsart} The magnetic dependencies of the critical current $I_{c+}(\Phi /\Phi _{0})$ of aluminum ring with asymmetric link-up of current leads measured at the temperature $T \approx 0.900T_{c}$; $T \approx 0.933T_{c}$; $T_{c} \approx 1.52 \ K$. }
\end{figure}
\section {The jump absence at measurements of the critical current with different section of ring halves}
The jump could be expected at measurement of the oscillations of the critical current of a superconducting ring with different section of ring halves $ s_{w} > s_{n}$. The critical current measured in the one $I_{c+}$ or other $I_{c-}$ direction should be equal \cite{Nikulov09}
$$I_{c+}, I_{c-} = I_{c0} - I_{p,A}2|n - \frac{\Phi }{\Phi _{0}}|(1 + \frac{s_{w}}{s_{n}}) \eqno{(1a)}$$
when the external current $I_{ext}$ measuring $I_{c+}$, $I_{c-}$ is added with the persistent current $I_{p}$ in the narrow ring half $ s_{n}$and
$$I_{c+}, I_{c-} = I_{c0} - I_{p,A}2|n - \frac{\Phi }{\Phi _{0}}|(1 + \frac{s_{n}}{s_{w}}) \eqno{(1b)}$$
when the currents are added in the wide ring half $ s_{w}$. The jump $\Delta I_{c} = I_{p,A}( s_{w}/s_{n} - s_{n}/s_{w})$ should be observed because of the change of the persistent current direction with the $n$ change at $ \Phi = (n+0.5)\Phi_{0}$. But the measurements \cite{Nikulov09,Nikulov10,Nikulov11} have revealed the absence of the expected jump in spite of the enough large amplitude $ I_{p,A}$ of the persistent current and the relation $s_{w}/s_{n} \approx 2$. According to the universally recognised quantum formalism the anisotropy $I_{c,an}(\Phi /\Phi _{0}) = I_{c+}(\Phi /\Phi _{0}) - I_{c-}(\Phi /\Phi _{0})$ of the critical current measured in the opposite directions should appear in the asymmetric ring $s_{w}/s_{n} > 1$ because of the change of the functions $I_{c+}(\Phi /\Phi _{0})$, $I_{c-}(\Phi /\Phi _{0})$ describing its magnetic dependencies, see Fig.19 \cite{Nikulov09} and Fig.3 \cite{Nikulov11}. But it was found \cite{Nikulov10} that the anisotropy $I_{c,an}(\Phi /\Phi _{0}) = I_{c}(\Phi /\Phi _{0}+ \Delta \phi) - I_{c}(\Phi /\Phi _{0} - \Delta \phi)$ appears because of the changes in the arguments $ I_{c+}(\Phi /\Phi _{0}) \approx I_{c}(\Phi /\Phi _{0}+ \Delta \phi)$, $ I_{c-}(\Phi /\Phi _{0}) \approx I_{c}(\Phi /\Phi _{0}- \Delta \phi)$ of the functions rather than the functions themselves. The shift of the periodic dependencies of the critical current on the quarter of the flux quantum $\Delta \phi \approx 1/4$ \cite{Nikulov10} is very paradoxical phenomenon which can not be explained now as well as the absence of the $I_{c+}$, $I_{c-}$ jump with the $n$ change.
\section {Measurements of the critical current of superconducting rings with asymmetric link-up of current leads}
In order to clear up the essence of the second puzzle we have measured the magnetic dependencies of the critical current of aluminum ring with asymmetric link-up of current leads, Fig.1. The critical current should determined with the transition into the resistive state only of the shot segment $l_{sh} \approx 0.35 \times 2\pi r$ at the relation of the persistent current amplitude and the critical current $I_{p,A}/I_{c0} < 0.25$ measured on the ring, Fig.2, with the segment relation $l_{long}/l_{sh} \approx 0.65/0.35 \approx 1.8$. The critical current dependencies should describe with the relations
$$I_{c+} = I_{c0} - I_{p,A}(n - \frac{\Phi }{\Phi _{0}}) \eqno{(2a)}$$
$$I_{c-} = I_{c0} + I_{p,A}(n - \frac{\Phi }{\Phi _{0}}) \eqno{(2b)}$$
implying the jump $\Delta I_{c} = I_{p,A}$ at $ \Phi = (n+0.5)\Phi_{0}$, Fig.2. The measured dependencies $I_{c+}(\Phi /\Phi _{0})$, $I_{c-}(\Phi /\Phi _{0})$ correspond to the expected one (2) in the region near $ \Phi = n\Phi_{0}$, Fig.2. Nevertheless the jump, which must be because of the quantum number $n$ change according to (2), is not observed, Fig.2. The measured values of the critical current $I_{c+}(\Phi /\Phi _{0})$, $I_{c-}(\Phi /\Phi _{0})$ vary smoothly with magnetic flux value near $ \Phi = (n+0.5)\Phi_{0}$, Fig.2, in defiance of the quantum formalism predicting the $I_{c}$ jump (2). The aspiration of Nature to avoid the jump surprises. It is puzzle that the experimental and theoretical dependencies agree near $ \Phi \approx n\Phi_{0}$ and disagree near $ \Phi \approx (n+0.5)\Phi_{0}$.
|
\section{Introduction}
At high latitudes, the large-scale electric field,
$\vec{E}_{0}$, from the solar wind and Earth's magnetosphere
penetrates to the ionosphere. Across the
\emph{E}/upper\emph{D}-region altitudes, roughly between 80 and
130 km, the electrons are strongly magnetized, while the ions
become at least partially demagnetized due to frequent
collisions with the neutral atmosphere. This demagnetization
slows the drifting ions down, making the
$\vec{E}_{0}\times\vec{B}_{0}$ drift of electrons into strong
Hall currents named high-latitude electrojets. In the global
picture of magnetosphere-ionosphere (MI) coupling, this is the
region where field-aligned magnetospheric currents close and
dissipate energy.
Electrojet currents often drive plasma instabilities that
generate electrostatic field fluctuations coupled to plasma
density irregularities. These fluctuations have low
frequencies, usually smaller than the average frequencies of
electron and ion collisions with neutrals, $\nu_{e,i}$, while
the corresponding wavelengths in the unstable range exceed the
ion mean free path. Density irregularities, usually in the
range from tens of centimeters to tens of meters, are routinely
detected as strong coherent radar echoes
\citep[e.g.,][]{Cohen:Secondary67,Balsley:Radar71,Crochet:HF79,Kudeki:Condor87}.
Rocket flights through the lower ionosphere have detected also
electrostatic field fluctuations
\citep[e.g.,][]{Pfaff:Electric87b,Pfaff:ERRRIS,Pfaff:97a,Rose:92,Fukao:SEEK}.
These irregularities and fluctuations are caused by variety of
instabilities including the Farley-Buneman (FB)
\citep{Farley:Ejet63,Buneman:Ejet63}, gradient drift
\citep{Hoh:63,Maeda:Theoretical63}, and thermal instabilities
\citep{Dimant:95c,Dimant:Physical97,Kagan:thermal00,DimOppen2004:ionthermal1}.
The FB instability is excited when the relative velocity
between the average electron and ion streams exceeds the local
ion-acoustic speed. At high latitudes, this usually occurs when
$E_{0}\equiv |\vec{E}_{0}|$ exceeds the threshold value of
about $20$~mV/m. These and much stronger fields are not
uncommon in the sub-auroral, auroral, and polar cap areas,
especially during magnetospheric storms and substorms.
Small-scale fluctuations generated by these instabilities can
cause enormous anomalous electron heating (AEH). For about
thirty years, radars have observed strong electron temperature
elevations from 300-500~K up to more than 4000~K, correlating
with strong convection fields $\vec{E}_{0}$
\citep{Schlegel:81a,Providakes:88,Stauning:89,St.-Maurice:90a,Williams:1992,
Foster:Simultaneous00,Bahcivan:Plasma2007}. Simple estimates
show that regular ohmic heating by $\vec{E}_{0}$ alone cannot
account in full measure for such huge temperature elevations. A
strong correlation between AEH and $E_0$, as well as other
physical arguments, shows that average heating by FB-generated
turbulent electric fields causes AEH
\citep{StMaLaher:85,Robinson:Towards86,Robinson:92,Providakes:88,StMaurice:Unified87,
St.-Maurice:90b,DimantMilikh:JGR03}. AEH occurs largely because
the turbulent electrostatic field,
$\delta\vec{E}=-\nabla\delta\Phi$, has a small component
$\delta\vec{E}_{||}$ parallel to the geomagnetic field
$\vec{B}_0$
\citep{StMaLaher:85,Providakes:88,DimantMilikh:JGR03,MilikhDimant:JGR03,Bahcivan:Parallel2006}.
The importance of $\delta\vec{E}_{||}$ makes the entire process
fully 3-D.
Anomalous electron heating can modify ionospheric conductances
and hence affect the coupling between the magnetosphere and
ionosphere. Any electron heating directly affects the
temperature-dependent electron-neutral collision frequency and,
hence, the electron part of the Pedersen conductivity. This
part, however, is usually small compared to the electron Hall
and ion Pedersen conductivities. However, AEH causes a gradual
elevation of the mean plasma density within the anomalously
heated regions by reducing the local plasma recombination rate
\citep{Gurevich:78,St.-Maurice:90b,DimantMilikh:JGR03,MilikhGoncharenko:Anom2006}.
The AEH-induced plasma density elevations increase all
conductivities in proportion. However, this mechanism requires
tens of seconds or even minutes because of the slow development
of the ionization-recombination equilibrium. If $\vec{E}_0$
changes faster than the characteristic recombination timescale
then its time-averaged effect on density will be smoothed and
reduced.
\begin{figure}
\noindent\includegraphics[width=21pc]{NC_sketch}
\caption{Formation of a net non-linear current (NC) at a given wave of
plasma compression/decompression with the wavevector $\vec{k}$ in the
$\vec{E}_0\times \vec{B}_0$-direction. The wave electrostatic field,
$\delta \vec{E} \| \vec{k}$, has opposite directions in the plasma
density maxima ($\delta n>0$) and minima ($\delta n<0$), resulting
in the oppositely directed $\delta\vec{E}\times \vec{B}_0$ drifts
of magnetized electrons. More negatively charged particles move in
the $-\vec{E}_0$ than those in the opposite direction, resulting in
formation of the net positive current, $\vec{j}_{\mathrm{NC}}$,
parallel to $\vec{E}_0$.}
\label{Fig:NC_sketch}
\end{figure}
Plasma turbulence, however, can directly modify local
ionospheric conductivities via a wave-induced non-linear
current (NC) associated with plasma density irregularities
\citep{RogisterJamin:1975,Oppenheim:Evidence97,Buchert:Effect2006}.
The physical nature of the NC is explained in
Fig.~\ref{Fig:NC_sketch} for magnetized electrons and
unmagnetized ions. The wave field, $\delta\vec{E}$, has
different signs in the wave density maxima and minima, so that
the corresponding $\delta\vec{E}\times\vec{B}_{0}$-drifts of
magnetized electrons have opposite directions. As a result,
more electrons drift along the maxima than in the opposite
direction along the minima, producing a net non-linear current
density, $\vec{j}_{\mathrm{NC}}$. At higher \emph{E}-region
altitudes, partially magnetized ions can also contribute to
this current. Relative density perturbations in saturated FB
turbulence may reach at most tens percent, while the rms
turbulent field, $\langle\delta\vec{E}^{2}\rangle^{1/2}$, is
comparable to $E_{0}$ (the angular brackets here and below
denote spatial-temporal averaging). As a result, the total NC,
considered as a plasma response to the external electric field
$\vec{E}_{0}$, amounts to only a fraction of the regular
electrojet Hall current. However, in most of the electrojet the
NC is directed largely parallel to $\vec{E}_{0}$, so that it
may increase significantly the much smaller Pedersen
conductivity. This is critically important because the Pedersen
conductivity allows the MI field-aligned currents to close and
dissipate energy. The combined effect of the NC and AEH makes
the ionosphere less resistive. The anomalous conductance may,
at least partially, account for systematic overestimates of the
total cross-polar cap potential in global MHD models that
employ laminar conductivities
\citep[e.g.,][]{winglee:1997,raeder:1998,raeder.monograph:2001,Siscoe:Hill2002,
Ober:Testing2003,
Merkin:Global2005,Merkin:Anomalous2005,Merkin:Predicting2007,guild:2008a,wang.h:2008}.
Furthermore, the NC and AEH are intrinsically related. As
discussed in this paper, the work by the external DC electric
field on the total current equals the total field energy input
to ionosphere (total Joule heating), including that responsible
for AEH. On the other hand, the total frictional heating of
electrons and ions effectively increases wave dissipation and
reduces the turbulence intensity, thus affecting the NC.
The process of anomalous energy deposition from the
magnetosphere to ionosphere has been recently studied by
\citet{Buchert:Effect2006}. They showed that the apparent
average turbulent energy deposition per unit volume and time,
$\langle\delta\vec{E}\cdot\delta\vec{j}\rangle$, equals zero
(here $\delta\vec{j}$ is the fluctuation of the total current
density $\vec{j}$; a number of notations in this paper differ
from those of \citet{Buchert:Effect2006}). The authors deduced
that the total energy per unit volume and time lost by the
external field $\vec{E}_{0}$ to particles equals the total work
performed by this field on the total average current,
$\vec{j}=\vec{j}_{0}+\vec{j}_{\mathrm{NC}}$ ($\vec{j}_{0}$ is
the laminar current density). They suggested that the actual
energy input for the instability development and
turbulence-induced energy losses equals
$\vec{E}_{0}\cdot\vec{j}_{\mathrm{NC}}$. Their calculations
supported this remarkable conjecture, but provided no complete
proof since their quantitative analysis was based on an
oversimplified and restrictive model. First and most
importantly, they studied only 2-D turbulence in a
perpendicular to $\vec{B}_{0}$ plane. The authors did not state
the 2-D restriction explicitly, but it becomes evident from
their Eqs.~(17) and (18) which effectively exclude
$\vec{k}_{\parallel}$. Due to this, their treatment leaves out
the key 3-D energy conversion process primarily responsible for
AEH. Secondly, they employed the simplest two-fluid model with
fully demagnetized ions. This widely used approximation is good
for most of the electrojet but fails at its top part, right
where the Pedersen conductivity reaches its maximum. Thirdly,
their calculations of NC and other average terms were fully
based on a quasilinear, narrowband approximation for FB waves.
In this approximation, turbulence consists of relatively
small-amplitude waves which are described by linear relations
with narrowband wave frequencies. In many cases, this is a
reasonable approach, although its applicability is less
justified for the driving electric field well above the
instability threshold, when the effect of AEH is especially
strong.
These restrictions and some inconclusiveness associated with
them raise a number of important questions. First and foremost,
is the conjecture about $\vec{E}_{0}\cdot\vec{j}_{\mathrm{NC}}$
as the energy input for plasma turbulence really true? If so,
then this fundamental fact should follow directly from the
first principles and be universally applicable. In particular,
it has to be valid for arbitrarily magnetized particles and
strongly non-linear processes. Then, is it possible to deduce
this fact from a general viewpoint with no or minimal
approximations? Second, presuming that
$\vec{E}_{0}\cdot\vec{j}_{\mathrm{NC}}$ is the correct energy
input, then how is it compatible with the 3-D effect of AEH
largely associated with small $\delta\vec{E}_{\parallel}$? This
question arises because $\vec{j}_{\mathrm{NC}}$, as well as
$\vec{E}_{0}$, lies almost precisely in the plane perpendicular
to $\vec{B}_{0}$ with no average contribution from
$\vec{k}_{\parallel}$. Recently, we performed a number of 2-D
and 3-D\ particle-in-cell (PIC) simulations in big periodic
boxes with unprecedentedly dense meshes and large numbers of
PIC particles \citep{Oppenheim:Fully2011}. Comparing the 2-D
and 3-D results for the same background parameters, we have
obtained compelling evidence that anomalous electron heating
related to plasma structuring along $\vec{B}_{0}$ does exist.
Then, how could essentially 3-D heating originate from the 2-D
energy input? Further, how this energy input is distributed
between various groups of particles? This is important for
making accurate estimates of anomalous heating for different
particle groups. During strong magnetospheric events, what
feedback of developed \emph{E}-region turbulence on global MHD
behavior of the magnetosphere might be expected? How to
quantify this effect for including it in global MHD codes
intended for space weather predictions? Finally, what channels
provide the corresponding energy flow between the magnetosphere
and \emph{E}/\emph{D}-region ionosphere?
In this paper, we address these issues and create a rigorous
basis for calculating anomalous conductivities in the companion
paper \citep{Dimant:Magnetosphere2011}. We start by confirming
from first principles that fully saturated turbulence does
yield $\langle\delta\vec{E}\cdot\delta\vec{j}\rangle=0$, then
confirm the \citet{Buchert:Effect2006} deduction regarding the
turbulent energy input and establish its universal validity. In
order to quantitatively develop a 3-D model of AEH and resolve
the apparent contradiction between this interpretation and the
2-D nature of the energy input, we perform specific
calculations for the case of arbitrary particle magnetization,
using a quasilinear approximation. We calculate the non-linear
current, total energy input, and partial average frictional
heating sources for both electrons and ions in terms of a given
spectrum of density irregularities. We show that the major
quantitative difference between 2-D and 3-D developed
turbulence lies in the magnitude of density perturbations.
These perturbations and the non-linear current proportional to
them are noticeably larger in 3-D than in 2-D. This difference
explains the larger energy input in 3-D and is responsible for
AEH caused by turbulent fields, thus resolving the
above-mentioned contradiction.
The paper is organized as follows. In Sect.~\ref{General
consideration}, using only first principles with no
approximations, we confirm for the general case the
\citet{Buchert:Effect2006} findings regarding the energy input.
In Sect.~\ref{Partial energy deposit: quasilinear approach}, we
develop a quasilinear approach, similar to that of
\citet{Buchert:Effect2006}, but for the general 3-D case of
arbitrarily magnetized particles. This allows us to calculates
partial non-linear currents, relevant energy inputs, and
frictional heating sources. In Sect.~\ref{Global Energy Flow},
we discuss global energy flow between the magnetosphere and
ionosphere. In the appendix, we check the validity of the
conventional electrostatic approximation for lower-ionosphere
wave processes.
\section{Energy Conversion: First Principle Consideration\label{General consideration}}
In this section, we derive general relations regarding the
average energy input in quasi-periodic systems and show how the
\citet{Buchert:Effect2006} deductions follow from fundamental
electrodynamic and plasma kinetics principles with no
approximations like electrostatics, quasi-neutrality,
fluid-model description, etc.
First, we consider the evolution of the field energy in plasmas
by considering the exact electrodynamics, starting with
Ampere's and Faraday's laws,
\begin{subequations}
\label{Maxwell's
\begin{align}
\partial_{t}\vec{E} & =c^{2}\left( \nabla\times\vec{B}\right) -\frac
{\vec{j}}{\varepsilon_{0}},\label{Ampere's}\\
\partial_{t}\vec{B} & =-\nabla\times\vec{E}
\label{Faraday's}
\end{align}
\end{subequations}
where $\varepsilon_{0}$ is the permittivity of free space, $c$
is the speed of light in vacuum, $\vec{E}$ and $\vec{B}$ are
the electric field and magnetic induction, and $\vec{j}$ is the
total current density.
Taking scalar products of Eq.~(\ref{Ampere's}) with $\varepsilon_{0}\vec{E}$,
Eq.~(\ref{Faraday's}) with $\varepsilon_{0}c^{2}\vec{B
$ and adding the results, we obtain the standard energy balance
equation (aka Poynting's theorem):
\begin{equation}
\partial_{t}U+\nabla\cdot\vec{S}=-\vec{E}\cdot\vec{j}, \label{Poynting's
\end{equation}
where
\begin{equation}
U=\frac{\varepsilon_{0
}{2}\left( E^{2}+c^{2}B^{2}\right) ,\qquad\vec{S}\equiv\frac{\vec{E
\times\vec{B}}{\mu_{0}}
\label{US}
\end{equation}
are the field energy density and the corresponding flux (the Poynting vector),
respectively; $\mu_{0}=(\varepsilon_{0}c^{2})^{-1}$ is the permeability of
free space.
Now we need to find $\vec{j}$ from the plasma. The dynamics of
individual particles of type $s$, such as electrons or ions
($s=e,i$), is accurately described by Boltzmann's kinetic
equation, which can be written in the 6-D divergence form as
\begin{equation}
\partial_{t}f_{s}+\nabla\cdot\left( \vec{v}_{s}f_{s}\right)
+\partial_{\vec{v}_{s}}\cdot\left[ \frac{q_{s}}{m_{s}}(\vec
{E}+\vec{v}_{s}\times\vec{B})f_{s}\right] =S_{s}.
\label{Kinetic_alpha
\end{equation}
Here $\vec{v}_{s}$ is the kinetic velocity of the
particles-$s$, while $f_{s}(\vec{r},t,\vec{v}_{s})$ is their
single-particle velocity distribution function normalized to
the particle-$s$ density, $n_{s}\equiv\int f_{s}d^{3}v_{s}$
(the integration here and below is performed over the entire
3-D velocity space); $q_{s}$ and $m_{s}$ are the particle
charge and mass; $S_{s}$ is the collisional operator which
includes particle-$s$ collisions and can also include the
ionization sources and recombination losses; for simplicity, we
disregard effects of gravity. Multiplying
Eq.~(\ref{Kinetic_alpha}) by $m_{s}v_{s}^{2}/2$, integrating
over the velocity space, and adding the results for all plasma
particles, we obtain the energy balance relation for plasma,
\begin{equation}
\partial_{t}\sum_{s}\mathcal{E}_{s}+\nabla\cdot\sum_{s}\vec
{K}_{s}=\vec{E}\cdot\vec{j}+\sum_{s}L_{s}.
\label{particle_energy _conserva
\end{equation}
Here $\mathcal{E}_{s}\equiv\int(m_{s}v_{s}^{2}/2)\ f_{s
}d^{3}v_{s}\ $and $\vec{K}_{s}\equiv\int(m_{s}v_{s
^{2}/2)\vec{v}_{s}f_{s}d^{3}v_{s}$ are the average particle
energy and energy-flux densities, respectively; $L_{s}\equiv
\int(m_{s}v_{s}^{2}/2)\allowbreak S_{s}d^{3}v_{s}$
combines all collisional energy gains and losses; $\vec{j}\equiv\sum_{s
}q_{s}n_{s}\vec{V}_{s}$, is the total current density, the same
as in Eq.~(\ref{Poynting's}); $\vec{V}_{s}\equiv\int\vec{v}_{s
}f_{s}d^{3}v_{s}/n_{s}$ is the particle-$s$ mean fluid
velocity. In the general case, the particle energy density $\mathcal{E
_{s}$ combines the mean thermal energy, $3n_{s}T_{s}/2$, where
$T_{s}\equiv\int[m_{s}(v_{s}-\vec{V}_{s})^{2}/3]f_{s
}d^{3}v_{s}/n_{s}$ is the effective particle temperature, with the
kinetic energy density of the mean particle flow, $n_{s}m_{s
}V_{s}^{2}/2$.
Comparison of Eqs.~(\ref{Poynting's}) and (\ref{particle_energy
_conserva}) shows that $\vec{E}\cdot\vec{j}$ is the total
energy input from the fields to particles per unit volume and
time. The energy deposited in an individual group of particles
at a given location may be then slightly redistributed via
Coulomb collisions and transported to other locations.
Eventually, this energy becomes lost to the abundant neutral
atmosphere via predominantly inelastic plasma-neutral
collisions.
General Eqs.~(\ref{Poynting's}) and (\ref{particle_energy
_conserva}) apply to all plasma processes, linear or
non-linear. To specifically discuss turbulent processes, we
need to separate them from the slowly evolving and large-scale
regular background structures and processes. This can be done
by using a conventional two-scale approach in which macroscopic
background structures and processes are presumed to have much
longer characteristic spatial and temporal scales than does the
turbulence. The physical conditions in the \emph{E}-region
plasma make this procedure more applicable to plasma
irregularities generated by local instability mechanisms, such
as the Farley-Buneman (FB) and thermal-driven instabilities
\citep{Dimant:Physical97,Kagan:thermal00,DimOppen2004:ionthermal1}
than for relatively large-scale irregularities generated by
plasma gradients.
We assume that plasma turbulence generated by instabilities
consists mostly of waves whose wavelengths $\lambda_{i}$ in
each direction $i$ are much smaller than the corresponding
typical scales of spatial variation of the macroscopic
background parameters, $\Lambda_{i}$. Then, for any given
location $\vec{r}$, we can build an imaginary rectilinear box
centered around $\vec{r}$ with the sizes $L_{i}$ that satisfy
the conditions $\lambda_{i}\ll L_{i}\ll\Lambda_{i}$. The
characteristic scales in different directions $i$ can differ
dramatically. For example, the wavelengths of nearly
field-aligned irregularities parallel to $\vec{B}_{0}$ are
generally orders of magnitude larger that those perpendicular
to $\vec{B}_{0}$, so that the corresponding box sizes would
also be vastly different. Within the imaginary box, we can
extend the background parameters from the box center $\vec{r}$
uniformly to the entire box and impose periodic boundary
conditions. This approach is widely employed in computer
simulations.
We can similarly describe the temporal evolution of the process
by presuming long-lasting non-linearly saturated turbulence in
a quasi-stationary background. This implies a continuous
driving of instability and allows one to introduce a
quasi-period $T$, analogous to $L_{i}$, which is much longer
than typical timescales of turbulent variations but is much
shorter than a characteristic timescale of slow evolution of
the macroscopic background.
To implement the two-scale procedure, we introduce the
following convention. We denote the short-scale and fast
periodic variables for turbulent processes by
$\vec{x}=(x_{1},x_{2},x_{3})$ and $t$, while denoting the
corresponding coordinates and time of large-scale and slow
background variations by $\vec{r}$ and $\tau$, respectively.
With respect to the periodic coordinates $\vec{x}$ and time
$t$, we define a spatial-temporal average as
\begin{equation}
\left\langle \cdots\right\rangle \equiv\frac{1}{L_{1}L_{2}L_{3}T}\!\int
_{-L_{1}/2}^{L_{1}/2}\!\!\!dx_{1}\!\!\int_{-L_{2}/2}^{L_{2}/2}\!\!\!dx_{2}\!
\int_{-L_{3}
/2}^{L_{3}/2}\!\!\!dx_{3}\!\int_{\tau-T/2}^{\tau+T/2}
\!\!\!\!\!\!\!\!\left(\cdots\right)
dt.
\label{spatial_average
\end{equation}
In general, such averages can remain functions of the slow variables $\vec{r}$ and
$\tau$. This description of turbulence does not require the ensemble
averaging employed by \citet{Buchert:Effect2006}.
Having implemented this procedure, we split the quasi-periodic
fields and currents into their $\vec{r}$,$\tau$-dependent
average parts and the corresponding local periodic
perturbations,
\begin{equation}
\vec{E}=\langle\vec{E}\rangle+\delta\vec{E},\qquad\vec{B}=\langle\vec
{B}\rangle+\delta\vec{B},\qquad\vec{j}=\langle\vec{j}\rangle+\delta\vec{j},
\label{perturbations
\end{equation}
Using $\langle\delta\vec{E}\rangle=\langle\delta\vec{j}\rangle=0$, we obtain
for the average energy deposition ter
\begin{equation}
\langle\vec{E}\cdot\vec{j}\rangle=\langle\vec{E}\rangle\cdot\langle\vec
{j}\rangle+\langle\delta\vec{E}\cdot\delta\vec{j}\rangle. \label{split_EJ}
\end{equation}
One might naively interpret the first term on the right-hand
side (RHS) of Eq.~(\ref{split_EJ}) as the energy input from the
background fields to the undisturbed plasma, while the second
term as the corresponding average contribution from the
turbulent fields. As mentioned in the Introduction,
\citet{Buchert:Effect2006} found, using a simplified model,
that $\langle\delta\vec{E}\cdot\delta\vec{j}\rangle$ equals
zero by the following formal mathematical reason. If one
expands this term further in terms of the mean particle fluid
velocities to the lowest-order quadratic non-linearity then the
expected total turbulent frictional heating term,
$\sum_{s}q_{s}n_{s0}\langle\delta\vec{E}\cdot\delta\vec
{V}_{s}\rangle$, turns out to be automatically canceled by a
density perturbation term, $\sum_{s}q_{s}\langle\delta
n_{s}\delta\vec{E} \rangle\cdot\vec{V}_{s0}$ (here $\delta
n_{s}$ and $\delta\vec {V}_{s}$ are the perturbations of the
densities and particle fluid velocities). Right below we show
that for purely periodic and spatially homogeneous turbulence
the equality $\langle\delta\vec{E} \cdot\delta\vec{j}\rangle=0$
is a natural and universal constraint, required merely by the
imposed periodicity. This constraint follows directly from
exact Maxwell's equations without invoking specific plasma
models.
Indeed, Maxwell's equations are linear, allowing the separation
of the average quantities from wave perturbations. The
perturbations depend on the small-scale, quasi-periodic
variables, $\vec{x}$, $t$, and can also have an adiabatically
slow $\vec{r}$, $\tau$-dependence due to background
inhomogeneities and evolution. In principle, the inhomogeneous
background parameters may evolve in such a way that an
instability threshold is crossed, resulting in sudden onset or
disappearance of instability in some locations at certain
moments of time. Such instances, analogous to second-order
phase transitions \citep{LandauStatistical_1_2000}, break the
validity of our two-scale approximation and deserve a special
treatment that lies beyond the framework of this paper. Apart
from these special occasions, we can apply to the perturbations
of the fields and current the same steps that lead to
Eq.~(\ref{Poynting's}). Averaging the result in accord with the
definition of Eq.~(\ref{spatial_average}), we obtain
\begin{equation}
\partial_{\tau}\left[ \frac{\varepsilon_{0}\left( \left\langle \delta
E^{2}\right\rangle +c^{2}\left\langle \delta B^{2}\right\rangle \right)}
{2}\right] +\nabla_{\vec{r}}\cdot\frac{\langle\delta\vec{E}\times\delta
\vec{B}\rangle}{\mu_{0}}=-\langle\delta\vec{E}\cdot\delta\vec{j}\rangle.
\label{Poynting's_perturb
\end{equation}
Then for saturated turbulence with constant average
characteristics in a strictly periodic box, the left-hand side
(LHS) of Eq.~(\ref{Poynting's_perturb}) disappears, yielding
\begin{equation}
\langle\delta\vec{E}\cdot\delta\vec{j}\rangle=0, \label{so_that_really
\end{equation}
regardless of the specific fluid or kinetic models employed for
the plasma description. For strictly periodic processes, this
result is exact. One can also obtain it directly from the
discrete Fourier harmonics of the electric field and current.
We do this right below and also demonstrate that electrostatic
and quasi-neutral approximations do not lift this exact
electrodynamic constraint.
We start by introducing spatial and temporal Fourier harmonics
of wave perturbations for general, linear or non-linear,
periodic processes. Since we imply saturated turbulence in a
4-D box (3-D space + time) with periodic boundary conditions,
it is logical to employ the 4-D discrete Fourier
transformation. For easy reading, we assign to the Fourier
transforms the notations of the original variables but with
additional subscripts $\vec {k},\omega$. For a scalar or vector
periodic perturbation $\delta F(\vec {x},t)$ as a function of
coordinates $x_{i}$ and time $t$, we define such
transformations within the 4-D box of the corresponding sizes
$L_{i}$ and the quasi-period $T$ as
\begin{subequations}
\label{Fourier_transform
\begin{align}
&\delta F(\vec{x},t) =\sum_{\vec{k},\omega\neq0}\delta F_{\vec{k},\omega
}\exp[i(\vec{k}\cdot\vec{x}-\omega t)],\label{Function}\\
&\delta F_{\vec{k},\omega} =\frac{1}{L_{1}L_{2}L_{3}T}\!\!\int\!\!\delta F(\vec
{x},t)\exp\!\left[-i\!\left(\sum_{i=1}^3k_{i}x_{i}-\omega t\right)\!\right]\!\!
d^{3}x_{i}dt
\label{Transform}
\end{align}
\end{subequations}
Here the $\vec{k},\omega\neq0$ summation is taken over all
discrete values of $\vec{k}$, $k_{i}=2\pi n_{i}/L_{i}$,
$\omega=2\pi m/T$, with integer $n_{i}$, $m$, while the
integration is performed over the entire 4-D box, as in
Eq.~(\ref{spatial_average}). Since we apply the discrete
Fourier transforms only to the perturbations, we exclude from
the summation over harmonics the average values corresponding
to $\vec{k},\omega =0$. In terms of their Fourier harmonics,
$\delta A_{\vec{k},\omega}$ and $\delta B_{\vec{k},\omega}$,
the spatial-temporal average of any product, scalar or vector,
of two functions $A(\vec{x},t)$ and $B(\vec{x},t)$, is given by
\begin{equation}
\left\langle \delta A(\vec{x},t)\delta B(\vec{x},t)\right\rangle =\sum
_{\vec{k},\omega\neq0}\delta A_{\vec{k},\omega}\delta B_{\vec{k},\omega
^{\ast}=\sum_{\vec{k},\omega\neq0}\delta A_{\vec{k},\omega}^{\ast}\delta
B_{\vec{k},\omega}.\label{Average_harmonics
\end{equation}
Our current discrete-transform normalization differs from the
continuous one in \citet{Buchert:Effect2006} and allows to
avoid the emergence of extraneous factors like $VT$
($V=L_1L_2L_3$) and $(2\pi)^4$ in the explicit physical
expressions for average quadratically non-linear quantities
\citep[e.g.,][Eqs.~(6)--(8), (19) --(26),
etc.]{Buchert:Effect2006}.
Using Fourier transforms, we now prove
Eq.~(\ref{so_that_really}) in a more direct way. Applying
Eq.~(\ref{Fourier_transform}) to Eq.~(\ref{Maxwell's}), for a
given Fourier harmonic, we obtain
\begin{equation}
ic^{2}\vec{k}\times\delta\vec{B}_{\vec{k},\omega}=\frac{1}{\varepsilon_{0
}\ \delta\vec{j}_{\vec{k},\omega}-i\omega\delta\vec{E}_{\vec{k},\omega},\qquad
i\vec{k}\times\delta\vec{E}_{\vec{k},\omega}=i\omega\delta\vec{B}_{\vec
{k},\omega}. \label{urki
\end{equation}
Making a cross-product of Eq.~(\ref{urki}) with $\vec{k}$\ and
using $\vec {k}\cdot\delta\vec{B}_{\vec{k},\omega}=0$, we
express the field perturbations in terms of
$\delta\vec{j}_{\vec{k},\omega}$,
\begin{align}
\delta\vec{B}_{\vec{k},\omega} & =\frac{i\vec{k}\times\delta\vec{j}_{\vec
{k},\omega}}{\varepsilon_{0}(k^{2}c^{2}-\omega^{2})},\label{del_B}\\
\delta\vec{E}_{\vec{k},\omega} & =\frac{i\left[ \omega^{2}\delta\vec
{j}_{\vec{k},\omega}-c^{2}(\vec{k}\cdot\delta\vec{j}_{\vec{k},\omega})\vec
{k}\right] }{\varepsilon_{0}\omega(k^{2}c^{2}-\omega^{2})}. \label{del_E
\end{align}
In a long-lived quasi-periodic non-linearly saturated state,
all wave frequencies $\omega$ must be real, so that
Eq.~(\ref{del_E}) yields
$\operatorname{Re}(\delta\vec{E}_{\vec{k},\omega}\cdot\delta\vec{j}_{\vec{k}
,\omega}^{\ast})=0$, i.e., Eq.~(\ref{so_that_really}).
General Eq.~(\ref{so_that_really}) represents a strict
constraint that follows from the full electrodynamics of
quasi-periodic processes, but it is not obvious that it should
hold for the electrostatic and quasi-neutrality approximations.
The following demonstrates that these approximations do not
lift this constraint.
Indeed, the electron and ion continuity equations combined
yield $e\partial_t (\delta n_e - \delta n_i) = \nabla
\delta\vec{j}$, so that $ e\omega(\delta n_{i\vec{k},\omega} -
\delta n_{e\vec{k},\omega}) = \vec{k}\cdot
\delta\vec{j}_{\vec{k},\omega} $. Combining this with Poisson's
equation, $\varepsilon_0\nabla\cdot\delta\vec{E}=e(\delta n_e -
\delta n_i)$, we obtain
\begin{equation}
i\vec{k}\cdot\delta\vec{E}_{\vec{k},\omega}=
\frac{\vec{k}\cdot\delta\vec{j}_{\vec{k},\omega}}{\varepsilon_0\omega}
\label{proshka1}
\end{equation}
For an electrostatic field, we have
$\delta\vec{E}_{\vec{k},\omega}=|\delta\vec{E}_{i\vec{k},\omega}|(\vec{k}/k)$,
so that Eqs.~(\ref{Average_harmonics}) and (\ref{proshka1})
yield
\begin{equation}
\langle\delta \vec{j}\cdot\delta\vec{E}\rangle = \mathrm{Re}
(\delta \vec{j}_{\vec{k},\omega}\cdot\delta\vec{E}_{\vec{k},\omega}^\ast)
= -\varepsilon_0 |\delta\vec{E}_{\vec{k},\omega}|^2 \,\mathrm{Im}\,\omega = 0
\label{ej=0_electrostat}
\end{equation}
If we add quasi-neutrality, $\nabla\cdot \delta\vec{j}=0$,
i.e., $\vec{k}\cdot\delta\vec{j}_{\vec{k},\omega}=0$ for
individual harmonics, then the satisfaction of
Eq.~(\ref{so_that_really}) becomes obvious even without
assumption of real $\omega$,
\begin{equation}
\langle \delta\vec{j}\cdot\delta\vec{E}\rangle =i\sum_{\vec{k},\omega
\neq0}(\vec{k}\cdot \delta\vec{j}_{\vec{k},\omega})
\delta\Phi_{\vec{k},\omega}^{\ast}=0
\label{ej=0_quasineutral}
\end{equation}
Equation (\ref{so_that_really}) does not mean, however, that
developed plasma turbulence makes no contribution to the
average particle heating. This would certainly contradict both
observations and PIC simulations. The paradox can be resolved
as follows. The average electric field is merely the external
field, $\langle\vec{E}\rangle=\vec{E}_{0}$, while $\langle\vec
{j}\rangle\neq\vec{j}_{0}$, where
$\vec{j}_{0}\equiv\sum_{s}q_{s }n_{s0}\vec{V}_{s0}$ is the
undisturbed current density determined by the laminar plasma
response to $\vec{E}_{0}$. In addition to $\vec{j}_{0}$, the
total average current density, $\langle\vec{j}\rangle$,
includes a wave-induced direct NC
\citep{RogisterJamin:1975,Oppenheim:Evidence97,Buchert:Effect2006},
\begin{equation}
\vec{j}^{\mathrm{NC}}=\langle\vec{j}\rangle-\vec{j}_{0}\equiv\sum_{s
}q_{s}\langle\delta n_{s}\delta\vec{V}_{s}\rangle. \label{J_NL
\end{equation}
The total average loss of field energy to particles per unit
volume and time is given by
$\langle\vec{E}\cdot\vec{j}\rangle$. The corresponding energy
dissipation in the laminar ionosphere with no instabilities is
given by $\vec{E}_{0}\cdot\vec{j}_{0}$. Then the total energy
dissipation exclusively due to plasma turbulence is
$L_{\mathrm{turb}}\equiv\langle\vec{E}\cdot\vec
{j}\rangle-\vec{E}_{0}\cdot\vec{j}_{0}$. Using
Eqs.~(\ref{so_that_really}) and (\ref{J_NL}), one can easily
establish that
\begin{equation}
L_{\mathrm{turb}}=P_{\mathrm{NC}}\equiv\vec{E}_{0}\cdot\vec{j}_{\mathrm{NC}}.
\label{L_turb
\end{equation}
Thus, it is formally the work by the external electric field on
the non-linear current, $P_{\mathrm{NC}}$, rather than
$\langle\delta\vec{E}\cdot\delta \vec{j}\rangle$, that provides
the required turbulent energy deposition for all kinds of
anomalous heating of plasma particles.
\citet{Buchert:Effect2006} showed this for the restricted case
of fluid plasmas with fully unmagnetized ions and a quasilinear
wave description, but did not establish it in the general case.
We have just demonstrated that this fundamental result follows
directly from the general field electrodynamics and no specific
plasma models. By their physical meaning and according to
Eq.~(\ref{L_turb}) each of the two equal quantities
$L_{\mathrm{turb}}$ and $P_{\mathrm{NC}}$ can be named
``turbulent Joule heating.''
Rigorously speaking, exact Eqs.~(\ref{so_that_really}) and
(\ref{L_turb}) apply only to a homogeneous and stationary
background. For mildly inhomogeneous and slowly evolving
background parameters, Eq.~(\ref{Poynting's_perturb}) can be
treated using a regular perturbation technique. The zero-order
approximation, corresponding to a given turbulence level within
an isolated box with a uniform background and periodic boundary
conditions and, hence, not affected by the outside
inhomogeneity, yields Eq.~(\ref{so_that_really}). To reach a
next-order accuracy, one has to establish the zero-order
parameter dependence of non-linearly saturated turbulence
characteristics, using, e.g., a series of computer simulations
with periodic boundary conditions but various background
parameters. Given the large-scale spatial dependence and slow
evolution of the background parameters, one could calculate
then the LHS of Eq.~(\ref{Poynting's_perturb}). This would
yield the first-order, potentially non-zero, values for
$\langle\delta\vec{E}\cdot\delta\vec{j}\rangle$ on the RHS.
Under applicability of our two-scale approach, however, such
possible finite values of $\langle\delta
\vec{E}\cdot\delta\vec{j}\rangle$ will automatically be small
compared to the nearly balancing each other leading terms in
the expanded form of $\langle\delta\vec
{E}\cdot\delta\vec{j}\rangle$ with $\delta\vec{j}=\sum_{s}q_{s}
(\vec{V}_{s0}\delta n+n_{0}\delta\vec{V}_{s}+\delta n\delta\vec
{V}_{s})$, creating only a relatively small mismatch between
$L_{\mathrm{turb}}$ and $P_{\mathrm{NC}}$.
Below we calculate energies deposited among individual particle
groups using an approach similar to that used by
\citet{Buchert:Effect2006}, except that we will apply it to a
system with arbitrarily magnetized electrons and ions. This
applies to all altitudes across the entire \textit{E} region
down to the upper \textit{D} region. As mentioned above,
\citet{Buchert:Effect2006} actually performed a 2-D treatment.
We consider here fully 3-D turbulence, which is crucial for
anomalous electron heating. Note that the total energy input
spent on anomalous heating of plasma particles is less than
$P_{\mathrm{NC}}$ because a fraction of the deposited energy
via collisions goes directly to colliding neutrals and has no
chance to heat the plasma.
\section{Partial Energy Deposit: Quasilinear
Approximation\label{Partial energy deposit: quasilinear
approach}}
In this section, we employ Fourier harmonics in a 3-D periodic
box in order to identify the effect of the parallel turbulent
electric field. Our results apply to all regions that
contribute to the total ionospheric conductances, from the top
electrojet down to the potentially unstable \emph{D}-region
\citep{Kelley:Ionosphere2009}. Similar processes can occur in
other plasma media, like the Solar chromosphere
\citep{Liperovsky:Generation2000,Fontenla:Chromospheric2008,Gogoberidze:Farley2009},
other planetary ionospheres, and laboratory plasma
\citep{Dangelo:1974,John:Observation1975,Koepke:Space2008},
despite the dramatic differences in parameters. Lastly, these
calculations for arbitrarily magnetized plasma serve as an
additional verification of the general relations obtained in
the previous section from first principles.
In this paper, we restrict ionospheric particles to electrons
and a single species of ions, adopting the quasi-neutrality,
$n_{e}\approx n_{i}=n$. Similarly to
\citet{Buchert:Effect2006}, we apply here a quasilinear
approximation, by which we imply that separate Fourier
harmonics of predominantly electrostatic field and plasma
density fluctuations are coupled through simple linear
relations. Using these relations, we calculate then
quadratically non-linear averages in terms of given
spatial-temporal turbulence spectra.
This section is organized as follows. In Sect.~\ref{Dispersion
relation and particle heating: first-order approximation}, we
obtain linear relationships to first-order accuracy and explain
what we mean by the different orders. These first-order
relations provide the linear wave frequencies and relations
between the electrostatic potential and density perturbations.
In Sect.~\ref{non-linear current} we calculate the partial and
total non-linear currents in terms of a given spectrum of
irregularities. In Sect.~\ref{Energy inputs and turbulent
heating}, we calculate partial energy inputs and turbulent
heating of electrons and ions and verify that general
Eqs.~(\ref{so_that_really}) and (\ref{L_turb}) remain in this
approximation exactly valid.
\subsection{First-Order Linear Wave Relations
\label{Dispersion relation and particle heating: first-order approximation}}
\citet{Fejer:Theory84} studied the linear theory of collisional
waves for strongly magnetized electrons and arbitrarily
magnetized ions. On the other hand, \citet{Buchert:Effect2006}
assumed arbitrarily magnetized electrons but unmagnetized ions.
Since no one has published the general 3-D linear relations
that cover all cases, we do this here and a more general
version in the appendix of the companion paper
\citep{Dimant:Magnetosphere2011}. In reasonable agreement with
our PIC simulations, we assume that most of developed
turbulence lies in the long-wavelength, low-frequency range of
$kl_{i}\ll 1$ and $\omega \ll \nu_{i}$, where $\omega$ is the
wave frequency and $l_{i}$ is the mean free path of ions with
respect to dominant ion-neutral collisions. This allows us to
employ a two-fluid, as opposed to kinetic, model and order
various terms in the momentum equations with respect to the
small parameters $kl_{i}$ and $\omega / \nu_{i}$. In this
ordering, particle inertia and pressure gradients are
second-order effects and can be neglected. This first-order
approximation yields a dispersion relation and the
corresponding relation between fluctuations of the plasma
density and electrostatic potential. These relations are common
for all \emph{E}/\emph{D}-region plasma instabilities
\citep{DimOppen2004:ionthermal1}. The neglected second-order
corrections are crucial for the linear wave dissipation and
instability driving, but they are of less importance to the
spatially/temporally averaged energy transfer and plasma
heating. To the second-order accuracy, the general linear wave
theory for arbitrarily magnetized plasmas is developed in the
appendix of \citet{Dimant:Magnetosphere2011}.
Though we discuss here the linear wave relationships, that does
not mean that we consider the linear stage of instability. On
the contrary, we assume a fully developed and non-linearly
saturated turbulence in which the linear wave growth is
balanced by non-linearities. However, we presume that these
non-linearities only weakly modify the linear wave
relationships, so that we include all non-linear terms, along
with the instability driving or damping terms, into the
second-order corrections and neglect their feedback on the
first-order relations.
Under the first-order approximation, each group of particles has only two
balancing forces: the Lorentz force and resistive collisional friction,
\begin{equation}
q_s(\vec{E}+\vec{V}_{s}\times\vec{B}) = m_{s}\nu_{s}\vec{V}_{s}
\label{Lorentz
\end{equation}
where $s=e,i$; we presume ${V}_{s}$ lies a neutral frame of
reference and an undisturbed magnetic field,
$\vec{B}=\vec{B}_{0}$. Introducing the conventional
magnetization parameters, $\kappa_{s}\equiv\Omega_{s}/\nu_{s}$,
where $\Omega_{s}=|q_s|B/m_s$ are the electron and ion
gyrofrequencies, we obtain from
Eq.~(\ref{Lorentz}
\begin{equation}
\vec{V}_{e\parallel}=-\ \frac{\kappa_{e}\vec{E}_{\parallel}}{B},\qquad\vec
{V}_{i\parallel}=\frac{\kappa_{i}\vec{E}_{\parallel}}{B}, \label{V_e,i_II
\end{equation
\begin{equation}
\vec{V}_{e\perp}=\frac{\kappa_{e}[-\vec{E}_{\perp}+\kappa_{e}(\vec{E
\times\hat{b})]}{\left( 1+\kappa_{e}^{2}\right) B},\ \ \ \vec{V}_{i\perp
}=\frac{\kappa_{i}[\vec{E}_{\perp}+\kappa_{i}(\vec{E}\times\hat{b})]}{\left(
1+\kappa_{i}^{2}\right) B}, \label{V_e,i
\end{equation}
with the subscripts $\parallel,\perp$ denoting the components parallel and
perpendicular to $\vec{B}$, respectively; here $B\equiv|\vec{B}|$ and $\hat
{b}\equiv\vec{B}/B$.
Further, we calculate the relative mean velocity between
electrons and ions, $\vec{U}\equiv\vec{V}_{e}-\vec{V}_{i}$,
which plays an important role in many relations.
Equations~(\ref{V_e,i_II}) and (\ref{V_e,i}) give
\begin{subequations}
\label{UUU
\begin{align}
\vec{U}_{\parallel} & =-\ \frac{\left( \kappa_{e}+\kappa_{i}\right)
\vec{E}_{\parallel}}{B},\label{U_II}\\
\vec{U}_{\perp} & =\frac{\left( \kappa_{e}+\kappa_{i}\right) [\left(
\kappa_{e}-\kappa_{i}\right) (\vec{E}\times\hat{b})-\left( 1+\kappa
_{i}\kappa_{e}\right) \vec{E}_{\perp}]}{\left( 1+\kappa_{e}^{2}\right)
\left( 1+\kappa_{i}^{2}\right) B}. \label{U_perp
\end{align}
\end{subequations}
Reversing Eq.~(\ref{UUU}), we obtain
\begin{eqnarray}
\frac{\vec{E}_{\parallel}}{B} & = &-\ \frac{\vec{U}_{\parallel}}{\kappa_{e}
+\kappa_{i}},\nonumber
\\
\frac{\vec{E}_{\perp}}{B} & = &-\,\frac{\left( \kappa_{e}
-\kappa_{i}\right) (\vec{U}\times\hat{b})+\left(
1+\kappa_{i}\kappa _{e}\right)
\vec{U}_{\perp}}{\kappa_{i}+\kappa_{e}}, \label{ExB_via_U}
\end{eqnarray}
so that Eqs.~(\ref{V_e,i_II}) and (\ref{V_e,i}) yield the following mean drift
velocities of electrons and ions in terms of their relative velocity
\begin{equation}
\vec{V}_{e\parallel}=\frac{\kappa_{e}\vec{U}_{\parallel}}{\kappa_{e
+\kappa_{i}},\qquad\vec{V}_{i\parallel}=-\ \frac{\kappa_{i}\vec{U}_{\parallel
}}{\kappa_{e}+\kappa_{i}}, \label{V_II_via_U
\end{equation}
\begin{subequations}
\label{V_perp_via_U
\begin{align}
\vec{V}_{e\perp} & =\frac{\kappa_{e}[\vec{U}_{\perp}-\kappa_{i}(\vec
{U}\times\hat{b})]}{\kappa_{e}+\kappa_{i}},\label{V_e_perp_via_U}\\
\vec{V}_{i\perp} & =-\ \frac{\kappa_{i}[\vec{U}_{\perp}+\kappa_{e}(\vec
{U}\times\hat{b})]}{\kappa_{e}+\kappa_{i}}. \label{V_i_perp_via_U
\end{align}
\end{subequations}
All these linear expressions can be applied separately to the undisturbed
quantities and wave perturbations.
Now we consider wave perturbations of the field, plasma
density, and fluid velocities and apply to them the discrete
Fourier transforms introduced by Eq.~(\ref{Fourier_transform}).
Note that the neglect of non-linear terms with respect to wave
perturbations leads to a unique $\vec{k}$-dependence of the
wave frequency, $\omega=\omega _{\vec{k}}(\vec{k})$, via the
corresponding first-order linear dispersion relation. Such
unique dependence may break the presumed discreteness of the
wave frequency, $\omega=2\pi m/T$. However, the quasi-period
$T$ can always be chosen sufficiently long so that the interval
between adjacent discrete frequencies, $\Delta\omega=2\pi/T$,
becomes much less than the finite spectral width around
$\omega=\omega_{\vec{k}}$; see also the discussion below in the
paragraph following Eq.~(\ref{for_kappa_e>>1}).
Expressing velocity perturbations,
$\delta\vec{V}_{e,i\vec{k},\omega}$, in terms of the
electrostatic potential perturbations,
$\delta\Phi_{\vec{k},\omega}$, via
$\delta\vec{E}_{\vec{k},\omega
}=-i\vec{k}\delta\Phi_{\vec{k},\omega}$, we obtain from
Eqs.~(\ref{V_e,i_II}) and (\ref{V_e,i}):
\begin{subequations}
\label{delta_Vs
\begin{align}
\delta\vec{V}_{e\parallel\vec{k},\omega} & =i\ \frac{\kappa_{e}\vec
{k}_{\parallel}\delta\Phi_{\vec{k},\omega}}{B},\qquad\delta\vec{V
_{i\parallel\vec{k},\omega}=-i\ \frac{\kappa_{i}\vec{k}_{\parallel}\delta
\Phi_{\vec{k},\omega}}{B},\label{delta_V_II}\\
\delta\vec{V}_{e\perp\vec{k},\omega} & =-i\ \frac{\kappa_{e}[-\vec{k
_{\perp}+\kappa_{e}(\vec{k}_{\perp}\times\hat{b})]\delta\Phi_{\vec{k},\omega
}{\left( 1+\kappa_{e}^{2}\right) B},\label{delta_V_perp_e}\\
\delta\vec{V}_{i\perp\vec{k},\omega} & =-i\ \frac{\kappa_{i}[\vec{k}_{\perp
}+\kappa_{i}(\vec{k}_{\perp}\times\hat{b})]\delta\Phi_{\vec{k},\omega
}{(1+\kappa_{i}^{2})B}. \label{delta_V_perp_i
\end{align}
\end{subequations}
Writing the quasi-neutral continuity equations as
\begin{equation}
\partial_{t}n+\nabla\cdot(n\vec{V}_{i})=0,\qquad\nabla\cdot(n\vec{U})=0,
\label{conti
\end{equation}
where we neglected wave variations of ionization-recombination
balance, and linearizing them with respect to density
perturbations, $\delta n_{\vec{k},\omega}$, we obtain
\begin{equation}
\frac{\delta n_{\vec{k},\omega}}{n_{0}}=\frac{\vec{k}\cdot\delta\vec{V
_{i\vec{k},\omega}}{\Omega_{\vec{k}}^{(i)}}=-\ \frac{\vec{k}\cdot\delta\vec
{U}_{\vec{k},\omega}}{\vec{k}\cdot\vec{U}_{0}}. \label{del_n_contin
\end{equation}
Here
$\Omega_{\vec{k}}^{(i)}\equiv\omega-\vec{k}\cdot\vec{V}_{i0}$
is the linear wave frequency in the ion frame;
$\delta\vec{U}_{\vec{k},\omega}=\delta\vec
{V}_{e\vec{k},\omega}-\delta\vec{V}_{i\vec{k},\omega}$, and,
according to Eq.~(\ref{UUU}),
\begin{equation}
\vec{U}_{0}=\frac{\left( \kappa_{i}+\kappa_{e}\right) [\left( \kappa
_{e}-\kappa_{i}\right) (\vec{E}_{0}\times\hat{b})-\left( 1+\kappa_{i
\kappa_{e}\right) \vec{E}_{0}]}{\left( 1+\kappa_{e}^{2}\right)
(1+\kappa_{i}^{2})B}. \label{U_0
\end{equation}
Using Eq.~(\ref{delta_Vs}), we obtain
\begin{subequations}
\label{k_Vushki
\begin{align}
\vec{k}\cdot\delta\vec{V}_{i\vec{k},\omega} & =-i\kappa_{i}\left(
\frac{k_{\perp}^{2}}{1+\kappa_{i}^{2}}+k_{\parallel}^{2}\right) \frac
{\delta\Phi_{\vec{k},\omega}}{B},\label{k_Vushki_i}\\
\vec{k}\cdot\delta\vec{V}_{e\vec{k},\omega} & =i\kappa_{e}\left(
\frac{k_{\perp}^{2}}{1+\kappa_{e}^{2}}+k_{\parallel}^{2}\right) \frac
{\delta\Phi_{\vec{k},\omega}}{B}, \label{k_Vushki_e
\end{align}
\end{subequations}
so that
\begin{equation}
\vec{k}\cdot\delta\vec{U}_{\vec{k},\omega}=\frac{i\kappa_{i}\kappa_{e}\left(
\kappa_{e}+\kappa_{i}\right) (1+\psi_{\vec{k}})k_{\perp}^{2}\delta\Phi
_{\vec{k},\omega}}{\left( 1+\kappa_{e}^{2}\right) (1+\kappa_{i}^{2})B},
\label{dul_ka
\end{equation}
where
\begin{subequations}
\label{psi
\begin{align}
\psi_{\vec{k}} & \equiv\psi_{\perp}\left[ 1+(1+\kappa_{e}^{2})(1+\kappa
_{i}^{2})\frac{k_{\parallel}^{2}}{k_{\perp}^{2}}\right] ,\label{psi_k}\\
\psi_{\perp} & \equiv\frac{1}{\kappa_{i}\kappa_{e}}=\frac{\nu_{e}\nu_{i
}{\Omega_{e}\Omega_{i}}. \label{psi_perp
\end{align}
\end{subequations}
The 2-D parameter $\psi_{\perp}$ is conventional. The newly
defined 3-D parameter $\psi_{\vec{k}}$ generalizes the
traditional 3-D parameter $\psi
=\psi_{\perp}(1+\kappa_{e}^{2}k_{\parallel}^{2}/k_{\perp}^{2})$
originally introduced for magnetized electrons,
$\kappa_{e}^{2}\gg1$, and unmagnetized ions,
$\kappa_{i}^{2}\ll1$ \citep[e.g.,][]{Farley:96}. In a more
general case of $\kappa_{e}^{2}\gg1$ but arbitrary
$\kappa_{i}$, the parameter $\psi_{\vec{k}}$ replaces the
product $(1+\kappa_{i}^{2})\psi$ \citep{Fejer:Theory84}. In the
lower ionosphere, the difference
between the two is negligible, $(1+\kappa_{i}^{2})\psi-\psi_{\vec{k}
=\kappa_{i}^{2}\psi_{\perp}=\Theta_{0}^{2}\equiv
m_{e}\nu_{e}/(m_{i}\nu _{i})\simeq1.8\times10^{-4}$
\citep{DimantMilikh:JGR03,DimOppen2004:ionthermal1}.
Using Eqs.~(\ref{k_Vushki}) and (\ref{dul_ka}), we obtain from
Eq.~(\ref{del_n_contin}) the first-order relation between the linear
fluctuations of the density and electrostatic potential
\begin{equation}
\delta\Phi_{\vec{k},\omega}=i\ \frac{(1+\kappa_{e}^{2})(1+\kappa_{i
^{2})B(\vec{k}\cdot\vec{U}_{0})}{\kappa_{i}\kappa_{e}\left( \kappa_{e
+\kappa_{i}\right) (1+\psi_{\vec{k}})k_{\perp}^{2}}\left( \frac{\delta
n_{\vec{k},\omega}}{n_{0}}\right) , \label{Phi->del_n_symmetric
\end{equation}
as well as the expression for the first-order wave frequency in the ion frame,
\begin{equation}
\Omega_{\vec{k}}^{(i)}=\frac{(1+\kappa
_{e}^{2})[1+(1+\kappa_{i}^{2})k_{\parallel}^{2}/k_{\perp}^{2}](\vec{k
\cdot\vec{U}_{0})}{\kappa_{e}\left( \kappa_{e}+\kappa_{i}\right)
(1+\psi_{\vec{k}})}. \label{Omega_K
\end{equation}
In the neutral frame, the corresponding frequency, $\omega_{\vec{k
}=\Omega_{\vec{k}}^{(i)}+\vec{k}\cdot\vec{V}_{i0}$, is
\begin{equation}
\omega_{\vec{k}}=\frac{(\kappa_{e}-\kappa_{i})(\vec{k}\cdot\vec{U}_{0
)}{(\kappa_{e}+\kappa_{i})(1+\psi_{\vec{k}})}-\frac{\kappa_{i}\kappa_{e
\vec{k}\cdot(\vec{U}_{0}\times\hat{b})}{\kappa_{e}+\kappa_{i}}.
\label{omega_k
\end{equation}
When obtaining these expressions, we used the easily derived relatio
\begin{equation}
\left( 1+\kappa_{i}\kappa_{e}\right) k_{\perp}^{2}+(1+\kappa_{e
^{2})(1+\kappa_{i}^{2})k_{\parallel}^{2}=\kappa_{i}\kappa_{e}k_{\perp
^{2}(1+\psi_{\vec{k}}). \label{used_here
\end{equation}
As might be expected for arbitrarily magnetized electrons and
ions, Eqs.~(\ref{Phi->del_n_symmetric}) and (\ref{omega_k}) are
symmetric with respect to the interchange between electrons and
ions that requires $\kappa_{i,e}\leftrightarrow\kappa_{e,i}$
and $\vec{U}_{0}\leftrightarrow -\vec{U}_{0}$. Equation
(\ref{Omega_K}) for the Doppler-shifted frequency in the ion
frame, $\Omega_{\vec{k}}^{(i)}$, is symmetric with respect to a
similar expression
for the wave frequency in the electron frame, $\Omega_{\vec{k}}^{(e)
\equiv\omega_{\vec{k}}-\vec{k}\cdot\vec{V}_{e0}=\Omega_{\vec{k}}^{(i)
-\vec{k}\cdot\vec{U}_{0}$,
\begin{equation}
\Omega_{\vec{k}}^{(e)}=-\ \frac{(1+\kappa_{i}^{2})[1+(1+\kappa_{e
^{2})k_{\parallel}^{2}/k_{\perp}^{2}](\vec{k}\cdot\vec{U}_{0})}{\kappa
_{i}\left( \kappa_{i}+\kappa_{e}\right) (1+\psi_{\vec{k}})}.
\label{Omega_k-kV_i0
\end{equation}
Notice that for prevalent waves with positive
$\vec{k}\cdot\vec{U}_{0}$ the shifted electron frequency
$\Omega_{\vec{k}}^{(e)}$ is always negative, while the
corresponding ion frequency $\Omega_{\vec{k}}^{(i)}$ is
positive. This reflects the fact that such waves, regardless of
the particle magnetization, always lag behind the streaming
electrons but move ahead of the ions.
Despite the formal symmetry in the general relations between
electrons and ions, their actual contributions are not
equivalent. In the lower ionosphere, since $m_{i}\simeq30$\,amu
and $\nu_{e}/\nu_{i}\simeq10$ \citep{Kelley:Ionosphere2009},
the ratio $\kappa_{e}/\kappa_{i}$ is huge,
$\kappa_{e}/\kappa_{i}\simeq5500$. Then practically everywhere,
except for the \emph{D}-region altitudes below 80~km, electrons
are strongly magnetized, $\kappa_{e}\gg1$, whilst in most of
the \emph{E}/\emph{D}-region electrojet ions are largely
unmagnetized, $\kappa _{i}\ll1$, reaching only a partial
magnetization, $\kappa_{i}\gtrsim1$, above 110~km. Apart from
the \emph{D}-region altitudes but practically throughout the
entire electrojet, setting $\kappa_{e}\gg1$ with arbitrary
$\kappa_{i}$, we reduce Eqs.~(\ref{psi}) to (\ref{omega_k}) to
simpler relations \citep{Fejer:Theory84}:
\begin{align}
\psi_{\vec{k}} & \approx\psi_{\perp}\left[ 1+(1+\kappa_{i}^{2})\frac
{\Omega_{e}^{2}k_{\parallel}^{2}}{\nu_{e}^{2}k_{\perp}^{2}}\right]
,\nonumber\\
\omega_{\vec{k}} & \approx\frac{\vec{k}\cdot\vec{U}_{0}}{1+\psi_{\vec{k}
}-\kappa_{i}\vec{k}\cdot(\vec{U}_{0}\times\hat{b}),\label{for_kappa_e>>1}\\
\delta\Phi_{\vec{k},\omega} & =i\ \frac{m_{i}\nu_{i}(1+\kappa_{i}^{2
)(\vec{k}\cdot\vec{U}_{0})}{ek_{\perp}^{2}(1+\psi_{\vec{k}})}\left(
\frac{\delta n_{\vec{k},\omega}}{n_{0}}\right) .\nonumber
\end{align}
For unmagnetized ions, $\kappa_{i}\ll1$, Eq.~(\ref{for_kappa_e>>1}) reduces
further to the conventional relations \citep{Farley:96}.
\begin{comment}
Equation~(\ref{Phi->del_n_symmetric}) and its simplified
version given by the last relation of
Eq.~(\ref{for_kappa_e>>1}) show that the spectra of density
irregularities, $\delta n_{\vec{k},\omega}$ scale as the
amplitude of the turbulent electric field,
$\delta\vec{E}_{\vec{k},\omega}=-i\vec{k}\delta\Phi_{\vec{k},\omega}$.
This means that the ratios of
$|\delta\vec{E}_{\vec{k},\omega}|$ to $\delta
n_{\vec{k},\omega}$ depend only on the direction of $\vec{k}$,
rather than on the wavenumber $k=|\vec{k}|$. For the major
perpendicular turbulent field,
$\delta\vec{E}_{\vec{k},\omega\perp}=-i\vec{k}_{\perp}\delta\Phi
_{\vec{k},\omega}$, we hav
\begin{align}
|\delta\vec{E}_{\vec{k},\omega\perp}| & =\frac{(1+\kappa_{e}^{2
)(1+\kappa_{i}^{2})BU_{0}\cos\chi_{\vec{k}}}{\kappa_{i}\kappa_{e}\left(
\kappa_{e}+\kappa_{i}\right) (1+\psi_{\vec{k}})}\frac{|\delta n_{\vec
{k},\omega}|}{n_{0}}\nonumber\\
& \approx\frac{m_{i}\nu_{i}(1+\kappa_{i}^{2})U_{0}|\cos\chi_{\vec{k}
|}{e(1+\psi_{\vec{k}})}\frac{|\delta n_{\vec{k},\omega}|}{n_{0}},
\label{del_E_perp
\end{align}
where the last approximate equality is for $\kappa_{e}\gg 1$;
$\chi_{\vec{k}}$ is the modified flow angle, i.e., the angle
between $\vec{k}$ and $\vec{U}_{0}$
\citep{DimOppen2004:ionthermal1}, while $\psi_{\vec{k}}$
depends also on $k_\parallel/k_\perp$. In the companion paper
\citep{Dimant:Magnetosphere2011}, this will allow us to express
the rms electric field perturbations in terms of the rms
density perturbations with no knowledge of the specific
$k$-spectra.
\end{comment}
Recall that all these relations represent only the first-order
approximation with respect to the small parameters $kl_{i}$ and
$\omega/\nu_{i}$. Neglect of pressure gradients, particle
inertia, etc., results in the highest-order dispersion relation
for the real part of the linear wave frequency,
Eq.~(\ref{omega_k}). It contains no $\pi/2$-phase shifted
corrections that determine the linear wave growth or damping.
For arbitrary particle magnetization, the corresponding
second-order corrections to the dispersion relation for the FB
and gradient drift instabilities are obtained in the appendix
of \citet{Dimant:Magnetosphere2011}. We should bear in mind,
however, that our quasi-periodic description of the
non-linearly saturated steady state requires all wave
frequencies to be strictly real. This implies that the
second-order destabilizing linear factors are on average
balanced by non-linearities. Furthermore, non-linearly
saturated turbulence itself has a strongly time-varying
character caused by dynamic mode-coupling
\citep[e.g.,][]{Hamza:turbulent93,Hamza:self93,Dimant:2000}.
All this breaks the narrowband approximation for which any wave
frequency $\omega$, via the linear dispersion relation, is
uniquely determined by the corresponding wavevector $\vec{k}$,
$\omega\approx\omega_{\vec{k}}$. This means that the
$\delta$-function dependence of the frequency spectrum is
actually spread over a finite band around $\omega _{\vec{k}}$.
Accurate theoretical description of such non-linearly saturated
states is extremely difficult, especially for strong turbulence
generated by the driving field well above the instability
threshold. To avoid serious difficulties and mathematical
complexities, we will continue using first-order linear
relations like Eqs.~(\ref{delta_Vs}) and
(\ref{Phi->del_n_symmetric}) and neglect next-order linear and
non-linear corrections. We expect the corresponding errors to
be reasonably small, although these expectations need
additional testing.
\subsection{Non-linear Currents\label{non-linear current}}
As shown in Sect.~\ref{General consideration}, the non-linear
current (NC) plays an important role in energy conversion. To
further clarify its physical meaning, we calculate separately
NCs for electrons and ions, $\vec{j}_{s}^{\mathrm{NC}}$
($s=e,i$), in terms of a given density-irregularity spectrum,
$\delta n_{\vec{k},\omega}$. These partial currents will be
used in Sect.~\ref{Energy inputs and turbulent heating} for
comparison of the partial turbulent Joule heating,
$\vec{E}_0\cdot\vec{j}_{s}^{\mathrm{NC}}$, with the
corresponding frictional heating terms.
Using Eqs.~(\ref{Average_harmonics}), (\ref{delta_Vs}), and
(\ref{Phi->del_n_symmetric}), we obtain the partial electron NC
density, $\vec {j}_{e}^{\mathrm{NC}}\equiv-e\langle\delta
n\delta\vec{V}_{e}\rangle =-e\sum_{\vec{k},\omega\neq0}\delta
n_{\vec{k}}^{\ast}\delta\vec{V}_{e\vec{k}}$:
\begin{align}
\vec{j}_{e}^{\mathrm{NC}} & =\frac{(1+\kappa_{e}^{2})\left( 1+\kappa_{i
^{2}\right) en_{0}}{\kappa_{i}\left( \kappa_{e}+\kappa_{i}\right)
\sum_{\vec{k},\omega\neq0}\frac{\vec{k}\cdot\vec{U}_{0}}{(1+\psi_{\vec{k
})k_{\perp}^{2}}\nonumber\\
& \times\left[ \vec{k}_{\parallel}+\frac{\vec{k}_{\perp}-\kappa_{e}(\vec
{k}_{\perp}\times\hat{b})}{1+\kappa_{e}^{2}}\right] \left\vert \frac{\delta
n_{\vec{k},\omega}}{n_{0}}\right\vert ^{2}.\label{j_NLe
\end{align}
Similarly, we obtain the partial ion NC density, $\vec{j}_{i}^{\mathrm{NC}}\equiv
e\langle\delta n\delta\vec{V}_{i}\rangle\allowbreak=\allowbreak e\sum_{\vec
{k},\omega\neq0}\delta n_{\vec{k}}^{\ast}\delta\vec{V}_{i\vec{k}}$, which
differs from Eq.~(\ref{j_NLe})
by the replacement $\kappa_{e,i}\leftrightarrow\kappa_{i,e}$ and
the ``plus'' sign in front of $(\vec{k}_{\perp}\times\hat{b})$.
\begin{comment}
\begin{align}
\vec{j}_{i}^{\mathrm{NC}} & =\frac{(1+\kappa_{e}^{2})(1+\kappa_{i}^{2
)en_{0}}{\kappa_{e}\left( \kappa_{e}+\kappa_{i}\right) }\sum_{\vec{k
,\omega\neq0}\frac{\vec{k}\cdot\vec{U}_{0}}{(1+\psi_{\vec{k}})k_{\perp}^{2
}\nonumber\\
& \times\left[ \vec{k}_{\parallel}+\frac{\vec{k}_{\perp}+\kappa_{i}(\vec
{k}_{\perp}\times\hat{b})}{1+\kappa_{i}^{2}}\right] \left\vert \frac{\delta
n_{\vec{k},\omega}}{n_{0}}\right\vert ^{2},\label{j_NLi
\end{align}
\end{comment}
The total NC, $\vec{j}^{\mathrm{NC}}\equiv-e\langle\delta n\delta\vec
{U}\rangle=-e\sum_{\vec{k},\omega\neq0}\delta n_{\vec{k}}^{\ast}\delta\vec
{U}_{\vec{k}}$, is then given by
\begin{align}
& \vec{j}^{\mathrm{NC}}=\vec{j}_{e}^{\mathrm{NC}}+\vec{j}_{i}^{\mathrm{NC
}=\frac{en_{0}}{\kappa_{i}\kappa_{e}}\sum_{\vec{k},\omega\neq0}
(\vec{k}\cdot\vec{U}_{0})\left\vert \frac{\delta n_{\vec{k},\omega}}{n_{0}}
\right\vert^{2}
\nonumber\\
&\times \frac{(1+\kappa_{e}^{2})\left(
1+\kappa_{i}^{2}\right) \vec{k}_{\parallel}+\left( 1+\kappa_{i}\kappa
_{e}\right) \vec{k}_{\perp}-\left( \kappa_{e}-\kappa_{i}\right) (\vec
{k}\times\hat{b})}{(1+\psi_{\vec{k
})k_{\perp}^{2}}.
\label{j_NL
\end{align}
At altitudes above $100$~km, where both
$\kappa_{e}\gg1\gtrsim\kappa_{i}$ and
$\kappa_{i}\kappa_{e}=\psi_{\perp}^{-1}\gg1$ hold together,
according to Eq.~(\ref{U_0}), we have
\begin{equation}
\vec{U}_{0}=\frac{\vec{E}_{0}\times\hat{b}-\kappa_{i}\vec{E}_{0}}{\left(
1+\kappa_{i}^{2}\right) B},\label{U_0_simple
\end{equation}
while the ion non-linear current turns out to be negligible compared to that of
electrons, $\vec{j}_{e}^{\mathrm{NC}}\approx\vec{j}^{\mathrm{NC}}$. We expect
no spectral asymmetry along $\vec{B}$, so that $\sum_{\vec{k},\omega\neq
0}f(k_{\perp}^{2})\vec{k}_{\parallel}=0$. As a result, the
NC density, $\vec{j}^{\mathrm{NC}}\perp\vec{B}$, reduces in this limit to
\begin{equation}
\vec{j}^{\mathrm{NC}}\approx-\ \frac{en_{0}}{\kappa_{i}}\sum_{\vec{k
,\omega\neq0}\frac{(\vec{k}\times\hat{b})(\vec{k}\cdot\vec{U}_{0})
{(1+\psi_{\vec{k}})k_{\perp}^{2}}\left\vert \frac{\delta n_{\vec{k},\omega
}{n_{0}}\right\vert ^{2}.\label{non-linear_simplest
\end{equation}
Within the bulk electrojet where ions are unmagnetized, $\kappa_{i}\ll1$, the
$\vec{k}$-spectrum of irregularities is largely perpendicular to $\vec{E}_{0
$, so that $\vec{j}^{\mathrm{NC}}$ has there a predominantly
Pedersen direction. This is of paramount importance for the
global MI coupling \citep{Dimant:Magnetosphere2011}.
\subsection{Partial Energy Inputs and Turbulent
Heating\label{Energy inputs and turbulent heating}}
Calculating the partial energy inputs and frictional heating
sources for a specific plasma model and given turbulence
spectrum allows to quantitatively understand how the turbulent
energy is distributed between different plasma components. Here
we obtain such expressions for arbitrarily magnetized two-fluid
plasmas in the quasilinear approximation and verify, in
particular, that exact Eqs.~(\ref{so_that_really}) and
(\ref{L_turb}) remain exactly valid. The turbulent frictional
heating sources found here could be included into
ionosphere-thermosphere computer models, as explained in the
companion paper \citep{Dimant:Magnetosphere2011}.
Multiplying Eq.~(\ref{Lorentz}) by the plasma density and
corresponding fluid velocities gives
\begin{equation}
\vec{E}\cdot\vec{j}_{s}=m_{s}\nu_{s}nV_{s}^{2}
\label{Heating
\end{equation}
($s=e,i$), where $\vec{j}_{e}\equiv-en\vec{V}_{e}$ and
$\vec{j}_{i}\equiv en\vec{V}_{i}$ are the electron and ion
partial current densities. These expressions relate the work
done by the field $\vec{E}$ on the \textit{s}-particle currents
to the corresponding sources of frictional heating. As
mentioned in Sect.~\ref{General consideration}, the actual
frictional heating of $s$-type particles is smaller than
$m_{s}\nu_{s n}nV_{s}^{2}$ because a fraction of the acquired
field energy equal to $m_s/(m_s + m_n)$ goes immediately to the
colliding neutrals without heating the plasma particles
\citep{Schunk:Ionospheres09}. This is especially true for ions
with $m_i\simeq m_n$, whose frictional heating source is nearly
half of $m_{i}\nu_{i}nV_{i}^{2}$. In what follows, we will
refer to $m_{s}\nu_{s n}nV_{s}^{2}$ as the \textit{s-n}
`heating term' with the caveat that the actual $s$-particle
frictional heating is described by $m_n/(m_s + m_n)$ of
$m_{s}\nu_{s}nV_{s}^{2}$, while the remaining fraction goes to
neutral ($n$) frictional heating
\citep{Dimant:Magnetosphere2011}.
Equation~(\ref{Heating}) includes total field energy losses and
plasma heating. The zero-order heating alone, i.e., that without any
plasma turbulence, is described by
\begin{equation}
\vec{E}_{0}\cdot\vec{j}_{s0}=m_{s}\nu_{s}n_{0}V_{s0}^{2}.
\label{Heating_0
\end{equation}
Subtracting Eq.~(\ref{Heating_0}) from Eq.~(\ref{Heating}) and averaging yields
the turbulent heating rates,
\begin{eqnarray}
\vec{E}_{0}\cdot\vec{j}_{s}^{\mathrm{NC}}
\!\!\!& + &\!\!\!\langle\delta\vec{E}\cdot\delta\vec{j}_{s}\rangle\nonumber
\\
& = &\!\! m_{s}\nu_{s}(n_{0}\langle \delta V_{s}^{2}\rangle +2\vec
{V}_{s0}\cdot\langle\delta\vec{V}_{s}\delta n\rangle+\!\langle \delta
V_{s}^{2}\delta n\rangle)
\label{HH
\end{eqnarray}
where $\vec{j}_{e}^{\mathrm{NC}}=-e\langle\delta
n\delta\vec{V}_{e}\rangle$ and
$\vec{j}_{i}^{\mathrm{NC}}=e\langle\delta
n\delta\vec{V}_{i}\rangle$ are the partial electron and ion
contributions to the total NC given by Eq.~(\ref{j_NL}). The
LHS of Eq.~(\ref{HH}) describes the additional average work by
electric fields in the turbulent plasma on a given plasma
species. The RHS represents the corresponding average turbulent
heating per unit volume. It is not obvious, however, that the
approximate quasilinear expressions obtained above ensure that
the two sides of Eq.~(\ref{HH}) are really equal. To verify the
equality and clarify the physical meaning of various terms, we
will calculate the two sides of Eq.~(\ref{HH}) separately.
We start by calculating the RHS of Eq.~(\ref{HH}) that
describes the frictional heating. The positively determined and
dominant term $m_s\nu_sn_{0}\langle\delta V_{s}^{2}\rangle$
describes energization of individual plasma particles, while
the other terms are associated with density variations at a
given location. Consistency of the quasilinear approximation
requires neglecting the last, cubically non-linear, term. To
calculate average quadratically non-linear quantities in terms
of given irregularity spectra, we use
Eq.~(\ref{Average_harmonics}).
By using Eq.~(\ref{delta_Vs}), for the dominant electron
and ion heating terms, we obtain
\begin{equation}
m_{s}\nu_{s}n_{0}\left\langle \delta V_{s}^{2}\right\rangle =\frac
{en_{0}\kappa_{s}}{B}\sum_{\vec{k},\omega\neq0}\left( \frac{k_{\perp}^{2
}{1+\kappa_{s}^{2}}+k_{\parallel}^{2}\right) |\delta\Phi_{\vec{k},\omega
}|^{2}.
\label{intermediate
\end{equation}
If $\kappa_{e}\gg\kappa_{i}$ and $\kappa_{e}\gg1$ then the ion
perpendicular heating dominates over the corresponding electron
one, while for the parallel heating the reverse is true. At
lower \emph{D}-region altitudes, where electrons are partially
demagnetized, $\kappa_{i}\ll\kappa_{e}\lesssim1$, the electron
heating always prevails. Using
Eq.~(\ref{Phi->del_n_symmetric}), we rewrite
Eq.~(\ref{intermediate}) in terms of a given
density-irregularity spectrum as
\begin{align}
m_{e}\nu_{e}n_{0}\left\langle \delta V_{e}^{2}\right\rangle & =\frac
{en_{0}B(1+\kappa_{e}^{2})(1+\kappa_{i}^{2})^{2}}{(\kappa_{e}+\kappa_{i
)^{2}\kappa_{i}^{2}\kappa_{e}}\nonumber\\
& \times\sum_{\vec{k},\omega\neq0}\frac{[1+(1+\kappa_{e}^{2})k_{\parallel
}^{2}/k_{\perp}^{2}](\vec{k}\cdot\vec{U}_{0})^{2}}{(1+\psi_{\vec{k}
)^{2}k_{\perp}^{2}}\left\vert \frac{\delta n_{\vec{k},\omega}}{n_{0
}\right\vert ^{2}
\label{diag_psi
\end{align}
\begin{comment}
\label{diag_psi_ee}
\\
m_{i}\nu_{i}n_{0}\left\langle \delta V_{i}^{2}\right\rangle & =\frac
{en_{0}B(1+\kappa_{i}^{2})(1+\kappa_{e}^{2})^{2}}{(\kappa_{e}+\kappa_{i
)^{2}\kappa_{e}^{2}\kappa_{i}}\nonumber\\
& \times\sum_{\vec{k},\omega\neq0}\frac{[1+(1+\kappa_{i}^{2})k_{\parallel
}^{2}/k_{\perp}^{2}](\vec{k}\cdot\vec{U}_{0})^{2}}{(1+\psi_{\vec{k}
)^{2}k_{\perp}^{2}}\left\vert \frac{\delta n_{\vec{k},\omega}}{n_{0
}\right\vert ^{2}. \label{diag_psi_ii
\end{subequations}
\end{comment}
and similar for ions with the symmetric replacement
$\kappa_{e,i}\leftrightarrow\kappa_{i,e}$. Using
Eqs.~(\ref{Phi->del_n_symmetric}) and (\ref{used_here}), for
the combined heating rate we obtain
\begin{subequations}
\label{diag_psi_summa_ei}
\begin{align}
& m_{e}\nu_{e}n_{0}\left\langle \delta V_{e}^{2}\right\rangle +m_{i}\nu
_{i}n_{0}\left\langle \delta V_{i}^{2}\right\rangle \nonumber\\
& =\frac{eBn_{0}\psi_{\perp}(1+\kappa_{e}^{2})(1+\kappa_{i}^{2})}{\kappa
_{e}+\kappa_{i}}\sum_{\vec{k},\omega\neq0}\frac{(\vec{k}\cdot\vec{U}_{0})^{2
}{(1+\psi_{\vec{k}})k_{\perp}^{2}}\left\vert \frac{\delta n_{\vec{k},\omega
}{n_{0}}\right\vert ^{2}\label{diag_psi_summa_ei_1}\\
& =\frac{en_{0}\left( \kappa_{i}+\kappa_{e}\right) }{B\psi_{\perp}\left(
1+\kappa_{e}^{2}\right) (1+\kappa_{i}^{2})}\sum_{\vec{k},\omega\neq0}(1+\psi
_{\vec{k}})k_{\perp}^{2}\left\vert \delta\Phi_{\vec{k},\omega}\right\vert
^{2}.\label{diag_psi_summa_ei_2}
\end{align}
\end{subequations}
Equation~(\ref{diag_psi_summa_ei}), presented here in terms of both $\delta
n_{\vec{k},\omega}$ and $\delta\Phi_{\vec{k},\omega}$, allows
one to understand the quantitative difference between 2-D and
3-D turbulent heating. The 3-D effect of parallel-field dominated turbulent heating
\citep{DimantMilikh:JGR03,MilikhDimant:JGR03} is entirely due
to the difference between $\psi_{\vec{k}}$ and $\psi_{\perp}$
in the multipliers $(1+\psi_{\vec{k}})$. While the total
perpendicular turbulent heating is
described by the $(1+\psi_{\perp})$ component, the total parallel turbulent
heating derives from the remainder, $\psi_{\vec{k}
-\psi_{\perp}=\psi_{\perp}(1+\kappa_{e}^{2})(1+\kappa_{i}^{2})k_{\parallel
}^{2}/k_{\perp}^{2}$, as seen from Eq.~(\ref{psi}). If $|\delta
n_{\vec{k},\omega}|^{2}$ were equal in 2-D and 3-D then,
according to Eq.~(\ref{diag_psi_summa_ei_1}), larger $(1+\psi
_{\vec{k}})$ would reduce total turbulent heating in 3-D
compared to 2-D. We have, however, the reverse by the following
reason. Heuristic arguments require the perpendicular rms
turbulent fields $\sim\sum_{\vec{k},\omega\neq0}
k_{\perp}^{2}|\delta\Phi_{\vec{k},\omega}|^{2}$ to be
approximately equal in 2-D and 3-D
\citep{DimantMilikh:JGR03,Dimant:Magnetosphere2011}. According
to Eq.~(\ref{diag_psi_summa_ei_2}), this leads to stronger
heating in 3-D compared to 2-D, entirely due to larger density
perturbations, $\sum_{\vec{k},\omega\neq0} |\delta
n_{\vec{k},\omega}|^{2}\propto\sum_{\vec{k},\omega\neq0}
(1+\psi_{\vec{k}})^{2}k_{\perp}^{2}|
\delta\Phi_{\vec{k},\omega}|^{2}$. The same pertains to the
non-linear currents discussed in Sect.~\ref{non-linear
current}. All these heuristic inferences have been confirmed by
our recent supercomputer PIC simulations
\citep{Oppenheim:Fully2011}.
The second term in the RHS of Eq.~(\ref{HH}), calculated
by expressing $\delta\vec{V}_{e,i}$ and $\vec{V}_{e,i0}$ in terms
of $\delta \vec{U}$ and $\vec{U}_{0}$, using Eq.~(\ref{V_perp_via_U}),
and $\vec{U}_{0}\times\hat{b}$ in terms of $\vec{E}_{0}$, using
Eq.~(\ref{ExB_via_U}), is
\begin{eqnarray}
2m_{i}&\!\!\!\!\!\!(&\!\!\!\!\!\!\nu_{i}\vec{V}_{i0}\cdot \langle\delta
n\delta\vec{V}_{i}\rangle\,)\nonumber\\
& \!= &\!\frac{2en_{0}(1+\kappa_{e}^{2})}{\kappa_{e}(\kappa_{i}+\kappa_{e})}
\sum
_{\vec{k},\omega\neq0}\frac{(\vec{k}\cdot\vec{E}_{0})(\vec{k}\cdot\vec{U}
_{0})}{(1+\psi_{\vec{k}})k_{\perp}^{2}}\left\vert \frac{\delta
n_{\vec {k},\omega}}{n_{0}}\right\vert ^{2
\label{``diag''_2}
\end{eqnarray}
and similar for electrons, with the symmetric replacement
$\kappa_{e,i}\leftrightarrow\kappa_{i,e}$. At higher altitudes,
where $\psi_{\perp}=(\kappa_{e}\kappa_{i})^{-1}<1$ and
$\kappa_{e}\gg1$, the ion component dominates over the electron
component, while at lower altitudes, $\psi_{\perp}>1$, the
reverse holds. Adding up the two plasma components gives
\begin{align}
&2m_{i}\nu_{i}(\vec{V}_{i0}\cdot\langle\delta n\delta\vec{V}_{i}\rangle) +
2m_{e}\nu_{e}(\vec{V}_{e0}\cdot\langle\delta n\delta\vec{V}_{e}
\rangle)\nonumber\\
& =2en_{0}\left( 1+\psi_{\perp}\right) \sum_{\vec{k},\omega\neq0}
\frac{(\vec{k}\cdot\vec{E}_{0})(\vec{k}\cdot\vec{U}_{0})}{(1+\psi_{\vec{k}
})k_{\perp}^{2}}\left\vert \frac{\delta n_{\vec{k},\omega}}{n_{0}}\right\vert
^{2}. \label{``diag''_combined
\end{align}
In the most of electrojet, these terms are small compared to
those in Eq.~(\ref{diag_psi_summa_ei}), although at the top of
the electrojet they can become comparable.
To verify that general Eqs.~(\ref{so_that_really}) and
(\ref{L_turb}) exactly hold in our quasilinear calculations, now we
proceed to calculating the LHS of Eq.~(\ref{HH}). Each of the
electron and ion terms,
$\langle\delta\vec{E}\cdot\delta\vec{j}_{e,i}\rangle$, can be
separated into two distinct parts,
\begin{equation}
\langle\delta\vec{E}\cdot\delta\vec{j}_{s}\rangle=\langle\delta\vec{E}
\cdot\delta\vec{j}_{s}\rangle_{1}+\langle\delta
\vec{E}\cdot\delta\vec{j}_{s}\rangle_{2}
\label{separa}
\end{equation}
where $\langle\delta\vec{E}\cdot\delta\vec{j}_{s}\rangle_{1}
\equiv-en_{0}
\langle\delta\vec{V}_{s}\cdot\delta\vec{E}\rangle$, $
\langle\delta\vec{E}\cdot\delta\vec{j}_{s}\rangle_{2}
\equiv-e\vec{V} _{e0}\cdot\langle\delta n\delta\vec{E}\rangle$,
and we neglect the cubically non-linear corrections,
$e\langle\delta n\delta\vec{V}_{s}\cdot\delta\vec{E}\rangle$.
Using Eqs.~(\ref{ExB_via_U}) and (\ref{delta_Vs}), we find that
the first term,
$\langle\delta\vec{E}\cdot\delta\vec{j}_{s}\rangle_{1}$, equals
the fixed-density frictional heating rate,
\begin{equation}
\langle\delta\vec{E}\cdot\delta\vec{j}_{s}\rangle_{1}=m_{s}\nu_{s
n_{0}\left\langle \delta V_{s}^{2}\right\rangle, \label{ravny
\end{equation}
while the combination of the electron and ion second terms in
Eq.~(\ref{separa}) yields
\begin{comment}
\begin{subequations}
\label{del_j.del_E_2
\begin{align}
& \langle\delta\vec{E}\cdot\delta\vec{j}_{e}\rangle_{2}=-\ \frac
{eBn_{0}(1+\kappa_{e}^{2})(1+\kappa_{i}^{2})}{\kappa_{i}\left( \kappa
_{e}+\kappa_{i}\right) ^{2}}\nonumber\\
& \times\sum_{\vec{k},\omega\neq0}\frac{(\vec{k}\cdot\vec{U}_{0})[\vec{k}\cdot
\vec{U}_{0}-\kappa_{i}\vec{k}\cdot(\vec{U}_{0}\times\hat{b})]}{(1+\psi
_{\vec{k}})k_{\perp}^{2}}\left\vert \frac{\delta n_{\vec{k}}}{n_{0
}\right\vert ^{2},\label{del_j.del_E_2_e}\\
& \langle\delta\vec{j}_{i}\cdot\delta\vec{E}\rangle_{2}=-\ \frac
{eBn_{0}(1+\kappa_{e}^{2})(1+\kappa_{i}^{2})}{\kappa_{e}\left( \kappa
_{e}+\kappa_{i}\right) ^{2}}\nonumber\\
& \times\sum_{\vec{k},\omega\neq0}\frac{(\vec{k}\cdot\vec{U}_{0})[\vec{k}\cdot
\vec{U}_{0}+\kappa_{e}\vec{k}\cdot(\vec{U}_{0}\times\hat{b})]}{(1+\psi
_{\vec{k}})k_{\perp}^{2}}\left\vert \frac{\delta n_{\vec{k}}}{n_{0
}\right\vert ^{2}. \label{del_j.del_E_2_i
\end{align}
\end{subequations}
The terms $\propto\vec{k}\cdot(\vec{U}_{0}\times\hat{b})$ in $\langle
\delta\vec{j}_{e}\cdot\delta\vec{E}\rangle_{2}$ and $\langle\delta\vec{j
_{i}\cdot\delta\vec{E}\rangle_{2}$ cancel each other, so tha
\end{comment}
\begin{align}
& \langle\delta\vec{E}\cdot\delta\vec{j}_{e}\rangle_{2}+\langle\delta\vec{E}
\cdot\delta\vec
{j}_{i}\rangle_{2} \equiv\langle\delta\vec{E}\cdot\delta\vec{j}\rangle_{2}
\nonumber\\
& =-\ \frac{eBn_{0}\left( 1+\kappa_{e}^{2}\right) \left( 1+\kappa_{i
^{2}\right) \psi_{\perp}}{\left( \kappa_{e}+\kappa_{i}\right) }\sum
_{\vec{k},\omega\neq0}\frac{(\vec{k}\cdot\vec{U}_{0})^{2}}{(1+\psi_{\vec{k}
)k_{\perp}^{2}}\left\vert \frac{\delta n_{\vec{k}}}{n_{0}}\right\vert ^{2}
\label{del_22
\end{align}
By combining Eqs.~(\ref{diag_psi_summa_ei}) and
(\ref{separa})--(\ref{del_22}), we verify that, in accord with
Eq.~(\ref{so_that_really}),
$\langle\delta\vec{E}\cdot\delta\vec{j}\rangle=0$.
\begin{comment}
\begin{equation}
\langle\delta\vec{j}_{e}\cdot\delta\vec{E}\rangle_{1}+\langle\delta\vec{j
_{i}\cdot\delta\vec{E}\rangle_{1}+\langle\delta\vec{j}_{e}\cdot\delta\vec
{E}\rangle_{2}+\langle\delta\vec{j}_{i}\cdot\delta\vec{E}\rangle_{2}=0.
\label{balancing
\end{equation}
\end{comment}
Due to this, turbulent Joule heating, $L_{\mathrm{turb}}\equiv
\langle \vec{E}\cdot\vec{j}\rangle - \vec{E}_0\cdot\vec{j}_0 =
\vec{E}_0\cdot\vec{j}^{\mathrm{NC}} +
\langle\delta\vec{E}\cdot\delta\vec{j}\rangle$, in accord with
Eq.~(\ref{L_turb}), equals the total average work of the
external electric field on the total non-linear current,
$P_{\mathrm{NC}}\equiv\vec{E}_0\cdot\vec{j}^{\mathrm{NC}}$. This
quantity, in turn, should equal the total frictional heating
source given by combining the RHS of Eq.~(\ref{HH}) for all
plasma species. Indeed, using Eq.~(\ref{j_NL}) for
$\vec{j}^{\mathrm{NC}}$, we obtain
\begin{comment}
\begin{subequations}
\label{P_NL_separate
\begin{align}
& P_{\mathrm{NC}e}=\frac{en_{0}\left( 1+\kappa_{i}^{2}\right) (\vec{k
\cdot\vec{U}_{0})}{\kappa_{i}(\kappa_{i}+\kappa_{e})(\kappa_{e}-\kappa_{i
)}\sum_{\vec{k},\omega\neq0}\frac{1}{(1+\psi_{\vec{k}})k_{\perp}^{2}}\left\vert
\frac{\delta n_{\vec{k}}}{n_{0}}\right\vert ^{2}\nonumber\\
& \times\left[ \frac{\kappa_{e}(1+\kappa_{i}^{2})(1+\kappa_{e}^{2})(\vec
{k}\cdot\vec{U}_{0})B}{\kappa_{e}+\kappa_{i}}+\left( 2\kappa_{e}-\kappa
_{i}+\kappa_{e}^{2}\kappa_{i}\right) (\vec{k}\cdot\vec{E}_{0})\right]
,\label{P_NL_e_separate}\\
& P_{\mathrm{NC}i}=-\ \frac{en_{0}(1+\kappa_{e}^{2})(\vec{k}\cdot\vec{U
_{0})}{\kappa_{e}(\kappa_{i}+\kappa_{e})(\kappa_{e}-\kappa_{i})}\sum_{\vec{k},
\omega\neq0}
\frac{1}{(1+\psi_{\vec{k}})k_{\perp}^{2}}\left\vert \frac{\delta
n_{\vec{k}}}{n_{0}}\right\vert ^{2}\nonumber\\
& \times\left[ \frac{\kappa_{i}\left( 1+\kappa_{i}^{2}\right)
(1+\kappa_{e}^{2})(\vec{k}\cdot\vec{U}_{0})B}{\kappa_{e}+\kappa_{i}}+\left(
2\kappa_{i}-\kappa_{e}+\kappa_{i}^{2}\kappa_{e}\right) (\vec{k}\cdot\vec
{E}_{0})\right] ,
\label{P_NL_i_separate
\end{align}
\end{subequations}
with the total
\end{comment}
\begin{align}
P_{\mathrm{NC}} &
=\frac{eBn_{0}
\psi_{\perp}(1+\kappa_{e}^{2})(1+\kappa_{i}^{2})}{\kappa_{e}+\kappa_{i}
\sum_{\vec{k},\omega\neq0}\frac{(\vec{k}\cdot\vec{U}_{0})^{2}}{(1+\psi_{\vec{k
})k_{\perp}^{2}}\left\vert \frac{\delta n_{\vec{k}}}{n_{0}}\right\vert
^{2}\nonumber\\
& +2(1+\psi_{\perp})en_{0}\sum_{\vec{k},\omega\neq0}\frac{(\vec{k}\cdot\vec{E
_{0})(\vec{k}\cdot\vec{U}_{0})}{(1+\psi_{\vec{k}})k_{\perp}^{2}}\left\vert
\frac{\delta n_{\vec{k}}}{n_{0}}\right\vert ^{2}, \label{total_P_NL_again
\end{align}
which equals the total turbulent heating found by combining
Eqs.~(\ref{diag_psi_summa_ei}) and (\ref{``diag''_combined}).
\begin{comment}
\begin{align}
P_{\mathrm{NC}} & =m_{e}\nu_{e}n_{0}\left\langle \delta V_{e}^{2
\right\rangle +m_{i}\nu_{i}n_{0}\left\langle \delta V_{i}^{2}\right\rangle
\nonumber\\
& +2m_{e}\nu_{e}(\vec{V}_{e0}\cdot\langle\delta n\delta\vec{V}_{e
\rangle)+2m_{i}\nu_{i}(\vec{V}_{i0}\cdot\langle\delta n\delta\vec{V
_{i}\rangle), \label{P_NL=H_NL
\end{align}
\end{comment}
Thus, our fluid-model calculations for arbitrarily magnetized
plasma particles in the quasilinear approximation have fully
confirmed the exact general relations derived in
Sect.~\ref{General consideration} exclusively from Maxwell's
equations. The partial relations for various energy conversion
terms show how the deposited field energy is divided between
the electron and ion heating channels and allow one to properly
interpret the corresponding terms.
\section{Global Energy Flow\label{Global Energy Flow}}
Now we discuss how the energy from the Earth's magnetosphere is deposited to
the \emph{E}/\emph{D}-region electrojet. From the above treatment it is clear
that the global energy flow should be analyzed based on the total average
currents that include the NCs. To obtain energy
balance equations similar to Eqs.~(\ref{Poynting's}) for spatially and
temporally averaged fields introduced in Eq.~(\ref{perturbations}), we use
separate linear Maxwell's equations for the average fields and for disturbances
caused by the total average currents, $\langle\vec{j}\rangle=\vec
{j}^{\mathrm{tot}}\equiv\vec{j}_{0}+\vec{j}_{\mathrm{NL}}$, as we did in
deriving Eq.~(\ref{Poynting's_perturb}) for the wave perturbations. As a
result, ignoring the induction component of $\langle\vec{E}\rangle=\vec{E
_{0}$, i.e., assuming $\nabla_{\vec{r}}\times\vec{E}_{0}=0$, we obtain
\begin{equation}
\partial_{\tau}\left[ \frac{\varepsilon_{0}\left( E_{0}^{2}+c^{2}\langle
B\rangle^{2}\right) }{2}\right] +\nabla_{\vec{r}}\cdot\Delta\vec{S}=-\vec
{E}_{0}\cdot\vec{j}^{\mathrm{tot}}, \label{field_energy_transport
\end{equation}
where $\tau$, $\vec{r}$ are the large-scale variables introduced in
Sect.~\ref{General consideration}, and
\begin{equation}
\Delta\vec{S}\equiv\frac{\vec{E}_{0}\times\Delta\vec{B}}{\mu_{0}}=\epsilon
_{0}c^{2}\vec{E}_{0}\times\Delta\vec{B}. \label{Delta_S
\end{equation}
Here
$\Delta\vec{B}\equiv\langle\vec{B}\rangle-\vec{B}_{0}$ is the quasi-stationary large-scale magnetic field disturbance, where by $\vec{B}_{0}$
we mean the geomagnetic field undisturbed by the electrojets, $\nabla_{\vec
{r}}\times\vec{B}_{0}=0$, so that $\nabla_{\vec{r}}\cdot\left( \vec{E
_{0}\times\vec{B}_{0}\right) =0$. The disturbance
$\Delta\vec{B}$ is caused by the total average electrojet
currents,
$\nabla\times\Delta\vec{B}\approx\mu_{0}\vec{j}^{\mathrm{tot}}$.
It is usually so small, $|\Delta\vec{B}|\lesssim10^{-3}B_{0}$,
that can be neglected in the expression for the magnetic energy
density, $\langle B\rangle^{2}\approx B_{0}^{2}$. However,
$\Delta\vec{B}$ is crucial for the Poynting flux
$\Delta\vec{S}$ that provides a downward flow of field energy
from the magnetosphere to the \emph{E}/\emph{D}-region
ionosphere. Further, bearing in
mind that $\vec{j}^{\mathrm{tot}}=\overleftrightarrow{\sigma^{\mathrm{tot}
}\cdot\vec{E}_{0}$, the work of the external field on the total current in the
RHS\ of Eq.~(\ref{field_energy_transport}) can be written as $\vec{E}_{0
\cdot\vec{j}^{\mathrm{tot}}=\sigma_{\mathrm{P}}^{\mathrm{tot}}E_{0}^{2}$, where $\overleftrightarrow{\sigma^{\mathrm{tot}}}$ and $\sigma_{\mathrm{P}}^{\mathrm{tot}}$ are the total conductivity tensor and Pedersen conductivity, respectively \citep{Dimant:Magnetosphere2011}. In the turbulent electrojet, the quantity $\sigma_{\mathrm{P}}^{\mathrm{tot}}E_{0}^{2}$ represents the combined laminar and turbulent Joule heating. Bearing in mind quasi-stationary
conditions, from Eq.~(\ref{field_energy_transport}) we estimate the magnitude
of the Poynting flux on top of the electrojet as $|\Delta\vec{S}|\sim
\sigma_{\mathrm{P}}^{\mathrm{tot}}E_{0}^{2}L_{\parallel}$, where
$L_{\parallel}$ is the characteristic size of the electrojet along the nearly
vertical magnetic field. The corresponding disturbance of the magnetic field,
$\Delta\vec{B}$, has a significant component perpendicular to both $\vec
{E}_{0}$ and $\vec{B}_{0}$ with the magnitude $\sim\mu_{0}\sigma_{\mathrm{P
}^{\mathrm{tot}}E_{0}L_{\parallel}$.
Now we do the same for the plasma. Averaging
Eq.~(\ref{particle_energy _conserva}) over small turbulent
scales and neglecting
$\langle\delta\vec{E}\cdot\delta\vec{j}\rangle$, in accord with
the analysis of Sect.~\ref{General consideration}, we obtain
\begin{equation}
\partial_{\tau}\sum_{s}\langle\mathcal{E}_{s}\rangle+\nabla_{r
\cdot\sum_{s}\langle\vec{K}_{s}\rangle=\vec{E}_{0}\cdot\vec
{j}^{\mathrm{tot}}+\sum_{s}\langle L_{s}\rangle.
\label{particle_energy_transport
\end{equation}
While the RHS of Eq.~(\ref{field_energy_transport}) has only one term
describing the total field energy loss to charged particles, the RHS\ of
Eq.~(\ref{particle_energy_transport}) has two terms. The first term, $\vec{E}_{0}\cdot\vec{j}^{\mathrm{tot}}=\sigma_{\mathrm{P}
^{\mathrm{tot}}E_{0}^{2}$, describes the total Joule heating
rate due to the total energy input from the fields. The second term describes
particle energy dissipation in the abundant neutral atmosphere. For
particle fluxes dominated by the thermal bulk, the two terms almost
cancel each other locally, so that the flux divergence term $\nabla_{r
\cdot\sum_{s}\langle\vec{K}_{s}\rangle$ is expected to be much
less than $\nabla_{\vec{r}}\cdot\Delta\vec{S}$ in
Eq.~(\ref{field_energy_transport}). This means that the energy
transport of ionospheric particles should be much weaker than
that of fields. The reason why the particle energy transport
plays a minor role is associated with relatively short mean
electron and ion free paths, even in the almost vertical
direction along $\vec{B}_{0}$. We should bear in mind, however,
that the fluxes $\langle\vec{K}_{s}\rangle$ can also include
high-energy precipitating particles that may comprise a
significant part of the field-aligned (Birkeland) currents
within auroral arcs. These occasions may break the nearly
perfect local balance between the Joule heating and atmospheric
cooling. Although in these cases $|\nabla_{r}\cdot\sum_{s
}\langle\vec{K}_{s}\rangle|$ should still remain less than
$|\nabla _{\vec{r}}\cdot\Delta\vec{S}|$, the two fluxes may be
closer to each other than those for the nearly local
thermal-bulk dominated balance.
Now we compare the total field and particle fluxes along the nearly vertical
magnetic field in greater detail. We can estimate the particle energy flux
magnitude as $| \sum_{s}\vec{K}_{s}| \sim
\langle\mathcal{E\rangle}j_{\parallel}$, where $\langle\mathcal{E\rangle}$ is
an effective energy of charged particles. Using the charge flow conservation,
$\nabla\cdot\vec{j}=0$, we estimate the parallel current density as
$j_{\parallel}\sim L_{\parallel}\sigma_{P}E_{0}/L_{\mathrm{P}}$, where
$L_{\mathrm{P}}$ is the characteristic scale of current density variations in
the Pedersen direction. As a result, we have $|\sum_{s}\langle\vec
{K}_{s}\mathcal{\rangle}|/|\Delta\vec{S}|\sim\langle\mathcal{E\rangle
}/(e|\Delta\Phi|)$, where $\Delta\Phi\sim L_{\mathrm{P}}E_{0}$ is the
characteristic cross-polar cap potential. Typical values of $\Delta\Phi$ are
$\sim100$~kV. If this current is mainly provided by thermal bulk particles
then their energy is many orders of magnitude smaller, $\langle
\mathcal{E\rangle}\lesssim1$~eV. Precipitating particles can have a rather
high energy, $\langle\mathcal{E\rangle}\lesssim30$~keV \citep[e.g.,][]{Ashrafi2005:Comparison}, so that the particle energy flux can sometimes be comparable to the Poynting flux.
\citet{Buchert:Effect2006} in their analysis of global energy flow between the
magnetosphere and \emph{E}-region ionosphere put more stress on the Birkeland
currents. These field-aligned currents are important for charge conservation
and they provide the MI coupling via particle
precipitation. Furthermore, electron fluxes along $\vec{B}_{0}$, because of
their high parallel mobility, ensure effective mapping of the electric field
from the magnetosphere to lower ionosphere. From the energy transport
viewpoint, however, our analysis gives preference to the Poynting flux.
It is interesting to note that 2-D or 3-D simulations of the \emph{E
/\emph{D}-region instabilities in purely periodic boxes have
always generated the DC currents $\vec{j}^{\mathrm{tot}}$ in a
plane perpendicular to $\vec {B}_{0}$. According to Maxwell's
equations, $\vec{j}^{\mathrm{tot}}$ must generate
quasi-stationary loop-like magnetic disturbances
$\Delta\vec{B}$, which in the employed electrostatic codes are
ignored by definition. Furthermore, such $\Delta\vec{B}$ would
even violate the imposed periodicity, unless one presumes on
the 3-D box boundaries surface currents that exactly balance
the volumetric currents. If, however, one did not ignore
$\Delta\vec{B}$ then the corresponding Poynting flux,
$c^{2}\epsilon_{0}\vec {E}_{0}\times\Delta\vec{B}$, would be
directed inward the box and provide the required energy input.
\section{Summary and Conclusions}
Plasma turbulence generated by \emph{E}/\emph{D}
-region instabilities can contribute significantly to the global energy
exchange between the magnetosphere and ionosphere. The spatially-temporally averaged energy deposit in the turbulent electrojet is
given by $\vec{E}_{0}\cdot\vec{j}^{\mathrm{tot}}$, where the total current
density $\vec{j}^{\mathrm{tot}}$, in addition to the regular current density, $\vec
{j}_{0}=\overleftrightarrow{\sigma_{\mathrm{P}}^{0}}\cdot\vec{E}_{0}$, includes
the non-linear current (NC) density, $\vec{j}^{\mathrm{NC}}$, caused by low-altitude plasma turbulence \citep{RogisterJamin:1975,Oppenheim:Evidence97}. The work of the external field on the NC, $\vec{E}_{0}\cdot\vec{j}^{\mathrm{NC}}$, provides the required energy input for anomalous (turbulent) heating of both electrons and ions \citep{Buchert:Effect2006}. In Sect.~\ref{General consideration} we prove this earlier conjecture using first principles with virtually no approximations. Specific fluid-model calculations for arbitrarily magnetized plasma in the quasilinear approximation show how exactly the deposited field energy is distributed between electrons and ions. This yields explicit NC expressions given by Eq.~(\ref{j_NL}), and turbulent sources of anomalous electron and ion heating given by Eq.~(\ref{diag_psi}). It has been known for a long time that anomalous electron heating (AEH) is largely a 3-D effect caused by turbulent electric fields parallel to $\vec{B}$ \citep{StMaLaher:85,Providakes:88,DimantMilikh:JGR03
Bahcivan:Parallel2006}. According to the new development, this
requires significantly larger
$\vec{E}_{0}\cdot\vec{j}^{\mathrm{NC}}$ in 3-D than in 2-D. Due
to the mirror symmetry along $\vec{B}$, neither $\vec{E}_0$ nor
$\vec{j}^{\mathrm{NC}}$ have any significant components in the
$\vec{B}$ direction. The difference in
$\vec{E}_{0}\cdot\vec{j}^{\mathrm{NC}}$ between the 3-D and 2-D
is entirely due to noticeably larger 3-D NC caused by larger
density perturbations, as explained in the text below
Eq.~(\ref{diag_psi_summa_ei}). This prediction has been
confirmed by our recent 2-D and 3-D PIC simulations
\citep{Oppenheim:Fully2011}. As discussed in the companion
paper \citep{Dimant:Magnetosphere2011}, a strong NC, either by
itself or combined with AEH
\citep{MilikhDimant:JGR03,MilikhGoncharenko:Anom2006}, can
significantly increase the global ionospheric conductances.
This may have serious implications for predictive modeling of
MI coupling and space weather.
\section*{Appendix: Validation of Electrostatic Appoximation\label{Direct Checking}}
Using some relations obtained in Sect.~\ref{General
consideration}, we now discuss the validity of the
electrostatic approximation in describing wave processes in the
lower ionosphere. To the best of our knowledge, the
electrostatic approximation for the \emph{E}-region
instabilities has always been employed but no detailed analysis
of its validity been done, especially for the general span of
magnetization conditions considered in this paper. There are
occasions when induction fields in the ionosphere play an
important role
\citep[e.g.,][]{Amm:Towards2008,Vanhamaki:Role2007}. The main
reason why one might question the validity of electrostatic
approximation in the description of
\textit{E}/\textit{D}-region wave processes is that even tiny
induction corrections to the wave electric field along
$\vec{B}$ could modify the small parallel turbulent electric
field largely responsible for AEH.
To estimate a possible non-electrostatic component of a
separate wave field harmonic, $\delta\vec{E}_{\vec{k},\omega}$,
we split it into the electrostatic (curl-free) part,
$\delta\vec{E}_{\vec{k},\omega}^{\mathrm{ES}}\equiv(\vec
{k}\cdot\delta\vec{E}_{\vec{k},\omega})\vec{k}/k^{2}$, and the
induction part,
$\delta\vec{E}_{\vec{k},\omega}^{\mathrm{IND}}\equiv\delta\vec{E}_{\vec
{k},\omega}-\delta\vec{E}_{\vec{k},\omega}^{\mathrm{ES}}=-\vec{k}\times
(\vec{k}\times\delta\vec{E}_{\vec{k},\omega})/k^{2}$. Then
Eq.~(\ref{del_E}) yields
\begin{subequations}
\label{E_split
\begin{align}
\delta\vec{E}_{\vec{k},\omega}^{\mathrm{ES}} & =-\ \frac{i(\vec{k
\cdot\delta\vec{j}_{\vec{k},\omega})\vec{k}}{\varepsilon_{0}\omega k^{2
},\label{E_ES}\\
\delta\vec{E}_{\vec{k},\omega}^{\mathrm{IND}} & =-\ \frac{i\omega\vec
{k}\times(\vec{k}\times\delta\vec{j}_{\vec{k},\omega})}{\varepsilon_{0
(k^{2}c^{2}-\omega^{2})k^{2}}. \label{E_Ind
\end{align}
\end{subequations}
The wave current density, $\delta\vec{j}_{\vec{k},\omega}$, is
determined by an anisotropic response to the total turbulent
electric field, $\delta\vec {E}_{\vec{k},\omega}$. Even if the
turbulent field is largely electrostatic,
$\delta\vec{E}_{\vec{k},\omega}\simeq\delta\vec{E}_{\vec
{k},\omega}^{\mathrm{ES}}\parallel\vec{k}$, this anisotropy can
give rise to a non-negligible component of
$\delta\vec{j}_{\vec{k},\omega}$ which is not parallel to
$\vec{k}$. According to Eq.~(\ref{E_Ind}), this component
generates non-zero $\vec{k}
\times(\vec{k}\times\delta\vec{j}_{\vec{k},\omega})=\vec{k}(\vec{k}\cdot
\delta\vec{j}_{\vec{k},\omega})-k^{2}\delta\vec{j}_{\vec{k},\omega}$,
i.e., finite $\delta\vec{E}_{\vec{k},\omega}^{\mathrm{IND}}$.
Using a perturbation technique in which the zero-order wave
field is electrostatic,
$\delta\vec{E}_{\vec{k},\omega}\approx\delta\vec{E}_{\vec{k},\omega
}^{\mathrm{ES}}=-i\vec{k}\delta\Phi_{\vec{k},\omega}$, one can
easily estimate the next-order induction component,
$\delta\vec{E}_{\vec{k},\omega }^{\mathrm{IND}}$. Considering
the wave with $\vec{k}$ in the $x,z$-plane
($\hat{z}||\vec{B}_0$), neglecting density perturbations, and
applying to $\delta\vec{E}_{\vec{k},\omega}$ the regular linear
conductivity, we obtain
\[
\delta\vec{j}_{\vec{k},\omega}\approx\overleftrightarrow{\sigma^{0}
\cdot\delta\vec{E}_{\vec{k},\omega}^{\mathrm{ES}}=-i(\overleftrightarrow
{\sigma^{0}}\cdot\vec{k})\delta\Phi_{\vec{k},\omega}=-i\delta\Phi_{\vec
{k},\omega}\left[
\begin{array}
[c]{c
\sigma_{\mathrm{P}}^{0}k_{\perp}\\
-\sigma_{\mathrm{H}}^{0}k_{\perp}\\
\sigma_{\parallel}^{0}k_{\parallel
\end{array}
\right] ,
\]
where we have restricted ourselves to the regular conductivity
tensor $\overleftrightarrow{\sigma^{0}}$
\citep{Kelley:Ionosphere2009},
\begin{subequations}
\label{sigma_II,P,H
\begin{align}
\sigma_{\parallel}^{0} & \equiv\frac{\vec{j}_{0\parallel}\cdot\vec
{E}_{0\parallel}}{E_{0\parallel}^{2}}=\frac{(\kappa_{e}+\kappa_{i})ne
{B},\label{sigma_II}\\
\sigma_{\mathrm{P}}^{0} & \equiv\frac{\vec{j}_{0\perp}\cdot\vec{E}_{0\perp
}{E_{\perp}^{2}}=\frac{(\kappa_{e}+\kappa_{i})(1+\kappa_{i}\kappa_{e
)ne}{(1+\kappa_{e}^{2})(1+\kappa_{i}^{2})B},\label{sigma_P}\\
\sigma_{\mathrm{H}}^{0} & \equiv\frac{\vec{j}_{0\perp}\cdot(\vec{E
_{0}\times\hat{b})}{E_{0\perp}^{2}}=-\ \frac{\left( \kappa_{e}^{2}-\kappa
_{i}^{2}\right) ne}{(1+\kappa_{e}^{2})(1+\kappa_{i}^{2})B}. \label{sigma_H
\end{align}
\end{subequations}
Since the anomalous conductivity discussed in the companion paper
is of the same order of magnitude at most, then according to
Eq.~(\ref{E_split}) we obtain
\begin{align}
\delta\vec{E}_{\vec{k},\omega}^{\mathrm{ES}} & \simeq-\ \frac
{(\sigma_{\mathrm{P}}^{0}+\sigma_{\parallel}^{0}k_{\parallel}^{2}/k_{\perp
}^{2})\delta\Phi_{\vec{k},\omega}}{\varepsilon_{0}\omega}\left[
\begin{array}
[c]{c
k_{\perp}\\
0\\
k_{\parallel
\end{array}
\right] ,\label{ES_est}\\
\delta\vec{E}_{\vec{k},\omega}^{\mathrm{IND}} & \simeq-\ \frac{\omega
\delta\Phi_{\vec{k},\omega}}{\varepsilon_{0}k^{2}c^{2}}\left[
\begin{array}
[c]{c
-\ \frac{\sigma_{\parallel}^{0}k_{\parallel}^{2}}{k_{\perp}}\\
-\sigma_{\mathrm{H}}^{0}k_{\perp}\\
\sigma_{\parallel}^{0}k_{\parallel
\end{array}
\right] , \label{Ind_est
\end{align}
where we have neglected $\sigma_{\mathrm{P}}^{0}$ compared to $\sigma
_{\parallel}^{0}$ and $k_{\parallel}^{2}$ compared to $k_{\perp}^{2}\approx
k^{2}$. We also have taken into account that in \textit{E/D}-region processes the wave phase speed, $V_{\mathrm{ph}}=\omega/k$, is many
orders of magnitude less than the speed of light, so that $\omega^{2}$ in the denominator of Eq.~(\ref{E_Ind}) can always be neglected compared to $k^{2}c^{2}$.
Comparing Eqs.~(\ref{ES_est}) and (\ref{Ind_est}) for both
parallel and perpendicular to $\vec{B}_{0}$ components and
presuming
$\sigma_{\parallel}^{0}k_{\parallel}^{2}/k_{\perp}^{2}\lesssim\sigma
_{\mathrm{P}}^{0}$, we see that the smallness of
$|\delta\vec{E}_{\vec{k},\omega}^{\mathrm{IND}}/\delta\vec{E}_{\vec{k},\omega}
^{\mathrm{ES}}|$
is ensured by
\begin{equation}
\frac{V_{\mathrm{ph}}}{c}\ll\left( \frac{\sigma_{\mathrm{P}}^{0}
{\sigma_{\parallel}^{0}}\right) ^{1/2},\qquad\frac{V_{\mathrm{ph}}}{c
\ll\left( \frac{\sigma_{\mathrm{P}}^{0}}{\sigma_{\mathrm{H}}^{0}}\right)
^{1/2}. \label{provided
\end{equation}
The first inequality pertains to the parallel components of the turbulent
electric field, $\delta\vec{E}_{\vec{k},\omega\parallel}$, while the second,
less restrictive, inequality pertains to the corresponding perpendicular
components, $\delta\vec{E}_{\vec{k},\omega\perp}$. As mentioned above, breaking
of the parallel-field condition may be of
importance because $\delta\vec{E}_{\vec{k},\omega\parallel}$ is crucial for
AEH. For magnetized electrons, up to the top electrojet where $\kappa_{e}\gg1$
and $\kappa_{i}\sim1$, we have $\sigma_{\mathrm{P}}^{0}\sim\kappa_{i
ne/B$, $\sigma_{\mathrm{H}}^{0}\sim ne/B$, $\sigma_{\parallel}^{0
\approx\kappa_{e}ne/B$, so that the most restrictive first
inequality becomes
\begin{equation}
\frac{V_{\mathrm{ph}}}{c}\ll\left( \frac{\kappa_{i}}{\kappa_{e}}\right)
^{1/2}=\Theta_{0}\simeq1.4\times10^{-2}. \label{most_restrictive
\end{equation}
Even for an extremely strong convection electric field of
$E_{0}\simeq 150$\ mV/m, the wave phase velocity may reach
$3$~km/s at most, $V_{\mathrm{ph}}/c\simeq10^{-5}$, so that
this condition holds with a big reserve. Thus for typical
\textit{E/D}-region wave processes the electrostatic
approximation is always valid even for small parallel
electric-field components, $\delta\vec{E}_{\vec{k},\omega||}$.
\begin{acknowledgments}
This work was supported by National Science Foundation
Ionospheric Physics Grants No.~ATM-0442075 and ATM-0819914.
\end{acknowledgments}
|
\section{Introduction}
In 1989, M. Fisher {\it et. al.} predicted that the homogeneous
Bose-Hubbard model (BH)
exhibits the Superfluid-Mott insulator
(SF-MI) quantum phase transition \cite{BB:Fisher}. In 2002 the
transition between these two phases were observed experimentally
for the first time with cold atomic gases in the group of I.
Bloch, T. Esslinger and T. H\" ansch \cite{BB:Bloch00}. The
experimental realization of a dipolar Bose-Einstein condensate
(BEC) of Chromium by the group of T. Pfau \cite{BB:Pfau00,
BB:Pfau01, BB:Kaz01}, and the recent progresses in trapping and
cooling of dipolar molecules by the groups of D. Jin and J. Ye
\cite{BB:Molecule01, BB:Molecule02, BB:Molecule03}, have opened
the path towards ultra-cold quantum gases with dominant dipole
interactions. A natural evolution, and present challenge, on the
experimental side is then to load dipolar BECs into optical
lattices and study strongly correlated ultracold dipolar lattice
gases.
Studies of BH models with interactions extended to nearest
neighbors had pointed out that novel quantum phases, like
supersolid (SS) and checker board phases (CB) are expected
\cite{BB:Goral02, BB:Kovrizhin, BB:Batrouni, BB:Sengupta}. Due to
the long-range character of the dipole-dipole interaction, which
decays as the inverse cubic power of the distance, it is necessary
to include more than one nearest neighbor to have a faithful
quantitative description of dipolar systems. In fact, longer-range
interactions tend to allow for and stabilize more novel phases.
In this work we first study BH models with dipolar interactions,
going beyond the ground state search. We consider a
two-dimensional (2D) lattice where the dipoles are polarized
perpendicularly to the 2D plane, resulting in an isotropic
repulsive interaction. We use the mean-field approximations and a
Gutzwiller Ansatz which are quite accurate and suitable to
describe this system. We find that dipolar bosonic gas in 2D
lattices exhibits a multitude of insulating metastable states,
often competing with the ground state, similarly to a disordered
system. We study in detail the fate of these metastable states, in
particular what is their lifetime due to tunneling.
Moreover, we find that the ground state is characterized by
insulating checkerboard-like states with fractional filling
factors $\nu$ (average number of particles per site) that depend
on the cut-off used for the interaction range. We confirm this
prediction by studying the same system with Quantum Monte Carlo
methods (the Worm algorithm). In this case no cut-off for the
dipolar interaction is used, and we find evidence for a Devil' s
staircase in the ground state, i.e. insulating phases which appear
at all rational $\nu$ of the underlying lattice. We also find
regions of parameters where the ground state is a supersolid,
obtained by doping the solids either with particles or vacancies.
Recently \cite{BB:Cooper}, a complete devil' s staircase has been
predicted in the phase diagram of a one-dimensional dipolar Bose
gas.
In this work, we also investigate how the previous scenario
changes by considering a multi-layer structure. We focus on the
simplest situation composed of two 2D layers in which the dipoles
are polarized perpendicularly to the planes; the dipolar
interaction is then repulsive for particles laying on the same
plane, while it is attractive for particles at the same lattice
site on different layers. Instead we consider inter-layer
tunneling to be suppressed, which makes the system analogous to a
bosonic mixture in a 2D lattice. Our calculations show that
particles pair into composites, and demonstrate the existence of
the novel Pair Super Solid (PSS) quantum phase.
Moreover, we study a 2D lattice where the dipoles are free to
point in both directions perpendicularly to the plane, which
results in a nearest neighbor repulsive (attractive) interaction
for aligned (anti-aligned) dipoles. We find regions of parameters
where the ground state of the system exhibits insulating phases
with ferromagnetic or anti-ferromagnetic ordering, as well as with
rational values of the average magnetization. Evidence for the
existence of a Counterflow Super Solid (CSS) quantum phase is also
presented.
Our predictions have direct experimental consequences, and we hope
that they will be soon checked in experiments with ultracold
dipolar atomic and molecular gases.
Although this paper reports on many novel results and predictions,
it is written in a tutorial form. The reader will have a
possibility to learn first the basic introduction to dilute Bose
gases, and in particular dipolar Bose gases. We will explain very
carefully various mean field methods that can be applied to
describe approximately diluted dipolar Bose gases in 2D and 3D.
These methods are elementary, but become more and more complex
for the extended Hubbard models, and in particular for the models
with infinite range interactions. In the chapters about multiple
layers and up-down mixtures we will present and explain in detail
methods of deriving effective low-energy Hamiltonians using
second order degenerate perturbation theory. Last, but not least,
we will present a very pedagogical introduction to QMC methods:
path integral Monte Carlo method and the Worm algorithm.
\section{Dipolar Bose gas in optical lattices}
\label{CH:Dipolar1}
\subsection{Optical lattices}
\label{SEC:OL} An optical lattice is an artificial crystal of
light, resulting from the interference pattern of two or more
counter propagating laser beams \cite{BB:Bloch}. The wavelengths
$\lambda_i$ of the laser beams determine the spatial periodicity
of the crystal; for example, two lasers of equal wavelengths
$\lambda_x$ propagating along $x$ but in opposite directions,
produce a standing wave with an intensity pattern $I(x)$ which is
spatially periodic with periodicity $\lambda_x/2$. An optical
lattice can trap neutral atoms by exploiting the energy shift
induced by the radiation on the atomic internal energy levels.
The electric field $\myvec{E}(\myvec{r},t) =
2E_0\cos\left(\myvec{k}\cdot\myvec{r} - \omega_L t\right)$ of a
monochromatic laser oscillating with frequency $\omega_L$,
interacts with a neutral atom, of spatial dimensions much smaller
compared to the wavelengths of the light $\lambda_i = 2\pi/k_i,\;
(i=x,y,z)$, through the Hamiltonian
\begin{equation}
\label{EQ:H_int}
\hat{H}_{\rm int}(t) = -\myvec{d} \cdot \myvec{E}(\myvec{r},t),
\end{equation}
where $\myvec{d} = -e\sum_i\myvec{r}_i$ is the electric dipole
moment of the atom, $\myvec{r}_i$ the positions of the atomic
electrons of charge $e$. With Hamiltonian (\ref{EQ:H_int}), one
can easily calculate the energy correction to the ground state of
the atom, by means of perturbation theory. The first order
correction vanishes because the dipole operator is odd with
respect to space inversion ($\myvec{r}_i \rightarrow -
\myvec{r}_i$), therefore the first non zero contribution is given
by the second order correction
\begin{equation}
\Delta E(\myvec{r}) = -\frac{1}{2}\alpha(\omega_L) \langle \myvec{E}(\myvec{r},t)^2\rangle_t
\end{equation}
where
\begin{equation}
\label{EQ:Polarizability}
\alpha(\omega_L) = \sum_\gamma \mymod{\bra{\gamma}\myvec{d}\cdot \myvec{\hat{\epsilon}} \ket{g}}^2\left(\frac{1}{E_\gamma - E_g + \hbar\omega_L} + \frac{1}{E_\gamma - E_g - \hbar\omega_L}\right),
\end{equation}
is the atomic polarizability \cite{BB:Pethick,BB:Maciek}, and
$\langle \cdots \rangle_t$ denotes a time average over one
oscillation period of the electric field \footnote{more
specifically $\langle \cdots \rangle_t = \frac{1}{t}\int_0^t
\cdots \mathrm{d} t$ where $t = n\pi/\omega_L$, $n=1,2,\cdots$}. In the
last expression the energies in the denominators are the
unperturbed energies of the atom, where $g$ is the ground state,
the sum runs over all excited states $\gamma$, and
$\myvec{\hat{\epsilon}}$ is the unit vector in the direction of
the electric field. In a typical experiment the laser light is far
off resonance, which means that the laser frequency is close to
one of the unperturbed excited states (e.g $E_e = \hbar
\omega_e$), but does not induce any real transition. In such a
situation, one can take only the smallest of the denominators
(\ref{EQ:Polarizability}), and the polarizability becomes
inversely proportional to the laser detuning from resonance
$\hbar\Delta = \hbar\omega_L - (E_e - E_g)$
\begin{equation}
\label{EQ:OmegaL}
\alpha(\omega_L) \simeq - \frac{\mymod{\bra{e}\myvec{d}\cdot \myvec{\hat{\epsilon}} \ket{g}}^2}{\hbar\Delta}.
\end{equation}
In this situation, the energy shift is given by
\begin{equation}
\label{EQ:Shift}
\Delta E(\myvec{r}) = -\frac{1}{2} \alpha(\omega_L) \langle \myvec{E}(\myvec{r},t)^2 \rangle_t \propto \frac{I(\myvec{r})}{\hbar\Delta},
\end{equation}
where $I(\myvec{r})$ is the intensity of the laser. In the dressed
atom picture, the energy shift (\ref{EQ:Shift}) is interpreted as
an effective potential $V_{\rm opt}(\myvec{r}) = \Delta
E(\myvec{r})$, that follows the spatial pattern of the laser field
intensity, in which the atom moves. In this picture, the atom then
feels a force
\begin{equation}
\myvec{F}_{\rm dipole} = -\myvec{\nabla} V_{\rm opt}(\myvec{r}),
\end{equation}
that attracts it towards the regions of high intensity for the so
called red-detuned lasers (i.e. $\Delta < 0$), while a
blue-detuned light (i.e. $\Delta > 0$) pushes the atom out of the
regions of high intensity. In the literature this force is called
the dipole force, as it is the resulting interaction of the
induced atomic dipole moment with the spatially varying electric
field of the light. Note that in order to reduce heating caused by
inelastic scattering, i.e. photon absorption and spontaneous
emission processes, a large detuning is required because the
photon scattering rate scales as $I(\myvec{r})/\Delta^2$. In the
limit of large detuning an optical lattice is therefore
non-dissipative, which makes it a basic tool to manipulate cold
neutral atoms.
For example, the simplest case of a one-dimensional lattice is
obtained by the superposition of two lasers propagating in
opposite directions, with electric fields linearly polarized, say
in the $z$ direction,and given by
\begin{eqnarray}
E_z(x,t) &=& 2E_0\cos\left(k_x x - \omega_L t\right) + 2E_0\cos \left(-k_x x - \omega_L t\right) \nonumber \\
&=& 4E_0\cos\left(k_x x\right)\cos(\omega_L t).
\end{eqnarray}
The time average over one period of oscillation of the electric field gives then $\langle E_z(x,t)^2 \rangle_t = 2E_0\cos^2(k_x x)$, which yields
to the spatially varying optical potential
\begin{equation}
V_{\rm opt}(x) = V_{0,x}\cos^2(k_x x),
\end{equation}
with periodicity $\lambda_x/2 = \pi/k_x$, and $V_{0,x}=2E_0\alpha(\omega_L)$ from Eq. (\ref{EQ:OmegaL}).
The generalization to the two dimensional (2D) or three dimensional case (3D) is straightforward (see e.g. \cite{BB:Pethick}). For
example in Fig. \ref{FIG:Lattice} two different geometries are shown.
\begin{center}
\begin{figure}[h]
\begin{center}
\includegraphics[width=0.6\linewidth]{Lattice.eps}
\end{center}
\caption{Picture of an optical potential. (a) 2D square lattice of quasi 1D traps; (b) a 3D cubic lattice, picture courtesy of I. Bloch \cite{BB:Bloch}.}
\label{FIG:Lattice}
\end{figure}
\end{center}
\subsection{Theory of dilute Bose gases}
\label{SEC:TDBG}
In this section we recall some basic theory of a dilute gas of neutral bosonic particles at temperature $T$ well below the degeneracy temperature. The type of particles we consider here can be atoms or molecules.
For a dilute gas, the interparticle separation (typically of the
order of $10^2$ nm for alkali atoms \cite{BB:Pethick}) is an order
of magnitude larger than the length scales associated with the
atom-atom interaction. In other words, a dilute gas of density $n$
is a very rarefied gas in which the "spatial extension" of an atom
is much smaller than the average volume per particle $n^{-1}$.
Because of this condition, the two-body interaction dominates the
physics while three-body or more are very unlikely and essentially
not important. The two-body interatomic potential $V(\myvec{r})$
depends on the type of particles one considers, the relative
distance between the atoms $\myvec{r} = \myvec{r}_1 - \myvec{r}_2$
and their internal states. For alkali atoms, the potential is
strongly repulsive for small atomic separations while for large
atomic distances it is dominated by the van der Waals attractions
that decay as $-C_6/\myvec{r}^6$, where the coefficient $C_6$
depends on the atomic species.
Here we will consider only elastic scattering, where the internal
states of the two atoms do not change in the collision process. If
the temperature of the gas is very low, i.e. $T \rightarrow 0$,
the kinetic energy of the particles is very small compared to the
centrifugal barrier and only $s$-wave scattering takes place.
Therefore, the only important parameter is the scattering length
given by
\begin{equation}
\label{EQ:ScatteringL}
a_s = \frac{m}{4\pi\hbar^2}\int \mathrm{d}^3r V(\myvec{r}),
\end{equation}
with $m$ being the mass of the atoms. This quantity has the
dimensions of a length, and for $a_s > 0$ has the physical
interpretation of the radius the atoms would have if they were
considered to be perfect billiard balls. The condition for the
diluteness of the gas then reads
\begin{equation}
\label{EQ:GasParameter}
n|a_s|^3 \ll 1
\end{equation}
where $n$ is the density of the gas and $n|a_s|^3$ is called the
gas parameter. One can invert expression (\ref{EQ:ScatteringL})
and think of an effective contact interaction between the two
particles proportional to the scattering length, and given by
\begin{equation}
\label{EQ:ContactI}
V_{\rm eff}(\myvec{r}) = g\, \delta^{(3)}(\myvec{r}),
\end{equation}
where $g$ is defined as
\begin{equation}
\label{EQ:g}
g = \frac{4\pi\hbar^2 a_s}{m},
\end{equation}
and $\delta$ is the Dirac delta function. Note also that since the
effective interaction depends only on the scattering length, it is
repulsive (attractive) for positive (negative) $a$, and it can be
dynamically modified for example in alkali atoms just by varying
an external magnetic field near a Feshbach resonance.
\subsubsection{The Gross-Pitaevskii equation}{--- }
The quantum state of a gas of $N$ particles is described by the
many-body wavefunction $\Psi(\myvec{r}_{\rm 1},\myvec{r}_{\rm
2},...,\myvec{r}_{\rm N})$, and the time evolution of the system
is determined by the Schr\"odinger equation. In a BEC, one can
describe the dynamics of the condensate just through the
Gross-Pitaevskii (GP) equation \cite{BB:Pitaevskii,BB:Pethick}
given by
\begin{equation}
\label{EQ:GP}
i\hbar \frac{\partial}{\partial t} \Psi_{\rm 0} (\myvec{r},t) = \left(-\frac{\hbar^2\myvec{\nabla}^2}{2m} + V_{\rm ext} (\myvec{r})
+ g |\Psi_{\rm 0} (\myvec{r},t)|^2\right) \Psi_{\rm 0} (\myvec{r},t),
\end{equation}
where $\Psi_{\rm 0} (\myvec{r},t)$ is the BEC wavefunction, also called the order parameter. The interaction between particles has been taken into account in a mean-field approximation by the term $g |\Psi_{\rm 0} (\myvec{r},t)|^2$, where $g$ is given in Eq. \ref{EQ:g}. $V_{\rm ext} (\myvec{r})$ is an external trapping potential, and
the order parameter is normalized to the total number of particles, i.e. $N = \int \mathrm{d}^3r |\Psi_{\rm 0} (\myvec{r},t)|^2$.
Equation (\ref{EQ:GP}) was independently derived by Gross and Pitaevskii in 1961, it is one of the main theoretical tools for investigating dilute weakly interacting Bose gases at low temperatures, and it has the typical form of a mean field equation where the order
parameter must be calculated in a self-consistent way.
The GP equation has proven to be a very useful tool to describe the physics of weakly
interacting Bose-Einstein atomic condensates in the early ages of this field. With this
formalisms, and its extension to include small fluctuations given by Bogoliubov theory,
one can describe accurately, among others, the collective excitations of the systems, the
response to rotations including the formation of vortices, the propagation of sound, the
presence of dynamical instabilities. Generally speaking, the GP treatment is well suited
in the regime of full coherence, when a single macroscopically occupied matterwave
correctly describes the system. At the end of the '90, few years after the creation of
the first alkali BECs in the lab, the need of "going beyond GP" started to be very
strongly felt, due to the theoretical interest and experimental possibility of going into
the strongly correlated regime. In fact, the presence of strong interactions, strong
rotations and/or special trapping potentials can limit the validity of the GP equation.
For
instance a strong confinement in one or two dimensions can reduce the system to an
effectively 2D or 1D one. A strong rotation combined with interactions can lead to
quantum Hall physics. Also the presence of a deep optical lattices,
when the combined effect of interactions and trapping potential leads to a
"fragmentation" of the condensate, requires more sophisticated descriptions.
In this tutorial we are interested in describing the physics of
Bosons trapped in a periodic optical potential ($V_{\rm opt}$) and
eventually also confined in a magnetic harmonic trap ($V_{\rm
ho}$), the total external field being given by the sum
\begin{eqnarray}
V_{\rm ext} (\myvec{r}) &=& V_{\rm opt} (\myvec{r}) + V_{\rm ho} (\myvec{r}) \nonumber \\
&=& \sum_{i=x,y,z} V_{0,i} \cos^2 (k_{i} r_{i}) + \frac{1}{2}m \sum_{i=x,y,z} \omega_{\rm i}^2 r_{i}^2,\label{EQ:TrapPotential}
\end{eqnarray}
where $(V_{0,x},V_{0,y},V_{0,z})$ is the depth of the optical lattice in the three spatial directions and $(\omega_{\rm x},\omega_{\rm y},\omega_{\rm z})$
the frequencies of the harmonic trap. In order to describe the physics of Bosons trapped in the potential (\ref{EQ:TrapPotential}), we need to
"go beyond" the GP equation, and we will devote the following sections to this purpose.
\subsubsection{Bose-Hubbard model}{--- }
\label{SUBSEC:BHM}
The starting point of our discussion is Hamiltonian (\ref{EQ:Hamiltomian2NDQ}), written in the second quantization formalism in terms of the creation
and annihilation operators for Bosons, $\hat{\psi}^\dag(\myvec{r})$ and $\hat{\psi}(\myvec{r})$ respectively, and given by the expression
\begin{equation}
\label{EQ:Hamiltomian2NDQ}
\hat{H} = \int \mathrm{d}^3r \hat{\psi}^\dag(\myvec{r}) \left[-\frac{\hbar^2 \nabla^2}{2m} + V_{\rm ext}(\myvec{r})
+ \frac{g}{2}\hat{\psi}^\dag(\myvec{r})\hat{\psi}(\myvec{r}) - \mu \right] \hat{\psi}(\myvec{r}),
\end{equation}
where the first term in square brackets is the kinetic energy, $V_{\rm ext}(\myvec{r}) = V_{\rm opt} (\myvec{r}) + V_{\rm ho} (\myvec{r})$ is the external trapping potential (\ref{EQ:TrapPotential}) and we have used the simplified contact interaction (\ref{EQ:ContactI}). We work in the grand canonical ensemble and the chemical
potential $\mu$ fixes the total number of particles. Additionally, we assume the harmonic confinement to change on a scale larger than the one of the optical lattice, such that we can consider the effect of the magnetic trapping to be constant over a single site of the lattice.
In this formalism, the field operators can be written in the basis of single-particle wave functions $\{\Phi_n(\myvec{r})\}_{n}$,
where $n$ is a complete set of single particle quantum numbers
\begin{equation}
\label{EQ:Fields}
\eqalign{
\hat{\psi}(\myvec{r}) &= \sum_{n} \Phi_n(\myvec{r}) \hat{a}_n \\
\hat{\psi}^\dag(\myvec{r}) &= \sum_{n} \Phi^*_n(\myvec{r}) \hat{a}^\dag_n,
}
\end{equation}
with $\hat{a}^\dag_n$ and $\hat{a}_n$ being the creation and annihilation operators for the mode $n$, i.e. $\hat{a}^\dag_n \ket{n} = \sqrt{n+1}\ket{n+1}$ and $\hat{a}_n \ket{n} = \sqrt{n}\ket{n-1}$. Also, the field operators satisfy the usual commutation relations for Bosons
\begin{equation}
\eqalign{
[\hat{\psi}(\myvec{r}),\hat{\psi}^\dag(\myvec{r^\prime})]
&= \sum_{n=0}^\infty\Phi_n(\myvec{r})\Phi^*_n(\myvec{r^\prime}) = \delta^{3}(\myvec{r} - \myvec{r^\prime}), \\
[\hat{\psi}(\myvec{r}),\hat{\psi}(\myvec{r^\prime})] &= [\hat{\psi}^\dag(\myvec{r}),\hat{\psi}^\dag(\myvec{r^\prime})] = 0.
}
\end{equation}
It is well known \cite{BB:Auerbach}, that the spectrum of a single particle in a periodic potential is
characterized by bands of allowed energies and energy gaps, and the single particle wave functions
are described by Bloch functions $\Phi_{\alpha\myvec{k}}(\myvec{r})$ with band index $\alpha$ and quasi-momentum $\hbar\myvec{k}$.
Alternatively, there exists a complementary single-particle basis given by the Wannier functions \cite{BB:Auerbach,BB:Korepin}
$w_\alpha(\myvec{r} - \myvec{R}_i)$, where $\myvec{R}_i$ is a lattice vector pointing at site $i$ and $w_\alpha(\myvec{r})$ are defined as
the Fourier transform of Bloch functions
\begin{equation}
w_\alpha(\myvec{r}) = \frac{1}{\sqrt{N_S}}\sum_{\myvec{k}}e^{-i\myvec{k}\cdot\myvec{r}}\Phi_{\alpha\myvec{k}}(\myvec{r}),
\end{equation}
where $N_S$, is the total number of sites in the lattice. The Wannier functions form a complete orthonormal set, so one may write the field operators
(\ref{EQ:Fields}) as
\begin{equation}
\label{EQ:OPWannier}
\eqalign{
\hat{\psi}(\myvec{r}) &= \sum_{\alpha\myvec{k}} \Phi_{\alpha\myvec{k}}(\myvec{r}) \hat{a}_{\alpha\myvec{k}} = \sum_{\alpha,i} w_\alpha(\myvec{r} - \myvec{R}_i) \hat{a}_{\alpha,i} \\
\hat{\psi}^\dag(\myvec{r}) &= \sum_{\alpha\myvec{k}} \Phi^*_{\alpha\myvec{k}}(\myvec{r}) \hat{a}^\dag_{\alpha\myvec{k}} = \sum_{\alpha,i} w^*_\alpha(\myvec{r} - \myvec{R}_i) \hat{a}^\dag_{\alpha,i}.
}
\end{equation}
Wannier functions are useful in the case of deep optical lattices where tight binding approximation apply. The big advantage of using Wannier functions
$w_\alpha(\myvec{r} - \myvec{R}_i)$ is that they are localized and centered around the lattice site pointed by $\myvec{R}_i$.
If the temperature of the system is low enough, and the interactions between the particles is not sufficient to induce transitions between the bands, one may restrict only to the first Bloch band because the particles have insufficient energy to overcome the gap that separates the first band from the others. This amounts to keep
in (\ref{EQ:OPWannier}) only the lowest of the $\alpha$ indices, which we omit for simplicity of notation in the following. Therefore the Hamiltonian (\ref{EQ:Hamiltomian2NDQ}) becomes
\begin{equation}
\hat{H} = -\sum_{i,j} J_{ij} \; \hat{a}^\dag_i \hat{a}_j + \sum_{i,j,k,l}\frac{U_{ijkl}}{2} \; \hat{a}^\dag_i \hat{a}^\dag_j \hat{a}_k \hat{a}_l - \sum_{i,j} \mu_{ij} \; \hat{a}^\dag_i \hat{a}_j,
\end{equation}
where the quantities in the sums are given by
\begin{eqnarray}
J_{ij} &=& -\int \mathrm{d}^3r w^*(\myvec{r} - \myvec{R}_i) \left[-\frac{\hbar^2 \nabla^2}{2m} + V_{\rm opt}(\myvec{r})\right]w(\myvec{r} - \myvec{R}_j) \label{EQ:Coeff01}\\
U_{ijkl} &=& g \int \mathrm{d}^3r w^*(\myvec{r} - \myvec{R}_i)w^*(\myvec{r} - \myvec{R}_j)w(\myvec{r} - \myvec{R}_k)w(\myvec{r} - \myvec{R}_l) \label{EQ:Coeff02} \\
\mu_{ij} &=& \int \mathrm{d}^3x w^*(\myvec{r} - \myvec{R}_i)\left[\mu - V_{\rm ho}(\myvec{r})\right]w(\myvec{r} - \myvec{R}_j).\label{EQ:Coeff03}
\end{eqnarray}
The Wannier functions are localized on the lattice sites, the deeper the lattice the more localized they are. For a sufficiently deep optical potential, in Eq. (\ref{EQ:Coeff02}) and (\ref{EQ:Coeff03}) the dominant contributions are given by $U_{iiii}$ and $\mu_{ii}$.
For the kinetic part (\ref{EQ:Coeff01}), there is a constant contribution given by $J_{ii}$ and due to the presence of the derivative in the integration, there is also a positive matrix element for nearest neighboring sites $J_{ij} > 0$. The two situations are qualitatively shown in Fig. (\ref{FIG:Wannier}) where we have approximated
the Wannier functions with two Gaussians respectively localized at site $i$ and $j$ of the lattice. However, we stress that the picture provided by Gaussian functions is only
qualitative. In fact, in order to be quantitatively correct, one needs to calculate the proper matrix elements with Wannier functions.
\begin{center}
\begin{figure}[h]
\begin{center}
\includegraphics[width=1\linewidth]{Wannier.eps}
\end{center}
\caption{(a) Two Gaussians localized on neighboring sites $i$ and $j$ of an optical lattice having negligible overlap. (b) The first derivatives of the
Gaussian functions instead, show a negative overlap in the region indicated by the arrow, which leads to a positive matrix element $J_{ij} > 0$.}
\label{FIG:Wannier}
\end{figure}
\end{center}
With the above considerations, we can now write the celebrated Bose-Hubbard Hamiltonian in the form
\begin{equation}
\label{EQ:HamiltonianBH}
\hat{H}_{\rm BH} = - J \sum_{\langle ij \rangle} \; \hat{a}^\dag_i \hat{a}_j + \frac{U}{2}\sum_i \hat{n}_i(\hat{n}_i-1) - \sum_i \mu_i \hat{n}_i,
\end{equation}
where $\langle ij \rangle$ indicates sum over nearest neighbors, the tunneling coefficient $J = J_{ij} = J_{ji}$ for hermiticity, the on-site interaction
$U = g\int \mathrm{d}^3r \mymod{w(\myvec{r})}^4$, $\hat{n}_i = \hat{a}^\dag_i \hat{a}_i$ is the number operator at site $i$, and we have neglected $J_{ii}$ since
it gives a constant contribution for each site.
The harmonic confinement, since it is assumed to be constant across one lattice site, has been taken into account in the chemical potential as
\begin{equation}
\label{EQ:LocalMu}
\mu_i = \mu - \frac{1}{2}m\vec{\omega}^{\;2}\cdot(\myvec{R}_i - \myvec{R_0})^2,
\end{equation}
where $\myvec{R_0}$ is the center of the harmonic trap with frequencies given by $\vec{\omega} = (\omega_{\rm x},\omega_{\rm y},\omega_{\rm z})$ in the three directions. The second
term on the right hand side of Eq. (\ref{EQ:LocalMu}) is practically a chemical potential that differs from site to site and it is often called the local chemical potential.
For a one dimensional optical lattice $V_{\rm opt}(x) = V_0\sin^2(kx)$ with wavevector $k = 2\pi/\lambda$, Fig. \ref{FIG:Jaksch} shows
both the on-site interaction $U$ (solid line) and the tunneling coefficient $J$ (dashed line) as a function of the optical lattice depth $V_0$, where all the
quantities are measured in terms of the recoil energy $E_R = \hbar^2k^2/2m$, that is the energy acquired by the atom after absorbing a photon with
momentum $\hbar k$. The lattice parameters $U$ and $J$ were calculated numerically in e.g. \cite{BB:Jacksch} for different values of $V_0$.
From Fig. \ref{FIG:Jaksch} (b), it is clear that it is possible to change the tunneling coefficient $J$ over a wide range, going from a situation of
practically isolated lattice sites at $V_0 = 25 E_R$ up to a regime in which particles can tunnel from site to site at $V_0 = 5 E_R$, only by changing
the optical potential depth by a few tens of recoil energies, and leaving the on-site interaction $U$ practically unchanged.
\begin{center}
\begin{figure}[h]
\begin{center}
\includegraphics[width=0.6\linewidth]{Jaksch.eps}
\end{center}
\caption{(a) Schematic representation of a 1D optical lattice, where $\varepsilon$ is a local chemical potential; (b) scaled on-site $U$ (solid line) and tunneling coefficient $J$ (dashed line)
dependence on the optical potential depth $V_0$. The on-site interaction is multiplied by $a/a_s (\gg 1)$, where $a = \lambda/2$ is the lattice period and $a_s$ is the
s-wave scattering length for atoms of equal mass $m$. Figure courtesy of D. Jaksch \cite{BB:Jacksch}.}
\label{FIG:Jaksch}
\end{figure}
\end{center}
\subsection{Dipolar Bose gas}
\label{SEC:DipolarBoseGas}
\subsubsection{Properties of the dipole-dipole interaction}{--- }
Two particles $1$ and $2$ in a three dimensional space, at relative distance $\myvec{r}$ and with dipole moments
along the unit vectors $\myvec{e_1}$ and $\myvec{e_2}$ as in Fig. \ref{FIG:Thierry} (a), interact through the dipole-dipole interaction such that
their interaction energy is given by
\begin{equation}
\label{EQ:Dipole}
U_{\rm dd} (\myvec{r}) = \frac{C_{\rm dd}}{4\pi} \frac{(\myvec{e}_1 \cdot \myvec{e}_2)r^2 - 3(\myvec{e}_1 \cdot \myvec{r})(\myvec{e}_2 \cdot \myvec{r})}{r^5},
\end{equation}
where $r = |\myvec{r}|$, and $U_{\rm dd} (\myvec{r}) = U_{\rm dd} (\myvec{-r})$.
The dipolar coupling constant $C_{\rm dd}$ is different for particles having a permanent magnetic dipole
moment $\mu$, and for particles having a permanent electric dipole moment $d$, and is respectively given by
\begin{equation}
\label{EQ:Cdd}
C_{\rm dd} = \left\{
\begin{tabular}{ll}
$\mu_0 \mu^2$ & magnetic \\
$d^2/\varepsilon_0$ & electric,
\end{tabular}
\right.
\end{equation}
where $\mu_0$ is the vacuum permeability, and $\varepsilon_0$ is the vacuum permittivity.
The dipole-dipole interaction (\ref{EQ:Dipole}) has a {\it long-range} character; this is because at large distances it decays as $U_{\rm dd} \sim 1/r^3$,
contrary to the typical van der Waals potential that behaves like $U_{\rm vdW} \sim -1/r^6$. Also, from (\ref{EQ:Dipole}) it is easy to see the
{\it anisotropic} property of this interaction; for polarized atoms, i.e. all dipoles pointing in the same direction, the interaction reduces to
\begin{equation}
\label{EQ:DipoleI}
U_{\rm dd} (\myvec{r}) = \frac{C_{\rm dd}}{4\pi} \frac{1-3\cos^2\theta}{r^3},
\end{equation}
where $\theta$ is the angle between the dipole and the relative distance of the particles, as in Fig \ref{FIG:Thierry} (b). The interaction is repulsive for
$\theta = \pi/2$ as the example of Fig \ref{FIG:Thierry} (c), and attractive for $\theta = 0$ as shown in Fig \ref{FIG:Thierry} (d). The situation is reversed
for anti-parallel dipoles, where a minus sign appears in front of Eq. (\ref{EQ:DipoleI}), and therefore the interaction is attractive for $\theta = \pi/2$
while $\theta = 0$ gives rise to repulsion.
The scattering properties of ultracold atoms, in the
simple case of isotropic van der Waals interactions, are entirely described by the s-wave scattering length and the potential can be replaced by
the effective contact interaction (\ref{EQ:ContactI}). In the presence of a dipolar interaction as (\ref{EQ:Dipole}), because
of its long range (decay as $1/r^3$) and anisotropic character (strong dependence on the relative angles between the dipoles), all partial waves
contribute to the scattering problem and also partial waves with different angular momenta couple with each other. While for Fermions,
replacing the real potential (\ref{EQ:Dipole}) with an effective dipolar interaction as (\ref{EQ:ContactI}) is reasonable \cite{BB:Baranov},
for Bosons this is not obvious, and in recent years it has been the subject of intensive studies \cite{BB:Derevianko, BB:Derevianko01, BB:Derevianko02, BB:Greg}.
In the presence of an optical lattice, it has been recently
argued \cite{BB:Yuan} that in a 1D geometry, replacing the real dipolar potential with an effective interaction as (\ref{EQ:ContactI}) is reasonable
as long as the optical lattice is shallow enough. However, in the most general case it is necessary to account for the full expression of the
dipole-dipole interaction potential (\ref{EQ:Dipole}).
\begin{center}
\begin{figure}[ht]
\begin{center}
\includegraphics[width=0.6\linewidth]{Dipoles.eps}
\end{center}
\caption{(a) Two dipoles, $1$ and $2$, directed along unit vectors $\myvec{e}_1$ and $\myvec{e}_2$ and separated by a distance $r$.
(b) Polarized dipoles, for which the interaction depends on the angle $\theta$ between the direction of the dipoles and the interparticle separation
$r$. This results in a repulsive interaction for $\theta = \pi/2$ (c), and attractive for $\theta = 0, \pi$ (d).
\label{FIG:Thierry}
\end{figure}
\end{center}
\subsubsection{Polarized dipoles in anisotropic harmonic traps}{--- }
\label{SEC:PolarizedDD}
We now move to the description of a BEC of polarized dipoles, pointing along the $z$ axis.
For polarized dipolar BECs, due to the anisotropy of the dipolar interactions, the geometry of the trapping potential plays a fundamental role,
first in determining the spatial distribution of the density, and second in the stability of the gas.
Qualitatively, there are two extreme scenarios depending on the shape of the confining potential, shown in Fig. \ref{FIG:DipolarScenario}:
(i) for a cigar-shaped trap elongated along the $z$ axis, i.e. with an aspect ratio between the axial $\omega_z$ and radial frequencies $\omega_\rho=\omega_x=\omega_y$ given by $\lambda = \omega_z/\omega_\rho \ll 1$, the density is mainly distributed along the polarization axis and the effect of dipole-dipole interaction is mostly attractive, which might lead to an instability of the gas even in the presence of a weak repulsive contact interaction; (ii), for a pancake-shaped trap, which is strongly confining along the $z$ axis, i.e. $\lambda \gg 1$, the dipolar interaction is mostly repulsive and the BEC is always stable for repulsive contact interactions and might be stable even for attractive contact interactions. In an intermediate situation in which the confining potential is perfectly spherical, the density distribution is then isotropic and the dipole-dipole interaction averages out to zero, which leads to a stable BEC for repulsive contact interactions. One can switch between one or the other scenario, just by adjusting the frequency of the confining potential along the $z$ axis with respect to the axial $x$ and $y$, and therefore it is natural to expect that
for any given $\lambda$ there is a critical value for the scattering length $a_{\rm crit}$ below which the BEC is unstable \cite{BB:Goral01}.
\begin{center}
\begin{figure}[h]
\begin{center}
\includegraphics[width=0.85\linewidth]{fig9.eps}
\end{center}
\caption{Polarized dipoles in anisotropic harmonic potentials. (a) in a cigar shaped trap elongated in the direction of polarization, the resulting dipolar interaction is attractive, and (b) in a pancake trap with a strong confinement in the direction of polarization, the dipolar interactions are repulsive.
\label{FIG:DipolarScenario}
\end{figure}
\end{center}
One can quantitatively describe the above scenarios starting from the Hamiltonian of the system, which in the presence of the dipole-dipole interaction (\ref{EQ:DipoleI}) reads
\begin{eqnarray}
\hat{H} &=& \int \mathrm{d}^3r \hat{\psi}^\dag(\myvec{r}) \left[-\frac{\hbar^2 \nabla^2}{2m} + V_{\rm ext}(\myvec{r})
+ \frac{g}{2}\hat{\psi}^\dag(\myvec{r})\hat{\psi}(\myvec{r}) - \mu \right] \hat{\psi}(\myvec{r}) \nonumber \\
&+& \frac{1}{2}\int \mathrm{d}^3r_1 \mathrm{d}^3r_2 \hat{\psi}^\dag(\myvec{r}_1)\hat{\psi}^\dag(\myvec{r}_2) U_{\rm dd}(\myvec{r}_1 - \myvec{r}_2)
\hat{\psi}(\myvec{r}_1)\hat{\psi}(\myvec{r}_2). \label{EQ:DipolarH2NDQ}
\end{eqnarray}
With the same approximations used to derive the Gross-Pitaevskii equation, one can write the Boson field operator
$\hat{\psi}(\myvec{r}) = \Psi_{\rm 0}(\myvec{r}) + \delta\hat{\psi}(\myvec{r})$ as a sum of a classical field $\Psi_{\rm 0}(\myvec{r})$, the condensate wave function, plus the non condensate component $\delta\hat{\psi}(\myvec{r})$ \cite{BB:Pitaevskii}. By neglecting the fluctuations $\delta\hat{\psi}(\myvec{r})$, one can calculate the energy
of the BEC given by
\begin{eqnarray}
E\big[\Psi_{\rm 0} \big] &=& \int \Big[ -\frac{\hbar^2}{2m} |\myvec{\nabla}\Psi_{\rm 0}(\myvec{r})|^2 + V_{\rm ext}(\myvec{r}) |\Psi_{\rm 0}(\myvec{r})|^2 + \frac{g}{2} |\Psi_{\rm 0}(\myvec{r})|^4 \nonumber \\
&& + \frac{1}{2} |\Psi_{\rm 0}(\myvec{r})|^2 \int U_{\rm dd}(\myvec{r} - \myvec{r}^\prime) |\Psi_{\rm 0}(\myvec{r}^\prime)|^2 \mathrm{d}^3r^\prime \Big]\mathrm{d}^3r,
\label{EQ:EFunctional}
\end{eqnarray}
where the macroscopic wavefunction $\Psi_{\rm 0}$ is normalized to the total number of particles $N$.
Within a variational Ansatz, we assume the condensate wave function to be a Gaussian of axial width $\sigma_z$ and radial width $\sigma_x = \sigma_y = \sigma_\rho$, normalized to the total number of particles $N$, namely
\begin{equation}
\label{EQ:Gaussian}
\Psi_{\rm 0}(z,\myvec{\rho}) = \sqrt{\frac{N}{\pi^{3/2}\sigma_\rho^2 \sigma_z a_{\rm ho}^3}}\exp \left[-\frac{1}{2a_{\rm ho}^2} \left( \frac{\rho^2}{\sigma_\rho^2} + \frac{z^2}{\sigma_z^2}\right) \right],
\end{equation}
where $a_{\rm ho} = \sqrt{\hbar/(m\bar{\omega})}$ is the harmonic oscillator length with average trap frequency $\bar{\omega} = (\omega_\rho^2\omega_z)^{1/3}$.
Therefore, inserting Ansatz (\ref{EQ:Gaussian}) into the energy functional Eq. (\ref{EQ:EFunctional}), after integration we find the energy of the BEC to be a function of the widths of the Gaussians, namely
\begin{equation}
\label{EQ:EnergyDD}
E_{0}(\sigma_z,\sigma_\rho) = E_{\rm kin} + E_{\rm trap} + E_{\rm contact} + E_{\rm dd},
\end{equation}
with the kinetic energy
\begin{equation}
\label{EQ:EKinDD}
E_{\rm kin} = \frac{N \hbar \bar{\omega}}{4} \left(\frac{2}{\sigma_\rho^2} + \frac{1}{\sigma_z^2}\right),
\end{equation}
the potential energy due to the trap
\begin{equation}
\label{EQ:ETrapDD}
E_{\rm trap} = \frac{N \hbar \bar{\omega}}{4\lambda^{2/3}} \left(2\sigma_\rho^2 + \lambda^2\sigma_z^2\right),
\end{equation}
the contact interaction energy given by
\begin{equation}
\label{EQ:EContactDD}
E_{\rm contact} = \frac{\hbar \bar{\omega}}{\sqrt{2\pi} a_{\rm ho}}\frac{1}{\sigma_\rho^2\sigma_z} a_s,
\end{equation}
and the contribution coming from the dipolar term
\begin{equation}
\label{EQ:EddDD}
E_{\rm dd} = -\frac{\hbar \bar{\omega} a_{\rm dd}}{\sqrt{2\pi} a_{\rm ho}}\frac{1}{\sigma_\rho^2\sigma_z} f(\kappa).
\end{equation}
We have introduced the {\it dipolar length} $a_{\rm dd} = \frac{C_{\rm dd}m}{12\pi\hbar^2}$, with $C_{\rm dd}$ given in Eq. (\ref{EQ:Cdd}), which measures the absolute strength of the dipolar interaction, $\kappa = \sigma_\rho/\sigma_z$ is the {\it aspect ratio} of the density distribution, and the function $f$ is given by
\begin{equation}
f(\kappa) = \frac{1+2\kappa^2}{1-\kappa^2} - \frac{3\kappa^2\textrm{artanh}{\sqrt{1-\kappa^2}}}{(1-\kappa^2)^{3/2}}.
\end{equation}
While the integrals needed to obtain (\ref{EQ:EKinDD},\ref{EQ:ETrapDD},\ref{EQ:EContactDD}) are easy to calculate since they contain only Gaussian functions and their derivatives, the integral to get (\ref{EQ:EddDD}) is not straightforward due to the presence of the dipolar potential $U_{\rm dd}(\myvec{r}_1 - \myvec{r}_2)$. See
section \ref{SUBSEC:DIESpherical} for more details.
In the left panel of Fig. \ref{FIG:FKappa}, we show the behavior of the function $f(\kappa)$ as $\kappa$ is continuously varied from $\kappa = 10^{-2}$ to
$\kappa = 10^{2}$. The function takes the asymptotic values of $f(0)=1$, $f(\infty)=-2$, and it vanishes for $\kappa = 1$, which implies that for a spherical
density distribution the dipole-dipole mean-field interaction (\ref{EQ:EddDD}) averages out to zero. Therefore
we notice that it is possible to control the strength and the sign of the mean-field dipolar interaction just by adjusting the aspect ratio $\lambda$ between the axial and
the radial frequencies of the confining trap. The total interaction energy is provided by the sum of the contact (\ref{EQ:EContactDD}) plus the dipolar interaction energy (\ref{EQ:EddDD}), given by
\begin{equation}
\label{EQ:OnSiteUGaussian}
E_{\rm int} = \frac{\hbar \bar{\omega} a_{\rm dd}}{\sqrt{2\pi} a_{\rm ho}}\frac{1}{\sigma_\rho^2\sigma_z} \left( \frac{a_s}{a_{\rm dd}} - f(\kappa)\right).
\end{equation}
The stability of the gas requires a repulsive interaction $E_{\rm int} > 0$, which
leads to the condition
\begin{equation}
\frac{a_s}{a_{\rm dd}} - f(\kappa) > 0,
\end{equation}
and can be adjusted ad-hoc by changing the frequencies of the trap in the three directions.
\begin{center}
\begin{figure}[h]
\begin{center}
\begin{tabular}{cc}
\includegraphics[width=0.5\linewidth]{FKappa.eps} & \includegraphics[width=0.45\linewidth]{fig13.eps}
\end{tabular}
\end{center}
\caption{(Left panel), $\kappa$ dependence of the $f(\kappa)$ function that appears in the mean-field dipolar interaction. (Right panel) Stability
diagram of a dipolar condensate: the thin line is the solution for $a_{\rm crit}(\lambda)/a_0$ calculated with the Gaussian Ansatz (\ref{EQ:Gaussian}),
where $a_0$ is the s-wave scattering length, while the thick line is
the numerical solution of the GP equation \cite{BB:BohnDipolar}. The dots with error bars are experimental data \cite{BB:Koch}.
\label{FIG:FKappa}
\end{figure}
\end{center}
To determine the stability threshold $a_{\rm crit}(\lambda)$, one needs to minimize the energy (\ref{EQ:EnergyDD}) with
respect to the variational parameters $\sigma_\rho$ and $\sigma_z$ for fixed values of $N$, $\lambda$ and $\bar{\omega}$. The results are
summarized in the right panel of Fig. \ref{FIG:FKappa} as a thin line, while the thick line represents more accurate results calculated from
solving numerically the Gross-Pitaevskii equation \cite{BB:BohnDipolar}. The dots with error bars correspond to experimental data \cite{BB:Koch}.
\subsubsection{Mean-field dipolar interaction in a spherical trap}{--- }
\label{SUBSEC:DIESpherical}
In order to calculate the mean-field dipolar interaction energy (\ref{EQ:EddDD}), we insert the Gaussian Ansatz (\ref{EQ:Gaussian}) into the
the second of the integrals (\ref{EQ:EFunctional}), and we get to the expression
\begin{equation}
E_{\rm dd} = \frac{1}{2}\int \mathrm{d}^3r_1 \mathrm{d}^3r_2 \rho(\myvec{r}_1) \rho(\myvec{r}_2) U_{\rm dd}(\myvec{r}_1 - \myvec{r}_2),
\end{equation}
with $\rho(\myvec{r}) = |\Psi_{\rm 0}(\myvec{r})|^2$ being the condensate density at $\myvec{r}$.
The last integral can be simplified by means of the convolution theorem \cite{BB:Goral01, BB:Goral} which states
\begin{equation}
\int \mathrm{d}^3r_2 U_{\rm dd}(\myvec{r}_1 - \myvec{r}_2) \rho(\myvec{r}_2) = \mathcal{F}^{-1}\left\{ \widetilde{U_{\rm dd}}(\myvec{k}) \; \widetilde{\rho}(\myvec{k})\right\},
\end{equation}
where $\widetilde{U_{\rm dd}}(\myvec{k})$ and $\widetilde{\rho}(\myvec{k})$ are the Fourier transform respectively of the dipole-dipole potential
and the density. $\mathcal{F}^{-1}$ indicates the inverse Fourier transform, and using its definition we can write
\begin{eqnarray}
E_{\rm dd} &=& \frac{1}{2}\int \mathrm{d}^3r_1 \rho(\myvec{r}_1) \frac{1}{(2\pi)^3}\int \mathrm{d}^3k \, \widetilde{U_{\rm dd}}(\myvec{k}) \;\widetilde{\rho}(\myvec{k}) e^{i\myvec{k}\cdot \myvec{r}_1} \nonumber \\
&=& \frac{1}{2(2\pi)^3} \int \mathrm{d}^3k \, \widetilde{U_{\rm dd}}(\myvec{k}) \;\widetilde{\rho}^{\;2}(\myvec{k}), \label{EQ:Edd}
\end{eqnarray}
where in the last step we have used the relation $\widetilde{\rho}(\myvec{k}) = \widetilde{\rho}(\myvec{-k})$.
The Fourier transform of the dipole-dipole interaction (\ref{EQ:DipoleI}) is given by
\begin{equation}
\widetilde{U_{\rm dd}}(\myvec{k}) = \int \mathrm{d}^3r U_{\rm dd}(\myvec{r}) e^{-i\myvec{k}\myvec{r}} = C_{\rm dd} (\cos^2\gamma - 1/3),
\end{equation}
where $\gamma$ is the angle between $\myvec{k}$ and the polarization direction, and $C_{\rm dd}$ is given by expression (\ref{EQ:Cdd}) \cite{BB:Goral01}.
In order to evaluate the integral of Eq. (\ref{EQ:Edd}) we need to insert the condensate wave function, which
in the simple case of isotropic potential ($\sigma_z = \sigma_\rho$) becomes a product of three Gaussian distributions with
equal widths $\sigma$. Therefore the Fourier transform of the condensate density is readily calculated as
\begin{equation}
\widetilde{\rho}(\myvec{k}) = \frac{N}{(\sqrt{\pi}\sigma a_{\rm ho})^3} \int \mathrm{d}^3r\, e^{-i\myvec{k}\cdot\myvec{r}}e^{-\frac{\myvec{r}^2}{\sigma^2 a_{\rm ho}^2}} = \exp \left[- \frac{\sigma^2 a_{\rm ho}^2}{4} \myvec{k}^2\right].
\end{equation}
This expression has to be inserted into the integral (\ref{EQ:Edd}), which can be easily evaluated in polar $(r,\gamma,\varphi)$ coordinates \footnote{remember, $\gamma$ is the angle between the polarization axis $z$ and $\myvec{k}$.}, giving
\begin{eqnarray}
E_{\rm dd} &=& \frac{NC_{\rm dd}}{2}\int \sin\gamma \mathrm{d}\gamma \mathrm{d}\varphi k^2\mathrm{d} k (\cos^2\gamma - 1/3)
\exp \left[- \frac{\sigma^2 a_{\rm ho}^2}{2} \myvec{k}^2\right] \nonumber \\
&=& NC_{\rm dd} 2\pi \int \mathrm{d} k k^2 \exp \left[- \frac{\sigma^2 a_{\rm ho}^2}{2} \myvec{k}^2\right] \int_{-1}^{+1} \mathrm{d} x (x^2 - 1/3) \nonumber\\
&=& 0,
\end{eqnarray}
where we have performed the change of variable $x = \cos\gamma$. The generalization to anisotropic density distributions is mathematically
more demanding but in principle straightforward, and leads to Eq. (\ref{EQ:EddDD}).
\subsubsection{Extended Bose-Hubbard model}{--- }
\label{SEC:EBHM}
As in section \ref{SUBSEC:BHM}, we expand the field operators
in the basis of Wannier functions (\ref{EQ:OPWannier}), and we keep only the lowest index corresponding to the first Bloch band.
Within this approximation the first line of Eq. (\ref{EQ:DipolarH2NDQ}) leads to the Bose-Hubbard Hamiltonian (\ref{EQ:HamiltonianBH}).
Instead the dipolar term gives rise to a further contribution
\begin{equation}
\label{EQ:HDip2Q}
\hat{H}_{\rm dd} = \frac{1}{2}\int \mathrm{d}^3r_1 \mathrm{d}^3r_2 \hat{\psi}^\dag(\myvec{r}_1)\hat{\psi}^\dag(\myvec{r}_2) U_{\rm dd}(\myvec{r}_1 - \myvec{r}_2)
\hat{\psi}(\myvec{r}_1)\hat{\psi}(\myvec{r}_2),
\end{equation}
which after expansion of the field operators in the basis of Wannier functions, becomes
\begin{equation}
\hat{H}_{\rm dd} = \sum_{i,j,k,l} \frac{V_{ijkl}}{2}\; \hat{a}^\dag_i \hat{a}^\dag_j \hat{a}_k \hat{a}_l,
\end{equation}
and the matrix elements $V_{ijkl}$ are given by the integral
\begin{equation}
\label{EQ:VIntegral}
\fl
V_{ijkl} = \int \mathrm{d}^3r_1 \mathrm{d}^3r_2 w^*(\myvec{r}_1 - \myvec{R}_i)w^*(\myvec{r}_2 - \myvec{R}_j)U_{\rm dd}(\myvec{r}_1 - \myvec{r}_2)
w(\myvec{r}_1 - \myvec{R}_k)w(\myvec{r}_2 - \myvec{R}_l).
\end{equation}
The Wannier functions are centered at the bottom of the optical lattice wells with a spatial localization that we assume to be $\sigma$. For deep enough optical potentials we can assume $\sigma$ to be much smaller than the optical lattice spacing $d$, i.e. $\sigma \ll d$. In this limit, each function $w(\myvec{r} - \myvec{R}_i)$ is
significantly non-zero for $\myvec{r} \sim \myvec{R}_i$, and the integral (\ref{EQ:VIntegral})
is significantly non-zero for the indices $i=k$ and $j=l$. Therefore there are two main contributions to the integral (\ref{EQ:VIntegral}):
the {\it off-site} matrix element $V_{ijij}$ corresponding to $k=i\neq j=l$, and the {\it on-site} $V_{iiii}$ when all the indices are equal.
Below we will explain the physical meaning of these two contributions.
\paragraph{Off-site ---}The dipolar potential $U_{\rm dd}(\myvec{r}_1 - \myvec{r}_2)$ changes slowly on the scale of $\sigma$, therefore one may approximate it with the constant $U_{\rm dd}(\myvec{R}_i - \myvec{R}_j)$ and take it out of the integration. Then the integral reduces to
\begin{equation}
V_{ijij} \simeq U_{\rm dd}(\myvec{R}_i - \myvec{R}_j) \int \mathrm{d}^3r_1 \mymod{w(\myvec{r}_1 - \myvec{R}_i)}^2
\int \mathrm{d}^3r_2 \mymod{w(\myvec{r}_2 - \myvec{R}_j)}^2,
\end{equation}
which leads to the off-site Hamiltonian
\begin{equation}
\hat{H}_{\rm dd}^{\rm off-site} = \sum_{i \neq j}\frac{V_{ij}}{2}\hat{n}_i \hat{n}_j.
\end{equation}
In the last expression $V_{ij} = U_{\rm dd}(\myvec{R}_i - \myvec{R}_j)$, $\hat{n}_i = \hat{a}^\dag_i \hat{a}_i$ is the bosonic number operator at site $i$, and the sum runs over all different sites of the lattice.
\paragraph{On-site ---}At the same lattice site $i$, where $|\myvec{r}_1 - \myvec{r}_2| \sim \sigma$, the dipolar potential changes very rapidly and
diverges for $|\myvec{r}_1 - \myvec{r}_2| \rightarrow 0$. Therefore the above approximation is not valid any more and the integral
\begin{equation}
\label{EQ:VInt}
V_{iiii} = \int \mathrm{d}^3r_1 \mathrm{d}^3r_2 \rho(\myvec{r}_1) U_{\rm dd}(\myvec{r}_1 - \myvec{r}_2) \rho(\myvec{r}_2),
\end{equation}
with $\rho(\myvec{r}) = \mymod{w(\myvec{r})}^2$ being the single particle density, has to be calculated taking into account the atomic spatial distribution
at the lattice site, similarly to what has been described in Sec. \ref{SEC:PolarizedDD} \footnote{Since $\myvec{R}_i$ is
a constant, we have renamed the variables as $\myvec{r}_u - \myvec{R}_i = \myvec{r}_u$ for $u = 1,2$.}.
We have already encountered this kind of integral in Sec. \ref{SEC:PolarizedDD}, and we have seen that, a part from a factor of $2$, the solution can be found by Fourier transforming, i.e.
\begin{eqnarray}
V_{iiii} = \frac{1}{(2\pi)^3} \int \mathrm{d}^3k \, \widetilde{U_{\rm dd}}(\myvec{k}) \;\widetilde{\rho}^{\;2}(\myvec{k}).
\end{eqnarray}
Which leads to an on-site dipolar contribution to the Hamiltonian of the type
\begin{equation}
\label{EQ:HddContact}
\hat{H}_{\rm dd}^{\rm on-site} = \sum_{i} \frac{V_{iiii}}{2} \hat{n}_i(\hat{n}_i-1).
\end{equation}
The extended Bose-Hubbard Hamiltonian is given by the sum
of the Bose-Hubbard (\ref{EQ:HamiltonianBH}) and the dipolar Hamiltonians calculated above, leading to the expression
\begin{equation}
\label{EQ:eBHH}
\fl
\hat{H}_{\rm eBH} = - J \sum_{\langle ij \rangle} \; \hat{a}^\dag_i \hat{a}_j + \frac{U}{2}\sum_i \hat{n}_i(\hat{n}_i-1) - \sum_i \mu_i \hat{n}_i + \sum_{i \neq j} \frac{V_{ij}}{2} \hat{n}_i\,\hat{n}_j,
\end{equation}
where $U$ is now taken into account as an effective on-site interaction
\begin{equation}
\label{EQ:OnSiteU}
U = g\int \mathrm{d}^3r \mymod{w(\myvec{r})}^4 + \frac{1}{(2\pi)^3} \int \mathrm{d}^3k \, \widetilde{U_{\rm dd}}(\myvec{k}) \;\widetilde{\rho}^{\;2}(\myvec{k}),
\end{equation}
which contains the contribution of the contact potential, with $g$ given in Eq. (\ref{EQ:g}), plus the dipolar contribution coming from (\ref{EQ:HddContact}).
Approximating each lattice site with a tiny harmonic trap, and approximating the atomic density distribution with Gaussians, $U$ looks like
Eq. (\ref{EQ:OnSiteUGaussian}), and one can see that the resulting on-site interaction can be increased or decreased by changing the lattice confinement.
\section{Hubbard models: theoretical methods}
\label{CH:Methods}
\subsection{Superfluid$-$Mott insulator quantum phase transition in the Bose-Hubbard model}
\label{SEC:SF_MI}
Consider the Bose-Hubbard Hamiltonian as derived in Sec. (\ref{SUBSEC:BHM}),
\begin{equation}
\label{EQ:BH2}
\hat{H}_{\rm BH} = - J \sum_{\langle ij \rangle} \; \hat{a}^\dag_i \hat{a}_j + \frac{U}{2}\sum_i \hat{n}_i(\hat{n}_i-1) - \mu \sum_i \hat{n}_i,
\end{equation}
with a uniform chemical potential $\mu$, and a total number of Bosons given by the expectation value of the operator $\hat{N} = \sum_i \hat{n}_i$. There are three parameters in this Hamiltonian, namely $J$, $U$ and $\mu$, but it is a convention to reduce the analysis of the phase diagram of $\hat{H}_{\rm BH}$ to the ratio of two of them over the third one, e.g. $J/U$ and $\mu/U$.
The ground state of Hamiltonian (\ref{EQ:BH2}) is easily understood for two opposite regimes of parameters: (i) for {\it shallow lattices}, i.e. $U/J \ll 1$, the system is in a gapless superfluid phase (SF) characterized by on-site density fluctuations and the particles delocalized over the whole lattice; (ii) for {\it deep lattices}, i.e $J/U \ll 1$, and commensurate filling, on-site density fluctuations are completely suppressed, each site is occupied by an integer number of atoms $\bar{n}$, and the ground state is a product of single-site Fock states
\begin{equation}
\label{EQ:GSFock}
\ket{GS} = \ket{\bar{n},\bar{n},\cdots}.
\end{equation}
This filling is energetically favorable in the range of chemical potential $\bar{n} - 1 < \mu/U < \bar{n}$. The system is gapped and incompressible, as beautifully explained in the famous paper of Fisher {\it et. al.} \cite{BB:Fisher}, and it is called a Mott insulator MI($\bar{n}$). For small values of $J/U$ the MI($\bar{n}$) phase persists in a closed and finite area of the $J$ vs. $\mu$ plane \cite{BB:Fisher}, which is called the Mott lobe for MI($\bar{n}$). The larger $J/U$ value of the lobe is called the tip of the lobe or also critical point $(J/U)_{\rm c}$.
The critical point changes with the dimensionality and geometry of the system. In Fig. \ref{FIG:MILobes} we plot the first $\bar{n}=0,1,2,3$ insulating lobes, calculated for an infinite optical lattice within the mean-field approximation, which will be discussed in Sec. \ref{SEC:Semi}. The thick black lines enclose the lobes and mark the boundaries between the MI and SF phases. Outside the insulating lobes, the phase is SF. The colored lines of Fig. \ref{FIG:MILobes}(a) indicate
a contour plot of constant fractional density, while the thick black lines departing from the tip of the lobes and extending into the SF region, correspond to an integer value $\bar{n}$ of the density.
\begin{center}
\begin{figure}[h]
\begin{center}
\includegraphics[width=0.7\linewidth]{BoseHubbard_Z.eps}
\end{center}
\caption{Mean-field phase diagram of the Bose Hubbard Hamiltonian. (a) Contour plot of the density per site; (b) contour plot of the order parameter (see Sec. \ref{SEC:GMF}). MI($\bar{n}$) indicate a Mott insulating phase with fixed $\bar{n}$ atoms per site.}
\label{FIG:MILobes}
\end{figure}
\end{center}
We will derive the mean-field Mott insulating lobes of Fig. \ref{FIG:MILobes} in a more rigorous way in Sec. \ref{SEC:Semi}, but for the moment we just list
the critical points $(J/U)_{\rm c}$, for the $\bar{n} = 1$ lobe, that have been estimated with different methods and for different dimensions of the lattice.
In one dimension, the critical point has been estimated to be $(J/U)_{\rm c} \simeq 0.29$ \cite{BB:Monien1} using Density Matrix Renormalization Group calculations (DMRG). In two dimensions, with quantum Monte Carlo calculations, the critical point has been estimated to be $(J/U)_{\rm c} \simeq 0.061$
\cite{BB:Wessel}, while in the three dimensional model the location of the critical point has been estimated with perturbative expansions \cite{BB:Monien2}, and quantum Monte Carlo simulations \cite{BB:Svistunov3D} to be at $(J/U)_{\rm c} \simeq 0.034$.
In the next section we will derive the mean-field lobes.
\subsection{The Gutzwiller mean-field approach}
\label{SEC:GMF}
The Gutzwiller mean-field approach provides an approximated many-body wave function for Hubbard-type Hamiltonians, of the form
\begin{equation}
\label{EQ:GW}
\ket{\Psi} = \prod_i \sum_{n=0}^{n_{\rm max}} f_n^{(i)} \ket{n}_i,
\end{equation}
where $\ket{n}_i$ represents the Fock state of $n$ atoms occupying the site $i$, $n_{\rm max}$ is a cut off in the maximum number of atoms per site, and $f_n^{(i)}$
is the probability amplitude of having the site $i$ occupied by $n$ atoms. The probability amplitudes are normalized to unity $\sum_n |f_n^{(i)}|^2 = 1$.
The Gutzwiller wave function (\ref{EQ:GW}) has been extensively used in the literature \cite{BB:Jacksch,BB:Krauth,BB:Rokhsar,BB:Jaksch02}, it predicts that there exists a critical value $(J/U)_{\rm mf}$ in a given range of $\mu$, below which the ground state is a product of single Fock states $f_n^{(i)} = \delta_{n,\bar{n}}$ with exactly $\bar{n}$ particles per site, as (\ref{EQ:GSFock}). Moreover, for $J/U > (J/U)_{\rm mf}$ the Gutwiller Ansatz predicts a superfluid ground state with fluctuating on-site particle number.
The Gutzwiller critical point, for $\bar{n} = 1$, is found to be $(J/U)_{\rm mf} = 1/5.8z$ \cite{BB:Zwerger}, where $z=\sum_{\langle j\rangle_i} 1$ is the number of nearest neighbor connections at each site of the lattice.
In table \ref{TAB:CriticalXXYY} we show the comparison of the critical points predicted by the Gutzwiller Ansatz for different dimensions of the system, with the more
precise ones discussed in Sec. (\ref{SEC:SF_MI}).
\begin{table}[htdp]
\caption{Comparison of the Gutzwiller critical points $(J/U)_{\rm mf}$ with the more precise, up to now, critical points $(J/U)_{\rm c}$, for different dimensions D of the system. }
\begin{center}
\begin{tabular}{ccccccc}
\br
D & & $z$ & & $(J/U)_{\rm mf}$ & & $(J/U)_{\rm c}$ \\
\mr
1 & & 2 & & 0.0862 & & 0.29 \\
2 & & 4 & & 0.0431 & & 0.061 \\
3 & & 6 & & 0.0287 & & 0.034 \\
\br
\end{tabular}
\end{center}
\label{TAB:CriticalXXYY}
\end{table}
From the comparison, one can deduce that the Guzwiller is unsatisfactory for 1D systems ($z=2$) while it is satisfactory for a 3D one ($z = 6$).
Also, in the limit of $J/U \rightarrow \infty$ the difference between the Gutzwiller predictions and the exact results are negligible \cite{BB:Zwerger}. Summarizing,
the Gutzwiller predictions are exact in the two limiting cases of $J/U \rightarrow 0$, and $J/U \rightarrow \infty$, while for intermediate cases the performance of the Gutzwiller approach strongly depends on the dimensionality of the lattice, since it does not correctly account for the quantum fluctuations at the phase transition.
Two important quantities are the {\it order parameter} $\varphi_i = \bra{\Psi} \hat{a}_i \ket{\Psi}$, namely the expectation value of the Bosonic annihilation
operator at the $i$-th site of the lattice, and the {\it density fluctuations} at the site $i$ given by $\delta n_i = \bra{\Psi} \hat{n}_i^2 \ket{\Psi} - \bra{\Psi} \hat{n}_i \ket{\Psi}^2$. By using the Gutzwiller wavefunction (\ref{EQ:GW}) one gets the following expressions for the order parameter
\begin{equation}
\varphi_i = \sum_n \sqrt{n+1} f_n^{*(i)} f_{n+1}^{(i)},
\end{equation}
and for the density fluctuations
\begin{equation}
\delta n_i = \sum_n n^2 |f_n^{(i)}|^2 - \Big[ \sum_n n |f_n^{(i)}|^2 \Big]^2.
\end{equation}
The order parameter $\varphi_i$ together with the density fluctuations $\delta n_i$ describe the phase of the system at the site $i$ of the lattice:
in the Mott phase $\varphi_i = 0$, and density fluctuations are suppressed $\delta n_i = 0$, while the superfluid phase is characterized by $\varphi_i \neq 0$ and presence of density fluctuations $\delta n_i \neq 0$. In the uniform system, the lattice is translationally invariant and therefore all sites are self-similar, which means that only one site is sufficient to determine the phase of the whole system. In Fig. \ref{FIG:MILobes}(b) we plot the absolute value of the order parameter $\varphi_i$ for such a system, in the $J/U$ vs. $\mu/U$ plane. The colored lines outside the insulating lobes correspond to a contour plot of constant non-zero value of $\varphi_i$ typical of the SF phase. Instead, in a non-uniform system as it is in the presence of an external confining harmonic potential, different phases can coexist. As an example, in Fig. \ref{FIG:BHTrap} we plot the density of the ground state (a), the order parameter at each site (b), and the density fluctuations (c) of a 2D lattice in the presence of a confining harmonic potential. Notice that the MI phase at the center of the harmonic trap $(0,0)$ is surrounded by a ring of SF phase, in a wedding-cake like structure, as first discussed in \cite{BB:Jacksch}. The ground state of Fig. \ref{FIG:BHTrap} was obtained within the mean-field approximation through the imaginary-time evolution technique, that will be discussed in Sec. \ref{SEC:DGA}.
\begin{center}
\begin{figure}[h]
\begin{center}
\includegraphics[width=1\linewidth]{BHTrap01.eps}
\end{center}
\caption{Mean-field ground state of a 2D optical lattice, calculated for $J=0.025U$, and $\mu_i = 0.45U - \Omega i^2$, where $\Omega = 8 \times 10^{-3} U/\hbar$ is the frequency of the harmonic oscillator confinement. (a) the vertical axis shows the value of the density at each site of the lattice, (b) the corresponding absolute value of the order parameter, and (c) on-site density fluctuations.}
\label{FIG:BHTrap}
\end{figure}
\end{center}
In the presence of dipolar interactions, as we shall see later on, it is also necessary to account for non-uniform quantum phases, because even in the uniform system the presence of dipolar interactions may lead to spontaneous symmetry breaking of translational invariance on a scale larger than the lattice constant.
\subsubsection{Dynamical Gutzwiller approach}{--- }
\label{SEC:DGA}
The time dependent version of the Gutzwiller wavefunction (\ref{EQ:GW}) is obtained by allowing the Gutzwiller amplitudes to depend on time $f_n^{(i)} (t)$ \cite{BB:Jaksch02}.
Then the equations of motion for the amplitudes are readily obtained by minimizing the action of the system, given by $S = \int \mathrm{d} t \mathcal{L}$, with respect to the variational parameters $f_n^{(i)}(t)$ and their complex conjugates $f_{n}^{*(i)}(t)$. The Lagrangian of the system in the quantum state $\ket{\Psi}$,
is given by \cite{BB:Perez}
\begin{equation}
\mathcal{L} = i\hbar\frac{\braket{\Psi}{\dot{\Psi}} - \braket{\dot{\Psi}}{\Psi}}{2} - \bra{\Psi} \hat{H}\ket{\Psi},
\end{equation}
where $\ket{\dot{\Psi}}$ is the time derivative of the wave function (\ref{EQ:GW}).
By equating to zero the variation of the action with respect to $f_{n}^{*(i)}$, one gets the equations
\begin{eqnarray}
i\hbar \frac{\mathrm{d}}{\mathrm{d} t} f_n^{(i)} &=& -J \Big[\bar{\varphi}_i \sqrt{n} f_{n-1}^{(i)} + \bar{\varphi}_i^* \sqrt{n+1} f_{n+1}^{(i)} \Big] \nonumber \\
& & + \Big[\frac{U}{2}n(n-1) + n \sum_{j\neq i} V_{ij} \langle \hat{n}_j\rangle - \mu_i n\Big] f_n^{(i)}, \label{EQ:FDynamics}
\end{eqnarray}
where $\bar{\varphi}_i = \sum_{\langle j \rangle_i} \varphi_j$, the sum runs over all nearest neighbors $j$ of site $i$, $\langle \hat{n}_j\rangle = \bra{\Psi} \hat{a}_j^\dag \hat{a}_j \ket{\Psi}$ is the average particle number at site $j$, and the total number of particles is given by
$N = \sum_i \langle \hat{n}_i\rangle$. It is not difficult to verify the commutation relation $[\hat{N},\hat{H}_{\rm BH}] = 0$,
which implies that the total number of Bosons is a conserved quantity for the dynamics in real time \cite{BB:Sachdev}.
These equations are of mean-field type, because for each site $i$ the influence of the neighboring sites is taken into account in a mean way
into the "field" $\bar{\varphi}_i$ together with $\sum_{j\neq i} V_{ij} \langle \hat{n}_j\rangle$, which have to be determined self-consistently. Eqs. (\ref{EQ:FDynamics}) are a set of coupled equations, the coupling arising from the tunneling part, and can be written in the matrix form
\begin{equation}
\label{EQ:FDinamics2}
i\hbar \frac{\mathrm{d}}{\mathrm{d} t} \vec{f} = \mathcal{M}[\vec{f},\mu,U,J] \cdot \vec{f},
\end{equation}
where $\vec{f} = \left[ f_0^{(1)}, f_1^{(1)}, \cdots, f_n^{(i)}, \cdots f_{n_{\rm max}}^{(N_S)} \right]^{\rm T}$, is the vector of the Gutzwiller amplitudes ordered from site $1$ to site $N_S$, the latter being the total number of sites. It is worth noticing that the matrix $\mathcal{M}[\vec{f},\mu,U,J]$ is itself a functional of the coefficients $\vec{f}$ through the fields $\bar{\varphi}_i$ and $\sum_{j\neq i} V_{ij} \langle \hat{n}_j\rangle$, which have to be calculated in a self-consistent way.
Let us clarify this point with an example. Suppose we want to solve Eq. (\ref{EQ:FDinamics2})
between an initial time $t_i=0$ and a final time $t_f$, with a given initial condition $\vec{f}(0)$. We discretize the time interval in $N$ steps of size $\Delta t$, with $N$ finite, and define $t_s = s \Delta t$ such that $t_{s=0} \equiv 0$ and $t_{s=N} \equiv t_f$. Therefore, to calculate the solution at a certain point in time $\vec{f}(t_{s+1})$
we need to know the solution right at the preceding time $\vec{f}(t_s)$, with which we can compute the fields that in turn determine $\mathcal{M}[\vec{f}(t_s),\mu,U,J]$,
and the solution is readily found to be
\begin{equation}
\label{EQ:SolTimeStep}
\vec{f}(t_{s+1}) = e^{-i \mathcal{M}[\vec{f}(t_s),\mu,U,J] \Delta t/\hbar} \vec{f}(t_s).
\end{equation}
Starting from $s=0$, in $N+1$ steps we have determined the solution at the desired time $t_f$. At the computational level, this is the
simplest procedure one can implement to calculate the dynamics of the system. However, one needs to be careful in the choice of the time step $\Delta t$, especially
for fast-oscillating dynamics. In such cases, a Runge-Kutta with adaptive stepsize control has proven to be more efficient. Instead in the case of imaginary time evolution, which requires solving stiff dynamics, the simple procedure described above has shown to be enough accurate and faster.
Equations (\ref{EQ:FDinamics2}) can be solved in real time $t$ or also imaginary time $\tau = it$. The imaginary time evolution is a standard technique that has been thoroughly used, because due to dissipation is supposed to converge to the ground state of the system.
Two things are worth to be noticed. First, because the imaginary time evolution is not unitary, it does not conserve the norm of the Gutzwiller wavefunction, which has to be renormalized after each time step. Second, the total number of particles is not a conserved quantity anymore.
For dipolar Hamiltonians the imaginary time evolution does not always converge to the true ground state and it gets blocked in configurations which are a local minimum of the energy. On the one hand this makes it a difficult task to identify the ground state of such systems, and on the other hand it is a signature of the existence of metastable states as we will discuss in details in Sec. \ref{CH:MS:DipolarBosons}.
\subsubsection{Perturbative mean-field approach}{--- }
\label{SEC:Semi}
A more convenient method to determine the insulating phases of a dipolar Hamiltonian is to use a mean-field approach perturbative in $\varphi_i$.
From statistical mechanics, the expectation value of the annihilation operator at the $i$-th site is given \cite{BB:Tannoudji01} by the trace
\begin{equation}
\label{EQ:Distribution}
\varphi_i = \langle \hat{a}_i \rangle = {\rm Tr} (\hat{a}_i\hat{\rho}),
\end{equation}
where $\hat{\rho} = Z^{-1}e^{-\beta \hat{H}}$ is the density matrix operator, $Z = {\rm Tr} (e^{-\beta \hat{H}})$ its normalization, and $\beta = 1/K_B T$ is the inverse temperature of the system. We write Hamiltonian (\ref{EQ:eBHH}) in the form $\hat{H} = \hat{H}_0 + \hat{H}_1$ where
\begin{eqnarray}
\hat{H}_0 &=& \frac{U}{2}\sum_i \hat{n}_i(\hat{n}_i-1) - \mu \sum_i \hat{n}_i + \sum_{i \neq j} \frac{V_{ij}}{2} \hat{n}_i\,\hat{n}_j \\
\hat{H}_1 &=& - J \sum_{\langle ij \rangle} \; \hat{a}^\dag_i \hat{a}_j \label{EQ:H1},
\end{eqnarray}
and we assume a uniform chemical potential $\mu$. The generalization to a site-dependent chemical potential is straightforward. Furthermore, we assume low temperatures $\beta \rightarrow \infty$, and the tunneling coefficient to be the smallest energy in the system, i.e $J \ll U,\mu,V_{ij} $ such that we can treat $\hat{H}_1$ as a small perturbation on $\hat{H}_0$, and use the Dyson expansion at the first order in $\hat{H}_1$ for all the exponential operators, so that one obtains
\begin{equation}
e^{-\beta(\hat{H}_0 + \hat{H}_1)} \simeq e^{-\beta \hat{H}_0}\Big[ \hat{\openone} - \int_0^\beta e^{\tau \hat{H}_0} \hat{H}_1 e^{-\tau \hat{H}_0} \mathrm{d} \tau \Big].
\end{equation}
We now write Hamiltonian (\ref{EQ:H1}) as a sum of single site Hamiltonians. Writing the annihilation operator as
$\hat{a}_i = \hat{A}_i + \varphi_i$, we can perform the mean field decoupling
on the tunneling term
\begin{eqnarray}
\hat{a}_i^\dag \hat{a}_j &=& \hat{A}_i^\dag \varphi_j + \hat{A}_j \varphi_i + \varphi_i \varphi_j + \hat{A}_i^\dag \hat{A}_j \nonumber \\
&\simeq& \hat{a}_i^\dag \varphi_j + \hat{a}_j \varphi_i - \varphi_i \varphi_j ,
\end{eqnarray}
where in the last step we have assumed small fluctuations,
characteristic of the Mott, or the deep superfluid
states, and replaced $\hat{A}_i^\dag \hat{A}_j \simeq 0$. In Hamiltonian (\ref{EQ:H1}) we
now replace $\hat{a}_i^\dag \hat{a}_j$ with the expression calculated above,
we neglect terms of the order of $\varphi^2$ and find the mean field tunneling Hamiltonian
\begin{equation}
\hat{H}_1^{\scriptscriptstyle \mathrm{MF}} = -J\sum_i \left( \hat{a}_i^\dag \bar{\varphi}_i + \bar{\varphi}_i^*\hat{a}_i \right).
\end{equation}
Given a classical distribution of atoms in a lattice such as
\begin{equation}
\ket{\Phi} = \prod_i\ket{n_i}_i,
\end{equation}
satisfying $\hat{H}_0 \ket{\Phi} = E_\Phi \ket{\Phi}$, let us suppose that this configuration is a local minimum of the
energy, it can be the ground state, namely the absolute minimum, or another local minimum.
We will be more
rigorous at the end of this section regarding the meaning of local minimum of energy but for the moment let us refer to the common picture of a local minimum.
In the basis of the eigenfunctions of $\hat{H}_0$, satisfying the relation $\hat{H}_0 \ket{\Upsilon} = E_\Upsilon \ket{\Upsilon}$ the partition function then takes the simple form
\begin{equation}
\label{EQ:PartitionF}
Z \simeq {\rm Tr} (e^{-\beta \hat{H}_0}) = \sum_{\ket{\Upsilon}} \bra{\Upsilon} e^{-\beta \hat{H}_0} \ket{\Upsilon} \;
\stackrel{\scriptscriptstyle{\beta \mapsto \infty}}{\longrightarrow } \; e^{-\beta E_\Phi},
\end{equation}
where the last limit holds because we do not trace over all the states of the basis but only around the state $\ket{\Phi}$ , which is assumed to be a local minimum of the energy.
Using again a Dyson expansion of the exponential of the density operator, we obtain the order parameter as
\begin{eqnarray}
\varphi_i &\simeq& - e^{\beta E_\Phi}\int_0^\beta {\rm Tr} \left[ \hat{a}_i \; e^{-(\beta - \tau) \hat{H}_0} \; \hat{H}_1^{\scriptscriptstyle \mathrm{MF}} \; e^{-\tau \hat{H}_0} \right] \mathrm{d} \tau \nonumber \\
&=& J \bar{\varphi}_i e^{\beta E_\Phi} \int_0^\beta \sum_{\ket{\Upsilon}} \bra{\Upsilon} \hat{a}_i \; e^{-(\beta - \tau) \hat{H}_0} \; \hat{a}_i^\dag \; e^{-\tau \hat{H}_0} \ket{\Upsilon}, \label{EQ:IntPhi}
\end{eqnarray}
which is easy to calculate. The trace is then non trivial only for
$\ket{\Upsilon} = \ket{\Phi}$ and $\ket{\Upsilon} = \frac{\hat{a}_i}{\sqrt{n_i}}\ket{\Phi}$, where $n_i$ is integer on $\ket{\Phi}$, and after the integration in the $\beta \mapsto \infty$ limit we are left with the result
\begin{equation}
\label{EQ:OrderParameter}
\varphi_i = J\bar{\varphi}_i \left[ \frac{n_i + 1}{E_{\scriptscriptstyle \mathrm{P}}^i} + \frac{n_i}{E_{\scriptscriptstyle \mathrm{H}}^i}\right],
\end{equation}
where the quantities $E_{\scriptscriptstyle \mathrm{P}}^i$, $E_{\scriptscriptstyle \mathrm{H}}^i$ are defined as
\begin{eqnarray}
\eqalign{
\label{EQ:PHExcitations}
E_{\scriptscriptstyle \mathrm{P}}^i &= - \mu + Un_i + V_{\rm dip}^{1,i} \\
E_{\scriptscriptstyle \mathrm{H}}^i &= \mu - U(n_i-1) - V_{\rm dip}^{1,i},
}
\end{eqnarray}
and are respectively the energy cost for a particle (P) and hole (H) excitation on top of the $\ket{\Phi}$ configuration. In the previous expressions $V_{\rm dip}^{1,i} = \sum_{j\neq i} V_{ij} n_j$
is the dipole-dipole interaction that one atom placed at site $i$ feels with the rest of the atoms in the lattice.
We performed the integral (\ref{EQ:IntPhi}) in the limit of $\beta \mapsto \infty$, and in such a limit one finds that the integral converges only for positive
values of the particle and hole excitation energies, namely
\begin{equation}
\label{EQ:Conditions01}
U(n_i-1) + V_{\rm dip}^{1,i} < \mu < Un_i + V_{\rm dip}^{1,i}.
\end{equation}
This requirements have to be fulfilled at every site $i$ of the lattice and they simply state that the configuration $\ket{\Phi}$ is a local minimum with respect to adding and removing particles at any site. In the light of this statement, the restriction on the trace of Eq. (\ref{EQ:PartitionF}) is now rigorous, and is in perfect agreement
with the treatment done in \cite{BB:Fisher}.
Notice that if $\ket{\Phi}$ is not a local minimum, then one finds that conditions (\ref{EQ:Conditions01}) are never satisfied and the integral (\ref{EQ:IntPhi}) indeed diverges.
This treatment is of course also valid for $\ket{\Phi}$, being in particular the ground state of the system.
One finds such an equation (\ref{EQ:OrderParameter}), and conditions (\ref{EQ:Conditions01}) for every site $i$ of the lattice.
The convergence conditions are simple and among them one has to choose the most stringent to find the boundary of the lobe at J = 0.
Instead the equations for the order parameters are coupled due to the $\bar{\varphi}_i$ term, they can be written in a matrix
form $\mathcal{M}(\mu,U, J) \cdot \vec{\varphi} = 0$, with $ \vec{\varphi} \equiv (\cdots \varphi_i \cdots )$, and have
a non trivial solution. For every $\mu$, the smallest $J$ for which ${\rm det} [\mathcal{M}(\mu,U, J)] = 0$ gives the lobe of the $\ket{\Phi}$ configuration in the
$J$ vs. $\mu$ plane.
\subsubsection{Perturbative mean-field vs. dynamical Gutzwiller approach}{--- }
\label{SS:Compare}
The predictions of the perturbative mean-field treatment are in perfect agreement with the results of the dynamical Gutzwiller approach,
since they both rely on the same mean-field approximation. The first looks at the stability of a given density distribution
$\ket{\Phi} = \prod_i\ket{n_i}_i$ of integer $n_i$ atoms per site, with respect to particle and hole excitations, while the latter minimizes the energy of a random initial configuration with respect to particle and hole excitations leading to the distribution $\ket{\Phi}$ if the initial condition is sufficiently close. However, the first method can only identify the phase boundaries of the insulating lobes without providing any further information on the SF phases outside the lobes, which can instead be explored with the imaginary time evolution. Nevertheless for dipolar Hamiltonians, due to the presence of many local minima of the energy, as we will see in the next part \cite{BB:MenottiPRL,BB:Trefzger01}, it is very difficult to identify the ground state with the dynamical Gutzwiller approach. This can be achieved more
efficiently through the perturbative mean-field approach. Therefore the two methods complement each other.
As an example, in Fig \ref{FIG:MILobes} (a,b) the black lines are calculated with the perturbative method ($V_{ij} = 0$) while the SF region outside the lobes is explored using imaginary time evolution showing perfect agreement between the two approaches.
\section{Dipolar Bosons in a 2D optical lattice}
\label{CH:MS:DipolarBosons}
\subsection{The model}
\label{SEC:TheModel01}
In \cite{BB:MenottiPRL,BB:Trefzger01}, we have studied the properties of dipolar Bosons in an infinite 2D optical lattice,
mimicked by an elementary cell of finite dimensions $L\times L$ ($N_S = L^2$ sites) satisfying periodic boundary conditions.
The dipoles are aligned and point perpendicularly out of the plane so that the dipole-dipole interaction (\ref{EQ:DipoleI}) between two particles at relative distance
$\myvec{r}$ becomes $U_{\rm dd} (\myvec{r}) = C_{\rm dd}/(4\pi r^3)$ repulsive and isotropic in the 2D plane of the lattice, where $C_{\rm dd}$ is given by
Eq. (\ref{EQ:Cdd}).
Furthermore for computational simplicity we truncate the range of the off-site interactions at a finite number of nearest neighbors, as shown in Fig. \ref{FIG:Truncation} up to
the range number four.
\begin{center}
\begin{figure}[h]
\begin{center}
\includegraphics[width=0.5\linewidth]{Truncation.eps}
\end{center}
\caption{Representation of the first four nearest neighbors of the site labeled as $0$ in the 2D lattice.}
\label{FIG:Truncation}
\end{figure}
\end{center}
We have studied the phase diagram of the system described by the Hamiltonian
\begin{equation}
\label{EQ:EBH2D}
\fl
\hat{H} = - J \sum_{\langle i j\rangle} \hat{a}^{\dag}_i \hat{a}_j - \sum_i \mu \hat{n}_i +\sum_i \frac{U}{2} \; \hat{n}_i(\hat{n}_i-1) +
\frac{U_{\rm NN}}{2} \sum_{\vec \ell} \sum_{\langle \langle ij \rangle \rangle_{\vec \ell}} \frac{1}{{|\;\vec{\ell}\phantom{1}| }^{3}} \; \hat{n}_i \hat{n}_j ,
\end{equation}
where $U_{\rm NN} = C_{\rm dd}/(4\pi d_{\scriptscriptstyle \rm 2D}^3)$ is the dipole-dipole interaction between nearest neighboring sites, $d_{\scriptscriptstyle \rm 2D}$ is the lattice period, and $\langle \langle ij \rangle \rangle_{\vec \ell}$ represents neighbors at relative distance $\vec{\ell}$ which is measured in units of $d_{\rm 2D}$.
All other quantities were introduced previously.
\subsection{Metastability}
\label{SEC:Metastability}
Let us start the discussion by introducing our definition of stability of a given classical distribution of atoms in the lattice, i.e. a product over single-site Fock states
\begin{equation}
\label{EQ:Metastable01}
\ket{\Phi} = \prod_i \ket{n_i}_i.
\end{equation}
At $J=0$, we define the state (\ref{EQ:Metastable01}) to be stable if there exists a finite interval $\Delta\mu = \mu_{\rm max} - \mu_{\rm min} > 0$
in the $\mu$ domain, in which the particle (P) and hole (H) excitations at each site
$i$ of the lattice are positive, and the system is gapped.
Using the dipolar Hamiltonian (\ref{EQ:EBH2D}), in Sec. \ref{SEC:Semi} we have calculated the particle and hole excitation energies of $\ket{\Phi}$ to be
\begin{eqnarray}
\eqalign{
\label{EQ:PHExcitations02}
E_{\scriptscriptstyle \mathrm{P}}^i &= - \mu + Un_i + V_{\rm dip}^{1,i} \\
E_{\scriptscriptstyle \mathrm{H}}^i &= \mu - U(n_i-1) - V_{\rm dip}^{1,i},
}
\end{eqnarray}
where we recall $V_{\rm dip}^{1,i}\geq 0$ to be the dipolar interaction experienced by one atom sitting at the site $i$ of the lattice.
From Eqs. (\ref{EQ:PHExcitations02}) it is then straightforward to find $\Delta\mu$, if it exists, given by the set of inequalities
\begin{equation}
\label{EQ:StabilityConditions}
U(n_i-1) + V_{\rm dip}^{1,i} < \mu < Un_i + V_{\rm dip}^{1,i}.
\end{equation}
This is consistent with the
stability conditions discussed in the seminal paper of Fisher {\it et. al.} \cite{BB:Fisher}. Indeed, in the absence of dipolar interactions $U_{\rm NN}=0$
into Eqs. (\ref{EQ:StabilityConditions}), one recovers the well known conditions $U(n_i-1) < \mu < Un_i$ for the stability of the MI($n_i$), with $n_i$ particles per site.
One can extend the stability analysis to small values of $J$, and for a given stable state calculate its insulating lobe with the perturbative
mean-field approach we have developed in Sec. \ref{SEC:Semi}. In this context, we therefore define a state like (\ref{EQ:Metastable01}) to be {\it metastable} if it satisfies two conditions: the first is that the state must have an insulating lobe inside which it is gapped, and the second is that the energy of the state must be higher than the ground state energy. In other words a metastable state is a local minimum of the energy.
In the absence of dipolar interactions $U_{\rm NN}=0$, no metastable states are found. In the low tunneling region the ground state of the system consists of Mott insulating lobes with integer filling factors $\nu = N_{\rm a}/N_{\rm S}$ (number of atoms$/$number of sites), while for large values of $J$ the system is superfluid.
In our treatment metastable states appear as soon as one introduces at least one nearest neighbor of the dipolar interaction. In fact, the imaginary time evolution, which for Bose Hubbard Hamiltonians with only on-site interactions converges unambiguously to the ground state, for the dipolar Hamiltonian (\ref{EQ:EBH2D}) often converges to different
metastable configurations depending on the exact initial condition. Moreover, in the real time evolution, their stability manifests as typical small oscillations of frequency $\omega_0$ around the local minimum of the energy.
Our main results are summarized in Figs. \ref{FIG:Lobes} (a,b,c), where we plot the phase diagram of the Hamiltonian (\ref{EQ:EBH2D}) for a $L=4$ elementary cell satisfying periodic boundary conditions, for different values of the cut off range of the dipolar interactions respectively at one (a), two (b), and four nearest neighbors. The on-site interaction is given by $U/U_{\rm NN} = 20$ and $U_{\rm NN}=1$ is the unit of energy.
\begin{center}
\begin{figure}[h!]
\begin{center}
\includegraphics[width=0.9\linewidth]{Lobes.eps}
\end{center}
\caption{(a), (b), (c) Phase diagram with a range of
the dipole-dipole interaction cut at the first, second, and fourth nearest
neighbor, respectively. The thick line is the ground state and the
other lobes correspond to the metastable states, the same color corresponding
to the same filling factor. In (c) filling factors range
from $\nu=1/8$ to $\nu=1$. In the right column we present metastable
configurations for $\nu = 1/2$ appearing at the first nearest neighbor (I), and second
(IIa, IIb), and the corresponding ground state (GS); those metastable
states remain stable for all larger ranges of the dipole-dipole interaction. White sites are empty, while gray sites are occupied by one atom.}
\label{FIG:Lobes}
\end{figure}
\end{center}
The thick lines correspond to the ground state while the thin lines correspond to metastable insulating states, with the color identifying the same filling
factor $\nu$. The difference with the Bose Hubbard phase diagram of Fig. \ref{FIG:MILobes},
where only integer filling factors $\nu = \bar{n}$ are present, is evident already with one nearest neighbor of the dipolar interaction (a). In fact the MI(1) lobe
undergoes a global shift of $zU_{\rm NN}$ ($z = 4$ in the figure) towards higher values of $\mu$, and the new fractional filling factor $\nu = 1/2$ appears with a ground state density distribution modulated in a checkerboard pattern, shown in Fig. \ref{FIG:Lobes} (GS).
Instead, the density distribution shown in Fig. \ref{FIG:Lobes} (I) is metastable with $\nu=1/2$, and its insulating lobe is given by the thin line extending from $1 < \mu/U_{\rm NN} < 3$.
Remarkably, in Fig. \ref{FIG:Lobes}(a) the two lobes extending from $0 < \mu/U_{\rm NN} < 1$ correspond to metastable configurations at filling factors $\nu = 1/4, 5/16$, while the two lobes between $3 < \mu/U_{\rm NN} < 4$ correspond to metastable states at filling factors $\nu = 3/4, 11/16$, but no ground state is found for these fillings. In the region immediately outside the ground state lobes, we found evidences of supersolid (SS) phases, where the order parameters $\varphi_i$ are different than zero and are spatially modulated, e.g. in a CB structure.
Before our work, studies of BH models with extended interactions have pointed out the existence of novel quantum phases, like the SS and checkerboard phases, but not the existence of the metastable states.
Increasing further the range of the dipolar interactions leads to the
appearance of more metastable states, as (IIa) and (IIb) found at $\nu=1/2$ for two nearest neighbors in the range of dipolar interactions. The MI(1) undergoes
a larger shift which is accompanied by
the emergence of other insulating fractional filling factors, as shown in Fig. \ref{FIG:Lobes} (c), where the dipole-dipole interaction is cut at the fourth nearest neighbor and the ground state is a series of lobes with
$\nu$ multiple of $\nu = 1/8$. The number of metastable states varies depending on the parameters of the Hamiltonian and the filling factor; it is found to be up to 400 for
$U/U_{\rm NN} = 20$ at filling $\nu=1/2$, and up to 1500 for $U/U_{\rm NN} = 2$ at unit filling \cite{BB:MenottiPRL}. With this picture in mind, it is now clear why the imaginary time evolution,
which often converges to different metastable configurations, is very inefficient both to find the ground state of the system and to compute the lobe boundaries of a given
metastable state. Instead, the mean-field perturbative approach we have derived in Sec. \ref{SEC:Semi} has proven to be satisfactory for this purpose but it also has
some limitations. In fact, all possible values of $\nu$ and the corresponding configurations which are detectable with this method is limited by the size of the elementary cell. Evidently the possible filling factors of an elementary cell of size $L$ are given by multiples of $\nu = 1/L^2$. It is worth to notice, that despite the inefficiency
of the imaginary time evolution in finding the insulating lobes of the system, the corresponding equations in real time turn out to be very useful, for example, to compute the excitation spectrum $\omega({\bf k})$ of the system. For a given metastable configuration, one can calculate $\omega({\bf k})$ from the small fluctuations $\delta f_{n}^{(i)}(t)$ around the unperturbed metastable state coefficients $\overline{f}_{n}^{(i)}$. Writing $f_{n}^{(i)} = \overline{f}_{n}^{(i)} + \delta f_{n}^{(i)}(t)$ into Eq. (\ref{EQ:FDynamics}), and taking into account only linear terms in the fluctuations, one obtains a set of coupled equations for $\delta f_{n}^{(i)}(t)$, which can be easily solved in the Fourier domain as explained in \cite{BB:Trefzger01}.
Finally, we find that there is usually a gap between the ground state and the lowest metastable state, which might allow to reach
the ground state by ramping up the optical lattice under some adiabaticity condition. However, this feature is strongly reduced in the case of larger elementary cells because the number of metastable configurations and the variety of their patterns increase very rapidly with the size of
the elementary cell $L$. Indeed, we have found that there exist many metastable configurations that differ from the ground state only by small localized defects, and the energy of these reduces the size of the gap.
\subsection{The lifetime}
\label{SEC:Lifetime}
We have studied the stability of the metastable states with a path integral formulation in imaginary time and a generalization of the instanton theory \cite{BB:Wen}. For any given initial metastable configuration $\ket{\Phi}_{\rm initial}$, we are able to estimate the time $T$ in which $\ket{\Phi}_{\rm initial}$ has tunneled completely into a different metastable state $\ket{\Phi}_{\rm final}$. We do this in analogy with the case of a classical particle tunneling through a potential barrier shown in Fig. \ref{FIG:Connection} (a), with the difference that we do not have any information a priory on the characteristics of the potential barrier separating initial and final state.
Nevertheless we can estimate the barrier and the time $T$ in three steps: (i) first we construct the imaginary time Lagrangian
of the system described by a quantum state $\ket{\Phi}$, (ii) we make use of a variational method on $\ket{\Phi}$ with only one variational parameter $q$, and its conjugate momenta $P$, that interpolate continuously between $\ket{\Phi}_{\rm initial}$ and $\ket{\Phi}_{\rm final}$, and (iii) through the variation of $q$ we calculate the minimal action $S_0$, with the imaginary time Lagrangian, along the stationary path starting at $\ket{\Phi}_{\rm initial}$; this path is called an {\it instanton path}, in short instanton.
It connects $\ket{\Phi}_{\rm initial}$ and $\ket{\Phi}_{\rm final}$ only if the two states are degenerate, otherwise the stationary path connects $\ket{\Phi}_{\rm initial}$ with
an intermediate state called the {bouncing point} $\ket{\Phi}_{\rm bounce}$. We get an estimate of the energy barrier separating the two states by evaluating the Lagrangian from $\ket{\Phi}_{\rm initial}$ to $\ket{\Phi}_{\rm final}$ and imposing zero "momentum" $P=0$, as one would do in the Lagrangian of a classical particle in a potential.
\begin{center}
\begin{figure}[h]
\begin{center}
\includegraphics[width=0.9\linewidth]{Instanton01.eps}
\end{center}
\caption{(a) Particle in a minimum of a potential barrier, the particle oscillates with frequency $\omega_0$ around the local minimum and tunnels
into the right well in a time $T$; (b) the instanton; and (c) the process for which a checkerboard state tunnels into the anti-checkerboard that
shown complete exchange of particle with holes and vice versa. The process happens in a time $T$ in analogy with (a).}
\label{FIG:Connection}
\end{figure}
\end{center}
Once the minimal action $S_0$ is known, the tunneling time $T$ is readily calculated \cite{BB:Wen} as
\begin{equation}
\omega_0 T = \frac{\pi}{2}e^{S_0},
\end{equation}
where $\omega_0$ is of the order of the frequency of the typical small oscillations of $\ket{\Phi}_{\rm initial}$ around the local minimum of the energy. In analogy to a classical particle tunneling through a barrier, the instanton has the nice interpretation of the stationary path connecting the two local minima in the inverted potential, as schematically represented in Fig. \ref{FIG:Connection} (b).
In units of $\hbar=1$, the imaginary time Lagrangian of a system \cite{BB:Perez} described by a quantum state $\ket{\Phi}$, is given by
\begin{equation}
\label{EQ:Lagrangian}
\mathcal{L} = \frac{\braket{\Phi}{\dot{\Phi}} - \braket{\dot{\Phi}}{\Phi}}{2} + \bra{\Phi}\hat{H}\ket{\Phi},
\end{equation}
with $\ket{\dot{\Phi}}$ indicating the time-derivative, and $\hat{H}$ being the Hamiltonian of the system. In the approximation where $\ket{\Phi}$ is the Gutzwiller wave function of a given metastable state, we write its amplitudes as
\begin{equation}
\label{EQ:GWAmplitudesStability}
f_{\rm n}^{(i)} = \frac{1}{\sqrt{2}} \left( x_{\rm n}^{(i)} + ip_{\rm n}^{(i)}\right),
\end{equation}
where $x_{\rm n}^{(i)}$, and $p_{\rm n}^{(i)}$ are real numbers which are going to be
related to the variational parameters and their conjugate momenta in the following.
For simplicity, we consider states with a maximum occupation number of
$n_{\rm max} = 1$ (i.e. $n=0,1$), and therefore we have a total of $4N_S$ parameters, $N_S$ being the total number of sites.
The Lagrangian (\ref{EQ:Lagrangian}) as well as the expectation value of the Hamiltonian $\bra{\Phi}\hat{H}\ket{\Phi}$, become functional of the $4N_S$ parameters, namely
\begin{equation}
\label{EQ:Lagrangian01}
\mathcal{L}[x_{\rm n}^{(i)},p_{\rm n}^{(i)}] = -i\sum_{i,n=0}^1 p_{\rm n}^{(i)} \dot{x}_{\rm n}^{(i)} + \bra{\Phi}\hat{H}\ket{\Phi}.
\end{equation}
After introducing the new coordinates $q_{\rm n}^{(i)} = x_{\rm n}^{(i)}$, and their conjugate momenta $P_{\rm n}^{(i)} = \partial \mathcal{L}/\partial \dot{q}_{\rm n}^{(i)} = -i p_{\rm n}^{(i)}$, we can put the Lagrangian (\ref{EQ:Lagrangian01}) in its canonical form
\begin{equation}
\label{EQ:Lagrangian02}
\mathcal{L}[q_{\rm n}^{(i)},P_{\rm n}^{(i)}] = \sum_{i,n=0}^1 P_{\rm n}^{(i)} \dot{q}_{\rm n}^{(i)} - \mathcal{H}[q_{\rm n}^{(i)},P_{\rm n}^{(i)}],
\end{equation}
where $\mathcal{H}[q_{\rm n}^{(i)},P_{\rm n}^{(i)}] = - \bra{\Phi}\hat{H}\ket{\Phi}$ is a constant of the motion \footnote{note that in the analogy of a classical particle in a potential $V(x)$, the conserved quantity in the imaginary time would be $\mathcal{H} = \frac{P^2}{2m} - V(x)$, which describes the particle's motion in the inverted potential.}. We now want to reduce the dynamic described by the Lagrangian (\ref{EQ:Lagrangian02}) to a one dimensional problem, described only by one variable $q$ and its conjugate momentum $P$. Through the variation of $(q,P)$ we want to describe the interchange between the state $\ket{\Phi}_{\rm initial}$ and $\ket{\Phi}_{\rm final}$, as for example the one represented in Fig. \ref{FIG:Connection}(c).
In \ref{SEC:Parametrization}, we show how to reduce the number of variational parameters to one, $q$, and its conjugate momentum $P$,
by making use of a variational Ansatz as well as the normalization condition on the coefficients (\ref{EQ:GWAmplitudesStability}) and the conservation of the total number of particles.
These conditions are enforced through Lagrange multipliers $\lambda_c$. Consequently the equations of motion
given by $\dot{q} = \partial \mathsf{H}/ \partial P$, and $\dot{P} = -\partial \mathsf{H}/ \partial q$
are governed by an Hamiltonian which also includes the constraints as follows
\begin{equation}
\label{EQ:HamiltonianStationary}
\mathsf{H} = \mathcal{H}[q,P] + \sum_c \lambda_c \mathcal{C}_c,
\end{equation}
where an explicit expression for the conditions $\mathcal{C}_c$ is given in \ref{SEC:Parametrization}.
The action is then readily calculated along the stationary path of Eq. (\ref{EQ:HamiltonianStationary}) as follows
\begin{equation}
S_0 = \int \mathcal{L}[q,P] \mathrm{d} \tau = \int_{\rm path} \mathcal{L}[q,P] \frac{\mathrm{d} q}{\dot{q}},
\end{equation}
with $\dot{q} = \partial \mathsf{H}/ \partial P$ from Eq. (\ref{EQ:HamiltonianStationary}).
\subsubsection{Action and tunneling time}{--- }
In Fig. \ref{FIG:Actions} (a,b) we plot the minimal action
divided by the total number of sites $N_S$ of the cell, as a function of the tunneling coefficient $J$, for two different processes. The first one (a,c), in which initial and final state are degenerate,
shows the exchange of particles with holes in the whole lattice, and is sketched in the lower part of Fig. \ref{FIG:Actions} (c). There we also plot
the potential barrier between initial and final state calculated as $-\mathsf{H}(q,P=0)$. Instead, in the second process (b,d) the final state is the ground state, i.e. deeper in energy with respect to the initial state, and only a few sites of the lattice exchange particles with holes during the process,
as sketched in the upper part of Fig. \ref{FIG:Actions} (d) along with the potential barrier.
A side remark, the point where the thick line of the barrier encounters the dashed line is called {\it bouncing point} $\ket{\Phi}_{\rm bounce}$.
\begin{center}
\begin{figure}[h!t]
\begin{center}
\includegraphics[width=0.7\linewidth]{Actions.eps}
\end{center}
\caption{(a,b) Action per site and (c,d) energy barrier for the process sketched in panels (c,d). In both cases the initial state is configuration (IIb) of Fig. \ref{FIG:Lobes} and the value $J = 0.12 U_{\rm NN}$ corresponds to the tip of its insulating lobe. The first one (a,c) is for degenerate initial
and final configurations while for the second one (b,d) the final configuration is energetically deeper. The difference in the two processes manifests also
in the height of the barrier which is smaller for the second case, leading to a smaller action and consequently a smaller life-time.}
\label{FIG:Actions}
\end{figure}
\end{center}
The action in general diverges for $J \rightarrow 0$ indicating a divergent tunneling time $T$, and then decreases monotonically up to a minimum value
in correspondence of the tip of the lobe, $J = 0.12 U_{\rm NN}$ here, signaling a minimum life-time at the tip of the lobe, as expected. In between these
two extreme behaviors, the action increases monotonically with the number of sites involved in the exchange of particles with holes; the more sites involved as in
the case of Fig. \ref{FIG:Actions} (a,c), the bigger the action is.
Summarizing, from the figures above, we conclude that small energy differences between the initial $\ket{\Phi}_{\rm initial}$
and the final states $\ket{\Phi}_{\rm final}$ and large regions of the lattice undergoing
particle-hole exchange in the tunneling process contribute to
large barriers, i.e. long life times $T$. On the contrary, for big energy differences and small regions
of the lattice undergoing particle-hole exchange, the barrier is small.
Hence, in general it is more likely for a
given state to tunnel into a state deeper in energy, e.g., the
ground state, than into its complementary, which implies the
exchange of particles with holes in the whole lattice.
\section{Multiple layers and up-down mixtures}
\subsection{Dipolar Bosons in a bilayer optical lattice}
\label{SEC:TheModel}
In \cite{BB:Trefzger02} we consider polarized dipolar particles in two decoupled
2D optical lattice layers (see Fig. \ref{FIG:Layers}), where the potential barrier between the two layers is large enough to prevent any inter-layer hopping.
This is the simplest multi-layer structure and can be
obtained by using anisotropic optical lattices or superlattices,
which can exponentially suppress tunneling in one
direction.
\begin{center}
\begin{figure}[h!]
\begin{center}
\includegraphics[width=0.6\linewidth]{Layers.eps}
\end{center}
\caption{Schematic representation of two 2D
optical lattice layers populated with dipolar bosons polarized
perpendicularly to the lattice plane. The particles feel repulsive
on site $U$ and nearest-neighbor $U_{\rm NN}$ interactions. Interlayer
tunneling is completely suppressed, while a nearest-neighbor
interlayer attractive interaction $W$ is present.}
\label{FIG:Layers}
\end{figure}
\end{center}
The in-plane dipolar interaction is isotropic
and repulsive. The interlayer interaction depends on the
relative position between the two dipoles, but is dominated
by the nearest-neighbor attractive interaction $W < 0$ between
two atoms at the same lattice site in different layers.
We include only nearest-neighbor (NN)
in-plane ($U_{\rm NN}$) and out-of-plane ($W$) dipolar interactions.
Since tunneling is suppressed between the layers particles
belonging to the different layers cannot mix and behave in
practice like two different species \footnote{Because of this analogy we will often refer
to the two layers as the two species and vice versa.}. The problem is analogous to that of two bosonic species on a 2D optical lattice
with an inter-species attraction $W < 0$ at the same lattice site, and intra-species repulsion $U_{\rm NN}$.
The relative strength between $U_{\rm NN}$ and $W$ can be tuned by
changing the spacing $d_\perp$ between the two layers, relative to
the 2D optical lattice spacing $d$. Because of the dependence
of the dipole-dipole interaction like the inverse cubic
power of the distance, the ratio $|W|/U_{NN}$ can be tuned
over a wide range. While it can be negligible for $d_\perp \gg d$
making the system asymptotically similar to a single 2D
lattice layer, it can also become relevant and give rise to
interesting physics, not existing in the single layer model as pointed out in \cite{BB:Arguelles,BB:Chen,BB:Wang,BB:Wang01,BB:Yi,BB:Wang02,BB:Klawunn,BB:Klawunn01}.
\newpage
The system is described by the Hamiltonian $\hat{H} = \hat{H}_0 + \hat{H}_1$, with
\begin{eqnarray}
\fl
\hat{H}_0 &=& \sum_{i,\sigma} \Big[ \frac{U}{2} \hat{n}_i^\sigma (\hat{n}_i^\sigma - 1) + \sum_{\langle j \rangle_i}\frac{U_{\rm NN}}{2} \hat{n}_i^\sigma \hat{n}_j^\sigma - \mu \hat{n}_i^\sigma \Big] + W \sum_i \hat{n}_i^a \hat{n}_i^b \label{EQ:H3Da}~\\
\fl
\hat{H}_1 &=& - J\sum_{\langle ij \rangle} [\hat{a}_i^\dag \hat{a}_j + \hat{b}_i^\dag \hat{b}_j ] \label{EQ:H3Db},
\end{eqnarray}
where $\sigma = \big(a, b\big)$ indicates the two species (which in the
specific case considered here are atoms in the lower and
upper 2D optical lattice layer, respectively), $U$ is the on-site
energy, $U_{\rm NN}$ the intralayer nearest neighbors repulsion,
$W$ the interlayer attraction, $J$ the intralayer tunneling parameter,
and $\mu$ the chemical potential, as schematically represented in Fig. \ref{FIG:Layers}.
The parameters $U$
and $J$ are equal for the upper and lower layers and the
chemical potentials $\mu$ are the same, since equal densities in
the two layers are assumed. Notice that since $W<0$, it is necessary to have $U+W > 0$ to avoid collapse. The symbols $\langle ij \rangle$ and $\langle j \rangle_i$ indicate nearest neighbors.
We focus on the physical situation in which the two layers are very close to one another ($d_\perp \ll d$) such that $(U+W) \ll U$, and small in-plane dipolar interactions $U_{\rm NN} \ll U$, because in these limits particles at the same lattice site $i$ of different layers pair into composites.
The composites localize in a MI state for small values of the tunneling
coefficient, while for larger values of $J$ the pairs hop around in the optical lattice forming a pair-superfluid (PSF) phase \cite{BB:Arguelles}.
Furthermore, the presence of the in-plane long-range interactions leads to the formation of a novel pair-supersolid phase (PSS), namely, a supersolid of composites \cite{BB:Mathey}. Finally, it is useful to introduce the operator of the sum of the two
species number operators at each site of the lattice, namely
\begin{equation}
\label{EQ:ms}
\hat{m}_i = \frac{\hat{n}_i^a + \hat{n}_i^b}{2},
\end{equation}
which is diagonal on a given Fock state $\ket{m_i}_i$, with $m_i = (n_i^a + n_i^b)/2$ being the average occupation per layer at the site $i$.
\subsubsection{Low-energy subspace and effective Hamiltonian}{--- }
\label{SEC:LES}
In the limit where $(U+W),U_{\rm NN},J\ll U$, there exist a low-energy subspace spanned by the states
\begin{equation}
\label{EQ:Alpha}
\ket{\alpha} = \prod_i \ket{n_i,n_i}_i,
\end{equation}
with equal occupation of the two species $a$ and $b$ at each site. These states can be equivalently written in terms of the pair occupations as
$\ket{\alpha} = \prod_i \ket{m_i}_i$, where $m_i$ is the number of pairs at site $i$, and is given by the eigenvalue of operator (\ref{EQ:ms}).
Single particle-hole excitations necessarily break one pair, and the energy cost of these excitations is
of the order of $U$. Therefore, in the limit where $(U+W),U_{\rm NN},J\ll U$ breaking one pair is
energetically very costly. In the limit $(U + W)/U \rightarrow 0$, asymptotically all states
$\ket{\alpha}$ become stable with respect to single-particle-hole excitations
and develop an insulating lobe at finite $J$, which tend to overlap as shown in Fig \ref{FIG:CB_2CB} (left) by the thin blue lobes.
However in this system, and for the above choice of parameters, single-particle-hole excitations are not the lowest lying ones. One
should also consider two-particles-two-holes excitations, which can be accounted for within a low-energy theory described by
an effective Hamiltonian acting on the low-energy subspace spanned by the states $\ket{\alpha}$-s. The validity of the effective
Hamiltonian relies on the existence of a low-energy subspace
well separated in energy from the subspace of virtual
excitations, to which it is coupled through the tunneling Hamiltonian $\hat{H}_1$ (\ref{EQ:H3Db}) via single-particle hopping.
The relevant virtual subspace is obtained from the states $\ket{\alpha}$ by breaking one composite, namely
\begin{equation}
\label{EQ:Subspace}
\eqalign{
\ket{\gamma_{ij}^{(a)}} &= \frac{\hat{a}_i^\dag \hat{a}_j }{\sqrt{n_j^{a}(n_i^{a} + 1 )}} \ket{\alpha} \\
\ket{\gamma_{ij}^{(b)}} &= \frac{\hat{b}_i^\dag \hat{b}_j }{\sqrt{n_j^{b}(n_i^{b} + 1 )}} \ket{\alpha},
}
\end{equation}
as schematically represented in Fig. \ref{FIG:Subspace}. All other states are not coupled to $|\alpha\rangle$ via single particle hopping and hence do not contribute.
\begin{center}
\begin{figure}[h!]
\begin{center}
\includegraphics[width=0.85\linewidth]{Subspace01.eps}
\end{center}
\caption{Schematic representation of the low-energy subspace. The state $\ket{\alpha}$ is uniformly occupied by one particle per site, and its energy is given by $E_\alpha$. A second order process in the tunneling connects the two states $\ket{\alpha}$, and $\ket{\beta}$, through the state $\ket{\gamma}$, which belongs to the virtual subspace. It is
straightforward to notice that in the limit of $(U + W), U_{\rm NN}, J \ll U$, the energies above satisfy the necessary condition $E_\alpha,E_\beta \ll E_\gamma$ for the existence of the subspace.}
\label{FIG:Subspace}
\end{figure}
\end{center}
The energy difference between the virtual states $\ket{\gamma}$, and the states $\ket{\alpha}$, is given by the sum of single-particle plus single hole-excitation energies of the states $\ket{\alpha}$, which are of the order of $U$ at $J=0$, and are minimized by the width of the lobes $\ket{\alpha}$ at finite $J$ (see, e.g., Fig. \ref{FIG:CB_2CB} (left)).
Slow processes drive the system through different states of the low energy subspace via second order tunneling; this happens through a fast coupling with the virtual subspace. Since we are interested in the long time physics of the system, we have to average out all the fast processes and therefore we write
an effective Hamiltonian in the subspace of pairs, and include tunneling through second order perturbation
theory \cite{BB:Tannoudji02,BB:Kuklov}. In the pair-state basis, the matrix elements of such a Hamiltonian in second order perturbation theory are given
by
\begin{eqnarray}
\bra{\alpha}\hat{H}_{\rm eff}\ket{\beta} &=& \bra{\alpha}\hat{H}_0 \ket{\beta} - \frac{1}{2} \sum_\gamma \bra{\alpha}\hat{H}_1 \ket{\gamma} \bra{\gamma}\hat{H}_1 \ket{\alpha}\nonumber \\
& & \times \left[ \frac{1}{E_\gamma - E_\alpha} + \frac{1}{E_\gamma - E_\beta} \right], \label{EQ:DefHeff}
\end{eqnarray}
where $\hat{H}_0$, given by the interaction terms (\ref{EQ:H3Da}), is diagonal on the states $\ket{\alpha}$, and the
single-particle tunneling term $\hat{H}_1$ in Eq. (\ref{EQ:H3Db}) is treated at second order.
For a given state $|\alpha\rangle$,
\begin{eqnarray}
E_{\gamma_{ij}}-E_\alpha=U+(U+W)(m_i-m_j)+ U_{\rm NN} \Delta
m^{ij}_{\rm NN}, \label{den}
\end{eqnarray}
with $\Delta m^{ij}_{\rm NN}=\sum_{\langle k\rangle_i}m_{k}-\sum_{\langle k \rangle_j} m_{k}-1$, where $m_i$
indicates the pair occupation number at site $i$ as defined in Eq. (\ref{EQ:ms}). For $(U+W),U_{\rm NN}\ll U$, the denominators
$E_{\gamma_{ij}}-E_\alpha$ are all of order $U$, which leads to
\begin{eqnarray}
\label{heff0} &&\hat{H}_{\rm eff}^{(0)} = \hat{H}_0 - \frac{2J^2}{U} \sum_{\langle ij \rangle} \left[ \hat{m}_i (\hat{m}_j+1) + \hat{c}_i^\dag \hat{c}_j \right],
\end{eqnarray}
where $\hat{c}_i$ and $\hat{c}_i^\dag$ are the pair destruction and creation operators such that
\begin{equation}
\eqalign{
\hat{c}_i |m_i\rangle &= m_i|m_i-1\rangle \\
\hat{c}_i^\dag|m_i\rangle &= (m_i+1)|m_i + 1\rangle.
}
\end{equation}
One can easily obtain corrections to $\hat{H}_{\rm eff}^{(0)}$ by expanding (\ref{den}) at higher orders in $(U+W)/U$ and $U_{\rm NN}/U$ but, as we will see,
the zeroth order is already quite accurate to describe the physics of the system for the range of parameters we consider.
\subsubsection{Insulating lobes}{--- }
We now make use of the effective Hamiltonian $\hat{H}_{\rm eff}^{(0)}$ derived above to study the ground state phase diagram of the system, starting from the insulating states.
For every classical distribution of pairs in the lattice we can calculate the pair order parameters $\psi_i = \langle \hat{c}_i \rangle$ with the perturbative mean-field method derived in Sec. \ref{SEC:Semi}, and get the expression
\begin{equation}
\label{Eqn:MF}
\psi_i = \frac{2J^2}{U}\left[ \frac{(m_i+1)^2}{E_{\scriptscriptstyle \mathrm{2P}}^i(J)}
+ \frac{m_i^2}{E_{\scriptscriptstyle \mathrm{2H}}^i(J)} \right]\bar{\psi_i},
\end{equation}
where $\bar{\psi_i} = \sum_{\langle j \rangle_i} \psi_j$, and the energy costs of adding a pair (2P) and removing a pair (2H) can be calculated from the diagonal part of Eq. (\ref{heff0}), and are respectively given by
\begin{equation}
\fl
\eqalign{
\label{EQ:PHE_2}
E_{\scriptscriptstyle \mathrm{2P}}^i(J) &= -2\mu + 2U m_i + (2m_i + 1)W + 2 V_{\rm dip}^{1,i} - \frac{2J^2}{U}\sum_{\langle k \rangle_i}(2m_k + 1)\\
E_{\scriptscriptstyle \mathrm{2H}}^i(J) &= 2\mu - 2U(m_i-1) - (2m_i - 1)W - 2 V_{\rm dip}^{1,i} + \frac{2J^2}{U}\sum_{\langle k \rangle_i}(2m_k + 1),
}
\end{equation}
with $V_{\rm dip}^{1,i} = U_{\rm NN} \sum_{\langle j \rangle_i} m_j$ being the in-plane dipolar interaction, i.e. the dipole-dipole interaction that one atom positioned
at site $i$ of the lattice, experiences with the rest of the particles belonging to the same plane.
Using Eq. (\ref{Eqn:MF}), one can calculate the
mean-field lobes for any given configuration of pairs in
the lattice. The ground state lobes for the checkerboard and doubly
occupied checkerboard are shown in Fig. \ref{FIG:CB_2CB} (right) for the 0th
(full lines) and 1st order (dashed lines) effective
Hamiltonians. The comparison between the two shows
that, for the parameters considered here, the 0th order
already captures the physics accurately. It is worth noticing that the $J^2$ dependence
of the energy of the elementary excitations is at the origin
of the reentrant behavior of the lobes, which was predicted by
exact matrix-product-state calculations for the 1D geometry
in \cite{BB:Arguelles}
\subsubsection{Pair superfluid/supersolid}{--- }
\label{SEC:BilayerSFSS}
While the MI phases are predictable through the perturbative mean-field approach of Eqs. (\ref{Eqn:MF}) for the pair order parameters, to identify
the SF phases, both PSF and PSS outside of the lobes, it is necessary to make use of the imaginary time evolution introduced in Sec. \ref{SEC:DGA} based on a Gutzwiller Ansatz for the pair wave function. Therefore, we need to calculate the dynamic equations equivalent to (\ref{EQ:FDynamics}) in the low energy subspace described by the effective Hamiltonian $\hat{H}_{\rm eff}^{(0)}$ of Eq. (\ref{heff0}).
The time dependent Gutzwiller Ansatz for the pairs, is given by
\begin{equation}
\label{EQ:GWPairs}
\ket{\Phi} = \prod_i \sum_m f_{\rm m}^{(i)} \ket{m}_i,
\end{equation}
where we allow the Gutzwiller amplitudes $f_{\rm m}^{(i)} (t)$ to depend on time. The order parameter is readily obtained from (\ref{EQ:GWPairs}) as
\begin{equation}
\psi_i = \bra{\Phi} \hat{c}_i \ket{\Phi} = \sum_m (m+1) f_{\rm m}^{*(i)} f_{\rm m+1}^{(i)}.
\end{equation}
As in Sec. \ref{SEC:DGA}, equating to zero the variation of the action with respect to $f_{\rm m}^{*(i)}$ leads to the equations
\begin{eqnarray}
\fl
i\hbar \frac{\mathrm{d}}{\mathrm{d} t} f_{\rm m}^{(i)} &=&
\Big[Um(m-1) -2\mu m + 2\Big(U_{NN} - \frac{2J^2}{U}\Big) m \sum_{\langle j \rangle_i} \langle \hat{m}_j \rangle
- \frac{2J^2}{U}z m \Big] f_{\rm m}^{(i)} \nonumber \\
\fl
&-&\frac{2J^2}{U} \Big[\bar{\psi}_i m f_{\rm m-1}^{(i)} + \bar{\psi}_i^* (m+1) f_{\rm m+1}^{(i)} \Big], \label{EQ:FDynamicsEff2}
\end{eqnarray}
where $\langle \hat{m}_i \rangle= \sum_m m|f_{\rm m}^{(i)}|^2$, the fields $\bar{\psi}_i = \sum_{\langle j \rangle_i} \psi_j$ and
$\sum_{\langle j \rangle_i} \langle \hat{m}_j\rangle$, have to be calculated self-consistently, and $z = \sum_{\langle j \rangle_i} 1$ is the coordination number in each lattice layer (here $z=4$). The solution of Eq. (\ref{EQ:FDynamicsEff2}) can be
easily obtained numerically, and by making use of the imaginary time evolution we get the ground state phase diagram of Fig. \ref{FIG:CB_2CB} (right).
\begin{figure}[h!]
\begin{center}
\begin{tabular}{cc}
\includegraphics[width=0.475\linewidth]{Validity.eps} & \includegraphics[width=0.475\linewidth]{PSSTRY.eps}
\end{tabular}
\end{center}
\caption{(left) Pair insulating lobes for $\nu=0,1/2,1,3/2$ (thick
lines); Lobes with respect to single particle-hole excitations
(thin blue lines) for the dominant configurations in the ground
state at $J=0.05U$ and $\mu=-0.4375U$, namely $m_i=0$ and $m_j=1,2$
(for $i,j$ nearest neighboring sites). The inset shows a zoom of
the pair phase diagram. (right) Phase diagram of the effective Hamiltonian.
The black full lines are the semi-analytic
solution of Eq. (\ref{Eqn:MF}) indicating the boundaries of the
insulating lobes for the checkerboard ($\nu=1/2$) and the doubly
occupied checkerboard ($\nu=1$). The black dashed lines are the
boundaries of the insulating lobes for 1st order expansion of
$\hat{H}_{\rm eff}$. The shaded area is the PSS phase predicted by the
Gutzwiller approach. The red line indicates the estimated limit of
validity of $\hat{H}_{\rm eff}^{(0)}$ (see text). The lattice parameters are $U_{\rm NN}=0.025U$ and $W = -0.95 U$, which can be obtained for
$d_{\perp}=0.37d$. }
\label{FIG:CB_2CB}
\end{figure}
To get reliable results, one should combine the Gutzwiller
predictions with an estimate of the limits of validity of $\hat{H}_{\rm
eff}^{(0)}$, beyond which the subspace of pairs looses its
meaning. Before starting the discussion on the validity
of the subspace, let us explain how we define the
dominant classical configurations of a given state $\ket{\Phi}$.
It is not difficult to see that Eq. (\ref{EQ:GWPairs}) can be equivalently written as
\begin{equation}
\label{EQ:GWPairs02}
\ket{\Phi} = \sum_{\{\vec{m}\}} g_{\vec{m}}\prod_i \ket{m_i}_i,
\end{equation}
where $\vec{m} = (m_1,...,m_i,..., m_{\scriptscriptstyle \mathrm{N_S}})$ is a collection of the indices $m$ at
each site, and we have introduced the notation such that
$g_{\vec{m}} = \prod_i f_{\rm m_i}^{(i)}$. The advantage of writing
the Gutzwiller state $\ket{\Phi}$ in the form (\ref{EQ:GWPairs02}), lays in the product over single-site Fock
states $\ket{\alpha} = \prod_i \ket{m_i}_i$, which is nothing but a classical distribution of atoms in the lattice.
Therefore we can rewrite Eq. (\ref{EQ:GWPairs02}) as
\begin{equation}
\label{EQ:GWAlphas}
\ket{\Phi} = \sum_{\{\alpha\}} g_{\rm \alpha}\ket{\alpha}.
\end{equation}
For each point of the phase diagram, from the ground
state Gutzwiller wavefunction, we define the dominant classical
configurations with the criteria
$|g_{\vec{m}}| = |\prod_i f_{\rm m_i}^{(i)}|>0.02$, and $|f_{\rm m_i}^{(i)}|^2>0.05$,
implying that each of the contributing $f_{\rm m_i}^{(i)}$
should also be sufficiently large.
For each of these configurations, we calculate the lobe with respect to single
particle-hole excitations \footnote{We have checked that the validity region is not
strongly modified upon small changes in these conditions.}. If the system at this given point of
the phase diagram turns out to be stable against all dominant
single particle-hole excitations (in other words, if this point is
inside all selected single particle-hole lobes), $\hat{H}_{\rm
eff}^{(0)}$ is considered valid at that point. This procedure is
shown for $J=0.05U$ and $\mu=-0.4375U$ in Fig.~\ref{FIG:CB_2CB} (left),
and gives the red line of Fig. \ref{FIG:CB_2CB} (right). On the right hand side of
this red line, the low energy subspace is not well defined and therefore the effective
Hamiltonian looses its meaning, leaving the description of the system to the
domain of single-particle single-hole excitation theory that predicts
SF and SS phases for each component separately.
To summarize, in the ground state phase diagram of Fig. \ref{FIG:CB_2CB} (right) we identify four different types of phases characterized by different values of the order parameters,
the single-particle order parameters for each species $\varphi_i^{a} = \langle \hat{a}_i \rangle$, $\varphi_i^{b} = \langle \hat{b}_i \rangle$,
and the pair order parameter $\psi_i = \langle \hat{c}_i \rangle$.
\begin{enumerate}
\item The Mott insulating checkerboard phases (MI) characterized by fractional filling factors $\nu$ and vanishing order parameters
$\varphi_i^{a} = \varphi_i^{b} = \psi_i = 0$ inside the lobes.
\item The pair-superfluid phase (PSF) in which the single particle order parameters are zero $\varphi_i^{a} = \varphi_i^{b} = 0$, while the pair order parameter is uniformly non-zero $\psi_i \neq 0$, signaling a finite fraction of superfluid density of the pairs.
\item The pair-supersolid (PSS) characterized by vanishing single-particle order parameters $\varphi_i^{a} = \varphi_i^{b} = 0$, and non vanishing pair order parameter $\psi_i \neq 0$, coexisting with broken translational symmetry, namely, a modulation of both density and order parameter on a scale larger than the one of the lattice
spacing, analogously to the supersolid phase.
\item We infer superfluid and supersolid phases of both components SF$_a$ and SF$_b$ (SS$_a$ and SS$_b$ respectively) with non vanishing single species order parameters $\varphi_i^{a} \neq 0$ and $\varphi_i^{b} \neq 0$, as well as non vanishing $\psi_i \neq 0$, on the right hand side of the red line of Fig. \ref{FIG:CB_2CB} (right), which we estimate to be the limit of validity of the low energy subspace for the parameters considered here.
\end{enumerate}
\subsection{Up-down mixture in a 2D optical lattice}
In most cases, if not all cases considered so far, dipolar gases are polarized, i.e. magnetic or electric dipoles point in the same direction.
This, however, does not have to be always the case. A prominent example concerns molecules that follow the Hund's rule (a) \cite{BB:Hund}. For such molecules the electric dipole is either parallel or anti-parallel to the direction of the magnetic moment. Thus, if one takes a sample of such molecules and polarizes it magnetically, say in the up direction, one will obtain in this way a dipolar gas with certain fraction of electric dipoles pointing up, and the remaining fraction pointing down. In fact there was recently a spectacular progress in cooling and trapping of magnetically confined neutral OH molecules. This progress opens important perspectives to study novel quantum phases with ultracold dipoles \cite{BB:Lev01, BB:Lev02}.
In \cite{BB:TrefzgerCSS} we consider a sample of dipoles in the presence of a 2D optical lattice, and an extra
confinement in the perpendicular direction. The dipoles are free to point in both directions normal to the lattice plane, which results in a nearest neighbor interaction either repulsive for aligned dipoles, or attractive for anti-aligned one, as shown in Fig. \ref{FIG:UpDown}.
The system is described by the Hamiltonian
\begin{eqnarray}
\hat{H} &=& \sum_{i,\sigma} \left[ \frac{U_\sigma}{2} \hat{n}_i^\sigma (\hat{n}_i^\sigma - 1) + \frac{U_{\sigma\sigma^\prime}}{2}\hat{n}_i^\sigma \hat{n}_i^{\sigma^\prime} - \mu_\sigma \hat{n}_i^\sigma \right]\nonumber \\
&+& \frac{1}{2}\sum_{ i\neq j,\sigma}\frac{U_{NN}}{|r_{ij}|^3} \left[ \hat{n}_i^\sigma \hat{n}_j^\sigma - \hat{n}_i^\sigma \hat{n}_j^{\sigma^\prime}\right] - J\sum_{\langle ij \rangle} [\hat{a}_i^\dag \hat{a}_j + \hat{b}_i^\dag \hat{b}_j ],\nonumber\\
\label{EQ:H3D}
\end{eqnarray}
where $\sigma = \big(a, b\big)$ indicates the two species, i.e. dipoles pointing in the up and
down direction perpendicular to the 2D plane of the lattice, respectively.
$U_{aa}$ and $U_{bb}$ are the on-site
energies for particles of the same species, while $U_{ab}$ is the on-site energy for different species.
The long-range dipolar interaction potential decays as the inverse cubic power of the relative distance $r_{ij}$, which we
express it in units of the lattice spacing. For computational reasons, in most theoretical approaches the range is cutoff at a certain neighbors.
In the present work, we consider a range of interaction up to the 4th nearest neighbor.
The first nearest neighbor dipolar interaction is attractive (repulsive) for particles of the same
(different) species, with strength $U_{NN}>0$ (or respectively $-U_{NN}$),
and we consider an equal tunneling coefficient $(J)$ for both species, while
the densities of the dipoles pointing upwards and downwards are fixed by the corresponding chemical potentials $\mu_\sigma$.
\begin{center}
\begin{figure}[t]
\begin{center}
\includegraphics[width=0.6\linewidth]{CSS.eps}
\end{center}
\caption{Schematic representation of a 2D
optical lattice populated with dipolar Bosons polarized
in both directions perpendicular to the lattice plane. The particles feel repulsive
intra-species $U_{aa}$, $U_{bb}$, and inter-species $U_{ab}$ repulsive on-site energies.
The nearest-neighbor interaction is repulsive $U_{\rm NN}>0$ for aligned dipoles, while it is attractive $-U_{\rm NN}$ for anti-aligned particles. The hopping term $J$ is equal for both species.}
\label{FIG:UpDown}
\end{figure}
\end{center}
The on-site interactions have two contributions (see Eq. \ref{EQ:OnSiteU}): one is arising from the s-wave scattering length, and the second one is due to the on-site dipole-dipole interaction. We consider that the s-wave scattering length is independent on the orientation of the dipoles. Instead, the on-site
dipolar contribution $U_{\rm dd}$ depends both on the orientation of the dipoles and on the geometry of the trapping potential, and it can be varied by changing the
ratio between the vertical to the axial confinement.
For simplicity we will focus on the specific case of a spherically symmetric confinement, where the on-site dipolar interactions
average out to zero $U_{\rm dd} = 0$, and the resulting on-site interactions are all equal to $U$.
We consider the case of dipole-dipole interactions to be 600 times weaker with respect to the on-site
interaction, i.e. $U_{\rm NN} = U/600$ \footnote{In standard experiments with $^{52}$Cr, which features a magnetic moment of $\mu = 6\mu_{\scriptscriptstyle \mathrm{B}}$,
this ratio is given by $U_{\rm NN} \simeq U/400$, for an optical lattice depth of $20E_{\rm R}$, where $E_{\rm R}$ is the recoil
energy at $\lambda = 500$nm.}.
The properties of the system are conveniently extracted using the operators given by the sum (filling factor) and by the difference (imbalance) of the two species number operators at each site of the lattice, namely by
\begin{equation}
\label{EQ:PMOperators}
\hat{\nu}_i = \frac{\hat{n}_i^a + \hat{n}_i^b}{2}, \;\;\; \hat{m}_i = \frac{\hat{n}_i^a - \hat{n}_i^b}{2},
\end{equation}
which are simultaneously diagonal on a given Fock state $\ket{\nu,m}_i$.
Notice that the eigenvalues of these two operators are not independent. In fact, by fixing $\nu$ the eigenvalues of $\hat{m}_i$ can only assume
$2\nu + 1$ values given by $m = \{-\nu, -\nu+1,..., +\nu\}$, in complete analogy with the angular momentum operator $\hat{S}_i^2$ and its projection along the
$z$ axis $\hat{S}_i^z$. It is also useful to introduce the average magnetization of the system, defined as
\begin{equation}
M = \frac{1}{N_{\scriptscriptstyle \mathrm{S}}}\sum_i m_i,
\end{equation}
where $N_{\scriptscriptstyle \mathrm{S}}$ is the total number of lattice sites.
Substituting Eqs. (\ref{EQ:PMOperators}) into Eq. (\ref{EQ:H3D}) allows us to express the system Hamiltonian as $\hat{H} = \hat{H}_{\rm 0}^\nu + \hat{H}_{\rm 0}^m + \hat{H}_{\rm 1}^{\nu m}$,
where
\begin{eqnarray}
\hat{H}_{\rm 0}^\nu &=& \sum_i \Big[-2\mu_{\scriptscriptstyle \mathrm{+}} \hat{\nu}_i + 2U\hat{\nu}_i\Big(\hat{\nu}_i - \frac{1}{2}\Big)\Big]\label{EQ:HNUM01} \\
\hat{H}_{\rm 0}^m &=&\sum_i \Big[-2\mu_{\scriptscriptstyle \mathrm{-}} \hat{m}_i + 2U_{NN}\sum_{j\neq i}\frac{\hat{m}_i\hat{m}_j}{|r_{ij}|^3} \Big]\label{EQ:HNUM02} \\
\hat{H}_{\rm 1}^{\nu m} &=& - J\sum_{\langle ij \rangle} [\hat{a}_i^\dag \hat{a}_j + \hat{b}_i^\dag \hat{b}_j ]. \label{EQ:HNUM03}
\end{eqnarray}
In Eqs. (\ref{EQ:HNUM01},\ref{EQ:HNUM02}), we have introduced the chemical potentials
\begin{equation}
\mu_\pm = \frac{\mu_a \pm \mu_b}{2},
\end{equation}
which respectively fix the eigenvalues of the filling factor ($+$), and the imbalance operators ($-$) in Eq. (\ref{EQ:PMOperators}).
In the following we consider $\hat{H}_{\rm 1}^{\nu m}$ to be a small perturbation on the interaction terms (\ref{EQ:HNUM01}) and (\ref{EQ:HNUM02}).
\subsubsection{Low-energy subspace and effective Hamiltonian}{--- }
In the limit where $U \gg U_{NN}$, there exist a low-energy subspace spanned by the states
\begin{equation}
\label{EQ:Alphas01}
\ket{\alpha} = \prod_i \ket{\nu,m_i}_i,
\end{equation}
with uniform total on-site occupation $2\nu$. Single particle hopping changes the total on-site population, the energy cost of these excitations is of the order of the on-site interaction energy $U$, and becomes very large in the limit where $U \gg \big(U_{NN},J\big)$.
Thus, a successful description of such a system is obtained through an effective Hamiltonian $\hat{H}_{\rm eff}$ restricted to the low-energy subspace, where single-particle hopping is suppressed and tunneling is included at second order in perturbation theory.
This situation is in fact similar to the one discussed in Sec. \ref{SEC:TheModel} for a bilayer optical lattice, and therefore we apply
the same technique to compute $\hat{H}_{\rm eff}$. In the basis of constant total on-site population $2\nu$, using Eq. (\ref{EQ:DefHeff}) we calculate the matrix elements of such a Hamiltonian at second order in perturbation theory, where the subspace of virtual excitations $\ket{\gamma}$ is obtained from the states $\ket{\alpha}$ via single particle hopping, as schematically represented in Fig. \ref{FIG:SubspacePH}.
%
\begin{center}
\begin{figure}[h!]
\begin{center}
\includegraphics[width=0.6\linewidth]{Excitations02.eps}
\end{center}
\caption{Schematic representation of a two-particle hopping between the states $\ket{\alpha}$ and $\ket{\beta}$, belonging to the low-energy subspace at
$\nu = 1/2$. These states are coupled through virtual excitations to the states $\ket{\gamma}$ by single-particle hopping.}
\label{FIG:SubspacePH}
\end{figure}
\end{center}
For a given state $|\alpha\rangle$,
\begin{eqnarray}
\label{EQ:Den02}
E_{\gamma_{ij}}-E_\alpha = U + U_{NN} \Delta m^{ij}_{\rm NN},
\end{eqnarray}
with $\Delta m^{ij}_{\rm NN}=\sum_{k\neq i}2m_k/|r_{ik}|^3 - \sum_{k\neq j} 2m_{k}/|r_{jk}|^3 - 1$, where $m_i$
indicates the population imbalance at site $i$ of Eq. (\ref{EQ:PMOperators}). For $U \gg U_{NN}$, the denominators
$E_{\gamma_{ij}}-E_\alpha$ are all of order $U$, which leads to
\begin{eqnarray}
\hat{H}_{\rm eff}^{(0)} &=& \hat{H}_{\rm 0}^\nu - \frac{2J^2}{U} \sum_{\langle ij \rangle} \hat{\nu}_i(\hat{\nu}_j+1) \nonumber \\
&+& \hat{H}_{\rm 0}^m - \frac{2J^2}{U} \sum_{\langle ij \rangle} \left[ \hat{m}_i \hat{m}_j + \hat{c}_i^\dag \hat{c}_j \right], \label{Heff0_num}
\end{eqnarray}
where $\hat{c}_i=\hat{a}_i \hat{b}_i^\dag$ and $\hat{c}_i^\dag = \hat{a}_i^\dag \hat{b}_i$ are composite operators, corresponding to the creation of a particle of one species and a hole of the other species, such that
\begin{equation}
\label{EQ:CreationAnn}
\eqalign{
\hat{c}_i |\nu,m_i\rangle &= \sqrt{\nu(\nu+1) - m_i(m_i-1)}|\nu,m_i - 1\rangle\\
\hat{c}_i^\dag|\nu,m_i\rangle &= \sqrt{\nu(\nu+1) - m_i(m_i+1)}|\nu,m_i + 1\rangle.
}
\end{equation}
\subsubsection{Insulating lobes}{--- }
In the limit where $U \gg U_{\rm NN}$, asymptotically all classical states
$\ket{\alpha}$ become stable with respect to single-particle-hole excitations
and develop an insulating lobe at finite $J$. As single particle hopping changes the total on-site population, it breaks the translational invariance of the ground state with respect to the total on-site occupation $2\nu$.
The energy cost of these excitations is of the order of the on-site interaction energy $U$, and is therefore very large in the limit where $U \gg \big(U_{NN},J\big)$.
Instead, the lowest-lying excitations within the subspace are obtained by adding (PH) or removing (HP) one composite, made of a particle of
the upper-polarized dipoles (species $a$) and a hole of the lower-polarized dipoles (species $b$), at the $i$-th site of the lattice.
For any given configuration $\ket{\alpha}$, one can calculate the corresponding energy costs from the diagonal terms of the effective Hamiltonian (\ref{Heff0_num}), which are respectively given by
\begin{equation}
\eqalign{
\label{EQ:PHE_CSS}
E_{\scriptscriptstyle \mathrm{PH}}^i(J) &=-2\mu_{\scriptscriptstyle \mathrm{-}} + 4U_{\rm NN}\sum_{k\neq i}\frac{m_k}{|r_{ik}|^3} - \frac{4J^2}{U}\sum_{\langle k \rangle_i}m_k,\\
E_{\scriptscriptstyle \mathrm{HP}}^i(J) &= 2\mu_{\scriptscriptstyle \mathrm{-}} - 4U_{\rm NN}\sum_{k\neq i}\frac{m_k}{|r_{ik}|^3} + \frac{4J^2}{U}\sum_{\langle k \rangle_i}m_k.
}
\end{equation}
The order parameters $\psi_i=\langle \hat{c}_i \rangle$
for $\ket{\alpha}$, are readily calculated as in Sec. \ref{SEC:Semi}, and satisfy the equations
\begin{eqnarray}
\fl
\psi_i = \frac{2J^2}{U}\Big[ \frac{\nu(\nu+1) - m_i(m_i+1)}{E_{\scriptscriptstyle \mathrm{PH}}^i(J)}
\; + \; \frac{\nu(\nu+1) - m_i(m_i-1)}{E_{\scriptscriptstyle \mathrm{HP}}^i(J)} \Big]\bar{\psi_i},\label{Eqn:MFCSS}
\end{eqnarray}
where $\bar{\psi_i} = \sum_{\langle j \rangle_i} \psi_j$. With Eqs. (\ref{Eqn:MFCSS}) one can calculate
the mean-field lobes of any distribution of atoms in the lattice $\ket{\alpha}$.
\begin{center}
\begin{figure}[h!]
\begin{center}
\includegraphics[width=0.6\linewidth]{Trefzger05.eps}
\end{center}
\caption{Ground state of a $4\times4$ square lattice satisfying periodic boundary conditions, for $\nu=1/2$ at
$\mu_{\scriptscriptstyle \mathrm{+}} = 0.5U$ (left), and $\nu=1$ at $\mu_{\scriptscriptstyle \mathrm{+}} = 1.4U$ (right),
and $U_{\rm NN} = U/600$. The text in parentheses $(\nu,M)$ indicate the filling factor $\nu$, and the average magnetization $M$, respectively.}
\label{FIG:GroundStateCSS}
\end{figure}
\end{center}
In Fig. \ref{FIG:GroundStateCSS} we plot the ground state insulating lobes calculated in this way for $\nu=1/2$ (left) and $\nu=1$ (right).
For all filling factors $\nu$, we find an anti-ferromagnetic (AM) ground state $(\nu,M=0)$. The AM insulating lobes are symmetric with respect to the $\mu_{\scriptscriptstyle \mathrm{-}} = 0$ axis, and present a spatial distribution of alternating
sites occupied by particles of species $a$ and $b$ resembling a checkerboard structure. Remarkably, the larger the $\nu$, the more stable is
the AM ordering with respect to flipping the direction of a dipole. By increasing the absolute value of $\mu_{\scriptscriptstyle \mathrm{-}}$ we find RM ground states with rational values of the average magnetization, corresponding to $M = (\pm 2\nu, \; \pm 4\nu, \; \pm 6\nu)/8$. The exact fractional values of $M$ in the ground state, depend on the cutoff range in the dipolar interactions, and on the size of the lattice. We have used a $4\times 4$ elementary cell with periodic boundary conditions, and dipolar interaction range cut at the 4th nearest neighbor. By considering more neighbors in the interactions, and larger lattices, we expect to find RM states appearing at all rational $M$, asymptotically approaching a Devil's staircase as recently shown in \cite{BB:Barbara2, BB:Cooper}. Finally we find a FM ground state $(\nu,M=\pm \nu)$, in which only particles of one type are present.
It is worth noticing that the insulating lobes calculated in this way, do not contain any dependence on $\mu_{\scriptscriptstyle \mathrm{+}}$,
which does not enter into Eqs. (\ref{Eqn:MFCSS}).
Therefore, for any given value of $\mu_{\scriptscriptstyle \mathrm{+}}$, in order to obtain the ground state phase diagram one has to compare the energies of the ground state configurations at different $\nu$. Using the effective Hamiltonian (\ref{Heff0_num}), for any value of $\mu_{\scriptscriptstyle \mathrm{+}}$, $J$, and $\mu_{\scriptscriptstyle \mathrm{-}}$, we compare the energies of the ground state configurations for different $\nu$, and select the state with the smaller energy. In this way we have obtained the phase diagram at $J=0$ shown in Fig. \ref{FIG:GSJ0}.
\begin{center}
\begin{figure}[h!]
\begin{center}
\includegraphics[width=0.6\linewidth]{Trefzger02.eps}
\end{center}
\caption{Ground state of the system at $J=0$, calculated for a $4\times 4$ elementary cell satisfying periodic boundary conditions, and
$U_{\rm NN} = U/600$. The text in parentheses $(\nu,M)$ indicates the filling factor $\nu$ and the average magnetization $M$.}
\label{FIG:GSJ0}
\end{figure}
\end{center}
\subsubsection{Counterflow superfluid/supersolid}{--- }
In the low-energy subspace at constant $\nu$, the Gutzwiller Ansatz on the
wave function of the system reads
\begin{equation}
\label{EQ:NuConstant}
\ket{\Phi} = \prod_i \sum_{\rm m=-\nu}^{\rm \nu} f_{\rm \nu,m}^{(i)} \ket{\nu,m}_i,
\end{equation}
where we allow the Gutzwiller amplitudes $f_{\rm \nu,m}^{(i)}(t)$ to depend on time.
We obtain the equations of motion for the amplitudes by minimizing the action of the system with respect to the variational parameters $f_{\rm \nu,m}^{*(i)}(t)$,
\begin{eqnarray}
i\hbar \frac{\mathrm{d}}{\mathrm{d} t} f_{\rm \nu,m}^{(i)} &=&
\Big[-2\mu_{\scriptscriptstyle \mathrm{-}} - \frac{4J^2}{U}\sum_{\langle j \rangle_i} \langle \hat{m}_j\rangle \nonumber \\&+& 4U_{NN}\sum_{j\neq i} \frac{\langle\hat{m}_j \rangle}{|r_{ij}|^3} \Big] m f_{\rm \nu,m}^{(i)} \nonumber \\
&-& \frac{2J^2}{U} \Big[\bar{\psi}_i \sqrt{\nu(\nu+1) - m(m-1)} \; f_{\rm \nu,m-1}^{(i)} \nonumber \\
&+& \bar{\psi}_i^*\sqrt{\nu(\nu+1) - m(m+1)} \; f_{\rm \nu,m+1}^{(i)} \Big] \label{EQ:FDynamicsEffCSS},
\end{eqnarray}
where $\langle \hat{m}_i \rangle= \sum_{\rm m=-\nu}^{\rm \nu} m |f_{\rm \nu,m}^{(i)}|^2$, the fields $\bar{\psi}_i = \sum_{\langle j \rangle_i} \psi_j$,
$\sum_{\langle j \rangle_i} \langle \hat{m}_j\rangle$, and $\sum_{j\neq i} \langle \hat{m}_j\rangle/|r_{ij}|^3$ have to be calculated in a self consistent way, and the
order parameter $\psi_i = \bra{\Phi} \hat{c}_i \ket{\Phi}$ is given by
\begin{equation}
\psi_i = \sum_{\rm m=-\nu}^{\rm \nu} \sqrt{\nu(\nu+1) - m(m+1)} \; f_{\rm \nu,m}^{*(i)} f_{\rm \nu,m+1}^{(i)}.
\end{equation}
We solve Eqs. (\ref{EQ:FDynamicsEffCSS}) in imaginary time $\tau=it$, and in Fig. \ref{FIG:GroundStateCSS} we show the ground state phase diagram of the system for $\nu = 1/2$ (left) computed in this way.
For $\nu=1/2$, in the region immediately outside the insulating AM lobe, depending on the values of $J$
and $\mu_{\scriptscriptstyle \mathrm{-}}$ we find either super-counter-fluid SCF or counterflow-supersolid CSS.
In the SCF phase, while the single-particle order parameters vanish $\langle \hat{a}_i \rangle = \langle \hat{b}_i \rangle = 0$,
the composite order parameters are non-zero $\langle \hat{c}_i \rangle \neq 0$, indicating the presence of counterflow \cite{BB:Kuklov}.
The CSS is characterized by vanishing single-particle order parameters $\langle \hat{a}_i \rangle = \langle \hat{b}_i \rangle = 0$,
and non-vanishing composite order parameters $\langle \hat{c}_i \rangle \neq 0$, coexisting with broken translational
symmetry, namely, a modulation of both $m_i$, and $\langle \hat{c}_i \rangle$ on a scale larger than the one of the lattice
spacing, analogously to the supersolid phase. Note that we don't find any evidence of CSS at
$\mu_{\scriptscriptstyle \mathrm{-}} = 0$, which indicates that a finite imbalance between the two
species is a necessary condition for the system in order to sustain CSS.
Finally, with a similar method described in Sec. \ref{SEC:BilayerSFSS}, we estimate the limits of validity of the effective Hamiltonian to be given by the vertical thick lines of Fig. \ref{FIG:GroundStateCSS}.
\section{Path Integral Monte Carlo and the Worm algorithm}
Any physical system consisting of $N$ non-relativistic particles can be in principle described by the many-body Schr{\"o}dinger equation. In three dimensions (3D),
the number of degrees of freedom in the Schr{\"o}dinger equation becomes 3 times $N$. For typical physical systems such as electrons in conducting materials or BEC that have a large number of constituents $N$, the Schr{\"o}dinger equation becomes difficult to solve exactly in a reasonable amount of time even for parallel computing. Monte Carlo methods overcome this problem, they allow for a description of the many-body system relying on repeated random sampling, at the cost of statistical uncertainty which can be reduced with more simulation time. The typical basic steps of a Monte Carlo algorithm can be summarized as follows
\begin{enumerate}
\item Define the {\it configuration space} (here by configuration we mean a collection of indices, see below).
\item Generate configurations randomly and accept them with a certain probability which depends on the specific problem. This is called
the {\it updating procedure}.
\item Perform a computation, i.e. calculate quantities of interest, based on the randomly generated configurations.
\item Add the result of the computation to the final result.
\end{enumerate}
In the spirit of the Metropolis-Hastings algorithm \cite{BB:Metropolis,BB:Hastings}, two requirements must be satisfied: a) ergodicity, that is, given an initial configuration it has to be possible to reach any other allowed configuration via the updating procedure; b) the probability of having a certain configuration appearing in sums which calculate quantities of interest has to be proportional to its Boltzmann weight.
There is a large class of Quantum Monte Carlo methods that can simulate quantum many-body systems, like for example the Variational Monte Carlo
\cite{BB:Polls, BB:Variational}, the Diffusion
Monte Carlo \cite{BB:Diffusion01, BB:Greg, BB:Greg01}, the Path Integral Monte Carlo \cite{BB:PathIN, BB:Worm,BB:Worm2}, auxiliary field Monte Carlo \cite{BB:Auxiliary01, BB:Auxiliary02}, etc. Most methods aim at computing the ground-state wavefunction of the system, with the exception of Path Integral Monte Carlo, and finite-temperature auxiliary field Monte Carlo, which calculate the density matrix. The results presented in this work are based on the Path Integral Monte Carlo (PIMC) and the Worm algorithm (WA), which was originally developed by Prokof'ev, Svistunov and Tupitsyn \cite{BB:Worm,BB:Worm2}.
\subsection{Path Integral Monte Carlo}
\label{SEC:PIMC}
Consider a system described by the Hamiltonian $\hat{H} = \hat{H}_0 + \hat{H}_1$, where $\hat{H}_0$ is diagonal in the basis
of eigenstates $\{\ket{\alpha}\}$ satisfying the eigenvalue equation
\begin{equation}
\hat{H}_0 \ket{\alpha} = E_\alpha \ket{\alpha},
\end{equation}
and $\hat{H}_1$ is non-diagonal. The thermodynamic properties of the system at equilibrium, can be derived from the partition function which is given by the
trace of the density matrix operator $Z = {\rm Tr} \left[e^{-\beta \hat{H}} \right]$, where $\beta = 1/K_BT$ is the inverse temperature and $K_B$
the Boltzmann constant. In the interaction picture \cite{BB:Tannoudji01} one may write
\begin{equation}
\label{EQ:Z1}
Z = {\rm Tr} \left[ e^{-\beta(\hat{H}_0 + \hat{H}_1)}\right] = {\rm Tr} \left[ e^{-\beta \hat{H}_0} \hat{\mathcal{T}}_\tau e^{-\int_0^\beta \mathrm{d} \tau \hat{H}_1(\tau)} \right],
\end{equation}
where $\hat{\mathcal{T}}_\tau$ is the time-ordering operator, $\hat{H}_1(\tau) = e^{\tau \hat{H}_0} \hat{H}_1 e^{-\tau \hat{H}_0}$ is the non-diagonal part of the Hamiltonian expressed in the interaction picture, and the variable $\tau$ is usually called the {\it imaginary time} \footnote{This is because by replacing $\tau = it$, with $t$ being the real time, the operator $e^{-\tau\hat{H}}$ becomes the usual time-evolution operator in quantum mechanics.}.
One can write the partition function using the Feynman path integral formulation, and by Taylor expanding the second exponent in the right-hand-side of Eq. (\ref{EQ:Z1}) one gets
\begin{eqnarray}
Z &=& \sum_{\alpha} e^{-\beta E_\alpha}\bra{\alpha} \hat{\openone} - \int_0^\beta \mathrm{d} \tau \hat{H}_1(\tau) + \nonumber \\
&+& \sum_{m=2}^\infty (-1)^m \int_0^\beta\mathrm{d} \tau_{\rm m}\;..\; \int_0^{\tau_2}\mathrm{d} \tau_1 \hat{H}_1(\tau_{\rm m}) \;..\; \hat{H}_1(\tau_1)\ket{\alpha}, \label{EQ:Z2}
\end{eqnarray}
where the integrals are ordered in time and the sum over the states $\ket{\alpha}$ comes from the trace.
Now we explicitly make use of the completeness property of the $\{\ket{\alpha}\}$ basis, and insert $m-1$ identity operators
$\hat{\openone} = \sum_{\alpha} \ket{\alpha} \bra{\alpha}$ between the products of $\hat{H}_1(\tau_{\rm m})$ operators, therefore we can write
\begin{eqnarray}
\fl
\bra{\alpha} \hat{H}_1(\tau_m) \;..\; \hat{H}_1(\tau_1) \ket{\alpha} = \sum_{\alpha_1,..,\alpha_{\rm m-1}} H_1^{\alpha \alpha_{\rm m-1}} (\tau_{\rm m})\;..\;
H_1^{\alpha_2 \alpha_1}(\tau_2)H_1^{\alpha_1 \alpha}(\tau_1),
\end{eqnarray}
where the matrix elements
\begin{equation}
\label{EQ:MatrixElements}
H_1^{\alpha^\prime \alpha}(\tau) = e^{\tau E_{\alpha^\prime}} H_1^{\alpha^\prime \alpha} e^{-\tau E_{\alpha}} = \bra{\alpha^\prime} \hat{H}_1 \ket{\alpha}
e^{-\tau(E_{\alpha} - E_{\alpha^\prime})},
\end{equation}
contain both diagonal ($E_\alpha$) and off-diagonal ($H_1^{\alpha^\prime \alpha}$) matrix elements. We now insert the last equation into expression
(\ref{EQ:Z2}) and get the final expression for the partition function
\begin{eqnarray}
\fl
Z &=& \sum_\alpha e^{-\beta E_\alpha}\Big\{1 - \int_0^\beta \mathrm{d} \tau H_1^{\alpha\alpha} (\tau) + \label{EQ:Z}\\
\fl
&+& \sum_{\rm m=2}^\infty(-1)^m \int_0^\beta \mathrm{d} \tau_{\rm m} \;..\; \int_0^{\tau_2} \mathrm{d} \tau_1 \sum_{\alpha_1,..,\alpha_{\rm m-1}}
H_1^{\alpha\alpha_{\rm m-1}} (\tau_{\rm m}) \;..\; H_1^{\alpha_1\alpha} (\tau_1) \Big\}, \nonumber
\end{eqnarray}
which contains only matrix elements of the operators $\hat{H}_0$ and $\hat{H}_1$. Therefore, by using this formalism of path integrals,
the calculation of the partition function reduces to a classical problem since only scalars enter into Eq. (\ref{EQ:Z}), but we have payed
the price of the extra dimension $\tau$. In other words, the original $d$-dimensional quantum system is equivalent to a ($d+1$)-dimensional
classical system.
It is worth noticing that since the partition function is a trace, periodic boundary conditions in the imaginary time $\tau$ must apply. This is
easily understood by looking at the $m$-th order term of $Z$, which contains the product of $m$ matrix elements
$H_1^{\alpha\alpha_{\rm m-1}} (\tau_{\rm m}) \;..\; H_1^{\alpha_1\alpha} (\tau_1)$ that are ordered in time from the first at $\tau_1$, to the last
at $\tau_m$. Therefore, for any given $\alpha$ in the trace, the first matrix element brings $\alpha$ to some $\alpha_1$ at time $\tau_1 \geq 0$,
while the last matrix element brings $\alpha_{\rm m-1}$ back to $\alpha$ at time $\tau_{\rm m} \leq \beta$.
All the possible configurations which are periodic in imaginary time and that enter into the expression for the partition function Eq. (\ref{EQ:Z}), define the configuration space spanned by a PIMC algorithm.
\subsubsection{Path Integral Monte Carlo and the 2D extended Bose-Hubbard model}{--- }
\label{SEC:PIMC_BH}
We now consider a 2D system of $L\times L$ sites filled with polarized dipolar Bosons, we assume spatial periodic boundary conditions and the dipoles to be polarized perpendicularly to the 2D plane as explained in Sec. \ref{CH:MS:DipolarBosons}. The system is therefore described by the extended Bose-Hubbard Hamiltonian (\ref{EQ:EBH2D}), which, to be consistent with the notations in our publication \cite{BB:Barbara2}, we rewrite in this form
\begin{equation}
\label{EQ:MCH}
\hat{H} = - J \sum_{\langle i j\rangle} \hat{b}^{\dag}_i \hat{b}_j + \sum_i \left[\frac{U}{2}\hat{n}_i(\hat{n}_i-1) -\mu_i \hat{n}_i\right] +
V\sum_{i<j} \frac{\hat{n}_i \hat{n}_j}{r_{ij}^3},
\end{equation}
where $\hat{b}^{\dag}_i$ ($\hat{b}_i$) is the boson creation (annihilation) operator at site $i$, $\hat{n}_i = \hat{b}^{\dag}_i \hat{b}_i$ is the number operator, $V = D/a^3 >0$
is the dipole-dipole interaction strength $D$ divided by the lattice spacing $a$, $r_{ij} = |i-j|$ is the distance between two sites of the lattice, and $\mu_i = \mu - \Omega i^2$
contains the chemical potential $\mu$ which fixes the number of particles, and the curvature $\Omega$ of an external harmonic confinement.
We choose to work in the basis of the interaction term of the Hamiltonian (\ref{EQ:MCH}), i.e. Fock states $\ket{\alpha} = \prod_i^{L^2} \ket{n_i}_i$ of localized particles in the $L\times L$ square lattice, where $n_i$ is the occupation number at site $i$. Therefore in this basis, the diagonal matrix elements entering Eq. (\ref{EQ:MatrixElements}) take the form
\begin{equation}
\label{EQ:Diagonal}
E_\alpha = \frac{U}{2}\sum_i n_i(n_i-1) - \sum_i \mu_in_i + V\sum_{i<j} \frac{n_i n_j}{r_{ij}^3}.
\end{equation}
The off-diagonal ones are given by the expression
\begin{equation}
\label{EQ:OffDiagonal}
-H_1^{\alpha^\prime\alpha} = J\bra{\alpha^\prime} \hat{b}^{\dag}_i \hat{b}_j \ket{\alpha} = J \sqrt{(n_i^{\alpha}+1)n_j^{\alpha}},
\end{equation}
and they connect states $\ket{\alpha^\prime}$ and $\ket{\alpha}$ that differ only in the occupation number of the two nearest neighboring sites $i$ and $j$, namely
$\ket{\alpha^\prime} \equiv \frac{\hat{b}^{\dag}_i \hat{b}_j}{\sqrt{(n_i^{\alpha}+1)n_j^{\alpha}}} \ket{\alpha} $ with $n_i^\alpha$ being the integer number of particles at the $i$-th site of the state $\ket{\alpha}$.
To write the partition function for the 2D extended Bose-Hubbard model, we notice that the first order term vanishes since the matrix elements (\ref{EQ:OffDiagonal}) are off-diagonal, i.e. $H_1^{\alpha\alpha} =0$, and due to the geometry of the system (2D square lattice) it is not difficult to see that all the terms with an odd value of $m$
also vanish, owing to periodic boundary conditions in imaginary time. Therefore by rearranging the exponentials and renaming $\alpha \equiv \alpha_0$, we get to the expression
\begin{eqnarray}
\label{EQ:ZeBH0}
\fl
Z_{\scriptscriptstyle \rm eBH} &=& \sum_{\alpha_0} e^{-\beta E_{\alpha_0}} + \sum_{\rm m=2}^\infty
J^m A_{\rm m} \times \\
\fl
&\times& \int_0^\beta \mathrm{d} \tau_{\rm m} \;..\; \int_0^{\tau_2} \mathrm{d} \tau_1 \sum_{\alpha_0,\alpha_1,..,\alpha_{\rm m-1}}
\exp \Big\{-\beta E_{\alpha_0} -\sum_{p=0}^{m-1} E_{\alpha_p}(\tau_{p+1} - \tau_p)\Big\}, \nonumber
\end{eqnarray}
where $A_{\rm m}$ is a product of $m$ square root factors coming from Eq. (\ref{EQ:OffDiagonal}) and we have introduced $\tau_0 = \tau_m$ to compact the notation.
We can compact further the notation by defining $A_{\rm m=0} = 1$, and keeping in mind that for $m=0$ there is no summation in the exponent of Eq. (\ref{EQ:ZeBH0})
we then write the partition function as follows
\begin{eqnarray}
\label{EQ:ZeBHCompact}
\fl
Z_{\scriptscriptstyle \rm eBH} &=& \sum_{\rm m=0}^\infty \; \sum_{\alpha_0,\alpha_1,..,\alpha_{\rm m-1}} J^m A_{\rm m} \times \\
\fl
&\times&\int_0^\beta \mathrm{d} \tau_{\rm m} \;..\; \int_0^{\tau_2} \mathrm{d} \tau_1
\exp \Big\{-\beta E_{\alpha_0} -\sum_{p=0}^{m-1} E_{\alpha_p}(\tau_{p+1} - \tau_p)\Big\}. \nonumber
\end{eqnarray}
From the last expression one can formally write
\begin{equation}
Z_{\scriptscriptstyle \rm eBH} = \sum_{\rm \nu} W_{\rm \nu},
\end{equation}
with $W_{\rm \nu}$ being the weight of each configuration $\nu \equiv \left[m,\alpha_0(\tau_1),\alpha_1(\tau_2),..,\alpha_{\rm m-1}(\tau_m)\right]$. A configuration can be
pictorially represented as in Fig. \ref{FIG:Configuration}, with imaginary time $\tau$ on the horizontal axis, and lattice sites on the vertical axis.
\begin{center}
\begin{figure}[h]
\begin{center}
\includegraphics[width=0.8\linewidth]{conf.eps}
\end{center}
\caption{Schematic representation of one configuration which enters the calculation of $Z_{\scriptscriptstyle \rm eBH}$. Each line is called a worldline and it represents a trajectory of a particle in imaginary time. Configurations have to fulfill periodic boundary conditions in the imaginary time $\tau$, owing to the definition of the partition function as a trace. Vertical arrows correspond to changes of the state of the system, and are called kinks. In the sketch, the thickness of intervals between the kinks shows the number of particles: the dashed black line is for $n$ particles while the solid and bold blue lines have occupation numbers equal to $n+1$ and $n+2$ respectively. }
\label{FIG:Configuration}
\end{figure}
\end{center}
Each line represents a trajectory of a particle in imaginary time and is called a {\bf worldline}. The latter has to close on itself owing to the fact that the partition function is a trace. Moreover if one assumes spatial periodic boundary conditions one can imagine the configuration of Fig. \ref{FIG:Configuration} to be wrapped on a torus (in the case of one dimensional systems). We call the phase space of all possible configurations the {\it closed path configuration space} (CP).
If we cut one configuration at a certain instant in imaginary time, we get the system in a particular quantum state. The points in imaginary time where the system changes state are called {\bf kinks}, which in Fig. \ref{FIG:Configuration} are represented by vertical arrows. A configuration with a number of kinks equal to $m$,
contributes to the $m$-th order term of the partition function Eq. (\ref{EQ:ZeBHCompact}), and it is straightforward to see that there exist an infinite number of different configurations with the same number of kinks, the difference being the time at which the kinks take place and/or the different states they connect. The updating procedure of a PIMC algorithm therefore consists of changing the number of kinks and/or their position in imaginary time. We will discuss the updating procedure specifically
for the Worm algorithm in the next section.
\subsection{The Worm algorithm}
\label{SEC:WAGF}
The Worm algorithm, originally developed by Prokof'ev, Svistunov and Tupitsyn \cite{BB:Worm,BB:Worm2}, works in an enlarged configuration space, in which
one allows one disconnected worldline, the worm, drawn as a red line in Fig. \ref{FIG:Configuration_Worm}. This is equivalent to work in the Grand-Canonical ensemble, as we shall discuss in Sec. \ref{SSEC:Updating}. The disconnected worldline allows to efficiently collect statistics for calculating the Matsubara Green function, defined as
\begin{equation}
\label{EQ:MatsubaraGreen}
G(j,\tau) = \langle \hat{\mathcal{T}}_\tau \hat{b}_{\rm i+j}(\tau_{\rm 0} + \tau)\hat{b}_{\rm i}^\dag (\tau_{\rm 0})\rangle,
\end{equation}
where $\hat{\mathcal{T}}_\tau$ is the time-ordering operator, $\tau_{\rm 0}$ and $\tau$ are two points in imaginary time, $i$ and $j$ are two sites of the lattice, and the symbol $\langle . \rangle$ stands for the statistical average of the expectation value of an operator. Due to space and imaginary time translational invariance of the system, the Green function Eq. (\ref{EQ:MatsubaraGreen}) does not depend on $i$ and $\tau_{\rm 0}$. The configuration space of the Matsubara Green function is called the $CP_{\rm g}$ space, and it is easy to see that the only difference between configurations contributing to the partition function $Z_{\scriptscriptstyle \rm eBH}$ and those contributing to the Green function $G$ is that, for the latter, one of the worldlines starts at $(i,\tau_0)$ and ends at $(i+j,\tau_0 + \tau)$, i.e. the worldline is disconnected.
\begin{center}
\begin{figure}[h]
\begin{center}
\includegraphics[width=0.8\linewidth]{conf_worm.eps}
\end{center}
\caption{Configuration of the $CP_{\rm g}$ space, the red disconnected line represents the worm.}
\label{FIG:Configuration_Worm}
\end{figure}
\end{center}
\subsubsection{Updating procedures}{--- }
\label{SSEC:Updating}
Let us now discuss the updating procedure of the Worm algorithm, that is when the system is in a certain configuration $\nu$ and the algorithm
has to generate randomly a new configuration $\nu^\prime$ to collect statistics for evaluating the observables of interest.
Apart from the creation of a worm, which is done in the $CP$ space, all other
updates are done in the $CP_{\rm g}$ space through the two ends of the worm. Nearly all updates are done in pairs. One can
picture the updating scheme as sequence of 'drawing' and 'erasing' procedures, happening
at the end points of the worm. Below we list and describe the four types of updates the Worm algorithm uses to simulate single component Bose-Hubbard models.
\paragraph{Creation of a worm --- }
Creating a worm is the only update performed in the $CP$ space, therefore the starting point is a configuration $\nu$ belonging to $CP$. Each configuration can be thought as divided into intervals, each interval being delimited by kinks shown in Fig. \ref {FIG:WormCreation} as crosses (or worm extremities for configurations in the $CP_{\rm g}$ space).
In the create worm update one of the intervals of $\nu$, delimited by $\tau_{\rm min}$ and $\tau_{\rm max}$ (see interval $n_1$ in Fig. \ref {FIG:WormCreation}), is randomly selected. Then the algorithm suggests at random two points $\tau_{\rm 1}$ and $\tau_{\rm 2}$ within $n_1$, which will be the worm extremities (indicated by plain dots in Fig. \ref {FIG:WormCreation}), with the constraint $\tau_{\rm min} < \tau_{\rm 1} < \tau_{\rm 2} < \tau_{\rm max}$. With equal probability one suggests to draw or delete the piece of worldline delimited by $\tau_{\rm 1}$ and $\tau_{\rm 2}$, with the constraints that the resulting configuration belongs to the Hilbert space, i.e. it is not possible to erase from an empty interval or to draw on an interval which has reached the maximum occupation number allowed, if any. The worm is therefore created and all other updates will take place through its two extremities.
\begin{center}
\begin{figure}[h]
\begin{center}
\includegraphics[width=0.9\linewidth]{WormCreation.eps}
\end{center}
\caption{Creation/Deletion of a worm. In the create worm update an interval is randomly selected (top), and two points $\tau_{\rm 1}$ and $\tau_{\rm 2}$, which will become the two extremities of the worm, are randomly chosen. Then one can either delete a piece of worldline (bottom left) or draw a piece of worldline (bottom right) with the constraint that the Hilbert space is observed.}
\label{FIG:WormCreation}
\end{figure}
\end{center}
\paragraph{Deletion of a worm --- }
In analogy, the update opposite to creation of a worm is the deletion of a worm. It can only take place in the $CP_{\rm g}$ space and only if the two extremities of
the worm belong to the same interval.
\paragraph{Time shift --- }
This is the simplest of the updates and it consists of moving one of the extremities of the worm in the imaginary time direction, such as to lengthen or shorten the size of the worm.
The algorithm selects at random the imaginary time instant to which the extremity of the worm will be moved.
\paragraph{Space shift --- }
This update changes the number of kinks and it consists of creating or deleting a kink to the left (space shift left) or to the right (space shift right) of the worm extremity. Fig. \ref{FIG:SpaceShift}(a) shows the creation/deletion of a kink backward in imaginary time, i.e. the space shift left. In the creation update a nearest neighbor of the site to which the worm extremity belongs is selected at random and the kink is inserted at an imaginary time instant within the interval delimited by $\tau_{\rm min}$ and $\tau_{\rm max}$ [see Fig. \ref{FIG:SpaceShift}(a)], i.e. with the requirement that the created (or deleted) kink does not interfere with any other interval.
\begin{center}
\begin{figure}[h]
\begin{center}
\includegraphics[width=0.9\linewidth]{SpaceShift.eps}
\end{center}
\caption{Sketch of space shift updates that create or delete kinks. In the space shift left (a), one kink is created/deleted backward in the imaginary time, i.e. to the left of the worm extremity; in the space shift right (b) the kink is created/deleted to the right, i.e. forward in the imaginary time.}
\label{FIG:SpaceShift}
\end{figure}
\end{center}
The last update, the space shift right shown in Fig. \ref{FIG:SpaceShift}(b), is equivalent to the left one with the only difference that the kink is inserted or
deleted to the right of the worm extremity, i.e. forward in imaginary time.
These are all the updates performed by the WA. It is straightforward to see that the WA works in the Grand Canonical ensemble, i.e. allows to change the number of worldlines present in the configurations. The chemical potential becomes an input parameter which fixes the average particle number. For example, suppose the algorithm starts with an initial configuration $\nu$ of zero particles
in the system. From this configuration the only possible update is to create a worm by drawing a piece of worldline. Then, trough the space shift and time shift updates, the worldline will eventually close on itself, corresponding to the insertion of one particle in the system.
\subsubsection{Advantages of the Worm algorithm}{--- }
\label{SEC:Advantages}
The updates described above are all \emph{local} and allow
to draw/erase any line, and create kinks between the sites.
Although only configurations belonging to the $CP$ space
contribute to the evaluation of the partition function,
by using the enlarged
configuration space $CP+CP_{\rm g}$ the intermediate configurations with one disconnected loop allow to
efficiently collect statistics for the Green function. For an
algorithm working in the $CP$ space only, instead, collecting
statistics for the Green function results computationally very
expensive.
Another advantage of the WA is that it does not suffer
from critical slowing down in the vicinity of a critical point. In
the critical region, a system develops long range
correlations, and in most cases an algorithm based on local updates
results very inefficient in simulating such a system for which the
relevant degrees of freedom are non-local. This results in the
divergence of the autocorrelation time with the system size.
Although the WA performs local updates, it overcomes this problem by using the drawing and erasing
updating procedures through the worm ends, which are directly linked to
the critical modes (long range order in $G(j,\tau)$).
As a result generating independent configurations in the critical region is very efficient.
The WA is also efficient in sampling
topologically different configurations and configurations which
are separated by an energy barrier. This property is a necessary condition in order to maintain ergodicity.
An example of two topologically different
configurations is shown in Fig. \ref{WN}, where a one-dimensional
system with one particle (worldline) is considered. Periodic
boundary conditions in time and space apply, i.e. the system is a
torus where the bottom and top facets of the cylinder are glued
together. Fig.~\ref{WN}(a) represents a configuration with zero
winding numbers, i.e. the worldline does not `wind' in space.
Fig.~\ref{WN}(d), instead, represents a configuration with
one winding number, i.e. the worldline winds once in space. An algorithm based on local updates which only works in the
$CP$ space would not allow to sample configurations with different
winding numbers, unless a global
update which introduces a winding number at once, is introduced.
The WA, instead, can easily go from
configuration of Fig.~\ref{WN}(a) to configuration
Fig. \ref{WN}(d) (see a sketch in Fig. \ref{WN}(b)-(c)).
Being able to sample configurations with different winding
numbers is crucial in order to simulate SF systems. It was shown
in \cite{BB:Ceperley}, that the superfluid stiffness can be
extracted from the statistics of winding numbers
\begin{equation}
\rho_s=\frac{T\langle\textbf{W}^2\rangle}{dL^{d-2}}\; ,
\label{SF_formula}
\end{equation}
where $T$ is the temperature, $L$ the system size, $d$ the
dimensionality, and $\textbf{W}^2=\sum_{i=1}^dW_i^2$, with $W_i$
being the winding number in the coordinate $i$.
\begin{figure}[h!]
\begin{center}
\includegraphics[width=1\linewidth]{WN01.eps}
\caption{One-dimensional system with (a) zero and
(d) one winding number(s). (b)-(c) sketch of how the WA is able to go from (a) to (d).}
\label{WN}
\end{center}
\end{figure}
\section{Quantum Monte Carlo studies of dipolar gases}
\label{CH:PaperQuantumPhase}
Quantum Monte Carlo is one of the most powerful
methods we have to study equilibrium properties of strongly
interacting many-body quantum systems.
In the literature, there is a large amount of work devoted to the study of dipolar gases
with Quantum Monte Carlo techniques.
From self-assembled floating lattices, provided by trapped polar molecules \cite{BB:Pupillo01},
to the possibility of tuning, and shaping the long-range interaction potential
of polar molecules \cite{BB:Pupillo02}, to self-organized mesoscopic structures of matter waves
in zigzag chains \cite{BB:Greg01}, to the spectrum of the elementary excitation that can exhibit a roton minimum \cite{BB:Boronat02, BB:Boronat01},
to the emergence of an emulsion phase in triangular lattices \cite{BB:Pollet}.
The ones listed above are just a few of the outstanding properties of dipolar gases, which have been investigated
with various Monte Carlo techniques.
Based on the Path Integral Monte Carlo and the Worm algorithm, in \cite{BB:Barbara2}, we have studied the ground state properties of dipolar hard-core Bosons confined in a 2D square lattice of linear size $L$, satisfying periodic boundary conditions. The system is described by the extended Bose-Hubbard Hamiltonian (\ref{EQ:MCH}), where no cut-off in the dipolar interaction potential is used.
\begin{center}
\begin{figure}[t]
\begin{center}
\includegraphics[width=0.9\linewidth]{PhaseDiagramMC.eps}
\end{center}
\caption{Phase diagram corresponding to
the Hamiltonian Eq. (\ref{EQ:MCH}) as a function of $\mu$ and
$J$ at zero temperature. Lobes: Mott solids (densities indicated);
SS: supersolid phase; SF: superfluid phase. DS: parameter
region where devil' s staircase is observed. Panels
(b-d): sketches of the groundstate configuration for the Mott
solids in panel (a), with $\rho = 1/2$, $1/3$ and $1/4$, respectively.}
\label{FIG:PhaseDiagramD}
\end{figure}
\end{center}
\subsection{Incompressible and supersolid phases}
\label{SEC:Incompressible}
The incompressible and supersolid phases are both characterized by
a finite value of the structure factor, defined as
\begin{equation}
S(\myvec{k}) = \sum_{\myvec{r},\myvec{r}^\prime} \frac{\langle n_{\myvec{r}} n_{\myvec{r}^\prime}\rangle}{N} e^{i \myvec{k}\dot (\myvec{r} - \myvec{r}^\prime)},
\end{equation}
with $\myvec{k}$ the reciprocal lattice vector, $n_{\myvec{r}}$ the density at position $\myvec{r}$, and $N$ the total number of particles.
While for the incompressible phases the superfluid fraction vanishes $\rho_s=0$, the supersolid phase is characterized by a finite value of $\rho_s$, indicating the presence of superfluid.
Our main results in the absence of harmonic confinement $\Omega = 0$, are summarized in Fig. \ref{FIG:PhaseDiagramD}, where we show the zero temperature phase diagram of the system, in the $J$ vs. $\mu$ plane,
in the range $J/V > 0.02$, and $1 < \mu/V < 6$ indicated by the unshaded area.
For finite $J$,
three main solid Mott lobes emerge with filling factor $\rho = 1/2$, $1/3$,
and $1/4$, named checkerboard (CB), stripe (ST), and star
(SR) solids, respectively. The corresponding groundstate
configurations are sketched in panels (b-d). We find that the CB solid is the most robust
against hopping and doping, and thus it extends furthest
in the $J$ vs. $\mu$ plane.
For large enough $J/V$, the low-energy phase is superfluid
(SF), for all $\mu$. At intermediate values of $J/V$,
however, we find that by doping the Mott solids either
with vacancies (removing particles) or interstitials
(adding extra particles) a supersolid phase (SS) can be
stabilized, with coexisting superfluid and crystalline orders.
Instead, we find no evidence of SS in the absence of doping.
The green shaded area above and below the CB lobe boundaries
in Fig. \ref{FIG:PhaseDiagramD},
correspond to a SS obtained by doping the CB crystal with interstitials, and vacancies respectively.
Remarkably, the long-range interactions stabilizes the supersolid in a
wide range of parameters, in fact for one, or two nearest neighbors in the dipolar interaction range, no stable CB
SS was found for $\rho<1/2$ \cite{BB:Sengupta}.
Interestingly, we find evidence for incompressible
phases in addition to those corresponding to the lobes
in Fig. \ref{FIG:PhaseDiagramD}. This is shown in Fig. \ref{FIG:Devil}, where the particle
density $\rho$ is plotted as a function of the chemical potential
$\mu$.
\begin{center}
\begin{figure}[h!]
\begin{center}
\includegraphics[width=0.8\linewidth]{FIG02_Submission_v2.eps}
\end{center}
\caption{$\rho$ vs. $\mu$. (a): Solids and SS for a
system with linear size $L = 12$ and $J/V = 0.05$. Some $\rho$ are
indicated. (b): SF and vacancy-SS for $L = 16$ and $J/V = 0.1$.}
\label{FIG:Devil}
\end{figure}
\end{center}
In Fig. \ref{FIG:Devil}, a continuous increase of $\rho$
as a function of $\mu$ signals a compressible phase, while a
solid phase is characterized by a constant $\rho$ for increasing
$\mu$. Panel (a), corresponding to $J/V = 0.05$, shows
a series of large constant-density plateaux connected by
a progression of smaller steps and regions of continuous
increase of $\rho$. Here, the main plateaux correspond to the
Mott lobes of Fig. \ref{FIG:PhaseDiagramD}, while the other steps correspond to
incompressible phases, with a fixed, integer, number of
particles. This progression of steps is an indication of a
Devil's-like staircase in the density, which was
discussed in \cite{BB:Cooper} for a one dimensional system.
Instead, for $J/V = 0.1$ in panel (b), no evidence of such a phase is found.
\section*{Acknowledgments}
This tutorial was supported by Spanish MEC (FIS2008-00784, QOIT),
EU projects AQUTE and NAMEQUAM, and ERC grant QUAGATUA. M.L.
acknowledges also Alexander von Humboldt Stiftung and Hamburg
Prize for Theoretical Physics. It is a great pleasure for us to
thank all the people in the quantum optics theory group of ICFO
for interesting discussions. We are especially grateful to K.
Rz{\c a}{\.z}ewski, who invited us to write this paper. C. T.
thanks C. Menotti, and M. Lewenstein for their constant support,
and patience, they showed him during these years, the result of
which is not only in this work. C. T. thanks B. Capogrosso-Sansone
and G. Pupillo, for their guide in the Quantum Monte Carlo work,
and Peter Zoller for the kind hospitality in Innsbruck. This work
was partially written at the Indian Association for the
Cultivation of Science, in Calcutta, while C. T. was visiting K.
Sengupta. The warm hospitality he showed him is unique, and C. T.
thanks him together with all the people of the theoretical physics
department.
|
\section{Introduction}
We recall that a Banach space is \emph{weakly sequentially complete} if every weakly Cauchy sequence is weakly convergent. Further, a Banach space has the \emph{Schur property} if any weakly convergent sequence is norm convergent. This is easily seen to be equivalent to the fact that any weakly Cauchy sequence is norm Cauchy (and hence norm convergent), and thus any space with the Schur property is weakly sequentially complete. A classical example of a Banach space with the Schur property is the space $\ell_1$ of all absolutely summable sequences. An example of a weakly sequentially complete Banach space without the Schur property is the Lebesgue space $L_1(0,1)$.
In the present paper we study a quantitative version of the Schur property and its relation to quantitative weak sequential completeness. A quantitative version of weak sequential completeness was studied in \cite{GoKaLi2,gode-handbook,ka-pf-sp}. The existence of a quantitative Schur property was pointed out to us by M.~Fabian who observed in 2005 that the space $\ell_1$ has, in the terminology defined below, the $5$-Schur property. The referee pointed out to us that in the paper \cite{GoKaLi1} the authors use the notion of ``$1$-strong Schur property'' which is another quantification of the Schur property. We explain the relation to our notions in the last section. Before stating the known results and our contribution we shall define the quantitative properties.
For a bounded sequence $(x_k)$ in a Banach space $X$, we write
$\clu{X}{x_k}$ for the set of all weak* cluster points of $(x_k)$ in $X^{**}$ and by $\de{x_k}$ we denote its diameter. This quantity measures in a way how far the sequence is from being weakly Cauchy. Similarly, the quantity
\[
\nc{x_k}=\inf_{n\in\en} \diam\{x_k\colon k\geq n\}
\]
measures how far the sequence is from being norm Cauchy.
Further, if $A$, $B$ are nonempty subsets of a Banach space $X$, then
\[
\dd(A,B)=\inf\{\|a-b\|\colon a\in A, b\in B\}
\]
denotes the
usual distance between $A$ and $B$ and the Hausdorff non-symmetrized distance from $A$ to $B$ is defined by
\[
\dh(A,B)= \sup\{\operatorname{d}(a,B): a\in A\}.
\]
We will say that a Banach space $X$ has the \emph{$C$-Schur property} (where $C\ge 0$) if
$$\nc{x_k}\le C \de{x_k}$$
for each bounded sequence $(x_k)$ in $X$. Similarly, $X$ is said to be \emph{$C$-weakly sequentially complete} if
$$\dh(\clu{X}{x_k},X)\le C \de{x_k}$$
for each bounded sequence $(x_k)$ in $X$. It is clear that any space with the $C$-Schur property has the Schur property and that any $C$-weakly sequentially complete space is weakly sequentially complete. It follows from Proposition~\ref{prop-easy} below that $C$-Schur property implies $C$-weak sequential completeness.
It is proved in \cite[Lemma IV.7]{GoKaLi2} that any $L$-embedded Banach space is $1$-weakly sequentially complete. We recall that a Banach space $X$ is called \emph{$L$-embedded} if there exists a projection $P:X^{**}\to X$ such that
\[
\|Px^{**}\|+\|(I-P)x^{**}\|=\|x^{**}\|,\quad x^{**}\in X^{**}.
\]
This result was mentioned in \cite[p.\ 829]{gode-handbook} together with the question which weakly sequentially complete Banach spaces satisfy a quantitative version of this property. Recently some new results were obtained in \cite{ka-pf-sp}. It is proved there that any $L$-embedded space is $\frac12$-weakly sequentially complete, where the constant $\frac12$ is optimal as witnessed by the space $\ell_1$. The quoted paper further contains an example of a~Schur space which is not $C$-weakly sequentially complete for any $C\ge0$. Conversely, $C$-weak sequential completeness does not imply Schur property (consider for example reflexive spaces or the space $L^1(0,1)$).
Inspired by the above mentioned results and a remark of M.\ Fabian, we investigate in this paper a quantification of the Schur property. We will assume that our Banach spaces are real. However, although some methods work only in real spaces, almost all the results are valid for complex spaces as well. We will discuss it in the final section.
The first proposition contains two easy inequalities and their consequence on the relationship of the quantitative Schur property and quantitative weak sequential completeness.
\begin{prop}\label{prop-easy} Let $(x_k)$ be a bounded sequence in a Banach space $X$. Then the following inequalities hold:
\begin{eqnarray}\label{ca-dh}
\dh(\clu{X}{x_k}, X)\leq \nc{x_k}, \\
\label{de-ca} \de{x_k}\leq \nc{x_k}.
\end{eqnarray}
In particular, if $X$ has the $C$-Schur property, then it is $C$-weakly sequentially complete.
\end{prop}
\begin{proof} We first observe that by the weak* lower semicontinuity of the norm we have
\[
\diam \{x_k\colon k\geq n\}=\diam \ov{\{x_k\colon k\geq n\}}^{w*},
\]
and thus
\[
\dd(x^{**},X)\leq \diam \ov{\{x_k\colon k\geq n\}}^{w*},\quad x^{**}\in \clu{X}{x_k}, n\in\en.
\]
From this we deduce that
\[
\nc{x_k}=\inf_{n\in\en} \diam\{x_k\colon k\geq n\}=\inf_{n\in\en}\diam \ov{\{x_k\colon k\geq n\}}^{w*}\geq \dh(\clu{X}{x_k},X),
\]
and
\[
\nc{x_k}\geq \diam \bigcap_{n=1}^\infty \ov{\{x_k\colon k\geq n\}}^{w*}=\de{x_k}.
\]
This proves \eqref{ca-dh} and \eqref{de-ca}. The remaining statement follows immediately from \eqref{ca-dh}.
\end{proof}
Inequalities \eqref{de-ca} and \eqref{ca-dh} are optimal. Indeed, let $X=c_0$ and $(y_k)$ be the summing basis (i.e., $y_k=(1,\dots,1,0,\dots)$ where the last `$1$' is on the $k$-th place). Denote by $(x_k)$ the sequence
$$y_1,0,y_2,0,y_3,0,\dots$$
Then clearly $\de{x_k}=\nc{x_k}=\dh(\clu{X}{x_k},X)=1$.
Inequality \eqref{de-ca} shows in particular that a nontrivial space cannot have the $C$-Schur property for any $C<1$. Another trivial consequence is that in spaces with the $C$-Schur property the quantities $\delta$ and $\operatorname{ca}$ are equivalent.
Let us remark that the situation with weak sequential completeness is different -- any $L$-embedded space is $\frac12$-weakly sequentially complete by \cite{ka-pf-sp} and reflexive spaces are trivially $0$-weakly sequentially complete. The constant $\frac12$ is optimal by the following proposition.
\begin{prop}\label{reflexive} Let $X$ be a Banach space which is $C$-weakly sequentially complete
for some $C<\frac12$. Then $X$ is reflexive.
Moreover, even a stronger version holds: Suppose that there is $C<\frac12$ such that
$$\dd(\clu{X}{x_k},X)\le C \de{x_k}$$
for each bounded sequence $(x_k)$ in $X$. Then $X$ is reflexive.
\end{prop}
\begin{proof} It is clear that first part follows from the stronger version, so
let us show the stronger statement. Let $X$ be a non-reflexive Banach space.
Let $\varepsilon>0$ be arbitrary.
It follows from \cite[Theorem 1]{GHP} that there is $x^*\in X^*$ such that any $x^{**}\in B_{X^{**}}$ such that $x^{**}(x^*)=\|x^*\|$ satisfies $\dd(x^{**},X)\ge 1-\varepsilon$.
Indeed, suppose that for any $y^*\in X^*$ there is some $x^{**}=x^{**}_{y^*}\in B_{X^{**}}$ such that $x^{**}(y^*)=\|y^*\|$ and $\dd(x^{**},X)< 1-\varepsilon$. Let $B=\{x^{**}_{y^*}: y^*\in X^*\}$.
Then $B$ is a boundary for $X^*$ and $\dh(B,X)<1$, which contradicts \cite[Theorem 1]{GHP}.
So, let $x^*$ be as in the first paragraph.
Let $(x_k)$ be a sequence in $B_X$ such that $x^*(x_k)\to \|x^*\|$. Then clearly
$\de{x_k}\le 2$. Moreover, $\dd(\clu{X}{x_k},X)\ge 1-\varepsilon$, as any
$x^{**}\in\clu{X}{x_k}$ satisfies $x^{**}(x^*)=\|x^*\|$.
This completes the proof.
\end{proof}
Let us remark, that the converse of the previous proposition is trivially valid as well, as any reflexive space is $0$-weakly sequentially complete.
We continue by our first main result which is an improvement of Schur's theorem for $\ell_1$.
\begin{thm}\label{positive} The space $\ell_1$ has the $1$-Schur property,
i.e.,
\[
\nc{x_k}= \de{x_k}.
\]
for any bounded sequence $(x_k)$ in $\ell_1$.
\end{thm}
It follows from \cite{ka-pf-sp} that there is a Banach space with the Schur property which fails its quantitative version. Indeed, it is constructed there a space with the Schur property which is not $C$-weakly sequentially complete for any $C\ge 0$. It follows from Proposition~\ref{prop-easy} that the same space fails the $C$-Schur property for each $C\ge0$.
Our second result shows that a quantitative version of the Schur property may fail even for an $L$-embedded space (which is $\frac12$-weakly sequentially complete).
\begin{example}\label{el-ex}
There exists a separable $L$-embedded Banach space with the Schur property
which fails the $C$-Schur property for every $C\ge0$.
\end{example}
Following the suggestion of the referee, we include the following diagramm of implications between the classes of Banach spaces we study in this paper (``wsc'' is an abbreviation of ``weakly sequentially complete'').
$$\begin{array}{ccccccc}
&&&&\mbox{Schur}&&\\
&&&\neimp&&\seimp&\\
C\mbox{-Schur}& \Rightarrow & C\mbox{-wsc and Schur} & \Rightarrow & C\mbox{-wsc} & \Rightarrow & \mbox{wsc}
\end{array}$$
All these implications are easy to check. If a constant $C$ is included both in the assumption and in the conclusion, the constant in the conclusion has the same value.
None of the displayed implications can be reversed, even if we allow the respective constants to change.
For the first implication from the left, it follows from Example~\ref{el-ex}. A reasoning for the implication $\neimp$ and the third one in the bottom line is a result of \cite{ka-pf-sp}. For the remaining two implications one can use the classical space $L^1(0,1)$.
\section{Proof of Theorem~\ref{positive}}
We start the proof with a simple fact based upon the standard ``sliding hump" argument. If $x\in\ell_1$ and $N\subset \en$, we denote by $x\r_N$ the element arising from $x$ by setting its coordinates outside $N$ to zero, i.e.,
$$x\r_N(n)=\begin{cases} x(n), & n\in N, \\ 0, & n\in\en\setminus N.\end{cases}$$
\begin{lemma}\label{hump}
Let $(y_n)$ be a sequence in $\ell_1$ converging pointwise to $y$ with $\|y\|<\ep$. Then there exists a subsequence $(y_{n_k})$ and indices $0=N_0<N_1<N_2<\cdots$ such that
\[
\|y_{n_k}\r_{(N_{k-1}, N_k]}\|>\|y_{n_{k}}\|-\ep,\quad k\in\en.
\]
\end{lemma}
\begin{proof}
First we set $N_0=0$, $n_1=1$ and choose $N_1>0$ such that $\|y_{n_1}\r_{(0, N_1]}\|>\|y_{n_{1}}\|-\ep$.
Suppose now that $k\in\en$ is such that we have already constructed $n_j$ for $1\le j\le k$ and $N_j$ for $0\le j\le k$.
Since $(y_n)$ converges pointwise to $y$, we can select $n_{k+1}>n_k$ so large that $\|y_n\r_{(N_0, N_k]}\|<\ep$ for every $n\geq n_{k+1}$. Thus, in particular
\[
\|y_{n_{k+1}}\r_{(N_k,+\infty)}\|>\|y_{n_{k+1}}\|-\ep,
\]
so we can find $N_{k+1}>N_{k}$ such that
$$\|y_{n_{k+1}}\r_{(N_{k},N_{k+1}]}\|>\|y_{n_{k+1}}\|-\ep.$$
This completes the construction.\end{proof}
Now we proceed with the proof of Theorem~\ref{positive}. Let $(x_k)$ be a bounded sequence in $X=\ell_1$ and $\ep>0$. We consider an arbitrary $c<\nc{x_k}$.
We extract subsequences $(a_n)$ and $(b_n)$ from $(x_k)$ such that $c<\|a_n-b_n\|$ for $n\in\en$. Denoting $y_n=a_n-b_n$, we pass to a subsequence if necessary and assume thus that $(y_n)$ pointwise converges to $y\in\ell_1$. Let
$m\in\en$ be chosen such that $\|y\r_{(m,+\infty)}\|<\ep$. By omitting finitely many elements of $(y_n)$ we achieve that
\begin{equation}\label{eNko}
\|(y_n-y)\r_{[1,m]}\|<\ep,\quad n\in\en.
\end{equation}
Hence
\begin{equation}
\label{cN}
c<\|y_n\r_{[1,m]}\|+\|y_n\r_{(m,+\infty)}\|\leq \|y\r_{[1,m]}\|+\ep+\|y_n\r_{(m,+\infty)}\|,\quad n\in\en.
\end{equation}
Using Lemma~\ref{hump} applied to $(y_n\r_{(m,+\infty)})$ we obtain a subsequence $(y_{n_k})$ and indices $m=N_0<N_1<\cdots$ such that
\begin{equation}
\label{hrb}
\|y_{n_k}\r_{(N_{k-1}, N_k]}\|>\|y_{n_{k}}\r_{(m,+\infty)}\|-\ep,
\quad k\in\en.
\end{equation}
Let $x^*\in\ell_\infty=\ell_1^*$ be defined as
\[
x^{*}(j)=\begin{cases} \operatorname{sign} y(j),& j\in[1,m],\\
\operatorname{sign} y_{n_k}(j),& j\in(N_{k-1}, N_k],\ k\in\en.
\end{cases}
\]
Then $\|x^*\|\le 1$, and, for each $k\in\en$, we have
using~\eqref{eNko}, \eqref{hrb} and \eqref{cN}
\[
\aligned
x^*(y_{n_k})&=x^*(y_{n_k}\r_{[1,m]})+x^*(y_{n_k}\r_{(N_{k-1}, N_k]})+
\sum_{j\in(m,+\infty)\setminus(N_{k-1},N_k]} x^*(j)y_{n_k}(j)\\
&> x^*(y\r_{[1,m]})-\ep+\|y_{n_{k}}\r_{(N_{k-1},N_k]}\|-\|y_{n_k}\r_{(m,+\infty)\setminus(N_{k-1},N_k]}\|\\
&> \|y\r_{[1,m]}\|+\|y_{n_k}\r_{(m,+\infty)}\|-3\ep\\
&> c-4\ep.
\endaligned
\]
Therefore we have
\[
x^*(a_{n_k}-b_{n_k})\geq c-4\ep.
\]
Up to a passing to a subsequence we can suppose that the sequence $(x^*(a_{n_k}))$ converges. Up to passing to a further subsequence we can assume that the sequence $(x^*(b_{n_k}))$ converges as well.
Let $a^{**}$ and $b^{**}$ be weak* cluster points of $(a_{n_k})$ and $(b_{n_k})$, respectively. Then $a^{**}, b^{**}\in\clu{X}{x_k}$ and
\[
\|a^{**}-b^{**}\|\geq (a^{**}-b^{**})(x^*)=\lim_{k\to\infty} x^*(a_{n_k}-b_{n_k}) \geq c-4\ep.
\]
Since $c$ and $\ep$ are arbitrary, $\nc{x_k}\leq \de{x_k}$.
Inequality \eqref{de-ca} from Proposition~\ref{prop-easy} finishes the proof.
\section{Construction of Example~\ref{el-ex}}
The construction is based upon Example~4 of \cite{ka-pf-sp}. It uses a standard renorming technique
(see, e.g. \cite[Proposition III.2.11]{hawewe}).
We recall that $\beta\en$ is the \v{C}ech--Stone compactification of $\en$ and $M(\beta\en)$ is the space of all signed Radon measures on $\beta\en$ considered as the dual of $\ell_\infty$.
Let us fix $\alpha>0$ and consider the space
\[
Y_\alpha=(\ell_1,\alpha\|\cdot\|_1)\oplus_1 (\ell_2,\|\cdot\|_2).
\]
Here $\|\cdot\|_1$ and $\|\cdot\|_2$ denote the usual norms on $\ell_1$ and $\ell_2$, respectively.
Note that we have the following canonical identifications:
$$\begin{aligned}
Y_\alpha^*&=(\ell_\infty,\tfrac1\alpha\|\cdot\|_\infty)\oplus_\infty (\ell_2,\|\cdot\|_2), \text{ and}\\
Y_\alpha^{**}&=(M(\beta\en),\alpha\|\cdot\|_{M(\beta\en)})\oplus_1 (\ell_2,\|\cdot\|_2).
\end{aligned}$$
For $k\in\en$, let $x_k=(e_k,e_k)\in Y_\alpha$, where $e_k$ denotes the $k$-th canonical basic vector.
Let $X_\alpha$ be the closed linear span of the set $\{x_k:k\in\en\}$.
We claim that
\begin{equation}\label{Xalpha}
X_\alpha=\left\{(x,y)\in Y_\alpha:
x(n)=y(n), n\in\en\right\}.
\end{equation}
Indeed, the set on the right-hand side is a closed linear subspace of $Y_\alpha$ containing $x_k$ for each $k\in\en$ which verifies the inclusion `$\subset$'. To prove the converse one, let us take any element $(z,y)\in Y_\alpha$ satisfying $z(n)=y(n)$ for all $n\in\en$. Since $z\in\ell_1$, we get
\[
(z,y)=\sum_{k=1}^\infty z(k)x_k\in X_\alpha
\]
as the series is absolutely convergent.
Let $T:\ell_1\to\ell_2$ denote the identity mapping. Since $T$ maps the unit ball of $\ell_1$ into the unit ball of $\ell_2$, we get $\|T\|\le 1$.
Therefore, for an arbitrary element $(x,y)\in X_\alpha$, we have
$$\alpha\|x\|_1 \le \|(x,y)\|_{X_\alpha}=\alpha\|x\|_1+\|Tx\|_2\le (\alpha+1)\|x\|_1.$$
Thus the projection on the first coordinate is an isomorphism of $X_\alpha$ onto $\ell_1$. In particular, $X_\alpha$ has the Schur property.
We further observe that $X_\alpha^{**}$ is canonically identified with the weak* closure of $X_\alpha$ in $Y_\alpha^{**}$, thus
\begin{equation}
\label{Xalpha**} X_\alpha^{**}=\{(\mu,y)\in M(\beta\en)\times \ell_2:
\mu(n)=y(n), n\in\en\}.
\end{equation}
Indeed, the set on the right-hand side is a weak* closed linear subspace of $Y_\alpha^{**}$ containing $X_\alpha$, which proves the inclusion `$\subset$'.
To prove the converse one, we fix $(\mu,y)$ in the set on the right-hand side.
Take a bounded net $(u_\tau)$ in $\ell_1$ which weak* converges to $\mu$. For each $\tau$ there is a unique $y_\tau\in \ell_2$ such that $(u_\tau,y_\tau)\in X_\alpha$. Then $(y_\tau)$ is clearly a bounded net in $\ell_2$. Moreover, we will show that $(y_\tau)$ weak* (i.e. weakly) converges to $y$. Since the weak topology on bounded sets in $\ell_2$ coincides with the topology of pointwise convergence, it suffices to show that $y_\tau$ pointwise converge to $y$. Indeed,
\[
y_\tau(n)=\mu_\tau(n) \to \mu(n)=y(n),\quad n\in\en.
\]
It follows that $X_\alpha$ is $L$-embedded in $X_\alpha^{**}$ because the projection $P:X_\alpha^{**}\to X_\alpha$ defined as
\[
P(\mu,y)=(\mu\r_\en,y),\quad (\mu,y)\in M(\beta\en)\times \ell_2,
\]
satisfies
\[
\|(I-P)(\mu,y)\|_{X_\alpha^{**}}+\|P(\mu,y)\|_{X_\alpha^{**}}=\|(\mu,y)\|_{X_\alpha^{**}},\quad (\mu,y)\in X_\alpha^{**}.
\]
Further, for the sequence $(x_k)$, its weak$^*$ cluster points in
$X_\alpha^{**}$
are equal to
\[
\{(\ep_t, 0): t\in \beta\en\setminus\en\},
\]
where $\ep_t$ denotes the Dirac measure at a point $t\in\beta\en$.
We claim that, for our sequence $(x_k)$, we have
\begin{equation}
\label{est-alpha}
\nc{x_k}=2\alpha+\sqrt{2}\quad\text{and}\quad \de{x_k}=2\alpha.
\end{equation}
To see the first inequality, we observe that for each distinct $k,k'\in\en$ we have
\[
\|x_k-x_{k'}\|_{X_\alpha}=\alpha\|e_k-e_{k'}\|_1+\|e_k-e_{k'}\|_{2}=2\alpha+\sqrt{2}.
\]
The second one follows from the fact that, given $t,t'\in\beta\en\setminus\en$ distinct, then
\[
\|(\ep_t,0)-(\ep_{t'},0)\|_{X_\alpha^{**}}=\|(\ep_t-\ep_{t'},0)\|_{X_\alpha^{**}}
=\alpha\|\ep_t-\ep_{t'}\|_{M(\beta\en)}=2\alpha.
\]
This verifies \eqref{est-alpha}.
Now we use the described procedure to construct the desired space $X$.
For $n\in\en$, let $\alpha_n=\frac1n$ and let $X_{\frac1n}$ be the space constructed for $\alpha_n$. Let
\[
X=\left(\sum_{n=1}^\infty X_{\frac1n}\right)_{\ell_1}
\]
be the $\ell_1$-sum of the spaces $X_{\frac1n}$. We claim that $X$ is the required space.
First, since each $X_{\frac1n}$ has the Schur property, $X$, as their $\ell_1$-sum, possesses this property as well (this follows
by a straightforward modification of the proof that $\ell_1$ has the Schur property, see \cite[Theorem~5.19]{fhhmpz}).
Similarly, as an $\ell_1$-sum of $L$-embedded spaces, $X$ is $L$-embedded as well (see \cite[Proposition~1.5]{hawewe}).
Finally, fix $n\in\en$. We consider a sequence $\wh{x}_k=(0,\dots,0,\stackrel{n\text{-th}}{x_k},0,\dots)$,
where the elements $x_k\in X_{\frac1n}$, $k\in\en$, are defined above.
Then, for any $k,k'\in\en$ distinct,
\[
\|\wh{x}_k-\wh{x}_{k'}\|_X= \|x_k-x_{k'}\|_{X_{\frac1n}},
\]
hence $\nc{\wh{x}_k}=\nc{x_k}>\sqrt2$ by \eqref{est-alpha}.
On the other hand,
\[
\de{\wh{x}_k}= \de{x_k}=\frac2n,
\]
again by \eqref{est-alpha}. So,
\[
\nc{\wh{x}_k}>\frac n{\sqrt2}\;\de{\wh{x}_k}.
\]
Since $n\in\en$ is arbitrary, the conclusion follows.
\section{Final remarks}
In \cite[p. 57]{GoKaLi1} a Banach space is said to have the {\it $1$-strong Schur property}, whenever for any $\delta\in(0,2]$, any $\varepsilon>0$ and any normalized $\delta$-separated sequence in $X$ there is a subsequence which is $(\frac2\delta+\varepsilon)$-equivalent to the standard $\ell_1$ basis.
Although this property has adjective ``strong'', it is in fact weaker than our $1$-Schur property, as witnessed by the following proposition. We do not know whether these two properties are in fact equivalent but we conjecture that it is not the case. The reason for this opinion is the fact that the $1$-Schur property is a property of all bounded sequences while the $1$-strong Schur property is a property of very special sequences.
\begin{prop}\label{posledni}
\begin{itemize}
\item Any Banach space with the $1$-Schur property has the $1$-strong Schur property.
\item Any Banach space with the $1$-strong Schur property has the $5$-Schur property.
\end{itemize}
\end{prop}
\begin{proof} Suppose that $X$ is a Banach space with the $1$-Schur property and fix
$\delta\in(0,2]$, $\varepsilon>0$ and a normalized $\delta$-separated sequence $(x_k)$ in $X$.
Since $(x_k)$ is $\delta$-separated, $\nc{x_{k_n}}\ge\delta$ for each subsequence $(x_{k_n})$ of $(x_k)$.
By the $1$-Schur property we get $\de{x_{k_n}}\ge\delta$ for each subsequence. By \cite[Theorem 3.2]{behrends} there is a subsequence $(x_{k_n})$ such that
$$\left\|\sum_{n=1}^N \alpha_n x_{k_n}\right\|\ge \frac1{\frac2\delta+\varepsilon} \sum_{n=1}^N |\alpha_n|$$
for each $N\in\en$ and each choice of real numbers $\alpha_1,\dots,\alpha_N$. Note, that we used that
$$\frac1{\frac2\delta+\varepsilon}<\frac\delta2.$$
It follows that $(x_{k_n})$ is $(\frac2\delta+\varepsilon)$-equivalent to the standard $\ell_1$ basis.
Conversely, suppose that $X$ is a Banach space which has the $1$-strong Schur property. Let $(x_k)$ be a bounded sequence in $X$ and $c>0$ be such that $\nc{x_k}>5c$. We will show that $\de{x_k}\ge c$.
We will distinguish two cases:
\begin{itemize}
\item[(a)] There is $n\in\en$ such that for each $k\in\en$ we have $\dd(x_k,\{x_1,\dots,x_n\})\le c$.
\item[(b)] For each $n\in\en$ there is $k>n$ such that $\dd(x_k,\{x_1,\dots,x_n\})> c$.
\end{itemize}
It is clear that exactly one of these two cases takes place.
First suppose that the case (a) occurs. As $\nc{x_k}>5c$, there are two subsequences $(u_k)$ and $(v_k)$ of $(x_k)$ such that $\|u_k-v_k\|>5c$ for each $k\in\en$. For each $k\in \en$ fix $p_k,q_k\in \{1,\dots,n\}$ such that $\|u_k-x_{p_k}\|\le c$ and $\|v_k-x_{q_k}\|\le c$. Up to taking a subsequence we can assume that the sequence $(p_k)$ is constant. Denote the common value by $p$. Up to taking a further subsequence we can suppose that the sequence $(q_k)$ is constant as well. Let $q$ be the common value. Set $a=x_p$ and $b=x_q$. Let $u^{**}$ be a weak* cluster point of $(u_k)$ in $X^{**}$.
Similarly, let $v^{**}$ be a weak* cluster point of $(v_k)$. Then
$$\begin{aligned}
\|v^{**}-u^{**}\|&\ge\|b-a\|-\|v^{**}-b\|-\|u^{**}-a\|
\ge\|b-a\|-2c\\&\ge\|v_1-u_1\|-\|b-v_1\|-\|a-u_1\|-2c>5c-4c=c.
\end{aligned}$$
Hence $\de{x_k}>c$.
Next suppose that the case (b) occurs and fix an arbitrary $\varepsilon>0$. It is easy to construct a $c$-separated subsequence $(y_k)$ of $(x_k)$. As $(y_k)$ is bounded, without loss of generality we can suppose that the sequence $(\|y_k\|)$ converges to some $\alpha\ge 0$. As the sequence is $c$-separated, necessarily $\alpha\ge\frac c2$. Up to omitting a finite number of $(y_k)$ we can suppose that
$|\|y_k\|-\alpha|<\varepsilon\alpha$ for each $k\in\en$. Then the sequence $(\frac{y_k}{\|y_k\|})$ is a normalized sequence which is $\frac c{\alpha}-2\varepsilon$-separated. Indeed, for any distinct $k,l\in\en$ we have
$$\begin{aligned}\left\|\frac{y_k}{\|y_k\|}-\frac{y_l}{\|y_l\|}\right\| &\ge
\left\|\frac{y_k}{\alpha}-\frac{y_l}{\alpha}\right\| -
\left\|\frac{y_k}{\|y_k\|}-\frac{y_k}{\alpha}\right\|-
\left\|\frac{y_l}{\|y_l\|}-\frac{y_l}{\alpha}\right\| \\ &
>\frac c\alpha - \left|\frac{\alpha-\|y_k\|}{\alpha}\right| -
\left|\frac{\alpha-\|y_l\|}{\alpha}\right|>\frac c\alpha-2\varepsilon.\end{aligned}$$
By the $1$-strong Schur property we can extract a subsequence of $\left(\frac{y_k}{\|y_k\|}\right)$
which is $\left(\frac{2}{\frac c\alpha-2\varepsilon}+\varepsilon\right)$-equivalent to the standard $\ell_1$-basis.
Thus
$$\de{\frac{y_k}{\|y_k\|}}\ge \frac{2}{\frac{2}{\frac c\alpha-2\varepsilon}+\varepsilon}$$
by \cite[Lemma 5]{ka-pf-sp}.
Since $\varepsilon>0$ is arbitrary, we get $\de{\frac{y_k}{\|y_k\|}}\ge \frac c\alpha$. It follows
that $\de{y_k}\ge c$, thus $\de{x_k}\ge c$. This completes the proof.
\end{proof}
As we remarked in the beginning of the paper, we worked only with real spaces. However, all the results except possibly for Proposition~\ref{posledni} are true also for complex spaces. Let us explain it.
Let $X$ be a complex Banach space. Denote by $X_R$ the real version of $X$, i.e., the same space considered over the reals. Given a bounded sequence $(x_k)$ in $X$, all the considered quantities ($\dd$, $\dh$, $\delta$, $\operatorname{ca}$) are the same for $X$ and for $X_R$. For the quantities $\dd$, $\dh$ and $\delta$ it is explained in \cite[Section 5]{ka-pf-sp}, for the quantity $\operatorname{ca}$ it is trivial. Hence, in particular, $X$ is $C$-weakly sequentially complete (or has the $C$-Schur property) if and only if $X_R$ has this property. Thus Propositions~\ref{prop-easy} and~\ref{reflexive} are valid for complex spaces as well.
The proof of Theorem~\ref{positive} can be easily adapted to the complex case. It is enough to define $x^*$ using complex signs (i.e., appropriate complex units) and to estimate from below the real part of $x^*(y_{n_k})$. The proof of Example~\ref{el-ex} works in the complex case without any change.
As for Proposition~\ref{posledni}, the second part works for complex spaces without any change. It is not clear whether the first part is valid because the proof uses a result of \cite{behrends} which works for real spaces and it is an open problem whether it is valid for complex spaces as well. It is also not clear, whether the $1$-strong Schur property of $X$ is equivalent to that of $X_R$.
\section*{Acknowledgement}
We are grateful to the referee for helpful comments which led to an improvement of our paper.
|
\section{Introduction}
\label{introduction}
Perturbative quantum gravity has been well-studied over many decades,
in the form of general relativity minimally coupled to matter particles
(see~\cite{Hamber:2007fk} for a recent review), as well as alternative
theories.
Although the pitfalls of such an approach are well-known -- chiefly that
quantum general relativity contains non-renormalizable ultraviolet
divergences -- it is nevertheless useful for a number of reasons.
Firstly, one may treat perturbative quantum gravity as an effective theory,
and calculate loop corrections to gravitational observables. Secondly, it
may well be the case that if a consistent quantum theory of gravity is found,
it shares features with quantized general relativity. This motivates the study
of features in this theory which, it may be argued, should be generically
shared by alternatives. A third reason for studying gravity is that there are
intriguing connections between the structure of perturbative scattering
amplitudes in gravity and non-abelian
gauge theories (e.g.~\cite{Hohm:2011dz,Bern:2010yg,Bern:2010ue,Bern:2002kj,
Bern:1999ji,ArkaniHamed:2008yf}). A possible
application of these ideas is, for example, that of settling the question of
whether $N=8$ supergravity is ultraviolet finite~\cite{Bern:1998ug,
Bern:2006kd,Bern:2007hh,Bern:2008pv,Bern:2009kd}.
In light of the above remarks (and as stressed recently
in~\cite{Naculich:2011ry}), an interesting property of perturbative
gravity is the structure of infrared singularities. One might hope, for
example, that the infrared behaviour of alternative (possibly UV finite)
quantum gravity theories shares at least some of the features that underly
quantum GR coupled to matter. Furthermore, the structure of IR singularities
in renormalizable gauge theories is well-understood. IR divergences
cancel between real and virtual contributions for sufficiently inclusive
observables~\cite{Bloch:1937pw}. The singularity structure nevertheless
becomes relevant in that it affects residual large
contributions to the perturbative expansion of physical cross-sections which
must be resummed to all orders in the coupling constant in order to obtain
sensible physical predictions. Resummation is by now a highly developed
subject in both abelian~\cite{Yennie:1961ad}
and non-abelian gauge theories. A number of approaches exist, such as
Feynman diagram methods~\cite{Sterman:1986aj,Catani:1989ne},
Wilson line techniques~\cite{Korchemsky:1992xv,Korchemsky:1993uz},
effective field theories~\cite{Beneke:2002ph,Bauer:2000yr,Bauer:2002nz,
Becher:2006nr,Becher:2006mr,Becher:2007ty},
and the recently developed path integral approach of~\cite{Laenen:2008gt}.
Crucial to all these approaches is the notion of {\it factorization},
namely that scattering amplitudes separate into a {\it hard interaction},
which is infrared finite, and a {\it soft function},
which contains all infrared
singularities. These are generated by the emission of soft (zero momentum)
gauge bosons between the external lines of the amplitude, referred to in
the literature as the {\it eikonal approximation}\footnote{When amplitudes
involve massless external particles, hard collinear singularities must
also be taken into account in so-called
{\it jet functions}~\cite{Mueller:1979ih,Collins:1980ih,Sen:1981sd,
Korchemsky:1988pn,Magnea:1990zb}. These are
not usually considered in gravity, as collinear
singularities cancel after summing over all emitting
particles~\cite{Weinberg:1965nx}.}. Such factorization properties are
important both for resummation applications, and also for understanding the
structure of infrared singularities to all
orders~\cite{Dixon:2008gr,Gardi:2009qi,Dixon:2009gx,Becher:2009cu,
Becher:2009qa,Becher:2009kw,Dixon:2009ur,Dixon:2010zz,Gardi:2009zv}.
The same physics also governs the infrared properties of perturbative
gravity amplitudes, as first analysed in~\cite{Weinberg:1965nx}. Recently,
Naculich and Schnitzer have reconsidered the structure of gravitational
IR divergences~\cite{Naculich:2011ry}. The main point of their paper is to
argue that infrared singularities at all orders are generated by the
exponentiation of the one-loop divergence i.e. that there are no subleading
divergences. Their argument is based upon the assumption that gravitational
amplitudes factorize in the same way that gauge theory amplitudes do, in
terms of hard and soft functions. That is, the $n$-graviton
scattering amplitude (adopting the notation of~\cite{Naculich:2011ry})
may be written
\begin{equation}
A_n=S_n\cdot H_n,
\label{Andef}
\end{equation}
where $H_n$ is IR-finite, and $S_n$ collects all IR singularities, generated
by soft graviton exchange. They make the further hypothesis that, by analogy
with gauge theory, the soft function can be expressed as a vacuum expectation
value:
\begin{equation}
S_n=\left\langle0\left|\prod_{i=1}^n\Phi_{p_i}(0,\infty)\right|0
\right\rangle,
\label{Snwilson}
\end{equation}
containing the Wilson line operators\footnote{These operators describe
the emission of soft gravitons from hard emitting particles, and should not
be confused with the parallel transport operator of general relativity,
defined in terms of the Christoffel symbol, which is also sometimes referred
to in Wilson line terms.}
\begin{equation}
\Phi_p(a,b)={\cal P}\exp\left(i\kappa\int_a^b ds\,p_{i\mu} p_{i\nu}
h^{\mu\nu}(sp)\right),
\label{wilsondef}
\end{equation}
where there is one such operator associated with each external line of the
amplitude (with momentum $p_i$). Here $h^{\mu\nu}$ is the graviton field
(we will define this more carefully in what follows), and
$\kappa=32\pi G$, where $G$ is Newton's constant. The line integral
in eq.~(\ref{wilsondef}) is over a straight line contour, parametrized by
$s$.
The aim of this paper is to examine these hypotheses within the path integral
resummation framework of~\cite{Laenen:2008gt}, which was developed in the
context of abelian and non-abelian gauge theory. The main purpose
of that paper was to address corrections to the eikonal approximation, and
the result was a systematic classification of the next-to-eikonal contributions
to scattering amplitudes -- that is, those which occur at first
subleading order in an expansion of the amplitude in the momenta of
emitted gauge bosons.
The results were subsequently confirmed using Feynman diagrammatic methods
in~\cite{Laenen:2010uz}, and have also been used to classify the structure
of (non-abelian) soft gluon corrections in multiparton processes,
generalising the concept of webs from two parton
scattering~\cite{Gatheral:1983cz,Frenkel:1984pz,Sterman:1981jc} to the
multiparton case~\cite{Gardi:2010rn,Gardi:2011wa}\footnote{See
also~\cite{Mitov:2010rp} for an alternative viewpoint on multiparton webs.
Other work on next-to-eikonal corrections can be found
in~\cite{Laenen:2008ux,Moch:2009hr,Soar:2009yh,Vogt:2010pe,Vogt:2010ik,
Grunberg:2009yi,Grunberg:2009vs,Grunberg:2010sw}.}. The results
of~\cite{Laenen:2008gt} show that the structure of a scattering
amplitude ${\cal A}$ subject to soft photon emissions has the schematic form
\begin{equation}
{\cal A}={\cal A}_0\exp\left[\sum_{G^{\E}}G^{\E}+\sum_{G^{\NE}}G^{\NE}
\right]\left(1+
{\cal A}_{rem.}\right).
\label{ampstruc}
\end{equation}
Here ${\cal A}_0$ is the hard interaction amplitude undressed by soft
photons, and the exponent contains connected subdiagrams (which span
the external lines) $G^{\E}$ and $G^{\NE}$ at eikonal and next-to-eikonal
order respectively. A set of effective Feynman rules for forming these
diagrams has been given in~\cite{Laenen:2008gt,Laenen:2010uz}, which
generalize the well-known eikonal Feynman rules to subleading order in
the momentum expansion. The term ${\cal A}_{rem.}$ in eq.~(\ref{ampstruc})
is also next-to-eikonal order, but does not formally
exponentiate. It consists of diagrams in which eikonal photons are
emitted from an external line and land inside the hard interaction. We
may call these {\it internal emission} diagrams to distinguish them
from the {\it external emission} diagrams which enter the exponent, and
their origin makes clear that they can be thought of directly as a breaking
of the factorization of the amplitude into hard and soft parts. Although
internal emission contributions do not formally exponentiate, they do have
an iterative structure to all orders in perturbation theory, which is
fixed by gauge invariance by a result known as the {\it Low-Burnett-Kroll}
theorem~\cite{Low:1958sn,Burnett:1967km} (see also~\cite{DelDuca:1990gz}
for a further generalization). Examples of internal and
external emission contributions are shown in figure~\ref{intext}.
\begin{figure}
\begin{center}
\scalebox{1.0}{\includegraphics{intext.eps}}
\caption{Examples of (a) a generic hard interaction with outgoing particles,
which may emit soft gauge bosons; (b) an external emission contribution;
(c) an internal emission contribution.}
\label{intext}
\end{center}
\end{figure}
Although the case of abelian gauge theory is described in the preceding
paragraph, a very similar structure occurs in non-Abelian theories
(essentially, the sum over connected diagrams in eq.~(\ref{ampstruc})
is replaced by a sum over so-called {\it webs}, as described
in~\cite{Gatheral:1983cz,Frenkel:1984pz,Sterman:1981jc,
Mitov:2010rp,Gardi:2010rn,Gardi:2011wa}).
We will see in this paper that the same general
structure is also observed in perturbative gravity, assuming that the
infrared properties can be analysed independently of the uncertain UV
completion of the theory. In particular, we will derive directly the
appropriate form of the Wilson line operator which describes soft
graviton emissions, as well as the appropriate gravitational generalization of
Low's theorem, which shows that internal emission contributions in gravity
are subleading with respect to the eikonal approximation, as is also the
case in gauge theory. These results strengthen the hypotheses made
in~\cite{Naculich:2011ry}, and may also shed more light on
the correspondence between gauge theory and gravity.
There are also phenomenological motivations for studying corrections to
eikonal gravity. The latter has been used in a variety of recent applications,
in particular the study of transplanckian scattering in different
gravitational theories. This has a number of potential uses, such as
investigating whether different gravity theories have the same long distance
behaviour~\cite{Giddings:2010pp}, or even what potential collider signatures
are in extra dimension scenarios~\cite{Stirling:2011mf}. The classification
of corrections to the eikonal approximation, as examined in this paper,
may allow further study of these and related subjects.
The structure of the paper is as follows. In section~\ref{sec:path} we
describe the path integral framework for scalar particles emitting soft
gravitons, adapting the approach used for abelian gauge theory
in~\cite{Laenen:2008gt}, and derive the form of the Wilson line operator
for soft gravitons. In section~\ref{sec:low} we discuss how gauge
invariance can be used to constrain factorization-breaking terms, and
discuss how this relates to Low's theorem~\cite{Low:1958sn} in QED.
In section~\ref{sec:discuss} we discuss our results before concluding.
\section{Path integral approach to soft graviton amplitudes}
\label{sec:path}
In this section, we will consider the scattering amplitude for $L$ scalar
particles, dressed by any number of soft gravitons. Although pure
multigraviton amplitudes were considered in~\cite{Naculich:2011ry},
scalar particles will be sufficient for our purposes,
due to the fact that eikonal gravitons are insensitive to the spin
of the particle which emits them~\cite{Weinberg:1965nx}.
The action for general relativity minimally coupled to a (complex) scalar field
is given by
\begin{equation}
S_{\rm grav}=S_{\rm E.H.}[g^{\mu\nu}]+S_{\rm mat}[\phi^*,\phi,g^{\mu\nu}],
\label{Stot}
\end{equation}
here $S_{\rm E.H.}$ is the Einstein-Hilbert action (which must be suitably
gauge-fixed in order to define the graviton propagator), and\footnote{Note
that we use the metric (-,+,+,+), as in~\cite{Laenen:2008gt}.}
\begin{equation}
S_{\rm mat}[\phi^*,\phi,g^{\mu\nu}]=\int d^d x\sqrt{-g}\left[-g^{\mu\nu}
\partial_\mu\phi^*\partial_\nu\phi-m^2\phi^*\phi\right],
\label{Smatdef}
\end{equation}
in $d$ dimensions, where $g$ is the determinant of the metric tensor
$g^{\mu\nu}$. Perturbation theory can then be defined after expanding the
metric tensor about the flat space Minkowski metric $\eta^{\mu\nu}$. As
explained in e.g.~\cite{Hamber:2007fk}, there is an ambiguity in how one
performs the weak field expansion. Here we adopt the approach
of~\cite{Capper:1973bk}, and define
\begin{equation}
\tilde{g}^{\mu\nu}=\sqrt{-g}g^{\mu\nu}.
\label{gtildedef}
\end{equation}
This introduces a Jacobian in principle in eq.~(\ref{Smatdef}), although
this may be taken to be one in dimensional
regularization~\cite{Hamber:2007fk}\footnote{Many other field
redefinitions are possible instead of that of eq.~(\ref{gtildedef}).
See~\cite{Hohm:2011dz} for a recent proposal in the context of exploring
the factorization of gravity amplitudes into gauge theory amplitudes.}.
We then define the graviton field $h^{\mu\nu}$ via
\begin{equation}
\tilde{g}^{\mu\nu}=\eta^{\mu\nu}+\kappa h^{\mu\nu},
\label{hdef}
\end{equation}
where $\kappa^2=32\pi G$ as in~\cite{Naculich:2011ry}. Note that symmetry
of the metric tensor implies $h^{\mu\nu}=h^{\nu\mu}$. One could also have
chosen, of course, to expand the metric $g^{\mu\nu}$ directly. However,
the choice of eq.~(\ref{hdef}) results in simpler expressions for the
scalar-graviton vertices. We will return to this
point later on when discussing next-to-eikonal corrections.
With the above definition for $h^{\mu\nu}$ one has~\cite{Capper:1973bk}
\begin{equation}
\sqrt{-g}=1+\frac{\kappa}{d-2}h^{\alpha}_{\alpha}+\kappa^2\left[\frac{(
h^{\alpha}_{\alpha})^2}{(d-2)^2}-\frac{h^{\alpha\beta}h_{\alpha\beta}}
{d-2}\right]+{\cal O}(\kappa^3).
\label{gexpand}
\end{equation}
Furthermore, the inverse of $\tilde{g}^{\mu\nu}$ is given
by~\cite{Capper:1973pv}
\begin{equation}
\tilde{g}_{\mu\nu}=\eta_{\mu\nu}-\kappa h_{\mu\nu}+\kappa^2
h_\mu^{\phantom{\nu}\alpha}h_{\alpha\nu}+{\cal O}(\kappa^2).
\label{gexpand2}
\end{equation}
The action of eq.~(\ref{Smatdef}), evaluated to quadratic order in $\kappa$,
is thus
\begin{align}
S_{\rm mat}[\phi^*,\phi,h^{\mu\nu}]&=\int d^d x\left\{-\eta^{\mu\nu}
\partial_\mu\phi^*\partial_\nu\phi-\kappa h^{\mu\nu}\partial_\mu\phi^*
\partial_\nu\phi\right.\notag\\
&\left.\quad-m^2\phi^*\phi-\frac{m^2\kappa}{d-2}\left[h^\alpha_\alpha+
\kappa\left(\frac{(h^\alpha_\alpha)^2}{(d-2)}-h^{\alpha\beta}
h_{\beta\alpha}\right)\right]\phi^*\phi\right\}.
\label{Smatdef2}
\end{align}
For what follows it is useful to write this as
\begin{equation}
S=-\int d^dx\phi^*\hat{S}\phi,
\label{Shat}
\end{equation}
where the quadratic operator $\hat{S}$ is given by
\begin{align}
\hat{S}&=-\eta^{\mu\nu}\partial_\mu\partial_\nu-\kappa(\partial_\mu h^{\mu\nu})
\partial_\nu-\kappa h^{\mu\nu}\partial_\mu\partial_\nu+m^2\notag\\
&\quad+\frac{m^2\kappa}
{d-2}\left[h^\alpha_\alpha+\kappa\left(\frac{
(h^\alpha_\alpha)^2}{(d-2)}-h^{\alpha\beta}h_{\beta\alpha}\right)\right]
\label{Shat2}
\end{align}
(n.b. we have integrated by parts where necessary, neglecting surface terms).
The starting point of the path integral approach of~\cite{Laenen:2008gt} is to
separate the gauge field (here the graviton) into hard and soft modes. That is,
in the path integral which defines the quantum gravity theory, one may write
\begin{equation}
\int{\cal D}h^{\mu\nu}=\int{\cal D}h^{\mu\nu}_h{\cal D}h^{\mu\nu}_s,
\label{hsep}
\end{equation}
where $h^{\mu\nu}_h$ and $h^{\mu\nu}_s$ contain only hard and soft modes in
momentum space respectively. The exact details of this definition (which
may proceed e.g. by constructing an explicit hypersurface in the
multigraviton momentum space which separates hard and soft parts) need not
concern us, as discussed in~\cite{Laenen:2008gt} for the QED case. However,
we note that this separation breaks the full gauge invariance
of the gravity theory. Under a gauge transformation, the graviton behaves as
\begin{equation}
h^{\mu\nu}(x)\rightarrow h^{\mu\nu}(x)+\partial^\mu\xi^\nu(x)
+\partial^\nu\xi^\mu(x)
\label{htrans}
\end{equation}
where $x$ denotes 4-position. A general gauge function $\xi^\mu$
(defined in position space) will in general have both soft and hard modes
in momentum space, so that gauge transformations exist which violate the
separation into soft and hard modes introduced in eq.~(\ref{hsep}).
However, a residual gauge invariance remains, namely that one may
transform $h^{\mu\nu}$ in momentum space according to any
transformation\footnote{We use the same symbol to denote
the position- and momentum-space graviton fields, where the argument
of the field removes any ambiguity.}
\begin{equation}
h_{h,s}^{\mu\nu}(k)\rightarrow h_{h,s}^{\mu\nu}(k)+k^\mu\xi_{h,s}^\nu(k)
+k^\nu\xi_{h,s}^\mu(k)
\label{htrans2}
\end{equation}
in which the gauge function $\xi^\mu_{h,s}$ contains only hard or soft modes
as required. This symmetry will be sufficient to derive the structure of
factorization-breaking corrections from internal emission graphs in
section~\ref{sec:low}.
We now consider the Green's function for the scattering of $L$ scalar
particles which, after performing the above separation of the graviton
field, may be written
\begin{equation}
G(p_1\ldots,p_L)=\int{\cal D}h^{\mu\nu}_s H(x_1,\ldots, x_L)
e^{iS_{\rm E.H.}[h^{\mu\nu}_s]}\prod_{j=1}^L\langle p_j|(\hat{S}
-i\epsilon)^{-1}|x_j\rangle.
\label{Greens}
\end{equation}
Here $H(x_1,\ldots,x_L)$ is the {\it hard interaction}, which produces scalar
particles at 4-positions $\{x_i\}$, whose external momenta are
given by $\{p_i\}$. This is directly analagous to the hard interaction given
for the QED case in~\cite{Laenen:2008gt}, except for the fact that this
will now consist of a sum of Feynman diagrams involving hard graviton
modes, rather than hard photon modes. There are also integrations
over the positions $\{x_i\}$, which we do not show explicitly in
eq.~(\ref{Greens}). Associated with each external line
is a propagator for a scalar particle in a background soft gravitational
field. According to the well-known definition of the propagator, this
is given by the inverse of the quadratic operator $\hat{S}$ for the scalar
field $\phi$, which includes the relevant gravitational interactions as
shown in eq.~(\ref{Shat2}). The propagator is sandwiched between states
of given initial position and final momentum. The latter is due to the fact
that Green's functions (and, consequently, scattering amplitudes) are
usually considered in momentum space. The former arises because, in what
follows, we will interpret soft graviton scattering in
terms of the space-time worldlines of the emitting scalars which participate
in the hard interaction.
The next step is to note that the propagator factors in eq.~(\ref{Greens})
can be expressed in terms of first quantized path integrals. Such a technique
arose some years ago~\cite{Strassler:1992zr,Schmidt:1994zj,vanHolten:1995ds},
motivated by string theoretic approaches to field theory
amplitudes~\cite{Bern:1991an,Bern:1991aq} (see also~\cite{Karanikas:2002sy}
for an example of worldline techniques applied in a resummation context).
Here we shall quote the result as written in~\cite{Laenen:2008gt}, that the
propagator factors appearing in eq.~(\ref{Greens}) may be written
\begin{align}
\langle p_j|(\hat{S}-i\epsilon)^{-1}|x_j\rangle&=\frac{1}{2}\int_0^\infty dT
\int_{x(0)=x_j}^{p(T)=p_j}{\cal D}p{\cal D}x\exp\left[\phantom{\int}
-ip(T)\cdot x(T)\right.\notag\\
&\left.\quad+i\int_0^Tdt(p\cdot\dot{x}-\hat{H}(p,x))\right],
\label{proppath}
\end{align}
where
\begin{equation}
\hat{H}=\frac{1}{2}\hat{S}.
\label{Hdef}
\end{equation}
Some explanatory comments are in order. Firstly, there is a double path
integral over the position and momentum trajectories of the particle,
which are parametrized in terms of a time-like variable $t$ (the upper
limit of $T$ will eventually be taken
to $\infty$, corresponding to the final state). The boundary conditions
are that the particle must be produced at position $x_j$ ($x(0)=x_j$), and
end up with final momentum $p_j$ ($p(T)=p_j$). Equation~(\ref{proppath})
arises as the solution of a Schr\"{o}dinger equation for a certain
evolution operator formed out of $\hat{S}$, and indeed has the
recognizable Feynman path integral representation, where the
second term in the exponent is the classical action formed out of
the ``Hamiltonian'' $\hat{H}$. The first term in the exponent arises due
to the fact that we are sandwiching the propagator between position and
momentum states, rather than two position states. Note that the path
integral over $x$ in eq.~(\ref{proppath}) has a clear physical
interpretation in terms of a sum over all possible spacetime
trajectories of the scalar particle whose final momentum is $p_j$.
This will be crucial to isolating the properties of the eikonal
approximation (and beyond) in what follows.
From eqs.~(\ref{Shat2}) and~(\ref{Hdef}), we find that the appropriate
Hamiltonian operator for a scalar particle coupled to gravity is
(dropping the subscript $s$ on the soft graviton field $h^{\mu\nu}_s$, which
we do from now on)
\begin{align}
\hat{H}&=-\frac{1}{2}\eta^{\mu\nu}\partial_\mu\partial_\nu-\frac{\kappa}{2}
(\partial_\mu h^{\mu\nu})\partial_\nu-\frac{\kappa}{2}h^{\mu\nu}
\partial_\mu\partial_\nu\notag\\
&\quad+\frac{1}{2}m^2+\frac{m^2\kappa}{2(d-2)}\left[h^\alpha_\alpha
+\kappa\left(\frac{
(h^\alpha_\alpha)^2}{(d-2)}-h^{\alpha\beta}h_{\beta\alpha}\right)\right],
\label{Hgrav}
\end{align}
which in momentum space ($p_\mu=i\partial_\mu$)\footnote{Care must be taken
with the sign of the momentum operator, which is here chosen to ensure
consistency with the diagrammatic calculation of appendix~\ref{app:diags}.}
becomes
\begin{align}
\hat{H}&=\frac{1}{2}(p^2+m^2)+\frac{\kappa}{2}p_\mu p_\nu h^{\mu\nu}
+\frac{i\kappa}{2}
p_\nu(\partial_\mu h^{\mu\nu})\notag\\
&\quad+\frac{m^2\kappa}{2(d-2)}\left[h^\alpha_\alpha+
\kappa\left(\frac{
(h^\alpha_\alpha)^2}{(d-2)}-h^{\alpha\beta}h_{\beta\alpha}\right)\right],
\label{Hgrav2}
\end{align}
so that the propagator function of eq.~(\ref{proppath}) becomes
\begin{align}
\langle p_j|(\hat{S}-i\epsilon)^{-1}|x_j\rangle&=\frac{1}{2}\int_0^\infty dT
\int_{x(0)=x_j}^{p(T)=p_j}{\cal D}p{\cal D}x\exp\left\{
-ip(T)\cdot x(T)+i\int_0^Tdt\left[\phantom{\int}p\cdot\dot{x}\right.\right.\notag\\
&\quad-\frac{1}{2}(p^2+m^2)-\frac{\kappa}{2}p_\mu p_\nu h^{\mu\nu}
-\frac{i\kappa}{2}p_\nu(\partial_\mu h^{\mu\nu})\notag\\
&\quad\left.\left.-\frac{m^2\kappa}{2(d-2)}\left[h^\alpha_\alpha
+\kappa\left(\frac{
(h^\alpha_\alpha)^2}{(d-2)}-h^{\alpha\beta}h_{\beta\alpha}\right)\right]\right]
\right\}.
\label{proppath2}
\end{align}
At this point, we can recover the eikonal approximation. Recalling
that this corresponds to the momentum $k$ of any emitted gravitons being
completely soft ($k\rightarrow0$), this means that the emitting scalar
particles do not recoil. Thus, they follow the straight line classical
trajectories
\begin{equation}
x(t)=x_j+p_j t.
\label{classtraj}
\end{equation}
Subeikonal corrections correspond to a systematic expansion about this
trajectory. That is, for each external line one writes
\begin{equation}
x(t)\rightarrow x_j+p_j t+x(t),\quad p(t)\rightarrow p_j+p(t),
\label{eikexpand}
\end{equation}
where the boundary conditions for the transformed variables are $x(0)=p(T)=0$.
One then finds~\cite{Laenen:2008gt}
\begin{align}
\langle p_j|(\hat{S}-i\epsilon)^{-1}|x_j\rangle&=\frac{1}{2}\int_0^\infty
dTe^{-ip_j\cdot x_j-\frac{1}{2}(p_j^2+m^2)T}f(T),
\label{proppath3}
\end{align}
where
\begin{align}
f(T)&=
\int_{x(0)=0}^{p(T)=0}{\cal D}p{\cal D}x\exp\left\{
i\int_0^Tdt\left[p\cdot\dot{x}-\frac{1}{2}p^2-\frac{\kappa}{2}
(p_{j\mu}+p_\mu) (p_{j\nu}+p_\nu)
h^{\mu\nu}\right.\right.\notag\\
&\left.\left.\quad-\frac{i\kappa}{2}(p_{j\nu}+p_\nu)
(\partial_\mu h^{\mu\nu})
-\frac{m^2\kappa}{2(d-2)}\left[h^\alpha_\alpha
+\kappa\left(\frac{
(h^\alpha_\alpha)^2}{(d-2)}-h^{\alpha\beta}h_{\beta\alpha}\right)\right]\right]
\right\}.
\label{proppath4}
\end{align}
The path integral over $p$ is Gaussian, and may be performed explicitly. First,
one writes eq.~(\ref{proppath4}) as
\begin{equation}
f(T)=\int_{x(0)=0}^{p(T)=0}{\cal D}p{\cal D}x\exp\left[
i\int_0^Tdt\left(-\frac{1}{2}p_\mu A^{\mu\nu}p_\nu+B^\mu p_\mu +C\right)
\right],
\label{proppath5}
\end{equation}
where we have defined
\begin{align}
A^{\mu\nu}&=\eta^{\mu\nu}+\kappa h^{\mu\nu};\notag\\
B^\mu&=\dot{x}^\mu-\kappa p_{j\nu} h^{\mu\nu}
-\frac{i\kappa}{2}\partial_\nu h^{\nu\mu};\notag\\
C&=-\frac{\kappa}{2}p_{j\mu}p_{j\nu}h^{\mu\nu}-\frac{i\kappa}{2}
p_{j\nu}\partial_\mu h^{\mu\nu}-\frac{m^2\kappa}{2(d-2)}
\left[h^\alpha_\alpha+\kappa\left(\frac{
(h^\alpha_\alpha)^2}{(d-2)}-h^{\alpha\beta}h_{\beta\alpha}\right)\right],
\label{ABC}
\end{align}
and also used the fact that $h^{\mu\nu}=h^{\nu\mu}$.
Using standard results for Gaussian integrals gives
\begin{equation}
f(T)=\int{\cal D}x\exp\left[
i\int_0^Tdt\left(\frac{1}{2}B^\mu (A^{-1})_{\mu\nu}B^\nu +C\right)\right],
\label{fT}
\end{equation}
where we have ignored an overall normalization constant (which is cancelled
upon correctly normalising the measure for the path integral).
Noting that $A^{\mu\nu}$ as defined in eq.~(\ref{ABC}) is just the metric
tensor
$\tilde{g}^{\mu\nu}$ of eq.~(\ref{gtildedef}), its inverse is simply
$(A^{-1})_{\mu\nu}=\tilde{g}_{\mu\nu}$, as given in eq.~(\ref{gexpand2}). Then
eq.~(\ref{fT}) gives (after some tedious algebra)
\begin{align}
f(T)&=\int{\cal D}x\exp\left\{i\int_0^Tdt\left[\frac{\dot{x}^2}{2}
-\frac{\kappa}{2}\dot{x}^\mu h_{\mu\nu}\dot{x}^\nu-\kappa
p_{j\alpha}h^{\mu\alpha}\dot{x}_\mu
-\frac{i\kappa}{2}(\partial_\alpha h^{\alpha\nu})\dot{x}_\nu\right.
\right.\notag\\
&\left.\left.\quad+\frac{\kappa^2}{2}\dot{x}^\mu\dot{x}^\nu
h_{\mu}^{\phantom{\mu}\alpha}h_{\alpha\nu}+\kappa^2p_{j\alpha}
h^{\mu\alpha}h_{\mu\nu}\dot{x}^\nu+\frac{i\kappa^2}{2}
(\partial_\beta h^{\beta\nu})h_{\mu\nu}\dot{x}^\mu
+\frac{\kappa^2}{2}p_{j\alpha}p_{j\beta}
h^{\mu\alpha}h_\mu^{\phantom{\mu}\beta}\right.\right.\notag\\
&\left.\left.\quad+\frac{i\kappa^2}{2}p_{j\alpha}
h_\nu^{\phantom{\nu}\alpha}(\partial_\beta h^{\beta\nu})
-\frac{\kappa^2}{8}(\partial_\alpha h^{\alpha\mu})(\partial_\beta
h^{\beta}_{\phantom{\beta}\mu})-\frac{\kappa}{2}p_{j\mu}p_{j\nu}h^{\mu\nu}
-\frac{i\kappa}{2}p_{j\nu}\partial_\mu
h^{\mu\nu}\right.\right.\notag\\
&\left.\left.\quad-\frac{m^2}{2}\frac{\kappa}{(d-2)}h^\mu_\mu
-\frac{m^2\kappa^2}{2}\left(\frac{(h^\alpha_\alpha)^2}{(d-2)^2}
-\frac{h^{\alpha\beta}h_{\beta\alpha}}{d-2}\right)\right]\right\}.
\label{fT2}
\end{align}
One may relate all this back to the eikonal scattering amplitude as follows.
After combining eq.~(\ref{fT2}) into eq.~(\ref{proppath3}) and
substituting the resulting propagator factors into the Green's function
of eq.~(\ref{Greens}), one finds (after truncating the external lines with
free scalar propagators as required by the LSZ formula) that
the scattering amplitude for $L$ scalar particles has the form
(see the derivation in~\cite{Laenen:2008gt})
\begin{equation}
{\cal A}(p_1,\ldots,p_L)=\int{\cal D}h^{\mu\nu} H(x_1,\ldots,x_L)
e^{iS_{\rm E.H.}[h^{\mu\nu}]}\prod_{j=1}^{\infty}e^{-ip_j\cdot x_j}
f_j(\infty).
\label{amp1}
\end{equation}
That is, each external line is associated with a factor $f_j(\infty)$, given
by eq.~(\ref{fT2}) with $T\rightarrow\infty$, which still implicitly
includes a path integral over all possible trajectories of particle $j$.
We will shortly analyse this result further, but first note that
eq.~(\ref{amp1}) has the form of a generating functional for a
quantum field theory, i.e. for the soft graviton field. The external
line factors act as source terms in this theory which, as can be seen
from eq.~(\ref{fT2}), generate soft graviton emission vertices
localized along the external lines. The Feynman diagrams generated
by eq.~(\ref{amp1}) are thus {\it subdiagrams} in the full theory, which
span the external lines. Given that connected diagrams exponentiate in
quantum field theory, we can immediately conclude that a large class of soft
graviton corrections (namely those consisting of connected subdiagrams
formed from the Feynman rules derivable from eq.~(\ref{fT2})) exponentiate.
This is a generalization of the original result in~\cite{Weinberg:1965nx},
which is based solely on the eikonal contributions. The argument given here
for the exponentiation of soft graviton corrections based on the textbook
exponentiation properties of connected diagrams in quantum field theory
is essentially exactly the same as the argument used for QED
in~\cite{Laenen:2008gt}\footnote{Strictly speaking, exponentiation of
connected diagrams holds only subject to
the condition that the source terms for the gauge field are mutually
commuting. This is true
for gravity and abelian gauge theories, but is not the case in
non-Abelian gauge theories, where a more sophisticated treatment of
exponentiation is then required~\cite{Laenen:2008gt,Gardi:2010rn,
Mitov:2010rp}.}.
Let us now physically interpret the various terms in the external line
factor of eq.~(\ref{fT2}) in more detail. First, note that the Feynman
rules generated by each term cannot be immediately read off, due to the
fact that for each external line the path integral over trajectories
$x(t)$ has yet to be carried out. However, as shown in~\cite{Laenen:2008gt},
this can be done by systematically expanding about the classical trajectory.
The leading term (as already remarked above) gives the eikonal approximation,
in which the particles do not recoil. The first subleading corrections give
the next-to-eikonal corrections, and further corrections are
next-to-next-to-eikonal etc. Put another way, an expansion about the classical
trajectories of the hard emitting particles (in position space) is entirely
equivalent to an expansion in the momenta of the emitted gravitons
(a momentum space expansion). We may evaluate the eikonal contribution in
eq.~(\ref{fT2}) by setting $x(t)=p(t)=0$, and also neglecting terms which
are quadratic in $h^{\mu\nu}$ (these lead to two-graviton emission vertices,
which are suppressed with respect to the eikonal limit, due to having
one less eikonal denominator compared with two separate graviton emissions).
One may also neglect any terms containing derivatives of the graviton field,
as these (in momentum space) will be suppressed by powers of the graviton
momentum. The result is
\begin{equation}
f^{\rm Eik.}_j(\infty)=\int {\cal D}x\exp\left[i\int_0^\infty dt
\left(-\frac{\kappa}{2}
p_{j\mu}p_{j\nu}h^{\mu\nu}-\frac{m^2}{2(d-2)}\eta_{\mu\nu}h^{\mu\nu}\right)\right].
\label{fT3}
\end{equation}
Carrying out the path integral over $x$ gives unity assuming the appropriate
normalization, and thus the eikonal approximation for the scattering amplitude
is
\begin{align}
{\cal A}(p_1,\ldots,p_L)&=\int{\cal D}h^{\mu\nu} H(x_1,\ldots,x_L)
e^{iS_{\rm E.H.}[h^{\mu\nu}]}\prod_{j=1}^{\infty}e^{-ip_j\cdot x_j}
\exp\left[i\int_0^\infty dt \left(-\frac{\kappa}{2}
p_{j\mu}p_{j\nu}h^{\mu\nu}\right.\right.\notag\\
&\left.\left.\quad-\frac{m^2}{2(d-2)}\eta_{\mu\nu}h^{\mu\nu}\right)\right].
\label{amp2}
\end{align}
In fact, up to next-to-eikonal corrections, one may set the initial
positions of the external lines to $x_j=0$~\cite{Laenen:2008gt}, so that
eq.~(\ref{amp2}) simplifies further to
\begin{align}
{\cal A}(p_1,\ldots,p_L)&=H(0,\ldots,0)\int{\cal D}h^{\mu\nu}
e^{iS_{\rm E.H.}[h^{\mu\nu}]}\prod_{j=1}^{\infty}
\exp\left[i\int_0^\infty dt \left(-\frac{\kappa}{2}
p_{j\mu}p_{j\nu}h^{\mu\nu}\right.\right.\\
&\left.\left.\quad-\frac{m^2}{2(d-2)}\eta_{\mu\nu}h^{\mu\nu}\right)\right],
\label{amp3}
\end{align}
where we have taken the hard interaction outside the path integral over the
soft gauge field. One thus sees that the amplitude has factorized explicitly
into a hard and a soft part, where the soft gravitons are described by
the Wilson line factors
\begin{equation}
\Phi_j=\exp\left[i\int_0^\infty dt \left(-\frac{\kappa}{2}
p_{j\mu}p_{j\nu}h^{\mu\nu}-\frac{m^2}{2(d-2)}\eta_{\mu\nu}h^{\mu\nu}
\right)\right].
\label{Wilson1}
\end{equation}
The generating functional for the soft graviton field theory indeed generates
vacuum expectation values of these operators, as expressed
in eq.~(\ref{Snwilson}). Let us now discuss the physics
of these operators in more detail.
Firstly, the first term in the exponent in eq.~(\ref{Wilson1}) agrees,
up to a normalization factor, with the hypothesized Wilson
line operator of~\cite{Naculich:2011ry} (which dealt with pure graviton
scattering, and thus massless external particles).
Note, however, that the Feynman rule
for the soft graviton generated by this term is derived from
\begin{equation}
-\frac{i\kappa}{2}\int_0^\infty dt p_{j\mu}p_{j\nu}h^{\mu\nu}(p_jt)=
\int\frac{d^dk}{(2\pi)^d}\left(\frac{\kappa}{2}\frac{p_\mu p_\nu}
{p\cdot k}h^{\mu\nu}(k)\right),
\label{feynrule}
\end{equation}
where $h^{\mu\nu}(k)$ is the momentum space graviton field. The
momentum space eikonal Feynman rule is thus
\begin{displaymath}
\frac{\kappa}{2}\frac{p_\mu p_\nu}{p\cdot k},
\end{displaymath}
whose normalization agrees with the result quoted
in~\cite{Naculich:2011ry}. Thus,
we maintain that eq.~(\ref{Wilson1}) is correctly normalized.
Secondly, note that there is an additional contribution in the case in which
the external particles are massive, as has been considered here, which
appears as the second term in the exponent of eq.~(\ref{Wilson1}). Wilson
line operators in conventional gauge theories are the same for massless and
massive particles, due to the fact that the mass does not source the
gauge field. Here, however, the mass plays the role of a gravitational
charge, and thus we expect such a dependence. Furthermore, such a
term can be motivated on general Lorentz invariance grounds, given that
the exponent of the Wilson line operator can only depend on the properties
of the emitting particle. That is, the prefactor which one contracts with
the graviton field can only depend upon $p^\mu$ and $m$, and the only two
combinations having two Lorentz indices and the correct mass dimension are
\begin{displaymath}
p_\mu p_\nu,\quad m^2\eta_{\mu\nu},
\end{displaymath}
which indeed both occur.
Thirdly, it is interesting to note the behaviour of eq.~(\ref{Wilson1})
under rescalings of external line momentum. In QED, the Wilson line operator
has the form
\begin{equation}
\Phi_{\rm QED}=\exp\left[ie\int dt p_\mu A^\mu\right],
\label{Wilson2}
\end{equation}
where $e$ is the electromagnetic charge. This leads to the momentum-space
eikonal Feynman rule
\begin{equation}
\frac{p_\mu}{p\cdot k}
\label{feynQED}
\end{equation}
for emission of a soft photon of momentum $k$, and this latter expression
is manifestly invariant under the transformation $p\rightarrow\lambda p$,
corresponding to the invariance of eq.~(\ref{Wilson2}) under reparametrizations of
the Wilson line. By contrast, the eikonal Feynman rules generated from
eq.(\ref{Wilson1}) do not have this property. Rather, they can be written
in the form
\begin{equation}
\frac{\kappa}{2}\left[p_{\mu}\left(\frac{p_\nu}{p\cdot k}\right)
+m\left(\frac{\eta_{\mu\nu}}{2(d-2)}\frac{m}{p\cdot k}\right)\right].
\label{feyngrav2}
\end{equation}
The quantities in the round brackets are both manifestly scale invariant
under rescalings of the external momentum. These are then multiplied
respectively by factors of $p_\mu$ and $m$, namely by gravitational charges.
Thus, the eikonal Feynman rules in gravity have the form of a scale-invariant
(in $p$) quantity multiplied by a charge, which is exactly what one expects
by analogy with abelian and non-abelian gauge theories.
The curious reader may wonder about the presence of poles in $d-2$
in the Wilson line operators. As discussed in e.g.~\cite{Hamber:2007fk},
perturbation theory becomes singular in $d=2$ dimensions, owing to the fact
that the Einstein-Hilbert action is then a topological invariant (i.e. there
is no non-trivial dynamics). Whether or not such poles occur in the Wilson
line operators is dependent on the choice of weak field approximation. If
one expands the metric itself rather than absorbing $\sqrt{-g}$, the $(d-2)$
poles instead occur in the graviton propagator, such that one can never
get rid of them, as expected from their physical origin.
Above, we have confirmed the factorization property of soft graviton
amplitudes in the eikonal approximation, in terms of vacuum expectation
values of appropriate Wilson line operators. Some comments are in order
regarding the plethora of additional terms in eq.~(\ref{fT2}). As
explained in~\cite{Laenen:2008gt}, the path integral over $x$ in
eq.~(\ref{fT2}) can be systematically expanded about the leading
eikonal term. This was done for the case of abelian and non-abelian
gauge theory by defining $p=\lambda n$, where $\lambda\rightarrow\infty$
isolates the eikonal term, and $n^2=0$ or $n^2=1$ for massless and massive
particles respectively.
The path integral could then be done perturbatively
in $\lambda^{-1}$ (also taking care to expand the argument of each gauge boson
field), leading to a set of next-to-eikonal Feynman rules.
There were two types of next-to-eikonal vertex: those involving a single
photon or gluon emission, and those involving two gauge bosons emitted from
the same position on the external line. These vertices were then confirmed
by an explicit diagrammatic treatment in~\cite{Laenen:2010uz}.
A similar procedure is possible in the present case.
Note, however, that the simple
scaling $p_j=\lambda n_j$ does not solely isolate the eikonal term as
$\lambda\rightarrow\infty$. Instead, eq.~(\ref{fT2}) becomes
(as $\lambda\rightarrow\infty$)
\begin{align}
\lim_{\lambda\rightarrow\infty}f(T)&=\int{\cal D}x\exp\left\{i\int_0^Tdt\left[
-\frac{\kappa\lambda^2}{2}n_{j\mu}n_{j\nu}h^{\mu\nu}
-\frac{\lambda^2m^2\kappa}{d-2}\eta_{\mu\nu}h^{\mu\nu}
+\frac{\lambda^2\kappa^2}{2}n_{j\alpha}n_{j\beta}
h^{\mu\alpha}h_\mu^{\phantom{\mu}\beta}\right.\right.\notag\\
&\left.\left.\quad
-\frac{\lambda^2m^2\kappa^2}{2}\left(\frac{(h^\alpha_\alpha)^2}{(d-2)^2}
-\frac{h^{\alpha\beta}h_{\beta\alpha}}{d-2}\right)\right]\right\}.
\label{fTeik}
\end{align}
The first two terms are the eikonal contribution of eq.~(\ref{Wilson1}),
whereas the last two term are two-graviton vertices, which have no
analogue in the case of abelian or non-abelian gauge theory. They are clearly
suppressed in the eikonal limit, as such vertices have one less eikonal
denominator than two separate eikonal emission vertices. The above remarks
imply that the $\lambda$ scaling of~\cite{Laenen:2008gt}
is not quite correct as a systematic method for expanding about the eikonal
approximation\footnote{We will see in what follows, however, that
the third term in eq.~(\ref{fTeik}) is cancelled upon performing
the path integral over the trajectory $x$.},
although does indeed lead to correct results for the gauge
theories studied in that paper. The extra terms arise here essentially because
of the weak field expansion, which allows the possibility to construct
terms which are quadratic in external momenta or masses,
and quadratic (or higher) in the graviton field. Such terms are absent in
QED, where no weak field expansion is needed.
In any case, it is straightforward to
isolate next-to-eikonal contributions from power-counting,
even if one does not use the $\lambda$-expansion (which
amounts to little more than a convenient
book-keeping device where applicable). To calculate the next-to-eikonal
graviton Feynman rules, one must perform the path integral perturbatively
as explained in appendix B of~\cite{Laenen:2008gt}. Here we perform this
calculation in appendix~\ref{app:path}, with the result that, up to
next-to-eikonal order, the external line factor of eq.~(\ref{fT2}) may be
written (taking $T\rightarrow\infty$)
\begin{align}
f(\infty)&=\exp\left\{\int\frac{d^dk}{(2\pi)^d}h^{\mu\nu}(k)
\left[\frac{\kappa}{2}\frac{p_\mu p_\nu}{p\cdot k}-\frac{\kappa}{4}
\frac{p_{(\mu}k_{\nu)}}{p\cdot k}+\frac{m^2}{2}\frac{\kappa}{(d-2)}
\frac{\eta_{\mu\nu}}{p\cdot k}+\frac{\kappa}{4}\frac{k^2}{(p\cdot k)^2}
p_\mu p_\nu\right.\right.\notag\\
&\left.\left.+\frac{m^2}{4}\frac{\kappa}{d-2}\frac{k^2}{(p\cdot k)^2}
\eta_{\mu\nu}\right]+\int\frac{d^d k}{(2\pi)^d}\int\frac{d^dl}{(2\pi)^d}
h^{\mu\nu}(k)\,h^{\alpha\beta}(l)\left[-\frac{\kappa^2}{16}
\left(\frac{p_\alpha \,p_\beta \,p_{(\mu}k_{\nu)}}{p\cdot k \,p\cdot (k+l)}
\right.\right.\right.\notag\\
&\left.\left.\left.
+\frac{p_\alpha \,p_\beta \,p_{(\mu}\,l_{\nu)}}{p\cdot l\,p\cdot (k+l)}
+(\mu\nu\leftrightarrow\alpha\beta)\right)+\frac{\kappa^2}{8}
\frac{p_\mu\,p_\nu\,p_\alpha\,p_\beta\,(k\cdot l)}{p\cdot k\,p\cdot l\,
p\cdot(k+l)}-\frac{m^2\kappa^2}{16(d-2)}\left(\frac{\eta_{\mu\nu}\,
p_{(\alpha}\,k_{\beta)}}{p\cdot k\,p\cdot(k+l)}\right.\right.\right.\notag\\
&\left.\left.\left.+\frac{\eta_{\mu\nu}\,
p_{(\alpha}\,l_{\beta)}}{p\cdot l\,p\cdot(k+l)}+(\mu\nu\leftrightarrow\alpha
\beta)\right)+\frac{m^2\kappa^2}{8(d-2)}\left(\frac{\eta_{\mu\nu}\,p_\alpha\,
p_\beta\,(k\cdot l)}{p\cdot k\,p\cdot l\,p\cdot(k+l)}+(\mu\nu
\leftrightarrow\alpha\beta)\right)\right.\right.\notag\\
&\left.\left.+\frac{m^4\kappa^2}{8(d-2)^2}\frac{\eta_{\mu\nu}\,\eta_{\alpha
\beta}\,(k\cdot l)}{p\cdot k\,p\cdot l\,p\cdot(k+l)}+\frac{m^2\kappa^2}{4
p\cdot(k+l)}\left(\frac{2\eta_{\mu\nu}\,\eta_{\alpha\beta}}{(d-2)^2}
-\frac{1}{d-2}(\eta_{\mu\beta}\,\eta_{\nu\alpha}+\eta_{\nu\beta}\,
\eta_{\mu\alpha})\right)\right]\right\}.
\label{expmom}
\end{align}
Here we have written the exponent explicitly in momentum space, where
we use the argument of the graviton field to identify its
Fourier-transform.
From eq.~(\ref{expmom}) one can easily read off the effective
Feynman rules for graviton emission up to NE order. There are eikonal and
NE one-graviton vertices given by
\begin{equation}
\scalebox{0.6}{\includegraphics{1gE.eps}}\quad\frac{\kappa}{2}\frac{1}
{p\cdot k}\left[p_\mu p_\nu+\frac{m^2}{d-2}\eta_{\mu\nu}\right]
\label{1gErule}
\end{equation}
and
\begin{equation}
\scalebox{0.6}{\includegraphics{1gNE.eps}}
\quad\frac{\kappa}{4}\left[
-\frac{p_{(\mu}k_{\nu)}}{p\cdot k}+\frac{k^2}{(p\cdot k)^2}
\left(p_\mu \,p_\nu+\frac{m^2}
{d-2}\eta_{\mu\nu}\right)\right]
\label{1gNErule}
\end{equation}
respectively, where the eikonal result for massless particles
is already well-known~\cite{Weinberg:1965nx}. There is also an
effective two-graviton vertex, given by
\begin{align}
\scalebox{0.6}{\includegraphics{2gNE.eps}}
\quad&\frac{\kappa^2}{16}\frac{1}{p\cdot k\,p\cdot l\,p\cdot(k+l)}\left\{
-\left[\left(p_\alpha p_\beta+\frac{m^2}{d-2}\eta_{\alpha\beta}\right)
\left(p_{(\mu}k_{\nu)}(p\cdot l)+p_{(\mu}l_{\nu)}(p\cdot k)\right)\right.
\right.\notag\\
&\left.\left.\quad
-\frac{2m^2}{d-2}\eta_{\mu\nu}p_\alpha p_\beta(k\cdot l)+(\mu\nu
\leftrightarrow\alpha\beta)\right]+2p_\mu\,p_\nu\,p_\alpha\,p_\beta(k\cdot l)
+\frac{2m^4}{(d-2)^2}\eta_{\mu\nu}\eta_{\alpha\beta}(k\cdot l)\right.\notag\\
&\left.\quad
+4m^2(p\cdot k)(p\cdot l)\left(\frac{2\eta_{\mu\nu}\eta_{\alpha\beta}}
{(d-2)^2}-\frac{1}{d-2}\left(\eta_{\mu\beta}\eta_{\nu\alpha}+\eta_{\nu\beta}
\eta_{\mu\alpha}\right)\right)\right\}.
\label{2gNErule}
\end{align}
One expects such a vertex for two reasons. Firstly, whilst successive graviton
emissions are completely decoupled in the eikonal approximation, this is not
expected to hold beyond the eikonal approximation, such that one expects
two-graviton correlations. Secondly, there is an exact two-graviton vertex in
the theory (as discussed in appendix~\ref{app:diags}), which forms part of the
result of eq.~(\ref{2gNErule}). It is useful to cross-check
the above results using diagrammatic methods,
as has been carried out for the case of QED and QCD in appendix B
of~\cite{Laenen:2008gt}. We present such a calculation for the present case
in appendix~\ref{app:diags} of this paper.
The above results imply that a subset of next-to-eikonal corrections
exponentiates at NE order in perturbative quantum gravity. Namely,
connected external emission graphs formed from effective eikonal
and NE Feynman rules, where each graph contains at most one NE vertex.
Some further comments are in order regarding the results we have obtained.
To start with, one may question how useful the NE Feynman rules are in
practice. Eikonal Feynman rules tell us about infrared singularities,
which may be expected to be somewhat universal in alternative field
theories of gravity. Next-to-eikonal corrections are non-singular, and thus
potentially tell us less about generic infrared features of gravity theories.
Furthermore, the fact that perturbative GR coupled to matter is not UV
renormalizable suggests that an expansion in the momenta of emitted gravitons
may cease to be meaningful at some point. Thus, one should be wary of
taking any NE corrections too seriously. What matters in the above analysis
is merely that in the soft limit, we have shown that the eikonal approximation
leads to Wilson line operators as hypothesized in~\cite{Naculich:2011ry},
where factorizable corrections are genuinely
subleading in that they do not produce
IR singularities. We have also clarified the situation when external particles
have non-zero mass.
Another important point is that unlike in abelian and non-abelian gauge
theory, the NE Feynman rules
are not unique. They depend on the nature of the weak field approximation
used to define the graviton field (e.g. whether one expands $g^{\mu\nu}$
or $\tilde{g}^{\mu\nu}$). This is not the case for the eikonal terms
in four dimensions. We have verified by explicit calculation (which we do
not consider worth recording here) that the
Wilson line operator which occurs upon expanding $g^{\mu\nu}$ rather than
$\tilde{g}^{\mu\nu}$ is
\begin{equation}
\Phi_j=\exp\left[i\int_0^\infty dt \left(-\frac{\kappa}{2}
p_{j\mu}p_{j\nu}h^{\mu\nu}-\frac{m^2}{4}\eta_{\mu\nu}h^{\mu\nu}
\right)\right].
\label{Wilson3}
\end{equation}
Comparing this with eq.~(\ref{Wilson2}), we see that the two Wilson line
operators are equal for $d=4$. They are also automatically equal for massless
external particles. The next-to-eikonal terms, however, would differ between
the two calculations, due to extra terms originating from the expansion of
$\sqrt{-g}$ in the kinetic energy term for the scalar field.
Such differences presumably cancel at the level of
full diagrams (as the choice of weak field approximation shuffles
contributions between different vertices), but it is still true that a level
of ambiguity occurs in thinking about NE Feynman rules that is absent in
conventional gauge theories, as a weak field expansion is not necessary in the
latter.
A further comment is in order regarding the weak field approximation.
In the above analysis, we have performed two expansions. Firstly,
we have separated the graviton field into hard and soft modes, and used this
as a basis for an expansion in the momenta of emitted gravitons. Secondly,
we have expanded the graviton field itself in the gravitational coupling
constant $\kappa$. In the latter expansion, we kept terms only up to
${\cal O}(\kappa^2)$. How can we be sure that by including higher order terms,
we do not induce further (next-to) eikonal corrections?
To see that this is not the
case, it suffices to note that any power of $\kappa$ is accompanied by a
further power of the graviton field $h^{\mu\nu}$, whose Lorentz indices
must be appropriately contracted with other factors of $h^{\mu\nu}$,
$\dot{x}^\mu$, $p_j^\mu$, $p^\mu$ or $\partial^\mu$ in eq.~(\ref{fT2}).
This always results in a multigraviton vertex, which is suppressed with
respect to the appropriate number of eikonal 1-graviton vertices due to the
fact that it lacks at least one eikonal denominator. The fact that
all vertices at ${\cal O}(\kappa^n)$ generate $n$-graviton vertices also
shows that for next-to-eikonal corrections it is sufficient to expand only
up to ${\cal O}(\kappa^2)$, as we have done above. This is itself interesting,
as it shows that the momentum expansion and weak field expansions are
partially correlated in a well-defined sense (although there is, of course,
a tower of subleading momentum terms at any given order in $\kappa$).
It follows from the above discussion that the NE Feynman rules will
contain 1-graviton and 2-graviton vertices. It then follows (as for
the eikonal corrections) that a large class of next-to-eikonal graviton
corrections exponentiates, namely those graphs containing connected
subdiagrams involving one next-to-eikonal Feynman rule. Examples are
shown in figure~\ref{NEexp}.
\begin{figure}
\begin{center}
\scalebox{1.0}{\includegraphics{NEexp.eps}}
\caption{Example external emission graphs which exponentiate at
next-to-eikonal order, where $\bullet$ represents a next-to-eikonal
vertex, the doubled gluon line a graviton, and all other vertices are
assumed to be eikonal.}
\label{NEexp}
\end{center}
\end{figure}
Thus, the exponent of the scattering amplitude for $L$-particle scattering
dressed by soft gravitons has the same schematic form as the QED case of
eq.~(\ref{ampstruc}). As in that case, this is not the whole story. One
may also get additional contributions from internal emission diagrams such
as that shown in figure~\ref{intext}(c), which explicitly break the
factorization of the scattering amplitude into hard and soft parts. These
are known to be subleading in the soft limit in conventional gauge
theories~\cite{Low:1958sn,Burnett:1967km,DelDuca:1990gz}. It is likely,
but not perhaps obvious, that the same is true for gravity. This is the subject
of the following section.
\section{Factorization breaking terms}
\label{sec:low}
In the previous section, we applied the path integral resummation method
of~\cite{Laenen:2008gt} to perturbative general relativity minimally coupled
to a complex scalar field. We found that the amplitude factorizes into hard
and soft parts, where the latter is decribed by a vacuum expectation value of
Wilson line operators as in eq.~(\ref{Wilson1}). Explicit calculation
yields diagrams formed out of eikonal Feynman rules, which then exponentiate.
Corrections to the exponent are subleading, and consist of diagrams containing
one next-to-eikonal Feynman rule, with all other vertices eikonal. The aim of
this section is to examine what happens when this factorization is broken,
and to argue that, as in QED, such corrections are next-to-eikonal, thus do
not affect the structure of IR singularities.
First let us briefly recap what happens in QED. We denote by
$H^{\mu}(p_1,\ldots p_L;k)$ the subamplitude for production of a soft photon
emerging from the hard interaction $H(p_1,\ldots p_L)$ with momentum $k$ and
with Lorentz index $\mu$. Then gauge invariance may be applied to show
that $H^\mu$ is related to the hard interaction itself, up to next-to-eikonal
order. More specifically, this relation has the following form:
\begin{equation}
H^\mu(p_1,\ldots p_L;k)=-\sum_{j=1}^Lq_j\frac{\partial}{\partial
p_{j\mu}} H(p_1,\ldots p_L),
\label{low1}
\end{equation}
where $q_j$ is the electric charge of particle $j$. This equation tells us
that factorization breaking terms are subleading with respect to the
eikonal approximation. Furthermore, their structure is completely fixed
up to next-to-eikonal order. Such contributions do not formally exponentiate
(as do external emission graphs), but they do nevertheless have an iterative
structure to all orders in perturbation theory, as given by eq.~(\ref{low1}).
This equation is an important ingredient in Low's
theorem~\cite{Low:1958sn,Burnett:1967km,DelDuca:1990gz}, which relates the
vertex function in gauge theory to the vertex function with an additional
emission of a gauge boson.
The above result suggests that we should attempt a similar strategy in quantum
gravity. That is, one may impose gauge invariance on the factorized form
for the soft graviton scattering amplitude, and check that this leads to
extra contributions which are suppressed with respect to the eikonal limit.
As discussed in the previous section, the separation of the graviton
field into hard and soft modes has broken the full gauge invariance of the
theory. Nevertheless, the residual gauge symmetry under transformations
of the type shown in eq.~(\ref{htrans2}) will be sufficient to derive a
gravitational analogue of eq.~(\ref{low1}).
Our starting point is the factorized form of the soft graviton amplitude
of eq.~(\ref{amp2}), where each external line factor is reduced to a
Wilson line operator. Under a gauge transformation (eq.~(\ref{htrans})),
the Wilson operators transform as
\begin{equation}
f_j(h^{\mu\nu}+\partial^\mu\xi^\nu+\partial^\nu\xi^\mu)
\rightarrow f(h^{\mu\nu})e^{-i\kappa p_j\cdot\xi_j}.
\label{ftrans}
\end{equation}
Requiring that eq.~(\ref{amp2}) be gauge invariant then implies that
the hard interaction transforms as
\begin{equation}
H(x_1,\ldots x_L;h^{\mu\nu}+\partial^\mu\xi^\nu+\partial^\nu\xi^\mu)
\exp\left[i\kappa\sum_jp_j\cdot\xi(x_j)\right].
\label{Htrans}
\end{equation}
Expanding both sides to first order in $h^{\mu\nu}$ and $\xi^\mu$ gives
\begin{align}
&\quad H(x_1,\ldots,x_L)+\int d^dxH_{\mu\nu}(x_1,\ldots x_L;x)\left[
h^{\mu\nu}(x)+\partial^\mu\xi^\nu(x)+\partial^\nu\xi^\mu(x)\right]\notag\\
&=H(x_1,\ldots x_L)+\int d^dxH_{\mu\nu}(x_1,\ldots,x_L;x)h^{\mu\nu}(x)\notag\\
&\quad+i\kappa\int d^dx\left[H(x_1,\ldots x_L)\sum_j p_{j\mu}
\delta(x-x_j)\right]\xi^\mu(x),
\label{Hexp1}
\end{align}
where all hard functions are evaluated with the untransformed graviton
field $h^{\mu\nu}$, and $H_{\mu\nu}(x_1,\ldots,x_L;x)$ is the
subamplitude for emission of a graviton from the hard interaction at position
$x$ (this is the analogue of $H^\mu(x_1,\ldots,x_L;x)$ in the QED case, whose
Fourier transform appears on the left-hand side of eq.~(\ref{low1})).
Equation~(\ref{Hexp1}) simplifies to
\begin{align}
&\quad\int d^dxH_{\mu\nu}(x_1,\ldots x_L;x)\left[
\partial^\mu\xi^\nu(x)+\partial^\nu\xi^\mu(x)\right]
\notag\\
&=i\kappa\int d^dxH(x_1,\ldots x_L)\sum_j p_{j\mu}\delta(x-x_j)\xi^\mu(x),
\label{Hexp2}
\end{align}
and integrating by parts on the left-hand side gives
\begin{equation}
-\int d^dx\xi^\mu\partial^\nu\bar{H}_{\mu\nu}
=i\kappa\int d^dxH(x_1,\ldots x_L)\sum_j p_{j\mu}\delta(x-x_j)\xi^\mu(x),
\label{Hexp3}
\end{equation}
where $\bar{H}_{\mu\nu}=H_{\mu\nu}+H_{\nu\mu}$. Due to the fact that
$\xi^\nu$ is arbitrary, one may write
\begin{equation}
-\partial^\nu\bar{H}_{\mu\nu}=i\kappa\sum_jp_{j\mu}H(x_1,\ldots x_L)
\delta(x-x_j).
\label{Hexp4}
\end{equation}
In momentum space this becomes
\begin{equation}
-ik^\nu\bar{H}_{\mu\nu}=i\kappa\sum_jp_{j\mu}H(p_1,\ldots,p_j+k,\ldots p_L),
\label{Hexpmom1}
\end{equation}
and expanding this to ${\cal O}(k)$ gives
\begin{equation}
-k^\nu\bar{H}_{\mu\nu}=\kappa\sum_j\left[p_{j\mu}+ p_{j\mu}k^\nu
\frac{\partial}{\partial p_j^\nu}\right]H(p_1,\ldots,p_L).
\label{Hexpmom2}
\end{equation}
The zeroth order term on the right-hand side cancels due to momentum
conservation
\begin{equation}
\sum_jp_{j\mu}=0,
\label{momcon}
\end{equation}
and one finally finds
\begin{equation}
\bar{H}_{\mu\nu}=-\kappa\sum_jp_{j\mu}\frac{\partial}{\partial p_j^\nu}
H(p_1,\ldots,p_L).
\label{gravlow}
\end{equation}
This is the gravitational equivalent of the QED result expressed by
eq.~(\ref{low1}). It forms part of a gravitational version of Low's theorem,
and describes the result of internal emission diagrams analagous
to that shown in
figure~\ref{intext}(c), in which a soft graviton emerges from the hard
interaction and lands on an external line. Clearly such diagrams are
next-to-eikonal from eq.~(\ref{gravlow}). To see this, imagine starting
with an external emission graph, and taking one of the soft gravitons off
an external line and into the hard interaction. Such a replacement replaces
an eikonal Feynman rule with the contribution of eq.~(\ref{gravlow}), and
the resulting diagram thus has one less eikonal denominator of form
$p\cdot k$, therefore is subleading with respect to the eikonal approximation.
It is worth noting that one may also derive eq.~(\ref{gravlow}) from a
traditional diagrammatic treatment - we present an example calculation
in appendix~\ref{app:low}.
There is also an additional contribution at NE order, corresponding to
the fact that the external lines do not start at the origin, but have
non-zero initial 4-positions $x_j$, as taken into account by the factors
$e^{-p_j\cdot x_j}$ in eq.~(\ref{amp1}). As in~\cite{Laenen:2008gt}, one may
evaluate this contribution by writing the eikonal one-graviton source
term for an external line of momentum $p$ as
\begin{equation}
\frac{\kappa}{2}\int\frac{d^dk}{(2\pi)^d}\frac{h^{\mu\nu}}{p\cdot k}\left(
p_\mu\,p_\nu+\frac{m^2}{d-2}\eta_{\mu\nu}\right)e^{ix\cdot k}
=\frac{\kappa}{2}\int\frac{d^dk}{(2\pi)^d}\frac{h^{\mu\nu}}{p\cdot k}\left(
p_\mu\,p_\nu+\frac{m^2}{d-2}\eta_{\mu\nu}\right)(1+ix\cdot k),
\label{jterm1}
\end{equation}
where we have expanded to NE order on the right-hand side. Combining the
second term on the right-hand side with the rest of eq.~(\ref{gravlow}),
the scattering amplitude is given by
\begin{align}
&{\cal A}(p_1,\ldots,p_L)=\int{\cal D}h\,e^{iS_{\rm E.H.}[h]}\left[
\left(-\sum_{j=1}^L p_{j\mu}\frac{\partial}{\partial p_j^\nu}H(p_1\ldots,p_L)
\right)\int\frac{d^dk}{(2\pi)^d}h^{\mu\nu}\right.\notag\\
&\left.+\int dx_1^d\ldots dx_L^d
H(x_1,\ldots x_n)\left(\sum_j\int\frac{d^dk}{(2\pi)^d}
(ix_j\cdot k)\frac{h^{\mu\nu}}{p_j\cdot k}\left(p_{j\mu}\,p_{j\nu}
+\frac{m^2}{d-2}
\eta_{\mu\nu}\right)\right)e^{-i\sum_jx_j\cdot p_j}
\right]\notag\\
&\times\prod_{j=1}^N f(0,p_j;h),
\label{jterm2}
\end{align}
where we have explicitly reinstated the integrals over the initial positions
of the external lines, and $f(0,p_j;h)$ is the external line factor starting
at the origin for line $j$. Performing the position integrals, one obtains
\begin{align}
{\cal A}(p_1,\ldots,p_L)&=\int{\cal D}h\left\{\int\frac{d^dk}{(2\pi)^d}
h^{\mu\nu}(k)\sum_{j=1}^L\left[-\left(p_{j\mu}\,p_{j\nu}+\frac{m^2}{d-2}
\eta_{\mu\nu}\right)k^\sigma\frac{\partial}{\partial p_j^\sigma}
-p_j^\mu\frac{\partial}{\partial p_j^\nu}\right]\right.\notag\\
&\left.\quad\times H(p_1,\ldots,p_L)
\right\}\prod_{j=1}^N f(0,p_j;h).
\label{jterm3}
\end{align}
One sees that the result of the non-zero initial positions of the
external lines is to add an extra NE term
involving momentum derivatives of the hard interaction. A similar result occurs
in QED and QCD, and it is useful to physically interpret this extra
contribution. In abelian and non-abelian gauge theory, massless particles
give rise to collinear as well as soft singularities, and one may write the
scattering amplitude in the factorized form (analagous to eq.~(\ref{Andef}))
\begin{equation}
{\cal A}=H\cdot S\cdot\prod_{j=1}^L J_j.
\label{ampfac}
\end{equation}
Here $H$ and $S$ are the hard and soft functions, where the latter collects
all singularities due to the emitted gauge boson momenta satisfying
$k\rightarrow 0$. Furthermore, $J_j$ is a {\it jet function}, and contains
all singularities arising from hard emissions which are collinear to
the outgoing particle of momentum $p_j$\footnote{Note
there is an overlap between
$S$ and $J_j$, namely when singularities become soft and collinear. This is
usually dealt with by dividing either the soft or jet functions by the
{\it eikonal jet functions} evaluated in the soft limit, which removes any
double counting.}. Del Duca has shown~\cite{DelDuca:1990gz} that
in such cases there are extra contributions to the Low-Burnett-Kroll theorem
due to emissions of soft gluons from inside the jet functions. This is
indeed the physical meaning of the additional term in eq.~(\ref{jterm3}).
One would not ordinarily write jet functions in perturbative gravity, due to
the fact that hard collinear singularities cancel after summing over all
external particle lines and using momentum conservation~\cite{Weinberg:1965nx}.
However, if one is interested in the structure of next-to-eikonal corrections,
there are such terms which do indeed come from emissions within jet functions,
and which do not cancel after summing over parton legs (due presumably to the
fact that the additional term in eq.~(\ref{jterm3}) is quadratic in each
external momentum). It may thus be useful to consider jet functions in gravity
after all.
In this section, we have investigated the structure of terms which break the
factorization of soft graviton amplitudes into hard and soft parts, and found
contributions whose physical meaning is that of emission of soft gravitons from
inside the hard and jet functions. All such contributions are subleading with
respect to the eikonal approximation, thus verifying the factorization of
graviton ampitudes at eikonal order. Furthermore, the structure of soft
graviton amplitudes up to NE order has the same schematic form as in QED or
QCD, namely that of eq.~(\ref{ampstruc}). External emission graphs (consisting
of connected subdiagrams formed out of eikonal and next-to-eikonal
Feynman rules) formally exponentiate, whereas internal emission graphs have
an iterative structure and do not formally enter the exponent.
\section{Discussion}
\label{sec:discuss}
In this paper, we have studied the recent hypothesis made
in~\cite{Naculich:2011ry}, that soft graviton amplitudes factorize into
hard and soft functions, as do gauge theory amplitudes. Our analysis
relied on path integral resummation techniques developed in the context
of abelian and non-abelian gauge theory~\cite{Laenen:2008gt}, which allow
an efficient classification of the structure of scattering amplitudes
including subleading corrections with respect to the eikonal approximation.
We find that gravity amplitudes have the same schematic structure as in gauge
theory, as written here in eq.~(\ref{ampstruc}). That is, eikonal graphs
exponentiate (as already known~\cite{Weinberg:1965nx}), and next-to-eikonal
graphs can be separated into external and internal emission contributions.
External emission graphs are connected subdiagrams formed out of effective
eikonal and next-to-eikonal Feynman rules, whose form we derived in this paper.
Internal emission graphs consist of soft gravitons which are emitted from
inside the hard interaction, or jet functions associated with each external
line. External emission graphs formally exponentiate, whereas internal emission
graphs have an iterative structure to all orders in perturbation theory.
Importantly, internal emission graphs are subleading with respect to the
eikonal approximation.
Note that, as in the cases of QED and QCD considered in~\cite{Laenen:2008gt},
one may quibble whether it is really valid to separate internal and external
emission contributions as far as exponentiation is concerned, or even what
is meant by next-to-eikonal exponentiation at all. After all, upon performing
the exponent of the NE external emission contributions, any term involving
products of two or more NE terms is NNE order or higher, such that
eq.~(\ref{ampstruc}) is formally equivalent to
\begin{equation}
{\cal A}={\cal A}_0\exp\left[\sum_{G^{\E}}G^{\E}
\right]\left(1+\sum_{G^{\NE}}G^{\NE}
+{\cal A}_{rem.}\right),
\label{ampstruc2}
\end{equation}
in which formula external and internal emission contributions are on the
same footing. Nevertheless, it is true that the former genuinely exponentiate,
whereas the latter have an iterative structure and thus must be calculated
differently. Furthermore, either eq.~(\ref{ampstruc}) or eq.~(\ref{ampstruc2})
allow NE contributions to be calculated to all orders in perturbation theory,
even if this is really a consequence of the exponentiation of eikonal terms,
as shown explicitly in eq.~(\ref{ampstruc2}).
Aside from the structure of NE effects, an important question
regarding these results is: have we formally proved
the factorization of graviton amplitudes (albeit those involving only
scalar external lines) into hard and soft parts? After all, the path
integral technique we have used starts off by assuming that we may separate
the gauge field into hard and soft modes, and that the resulting exponential
structure captures all singularities. It may perhaps be prudent to regard
the results we have obtained as being supporting evidence for the factorization
property of amplitudes, rather than a rigorous proof.
Nevertheless, it is {\it very strong} evidence, for a number
of reasons. Firstly, we have analysed corrections which explicitly break the
factorization property (i.e. internal emission graphs), and found these to be
subleading in the momenta of the emitted gluons. We showed this using both
the path integral framework, and also using diagrammatic methods (although this
is essentially the same calculation, as it is gauge invariance that ultimately
fixes these contributions). Secondly, we find that the structure of eikonal
and next-to-eikonal contributions is exactly analagous to that of QED and QCD.
This suggests that the qualitative structures one finds in the infrared
limits of both gauge and gravity theories are somewhat generic.
In gauge theories, the factorization property (together with the structure of
the NE Feynman rules) is indeed explicitly corroborated using diagram-based
methods, which gives us confidence that the same would be true for gravity.
A more formal proof of the factorization of gravity amplitudes may perhaps
proceed using a power-counting analysis based upon the Landau equations
associated with the singularities of Feynman diagrams, as
has been carried out in gauge theories (see e.g.~\cite{Sterman:1995fz}
for a pedagogical review).
We are also able to confirm the hypothesis of~\cite{Naculich:2011ry} that
the soft graviton part of the scattering amplitude is described by a vacuum
expectation value of certain Wilson line operators, with one such operator
associated with each external line. This in itself is not so surprising: in
the eikonal approximation external particles do not recoil, and thus can only
be changed by a phase. If this phase is to have the correct gauge
transformation properties to slot into an amplitude, then some sort of Wilson
line operator is inevitable. In gravity, the Wilson line operator can only
depend upon $p_\mu\,p_\nu$ and $m^2\eta_{\mu\nu}$ due to Lorentz invariance.
It is therefore reassuring to have verified that this is indeed the case.
The fact that massive external particles give an additional contribution to
the Wilson line operator is also expected, given that $p_\mu$ and $m$ play
the role of charges in gravity. The exponent should consist of charges
multiplying factors which are invariant under rescalings of the external
momenta, which is indeed the case.
In this paper we have considered scalar particles emitting soft gravitons.
One could also consider the more phenomenologically relevant examples of
fermions, gauge bosons or hard gravitons emitting gravitational radiation.
This should make no difference to the eikonal approximation, which is known
to be insensitive to the spin of the emitting particles~\cite{Weinberg:1965nx}.
Beyond the eikonal approximation, and by analogy with the QED case examined
in~\cite{Laenen:2008gt,Laenen:2010uz} one expects NE corrections which are
insensitive to the spin of the emitting particle, together with additional
contributions which couple graviton emissions to the spin of the emitter.
For fermions, this would need an appropriate curved space formulation of
the second order fermion action~\cite{Morgan:1995te}, which was used
in~\cite{Laenen:2008gt} to analyse fermions in the path integral approach
to soft gauge boson emission.
The results presented here provide further insight into the relationships
between gauge theories and perturbative quantum gravity. They may also be
useful in phenomenological applications of gravitational physics e.g.
extending results obtained in the eikonal approximation for
transplanckian scattering~\cite{Stirling:2011mf,Giddings:2010pp}.
The investigation of such higher order effects in both gravity and gauge
theories is ongoing.
\acknowledgments
CDW thanks Jack Laiho and David Miller for
many useful conversations, and is also grateful to Einan Gardi and Lorenzo
Magnea for discussions, and comments on the manuscript.
He is supported by the STFC Postdoctoral Fellowship
``Collider Physics at the LHC''. We have used JaxoDraw~\cite{Binosi:2008ig,Binosi:2003yf}
throughout the paper.
|
\section{Introduction}
The search for a mechanism for electroweak symmetry breaking, and in
particular for a standard model (SM) Higgs boson, has been a major
goal of particle physics for many years, and is a central part of the
Fermilab Tevatron physics program, and also that of the Large Hadron Collider.
Recently, the ATLAS and CMS
collaborations at the Large Hadron Collider have released results on searches for the
Standard Model Higgs boson
decaying to $W^+W^-$~\cite{atlasww,cmsww}, $ZZ$~\cite{atlaszz}, $\gamma\gamma$~\cite{atlasgammagamma}
Both the CDF and D0 collaborations
have performed new combinations~\cite{cdfHWW,DZHiggs} of multiple
direct searches for the SM Higgs boson.
The new searches cover a larger data sample, and incorporate
improved analysis techniques compared to previous analyses.
The sensitivities of these new combinations significantly exceed those
of previous combinations~\cite{prevhiggs,WWPRLhiggs}.
In this note, we combine the most recent results of
searches for the SM Higgs boson produced in $p\bar{p}$~collisions at~$\sqrt{s}=1.96$~TeV. The analyses combined
here seek signals of Higgs bosons produced in association with
vector bosons ($q\bar{q}\rightarrow W/ZH$), through gluon-gluon
fusion ($gg\rightarrow H$), and through vector boson fusion (VBF)
($q\bar{q}\rightarrow q^{\prime}\bar{q}^{\prime}H$) corresponding
to integrated luminosities up to 7.1~fb$^{-1}$~at CDF and up to
8.2~fb$^{-1}$~at D0.
The searches included here target Higgs boson decays to $W^+W^-$, although
acceptance for decays into $\tau^+\tau^-$ and $\gamma \gamma $ are included in the D0 channels.
The searches are separated into
46 mutually exclusive final states (12 for CDF and 34 for D0;
see Tables~\ref{tab:cdfacc} and~\ref{tab:dzacc}) referred to
as ``analysis sub-channels'' in this note. The selection procedures for each
analysis are detailed in Refs.~\cite{cdfHWW} through~\cite{dzHgg},
and are briefly described below.
\section{Channels Included in the Combination}
\label{sec:channels}
Event selections are similar for the corresponding CDF and D0 analyses.
\smallskip
For the $H\rightarrow W^+ W^-$~analyses which seek events in which both $W$ bosons decay leptonically,
signal events are characterized by large {\mbox{$E\kern-0.57em\raise0.19ex\hbox{/}_{T}$}}~and
two oppositely-signed, isolated leptons. The presence of neutrinos in the
final state prevents the accurate reconstruction of the candidate Higgs boson mass.
D0 selects events containing large {\mbox{$E\kern-0.57em\raise0.19ex\hbox{/}_{T}$}}\ and electrons and/or muons, dividing the data sample
into three final states: $e^+e^-$, $e^\pm \mu^\mp$, and $\mu^+\mu^-$.
The searches in the three final states each use 8.1 fb$^{-1}$ of data which is subdivided according to
the number of jets in the event: 0, 1, or 2 or more (``2+'') jets.
Decays involving tau leptons are included in two orthogonal ways. For the first time, a dedicated
analysis ($\mu\tau_{had}$) using 7.3 fb$^{-1}$
of data studying the final state involving a muon and a hadronic tau decay, is included in the
Tevatron combination. Final states involving
other tau decays and mis-identified hadronic tau decays are included in the $e^+e^-$, $e^\pm \mu^\mp$,
and $\mu^+\mu^-$ final state analyses.
CDF separates the $H\rightarrow W^+ W^-$\ events in five non-overlapping samples, split into
both ``high $s/b$'' and ``low $s/b$'' categories based on lepton types and
different categories based on the number of reconstructed jets: 0, 1, or 2+
jets. The sample with two or more jets is not split into low $s/b$ and
high $s/b$ lepton categories due to low statistics. A sixth CDF channel
is the low dilepton mass ($m_{\ell^+\ell^-}$) channel, which accepts events
with $m_{\ell^+\ell^-}<16$~GeV. This channel increases the sensitivity of
the $H\rightarrow W^+W^-$ analyses at low $m_H$, adding 10\% additional
acceptance at $m_H=120$~GeV.
The division of events into jet categories allows the analysis discriminants
to separate three different categories of signals from the backgrounds more
effectively. The signal production mechanisms considered are $gg\rightarrow
H\rightarrow W^+W^-$, $WH+ZH\rightarrow jjW^+W^-$, and vector-boson fusion.
The D0 $e^+e^-$, $e^\pm \mu^\mp$, and $\mu^+\mu^-$ final state channels use
boosted decision tree outputs as the final discriminants while
the $\mu\tau_{had}$ channel uses neural networks.
CDF uses neural-network outputs, including likelihoods constructed from
calculated matrix-element probabilities as additional inputs for the 0-jet bin.
D0 includes $VH \rightarrow \ell^{\pm}\ell'^{\pm} + X$\ analyses in which the associated vector boson and the $W$ boson from the Higgs
boson decay that has the same charge are required to decay leptonically,
thereby defining three like-sign dilepton final states ($e^\pm e^\pm$, $e^\pm
\mu^\pm$, and $\mu^{\pm}\mu^{\pm}$). The combined output of two decision
trees, trained against the instrumental and diboson backgrounds respectively,
is used as the final discriminant. CDF also includes a separate analysis
of events with same-sign leptons and large {\mbox{$E\kern-0.57em\raise0.19ex\hbox{/}_{T}$}}\ to incorporate additional
potential signal from associated production events in which the two leptons
(one from the associated vector boson and one from a $W$ boson produced in the
Higgs decay) have the same charge. CDF for the first time also incorporates
three tri-lepton channels to include additional associated production
contributions where leptons result from the associated $W$ boson and the two
$W$ bosons produced in the Higgs decay or where an associated $Z$ boson decays
into a dilepton pair and a third lepton is produced in the decay of either
of the $W$ bosons resulting from the Higgs decay. In the latter case, CDF
separates the sample into 1 jet and 2+ jet sub-channels to fully take
advantage of the Higgs mass constraint available in the 2+ jet case where
all of the decay products are reconstructed.
CDF also includes opposite-sign channels in which one of the
two lepton candidates is a hadronic tau. Events are separated into $e$-$\tau$
and $\mu$-$\tau$ channels. The final discriminants are obtained from boosted
decision trees which incorporate both hadronic tau identification and kinematic
event variables as inputs.
D0 includes channels in which one of the $W$ bosons in the $H \rightarrow W^+W^-$
process decays leptonically and the other decays hadronically.
Electron and muon final states
are studied separately, each with 5.4 fb$^{-1}$ of data. Random forests
are used for the final discriminants.
D0 has updated its search for
Higgs boson production in which the Higgs decays into a
pair of photons to include 8.2 fb$^{-1}$ of data and to use a boosted decision tree as the final
discriminant.
The D0 analysis of the $\tau^+\tau^-+2$~jets final state
has also been updated to include additional data (4.3 fb$^{-1}$) and both muonic and electronic tau decays.
The output of boosted decision trees is used as the final discriminant.
For both CDF and D0, events from QCD multijet (instrumental) backgrounds
are measured in independent data samples using different methods.
For CDF, backgrounds from SM processes with electroweak gauge bosons or
top quarks were generated using \textsc{pythia}~\cite{pythia}, \textsc{alpgen}~\cite{alpgen},
\textsc{mc@nlo}~\cite{MC@NLO}, and \textsc{herwig}~\cite{herwig} programs.
For D0, these backgrounds were generated using \textsc{pythia},
\textsc{alpgen}, and \textsc{comphep}~\cite{comphep}, with \textsc{pythia}
providing parton-showering and hadronization for all the generators. These
background processes were normalized using either experimental data or
next-to-leading order calculations (including \textsc{mcfm}~\cite{mcfm} for
the $W+$ heavy flavor process).
Tables~\ref{tab:cdfacc} and~\ref{tab:dzacc} summarize, for CDF and D0 respectively,
the integrated luminosities, the Higgs boson mass ranges over which the searches are performed,
and references to further details for each analysis.
\begin{table}[h]
\caption{\label{tab:cdfacc} Luminosity, explored mass range and references
for the different processes
and final states ($\ell=e, \mu$) for the CDF analyses.
}
\begin{ruledtabular}
\begin{tabular}{lccc} \\
Channel & Luminosity (fb$^{-1}$) & $m_H$ range (GeV/$c^2$) & Reference \\ \hline
$H\rightarrow W^+ W^-$ \ \ \ 2$\times$(0,1 jets)+(2+ jets)+(low-$m_{\ell\ell}$)+($e$-$\tau_{had}$)+($\mu$-$\tau_{had}$) & 7.1 & 130-200 & \cite{cdfHWW} \\
$WH \rightarrow WW^+ W^-$ \ \ \ (same-sign leptons 1+ jets)+(tri-leptons) & 7.1 & 130-200 & \cite{cdfHWW} \\
$ZH \rightarrow ZW^+ W^-$ \ \ \ (tri-leptons 1 jet)+(tri-leptons 2+ jets) & 7.1 & 130-200 & \cite{cdfHWW} \\
\end{tabular}
\end{ruledtabular}
\end{table}
\vglue 0.5cm
\begin{table}[h]
\caption{\label{tab:dzacc} Luminosity, explored mass range and references
for the different processes
and final states ($\ell=e, \mu$) for the D0 analyses.
Most analyses are in addition analyzed separately for Run~IIa and IIb.
}
\begin{ruledtabular}
\begin{tabular}{lccc} \\
Channel & Luminosity (fb$^{-1}$) & $m_H$ range (GeV/$c^2$) & Reference \\ \hline
$H\rightarrow W^+ W^- \rightarrow \ell^\pm\nu \ell^\mp\nu$ \ \ \ (0,1,2+ jet) & 8.1 & 130-200 & \cite{dzHWW}\\
$H\rightarrow W^+ W^- \rightarrow \mu\nu \tau_{had}\nu$ \ \ \ & 7.3 & 130-200 & \cite{dzHWWtau}\\
$H\rightarrow W^+ W^- \rightarrow \ell\bar{\nu} jj$ & 5.4 & 130-200 & \cite{dzHWWjj}\\
$VH \rightarrow \ell^\pm \ell^\pm\ + X $ \ \ \ & 5.3 & 130-200 & \cite{dzWWW} \\
$H$+$X$$\rightarrow$$ \ell^\pm \tau^{\mp}_{had}jj$ \ \ \ & 4.3 & 130-200 & \cite{dzVHt2} \\
$H \rightarrow \gamma \gamma$ & 8.2 & 130-150 & \cite{dzHgg} \\
\end{tabular}
\end{ruledtabular}
\end{table}
\section{Signal Predictions and Uncertainties}
\label{sec:theory}
We normalize our Higgs boson signal predictions to the most recent highest-order calculations
available, for all production processes considered. The largest production cross section, $\sigma(gg\rightarrow H)$,
is calculated at next-to-next-to-leading order (NNLO) in QCD with soft gluon resummation to
next-to-next-to-leading-log (NNLL) accuracy,
and also includes
two-loop electroweak effects and handling of the running $b$ quark
mass~\cite{anastasiou,grazzinideflorian}. The numerical values in Table~\ref{tab:higgsxsec} are
updates~\cite{grazziniprivate} of these predictions with $m_t$ set to 173.1~GeV/$c^2$~\cite{tevtop10}, and an exact treatment
of the massive top and bottom loop corrections up to next-to-leading-order (NLO) + next-to-leading-log (NLL) accuracy.
The factorization and renormalization scale
choice for this calculation is $\mu_F=\mu_R=m_H$.
These calculations are
refinements of earlier NNLO calculations of the $gg\rightarrow H$
production cross section~\cite{harlanderkilgore2002,anastasioumelnikov2002,ravindran2003}.
Electroweak corrections were computed in Refs.~\cite{actis2008,aglietti}.
Soft gluon resummation was introduced in the prediction of the
$gg\rightarrow H$ production cross section in Ref.~\cite{catani2003}.
The $gg\rightarrow H$ production cross section depends strongly on
the gluon parton density function, and the accompanying value of
$\alpha_s(q^2)$. The cross sections used here are calculated
with the MSTW 2008 NNLO PDF set~\cite{mstw2008}, as recommended by the PDF4LHC working group~\cite{pdf4lhc}.
The reason to use this PDF parameterization is that it is the only NNLO prescription that results from
a fully global fit to all relevant data. We follow the PDF4LHC working group's prescription
to evaluate the uncertainties on the $gg\rightarrow H$ production cross section
due to the PDFs. This prescription is to evaluate the predictions of $\sigma(gg\rightarrow H)$ at NLO
using the global NLO PDF sets CTEQ6.6~\cite{cteq66}, MSTW08~\cite{mstw2008}, and NNPDF2.0~\cite{nnpdf}, and to
take the envelope of the predictions and their uncertainties due to PDF+$\alpha_s$ for the three sets as the uncertainty range
at NLO. The ratio of the NLO uncertainty range to that of just MSTW08 is then used to scale the
NNLO MSTW08 PDF+$\alpha_s$ uncertainty, to estimate a larger uncertainty at NNLO. This procedure roughly doubles the
PDF+$\alpha_s$ uncertainty from MSTW08 at NNLO alone. Other PDF sets, notably ABKM09~\cite{abkm09} and
HERAPDF1.0~\cite{herapdf}, have been proposed~\cite{bd,bagliocrit} as alternate PDF sets for use in setting
the uncertainties on $\sigma(gg\rightarrow H)$. These two PDF sets fit HERA data only in the case of
HERAPDF1.0, and HERA, fixed-target deep-inelastic scattering (DIS), fixed-target Drell-Yan (DY), and neutrino-nucleon di-muon data
in the case of ABKM09, and do not
include Tevatron jet data~\cite{cdfinclusivejet}\cite{d0inclusivejet}.
We note that HERA, fixed-target DIS, fixed-target DY, and neutrino-nucleon di-muon data do not probe directly
the gluon PDF at the scales most relevant for our analyses, but are most sensitive to the quark distributions.
The high-$E_T$ Tevatron jet data contain gluon-initiated
processes at the relevant momentum transfer scales. The PDF constraints from the HERA, fixed-target DIS,
fixed-target DY, and di-muon data
are extrapolated to make predictions of high-$x$ gluons at the relevant energy scales. The reliability of these
extrapolations is checked for the HERAPDF1.0 PDF set
by showing the NLO prediction~\cite{fastnlo} of the inclusive jet spectrum
as a function
of jet $E_T$ and $\eta$, as compared with measurements by CDF and D0 as shown
in Figures~\ref{fig:d0herapdf} and~\ref{fig:cdfherapdf},
respectively.
The measured inclusive jet cross sections are significantly under-predicted using this PDF set. The corresponding comparisons using the same
NLO jet cross section calculation with the MSTW08 PDF set are shown in Figures~49 and~50 in Ref.~\cite{mstw2008}. The
agreement with the jet data is much better with the MSTW2008 PDF set, which is expected, since those data are included
in the MSTW08 PDF fit.
Consequently the high-$x$ gluon component is larger in the MSTW2008 PDF set than in the
HERAPDF1.0 PDF set. This forms an important experimental foundation for the prediction of
the gluon-fusion Higgs boson production cross section.
Recently, it has been pointed out~\cite{abm2011} that the gluon PDF in the ABKM09 fit, and thus the values
of $\sigma(gg\rightarrow H)$, depend on the handling of the data input from the New Muon Collaboration (NMC),
specifically whether the differential
spectrum as measured by NMC is included in the fit or whether the structure function $F_2$ reported by NMC is
used in the fit. It is verified by the NNPDF Collaboration~\cite{nnpdfnmc} that the NNPDF2.1 global PDF fit
is negligibly impacted by changing the handling of the input NMC data, indicating that the global fit,
which includes Tevatron jet data, is more solid and less sensitive to the precise handling of NMC data. It should be noted
however that this study is performed only at NLO.
We also include uncertainties on $\sigma(gg\rightarrow H)$ due to uncalculated higher order processes by following the
standard procedure of varying the factorization renormalization scales up and down together by a factor $\kappa=2$.
It had been proposed to vary $\mu_R$ and $\mu_F$ independently~\cite{bd}, but the maximum impact on $\sigma(gg\rightarrow H)$
is found when varying the two scales together.
At a central scale choice of $\mu_R=\mu_F=m_H/2$, the fixed-order calculations at NLO and NNLO converge rapidly and
the scale uncertainties generously cover the differences. Furthermore, the NNLO calculation using a scale choice
of $\mu_R=\mu_F=m_H/2$ is very close to the NNLL+NNLO
calculation that we use. The authors of the NNLL+NNLO calculation
recommend that we use the scale choice $\mu_R=\mu_F=m_H$ for the central value~\cite{grazzinideflorian}.
The uncertainty on $\sigma(gg\rightarrow H)$ is evaluated by varying $\mu_R=\mu_F$ from a lower value of $m_H/2$ to
an upper value of $2m_H$, the customary variation, and by evaluating the NNLO+NNLL cross sections at these scales.
A larger scale variation has been proposed, varying the scale by a factor $\kappa=3$~\cite{bd}. The original justification
for this proposal~\cite{bd} was for the LO calculation to cover the NNLO central value. The
LO calculation however does not have enough scale-dependent terms for its scale uncertainty to be fairly considered -- it is missing
significant QCD radiation and other terms that increase the scale dependence at NLO compared with that at LO.
The second justification~\cite{bagliocrit}
is simply that the higher-order QCD corrections are larger than those of other processes, such as Drell-Yan production.
The corrections are large due to the
fact that there are gluons in the initial state which radiate more gluons more frequently than the quarks that
initiate other processes. The scale dependence of the calculations of $\sigma(gg\rightarrow H)$ which include
this radiation and other terms is correspondingly
larger, and the customary range of scale variation provides uncertainties which are commensurate with the values of the corrections.
An independent calculation by Ahrens, Becher, Neubert, and Yang~\cite{Ahrens} using a renormalization-group improved
resummation of radiative corrections at NNLL accuracy gives an alternative
calculation of $\sigma(gg\rightarrow H)$ that converges more rapidly and requires even smaller scale uncertainties.
The central values of this calculation of $\sigma(gg\rightarrow H)$ are in excellent agreement with those used in our analysis
and which are given in Table~\ref{tab:higgsxsec}.
Because the $H\rightarrow W^+W^-$ analyses separate the data into categories based on the numbers of observed jets,
we assess factorization and renormalization scale and PDF$+\alpha_s$ variations separately for each
jet category as evaluated in Ref.~\cite{anastasiouwebber}.
This calculation is at NNLO for $H$+0~jets, at NLO for $H$+1~jet, and at LO for $H$+2 or more jets. A newer,
more precise calculation~\cite{campbell2j}
of the $H$+2 or more jets cross section at NLO is used to evaluate the uncertainties in this category.
These scale uncertainties are used instead of the inclusive NNLL+NNLO scale uncertainty because we require
them in each jet category, and the uncertainties we use are significantly larger than the inclusive scale uncertainty.
The scale choice affects the $p_T$ spectrum of the Higgs boson when produced in gluon-gluon fusion, and this
effect changes the acceptance of the selection requirements and also the shapes of the distributions of the final
discriminants. The effect of the acceptance change is included in the calculations of Ref.~\cite{anastasiouwebber}
and Ref.~\cite{campbell2j}, as the experimental requirements are simulated in these calculations.
The effects on the
final discriminant shapes are obtained by reweighting the $p_T$ spectrum of the Higgs boson production in our
Monte Carlo simulation to higher-order calculations. The Monte Carlo signal simulation used by CDF and D0
is provided by the LO generator {\sc pythia}~\cite{pythia} which includes a parton shower and fragmentation and
hadronization models. We reweight the Higgs boson $p_T$ spectra in our {\sc pythia} Monte Carlo samples
to that predicted by {\sc hqt}~\cite{hqt} when making predictions of differential distributions of $gg\rightarrow H$ signal events.
To evaluate the impact of the scale uncertainty on our differential spectra, we use the {\sc resbos}~\cite{resbos} generator, and
apply the scale-dependent differences in the Higgs boson $p_T$ spectrum to the {\sc hqt} prediction, and propagate these to
our final discriminants as a systematic uncertainty on the shape, which is included in the calculation of the limits.
We treat the scale uncertainties
as 100\% correlated between jet categories and between CDF and D0, and
also treat the PDF+$\alpha_s$ uncertainties in the cross section as correlated
between jet categories and between CDF and D0. We treat however the PDF$+\alpha_s$ uncertainty
as uncorrelated with the scale uncertainty. The main reason is that the PDF uncertainty arises
from experimental uncertainties and PDF parameterization choices, while the scale uncertainty
arises from neglected higher-order terms in the perturbative cross section calculations.
The PDF predictions do depend on the scale choice, however.
The dependence of the prediction of $\sigma(gg\rightarrow H)$ due to the scale
via the PDF dependence is included as part of the scale uncertainty and not as part of the PDF
uncertainty, to ensure that all scale dependence is considered correlated.
Furthermore, we have verified~\cite{anastasiouprivate} that the fractional change in the prediction of
$\sigma(gg\rightarrow H)$ due to the PDF variation depends negligibly on the value of the
scale choice, justifying the treatment of PDF and scale uncertainties as uncorrelated. As described
in Section~\ref{sec:combination}, systematic uncertainties arising from uncorrelated sources should not
be added linearly together, but instead should be considered to fluctuate independently of one another.
We have investigated the use of a Gaussian prior for the scale uncertainty using the $\kappa=2$ variation
as the $\pm 1\sigma$ variation, and compared it with the use of a flat prior between the low edge and the
upper edge, and have found a negligible impact on our sensitivity and our observed limits. The main reason
for this is that while the flat prior has a higher density near the edges, the Gaussian prior has longer tails, and
both effects essentially compensate.
The use of a flat prior with edges for the scale uncertainty does not affect the choice of whether to add this
uncertainty linearly or not with the PDF uncertainty; that choice is determined by the correlations of the
two sources of uncertainty.
Another source of uncertainty in the prediction of $\sigma(gg\rightarrow H)$ is the extrapolation of the
QCD corrections computed for the heavy top quark loops to the light-quark loops included as part of the electroweak
corrections. Uncertainties at the level of 1-2\% are already included in the cross section values we
use~\cite{grazzinideflorian,anastasiou}. In Ref.~\cite{anastasiou}, it is argued that the factorization of QCD
corrections is known to work well for Higgs boson masses many times in excess of the masses of the loop particles.
A 4\% change in the predicted cross section is seen when all QCD corrections are removed from the diagrams
containing light-flavored quark loops, which is too conservative. For the $b$ quark loop, which is computed
separately in Ref.~\cite{anastasiou}, the QCD corrections are much smaller than for the top loop, further giving
confidence that it does not introduce large uncertainties.
\begin{figure}
\begin{center}
\includegraphics[width=\textwidth]{herapdf_d0experr.eps}
\caption{\label{fig:d0herapdf}
Inclusive jet cross section measurements by D0 as a
function of jet $p_T$ in six bins of jet rapidity $y$, divided by the theoretical prediction.
The data systematic uncertainties are displayed
by the full shaded band. The theoretical prediction is an
NLO perturbative QCD calculation~\protect\cite{fastnlo}, with renormalization and
factorization scales set to jet $p_T$ using the HERAPDF1.0 PDFs~\protect{\cite{herapdf}\cite{herapdffigs}}
and including nonperturbative corrections.
The PDF uncertainties are shown as hatched regions -- the blue hatching shows the total PDF uncertainty,
and the red hatching shows the portion of the total due to experimental uncertainties.
This figure is obtained using the same
data and methods described in Ref.~\protect{\cite{d0inclusivejet}}.}
\end{center}
\end{figure}
\begin{figure}
\begin{center}
\includegraphics[width=\textwidth]{herapdf_cdf_noexperr.eps}
\caption{\label{fig:cdfherapdf}
Inclusive jet cross section measurements by CDF as a
function of jet $p_T$ in six bins of jet rapidity $y$, divided by the theoretical prediction.
The theoretical prediction is an
NLO perturbative QCD calculation~\protect\cite{fastnlo}, with renormalization and
factorization scales set to jet $p_T$ using the HERAPDF1.0 PDFs~\protect{\cite{herapdf}\cite{herapdffigs}}
and including nonperturbative corrections.
The PDF uncertainties are shown as hatched regions -- the blue hatching shows the total PDF uncertainty,
and the red hatching shows the portion of the total due to experimental uncertainties. The
CDF data are published in Ref.~\protect{\cite{cdfinclusivejet}}. Experimental uncertainties on the CDF
data are comparable to those of the D0 data, shown in Fig.~\protect{\ref{fig:d0herapdf}}.
}
\end{center}
\end{figure}
We include all significant Higgs boson production modes in our searches.
Besides gluon-gluon fusion through virtual quark loops
(GGF), we include Higgs boson production in association with a $W$
or $Z$ vector boson, and vector boson fusion (VBF). We use the $WH$ and $ZH$
production cross sections computed
at NNLO in Ref.~\cite{djouadibaglio}. This calculation starts with the NLO calculation of
{\sc v2hv}~\cite{v2hv} and includes NNLO QCD contributions~\cite{vhnnloqcd}, as well
as one-loop electroweak corrections~\cite{vhewcorr}.
We use the VBF cross section computed at NNLO in QCD in Ref.~\cite{vbfnnlo}. Electroweak corrections
to the VBF production cross section are computed with the {\sc hawk} program~\cite{hawk}, and
are small and negative (-2\% to -3\%) for the Higgs boson mass range considered here. They are not
included in this update but will be incorporated in future results.
In order to predict the distributions of the kinematics of Higgs boson signal events, CDF
and D0 use the \textsc{pythia}~\cite{pythia} Monte Carlo program,
with \textsc{CTEQ5L} and \textsc{CTEQ6L}~\cite{cteq} leading-order
(LO) parton distribution functions. The Higgs boson decay branching
ratio predictions are calculated with \textsc{hdecay}~\cite{hdecay}, and are
also listed in Table~\ref{tab:higgsxsec}. We use \textsc{hdecay} Version 3.53.
While the $HWW$ coupling is well predicted, $Br(H\rightarrow W^+W^-)$ depends on the
partial widths of all other Higgs boson decays. The partial width $\Gamma(H\rightarrow b{\bar{b}})$ is sensitive
to $m_b$ and $\alpha_s$, $\Gamma(H\rightarrow c{\bar{c}})$ is sensitive to $m_c$ and $\alpha_s$, and $\Gamma(H\rightarrow gg)$
is sensitive to $\alpha_s$.
The impacts of these uncertainties on $Br(H\rightarrow W^+W^-)$ depend on $m_H$ due to the fact that
$Br(H\rightarrow b{\bar{b}})$, $Br(H\rightarrow c{\bar{c}})$, $Br(H\rightarrow gg)$ become very small for Higgs boson
masses above 160~GeV$/c^2$, while they have a larger impact for lower $m_H$. We use the uncertainties
on the branching fraction $Br(H\rightarrow W^+W^-)$ from Ref.~\cite{bagliodjouadilittlelhc}.
At $m_H=130$~GeV$/c^2$, for example, the $m_b$ variation gives a $^{-4.89}_{+1.70}\%$ relative variation
in $Br(H\rightarrow W^+W^-)$, $\alpha_s$ gives a $^{-1.02}_{+1.09}\%$ variation, and
$m_c$ gives a $^{-0.45}_{+0.51}\%$ variation. At $m_H=165$~GeV$/c^2$, all three of these
uncertainties are below 0.1\% and remain small for higher $m_H$.
\begin{table}
\begin{center}
\caption{
The production cross sections and decay branching fractions
for the SM Higgs boson assumed for the combination.}
\vspace{0.2cm}
\label{tab:higgsxsec}
\begin{ruledtabular}
\begin{tabular}{ccccccccc}
$m_H$ & $\sigma_{gg\rightarrow H}$ & $\sigma_{WH}$ & $\sigma_{ZH}$ & $\sigma_{VBF}$ & $B(H\rightarrow \tau^+{\tau^-})$ & $B(H\rightarrow W^+W^-)$ & $B(H\rightarrow ZZ)$ & $B(H\rightarrow \gamma\gamma)$ \\
(GeV/$c^2$) & (fb) & (fb) & (fb) & (fb) & (\%) & (\%) & (\%) & (\%) \\ \hline
130 & 842.9 & 112.00 & 68.5 & 62.1 & 5.305 & 29.43 & 3.858 & 0.2182 \\
135 & 750.8 & 97.20 & 60.0 & 57.5 & 4.400 & 39.10 & 5.319 & 0.2077 \\
140 & 670.6 & 84.60 & 52.7 & 53.2 & 3.472 & 49.16 & 6.715 & 0.1897 \\
145 & 600.6 & 73.70 & 46.3 & 49.4 & 2.585 & 59.15 & 7.771 & 0.1653 \\
150 & 539.1 & 64.40 & 40.8 & 45.8 & 1.778 & 68.91 & 8.143 & 0.1357 \\
155 & 484.0 & 56.20 & 35.9 & 42.4 & 1.057 & 78.92 & 7.297 & 0.09997 \\
160 & 432.3 & 48.50 & 31.4 & 39.4 & 0.403 & 90.48 & 4.185 & 0.05365 \\
165 & 383.7 & 43.60 & 28.4 & 36.6 & 0.140 & 95.91 & 2.216 & 0.02330 \\
170 & 344.0 & 38.50 & 25.3 & 34.0 & 0.093 & 96.39 & 2.351 & 0.01598 \\
175 & 309.7 & 34.00 & 22.5 & 31.6 & 0.073 & 95.81 & 3.204 & 0.01236 \\
180 & 279.2 & 30.10 & 20.0 & 29.4 & 0.059 & 93.25 & 5.937 & 0.01024 \\
185 & 252.1 & 26.90 & 17.9 & 27.3 & 0.046 & 84.50 & 14.86 & 0.008128 \\
190 & 228.0 & 24.00 & 16.1 & 25.4 & 0.038 & 78.70 & 20.77 & 0.006774 \\
195 & 207.2 & 21.40 & 14.4 & 23.7 & 0.033 & 75.88 & 23.66 & 0.005919 \\
200 & 189.1 & 19.10 & 13.0 & 22.0 & 0.029 & 74.26 & 25.33 & 0.005285 \\
\end{tabular}
\end{ruledtabular}
\end{center}
\end{table}
\section{Distributions of Candidates}
All analyses provide binned histograms of the final discriminant variables
for the signal and background predictions, itemized separately for each
source, and the observed data.
The number of channels combined is large, and the number of bins
in each channel is large. Therefore, the task of assembling
histograms and checking whether the expected and observed limits are
consistent with the input predictions and observed data is difficult.
We thus provide histograms that aggregate all channels' signal,
background, and data together. In order to preserve most of the
sensitivity gain that is achieved by binning the data
instead of collecting them all together and counting, we aggregate the
data and predictions in narrow bins of signal-to-background ratio,
$s/b$. Data with similar $s/b$ may be added together with no loss in
sensitivity, assuming similar systematic errors on the predictions.
The aggregate histograms do not show the effects of systematic
uncertainties, but instead compare the data with the central
predictions supplied by each analysis.
The range of $s/b$ is quite large in each analysis, and so
$\log_{10}(s/b)$ is chosen as the plotting variable. Plots of the
distributions of $\log_{10}(s/b)$ are shown for $m_H=150$ and 165~GeV/$c^2$ in Figure~\ref{fig:lnsb}. These
distributions can be integrated from the high-$s/b$ side downwards,
showing the sums of signal, background, and data for the most pure
portions of the selection of all channels added together. These
integrals can be seen in Figure~\ref{fig:integ}. The most significant
candidates are found in the bins with the highest $s/b$; an excess
in these bins relative to the background prediction drives the Higgs
boson cross section limit upwards, while a deficit drives it downwards.
The lower-$s/b$ bins show that the modeling of the rates and kinematic
distributions of the backgrounds is very good.
We also show the distributions of the data after subtracting the
expected background, and compare that with the expected signal yield
for a standard model Higgs boson, after collecting all bins in all
channels sorted by $s/b$. These background-subtracted distributions
are shown in Figure~\ref{fig:bgsub}. These graphs also show the
remaining uncertainty on the background prediction after fitting the
background model to the data within the systematic uncertainties on
the rates and shapes in each contributing channel's templates.
\begin{figure}[t]
\begin{centering}
\includegraphics[width=0.4\textwidth]{lnsbTEV150.eps}\includegraphics[width=0.4\textwidth]{lnsbTEV165.eps}
\caption{
\label{fig:lnsb} Distributions of $\log_{10}(s/b)$, for the data from all contributing channels from
CDF and D0, for Higgs boson masses of 150 and 165~GeV/$c^2$. The
data are shown with points, and the expected signal is shown stacked on top of
the backgrounds. Underflows and overflows are collected into the
bottom and top bins. }
\end{centering}
\end{figure}
\begin{figure}[t]
\begin{centering}
\includegraphics[width=0.4\textwidth]{tevinteg150.eps}\includegraphics[width=0.4\textwidth]{tevinteg165.eps}
\caption{
\label{fig:integ} Integrated distributions of $s/b$, starting at the high $s/b$ side, for Higgs boson
masses of 150 and 165~GeV/$c^2$. The total signal-plus-background
and background-only integrals are shown separately, along with the
data sums. Data are only shown for bins that have
data events in them.}
\end{centering}
\end{figure}
\begin{figure}[t]
\begin{centering}
\includegraphics[width=0.45\textwidth]{tevsm150bgsub.eps}\includegraphics[width=0.45\textwidth]{tevsm165bgsub.eps}
\caption{
\label{fig:bgsub} Background-subtracted data distributions for all channels, summed in bins of $s/b$,
for Higgs boson masses of 150, and 165~GeV/$c^2$. The background has been fit, within its systematic
uncertainties, to the data. The points with error bars indicate the background-subtracted data; the sizes of the
error bars are the square roots of the predicted background in each bin. The unshaded (blue-outline) histogram shows the systematic
uncertainty on the best-fit background model, and the shaded histogram shows the expected signal for a Standard
Model Higgs boson.}
\end{centering}
\end{figure}
\section{Combining Channels}
\label{sec:combination}
To gain confidence that the final result does not depend on the
details of the statistical formulation, we perform two types of
combinations, using Bayesian and Modified Frequentist approaches,
which yield limits on the Higgs boson production rate that agree
within 10\% at each value of $m_H$, and within 1\% on average. Both
methods rely on distributions in the final discriminants, and not just
on their single integrated values. Systematic uncertainties enter on
the predicted number of signal and background events as well as on the
distribution of the discriminants in each analysis (``shape
uncertainties''). Both methods use likelihood calculations based on
Poisson probabilities.
Both methods treat the systematic uncertainties in a Bayesian fashion,
assigning a prior distribution to each source of uncertainty,
parameterized by a nuisance parameter, and propagating the impacts of
varying each nuisance parameter to the signal and background
predictions, with all correlations included. A single nuisance
parameter may affect the signal and background predictions in many
bins of many different analyses. Independent nuisance parameters are
allowed to vary separately within their prior distributions.
Both methods use the data to constrain the values of the nuisance
parameters, one by integration, the other by fitting. These methods
reduce the impact of prior uncertainty in the nuisance parameters thus
improving the sensitivity. Because of these constraints to the data,
it is important to evaluate the uncertainties and correlations
properly, and to allow independent parameters to vary separately,
otherwise a fit may overconstrain a parameter and extrapolate its use
improperly. The impacts of correlated uncertainties add together
linearly on a particular prediction, while those of uncorrelated
uncertainties are convoluted together, which is similar to adding in
quadrature. Adding uncorrelated uncertainties linearly implies a
correlation which is not present and which may result in incorrect
results.
Both methods set limits at the 95\% confidence (or ``credibility'' in
the case of the Bayesian method). This is a probabalistic statement.
To be consistent and accurate, we give equal treatment to all sources
of uncertainty, both theoretical and experimental.
Ref.~\cite{bagliocrit} proposes testing the minimum possible
production cross section and quoting our exclusion limits based on
only this prediction. We do not claim that our exclusion limits are
independent of the model choices that are made, and hence we are not
required to only test the case in which all values of all uncertain
parameters fluctuate simultaneously to the values that produce the
weakest results. To do so would provide inconsistent results when
computing limits and when attempting to discover. A set of values for
the nuisance parameters which weakens the limit may strengthen a
discovery and vice versa, particularly those parameters that affect
the background rate and shape predictions. By always setting
uncertain parameters to their most extreme values, we could find we
have an excess of data over the background prediction when we set our
limits, and a deficit of data with respect to the background in the
same sample when attempting to make a discovery. The prescriptions
described below provide a consistent method for both tasks.
\subsection{Bayesian Method}
Because there is no experimental information on the production cross
section for the Higgs boson, in the Bayesian
technique~\cite{prevhiggs}\cite{pdgstats} we assign a flat prior for
the total number of selected Higgs events. For a given Higgs boson
mass, the combined likelihood is a product of likelihoods for the
individual channels, each of which is a product over histogram bins:
\begin{equation}
{\cal{L}}(R,{\vec{s}},{\vec{b}}|{\vec{n}},{\vec{\theta}})\times\pi({\vec{\theta}})
= \prod_{i=1}^{N_C}\prod_{j=1}^{N_b} \mu_{ij}^{n_{ij}} e^{-\mu_{ij}}/n_{ij}!
\times\prod_{k=1}^{n_{np}}e^{-\theta_k^2/2}
\end{equation}
\noindent where the first product is over the number of channels
($N_C$), and the second product is over $N_b$ histogram bins containing
$n_{ij}$ events, binned in ranges of the final discriminants used for
individual analyses, such as the dijet mass, neural-network outputs,
or matrix-element likelihoods.
The parameters that contribute to the
expected bin contents are $\mu_{ij} =R \times s_{ij}({\vec{\theta}}) + b_{ij}({\vec{\theta}})$
for the
channel $i$ and the histogram bin $j$, where $s_{ij}$ and $b_{ij}$
represent the expected background and signal in the bin, and $R$ is a scaling factor
applied to the signal to test the sensitivity level of the experiment.
Truncated Gaussian priors are used for each of the nuisance parameters
$\theta_k$, which define the sensitivity of the predicted signal and
background estimates to systematic uncertainties. These can take the
form of uncertainties on overall rates, as well as the shapes of the
distributions used for combination. These systematic uncertainties
can be far larger than the expected SM Higgs boson signal, and are
therefore important in the calculation of limits. The truncation is
applied so that no prediction of any signal or background in any bin
is negative. The posterior density function is then integrated over
all parameters (including correlations) except for $R$, and a 95\%
credibility level upper limit on $R$ is estimated by calculating the
value of $R$ that corresponds to 95\% of the area of the resulting
distribution.
\subsection{Modified Frequentist Method}
The Modified Frequentist technique relies on the ${\rm CL}_{\rm s}$
method, using a log-likelihood ratio (LLR) as test
statistic~\cite{DZHiggs}:
\begin{equation}
LLR = -2\ln\frac{p({\mathrm{data}}|H_1)}{p({\mathrm{data}}|H_0)},
\end{equation}
where $H_1$ denotes the test hypothesis, which admits the presence of
SM backgrounds and a Higgs boson signal, while $H_0$ is the null
hypothesis, for only SM backgrounds. The probabilities $p$ are
computed using the best-fit values of the nuisance parameters for each
pseudo-experiment, separately for each of the two hypotheses, and
include the Poisson probabilities of observing the data multiplied by
Gaussian priors for the values of the nuisance parameters. This
technique extends the LEP procedure~\cite{pdgstats} which does not
involve a fit, in order to yield better sensitivity when expected
signals are small and systematic uncertainties on backgrounds are
large~\cite{pflh}.
The ${\rm CL}_{\rm s}$ technique involves computing two $p$-values,
${\rm CL}_{\rm s+b}$ and ${\rm CL}_{\rm b}$. The latter is defined by
\begin{equation}
1-{\rm CL}_{\rm b} = p(LLR\le LLR_{\mathrm{obs}} | H_0),
\end{equation}
where $LLR_{\mathrm{obs}}$ is the value of the test statistic computed
for the data. $1-{\rm CL}_{\rm b}$ is the probability of observing a
signal-plus-background-like outcome without the presence of signal,
i.e. the probability that an upward fluctuation of the background
provides a signal-plus-background-like response as observed in data.
The other $p$-value is defined by
\begin{equation}
{\rm CL}_{\rm s+b} = p(LLR\ge LLR_{\mathrm{obs}} | H_1),
\end{equation}
and this corresponds to the probability of a downward fluctuation of the sum
of signal and background in
the data. A small value of ${\rm CL}_{\rm s+b}$ reflects inconsistency with $H_1$.
It is also possible to have a downward fluctuation in data even in the absence of
any signal, and a small value of ${\rm CL}_{\rm s+b}$ is possible even if the expected signal is
so small that it cannot be tested with the experiment. To minimize the possibility
of excluding a signal to which there is insufficient sensitivity
(an outcome expected 5\% of the time at the 95\% C.L., for full coverage),
we use the quantity ${\rm CL}_{\rm s}={\rm CL}_{\rm s+b}/{\rm CL}_{\rm b}$. If ${\rm CL}_{\rm s}<0.05$ for a particular choice
of $H_1$, that hypothesis is deemed to be excluded at the 95\% C.L. In an analogous
way, the expected ${\rm CL}_{\rm b}$, ${\rm CL}_{\rm s+b}$ and ${\rm CL}_{\rm s}$ values are computed from the median of the
LLR distribution for the background-only hypothesis.
Dividing CL$_{\rm s+b}$ by CL$_{\rm b}$ incurs a median penalty of a factor of two in the expected $p$-value,
as the distribution of CL$_{\rm b}$ is uniform under the null hypothesis. One may use CL$_{\rm s+b}$ directly
to set limits, and solve the problem of excluding non-testable models by introducing a
``power-constraint''. The power-constrained limit (PCL) method~\cite{atlasww,atlaszz,atlasgammagamma}
involves excluding models
for which CL$_{\rm s+b}<0.05$, but if the limit thus obtained makes an excursion below a previously chosen
lower bound, to quote the lower-bound limit instead. The ATLAS collaboration sets this lower bound at $-1\sigma$
in the distribution of limits expected from background-only outcomes. The PCL method thus retains
the desired coverage rate and does not exclude untestable models, and it
provides by construction stronger expected and observed limits than CL$_{\rm s}$. The expected and observed PCL
limits with Tevatron data may be estimated from Figure~\ref{fig:comboCLSB} and Table~\ref{tab:clsbVals} which show and list the observed and expected values
of CL$_{\rm s+b}$ as functions of $m_H$. Nowhere in the range is there more than a $1\sigma$ downward fluctuation
relative to the background prediction, and so a similar power constraint to the one ATLAS applies would not have
an effect on the observed limits. With the PCL method, the region of $m_H$ expected to be excluded at the 95\% C.L.
grows by $\sim$40\% compared to that obtained with the CL$_{\rm s}$ method.
The PCL method has a larger false exclusion rate than the CL$_{\rm s}$ and Bayesian
methods that we use to quote our results,
the distribution of possible limits under the null hypothesis is highly asymmetrical,
and this distribution depends on the arbitrary choice of the location of the power constraint.
We decided to continue quoting CL$_{\rm s}$ and Bayesian limits also because they provide a strong numerical cross-check
of each other, and they are comparable to those of LEP~\cite{lephiggs}, CMS~\cite{cmsww}, and our own
previous results~\cite{prevhiggs}.
Systematic uncertainties are included by fluctuating the predictions for
signal and background rates in each bin of each histogram in a correlated way when
generating the pseudo-experiments used to compute ${\rm CL}_{\rm s+b}$ and ${\rm CL}_{\rm b}$.
\subsection{Systematic Uncertainties}
\label{systematics}
Systematic uncertainties differ
between experiments and analyses, and they affect the rates and shapes of the predicted
signal and background in correlated ways. The combined results incorporate
the sensitivity of predictions to values of nuisance parameters,
and include correlations between rates and shapes, between signals and backgrounds,
and between channels within experiments and between experiments.
More discussion on this topic can be found in the
individual analysis notes~\cite{cdfHWW} through~\cite{dzHgg}. Here we
consider only the largest contributions and correlations between and
within the two experiments.
\subsubsection{Correlated Systematics between CDF and D0
The uncertainties on the measurements of the integrated luminosities are 6\%
(CDF) and 6.1\% (D0).
Of these values, 4\% arises from the uncertainty
on the inelastic $p\bar{p}$~scattering cross section, which is correlated
between CDF and D0.
CDF and D0 also share the assumed values and uncertainties on the production cross sections
for top-quark processes ($t\bar{t}$~and single top) and for electroweak processes
($WW$, $WZ$, and $ZZ$). In order to provide a consistent combination, the values of these
cross sections assumed in each analysis are brought into agreement. We use
$\sigma_{t\bar{t}}=7.04^{+0.24}_{-0.36}~{\rm (scale)}\pm 0.14{\rm (PDF)}\pm 0.30{\rm (mass)}$,
following the calculation of Moch and Uwer~\cite{mochuwer}, assuming
a top quark mass $m_t=173.0\pm 1.2$~GeV/$c^2$~\cite{tevtop10},
and using the MSTW2008nnlo PDF set~\cite{mstw2008}. Other
calculations of $\sigma_{t\bar{t}}$ are similar~\cite{otherttbar}.
For single top, we use the NLL $t$-channel calculation of Kidonakis~\cite{kid1},
which has been updated using the MSTW2008nnlo PDF set~\cite{mstw2008}~\cite{kidprivcomm}.
For the $s$-channel process we use~\cite{kid2}, again based on the MSTW2008nnlo PDF set.
Both of the cross section values below are the sum of the single $t$ and single ${\bar{t}}$
cross sections, and both assume $m_t=173\pm 1.2$ GeV.
\begin{equation}
\sigma_{t-{\rm{chan}}} = 2.10\pm 0.027 {\rm{(scale)}} \pm 0.18 {\rm{(PDF)}} \pm 0.045 {\rm{(mass)}} {\rm {pb}}.
\end{equation}
\begin{equation}
\sigma_{s-{\rm{chan}}} = 1.046\pm 0.006 {\rm{(scale)}} \pm 0.059~{\rm{(PDF)}} \pm 0.030~{\rm{(mass)}}~{\rm {pb}}.
\end{equation}
Other calculations of $\sigma_{\rm{SingleTop}}$ are
similar for our purposes~\cite{harris}.
mcfm~\cite{mcfm} has been used to compute the NLO cross sections for $WW$, $WZ$,
and $ZZ$ production~\cite{dibo}. Using a scale choice $\mu_0=M_V^2+p_T^2(V)$ and
the MSTW2008 PDF set~\cite{mstw2008}, the cross section for inclusive $W^+W^-$
production is
\begin{equation}
\sigma_{W^+W^-} = 11.34^{+0.56}_{-0.49}~{\rm{(scale)}}~^{+0.35}_{-0.28} {\rm(PDF)} {\rm{pb}}
\end{equation}
and the cross section for inclusive $W^\pm Z$ production is
\begin{equation}
\sigma_{W^\pm Z} = 3.22^{+0.20}_{-0.17}~{\rm{(scale)}}~^{+0.11}_{-0.08}~{\rm(PDF)}~{\rm{pb}}
\end{equation}
For the $Z$, leptonic decays are used in the definition of the cross section, and
we assume both $\gamma$ and $Z$ exchange. The $W^\pm Z$ cross section quoted above involves the requirement
$75\leq m_{\ell^+\ell^-}\leq 105$~GeV for the leptons from the neutral current
exchange. The same dilepton invariant mass requirement is applied to both
sets of leptons in determining the $ZZ$ cross section which is
\begin{equation}
\sigma_{ZZ} = 1.20^{+0.05}_{-0.04}~{\rm{(scale)}}~^{+0.04}_{-0.03}~{\rm(PDF)}~{\rm{pb}}
\end{equation}
For the diboson cross section calculations, we use $|\eta_{\ell}|<5$
for all calculations. Loosening this requirement to include all
leptons leads to $\sim$+0.4\% change in the predictions. Lowering the
factorization and renormalization scales by a factor of two increases
the cross section, and raising the scales by a factor of two decreases
the cross section. The PDF uncertainty has the same fractional impact
on the predicted cross section independent of the scale choice. All
PDF uncertainties are computed as the quadrature sum of the twenty
68\% C.L. eigenvectors provided with MSTW2008 (MSTW2008nlo68cl). We
furthermore constrain $\sigma_{W^+W^-}$ in the signal regions of the
channels in the process of setting our limits, either by integration
over the uncertain parameters, or by a direct fit, depending on the
method. Our posterior constraint on $\sigma_{W^+W^-}$ has a
fractional uncertainty of $\pm 4$\%, which includes the prior
theoretical constraint of $\pm 6$\%, indicating that the data and the
theoretical prediction are approximately equally constraining on
$\sigma_{W^+W^-}$.
In many analyses, the dominant background yields are calibrated with data control
samples. Since the methods of measuring the multijet (``QCD'') backgrounds differ
between CDF and D0, and even between analyses within the collaborations, there is
no correlation assumed between these rates. Similarly, the large uncertainties on
the background rates for $W$+heavy flavor (HF) and $Z$+heavy flavor are considered
at this time to be uncorrelated, as both CDF and D0 estimate these rates using data
control samples, but employ different techniques. The calibrations of fake leptons,
unvetoed $\gamma\rightarrow e^+e^-$ conversions, $b$-tag efficiencies and mistag
rates are performed by each collaboration using independent data samples and
methods, and are therefore also treated as uncorrelated.
\subsubsection{Correlated Systematic Uncertainties for CDF
The dominant systematic uncertainties for the CDF analyses are shown
in the Appendix
in Tables~\ref{tab:cdfsystww0}, \ref{tab:cdfsystww4}, and~\ref{tab:cdfsystww5}
for the $H \rightarrow W^+W^-\rightarrow \ell^{\prime \pm}\nu \ell^{\prime\mp}\nu$ channels,
in Table~\ref{tab:cdfsystwww} for the $WH \rightarrow WWW \rightarrow\ell^{\prime \pm}\ell^{\prime \pm}$ and
$WH\rightarrow WWW \rightarrow \ell^{\pm}\ell^{\prime \pm} \ell^{\prime \prime \mp}$ channels, and
in Table~\ref{tab:cdfsystzww} for the $ZH \rightarrow ZWW \rightarrow \ell^{\pm}\ell^{\mp} \ell^{\prime \pm}$ channels.
Each source induces a
correlated uncertainty across all CDF channels' signal and background
contributions which are sensitive to that source.
Shape dependencies of templates on jet energy scale
and ``FSR'') are taken into account in the analyses (see tables).
For $H\rightarrow W^+ W^-$, the largest uncertainties on signal acceptance originate from
Monte Carlo modeling. Uncertainties on background event rates vary
significantly for the different processes. The backgrounds with the
largest systematic uncertainties are in general quite small. Such
uncertainties are constrained by fits to the nuisance parameters, and
they do not affect the result significantly. Because the largest
background contributions are measured using data, these uncertainties
are treated as uncorrelated for the $H\rightarrow b\bar{b}$~channels. The differences in
the resulting limits when treating the remaining uncertainties as either
correlated or uncorrelated, is less than $5\%$.
\subsubsection{Correlated Systematic Uncertainties for D0
The dominant systematic uncertainties for the D0 analyses are shown in the Appendix, in
Tables~\ref{tab:d0systww}, \ref{tab:d0systwwtau}, \ref{tab:d0systwww}, \ref{tab:d0lvjj},
\ref{tab:d0sysVHtau}, and \ref{tab:d0systgg}. Each source induces a correlated uncertainty
across all D0 channels sensitive to that source. Wherever appropriate the impact of systematic
effects on both the rate and shape of the predicted signal and background is included.
For the $H\rightarrow W^+ W^-$ and $VH \rightarrow \ell^{\pm}\ell'^{\pm} + X$ analyses, a significant source
of uncertainty is the measured efficiencies for selecting leptons. Significant sources for
all analyses are the uncertainties on the luminosity and the cross sections for the simulated
backgrounds.
For analyses involving jets the determination of the jet energy scale, jet
resolution and
the multijet background contribution are significant sources of uncertainty.
All systematic uncertainties arising from the same source are taken to be
correlated among the different backgrounds and between signal and background.
\begin{figure}[t]
\begin{centering}
\includegraphics[width=14.0cm]{tevatronLLR_Mar7.eps}
\caption{
\label{fig:comboLLR} {
Distributions of the log-likelihood ratio (LLR) as a function of Higgs mass obtained with the ${\rm CL}_{\rm s}$ method for the combination of all CDF and D0 analyses.
}}
\end{centering}
\end{figure}
\vspace*{1cm}
\section{Combined Results}
Before extracting the combined limits we study the distributions of the
log-likelihood ratio (LLR) for different hypotheses, to quantify the expected
sensitivity across the mass range tested.
Figure~\ref{fig:comboLLR} displays the LLR distributions for the combined
analyses as functions of $m_{H}$. Included are the median of the LLR distributions for the
background-only hypothesis (LLR$_{b}$), the signal-plus-background
hypothesis (LLR$_{s+b}$), and the observed value for the data (LLR$_{\rm{obs}}$). The
shaded bands represent the one and two standard deviation ($\sigma$)
departures for LLR$_{b}$ centered on the median. Table~\ref{tab:llrVals} lists the observed
and expected LLR values shown in Figure~\ref{fig:comboLLR}.
\begin{table}[htpb]
\caption{\label{tab:llrVals}
Log-likelihood ratio (LLR) values for the combined CDF + D0 Higgs boson search obtained using the {\rm CL}$_{S}$ method.}
\begin{ruledtabular}
\begin{tabular}{lccccccc}
$m_{H}$ (GeV/$c^2$ & LLR$_{\rm{obs}}$ & LLR$_{S+B}^{\rm{med}}$ &
LLR$_{B}^{-2\sigma}$ & LLR$_{B}^{-1\sigma}$ & LLR$_{B}^{\rm{med}}$ & LLR$_{B}^{+1\sigma}$ & LLR$_{B}^{+2\sigma}$ \\
\hline
130 & 0.01 & -1.07 & 4.62 & 2.73 & 0.78 & -1.18 & -3.08 \\
135 & 0.52 & -1.62 & 5.92 & 3.58 & 1.23 & -1.27 & -3.77 \\
140 & 1.39 & -2.12 & 7.17 & 4.53 & 1.77 & -1.07 & -4.22 \\
145 & 1.71 & -2.77 & 8.47 & 5.42 & 2.33 & -0.93 & -4.28 \\
150 & 2.95 & -3.67 & 10.18 & 6.78 & 3.23 & -0.53 & -4.53 \\
155 & 0.89 & -5.17 & 12.57 & 8.53 & 4.42 & 0.23 & -4.33 \\
160 & 6.96 & -9.22 & 17.27 & 12.38 & 7.42 & 2.17 & -3.48 \\
165 & 10.93 & -10.57 & 18.07 & 13.22 & 8.03 & 2.62 & -3.52 \\
170 & 5.90 & -7.62 & 15.53 & 10.93 & 6.22 & 1.23 & -3.98 \\
175 & 2.33 & -5.88 & 12.97 & 8.97 & 4.88 & 0.42 & -4.47 \\
180 & 3.78 & -4.08 & 10.68 & 7.17 & 3.42 & -0.38 & -4.22 \\
185 & 1.26 & -2.67 & 8.07 & 5.22 & 2.23 & -0.93 & -4.08 \\
190 & -0.56 & -1.93 & 6.53 & 4.08 & 1.52 & -1.07 & -3.88 \\
195 & -0.08 & -1.52 & 6.17 & 3.52 & 1.18 & -1.12 & -3.58 \\
200 & -1.30 & -1.07 & 4.67 & 2.77 & 0.82 & -1.18 & -3.23 \\
\end{tabular}
\end{ruledtabular}
\end{table}
These distributions can be interpreted as follows: The separation
between the medians of the LLR$_{b}$ and LLR$_{s+b}$ distributions
provides a measure of the discriminating power of the search. The
sizes of the one- and two-$\sigma$ LLR$_{b}$ bands indicate the width
of the LLR$_{b}$ distribution, assuming no signal is truly present and
only statistical fluctuations and systematic effects are present. The
value of LLR$_{\rm{obs}}$ relative to LLR$_{s+b}$ and LLR$_{b}$
indicates whether the data distribution appears to resemble what we
expect if a signal is present (i.e. closer to the LLR$_{s+b}$
distribution, which is negative by construction) or whether it
resembles the background expectation more closely; the significance of
any departures of LLR$_{\rm{obs}}$ from LLR$_{b}$ can be evaluated by
the width of the LLR$_{b}$ bands.
Using the combination procedures outlined in Section III, we extract
limits on SM Higgs boson production $\sigma \times B(H\rightarrow X)$
in $p\bar{p}$~collisions at $\sqrt{s}=1.96$~TeV for $130\leq m_H \leq 200$
GeV/$c^2$. To facilitate comparisons with the SM and to
accommodate analyses with different degrees of sensitivity, we present
our results in terms of the ratio of obtained limits to the SM Higgs
boson production cross section, as a function of Higgs boson mass, for
test masses for which both experiments have performed dedicated
searches in different channels. A value of the combined limit ratio
which is equal to or less than one indicates that that particular
Higgs boson mass is excluded at the 95\% C.L. A value less than one indicates
that a Higgs boson of that mass is excluded with a smaller cross section than
the SM prediction, and that the SM prediction is excluded with more
certainty than 95\% C.L.
The combinations of results~\cite{cdfHWW,DZHiggs} of each single
experiment, as used in this Tevatron combination, yield the following
ratios of 95\% C.L. observed (expected) limits to the SM cross
section: 0.92~(0.93) for CDF and 0.75~(0.92) for D0 at
$m_{H}=165$~GeV/$c^2$. Both collaborations independently exclude
$m_{H}=165$~GeV/$c^2$ at the 95\% C.L.
The ratios of the 95\% C.L. expected and observed limit to the SM
cross section are shown in Figure~\ref{fig:comboRatio} for the
combined CDF and D0 analyses. The observed and median expected ratios
are listed for the tested Higgs boson masses in Table~\ref{tab:ratios}
for $m_{H}
\leq 150$~GeV/$c^2$, and in Table~\ref{tab:ratios-3} for $m_{H} \geq
155$~GeV/$c^2$, as obtained by the Bayesian and the ${\rm CL}_{\rm s}$
methods. In the following summary we quote only the limits obtained
with the Bayesian method, which was chosen {\it a priori}. It
turns out that the Bayesian limits are slightly less stringent. The
corresponding limits and expected limits obtained using the ${\rm
CL}_{\rm s}$ method are shown alongside the Bayesian limits in the
tables. We obtain the observed (expected) values of 1.31~(0.92) at
$m_{H}=155$~GeV/$c^2$, 0.54~(0.65) at $m_{H}=165$~GeV/$c^2$,
1.13~(0.85) at $m_{H}=175$~GeV/$c^2$ and 1.49~(1.30) at
$m_{H}=185$~GeV/$c^2$.
\begin{figure}[hb]
\begin{centering}
\includegraphics[width=16.5cm]{tevsmh7mar2011.eps}
\caption{
\label{fig:comboRatio}
Observed and expected (median, for the background-only hypothesis)
95\% C.L. upper limits on the ratios to the SM cross section, as
functions of the Higgs boson mass for the combined CDF and D0
analyses. The limits are expressed as a multiple of the SM prediction
for test masses (every 5 GeV/$c^2$) for which both experiments have
performed dedicated searches in different channels. The points are
joined by straight lines for better readability. The bands indicate
the 68\% and 95\% probability regions where the limits can fluctuate,
in the absence of signal. The limits displayed in this figure are
obtained with the Bayesian calculation. }
\end{centering}
\end{figure}
\begin{table}[ht]
\caption{\label{tab:ratios} Ratios of median expected and observed 95\% C.L.
limit to the SM cross section for the combined CDF and D0 analyses as a function
of the Higgs boson mass in GeV/$c^2$, obtained with the Bayesian and with the ${\rm CL}_{\rm s}$ method.}
\begin{ruledtabular}
\begin{tabular}{lccccc}\\
Bayesian & 130 & 135 & 140 & 145 & 150 \\ \hline
Expected & 2.18 & 1.72 & 1.46 & 1.31 & 1.09 \\
Observed & 2.69 & 2.05 & 1.65 & 1.42 & 1.18 \\
\hline
\hline\\
${\rm CL}_{\rm s}$ & 130 & 135 & 140 & 145 & 150 \\ \hline
Expected: &2.14 &1.72 &1.49 &1.29 &1.14 \\
Observed: &2.57 &1.98 &1.60 &1.42 &1.16 \\
\end{tabular}
\end{ruledtabular}
\end{table}
\begin{table}[ht]
\caption{\label{tab:ratios-3}
Ratios of median expected and observed 95\% C.L.
limit to the SM cross section for the combined CDF and D0 analyses as a function
of the Higgs boson mass in GeV/$c^2$, obtained with the Bayesian and with the ${\rm CL}_{\rm s}$ method.}
\begin{ruledtabular}
\begin{tabular}{lccccccccccc}
Bayesian & 155 & 160 & 165 & 170 & 175 & 180 & 185 & 190 & 195 & 200 \\ \hline
Expected & 0.92 & 0.68 & 0.65 & 0.75 & 0.85 & 1.06 & 1.30 & 1.59 & 1.83 & 2.13 \\
Observed & 1.31 & 0.72 & 0.54 & 0.82 & 1.13 & 1.04 & 1.49 & 2.13 & 2.20 & 3.22 \\
\hline
\hline\\
${\rm CL}_{\rm s}$ & 155 & 160 & 165 & 170 & 175 & 180 & 185 & 190 & 195 & 200 \\ \hline
Expected & 0.92 &0.68 &0.64 &0.77 &0.87 &1.05 &1.32 &1.60 &1.82 &2.09 \\
Observed & 1.28 &0.70 &0.52 &0.80 &1.09 &1.03 &1.49 &2.13 &2.22 &3.13 \\
\end{tabular}
\end{ruledtabular}
\end{table}
We show in Figure~\ref{fig:comboCLS} and list in
Table~\ref{tab:clsVals} the observed 1-${\rm CL}_{\rm s}$ and its
expected distribution for the background-only hypothesis as a function
of the Higgs boson mass. This is directly interpreted as the level of
exclusion of our search using the ${\rm CL}_{\rm s}$ method. The
region excluded at the 95\% C.L. agrees very well with that obtained
via the Bayesian calculation.
In addition, we provide in Figure~\ref{fig:comboCLSB} (and listed in
Table~\ref{tab:clsbVals}) the values for the observed 1-${\rm CL}_{\rm s+b}$
and its expected distribution as a function of $m_H$. The value
${\rm CL}_{\rm s+b}$ is the $p$-value for the signal-plus-background
hypothesis. These values can be used to obtain alternative
upper limits that are more constraining compared to those obtained via the ${\rm CL}_{\rm s}$
method. In such a formulation, the power of the search is limited
at a level chosen {\it a priori} to avoid setting limits when the
background model grossly overpredicts the data or the data exhibit a
large background-like fluctuation ({\it e.g.}, limit at the -1$\sigma$
background fluctuation level.). Within Figure~\ref{fig:comboCLSB}, 95\%
C.L. power-constrained limits can be found at the points for which
1-${\rm CL}_{\rm s+b}$ exceeds 95\%. The expected range of exclusion is
$\sim$40\% larger using PCL than the Bayesian and CL$_{\rm s}$ limits quoted
here. We continue our convention of quoting Bayesian and CL$_{\rm s}$ limits
however.
In summary, we combine CDF and D0 results on SM Higgs boson searches,
based on luminosities up to 8.2 fb$^{-1}$.
Compared to our previous combination, more data have been added to the existing
channels, additional channels have been included, and analyses have been further
optimized to gain sensitivity. We use the recommendation of the PDF4LHC
working group for the central value of the parton distribution functions and uncertainties~\cite{pdf4lhc}.
We use the highest-order calculations of $gg \rightarrow H$, $WH$, $ZH$, and VBF theoretical cross
sections when comparing our limits to the SM predictions at high mass. We include consensus estimates
of the theoretical uncertainties on these production cross sections and the decay
branching fractions in the computations of our limits.
The 95\% C.L. upper limit on Higgs boson production is a factor of 0.54
times the SM cross section for a Higgs boson mass of $m_{H}=$165~GeV/$c^2$.
Based on simulation, the corresponding median expected upper limit is 0.65 times the SM cross section.
Standard Model branching ratios, calculated as functions of the Higgs boson mass, are assumed.
We choose to use the intersections of piecewise linear interpolations of our observed and expected
rate limits in order to quote ranges of Higgs boson masses that are excluded and that are expected
to be excluded. The sensitivities of our searches to Higgs bosons are smooth functions of the Higgs
boson mass and depend most strongly on the predicted cross sections and the decay branching ratios
(the decay $H\rightarrow W^+W^-$ is the dominant decay for the region of highest sensitivity). The
mass resolution of the channels is poor due to the presence of two highly energetic neutrinos in
signal events. We therefore use the linear interpolations to extend the results from the 5~GeV/$c^2$
mass grid investigated to points in between. This procedure yields higher expected and observed
interpolated rate limits than if the full dependence of the cross section and branching ratio were
included as well, since the latter produces limit curves that are concave upwards.
We exclude in this way the region $158<m_{H}<173$~GeV/$c^{2}$ at the the 95\% C.L.
The expected exclusion region, given the current sensitivity, is
$153<m_{H}<179$~GeV/$c^{2}$. The excluded region obtained by finding the intersections of the
linear interpolations of the observed $1-{\rm CL}_{\rm s}$ curve shown in Figure~\ref{fig:comboCLS} is
slightly larger than that obtained with the Bayesian calculation. As previously stated, and following
the procedure used in previous combinations~\cite{prevhiggs}, we make
the {\it a priori} choice to quote the exclusion region using the Bayesian calculation.
The results presented in this paper significantly extend the individual limits of each
collaboration and those obtained in our previous combination. The sensitivity of our combined search is
sufficient to exclude a Higgs boson at high mass and is expected to grow substantially
in the future as more data are added and further improvements are made to the analysis
techniques.
\begin{figure}[t]
\begin{centering}
\includegraphics[width=14.0cm]{tevatronCLS_Mar7.eps}
\caption{
\label{fig:comboCLS}
The exclusion strength 1-${\rm CL}_{\rm s}$ as a function of the Higgs boson mass
(in steps of 5 GeV/$c^2$), as obtained with ${\rm CL}_{\rm s}$ method
for the combination of the
CDF and D0 analyses. }
\end{centering}
\end{figure}
\begin{table}[htpb]
\caption{\label{tab:clsVals}
The observed and expected 1-{\rm CL}$_{\rm s}$ values as functions of $m_H$, for the combined
CDF and D0 Higgs boson searches.}
\begin{ruledtabular}
\begin{tabular}{lcccccc}
$m_H$ (GeV/$c^2$) & 1-{\rm CL}$_{\rm s}^{\rm{obs}}$ &
1-{\rm CL}$_{\rm s}^{-2\sigma}$ &
1-{\rm CL}$_{\rm s}^{-1\sigma}$ &
1-{\rm CL}$_{\rm s}^{\rm{median}}$ &
1-{\rm CL}$_{\rm s}^{+1\sigma}$ &
1-{\rm CL}$_{\rm s}^{+2\sigma}$ \\ \hline
130 & 0.530 & 0.929 & 0.816 & 0.638 & 0.366 & 0.130\\
135 & 0.654 & 0.925 & 0.888 & 0.729 & 0.466 & 0.167\\
140 & 0.790 & 0.923 & 0.919 & 0.811 & 0.549 & 0.233\\
145 & 0.840 & 0.956 & 0.950 & 0.865 & 0.636 & 0.283\\
150 & 0.909 & 0.985 & 0.980 & 0.916 & 0.737 & 0.398\\
155 & 0.864 & 0.987 & 0.984 & 0.951 & 0.834 & 0.545\\
160 & 0.991 & 0.995 & 0.992 & 0.990 & 0.952 & 0.795\\
165 & 0.998 & 1.000 & 0.999 & 0.995 & 0.967 & 0.839\\
170 & 0.982 & 0.997 & 0.994 & 0.987 & 0.923 & 0.727\\
175 & 0.933 & 0.999 & 0.995 & 0.970 & 0.876 & 0.614\\
180 & 0.941 & 0.989 & 0.980 & 0.929 & 0.785 & 0.476\\
185 & 0.820 & 0.947 & 0.943 & 0.866 & 0.650 & 0.311\\
190 & 0.598 & 0.918 & 0.914 & 0.775 & 0.519 & 0.227\\
195 & 0.602 & 0.931 & 0.899 & 0.751 & 0.469 & 0.203\\
200 & 0.370 & 0.925 & 0.823 & 0.647 & 0.387 & 0.121\\
\end{tabular}
\end{ruledtabular}
\end{table}
\begin{figure}[t]
\begin{centering}
\includegraphics[width=14.0cm]{tevatronCLSB_Mar7.eps}
\caption{
\label{fig:comboCLSB}
The signal exclusion $p$-value 1-${\rm CL}_{\rm s+b}$ as a function of the Higgs boson mass
(in steps of 5 GeV/$c^2$) for the combination of the
CDF and D0 analyses. }
\end{centering}
\end{figure}
\begin{table}[htpb]
\caption{\label{tab:clsbVals}
The observed and expected 1-{\rm CL}$_{\rm s+b}$ values as functions of $m_H$, for the combined CDF and D0 Higgs boson searches.}
\begin{ruledtabular}
\begin{tabular}{lcccccc}
$m_H$ (GeV/$c^2$) & 1-{\rm CL}$_{\rm s+b}^{\rm{obs}}$ &
1-{\rm CL}$_{\rm s+b}^{-2\sigma}$ &
1-{\rm CL}$_{\rm s+b}^{-1\sigma}$ &
1-{\rm CL}$_{\rm s+b}^{\rm{median}}$ &
1-{\rm CL}$_{\rm s+b}^{+1\sigma}$ &
1-{\rm CL}$_{\rm s+b}^{+2\sigma}$ \\ \hline
130 & 0.69856 & 0.99833 & 0.97078 & 0.81544 & 0.47078 & 0.16344\\
135 & 0.79378 & 0.99800 & 0.97978 & 0.86800 & 0.55056 & 0.20189\\
140 & 0.88293 & 0.99840 & 0.99000 & 0.91040 & 0.63680 & 0.23773\\
145 & 0.90322 & 0.99922 & 0.99189 & 0.92922 & 0.70000 & 0.32522\\
150 & 0.95089 & 0.99922 & 0.99533 & 0.95744 & 0.78011 & 0.40944\\
155 & 0.89711 & 0.99956 & 0.99811 & 0.97967 & 0.86833 & 0.56667\\
160 & 0.99551 & 0.99956 & 0.99878 & 0.99533 & 0.96289 & 0.81144\\
165 & 0.99951 & 0.99993 & 0.99993 & 0.99746 & 0.97289 & 0.83888\\
170 & 0.98971 & 0.99956 & 0.99878 & 0.99022 & 0.93189 & 0.72667\\
175 & 0.94778 & 0.99967 & 0.99822 & 0.98444 & 0.89267 & 0.60356\\
180 & 0.97163 & 0.99928 & 0.99763 & 0.96493 & 0.80864 & 0.48246\\
185 & 0.88767 & 0.99767 & 0.99256 & 0.93656 & 0.70467 & 0.34278\\
190 & 0.68256 & 0.99844 & 0.98589 & 0.89300 & 0.61711 & 0.24367\\
195 & 0.71978 & 0.99700 & 0.98100 & 0.86678 & 0.55878 & 0.20411\\
200 & 0.45800 & 0.99778 & 0.97289 & 0.82689 & 0.47867 & 0.15800\\
\end{tabular}
\end{ruledtabular}
\end{table}
\clearpage
|
\section{Introduction}
\label{intro}
One of the primary challenges faced by liquid
xenon WIMP dark matter experiments
is the presence of trace amounts of
radioactive krypton.
Xenon itself has no long-lived
radioactive isotopes which might
act as background sources, but krypton
includes the troublesome anthropogenic isotope
$^{85}$Kr, a beta emitter with a Q value of 687 keV and a
half-life of 10.76 years. $^{85}$Kr is created
in nuclear power plants and released
into the earth's atmosphere during
fuel reprocessing,
and its isotopic fraction at present
is about $2 \times 10^{-11}$ mol/mol
($^{85}$Kr/$^{nat}$Kr)\cite{kr85_review}.
Xenon, on the other hand, is extracted from the
atmosphere with a residual krypton concentration
typically ranging from $10^{-9}$
to $10^{-6}$ mol/mol
($^{nat}$Kr/Xe).
Although this implies that the absolute
concentration of $^{85}$Kr in xenon is rather small,
the $^{85}$Kr beta decay is nevertheless
highly problematic for dark matter experiments because
these decays are not suppressed by self-shielding
and because krypton cannot be separated from xenon
with conventional chemical purifiers.
The acceptable krypton concentration for a particular experiment is determined by its design sensitivity and by its nuclear recoil discrimination factor. As an example, the LUX dark matter experiment, a dual phase liquid xenon TPC with a recoil discrimination factor of 99.5\%, requires that the residual krypton concentration of the xenon target material be no more than $\sim 3 \times 10^{-12}$ mol/mol ($^{nat}$Kr/Xe)\footnote{All concentrations in this article refer to the natural krypton to xenon ratio, measured in units of mol/mol, unless otherwise indicated.}\cite{lux,lux-simon} in order to be sensitive to a 100 GeV WIMP with a cross section as small as $7 \times 10^{-46}$ cm$^2$. Other liquid xenon detectors, such as XMASS\cite{xmass}, and XENON100\cite{xenon100_first_results, xenon100_backgrounds}, also have demanding krypton goals, and future upgrades of these experiments will require reducing the krypton concentration even further.
To achieve these ultra-low krypton concentrations, commercially procured xenon must undergo additional processing via distillation or gas chromatography. Once this processing is complete, the residual krypton content of the xenon can be determined by low background counting of the $^{85}$Kr beta decays\cite{distillation,xenon100_first_results}, by chromatography\cite{chromotography}, or by atmospheric pressure ionization mass spectroscopy (API-MS)\cite{distillation}\footnote{It has also been suggested that atomic trap trace analysis may be sensitive to krypton at the level of $3 \times 10^{-14}$ mol/mol, but this method has not yet been demonstrated for this application\cite{atta}.}. These methods have achieved a sensitivity of about $\sim 10^{-12}$ mol/mol. Krypton monitoring is useful because it can confirm that the processing has been successful prior to full detector operations, and because it can constrain the background count rate due to $^{85}$Kr in the WIMP search data.
In this article we show that very small krypton
concentrations can be observed in xenon gas using
a mass spectrometry technique which we previously
developed to detect electronegative impurities in
xenon\cite{coldtrap}. The method is
inexpensive, highly sensitive, and it could be quickly adopted and
applied by many working dark matter experiments.
\section{Xenon cold trap mass spectrometry}
We use a residual gas analyzer (RGA)
mass spectrometer to analyze
our xenon by introducing a
small quantity of the gas into the
RGA's vacuum enclosure through a leak valve.
Since the partial pressure of each
component species is proportional to both
its absolute concentration
and to the flow rate through the analysis system,
by controlling for the flow rate
the partial pressures can be interpreted in terms
of the absolute concentrations. The measurement
is calibrated by preparing
samples of xenon gas with known impurity concentrations for
the various species of interest by directly mixing known quantities
of impurities with a known amount of xenon.
The flow rate can be controlled either by
measuring the actual flow rate in real time or by
using a standard leak-valve setting whose flow rate
was previously calibrated.
Since the partial pressures of all species
are proportional to the flow rate, the RGA signals
can be vastly increased simply by opening the leak valve further.
However, the RGA cannot be operated above some maximum total
pressure, typically about $10^{-5}$ Torr,
and the total pressure is dominated
by the xenon present in the gas sample.
This limits the maximum flow rate that can be used.
For example, if krypton can be detected by the RGA
at a partial pressure of $\sim 10^{-12}$ Torr, and
the xenon pressure is $10^{-5}$ Torr, then
the limit of detection is about one part in $10^7$.
Since we are interested in krypton
concentrations at the level of
$10^{-12}$ mol/mol, this is
inadequate for our purposes.
We solve the saturation problem simply
by removing most of the xenon from
the gas sample with a liquid nitrogen cold trap placed
between the leak valve and the RGA which allows the flow rate to be vastly
increased without saturating the RGA. For the bulk xenon, the pressure is adequate for xenon ice to form ($>$1.8 mTorr at 77 K) \cite{xeice}. So the xenon pressure is held fixed at its vapor pressure. For many common species however, the partial pressure is below the solid-vapor or liquid-vapor equilibrium. This prevents the impurities from becoming trapped.
In our previous paper, we showed that
impurity species such as oxygen, nitrogen,
and methane pass through the cold trap in large quantities,
and that their partial pressures, corrected
for flow rate, remain proportional to their absolute
concentrations. The sensitivity to oxygen, nitrogen,
and methane was found to be $0.66 \times 10^{-9}$,
$9.4 \times 10^{-9}$, and $0.49 \times 10^{-9}$
(mol/mol), respectively \cite{coldtrap}.
Here we extend the technique to observe krypton
in xenon. We expect that
krypton could be observed in very small quantities
by the RGA because there are very
few background species which could obscure the
krypton signal. In fact, we find that we are able
to detect krypton in xenon at a concentration of
$0.5 \times 10^{-12}$ mol/mol,
which makes this technique
better than or comparable to existing methods,
and sensitive enough to be useful for working
dark matter experiments.
\section{Apparatus and procedures}
\begin{figure}[t!]\centering
\includegraphics[width=100mm]{Fig1.eps}
\caption{Diagram of the xenon handling system. The flow pattern shown is typical of a measurement described in section 5.}
\label{fig:diagram}
\end{figure}
\begin{figure}[t!]\centering
\includegraphics[width=100mm]{Fig2.eps}
\caption{Diagram of the cold trap mass spectrometer analysis system.}
\label{fig:coldtrap_diagram}
\end{figure}
A diagram of our analysis apparatus and our xenon handling system
is shown in Figure \ref{fig:diagram} and Figure \ref{fig:coldtrap_diagram}.
The xenon of interest is admitted
into the analysis system through an
ultra-high vacuum leak valve (Kurt
Lesker part number VZLVM940R). It
passes through a liquid nitrogen cold trap
and a section of low-conductance
plumbing before reaching an SRS RGA200
mass spectrometer.
The low-conductance plumbing is necessary
to reduce the xenon partial
pressure from $1.8 \times 10^{-3}$ Torr
(its vapor pressure at liquid nitrogen
temperature)
to $< 10^{-5}$ Torr. This
ensures that the RGA, with the electron multiplier on, remains
unsaturated. We use a fully open
hand valve for the low-conductance element.
The cold trap is constructed from 1.5''
OD stainless-steel tubing, welded into
a U shape with a radius of 2.5''
and 12.5'' linear inlet and outlet
legs\footnote{We have also constructed
working cold traps from standard vacuum
plumbing components with 2.75"
CF flanges.}.
The tubing diameter is chosen to allow
a significant amount of xenon to be analyzed
before the growth of xenon ice blocks
the flow of gas through the analysis
system. We find that the liquid nitrogen
level must be high enough to submerse
the bottom of the cold trap U, but otherwise
its level is not critical.
We use several procedures to calibrate and monitor
the flow rate through the leak valve. First,
we calibrate the flow rate directly for a
variety of leak valve settings using a fixed
volume of xenon gas ($\sim$ 1 liter) at the
leak valve input. We measure the pressure
drop in this volume with a capacitive manometer (Type 627D Baratron)
as the gas flows through the leak valve to
infer the flow rate for each leak valve setting.
We find that the flow rate is repeatable
to within 10\% simply by returning the leak
valve to the same indicator marking on its dial.
Second, we also measure the flow
rate directly using a MKS model 179A
mass flow meter which has been
calibrated for use with xenon gas.
Third, when the xenon under analysis
contains a small, constant concentration of
a tracer gas which is unaffected by the
cold trap, such as argon or helium, then the
partial pressure of the tracer can
be used to accurately monitor the leak rate
in real time with the RGA itself. This
eliminates the systematic error due to
the leak valve dial setting and RGA gain
drift. We use this method in
Section \ref{sec:constant-pressure}.
To calibrate the partial pressure measurements
of the RGA in terms of the true krypton
concentration, we insert known quantities of
krypton into the xenon using a krypton gas
cylinder (99.999\% krypton purity). The
injection volume is $13.8\pm 0.1$ cc of plumbing
monitored by a pressure gauge, accurate within 0.1 Torr, and isolated
by two valves. We inject krypton with pressures above 20 Torr, and further reduce the pressure by volume sharing. The
krypton is combined with the xenon
by flowing the xenon through
the injection volume and collecting the
gases in a recovery bottle where they mix.
To perform a measurement, first we submerge
the cold trap in liquid nitrogen while it
is pumped to ultra-high vacuum by the
turbo-molecular pump, typically reaching a vacuum of $4\times 10^{-8}$ Torr. We then
open the leak valve in two steps. In the
first step, we use a very small leak rate,
less than $10^{-4}$ standard liters per minute
(SLPM), which allows xenon ice
to form in the cold trap, establishes
the fixed xenon partial pressure, and flushes
some trace background gases out of
the analysis system plumbing. We wait for
several minutes for the partial pressures
of all species to stabilize, and then we
open the leak valve to the desired flow
rate for purity analysis. In general, the
best sensitivity is obtained by using the
maximum possible flow rate.
In some cases the flow rate is
limited by the partial pressure of
non-xenon impurity species such as
oxygen, nitrogen, or argon. Since these
species are not removed by the cold trap,
their presence in the xenon gas will eventually
cause the RGA to saturate as the flow
rate is increased. For the very best
sensitivity, the xenon should be free
from extraneous impurity species.
\section{Response of the analysis system to krypton}
\label{sec:constant-pressure}
In our first series of experiments, we confirm
that the krypton partial pressure
observed by the RGA is indeed linear in the
true concentration by injecting known amounts
of krypton into our 2.8 kg xenon supply. For these
measurements, the xenon supply bottle continuously
feeds xenon into the system through a regulator,
maintaining a constant pressure at the leak valve
input. This insures that the leak rate into the
analysis system is nearly constant throughout the
measurement, which simplifies the data analysis.
To precisely monitor the leak valve flow rate in
real time, we use argon as a tracer gas. Our
2.8 kg xenon supply contains an argon
concentration of about $10^{-6}$ mol/mol,
and since the argon level is constant
from one injection experiment to the next (because
the injected gas is 99.999\% krypton, with only
trace quantities of argon), the argon partial
pressure serves as a convenient proxy for the gas
flow rate through the analysis system. Under these
conditions, the leak valve can be opened to an
arbitrary setting, and the krypton-to-argon
partial pressure ratio should be proportional
to the true krypton concentration.
\begin{figure}[t!]\centering
\includegraphics[width=100mm]{Fig3.eps}
\caption{The results for a typical purity
measurement with constant
flow rates.
At t = 10 minutes the leak valve is opened and a
measurement is made with a flow rate of 0.10 SLPM.
This sample of xenon contained
$40 \times 10^{-9}$ mol/mol nitrogen,
$4 \times 10^{-6}$ mol/mol argon, and $66.4 \times 10^{-9}$
mol/mol krypton. At t = 28 minutes
the leak valve is closed.}
\label{fig:ConstantFlow}
\end{figure}
A typical measurement is shown in Figure
\ref{fig:ConstantFlow}. At t = 0, xenon ice
has already been established in the cold trap and backgrounds have stabilized,
and the measurement begins at t = 10 minutes, with a flowrate of 0.10 SLPM.
Krypton, argon, and nitrogen are clearly
present in the sampled gas, while the
oxygen concentration is less than
$0.7 \times 10^{-9}$ mol/mol.
At t = 28 minutes the leak valve is closed.
\begin{table}[t!]
\begin{centering}
\begin{tabular}{|c|c|c|c|}
\hline
$\Delta \rho$(Kr) & $P_{Kr}$ (84 u) &
$P_{Ar}$ (40 u) & $P_{Kr}/P_{Ar} $ \\
($10^{-9}$ mol/mol) & ($10^{-9}$ Torr)
& ($10^{-9}$ Torr) & (Torr/Torr) \\
\hline
0 & 4.11 & 70.3 & 0.0584 \\ \hline
7.37 & 5.90 & 91.2 & 0.0647 \\
& 6.02 & 91.9 & 0.0655 \\ \hline
18.4 & 7.00 & 93.7 & 0.0747 \\
& 6.60 & 87.8 & 0.0752 \\ \hline
33.1 & 7.88 & 89.6 & 0.0879 \\
& 7.51 & 86.0 & 0.0873 \\
& 8.05 & 92.0 & 0.0876 \\
& 7.76 & 89.1 & 0.0871 \\ \hline
\end{tabular}
\caption{Krypton and argon partial pressures as a function
of the injected krypton concentration ($\Delta \rho$(Kr)).
Each gas sample was measured at least twice to
confirm the repeatability of the Kr-to-Ar ratio. The uncertainty is the amount of krypton injected depends on the error in the injection volume and the error in the pressure gauge, combined they are less than 1\%. The partial pressure recorded by the RGA is averaged during the measurement yielding a statistical uncertainty of 1\% in partial pressure.}
\label{tab:natural_xe_tab}
\end{centering}
\end{table}
In Figure \ref{fig:natural_xe}
and Table \ref{tab:natural_xe_tab} we show the
krypton-to-argon partial pressure ratio as
a function of the amount of krypton which
we inject into our xenon supply. The partial pressure recorded by the RGA is averaged during the measurement yielding a statistical uncertainty of less than 1\%. The uncertainty in the krypton to xenon ratio after an injection is 1\%, from the uncertainty in the xenon mass (2800$\pm 20$g), the injection volume, and the error on the pressure gauge.
As shown in Table \ref{tab:natural_xe_tab} each gas sample was measured at least twice to
study the repeatability of the krypton-to-argon ratio for a fixed purity concentration. We find the ratio repeatable to about 1\%. The absolute concentration of the argon is known only to within 50\% ($1\pm 0.5$ ppm), however, the absolute concentration is not relevant, since we only use the argon as a flow rate standard.
The krypton-to-argon ratio
is found to be linear in the injected concentration,
which confirms that the cold trap allows the
krypton to pass through as desired. Additional data confirms that it is also linear in the flow rate as observed for other species \cite{coldtrap}. In total
we injected $33.1 \times 10^{-9}$ mol/mol
of krypton, which
resulted in a total increase in the
krypton-to-argon ratio of a factor of
1.50 relative to the
vendor-supplied xenon. From this we infer
that the krypton concentration
was $(66.4\pm 4) \times 10^{-9}$ mol/mol
before our injections, and
$(99.5\pm 4) \times 10^{-9}$
mol/mol after injections.
\begin{figure}[t!]\centering
\includegraphics[width=100mm]{Fig4.eps}
\caption{Krypton-to-argon partial pressure
ratio versus the injected krypton concentration
($\Delta \rho$(Kr)). Repeated measurements are
shown as separate data points.
The non-zero y-intercept value is due to the
krypton present in our vendor supplied xenon
before our injections. Each gas sample was measured at least twice to
gauge the systematic uncertainty in the RGA's partial pressure
measurements for a fixed concentration of krypton.
We infer an uncertainty of 1\% in the krypton-to-argon
partial pressure ratio for a fixed concentration of krypton. Error bars are not plotted as they are too small to be seen on the graph}
\label{fig:natural_xe}
\end{figure}
\section{Detection of krypton at the $10^{-12}$ mol/mol level}
The measurements described in Section \ref{sec:constant-pressure} allow us to quantify the response of the analysis system to krypton. For example, the data in Table \ref{tab:natural_xe_tab} shows a krypton partial pressure of $7.9\times 10^{-9}$ Torr, for a flow rate of 0.1 SLPM, and a krypton concentration of $99.5\times 10^{-9}$ mol/mol. Therefore, the analysis system response to krypton is
$0.79$ Torr/(SLPM $\cdot$ mol/mol).
Since the fluctuations in the RGA reading
at 84 AMU (Atomic Mass Unit) are $\sim 3 \times 10^{-13}$ Torr,
we expect that a concentration
of $1 \times 10^{-12}$ mol/mol
krypton could be detectable at a
flow rate of $\sim 0.4$ SLPM. However, as shown
in Figure \ref{fig:ConstantFlow}, our xenon
supply contains significant argon and nitrogen.
Since these trace gases are not removed by
the cold trap, they will cause the RGA to
saturate at flow rates above 0.1 SLPM,
with the argon being the leading problem. This would
limit our krypton sensitivity
to about $4 \times 10^{-12}$ mol/mol.
To detect krypton at lower concentrations
it is necessary to remove these trace impurities.
Nitrogen can be removed from xenon using standard
getters \cite{purifier}, but argon cannot.
We first tried to remove the argon
by freezing the xenon in its supply bottle
with liquid nitrogen and pumping on the vapor with
the turbo-molecular pump. This strategy proved
to be inefficient, probably
because the argon is trapped in the xenon ice requiring long diffusion times to escape.
However, we successfully purified a small quantity
of xenon using the cold trap itself. We allowed approximately seven standard
liters of xenon to slowly leak
into the cold trap while continuously pumping with a turbo-pump.
We repeated this process several times and during each cycle we measured the remaining krypton concentration with the RGA. We found that the krypton and argon levels were significantly reduced after each pass. We guess that the reduction comes about because the turbo pump is able to remove the krypton and argon before they become permanently trapped in the slowly forming xenon ice.
The average argon and krypton one-pass purification efficiencies were determined to be 99.4\% and 87\% respectively. We find that the krypton purification efficiency remains roughly constant vs. initial concentration over five orders of magnitude, shown in Figure \ref{fig:KrRemoval}. The removal efficiencies are derived assuming that the RGA's response to krypton is linear in concentration, which was later confirmed (see Figure \ref{fig:oneppt}). The uncertainty in the purification efficiency is derived from the maximum deviation from the linear response reported in Table \ref{tab:oneppt}. We do not include the argon removal efficiency in Figure \ref{fig:KrRemoval} because we only have one measurement for argon. After the second pass, argon was not detectable by the RGA due to interference from doubly ionized krypton. This procedure ultimately produced a 40 gram sample of xenon with much
less than $1 \times 10^{-12}$ mol/mol argon
and krypton.
\begin{figure}[t!]\centering
\includegraphics[width=100mm]{Fig5.eps}
\caption{One pass krypton removal efficiency of the coldtrap vs. initial concentration, for several purification cycles.}
\label{fig:KrRemoval}
\end{figure}
Starting with this 40 gram sample of
de-argonated and de-kryptonated xenon gas,
we created xenon with known
krypton concentration by mixing it with
small quantities of the
2.8 kg xenon supply,
which contained $99.5 \times
10^{-9}$ mol/mol of krypton after the
experiments described in
Section \ref{sec:constant-pressure}.
For example, to achieve
$0.5 \times 10^{-12}$ mol/mol of
krypton, we added 0.19 milligrams of our
krypton-rich xenon supply to the
40 grams of de-kryptonated xenon. The sample was created by filling a volume of $13.8\pm 0.1$ cc with $215.9\pm 0.1$ Torr of xenon containing $(99.5\pm 4) \times 10^{-9}$ mol/mol of krypton and then volume sharing the gas with a $1.5222\pm 0.0005$ L volume to further reduce the pressure. With this simple method xenon samples could be prepared with krypton concentrations known to within 5\% of the injection amount.
The xenon was then analyzed
by the cold trap mass spectrometry
technique. Between each run we removed
the extra krypton, and a new sample was mixed
starting again with de-kryptonated xenon.
\begin{figure}[t!]\centering
\includegraphics[width=60mm]{Fig6a.eps}
\includegraphics[width=60mm]{Fig6b.eps}
\includegraphics[width=60mm]{Fig6c.eps}
\includegraphics[width=60mm]{Fig6d.eps}
\caption{RGA response to the smallest
concentrations of krypton. The krypton
signals decay in time because the
flow rate is decreasing due to the small
amount of xenon available for these
measurements. Upper left:
$0.2 \times 10^{-12}$ mol/mol. Upper right:
$0.5 \times 10^{-12}$ mol/mol. Lower left:
$0.9 \times 10^{-12}$ mol/mol. Lower right:
$1.3 \times 10^{-12}$ mol/mol. }
\label{fig:peakplots}
\end{figure}
RGA partial pressure plots are shown in
Figure \ref{fig:peakplots} for xenon with 0.2, 0.5,
0.9, and 1.3 $\times 10^{-12}$ mol/mol of krypton.
For the data
shown in these plots we open the
leak valve to its maximum setting at
t = 0 seconds. Since the total amount of xenon
available for these measurements is
modest, the flow rate immediately peaks at 1.5 SLPM
and then decreases during the
measurement due to the decreasing pressure
at the leak valve input. The resulting
partial pressure plots follow this pattern,
as shown in Figure \ref{fig:peakplots}.
\begin{table}[t!]
\begin{centering}
\begin{tabular}{|c|c|c|c|c|}
\hline
$\rho$(Kr) & Avg. $P_{Kr}$ & Avg. Flow & Kr Fig. of Merit & Dev. from Fit \\
($10^{-12}$ mol/mol) & ($10^{-12}$ Torr) & (SLPM) & ($10^{-12}$ Torr/SLPM) & (\%) \\ \hline
$0.5\pm 0.2$ & 0.406 & 1.32 & $0.308\pm 0.0154$ & 3.8 \\ \hline
$0.9\pm 0.2$ & 0.646 & 1.27 & $0.507\pm 0.0254$ & 5.1 \\ \hline
$1.3\pm 0.2$ & 1.15 & 1.28 & $0.901\pm 0.0450$ & -12.4 \\ \hline
$1.7\pm 0.2$ & 1.37 & 1.28 & $1.07 \pm 0.0528$ & 1.2 \\ \hline
$5.1\pm 0.3$ & 3.81 & 1.30 & $2.93 \pm 0.147$ & 8.6 \\ \hline
$17.1\pm 0.9$ & 15.7 & 1.29 & $12.2 \pm 0.608$ & -13.7 \\ \hline
$171.1\pm 8.8$ & 127 & 1.27 & $99.5 \pm 4.97$ & 6.9 \\ \hline
$^*1711\pm 88$ & 1450 & 1.31 & $1110 \pm 55.3$ & -3.5 \\
$1711\pm 88$ & 1330 & 1.29 & $1030 \pm 51.3$ & 4.1 \\ \hline
\end{tabular}
\caption{Results of krypton detection experiments with
40 grams of highly purified xenon. The
various krypton concentrations
were created by mixing with xenon containing
$99.5 \times 10^{-9}$ mol/mol krypton, except the
sample labeled ($^*$), which was created by injecting
99.999\% krypton from a krypton gas cylinder.
For prepared samples of $1.7 \times 10^{-12}$ mol/mol
or less the uncertainty in the concentration is
dominated by the minimum sensitivity
to krypton in the highly purified xenon, which we take
to be $0.2 \times 10^{-12}$ mol/mol.
For concentrations above $1.7 \times 10^{-12}$
the uncertainty in the concentration is 5\%,
dominated by the uncertainty of the concentration of the
krypton-rich xenon supply.
The krypton figure of merit is the average partial
pressure divided by the average flow rate.
The last column shows the deviation from the
linear fit shown in Figure \ref{fig:oneppt}.}
\label{tab:oneppt}
\end{centering}
\end{table}
Clear krypton signals are seen for concentrations
of $0.9 \times 10^{-12}$ mol/mol and larger, while
the $0.5 \times 10^{-12}$ mol/mol sample
gives a marginal signal. No significant deviation
from background is seen for the
$0.2 \times 10^{-12}$ mol/mol sample, which
agrees with our expected sensitivity of
$0.25 \times 10^{-12}$ mol/mol (inferred by
assuming partial pressure fluctuations of
$3 \times 10^{-13}$ Torr and a
1.5 SLPM flow rate.)
\begin{figure}[t!]\centering
\includegraphics[width=100mm]{Fig7.eps}
\caption{Krypton detection figure-of-merit
as a function of true krypton concentration.
The krypton partial pressure
data (at 84 AMU) is averaged in a 60 second
window around its peak value and
then normalized to the average
flow rate. The maximum deviation from a
linear fit over three orders of magnitude
is 13.7\%. The solid line indicates the linear fit to the data.}
\label{fig:oneppt}
\end{figure}
To quantify the krypton concentration, we
calculate the average partial pressure in a
60 second window around its maximal value
after subtracting the background level,
and we divide by the average flow rate measured
by the MKS mass flow meter. Since the
same leak valve setting (fully open) was
used in each dataset, the average flow rate
varies by less than 2\% in all datasets.
As shown in Figure \ref{fig:oneppt} and
Table \ref{tab:oneppt}, the
ratio of average pressure to average flow rate
is proportional
to the true krypton concentration over four
orders of magnitude. The linear dependence was expected as the peak partial pressure, for a fixed flow rate, should be proportional to the number of particles passing by the RGA per unit time. The largest deviation
from the fitted line is 13.7\%, and the
dataset at $0.5 \times 10^{-12}$ mol/mol
deviates from the fitted line by only 4\%.
This indicates that the small krypton
signal at $0.5 \times 10^{-12}$ mol/mol
is likely to be genuine, and demonstrates
sensitivity to krypton at concentrations
less than $1 \times 10^{-12}$. Previous methods have achieved sensitivities of about $\sim 10^{-12}$ mol/mol \cite{distillation,chromotography}.
The linear dependence of the krypton figure-of-merit to concentration down to $0.5 \times 10^{-12}$ mol/mol also demonstrates that the de-kryptonated xenon truly had an initial concentration less than $0.5 \times 10^{-12}$ mol/mol. Had there been residual krypton in the xenon before the krypton injections, the data in Figure \ref{fig:oneppt} would level off and not continue its linear decline at lower concentrations.
To confirm that the absolute krypton
concentration of our experiments
is not in error, we performed one final
krypton injection from the krypton cylinder
into our 40 grams of de-kryptonated xenon
at a concentration of
$1.7 \times 10^{-9}$ mol/mol. This dataset
gives a krypton figure of merit which agrees
with the equivalent dataset produced by mixing
to within 8\%.
A second cross-check on our krypton concentration scale is provided by
the EXO-200 double beta decay experiment. As reported in Ref.
\cite{steveherrin}, EXO-200 has observed $^{85}$Kr in its natural
(unenriched) xenon gas supply at a decay rate consistent with that
inferred from our mass spectrometry technique, assuming that the
$^{85}$Kr isotopic abundance is $\sim 10^{-11}$, as expected. This
confirms our absolute scale to within a factor of two.
\section{Conclusion}
We have extended the xenon cold trap mass spectrometry
technique to detect trace quantities of krypton in
xenon gas. We find that krypton passes through the
cold trap largely undisturbed, and that the
resulting partial pressure is proportional to
the true concentration after accounting for the flow rate.
Using this method we have detected krypton concentrations
as low as $(0.5\pm 0.2) \times
10^{-12}$ mol/mol $^{nat}$Kr/Xe.
Compared to the previously reported methods for krypton detection in
xenon, our technique is rather simple and inexpensive, yet it achieves
better sensitivity. It does not require any specialized equipment beyond an RGA, which most labs use routinely anyway.
We believe our sensitivity
could be significantly improved by using faster flow rates,
which could be achieved by using a larger sample of
highly purified xenon. In principle we see no reason
why an additional factor of ten improvement could
not be achieved by using ten times the amount of xenon,
which would make the technique useful for future
WIMP dark matter experiments. In any case, the sensitivity
demonstrated here will already be useful for krypton
monitoring programs at existing detectors.
\section{Acknowledgments}
\label{sec:Acknowledgments}
This work was supported by the National Science Foundation
under award number PHY0810495.
\bibliographystyle{elsarticle-num}
|
\section{Introduction}
The existence of particle DM is a widely accepted astrophysical concept that is used to explain, e.g., the formation of large scale structure during the evolution of the universe. In the present universe, galaxies, among others, are believed to be embedded in DM halos, which have density profiles that are much more extended than the profiles of visible matter and also exhibit the highest DM density in their centers (e.g. \cite{NFW:1996,Springel:2008,Diemand:2008}). This fact makes galactic centers promising targets to search for signals from annihilation or decay of hypothetical DM particles. In simulations of cosmological structure formation, cold DM, which, among others, might consist of non-baryonic, weakly interacting massive particles ($m_{\mathrm{\chi}} \sim$ 10~GeV up to a few TeV) \cite{Griest:1988}, is able to reproduce the observed large scale structure of the universe \cite{Springel:2008,Diemand:2008}. They are expected to either annihilate or decay into standard model particles, producing, among other particles, photons in the final state. These photons are predicted to exhibit a continuous spectrum and to cover a broad energy range up to the DM mass, with possibly narrow lines from direct $\chi\chi\rightarrow\gamma\gamma / \gamma Z$ annihilations or other spectral features overlaid (see, e.g., \cite{Srednicki:1986, Bergstroem:1988, Bergstroem:2007}). Therefore, $\gamma$-ray observations, in particular at very high energies where DM masses of a few 100~GeV and above are probed, are a promising way to detect DM in space and to constrain its particle physics properties.
A successful DM observation strategy has to avoid sky regions with strong astrophysical $\gamma$-ray signals, and should focus at the same time on regions with an expectedly large DM density. Besides dwarf galaxies, from which limits on $\langle\sigma v\rangle$ have been derived recently from observations at high (100~MeV-20~GeV) \cite{Scott:2010,Abdo:2010} and very high energies \cite{Aharonian:2008dm,Wood:2008,Aharonian:2010dmerr,Albert:2008,Aharonian:2009dm,Veritas:2010dm}, the Galactic Center (GC) region is a prime target for DM search, both because of its proximity and its predicted large DM concentration (see, e.g., \cite{NFW:1996}). As opposed to dwarf galaxies, however, the search for DM induced $\gamma$-rays in the GC is hampered by a strong astrophysical background. Especially at the very center, there is the compact $\gamma$-ray source HESS~J1745-290, coincident with the position of the supermassive black hole Sgr~A* and a nearby pulsar wind nebula \cite{Aharonian:2004wa,GCSpectrum,vanEldik:2010}, which makes the detection of $\gamma$-rays from DM annihilation difficult \cite{Aharonian:2006wh}. Furthermore, there is a band of diffuse emission along the GC ridge \cite{Aharonian:2006au}, which has already been used to derive upper limits on a DM induced signal \cite{Crocker:2010, Bertone:2009, Meade:2010}.
The $\gamma$-ray flux expected from annihilations of DM particles of mass $m_\chi$ is given by the product of the integral $J$ of the squared DM mass density $\rho$ along the line-of-sight, the velocity-weighted annihilation cross section $\langle\sigma v\rangle$, and the average $\gamma$-ray spectrum $dN/dE_\gamma$ produced in a single annihilation event \cite{Bertone:2005}.
For a given DM mass and given branching ratios for annihilation into SM particles, the shape of the expected $\gamma$-ray spectrum, determined by $dN/dE_{\gamma}$, can be rather accurately computed, since only SM particle physics processes are involved. Measuring the velocity-weighted annihilation cross section $\langle\sigma v\rangle$, however, suffers from large systematic uncertainties on the astrophysical factor $J$. This is because in most indirect searches the main contribution of the total DM signal arises from the very central region of the object under study, where the DM density profile peaks. In these inner regions, however, the profile is so far only poorly known. In particular, for a Milky-Way sized DM halo, the radial DM density profiles obtained by the Aquarius \citep{Springel:2008} and Via Lactea II \citep{Diemand:2008} simulations can be described by Einasto and NFW parametrizations, respectively \citep{Pieri:2009}. These profiles are shown in Fig.~\ref{Fig:Profiles} as a function of galactocentric distance $r$. Large differences between both parametrization s occur if they are extrapolated down to the very center of the halo, where the NFW profile is more strongly peaked. At distances $>10$~pc, however, the difference is merely a factor of two, allowing one to put limits on $\langle\sigma v\rangle$ which do not depend strongly between either of the two parametrizations.
\begin{figure}
\includegraphics[width=0.5\textwidth]{./DensityProfiles.eps}
\caption{Comparison of the Galactic DM halo profiles used in this analysis. The parameters for the NFW and Einasto profiles are taken from \cite{Pieri:2009}. An isothermal profile \citep{Bertone:2005}, exhibiting a flat DM density out to a galactocentric distance of ~1~kpc, is shown for comparison. All profiles are normalized to the local DM density ($\rho_0 = 0.39$~GeV/cm$^{3}$ \cite{Catena:2009} at a distance of 8.5~kpc from the GC). The source region and the region used for background estimation are indicated. Note that the predicted DM density is always larger in the source region, except for the isothermal profile, which is included for completeness.}
\label{Fig:Profiles}
\end{figure}
Here we exploit this fact by searching for a VHE $\gamma$-ray signal from DM annihilation in our own Galaxy, in a region with a projected galactocentric distance of $45$~pc~$-150$~pc (corresponding to an angular distance of $0.3^\circ - 1.0^\circ$)\footnote{Here and in the following a distance of the GC to the observer of 8.5~kpc is assumed.}, excluding the Galactic plane. In this way, contamination from $\gamma$-ray sources in the region is naturally avoided as well.
\section{Methodology\label{sec:Methodology}}
The analysis is carried out using 112~h (live time) of GC observations with the H.E.S.S. VHE $\gamma$-ray instrument (see \cite{Crab:2006} and references therein) taken during the years 2004-2008. For minimum energy threshold, only observations with zenith angles smaller than $30^\circ$ are considered. The mean zenith angle is $14^\circ$. To avoid possible systematic effects, pointing positions were chosen fairly symmetric with regard to the Galactic plane. The mean distance between the pointing position and the GC is $0.7^\circ$, with a maximum of $1.5^circ$. Events passing H.E.S.S. standard cuts defined in \cite{Crab:2006} are selected for analysis. To minimize systematic uncertainties due to reduced $\gamma$-ray efficiency at the edges of the $\sim 5^\circ$ diameter field-of-view (FoV) of H.E.S.S., only events reconstructed within the central $4^\circ$ are considered. The effective $\gamma$-ray collection area of the chosen event selection is $\approx 1.7 \times 10^{5} \text{ m}^{2}$ at 1 TeV. The total effective exposure of the utilized dataset at 1 TeV amounts to $\approx 2.4 \times 10^{7} \text{ m}^{2}$ sr s.
$\gamma$-rays from DM annihilations are searched for in a circular \emph{source region} of radius $R_\mathrm{on}=1.0^\circ$ centered at the GC. Contamination of the DM signal by local astrophysical $\gamma$-ray sources is excluded by restricting the analysis to Galactic latitudes $|b|>0.3^\circ$, effectively cutting the source region into two segments above and below the Galactic plane (see Fig. \ref{Fig:ReflectedPixel}). Simulations show that both for Einasto and NFW parametrizations the sensitivity varies only within a few percent when varying the source region size in the range $0.8^\circ \leq R_\mathrm{on} \leq 1.2^\circ$.
For ground-based VHE $\gamma$-ray instruments like H.E.S.S., any $\gamma$-ray signal is accompanied by a sizable number of cosmic-ray induced background events, which are subtracted from the source region using control regions in the FoV of the observation. Fig.~\ref{Fig:ReflectedPixel} visualizes details of the method, which is an evolution of the standard \emph{reflected background} technique \cite{Berge:2006ae} adjusted for this particular analysis.
By construction, background regions are located further away from the GC than the source region. This is an important aspect, since, unavoidably, a certain amount of DM annihilation events would be recorded in the background regions, too, reducing a potential excess signal obtained in the source region. For the NFW and Einasto profiles, the expected DM annihilation flux is thus smaller in the background regions than in the source region (cf. Fig.~\ref{Fig:Profiles}), making the measurement of a residual annihilation flux possible. Note, however, that for an isothermal halo profile, the signal would be completely subtracted. As far as the background from Galactic diffuse emission is concerned, its predicted flux \cite{GDE:2005} is significantly below the current analysis sensitivity, thus its contribution is not further considered in the analysis. In any case, since its intensity is believed to drop as a function of Galactic latitude, $\gamma$-rays from Galactic diffuse emission would be part of a potential signal, and therefore lead to more conservative results for the upper limits derived in this analysis.
\begin{figure}
\includegraphics[width=0.5\textwidth]{./ReflectedPixel.eps}
\caption{Illustration of the cosmic ray background subtraction technique for a single telescope pointing position (depicted by the star). Note that this position is only one of the several different pointing positions of the dataset. The DM source region is the green area inside the black contours, centered on the GC (black triangle). Yellow regions are excluded from the analysis because of contamination by astrophysical sources. Corresponding areas for background estimation (red regions) are constructed by rotating individual pixels of size $0.02^\circ\times 0.02^\circ$ of the source region around the pointing position by $90^\circ$, $180^\circ$, and $270^\circ$. This choice guarantees similar $\gamma$-ray detection efficiency in both the source and background regions. As an example, pixels labeled \emph{1} and \emph{2} serve as background control regions for pixel \emph{0}. Pixel \emph{3} is not considered for background estimation because it is located in an excluded region. Pixels in the source region, for which no background pixels can be constructed, are not considered in the analysis for this particular pointing position and are left blank.}
\label{Fig:ReflectedPixel}
\end{figure}
\section{Results}
Using zenith angle-, energy- and offset-dependent effective collection areas from $\gamma$-ray simulations, flux spectra shown in Fig. \ref{Fig:Spectra} are calculated from the number of events recorded in the source and background regions\footnote{The background spectrum is rescaled by the ratio of the areas covered by source and background regions (cf. also \cite{Berge:2006ae}).}. It should be stressed that these spectra consist of $\gamma$-ray-like cosmic-ray background events. Both source and background spectra agree well within the errors, resulting in a null measurement for a potential DM annihilation signal, from which upper limits on $\langle\sigma v\rangle$ can be determined.
\begin{figure}
\includegraphics[width=0.5\textwidth]{./SpectraAndResidua.eps}
\caption{Top panel: Reconstructed differential flux $F_{\text{Src/Bg}}$, weighted with $E^{2.7}$ for better visibility, obtained for the source and background regions as defined in the text. The units are TeV$^{1.7}$~m$^{-2}$~s$^{-1}$~sr$^{-1}$. Due to an energy-dependent selection efficiency and the use of effective areas obtained from $\gamma$-ray simulations, the reconstructed spectra are modified compared to the cosmic-ray power-law spectrum measured on Earth. Bottom panel: Flux residua $F_{\text{res}}/ \Delta F_{\text{res}}$, where $F_{\text{res}}=F_{\text{Src}}-F_{\text{Bg}}$ and $\Delta F_{\text{res}}$ is the statistical error on $F_{\text{res}}$. The residual flux is compatible with a null measurement. Comparable null residuals are obtained when varying the radius of the source region, subdividing the data set into different time periods or observation positions, or analyzing each half of the source region separately.}
\label{Fig:Spectra}
\end{figure}
The mean astrophysical factors $\bar{J}_\mathrm{src}$ and $\bar{J}_\mathrm{bg}$ are calculated for the source and background regions, respectively. The density profiles are normalized to the local DM density $\rho_0 = 0.39$~GeV/cm$^{3}$ \cite{Catena:2009}. Assuming an Einasto profile, $\bar{J}_\mathrm{src}=3142 \times \rho_\mathrm{E}^2 \times d_\mathrm{E}$ and $\bar{J}_\mathrm{bg}=1535 \times \rho_\mathrm{E}^2 \times d_\mathrm{E}$, where $\rho_\mathrm{E} = 0.3 \text{ GeV}/\text{cm}^3$ is the conventional value for the local DM density and $d_\mathrm{E} = 8.5$~kpc the distance of Earth to the GC. For a NFW profile, $\bar{J}_\mathrm{src}=1604 \times \rho_\mathrm{E}^2 \times d_\mathrm{E}$ and $\bar{J}_\mathrm{bg}=697 \times \rho_\mathrm{E}^2 \times d_\mathrm{E}$ are obtained. This means that for an assumed Einasto (NFW) profile, background subtraction reduces the excess DM annihilation flux in the source region by 49~\% (43~\%), which is taken into account in the upper limit calculation.
Under the assumption that DM particles annihilate into quark-antiquark pairs and using a generic parametrization for a continuum spectrum of $\gamma$-rays created during the subsequent hadronization \cite{Tasitsiomi:2002, Hill:1987}, limits on $\langle \sigma v \rangle$ as a function of the DM particle mass are calculated for both density profiles (see Fig. \ref{Fig:UpperLimits}). These limits are among the most sensitive so far at very high energies, and in particular are the best for the Einasto density profile, for which at $\sim1$ TeV values for $\langle\sigma v\rangle$ above $3 \times 10^{-25} \text{ cm}^{3}\text{ s}^{-1}$ are excluded. As expected from the astrophysical factors, the limits for the Einasto profile are better by a factor of two compared to those for the NFW profile. Still, the current limits are one order of magnitude above the region of the parameter space where supersymmetric models provide a viable DM candidate (see Fig. \ref{Fig:UpperLimits}). Apart from the assumed density parametrizations and the shape of the $\gamma$-ray annihilation spectrum, the limits can shift by 30\% due to both the uncertainty on the absolute flux measurement \cite{Crab:2006} and the uncertainty of 15\% on the absolute energy scale. For the latter case, apart from a displacement with regard to the DM particle mass scale, the limits shift up (down) if the $\gamma$-ray energy is overall under(over)estimated.
\begin{figure}
\includegraphics[width=0.5\textwidth]{./NewUpperLimitsWithDarkSUSYPoints.eps}
\caption{Upper limits (at 95\% CL) on the velocity-weighted annihilation cross-section $\langle \sigma v \rangle$ as a function of the DM particle mass $m_{\chi}$ for the Einasto and NFW density profiles. The best sensitivity is achieved at $m_{\chi}\sim 1$~TeV.
For comparison, the best limits derived from observations of dwarf galaxies at very high energies, i.e. Sgr Dwarf \cite{Aharonian:2008dm}, Willman 1, Ursa Minor \cite{Veritas:2010dm} and Draco \cite{Abdo:2010}, using in all cases NFW shaped DM profiles, are shown. Similar to source region of the current analysis, dwarf galaxies are objects free of astrophysical background sources. The green points represent DarkSUSY models \cite{DarkSusy:2004}, which are in agreement with WMAP and collider constraints and were obtained with a random scan of the mSUGRA parameter space using the following parameter ranges: $10 \text{ GeV} < M_0 < 1000 \text{ GeV}$, $10 \text{ GeV} < M_{1/2} < 1000 \text{ GeV}$, $A_0 = 0$, $0 < \text{tan} \beta < 60$, $\text{sgn} (\mu) = \pm 1$. }
\label{Fig:UpperLimits}
\end{figure}
\section{Summary}
A search for a VHE $\gamma$-ray signal from DM annihilations was conducted using H.E.S.S. data from the GC region. A circular region of radius $1^\circ$ centered at the GC was chosen for the search, and contamination by astrophysical $\gamma$-ray sources along the Galactic plane was excluded. An optimized background subtraction technique was developed and applied to extract the $\gamma$-ray spectrum from the source region. The analysis resulted in the determination of stringent upper limits on the velocity-weighted DM annihilation cross-section $\langle \sigma v \rangle$, being among the best so far at very high energies. At the same time, the limits do not differ strongly between NFW and Einasto parametrizations of the DM density profile of the Milky-Way.
\begin{acknowledgments}
The support of the Namibian authorities and of the University of Namibia
in facilitating the construction and operation of H.E.S.S. is gratefully
acknowledged, as is the support by the German Ministry for Education and
Research (BMBF), the Max Planck Society, the French Ministry for Research,
the CNRS-IN2P3 and the Astroparticle Interdisciplinary Programme of the
CNRS, the U.K. Science and Technology Facilities Council (STFC),
the IPNP of the Charles University, the Polish Ministry of Science and
Higher Education, the South African Department of
Science and Technology and National Research Foundation, and by the
University of Namibia. We appreciate the excellent work of the technical
support staff in Berlin, Durham, Hamburg, Heidelberg, Palaiseau, Paris,
Saclay, and in Namibia in the construction and operation of the
equipment.
\end{acknowledgments}
|
\section{Introduction}
High-order harmonic generation (HHG) is one of the most interesting nonlinear
phenomena when atoms or molecules are exposed to an intense laser field
\cite{Winter-RMP-2008,Krausz-RMP-2000,Agostini-RPP-2004}. Experimentally the
harmonic fields generated by all atoms or molecules within the laser focus
co-propagate with the fundamental laser field till they reach the detector. To
compare with experimental measurements, theoretical harmonics from each atom or
molecule first have to be calculated. The induced dipoles are then fed into the
propagation equation, taking into account the focusing condition, the nature of
the gas medium, and the equation is finally integrated till the harmonics are
collected
\cite{Mette-jpb,Priori-pra-2000,tosa-pra-2005,Geissler-prl-99,jin-pra-2011}.
HHG from randomly distributed molecules have been studied since the
1990's
\cite{Altucci-pra-2006,Wong-pra-2010,Wong-ol-2010,Trallero-CP-2009}.
In recent years, experimental HHG studies tend to focus on partially
aligned molecules
\cite{stanford-Sci-N2,Haessler-NatPhys-2010,Mairesse-prl-2010,Lee-jpb-2010}.
Among the molecules, CO$_2$ is likely the most extensively studied
system so far
\cite{Boutu-NatPhys-2008,Kanai-pra-2008,Wagner-pra-2007,Zhou-prl-2008}.
Initially the interest was focused on the observation of the minimum
in the HHG spectra of CO$_2$
\cite{olga-nature-2009,Vozzi-prl-2005,Torres-pra-2010}. The
positions of the minima from different experiments, however, are
often vastly different. According to the three-step model
\cite{Corkum-prl-1993,Krause-prl-1992}, HHG is generated through the
recombination of the returning electrons with the molecular ion. The
interference of electron waves from the two atomic centers, under
some conditions, results in a minimum in the transition dipole, the
simplest is the two-center interference model
\cite{Kanai-nature-2005,Lein-prl-2002,Le-pra-2006}. In fact,
recombination is an inverse process of photoionization. Thus the
transition dipole is the same as that in photoionization. Any such
minima have not been observed in photoionization, however, since in
these measurements molecules are isotropically distributed.
From the theoretical side, the alignment dependence of HHG was first
studied using the strong-field approximation (SFA) (or the
Lewenstein model)
\cite{Lewen-pra-1994,Zhou-pra-2005a,Zhou-pra-2005b}. Subsequently we
developed a quantitative rescattering (QRS) theory
\cite{lin-jpb-10,toru-2008,at-pra-2009} for HHG which was a
significant improvement on the Lewenstein model. In QRS the accuracy
of recombination is treated at the same level as in the
photoionization process. Using QRS, the HHG minimum is attributed to
the minimum in the photoionization transition dipole between the
highest occupied molecular orbital (HOMO) and continuum molecular
wave functions \cite{at-pra-2009,le-prl-2009}. For fixed-in-space
CO$_2$ molecules the photoionization cross sections of HOMO indeed
exhibit minima at small alignment angles. The experimental HHG
spectra from partially aligned CO$_2$ have been explained reasonably
well by QRS, including the harmonic intensities and the phase, so
are the polarization and ellipticity of the harmonics
\cite{le-prl-2009,at-pra-2010}.
If HHG is generated from HOMO only, then one expects that the
position of the minimum does not significantly change with laser
intensity according to QRS theory. Indeed, strong field ionization
depends exponentially on the ionization potential I$_{\text p}$. The
HOMO-1 and HOMO-2 orbitals in CO$_2$ are 4 and 4.4 eV more deeply
bound than the HOMO \cite{Turner}, thus they are not expected to
contribute significantly to the HHG spectra. However, it is also
well known that tunneling ionization rates depend sensitively on the
symmetry of the molecular orbital \cite{lin-jmo-2006}. The HOMO is a
$\pi_{g}$ orbital. It means that at small alignment angles the
ionization rates are small. For HOMO-2, on the other hand, it is a
$\sigma_{g}$ orbital, thus it has large ionization rate when CO$_2$
molecules are parallel aligned. Thus for small alignment angles,
HOMO-2 may become important even though it is bound 4.4 eV deeper
than the HOMO. (HOMO-1 is a $\pi_{u}$ orbital and thus not expected
to contribute significantly to the HHG.)
The first step of HHG process is tunneling ionization. The alignment
dependence of tunneling ionization rate is usually calculated using
molecular Ammosov-Delone-Krainov (MO-ADK) theory
\cite{tong-pra-2002,zhao-pra-10} or SFA. For most molecules that
have been studied the two models give nearly identical alignment
dependence (after normalization). However, this is not the case for
CO$_2$. Experimentally, the alignment dependence of CO$_2$
ionization reported by Pavi\u{c}i\'{c} {\it et al.}
\cite{Pavi-prl-2007} is very narrowly peaked near alignment angle of
46$^{\circ}$. It differs significantly from the predictions of
MO-ADK and SFA \cite{hoang-jpb-2008}. In fact, so far all
theoretical attempts
\cite{Spanner-pra-2009,Zhao-pra-2009,madsen-pra-2009,chu-pra-2009,Petretti-prl-2010}
have not been able to confirm the sharp alignment dependence
reported in the experiment. Furthermore, the observed HHG spectra
from aligned molecules are inconsistent with the reported
experimental alignment dependence of ionization
\cite{at-pra-2010,at-jpb-2009}.
The HHG spectra of parallel aligned CO$_2$ have been studied in a
number of experiments, with 800-nm lasers \cite{olga-nature-2009} as
well as lasers of longer wavelengths
\cite{Torres-pra-2010,hans-prl-10}. It was observed that the
position of the HHG minimum moved to higher photon energies as the
laser intensity was increased. In Smirnova {\it et al.}
\cite{olga-nature-2009}, such effects were interpreted in terms of
the multi-channel interference (between HOMO and HOMO-2) and the
intricate hole dynamics. Their calculations were based on HHG
generated by a single-molecule response. They introduced filters
into the theory such that only short trajectories contributed to the
signals. In their calculation, a relative phase between different
channels due to strong-field ionization step was introduced ``by
hand'' in order to obtain the good agreement with the measurement.
As noted at the beginning, to compare theoretically simulated HHG
spectra of molecules with experimental ones, the effect of
macroscopic propagation should be considered. Practically, this has
rarely been done. For atomic targets, propagation effect is usually
included with single-atom induced dipoles calculated using the
Lewenstein model. Only in a few rare occasions the atomic response
is calculated more accurately by solving the time-dependent
Schr\"odinger equation (TDSE) \cite{mette-pra-2006,mette-pra-2002}.
It is also well-known that the Lewenstein model does not predict the
HHG spectra (the intensity) precisely, but the phases of the
harmonics are relatively accurate. This fact is used in the QRS
model, which can be understood as simply replacing the transition
dipole calculated using plane waves in the Lewenstein model by one
calculated using accurate molecular continuum wave functions. The
improvement of single-molecule HHG spectra calculated using QRS has
been well documented in our previous publications
\cite{at-pra-2009,at-jpb-2008}. Computationally, QRS is nearly as
simple as the Lewenstein model (actually QRS is even simpler than
SFA for polyatomic molecules, see Ref. \cite{Zhao-pra-2011}), thus
induced dipole can be easily obtained from QRS to feed into the
propagation equations to account for medium propagation effects.
This has been done for atomic targets for low-laser-intensity and
low-gas-pressure conditions \cite{jin-pra-2009}. It has been
extended recently to the conditions of high intensity and high
pressure for Ar and to molecular targets
\cite{jin-pra-2011,jin-jpb-2011}, including polyatomic molecules
\cite{Zhao-pra-2011}.
In this paper, we report theoretical studies of the propagation
effect on the HHG of CO$_2$, including contributions from HOMO and
HOMO-2. Note that experimentally the effect of propagation on the
HHG spectra has rarely been investigated either, in particular, its
effect on the minimum of the HHG spectra. However, this has changed
recently with the reports of such studies on Ar
\cite{jin-jpb-2011,Farrell-pra-2011,Higuet-2011}. The rest of this
paper is arranged as follows. In Sec. II, we first summarize the
equations used for the propagation calculations. We limit ourselves
to low laser intensity and low pressure. To extend the theory to
high intensity and high pressure, the optical properties of CO$_2$
have to be used and they are not available over a broad range of
frequencies. We then summarize how the single-molecule response from
partially aligned molecules is calculated. In Sec. III the
calculated results are presented. Different factors that can
influence the precise positions of the HHG minima are examined and
reported. These results show that precise theoretical predictions of
the positions of HHG minima in a given experiment is difficult, but
the trend (the direction of the shift of the minimum positions) can
be predicted (or explained). A short summary at the end concludes
this paper.
\section{Theoretical method}
We first briefly describe the theory of propagation of high harmonics in a
macroscopic medium, and in free space, till they reach the detector. To
calculate the induced dipole for individual molecules, we include the
contributions from multiple molecular orbitals. Our method is based on
extending the QRS theory \cite{at-pra-2009}.
\subsection{Propagation of harmonics in the medium}
In general, both the fundamental laser field and the harmonic field are
modified when they co-propagate through a macroscopic medium. In this paper, we
limit ourselves to experiments carried out under the conditions of low laser
intensity and low gas pressure, where the effects of dispersion, Kerr
nonlinearity, and plasma defocusing on the fundamental laser field can be
neglected \cite{jin-pra-2009,jin-pra-2011}. The fundamental laser field is
assumed to be a Gaussian beam in space. Its spatial and temporal dependence can
be expressed in an analytical form \cite{jin-pra-2009}. For high harmonics,
dispersion and absorption effects from the medium are not included. The
dispersion and absorption coefficients depend linearly on gas pressure and
could be ignored under low pressure \cite{jin-pra-2011}. The free-electron
dispersion is also neglected since the plasma frequency is much smaller than
the frequencies of high harmonics \cite{Priori-pra-2000}.
The three-dimensional propagation equation of the harmonic field is
described as
\cite{jin-pra-2011,jin-pra-2009,Priori-pra-2000,Mette-jpb,tosa-pra-2005}
\begin{eqnarray}
\label{harm-freq}\nabla_{\bot}^{2}\tilde{E}_{\text
h}^{\parallel}(r,z',\omega,\alpha)-&&\frac{2i\omega}{c}\frac{\partial
\tilde{E}_{\text h}^{\parallel}(r,z',\omega,\alpha)}{\partial
z'}\nonumber\\&&
=-\omega^{2}\mu_{0}\tilde{P}_{\text{nl}}^{\parallel}(r,z',\omega,\alpha),
\end{eqnarray}
where
\begin{eqnarray}
\tilde{E}_{\text h}^{\parallel}(r,z',\omega,\alpha)=\hat{F}[E_{\text
h}^{\parallel}(r,z',t',\alpha)],
\end{eqnarray}
and
\begin{eqnarray}
\tilde{P}_{\text
{nl}}^{\parallel}(r,z',\omega,\alpha)=\hat{F}[P_{\text
{nl}}^{\parallel}(r,z',t',\alpha)].
\end{eqnarray}
Here $\hat{F}$ is the Fourier transform operator acting on the temporal
coordinate. Eq.~(\ref{harm-freq}) is written in a moving coordinate frame
($z^{\prime}=z$ and $t^{\prime}=t-z/c$) and the term $\partial^{2}E_{\text
h}^{\parallel}/\partial z^{\prime 2}$ is neglected. $\tilde{E}_{\text
h}^{\parallel}(r,z',\omega,\alpha)$ is the parallel component of the harmonic
field with respect to the polarization direction of the probe (or generating)
laser, and $\tilde{P}_{\text {nl}}^{\parallel}(r,z',\omega,\alpha)$ is the
parallel component of the nonlinear polarization caused by the generating laser
field, where $\alpha$ is pump-probe angle, i.e., the angle between the aligning
laser and the harmonic generating laser polarizations.
The nonlinear polarization term can be expressed as
\begin{eqnarray}
\label{pola}\tilde{P}_{\text{nl}}^{\parallel}(r,z',\omega,\alpha)=\hat{F}{\{[n_{0}-n_{\text
e}(r,z',t',\alpha)]D^{\parallel,\text{tot}}(r,z',t',\alpha)\}},
\nonumber\\
\end{eqnarray}
where $n_{0}$ is the density of neutral molecules,
$D^{\parallel,\text{tot}}(t',\alpha)$ is the parallel component of
the induced single-molecule dipole over a number of active electrons
[see Eq.~(\ref{total-dip}) below], and $n_{\text e}(t',\alpha)$ is
the free-electron density, which can be calculated as following:
\begin{eqnarray}
\label{free-afa}n_{\text e}(t',\alpha)=\int^{\pi}_{0}n_{\text
e}(t',\theta)\rho(\theta,\alpha)\sin\theta d\theta.
\end{eqnarray}
Here $n_{\text e}(t',\theta)$ is the alignment-dependent
free-electron density, obtained from
\begin{eqnarray}
\label{free-electron}n_{\text
e}(t',\theta)=n_{0}\{1-\exp[-\int_{-\infty}^{t'}\gamma(\tau,\theta)d\tau]\},
\end{eqnarray}
where $\gamma(\tau,\theta)$ is the alignment-dependent ionization
rate, which can be calculated by MO-ADK theory
\cite{tong-pra-2002,zhao-pra-10} for different molecular orbitals.
In Eq.~(\ref{free-afa}), $\theta$ is the alignment angle with
respect to the polarization direction of the probe laser, and
$\rho(\theta,\alpha)$ is the alignment distribution with pump-probe
angle $\alpha$ \cite{jin-pra-2010,at-pra-2009,lein-jmo-2005}.
After the propagation in the medium, we obtain the parallel
component of near-field harmonics on the exit face of the gas jet
($z'=z_{\text {out}}$). For isotropically distributed molecules and
partially aligned molecules with $\alpha=0^{\circ}$ or $90^{\circ}$,
by symmetry there are only parallel harmonic components. The
perpendicular components, which are usually much smaller, would
appear for partially aligned molecules and the harmonics will be
elliptically polarized in general \cite{at-pra-2010}. Generalization
of Eq.~(\ref{harm-freq}) to the perpendicular component is
straightforward, but we restrict ourselves to parallel component
only. Eq.~(\ref{harm-freq}) is solved numerically using a
Crank-Nicholson routine for each value of $\omega$. Typical
parameters used in the calculations are 300 grid points along the
radial direction and 400 grid points along the longitudinal
direction.
\subsection{Harmonics in the far field}
Once the near-field harmonics are obtained on the exit face of the
medium, they further propagate in free space before detected by the
spectrometer. In the meanwhile, they may pass through a slit, an
iris, or be reflected by a mirror. The far-field harmonics can be
obtained from near-field harmonics through a Hankel transformation
\cite{far-field,L'Huillier-1992,tosa-2009}
\begin{eqnarray}
\label{far-hhg}{E}_{\text h}^{\text f}(r_{\text f},z_{\text
f},\omega,\alpha)=&&-ik\int\frac{\tilde{E}_{\text
h}^{\parallel}(r,z',\omega,\alpha)}{z_{\text f}-z'}J_{0}(\frac{k
rr_{\text f}}{z_{\text f}-z'})\nonumber\\&&\times \exp[\frac{i
k(r^{2}+r_{\text f}^{2})}{2(z_{\text f}-z')}] r dr,
\end{eqnarray}
where $J_{0}$ is the zero-order Bessel function, $z_{\text f}$ and $r_{\text f}$
are the coordinates of the far-field points. The wave vector $k$
is given by $k=\omega/c$.
We assume that the harmonics in the far field are collected from an
extended area. By integrating harmonic yields over the area, the
power spectrum of the macroscopic harmonics is obtained by
\begin{eqnarray}
\label{total-hhg}S_{\text
h}(\omega,\alpha)\propto\int\int|{E}_{\text h}^{\text f}(x_{\text
f},y_{\text f},z_{\text f},\omega,\alpha)|^{2}dx_{\text f}dy_{\text
f},
\end{eqnarray}
where $x_{\text f}$ and $y_{\text f}$ are the Cartesian coordinates
on the plane perpendicular to the propagation direction, and
$r_{\text f}=\sqrt{x_{\text f}^{2}+y_{\text f}^{2}}$. To simulate
experimental HHG spectra quantitatively, besides laser parameters,
detailed information on the experimental setup is needed.
\subsection{Quantitative rescattering theory for a multi-orbital molecular system}
In Eq.~(\ref{pola}), laser induced single-molecule dipole moment
$D(t')$ is calculated quantum mechanically using the QRS theory,
which has been discussed in detail for molecules in Ref.
\cite{at-pra-2009}. Within QRS, laser induced dipole moment
$D(\omega,\theta)$ for a molecule at a fixed angle $\theta$
(measured with respect to laser polarization) is given by
\begin{eqnarray}
\label{mol-qrs}D^{\parallel,\perp}(\omega,\theta)=N(\theta)^{1/2}
W(\omega)d^{\parallel,\perp}(\omega,\theta),
\end{eqnarray}
where $N(\theta)$ is the alignment-dependent ionization probability,
$W(\omega)$ is the recombining electron wave packet, and
$d^{\parallel,\perp}(\omega,\theta)$ is the parallel component (or
perpendicular component) of the photorecombination (PR) transition dipole
(complex in general). Here we only consider linearly polarized lasers and
linear molecules. $W(\omega)$ is independent of the alignment angle $\theta$.
From Eq. (9), it can be expressed as
\begin{eqnarray}
\label{mol-wave}W(\omega)=\frac{D^{\parallel,\perp}(\omega,\theta)}
{N(\theta)^{1/2}d^{\parallel,\perp}(\omega,\theta)}.
\end{eqnarray}
In QRS, $W(\omega)$ is usually calculated only once for a given
angle $\theta$ using SFA, where all the elements on the right-hand
side of Eq.~(\ref{mol-wave}) are obtained by replacing the continuum
waves with plane waves. In QRS, accurate
$d^{\parallel,\perp}(\omega,\theta)$ are obtained from quantum
chemistry code \cite{Lucchese-co2-82,Lucchese-n2-82} and tunneling
ionization probability $N(\theta)$ are obtained either from MO-ADK
\cite{tong-pra-2002,zhao-pra-10} or from SFA. Put all of these
together into Eq.~(\ref{mol-qrs}), laser induced dipole moment
$D(\omega,\theta)$ for each orbital is obtained. Note that in QRS,
the wave packet $W(\omega)$ automatically includes the phase which
is dependent on the molecular orbital. Thus there is no need to
introduce the relative phase between orbitals, in contrast to the
approach in Ref. \cite{olga-nature-2009}. SFA is used to calculate
the ionization probability in Eq.~(\ref{mol-qrs}) in this paper
unless otherwise stated.
We assume that the degree of molecular alignment is not varied
spatially within the medium because the molecules are usually
partially aligned by a weak and loosely focused laser beam
\cite{jin-pra-2010}. By coherently averaging the induced dipole
moments over the molecular angular distribution, we obtain the
averaged induced dipole of partially aligned molecules at each
point inside the medium:
\begin{eqnarray}
\label{avg-in-dip}D^{\parallel,\text
{avg}}(\omega,\alpha)=\int^{\pi}_{0}D^{\parallel}(\omega,\theta)\rho(\theta,\alpha)\sin\theta
d\theta,
\end{eqnarray}
where $\rho(\theta,\alpha)$ is again the angular (or alignment)
distribution of the molecules with respect to the polarization
direction of the probe laser. For randomly distributed molecules,
$\rho(\theta,\alpha)$ is a constant. Note that the above procedure
is only for the specified molecular orbital.
For the specific problem addressed in this paper, we consider electrons either
in the HOMO (1$\pi_{g}$) or in the HOMO-2 (3$\sigma_{u}$) of CO$_{2}$. The
electrons are freed and then recombine back to the same orbital. As discussed
above, the multiple orbital effects are important at small alignment angles
only due to symmetry consideration. At these angles HOMO-1 (1$\pi_{u}$) is
ignored since the ionization rate of HOMO-1 is quite small compared with HOMO
and HOMO-2 \cite{zhao-pra-10,Spanner-pra-2009,olga-nature-2009}. At large
alignment angles, only HOMO becomes important. The total laser induced dipole
over a number of active electrons can be written as
\cite{Madsen-2006,Faria-pra-2010}
\begin{eqnarray}
\label{total-dip}D^{\parallel,\text
{tot}}(\omega,\alpha)=\sum_{j,n}D_{j,n}^{\parallel,\text
{avg}}(\omega,\alpha),
\end{eqnarray}
where index $j$ refers to the HOMO and HOMO-2 of CO$_{2}$, and $n$ is an index
to account for degeneracy in each molecular orbital.
Before proceeding, we comment that the QRS theory is formulated in
the energy (or frequency) domain. There is no explicit treatment of
``core dynamics" as in Ref. \cite{olga-nature-2009}. The time
evolution of the core is reflected only by the energy of the core in
the free propagation of the electron wave packet. In other words,
any possible many-body interchannel couplings between HOMO and
HOMO-2 are not included in the present treatment. Before such
effects are addressed, other factors that are more important on the
HHG as discussed here have to be treated first.
For the propagation of harmonics in the medium, we need to obtain
hundreds of the total induced dipoles $D^{\parallel,\text
{tot}}(\omega,\alpha)$ for different laser intensities, and then
they are fed into Eq.~(\ref{harm-freq}). The same procedure is used
in Jin {\it et al.} \cite{jin-pra-2009} for atomic targets. Note
that in the present model, dielectric properties of molecules due to
non-isotropic distributions are also neglected.
\section{Results and discussion}
\subsection{HHG spectra of randomly distributed CO$_{2}$: theory vs experiment}
\begin{figure*}
\mbox{\rotatebox{270}{\myscaleboxa{
\includegraphics{Fig1.eps}}}}
\caption{(Color online) Comparison of theoretical (green lines) and
experimental (red lines) HHG spectra of random and aligned CO$_{2}$
molecules, in an 800-nm laser shown in (a) and (d), and in a 1200-nm
laser shown in (b), (c), (e), and (f). Laser intensities are
indicated where I$_{\circ}$=10$^{14}$ W/cm$^{2}$. Experimental data
are from Ref. \cite{hans-prl-10}. Arrows indicate the positions of
minima. Pump-probe angle $\alpha$=0$^{\circ}$. See text for
additional laser parameters and experimental arrangements.
\label{Fig1}}
\end{figure*}
HHG spectra by 800-nm and 1200-nm lasers have been reported for
isotropically distributed and partially aligned N$_{2}$ and CO$_{2}$
molecules \cite{hans-prl-10}. The spectra of N$_{2}$ have been
analyzed by Jin {\it et al.} \cite{jin-pra-2011,jin-jpb-2011}
recently including only the HOMO.
In Figs.~\ref{Fig1}(a)-\ref{Fig1}(c), we show the HHG spectra for
isotropically distributed CO$_{2}$ molecules by 800-nm and 1200-nm
lasers. To obtain good agreement between theory and experiment,
especially in the cutoff region, the intensity used in the theory is
adjusted from the value given in the experiment. In
Figs.~\ref{Fig1}(a)-\ref{Fig1}(c), the intensities in theory
(experiment) are 1.9 (2.1), 0.8 (1.0), 1.0 (1.2), in units of
10$^{14}$W/cm$^{2}$, respectively. Other parameters used in the
simulation are the same as those given in the experiment
\cite{hans-prl-10}. The laser parameters are: pulse duration is
$\sim$ 32 fs (800 nm) or $\sim$ 44 fs (1200 nm), beam waist at the
focus is $\sim$ 40 $\mu$m. A 0.6-mm-wide gas jet is located 3 mm
(800 nm) or 3.5 mm (1200 nm) after the laser focus, and a slit with
a width of 100 $\mu$m is placed at 24 cm after the gas jet.
Figs.~\ref{Fig1}(a)-\ref{Fig1}(c) clearly show the good overall
agreement between experiment and theory for randomly distributed
CO$_{2}$ molecules. We have checked that HOMO is dominant for
randomly distributed CO$_2$, with negligible contributions from
inner orbitals. The HHG spectra do not exhibit any minima, as
opposed to the spectra of random N$_{2}$ molecules when they are
generated under the same experimental conditions \cite{hans-prl-10}.
For randomly distributed CO$_{2}$, there was no minimum found in HHG
spectra using an 800-nm laser by Vozzi {\it et al.}
\cite{Vozzi-prl-2005}. However, for the 1300-nm lasers, the data
from Torres {\it et al.} \cite{Torres-pra-2010,Torres-oe-2010}
appear to show a very weak minimum at photon energy near 45 eV.
Without more careful study including different intensities and
wavelengths, however, this is not conclusive.
\subsection{ HHG spectra of aligned CO$_{2}$: theory vs experiment}
Experimentally HHG spectra have also be reported from aligned
CO$_{2}$ molecules. A relatively weak and short laser pulse was used
to impulsively align molecules, and the HHG spectra were taken at
half-revival ($\sim$ 21.2 ps in CO$_{2}$) when the molecules were
maximally aligned \cite{hans-prl-10}. The angular distributions of
the aligned molecules are obtained by solving the TDSE of rotational
wave packet \cite{jin-pra-2010}. The degree of alignment is
$\langle\cos^{2}\theta\rangle$=0.60 in Fig.~\ref{Fig1}(d), and
$\langle\cos^{2}\theta\rangle$=0.50 in Figs.~\ref{Fig1}(e)
and~\ref{Fig1}(f). The polarizations of the pump and probe lasers
are parallel to each other.
The HHG spectra of partially aligned CO$_{2}$ molecules are shown in
Figs.~\ref{Fig1}(d)-\ref{Fig1}(f), which are obtained under the same
generating lasers and experimental arrangements as those in
Figs.~\ref{Fig1}(a)-\ref{Fig1}(c), respectively. The simulation and
experimental data agree well with each other in general. In
Fig.~\ref{Fig1}(e), the discrepancy is a little bigger, showing the
drop near 40 eV is larger in the experiment than in the theory. But
we note that in Fig.~\ref{Fig1}(f), the experimental data do not
drop as rapidly, in agreement with the theoretical simulation.
The minima in the HHG spectra of CO$_2$ and their dependence on laser intensity
have been widely discussed in the literature
\cite{Torres-pra-2010,olga-nature-2009}. In Fig.~\ref{Fig1}(d), for an 800-nm
laser, experiment gives a strong minimum at 42$\pm$2 eV, our simulation
predicts a minimum around 42 eV. For the 1200-nm laser, in Fig.~\ref{Fig1}(e),
experiment shows a minimum at 51$\pm$2 eV, theory predicts a minimum around 50
eV. In Fig.~\ref{Fig1}(f), the experimental minimum is shifted to 57$\pm$2 eV,
and the theoretical one is moved to around 53.5 eV. Thus our simulation also
shows the shift of the minimum from low to high harmonic orders as laser
intensity is increased. Below we interpret the origin of the shift.
\subsection{Origin of minimum in the HHG spectra of aligned CO$_{2}$: Type I and Type II}
\begin{figure*}
\mbox{\rotatebox{270}{\myscaleboxa{
\includegraphics{Fig2.eps}}}}
\caption{(Color online) (a)-(c) Macroscopic HHG spectra (envelope
only) corresponding to Figs.~\ref{Fig1}(d)-\ref{Fig1}(f),
respectively. Total (HOMO and HOMO-2 together) spectra and the
spectra of individual HOMO and HOMO-2 are shown. (d)-(f) Averaged
photorecombination transition dipoles (parallel component, the
square of magnitude) of HOMO and HOMO-2 corresponding to (a), (b),
and (c), respectively. Laser intensities are indicated where
I$_{\circ}$=10$^{14}$ W/cm$^{2}$. Arrows indicate the positions of
minima. Pump-probe angle $\alpha$=0$^{\circ}$. \label{Fig2}}
\end{figure*}
We next analyze the origin of the minimum in the HHG spectra seen in
Figs.~\ref{Fig1}(d)-\ref{Fig1}(f), and consider the dominant contributions from
HOMO and HOMO-2 only. First we define the averaged PR transition dipole for
each molecular orbital by
\begin{eqnarray}
\label{dip-avg}d^{\parallel,\text
{avg}}(\omega,\alpha)=\int^{\pi}_{0}N(\theta)^{1/2}
d^{\parallel}(\omega,\theta)\rho(\theta,\alpha)\sin\theta
d\theta.
\end{eqnarray}
This gives a measure of the relative contribution of each molecular
orbital to the HHG, which is obtained by averaging over the angular
(or alignment) distribution of the partially aligned molecules,
weighted by the square root of the tunneling ionization probability.
The relative ionization rates between HOMO and HOMO-2 change with
laser intensity.
Figs.~\ref{Fig2}(a)-\ref{Fig2}(c) show the envelopes of the HHG
spectra from individual molecular orbitals together with the total
ones, each obtained after propagation in the medium. In the
meanwhile, the averaged PR transition dipoles of HOMO and HOMO-2
under different generating lasers and alignment distributions are
shown in Figs.~\ref{Fig2}(d)-\ref{Fig2}(f), respectively.
In Figs.~\ref{Fig2}(a) and~\ref{Fig2}(b), there are no minima in the
HHG spectra of HOMO or HOMO-2, but the minimum shows up in the total
spectra. This is due to the interference between HOMO and HOMO-2. We
call this type I minimum. Clearly the minimum position will change
with laser intensity since the relative ionization rates between
HOMO and HOMO-2 change with intensity [also see
Figs.~{\ref{Fig5}}(c) and {\ref{Fig5}}(d)]. Similar analysis can be
found in Ref. \cite{olga-nature-2009}. In Fig.~\ref{Fig2}(c), there
is a minimum in the HOMO spectra at 52.6 eV. This minimum is shifted
to 53.6 eV in the total spectra due to the interference with the
HOMO-2. This is categorized as type II minimum. Similar analysis of
this type can be found in Refs. \cite{Torres-pra-2010,olga-pnas-09}.
The minimum in the HOMO spectra is due to the minimum in the
averaged PR transition dipole of HOMO shown in Fig.~\ref{Fig2}(f).
But their positions differ due to modification of the macroscopic
wave packet (MWP). In this connection we mention that the earlier
calculations \cite{le-prl-2009,at-pra-2009} with an 800-nm laser
showed minimum in HHG spectra at small pump-probe angles due to the
contribution from the HOMO only. These calculations were carried out
with a higher degree of alignment and higher laser intensities as
compared to our present study. This is expected as the minimum in
the averaged PR transition dipole from HOMO in Fig.~\ref{Fig2}(d)
becomes deeper and is slightly shifted away from the cutoff to lower
energies with increased degree of alignment [see Fig.~\ref{Fig2}(e)
and \ref{Fig2}(f)]. Furthermore, as shown in Ref.
\cite{jin-jpb-2011}, the minimum in the HOMO spectra could also
result from the multiplication of MWP and averaged PR transition
dipole even when neither has minimum. When a minimum occurs in the
dominant orbital, its position will not change much if it remains
the dominant one when the laser intensity changes. The little bump
around 36 eV in the HOMO spectra as well as in the total spectra can
be seen due to the bump in the HOMO curves in
Figs.~\ref{Fig2}(d)-\ref{Fig2}(f). Its position does not change much
since the HOMO-2 remains small.
Previously in Refs. \cite{jin-jpb-2011,jin-pra-2011}, we have shown
that the macroscopic HHG spectrum is the product of a MWP and an
averaged PR transition dipole for each individual molecular orbital.
Since the ionization rate for each orbital has been incorporated in
the averaged PR transition dipole, the MWP is mostly identical
except for the phase due to ionization potential. The averaged PR
transition dipole is very sensitive to ionization rates. The
relative magnitude changes rapidly with the increase of laser
intensity. Thus when two averaged PR transition dipoles are
comparable [see Fig.~\ref{Fig2}(d)], the position of the minimum
changes rapidly with the laser intensity. The averaged PR transition
dipole is also sensitive to alignment distributions [see
Figs.~\ref{Fig2}(d)-\ref{Fig2}(f)]. At low laser intensity, HOMO-2
is small, interference often occurs in a narrow region only where
the two amplitudes are comparable, see Figs.~\ref{Fig2}(b) and
\ref{Fig2}(c). In comparison, in Smirnova {\it et al.}
\cite{olga-nature-2009}, HOMO and HOMO-2 tend to interfere over a
broad photon-energy region. The ionization rates and transition
dipoles used in their calculations are different from ours.
\subsection{Progression of harmonic minimum vs laser intensity}
\begin{figure}
\mbox{\rotatebox{270}{\myscaleboxa{
\includegraphics{Fig3.eps}}}}
\caption{(Color online) Laser intensity dependence of macroscopic HHG spectra
(envelope only) (a) in an 800-nm laser and (b) in a 1200-nm laser.
Intensities are shown in units of I$_{\circ}$=10$^{14}$ W/cm$^{2}$. Arrows
indicate the positions of minima. Degree of alignment:
$\langle\cos^{2}\theta\rangle$=0.60. Pump-probe angle $\alpha$=0$^{\circ}$.
\label{Fig3}}
\end{figure}
In Figs.~{\ref{Fig3}}(a) and {\ref{Fig3}}(b) we show the envelope of
the calculated HHG spectra for four different peak intensities with
an 800-nm laser and a 1200-nm laser, respectively. For the 800-nm
spectra, the lowest intensity does not have a minimum. For the
higher ones, each spectrum has a type I minimum, with its position
shifts to lower photon energy as the laser intensity is decreased.
The degree of alignment of molecules used in the calculation is
$\langle\cos^{2}\theta\rangle$=0.60. We find that the shift cannot
be attributed to either MWP or the averaged PR transition dipole
alone. For the 1200-nm data, also with
$\langle\cos^{2}\theta\rangle$=0.60, which is different from
Figs.~{\ref{Fig1}}(e) and {\ref{Fig1}}(f), we find that the minimum
is type II, where the averaged PR transition dipole of the HOMO has
a minimum. The minimum in the HHG spectra of the HOMO shifts to
higher photon energy as the intensity increases, but the
interference with HOMO-2 shifts the minimum to even higher energies.
In other words, the shift of the position of the HHG minimum vs
intensity cannot be attributed to a single factor alone.
\subsection{Other factors that influence the precise positions of HHG minima}
\begin{figure*}
\mbox{\rotatebox{270}{\myscaleboxa{
\includegraphics{Fig4.eps}}}}
\caption{(Color online) Dependence of macroscopic HHG spectra
(envelope only) with degrees of molecular alignment distributions
for (a) an 800-nm laser with intensity of 1.8$\times$10$^{14}$
W/cm$^{2}$, and (b) a 1200-nm laser with intensity of
1.0$\times$10$^{14}$ W/cm$^{2}$. The weighted angular distributions
of the molecules are shown in (c). Arrows indicate the positions of
minima. Pump-probe angle $\alpha$=0$^{\circ}$.\label{Fig4}}
\end{figure*}
In our analysis, the averaged PR transition dipole is calculated over the
angular distribution of the molecules and thus depend on the degrees of
alignment. Since the latter cannot be accurately measured, we check how
sensitive the calculated spectra is with respect to the assumed alignment
distribution. In Fig.~{\ref{Fig4}}(c), four different alignment distributions
are shown. The distributions are multiplied by the volume element $\sin\theta$
for easy comparison. Three of them are obtained from the calculated rotational
wave packets \cite{jin-pra-2010}, with $\langle\cos^{2}\theta\rangle$ as 0.63,
0.60, and 0.55, respectively. The other is the commonly used $\cos^{4}\theta$
distribution. For 800-nm and 1200-nm lasers, the envelopes of the calculated
HHG spectra are shown in Figs.~{\ref{Fig4}}(a) and {\ref{Fig4}}(b),
respectively. The precise position of the minimum changes slightly except for
the one from the $\cos^{4}\theta$ distributions. However, change of a couple of
eV's is seen.
\begin{figure*}
\mbox{\rotatebox{0}{\myscaleboxc{
\includegraphics{Fig5.eps}}}}
\caption{(Color online) Dependence of macroscopic HHG spectra
(envelope only) on the ionization probabilities calculated from
MO-ADK or SFA in (a) an 800-nm laser, and (b) a 1200-nm laser. Laser
intensities are indicated where I$_{\circ}$=10$^{14}$ W/cm$^{2}$.
Arrows indicate the positions of minima. Degree of alignment:
$\langle\cos^{2}\theta\rangle$=0.60. Pump-probe angle
$\alpha$=0$^{\circ}$. (c) and (d) Alignment-dependent ionization
probabilities of HOMO and HOMO-2 calculated using MO-ADK and SFA.
Laser parameters are the same as (a) and (b). Ionization
probabilities of HOMO-2 in (d) are multiplied by 5. \label{Fig5}}
\end{figure*}
To precisely determine the minimum in the HHG spectra, accurate
alignment-dependent ionization probability $N(\theta)$ for each
molecular orbital is needed. For CO$_{2}$, even for HOMO, different
theories in the literature
\cite{zhao-pra-10,Spanner-pra-2009,olga-nature-2009,Petretti-prl-2010,
Pavi-prl-2007,hoang-jpb-2008,madsen-pra-2009,chu-pra-2009} show
non-negligible differences, and they do not agree with the
experimental data \cite{Pavi-prl-2007}. Here we examine how the HHG
spectra change with the different ionization rates used. The
ionization rates for both HOMO and HOMO-2 can all be easily
calculated from SFA or from MO-ADK theory. Figs.~{\ref{Fig5}}(a) and
{\ref{Fig5}}(b) show the HHG spectra calculated using the ionization
probabilities shown in Figs.~{\ref{Fig5}}(c) and {\ref{Fig5}}(d).
Other laser parameters used in the calculation are given in the
figure captions. The difference of the position of the minimum is 3
eV in Fig.~{\ref{Fig5}}(a) and 2 eV in Fig.~{\ref{Fig5}}(b). Note
that the ionization probabilities from SFA and MO-ADK are normalized
at the peak of the HOMO curve. In Fig.~{\ref{Fig5}}(a) the spectra
are normalized at H33 (51 eV) and in Fig.~{\ref{Fig5}}(b) at H65 (67
eV).
\begin{figure}
\mbox{\rotatebox{270}{\myscaleboxa{
\includegraphics{Fig6.eps}}}}
\caption{(Color online) Dependence of macroscopic HHG spectra
(envelope only) on experimental arrangements (a) for an 800-nm laser
with intensity of 2.1$\times$10$^{14}$ W/cm$^{2}$, and (b) for a
1200-nm laser with intensity of 1.2$\times$10$^{14}$ W/cm$^{2}$. The
arrangements are: (1) gas jet after focus and slit is used (solid
lines); (2) gas jet at the focus and slit is used (dashed lines);
and (3) gas jet is after the focus but without the slit (dot-dashed
lines). Arrows indicate the positions of minima. Degree of
alignment: $\langle\cos^{2}\theta\rangle$=0.60. Pump-probe angle
$\alpha$=0$^{\circ}$. \label{Fig6}}
\end{figure}
The HHG spectra are also sensitive to the experimental arrangement
and thus can also move the position of the HHG minimum. To
demonstrate this, we (i) move the gas jet to the laser focus and
collect the signal using a slit; (ii) put the gas jet after the
laser focus, and collect HHG signal without slit (total signal).
These two will be compared to the arrangement used in this paper:
gas jet is after the laser focus and the HHG is collected with a
slit. The results are compared in Fig.~{\ref{Fig6}}. Note that the
spectra are normalized at H17 (26 eV) in Fig.~{\ref{Fig6}}(a) and at
H35 (36 eV) in Fig.~{\ref{Fig6}}(b). Not only the spectra change
quite significantly, but also the position of the HHG minimum. This
illustrates that it is very difficult to compare the position of the
HHG minimum from different experiments. In this comparison, the
change of the HHG spectra is due to the change of MWP which depends
on the experimental setup. The averaged PR transition dipoles of
HOMO and HOMO-2 are the same in the three calculations.
\subsection{The dependence of the HHG minimum on the pump-probe angle}
\begin{figure*}
\mbox{\rotatebox{270}{\myscaleboxa{
\includegraphics{Fig7.eps}}}}
\caption{(Color online) Dependence of macroscopic HHG spectra
(envelope only) on pump-probe angles $\alpha$=10$^{\circ}$,
20$^{\circ}$, 30$^{\circ}$, and 40$^{\circ}$ (a) for an 800-nm
laser, (b) and (c) for a 1200-nm laser. Laser intensities are given
in units of I$_{\circ}$=10$^{14}$ W/cm$^{2}$. Arrows indicate the
positions of minima. Degree of alignment:
$\langle\cos^{2}\theta\rangle$=0.60. (d)-(f) Photorecombination
transition dipoles (parallel component, the square of magnitude) of
HOMO and HOMO-2 in terms of photon energy at alignment angles
$\theta$=30$^{\circ}$, 60$^{\circ}$, and 80$^{\circ}$. \label{Fig7}}
\end{figure*}
In W\"{o}rner {\it et al.} \cite{hans-prl-10}, it was found that the
minimum in the HHG spectra of aligned CO$_{2}$ moved to higher
photon energy with increasing pump-probe angle $\alpha$, i.e., the
angle between aligning pump beam and the HHG generating probe beam
polarizations. This phenomenon has also been observed in other
measurements \cite{Mairesse-jmo-2008,olga-nature-2009}. In contrast,
the minimum in the HHG spectra of aligned N$_{2}$ has been found to
stay at the same position for all pump-probe angles
\cite{Mairesse-jmo-2008,Haessler-NatPhys-2010}.
We show the calculated HHG spectra at four pump-probe angles in an
800-nm laser in Fig.~{\ref{Fig7}}(a), and in a 1200-nm laser in
Figs.~{\ref{Fig7}}(b) and {\ref{Fig7}}(c). Laser parameters and
experimental arrangements are the same as in Fig.~{\ref{Fig1}} and
the degree of alignment is $\langle\cos^{2}\theta\rangle$=0.60. The
HHG spectra for $\alpha$=0$^{\circ}$ have been shown in
Fig.~{\ref{Fig3}}. These figures show that the minimum in the HHG
spectra moves to higher photon energies with increasing $\alpha$ ,
and the minimum disappears at large angles. At larger pump-probe
angles, contributions from large alignment angles increase. Since
HOMO dominates over HOMO-2 at large angles in both the ionization
rates [see Figs.~{\ref{Fig5}}(c) and {\ref{Fig5}}(d)] and the PR
transition dipoles [see Figs.~{\ref{Fig7}}(d)-{\ref{Fig7}}(f)], thus
HHG at large pump-probe angles has essentially no contributions from
HOMO-2. Also note that at small $\alpha$, the total HHG yield is
much smaller \cite{at-pra-2009}. In fact, the total HHG spectra for
randomly distributed CO$_2$ have little contributions from molecules
that are aligned nearly parallel to the polarization axis of the
laser. The HHG spectra of CO$_2$ are complex only in the region
where HHG yields are small. This is generally true -- interpretation
of small processes always requires careful and detailed theories.
\section{Summary and outlook}
In this paper we have analyzed the multiple orbital contribution to
HHG in CO$_2$ with the inclusion of macroscopic propagation effect.
In the past few years, there have been many experimental and
theoretical studies on the HHG of CO$_2$ from many laboratories,
using lasers with different wavelengths and intensities, for CO$_2$
molecules that are randomly distributed or partially aligned. In
particular, for CO$_2$ molecules that are partially aligned along
the polarization axis of the probe laser, many experiments have
shown that the HHG spectra exhibit minima and the positions of the
minima shift with laser intensities
\cite{olga-nature-2009,hans-prl-10,Torres-pra-2010}. The shift of
the minimum position with laser intensities has been attributed to
the interference between the contributions to HHG from the HOMO-2
with the one from the HOMO, despite HOMO-2 lying at 4.4 eV deeper
than the HOMO. Since HHG is a nonlinear process, these observations
posed a great challenge to the theory, especially for the prediction
of the position of the minimum and how it changes with the
experimental conditions. Since all experimental HHG spectra include
macroscopic propagation effect, comparison of theory with experiment
is incomplete unless the theory also has included the propagation
effect. Our analysis in this paper is based on the macroscopic
propagation code extended for aligned molecules, and the recently
developed quantitative rescattering theory. We find that although
HHG spectra change significantly under different experimental
parameters such as degree of alignment, focussing condition, and the
use of a slit, the position of the minimum in the HHG spectra
behaves in a similar trend as laser intensity and pump-probe angle
vary. This trend has been found to be consistent with the recent
experimental measurements from different groups.
We comment that the present theory and the earlier one by Smirnova
{\it et al.} \cite{olga-nature-2009} both explain the intensity
dependence of the change of HHG minima, but the details between the
two theories are quite different. The alignment dependence of the
ionization rates, the recombination dipole matrix elements and their
phases entering the two theories are not the same, for both the HOMO
and HOMO-2. As illustrated in this paper, these parameters can all
affect the position of the predicted interference minimum.
Furthermore, in Smirnova {\it et al.} \cite{olga-nature-2009} the
interference is attributed to the importance of hole dynamics in the
ion core. Our approach is formulated in the time-independent
fashion, no hole dynamics is included. Since HHG spectra are taken
without explicit observation of electron dynamics, the difference
between the two models cannot be settled. Despite these differences,
our understanding of the HHG spectra of CO$_2$ has come a long way
since 2005 \cite{Kanai-nature-2005}. With the possibility of
including macroscopic propagation effect ``routinely" in the HHG
theory for molecular targets, further experimental studies should
explore the effects of laser focusing condition and gas pressure,
for lasers extending to longer wavelengths. Such studies would
further our basic understanding of strong-field physics of molecules
to the next level, and eliminate the need of introducing extraneous
assumptions in the interpretation of experimental data.
\section{Acknowledgments}
This work was supported in part by Chemical Sciences, Geosciences
and Biosciences Division, Office of Basic Energy Sciences, Office of
Science, U.S. Department of Energy.
|
\section{Introduction and Main results}
The methods employed to study the convergence of iterates of some operators
include Matrix Theory methods, like stochastic matrices, Korovkin-type
theorems, quantitative results about the approximation of functions by
positive linear operators, point theorems, or methods from the theory of
$C_{0}$-semigroups, like Trotter's approximation theorem, see \cite{Alt1
-\cite{wenz}. However, these results fail to calculate the iterate limit and
give the quantitative estimates of iterates for many classical operators. In
respect to this, we mention recent works by I. Gavrea and M. Ivan
\cite{gavrea1}, \cite{gavrea2} and I. Rasa \cite{rasa1}, \cite{rasa2}.
In this paper we establish quantitative Korovkin type theorem for the iterates
of certain positive linear operators $T:C\left[ 0,1\right] \rightarrow
C\left[ 0,1\right] $ satisfying $T\left( e_{0}\right) =e_{0},\ \ T\left(
e_{1}\right) -e_{1}\leq0$ (or $\geq0$). As a consequence of our results, we
obtain the quantitaive estimates for the iterates of almost all classical and
new positive linear operators. Notice that quantitative Korovkin type theorems
for a sequence of positive linear operators are studied in \cite{wang2},
\cite{mah2}.
Let $C\left[ 0,1\right] $ be the set of all real-valued and continuous
functions defined on the compact interval $\left[ 0,1\right] $ endowed with
the $\sup$-norm $\left\Vert f\right\Vert :=\sup\left\{ \left\vert f\left(
x\right) \right\vert :x\in\left[ 0,1\right] \right\} $. $W_{2,\infty
}\left[ 0,1\right] $ is defined as follows
\begin{align*}
W_{2,\infty}\left[ 0,1\right] & :=\left\{ f\in C\left[ 0,1\right]
:f^{\prime}\text{ absolutely continuous, }\left\Vert f^{\prime\prime
}\right\Vert _{L_{\infty}}<\infty\right\} ,\\
\left\Vert f\right\Vert _{L_{\infty}} & :=\text{vrai}\max\left\{ \left\vert
f^{\prime\prime}\left( x\right) \right\vert :0\leq x\leq1\right\} ,
\end{align*}
and $L_{\infty}$ is the space of essentially bounded measurable functions
endowed with $\left\Vert \cdot\right\Vert _{L_{\infty}}$ norm.
The main tools to measure the degree of convergence of the powers of positive
linear operators are the moduli of smoothness of first and second order. For
$f\in C\left[ 0,1\right] $ and $\delta\geq0$ we hav
\begin{align*}
\omega_{1}\left( f;\delta\right) & :=\sup\left\{ \left\vert f\left(
x+h\right) -f\left( x\right) \right\vert :x,x+h\in\left[ 0,1\right]
,0\leq h\leq\delta\right\} ,\\
\omega_{2}\left( f;\delta\right) & :=\sup\left\{ \left\vert f\left(
x+h\right) -2f\left( x\right) +f\left( x-h\right) \right\vert :x,x\pm
h\in\left[ 0,1\right] ,0\leq h\leq\delta\right\} .
\end{align*}
For $f\in C\left[ 0,1\right] $ we define the extension $f_{h}:\left[
-h,1+h\right] \rightarrow\mathbb{R},$ with $h>0$, b
\[
f_{h}\left( x\right) :=\left\{
\begin{array}
[c]{c
P_{-}\left( x\right) ,\ \ \ -h\leq x\leq0,\\
f\left( x\right) ,\ \ \ 0\leq x\leq1,\\
P_{+}\left( x\right) ,\ \ \ 1\leq x\leq1+h,
\end{array}
\right.
\]
where $P_{-}$ and $P_{+}$ are at most order best approximants to $f$ on the
indicated intervals. Then Zhuk's function $Z_{h}f$ is defined by means of the
second order Steklov mean
\[
Z_{h}f\left( x\right) :=\frac{1}{h}\int_{-h}^{h}\left( 1-\frac{\left\vert
t\right\vert }{h}\right) f_{h}\left( x+t\right) dt,\ \ 0\leq x\leq1.
\]
It can be shown that $Z_{h}f\in W_{2,\infty}\left[ 0,1\right] .$ It is known
that for sufficiently large $l$ and a fixed $\varepsilon>0$ we hav
\begin{align}
\left\Vert f-g\right\Vert & \leq\left\Vert f-Z_{h}f\right\Vert +\left\Vert
B_{l}\left( Z_{h}f\right) -Z_{h}f\right\Vert \leq\frac{3}{4}\omega
_{2}\left( f;h\right) +\varepsilon,\nonumber\\
\left\Vert g^{\prime}\right\Vert & \leq\left\Vert \left( Z_{h}f\right)
^{\prime}\right\Vert \leq\frac{1}{h}\left( 2\omega_{1}\left( f;h\right)
+\frac{3}{2}\omega_{2}\left( f;h\right) \right) ,\nonumber\\
\left\Vert g^{\prime\prime}\right\Vert & \leq\left\Vert \left( Z_{\delta
}f\right) ^{\prime\prime}\right\Vert _{L_{\infty}}\leq\frac{1}{\delta^{2
}\frac{3}{2}\omega_{2}\left( f;\delta\right) .\label{zhuk
\end{align}
For any positive linear operator $T:C\left[ 0,1\right] \rightarrow C\left[
0,1\right] $, we define the powers of $T$ b
\[
T^{0}=I,\ \ \ T^{1}=T,\ \ \ T^{m+1}=T\circ T^{m},\ \ \ m\in\mathbb{N}.
\]
Let $e_{i}:\left[ 0,1\right] \rightarrow R$ be the monomial functions
$e_{i}\left( x\right) =x^{i},$ $i=0,1,2.$
Now we formulate the main results of the paper. It shows that under the
conditions $T\left( e_{0}\right) =1\ $and $T\left( e_{1}\right) -e_{1}$
does not change the sign, the iterates of $T:C\left[ 0,1\right] \rightarrow
C\left[ 0,1\right] $ converges to some linear positive operator $T^{\infty
}:C\left[ 0,1\right] \rightarrow C\left[ 0,1\right] .$
\begin{theorem}
\label{thm0} Suppose that $T:C\left[ 0,1\right] \rightarrow C\left[
0,1\right] $ is a positive linear operator such that $T\left( e_{0}\right)
=e_{0}.$
\begin{enumerate}
\item[(a)] If $T\left( e_{1}\right) \leq e_{1}$, then there exists a linear
positive positive operator $T^{\infty}:C\left[ 0,1\right] \rightarrow
C\left[ 0,1\right] $ such that
\[
\lim_{n\rightarrow\infty}T^{m}\left( f\right) =T^{\infty}\left( f\right)
,\ \ \ f\in C\left[ 0,1\right] ,
\]
and the following pointwise estimat
\begin{align}
\left\vert T^{\infty}\left( f;x\right) -T^{m}\left( f;x\right)
\right\vert & \leq3\omega_{2}\left( f;\left\vert \left( T^{m}-T^{\infty
}\right) \left( e_{1};x\right) \right\vert \right) +2\omega_{1}\left(
f;\left\vert \left( T^{m}-T^{\infty}\right) \left( e_{1};x\right)
\right\vert \right) \nonumber\\
& +\frac{3}{4}\omega_{2}\left( f;\sqrt{\left\vert \left( T^{\infty
-T^{m}\right) \left( e_{2};x\right) \right\vert +2\left\vert \left(
T^{m}-T^{\infty}\right) \left( e_{1};x\right) \right\vert }\right)
\label{ss1
\end{align}
holds true \textit{for }$x\in\left[ 0,1\right] $\textit{ and }$f\in C\left[
0,1\right] .$
\item[(b)] If $T\left( e_{1}\right) \geq e_{1}$, then there exists a linear
positive positive operator $T^{\infty}:C\left[ 0,1\right] \rightarrow
C\left[ 0,1\right] $ such that
\[
\lim_{n\rightarrow\infty}T^{m}\left( f\right) =T^{\infty}\left( f\right)
,\ \ \ f\in C\left[ 0,1\right] ,
\]
and the following pointwise estimat
\begin{align}
\left\vert T^{\infty}\left( f;x\right) -T^{m}\left( f;x\right)
\right\vert & \leq3\omega_{2}\left( f;\left\vert \left( T^{m}-T^{\infty
}\right) \left( e_{1};x\right) \right\vert \right) +2\omega_{1}\left(
f;\left\vert \left( T^{m}-T^{\infty}\right) \left( e_{1};x\right)
\right\vert \right) \nonumber\\
& +\frac{3}{4}\omega_{2}\left( f;\sqrt{\left\vert \left( T^{\infty
-T^{m}\right) \left( e_{2};x\right) \right\vert }\right) \label{sss1
\end{align}
holds true \textit{for }$x\in\left[ 0,1\right] $\textit{ and }$f\in C\left[
0,1\right] .$
\end{enumerate}
\end{theorem}
\begin{corollary}
\label{thm1} Suppose that $T:C\left[ 0,1\right] \rightarrow C\left[
0,1\right] $ is a positive linear operator such that
\begin{equation}
T\left( e_{0}\right) =e_{0},\ \ T\left( e_{1}\right) =e_{1}. \label{cc10
\end{equation}
Then there exists a linear positive operator $T^{\infty}:C\left[ 0,1\right]
\rightarrow C\left[ 0,1\right] $ such that
\[
\lim_{m\rightarrow\infty}\left\Vert T^{\infty}\left( f\right) -T^{m}\left(
f\right) \right\Vert =0,\ \ \ \mathit{\ }f\in C\left[ 0,1\right] .
\]
Furthermore the following pointwise estimat
\begin{equation}
\left\vert T^{\infty}\left( f;x\right) -T^{m}\left( f;x\right) \right\vert
\leq\frac{3}{4}\omega_{2}\left( f;\sqrt{\left\vert \left( T^{\infty
-T^{m}\right) \left( e_{2};x\right) \right\vert }\right) \label{rr1
\end{equation}
holds true \textit{for }$x\in\left[ 0,1\right] $ \textit{and }$f\in C\left[
0,1\right] .$
\end{corollary}
The second result shows that under the conditons $T\left( e_{0}\right)
=e_{0},\ T\left( e_{1}\right) =e_{1},$ $T^{\infty}\left( e_{2}\right)
=e_{1}$ the limit of the iteraes $T^{m}$ is exactly the operator $P\left(
f;x\right) :=\left( 1-x\right) f\left( 0\right) +xf\left( 1\right) ,$
and under the conditions $T\left( e_{0}\right) =e_{0},\ T\left(
e_{2}\right) =e_{2},$ $T^{\infty}\left( e_{1}\right) =e_{2}$ the limit of
the iteraes $T^{m}$ is exactly the operator $V\left( f;x\right) :=\left(
1-x^{2}\right) f\left( 0\right) +x^{2}f\left( 1\right) .$
\begin{theorem}
\label{t:11}Suppose that $T:C\left[ 0,1\right] \rightarrow C\left[
0,1\right] $ is a positive linear operator.
\begin{enumerate}
\item[(a)] If $T\left( e_{0}\right) =e_{0},\ T\left( e_{1}\right)
=e_{1},T^{\infty}\left( e_{2}\right) =e_{1},$ then $T^{\infty}=P$ and
\begin{equation}
\left\vert P\left( f;x\right) -T^{m}\left( f;x\right) \right\vert
\leq\frac{3}{4}\omega_{2}\left( f;\sqrt{\left\vert x-T^{m}\left(
e_{2};x\right) \right\vert }\right) . \label{ee1
\end{equation}
\item[(b)] If $T\left( e_{0}\right) =e_{0},\ \ \ T\left( e_{1}\right) \leq
e_{1},\ \ \ T\left( e_{2}\right) =e_{2},T^{\infty}\left( e_{1}\right)
=e_{2},$ then $T^{\infty}=V$ and
\begin{align*}
\left\vert V\left( f;x\right) -T^{m}\left( f;x\right) \right\vert &
\leq3\omega_{2}\left( f;\left\vert T^{m}\left( e_{1};x\right)
-x^{2}\right\vert \right) +2\omega_{1}\left( f;\left\vert T^{m}\left(
e_{1};x\right) -x^{2}\right\vert \right) \\
& +\frac{3}{4}\omega_{2}\left( f;\sqrt{2\left\vert T^{m}\left(
e_{1};x\right) -x^{2}\right\vert }\right) .
\end{align*}
\item[(c)] If $T\left( e_{0}\right) =e_{0},\ \ \ T\left( e_{1}\right) \geq
e_{1},\ \ \ T\left( e_{2}\right) =e_{2},T^{\infty}\left( e_{1}\right)
=e_{2},$ then $T^{\infty}=V$ and
\[
\left\vert V\left( f;x\right) -T^{m}\left( f;x\right) \right\vert
\leq3\omega_{2}\left( f;\left\vert T^{m}\left( e_{1};x\right)
-x^{2}\right\vert \right) +2\omega_{1}\left( f;\left\vert T^{m}\left(
e_{1};x\right) -x^{2}\right\vert \right) .
\]
\end{enumerate}
\end{theorem}
\begin{remark}
(i) Results similar to that of Theorem \ref{t:11} (a) without the estimation
(\ref{ee1}) was obtained in \cite{rasa1} and \cite{gavrea2}. (ii) Theorem
\ref{t:11} (b) and \ref{t:11} (c) are new. They cover positive linear
operators which preserve $e_{0}$ and $e_{2}$. Convergence of overiterates of
Bernstein operators and discrete type positive linear operators preserving
$e_{0}$ and $e_{2}$ is studied in \cite{agr}, \cite{gonska6}.
\end{remark}
\begin{corollary}
\label{t:iterate}Let $T:C\left[ 0,1\right] \rightarrow C\left[ 0,1\right]
$ be a positive linear operator such tha
\begin{equation}
T\left( e_{0}\right) =e_{0},\ \ T\left( e_{1}\right) =e_{1},\ T\left(
e_{2}\right) \leq ae_{2}+be_{1},a,b\in R\backslash\left\{ 0\right\}
,\ a+b=1. \label{cc11
\end{equation}
Then the pointwise approximation
\begin{equation}
\left\vert T^{m}\left( f;x\right) -P\left( f;x\right) \right\vert
\leq\frac{3}{4}\omega_{2}\left( f;\sqrt{a^{m}x\left( 1-x\right) }\right)
\label{est5
\end{equation}
holds true for all $x\in\left[ 0,1\right] $ and $f\in C\left[ 0,1\right] $.
\end{corollary}
\begin{remark}
It is worth mentioning that the conditions
\[
T\left( e_{0}\right) =e_{0},\ \ T\left( e_{1}\right) =e_{1},\ T\left(
e_{2}\right) =ae_{2}+be_{1},a,b\in R\backslash\left\{ 0\right\} ,\ a+b=1.
\]
are satisfied by the many classical positive linear operators defined on
$C\left[ 0,1\right] $ and convergence of overiterates under these conditions
was studied in \cite{gonska3}, \cite{rasa1}, \cite{gavrea2}. The conditions
(\ref{cc11}) cover the classical MKZ and $q$-MKZ operators. The problem of
convergence of overiterates of MKZ operators without quantitative estimate was
studied in \cite{gavrea1}. In Corollary \ref{t:iterate} we give quantitative
estimate for convergence.
\end{remark}
\begin{corollary}
\label{c:11}Suppose that $T:C\left[ 0,1\right] \rightarrow C\left[
0,1\right] $ is a positive linear operator.
\begin{enumerate}
\item[(a)] If $T\left( e_{0}\right) =e_{0},\ \ \ T\left( e_{1}\right) \leq
e_{1},\ \ T^{\infty}\left( e_{1}\right) =T^{\infty}\left( e_{2}\right)
=0,$ then $T^{\infty}f=f\left( 0\right) $ an
\begin{align*}
\left\vert f\left( 0\right) -T^{m}\left( f;x\right) \right\vert &
\leq3\omega_{2}\left( f;\left\vert T^{m}\left( e_{1};x\right) \right\vert
\right) +2\omega_{1}\left( f;\left\vert T^{m}\left( e_{1};x\right)
\right\vert \right) \\
& +\frac{3}{4}\omega_{2}\left( f;\sqrt{\left\vert T^{m}\left(
e_{2};x\right) \right\vert +2\left\vert T^{m}\left( e_{1};x\right)
\right\vert }\right) .
\end{align*}
\item[(b)] If $T\left( e_{0}\right) =e_{0},\ \ \ T\left( e_{1}\right) \geq
e_{1},\ \ T^{\infty}\left( e_{1}\right) =T^{\infty}\left( e_{2}\right)
=1,$ then $T^{\infty}f=f\left( 1\right) $ and
\begin{align*}
\left\vert f\left( 1\right) -T^{m}\left( f;x\right) \right\vert &
\leq3\omega_{2}\left( f;\left\vert T^{m}\left( e_{1};x\right) -1\right\vert
\right) +2\omega_{1}\left( f;\left\vert T^{m}\left( e_{1};x\right)
-1\right\vert \right) \\
& +\frac{3}{4}\omega_{2}\left( f;\sqrt{\left\vert T^{m}\left(
e_{2};x\right) -1\right\vert }\right) .
\end{align*}
\end{enumerate}
\end{corollary}
\begin{remark}
Corollary \ref{c:11} covers the Berstein-Stancu operators $S_{n}^{\left\langle
\alpha,\beta,\gamma\right\rangle }$ for some values of $\alpha,\beta,\gamma,$
see \cite{gonska5}.
\end{remark}
\section{Proofs of the main results}
\begin{proof}
[Proof of Theorem \ref{thm0}](b) For every convex $g\in C^{2}\left[
0,1\right] ,$ we have
\begin{equation}
g\left( t\right) \geq g\left( x\right) +g^{\prime}\left( x\right)
\left( t-x\right) . \label{b0
\end{equation}
It follows that for any nondecreasing convex $g\in C^{2}\left[ 0,1\right] $
\begin{align*}
T\left( g;x\right) & \geq g\left( x\right) +g^{\prime}\left( x\right)
\left( T\left( e_{1};x\right) -x\right) \geq g\left( x\right) ,\\
g\left( x\right) & \leq T^{m}\left( g;x\right) \leq T^{m+1}\left(
g;x\right) \leq\left\Vert g\right\Vert .
\end{align*}
In other words the sequence continuous functions $\left\{ T^{m}\left(
g;\cdot\right) \right\} $ is nondecreasing for any nondecreasing convex
function $g\in C^{2}\left[ 0,1\right] .$ By Dini's theorem there exists
$T^{\infty}\left( g;\cdot\right) $ such that $T^{m}\left( g;\cdot\right)
\rightarrow T^{\infty}\left( g;\cdot\right) $ uniformly on $\left[
0,1\right] ,$ for any nonincreasing convex function $g\in C^{2}\left[
0,1\right] .$In particular
\begin{align*}
\lim_{m\rightarrow\infty}\left\Vert T^{m}\left( e_{1}\right) -T^{\infty
}\left( e_{1}\right) \right\Vert & =0,\\
\lim_{m\rightarrow\infty}\left\Vert T^{m}\left( e_{2}\right) -T^{\infty
}\left( e_{2}\right) \right\Vert & =0.
\end{align*}
Let $g\in C^{2}\left[ 0,1\right] $ be arbitrary. Introduce the following
auxiliary function
\[
g_{\pm}\left( t\right) =\frac{1}{2}\left\Vert g^{\prime\prime}\right\Vert
t^{2}+\left\Vert g^{\prime}\right\Vert t\pm g\left( t\right) .
\]
It is clear that
\[
g_{\pm}^{\prime}\left( t\right) =\left\Vert g^{\prime\prime}\right\Vert
t+\left\Vert g^{\prime}\right\Vert \pm g\left( t\right) \geq0,\ \ \ g_{\pm
}^{\prime\prime}\left( t\right) =\left\Vert g^{\prime\prime}\right\Vert \pm
g\left( t\right) \geq0.
\]
Therefore the functions $g_{\pm}\left( t\right) $ are nondecreasing convex
for both choices of the sign. We hav
\begin{align*}
0 & \leq T^{m+p}\left( g_{\pm};x\right) -T^{m}\left( g_{\pm};x\right)
=\frac{1}{2}\left\Vert g^{\prime\prime}\right\Vert \left( T^{m+p}\left(
e_{2};x\right) -T^{m}\left( e_{2};x\right) \right) \\
& +\left\Vert g^{\prime}\right\Vert \left( T^{m}\left( e_{1};x\right)
-T^{m+p}\left( e_{1};x\right) \right) \pm\left( T^{m+p}\left( g;x\right)
-T^{m}\left( g;x\right) \right) .
\end{align*}
It follows tha
\begin{equation}
\left\vert T^{m+p}\left( g;x\right) -T^{m}\left( g;x\right) \right\vert
\leq\frac{1}{2}\left\Vert g^{\prime\prime}\right\Vert \left\vert
T^{m+p}\left( e_{2};x\right) -T^{m}\left( e_{2};x\right) \right\vert
+\left\Vert g^{\prime}\right\Vert \left\vert T^{m}\left( e_{1};x\right)
-T^{m+p}\left( e_{1};x\right) \right\vert . \label{b1
\end{equation}
So $\left\{ T^{m}\left( g;x\right) \right\} $ is a Cauchy sequence in
$C\left[ 0,1\right] .$ Since $C\left[ 0,1\right] $ is complete there is a
function $f^{\infty}$ such tha
\[
\lim_{m\rightarrow\infty}\left\Vert T^{m}\left( g\right) -f^{\infty
}\right\Vert =0.
\]
Although this limit has been obtained for $g\in C^{2}\left[ 0,1\right] $
only, it extends to all $f\in C\left[ 0,1\right] $ by the Banach--Steinhaus
theorem. Hence we find an operator $T^{\infty}:C\left[ 0,1\right]
\rightarrow C\left[ 0,1\right] $ say, such that $T^{\infty}f:=\lim
_{m\rightarrow\infty}T^{m}\left( f\right) =f^{\infty},$ $f\in C\left[
0,1\right] $. Clearly, this operator is linear and positive.
Taking the limit as $p\rightarrow\infty$ in (\ref{b1}) we hav
\[
\left\vert \left( T^{\infty}-T^{m}\right) \left( g;x\right) \right\vert
\leq\frac{1}{2}\left\Vert g^{\prime\prime}\right\Vert \left\vert \left(
T^{\infty}-T^{m}\right) \left( e_{2};x\right) \right\vert +\left\Vert
g^{\prime}\right\Vert \left\vert \left( T^{m}-T^{\infty}\right) \left(
e_{1};x\right) \right\vert .
\]
Let $f\in C\left[ 0,1\right] $. For $g\in C^{2}\left[ 0,1\right] $
arbitrarily chosen we have the following estimat
\begin{gather}
\left\vert T^{\infty}\left( f;x\right) -T^{m}\left( f;x\right) \right\vert
\leq\left\vert \left( T^{\infty}-T^{m}\right) \left( f-g;x\right)
\right\vert +\left\vert T^{\infty}\left( g;x\right) -T^{m}\left(
g;x\right) \right\vert \nonumber\\
\leq2\left\Vert f-g\right\Vert +\frac{1}{2}\left\Vert g^{\prime\prime
}\right\Vert \left\vert \left( T^{\infty}-T^{m}\right) \left(
e_{2};x\right) \right\vert +\left\Vert g^{\prime}\right\Vert \left\vert
\left( T^{m}-T^{\infty}\right) \left( e_{1};x\right) \right\vert
\nonumber\\
=2\left\Vert f-g\right\Vert +\left\Vert g^{\prime}\right\Vert \left\vert
\left( T^{m}-T^{\infty}\right) \left( e_{1};x\right) \right\vert +\frac
{1}{2}\left\Vert g^{\prime\prime}\right\Vert \left\vert \left( T^{\infty
}-T^{m}\right) \left( e_{2};x\right) \right\vert . \label{b2
\end{gather}
We substitute now $g:=B_{n}\left( Z_{h}f\right) \in C^{2}\left[ 0,1\right]
,$ where $Z_{h}f$ is Zhuk's function. In (\ref{b2}) using the inequalities
(\ref{zhuk}) and letting $\varepsilon\rightarrow0$ we arrive a
\begin{align*}
\left\vert T^{\infty}\left( f;x\right) -T^{m}\left( f;x\right)
\right\vert & \leq\frac{3}{2}\omega_{2}\left( f;h\right) +\frac{1
{h}\left( 2\omega_{1}\left( f;h\right) +\frac{3}{2}\omega_{2}\left(
f;h\right) \right) \left\vert \left( T^{m}-T^{\infty}\right) \left(
e_{1};x\right) \right\vert \\
& +\frac{1}{\delta^{2}}\frac{3}{4}\omega_{2}\left( f;\delta\right)
\left\vert \left( T^{\infty}-T^{m}\right) \left( e_{2};x\right)
\right\vert
\end{align*}
with $h>0$ and $\delta>0$. I
\[
\left\vert \left( T^{m}-T^{\infty}\right) \left( e_{1};x\right)
\right\vert >0,\ \ \ \left\vert \left( T^{m}-T^{\infty}\right) \left(
e_{2};x\right) \right\vert >0
\]
taking $h=\left\vert \left( T^{m}-T^{\infty}\right) \left( e_{1};x\right)
\right\vert $ and $\delta^{2}=\left\vert \left( T^{\infty}-T^{m}\right)
\left( e_{2};x\right) \right\vert $ yields the desired result. If
$\left\vert \left( T^{m}-T^{\infty}\right) \left( e_{1};x\right)
\right\vert =0$ and $\left\vert \left( T^{\infty}-T^{m}\right) \left(
e_{2};x\right) \right\vert >0$, then
\[
\left\vert T^{\infty}\left( f;x\right) -T^{m}\left( f;x\right) \right\vert
\leq\frac{3}{2}\omega_{2}\left( f;h\right) +\frac{1}{\delta^{2}}\frac{3
{4}\omega_{2}\left( f;\delta\right) \left\vert \left( T^{m}-T^{\infty
}\right) \left( e_{2};x\right) \right\vert
\]
for all $h>0.$ Taking $\delta=\left\vert \left( T^{\infty}-T^{m}\right)
\left( e_{2};x\right) \right\vert ,$ $h\rightarrow0$ yields the desired
result. If $\left\vert \left( T^{m}-T^{\infty}\right) \left( e_{1
;x\right) \right\vert >0$ and $\left\vert \left( T^{\infty}-T^{m}\right)
\left( e_{2};x\right) \right\vert =0$, then
\[
\left\vert T^{\infty}\left( f;x\right) -T^{m}\left( f;x\right) \right\vert
\leq\frac{3}{2}\omega_{2}\left( f;h\right) +\frac{1}{h}\left( 2\omega
_{1}\left( f;h\right) +\frac{3}{2}\omega_{2}\left( f;h\right) \right)
\left\vert \left( T^{m}-T^{\infty}\right) \left( e_{1};x\right)
\right\vert
\]
for all $h>0.$ Taking $h=\left\vert \left( T^{\infty}-T^{m}\right) \left(
e_{1};x\right) \right\vert ,$ yields the desired result. If $\left(
T^{m}-T^{\infty}\right) \left( e_{1};x\right) =0$ and $\left(
T^{m}-T^{\infty}\right) \left( e_{2};x\right) =0$, then
\[
\left\vert T^{\infty}\left( f;x\right) -T^{m}\left( f;x\right) \right\vert
\leq\frac{3}{2}\omega_{2}\left( f;h\right) .
\]
For $h\rightarrow0$ we obtain $T^{\infty}\left( f;x\right) =T^{m}\left(
f;x\right) $ for all $0\leq x\leq1.$
(a) It follows that from (\ref{b0}) that any nonincreasing convex $g\in
C^{2}\left[ 0,1\right] $
\begin{align*}
T\left( g;x\right) & \geq g\left( x\right) +g^{\prime}\left( x\right)
\left( T\left( e_{1};x\right) -x\right) \geq g\left( x\right) ,\\
g\left( x\right) & \leq T^{m}\left( g;x\right) \leq T^{m+1}\left(
g;x\right) \leq\left\Vert g\right\Vert .
\end{align*}
In other words the sequence continuous functions $\left\{ T^{m}\left(
g;\cdot\right) \right\} $ is nondecreasing for any nonincreasing convex
function $g\in C^{2}\left[ 0,1\right] .$ By Dini's theorem there exists
$T^{\infty}\left( g;\cdot\right) $ such that $T^{m}\left( g;\cdot\right)
\rightarrow T^{\infty}\left( g;\cdot\right) $ uniformly on $\left[
0,1\right] ,$ for any nonincreasing convex function $g\in C^{2}\left[
0,1\right] .$In particular
\begin{align*}
\lim_{m\rightarrow\infty}\left\Vert T^{m}\left( -e_{1}\right) -T^{\infty
}\left( -e_{1}\right) \right\Vert & =0,\\
\lim_{m\rightarrow\infty}\left\Vert T^{m}\left( \left( e_{0}-e_{1}\right)
^{2}\right) -T^{\infty}\left( \left( e_{0}-e_{1}\right) ^{2}\right)
\right\Vert & =0,
\end{align*}
and $\lim_{m\rightarrow\infty}\left\Vert T^{m}\left( e_{2}\right)
-T^{\infty}\left( e_{2}\right) \right\Vert =0$ since $T^{m}\left( \left(
e_{0}-e_{1}\right) ^{2}\right) =T^{m}\left( e_{0}\right) -2T^{m}\left(
e_{1}\right) +T^{m}\left( e_{2}\right) $
Let $g\in C^{2}\left[ 0,1\right] $ be arbitrary. Introduce the following
auxiliary function
\[
g_{\pm}\left( t\right) =\frac{1}{2}\left\Vert g^{\prime\prime}\right\Vert
\left( 1-t\right) ^{2}+\left\Vert g^{\prime}\right\Vert \left( 1-t\right)
\pm g\left( t\right) .
\]
It is clear that
\[
g_{\pm}^{\prime}\left( t\right) =-\left\Vert g^{\prime\prime}\right\Vert
\left( 1-t\right) -\left\Vert g^{\prime}\right\Vert \pm g\left( t\right)
\leq0,\ \ \ g_{\pm}^{\prime\prime}\left( t\right) =\left\Vert g^{\prime
\prime}\right\Vert \pm g\left( t\right) \geq0.
\]
Therefore the functions $g_{\pm}\left( t\right) $ are nonincreasing convex
for both choices of the sign. We hav
\begin{align*}
0 & \leq T^{m+p}\left( g_{\pm};x\right) -T^{m}\left( g_{\pm};x\right)
=\frac{1}{2}\left\Vert g^{\prime\prime}\right\Vert \left( T^{m+p}\left(
e_{2};x\right) -T^{m}\left( e_{2};x\right) \right) \\
& +\left( \left\Vert g^{\prime\prime}\right\Vert +\left\Vert g^{\prime
}\right\Vert \right) \left( T^{m}\left( e_{1};x\right) -T^{m+p}\left(
e_{1};x\right) \right) \pm\left( T^{m+p}\left( g;x\right) -T^{m}\left(
g;x\right) \right) .
\end{align*}
The rest of the proof is similar to that of part (b).
\end{proof}
\begin{proof}
[Proof of Theorem \ref{t:11}](a) It remains to show that $T^{\infty}\left(
f\right) =P\left( f\right) $ for all $f\in C\left[ 0,1\right] $. It is
clear that it is enough to show this equality in $C^{2}\left[ 0,1\right] $.
Let $g\in C^{2}\left[ 0,1\right] $. Define the following auxiliary
functions
\[
G\left( x\right) :=g\left( x\right) -P\left( g;x\right) ,\ \ \ l:=\frac
{1}{2}\left\Vert G^{\prime\prime}\right\Vert =\frac{1}{2}\left\Vert
g^{\prime\prime}\right\Vert ,\ \ g_{\pm}\left( x\right) :=-lx^{2}+lx\pm
G\left( x\right) .
\]
It is clear that $g_{\pm}$ is concave and nonnegative, since
\[
g_{\pm}^{\prime\prime}\left( x\right) =-\left\Vert G^{\prime\prime
}\right\Vert \pm G^{\prime\prime}\left( x\right) \leq0,\ \ \ G\left(
0\right) =G\left( 1\right) =0.
\]
It follows that
\[
-l\left( x-x^{2}\right) \leq G\left( x\right) \leq l\left( x-x^{2
\right) ,\ \ \ \ 0\leq x\leq1.
\]
Applying the positive operator $T^{\infty}$ we ge
\[
-l\left( T^{\infty}\left( e_{1};x\right) -T^{\infty}\left( e_{2};x\right)
\right) \leq T^{\infty}\left( G;x\right) =T^{\infty}\left( g;x\right)
-P\left( g;x\right) \leq l\left( T^{\infty}\left( e_{1};x\right)
-T^{\infty}\left( e_{2};x\right) \right) ,\ \ \ \
\]
for all $0\leq x\leq1,$ and consequentl
\[
T^{\infty}\left( g\right) =P\left( g\right)
\]
for all $g\in C^{2}\left[ 0,1\right] $, which completes the proof.
(b) The operator $T^{\infty}$ of Theorem \ref{thm1} satisfie
\[
T^{\infty}\left( e_{0}\right) =e_{0},\ \ T^{\infty}\left( e_{1}\right)
=e_{2},\ \ \ T^{\infty}\left( e_{2}\right) =e_{2}.
\]
It remains to show that $T^{\infty}\left( f\right) =V\left( f\right) $ for
all $f\in C\left[ 0,1\right] $. It is clear that it is enough to show this
equality in $C^{2}\left[ 0,1\right] $.
Let $g\in C^{2}\left[ 0,1\right] $. Define the following auxiliary
functions
\begin{align*}
G\left( x\right) & :=g\left( x\right) -V\left( g;x\right) =g\left(
x\right) -\left( 1-x^{2}\right) g\left( 0\right) -x^{2}g\left( 1\right)
,\ \ \ \\
l & :=\frac{1}{2}\left\Vert g^{\prime\prime}+2g\left( 0\right) -2g\left(
1\right) \right\Vert ,\ \ G^{\prime\prime}\left( x\right) =g^{\prime\prime
}\left( x\right) +2g\left( 0\right) -2g\left( 1\right) ,\ \\
g_{\pm}\left( x\right) & =-lx^{2}+lx\pm G\left( x\right) .
\end{align*}
It is clear that $g_{\pm}$ is concave and nonnegative, sinc
\[
g_{\pm}^{\prime\prime}\left( x\right) =-\left\Vert G^{\prime\prime
}\right\Vert \pm G^{\prime\prime}\left( x\right) \leq0,\ \ \ G\left(
0\right) =G\left( 1\right) =0.
\]
It follows that
\[
-l\left( x-x^{2}\right) \leq G\left( x\right) \leq l\left( x-x^{2
\right) ,\ \ \ \ 0\leq x\leq1.
\]
Application of the positive operator $T^{\infty}$ give
\[
-l\left( T^{\infty}\left( e_{1};x\right) -T^{\infty}\left( e_{2};x\right)
\right) \leq T^{\infty}\left( G;x\right) =T^{\infty}\left( g;x\right)
-V\left( g;x\right) \leq l\left( T^{\infty}\left( e_{1};x\right)
-T^{\infty}\left( e_{2};x\right) \right) ,\ \ \ \ 0\leq x\leq1,
\]
and $T^{\infty}\left( g\right) =V\left( g\right) $ for all $g\in
C^{2}\left[ 0,1\right] .$
\end{proof}
\begin{proof}
[Proof of Corollary \ref{t:iterate}]By the induction we have
\[
x^{2}\leq T^{m}\left( e_{2};x\right) \leq a^{m}x^{2}+b\left(
1+a+...+a^{m-1}\right) x=a^{m}x^{2}+\left( 1-a^{m}\right) x.
\]
So
\[
0\leq x-T^{m}\left( e_{2};x\right) \leq a^{m}x\left( 1-x\right) .
\]
\end{proof}
\begin{proof}
[Proof of Corollary \ref{c:11}] The proof is based on the Taylor formula about
the point $0$ for part (a) and about the point $1$ for part $(b).$
\end{proof}
\bigskip
|
\section{Introduction}
The spatial distribution of heavy element abundances is a unique tool to trace the structure formation history of disk galaxies. Since the very early work of \citet{Searle71}, \citet{Shields74}, and \citet{Zaritsky94}, it has been well established that there is a negative radial metallicity gradient in spiral galaxies such that the central disk region is more metal-enriched than the outer regions. The evolution of this gradient over cosmic time provides a powerful constraint on disk formation and evolution models.
Sophisticated chemical evolution models(CEMs) have been successful in reproducing the current gradient of local galaxies, but different models disagree on the time variation of the radial metallicity gradient \citep[e.g.,][]{Molla96,Chiappini97,Magrini07,FuJ09,Marcon10}. Some models predict that the radial gradients steepen with time \citep{Chiappini97,Chiappini01}, while other models predict that the radial gradients flatten with time \citep{Molla97,Hou00,Prantzos00,FuJ09}. The reason for these opposite predictions is that CEM models are very sensitive to the prescriptions of the detailed physical processes that lead to the chemical enrichment of inner and outer disks. Observational constraints on these physical
processes are lacking.
Metallicity gradient evolution is unknown due to the small angular sizes and the strong decline of surface brightness of distant galaxies.
Before the next generation of telescopes, the only way to overcome this observational hurdle is to combine the magnification power of gravitational lensing with the Laser Guide Star Adaptive Optics (LGSAO) aided integral field spectrograph units (IFUs) on the world's largest telescopes \citep{Jones10a,Stark08,Swinbank07b}.
Recently, \citet{Jones10b} reported the metallicity gradient of a strongly lensed dispersion-dominated galaxy at high-$z$ (``the Clone arc" at $z=2.00$).
\citet{Jones10b} measured the slope of the gradient for the inner $\sim$ 1 kpc and found it to be considerably steeper than local disk galaxies. The steep gradient at high-$z$
is consistent with the ``inside-out" disk formation scenarios. However, in a low resolution study, \citet{Cresci10} reported ``inverted" metallicity gradients for three non-lensed $z\sim3$ Lyman-break galaxies (LBGs), arguing for a ``cold-flow" of primordial gas to the galactic core regions. Clearly, a large sample of galaxies with chemical abundance gradients across cosmic time are required to place robust constraints on disk formation theories.
In this Letter, we report the first metallicity gradient measurement of a lensed {\it spiral} galaxy at $z=1.49$ based on our
OH-Suppressing Infra-Red Imaging Spectrograph (OSIRIS) observations at KECK II.
Throughout this Letter we use a $\Lambda$CDM cosmology with $H_0$= 70 km s$^{-1}$
Mpc$^{-1}$, $\Omega_M$=0.30, and $\Omega_\Lambda$=0.70. At $z=1.5$, 1 arcsec corresponds to 8.5 kpc and a look-back time of 9.3 Gyr. We use solar oxygen abundance 12 + log(O/H)$_{\odot}$=8.66 \citep{Asplund05}.
\section{Observations and Data Reduction}\label{obs}
The spiral galaxy is a four-image lensed system first identified by \citet{Smith09}, located behind the massive, X-ray luminous cluster MACS\,J$1149.5$$+$2223.
We observed the largest of these four images at $(\alpha_{2000},\delta_{2000})$\,=\,(11$^{\text h}$49$^{\text m}$ 35.$\!\!^{\text s}$284, +22$^{\circ}$23$^{\prime}$45.$\!\!^{\prime\prime}$86) (called Sp1149 hereafter). \citet{Smith09} construct a detailed mass model for the galaxy cluster and show that Sp1149 is
one of the largest recorded lensed images of a single galaxy at $z>1$. Owing to the lensing magnification and fortuitous face-on orientation, this galaxy
exhibits more than 10 locally resolved H\,{\sc ii}\rm\ regions (Figure~\ref{fig:hst}), ideal for spatial chemical abundance studies.
The star formation rate from the $V_{555}$ band is $\sim6~M_{\odot}$ yr$^{-1}$ \citep{Smith09}.
\begin{figure*}[!ht]
\begin{center}
\vspace{-0.8cm}
\includegraphics[scale=0.56]{fig1.eps}
\caption{Left: HST F814W, F555W-band color image of Sp1149 behind the lensing cluster MACS\,J1149.5$+$2223. North is 126$^{\circ}$ counterclockwise to the $y+$ direction. Sp1149 is the largest of the four multiple images of a spiral galaxy at $z=1.49$. The white box shows the
4.$\!\!^{\prime\prime}$8$\times$ 6.$\!\!^{\prime\prime}$4
FOV of OSIRIS Hn3 filter with the 0.$\!\!^{\prime\prime}$1 per pixel plate scale. Right: wavelength-collapsed image of the {H$\alpha$}\ emission from our OSIRIS datacube. Contours in black denote observed {H$\alpha$}\ flux levels of 0.6, 1.3 in units of 10$^{-17}$ ergs~s$^{-1}$~cm$^{-2}$~sec$^{-2}$. Labels on the left and bottom axes indicate the imaging plane coordinates while the right and top axes show the corresponding physical scale on the source plane. The white circles mark the three annuli $(r1, r2, r3)$ we use for deriving metallicity gradient (Section~\ref{gradient}). The {H$\alpha$}\ morphology aligns generally well with the rest-frame UV morphology.
}
\label{fig:hst}
\end{center}
\end{figure*}
We observed Sp1149 OSIRIS \citep{Larkin06} in conjunction with the LGSAO
system on KECK II telescope on 2010 March 3rd. The weather conditions were moderate with the optical FWHM ranging
between 0.$\!\!^{\prime\prime}$6 and 1.$\!\!^{\prime\prime}$5. The average tip/tilt corrected FWHM during the observation was 0.$\!\!^{\prime\prime}$1. To capture {H$\alpha$}\ and {[N\,{\sc ii}]}, we used the narrow band Hn3 filter (spectral coverage: 1.594$-$1.676$\micron$; spectral resolution: $\sim$3400), and a plate scale of 0.$\!\!^{\prime\prime}$1 per lenslet. The observations were conducted in an ABAA dithering sequence with a 10.$\!\!^{\prime\prime}$7 West-North chop onto an object-free sky region. A total of 19 exposures were obtained for the target, with 900 s of exposure for each individual frame. The net on-target exposure time is therefore 4.75 hr. Figure~\ref{fig:hst} shows the OSIRIS field of view (FOV) and locations of Sp1149 on an HST F814W, F555W color image.
Individual exposures were first reduced using the OSIRIS data reduction pipeline \citep{Larkin06}. We used
the sky subtraction IDL code of \citet{Davies07}, which we have modified to optimize the sky subtraction
specifically around the wavelength region of {H$\alpha$}\ and {[N\,{\sc ii}]}. Since we are mainly concerned with the emission lines,
a first order polynomial function was fit to each spatial sample pixel (spaxel) to improve the removal of lenslet to lenslet variations. The final datacube was constructed by aligning the sub-exposures with the centroid of the {H$\alpha$}\ images and
combing using a 3$\sigma$ mean clip to reject cosmic rays and bad lenslets. Telluric correction and flux calibration was performed by observing a few A0V standard stars immediately after the science exposure. The typical uncertainty in flux calibration is $\sim$10\%.
Gaussian profiles were fitted simultaneously to the three emission lines: {[N\,{\sc ii}]}\lam6548, 6583 and {H$\alpha$}.
The line profile fitting was conducted using a $\chi^2$ minimization procedure which takes into account the greater noise level close to atmospheric OH emissions.
The centroid and velocity width of {[N\,{\sc ii}]}\llam6548, 6583 lines were constrained by the velocity width of {H$\alpha$}\lam6563, and
the ratio of {[N\,{\sc ii}]}\lam6548 and {[N\,{\sc ii}]}\lam6583 is constrained to the theoretical value \citep{Osterbrock89}.
An instrumental profile of 5.9\AA\ calculated from sky lines has been subtracted from the line widths.
We demand a minimum signal-to-noise (S/N) of 3 to detect each emission line and a minimum S/N of 5 for kinematic and metallicity analysis.
\section{ Kinematics and Lensing Modeling}\label{mor}
To align the astrometry to the IFU datacube, we first assign the brightest pixel of the {H$\alpha$}\ narrow-band
image with the coordinates of the galactic core from the {\it Hubble Space Telescope (HST)} image, and then rotate the IFU datacube with the observing position angle. The {H$\alpha$}\ image can
then be aligned with the {\it HST} image by matching coordinates. The upper right panel of Figure~\ref{fig:hst} shows our observed {H$\alpha$}\ emission line image, collapsed in wavelength, and smoothed with a Gaussian kernel of FWHM = 2.5 pixels. The {H$\alpha$}\ morphology aligns generally well with the rest-frame NUV/$U$-band image, especially some of the H\,{\sc ii}\rm\ regions in the west. The SFR and size relation of individual H\,{\sc ii}\rm\ regions will be reported in a following paper (R. C. Livermore et al. 2011, in preparation).
We fit the velocity field with rotation disk models as described in \citet{Jones10a}.
The best fit yields a rotation velocity of $V_{\text rot} \sin{i}=150\pm15~$km s$^{-1}$, with inclination angle $i=45^{\circ} \pm 10^{\circ}$.
Using $M_{\text dyn}= V_{\text rot}^2\,R / G$, the derived
dynamical mass is $M_{\text dyn}= 2.6 \pm0.4 \times10^{10} M_{\odot}$ within R=2.5 kpc.
Sp1149 is a rotation dominated disk with $V_{\text rot}$/ $\sigma \sim 4$.
To reconstruct the source-plane morphology and obtain flux magnification for Sp1149,
we apply the most recent lens model for cluster MACS\,J1149.5$+$2223 as described in \citet{Smith09}.
The derived luminosity weighted magnification for Sp1149 is $\mu=22 \pm 2$.
Accounting for the lensing amplification, the intrinsic magnitude in the $I-$band is $I\simeq23.4\pm0.3$, corresponding to
$M_B\simeq-20.7$.
Sp1149 is stretched $\sim$ 5 on each side $(x, y)$ of the two-dimentional image. Because of this fortuitous uniform magnification, we can work with the observed datacube and divide a factor of 5 linearly and a factor of 22 in flux to recover the intrinsic linear geometry and luminosity of the galaxy. Our results are
therefore unchanged when we use the reconstructed source-plane datacube.
\section{Metallicity Gradient}\label{gradient}
For the metallicity analysis in this work, we use the empirical strong line diagnostic --- the
N2\,=\,log({[N\,{\sc ii}]}\lam6583/{H$\alpha$}) index, as calibrated by \citet[][hereafter PP04]{Pettini04}:
\begin{equation}
12+{\rm log (O/H)}=8.90+0.57\times {\text N2}
\end{equation}
We note that the N2 ratio is sensitive to shock excitation, the ionization state of the gas, and
the hard ionization radiation field from an active galactic nucleus
\citep{Kewley02, Kewley06, Rich10}. Our N2 ratios are consistent with star-formation throughout the disk. The ratio of {[N\,{\sc ii}]}\,/{H$\alpha$}\ for the whole
galaxy is 0.112$\pm$0.050.
Due to its relatively low metallicity (compared with local spirals), the {[N\,{\sc ii}]}\ line is too weak to reach a S/N of 5 in individual OSIRIS pixels, but is clearly detected.
We therefore integrate the spectrum within three annuli (see Figure~\ref{fig:hst}, right panel). The physical lengths of the three outer radii of each annulus in the source plane are $r1=0.72\pm0.1$ kpc, $r2=2.34\pm0.2$ kpc, $r3=4.5\pm0.4$ kpc.
The {[N\,{\sc ii}]}\ line is robustly detected at S/N $>$ 5 for the inner two annuli and is a 3$\sigma$ detection for the outer annulus which we give as an upper limit. The final spectra for the three annuli are presented in Figure~\ref{fig:spec}.
\begin{figure}[!ht]
\vspace{-0.8cm}
\begin{center}
\includegraphics[scale=1.]{fig2.eps}
\caption{Integrated spectra for the three annuli defined in Figure~\ref{fig:hst}. Vertical dashed lines indicate the zero-velocity positions of {[N\,{\sc ii}]}\ and {H$\alpha$}. The $y-$axis is the observed intensity in units of 10$^{-17}$ ergs~s$^{-1}$~cm$^{-2}$~\AA$^{-1}$. The annuli radii $r1, r2, r3$
are chosen to ensure a S/N ({[N\,{\sc ii}]}) to be $>$5 for the inner two radii. Blue curves are the Gaussian fits of the spectra. Note that the outer annulus $r2:r3$ is a 3$\sigma$ upper limit for {[N\,{\sc ii}]}.
The declining ratio of {[N\,{\sc ii}]}\,/{H$\alpha$}\ from the central to the outer disk infers a decreasing metallicity with galactocentric distance.
}
\label{fig:spec}
\end{center}
\vspace{-0.4cm}
\end{figure}
The metallicities converted from the N2 calibrator for the three annuli are
\begin{align}
\text{12\,+\,log(O/H)\,} = \begin{cases}
8.54\pm0.04 &\text{for} \quad r<r1 \\
8.38\pm0.05 &\text{for} \quad r1<r<r2 \\
<8.05 &\text{for} \quad r2<r<r3 \quad (\text{3$\sigma$ upper limit}) \\
\end{cases}
\end{align}
The ``global" metallicity measured from the integrated spectrum of all three regions is 8.36$\pm$0.04. Note that
the errors of line flux measurements are smaller than the intrinsic 0.18 dex scatter pertinent to the N2 calibration \citep{Erb06}.
The systematic uncertainty of the N2 calibration with respect to other line diagnostics is about 0.04 dex \citep{Kewley08}.
A linear regression with censored data yields a metallicity gradient for the central 4.5 kpc of $\frac{\Delta \log({\rm O/H})}{\Delta {\rm \text R}}= -0.16 \pm 0.02$ dex kpc$^{-1}$, or -0.15 dex kpc$^{-1}$ if we treat the last data point as a measurement rather than an upper limit. The true gradient may be steeper than this value given the
3$\sigma$ upper limit in the outer annulus.
\section{Metallicity Gradient Compared with Local and Higher Redshift}
In Figure~\ref{fig:gr} we compare the metallicity gradient of Sp1149 with galaxies in the local and higher redshift universe, as well as the model predictions from \citet{Prantzos00}.
To avoid systematic effects caused by different calibration indicators, we use the oxygen abundance of H\,{\sc ii}\rm\ regions for comparison.
\citet{Kewley08} showed that metallicities can be compared if the same metallicity calibration is applied consistently across all sample galaxies and H\,{\sc ii}\rm\ regions.
We convert all metallicities to the PP04 calibration using the calibrations in \citet{Kewley08}.
\citet{Prantzos00} modeled metallicity gradients of spiral galaxies at ages $t$\,=\,3, 7.5, and 13.5 Gyr, for different velocities ($V_c$) and spin parameters ($\lambda$). Since Sp1149 is at an age of $\sim$4.3 Gyr and has $V_c$=212$\pm$50~km~s$^{-1}$, we adopt model grids with $V_c$\,=\,220~km~s$^{-1}$, $t$\,=\,3, 13.5 Gyrs, and $\lambda$\,=\,0.03, 0.05, 0.07 for our comparison. The model metallicities have been calibrated to the PP04 metallicity diagnostic. Sp1149 shows a steeper gradient than predicted by \citet{Prantzos00} for $V_c$\,=\,220~km~s$^{-1}$. However, the steep gradient of Sp1149 is broadly consistent with the $V_c$\,=\,150~km~s$^{-1}$ model (the dash-dotted model line in Figure~\ref{fig:gr}).
Metallicity gradients in local galaxies are strongly correlated with Hubble type, bar strength, and merging events; with more flattened gradients for barred galaxies and merging pairs \citep{Kewley10, Rupke10a}. Therefore it is significant that the metallicity gradient of Sp1149 is substantially steeper than local late-type galaxies (see Figure~\ref{fig:gr}).
For the local galaxy gradients, we adopt the isolated spiral control sample of \citet{Rupke10b} comprised of 11 isolated mid-to-late type local galaxies with a mix of bar-strengths. The median slope for the local control sample is -0.041$\pm$0.009 dex kpc$^{-1}$, with a standard deviation of 0.03 dex kpc$^{-1}$.
The median slope of gradients from the Rupke et al. (2010) sample is consistent with the values of \citet{Zaritsky94} and \citet{vanZee98} samples.
The \citet{Zaritsky94} gradient sample composed of 39 local disk galaxies, and the \citet{vanZee98} sample added another 11 local spirals.
Local early-type galaxies have even shallower gradients \citep{HenryR99}.
A similar (and even steeper) gradient has been reported in a recent work of \citet{Jones10b} for a
dispersion-dominated lensed system within only $\sim$1 kpc at $z$=2.
\citet{Cresci10} recently reported reversed metallicity gradients (i.e., the core is less enriched than the outer disk) for three un-lensed LBGs at $z$$\sim$ 3. They explained the inverted gradients by inflow of pristine cold gas to the galactic center. Due to their low resolution ($\sim$ 4-6 kpc), the LBG selection, and potential presence of mergers and galactic winds which may contaminate metallicity gradient, it is not clear how the \citet{Cresci10} results relate to the work presented here.
In the right panel of Figure~\ref{fig:gr} we show that the metallicity gradient of Sp1149 is still considerably steeper than local galaxies when expressed in units of dex/($R$/$R$$_{25}$),
where $R_{25}$ is the $B-$band isophotal radius at a surface brightness of 25 mag arcsec$^{-2}$.
We estimate the $R_{25}$ for Sp1149 by fitting the isophotal profile of the source-plane F814W-band image assuming
a color difference of $U-B= -0.1$ mag for Sc type galaxies \citep{Fukugita95}. The resulting $R_{25}$ for Sp1149 is 12$\pm$2kpc, comparable to the range of $R_{25}$
in our local control sample \citep{Kewley10,Rupke10b}.
To analyze the effect of size and luminosity on the metallicity gradients,
in Figure~\ref{fig:grdiag}, we plot the metallicity gradients (in units of $R_{25}$) v.s. the absolute $B-$band magnitude $M_B$ for local late-type galaxies and Sp1149.
Sp1149 is comparable to the brightest local late-type disks, while ``the Clone Arc" is much more luminous ($M_B=-22.12$). We see from Figure~\ref{fig:grdiag} that
the scaled gradient of Sp1149 is steeper than in typical local galaxies.
\begin{figure*}[!ht]
\begin{center}
\includegraphics[scale=0.5,angle=0]{fig3.eps}
\caption{
Left: metallicity vs. galactocentric radius.
Red lines are the measurements for Sp1149 at z=1.49 from this work.
The gradient within the central 4.5 kpc is $-0.16\pm0.02$ dex kpc$^{-1}$.
Vertical red dotted lines show the annulus used to average/sum the spectra.
Purple dashed lines show the typical gradients of local isolated late-type galaxies, using the control sample of \citet{Rupke10b}. The orange dotted line represents the mean gradient
of local early-type galaxies, which is typically $\sim$ 3 times shallower than local late-type galaxies \citep{HenryR99}.
Blue lines show the work of \citet{Jones10b} for a dispersion-dominated lensed galaxy the ``Clone Arc" at $z=2.0$
who report a even steeper gradient within a size of $\sim$ 1 kpc.
It is interesting to see that two $z>1$ galaxies have considerably steeper gradients than local galaxies.
Black lines are the model predictions from \citet{Prantzos00} for rotational velocity $V_{c}$=220km~s$^{-1}$ at age t=3Gyr (dashed curves) and $t=13.5$ Gyr (solid curves), with spin parameter $\lambda$=0.03,0.05,0.07. The black dash-dotted line is the model grid with velocity $V_{c}$=150km~s$^{-1}$ at $t=3$ Gyr and $\lambda$=0.03. The steep gradients of Sp1149
and ``Clone Arc" are broadly consistent with the shape of the $V_{c}$=150km~s$^{-1}$ model grids. Right: The same as the left panel, except that the $x-$axis is expressed in the scaled radius $R_{25}$.
}
\label{fig:gr}
\end{center}
\vspace{-0.4cm}
\end{figure*}
\begin{figure}[!ht]
\begin{center}
\includegraphics[scale=0.6,angle=90]{fig4.eps}
\caption{Metallicity gradient (dex/$R_{25}$) vs absolute B band magnitude $M_{B}$.
The red square represents the measurement for Sp1149 at $z$=1.49.
Black filled circles indicate the locations of the local late-type galaxies of \citet{Rupke10b}.
Black empty circles are from the sample of \citet{Zaritsky94}, where the metallicity gradients are measured for 39 local disk galaxies.
Black empty triangles come from the sample of \citet{vanZee98}, where the metallicity gradients are provided for another 11 local spiral galaxies.
Sp1149 has significantly steeper gradient than local galaxies of comparable luminosity, even when expressed in units of scaled radius.
}
\label{fig:grdiag}
\end{center}
\vspace{-0.4cm}
\end{figure}
\section{Discussion}\label{discuss}
The redshift range of $z$=1---3 is the period when star-formation, mass assembly and chemical enrichment activity peaks \citep[e.g.,][]{Fan01,Dickinson03, Chapman05,HopkinsAM06,Conselice07}. Metallicity gradient studies of galaxies in this redshift range therefore give important insights into disk galaxy formation. The picture that the metallicity gradient flattens with cosmic time is qualitatively consistent with
the ``inside-out" disk formation scenario \citep[e.g.,][]{Prantzos00, Bouwens97,Benson10}.
This scenario predicts that in the early stages of
galaxy evolution, the inner galactic bulge undergoes a rapid collapse with vigorous star formation at the center,
building a steep radial metallicity gradient from the core to the outer disk. Violent gas dynamics and accretion events marked these early stages of disk formation. Subsequent radial mixing and infall cause the gradient to flatten over the following Gyrs, resulting in a weaker
gradient on average in late-type galaxies today.
In this work, we have presented an OSIRIS IFU study of a grand-design lensed spiral galaxy at $z=1.49$. The 22$\times$ magnification allows us to resolve individual H\,{\sc ii}\rm\ regions at an intrinsic resolution of 170 parsec. We measure the radial distribution of chemical abundances in three annuli and find that the metallicity decreases from
a central value of 12+{\rm log (O/H)}=8.54 to less than 8.05 in the outer disk. The metallicity slope is $-0.16\pm0.02$ dex kpc$^{-1}$ for the inner 4.5 kpc, much steeper than that of late-type galaxies of comparable luminosities in the local universe, even when expressed in $R_{25}$ units. Our results are broadly consistent with the ``inside-out" model of disk formation.
Metallicity gradient observations for a larger number of star-forming galaxies at different redshifts are required to build a well-defined sample to form a solid observational picture of the metallicity gradient evolution as a function of cosmic time. Before the next generation of telescopes, we are aiming to secure a few tens of gravitationally lensed systems
as a pioneer sample to study the gradient evolution of galaxies at high-z. Future instruments such as the {\it JWST}, TMT and GMT will improve the current resolution limits by $\sim$
1 order of magnitude, revolutionizing chemical gradient evolution studies at high-$z$.
\acknowledgments
We thank the referee for his/her valuable comments which have helped a lot to improve this Letter.
This work is based on data obtained at the W. M. Keck
Observatory. We are grateful to the Keck Observatory staff for assistance with our observations, especially Jim Lyke, Hien Tran, and
Randy Campbell. T.-T.-Y. thanks the hospitality of ICC,
Durham University, where a large portion of the work was done.
A.-M.-S. acknowledges a STFC Advanced Fellowship. R.-C.-L. acknowledges a studentship from STFC.
The authors recognize the very significant cultural role that the summit of Mauna Kea has within the indigenous Hawaiian community.
|
\section{Introduction}\label{sec:intro}
A large variety of new physics scenarios feature the presence of charged particles with peculiar properties, that can lead to a systematic mis-reconstruction of their tracks by the standard algorithms. These properties can induce mismeasurements of the $p_T\,$ or even a failure to reconstruct tracks. One of the simplest examples is a particle that decays in flight inside the tracker, but in the following we list other possibilities and refer to them generically as New Odd Tracks (NOTs). The systematic mis-reconstruction of NOTs implies that such theories may evade detection, even if they are produced at surprisingly high rates. Consequently, particles of this kind can be very light, and may require dedicated studies for discovery. In this note we argue that there are viable models with very light colored states that would have gone unnoticed. As a motivating example, we consider an anomaly in a recent measurement based on minimum bias events, and provide viable explanations using NOTs.
In~\cite{weirdcdf} the single charged particle inclusive distribution
was measured by the CDF collaboration. The measurement was
found to be inconsistent with the QCD prediction at high $p_T\,$~\cite{1003.1854, 1003.2963, 1003.3433} by
a factor of $10^4$. Subsequent measurements by ATLAS~\cite{atlas} and CMS~\cite{cms} found no evidence for an anomaly at high $p_T\,$. Most recently, CDF released an erratum~\cite{errata} where they changed their track selection to remove the high-$p_T\,$ tracks they had previously measured. While the original result
could
in principle come from an experimental mismeasurement/unaccounted for background, there is an intriguing possibility that
it was due to new physics that can also account for all subsequent findings.
The main difference between the original CDF analysis and the subsequent reanalysis lies in demanding a higher track quality. Unfortunately, the reanalysis does not quantitatively find a SM explanation of the original excess tracks, which are simply removed by the quality cut. The measurements by ATLAS and CMS also require much more stringent track quality cuts than the original CDF measurement. To address the original anomaly, together with the null results, one is therefore required to introduce new particles which appear fundamentally different in the tracker, i.e. NOTs. The
more stringent cuts on the CDF data reconcile the apparent tension, however, they do so at the cost of losing the sensitivity to new physics of this kind.
The models presented are interesting in their own right. The take-home message, however, is not the specific model. Rather, we stress the existence of a large variety of theories that exhibit NOTs, which would be misinterpreted or even missed at the LHC unless specifically searched for. Our work aims, in part, at motivating additional studies to ensure NOTs will not escape detection.
\section{New Odd Tracks}
Before discussing specific models, we briefly discuss the spectrum of possibilities for theories that exhibit NOTs. It is useful to classify the possible effects of new physics on standard track signatures. Typically, a given model exhibits more than one signature, which may simplify its identification~\cite{simplified}:
\squishlisttwo
\item {\bf Kinks}. Tracks that appear to change direction, without a secondary vertex. Typically produced by one-prong decays.
\item {\bf Displaced vertices}. Tracks appearing to emanate not from the PV.
\item {\bf Anomalous $dE/dx$}. Tracks may have lower or higher ionization loss. Standard Heavy Stable Charged Particle (HSCP) searches typically look for the latter.
\item {\bf Anomalous timing}. Slowly moving tracks as measured via the timing module at the calorimeter, but not necessarily with a larger $dE/dx$.
\item {\bf Intermittent hits}. Otherwise normal tracks that leave fewer hits than expected.
\item {\bf Anomalous curvature}. Tracks that appear to bend anomalously in the tracker.
\item {\bf Stub Tracks}. Tracks that seem to disappear inside the tracking volume.
\squishend
We note that it is possible to misidentify some of the signatures above. For instance, as we discuss below, tracks with kinks may be misinterpreted as tracks with anomalous curvature.
Several of these possibilities have been explored before in the context of models of new physics. In Gauge Mediated Supersymmetry Breaking (GMSB, see~\cite{Giudice:1998bp} and references therein), there are generically two types of NOTs. Models with a long lived slepton as the NLSP, admit tracks that can have kinks, anomalously high $dE/dx$ and anomalous timing. Models with a long lived netrualino NLSP would typically give rise to displaced vertices. Stub tracks or kinks can arise in models of Anomaly Mediated Supersymmetry Breaking (AMSB)~\cite{9904250}. Models of SUSY with R-parity violation~\cite{rpvreview} can also give rise to kinks, displaced vertices, anomalously high $dE/dx$ and anomalous timing. Quirks~\cite{quirks} give rise to tracks with either anomalous curvature or anomalously high $dE/dx$.
It is important to note that, while in many of these examples the existence of new physics can be established by other means, it could prove to be significantly harder to identify the model without studying some of the above signatures. For instance, even if supersymmetry is discovered, identifying the breaking mechanism may require the study of kinks.
There are, in addition, several interesting possibilities for NOTs that remain altogether unexplored at the Tevatron and LHC. In particular, intermittent hits and anomalously low $dE/dx$ are possibilities that have not been investigated. One reason for this, is that such signatures are caused by particles that leave less energy than a Minimally Ionizing Particle (MIP) in the tracking system, thereby deteriorating their reconstruction efficiency. Similarly, models with kinked tracks may not be reconstructed, albeit having regular $dE/dx$ signature, or may present significant backgrounds from detector material effects.
In this note we will give examples of some theories of NOTs that can explain the original CDF anomaly. These models may serve as benchmarks for classes of models that will not be found through standard tracking algorithms. A more comprehensive study of benchmark models and their prospective discovery will be presented in future work~\cite{futurework}.
For now, our hope is that the models presented here will serve, in addition to explaining the CDF data, as motivating examples to study irregular tracking at the LHC.
\section{The CDF Anomaly}
It is useful to recall why the original CDF results are non-trivial to explain with new physics. In~\cite{weirdcdf}, CDF looked at the $p_T\,$ distribution of all charged tracks examined through the minimum bias trigger path. The dominant contribution to this distribution at high $p_T\,$ is the single jet inclusive channel. The latter, however, was found to be saturated above 100 GeV by the single charged particle distribution~\cite{Abulencia:2007ez,1003.3433}, thereby signaling the breakdown of QCD factorization. On the other hand, the high-$p_T\,$ tracks may come from new massive particles of mass $M$. However, on average, $p_T\lesssim M$, and therefore to account for the high-$p_T\,$ spectrum one requires particles with mass of order $100$ GeV. In turn, the production rate for a particle of that mass is typically too low to explain the data, even if charged under QCD. This conflict between the $p_T\,$ scale and the cross section represents the inherent difficulty in explaining this data with new physics.
Even if a scenario predicting high $p_T\,$ tracks with a large enough rate were possible, additional constraints must be satisfied. In particular, the new physics:
\begin{itemize}
\item Must not substantially affect the inclusive jet cross section which is well measured~\cite{Abulencia:2007ez}.
\item Can not be a new resonance that decays only into a pair of charged tracks or jets~\cite{Khachatryan:2010jd}.
\item Must not have collider-stable particles~\cite{Aaltonen:2009kea}.
\end{itemize}
With these basic restrictions the difficulty to describe the measurement with NP is understood~\cite{1003.3433}.
The tension described above can be ameliorated if the transverse momentum of new particles is mismeasured. The presence of NOTs found in a variety of models, may thus provide an explanation to the anomaly. A first example which would account for a systematic mismeasurement of $p_T\,$ is a fractionally charged particle.
Indeed, the analysis~\cite{weirdcdf} assumes that the tracks have charge one
and therefore a particle of charge $q$ would have its $p_T\,$ measured as $p_T/q$. A sufficiently light new particle of this kind, interacting with QCD strength, could have a large cross section and {\em still} produce high $p_T\,$ tracks which would account for the CDF data.
Another example of a NOT that may cause a systematic mismeasurement of $p_T\,$ is a track with a kink. An attractive possibility is a light mass sparticle such as a light sbottom that decays through an RPV operator. Such a particle can be produced with a large cross-section without being detected due to the kink or the displaced vertex.
In a standard reconstruction algorithm these tracks {\em could} in principle, be reconstructed as a single track with a high or low $p_T\,$ and large $\chi^2$. Only those tracks that are reconstructed with a high $p_T\,$ would rise above the background, thereby addressing the measurement. Much like the fractionally charged particles, a model of the above kind could escape detection unless specifically searched for.
As discussed above, whether or not the original CDF data turns out to be attributed to new physics, it is important that the LHC looks for NOTs so that this window into new physics is not missed. In fact, since the examples above cause systematic mismeasurements in the tracker, they may not even be triggered on at the LHC. If the CDF data is indeed a measurement of such a NOT model, the looser track quality selection criteria together with the MB triggering path, may be the only reason that these particles were observed.
\section{Light Colored Particles}
\label{sec:fractional-charges}
As discussed in the previous section, a light colored particle would have a large enough cross-section to reproduce the CDF anomaly, if the $p_T\,$ of the resulting tracks were mismeasured. As an example, in this section we describe a viable and concrete model which exhibits fractionally charged particles. The possibility of light sbottoms with RPV discussed above, will be presented elsewhere.
\subsection{A Model}
While fundamental particles with fractional electric charges are very constrained, composite fractionally charged particles can more easily escape detection. For instance, let us
introduce
vector-like fermionic fields, $X+\bar X$, charged under the SM as
$(\bf 3,\bf 1)_0+(\bar\bf 3,\bf 1)_0$ ~\cite{Caldi:1982dj} and with a mass
$m_X\simeq\mathcal{O}(10\,\mathrm{GeV})$.
Once produced, $X,\bar
X$ hadronize to form mesons $M_X$ and baryons $B_X$, both carrying
fractional charges. Since the probability of hadronizing into baryons
is suppressed by an order of magnitude compared to that of mesons~\cite{pdg},
below we consider only the meson case, with charges $\pm 1/3$ and $\pm
2/3$.
As we discuss below, if $X$, $\bar X$ were stable, they would be
excluded in Charged Massive long-lived
Particle (CHAMP) searches~\cite{Aaltonen:2009kea} by many orders of
magnitude~\cite{crap1}. Consequently, $X$ must decay sufficiently fast and we
are therefore led to introduce additional scalar fields, $Y+\bar Y$, with quantum
numbers $(\bf 1,\bf 1)_{-1/9}+(\bf 1,\bf 1)_{1/9}$, mass $m_Y<m_X$ and
non-renormalizable couplings,
\begin{eqnarray}
\label{eq:1}
\frac{1}{\Lambda^2} X \bar d_R Y^3\,.
\end{eqnarray}
Here, $\bar d_R$ is the SM right handed down quark. Different charge assignments for $Y$ can be accommodated, implying a corresponding dimension in the operator above. We stress, however, that some of the bounds discussed below change for different charges, and may require additional structure (such as additional fractionally charged particles).
The virtue of
the above setup is that the colored $X$ particles are produced copiously
at hadron colliders but have suppressed production rate at $e^{+}e^{-}$
colliders. On the other hand, the production rate of the fractionally charged $Y$ particles
are suppressed in both colliders due to their small EM charge.
Furthermore, as we discuss, such particles are invisible at the Tevatron since they
rarely leave ionization signals in the detectors. Below, we study the
various constraints on this model and establish the predictions for
the Tevatron and the LHC.
\subsection{Constraints}
New light strongly interacting particles with a fractional charge are potentially bounded by many different experiments. Here we identify the most stringent bounds on these new states. We find that the model above, while only marginally in some cases, evades all experimental bounds. This is astonishing, given the lightness and strong coupling of these new states to SM particles. Such an example demonstrates the need to carefully search for NOTs as they can easily go unnoticed, certainly in less radical scenarios.
\mysubsection{CHAMPs.}
The most straightforward way to look for new heavy states,
is by searching for slowly moving particles through time-of-flight measurements. Such a search has been
performed by CDF~\cite{Aaltonen:2009kea} where events with isolated
muon candidates were studied as possible
CHAMPs. Good agreement with the SM was found,
thereby strongly excluding the possibility of a stable $X$. For instance,
the existence of a stable $X$ with $m_X =10$ GeV, predicts ${\cal O}(10^7)$ events that pass all
cuts, while only ${\cal O}(100)$ can be tolerated. As a consequence, the lifetime for the $X$ decays induced by Eq.~\ref{eq:1} is strongly constrained. For the above $X$ mass, we find the proper lifetime to be $c\tau_X \lesssim 25$~cm, corresponding to the cutoff scale, $\Lambda\simeq 3$ TeV in Eq.~\ref{eq:1}. Interestingly, it follows that $X$ produced in colliders would typically decay inside the tracker.
\mysubsection{Monojets.}
Since the $Y$'s do not significantly ionize, they will be registered as missing energy in events. Consequently, one can produce a monojet by having an $X$ particle depositing much more visible energy than the other or by recoiling the $X\bar X$ pair against a gluon. We checked the available Tevatron monojet searches~\cite{monojets}. The requirement of a jet with relatively high-$p_T\,$, missing energy separation and track isolation rejection are sufficient to greatly suppress the number of events passing the search criteria, to a level comparable but slightly larger than their $95\%$ CL. Since some of the cuts depend on the analysis response to the presence of the in-flight decays, one cannot asses whether the model is ruled out by this searches without a proper simulation.
\mysubsection{LEP and $e^+e^-$ colliders.} Since $X$ couples to the SM through strong interactions, its production rate at LEP is suppressed. The only relevant production comes through gluon radiation followed by a splitting to $X$-particles and as such is not constrained~\cite{lepstudies}. Similarly virtual corrections to QCD observables are not constraining enough~\cite{Kaplan:2008pt}. On the other hand, $Y$-particles couple to the $Z$ boson and may therefore be constrained by the $Z$-width measurement at LEP-I. Its small charge implies a contribution to the invisible $Z$-width, which is found to be $0.88$~MeV. Thus even though the LEP bound is particularly strong due to a downward fluctuation, $\Gamma_{\rm inv}^{\rm NP} < 2$~MeV at $95\%$ CL~\cite{0509008}, $Y$ easily evades it. In fact, even a $1/6$ charge is not excluded. Finally, constraints from lower energy $e^+e^-$ colliders based on $dE/dx$ do not constrain the $Y$ particles either~\cite{lowenergy}.
\mysubsection{Cosmology.}
There are no cosmological constraints on $X$ particles since they are unstable and decay almost promptly. However, for any stable fractionally charged relic, such as the $Y$ particle, there are severe constraints on its present abundance coming from a number of searches. The strongest comes from liquid drop experiments with mineral and silicon oil (for a review and references see~\cite{Perl:2009zz}), requiring concentrations smaller than $O(10^{-17})$. In principle these limits do not directly translate into a relic density bound, due to large ``environmental'' uncertainties from the chemistry of the $Y$'s~\cite{Lackner:1982rq} and from additional reprocessing stages in the core of the first stars~\cite{Goldberg:1982af}, both diluting the Y's. A conservative stance would be to require a relic abundance directly compatible with such limits, $\Omega_Y h^2 \lesssim 10^{-18}$. This is clearly impossible in a standard thermal history with high reheating temperature. Lowering the reheating temperature as much as allowed by nucleosynthesis still does not help since the electromagnetic interactions are strong enough to thermally populate the Y's during reheating but not strong enough to sufficiently deplete them during the freeze-out phase. Modifying the model Lagrangian by allowing additional annihilation channels for the Y's into particles not directly coupled to the plasma lowers the abundance down to $10^{-11}\div 10^{-12}$~\cite{Giudice:2000ex}, still not sufficient to respect the bound. One possibility would be to also lower $T_{max}$, the maximum temperature reached during the reheating era, far below $m_Y$, in order to gain a further Boltzmann suppression at the price of an extremely unnatural shape of the inflaton potential. A second, more natural, way out is to allow $Y$ to further decay to a lighter particle $Z$ with an even smaller electric charge, further relaxing the bounds from liquid drop searches.
\mysubsection{Cosmic Rays.}
$X$ and $Y$ particles are regularly produced through Cosmic Ray (CR) interactions in the atmosphere. A flux of $Y$ particles can then be searched for in underground experiments. The most stringent constraint, derived by the MACRO experiment at Gran Sasso, is only sensitive to a $1/6$ charge and therefore irrelevant for the above model. Nonetheless, it is interesting to study the bounds of $Y$ with a $1/6$ charge.
The experiment places a $90\%$ CL bound of $5\times 10^{-15}\textrm{ cm}^{-2}{\rm s}^{-1}{\rm sr}^{-1}$ on the flux at the detector, assuming an isotropic production on the rocks around it~\cite{Ambrosio:2004ub}. Unfortunately, most $Y$ particles are produced at the top of the atmosphere and translating the above bound to a bound on the flux at the surface of the Earth is nontrivial and requires a precise modeling of the surrounding. We compute the $Y$ flux at the detector taking into account the measured CR flux at different altitudes and considering severals models of the rocky terrain above the detector. We find the $Y$ flux to be within a factor of two of the bound, consistent with the $90\%$ CL bound when taking systematic uncertainties into account. A more accurate constraint could only be derived with a better study of the environment.
Fractionally charged particles produced by CRs can also be searched for the same liquid drop experiments constraining the cosmological abundance. We compute the density of $Y$ particles accumulated at the Earth and find the model to be comfortably within current bounds.
\subsection{Predictions}
\label{sec:predictions}
The difficulty with making specific predictions for NOTs, is that it requires a detailed understanding of both the detector components and the algorithms used to reconstruct physics objects. For instance it would be nearly impossible for us to quantify how often a moderately long lived sbottom, that decays in the tracker, would be reconstructed as a single high $p_T\,$ track, given the sizable number of tracks in the decay. Nonetheless, for the case of a fractionally charged particle we can make a sensible set of estimates.
To make a prediction for CDF with the above model, we require that a measured track leaves at least 15 hits in the COT layers, and survives more than halfway through the COT before decaying. To estimate the number of hits in the COT we use both a Landau and Bichsel parametrizations~\cite{pdg} for the tail of the energy loss of the fractionally charged mesons, and define a hit to occur when at least 15\% of a MIPs energy loss is deposited within a layer. Using this minimal track definition, a prediction for the $p_T$ distribution of the
$X$-$Y$ model presented above, is compared to the data in Fig.~\ref{cdf}. Background was simulated using the DW tune in {\tt Pythia6.423}~\cite{pythia}.
Given that $m_X$ sets the rate, adjusting $m_X$ accordingly we find a reasonable agreement with the original data published by CDF. Based on the nature of new physics, this model predicts that at high $p_T$, tracks have fewer hits with almost no hits in the silicon tracker, and very likely a bad $\chi^2$ fit for the track. These tracks are precisely the types of tracks thrown out by CDF in their re-analysis~\cite{errata}. It would therefore be very useful to analyze these tracks in more detail. Of course we stress that a more accurate study of this signal is required, in order to take into account detector effects and the tracking algorithms. Nevertheless, it clearly shows that a new physics is an intriguing and viable possibility even in light of the errata~\cite{errata}.
\begin{figure}[t]
\begin{center}
\includegraphics[width=3.2in]{pTdistCDF}
\caption{
Charged track $p_T$ distribution. The dashed black line is the QCD prediction estimated using the DW tune in {\tt Pythia6.423}, while the dashed blue line is the prediction for the $X$-$Y$ model described in the text, with a best-fit value of $m_X=7$ GeV and $m_Y = 1$ GeV. The green curve is the sum of the two contributions, to be compared with the CDF data in red. }
\label{cdf}
\end{center}
\end{figure}
The cross section at ATLAS and CMS will be even higher than at the Tevatron, $\mathcal{O}(10 \mu b)$, so it would be useful to investigate how these NOTs from the model we study could show up at the LHC. Finding kinks may be more difficult than at the Tevatron given the larger amount of detector material in the tracker, which increases the probability of multiple scattering. To date, in order to manage backgrounds, ATLAS and CMS searches have required stringent track quality cuts. Consequently,
\cite{atlas,cms} would have missed NOTs of the type studied here. Given the large production cross section, it seems advantageous to expand the current searches by loosening the track quality cuts. In particular the nature of the silicon trackers at ATLAS and CMS allow for a lower threshold for tracker hits and may be well suited for discovering NOTs with intermittent hits.
As discussed above, while we have focused on the model parameters that could explain the original CDF data, there is a wide range of NOT phenomenology. In particular, the production rates may be significantly lower, thereby easing the tension with existing constraints. Developing new techniques to search for NOTs and expanding the benchmarks beyond those given in this paper, thus provide important directions for future theoretical and experimental studies.
\section*{Acknowledgement}
We thank L.~Dixon, S.~Nussinov, M.~Reece, R.~Snider, Y.~Shadmi, G.~Sterman, T.~Phillips and L-T.~Wang for useful discussions and G.~Elor for useful discussions and comments on the manuscript. We especially thank N.~Moggi, M.~Mussini and F.~Rimondi for valuable discussions on the CDF measurement. The work of PM was supported in part by NSF grant PHY-0653354. The work of MP and TV were supported in part by the Director, Office of Science, Office of High Energy and Nuclear Physics, of the US Department of Energy under Contract DE-AC02-05CH11231.
|
\section{Introduction}
The two-coupled chains model (TCCM) is a quasi-one-dimensional
system of interacting fermions located along two
Tomonaga-Luttinger chains coupled by a interchain hopping (usually
denoted by $t_{\perp}$). In the last two decades the TCCM has also
been investigated as a prototype model to test the possibility of
having a higher dimensional Luttinger liquid state
\cite{AN87}-\cite{CL95}. Moreover there were important predictions
indicating that the TCCM might also manifest quantum confinement
regime \cite{FA93}-\cite{LE07} at strong coupling. This quantum
transition occurs when the electronic motion becomes strictly one
dimensional and, as a result of the interaction, the renormalized
interchain hopping $t_{\perp}^{R}$ becomes nullified. Since the
strong coupling regime is very difficult to deal with even in such
a simplified system, this problem remains open to this date, even
if several perturbative and non-perturbative methods have been
used to treat the TCCM \cite{SO79}-\cite{PA93}.
A very powerful method in the analysis of low dimensional
fermionic systems is bosonization, used by Mattis and Lieb
\cite{MA65} in their exact solution of the Tomonaga-Luttinger
chain, a model obtained from a system of non relativistic
fermions linearizing the energy relations in the vicinity of the
Fermi points (see Tomonaga \cite{TO50}). The presence of the Dirac
sea generates {\it anomalous density commutators} of the
associated Kac-Moody algebra in the Tomonaga-Luttinger model which
are at the core of the bosonization method. Indeed, it was well known,
as was pointed out earlier by Jordan \cite{JO35} and Schwinger
\cite{SC59}, that when the number of degrees of freedom in the
ground state of a physical system becomes infinite, anomalous
non-zero commutators of currents or density operators are
naturally produced by such a state.
However, the solvability of the Tomonaga-Luttinger chain is due to
certain special features of the system which allow us to reduce it
into harmonic oscillator modes. This feature is absent in more realistic
models, like the TCCM model, at least for general interactions. One way
to proceed is to combine approximate bosonization schemes with the
Renormalization Group method (see \cite{VO92}). Another
possible promising approach is based on the combination of
Renormalization Group methods with the Ward Identities (WI), which
must however be derived with care. Indeed even in the single chain
problem the presence of anomalous commutators reflect themselves in
the presence of quantum anomalies in the WI; the
(formal) WI derived neglecting the ultraviolet (UV) cutoffs which are necessary
to regularize the theory are not correct and extra terms must be
added to it. This was pointed out by Johnson \cite{J61} and Georgi and
Rawls \cite{GR73} (in order to verify the Adler result in d=1+1 \cite{AB69}), by
deriving the correct WI for the massless Thirring model making use of the
exact solution. The presence of anomalies in the WI on the other
hand can be easily overlooked in a formal diagrammatic approach, as in
the Dzyaloshinkii and Larkin solution of the Tomonaga-Luttinger
model \cite{DL73} (see also \cite{SO79}); an anomaly is present
in the infinite bandwidth limit and the non-regularized WI is no
longer valid in that case.
The derivation of the WI for the TCCM is more tricky, as in
such a case there is no exact solution from which the WI can be
derived, and the presence of anomalies can be easily overlooked.
In this paper we derive, in section II, the (non-regularized) WI
associated with the existing chiral symmetry of the TCCM
proceeding formally from the functional integrals with no UV
cutoff; in section III we show that such an identity is
{\it violated}, by an explicit one-loop perturbative
calculation. We then consider in section IV the theory with an
ultraviolet momentum cutoff and we derive, following \cite{BE05},
the WI with an extra term generated by the presence of the cutoff
which actually breaks the local invariance of the model. This extra term in
the WI is written as a functional integral and in Section V we
show that its lowest order contribution in the
WI is {\it non vanishing} even when the UV cutoff is taken to infinity;
this signals the presence of an anomaly in the TCMM in this limit.
Benfatto and Mastropietro obtained for systems in the Tomonaga-Luttinger
model universality class in \cite{BE05} (see
also \cite{M07},\cite{BFM07}) a full non-perturbative control of
this extra term in the WI in the limit of a removed UV
cutoff, using the methods of Constructive Quantum Field Theory;
it turns out that such a correction is proportional, in the limit of
a removed cutoff, to the vertex function and, in the case of the
Tomonaga-Luttinger model, it reproduces the same WI given by the
anomalous commutators.
In the case of the TCCM only if the backscattering and the umklapp
interactions are neglected the model becomes solvable and the WI
correction term reduces to the Tomonaga-Luttinger model result;
with general interactions there is no reason for this to happen.
While the non-perturbative limit of the correcting term in the
Tomonaga-Lutinger model is probably technically quite hard to implement, we
believe that this correction term, which is found here for the first time, plays
an important role in the physical properties of the TCCM.
\section{The Two-Coupled-Chains Model}
The TCCM describes two spinless Tomonaga-Luttinger chains coupled together
by interchain hopping. After diagonalization the spinless electrons are
distributed among the bonding (b) and anti-bonding (a) bands with a Fermi
surface (FS) consisting of four points, as displayed in our Figure 1. The
corresponding bare free action $S_f$ for single particles located in the
vicinities of those Fermi points of the TCCM is, in this way, given by
\begin{eqnarray}
S_f = \sum_{j=a,b \atop \alpha= \pm} \int\limits_{{\rm x}}
{\psi^+}_{\alpha}^{j}({\rm x}) (i (\partial_{t} + \alpha
v_{F}^{j}\partial_{{x}}) + \gamma^{j} )\psi_{\alpha}^{j}({\rm x}),
\end{eqnarray}
where ${\rm x} =(t,x)$, the flavor index {\it j} refers to the
bonding and antibonding bands and the index $\alpha$ denotes the right
(+) and left (-) sides of the FS. Here $v_{F}^{j}$ is the Fermi velocity,
$k_{F}^{j}$ is the Fermi wave vector and $\gamma^{j} = v_{F}^{j} k_{F}^{j}$
is a characteristic energy for $j$-th ``flavor'', used to clarify Figure 1.
\begin{figure}[h]
\includegraphics[width=0.36\textwidth]{fig1.ps}
\caption{Fermi surface representation of TCCM, formed by the
location of four bare momenta ${\pm k_{F}^{a}}$ (antibonding band)
and ${\pm k_{F}^{b}}$ (boundig band).}
\end{figure}
Following common practice among workers in the field \cite{FA93}
we next define the corresponding bare interacting action $S_{int}$ by
\begin{eqnarray}
S_{int} &=& \sum\limits_{j,\alpha}\int\limits_{\rm x, x'}\bigg\{g_{0}{\psi^+}_{\alpha}^{j}({\rm x})
{\psi^+}_{-\alpha}^{j}({\rm x'})\psi_{-\alpha}^{j}({\rm x'})\psi_{\alpha}^{j}({\rm x}) +
\sum\limits_{i\neq j}\bigg[g_{bs} {\psi^+}_{\alpha}^{j}({\rm x}){\psi^+}_{-\alpha}^{i}({\rm x'})
\psi_{-\alpha}^{j}({\rm x'})\psi_{\alpha}^{i}({\rm x})
+ \nonumber \\
&+& g_{f} {\psi^+}_{\alpha}^{j}({\rm x}) {\psi^+}_{-\alpha}^{i}({\rm x'})\psi_{-\alpha}^{i}({\rm x'})
\psi_{\alpha}^{j}({\rm x}) +
g_{u} {\psi^+}_{\alpha}^{j}({\rm x}) {\psi^+}_{-\alpha}^{j}({\rm x'})
\psi_{-\alpha}^{i}({\rm x'})\psi_{\alpha}^{i}({\rm x})\bigg]\bigg\},
\end{eqnarray}
with $g_{0}\equiv g_{0}(x-x')$ being the intraband forward short
range interaction, and $g_{bs}\equiv g_{bs}(x-x')$
(backscattering), $g_{f}\equiv g_{f}(x-x')$ (forward) and
$g_{u}\equiv g_{u}(x-x')$ (umklapp) the three interband short
range interactions. See Figure 2, for illustration. Notice that
although $g_u$ ($g_{bs}$) doesn't preserve flavor globally
(locally), $S_{int}$ is fully symmetric under chiral ($+$,$-$)
transformation. If $g_u$ and $g_{bs}$ are vanishing the model becomes
solvable and equivalent to a Tomonaga-Luttinger model, while in the
presence of non vanishing $g_u$ and $g_{bs}$ an exact solution is
unknown.
\begin{figure}[h]
\includegraphics[width=0.48\textwidth]{fig2.ps}
\caption{Diagrammatic representation for the TCCM interactions for
right- and left-electrons, respectively, with $j \neq j' $. Full and
dashed lines represent $+$ and $-$ electrons, respectively, in the vicinities of $ k_{F}^{j}$
and $ - k_{F}^{j}$ respectively.}
\end{figure}
Using the total bare action $S= S_f + S_{int}$, we define next the functional generators $Z$
and $W$ by
\begin{eqnarray}
Z [\xi, {{\xi^+}}, J] = {\rm e}^{i W [\xi, {\xi^+}, J]} = \int [d {{\psi^+}}] [d \psi] \;
\exp{ [ i ( S_f + S_{int} + \xi \cdot \psi + {{\psi^+}} \cdot {{\xi^+}}
+ J \cdot {{\psi^+}} \psi ) ]},
\end{eqnarray}
where, for simplicity, we use the notation $ \xi \cdot \psi =
\sum\limits_{j, \alpha} \int\limits_{\rm x} \; \xi^{j}_{\alpha}
({\rm x}) \psi^{j}_{\alpha} ({\rm x}) $.
Let us next apply the local phase transformation to the fermionic fields:
\begin{eqnarray}
\psi_{\alpha}^{j}({\rm x})\rightarrow e^{-i\lambda({\rm x})}\psi_{\alpha}^{j}({\rm x}), \\
{\psi^+}_{\alpha}^{j}({\rm x})\rightarrow e^{i\lambda({\rm x})} {\psi^+}_{\alpha}^{j}({\rm x}).
\end{eqnarray}
Under these $U(1)$ transformations, the action S becomes
\0
S \rightarrow S - \sum\limits_{\alpha, j} \int\limits_{{\rm x}} \lambda ({\rm x})
(\partial_t + \alpha v_{F}^{j} \partial_{x}) {\psi^+}_{\alpha}^{j} ({\rm x}) \psi_{\alpha}^{j} ({\rm x}),
\1
and the source terms transform as
\0
\xi \cdot \psi \rightarrow \xi \cdot \psi - i \sum\limits_{\alpha, j}
\int\limits_{{\rm x}} \lambda ({\rm x}) \xi_{\alpha}^{j} ({\rm x}) \psi_{\alpha}^{j} ({\rm x}),
\1
\0
{\psi^+} \cdot {\xi^+} \rightarrow {\psi^+} \cdot {\xi^+} + i \sum\limits_{\alpha, j}
\int\limits_{{\rm x}} \lambda ({\rm x}) {\psi^+}_{\alpha}^{j} ({\rm x}) {\xi^+}_{\alpha}^{j} ({\rm x}).
\1
Since these phase transformations leave $Z$ invariant, we immediately arrive at the WIs
\begin{eqnarray}
\sum\limits_{j, \alpha} \left[ i ( \partial_t + \alpha v_{F}^{j} \partial_{x} ) \frac{\delta }{\delta J_{\alpha}^{j} ({\rm x})}
- \xi_{\alpha}^{j} ({\rm x}) \frac{\delta }{\delta \xi_{\alpha}^{j}({\rm x}) } +
{{\xi^+}}_{\alpha}^{j}({\rm x}) \frac{\delta }{\delta {{\xi^+}}_{\alpha}^{j}({\rm x}) } \right] Z = 0,
\end{eqnarray}
or, equivalently
\begin{eqnarray}
\sum\limits_{j, \alpha} \left[ i ( \partial_t + \alpha v_{F}^{j} \partial_{x} ) \frac{\delta }{\delta J_{\alpha}^{j} ({\rm x})}
- \xi_{\alpha}^{j} ({\rm x}) \frac{\delta }{\delta \xi_{\alpha}^{j}({\rm x}) } +
{\bar{\xi}}_{\alpha}^{j}({\rm x}) \frac{\delta }{\delta {\bar{\xi}}_{\alpha}^{j}({\rm x}) } \right] W = 0.
\end{eqnarray}
Alternatively we can rewrite the WIs in terms of $\Gamma[\< {\psi^+} \>, \< \psi \>, J]$, the
generator of the one-particle irreducible functions. Defining
\begin{eqnarray}
\Gamma [\langle \psi \rangle, \langle \psi^+ \rangle, J] = W
[J, \xi, {\bar{\xi}}] - {{\xi^+}}_{\alpha}^{j} \cdot \langle \psi_{\alpha}^{j} \rangle +
\langle {{\psi}^+}_{\alpha}^{j} \rangle \cdot \xi_{\alpha}^{j} ,
\end{eqnarray}
with
\begin{eqnarray}
\langle \psi_{\alpha}^{j} ({\rm x}) \rangle = \frac{\delta W}{\delta {{\xi^+}}_{\alpha}^{j} ({\rm x})}, \;\;\;\;\;\;\;\;\;\;\;\;
\langle {{\psi^+}}_{\alpha}^{j} ({\rm x}) \rangle = - \frac{\delta W}{\delta \xi_{\alpha}^{j} ({\rm x})},
\end{eqnarray}
we find
\0
\sum\limits_{j, \alpha} i ( \partial_t + \alpha v_{F}^{j} \partial_{x} ) \frac{\delta \Gamma}{\delta J_{\alpha}^{j} ({\rm x})}
= \sum\limits_{j, \alpha} \left( \< {\bar \psi}_{\alpha}^{j}({\rm x}) \>
\frac{ \delta }{ \delta \< {\psi^+}_{\alpha}^{j}({\rm x}) \>}
- \< \psi_{\alpha}^{j} ({\rm x}) \> \frac{\delta }{\delta \< \psi_{\alpha}^{j} ({\rm x}) \>} \right) \Gamma.
\1
Consequently, if we differenciate this last equation with respect to
\0
\frac{ \delta^{2} }{ \delta \< \psi_{\beta}^{k} ({\rm x}')\>
\delta \< {\psi^+}_{\beta}^{k} ({\rm x}'')\> },
\1
we obtain, after setting the fields equal to zero,
\2
\sum\limits_{j, \alpha} &i& ( \partial_t + \alpha v_{F}^{j} \partial_{x} )
\frac{\delta^{3} \Gamma }{\delta \< \psi_{\beta}^{k} ({\rm x}')\>
\delta \< {\psi^+}_{\beta}^{k} ({\rm x}'')\> \delta J_{\alpha}^{j} ({\rm x})}
= \nonumber \\
&=& \delta ({\rm x - x'}) \frac{ \delta^{2} \Gamma}{ \delta \< \psi_{\beta}^{k} ({\rm x})\>
\delta \< {\psi^+}_{\beta}^{k} ({\rm x}'')\> } - \delta({\rm x - x''})
\frac{ \delta^{2} \Gamma}{ \delta \< \psi_{\beta}^{k} ({\rm x}')\> \delta \< {\psi^+}_{\beta}^{k} ({\rm x})\> },
\3
which, in momenta space, reduces to the more conventional form
\begin{equation}
\sum\limits_{{\alpha}, j} \left( q_0 - {\alpha} v_{F}^{j} q \right)
\Gamma_{\beta, {\alpha}}^{k, j (2,1)} ({\rm p + q, p}) = \Gamma_{\beta}^{k (2)} ({\rm p + q}) - \Gamma_{\beta}^{k (2)} ({\rm p}),
\end{equation}
with ${\rm p} = (p_0, p) $ and $\Gamma_{\beta}^{k (2)} ({\rm p}) =
(G_{\beta}^{k} ({\rm p}) )^{-1}$ being the inverse of the $(k, \beta)$-th
single particle propagator. Writing
\0
\Gamma_{\beta, {\alpha}}^{k, j (2,1)} ({\rm p + q, p}) =
\delta_{j}^{k} \delta_{\alpha}^{\beta} + \Lambda_{\beta, \alpha}^{k, j} ({\rm p + q, p}),
\1
this can be expressed in a more convenient form:
\0
\sum\limits_{{\alpha}, j} ( q_0 - {\alpha} v_{F}^{j} q )
\Lambda_{\beta, {\alpha}}^{k, j} ({\rm p + q, p}) = \Sigma_{\beta}^{k} ({\rm p}) - \Sigma_{\beta}^{k} ({\rm p + q}),
\1
where $\Sigma_{\beta}^{k} ({\rm p})$ is the self-energy associated with the $(k, \beta)$-th particle.
\section{Perturbative Test of the Ward Identity}
The above derivation of the WI is purely formal; the functional
integrals have just a formal meaning so that the results of their
manipulations may not be true. This is precisely what happens here
since, as we can see from perturbation theory, this WI is no longer valid.
In zero-th order the previously derived WI is trivially
satisfyed. Choosing, for convenience, $k = b$ and $ \beta = +$, we have
simply in the zero-th order: \0 \Gamma_{+}^{b (2)} ({\rm p + q} )
- \Gamma_{+}^{b (2)} ({\rm p}) = q_0 - v_{F}^{b} q \1 and \0
\Gamma_{+, \alpha}^{b, j} ({\rm p+q, q} ) = \delta_{+, \alpha}
\delta_{b, j}, \1 in agreement with equation (16). In 1-loop order
there are just two tadpole contribuctions to the self-energy, as
indicated in Figure (3).
\begin{figure}[h]
\includegraphics[width=0.21\textwidth]{fig3.ps}
\caption{One-loop self-energy diagrams for a
bonding electron in the $ k_{F}^{b} $ branch. The couplings are $g_0$ and $g_f$, respectively.}
\end{figure}
It is straightforward to calculate these two diagrams and we immediatly find
\0
\Sigma_{+}^{b} ({\rm p} ) - \Sigma_{+}^{b} ({\rm p+q} ) = 0.
\1
There are two vertex function in 1-loop order associated with those self-energy diagrams:
$\Lambda_{+,-}^{b, b}$ and $\Lambda_{+,-}^{b, a}$. They are displayed in Figure (4).
\begin{figure}[h]
\includegraphics[width=0.21\textwidth]{fig4.ps}
\caption{The two vertices corresponding to the self-energies diagrams of FIG.3. }
\end{figure}
Calculating their contributions we find
\0
\Lambda_{+,-}^{b, b} ({\rm p+q, p}) = \frac{g_0 (q)}{2 \pi} \vert q \vert G_{-}^{b} ({\rm q}),
\1
\0
\Lambda_{+,-}^{b, a} ({\rm p+q, p}) = \frac{g_f (q)}{2 \pi} \vert q \vert G_{-}^{a} ({\rm q}),
\1
with $G_{-}^{b} (G_{-}^{a})$ being the noninteracting Green's function for the bonding (anti-bonding)
spinless ``$-$'' electrons. As a result it follows from that
\0
(q_0 + v_{F}^{b} q) \Lambda_{+, -}^{b, b} + (q_0 + v_{F}^{a} q) \Lambda_{+, -}^{b, a} = \frac{ \vert q \vert }{2 \pi} ( g_0 (q) + g_f (q) ),
\1
in clear disagreement with our WI, for $ q \ne 0 $. If we take the ultraviolet limit $\Lambda \rightarrow \infty$
this WI violation reflects the chiral symmetry breaking produced by the Dirac seas for the bonding and
anti-bonding bands. However if $\Lambda$ is finite our derivation of the WI should also be modified to
take the presence of such a cuttoff explicitly into account. We will do that next.
\section{Ward Identity in the presence of the cutoff}
The WI in the presence of a finite cutoff for the Luttinger liquid
was obtained by Benfatto and Mastropietro \cite{BE05} and here we
adopt their derivation for our purposes, considering imaginary
times for convenience. Let us therefore define the functional
generator $W$ in the presence of the cutoff $\Lambda$ as
\0
{\rm exp} (- W) = \int {\cal D} [ {\psi^+} \psi] {\rm exp} \;
\left( S_{\Lambda} + {\xi^+} \cdot \psi + {\psi^+} \cdot \xi
+ J \cdot {\psi^+} \psi \right).
\1
The cutoff $\Lambda$ manifests itself in the regulators
$C_{j}^{\alpha} ( {\rm x} , \Lambda)$'s which now appears in our
definition of the non-interacting action $S_{f}^{\Lambda}$:
\0
S_{f}^{\Lambda} =\sum\limits_{\alpha, j} \int\limits_{{\rm k}}
{\psi^+}_{\alpha}^{j} ({\rm k}) (C_{\alpha}^{j} (k,
\Lambda)^{-1} (i k_0-\alpha v_{F}^{j} k + \gamma^{j} )
\psi_{\alpha}^{j} ({\rm k}) ),
\1
where $C_{\alpha}^{j} ( p ; \Lambda)$ can either be a step function or any other smoother
function we wish. This non-interacting action is now consistent
with the definition of a non-interacting single particle
propagator $G_{\alpha}^{j} ( {\rm p} )$
$$G_{\alpha}^{j} ({\rm p}) = \frac{ C_{\alpha}^{j} ( p ; \Lambda)}{
-i p_0 + \alpha v_{F} (p - \alpha k_{F}^{j})}. $$
If we now follow the derivation of the WI as before and work in energy-momentum space we find
\2
& \int & {\cal D} [{\psi^+}, \psi ] \sum\limits_{\alpha, j} \left\{ ( i
q_0 - \alpha v_{F}^{j} q) \int\limits_{{\rm k}} {\psi^+}_{\alpha}^{j}
({\rm k+q}) \psi_{\alpha}^{j} ({\rm k}) +
\int\limits_{{\rm k}} {\xi^+}_{\alpha}^{j} ({\rm k + q})
\psi_{\alpha}^{j}
- \int\limits_{{\rm k}} {\psi^+}_{\alpha}^{j} ({\rm k + q}) \xi_{\alpha}^{j} ({\rm k}) \right. \nonumber \\
&+& \left. \int\limits_{{\rm k}} {\psi^+}_{\alpha}^{j} ({\rm k
+ q}) ( C_{\alpha}^{j} ({\rm k, q}))^{-1} \psi_{\alpha}^{j} ({\rm
k}) \right\} {\rm exp} \; i (S + {\xi^+} \cdot \psi + {\psi}
\cdot \xi + J \cdot {\rm \psi} \psi ) = 0,
\3
where
$\int\limits_{{\rm k}} = (2 \pi)^{-2} \int dk_0 \int dk $ and
\2
(C_{\alpha}^{j} ({\rm k, q}))^{-1} &=& \left[ (C_{\alpha}^{j} ({k +
q}, \Lambda ))^{-1} - 1 \right] \left[ i(k_0 + q_0) - \alpha
v_{F}^{j} (k + q) \right] - \nonumber \\
&-& \left[ (C_{\alpha}^{j} ( {k}, \Lambda
))^{-1} - 1 \right] \left[ i k_0 - \alpha v_{F}^{j} k \right].
\3
That is, the presence of the cutoff produces a new $\Lambda$-dependent term in the WI.
For convenience we introduce a new source term to our functional generator $W_{\Lambda}$, namely
\0 \sum\limits_{\alpha, j } \int\limits_{{\rm q, k}}
\chi_{\alpha}^{j} (q) {\psi^+}_{\alpha}^{j} ({\rm k + q}) (
C_{\alpha}^{j} ({k, q}))^{-1} \psi_{\alpha}^{j} ({\rm k}).
\1
In this way, our generalized cutoff dependent WI becomes
\2
\sum_{\alpha, j} &(i q_0 - \alpha v_{F}^{j} q )& \frac{\delta W_{\Lambda}}{ \delta J_{\alpha}^{j} ({\rm q})} \nonumber \\
&=& \sum_{\alpha, j} \left\{ \int_{\rm k} \left( \xi_{\alpha}^{j} (\rm k) \frac{\delta}{\delta \xi_{\alpha}^{j} ({\rm k + q })} -
{\xi^+}_{\alpha}^{j} (\rm k) \frac{\delta}{\delta {\xi^+}_{\alpha}^{j} ({\rm k + q })} \right) -
\frac{\delta}{\delta \chi_{\alpha}^{j} ({\rm q})} \right\} W_{\Lambda}.
\3
Defining the 1PI functional generator $\Gamma$ as before we can, equivalently, rewrite this WI as
\2
\sum_{\alpha, j} &\; & ( i q_0 - \alpha v_{F}^{j} q ) \frac{\delta \Gamma}{ \delta J_{\alpha}^{j} ({\rm q})} \nonumber \\
&=& \sum_{\alpha, j} \left\{ \int_{\rm k} \left( \< \psi_{\alpha}^{j} (\rm k) \> \frac{\delta}{\delta \< \psi_{\alpha}^{j} ({\rm k + q }) \> } -
\< {\psi^+}_{\alpha}^{j} (\rm k) \> \frac{\delta}{\delta \< {\psi^+}_{\alpha}^{j} ({\rm k + q }) \> } \right) -
\frac{\delta}{\delta \chi_{\alpha}^{j} ({\rm q})} \right\} \Gamma.
\3
If we functionally differentiate this last result with respect
\0
\frac{ \delta^{2} }{ \delta \< \psi_{\alpha}^{j} ({\rm k}')\>
\delta \< {\psi^+}_{\alpha}^{j} ({\rm k}'')\> },
\1
we obtain
\begin{equation}
\sum\limits_{{\alpha}' , j'} \left( i q_0 - {\alpha}' v_{F}^{j'} q
\right) \Gamma_{\alpha, \alpha'}^{j, j' (2,1)} ({\rm k + q, k}) =
\Gamma_{\alpha}^{j (2)} ({\rm k + q}) - \Gamma_{\alpha}^{j (2)}
({\rm k}) + \sum\limits_{\alpha', j'} \Delta_{\alpha, \alpha'}^{j,
j'} ({\rm k, q}),
\end{equation}
where
\0
\Delta_{\alpha, \alpha'}^{j, j'} ({\rm k, q}) = - \frac{\delta^3 \Gamma }{ \delta \< \psi_{\alpha}^{j} ({\rm k})\>
\delta \< {\psi^+}_{\alpha}^{j} ({\rm k + q})\> \delta \chi_{\alpha'}^{j'} ({\rm q}) }.
\1
These $\Delta$'s are precisely the new terms produced by the finite cutoff $\Lambda$.
\section{Ward Identity restauration in the presence of a cutoff}
In the previous section we derived the exact WI for the theory with a cutoff;
the functional integrals are well defined and the throughout resulting
WI differs with respect to its non-regularized counterpart by a
correcting term. Of course, now the WI should be true order by order in
perturbation theory. Indeed let us calculate this cutoff
contribution in one-loop order.
\begin{figure}[h]
\includegraphics[width=0.18\textwidth]{fig5.ps}
\caption{$\Delta_{+,-}^{j, j'}$ in one-loop order.}
\end{figure}
Since at this order we have that
\0
(C_{-}^{j'} ( {\rm k', q},
\Lambda ) )^{-1} = - (i q_0 + v_{F}^{j'} q) + (C_{-}^{j'} ({k' +
q}))^{-1} ( i(k_0' + q_0) + v_{F}^{j'} (k' + q) ) - (C_{-}^{j'}
(k))^{-1} (i k_0' + v_{F}^{j'} k' ),
\1
it follows immediately that
{\small{\2
\sum\limits_{j'} \int\limits_{k'} &\;& \left[ (
C_{-}^{j'}({\rm k' + q}))^{-1} \left[ i(k_0' + q_0) + v_{F}^{j'}
(k' + q) \right] - (C_{-}^{j'} ({k}))^{-1} \left[ i k_0' +
v_{F}^{j'} k' \right] \right]
G_{-}^{j'} ({\rm k'}) G_{-}^{j'} ({\rm k' + q} ) \nonumber \\
&=& \sum\limits_{j'} \int\limits_{k'} \left[ G_{-}^{j'} ({\rm
k'}) - G_{-}^{j'} ({\rm k' + q} ) \right] = 0.
\3 }}
As a result, we obtain the trivial identity
\0
\sum\limits_{j'= b,a} ( i q_0 +
v_{F}^{j'} q) \Lambda_{+,-}^{b, j'} = \sum\limits_{j'= b,a} (i q_0
+ v_{F}^{j'} q) \Lambda_{+,-}^{b, j'}
\1
and, as expected, the WI is fully restored for a finite cutoff. In the limit
$\Lambda\to\infty$, $\Lambda_{+,-}^{b, j'}$ is given by (22),(23)
times an extra $i$.
Note that the lowest order contribution to $\Delta^{jj'}_\pm$ is
non vanishing even in the limit $\Lambda\to\infty$. A genuine
anomaly is present in this case since the regularized WI differs from its
non-regularized counterpart. Note also that, even if the lowest order contribution
of $\Delta^{jj'}_\pm$ depends only on $g_0$ and $g_f$, the higher order
terms depend on the other couplings as well.
\section{Conclusion}
In the absence of backscattering and umklapp interactions the TCCM
can be solved exactly by bosonization methods but such
interactions cannot be mapped into a quadratic bosonic expression;
therefore no solution exists in the general case. In order to
derive the WI, we perform first a phase transformation in the
functional integrals with no cutoff to derive identities
connecting the vertex function with the self energy. The non-regularized
formal WI's obtained in this way, even if widely accepted, are however not
always true, as an explicit one-loop calculation demonstrates.
Indeed, without regularizations the functional integrals are
ill-defined and one should not take for granted the WI which result from their
manipulation. On the other hand, if a momentum cutoff is
imposed from the start the functional integrals are well defined and an exact WI
is derived, with a new extra term $\Delta^{jj'}_\pm$ with
respect to the non-regularized result. This extra term can be written as a
functional integral and a perturbative computation shows that its
first order contribution is already {\it non} vanishing, signaling the presence of the anomaly
in the infinite bandwidth limit.
Note that the correction term $\Delta^{jj'}_\pm$ differs from result obtained for
the Tomonaga-Luttinger model, since it depends on {\it all} the interactions appearing in
the TCCM. Only if $g_b$ and $g_u$ are nullified one can apply the non
perturbative analysis of \cite{BE05} which gives $\Delta^{jj'}_\pm$ proportional to
the vertex function, in the limit of a removed cutoff, in agreement with the result
obtained from the anomalous commutators. In the general case, however, there is
no reason to expect that the WI for the TCCM has the same form as in the
Tomonaga-Luttinger model. While the non-perturbative limit of the correcting
term in the Tomonaga-Luttinger model is probably technically quite hard to analyze,
we believe that this correcting term, plays an important role in the physical
properties of the TCCM as well.
\begin{acknowledgments}
We acknowledge the comments we had from E. Kochetov and P. Kopietz. LCC would also thanks
G. Bonneau, L. Bartosch, A. Isidori, A. Kreisel and E. Correa for many valuable discussion. This
work was supported by the CNPq, CAPES (Probral) and by the Brazilian Min. of Science and Technology.
\end{acknowledgments}
|
\section{Introduction}
An\hspace{0.1cm}outstanding question regarding solar flares is where, when, and how electrons
are accelerated. The\hspace{0.1cm}direct detection of X-ray and radio emission from an acceleration region has
proved difficult. The\hspace{0.1cm}detection of X-rays from the\hspace{0.1cm}acceleration site
is challenging due to firstly a\hspace{0.1cm}relatively low density of the\hspace{0.1cm}surrounding
coronal plasma and secondly due to the\hspace{0.1cm}presence of competing emissions,
i.e., emission from hot flare loop plasma and trapped electron populations.
In addition, as the\hspace{0.1cm}hard X-ray (HXR)\hspace{0.1cm}flux is proportional to the\hspace{0.1cm}plasma density,
the\hspace{0.1cm}bulk of HXRs are emitted in the\hspace{0.1cm}dense plasma of the\hspace{0.1cm}chromosphere
(HXR footpoints) making X-ray imaging of tenuous coronal emission
problematic \citep{2003ApJ...595L..69L,emslie2003,2007LNP...725...65B}.
Studies of flares with footpoints occulted by the\hspace{0.1cm}solar disk \citep{krucker_lin_2008, krucker_etal_2010} provide direct imaging of the\hspace{0.1cm}looptop X-ray
emission, but are hampered because essential information on the\hspace{0.1cm}flare energy release contained in the\hspace{0.1cm}precipitating electrons becomes unavailable. What is
needed is to cleanly separate the\hspace{0.1cm}acceleration and precipitation regions while retaining observations of both. Having both radio and X-ray observations of
a flare without significant plasma heating and without noticeable magnetic trapping would provide the\hspace{0.1cm}needed information on both components to make
characterization of the\hspace{0.1cm}acceleration region possible.
Recently \citet{Bastian_etal_2007} have reported a\hspace{0.1cm}cold flare observed from a\hspace{0.1cm}very dense loop, where no significant heating occurred simply because the\hspace{0.1cm}flare energy was deposited to denser than usual plasma resulting in lower
than usual flaring plasma temperature. Although in their case highly important
implications for the\hspace{0.1cm}plasma heating and electron acceleration have been obtained,
the\hspace{0.1cm}strong Coulomb losses in the\hspace{0.1cm}dense coronal loop did not allow observing the\hspace{0.1cm}acceleration and thick-target regions separately.
From this perspective it would be better to have a\hspace{0.1cm}cold and \emph{tenuous}, rather than dense, flare.
However, such cold, tenuous flares seem to be unlikely because the\hspace{0.1cm}energy deposition from non-thermal
particles should result in even greater plasma heating in the\hspace{0.1cm}case of tenuous than dense plasma.
Nevertheless, inspection of available X-ray and radio databases reveals a\hspace{0.1cm}number of cold flare
candidates, one of the\hspace{0.1cm}most vivid examples of which is presented in this letter.
Specifically, we present and discuss an event (i)\hspace{0.1cm}lacking any noticeable GOES enhancement, (ii)\hspace{0.1cm}not showing any coronal X-ray source between the\hspace{0.1cm}footpoint sources, (iii)\hspace{0.1cm}with X-ray spectra
well fitted by a\hspace{0.1cm}thick-target model without any thermal component, and (iv)\hspace{0.1cm}producing relatively
low-frequency GS microwave continuum emission, which all together proves the\hspace{0.1cm}phenomenon
of the\hspace{0.1cm}cold, tenuous flare in the\hspace{0.1cm}case under study. The\hspace{0.1cm}available data offer evidence
that the\hspace{0.1cm}observed microwave GS emission is produced directly in the\hspace{0.1cm}acceleration
region of the\hspace{0.1cm}flare, and hence, parameters derived from microwave spectrum pertain to the\hspace{0.1cm}directly accelerated electron population.
\section{ X-ray and radio observations}
\label{S_XR_observ}
The\hspace{0.1cm}July 30, 2002 flare appeared close to the\hspace{0.1cm}disk center (W10S07, NOAA AR10050)\hspace{0.1cm}and
is brightly visible in HXR images obtained by the\hspace{0.1cm}RHESSI spacecraft \citep{lin2002}.
Although the\hspace{0.1cm}radio emission recorded by Phoenix-2 spectrometer at 0.1--4\hspace{0.1cm}GHz \citep{1999SoPh..187..335M}
and Owens Valley Solar Array (OVSA) at 1--18\hspace{0.1cm}GHz also indicate an abundance of non-thermal
electrons, the\hspace{0.1cm}flare has weak or nonexistent thermal soft X-ray emission (Figures\hspace{0.15cm}\ref{fig_30_jul_2002_over}--\ref{fig:Ximage}).
The\hspace{0.1cm}GOES light curves are almost flat at the\hspace{0.1cm}C2.2 level and the\hspace{0.1cm}temperature inferred
does not exceed $6.1$\hspace{0.1cm}MK. The\hspace{0.1cm}OVSA dynamic spectrum (4\hspace{0.1cm}s time resolution) displays
a low-frequency microwave continuum burst with a\hspace{0.1cm}peak frequency at 2--3\hspace{0.1cm}GHz.
The\hspace{0.1cm}Phoenix-2 dynamic spectrum obtained with higher (0.1\hspace{0.1cm}s) time resolution
shows that a\hspace{0.1cm}few type III-like features are superimposed on this
continuum at 1--4\hspace{0.1cm}GHz. There are also nonthermal emissions below 1\hspace{0.1cm}GHz.
\begin{figure}\centering
\includegraphics[width=0.9\columnwidth]{overview.eps}\\
\caption{\label{fig_30_jul_2002_over} Overview of July 30, 2002 flare: Phoenix-2 and OVSA dynamic spectra, top panels. GOES (3\hspace{0.1cm}s) lightcurves and temperature
as measured by GOES-10 spacecraft assuming photospheric abundances
from CHIANTI v5.2 (middle panels). RHESSI (2 second bins) lightcurves (bottom panel)
in: 3-9\hspace{0.1cm}keV (black), 9-15\hspace{0.1cm}keV (red), 15-30\hspace{0.1cm}keV (orange), 30-100\hspace{0.1cm}keV (blue). }
\end{figure}
The\hspace{0.1cm}spatially integrated RHESSI \citep{lin2002} X-ray spectrum
(Figure\hspace{0.15cm}\ref{fig:peak_spectr}) over the\hspace{0.1cm}duration of the\hspace{0.1cm}peak of the\hspace{0.1cm}flare indicates
strong non-thermal emission above $6$\hspace{0.1cm}keV and very weak or no thermal emission.
Spectral analysis (for spectroscopy, detectors 2 and 7 were avoided) was done using
OSPEX \citep{Schwartz_etal2002} with systematic errors
set to zero \citep[e.g.][]{Su_etal2009}. Since the\hspace{0.1cm}flare was close to the\hspace{0.1cm}disk center
(heliocentric angle $\sim20^o$), albedo correction was applied
\citep{kontar2006} assuming isotropic emission and hence minimum correction.
The\hspace{0.1cm}spectrum was fitted with the\hspace{0.1cm}standard thermal plus non-thermal thick-target
model \citep{brown1971} with $\chi^2{\sim}0.7$ (Figure\hspace{0.15cm}\ref{fig:peak_spectr}).
Bremsstrahlung cross-section following \citet{haug1997} has been utilized \citep{2006ApJ...643..523B}.
However, given the\hspace{0.1cm}clear lack of the\hspace{0.1cm}thermal emission attributes in the\hspace{0.1cm}event,
we also used a\hspace{0.1cm}purely non-thermal fit and found that the\hspace{0.1cm}count spectrum can
be nearly as well-fitted ($\delta_x{\sim}3.55$) without any thermal component,
$\chi^2\sim1.8$ (Figure\hspace{0.15cm}\ref{fig:peak_spectr}).
\begin{figure}\centering
\includegraphics[width=0.9\columnwidth]{sp_fit20020730_173730_173744.eps} \\
\includegraphics[width=0.9\columnwidth]{sp_fit_noth20020730_173730_173744.eps}\\
\caption{\label{fig:peak_spectr} RHESSI X-ray spectrum (black) above 6\hspace{0.1cm}keV near the\hspace{0.1cm}peak of the\hspace{0.1cm}event
from $17:37:30-17:37:44$\hspace{0.1cm}UT using thermal (red-dashed) and thick-target model (red-dashed)
(top panel) with best fit (blue) parameters: emission measure $EM=9.7\times10^{45}$\hspace{0.1cm}cm$^{-3}$, temperature $16$\hspace{0.1cm}MK, $N_e(>6keV)=2.7\times10^{35}$\hspace{0.1cm}s$^{-1}$,
$\delta_{x}=3.56$. Thick-target model only (bottom panel): $N_e(>6keV)=2.8\times10^{35}$\hspace{0.1cm}s$^{-1}$, $\delta_{x}=3.56$. The pre-flare background is shown in brown.
}
\end{figure}
We have considered the\hspace{0.1cm}variation of spectral parameters over time using the\hspace{0.1cm}thick-target
model and fitting data above 10\hspace{0.1cm}keV (strong background below 10\hspace{0.1cm}keV does not allow reliable time-dependent fit at lower energies, Figure\hspace{0.15cm}\ref{fig:peak_spectr}). Both the\hspace{0.1cm}number
of accelerated electrons and the\hspace{0.1cm}spectral index demonstrate typical
soft-hard-soft behavior \citep[e.g.][]{1985SoPh..100..465D}. The\hspace{0.1cm}hardest
electron spectra $\delta_x{\sim}3.5$ are reached around
$17:37:40$\hspace{0.1cm}UT. At the\hspace{0.1cm}same time, the\hspace{0.1cm}electron acceleration rate
has its maximum $F_{e\max}(>10keV){\simeq}10^{35}$\hspace{0.1cm}electrons
per second.
\begin{figure}\centering
\qquad\includegraphics[width=0.9\columnwidth]{eit_rhessi_im.eps}
\includegraphics[width=0.8\columnwidth]{maps02.eps}\\
\includegraphics[width=0.8\columnwidth]{maps0.eps}
\caption{\label{fig:Ximage} Top panel: Spatial distribution of X-ray emission from July 30, 2002 flare
in various energy ranges contours at 30, 50, 70, 90\% levels: $9-15$\hspace{0.1cm}keV (red) $15-30$\hspace{0.1cm}keV (orange), $30-100$\hspace{0.1cm}keV (blue).
Time accumulation interval for RHESSI images is $17:37-17:38$\hspace{0.1cm}UT. Background image is SoHO EIT\hspace{0.1cm}195
taken just before the\hspace{0.1cm}flare at $17:36$\hspace{0.1cm}UT. Middle and bottom panels: the\hspace{0.1cm}full and close-up view of the\hspace{0.1cm}active region and a\hspace{0.1cm}potential flux tube (green)
connecting two X-ray footpoints (blue contours), $2.6-3.2$\hspace{0.1cm}GHz radio image (red contours)
and $4.2-8.2$\hspace{0.1cm}GHz (yellow contours). Magenta plus signs mark the\hspace{0.1cm}spatial peaks of the\hspace{0.1cm}HXR and radio sources. Dashed ellipses display the\hspace{0.1cm}sizes of the\hspace{0.1cm}synthesized beams.}
\end{figure}
X-ray image reconstruction \citep{hurford2002} performed
with Pixon algorithm \citep{PinaPuetter1993} shows that the\hspace{0.1cm}flare
has two well defined footpoints (Figure\hspace{0.15cm}\ref{fig:Ximage}),
which are well visible over the\hspace{0.1cm}entire range of the\hspace{0.1cm}X-ray
spectrum $6-80$\hspace{0.1cm}keV. The\hspace{0.1cm}imaging below $10$\hspace{0.1cm}keV does not demonstrate
any thermal component in a\hspace{0.1cm}separate location as is often seen
at the\hspace{0.1cm}top of a\hspace{0.1cm}loop in flares \citep{Kosugi1992yohkoh,aschwanden2002,emslie2003,Kontar_etal08},
so all the\hspace{0.1cm}detectable X-ray emission down to the\hspace{0.1cm}lowest energy $\sim6$\hspace{0.1cm}keV comes from the\hspace{0.1cm}footpoints. The\hspace{0.1cm}flare occurred
at the\hspace{0.1cm}extreme eastern edge of the\hspace{0.1cm}active region (Figure\hspace{0.15cm}\ref{fig:Ximage}) with the\hspace{0.1cm}weaker
X-ray source projected onto the\hspace{0.1cm}photosphere in a\hspace{0.1cm}region of strong positive magnetic field,
while the\hspace{0.1cm}stronger X-ray source projects onto a\hspace{0.1cm}small region of weaker
negative magnetic field, as has commonly been observed from asymmetric
flaring loops.
OVSA radio imaging for this flare is limited because only four (of six)
antennas (two big and two small antennas) recorded the\hspace{0.1cm}radio emission
at the\hspace{0.1cm}time of flare. To improve the\hspace{0.1cm}image quality a\hspace{0.1cm}frequency
synthesis in two separate bands,
$2.6-3.2$\hspace{0.1cm}GHz and $4.2-8.2$\hspace{0.1cm}GHz, with the\hspace{0.1cm}synthesized beam sizes of $19''\times81''$ and $10''\times58''$, respectively, was used. This suggests that the\hspace{0.1cm}both low- and high-frequency radio sources are unresolved.
The\hspace{0.1cm}corresponding radio images (Figure\hspace{0.15cm}\ref{fig:Ximage})
are located between the\hspace{0.1cm}X-ray footpoints although with an offset
from their connecting line, which is consistent with the\hspace{0.1cm}radio sources placement
in a\hspace{0.1cm}coronal part of a\hspace{0.1cm}magnetic loop connecting the\hspace{0.1cm}X-ray footpoints.
The\hspace{0.1cm}higher-frequency radio source is displaced compared with the\hspace{0.1cm}lower-frequency
one towards the\hspace{0.1cm}stronger magnetic field (weaker X-ray) footpoint.
No spatial displacement with time is detected for either of the\hspace{0.1cm}radio sources.
Based on the\hspace{0.1cm}source separation, implied magnetic topology, and the\hspace{0.1cm}northern HXR source size, in
what follows we adopt the\hspace{0.1cm}area of $10''$(transverse the\hspace{0.1cm}loop)$\times15''$(along the\hspace{0.1cm}loop),
and the\hspace{0.1cm}depth of $10''$ for the\hspace{0.1cm}lower-frequency radio source, which suggests the\hspace{0.1cm}radio source volume
of $V_{radio}\sim6\times10^{26}$\hspace{0.1cm}cm$^3$, and roughly half of that for transverse sizes
of the\hspace{0.1cm}higher-frequency source.
\section{Data Analysis}
Generally, GS continuum radio emission can be produced by any of (i)\hspace{0.1cm}a
magnetically trapped component \citep[trap-plus-precipitation model,][]{Melrose_Brown_1976},
or (ii)\hspace{0.1cm}a precipitating component, or (iii)\hspace{0.1cm}the\hspace{0.1cm}primary component within the\hspace{0.1cm}acceleration region.
These three populations of fast electrons produce radio emission with
distinctly different characteristics. Indeed, (i)\hspace{0.1cm}in the\hspace{0.1cm}case of magnetic trapping
the\hspace{0.1cm}electrons are accumulated at the\hspace{0.1cm}looptop \citep{melnikov_etal_2002}, and the\hspace{0.1cm}radio light curves are delayed
by roughly the\hspace{0.1cm}trapping time $\tau_{trap}$ relative to accelerator/X-ray light curves.
(ii)\hspace{0.1cm}In the\hspace{0.1cm}case of free electron propagation,
untrapped precipitating electrons are more evenly
distributed in a\hspace{0.1cm}tenuous loop, and no delay longer than $L/v$ is expected.
However, even with a\hspace{0.1cm}roughly uniform electron distribution, most of the\hspace{0.1cm}radio emission
comes from loop regions with the\hspace{0.1cm}strongest magnetic field.
Spectral indices of the\hspace{0.1cm}radio- and X-ray- producing fast electrons
differ by 1/2 from each other, because $L/v{\propto}E^{-1/2}$.
(iii)\hspace{0.1cm}In the\hspace{0.1cm}case of radio emission from the\hspace{0.1cm}acceleration region,
even though the\hspace{0.1cm}residence time ($\tau_l$)\hspace{0.1cm}that fast electrons spend in the\hspace{0.1cm}acceleration region can be relatively long,
the\hspace{0.1cm}radio and X-ray light curves are proportional
to each other simply because the\hspace{0.1cm}flux of the\hspace{0.1cm}X-ray producing electrons is equivalent to the\hspace{0.1cm}electron loss rate
from the\hspace{0.1cm}acceleration region, $F_e(t)=N_r(t)/\tau_l$.
\begin{figure}\centering
\includegraphics[width=0.9\columnwidth]{rhessi-zurich.eps}\\
\caption{\label{fig_r_z_corr} RHESSI 30-100\hspace{0.1cm}keV (thick line) and Phoenix-2 3.2 to 3.6\hspace{0.1cm}GHz (thin line) lightcurves of July 30, 2002 flare with 2\hspace{0.1cm}s time resolution both. The\hspace{0.1cm} light curves are highly correlated; no significant delay is present: the\hspace{0.1cm}lag correlation plot is given in the\hspace{0.1cm}insert; negative delay means the\hspace{0.1cm}radio emission comes first.}
\end{figure}
The\hspace{0.1cm}analysis of the\hspace{0.1cm}radio data requires, therefore, (in addition to the\hspace{0.1cm}electron injection
rate and spectrum derived from {\it RHESSI}) some information of the\hspace{0.1cm}fast electron residence
time at the\hspace{0.1cm}radio source. To address the\hspace{0.1cm}timing, we use Phoenix-2 (rather than the\hspace{0.1cm}OVSA)
data because of its higher time resolution. We select the\hspace{0.1cm}frequency range of 3.2 to 3.6\hspace{0.1cm}GHz
corresponding to the\hspace{0.1cm}optically thin part of the\hspace{0.1cm}radio spectrum and almost free of
fine structures and interference, see Figure\hspace{0.15cm}\ref{fig_30_jul_2002_over}.
The\hspace{0.1cm}cross-correlation (Figure\hspace{0.15cm}\ref{fig_r_z_corr}) displays clearly that the\hspace{0.1cm}radio
and HXR light curves are very similar to each other and there is no measurable
delay in the\hspace{0.1cm}radio component. In fact, the\hspace{0.1cm}cross-correlation is consistent with the\hspace{0.1cm}radio emission peaking $\sim130$\hspace{0.1cm}ms earlier. The\hspace{0.1cm}lack of noticeable delay between the\hspace{0.1cm}radio
and X-ray light curves is further confirmed by considering the\hspace{0.1cm}OVSA light
curves (4\hspace{0.1cm}s time resolution) at different optically thin frequencies.
Therefore, the\hspace{0.1cm}magnetically trapped electron component appears to be absent,
and the\hspace{0.1cm}radio emission is formed by either (ii)\hspace{0.1cm}precipitating electrons or
(iii)\hspace{0.1cm}electrons in the\hspace{0.1cm}acceleration region or both.
With the\hspace{0.1cm}spectrum of energetic electrons from HXR data, it is easy to estimate
the\hspace{0.1cm}radio emission produced by the\hspace{0.1cm}precipitating electron component.
Taking the\hspace{0.1cm}electron flux, the\hspace{0.1cm} spectral index of the\hspace{0.1cm}radio-producing
electrons $\delta_r{\approx}\delta_x+1/2$, and the\hspace{0.1cm}electron lifetime at the
loop $L/v$ (the\hspace{0.1cm}time of flight), we can vary the\hspace{0.1cm}magnetic field at the\hspace{0.1cm}source
in an attempt to match the\hspace{0.1cm}spectrum shape and flux level. However, if we match
the\hspace{0.1cm}spectrum peak position, we strongly underestimate the\hspace{0.1cm}radio flux, while if we match
the\hspace{0.1cm}flux level at the\hspace{0.1cm}peak frequency or at an optically thin frequency, we
overestimate the\hspace{0.1cm}spectrum peak frequency; examples of such spectra are given
in Figure\hspace{0.15cm}\ref{fig:fit_tt} by the\hspace{0.1cm}dotted curves. We conclude that precipitating
electrons only [option (ii)] cannot make a\hspace{0.1cm}dominant contribution to the\hspace{0.1cm}observed
radio spectrum.
\begin{figure}\centering
\includegraphics[width=0.9\columnwidth]{Spec_fit_radio_6keV_2components.eps}\\
\caption{\label{fig:fit_tt}
OVSA radio spectra obtained by two small antennas (pluses and asterisks; differences between them offer an idea about the\hspace{0.1cm}data scatter vs frequency) and model GS emission from the\hspace{0.1cm}acceleration region (dashed lines), precipitating electrons (dotted lines), and sum of these components (solid line). Total number and number density of the\hspace{0.1cm}fast electrons at the\hspace{0.1cm}radio source as derived from the\hspace{0.1cm}OVSA radio spectrum.}
\end{figure}
To quantify the\hspace{0.1cm}third option, we have to consider further constraints
to estimate the\hspace{0.1cm}electron residence time in the\hspace{0.1cm}main radio source. On one hand, this residence
time must be shorter than the\hspace{0.1cm}radio light curve decay time, $\sim10$\hspace{0.1cm}s; otherwise, the\hspace{0.1cm}decay
of the\hspace{0.1cm}radio emission would be longer than observed. On the\hspace{0.1cm}other hand, the\hspace{0.1cm}extremely
low frequency of the\hspace{0.1cm}microwave spectrum peak implies the\hspace{0.1cm}magnetic field is well
below $100$\hspace{0.1cm}G at the\hspace{0.1cm}radio source, i.e., much smaller than the\hspace{0.1cm}footpoint magnetic field values
($\sim-130$\hspace{0.1cm}G and $\sim+1000$\hspace{0.1cm}G). This implies that the\hspace{0.1cm}residence time in the\hspace{0.1cm}main radio source
must be noticeably larger than the\hspace{0.1cm}time of flight, which is a\hspace{0.1cm}fraction
of second: otherwise, the\hspace{0.1cm}fast electron density would be evenly
distributed over the\hspace{0.1cm}loop and the\hspace{0.1cm}gyrosynchrotron (GS) microwave emission would
be dominantly produced at the\hspace{0.1cm}large field regions, resulting in a\hspace{0.1cm}spectrum
with much higher peak frequency than the\hspace{0.1cm}observed one (as in the\hspace{0.1cm}already considered
case of the\hspace{0.1cm}precipitating electron population). Thus, a\hspace{0.1cm}reasonable estimate
of this lifetime is somewhere between those two extremes, $\tau_{l}\sim 3$\hspace{0.1cm}s.
The\hspace{0.1cm}quality of the\hspace{0.1cm}OVSA data available for this event appears insufficient
to perform a\hspace{0.1cm}complete forward fit with all model parameters being free \citep{Fl_etal_2009},
which would require better calibrated imaging spectroscopy data.
Instead, we have to fix as many parameters as possible \citep{Bastian_etal_2007, Altyntsev_etal_2008}
and estimate one or two remaining parameters from the\hspace{0.1cm}fit. To do so we adopt the\hspace{0.1cm}maximum total
electron number $N_{r\max}(>6keV)=\tau_{l}F_{e\max}(>6keV)$ with a value of $F_{e\max}(>6keV)=2\times10^{35}$ electrons/s and $\tau_{l}=3$\hspace{0.1cm}s, and the\hspace{0.1cm}time evolution
of $N_{r}(>6keV)$ to be proportional to an optically thin gyrosynchrotron light curve.
The\hspace{0.1cm}total lifetime of the\hspace{0.1cm}electrons in the\hspace{0.1cm}flaring loop is a\hspace{0.1cm}sum of the\hspace{0.1cm}residence time
at the\hspace{0.1cm}radio source and the\hspace{0.1cm}time of flight between the\hspace{0.1cm}radio source and the\hspace{0.1cm}footpoints.
As we adopted a\hspace{0.1cm}constant, energy-independent, electron lifetime $\tau_{l}$ to be much larger
than the\hspace{0.1cm}time of flight ${\sim}L/v$
we have to accept for consistency that the\hspace{0.1cm}spectral index of the\hspace{0.1cm}radio emitting fast electrons
is roughly the\hspace{0.1cm}same as the\hspace{0.1cm}spectral index of HXR emitting fast electrons determined
above, $\delta_r=3.5$. The\hspace{0.1cm}radio source sizes are taken as estimated in \S~\ref{S_XR_observ}.
The\hspace{0.1cm}thermal electron number density is adopted to be $n_{th}=1.5\times10^9$\hspace{0.1cm}cm$^{-3}$
(see the\hspace{0.1cm}next section); the\hspace{0.1cm}GS spectra
are not sensitive to this parameter until $n_{th}\lesssim2\times10^9$\hspace{0.1cm}cm$^{-3}$.
The\hspace{0.1cm}remaining radio source
parameter, not constrained by other observations, is the\hspace{0.1cm}magnetic field $B$, which is determined by
comparing the\hspace{0.1cm}observed (symbols) microwave and calculated (dashed curves)
GS spectra \citep{Fl_Kuzn_2010} in Figure\hspace{0.15cm}\ref{fig:fit_tt}. Remarkably, that the\hspace{0.1cm}whole time sequence of the\hspace{0.1cm}radio spectra
is reasonably fitted with a\hspace{0.1cm}single magnetic field strength of $B\approx60$\hspace{0.1cm}G; the\hspace{0.1cm}only source
parameter changing with time is the\hspace{0.1cm}instantaneous number of the\hspace{0.1cm}fast electrons, see Figure\hspace{0.15cm}\ref{fig:fit_tt}.
The\hspace{0.1cm}OVSA spectra, however, deviate from the\hspace{0.1cm}model dashed curves by the\hspace{0.1cm}presence of a\hspace{0.1cm}higher-frequency
bump at $f\sim4-8$\hspace{0.1cm}GHz. Nevertheless, adding the\hspace{0.1cm}contribution produced by precipitating
electrons (dotted curves) at a\hspace{0.1cm}larger magnetic field strength $B_{leg}$ peaking somewhere at the\hspace{0.1cm}western leg (60G$<B_{leg}<$1000G) of the\hspace{0.1cm} loop (perhaps, around the\hspace{0.1cm}mirror point) as follows from the\hspace{0.1cm}OVSA imaging, Figure\hspace{0.15cm}\ref{fig:Ximage}, offers a\hspace{0.1cm}nice, consistent overall fit (solid curves) to the\hspace{0.1cm}spectra.
We conclude that the\hspace{0.1cm}radio spectrum is dominated by the\hspace{0.1cm}GS emission
from the\hspace{0.1cm}electron acceleration region with a\hspace{0.1cm}distinct weaker contribution
from the\hspace{0.1cm}precipitating electrons.
\section{Discussion}
In order to complete a\hspace{0.1cm}model for this cold flare event we have to estimate the\hspace{0.1cm}flaring
loop geometry. As a\hspace{0.1cm}zero-order approximation we utilize a\hspace{0.1cm}potential field extrapolation \citep{Rudenko_2001}
based on the\hspace{0.1cm}SoHO/MDI line-of-sight magnetogram. Figure\hspace{0.15cm}\ref{fig:Ximage} shows
that there is a\hspace{0.1cm}flux tube connecting the\hspace{0.1cm}two X-ray footpoints, which confirms the
existence of the\hspace{0.1cm}required magnetic connectivity. This magnetic loop is highly asymmetric
with the\hspace{0.1cm}magnetic field reaching its minimum value (around 130\hspace{0.1cm}G) at the\hspace{0.1cm}northern
footpoint (stronger X-ray source). The\hspace{0.1cm}length of the\hspace{0.1cm}central field line is
about $55''\approx4\times10^9$\hspace{0.1cm}cm. We know, however, that the\hspace{0.1cm}flare
phenomenon requires a\hspace{0.1cm}non-potential magnetic loop. Furthermore, the\hspace{0.1cm}magnetic field at the\hspace{0.1cm}radio
source (which is likely to belong to the\hspace{0.1cm}same magnetic structure because of excellent
light curve correlation) is below 100\hspace{0.1cm}G, which is likely to be located higher
than the\hspace{0.1cm}potential loop in Figure\hspace{0.15cm}\ref{fig:Ximage}. We, therefore, propose that
the\hspace{0.1cm}true flaring loop is higher and the\hspace{0.1cm}length of the\hspace{0.1cm}central field line
is somewhat longer; for the\hspace{0.1cm}estimate we adopt $L\sim100''\sim7\times10^9$\hspace{0.1cm}cm that
yields the\hspace{0.1cm}loop volume $V_{loop}\sim4\times10^{27}$\hspace{0.1cm}cm$^3$ roughly 5 times larger
than the\hspace{0.1cm}radio source volume.
The\hspace{0.1cm}next needed step is an estimate of the\hspace{0.1cm}thermal density in the\hspace{0.1cm}flaring loop.
From the\hspace{0.1cm}radio spectra and from the\hspace{0.1cm}absence of any coronal X-ray source we already know that this density is low. In fact, both radio and X-ray data can
be fit without any thermal plasma at all.
Let us consider the\hspace{0.1cm}10keV electron Coulomb losses
to quantify the\hspace{0.1cm}thermal density. The\hspace{0.1cm}collisional lifetime of the\hspace{0.1cm}fast
electrons is $t_E\approx20E_{100}^{3/2}n_{10}^{-1}$\hspace{0.1cm}s \citep{Bastian_etal_2007}, which
yields $t_{10keV}\lesssim3$\hspace{0.1cm}s for the\hspace{0.1cm}background plasma density $n_{th}\gtrsim 2\times10^{9}$\hspace{0.1cm}cm$^{-3}$,
which would imply the\hspace{0.1cm}presence of a\hspace{0.1cm}coronal 10keV X-ray component in contradiction
with observations. We conclude that $2\times10^{9}$\hspace{0.1cm}cm$^{-3}$ is an upper bound
for the\hspace{0.1cm}thermal electron density. To estimate the\hspace{0.1cm}lower bound of the\hspace{0.1cm}density, we consider
the\hspace{0.1cm}efficiency of the\hspace{0.1cm}electron accelerator. From the\hspace{0.1cm}X-ray fit we know that the\hspace{0.1cm}peak acceleration
efficiency is at least $F_{e\max}(>6keV)\sim 2{\times}10^{35}$ electrons per second.
The\hspace{0.1cm}duration of the\hspace{0.1cm}flare at the\hspace{0.1cm}half of the\hspace{0.1cm}peak level is about $t_{1/2}\approx25$\hspace{0.1cm}s,
thus the\hspace{0.1cm}total number of accelerated electrons is $N_{tot}{\sim}t_{1/2}F_{e\max}\sim\hspace{0.1cm}5\times10^{36}$.
These electrons must apparently be taken from the\hspace{0.1cm}thermal electrons available in the\hspace{0.1cm}flaring
loop prior to the\hspace{0.1cm}flare, thus, the\hspace{0.1cm}ratio of $N_{tot}/V_{loop}{\sim}10^9$\hspace{0.1cm}cm$^{-3}$
represents a\hspace{0.1cm}lower bound of the\hspace{0.1cm}thermal electron density. These estimates justify
the\hspace{0.1cm}$n_{th}$ value adopted in the\hspace{0.1cm}previous section.
Let us proceed now to the\hspace{0.1cm}energy release and plasma heating efficiency. The\hspace{0.1cm}energy release
rate $dW/dt$ is estimated as the\hspace{0.1cm}product of the\hspace{0.1cm}minimum energy (6keV) and the\hspace{0.1cm}acceleration
rate $F_{e}(>6keV)$, which yields $\sim 2\times10^{27}$\hspace{0.1cm}ergs/s at the\hspace{0.1cm}flare peak time. Being evenly
distributed over the\hspace{0.1cm}loop volume this corresponds to the\hspace{0.1cm}averaged density of the\hspace{0.1cm}energy release of $\sim0.5$\hspace{0.1cm}erg/cm$^{-3}$/s and, being multiplied by the\hspace{0.1cm}effective duration $t_{1/2}$, the\hspace{0.1cm}energy
density deposition of $w\sim12$\hspace{0.1cm}erg/cm$^{-3}$. Most of this energy is produced in the\hspace{0.1cm}form
of accelerated electrons around 10\hspace{0.1cm}keV. During the\hspace{0.1cm}time of flight in the\hspace{0.1cm}loop
(with density $1.5\times10^{9}$\hspace{0.1cm}cm$^{-3}$ and half length $\sim3\times10^9$\hspace{0.1cm}cm) these electrons
lose about $\Delta\simeq15\%$ of their initial energy. Thus, we can estimate the\hspace{0.1cm}plasma heating
by the\hspace{0.1cm}accelerated electrons up to $T{\simeq}w/(k_Bn_{th})\times\Delta\sim7$\hspace{0.1cm}MK,
where $k_B$ is the\hspace{0.1cm}Boltzman constant. Combined with a\hspace{0.1cm}relatively low emission measure
of the\hspace{0.1cm}tenuous loop, this heating is undetectable by GOES and RHESSI, even though the\hspace{0.1cm}acceleration
efficiency is extremely high.
Many acceleration mechanisms require the\hspace{0.1cm}particles to be confined
in the\hspace{0.1cm}acceleration region.
In particular, this is the\hspace{0.1cm}case for acceleration in collapsing
magnetic traps \citep[e.g.][]{1975ApJ...200..734B,1997ApJ...485..859S,2004A&A...419.1159K}
and for stochastic acceleration \citep[e.g.][]{1993ApJ...418..912L,1996ApJ...461..445M,Byk_Fl_2009,2010A&A...519A.114B}, but not for a\hspace{0.1cm}DC field acceleration.
This flare does not display
any characteristics expected for the\hspace{0.1cm}collapsing trap scenario
\citep[e.g., a spatial displacement of the\hspace{0.1cm}source, the\hspace{0.1cm}magnetic field growth, the\hspace{0.1cm}radio peak frequency increase, or a time delay between radio light curves at different frequencies;][]{Li_Fl_2009},
while it is consistent with the\hspace{0.1cm}stochastic acceleration \citep{Park_Fl_2010} in a\hspace{0.1cm}magnetic loop, when a standard, relatively narrowband, GS emission is produced at a given volume (permitted with a loop magnetic field) by the\hspace{0.1cm}electrons accelerated there by a turbulence, whose side effect is to enhance the\hspace{0.1cm}electron trapping and so increase, as observed, their residence time at the\hspace{0.1cm} acceleration region.
We conclude that detection of GS radio emission from a\hspace{0.1cm}region of stochastic
acceleration of electrons is likely in the\hspace{0.1cm}event
under study. Various stochastic acceleration scenarios differ from each other in some predictions,
and so in principle are distinguishable by observations. One of them is the\hspace{0.1cm}energy dependence of the\hspace{0.1cm}fast
electron lifetime at the\hspace{0.1cm}acceleration region. Our data are consistent with $\tau_l(E)=const$,
which is the\hspace{0.1cm}case, in particular, for acceleration by strong turbulence \citep{Byk_Fl_2009}.
However, the\hspace{0.1cm}precision of this constancy is insufficient in our observations to exclude competing
versions of the\hspace{0.1cm}stochastic acceleration.
What is special about this flare, which allowed detecting the\hspace{0.1cm}radio emission
from the\hspace{0.1cm}acceleration region, compared with other flares where the\hspace{0.1cm}radio emission is
often dominated by trapped population of fast electrons?
Possibly, an important aspect is the\hspace{0.1cm}exceptional asymmetry of the\hspace{0.1cm}flare loop; recall that the\hspace{0.1cm}corresponding
potential loop displays the\hspace{0.1cm}minimum strength of the\hspace{0.1cm}magnetic field absolute value at the\hspace{0.1cm}northern
photospheric footpoint (rather than coronal level), making the\hspace{0.1cm}magnetic trapping of fast electrons
totally inefficient in this case.
A\hspace{0.1cm}further important point is the\hspace{0.1cm}somewhat low initial plasma density of the\hspace{0.1cm}loop. This low density
maintains relatively low coronal Coulomb losses of the\hspace{0.1cm}accelerated electrons, which allows
them to freely stream towards the\hspace{0.1cm}chromosphere to produce the\hspace{0.1cm}X-ray emission. In addition,
the\hspace{0.1cm}low density cannot supply the\hspace{0.1cm}accelerator with seed electrons longer than half a\hspace{0.1cm}minute
or so, thus, the\hspace{0.1cm}accelerator becomes exhausted very quickly, before the\hspace{0.1cm}energy needed
for substantial (measurable) plasma heating has been released. This explains why the\hspace{0.1cm}main
plasma remains relatively cold although the\hspace{0.1cm}acceleration efficiency is exceptionally
high (almost all available electrons are being eventually accelerated).
\section{Conclusions}
We have reported the\hspace{0.1cm}observations and analysis of a\hspace{0.1cm}cold, tenuous flare, which displays prominent and numerous non-thermal signatures,
while not showing any evidence for thermal plasma heating or chromospheric evaporation.
A\hspace{0.1cm}highly asymmetric flaring loop, combined with rather low thermal electron density, made
it possible to detect the\hspace{0.1cm}GS radio emission directly from the\hspace{0.1cm}acceleration site.
We found that the\hspace{0.1cm}electron acceleration efficiency is very high in the\hspace{0.1cm}flare,
so almost all available thermal electrons are eventually accelerated. Some sort of stochastic acceleration process is needed to account for an approximately energy-independent lifetime of about 3\hspace{0.1cm}s for the\hspace{0.1cm}electrons in the\hspace{0.1cm}acceleration region.
We emphasize that the\hspace{0.1cm}numbers derived for the\hspace{0.1cm}quantity of accelerated electrons
offer a\hspace{0.1cm}lower bound for this measure because for our estimates we adopted a\hspace{0.1cm}lower bound of the\hspace{0.1cm}electron
flux ($F_{e\max}(>6keV)\simeq2\times10^{35}$\hspace{0.1cm}electrons
per second) and lowest electron spectral index ($\delta_r=3.5$) compatible with the\hspace{0.1cm}HXR fit;
taking mean values would result in even more powerful acceleration. However, given a\hspace{0.1cm}relatively
small flaring volume and rather low thermal density at the\hspace{0.1cm}flaring loop, the\hspace{0.1cm}total energy
release turned out to be insufficient for a\hspace{0.1cm}significant heating of the\hspace{0.1cm}coronal
plasma or for a\hspace{0.1cm}prominent chromospheric response giving rise to chromospheric
evaporation.
\acknowledgments
This work was supported in part by NSF grants
AGS-0961867, AST-0908344, and NASA grants NNX10AF27G and NNX11AB49G to New Jersey
Institute of Technology,
and by the\hspace{0.1cm}RFBR grants 09-02-00226 and 09-02-00624.
This work was supported by a\hspace{0.1cm}UK STFC rolling grant, STFC/PPARC Advanced
Fellowship, and the\hspace{0.1cm}Leverhulme Trust, UK. Financial support by the
European Commission through the\hspace{0.1cm}SOLAIRE and HESPE Networks is
gratefully acknowledged. The\hspace{0.1cm}authors are greatly thankful to George Rudenko and Vasil Yurchishin for valuable discussions of the\hspace{0.1cm}coronal magnetic extrapolations.
|
\section{Introduction}
In the recent Hollywood movie \emph{The Dark Knight} (2008) the comic
character known as the Joker jeopardizes a whole society spreading
chaos and destruction with no aim of benefit at it. The situation is
so critical that even the mob is willing to cooperate with honest
people to stop this nonsensical catastrophe. This fiction provides
a visual metaphor of how an event like this can force exploiters of
society to collaborate temporarily to fight the common enemy. Society
is an emergent structure resulting from the cooperation among its
members, and exploiters need society to survive, even if they do not
contribute to it. Thus they are specially sensitive to the destruction
of society precisely because, being selfish agents, society is their
only source of survival. The appearance of
the Joker provides a strong incentive for cooperation.
Beside situations like the one depicted by the Joker metaphor,
the importance of the inclusion of malicious agents on
the game is also illustrated in other scenarios.
Here are a few examples. Temporary coalitions
of rival parties are constantly formed whenever a common enemy arises,
only to restore their old rivalry once this enemy has been
wiped out. During the Second World War U.S.A.\ and U.S.S.R.\ were
allied in fighting Hitler, but they got engaged in the Cold War for
decades after the danger of Nazism had been ruled out. It is also well
known that strong affective links between humans are created when they
face a common difficult situation. Biology is another
source of potential examples. For instance, it has been shown that the
perception of an increase in the risk of predation can induce cooperative
behavior in some bird species \citep{krams:2010}.
Indeed, prey species frequently form groups to
increase the survival rate against predator attacks
\citep{hamilton:1971,krebs:1993}.
In some cases, this has been proven to happen even in the absence of
kinship among its members, as in the collective defense of spiny
lobsters \citep{lavalli:2009}.
The existence of these temporary coalitions for defense against a common
danger in rational and irrational agents alike calls for an evolutionary
explanation. In this article we propose a stylized evolutionary game
\citep{hofbauer:1998} aimed at studying theoretically this enhancement
of cooperation driven by the emergence of purely destructive agents.
The game does not try to model any specific situation, but it proposes
an abstract setting in which the role of the indiscriminate destructive
action of these agents in enhancing cooperation is made clear.
Our model is a modification
of the standard Public Goods (PG) game \citep{groves:1977}, the
$n$-players version of Prisoner's Dilemma and a paradigm of the risk
of exploitation faced by cooperative behavior \citep{hardin:1968}.
It has been shown that several mechanisms involving reputation \citep{milinski:2005},
allowing for volunteer participation \citep{hauert:2002b,hauert:2002c}, punishing
defectors \citep{fehr:1999,fehr:2000}, rewarding
cooperators \citep{sigmund:2001} or structuring agents
\citep{szabo:2002,wakano:2009,hauert:2008}, can enhance cooperation.
Here, we present a different mechanism
for the enhancement of cooperation based on the existence of evil agents.
The game involves $n$
players who belong to one out of three different types: cooperators,
who contribute to the public good at a cost for themselves; defectors,
who free-ride the public good at no expense; and jokers, who do not
participate in the public good ---hence obtain no benefit whatsoever---
and only inflict damage to the public good. Groups are formed randomly,
and each player's strategy is established before the group is selected.
Hence, players have no memory. Remarkably, the appearance
of jokers promotes a rock-paper-scissors dynamics, where jokers outbeat
defectors and cooperators outperform jokers, which are subsequently
invaded by defectors. In contrast to previous models \citep{hauert:2002b,hauert:2002c},
the cycles induced by jokers are limit cycles, i.e. attractors of the dynamics,
and exist in the presence of mutations; these properties make them
robust evolutionary outcomes.
Therefore, paradoxically, the existence of destructive
agents acting indiscriminately promotes cooperation.
The paper is organized as follows. Section 2 exposes the model and shows the
existence of cycles. Section 3 analyzes the dynamics for infinite populations,
and section 4 compares the joker model with other RPS dynamics.
\section{A Public Good game with jokers: existence of limit cycles}
The PG game works as usual: every cooperator yields a benefit $b=rc$
($r>1$) to be shared by cooperators and defectors alike,
at a cost $c$ for herself (this cost can be set to $c=1$
without loss of generality: all other payoffs are given in units of
$c$), and defectors produce no benefit at all but get their share of the
public good. As for the new agents (jokers), every joker inflicts a
damage $-d < 0$ to be shared equally by all non-jokers
and gets no benefit. In a given game $0\le m \le n$ denotes the number of
cooperators, $0\le j\le n$ the number of jokers, and $n-m-j\ge 0$
the number of defectors; $S=n-j$ expresses the
number of non-jokers. In this group, the payoff of a defector will be
$\Pi_{\rm D}(m,j)= (rm-dj)/S$, and that of a cooperator
$\Pi_{\rm C} =\Pi_{\rm D}-1$. Then, in each group, defectors will always
do better than cooperators. Jokers' payoff is always $0$.
A usual requirement of PG games is that $r<n$. Without this requirement
the solution in which all $n$ players are defectors is no longer a Nash
equilibrium ---hence the dilemma goes away. As shown later, the evolutionary
dynamics for infinite populations yields the same constraint, i.e., if
$r<n$ the dynamics asymptotically approaches the tragedy of the commons.
However this is no longer true for finite populations, where the upper
bound of $r$ for which the tragedy of the commons takes place grows as $M$,
the population size, decreases. In this case the
tragedy of the commons arises whenever $r<r_{\rm max}=n(M-1)/(M-n)$
(see~\ref{app:A}; notice in passing that for a population of $M=n$ individuals,
the evolutionary dynamics yields a tragedy of the commons for every $r>1$).
An invasion analysis provides the clue as to why a rock-paper-scissors
(RPS) cycle is to be expected when jokers intervene in the game. We
shall assume that we have a population of $M$ players of the same type
and will consider putative mutations of one individual to any of the
other two types. The mutation will thrive if the average payoff of the
mutant after many interactions overcomes the average payoff of a
non-mutant player. The result of this analysis (see~\ref{app:A}) is summarized
in Fig.~\ref{fig:diagram}, which represents the three different patterns
of invasion that can be observed within the region of interest
$1<r<r_{\rm max}$, $d>0$:
\begin{enumerate}[I.]
\item \textbf{Rock-paper-scissors cycle:} It arises whenever
$r>1+d(n-1)$. This condition expresses the fact that a single cooperator gets
a positive payoff in spite of the damage inflicted by $n-1$ jokers and
therefore being a cooperator pays (jokers get no payoff whatsoever).
\item \textbf{Joker-cooperator bistability:} If $1+d/(M-1)<r<1+(n-1)d$
neither jokers nor cooperators can invade each other. Nonetheless
defectors always invade cooperators, and jokers always invade defectors,
so eventually only jokers survive, either because they are initially
a majority or indirectly through the emergence of defectors.
\item \textbf{Joker invasion:} If $r<1+d/(M-1)$ jokers will invade any
homogeneous population, so a homogeneous population of jokers is
the only stable solution. Notice that this region disappears for
large populations ($M\to\infty$) because $r>1$.
\end{enumerate}
The RPS cycle C$\to$D$\to$J$\to$C
occurring in region I is the essence of the Joker effect.
\begin{figure}
\begin{center}
\includegraphics[width=88mm]{sketch.eps}
\end{center}
\caption{{\bf Dynamics of invasions in a Public Goods game with jokers.}
The axes represent the gain factor $r$ of the Public Goods game (i.e.,
the payoff each cooperator yields to the public good) and the ``damage''
$d>0$ that every joker inflicts on the public good. The tragedy of the commons
occurs for $1<r<r_{max}=n(M-1)/(M-n)$ (see text), which includes the dilemmatic region
$1<r<n$ characteristic of PG games. Different colors are assigned to different
invasion patterns: Light blue corresponds to a region
where J invades both C and D (III); light green corresponds to a region where
neither C nor J invades each other (there is bistability on the J--C line)
but D invades C and is in turn invaded by J, so again everything ends up
in J (II); finally, light yellow corresponds to a region where D invades C,
J invades D, but C invades J back, thus generating a rock-paper-scissors
cycle (I). The latter behavior is the essence of the \emph{Joker effect.}
The equations of the straight lines separating the three regions
are (from top to bottom) $r=1+d(n-1)$ and $r=1+d/(M-1)$. Notice that this
scheme is valid for arbitrary $n>1$. Also, for fixed $r$, all three
regions are crossed upon varying $d$, whereas vice versa is only true
provided $d<d_1=M/(M-n)$. The Joker effect does not occur if $d>d_1$.
For large populations, $M\gg 1$, the region for the rock-paper-scissors
cycle simplifies to $n>r>1+(n-1)d$ and $d<1$.
\label{fig:diagram}}
\end{figure}
\section{Infinite populations}
We can gain further insight into this effect by studying a replicator-mutator
dynamics \citep{maynard-smith:1982}. We assume a very large population in which
the three types are present at time $t$ in fractions $x$ (cooperators),
$y$ (defectors), and $z=1-x-z$ (jokers). Agents interact
with the whole population by engaging in the above described game within
groups of $n$ randomly chosen individuals \citep{hauert:2006}. Average
payoffs of a cooperator, a defector, and a joker are denoted $P_{\rm C}(x,z)$,
$P_{\rm D}(x,z)$, and $P_{\rm J}(x,z)$, respectively. Assuming individuals
of a given type mutate to any other type at a rate $\mu\ll 1$,
the replicator-mutator equations for this system will be
\begin{equation}
\begin{split}
\dot x &= x (P_{\rm C} - \bar{P}) + \mu (1-3x) ,\\
\dot y &= y (P_{\rm D} - \bar{P}) + \mu (1-3y), \\
\dot z &= z (P_{\rm J} - \bar{P}) + \mu (1-3z),
\end{split}
\label{eq:RM}
\end{equation}
where $\bar{P}=xP_{\rm C}+ yP_{\rm D}+zP_{\rm J}$ is the mean payoff
of the population at a given time. Explicit expressions for $P_{\rm C}$,
$P_{\rm D}$, and $P_{\rm J}$ can be obtained by averaging over all
samples of groups of $n$ players extracted from a population containing
$Mx$ cooperators, $My$ defectors, and $Mz$ jokers, in the limit of very
large populations ($M\to\infty$); the derivation can be found in~\ref{app:B}.
Let us recall that the parameters of the game in the infinite population limit satisfy $1 < r < n$ and $d > 0$;
the first condition enforces the public goods dilemma,
and the second one implies that jokers beat defectors in the absence of cooperators, because
defectors receive the damage inflicted by jokers thus obtaining a negative payoff.
The stability analysis of the dynamical system (\ref{eq:RM}) recovers the
picture displayed in Fig.~\ref{fig:diagram} (taking $M\to\infty)$. When
$r<1+(n-1)d$ the system is in region II. The only stable equilibrium is
a population of only jokers and any trajectory of (\ref{eq:RM})
is asymptotically attracted to it. Thus, in this region the destructive
power of jokers is high enough to wipe out the populations of both
cooperators and defectors. But the most interesting situation takes
place when
\begin{equation}
\label{ineq}
r>1+(n-1)d,
\end{equation}
i.e., in region I. In the absence of mutations the dynamical system (\ref{eq:RM})
has three saddle points at the corners of the simplex
as well as an unstable mixed equilibrium (see~\ref{app:C}). As a
consequence, the attractor of the system is the heteroclinic orbit
${\rm C}\to {\rm D}\to {\rm J}\to {\rm C}$. The period is infinite because
the system delays more and more around the saddle points.
When mutations occur the corners of the simplex are no longer equilibria,
and one is left with the interior fixed point, which for small mutations is
a repeller (see~\ref{app:C}). Since trajectories are confined within the closed
region of the simplex, they are attracted to a stable limit cycle for
any $r>1$ (a direct consequence of the Poincar\'e-Bendixon
theorem \citep{simmons:2006}), as shown in Fig.~\ref{fig:simplex}.
\begin{figure}
\begin{center}
\includegraphics[width=88mm]{simplex.eps}
\end{center}
\caption{{\bf The Joker effect in public goods games for large, well-mixed
populations.} The simplex describes the replicator-mutator dynamics,
Eq.~\ref{eq:RM}, for a population of cooperators, defectors and jokers
with parameter values satisfying $n>r>1+d(n-1)$, for which a rock-paper-scissor
dynamics is expected (yellow region in Fig.~\ref{fig:diagram}). When mutation
rates are small, the only equilibrium is a repeller (white dot), and
trajectories end up in a stable limit cycle (black line). Thus the presence
of jokers induces periodically a burst of cooperators. Cooperators abound
during short time spans, as shown by the small fraction of cooperators in
the equilibrium point.
Parameters: $n=5$, $r=3$, $d=0.4$ and $\mu=0.005$.
(Images generated using a modified version of the Dynamo Package
\citep{sandholm:2007}).
\label{fig:simplex}}
\end{figure}
\begin{figure}
\centering
\includegraphics[width=88mm]{FigureSI1.eps}
\caption{{\bf Replicator-mutator dynamics as a function of the damage $d$ inflicted by jokers}.
For a fixed mutation rate, the size of the cycles increases as the damage increases.
Parameters: $n=5$, $r=3$ and $\mu=0.001$.
\label{FigureSI1}}
\end{figure}
\begin{figure}
\centering
\includegraphics[width=88mm]{FigureSI2.eps}
\caption{{\bf Replicator-mutator dynamics as a function of the mutation rate
$\mu$}. (a) For very small mutation rates cycles approach the boundary of the
simplex. (b) As $\mu$ increases, the cycle amplitude decreases and, above a
critical value (typically, $\mu_c \simeq 0.01$), cycles disappear in a Hopf
bifurcation yielding a stable mixed equilibrium (c).
Parameters: $n=5$, $r=3$ and $d=0.4$.
\label{FigureSI2}}
\end{figure}
The size of the cycle depends on the parameter values. It grows as $d$ increases
---i.e., when jokers play a more important role
(Fig.~\ref{FigureSI1})--- and as the mutation rate
decreases (Fig.~\ref{FigureSI2}). For both, large values of $d$ [compatible with condition \eqref{ineq}]
and very small mutations, the cycle closely follows the boundaries of the
simplex (see Fig.~\ref{FigureSI2}a). By increasing the mutation rate (typically over $0.01$),
cycles disappear in a Hopf bifurcation yielding a stable mixed equilibrium (Figs.~\ref{FigureSI2}b-c).
\section{Discussion and conclusions}
This evolution has some resemblances with the effect of volunteering in
a PG game \citep{hauert:2002b,hauert:2002c}, but the two games are fundamentally
different. This can be told from the dynamic behavior of the system.
In both cases, the existence of a third agent which does not participate in
the game is the ultimate reason why cooperators periodically thrive
through a Rock-Paper-Scissor dynamics. However, while the loners game
leads to neutrally stable cycles around a center, trajectories in the
Joker model are attracted by the heteroclinic cycle C--D--J--C. The
difference is even more striking if mutations are included. Mutations
replace the cycles in the loner model by a stable mixed equilibrium.
In contrast, in the Joker model mutations substitute the heteroclinic
orbit by a stable limit cycle, which undergoes a transition (Hopf
bifurcation) to a stable mixed equilibrium above a threshold mutation rate.
These two scenarios can be understood from the analysis of general RPS
games \citep{hofbauer:1998}. There are three situations: (a) orbits are
attracted towards an asymptotically stable mixed equilibrium (the case
of the loners game with mutations), (b) orbits cycle around a neutrally
stable mixed equilibrium (the case of the loners game without mutations),
and (c) orbits go away from an unstable mixed equilibrium and approach
the heteroclinic orbit defined by the border of the simplex (the case
of the Joker game without mutations). If mutations are added to the
latter type of RPS games, limit cycles and a Hopf bifurcation upon
increasing the mutation rate are also found \citep{mobilia:2010}.
Limit cycles are robust to perturbations and have a well defined amplitude
irrespective of the initial fractions of players (as long as it is
not at the border of the simplex). Therefore, they are true attractors of
the dynamics, and can thus be regarded as a robust evolutionary outcome,
in contrast to neutrally stable cycles.
In contrast to loners, which do not participate in the game but receive
a benefit outside of it, jokers do not receive any benefit at all and
cause damage to players. Both loner and joker models coincide ---in
the absence of mutations--- when the damage inflicted by jokers and
the benefit obtained by loners are both zero. In this case both become
simply non-participants in the game, and the only effect they produce
is a reduction in the effective number of players in the game, which
is not enough to induce an oscillatory dynamics (see Fig.~\ref{FigureSI3}).
In other words, the appearance of the RPS cycle which periodically
increases the population of cooperators in the presence of jokers can
only happen, remarkably, provided $d>0$, i.e., if jokers are truly
destructive agents.
\begin{figure}
\centering
\includegraphics[width=88mm]{FigureSI3.eps}
\caption{{\bf Replicator-mutator dynamics for $d=0$}. If jokers are just
passive agents cooperators go extinct. (a) $\mu=0$. The system ends up in a
point of the line DJ with a majority of defectors. (b) $\mu=0.001$. Mutation
generates one single stable state made up mostly by defectors.
Parameters: $n=5$, $r=3$ and $d=0$.
\label{FigureSI3}}
\end{figure}
In this letter we have shed light on a still unexplored aspect of evolutionary
game theory (the presence of a destructive strategy) in the prototypical
PG game. We have shown, both theoretically and by numerical
simulations, that the addition of purely destructive agents (jokers)
to a standard PG game has, paradoxically, a positive effect on cooperation.
Bursts of cooperators are induced through the appearance of a RPS cycle in
which jokers beat defectors, who beat cooperators, who beat jokers in
succession.
The evolutionary dynamics provoked by the Joker, with periods of cooperation,
defection and destruction of the PG, may help understand the appearance of
cognitive abilities that allow individuals to foresee the destructive periods,
promoting in advance the necessary cooperation to avoid them.
We have proven this ``Joker effect'' to occur both in
finite and infinite populations, discarding the possibility of its being
an artificial size-depending phenomenon. Further research is required to ascertain
the scope of the constructive role of destruction in general settings.
This provides a new framework for
the evolution of cooperation that may find important implications in social,
biological, economical, and even philosophical contexts, and that is worth
exploring either with different variants of this game or with new, more
specific games accounting for indiscriminate destruction.
\section*{Acknowledgments}
Financial support from Ministerio de Ciencia y Tecnolog\'{\i}a
(Spain) under projects FIS2009-13730-C02-02 (A.A.), FIS2009-13370-C02-01
(J.C. and R.J.R.) and MOSAICO (J.A.C.); from the Director, Office of Science,
Computational and Technology Research, U.S. Department of Energy under
Contract No.\ DE-AC02-05CH11231 (A.A.); from the Barcelona Graduate School of Economics and of the Government of Catalonia (A.A.); from the Generalitat de Catalunya under project 2009SGR0838 (A.A.)
2009SGR0164 (J.C. and R.J.R.) and from Comunidad de Madrid under project MODELICO-CM
(J.A.C.). R.J.R. acknowledges the financial support of the Universitat
Aut\`onoma de Barcelona (PIF grant) and the Spanish government (FPU grant).
|
\section{INTRODUCTION}
The intercalation of carbon nanomaterials with electron donors and acceptors is an active research area in which much effort is channelled towards the understanding and controlling of the electronic properties of graphene and graphite. Numerous potential applications such as sensors, electronic display panels, hydrogen storage and supercapacitors \cite{Seung-Hoon2002} have been suggested for such intercalated materials. The layered structure of graphites plays an important role in charge transfer reactions. Acceptor species can intercalate between graphitic layers, expanding the graphite with the resultant hybrids known as graphite intercalated compounds \cite{Dresselhaus2002}.
Bromine acts as an acceptor when intercalated in materials such as graphite or nanotubes, and has been proposed experimentally as a way to open a band gap in 3- or 4- layer graphene\cite{Jung2009}. The in-
plane electrical conductivity of graphite increases from $2.4 \times 10^{4}\ \Omega ^{-1} $cm$^{-1}$ at room temperature to $2.2 \times 10^{5}\ \Omega ^{-1} $cm$^{-1}$ after intercalation with bromine\cite{Dresselhaus2002}.
Bromine forms many ordered phases in graphites and undergoes an order-disorder phase transition as the
amount of bromine or temperature changes \cite{Dresselhaus2002,Bandow2002,Yoichi2003,Eeles1964, Bach1963,Bardhan1980}. Bromine intercalated graphite forms
stage-$n$ compounds where $n$ is the number of graphitic layers between planes of Br$_2$ ($n>1$ \cite{Sasa1971}). Extended X-ray Absorption Fine Structure (EXAFS) spectroscopy at
room temperature and 100K \cite{Heald1778} showed that intercalated bromine molecules lie parallel to the basal
plane, with an expansion of the Br-Br distance by 0.03 \AA\ to accommodate the lattice mismatch
between the free molecule and the 2.46 \AA\ spacing between graphite hexagons. X-ray diffraction and
electron microscopy studies \cite{Eeles1964} suggest that intercalated Br$_2$ at lower concentration is composed of
chains of Br$_2$ molecules in which the intermolecular distances is identical to that of solid bromine.
Graphite Raman spectra associated with Br$_2$ intercalation show a strong peak at 242 cm$^{-1}$ \cite{Ecklund1978, Erbil1983,Erbil1982,Defang2007} assigned to the intercalated
Br$_2$ stretch mode. The frequency is downshifted from 320 cm$^{-1}$ for gaseous Br$_2$ and 295 cm$^{-1}$ for solid
molecular bromine\cite{Cahill1966}. There have been limited density functional studies of
brominated graphite\cite{Widenkist2009} and graphene\cite{Seung-Hoon2002,Rudenko2009}.
In this paper we examine low density bromination of graphene and graphite using density
functional (DFT) calculations within the local density approximation\cite{Briddon2000}. The method has
been successfully used to study intercalated boron in graphite\cite{Suarez-Martinez2007}. A localised Gaussian basis set is used with a large
number of fitting functions per atom (22 for each C atom and 50 for each Br), with angular
momenta up to $l$=2 for C and $l$=3 for Br. A finite temperature electron level filling of kT=0.04eV is
used to improve convergence. Core electrons were eliminated using norm-conserving relativistic
pseudopotentials of Hartwigsen, Goedecker and Hutter\cite{Hartwigsen1998}. A cut-off energy of 150 Hartrees was
used to obtain convergence of the charge density.
Isolated Br$_2$ was calculated in a 13.23 \AA\ cubic supercell. Hexagonal $4\times4$ graphene supercells
containing C$_{32}$Br$_2$ were used with a large vacuum spacing of 31 \AA\ between layers to ensure no inter-
layer interaction, and a $4 \times 4 \times 1$ Monkhorst-Pack k-point grid\cite{Monkhorst1976}. Graphite calculations used $3 \times 3 \times n$
layer supercells (C$_{18}$)$_n$(Br$_2$)$_m$, $n$=1-4, $m$=1-2) for different layer stackings, with $4 \times 4 \times 1$ or $4 \times 4 \times 2$ k-point grids depending on cell size.
All structures were fully geometrically optimised with no constraints of symmetry, allowing both
atomic positions and cell dimensions to vary freely. Atomic charge states were obtained by
summing Mulliken population analysis over all the filled electronic states. Vibrational frequencies
were calculated by determining the energy and forces for $\pm$ 0.2 au displacements of the Bromine
atoms. The second derivatives on the displaced atoms can then be found from the two-sided
difference formula for the second derivative. All results are spin averaged, test calculations with spin polarisation all gave zero spin solutions as the most thermodynamically stable.
Our calculated bond length, 2.29\AA, and stretching frequency, 326 cm$^{-1}$, for isolated Br$_2$ show excellent
agreement with experiment (2.27\AA\ and 323 cm$^{-1}$\cite{Ecklund1978}, 2.283\AA\ and 320 cm$^{-1}$ \cite{Erbil1982}) and literature DFT/LDA values (2.263\AA \cite{Miao2009}, 2.244\AA\ and 324 cm$^{-1}$ \cite{Seung-Hoon2002}).
\begin{figure}
\begin{tabular}{cccc}
(a) 0.00 & (b) +0.02 & (c) +0.05 & (d) +0.11 \\
\includegraphics[width=1.7cm]{1a} & \includegraphics[width=1.7cm]{1b} &
\includegraphics[width=1.7cm]{1c} & \includegraphics[width=1.7cm]{1d} \\
(e) +0.14 & (f) +0.14 & (g) +0.15 & (h) +0.15 \\
\includegraphics[width=1.7cm]{1e} & \includegraphics[width=1.7cm]{1f} &
\includegraphics[width=1.7cm]{1g} & \includegraphics[width=1.7cm]{1h}
\end{tabular}
\caption{Optimised geometries of Br$_2$ molecule perpendicular (a-c) and parallel (d-h) to a
graphene sheet. Relative stabilities are quoted in eV (binding energy of Br$_2$ in structure (a) is 0.40eV). The text
refers to (a) Br$_2^\perp$\ and (d) Br$_2^\Vert$\ \label{brominepic}}
\end{figure}
\begin{table}
\caption{Calculated parameters for the most stable Br$_2^\perp$\ and Br$_2^\Vert$\ on graphene compared with literature DFT calculations.\label{strucparams}}
\begin{ruledtabular}
\begin{tabular}{lcc|c}
& Br$_2^\perp$\ & Br$_2^\Vert$\ & Literature \\ \hline
Orientation & Perp. & Parallel & Parallel \\
Binding energy / Br$_2$ (eV) & -0.40 & -0.29 & -0.20\cite{Seung-Hoon2002}\\&&& -0.29\cite{Rudenko2009} \\
Br-Br Stretch Frequency (cm$^{-1}$) & 288 & 270&311\cite{Seung-Hoon2002}\\
Br-Br bond length (\AA) &2.33&2.31&2.245\cite{Seung-Hoon2002}\\
Br-C distance (\AA) & 2.74&3.45&3.375\cite{Seung-Hoon2002} \\&&&3.74\cite{Rudenko2009} \\
Charge state of Br atoms (e)
&+0.48 & -0.04& \\
&-0.61&-0.04 \\
Total Charge transfer / Br$_2$ (e) &
-0.13 & -0.08&
\end{tabular}
\end{ruledtabular}
\end{table}
Figure \ref{brominepic} shows the structures of Br$_2$ over graphene in both perpendicular and parallel
orientations after geometry optimisation, with calculated properties for Br$_2^\perp$\ and Br$_2^\Vert$\ (structures (a) and (d) respectively in Figure \ref{brominepic}) in Table \ref{strucparams}.
The most thermodynamically stable arrangement is Br$_2$ oriented perpendicular to the graphene sheet above a
carbon atom (Figure \ref{brominepic}a, referred to hereafter as Br$_2^\perp$\ ).
Its binding energy of 0.40eV shows it will be strongly
physisorbed at room temperature. However the small variations in binding energy
between structures suggests that low density Br$_2$ binding to graphene will be largely orientation
independent; indeed and at these densities at room temperature Br$_2$ should be in constant tumbling motion
(this also holds for the results of\cite{Rudenko2009}).
Increasing Br$_2$ concentration did not significantly change the relative energies of perpendicular and parallel
orientations. However it is possible to obtain twice the maximum
surface density for Br$_2^\perp$\ as for Br$_2$ in the parallel orientation (Figure \ref{brominepic}d, referred to hereafter as Br$_2^\Vert$\ ).
Thus for the limit of high surface concentrations we expect Br$_2^\perp$\ to dominate.
Br$_2^\perp$\ represents a very unusual configuration for bromine.
It shows strong charge transfer (0.129e) from the graphene, with a very strong induced molecular dipole (Br$^{+0.480}$-Br$^{-0.609}$). The singly occupied p$_z$-orbital of the lower Br atom depopulates into the p$_z$ of the upper Br atom
forming a nascent Bromonium and Bromide ion pair. In this way the emptied lower p$_z$-orbital can sit
within the graphene $\pi$-cloud (Br only 2.74 \AA\ above the graphene).
This behaviour is reminiscent of the well-known reaction between Br$_2$ and unsaturated bonds in organic chemistry. However in these cases this
dipolar form of Br$_2$ is an unstable transient state that immediately saturates the C=C bond, forming two Br-C bonds. In graphene this final step would be endothermic due to steric hindrance between
the Br atoms as a result of the mechanical confinement of the lattice. Indeed our attempts to stabilise C-Br pairs on graphene in both neighbouring (1,2) and cross-hexagon (1,4) configurations both resulted in Br spontaneously reconstructing into a Br$_2$ molecule (placing Br-C in a (1,4) configuration with Br atoms on opposite sides of the graphene is 1.68eV less stable than Br$_2^\perp$\ ).
The unusual Br$_2^\perp$\ configuration is reflected in the band structure (Figure \ref{bandstructures}a).
The strong coupling between the graphene LUMO and the Br$_2$ antibonded state at around +0.6 eV reflects the interaction between the empty Br p$_z$-orbital and the graphene $\pi$-cloud, with the resultant low density Br$_2$ layer
opening a small 86meV band gap.
In contrast Br$_2^\Vert$\ has no induced dipole, with weaker charge transfer (0.084e) from the graphene. The molecule sits
3.254 \AA\ above the graphene {\it i.e.} above the $\pi$-cloud, with the additional charge occupying the
Br$_2$ $pp\sigma^{*}$ anti-bonding state. The band structure (Figure \ref{bandstructures}b) shows the
Br$_2$ states lie lower than those of Br$_2^\perp$\ by $\sim$0.7eV.
The Br$_2$ anti-bonded state pins the Fermi level $\sim$0.2eV lower than in the pristine case, leaving graphene states around the K-point depopulated, indicating charge transfer from graphene to Br$_2$.
The Bromine states are flat and largely decoupled from the graphene bands since
there is only weak interaction between Br$_2$ orbitals and the graphene $\pi$-cloud in this orientation.
\begin{figure}
(a) \includegraphics[width=5.0cm,angle=-90]{2a} \\
(b) \includegraphics[width=5.0cm,angle=-90]{2b}
\caption{Band structure of (a) Br$_2^\perp$\ and (b) Br$_2^\Vert$\ on graphene (eV), faded dotted lines indicate the same supercell of pristine graphene for comparison. Arrows indicate bromine related states. \label{bandstructures}}
\end{figure}
Our results are consistent with experimental Raman observations of brominated graphene\cite{Jung2009}.
While for graphene with three or more layers a resonant Raman signal was observed for Br$_2$ at
around 240cm$^{-1}$, for mono- and bi-layer graphene no bromine signal was observed. This could be
an orientation effect, since if the Br$_2$ sits perpendicular to the surface as we propose and orthogonal
Raman is used, then the molecules will be aligned with the beam and there will be no interaction
and hence no signal.
In addition the bromine HOMO/LUMO states are further apart for Br$_2^\perp$\ ($\sim$2.2eV) than for Br$_2^\Vert$\ ($\sim$1.96)
or the various graphite structures we have examined (1.7-2.0eV). Given the excitation laser used
(633nm, 1.96eV), Br$_2^\perp$\ may not be in resonance as the authors suggested \cite{Jung2009}.
Br$_2^\Vert$\ would have an associated Raman signal, and since none is observed this allows us to exclude this configuration.
We note that the strong dipole of Br$_2^\perp$\ will make the molecule infra-red active.
We now turn our attention to graphite.
Our calculated energy to separate AB graphite layers is 36.7 meV/atom, with AA stacked
graphite 12.0meV/atom less stable than AB stacked graphite. These figures are in good
agreement with experiment (35meV/atom\cite{Benedict1998}) and previous calculations (9.68-9.70meV/atom
AA/AB energy difference\cite{Telling2003,Charlier1994}). Our interlayer spacings of 3.39\AA\ and 3.50\AA\
for AB and AA-stacked graphite respectively are also in reasonable agreement with previous calculations\cite{Charlier1994}.
We placed Br$_2$ in a variety of different orientations and locations including above $\alpha$- and $\beta$- carbon
atom sites and hexagon centers, in graphite of various layer stackings.
Unlike graphene, Br$_2$ in graphite is more stable parallel
to the graphitic layers with Br atoms above hexagon centers (see Table \ref{graphiteparams}), in
agreement with experiment\cite{Eeles1964,Heald1778}.
Perpendicular oriented Br$_2$ structures are much less stable (by typically 0.52eV).
The results are summarised and compared with experiment in Table\ref{graphiteparams}.
We note that the charge transfer value from Raman was reported with a large uncertainty\cite{Erbil1982}.
\begin{table}
\caption{Calculated and experimental results for Br$_2$ intercalated graphite.
The C-C layer distance refers to the
layers separated by bromine. Experimental values from EXAFS \cite{Heald1778}, Raman \cite{Ecklund1978,Erbil1982} and XRD \cite{Eeles1964,Erbil1983,Sasa1971,Rudorff1941}.
At these low concentrations Br$_2$ intercalation is endothermic. \label{graphiteparams}}
\begin{ruledtabular}
\begin{tabular}{l|ccc}
& Stage-1 & Stage-2 & Experiment \\
Cell used & C$_{18}$Br$_2$ & C$_{72}$Br$_4$ & \\
Binding Energy/Br$_2$ (eV) & +0.08 & +0.08 & \\
C-C layer distance (\AA) & 6.45 & 6.47 & 7.0\cite{Sasa1971,Erbil1983} 7.05\cite{Rudorff1941} \\
$c$-axis (\AA) & 9.80 & 9.82 & 10.7\cite{Eeles1964} \\
Br-Br bond length (\AA) & 2.30 & 2.30 & 2.34 \cite{Heald1778,Heald1980}\cite{Erbil1983} \\
Br-C distance (\AA) & 3.51 & 3.46 & 2.9\cite{Heald1778} \\
Br-Br frequency (cm$^{-1}$) & 287 & 274 & 242-258 \cite{Ecklund1978,Erbil1982} \\
Charge transfer / Br$_2$ (e) & -0.10 & -0.12 & -0.16\cite{Heald1980} -0.34 \cite{Erbil1982} \\
\end{tabular}
\end{ruledtabular}
\end{table}
Besides the lowest energy structures quoted here we obtained many metastable structures.
While their energies were all within 0.01-0.05eV of
those structures discussed here, they show significant variation in the Br$_2$ stretch frequency (250-278cm$^{-1}$), and slight variation in position and Br-Br bond length (2.31-2.33\AA). This suggests that at room temperature Br$_2$ in graphite will be mobile, and is consistent with the observation
of a broad and somewhat complicated Raman peak \cite{Ecklund1978}.
We explored all possible layer stacking combinations for stage-1 and stage-2 intercalated graphites.
Bromine molecules are most stable with AA-stacked graphite each side, while unbrominated graphite layers
preferentially are AB stacked.
Thus (indicating layers of bromine molecules by X) we find the most stable stage-1 phase to be [AX]$_n$,
and stage-2 to be [AXABXB]$_n$, suggesting [AXAB]$_n$ and [AXABABXBAB]$_n$ for
stage-3 and stage-4 respectively.
Our calculated bromine intercalation energy is weakly endothermic, since at these low densities the
energetic cost associated with separating graphite planes (a cost per unit area) is not sufficiently
offset by the binding energy of Br$_2$ to the layers.
Subtracting the energy to separate graphite layers from the Br$_2$ intercalation energy
gives an energy for intercalation of Br$_2$ into `pre-separated stage-1 graphite' of 0.581eV/Br$_2$.
This implies that the {\it minimum} Br$_2$ concentration for exothermic intercalation in stage-$n$ graphite will be C$_{16n}$Br$_2$. Indeed a fixed-Br calculation for a C$_8$Br$_2$ stage-1 high coverage structure gives Br$_2$ intercalation as exothermic. Thus Br$_2$ will aggregate in the same inter-layer space and should be considered as a layer rather than individual molecules.
This minimum required concentration also suggests intercalation will be a slow diffusion process with an
abrupt diffusion front. This is consistent with long experimental intercalation times \cite{Fischer2002},
despite the high bromine inter-layer mobility. It also explains why,
on out-gassing bromine, the material switches from a stage-2 to compound to
stage-$n$ ($n$=3, 4, ...) rather than remaining stage-2 with lower bromine density per layer\cite{Dresselhaus2002, Ecklund1978,Fischer2002}.
We note that frequency calculations incorporating the energetic double derivatives of surrounding carbon atoms gave identical values to within 1 cm$^{-1}$, showing that the Br$_2$ stretch mode is decoupled
from the surrounding carbon lattice consistent with the literature\cite{Seung-Hoon2002}.
Band structure calculations of Br$_2$ layers between graphite sheets (not shown here) give a bromine-related state which pins the Fermi level just below that of perfect graphite ($\sim$0.1eV) indicating charge transfer from graphite to Br$_2$. This state shows some dispersion indicating weak coupling with the underlying graphite. In other respects the graphite band structures are barely perturbed.
Our results on graphite and graphene can explain the anomalously large G peak shift for single graphene sheets in comparison with few-layer graphene\cite{Jung2009}.
Since the higher maximum Br$_2$ surface density for Br$_2^\perp$\ than for inter-layer
Br$_2$ means charge transfer
per unit area will be higher. Additionally Br$_2$ can attach to both sides of graphene.
We find a binding energy of -0.38eV /Br$_2$ for two
Br$_2^\perp$\ either side of the same C atom, with associated charge transfer of -0.12e/Br$_2$. Thus the net total charge transfer per
unit area will indeed be significantly higher for monolayer graphene than multi-layer systems.
Test calculations for a (5,5) single walled nanotube, either isolated or in bundles, gave similar structural behaviour, {\it i.e.} Br$_2$ on the surface of the isolated tube adopts a
perpendicular orientation, while intercalated Br$_2$ sits parallel to the tube walls. This will
be explored further in a later publication.
In summary, we have examined low density Br$_2$ adsorption in graphene and graphite.
On graphene Br$_2$ adopts an unusual perpendicular orientation, opening a small band gap ($\sim$86meV) in the
graphene with strong charge transfer.
The molecule forms a Br$^{+}+$Br$^{-}$ pair, rendering it infra-red active. This is a new form of Br$_2$ previously only considered as an unstable intermediate to bromine induced carbon bond saturation. Such graphene-induced stabilisation behaviour may be mirrored in other molecular species, enabling study of otherwise unstable reactive molecular forms.
In graphite Br$_2$ adopts a parallel orientation to the sheets with an associated charge transfer. Our
calculations are in good agreement with experimental data where available. At high bromine
concentrations and low temperatures there is some evidence of bromine chain structure formation
\cite{Chung1986,Chung1977} in graphite, and we are currently investigating this further.
We note that high density bromination of graphite leads to stage-2 compounds, and in conjunction
with an appropriate secondary surfactant this may be a promising way to produce bilayer graphene.
\begin{acknowledgments}
CPE and PW thank the ``NANOSIM\_ GRAPHENE'' project
ANR-09-NANO-016-01 funded by the French National Agency (ANR) within the P3N2009 programme.
\end{acknowledgments}
\bibliographystyle{apsrev4-1}
|
\section{Introduction}
Let $F = F(A)$ be a free group of rank $r \geq 2$ with basis $A = \{a_1, \dots, a_r\}$. A subgroup $V$ of $F$ is \emph{fully invariant} if, for all endomorphisms $f: F \to F$, $f(V) \subseteq V$. Fully invariant subgroups of $F$ include the trivial subgroup, the derived subgroups, the subgroups in the lower central series of $F$, and, more generally, verbal subgroups of $F$. Every fully invariant subgroup is normal in $F$, and the quotients of the form $F/V$ include the free group, free nilpotent groups, and free solvable groups. A subset $S \subset F$ is \emph{primitive in $F$ mod $V$} if the corresponding set of cosets $SV$ of $V$ can be extended to a basis of $F/V$ (i.e. to a generating set for $F/V$ such that each mapping of this set to $F/V$ extends to an endomorphisms of $F/V$).
We wish to study the following situation. Let $V$ be a fully invariant subgroup of $F$ and let $S$ be a primitive set mod $V$. Assume that there exists some index $l \in \{1, \dots, r-1\}$ such that, for each $s \in S$, the generators $a_t, t=l+1,\ldots,r$ of $F(A)$ are not used to express $s$ as a reduced word in the alphabet $A$. In short, $S \subset \hat{F} \subset F$, where $\hat{F} := F(\{a_1, \ldots, a_{l}\})$. Then $\hat V=V\cap\hat F$ is a fully invariant subgroup of $\hat F$ (as every endomorphism of $\hat F$ can be extended to an endomorphism of $F$). Our main question is:
\begin{mainquestion}
Under what conditions is it true that $S$ is also primitive in $\hat{F}$ mod $\hat V$?
\end{mainquestion}
If $V$ is trivial -- that is, if we are considering $S$ to be primitive in $F$ -- and $|S| = r-1$ (where $|S|$ denotes the size of set $S$), then it is not difficult to show that $S$ must be primitive in $\hat{F}$. The question becomes more interesting when $|S|$ is allowed to vary, and when $V$ is allowed to vary. Our main results answer this question for different choices of $V$, $|S|$ and $l$. Let $\mathbf A_m$ denote the variety of all abelian groups whose exponent divides $m$. In particular, $\mathbf A:=\mathbf A_0$ is the variety of all abelian groups. Then the product variety $\mathbf A_m\mathbf A_n$ consists of extensions of abelian groups of exponent dividing $m$ by abelian groups of exponent dividing $n$.
\pagebreak
\begin{theorem}\label{thm:main} Let $S \subset \hat{F}$ be primitive in $F$ mod $V$. Then $S$ is primitive in $\hat{F}$ mod $\hat{V}$ if $F/V$ is:
\begin{enumerate}
\item free,
\item free abelian of arbitrary exponent (including 0),
\item free nilpotent,
\item free in the abelian-by-abelian variety $\mathbf A_m\mathbf A_n$ and at least one of the following conditions fails: $(a)$ $m=0$, $(b)$ $n>0$, or $(c)$ either $|S| = r-1$ or $|S|=l-1$; or
\item free solvable and $|S|=r-1$.
\end{enumerate}
\end{theorem}
The motivation for this theorem came from the statement in the free case, which was initially proved by the authors as a lemma pertaining to the work in \cite{SabalkaSavchuk}. The result was a generalization of one case of a result describing the structure of subwords of primitive elements in $F$ that do not involve one of the generators of $F$. Though it was eventually realized that this lemma was unnecessary for the paper \cite{SabalkaSavchuk}, the authors felt the lemma and proof interesting enough to generalize and publish.
We conclude the introduction with the following natural question.
\begin{question}
Does there exist a variety of groups for which the Main Question has a negative answer? In particular, are there varieties where a stably primitive element is not necessarily primitive?
\end{question}
The structure of the paper is as follows. We begin by examining the case where $V$ is trivial in Section~\ref{sec:free}. Then we proceed to cases of the free abelian groups of arbitrary exponent in Section~\ref{sec:abelian} and the free nilpotent and free solvable groups in Section~\ref{sec:nilpotent}. We conclude the paper by the cases of free abelian-by-abelian of arbitrary exponents in Section~\ref{sec:product_variety} and free solvable groups in Section~\ref{sec:solvable}.
The authors would like to thank Mark Sapir for helpful advice on this paper, and Michael Handel, Ilya Kapovich, Olga Kharlampovich, Marcin Mazur, and Benjamin Steinberg for useful conversations on this material. And we are especially grateful to Martin Kassabov for providing a short proof of the free nilpotent case that replaced a longer argument involving `lifting primitivity' and to Vitali\u{\i} Roman'kov for pointing out an error in an earlier version of the paper. Finally, we greatly appreciate the comments and suggestions of the referee that enhanced the paper.
\section{The Free Case}\label{sec:free}
In the free case, the answer to our Main Question is: always. Our proof uses an interpretation of primitivity using Fox derivatives via a theorem of Umirbaev.
We begin by recalling the definition of free Fox derivatives and establish notation for them. Let $\mathbb ZF_r$ denote the integral ring over the free group $F_r$. For each $j=1,2,\ldots,r$ define a linear map $\partial_j\colon \Z F_r\to\Z F_r$ recursively by
\[\partial_j(a_j)=1,\qquad \partial_j(a_i)=0, i\neq j\]
and
\[\partial_j(uv)=\partial_j(u)+u\partial_j(v) \text{ for all } u,v\in F_r.\]
The map $\partial_j$ is called the (left) free Fox derivative associated with $a_j$.
\begin{theorem}\cite{umirbaev:primitive}\label{thm:Umirbaev}
A subset $\{x_1,\ldots,x_k\}$ is primitive in $F_r$ if and only if the $k\times r$ Jacobian matrix $J=\bigl[\partial_j(x_i)\bigr]_{1\leq i\leq k, 1\leq j\leq r}$, is right invertible in the integral ring $\mathbb ZF_r$.
\end{theorem}
We note that the ``only if'' part in this theorem was proven by Birman in \cite{birman:primitive}.
\begin{theorem} \label{thm:free}
Let $S=\{x_1,x_2,\ldots, x_k\}\subset \hat{F}$ be primitive in $F$. Then $S$ is primitive in $\hat{F}$.
\end{theorem}
\begin{proof}
By Theorem \ref{thm:Umirbaev}, since the set $S$ is primitive in $F$, the associated Jacobian $J$ is such that there exists an $r\times k$ matrix $P=[p_{jl}]$ with $p_{jl}\in \mathbb ZF_r$
satisfying
\begin{equation}
\label{eqn_JP}
JP=I_k,
\end{equation}
where $I_k$ is the $k\times k$ identity matrix over $\mathbb ZF_r$.
For $m > l$, since elements of $S$ do not involve $a_m$, the $m^{th}$ column of the matrix $J$ consists of zeros and all other entries of $J$ do not involve $a_m$. Each entry of matrix $P$ can be uniquely written in the form $p_{ij}=q_{ij}+r_{ij}$, where $q_{ij}$ represents the sum of all terms in $p_{ij}$ involving some $a_m$ with $m>l$, and $r_{ij}$ is an element of $\mathbb ZF_{l}$ not involving any $a_m$ with $m>l$. Then for matrices $Q=[q_{ij}]$ and $R=[r_{ij}]$ we have $P=Q+R$, and by Equation~\eqref{eqn_JP}, $JQ+JR=I_k$. But each entry of matrix $JQ$ is either $0$ or involves $a_m$ for some $m>l$. Since neither $I_k$ nor $JR$ involves $a_m$ for $m>l$ we must have that $JQ=0$, yielding
\begin{equation}
\label{eqn_JR}
JR=I_k
\end{equation}
Now let $\hat J$ and $\hat R$ be the matrices obtained from $J$ and $R$, respectively, by deleting last $r-l$ columns and rows, respectively. Then $\hat J$ is the Jacobian matrix of the set $S$ seen as a subset of $\hat F=F_{l}$ and $\hat R$ is a matrix over $\mathbb Z\hat F=\mathbb ZF_{l}$. Also, equation~\eqref{eqn_JR} implies that
\begin{equation*}
\hat J\hat R=I_{k}.
\end{equation*}
Thus by Umirbaev's criterion the set $S$ is a subset of the basis of $\hat{F}$.
\end{proof}
Via private communication, Ilya Kapovich has detailed an alternative proof of Theorem \ref{thm:free}, using Gersten's characterization of Whitehead's algorithm for subgroups \cite{Gersten}. Olga Kharlampovich has also suggested a proof using Bass-Serre theory. The benefit of the above proof is that it generalizes to the case of free abelian-by-abelian groups of arbitrary exponents in Section~\ref{sec:product_variety} (and also to the case of free metabelian groups and partially to the case of free solvable groups, though shorter proofs in these cases are given here).
\section{The Abelian Case}\label{sec:abelian}
The answer to our Main Question in the free abelian case is also: always. Our proof works for free abelian groups of arbitrary exponent and uses linear algebra.
For any $n> 0$, Let $\Z_n := \Z/n\Z$ denote the free group of rank 1 in the variety $\mathbf A_n$ of all abelian groups of exponent dividing $n$, and let $F^n$ denote the subgroup of $F$ generated by all $n^{th}$ powers and similarly define $\hat{F}^n$. Let $F'$ denote the commutator subgroup of $F$ and similarly define $\hat{F}'$. Note $(\hat F)'=\widehat{(F')}$, so $\hat F'$ is well defined. Further, for each $x\in F$ we denote by $\bar x$ the image of $x$ in $F/F'$ under the canonical epimorphism. For a subset $A$ of $F$ define $\bar A=\{\bar a\colon a\in A\}$.
We start with the proof for free abelian groups in the case of exponent $0$.
\begin{proposition}\label{prop:abelian}
Let $S=\{x_1,x_2,\ldots, x_k\}\subset \hat F$ be primitive in $F \mod F'$. Then $S$ is primitive in $\hat F\mod \hat F'$.
\end{proposition}
Before we proceed to our proof, we note that this Proposition follows from a well-known fact (see, for example,~\cite{kargapolov_m:fundamentals_book_eng}) that the basis of any subgroup of a finitely generated free abelian group can be extended to a basis of the whole group. Indeed, since $\bar S$ can be extended to a basis in $F/F'$, we get that $\bar S$ is linearly independent. Hence, $\bar S$ is a basis of a subgroup $\langle \bar S\rangle<\hat F/\hat F'$. We give a proof of Proposition~\ref{prop:abelian} that extends to the case of free abelian groups of finite exponent in Theorem~\ref{thm:abelian_exponent}.
\begin{proof}
There is a canonical embedding $\Z^{l} \cong \hat F/\hat F'\hookrightarrow F/F' \cong \Z^r$ and with a slight abuse of notation we will sometimes identify $\hat F/\hat F'$ with its image under this embedding.
The standard basis of $\Z^r\cong F/F'$ is $\{\bar a_1, \dots, \bar a_r\}$, whereas $\{\bar a_1, \dots, \bar a_{l}\}$ is the standard basis of $\Z^{l} \cong \hat F/\hat F'$. Extend $\bar S$ to a basis $B$ of $\Z^r$. Express $B$ with respect to the standard basis as a matrix $M$, where the elements of $\bar S$ correspond to the first $k$ columns of $M$. Let $\hat M$ denote the submatrix of $M$ obtained by deleting the last $r-l$ rows and columns. Then $\hat M$ corresponds to a set $\hat B$ of $l$ elements of $\Z^{l}$. Since the elements of $\bar S$ do not involve generators $\bar a_{l+1},\ldots,\bar a_r$, the matrix $M$ has a block triangular shape:
\[M=\left[\begin{array}{cc}
\hat M&P\\
0&Q\\
\end{array}\right],\]
for some $l\times (r-l)$-matrix $P$ and $(r-l)\times(r-l)$-matrix $Q$. Then since $M$ is in $SL_r(\mathbb Z)$ we have
\[\det\hat M\cdot \det Q=\det M=\pm1.\]
This is possible only if $\det\hat M=\pm 1$, and $\hat M$ is invertible. This shows $\hat B$ is a basis of $\Z^{l} \cong \hat{F}/\hat{F}'$. But $\bar S \subset \hat B$, so $\bar S$ is primitive in $\hat{F}/\hat{F}'$ and thus $S$ is primitive in $\hat F\mod \hat F'$.
\end{proof}
Now we proceed to the case of free abelian groups with exponent $n > 0$. First we note that the free group of rank $r$ in this variety is isomorphic to $F/(F'F^n)$.
\begin{theorem}\label{thm:abelian_exponent}
Let $S=\{x_1,x_2,\ldots, x_k\}\subset \hat F$ be primitive in $F \mod F'F^n$. Then $S$ is primitive in $\hat F\mod \hat F'\hat F^n$.
\end{theorem}
\begin{proof}
The proof is almost identical to that of Proposition~\ref{prop:abelian}. We replace $F'$ with $F'F^n$, $\hat F'$ with $\hat F'\hat F^n$, and $\Z$ with $\Z_n$, but otherwise define all other objects in the same way. The only other change to make is the following argument to show that $\hat M$ is invertible. The matrix $M$ is invertible in $SL_r(\Z_n)$ if and only if $\det M$ is a unit in $\Z_n$. But a product of elements in $\Z_n$ is a unit if and only if each of the elements is also a unit. Thus, $\det M = \det \hat M \cdot \det Q$ is a unit in $\Z_n$ if and only if $\det \hat M$ is a unit in $\Z_n$, if and only if $\hat M$ is invertible in $SL_r(\Z_n)$.
\end{proof}
\section{The Free Nilpotent Case}
\label{sec:nilpotent}
We now turn to the free nilpotent and free solvable cases. A \emph{commutator} of \emph{weight} $c$ is an expression recursively defined by $[y_1, \dots, y_{c-1}, y_c] := [[y_1, \dots, y_{c-1}], y_c]$ and $[y_1, y_2] := y_1^{-1}y_2^{-1}y_1y_2$. Fix a value for $c\geq2$, and let $\gamma_c(F)$ be the normal closure of subset of $F$ consisting of all commutators of weight $c$ (i.e. the $c^{th}$ term in the lower central series of $F$). Then $F/\gamma_{c+1}(F)$ is the free nilpotent group of class $c$. Also define $\gamma_1(F)=F$. Further, we write $F^{(t)}$ for the $t^{th}$ derived subgroup of $F$. Then $F/F^{(t)}$ is a free solvable group of derived length $t$. Note that $F' = \gamma_2(F) = F^{(1)}$, so the free abelian group is the free nilpotent group of class $1$ and the free solvable group of derived length $1$.
\begin{theorem}\label{thm:nilpotent}~\\\vspace{-.4cm}
\begin{enumerate}
\item Let $S=\{x_1,x_2,\ldots, x_k\}\subset \hat F$ be primitive in $F \mod \gamma_{t+1}(F)$ for some $t\geq 1$. Then $S$ is primitive in $\hat F\mod \gamma_{t+1}(\hat F)$.
\end{enumerate}
\end{theorem}
The following proof is based on suggestions and references from Martin Kassabov and the referee. In the nilpotent case, our original proof used a primitivity lifting technique \cite{gupta_g:lifting_primitivity_nilpotent}. In the solvable case, our original proof was based on a generalization of our approach in the free case, and used Fox derivatives and criteria for primitivity by Roman'kov and Timoshenko in the free metabelian case \cite{romankov:primitivity,timoshenko:algorithmic}, and by Krasnikov in the free solvable case~\cite{krasnikov:generating_elements_eng}. However, our original proofs were longer and less general.
\begin{proof}
It is a well-known fact (see, for example,~\cite[Lemma 5.9]{magnus_ks:combinatorial_group_theory}) that a set of elements generates a nilpotent group $N$ if and only if the projection of these elements to the abelianization $N/N'$ generates the abelianization. Assume a subset $S = \{x_1, \dots, x_k\}$ is primitive in $F$ mod $\gamma_{t+1}(F)$ and $S$ does not involve generators $a_{l+1},\ldots,a_r$. Then $S$ is also primitive over $F'$, which implies by Proposition~\ref{prop:abelian} that $S$ is primitive in $\hat F$ mod $\hat F'$. Now by the above fact $S$ must be primitive in $\hat F\mod \gamma_{t+1}(\hat F)$.
\end{proof}
\section{The Abelian-by-Abelian Case}
\label{sec:product_variety}
For integers $m, n \geq 0$, we consider the case of free groups in the variety $\mathbf A_m \mathbf A_n$ of all extensions of abelian groups of exponent dividing $m$ by abelian groups of exponent dividing $n$. Let $V_{m,n}$ and $\hat{V}_{m,n}$ denote the corresponding fully invariant subgroups of $F$ and $\hat F$, respectively. The free Fox derivatives $\partial_i$ induce Fox derivatives $\partial^0_i\colon \Z F\to \Z_m(F/F'F^n)$ (see Gupta and Timoshenko~\cite{gupta_t:primitivity_in_AnAm96}) by
\[\Z F\stackrel{\partial_i}{\longrightarrow}\Z F\stackrel{\alpha^*}{\longrightarrow}\Z (F/F'F^n)\stackrel{\gamma^*}{\longrightarrow}\Z_m(F/F'F^n),\]
where $\alpha^*\colon\Z F\to\Z(F/F'F^n)$ and $\gamma^*\colon \Z(F/F'F^n)\to\Z_m(F/F'F^n)$ are linear extensions of the canonical epimorphisms $\alpha\colon F\to F/F'F^n$ and $\gamma\colon \Z\to\Z_m$ respectively.
In \cite{gupta_t:primitive_systems_in_AmAn99}, Gupta and Timoshenko prove the following theorem.
\begin{theorem}\cite{gupta_t:primitive_systems_in_AmAn99}\label{thm:AmA}
Let $S=\{x_1,x_2,\ldots, x_k\}\subset F$. Let $J = (\partial^0_j x_i), i=1,\ldots,k, j=1,\ldots,r$ denote the $k \times r$ Jacobian matrix of $S$ over $\Z_m(F/F'F^n)$. Assume that at least one of the conditions $m=0$, $n>0$, or $k=r-1$ fails. Then $S$ is primitive mod $V_{m,n}$ if and only if:
\begin{enumerate}
\item[$(i)$] the ideal in the ring $\Z_m(F/F'F^n)$ generated by $k^{th}$ order minors of $J$ is the whole ring, and
\item[$(ii)$] $S$ is primitive mod $F'F^n$.
\end{enumerate}
\end{theorem}
Using the above theorem we deduce:
\begin{theorem}
Let $S=\{x_1,x_2,\ldots, x_k\}\subset \hat F$ be primitive in $F \mod V_{m,n}$ and assume at least one of the following conditions fails:
$(a)$ $m=0$, $(b)$ $n>0$, or $(c)$ either $k = r-1$ or $k=l-1$. Then $S$ is primitive in $\hat F\mod \hat V_{m,n}$.
\end{theorem}
\begin{proof}
Let $m_1, \dots, m_t$ denote the $k^{th}$ order minors of $J$ in $\Z_m(F/F'F^n)$, and let $J_m$ denote the ideal of $\Z_m(F/F'F^n)$ generated by $m_1, \dots, m_t$. As $S \subset \hat F$, $J$ is a matrix over $\Z_m(\hat{F}/\hat{F}'\hat F^n)$ whose last $r-l$ columns are zero columns. Therefore the $l\times k$ Jacobian matrix $\hat J=(\partial_j x_i)$ formed by viewing $S$ as a subset of $\hat F$ is obtained from $J$ by simply removing last $r-l$ columns. But removing columns of zeros does not change the $k^{th}$ order minors of a matrix, so the $k^{th}$ order minors of $\hat{J}$ are also $m_1, \dots, m_t$, considered now as elements in $\Z_m(\hat{F}/\hat{F}' \hat{F}^n)$ (via the canonical embedding of $\Z_m(\hat{F}/\hat{F}'\hat F^n)$ into $\Z_m(F/F'F^n)$). Let $\hat{J}_m$ denote the ideal generated by $m_1, \dots, m_t$ in the ring $\Z_m(\hat{F}/\hat{F}'\hat{F}^n)$.
Since $S$ is primitive in $F \mod V_{m,n}$, by Theorem \ref{thm:AmA} we know that: ($i$) $J_m = \Z_m(F/F'F^n)$, and ($ii$) $S$ is primitive mod $F'F^n$. It follows from ($i$) that $J_m$ contains the identity in $\Z_m(F/F'F^n)$. Therefore for some $p_i\in\Z_m(F/F'F^n)$ we have
\begin{equation}
\label{eqn:ideal}
1=\sum_{i=1}^{t}m_ip_i.
\end{equation}
We can decompose each $p_i$ as
\[p_i=q_i+r_i,\]
where $q_i$ is the sum of terms in $p_i$ that do not involve $a_{l+1},\ldots,a_r$, and $r_i$ is the sum of terms in $p_i$ that do involve at least one of the $a_{l+1},\ldots,a_r$ in non-zero power. Then equation~\eqref{eqn:ideal} can be rewritten as
\begin{equation*}
\label{eqn:ideal2}
1-\sum_{i=1}^{t}m_iq_i=\sum_{i=1}^{t}m_ir_i
\end{equation*}
The left-hand side of the above equation does not have terms involving $a_{l+1},\ldots,a_r$. However, since every term of each $m_i$ does not involve any of $a_{l+1}, \dots, a_r$ but every term of each $r_i$ does, every term of $m_ir_i$ involves at least one of the elements $a_{l+1}, \dots, a_r$. Some of the terms of $m_ir_i$ might cancel, but it follows that every surviving term of $m_ir_i$ must involve at least one of the elements $a_{l+1}, \dots, a_r$. Further, when we take the sum of $m_ir_i$ we cannot create terms that do not involve $a_{l+1},\ldots,a_r$ as the support of $\sum_{i=1}^{t}m_ir_i$ viewed as a function from $F/F'F^n$ to $\Z_m$ is included into the union of supports of $m_ir_i$, which, in turn, are included in the set of all elements of $F/F'F^n$ that involve $a_{l+1},\ldots,a_r$. Thus, for the two sides to be equal, the only possibility is if both sides are equal to zero. But this yields that
\begin{equation*}
1=\sum_{i=1}^{t}m_iq_i
\end{equation*}
with $m_i, q_i\in \Z_m(\hat{F}/\hat{F}'\hat{F}^n)$. Thus the identity in $\Z_m(\hat{F}/\hat{F}'\hat{F}^n)$ belongs to the ideal $\hat{J}_m$, and so condition $(i)$ holds for $S$ with respect to the ring $\Z_m(\hat{F}/\hat{F}'\hat{F}^n)$.
It follows from ($ii$), the fact that $S \subset \hat F$, and Proposition~\ref{prop:abelian} and Theorem~\ref{thm:abelian_exponent} that $S$ is primitive mod $\hat{F}'\hat{F}^n$. Thus, applying Theorem~\ref{thm:AmA} to $S$ thought of as a subset of $\hat F$, we see that $S$ is primitive in $\hat{F} \mod \hat{V}_{m,n}$. Note that Theorem~\ref{thm:AmA} applies in this situation if at least one of the conditions $m=0$, $n>0$, or $k=rank(\hat F)-1=l-1$ fails, which is true by assumption.
\end{proof}
Vitali\u{\i} Roman'kov suggested that one could potentially use the basis cofinality property of free groups of countable rank in subvarieties $\mathbf B$ of $\mathbf N_c \mathbf A$ (where $\mathbf N_c$ denotes the variety of nilpotent groups of class $c$) to generalize our results in this section. This property was proved in this case by Bryant and Roman'kov in~\cite{bryant_r:automorphism_group_of_rel_free_groups99} (see also~\cite{bryant_e:small_index_property97} for definitions). To do so, one must show that primitive systems in a finitely generated relatively free group in $\mathbf B$ that miss certain generators can be lifted to primitive systems in $F$ with the same property. For arbitrary varieties this statement is false, but it could be true for the varieties above.
\section{The Free Solvable Case}
\label{sec:solvable}
The final case we consider is the case of free solvable groups. First of all, the free metabelian case is covered by Section~\ref{sec:product_variety}. Alternatively, for the free metabelian case one can use the primitivity criteria developed by Timoshenko~\cite{timoshenko:on_the_inclusion_elements_into_a_basis_of_free_metabelian_group88,timoshenko:primitive_systems98}
and Roman'kov~\cite{romanov:primitive_elements91,romankov:primitivity}, which use Fox derivatives similar in spirit to Umirbaev's criterion~\cite{umirbaev:primitive} stated in Theorem~\ref{thm:Umirbaev}.
Unfortunately, there is no such criterion for free solvable groups of arbitrary derived length. However, Krasnikov~\cite{krasnikov:generating_elements} has obtained a result for groups of the form $F/[N,N]$ where $N$ is a normal subgroup of $F$, which works only in the case $k=r$. The Fox derivatives used here are the free Fox derivatives of Section \ref{sec:free}.
\begin{theorem}\cite{krasnikov:generating_elements}\label{thm:krasnikov}
A set $\{x_1,\ldots,x_r\}$ in $F/[N,N]$ is a generating set of this group if and only if the $r\times r$ Jacobian matrix $J=\bigl[D_j(x_i)\bigr]_{1\leq i,j \leq r}$ of free Fox derivatives, projected to $Z(F/N)$, is left invertible in the integral ring $\mathbb Z(F/N)$.
\end{theorem}
With this criterion we obtain an analog of Theorem~\ref{thm:free} in the case of free solvable groups with a restriction on the size of $S$.
\begin{theorem}\label{thm:solvable}
Let $S=\{x_1,x_2,\ldots,x_{r-1}\} \subset \hat{F}$ be primitive in $F$ mod $F^{(t)}$, where $F^{(t)}$ is the $t$-th derived subgroup of $F$, $t > 0$. Then $S$ is a basis for $\hat{F}/\hat{F}^{(t)}$.
\end{theorem}
\begin{proof}
Since $S$ is primitive in $F\mod F^{(t)}$, we can extend $S$ to a set $\tilde S=\{x_1,x_2,\ldots,x_r\}\subset F$ such that the images of elements of $\tilde S$ in $F/F^{(t)}$ under the canonical epimorphism form a basis. By Theorem~\ref{thm:krasnikov} we have that the associated to $\tilde S$ Jacobian matrix $J=\bigl[D_j(x_i)\bigr]_{1\leq i, j \leq r}$ is left invertible over $\mathbb{Z}(F/F^{(t-1)})$, so there exists an $r\times r$ matrix $P=\bigl[p_{ij}]_{1 \leq i, j \leq r}$ with entries over $\mathbb Z(F/F^{(t-1)})$ such that
\[PJ=I_r,\]
where $I_r$ is an $r\times r$ identity matrix over $\mathbb Z(F/F^{(t-1)})$.
Since elements of $S$ do not involve the generator $a_r$, we have
\[D_r(x_j)=0, \quad j=1,\ldots,r-1,\]
and $D_i(x_j)$ does not involve $a_r$ for $1\leq j\leq r-1$. Therefore, matrix $J$ can be written in the form
\[J=\left[
\begin{array}{c|c}
\hat J&\begin{array}{c}0\\ \vdots\\ 0
\end{array}\\\hline
D_1(x_r)\cdots D_{r-1}(x_r)&D_{r}(x_r)
\end{array}\right],\]
where $\hat J=\bigl[D_i(x_j)\bigr]_{1\leq i, j\leq r-1}$ is the Jacobian martix of $S$ in $\hat F/\hat F^{(t-1)}$.
Then
\[I_r=PJ=\left[
\begin{array}{c|c}
*&\begin{array}{c}p_{1r}D_r(x_r)\\ \vdots\\ p_{r-1,r}D_r(x_r)
\end{array}\\\hline
*\cdots *&p_{rr}D_{r}(x_r)
\end{array}\right].\]
Hence, $p_{rr}D_r(x_r)=1$ in $\mathbb Z(F/F^{(t-1)})$ and, in particular, $D_r(x_r)\neq 0$. On the other hand,
\[p_{in}D_r(x_r)=0, \quad i=1,\ldots,r-1,\]
so we must have $p_{in}=0$ in $\mathbb Z(F/F^{(t-1)})$. This follows from the fact that the free solvable group $F/F^{(t-1)}$ is a torsion-free elementary amenable group, and for such groups the Kaplansky conjecture on zero divisors holds true~\cite{linnell:zero_divisors}. In particular, this implies that the ring $\mathbb Z(F/F^{(t-1)})$ does not have zero divisors.
Now we have
\begin{eqnarray*}
I_r=PJ
&=&\left[\begin{array}{c|c}
\hat P&\begin{array}{c}0\\ \vdots\\0
\end{array}\\\hline
p_{r1}\cdots p_{r,r-1}&p_{rr}
\end{array}\right] \left[\begin{array}{c|c}
\hat J&\begin{array}{c}0\\ \vdots\\ 0
\end{array}\\\hline
D_1(x_r)\cdots D_{r-1}(x_r)&D_{r}(x_r)
\end{array}\right]\\
&=&\left[\begin{array}{c|c}
\hat P\hat J&\begin{array}{c}0\\ \vdots\\ 0
\end{array}\\\hline
*\cdots *&p_{rr}D_{r}(x_r)
\end{array}\right].
\end{eqnarray*}
Therefore we must have $\hat P\hat J=I_{r-1}$, where $I_{r-1}$ is the $(r-1)\times(r-1)$ identity matrix over $\mathbb Z(F/F^{(t-1)})$. Similarly to the proof in the free case, it follows that $\hat P$ can be chosen to be in $\mathbb Z(\hat F/\hat F^{(t-1)})$. Finally, applying Theorem~\ref{thm:krasnikov} again, we obtain that $S$ is a basis for $\hat F/\hat F^{(t)}$.
\end{proof}
\def$'${$'$} \def$'${$'$} \def$'${$'$} \def$'${$'$}
\def$'${$'$}
|
\section{Introduction}
Being related to almost all stable elementary particles such as
proton, electron and neutrino, spinor field, especially Dirac
spin-$1/2$ play a principal role at the microlevel. However, in
cosmology, the role of spinor field was generally considered to be
restricted. Only recently, after some remarkable works by different
authors (e.g., Henneaux, 1980; Saha, 2001a, 2004a, 2004b, 2006c,
2006d; Armend$\acute a$riz-Pic$\acute o$n and Greene, 2003; Ribas
{\it et al}, 2005; Souza and Kremer, 2008), showing the important
role that spinor fields play on the evolution of the Universe, the
situation began to change. This change of attitude is directly
related to some fundamental questions of modern cosmology: (i)
problem of initial singularity; (ii) problem of isotropization and
(iii) late time acceleration of the Universe.
{\bf (i) Problem of initial singularity:} One of the problems of
modern cosmology is the presence of initial singularity, which means
the finiteness of time. The main purpose of introducing a nonlinear
term in the spinor field Lagrangian is to study the possibility of
the elimination of initial singularity. In a number of papers it was
shown that the introduction of nonlinear spinor field into the
system indeed gives rise to singularity-free models of the Universe
(Saha, 2001a, 2001b, 2004a), (Saha and Shikin, 1997a, 1997b).
{\bf (ii) problem of isotropization:} Although the Universe seems
homogenous and isotropic at present, it does not necessarily mean
that it is also suitable for a description of the early stages of
the development of the Universe and there are no observational data
guaranteeing the isotropy in the era prior to the recombination. In
fact, there are theoretical arguments that support the existence of
an anisotropic phase that approaches an isotropic one (Misner 1968).
The observations from Cosmic Background Explorer's differential
radiometer have detected and measured cosmic microwave background
anisotropies in different angular scales. These anisotropies are
supposed to hide in their fold the entire history of cosmic
evolution dating back to the recombination era and are being
considered as indicative of the geometry and the content of the
universe. More about cosmic microwave background anisotropy is
expected to be uncovered by the investigations of microwave
anisotropy probe. There is widespread consensus among the
cosmologists that cosmic microwave background anisotropies in small
angular scales have the key to the formation of discrete structure.
It was found that the introduction of nonlinear spinor field
accelerates the isotropization process of the initially anisotropic
Universe (Saha, 2001a, 2004a, 2006c).
{\bf (iii) late time acceleration of the Universe:} Detection and
further experimental reconfirmation of current cosmic acceleration
pose to cosmology a fundamental task of identifying and revealing
the cause of such phenomenon. This fact can be reconciled with the
theory if one assumes that the Universe id mostly filled with
so-called dark energy. This form of matter (energy) is not
observable in laboratory and it does not interact with
electromagnetic radiation. These facts played decisive role in
naming this object. In contrast to dark matter, dark energy is
uniformly distributed over the space, does not intertwine under the
influence of gravity in all scales and it has a strong negative
pressure of the order of energy density. Based on these properties,
cosmologists have suggested a number of dark energy models, those
are able to explain the current accelerated phase of expansion of
the Universe. In this connection a series of papers appeared
recently in the literature, where a spinor field was considered as
an alternative model for dark energy (Ribas {\it et al}, 2005; Saha,
2006d, 2006e, 2007).
It should be noted that most of the works mentioned above were
carried out within the scope of Bianchi type-I cosmological model.
Results obtained using a spinor field as a source of Bianchi type-I
cosmological field can be summed up as follows: A suitable choice of
spinor field nonlinearity\\
(i) {\it accelerates the isotropization process} (Saha, 2001a,
2004a,
2006c);\\
(ii) {\it gives rise to a singularity-free Universe} (Saha, 2001a,
2004a,
2004b, 2006c);\\
(iii) {\it generates late time acceleration} (Ribas {\it et al},
2005; Saha, 2006d; Souza and Kremer, 2008).
Given the role that spinor field can play in the evolution of the
Universe, question that naturally pops up is, if the spinor field
can redraw the picture of evolution caused by perfect fluid and dark
energy, is it possible to simulate perfect fluid and dark energy by
means of a spinor field? Affirmative answer to this question was
given in the a number of papers (Krechet {\it et al}, 2008; Saha,
2010a, 2010b). In those papers the authors have shown that different
types of perfect fluid and dark energy can be described by nonlinear
spinor field. In (Saha, 2010a) we used two types of nonlinearity,
one occurs as a result of self-action and the other resulted from
the interaction between the spinor and scalar field. It was shown
that the case with induced nonlinearity is the partial one and can
be derived from the case with self-action. In (Saha, 2010b, 2011) we
give the description of generalized Chaplygin gas and modified
quintessence in terms of spinor field and study the evolution of the
Universe filled with nonlinear spinor field within the scope of a
Bianchi type-I and FRW cosmological model. The purpose of this paper
is to extend that study within the framework of other Bianchi
models.
\section{Simulation of perfect fluid with nonlinear spinor field}
Nonlinear quantum Dirac fields were used by Heisenberg (1953, 1957)
in his ambitious unified theory of elementary particles. They are
presently the object of renewed interest since the widely known
paper by Gross and Neveu (1974). A nonlinear spinor field, suggested
by the symmetric coupling between nucleons, muons, and leptons, has
been investigated by Finkelstein {\it et al.} (1951) in the
classical approximation.
In this section we simulate different types of perfect fluid and
dark energy by means of a nonlinear spinor field.
\subsection{Spinor field Lagrangian}
For a spinor field $\p$, the symmetry between $\p$ and $\bp$ appears
to demand that one should choose the symmetrized Lagrangian (Kibble,
1961). Keeping this in mind we choose the spinor field Lagrangian as
(Saha, 2001a):
\begin{equation}
L_{\rm sp} = \frac{i}{2} \biggl[\bp \gamma^{\mu} \nabla_{\mu} \psi-
\nabla_{\mu} \bar \psi \gamma^{\mu} \psi \biggr] - m_{\rm sp}\bp
\psi + F, \label{lspin}
\end{equation}
where the nonlinear term $F$ describes the self-action of a spinor
field and can be presented as some arbitrary functions of invariant
generated from the real bilinear forms of a spinor field. For
simplicity we consider the case when $F = F(S)$ with $S = \bp \psi$.
Here $\nabla_\mu$ is the covariant derivative of spinor field:
\begin{equation}
\nabla_\mu \psi = \frac{\partial \psi}{\partial x^\mu} -\G_\mu \psi,
\quad \nabla_\mu \bp = \frac{\partial \bp}{\partial x^\mu} + \bp
\G_\mu, \label{covder}
\end{equation}
with $\G_\mu$ being the spinor affine connection. In \eqref{lspin}
$\gamma$'s are the Dirac matrices in curve space-time and obey the
following algebra
\begin{equation}
\gamma^\mu \gamma^\nu + \gamma^\nu \gamma^\mu = 2 g^{\mu\nu}
\label{al}
\end{equation}
and are connected with the flat space-time Dirac matrices $\bg$ in
the following way
\begin{equation}
g_{\mu \nu} (x)= e_{\mu}^{a}(x) e_{\nu}^{b}(x) \eta_{ab},
\quad \gamma_\mu(x)= e_{\mu}^{a}(x) \bg_a, \label{dg}
\end{equation}
where $\eta_{ab}= {\rm diag}(1,-1,-1,-1)$ and $e_{\mu}^{a}$ is a set
of tetrad 4-vectors. The spinor affine connection matrices $\G_\mu
(x)$ are uniquely determined up to an additive multiple of the unit
matrix by the equation
\begin{equation}
\nabla_\mu \gamma_\nu = \frac{\pr \gamma_\nu}{\pr x^\mu} -
\G_{\nu\mu}^{\rho}\gamma_\rho - \G_\mu \gamma_\nu + \gamma_\nu
\G_\mu = 0, \label{afsp}
\end{equation}
with the solution
\begin{equation}
\Gamma_\mu = \frac{1}{4} \bg_{a} \gamma^\nu \partial_\mu e^{(a)}_\nu
- \frac{1}{4} \gamma_\rho \gamma^\nu \Gamma^{\rho}_{\mu\nu},
\label{sfc}
\end{equation}
Varying \eqref{lspin} with respect to $\bp (\psi)$ one finds the
spinor field equations:
\begin{subequations}
\label{speq}
\begin{eqnarray}
i\gamma^\mu \nabla_\mu \psi - m_{\rm sp} \psi + \frac{dF}{dS} \psi &=&0, \label{speq1} \\
i \nabla_\mu \bp \gamma^\mu + m_{\rm sp} \bp - \frac{dF}{dS} \bp
&=& 0, \label{speq2}
\end{eqnarray}
\end{subequations}
The energy-momentum tensor of the spinor field is given by
\begin{equation}
T_{\mu}^{\rho}=\frac{i}{4} g^{\rho\nu} \biggl(\bp \gamma_\mu
\nabla_\nu \psi + \bp \gamma_\nu \nabla_\mu \psi - \nabla_\mu \bar
\psi \gamma_\nu \psi - \nabla_\nu \bp \gamma_\mu \psi \biggr) \,-
\delta_{\mu}^{\rho} L_{\rm sp} \label{temsp}
\end{equation}
where $L_{\rm sp}$ in account of spinor field equations
\eqref{speq1} and \eqref{speq2} takes the form
\begin{equation}
L_{\rm sp} = - S \frac{dF}{dS} + F(S). \label{lsp}
\end{equation}
We consider the case when the spinor field depends on $t$ only. In
this case for the components of energy-momentum tensor we find
\begin{subequations}
\begin{eqnarray}
T_0^0 &=& m_{\rm sp}S - F, \label{t00s}\\
T_1^1 = T_2^2 = T_3^3 &=& S \frac{dF}{dS} - F. \label{t11s}
\end{eqnarray}
\end{subequations}
A detailed study of nonlinear spinor field was carried out in Saha
(2001a, 2004a, 2006c). In what follows, exploiting the equation of
states we find the concrete form of $F$ which describes various
types of perfect fluid and dark energy.
\subsection{perfect fluid with a barotropic equation of state}
First of all let us note that one of the simplest and popular model
of the Universe is a homogeneous and isotropic one filled with a
perfect fluid with the energy density $\ve = T_0^0$ and pressure $p
= - T_1^1 = -T_2^2 = -T_3^3$ obeying the barotropic equation of
state
\begin{equation}
p = W \ve, \label{beos}
\end{equation}
where $W$ is a constant. Depending on the value of $W$ \eqref{beos}
describes perfect fluid from phantom to ekpyrotic matter, namely
\begin{subequations}
\label{zeta}
\begin{eqnarray}
W &=& 0, \qquad \qquad \qquad {\rm (dust)},\\
W &=& 1/3, \quad \qquad \qquad{\rm (radiation)},\\
W &\in& (1/3,\,1), \quad \qquad\,\,{\rm (hard\,\,Universe)},\\
W &=& 1, \quad \qquad \quad \qquad {\rm (stiff \,\,matter)},\\
W &\in& (-1/3,\,-1), \quad \,\,\,\,{\rm (quintessence)},\\
W &=& -1, \quad \qquad \quad \quad{\rm (cosmological\,\, constant)},\\
W &<& -1, \quad \qquad \quad \quad{\rm (phantom\,\, matter)},\\
W &>& 1, \quad \qquad \quad \qquad{\rm (ekpyrotic\,\, matter)}.
\end{eqnarray}
\end{subequations}
The barotropic model of perfect fluid is used to study the evolution
of the Universe. Most recently the relation \eqref{beos} is
exploited to generate a quintessence in order to explain the
accelerated expansion of the Universe (Saha, 2005; Zlatev 1999).
In order to describe the matter given by \eqref{zeta} with a spinor
field let us now substitute $\ve$ and $p$ with $T_0^0$ and $-T_1^1$,
respectively. Thus, inserting $\ve = T_0^0$ and $p = - T_1^1$ from
\eqref{t00s} and \eqref{t11s} into \eqref{beos} we find
\begin{equation}
S \frac{dF}{dS} - (1+W)F + m_{\rm sp} W S= 0, \label{eos1s}
\end{equation}
with the solution (Saha, 2010a, 2010b, 2011)
\begin{equation}
F = \lambda S^{1+W} + m_{\rm sp}S. \label{sol1}
\end{equation}
Here $\lambda$ is an integration constant. Taking into account that
the energy density should be non-negative, we conclude that
$\lambda$ is a negative constant, we write the components of the
energy momentum tensor
\begin{subequations}
\begin{eqnarray}
T_0^0 &=& \nu S^{1+W}, \label{t00sf}\\
T_1^1 = T_2^2 = T_3^3 &=& - \nu W S^{1+W}, \label{t11sf}
\end{eqnarray}
\end{subequations}
where $\lambda = - \nu$, with $\nu$ being a positive constant. As
one sees, the energy density $\ve = T_0^0$ is always positive, while
the pressure $p = - T_1^1 = \nu W S^{1+W}$ is positive for $W
> 0$, i.e., for usual fluid and negative for $W < 0$, i.e. for dark
energy.
In account of it the spinor field Lagrangian now reads
\begin{equation}
L_{\rm sp} = \frac{i}{2} \biggl[\bp \gamma^{\mu} \nabla_{\mu} \psi-
\nabla_{\mu} \bar \psi \gamma^{\mu} \psi \biggr] - \nu S^{1+W},
\label{lspin1}
\end{equation}
Thus a massless spinor field with the Lagrangian \eqref{lspin1}
describes perfect fluid from phantom to ekpyrotic matter. Here the
constant of integration $\nu$ can be viewed as constant of
self-coupling. A detailed analysis of this study was given in
Krechet (2008).
\subsection{Chaplygin gas}
An alternative model for the dark energy density was used by
Kamenshchik {\it et al.} (2001), where the authors suggested the use
of some perfect fluid but obeying "exotic" equation of state. This
type of matter is known as {\it Chaplygin gas}. The fate of density
perturbations in a Universe dominated by the Chaplygin gas, which
exhibit negative pressure was studied by Fabris {\it et al.} (2002).
Model with Chaplygin gas was also studied in the Refs. (Dev {\it et
al}, 2003; Bento {\it et al}, 2002).
Let us now generate a Chaplygin gas by means of a spinor field. A
Chaplygin gas is usually described by a equation of state
\begin{equation}
p = -A/\ve^\gamma. \label{chap}
\end{equation}
Then in case of a massless spinor field for $F$ one finds
\begin{equation}
\frac{(-F)^\gamma d(-F)}{(-F)^{1+\gamma} - A} = \frac{dS}{S},
\label{eqq}
\end{equation}
with the solution (Saha, 2010a, 2010b, 2011)
\begin{equation}
-F = \bigl(A + \lambda S^{1+\gamma}\bigr)^{1/(1+\gamma)}.
\label{chapsp}
\end{equation}
On account of this for the components of energy momentum tensor we
find
\begin{subequations}
\begin{eqnarray}
T_0^0 &=& \bigl(A + \lambda S^{1+\gamma}\bigr)^{1/(1+\gamma)}, \label{edchapsp}\\
T_1^1 = T_2^2 = T_3^3 &=& A/\bigl(A + \lambda
S^{1+\gamma}\bigr)^{\gamma/(1+\gamma)}. \label{prchapsp}
\end{eqnarray}
\end{subequations}
As was expected, we again get positive energy density and negative
pressure. Analogical results were obtained in (Cai {\it et al}, 2008
).
Thus the spinor field Lagrangian corresponding to a Chaplygin gas
reads
\begin{equation}
L_{\rm sp} = \frac{i}{2} \biggl[\bp \gamma^{\mu} \nabla_{\mu} \psi-
\nabla_{\mu} \bar \psi \gamma^{\mu} \psi \biggr] - \bigl(A + \lambda
S^{1+\gamma}\bigr)^{1/(1+\gamma)}. \label{lspin2}
\end{equation}
Setting $\gamma = 1$ we find the result obtained in Saha (2010a).
\subsection{Modified quintessence}
Finally, we simulate modified quintessence with a nonlinear spinor
field. It should be noted that one of the problems that face models
with dark energy is that of eternal acceleration. One of the
possible way to avoid this problem is to introduce a negative
$\Lambda$-term together with a quintessence, which gives rise to a
oscillatory mode of expansion (Cardenas {\it et. al.}, 2003, Saha,
2006a) In order to get rid of that problem quintessence with a
modified equation of state was proposed which is given by (Saha,
2006b)
\begin{equation}
p = W (\ve - \ve_{\rm cr}), \quad W \in (-1,\,0), \label{mq}
\end{equation}
Here $\ve_{\rm cr}$ some critical energy density. Setting $\ve_{\rm
cr} = 0$ one obtains ordinary quintessence. It is well known that as
the Universe expands the (dark) energy density decreases. As a
result, being a linear negative function of energy density, the
corresponding pressure begins to increase. In case of an ordinary
quintessence the pressure is always negative, but for a modified
quintessence as soon as $\ve$ becomes less than the critical one,
the pressure becomes positive.
Inserting $\ve = T_0^0$ and $p = - T_1^1$ into \eqref{mq} we find
\begin{equation}
F = - \nu S^{1+W} + m_{\rm sp}S - \frac{W}{1+W}\ve_{\rm cr},
\label{Fmq}
\end{equation}
with $\eta$ being a positive constant. On account of this for the
components of energy momentum tensor we find
\begin{subequations}
\begin{eqnarray}
T_0^0 &=& \nu S^{1+W} + \frac{W}{1+W}\ve_{\rm cr}, \label{edmq}\\
T_1^1 = T_2^2 = T_3^3 &=& - \nu W S^{1+W} + \frac{W}{1+W}\ve_{\rm
cr}. \label{prmq}
\end{eqnarray}
\end{subequations}
Lagrangian for spinor field describing perfect fluid and modified
quintessence can be written in the following way (Saha, 2011)
\begin{equation}
L_{\rm sp} = \frac{i}{2} \biggl[\bp \gamma^{\mu} \nabla_{\mu} \psi-
\nabla_{\mu} \bar \psi \gamma^{\mu} \psi \biggr] - \nu S^{1+W}
-\frac{W}{1+W}\ve_{\rm cr}. \label{lspin2n}
\end{equation}
One can easily verify, in case of $\ve_{\rm cr} = 0$ \eqref{lspin2n}
corresponds to a perfect fluid, while nontrivial $\ve_{\rm cr}$ with
$W \in (-1,\,0)$ generates modified quintessence. It should be noted
that the restriction $W > -1$ is very important, as it does not
allow the system to cross over phantom divide barrier.
We see that a nonlinear spinor field with specific type of
nonlinearity can substitute perfect fluid and dark energy, thus give
rise to a variety of evolution scenario of the Universe.
\section{Cosmological models with a spinor field}
In the previous section we showed that the perfect fluid and the
dark energy can be simulated by a nonlinear spinor field. In the
section II the nonlinearity was the subject to self-action. In
(Saha, 2010a) we have also considered the case when the nonlinearity
was induced by a scalar field. It was also shown the in our context
the results for induced nonlinearity is some special cases those of
self-interaction. Taking it into mind we study the evolution an
Universe filled with a nonlinear spinor field given by the
Lagrangian \eqref{lspin}, with the nonlinear term $F$ is given by
\eqref{sol1}, \eqref{chapsp} and \eqref{Fmq}.
\subsection{Bianchi type anisotropic cosmological model}
Let us study the evolution of an anisotropic Bianchi type
cosmological model filled with spinor field. In this report we
consider Bianchi type-VI, VI$_0$, V, III, I and FRW models.
We choose the Bianchi type-VI cosmological model in the form (Saha,
2004b)
\begin{equation}
ds^2 = dt^2 - a_1^2 e^{-2mz} dx^2 - a_2^2 e^{2nz} dy^2 - a_3^2 dz^2,
\label{bvi}
\end{equation}
with $a_1,\,a_2,\,a_3$ being the function of $t$ only. A suitable
choice of $m,\,n$ as well as the metric functions $a_1,\,a_2,\,a_3$
in the BVI given by \eqref{bvi} evokes the following Bianchi-type
universes:
\begin{itemize}
\item
for $m = n$ the BVI metric transforms to a Bianchi-type VI$_0$
(BVI$_0$) one, i.e., $m = n$, BVI $\Longrightarrow$ BVI$_0$ $\in$
open FRW with the line elements
\begin{equation}
ds^2 = dt^2 - a_1^{2} e^{-2mz}\,dx^{2} - a_2^{2} e^{2mz}\,dy^{2} -
a_3^{2}\,dz^2; \label{bvi0}
\end{equation}
\item
for $m = - n$ the BVI metric transforms to a Bianchi-type V (BV)
one, i.e., $m = n$, BVI $\Longrightarrow$ BV $\in$ open FRW with the
line elements
\begin{equation}
ds^2 = dt^2 - a_1^{2} e^{2mz}\,dx^{2} - a_2^{2} e^{2mz}\,dy^{2} -
a_3^{2}\,dz^2; \label{bv}
\end{equation}
\item
for $n = 0$ the BVI metric transforms to a Bianchi-type III (BIII)
one, i.e., $n = 0$, BVI $\Longrightarrow$ BIII with the line
elements
\begin{equation}
ds^2 = dt^2 - a_1^{2} e^{-2mz}\,dx^{2} - a_2^{2} \,dy^{2} -
a_3^{2}\,dz^2; \label{biii}
\end{equation}
\item
for $m = n = 0$ the BVI metric transforms to a Bianchi-type I (BI)
one, i.e., $m = n = 0$, BVI $\Longrightarrow$ BI with the line
elements
\begin{equation}
ds^2 = dt^2 - a_1^{2} \,dx^{2} - a_2^{2} \,dy^{2} - a_3^{2}\,dz^2;
\label{bi}
\end{equation}
\item
for $m=n=0$ and equal scale factor in all three directions the BVI
metric transforms to a Friedmann-Robertson-Walker (FRW) universe,
i.e., $m = n = 0$ and $a=b=c$, BVI $\Longrightarrow$ FRW with the
line elements.
\begin{equation}
ds^2 = dt^2 - a^{2} \bigl(dx^{2} + \,dy^{2} + \,dz^2\bigr).
\label{frw}
\end{equation}
\end{itemize}
Let us go back to the metric Bianchi-type VI. The metric \eqref{bvi}
possesses following nontrivial covariant and contravariant
components:
\begin{eqnarray}
g_{00} = 1, \quad g_{11} = a_1^2 e^{-2mz}, \quad g_{22} = a_2^2
e^{2nz}, \quad g_{33} = a_3^2, \nonumber\\
\nonumber \\
g^{00} = 1, \quad g^{11} =\frac{1}{a_1^2 e^{-2mz}}, \quad g^{22} =
\frac{1}{a_2^2 e^{2nz}}, \quad g^{33} = \frac{1}{a_3^2}. \nonumber
\end{eqnarray}
The nontrivial Christoffel symbols for \eqref{bvi} are
\begin{eqnarray}
\G_{01}^{1} &=& \frac{\dot{a_1}}{a_1},\quad \G_{02}^{2} =
\frac{\dot{a_2}}{a_2},\quad
\G_{03}^{3} = \frac{\dot{a_3}}{a_3}, \nonumber\\
\G_{11}^{0} &=& a_1 \dot{a_1} e^{-2mz},\quad \G_{22}^{0} = a_2
\dot{a_2} e^{2nz},\quad
\G_{33}^{0} = a_3 \dot{a_3},\label{Chrysvi}\\
\G_{31}^{1} &=& -m,\quad \G_{32}^{2} = n,\quad \G_{11}^{3} = \frac{m
a_1^2}{a_3^2} e^{-2mz},\quad \G_{22}^{3} = -\frac{n a_2^2}{a_3^2}
e^{2nz}. \nonumber
\end{eqnarray}
In view of \eqref{dg} we choose the tetrad as follows:
\begin{equation}
e_0^{(0)} = 1, \quad e_1^{(1)} = a_1 e^{-mz}, \quad e_2^{(2)} = a_2
e^{nz}, \quad e_3^{(3)} = a_3. \label{tetradvi}
\end{equation}
From
\begin{equation} \gamma_\mu = e^{(a)}_\mu \bg_a.
\label{gammabg}
\end{equation}
one now finds
\begin{equation}
\gamma_0 = \bg_0, \quad \gamma_1 = a_1 e^{-mz} \bg_1, \quad
\gamma_2 = a_2 e^{nz} \bg_2, \quad \gamma_3 = a_3 \bg_3.
\label{gbgvi}
\end{equation}
Taking into account that in our case
$$\bg^0 = \bg_0, \quad \bg^1 = -\bg_1, \quad \bg^2 = -\bg_2, \quad
\bg^3 = -\bg_3,$$ one also finds
\begin{equation}
\gamma^0 = \bg^0, \quad \gamma^1 = \frac{e^{mz}}{a_1} \bg^1, \quad
\gamma^2 = \frac{e^{-nz}}{a_2} \bg^2, \quad \gamma^3 = \frac{1}{a_3}
\bg^3. \label{gbgviup}
\end{equation}
Now we are ready to compute spinor affine connections using
\eqref{sfc} which in our particular case gives
\begin{subequations}
\label{sacvi}
\begin{eqnarray}
\Gamma_0 &= & \frac{1}{4} \bg_{a} \gamma^\nu \partial_t e^{(a)}_\nu
- \frac{1}{4} \gamma_\rho \gamma^\nu \Gamma^{\rho}_{0\nu} = 0,\label{G0vi}\\
\Gamma_1 &=& - \frac{1}{4} \gamma_\rho \gamma^\nu
\Gamma^{\rho}_{1\nu} = \frac{1}{2}\Bigl(\dot a_1 \bg^1\bg^0 -
m\frac{a_1}{a_3} \bg^1\bg^3\Bigr) e^{-mz}, \label{G1vi}\\
\Gamma_2 &=& - \frac{1}{4} \gamma_\rho \gamma^\nu
\Gamma^{\rho}_{2\nu} = \frac{1}{2}\Bigl(\dot a_2 \bg^2\bg^0 +
n\frac{a_2}{a_3} \bg^2\bg^3\Bigr) e^{nz}, \label{G2vi}\\
\Gamma_3 &= & \frac{1}{4} \bg_{a} \gamma^\nu \partial_z e^{(a)}_\nu
- \frac{1}{4} \gamma_\rho \gamma^\nu \Gamma^{\rho}_{3\nu} =
\frac{1}{2} \dot a_3 \bg^3\bg^0. \label{G3vi}
\end{eqnarray}
\end{subequations}
From \eqref{sacvi} one finds
\begin{equation}
\gamma^\mu \Gamma_\mu = - \frac{1}{2} \Bigl(\frac{\dot a_1}{a_1} +
\frac{\dot a_2}{a_2} + \frac{\dot a_3}{a_3}\Bigr) \bg^0 +
\frac{m-n}{2a_3} \bg^3. \label{gGvi}
\end{equation}
Choosing $m,\,n$ we can thus write the spinor affine connections for
other Bianchi type metrics.
\subsubsection{Bianchi type-VI anisotropic cosmological model}
Let us now study the evolution of the Universe given by a Bianchi
type-VI cosmological model. The Einstein equations corresponding to
the metric \eqref{bvi} take the form:
\begin{subequations}
\label{einbvi}
\begin{eqnarray}
\frac{\ddot a_2}{a_2} +\frac{\ddot a_3}{a_3} +\frac{\dot
a_2}{a_2}\frac{\dot
a_3}{a_3} - \frac{n^2}{a_3^2} &=& \kappa T_{1}^{1}, \label{11bvi}\\
\frac{\ddot a_3}{a_3} +\frac{\ddot a_1}{a_1} +\frac{\dot
a_3}{a_3}\frac{\dot
a_1}{a_1} - \frac{m^2}{a_3^2} &=& \kappa T_{2}^{2}, \label{22bvi} \\
\frac{\ddot a_1}{a_1} +\frac{\ddot a_2}{a_2} +\frac{\dot
a_1}{a_1}\frac{\dot
a_2}{a_2} + \frac{m n}{a_3^2} &=& \kappa T_{3}^{3}, \label{33bvi}\\
\frac{\dot a_1}{a_1}\frac{\dot a_2}{a_2} +\frac{\dot a_2}{a_2}
\frac{\dot a_3}{a_3} + \frac{\dot a_3}{a_3}\frac{\dot a_1}{a_1} -
\frac{m^2 - m n + n^2}{a_3^2} &=&
\kappa T_{0}^{0}, \label{00bvi}\\
m \frac{\dot a_1}{a_1} - n \frac{\dot a_2}{a_2} - (m - n) \frac{\dot
a_3}{a_3} &=& \kappa T_{3}^{0}. \label{03bvi}
\end{eqnarray}
\end{subequations}
Taking into account that we are dealing with massless spinor field,
from \eqref{t00s} and \eqref{t11s} one finds $T_0^0 = - F$,\,\, $
T_1^1 = T_2^2 = T_3^3 = S \frac{dF}{dS} - F$ and $T_{3}^{0} = 0$.
Then the Eq. \eqref{03bvi} immediately gives
\begin{equation}
\Bigl(\frac{a_1}{a_3}\Bigr)^m = {\cal N}_1
\Bigl(\frac{a_2}{a_3}\Bigr)^n, \quad {\cal N}_1 = {\rm const.}.
\label{abcrel}
\end{equation}
Let us now define
\begin{eqnarray}
V = a_1 a_2 a_3. \label{vdef}
\end{eqnarray}
Summation of \eqref{11bvi}, \eqref{22bvi}, \eqref{33bvi} and three
times \eqref{00bvi} gives
\begin{eqnarray}
\frac{\ddot V}{V} = 2 \frac{m^2 - mn + n^2}{a_3^2} + \frac{3
\kappa}{2} \Bigl[S \frac{dF}{dS} - 2F\Bigr]. \label{vdefeq}
\end{eqnarray}
Before solving this equation let us write $S$ in terms of $v$. The
spinor field equations in this case read
\begin{subequations}
\label{spinv}
\begin{eqnarray}
\bg^0\Bigl(\dot \psi + \frac{\dot V}{2 V} \psi\Bigr)
-\frac{m-n}{2 a_3} \bg^3 \psi - i\frac{dF}{dS} \psi &=& 0, \\
\Bigl(\dot{\bp} + \frac{\dot v}{2 v}\bp\Bigr)\bg^0 -\frac{m-n}{2
a_3} \bp \bg^3 + i\frac{dF}{dS} \bp &=& 0.
\end{eqnarray}
\end{subequations}
From \eqref{spinv} one finds
\begin{equation}
\frac{d (V S)}{d t}= 0, \label{Svid}
\end{equation}
which gives
\begin{equation}
S = \frac{C_0}{V}, \quad C_0 = {\rm const.} \label{Svi}
\end{equation}
Thus we see that the equation for defining $V$ explicitly depends on
$a_3$. Here we assume that the expansion $\theta$ is proportional to
the eigenvalue $\sigma_1^1$ of shear tensor $\sigma_\mu^\nu$. In a
comoving system of reference with $U^\mu = (1,\,0,\,0,\,0)$ for
Bianchi type-VI metric we find
\begin{equation}
\theta = U^{\mu}_{;\mu} = \frac{\dot a_1}{a_1} + \frac{\dot
a_2}{a_2} + \frac{\dot a_3}{a_3}. \label{expan}
\end{equation}
From
\begin{equation}
\sigma_{\mu \nu} = \frac{1}{2}\bigl(U_{\mu;\rho} P^{\rho}_{\nu} +
U_{\nu;\rho} P^{\rho}_{\mu}\bigr) - \frac{1}{3} \theta P_{\mu\nu},
\label{sigma}
\end{equation}
we find
\begin{equation}
\sigma_{1}^{1} = -\frac{1}{3}\bigl(-2\frac{\dot a_1}{a_1} +
\frac{\dot a_2}{a_2} + \frac{\dot a_3}{a_3}\bigr).\label{sigma11}
\end{equation}
As one sees, $S$, $\theta$ and $\sigma_1^1$ (it is true for
$\sigma_2^2$ and $\sigma_3^3$ as well) do not depend on $m$ and $n$.
Hence, they will remain unaltered for other Bianchi models following
from \eqref{bvi}. Now setting
\begin{equation}
\sigma_1^1 = q_1 \theta, \label{prop}
\end{equation}
where $q_1$ is a constant, one finds
\begin{equation}
a_1 = (a_2 a_3)^{{\cal N}_2}, \quad {\cal N}_2 = (1 + 3 q_1)/(2- 3
q_1). \label{abcrel1}
\end{equation}
Inserting this into \eqref{abcrel} and using \eqref{vdef} one
finally finds:
\begin{subequations}
\label{a123vi}
\begin{eqnarray}
a_1 &=& V^{{\cal N}_2/(1 + {\cal N}_2)}, \label{a1vi}\\
a_2 &=& {{\cal N}_1}^{-1/(2n - m)} V^{(m - n -
{\cal N}_2 m)/({\cal N}_2 + 1)(m - 2n)}, \label{a2vi}\\
a_3 &=& {{\cal N}_1}^{1/(2n - m)} V^{({\cal N}_2 m - n)/({\cal N}_2
+ 1)(m - 2n)}. \label{a3vi}
\end{eqnarray}
\end{subequations}
Let us note that perfect fluid satisfying the barotropic equation of
state, as well as Chaplygin gas are described by a massless spinor
field Lagrangian. Moreover, $S$, as well as $F(S)$ are the functions
of $V$. Therefore, we can rewrite the equations for defining $V$ as
\begin{eqnarray}
\ddot V = 2 \frac{m^2 - mn + n^2}{{\cal N}_1^{2/(2n - m)}} V^{q_2}+
{\cal D} (V), \quad {\cal D} (V) = \frac{3 \kappa}{2} \Bigl[S
\frac{dF}{dS} - 2F\Bigr]V, \label{vdefeqvi}
\end{eqnarray}
with
$$q_2 =
\frac{2(n-{\cal N}_2 m)}{(1 + {\cal N}_2)(m - 2n)} + 1 = \frac{1}{3}
- \frac{2q_1 (m + n)}{m - 2n}.$$
In case of spinor field nonlinearity given by \eqref{lspin1}, Eq.
\eqref{vdefeqvi} takes the form
\begin{eqnarray}
\ddot V = 2 \frac{m^2 - mn + n^2}{{\cal N}_1^{2/(2n - m)}} V^{q_2}+
\frac{3 \kappa \nu C_0^{1+W} (1-W)}{2} V^{-W}, \label{vdefeqvi1}
\end{eqnarray}
with the solution in quadrature
\begin{equation}
\int\frac{dV}{\sqrt{q_3 V^{q_2 +1} +3 \kappa \nu C_0^{1+W} V^{1-W} +
C_1}} = t + t_0, \quad q_3 = \frac{4(m^2 - mn + n^2)}{(q_2 +
1){\cal N}_1^{2/(2n - m)}}. \label{Vquadq}
\end{equation}
Here $C_1$ is some integration constant.
Let us consider the case when the spinor field is given by the
Lagrangian \eqref{lspin2}. The equation for $V$ now reads
\begin{equation}
\ddot V = 2 \frac{m^2 - mn + n^2}{{\cal N}_1^{2/(2n - m)}} V^{q_2} +
\frac{3\kappa}{2} \Biggl[ \bigl(AV^{1+\gamma} + \lambda
C_0^{1+\gamma}\bigr)^{1/(1+\gamma)} + \frac{A
V^{1+\gamma}}{\bigl(AV^{1+\gamma} + \lambda
C_0^{1+\gamma}\bigr)^{\gamma/(1+\gamma)}}\Biggr], \label{vdefeqvi2}
\end{equation}
with the solution
\begin{equation}
\int \frac{dV}{\sqrt{C_1 + q_3 V^{q_2 +1} + 3 \kappa V
\bigl(AV^{1+\gamma} + \lambda C_0^{1+\gamma}\bigr)^{1/(1+\gamma)}}}
= t + t_0, \quad C_1 = {\rm const}. \quad t_0 = {\rm const}.
\label{Vquadch}
\end{equation}
Inserting $\gamma = 1$ we come to the result obtained in (Saha,
2005).
Finally we consider the case with modified quintessence. In this
case for $V$ we find
\begin{eqnarray}
\ddot V = 2 \frac{m^2 - mn + n^2}{{\cal N}_1^{2/(2n - m)}} V^{q_2}+
\frac{3\kappa}{2} \Bigl[\nu C_0^{1 + W} (1 - W) V^{-W} + 2W \ve_{\rm
cr}V/(1 + W)\Bigr], \label{vdefeqvi1mq}
\end{eqnarray}
with the solution in quadrature
\begin{equation}
\int\frac{dV}{\sqrt{q_3 V^{q_2 +1} +3 \kappa \bigl[\nu C_0^{1 + W}
V^{1 - W} + W\ve_{\rm cr}V^2/(1 + W)\bigr] + C_1}} = t + t_0.
\label{Vquadmq}
\end{equation}
Recalling that in case of modified quintessence $W \in (-1,\,0)$.
So the model allows cyclic mode of expansion, only when $q_2 < 1$.
\subsubsection{Bianchi type-VI$_0$ anisotropic cosmological model}
Setting $m = n$ from \eqref{bvi} we get Bianchi type-VI$_0$
cosmological model given by \eqref{bvi0}. In this case equation
\eqref{03bvi} gives
\begin{equation}
\frac{\dot a_1}{a_1} - \frac{\dot a_2}{a_2} = 0, \label{03bvi0}
\end{equation}
with the solution
\begin{equation}
a_2 = {\cal N}_1 a_1, \label{a12bvi0}
\end{equation}
Assuming that $\sigma_1^1 \propto \theta$, on account of
\eqref{abcrel1} in this case we find
\begin{subequations}
\label{a123vi0}
\begin{eqnarray}
a_1 &=& V^{{\cal N}_2/(1 + {\cal N}_2)}, \label{a1vi0}\\
a_2 &=& {\cal N}_1 V^{{\cal N}_2/(1 + {\cal N}_2)}, \label{a2vi0}\\
a_3 &=& \frac{1}{{\cal N}_1} V^{(1 - {\cal N}_2)/(1 + {\cal N}_2)}.
\label{a3vi0}
\end{eqnarray}
\end{subequations}
The equation for $V$ in this case reads
\begin{eqnarray}
\ddot V = 2 m^2{\cal N}_1^2 V^{q_2} + {\cal D} (V).
\label{vdefeqvi0}
\end{eqnarray}
In this case for VI$_0$ we find the solution in quadrature analogous
to \eqref{Vquadq}, \eqref{Vquadch} and \eqref{Vquadmq}, for
quintessence, Chaplygin gas and modified quintessence, respectively,
with
$$q_2 = \frac{2({\cal N}_2 - 1)}{{\cal
N}_2 + 1} + 1 = \frac{1}{3} + 4 q_1, \qquad q_3 = \frac{4 m^2 {\cal
N}_1^2}{q_2 + 1} = \frac{3 m^2 {\cal N}_1^2}{1 + 3 q_1}.$$
As we see, in presence of a modified quintessence, BVI$_0$
cosmological model allows cyclic mode of evolution if $q_1 < 1/6$.
In case of $q_1 > 1/6$ we find the model is not bound from above as
is seen from figure \ref{BVI0mq}.
\subsubsection{Bianchi type-V anisotropic cosmological model}
Setting $m = - n$ from \eqref{bvi} we get Bianchi type-V
cosmological model given by \eqref{bv}. In this case equation
\eqref{03bvi} gives
\begin{equation}
\frac{\dot a_1}{a_1} + \frac{\dot a_2}{a_2} - 2\frac{\dot
a_3}{a_3}= 0, \label{03bv}
\end{equation}
with the solution
\begin{equation}
a_1 a_2 = {\cal N}_1 a_3^2, \label{a12bv}
\end{equation}
Assuming that $\sigma_1^1 \propto \theta$, on account of
\eqref{abcrel1} in this case we find
\begin{subequations}
\label{a123v}
\begin{eqnarray}
a_1 &=& V^{{\cal N}_2/(1 + {\cal N}_2)}, \label{a1v}\\
a_2 &=& {\cal N}_1^{1/3} V^{(2 - {\cal N}_2)/3(1 + {\cal N}_2)}, \label{a2v}\\
a_3 &=& \frac{1}{{\cal N}_1^{1/3}} V^{1/3}. \label{a3v}
\end{eqnarray}
\end{subequations}
The equation for $v$ in this case reads
\begin{eqnarray}
\ddot V = 6 m^2 {\cal N}_1^{2/3} V^{q_2} + {\cal D} (V).
\label{vdefeqv}
\end{eqnarray}
In this case for V we find the solution in quadrature analogous to
\eqref{Vquadq}, \eqref{Vquadch} and \eqref{Vquadmq}, for
quintessence, Chaplygin gas and modified quintessence, respectively,
with
$$q_2 = \frac{1}{3}, \qquad q_3 = \frac{12 m^2 {\cal N}_1^{2/3}}{q_2 + 1} = 9 m^2 {\cal N}_1^{2/3}.$$
We can conclude that in presence of modified quintessence BV model
always undergoes a cyclic mode of expansion.
\subsubsection{Bianchi type-III anisotropic cosmological model}
Setting $n = 0$ from \eqref{bvi} we get Bianchi type-VI$_0$
cosmological model given by \eqref{biii}. In this case equation
\eqref{03bvi} gives
\begin{equation}
\frac{\dot a_1}{a_1} - \frac{\dot a_3}{a_3} = 0, \label{03biii}
\end{equation}
with the solution
\begin{equation}
a_3 = {\cal N}_1 a_1, \label{a12biii}
\end{equation}
Assuming that $\sigma_1^1 \propto \theta$, on account of
\eqref{abcrel1} in this case we find
\begin{subequations}
\label{a123biii}
\begin{eqnarray}
a_1 &=& V^{{\cal N}_2/(1 + {\cal N}_2)}, \label{a1biii}\\
a_2 &=& \frac{1}{{\cal N}_1} V^{(1 - {\cal N}_2)/(1 + {\cal N}_2)}, \label{a2biii}\\
a_3 &=& {\cal N}_1 V^{{\cal N}_2/(1 + {\cal N}_2)}. \label{a3biii}
\end{eqnarray}
\end{subequations}
The equation for $V$ in this case reads
\begin{eqnarray}
\ddot V = 2 \frac{m^2}{{\cal N}_1^2} V^{q_2} + {\cal D} (V).
\label{vdefeqbiii}
\end{eqnarray}
In this case for V we find the solution in quadrature analogous to
\eqref{Vquadq}, \eqref{Vquadch} and \eqref{Vquadmq}, for
quintessence, Chaplygin gas and modified quintessence, respectively,
with
$$q_2 = -\frac{2{\cal N}_2}{{\cal
N}_2 + 1} + 1 = \frac{1}{3} - 2 q_1, \qquad q_3 = \frac{4 m^2}{(q_2
+ 1){\cal N}_1^2} = \frac{6 m^2}{(2 - 3 q_1){\cal N}_1^2}.$$
As we see, in presence of a modified quintessence, BVI$_0$
cosmological model allows cyclic mode of evolution if $q_1 > - 1/3$.
\myfigures{BVI-Iq}{0.45}{Evolution of the Universe filled with
quintessence for different Bianchi models}
{0.45}{BVI-III-Imq}{0.45}{Evolution of the Universe filled with
modified quintessence for different Bianchi models}{0.45}
\myfigures{BVI0mq}{0.45}{Evolution of the Universe filled with
modified quintessence for Bianchi type-VI$_0$ model}
{0.45}{BVmq}{0.45}{Evolution of the Universe filled with modified
quintessence for Bianchi type-V model}{0.45}
In the figures \ref{BVI-Iq}, \ref{BVI-III-Imq}, \ref{BVI0mq} and
\ref{BVmq} we have plotted the evolution of $V$ for different
Bianchi models filled with quintessence and modified quintessence.
\subsubsection{Bianchi type-I anisotropic cosmological model}
Bianchi type-I (BI) model is the simplest anisotropic cosmological
model and gives an excellent scope to take into account the initial
anisotropy of the Universe. Given the importance of BI model to
study the effects of initial anisotropy in the evolution of he
Universe, we study this models in details. Unlike the models
considered previously, the system of Einstein's equations for BI
model does not contain off-diagonal component \eqref{03bvi},
explicitly relating metric functions between themselves. The system
of Einstein equations in this case reads
\begin{subequations}
\label{BIE}
\begin{eqnarray}
\frac{\ddot a_2}{a_2} +\frac{\ddot a_3}{a_3} + \frac{\dot
a_2}{a_2}\frac{\dot a_3}{a_3}&=& \kappa T_{1}^{1},\label{11bi}\\
\frac{\ddot a_3}{a_3} +\frac{\ddot a_1}{a_1} + \frac{\dot
a_3}{a_3}\frac{\dot a_1}{a_1}&=& \kappa T_{2}^{2},\label{22bi}\\
\frac{\ddot a_1}{a_1} +\frac{\ddot a_2}{a_2} + \frac{\dot
a_1}{a_1}\frac{\dot a_2}{a_2}&=& \kappa T_{3}^{3},\label{33bi}\\
\frac{\dot a_1}{a_1}\frac{\dot a_2}{a_2} +\frac{\dot
a_2}{a_2}\frac{\dot a_3}{a_3} +\frac{\dot a_3}{a_3}\frac{\dot
a_1}{a_1}&=& \kappa T_{0}^{0}. \label{00bi}
\end{eqnarray}
\end{subequations}
Solving the Einstein equation for the metric functions one finds
(Saha, 2001a)
\begin{eqnarray}
a_i = D_i V^{1/3} \exp{\Bigl(X_i \int \frac{dt}{V}\Bigr)}, \quad
\prod_{i=1}^{3} D_i = 1, \quad \sum_{i=1}^{3} X_i = 0,
\label{metricf}
\end{eqnarray}
with $D_i$ and $X_i$ being the integration constants.
The equation for $v$ in this case takes the form (Saha, 2001a)
\begin{eqnarray}
\ddot V = {\cal D} (V). \label{detvbi}
\end{eqnarray}
In case of \eqref{lspin1} Eq. \eqref{detvbi} takes the form
\begin{equation}
\ddot V = (3/2) \kappa \nu C_0^{1+W} (1-W) V^{-W}
\end{equation}
with the solution in quadrature
\begin{equation}
\int\frac{dV}{\sqrt{3 \kappa \nu C_0^{1+W} V^{1-W} + C_1}} = t +
t_0.
\end{equation}
Here $C_1$ and $t_0$ are the integration constants.
\myfigures{spinpf_pf1}{0.45}{Evolution of the Universe filled with
perfect fluid.} {0.45}{spinpf_de1}{0.45}{Evolution of the Universe
filled with dark energy.}{0.45}
In the Figs. \ref{spinpf_pf1} and \ref{spinpf_de1} we have plotted
the evolution of the Universe defined by the nonlinear spinor field
corresponding to perfect fluid and dark energy (Saha, 2010b).
Let us consider the case when the spinor field is given by the
Lagrangian \eqref{lspin2}. The equation for $V$ now reads
\begin{equation}
\ddot V = (3/2) \kappa \Biggl[ \bigl(AV^{1+\gamma} + \lambda
C_0^{1+\gamma}\bigr)^{1/(1+\gamma)} + A
V^{1+\gamma}/\bigl(AV^{1+\gamma} + \lambda
C_0^{1+\gamma}\bigr)^{\gamma/(1+\gamma)}\Biggr],
\end{equation}
with the solution
\begin{equation}
\int \frac{dV}{\sqrt{C_1 + 3 \kappa V \bigl(AV^{1+\gamma} + \lambda
C_0^{1+\gamma}\bigr)^{1/(1+\gamma)}}} = t + t_0, \quad C_1 = {\rm
const}. \quad t_0 = {\rm const}.
\end{equation}
Inserting $\gamma = 1$ we come to the result obtained in (Saha,
2005).
Finally we consider the case with modified quintessence. In this
case for $V$ we find
\begin{equation}
\ddot V = (3/2) \kappa \Bigl[\eta C_0^{1-W} (1+W) V^{W} - 2W
\ve_{\rm cr}V/(1-W)\Bigr],
\end{equation}
with the solution in quadrature
\begin{equation}
\int \frac{dV}{\sqrt{3 \kappa \bigl[\eta C_0^{1-W} V^{1+W} -
W\ve_{\rm cr}V^2/(1-W)\bigr] + C_1}} = t + t_0. \label{qdmq}
\end{equation}
Here $C_1$ and $t_0$ are the integration constants. Comparing
\eqref{qdmq} with those with a negative $\Lambda$-term we see that
$\ve_{\rm cr}$ plays the role of a negative cosmological constant.
\myfigures{spinpf_mqep1}{0.45}{Dynamics of energy density and
pressure for a modified quintessence.}
{0.45}{spinpf_mq1}{0.45}{Evolution of the Universe filled with a
modified quintessence.}{0.45}
In the Fig. \ref{spinpf_mqep1} we have illustrated the dynamics of
energy density and pressure of a modified quintessence. In the Fig.
\ref{spinpf_mq1} the evolution of the Universe defined by the
nonlinear spinor field corresponding to a modified quintessence has
been presented. As one sees, in the case considered, acceleration
alternates with declaration. In this case the Universe can be either
singular (that ends in Big Crunch) or regular.
\subsubsection{FRW cosmological models with a spinor field}
Since our Universe is almost isotropic at large scale it would be
fitting to study evolution of the FRW model within the scope of
spinor description of matter. The Einstein equations read
corresponding to the FRW model \eqref{frw} reads
\begin{subequations}
\label{EFRW}
\begin{eqnarray}
2 \frac{\ddot a}{a} + \frac{\dot a^2}{a^2}&=& \kappa T_{1}^{1}
\label{FRW11}\\
3\frac{\dot a^2}{a^2}&=& \kappa T_{0}^{0}. \label{FRW00}
\end{eqnarray}
\end{subequations}
From the spinor field equations in this case we find
\begin{equation}
S = \frac{C_0}{a^3}, \quad C_0 = {\rm const.} \label{SFRW}
\end{equation}
In order to find the solution that satisfies both \eqref{FRW11} and
\eqref{FRW00} we rewrite \eqref{FRW11} in view of \eqref{FRW00} in
the following form:
\begin{equation}
\ddot a = \frac{\kappa}{6}\Bigl(3 T_1^1 - T_0^0\Bigr) a. \label{dda}
\end{equation}
Further we solve this equation for concrete choice of source field.
Let us consider the case of perfect fluid given by the barotropic
equation of state. In account of \eqref{t00sf}, \eqref{t11sf} and
\eqref{SFRW}, \eqref{dda} takes the form
\begin{equation}
\ddot a = \frac{\kappa \nu (1+3W)C_0^{1+W}}{2} a^{-(2+3W)},
\label{ddasf}
\end{equation}
that admits the first integral
\begin{equation}
\dot a^2 = \frac{\kappa}{3} \nu C_0^{1+W} a^{-(1+3W)} + E_1, \quad
E_1 = {\rm const}. \label{dda1}
\end{equation}
In Fig. \ref{spinpf_pf_FRW} and \ref{spinpf_phan_FRW} we plot the
evolution of the FRW Universe for different values of $W$.
\myfigures{spinpf_pf_FRW}{0.45}{Evolution of the Universe filled
with perfect fluid and dark energy.}
{0.45}{spinpf_phan_FRW}{0.45}{Evolution of the Universe filled with
phantom matter.}{0.45}
As one sees, equation \eqref{dda1} imposes no restriction on the
value of $W$. But it is not the case, when one solves \eqref{FRW00}.
Indeed, inserting $T_0^0$ from \eqref{t00sf} into \eqref{FRW00} one
finds
\begin{equation}
a = (A_1 t + C_1)^{2/3(1+W)}, \label{afrw}
\end{equation}
where $A_1 = (1+W)\sqrt{3 \kappa \nu C_0^{1+W}/4}$ and $C_1 = 3 (1 +
W) C/2$ with $C$ being some arbitrary constant. This solution
identically satisfies the equation \eqref{FRW11}. As one sees, case
with $W = -1$, cannot be realized here. In that case one has to
solve the equation \eqref{FRW00} straight forward. As far as phantom
matter ($W < -1$) is concerned, there occurs some restriction on the
value of $C$, as in this case $A_1$ is negative and for the $C_1$ to
be positive, $C$ should be negative. As one can easily verify, in
case of cosmological constant with $W = -1$ Eqn. \eqref{FRW00} gives
\begin{equation}
a = a_0 e^{\pm \sqrt{\kappa \nu/3}\,t}. \label{FRWlambda}
\end{equation}
Inserting \eqref{edchapsp} and \eqref{prchapsp} into \eqref{dda} in
case of Chaplygin gas we have the following equation
\begin{equation}
\ddot a = \frac{\kappa}{6}\frac{2A a^{3(1+\gamma)} - \lambda
C_0^{1+\gamma}}{a^2\Bigl(Aa^{3(1+\gamma)} + \lambda
C_0^{1+\gamma}\Bigr)^{\gamma/(1+\gamma)}}. \label{FRWchap}
\end{equation}
We solve this equation numerically. The corresponding solution has
been illustrated in Fig. \ref{spinpf_pf_FRW}.
Finally we consider the case with modified quintessence. Inserting
\eqref{edmq} and \eqref{prmq} into \eqref{dda} in this case we find
\begin{equation}
\ddot a = \frac{\kappa}{6}\Bigr[(3W-1)\eta C_0^{1-W} a^{3W-2} -
\frac{2W}{1-W}\ve_{\rm cr} a\Bigr], \label{FRWmq}
\end{equation}
with he solution
\begin{equation}
\dot a^2 = \frac{\kappa}{3}\Bigl[\nu C_0^{1-W} a^{3W - 1} -
\frac{W}{1-W} \ve_{\rm cr} a^2 + E_2\Bigr], \quad E_2 = {\rm const}.
\label{dda2}
\end{equation}
It can be shown that in case of modified quintessence the pressure
is sign alternating. As a result we have a cyclic mode of evolution.
\section{Singularity problem}
On the the main problems of modern cosmology is the singularity
problem. Let us study this problem within the scope of models
discussed above. As we see, the components of the spinor field and
metric functions are expressed in terms of $v$. It should be noted
that $v$ plays one of the central roles in studying the singular
space-time points of BI cosmological models. Here we describe it in
brief. In doing so let us first write the Kretschmann scalar for the
Bianchi type-VI metric. Taking into account that the metric
\eqref{bvi} possesses the following non-trivial components of
Riemann tensor:
\begin{eqnarray}
R_{\,\,\,01}^{01} &=& -\frac{\ddot a_1}{a_1}, \quad
R_{\,\,\,02}^{02} = -\frac{\ddot a_2}{a_2}, \quad
R_{\,\,\,03}^{03} = -\frac{\ddot a_3}{a_3}, \nonumber \\
R_{\,\,\,12}^{12} &=& -\frac{mn}{a_3^2} - \frac{\dot
a_1}{a_1}\frac{\dot a_2}{a_2},\quad R_{\,\,\,13}^{13} =
\frac{m^2}{a_3^2} - \frac{\dot a_3}{a_3}\frac{\dot a_1}{a_1}, \quad
R_{\,\,\,23}^{23} = \frac{n^2}{a_3^2} - \frac{\dot
a_2}{a_2}\frac{\dot a_3}{a_3}, \nonumber\\
R_{\,\,\,10}^{31} &=& \frac{m}{a_3^2} \Bigl(\frac{\dot a_1}{a_1} -
\frac{\dot a_3}{a_3} \Bigr),\quad R_{\,\,\,13}^{01} = m
\Bigl(\frac{\dot a_3}{a_3} - \frac{\dot a_1}{a_1} \Bigr),\nonumber\\
R_{\,\,\,20}^{32} &=& \frac{n}{a_3^2} \Bigl(\frac{\dot a_3}{a_3} -
\frac{\dot a_2}{a_2} \Bigr),\quad R_{\,\,\,23}^{02} = n
\Bigl(\frac{\dot a_2}{a_2} - \frac{\dot a_3}{a_3} \Bigr), \nonumber
\end{eqnarray}
for the Kretschmann scalar we find
\begin{eqnarray}
{\cal K} &=& R_{\mu\nu\alpha\beta} R^{\mu\nu\alpha\beta} =
R^{\mu\nu}_{\,\,\,\,\,\,\,\alpha\beta} R_{\,\,\,\,\,\,\,\mu\nu}^{\alpha\beta} \nonumber \\
& = & 4\Bigl[\Bigl(\frac{\ddot a_1}{a_1}\Bigr)^2 + \Bigl(\frac{\ddot
a_2}{a_2}\Bigr)^2 + \Bigl(\frac{\ddot a_3}{a_3}\Bigr)^2 +
\Bigl(\frac{\dot a_1}{a_1}\frac{\dot a_2}{a_2}\Bigr)^2 +
\Bigl(\frac{\dot a_2}{a_2}\frac{\dot a_3}{a_3}\Bigr)^2 +
\Bigl(\frac{\dot a_3}{a_3}\frac{\dot a_1}{a_1}\Bigr)^2\Bigr]
\label{Kretsch}\\
& + & \frac{4}{a_3^4}\Bigl[(m^4 + m^2n^2 + n^4) - (m^2 + n^2){\dot
a_3}^2 - a_3^2 \Bigl(m \frac{\dot a_1}{a_1} - n \frac{\dot
a_2}{a_2}\Bigr)^2\Bigr]. \nonumber
\end{eqnarray}
The metric functions for BVI, BVI$_0$, BV and BIII can be expressed
as
\begin{equation}
a_i = {\cal B}_i V^{s_i}, \label{Vgen}
\end{equation}
which gives
\begin{subequations}
\label{singvis}
\begin{eqnarray}
\frac{\dot a_i}{a_i} &=& {\cal B}_i \frac{\dot{V}}{V},\\
\frac{\ddot a_i}{a_i} &=& {\cal B}_i \frac{\ddot{V}}{V} + {\cal B}_i
({\cal B}_i -1) \Bigl(\frac{\dot{V}}{V}\Bigr)^2.
\end{eqnarray}
\end{subequations}
The metric functions and their derivatives for the BI take the
following form:
\begin{subequations}
\label{singb1}
\begin{eqnarray}
\frac{\dot a_i}{a_i} &=& \frac{\dot{V}}{3 V} + \frac{X_i}{3V},\\
\frac{\ddot a_i}{a_i} &=& \frac{\ddot{V}}{3V} - \frac{2}{9}
\Bigl(\frac{\dot{V}}{V}\Bigr)^2 - \frac{X_i}{9}\frac{ \dot{V}}{V^2}
+ \frac{X_i^2}{9 V^2}.
\end{eqnarray}
\end{subequations}
We study the singularity on the basis of Kretschmann scalar and in
doing so we follow the criteria given in (Bronnikov {\it et al},
2004):
(i) For any finite $t$ some $a_i \to 0$. \eqref{Kretsch} shows that
if more than one scale factor becomes trivial at finite $t$, then it
is a singularity. As is seen from \eqref{singb1} as $V \to 0$ (i)
all $a_i \to 0$ if $X_1 = X_2 = X_3 = 0$, i.e., it is a singularity;
(ii) more than one $a_i \to 0$ if more than one $X_i < 0$ and in
this case we have singularity; (iii) only one $a_i \to 0$ if only
one $X_i < 0$, i.e., the space-time can be non-singular.
Note that this criteria of singularity always fulfills at the point
where $V = 0$.
As far as $t \to \infty$ is concerned, the corresponding asymptote
can be singular if at least one $a_i$ vanishes faster than
exponentially. As $X_1 + X_2 + X_3 = 0$, it means one or more $X_i$
is negative, and from \eqref{singvis} it follows that at least one
function $a_i$ vanishes faster than exponentially. Hence it is a
singularity.
As far as BVI, BVI$_0$, BV, BIII and FRW models are concerned, they
are all singular at a space-time point, where $V = 0$. Moreover, in
cases of BVI, BVI$_0$, BV and BIII at least two scale factor $a_i$'s
are directly related, hence there is always a possibility of more
than one one scale factor becoming trivial, thus giving rise to a
singularity at finite $t$.
Moreover, components of the spinor field as well as the physically
observable quantities such as charge, current, spin etc. constructed
from them are inverse functions of $V$ (Saha, 2001a). Hence we
conclude that within the scope of the model considered here at any
point where $V = 0$ there occurs a space-time singularity.
\section{Problem of isotropization}
Since the present-day Universe is surprisingly isotropic, it is
important to see whether our anisotropic BI model evolves into an
isotropic FRW model. Isotropization means that at large physical
times $t$, when the volume factor $V$ tends to infinity, the three
scale factors $a_i(t)$ grow at the same rate. Two wide-spread
definition of isotropization read
\begin{subequations}
\label{aniso}
\begin{eqnarray}
{\cal A} &=& \frac{1}{3} \sum\limits_{i=1}^{3} \frac{H_i^2}{H^2} - 1 \to 0,\\
\Sigma^2 &=& \frac{1}{2} {\cal A} H^2 \to 0.
\end{eqnarray}
\end{subequations}
Here ${\cal A}$ and $\Sigma^2$ are the average anisototropy and
shear, respectively. $H_i = \dot{a_i}/a_i$ is the directional Hubble
parameter and $H = \dot{a}/a$ average Hubble parameter, where $a(t)
= V^{1/3}$ is the average scale factor. Here we exploit the
isotropization condition proposed proposed in Bronnikov {\it et al}
(2004):
\begin{equation}
\frac{a_i}{a}\Bigl|_{t \to \infty} \to {\rm const.} \label{isocon}
\end{equation}
Then by rescaling some of the coordinates, we can make $a_i/a \to
1$, and the metric will become manifestly isotropic at large $t$. It
can be shown that in case of BVI, BVI$_0$, BV and BIII models the
criteria \eqref{isocon} does not hold, hence the isotropization
process for these models does not take place. So we consider the BI
model and study the problem od isotropization for this model in
details.
From \eqref{metricf} we find
\begin{eqnarray}
\frac{a_i}{a} = \frac{a_i}{V^{1/3}} = D_i \exp{\Bigl(X_i \int
\frac{dt}{\tau}\Bigr)}.\label{isocon1}
\end{eqnarray}
As is seen from \eqref{metricf} in our case $a_i / a \to D_i =$
const as $v \to \infty$. Recall that the isotropic FRW model has
same scale factor in all three directions, i.e., $a_1(t) = a_2(t) =
a_3(t) = a(t)$. So for the BI universe to evolve into a FRW one the
constants $D_i$'s are likely to be identical, i.e., $D_1 = D_2 = D_3
= 1$. Moreover, the isotropic nature of the present Universe leads
to the fact that the three other constants $X_i$ should be close to
zero as well, i.e., $|X_i| << 1$, ($i = 1,2,3$), so that $X_i \int
[v (t)]^{-1}dt \to 0$ for $t < \infty$ (for $V (t) = t^n$ with $n
> 1$ the integral tends to zero as $t \to \infty$ for any $X_i$). It
can be concluded that the spinor field Lagrangian with $W < 1$ leads
to the isotropization of the Universe as $t \to \infty$, moreover,
in case of $W < 0$ the system undergoes an earlier isotropization.
\section{Discussion}
Let us now examine what kind of advantage one gets exploiting the
spinor description of matter. We do it within the scope of a BI
cosmological model. In doing so, we consider the case when the
Universe is filled with, say, Van-der-Waals fluid, radiation and
quintessence. In this case we have
\begin{subequations}
\label{3comp}
\begin{eqnarray}
T_0^0 &=& \ve_{\rm v} + \ve_{\rm r} + \ve_{\rm q},\\
T_1^1 &=& - p_{\rm v} - p_{\rm r} - p_{\rm q}.
\end{eqnarray}
\end{subequations}
To solve \eqref{vdefeqvi} one has to know $T_0^0$ and $T_1^1$ in
terms of $v$. It can be done exploiting Bianchi identity
\begin{equation}
\dot T_0^0 + \frac{\dot V}{V} \Bigl(T_0^0 - T_1^1\Bigr) = 0.
\label{bianid}
\end{equation}
The dark energy is supposed to interact with itself only, so it is
minimally coupled to the gravitational field. As a result, the
evolution equation for the energy density decouples from that of the
perfect fluid. Taking this into account and inserting \eqref{3comp}
into \eqref{bianid}, we obtain two balance equations
\begin{subequations}
\label{2eq}
\begin{eqnarray}
\dot \ve_{\rm v} + \frac{\dot V}{V} \Bigl(\ve_{\rm v} + p_{\rm
v}\Bigr) + \dot \ve_{\rm r} + \frac{\dot V}{V} \Bigl(\ve_{\rm
r} + p_{\rm r}\Bigr) &=& 0, \label{2eq1}\\
\dot \ve_{\rm q} + \frac{\dot V}{V} \Bigl(\ve_{\rm q} + p_{\rm
q}\Bigr) &=& 0. \label{2eq2}
\end{eqnarray}
\end{subequations}
In usual approach we are in trouble, as neither $\ve_{\rm v}$ nor
$\ve_{\rm r}$ cannot be expressed in terms of $v$ from \eqref{2eq1}.
But if one uses spinor description for radiation, thanks to spinor
field equation one finds
\begin{equation}
\dot \ve_{\rm r} + \frac{\dot V}{V} \Bigl(\ve_{\rm r} + p_{\rm
r}\Bigr) \equiv 0. \label{rad}
\end{equation}
As a result, in place of \eqref{2eq1} we now have
\begin{equation}
\dot \ve_{\rm v} + \frac{\dot V}{V} \Bigl(\ve_{\rm v} + p_{\rm
v}\Bigr) = 0, \label{2eq1n}
\end{equation}
which is quite computable for the given equation of state. For
Van-der-Waals fluid we use the following equation of state
\begin{equation}
p_{\rm v} = \frac{8 W_1 \ve_{\rm v}}{3 - \ve_{\rm v}} - 3 \ve_{\rm
v}^2, \label{vwf}
\end{equation}
where $W_1$ is a constant. The dark energy density $\ve_{\rm q}$ can
be expressed in terms of $V$ either solving \eqref{2eq2} or using
the spinor description. Thus using the spinor description
\eqref{t00sf} and \eqref{t11sf} for radiation and dark energy, and
defining the Hubble parameter, we find the following system of
equations
\begin{subequations}
\label{three}
\begin{eqnarray}
\dot V &=& 3 H V, \label{tau1}\\
\dot H &=& -3H^2 + \frac{\kappa}{2} \Bigl[\ve_{\rm v} -
\frac{8W_1\ve_{\rm v}}{3-\ve_{\rm v}} + 3 \ve_{\rm v}^2 + \frac{2
\ve_{\rm r0}}{3}V^{-4/3} + (1-W) \ve_{\rm q0}V^{-1 -W}\Bigr],
\label{H}\\
\dot \ve_{\rm v} &=& - 3 \Bigl[\ve_{\rm v} + \frac{8W_1 \ve_{\rm
v}}{3 - \ve_{\rm v}} - 3 \ve_{\rm v}^2\Bigr] H. \label{vwf1}
\end{eqnarray}
\end{subequations}
For simplicity, we set $\ve_{\rm r0} = 1$ and $\ve_{\rm q0} = 1$.
The equation \eqref{vwf1} can be exactly solved for $W_1 = 1/2$. In
this case we find
\begin{equation}
\frac{\ve_{\rm v}^6 (3 \ve_{\rm v} - 7)}{(\ve_{\rm v} - 1)^7} =
\frac{C_0}{V^{14}}, \quad C_0 = const.\label{vwr}
\end{equation}
As one sees, it is quite a complicate expression. In what follows,
we solve the system \eqref{three} numerically. For this purpose we
also set $W = -1/2$ for quintessence.
\myfigures{VEPH1}{0.45}{Evolution of the $V$, $H$ and $\ve_{\rm v}$
and $p$ for $V (0) = 1$, $H (0) = 0.0000945$ and $\ve_{\rm v} =
0.957$}{0.45}{VEPH2}{0.45}{Evolution of the $V$, $H$, $\ve_{\rm v}$
and $p$ for $V (0) = 1$, $H (0) = -0.9$ and $\ve_{\rm v} =
0.7$}{0.45}
In the figures \ref{VEPH1} and \ref{VEPH2} we plot the evolution of
$V$, $H$ and $\ve_{\rm v}$ and $p$ for different initial values. As
one sees, the pressure has negative value at the initial stage of
evolution, then it begin to increase, thus giving rise to
deceleration and finally becomes negative that results in the late
time acceleration of the Universe.
\section{Conclusion}
Within the framework of cosmological gravitational field equivalence
between the perfect fluid (and dark energy) and nonlinear spinor
field has been established. It is shown that different types of dark
energy can be simulated by means of a nonlinear spinor field. Using
the new description of perfect fluid or dark energy evolution of the
Universe has been studied within the scope of a BVI, BVI$_0$, BV,
BIII and BI anisotropic models as well as isotropic FRW model. The
corresponding Einstein equations have been solved. It is shown that
all the models give rise to a space-time singularity where $V$ is
trivial. Among the Bianchi models considered only the BI allows
isotropization of the initially anisotropic space-time. It is shown
that the spinor description of fluid allows one to solve the
two-fluid system without without imposing any additional condition.
|
\section{Introduction}
Analytic results for one-scale multiloop Feynman integrals in a Laurent expansion
in $\epsilon=(4-d)/2$
are expressed as linear combinations of transcendental constants with rational
coefficients. The set of these constants essentially depends on the type of
Feynman integrals. Probably, the simplest type of one-scale Feynman integrals
are massless propagator integrals depending on one external momentum.
Here the world record is set at four loops --- see Ref.~\cite{Baikov:2010hf}
where all the corresponding master integrals were analytically evaluated in
an epsilon expansion up to transcendentality weight seven.
Practical calculations show that only multiple zeta values (MZV) (see, e.g., \cite{BBV})
appear in results.
Brown proved \cite{Brown:2008um} that convergent scalar massless planar propagator
diagrams with the degree of divergence $\omega\equiv 4 h-2 L=-2$ (where $h$ and $L$ are
numbers of loops and edges, correspondingly) up to five loops contain only MZV
in their epsilon expansions. (In two loops, a proof was earlier presented in Ref.~\cite{BW}.)
He also proved that
for the three non-planar diagrams depicted in Fig.~\ref{Fig:integrals}, every coefficient in a
Taylor expansion in $\epsilon$ is a rational linear combination of
MZV and Goncharov's polylogarithms \cite{Goncharov} with
sixth roots of unity as arguments.
\begin{figure}
\centering
\includegraphics[width=7cm]{Fig1.eps}\\
\caption{Diagrams considered in Ref.~\cite{Brown:2008um} as possible candidates for the presence of sixth
roots of unity in the $\epsilon$-expansion.}\label{Fig:integrals}
\end{figure}
The goal of this brief communication is to study these diagrams experimentally. We present
results in an epsilon expansion up to transcendentality weight twelve. To do this we
apply the DRA method recently suggested by one of the authors (R.L.), Ref. \cite{Lee:2009dh}.
The method is based on the use of dimensional recurrence relations (DRR) \cite{Tarasov1996} and analytic
properties of Feynman integrals as functions of the parameter of
dimensional regularization, $d$, and was already successfully applied in previous calculations
\cite{Lee:2010cga,Lee:2010ug,Lee:2010wea,Lee:2010hs,Lee:2010ik}.
To apply this method it is essential to perform an integration by parts (IBP)
\cite{IBP} reduction of
integrals that participate in dimensional recurrence relations to master integrals.
To do this, we use the {\tt C++} version of the code {\tt FIRE} \cite{FIRE}.
To study analytic properties of solutions of dimensional recurrence relations,
i.e. to reveal the position and the order of poles in $d$
in a basic stripe, we used a sector decomposition \cite{BH,BognerWeinzierl,FIESTA}
implemented in the code {\tt FIESTA} \cite{FIESTA,FIESTA2}.
To fix remaining constants in the homogenous solutions of dimensional
recurrence relations it was quite sufficient for us to use analytic
results for the four-loop massless propagators master integrals \cite{Baikov:2010hf}
(confirmed numerically by {\tt FIESTA} \cite{Smirnov:2010hd}).
Finally, after obtaining results for master integrals in terms of multiple series
we calculated resulting coefficients at powers of $\epsilon$ numerically with a high
precision and then applied the PSLQ algorithm \cite{PSLQ}.
We also applied the code {\tt HPL} \cite{Maitre:2005uu}, a Mathematica code dealing with harmonic polylogarithms \cite{RV:1999ew}
and MZV.
For all the three diagrams of Fig.~1, we performed evaluation up to transcendentality
weight twelve where the basis of transcendental numbers we used includes 48 constants.
Coefficients in our results turn out to be cumbersome, in particular, such are
the coefficients at $\pi^{12}$ in subsequent formulae. Therefore, to achieve
a successful calculation by PSLQ, we were forced to perform
numerical calculations with a high accuracy. In fact, the accuracy of 800 digits
was enough to obtain results by PSLQ. However, to be on the safe side, we checked
our results by numerical evaluation with the accuracy of 1500 digits.
The first diagram of Fig.~1 is a master diagram. This is nothing but $M_{45}$ in Fig.~2 of
Ref.~\cite{Baikov:2010hf} where all the master integrals for four-loop massless
propagators are shown. Following the method of \cite{Lee:2009dh} we needed, first to
calculate lower master integrals $ M_{01}, M_{11}, M_{13}, M_{14}, M_{21}, M_{27}, M_{35}$.
Eventually, we arrived at the following result which is made homogeneously transcendental
by pulling out an appropriate rational function of $\epsilon$ and normalizing it at
the fourth power of the one-loop integral (i.e. $M_{31}$ according to the notation of \cite{Baikov:2010hf}):
\begin{align}
&\frac{M_{45}(4-2\epsilon)}{\epsilon^4 M_{31}(4-2\epsilon)}
=
\frac{(1-2 \epsilon )^3}{1-6 \epsilon}
\Biggl\{
36 \zeta _3^2-\biggl(-\frac{6}{5} \pi ^4 \zeta _3+378 \zeta _7\biggr) \epsilon
\nonumber\\
&+\biggl(-\frac{427 \pi ^8}{1500}+2844 \zeta _3 \zeta _5+\frac{3024 \zeta _{5,3}}{5}\biggr)\epsilon ^2-\biggl(-\frac{22}{3} \pi ^6 \zeta _3
\nonumber\\
&+732 \zeta _3^3+3 \pi ^4 \zeta _5+\frac{42458 \zeta _9}{3}\biggr) \epsilon ^3
+\biggl(-\frac{60329 \pi ^{10}}{24948}-\frac{183}{5} \pi ^4 \zeta _3^2
\nonumber\\
&+62403 \zeta _5^2+149895 \zeta _3 \zeta _7-\frac{58563 \zeta _{8,2}}{2}\biggr)\epsilon ^4
-\biggl(-\frac{19817}{500} \pi ^8 \zeta _3
\nonumber\\
&+\frac{29578 \pi ^6 \zeta _5}{315}+101952 \zeta _3^2 \zeta _5-\frac{50761 \pi ^4 \zeta _7}{50}-325152 \pi ^2 \zeta _9
\nonumber\\
&+\frac{73041423 \zeta _{11}}{20}+\frac{175392}{5} \zeta _3 \zeta _{5,3}-\frac{216768}{5} \zeta _{5,3,3}\biggr)\epsilon ^5
\nonumber\\
&+\biggl(-\frac{98988919597 \pi ^{12}}{17027010000}-\frac{19242}{35} \pi ^6 \zeta _3^2+\frac{37216 \zeta _3^4}{3}
\nonumber\\
&-\frac{10280}{9} \pi ^4 \zeta _3 \zeta _5+\frac{210112}{3} \pi ^2 \zeta _5^2+\frac{600320}{3} \pi ^2 \zeta _3 \zeta _7
\nonumber\\
&+2100799 \zeta _5 \zeta _7+\frac{9455470 \zeta _3 \zeta _9}{9}+\frac{160424}{75} \pi ^4 \zeta _{5,3}
\nonumber\\
&+60032 \pi ^2 \zeta _{7,3}-\frac{1132952 \zeta _{9,3}}{3}-120064 \zeta _{6,4,1,1}\biggr)\epsilon ^6
\nonumber\\
&+O\left(\epsilon ^7\right)
\Biggr\}\,.
\end{align}
Here $\zeta_{m_1,\dots,m_k}$ are MZV given by
\begin{equation}\label{MZVdef}
\zeta(m_1,\dots,m_k)=
\sum\limits_{i_1=1}^\infty\sum\limits_1^{i_1-1}
\dots\sum\limits_1^{i_{k-1}-1}\prod\limits_{j=1}^k\frac{\mbox{sgn}(m_j)^{i_j}}{i_j^{|m_j|}}\,.
\end{equation}
The other two diagrams of Fig.~1 {\em are not} master integrals and, consequently,
they do not appear in Fig.~2 of Ref.~\cite{Baikov:2010hf}. We found their reduction to lower
master integrals by FIRE and then evaluated resulting master integrals by our technique.
Let us stress that the master integrals involved in the IBP reduction of
the second and the third non-planar diagrams of Fig.~\ref{Fig:integrals} are all {\em planar}
diagrams. However, the result of Brown \cite{Brown:2008um} quoted in the beginning of
our paper is not applicable to such planar diagrams.
(Two of them are $M_{35}$ and $M_{36}$ in Fig.~2 of Ref.~\cite{Baikov:2010hf}, with
$\omega=0$, rather than $-2$.) So, {\em a priori},
one could admit the presence of something additional to MZV in their epsilon expansions.
These are our results for them in the same normalization where we again managed to reveal
homogenous transcendentality:
\begin{align}
&\frac{M_{4B2}(4-2\epsilon)}{\epsilon^4 M_{31}(4-2\epsilon)}
=
\frac{(1-2 \epsilon )^3}{1
-\epsilon }
\Biggl\{
36 \zeta _3^2+\biggl(\frac{6 \pi ^4 \zeta _3}{5}+\frac{189 \zeta _7}{2}\biggr) \epsilon
\nonumber\\
&+\biggl(\frac{1359 \pi ^8}{7000}+144 \zeta _3 \zeta _5-\frac{2916 \zeta _{5,3}}{5}\biggr)\epsilon ^2+\biggl(\frac{4 \pi ^6 \zeta _3}{21}-1392 \zeta _3^3
\nonumber\\
&+51 \pi ^4 \zeta _5+\frac{25549 \zeta _9}{6}\biggr) \epsilon ^3
+\biggl(\frac{146255 \pi ^{10}}{99792}-\frac{348}{5} \pi ^4 \zeta _3^2
\nonumber\\
&-\frac{75843 \zeta _5^2}{4}-\frac{206655 \zeta _3 \zeta _7}{4}+\frac{152163 \zeta _{8,2}}{8}\biggr)\epsilon ^4
\nonumber\\
&+\biggl(-\frac{313039 \pi ^8 \zeta _3}{31500}+\frac{25982 \pi ^6 \zeta _5}{315}-39372 \zeta _3^2 \zeta _5+\frac{234339 \pi ^4 \zeta _7}{200}
\nonumber\\
&+29412 \pi ^2 \zeta _9-\frac{26995129 \zeta _{11}}{160}+\frac{71928}{5} \zeta _3 \zeta _{5,3}+\frac{19608}{5} \zeta _{5,3,3}\biggr)\epsilon ^5
\nonumber\\
&+\biggl(\frac{107930288857 \pi ^{12}}{68108040000}-\frac{42968}{315} \pi ^6 \zeta _3^2+\frac{85696 \zeta _3^4}{3}
\nonumber\\
&-\frac{23510}{9} \pi ^4 \zeta _3 \zeta _5+\frac{31822}{3} \pi ^2 \zeta _5^2+\frac{90920}{3} \pi ^2 \zeta _3 \zeta _7+165244 \zeta _5 \zeta _7
\nonumber\\
&-\frac{6321395 \zeta _3 \zeta _9}{9}+\frac{45614}{75} \pi ^4 \zeta _{5,3}+9092 \pi ^2 \zeta _{7,3}-\frac{803569 \zeta _{9,3}}{6}
\nonumber\\
&-18184 \zeta _{6,4,1,1}\biggr)\epsilon ^6+O\left(\epsilon ^7\right)
\Biggr\}\,,
\end{align}
\begin{align}
&\frac{M_{4B3}(4-2\epsilon)}{\epsilon^4 M_{31}(4-2\epsilon)}
=
\frac{(1-2 \epsilon )^3}{1+4 \epsilon}
\Biggl\{
36 \zeta _3^2+\biggl(\frac{6 \pi ^4 \zeta _3}{5}+567 \zeta _7\biggr) \epsilon
\nonumber\\
&+ \biggl(\frac{211 \pi ^8}{1750}+3744 \zeta _3 \zeta _5+\frac{1944 \zeta _{5,3}}{5}\biggr)\epsilon ^2+\biggl(\frac{68 \pi ^6 \zeta _3}{7}+288 \zeta _3^3
\nonumber\\
&+30 \pi ^4 \zeta _5+22094 \zeta _9\biggr) \epsilon ^3
+ \biggl(\frac{1255 \pi ^{10}}{4158}+\frac{72}{5} \pi ^4 \zeta _3^2
\nonumber\\
&+51318 \zeta _5^2+95490 \zeta _3 \zeta _7-11799 \zeta _{8,2}\biggr)\epsilon ^4
+ \biggl(\frac{3283 \pi ^8 \zeta _3}{125}
\nonumber\\
&+\frac{12484 \pi ^6 \zeta _5}{105}-65952 \zeta _3^2 \zeta _5+\frac{17538 \pi ^4 \zeta _7}{25}+65232 \pi ^2 \zeta _9
\nonumber\\
&-\frac{360909 \zeta _{11}}{10}-\frac{54432}{5} \zeta _3 \zeta _{5,3}+\frac{43488}{5} \zeta _{5,3,3}\biggr)\epsilon ^5
\nonumber\\
&+\biggl(\frac{2972813873 \pi ^{12}}{1064188125}-\frac{15056}{105} \pi ^6 \zeta _3^2-29608 \zeta _3^4-\frac{3640}{3} \pi ^4 \zeta _3 \zeta _5
\nonumber\\
&-8176 \pi ^2 \zeta _5^2-23360 \pi ^2 \zeta _3 \zeta _7+1091724 \zeta _5 \zeta _7+\frac{4198640 \zeta _3 \zeta _9}{3}
\nonumber
\end{align
\begin{align}
&-\frac{6752}{25} \pi ^4 \zeta _{5,3}-7008 \pi ^2 \zeta _{7,3}+95936 \zeta _{9,3}+14016 \zeta _{6,4,1,1}\biggr)\epsilon ^6
\nonumber\\
&+O\left(\epsilon ^7\right)
\Biggr\}\,.
\end{align}
We see that only MZV are present in our results.
Although Goncharov's polylogarithms at sixth roots of unity were allowed
to appear according to the analysis of Ref.~\cite{Brown:2008um} they have not
appeared. In fact, such transcendental numbers do appear in epsilon expansions
of some classes of {\em massive} Feynman integrals \cite{B99,MK1,MK2,MK3}.
Taking our results into account one could try to prove that
that there are only MZV in massless propagator diagrams.
Unfortunately, the way Brown proceeded ~\cite{Brown:2008um} in the case of
a few concrete
diagrams can hardly be generalized to other situations because his analysis
was based on the possibility to perform recursive integrations in Feynman parameters,
as long as the denominator of the integrand is linear with respect to them, so that
some other approach is needed. On the other hand, attempts to discover new types of
transcendental numbers, in addition to MZV, meet more and more difficulties because we
have to turn to more loops and more terms of $\epsilon$-expansions.
Still this is certainly possible at the current level of our possibilities so that
we are planning to do this in the nearest future.
\vspace{0.2 cm}
{\em Acknowledgments.}
This work was supported by the Russian Foundation for Basic Research through grant
11-02-01196 and by DFG through SFB/TR~9 ``Computational Particle Physics''.
The work of R.L. was also supported through Federal special-purpose program
``Scientific and scientific-pedagogical personnel of innovative Russia''. R.L. gratefully acknowledges
Karlsruhe Institut f\"{u}r Theoretische Teilchenphysik for warm hospitality and
financial support during his visit. We are grateful to David Broadhurst and Mikhail Kalmykov for
instructive discussions.
|
\section{#1}}
\newcommand{\PRD}[3]{Phys. Rev. {\bf D#1}, #2 (#3)}
\newcommand{\PLB}[3]{Phys. Lett. {\bf B#1}, #2 (#3)}
\newcommand{\PRL}[3]{Phys. Rev. Lett. {\bf#1}, #2 (#3)}
\newcommand{\NPB}[3]{Nucl. Phys. {\bf B#1}, #2 (#3)}
\newcommand{\begin{equation}}[1] {\begin{equation}\label{#1} }
\newcommand{\end{equation}} {\end{equation} }
\newcommand{\begin{equation}}[1]{\begin{eqnarray}\label{#1} }
\newcommand{\end{equation}}{\end{eqnarray}}
\newcommand{\,{\rm GeV}}{\,{\rm GeV}}
\newcommand{{\vec{n}}}{{\vec{n}}}
\newcommand{{\vec{m}}}{{\vec{m}}}
\newcommand{\sigma}{\sigma}
\newcommand{{\hat\mu}}{{\hat\mu}}
\newcommand{{\hat\nu}}{{\hat\nu}}
\newcommand{{\hat\rho}}{{\hat\rho}}
\newcommand{{\hat{h}}}{{\hat{h}}}
\newcommand{{\hat{g}}}{{\hat{g}}}
\newcommand{{\hat\kappa}}{{\hat\kappa}}
\def{\bf F}{{\bf F}}
\def{\bf A}{{\bf A}}
\def{\bf J}{{\bf J}}
\def{\bf \alpha}{{\bf \alpha}}
\def\begin{eqnarray}{\begin{eqnarray}}
\def\end{eqnarray}{\end{eqnarray}}
\def\baselineskip = .30in{\baselineskip = .30in}
\def\begin{equation}{\begin{equation}}
\def\end{equation}{\end{equation}}
\def\begin{equation}{\begin{equation}}
\def\end{equation}{\end{equation}}
\def\partial{\partial}
\def\nabla{\nabla}
\def\alpha{\alpha}
\def\beta{\beta}
\def\Gamma{\Gamma}
\def\gamma{\gamma}
\def\delta{\delta}
\def\Delta{\Delta}
\def\dagger{\dagger}
\def\kappa{\kappa}
\def\sigma{\sigma}
\def\Sigma{\Sigma}
\def\theta{\theta}
\def\Lambda{\Lambda}
\def\lambda{\lambda}
\def\Omega{\Omega}
\def\omega{\omega}
\def\epsilon{\epsilon}
\def\nonumber{\nonumber}
\def\sqrt{\sqrt}
\def\sqrt{G}{\sqrt{G}}
\def\supset{\supset}
\def\subset{\subset}
\def\left ({\left (}
\def\right ){\right )}
\def\left [{\left [}
\def\right ]{\right ]}
\def\frac{\frac}
\def\label{\label}
\def\hspace{\hspace}
\def\vspace{\vspace}
\def\infty{\infty}
\def\rangle{\rangle}
\def\langle{\langle}
\def\overline{\overline}
\def\tilde{\tilde}
\def\times{\times}
\def\leftrightarrow{\leftrightarrow}
\def\overrightarrow{\overrightarrow}
\renewcommand{\baselinestretch}{1.0}
\title{Recent Progress in SUSY GUTs}
\ShortTitle{SUSY GUTs}
\author{\speaker{K.S. Babu}\thanks{Work is supported in part by the US Department of Energy Grants
DE-FG02-04ER41306 and DE-FG02-ER46140; OSU-HEP-11-05.}\\
Department of Physics, Oklahoma State University, Stillwater, OK 74078, USA\\
E-mail: \email{<EMAIL>}}
\abstract{After a brief review of the motivations for grand unification, I discuss the main
challenges facing realistic SUSY GUT model building. Achieving doublet--triplet splitting without
fine--tuning is chief among them. Symmetry breaking should occur consistently
without unwanted Goldston bosons, $\mu$ term of order TeV for the MSSM Higgs fileds should emerge
naturally, and realistic fermion masses with small quark mixing angles and large lepton mixing angles should be
generated with some predictivity. Significant progress has been made over the years
towards achieving these goals in the context of supersymmetric $SO(10)$ GUT. A complete
$SO(10)$ model is presented along this line wherein, somewhat surprisingly,
the GUT scale threshold corrections to the gauge couplings are found to be small.
This results in a predictive scenario for proton lifetime. An interesting correlation between the
$d=6$ ($p\to e^+\pi^0$) and $d=5$ ($p\to \overline{\nu }K^+$) decay amplitudes is observed.
This class of models predicts that both proton decay modes should be observable with an improvement
in the current sensitivity by about a factor of five to ten.}
\FullConference{35th International Conference of High Energy Physics - ICHEP2010,\\
July 22-28, 2010\\
Paris France}
\begin{document}
\section*{}
\vspace*{-0.5in}
Motivations for unifying the strong, weak and electromagnetic forces are manyfold \cite{ps,gg,georgi}. The experimental
observation that electric charges are quantized ($|Q_{\rm proton}| = |Q_{\rm electron}|$ to better than 1 part in $10^{21}$)
has a natural explanation in grand unified theories (GUT) owing to their non--Abelian nature.
The miraculous cancelation of chiral anomalies that occurs among each family of
the SM fermions has a symmetry--based
explanation in GUTs. $SO(10)$ GUT, for example, is automatically free of such anomalies \cite{georgi}. GUTs provide a natural understanding of the quantum
numbers of quarks and leptons. This point is worth emphasizing further. All quarks and leptons of a family,
including the right--handed neutrino ($\nu^c$) needed for generating small neutrino masses via the seesaw mechanism, are organized into
a {\bf 16}--dimensional spinor representation of $SO(10)$, as shown in Table 1. The gauge symmetry $SO(10)$ contains five independent
internal spins, denoted as $+$ or $-$ signs (for spin--up and spin--down) in Table 1. Subject to the condition that the number of
down spins must be even, there are 16 combinations, which form the irreducible ${\bf 16}$ dimensional spinor of $SO(10)$.
The first three spins denote color charges, while the last two are weak charges. There are three independent combinations of color spins,
identified as the color degrees of freedom ($r,b,g$). Going top down in each column of Table 1, one sees
that in addition, there is a fourth color, identified as lepton number \cite{ps}. Thus quarks and leptons
are unified under the GUT symmetry. The first and the third columns (and similarly the second and the fourth) are left--right conjugates.
Thus $SO(10)$ contains Parity as part of the gauge symmetry. Furthermore, the same {\bf 16} multiplet unifies quarks with anti-quarks, and leptons
with anti-leptons. In fact, $SO(10)$ symmetry is the maximal gauge symmetry that is chiral with sixteen particles (members of one
family). Hypercharge (and thus electric charge) of each fermion follows from the formula $Y = \frac{1}{3} \Sigma (C) - \frac{1}{2} \Sigma (W)$,
where $\Sigma (C)$ is the summation of color spins (first three entries) and $\Sigma (W)$ is the sum of weak spins (last two entries). Thus
$Y$ for the $e^c$ field is $Y(e^c) = \frac{1}{3} (3) - \frac{1}{2} (-2) = 2$. Note that $Y$ (and thus $Q$) must be quantized.
Such a simple organization of matter is remarkably beautiful and can be argued as a strong hint for GUTs. SUSY GUTs have further
empirical support
from the observed unification of gauge couplings at a high energy scale $M_X \approx 10^{16}$ GeV. In Fig. 1, left panel, we demonstrate this unity
of forces in a fully realistic $SO(10)$ SUSY GUT \cite{bpt}. Another remarkable feature of $SO(10)$ GUTs is that the small neutrino masses inferred from
neutrino oscillation data suggest the scale of new physics ($\nu^c$ mass scale) to be $M_{\nu^c} \sim 10^{14}$ GeV, which is close to $M_X$.
$M_{\nu^c}$ is inferred from the effective neutrino mass operator ${\cal L}_{\rm \nu} = LLH_u H_u/M_{\nu^c}$ ($L$ is the lepton doublet and $H_u$
is the Higgs doublet), using $m_\nu \sim 0.05$ eV and $\left\langle H_u \right\rangle \sim 246$ GeV as inputs. In a class of $SO(10)$ models discussed further here, $M_{\nu^c} \sim M_X^2/M_{\rm Pl} \sim 10^{14}$ GeV quite naturally \cite{bpw}. The decay of $\nu^c$ can elegantly explain the observed baryon asymmetry of the universe via leptogenesis. Finally, as exemplified later, the unification of quarks and leptons into GUT multiplets can be quite powerful in realizing predictive
frameworks for fermion masses, perhaps in association with flavor symmetries.
\begin{table}[h]
{\footnotesize
\begin{center}
\begin{tabular}{||c|c|c|c||}\hline\hline
$u_r:~\{-++~+-\}$ & $d_r:~ \{-++~-+\}$ & $u^c_r:~\{+--~++\}$ & $d^c_r:~ \{+--~--\}$ \\
$u_b:~\{+-+~+-\}$ & $d_b:~ \{+-+~-+\}$ & $u^c_b:~\{-+-~++\}$ & $d^c_b:~ \{-+-~--\}$ \\
$u_g:~\{++-~+-\}$ & $d_g:~\{++-~-+\}$ & $u^c_g:~\{--+~++\}$ & $d^c_g:~ \{--+~--\}$ \\
$~\nu:~\{---~+-\}$ & $~e:~~ \{---~-+\}$ & $~\nu^c:~\{+++~++\}$ & $~e^c:~ \{+++~--\}$ \\ \hline\hline
\end{tabular}
\end{center}
\vspace*{-0.2in}
\caption{Quantum numbers of quarks and leptons. The first 3 signs refer to color charge, and the last three to weak charge.
To obtain hypercharge, use $Y = \frac{1}{3}\Sigma(C)-\frac{1}{2}\Sigma (W)$.}
}
\end{table}
While extremely well motivated, constructing fully realistic SUSY GUTs is not so trivial. Chief among the challenges is
the so--called doublet--triplet (DT) splitting problem. Even in the simplest of GUTs, based on $SU(5)$ gauge symmetry \cite{gg},
the smallest irreducible representation is 5 dimensional, which contains the SM Higgs doublet. This means that the
Higgs doublet will be accompanied by a GUT partner, which is a color triplet scalar ($5=2+3$ is the relevant
math). This state must have a mass of order
$10^{16}$ GeV, or else it would lead to rapid proton decay. The DT splitting challenge is to naturally make the doublet
component of the ${\bf 5}$--plet light (of order $10^2$ GeV), while maintaining its color triplet partner superheavy.
In minimal SUSY $SU(5)$, this is done by extreme fine--tuning, of order one part in $10^{14}$. The relevant superpotential
is $W = \overline{5}_H (\lambda 24_H + M)5_H$, where the $(\overline{5}_H, \,5_H$) contain the MSSM Higgs doublets ($H_d,\,H_u$),
and where $\left\langle 24_H \right\rangle = V. \,{\rm diag}(1,\,1,\,1,\,-3/2,\,-3/2)$ breaks the gauge symmetry down to that of
the SM. The masses of the color triplet and $SU(2)_L$ doublet fields are then $m_T = \lambda V + M$, $m_D = \lambda V - (3/2)\,M$.
One chooses $\lambda V$ and $M$ to be both of order $10^{16}$ GeV, but with the condition $\lambda V = (3/2)\,M
+ {\cal O} (10^2)$ GeV, so that the doublet remains light. Such a severe fine--tuning raises questions about the naturalness of the
model.
In SUSY $SO(10)$ the situation for DT splitting is much better. The adjoint Higgs $A(45)$ in $SO(10)$ can acquire a vacuum expectation value
(VEV) $\langle A\rangle ={\rm i}\sigma_2\otimes {\rm Diag}\left ( a, ~a,~a,~0,~0\right )$. The coupling $H(10)A(45)H'(10)$ of two
$10$--plets would result in heavy color triplets with massless $SU(2)_L$ doublets -- without fine--tuning \cite{dw}. There are
a variety of issues that need to be addressed. Symmetry breaking must be complete without unwanted pseudo--Goldstone bosons,
the VEV of the adjoint should be stable against higher dimensional operators, unification of gauge couplings should be maintained,
and the theory should be consistent with proton
lifetime limits. All these issues have been successfully addressed recently \cite{bpt} by making use a set of small dimensional
Higgs fields, as shown in Table 2. An anomalous ${\cal U}(1)_A$ symmetry of possible string origin and a $Z_2$ symmetry are also used, for
stabilizing the doublet mass. In Table 2, $k$ is a positive integer, which will be taken to be 5.
\begin{table}[h]
\vspace{-0.5cm}
\label{t:U-Z-charges} $$\begin{array}{|c||c|c|c|c|c|c|c|c|c|c|c|}
\hline
\vspace{-0.4cm}
& & & & & & & & & &&\\
\vspace{-0.5cm}
& ~A(45)~& ~H(10)~&~ H'(10)~& ~C(16) ~& ~ \bar C(\overline{16})~&~Z ~&~ S ~ & ~C'(16) ~ & ~\bar C'(\overline{16}) & 16_{1,2}&16_3\\
& & & & & & & & & &&\\
\hline
\vspace{-0.5cm}
& & & & & & & & & &&\\
~Q~& 0 &1 &-1 &{(k+4)}/{2k} & -{1}/{2} &{2}/{k} & {2}/{k} & {(k-4)}/{2k} & -{(k+8)}/{2k} &q_{1,2}&-{1}/{2}\\
\vspace{-0.5cm}
& & & & & & & & & &&\\
\hline
\vspace{-0.5cm}
& & & & & & & & &&&\\
\vspace{-0.5cm}
~\omega ~& 1 & 0 & 1 & 0 & 0 & 1 & 0 & 0 & 0 &P_{1,2}&0\\
& & & & & & & & & &&\\
\hline
\end{array}$$
\vspace*{-0.2in}
\caption{${\cal U}(1)_A$ and $Z_2$ charges $Q_i$ and $\omega_i$ of the superfield $\phi_i$.}
\end{table}
\noindent The superpotential consistent
with the symmetries is
{\small
\begin{eqnarray}
W &=& M_A{\rm tr}A^2+\frac{\lambda_A}{M_*}\left ( {\rm tr}A^2\right )^2+\frac{\lambda_A'}{M_*} {\rm tr}A^4
+ C\left ( \frac{a_1}{M_*}ZA+\frac{b_1}{M_*}C\bar C+
c_1S\right ) \bar C' +C'\left ( \frac{a_2}{M_*}ZA+\frac{b_2}{M_*}C\bar C+c_2S\right ) \bar C
\label{sup-ACCbar}\nonumber \\
&+& \lambda_1HAH'+\left ( \lambda_{H'}SZ^{k-1}+ {\lambda'}_{H'}Z^k\right ) \frac{(H')^2}{M_*^{k-1}}
+\lambda_2H\bar C\bar C +\frac{\lambda_3}{M_*}AH'CC'~.
\label{W-H}
\end{eqnarray}
}
This superpotential consistently breaks the $SO(10)$ gauge symmetry down to the SM symmetry in the SUSY limit without generating
unwanted Goldstone bosons. The first three terms of $W$ induce the VEV for $A$, the fourth and fifth terms guarantee absence of
Goldstones \cite{br}, and the remaining terms achieve doublet--triplet splitting without fine--tuning \cite{dw}. At the minimum we have $\left \langle C \right
\rangle = \left \langle \bar{C} \right \rangle = c$, $\left \langle C' \right
\rangle = \left \langle \bar{C}' \right \rangle = 0$, with $c$ determined by the Fayet--Iliopoulos term of ${\cal U}(1)_A$ symmetry. The
Higgs doublet mass is zero in the SUSY limit to all orders. Once SYSY breaking is turned on, the VEV of the $A(45)$ no longer has the zeros, which
are modified to be entries of order $m_{\rm SUSY}$. This in turn induces $\mu$ term of order $m_{\rm SUSY}$ for the MSSM Higgs fields.
In Fig. 1 (left panel), the evolution of the three gauge couplings is displayed, which takes into account the threshold corrections of
the model. Interestingly, the threshold corrections in the model in the $10+ \bar{10}$ sector (of the $SU(5)$ subgroup) cancel between
the matter fields and the gauge boson fields. Owing to this cancelation, the model becomes very predictive for proton lifetime. We find
a correlation between the $d=6$ gauge boson mediated $p \rightarrow e^+ \pi^0$ and the $d=5$ Higgsino--mediated $p \rightarrow \overline{\nu} K^+$
decay amplitudes: $M_{\rm eff}\simeq 10^{19}{\rm GeV} \cdot \left ( \frac{10^{16}{\rm GeV}}{M_X}\right )^3
\left ( \frac{1/100}{r}\right ) \left ( \frac{3}{\tan \beta }\right ) $. $M_X$ controls $ p \rightarrow e^+ \pi^0$, while $M_{\rm eff}$ controls
$p \rightarrow \overline{\nu}K^+$. This is plotted in Fig. 1 (right panel) for varying $r=M_\Sigma/M_X$ ($\Sigma$ is a color
octet Higgs field). Also plotted are the current experimental limits from these decays. One concludes that both modes
should be observable with an improved sensitivity of about five to ten.
\begin{figure}
\centering
\includegraphics[scale=0.35]{spectrum4Cp4unification.eps}
\hspace{0.5cm}
\includegraphics[scale=0.37]{smallRcorel4C-Feb22.eps}
\caption{Left panel: Gauge coupling evolution including threshold corrections. Right panel:
Correlations between $M_{\rm eff}$ and $M_X$ for $m_{\tilde{q}} = 1.5$ TeV, $m_{\tilde{W}} = 130$ GeV,
and $\alpha_3(M_Z)= 0.1176$. {\bf (a)}: $r=1/200$. {\bf (b)}: $r=1/250$. {\bf (c)}: $r=1/300$. The vertical
and horizontal dashed lines correspond to the experimentally allowed lowest values of $M_X$ and $M_{\rm eff}$
which arise from limits on $\Gamma^{-1}(p\to e^+\pi^0)$ and $\Gamma^{-1}(p\to \bar{\nu }K^+)$.}
\end{figure}
Realistic and predictive fermion masses and mixings can be obtained within this framework by assuming a flavor $Q_4$ symmetry, under
which the first two families of $16$ form a doublet. The mass matrices for up and down quarks, charged leptons and Majorana
neutrinos have the form \cite{bpt,bpw}:
\begin{eqnarray}
M_u = m_U^0\left(\begin{matrix}
0 & \epsilon' & 0 \cr -\epsilon' & 0 & \sigma \cr 0 & \sigma & 1
\end{matrix}\right)\hspace{-0.1cm},\,M_{d,e} = m_D^0\left(\begin{matrix}
0 & \kappa_{d,e}\epsilon'+\eta' & 0 \cr -\kappa_{d,e}\epsilon'-\eta' & \kappa_{d,e}\xi_{22}^d & \sigma+\kappa_{d,e}\epsilon \cr 0 & \sigma +
\kappa_{d,e}\bar{\epsilon}& 1
\end{matrix}\right)\hspace{-0.1cm},\,M_R = m_R^0\left(\begin{matrix}
b & 0 & 0 \cr 0 & b & a \cr 0 & a & 1
\end{matrix}\right)\,\,\,
\end{eqnarray}
where $\kappa_{d}=1$ and $\kappa_e=3$. There are fewer parameters than observables in this setup, which results in predictions. A consistent
fit for all masses and mixing parameters is obtained with the choice
$\sigma =0.0508, \epsilon =-0.0188+0.0333i, \overline{\epsilon }=0.106+0.0754i ,
\epsilon'=1.56\cdot 10^{-4}, \eta'=-0.00474+0.00177i,
\xi_{22}^d=0.014 e^{4.1i}$ at the GUT scale. Along with central values of charged lepton masses, we obtain for the quarks,
$m_u(2~{\rm GeV})=3.55~{\rm MeV},~m_c(m_c)=1.15~{\rm GeV},\,
m_d(2~{\rm GeV})=6.45~{\rm MeV}$, $m_s(2~{\rm GeV})=137.6 {\rm MeV},~m_b(m_b) = 4.67~{\rm GeV}$.
For the CKM mixings we obtain:
$|V_{us}|=0.225$, $|V_{cb}|=0.0414~,|V_{ub}|=0.0034~,|V_{td}|=0.00878~,
\overline{\eta }=0.334~,\overline{\rho }=0.12$, and thereby $\sin 2\beta =0.663$. All these are in a good agreement with experiments.
For neutrinos, the Dirac mass matrix is obtained from $M_u$ by replacing $\epsilon'\to -3\epsilon'$.
With $\theta_{12}\simeq 30^o$ and $\theta_{23}\simeq 43^o$ as inputs, we obtain $m_2/m_3\simeq 0.13$
and $\theta_{13}\simeq 3.6^o$ as predictions.
Such a fit is realized by choosing $a=0.0252e^{-0.018i}$, $b=1.61\cdot 10^{-6}e^{-1.592i}$,
and $M_0=1.89\cdot 10^{13}$~GeV. One sees broad, although not precise, agreement with
data. The model succeeds in obtaining large neutrino mixings along with small quark mixings.
|
\section{Introduction} \label{sec:intro}
The discovery of a strong correlation between the stellar bulge mass and the
central super-massive black hole (SMBH) mass of galaxies
\citep[e.g.,][]{1998AJ....115.2285M} has led to detailed theoretical models in
which the growth of SMBHs and their host galaxies occur (nearly) simultaneously
during a brief period of intense, merger-driven activity
\citep[e.g.,][]{2006ApJS..163....1H}. In these models, the nature of the
connection between SMBHs and their host galaxies has important implications for
the evolution of massive galaxies.
The observational foundation of this evolutionary link between SMBHs and their
host galaxies was established by studies of ultra-luminous infrared galaxies (ULIRGs)
identified in the local universe using
{\it InfraRed Astronomical Satellite} \citep[see, e.g.,][]{1984ApJ...278L...1N,1996ARA&A..34..749S} data. ULIRGs
are systems whose spectral energy distributions (SEDs) are dominated by dust
emission at infrared (IR) wavelengths \citep{1986ApJ...303L..41S} and whose
morphologies tend to show evidence for recent or on-going major merger activity
that has been linked to the formation of active galactic nuclei (AGN) and
quasars \citep{1988ApJ...325...74S}. Although ULIRGs in the local universe are
too rare to contribute significantly to the bolometric luminosity density,
recent studies with the {\it Spitzer Space Telescope} have shown that they
become increasingly important at higher redshifts
\citep[e.g.][]{2005ApJ...632..169L,2009A&A...496...57M}. To understand the
physical mechanisms that drive massive galaxy evolution, it is essential to
identify and study high-redshift ($z > 1$), dusty, luminous galaxies that show
signs of concurrent AGN and starburst activity.
Efforts to identify high-redshift ULIRGs have been increasingly fruitful over the
last two decades. In particular, blank-field surveys at
sub-millimeter or millimeter wavelengths have identified dusty and rapidly star-forming galaxies, the so-called sub-millimeter
galaxies \citep[SMGs; e.g.][]{1997ApJ...490L...5S,2006MNRAS.372.1621C}. More
recently, the advent of the Multiband Imaging Photometer for Spitzer
\citep[MIPS;][]{2004ApJS..154...25R} on board the {\it Spitzer Space Telescope}
has allowed for the identification of sources which are bright at mid-IR
wavelengths but faint in the optical
\citep[e.g.][]{2004ApJS..154...60Y,2008ApJ...672...94F,2008ApJ...677..943D,2009ApJ...692..422L}.
Follow-up spectroscopy and clustering measurements of both the
sub-millimeter-selected and the {\it Spitzer}-selected populations has
demonstrated that they have similar number densities, redshift distributions,
and clustering properties that indicate they are undergoing an extremely
luminous, short-lived phase of stellar bulge and nuclear black hole growth and
may be the progenitors of the most luminous ($\sim$4$L^*$) present-day galaxies
\citep{2004ApJ...611..725B,2005ApJ...622..772C,2007ApJ...658..778Y,2006ApJ...641L..17F,2008ApJ...677..943D,2008ApJ...687L..65B}.
One intriguing difference between the the ULIRG samples selected at different wavelengths (as might be expected
based on the selection criteria) is that the mid-IR selected ULIRGs have hotter
dust than the far-IR selected SMGs
\citep{2006ApJ...650..592K,2008MNRAS.384.1597C,2008ApJ...683..659S,2009MNRAS.394.1685Y,2009ApJ...692..422L,2009ApJ...705..184B,2009A&A...508..117F}.
This distinction may be analogous to the warm-dust/cool-dust dichotomy seen in
local ULIRGs, where it has been suggested that warm ULIRGs represent an
important transition stage between cold ULIRGs and quasars
\citep{1988ApJ...328L..35S}. Furthermore, the mid-IR-selected population shows a range of spectral energy distributions (SEDs), with the brighter sources showing power-law SEDs in the mid-IR (``power-law'' sources), and the fainter ones exhibiting peaks at rest-frame wavelengths near 1.6$\mu$m (the ``bump'' sources). The bump is generally attributed to starlight and 1.2~mm photometry suggests that the ``bump'' sources are dominated by cooler dust than the power-law sources \citep{2005ApJ...632L..13L,2008ApJ...683..659S,2009MNRAS.394.1685Y,2009ApJ...692..422L,2009ApJ...705..184B,2009A&A...508..117F}.
\citep{2005ApJ...632L..13L,2008ApJ...683..659S,2009MNRAS.394.1685Y,2009ApJ...692..422L,2009ApJ...705..184B,2009A&A...508..117F}.
Also, the mid-IR spectra of bump sources show strong polycyclic aromatic
hydrocarbon (PAH) emission features typical of star-forming regions
\citep{2007ApJ...658..778Y,2009ApJ...700.1190D,2009ApJ...700..183H}, while
power-law sources have silicate absorption features or are dominated by
continuum emission consistent with obscured AGN
\citep{2005ApJ...622L.105H,2006ApJ...651..101W,2007ApJ...658..778Y}. These
results suggest that there may be a connection between the power source responsible for the
bolometric luminosity of a system and its globally averaged dust temperature.
Efforts to understand this connection between mid-IR and far-IR selected
high-$z$ ULIRGs within the context of an evolutionary paradigm have been recently
advanced by numerical simulations of galaxy mergers
\citep[e.g.][]{1996ApJ...464..641M,2009arXiv0910.2234N}. In these models, when
the merging system approaches final coalescence, the star-formation rate (SFR)
spikes and, because it is enshrouded in cold-dust, the system is observed as an
SMG. As time proceeds, feedback from the growth of a central super-massive
black hole warms the ambient dust and ultimately quenches star-formation. It
is during this critical period of galaxy evolution when the system is
observable as a {\it Spitzer}-selected ULIRG. The models predict observable morphological differences
between the various phases of the merger and, in particular, suggest that mergers occupy a distinct morphological phase space during the ``final coalescence" period when the SFR peaks \citep{2008MNRAS.391.1137L,2009arXiv0912.1590L,2009arXiv0912.1593L}. To test these predictions, and in general to understand the physical processes governing galaxy
evolution, it is essential to study the {\it Spitzer}-selected and SMG
populations in detail.
We have embarked on a detailed study of a large sample of extremely
dust-obscured, high-redshift ULIRGs with the goal of understanding
their evolutionary history. Our sample is selected using {\it Spitzer} and
ground-based optical imaging of the Bo\"otes field of the NOAO Deep
Wide-Field Survey \citep[NDWFS\footnote{http://www.noao.edu/noaodeep}; Jannuzi et al., in prep.; Dey et al., in prep.][]{1999ASPC..191..111J} to have extreme optical-to-mid-IR
colors $R-[24]\ge 14$ Vega mag ($\approx F_\nu(24\mu{\rm m})/F_\nu(R)
\ge 1000$) and are called Dust-Obscured Galaxies (DOGs). Spectroscopic
redshifts for a subset of the DOGs have been measured using the
Infrared Spectrometer \citep[IRS;][]{2004ApJS..154...18H} on {\it Spitzer} and optical and
near-IR spectrographs at the W. M. Keck Observatory
\citep{2005ApJ...622L.105H, 2006ApJ...651..101W, 2009ApJ...700.1190D}. DOGs satisfying
$F_\nu(24\mu{\rm m})\ge 0.3$~mJy have a fairly narrow distribution
in redshift ($z\approx 2.0\pm0.5$) and a space density of $\approx
2.8\times 10^{-5} h_{70}^3 {\rm Mpc}^{-3}$ \citep{2008ApJ...677..943D}.
Although rare, these sources are sufficiently luminous that they
contribute up to one-quarter of the total IR luminosity density at
redshift $z\sim 2$ and constitute a substantial fraction of the
ULIRG population at this redshift.
DOGs are the most dust-reddened ULIRGs at $z\approx 2$; similar to
the broader ULIRG population, DOGs exhibit a wide range in SED
stretching from power-law dominated mid-IR SEDs (i.e., ``power-law
DOGs'') to SEDs which exhibit bumps (i.e, ``bump DOGs''). In
\citet[][hereafter Paper~I]{2009ApJ...693..750B}, we analyzed the
morphologies of 31 of the brightest 24$\mu$m-selected DOGs (all
with $F_{\rm 24\mu m} > 0.8$ mJy) that have power-law mid-IR SEDs.
All of these objects had spectroscopic redshifts and most exhibit
strong 9.7$\mu$m silicate absorption in their IRS spectra
\citep{2005ApJ...622L.105H, 2006ApJ...651..101W, 2009ApJ...700.1190D}.
The power-law DOGs are nearly always spatially resolved, with
effective radii of $1-5$~kpc, although a few show obvious signs of
merger activity \citep{2008ApJ...680..232D,2009ApJ...693..750B}. $K$-band
adaptive optics imaging (from Keck) of 15 objects has revealed an intriguing
dependence of size on SED shape: power-law dominated sources are
more compact than 24$\mu$m-faint bump-dominated sources
\citep{2008AJ....136.1110M,2009AJ....137.4854M}. This is consistent with
the idea of the bright, power-law DOGs being more AGN dominated.
The primary goal of this paper is to identify any quantifiable morphological
differences between SMGs, {\it Spitzer}-selected bump ULIRGs and other {\it
Spitzer}-selected power-law ULIRGs. We present and analyze new {\it HST} Wide-Field
Planetary Camera 2 \citep[WFPC2][]{1994ApJ...435L...3T} and Near-IR Camera and Multi-object Spectrometer \citep[NICMOS][]{1998ApJ...492L..95T} observations of 19 bump DOGs and 3 more
power-law DOGs. We also assemble a larger sample of $z\approx 2$ ULIRGs,
drawn from Paper I (power-law DOGs) and the literature, with high
spatial resolution imaging data appropriate for morphological
analyses. In particular, we include a large sample of SMGs
\citep[from the stufy of][]{2010MNRAS.405..234S} and expand the sample of
{\it Spitzer}-selected ULIRGs by including those from the eXtragalactic
First Look Survey \citep[XFLS;][]{2008ApJ...680..232D}.
Our combined dataset contains 103 high-redshift ULIRGs with available
and fairly comparable {\it HST} data. We present a uniform morphological analysis of these objects and compare the results to the expectations from models for the formation and evolution of these systems.
In section~\ref{sec:data} we detail the sample selection, observations, and
data reduction. In section~\ref{sec:methods}, we describe our methodology for
measuring photometry and morphologies, including a visual classification
experiment, non-parametric quantities, and GALFIT modeling.
Section~\ref{sec:results} contains the results of this analysis, including a
comparison of SMG, DOG, and simulated merger morphologies. In
section~\ref{sec:disc}, we discuss the implications of our results. We summarize
our conclusions in section~\ref{sec:conclusions}.
Throughout this paper we assume $H_0=$70~km~s$^{-1}$~Mpc$^{-1}$, $\Omega_{\rm
m} = 0.3$, and $\Omega_\lambda = 0.7$. At $z=2$, this results in a spatial
scale of 8.37~kpc/$\arcsec$.
\section{Data}\label{sec:data}
In this section, we describe the new {\it HST} observations of bump DOGs and
the procedure used to reduce them. We also detail the archival datasets of
power-law DOGs, SMGs, and XFLS ULIRGs used in subsequent sections of this paper.
\subsection{Bump DOGs}\label{sec:bumpdog}
The 22 DOGs presented in this paper were observed with {\it HST} from 2007
December to 2008 May. All were observed with WFPC2 through the F814W filter
and with the NICMOS NIC2 camera through the F160W filter.
Table~\ref{tab:observations} summarizes the details of the observations. All
data were processed using IRAF\footnote{IRAF is distributed by the National
Optical Astronomy Observatory, which is operated by the Association of
Universities for Research in Astronomy, Inc., under cooperative agreement with
the National Science Foundation. http://iraf.noao.edu/}. The following
sections provide more details about the sample selection and processing of the
WFPC2 and NICMOS images used in this paper.
\begin{deluxetable*}{lllllllll}
\tabletypesize{\footnotesize}
\tablecolumns{7}
\tablewidth{6.5in}
\tablecaption{Observations}
\tablehead{
\colhead{Source Name} &
\colhead{ID\tablenotemark{a}} &
\colhead{RA (J2000)} &
\colhead{DEC (J2000)} &
\colhead{$z$\tablenotemark{b}} &
\colhead{WFPC2/F814W} &
\colhead{NIC2/F160W}
}
\startdata
SST24 J142637.3+333025 & 1 & +14:26:37.397 & +33:30:25.82 & 3.200\tablenotemark{c} & 2008-02-11 & 2007-12-31 \\
SST24 J142652.4+345504 & 12 & +14:26:52.555 & +34:55:05.53 & 1.91 & 2008-03-28 & 2008-01-01 \\
SST24 J142724.9+350823 & 4 & +14:27:25.016 & +35:08:24.20 & 1.71 & 2008-07-02 & 2008-01-14 \\
SST24 J142832.4+340850 & 8 & +14:28:32.476 & +34:08:51.23 & 1.84 & 2008-07-03 & 2008-01-15 \\
SST24 J142920.1+333023 & 17 & +14:29:20.164 & +33:30:23.59 & 2.01 & 2008-02-01 & 2008-05-26\tablenotemark{d} \\
SST24 J142941.0+340915 & 13 & +14:29:41.085 & +34:09:15.61 & 1.91 & 2008-05-21 & 2008-03-15 \\
SST24 J142951.1+342041 & 5 & +14:29:51.163 & +34:20:41.33 & 1.76 & 2008-01-28 & 2008-01-14 \\
SST24 J143020.4+330344 & 11 & +14:30:20.537 & +33:03:44.45 & 1.87 & 2008-03-21 & 2008-04-11 \\
SST24 J143028.5+343221 & 21 & +14:30:28.534 & +34:32:21.62 & 2.178\tablenotemark{e} & 2008-05-07 & 2008-01-15 \\
SST24 J143137.1+334500 & 7 & +14:31:37.080 & +33:45:01.26 & 1.77 & 2008-05-20 & 2008-04-12 \\
SST24 J143143.3+324944 & 2 & +14:31:43.400 & +32:49:44.38 & --- & 2008-02-10 & 2008-03-15 \\
SST24 J143152.4+350029 & 3 & +14:31:52.463 & +35:00:29.44 & 1.50 & 2008-01-24 & 2008-05-22 \\
SST24 J143216.8+335231 & 6 & +14:32:16.904 & +33:52:32.18 & 1.76 & 2008-02-01 & 2008-03-16 \\
SST24 J143321.8+342502 & 18 & +14:33:21.890 & +34:25:02.62 & 2.10 & 2008-05-21 & 2008-01-15 \\
SST24 J143324.2+334239 & 14 & +14:33:24.269 & +33:42:39.55 & 1.91 & 2008-02-02 & 2008-01-17 \\
SST24 J143331.9+352027 & 15 & +14:33:31.945 & +35:20:27.28 & 1.91 & 2007-12-25 & 2008-01-14 \\
SST24 J143349.5+334602 & 10 & +14:33:49.585 & +33:46:02.00 & 1.86 & 2008-03-18 & 2008-01-07 \\
SST24 J143458.8+333437 & 20 & +14:34:58.953 & +33:34:37.57 & 2.13 & 2008-07-03 & 2008-05-21 \\
SST24 J143502.9+342657 & 19 & +14:35:02.930 & +34:26:58.88 & 2.10 & 2008-05-09 & 2008-01-15 \\
SST24 J143503.3+340243 & 16 & +14:35:03.336 & +34:02:44.16 & 1.97 & 2008-02-29 & 2008-01-07 \\
SST24 J143702.0+344631 & 22 & +14:37:02.018 & +34:46:30.93 & 3.04 & 2008-03-28 & 2007-12-28 \\
SST24 J143816.6+333700 & 9 & +14:38:16.714 & +33:37:00.94 & 1.84 & 2008-07-03 & 2008-01-14 \\
\tablenotetext{a}{Panel number in Figure~\ref{fig:cutouts1}}
\tablenotetext{b}{Redshift from {\it Spitzer}/IRS \citep{2009ApJ...700.1190D} unless otherwise noted}
\tablenotetext{c}{Redshift from Keck LRIS (Soifer et al., in prep.)}
\tablenotetext{d}{This observation provided no usable data}
\tablenotetext{e}{Redshift from Keck NIRSPEC \citep{2007ApJ...663..204B}}
\enddata
\label{tab:observations}
\end{deluxetable*}
\subsubsection{Sample Selection} \label{sec:sample}
A sample of 2603 DOGs was identified by \citet{2008ApJ...677..943D} from the
9.3 deg$^2$ Bo\"{o}tes Field of the NDWFS. Keck and {\it Spitzer} spectroscopy
have resulted in redshifts of $\approx~100$ DOGs, approximately 60\% of which
have power-law dominated mid-IR SEDs and 40\% have bump SEDs. These are
objects which have very high intrinsic to observed UV luminosity ratios, on par
with or beyond the most extreme starbursts studied by {\it Spitzer} in the
local universe \citep{2010ApJ...715..986S}.
In \citet[][hereafter Paper~I]{2009ApJ...693..750B}, we analyzed {\it HST}
imaging (program HST-GO10890) of 31 of the brightest DOGs at 24$\mu m$ (all
have $F_{24\mu {\rm m}} > 0.8 \:$mJy) that have power-law mid-IR SEDs and
spectroscopic redshifts based on the 9.7$\mu$m silicate absorption feature,
most likely due to the presence of warm dust heated by an AGN
\citep{2006ApJ...653..101W,2007ApJ...660..167D,2008ApJ...675..960P,2008ApJ...680..119B}.
In this paper, we analyze {\it HST} imaging (program HST-GO11195) of 22 DOGs
that show a bump in their rest-frame mid-IR SED \citep[selected using Arp~220
as a template; for details see][]{2009ApJ...700.1190D}. This feature indicates
that the mid-IR light is dominated by stellar emission in these sources.
Furthermore, {\it Spitzer} mid-IR spectroscopy has provided redshifts for 20/22
of these sources via identification of PAH emission features commonly
associated with on-going star-formation \citep{2009ApJ...700.1190D}.
Subsequent deeper mid-IR imaging from the {\it Spitzer} Deep Wide-Field Survey
\citep{2009ApJ...701..428A} has revealed that the two sources lacking PAH
features have power-law mid-IR SEDs. One additional target has a power-law
mid-IR SED (SST24~J143028.5+343221) and was observed by {\it HST} because the
bump source it replaced could not be observed due to scheduling constraints.
Figure~\ref{fig:samplecmd} shows the $R - [24]$ color and $R$-band magnitude
(Vega system) for the following sources with {\it HST} imaging: bump and
power-law DOGs, SMGs, and XFLS ULIRGs at high redshift. Following careful
reanalysis of the $R$-band photometry \citep[compared to][with the main
difference being a revised estimate of the sky background
level]{2008ApJ...677..943D}, a few DOGs show $R-[24]$ colors $\approx 0.1~$mag
below the nominal DOG threshold. We refer to these objects as DOGs in this
paper because they satisfy the essential physical characteristics of DOGs: they
are $z \sim 2$ ULIRGs that are likely to be a highly obscured stage in massive
galaxy evolution. The bump DOGs in this sample have fainter 24$\mu$m flux
densities and less extreme $R - [24]$ colors than the power-law DOGs. These
distinctions are qualitatively representative of the photometric properties of
the full sample of 2603 DOGs in the Bo\"otes Field.
\begin{figure}[!tbp]
\epsscale{1.00}
\plotone{f1.eps}
\caption{$R - [24]$ color vs. 24$\mu$m magnitude distribution for all DOGs in
the NDWFS Bo\"{o}tes field (gray dots). Arrows indicate $R$-band
non-detections (2$\sigma$ level), and cross symbols highlight power-law
dominated sources. Also shown are the samples with high-spatial resolution
imaging studied in this paper: power-law DOGs (red circles), bump DOGs (black
squares), SMGs (blue stars), and XFLS ULIRGs (purple triangles). Power-law
sources tend to be the brightest at 24$\mu$m and the most heavily obscured.
\label{fig:samplecmd}}
\end{figure}
Figure~\ref{fig:zdist} shows the redshift distributions of bump DOGs, power-law
DOGs, SMGs, and XFLS ULIRGs with {\it HST} data in comparison to all DOGs in
Bo\"otes with spectroscopic redshifts. Bump DOGs predominantly lie in a
relatively narrow redshift range of $1.5 < z < 2.1$. Briefly, this is because
at $z = 1.9$, the strong 7.7$\mu$m PAH feature boosts the 24$\mu$m flux,
pushing sources with weaker continuum into the flux-limited bump DOG
sample \citep[for additional details, see][]{2009ApJ...700.1190D}.
\begin{figure}[!bp]
\epsscale{1.00}
\plotone{f2.eps}
\caption{ {\it Left}: Redshift distribution of DOGs in the Bo\"{o}tes Field
with spectroscopic redshifts (gray histogram; either from $Spitzer/IRS$ or Keck
DEIMOS/LRIS, Soifer et al. in prep.). The hatched histograms show the redshift
distributions of the subset of power-law DOGs (red) and bump DOGs (black) studied
in this paper. The redshift distribution of bump DOGs is relatively narrow due to selection effects \citep[for details
see][]{2009ApJ...700.1190D}, while power-law DOGs are weighted
towards slightly larger redshifts. {\it Right}: Redshift distribution of SMGs
(blue histogram) and XFLS ULIRGs (purple histogram) at $z > 1.4$ studied in
this paper. Hatched regions denote the sub-sample qualifying as power-law
dominated in the mid-IR. }
\label{fig:zdist}
\end{figure}
\subsubsection{WFPC2 Data} \label{sec:wfpc2} The Wide Field Camera CCD 3 of
WFPC2 was used to image the 22 DOGs in this study. These observations
consisted of double-orbit data with the F814W filter. We used a three point
dither pattern (WFPC2-LINE) with a point and line spacing of 0$\farcs$3535 and
a pattern orientation of 45$^\circ$. Total exposure duration at the nominal
pixel scale of 0$\farcs$1~pix$^{-1}$ was $\approx$3800~sec. The standard WFPC2
pipeline system was used to bias-subtract, dark-subtract, and flat-field the
images (Mobasher et al., 2002). MultiDrizzle was then used to correct for
geometric distortions, perform sky-subtraction, image registration, cosmic ray
rejection and final drizzle combination \citep{2002hstc.conf..337K}. We used a
square interpolation kernel and output pixel scale of
0.075$\arcsec$~pix$^{-1}$, leading to a per-pixel exposure time in the drizzled
image of $\approx$2200~sec. Typically, a point source with an F814W AB
magnitude of 26.1 may be detected at the 5$\sigma$ level by using a 0\farcs3
diameter aperture.
\subsubsection{NICMOS Data} \label{sec:nicmos}
Single-orbit data of the DOGs were acquired with the NIC2 camera and the F160W
filter. We used a two-point dither pattern (NIC-SPIRAL-DITH) with a point
spacing of 0.637$\arcsec$. The total exposure time per source was $\approx$2700~s.
We followed the standard data reduction process outlined in the NICMOS data
handbook \citep{nicmos}. We used the IRAF routine {\tt nicpipe} to pre-process
the data, followed by the {\tt biaseq} task to correct for non-linear bias
drifts and spatial bias jumps. We then used {\tt nicpipe} a second time to do
flat-fielding and initial cosmic-ray removal. The IRAF task {\tt pedsky} was
used to fit for the sky level and the quadrant-dependent residual bias.
Significant residual background variation remained after this standard
reduction process. To minimize these residuals, we followed the procedure
outlined in Paper~I: we constructed an object-masked median sky image based on
all of our NIC2 science frames, scaled it by a spatially constant factor and
subtracted it from each science image. The scaling factor was computed by
minimizing the residual of the difference between the masked science image and
the scaled sky image. Mosaicing of the dithered exposures was performed using
{\tt calnicb} in IRAF, resulting in a pixel scale of 0.075$\arcsec$~pix$^{-1}$.
Although the noise varies from image to image, typically a point source with an
F160W AB magnitude of 25.2 may be detected at the 5$\sigma$ level by using a
0\farcs3 diameter aperture.
\subsubsection{Astrometry}\label{sec:astro}
Each WFPC2 and NICMOS image is aligned to the reference frame of the NDWFS,
which itself is tied to the USNO A-2 catalog. We identify well-detected,
unsaturated sources in the $I$-band NDWFS data overlapping the field of view
(FOV) of each WFPC2/F814W image using Source Extractor
\citep[SExtractor, version 2.5.0,][]{1996A&AS..117..393B}. The IRAF tasks {\tt wcsctran} and
{\tt imcentroid} are used to convert the RA and DEC values of this list of
comparison sources into WFPC2 pixel coordinates. Finally, the IRAF task {\tt
ccmap} is used to apply a first order fit which corrects the zero point of the
astrometry and updates the appropriate WCS information in the header of the
WFPC2 image. The aligned WFPC2 image serves as a reference frame for
correcting the astrometry of the NICMOS image as well as the IRAC images
\citep[since the IRAC images of the Bo\"{o}tes Field are not tied to the USNO
A-2 catalog, but instead to the 2$\mu$m All-Sky Survey frames,
see][]{2009ApJ...701..428A} using a similar procedure. The properly aligned,
multi-wavelength dataset generally allows for straightforward identification of
the proper counterpart to the MIPS source, since inspection of the four IRAC
channels reveals a single source associated with the 24$\mu$m emission for all
sources. The absolute uncertainty in the centroid of the IRAC 3.6$\mu$m
emission ranges from 0$\farcs$2-0$\farcs$4.
\subsection{Power-law DOGs}\label{sec:pldogs}
In Paper~I we analyzed {\it HST} imaging
of 31 power-law DOGs at $z > 1.4$. Although these sources have mid-IR SED
features indicative of obscured AGN, their rest-frame optical morphologies
nearly all show minor ($<30\%$) point-source contributions and significant
emission on scales of 1-5~kpc. This indicates that the rest-frame optical
light of these sources is produced from stars, rather than AGN.
The NICMOS exposure times and $H$-band luminosities of these sources are
similar to the bump DOGs, facilitating a comparison between the two
populations. This particular comparison --- between distinct sub-classes of
the most extreme dust-obscured ULIRGs --- is a major aspect of this study.
\subsection{SMG Data}\label{sec:smgdata}
The SMG data used in this paper are {\it HST} NICMOS/F160W imaging of a sample
of 25 SMGs selected from a catalog of 73 SMGs with spectroscopic redshifts
\citep{2005ApJ...622..772C} and were first presented by
\citet{2010MNRAS.405..234S}. Of the 25 SMGs, 23 have single-orbit NIC2 imaging
from Cycle~12 {\it HST} program GO-9856 \citep{2010MNRAS.405..234S} and an
additional 6 have multi-orbit NIC3 imaging from GOODS-N (Conselice et al. 2010
in prep.). {\it HST} optical imaging in the F814W filter is also available for
all of these objects.
In this paper, we focus on the subset of 18 SMGs at $z>1.4$. Of these 18, all
have NIC2 imaging and three (SMM~J123622.65+621629.7,
SMM~J123632.61+620800.1, and SMM~J123635.59+621424.1) have NIC3 imaging as well. Although the NIC3
images are significantly deeper, we prefer to use the NIC2 data (each of these
sources is well-detected at S/N$>2$) because of the superior pixel scale of
NIC2 and the unusual shape of the NIC3 PSF. Some of these sources have optical
{\it HST} imaging with the Advanced Camera for Surveys (ACS), but the S/N
levels are generally insufficient for quantitative analysis and so are not used
in this study.
We obtained the NIC2 images of SMGs from the {\it HST} data archive and reduced
them following the same procedure that is outlined in section~\ref{sec:nicmos}.
Most importantly, the methodology used to analyze the photometry and morphology
of both SMGs and DOGs in this study is identical and is described in
section~\ref{sec:methods}.
\subsection{XFLS Data}\label{sec:xflsdata}
A sample of 33 XFLS ULIRGs at $z > 1.4$ was imaged with {\it HST} NICMOS/F160W
in Cycle~15 as part of program GO10858. These data and a morphological
analysis of the imaging was presented in \citet{2008ApJ...680..232D}. We note
that in our study, we use only single-orbit NIC2 data of these objects to
facilitate comparison with the NIC2 images of the other high-$z$ ULIRG
populations studied here, which all have only single-orbit NIC2 data.
Double-orbit imaging is available for nearly 50\% of the sample and in
principle could be used to measure more accurate morphologies of the fainter
objects as well as test for systematic errors in the morphologies resulting
from low S/N. The data were obtained from the {\it HST} data archive, reduced,
and analyzed using the same methodology that was applied to DOGs and SMGs.
\section{Methodology}\label{sec:methods}
In this section, we describe our methods to measure photometry as well as
visual, non-parametric, and GALFIT morphologies.
\subsection{Photometry}\label{sec:photo}
We perform 2$\arcsec$ diameter aperture photometry on each DOG in both the
NICMOS and WFPC2 images, choosing the center of the aperture to be located at
the peak flux pixel in the NICMOS images. Foreground and background objects
are identified and removed using SExtractor (see Section~\ref{sec:nonpar}).
The sky level is derived using an annulus with an inner diameter of 2$\arcsec$
and a width of 2$\arcsec$. In cases where the flux density radial profile did
not flatten at large radii, the appropriate sky value was determined by
trial-and-error. Photometric uncertainty was computed by measuring the
sigma-clipped root-mean-square of fluxes measured in $N$ $2\arcsec$ diameter
apertures, where $N \approx 10$ and $N \approx 100$ for the NICMOS and WFPC2
images, respectively. We verified the accuracy of our WFPC2 photometric
zeropoints by comparing well-detected, non-saturated sources common to both the
WFPC2/F814W and NDWFS/$I$-band imaging. Photometric measurements of the DOGs
are presented in Table~\ref{tab:sources}.
\subsection{Morphology}\label{sec:morphmeth}
To analyze the morphologies of the bump DOGs, we follow a similar procedure to
that outlined in Paper~I. Here we summarize the three
different, complementary approaches used in analyzing the morphology of the
DOGs in our sample: a visual classification experiment, multi-component GALFIT
modeling, and non-parametric quantification.
\subsubsection{Visual Classification} \label{sec:visualmeth}
For this paper, our visual classification experiment differed significantly from
Paper~I. The goal of the original experiment outlined in Paper~I was to
determine if DOGs could be distinguished from normal field galaxies based on a
visual classification. This proved difficult to quantify due to the faintness
of DOGs in the rest-frame UV (ACS/WFPC2 images) and the small number of field
galaxies in the rest-frame optical (NICMOS images).
Our new classification experiment is designed specifically to identify
morphological differences found in the NICMOS imaging of bump and power-law
DOGs. We generated a 5$\arcsec$x5$\arcsec$ cutout image of every DOG with
NICMOS data (both power-law and bump sources, a total of 53 objects) and
arranged them randomly. Seven of the coauthors classified these objects into
``Regular'', ``Irregular'', or ``Too Faint To Tell''. In addition to probing
for a difference between bump and power-law DOGs, the mode of the
classifications for each DOG as well as the number of coauthors in agreement
with the mode is useful as a qualititative assessment of the morphology for
comparison with the more quantitative methods discussed below.
Results are presented in
Table~\ref{tab:vismorph} and discussed in section~\ref{sec:visual}.
\subsubsection{Non-parametric Classification} \label{sec:nonparmeth}
A wide variety of tools now exist to quantify the morphologies of galaxies.
Five which frequently appear in the literature are the concentration
index $C$ \citep{1994ApJ...432...75A}, the rotational asymmetry $A$
\citep{1995ApJ...451L...1S}, the residual clumpiness, $S$
\citep{2003ApJS..147....1C}, the Gini coefficient $G$
\citep{2003ApJ...588..218A}, and $M_{20}$ parameter
\citep{2004AJ....128..163L}. Of these five, $A$ and $S$ have S/N and spatial
resolution requirements that are not satisfied by the existing imaging of the DOGs in
this sample \citep[e.g.,][show that significant type-dependent systematic
offsets in $A$ arise at per-pixel S/N$<5$]{2004AJ....128..163L}. Therefore,
this analysis is focused on $C$, $G$, and $M_{20}$.
The concentration index $C$ is defined as
\citep{2000AJ....119.2645B}:
\begin{equation}
C = 5 {\rm log10} \left( \frac{ r_{80} }{ r_{20} } \right),
\end{equation}
\noindent where $r_{80}$ and $r_{20}$ are the radii of circular apertures containing
80\% and 20\% of the total flux, respectively. $G$ was originally introduced
to measure how evenly the wealth in a society is distributed
\citep{glasser1962}. Recently, \citet{2003ApJ...588..218A} and
\citet{2004AJ....128..163L} applied this method to aid in galaxy
classification: low values imply a galaxy's flux is well-distributed among many
pixels, while high values imply a small fraction of the pixels within a galaxy
account for the majority of the total flux. $M_{20}$ is the logarithm of the
second-order moment of the brightest 20\% of the galaxy's flux, normalized by
the total second-order moment \citep{2004AJ....128..163L}. Higher values of
$M_{20}$ indicate multiple bright clumps offset from the second-order moment
center. Lower values are typical of centrally-dominated systems.
Prior to computing $G$ or $M_{20}$, we first generate a catalog of objects
using SExtractor \citep{1996A&AS..117..393B}. We use a detection threshold of
3$\sigma$ (corresponding to 23.7~mag~arcsec$^{-2}$) and a minimum detection
area of 15~pixels. The number of deblending thresholds was 32, and the minimum
contrast parameter for deblending was 0.1. We found by trial and error that
these parameters minimized the separation of a single galaxy into multiple
components.
For SST24 J143349.5+334602 and SST24 J142652.4+345504, examination of the
F814W-F160W color indicated that a nearby neighbor with similar color should
not be excluded as a foreground/background object. For both DOGs, we modified
the segmentation map to reflect this.
The final segmentation map produced by SExtractor (and modified in two cases)
is used to mask out foreground/background objects (pixels that are masked out
are simply not used in the remainder of the analysis).
The center of the image, the ellipticity, and position angle computed by
SExtractor are used as inputs to our morphology code. This code is written by
J.~Lotz and described in detail in \citet{2004AJ....128..163L}. Here, we
summarize the relevant information.
Postage stamps of each object in the SExtractor catalog (and the associated
segmentation map) are created with foreground and background objects masked
out. For each source, we adopt the sky value computed in our photometric
analysis. Since the isophotal-based segmentation map produced by SExtractor is
subject to the effects of surface brightness dimming at high redshift, pixels
belonging to the galaxy are computed based on the surface brightness at the
elliptical Petrosian radius, $\mu(r_{\rm P})$. We adopt the usual generalized
definition for $r_{\rm P}$ as the radius at which the ratio of the surface
brightness at $r_{\rm P}$ to the mean surface brightness within $r_{\rm P}$ is
equal to 0.2 \citep{1976ApJ...209L...1P}. The elliptical $r_{\rm P}$ is
derived from surface brightness measurements within elliptical apertures and
represents the length of the major axis. Studies have shown that using the
Petrosian radius to select pixels associated with a galaxy provides the most
robust morphological measurements
\citep{2004AJ....128..163L,2008ApJS..179..319L}. Pixels with surface
brightness above $\mu(r_{\rm P})$ are assigned to the galaxy while those below
it are not.
Using the new segmentation map, we recompute the galaxy's center by minimizing
the total second-order moment of the flux. A new value of $r_{\rm P}$ is then
computed and a revised segmentation map is used to calculate $G$ and $M_{20}$.
Finally, the morphology code calculates an average S/N per pixel value using
the pixels in the revised segmentation map \citep[Eqs.~1 through 5
in][]{2004AJ....128..163L}. The S/N per pixel and spatial resolution of each
image is used to estimate the uncertainties in the morphological parameters of
each galaxy. The uncertainties are derived from the rms variation between
measurements of the same galaxies in GOODS images compared to UDF images
\citep{2006ApJ...636..592L} and assumes that the UDF morphology measurements
are ``truth''. Results of this analysis are presented in Table~\ref{tab:morph}
and will be discussed in section~\ref{sec:nonpar}.
\subsubsection{GALFIT Modeling}\label{sec:parmeth}
In Paper~I, we reported the existence of a centrally located, compact component
that was present in the NICMOS images of power-law DOGs but absent in the
ACS/WFPC2 images, signifying the presence of strong central obscuration. To
quantify this feature, we used GALFIT \citep{2002AJ....124..266P} to model the
2-D light profile of the DOGs. In this paper, we repeat this procedure on the
bump DOGs with {\it HST} NICMOS data. Here, we review our methodology.
We choose the size of the fitting region to be 41$\times$41 pixels
(corresponding to angular and physical sizes of 3$\arcsec$ and $\approx$24~kpc,
respectively) because the DOGs are small and have low S/N compared to more
typical applications of GALFIT. For the same reason, we wish to include only
the minimum necessary components in our model. We model the observed emission
with three components which are described by a total of 10 free parameters.
The number of degrees of freedom, $N_{\rm DOF}$, is calculated as the
difference of the number of pixels in the image and the number of free
parameters. Thus, the maximum $N_{\rm DOF}$ is 1671. Cases where $N_{\rm DOF}
< 1671$ are associated with images where some pixels were masked out because
they were associated with obvious residual instrumental noise. NIC2 is a
Nyquist-sampled array (0.075$\arcsec$~pix$^{-1}$ compared to 0.16$\arcsec$ FWHM
beam), so the pixels in our image are not completely independent and the
$\chi^2_\nu$ values should be interpreted in a relative sense rather than an
absolute one.
The first element in our GALFIT model is a sky component whose amplitude is
held constant at a value derived from the photometry to yield flat radial
profiles. The second is an instrumental PSF generated from the TinyTim
software assuming a red power-law spectrum ($F_\nu \propto \nu^{-2}$) as the
object spectrum (Krist and Hook 2004), which can simulate a PSF for NICMOS,
WFPC2, and ACS. For the NICMOS and WFPC2 images, the DOG is positioned in
nearly the same spot on the camera. In the case of WFPC2 this is pixel
(132,144) of chip 3 and pixel (155, 164) for NICMOS. The PSF is computed out
to a size of 3.0$\arcsec$, and for the WFPC2 PSF we oversample by a factor of
1.3 to match the pixel scale of the drizzled WFPC2 images.
The final component is a S\'{e}rsic profile \citep{1968adga.book.....S} where
the surface brightness scales with radius as exp[$-\kappa ( (r/R_{\rm
eff})^{1/n}-1)$], where $\kappa$ is chosen such that half of the flux falls
within $R_{\rm eff}$. As few constraints as possible were placed so as to
optimize the measurement of the extended flux (i.e., non-point source
component). In certain cases, the S\'{e}rsic index had to be constrained to be
positive to ensure convergence on a realistic solution. When fitting the
NICMOS data, the uncertainty image from {\tt calnicb} provides the necessary
information required by GALFIT to perform a true $\chi^2$ minimization. The
TinyTim NIC2 PSF is convolved with the S\'ersic profile prior to performing the
$\chi^2$ minimization. The initial guesses of the magnitude, half-light
radius, position angle, and ellipticity were determined from the output values
from SExtractor. Varying the initial guesses within reasonable values (e.g.,
plus or minus two pixels for the half-light radius) yielded no significant
change in the best-fit model parameters. The NICMOS centroid was used as the
initial guess for the (x,y) position of both the PSF and extended components.
A degeneracy potentially exists between our estimates of the point-source
fraction (i.e., relative ratio of PSF component flux to S\'ersic component
flux) and the S\'ersic index. Fits using models without the PSF
component yield larger reduced $\chi_\nu^2$ values, especially when the point
source fraction in our three-component model was large (see further discussion
in section~\ref{sec:GALFIT}). In cases where the point source fraction was
small, the no-PSF model had similar parameter values as our fiducial
three-component model, as would be expected.
The results of this GALFIT analysis are presented in Table~\ref{tab:morph} and
will be discussed in section~\ref{sec:GALFIT}.
It is important to note here that NIC2 cannot spatially resolve objects smaller
than 1.3~kpc at $z \approx 2$. This limit is large enough to encompass a
compact stellar bulge as well as an active galactic nucleus, implying that we
cannot, from these data alone, distinguish between these two possibilities as
to the nature of any central, unresolved component.
\section{Results}\label{sec:results}
In this section, we present our photometry, visual classification,
non-parametric classification, GALFIT modeling, and stellar and dust mass
results.
\subsection{Photometry}\label{sec:photres}
Table~\ref{tab:sources} presents the photometric information derived from the
{\it HST} imaging. In Figure~\ref{fig:imh}, we show the $I-H$ vs. $H$
color-magnitude diagram for bump DOGs, power-law DOGs, XFLS ULIRGs, and a
sample of galaxies in the Hubble Deep Field (HDF) whose photometric redshifts
are comparable to DOGs ($1.5 < z_{\rm phot} < 2.5$). Power-law DOGs tend to be
the reddest sources ($I-H \approx 2-5$ AB mag), followed by bump DOGs ($I-H
\approx 2-3$ AB mag), XFLS ULIRGs ($I-H \approx 1.5-3$), and SMGs, which have
$I-H$ colors similar to high-$z$ galaxies in the HDF ($I-H \approx 0-2$ AB
mag). SMGs and DOGs (both bump and power-law varieties) are comparably bright
in $H$. The bluer color of SMGs relative to DOGs at a given $H$-band magnitude
suggests weaker UV flux from DOGs, either due to older stellar populations in
DOGs or a higher dust mass relative to stellar mass in DOGs.
\begin{deluxetable*}{lcccccc}[!tp]
\tabletypesize{\small}
\tablecolumns{7}
\tablewidth{6.5in}
\tablecaption{Photometric Properties}
\tablehead{
\colhead{} &
\colhead{$F_{\rm F814W}$} &
\colhead{$\sigma_{\rm F814W}$} &
\colhead{$F_{F160W}$} &
\colhead{$\sigma_{F160W}$} &
\colhead{$F_{24}$} &
\colhead{$R - [24]$} \\
\colhead{Source Name} &
\colhead{($\mu$Jy)} &
\colhead{($\mu$Jy)} &
\colhead{($\mu$Jy)} &
\colhead{($\mu$Jy)} &
\colhead{(mJy)} &
\colhead{(Vega)}
}
\startdata
SST24 J142637.3+333025 & 0.36 & 0.19 & 0.45 & 0.57 & 0.64 & $>$15.0 \\
SST24 J142652.4+345504 & 0.24 & 0.15 & 1.78 & 0.36 & 1.29 & 15.0 \\
SST24 J142724.9+350823 & 0.63 & 0.15 & 6.72 & 0.42 & 0.51 & 14.4 \\
SST24 J142832.4+340850 & 0.59 & 0.16 & --- & --- & 0.52 & 13.9 \\
SST24 J142920.1+333023 & 0.35 & 0.14 & 2.85 & 0.27 & 0.51 & $>$13.6 \\
SST24 J142941.0+340915 & 0.30 & 0.13 & 2.47 & 0.46 & 0.59 & $>$14.6 \\
SST24 J142951.1+342041 & 0.55 & 0.16 & 5.30 & 0.52 & 0.60 & $>$14.9 \\
SST24 J143020.4+330344 & 0.31 & 0.13 & 4.26 & 0.50 & 0.54 & $>$15.3 \\
SST24 J143028.5+343221 & 0.59 & 0.16 & 4.92 & 0.31 & 1.27 & 14.7 \\
SST24 J143137.1+334500 & 0.18 & 0.14 & 2.67 & 0.37 & 0.57 & 14.3 \\
SST24 J143143.3+324944 & 0.43 & 0.15 & 6.43 & 0.37 & 1.51 & 14.4 \\
SST24 J143152.4+350029 & 0.54 & 0.16 & 8.21 & 0.31 & 0.52 & 14.3 \\
SST24 J143216.8+335231 & 0.51 & 0.15 & 4.24 & 0.37 & 1.28 & $>$16.1 \\
SST24 J143321.8+342502 & 0.72 & 0.16 & 7.16 & 0.37 & 0.56 & 14.4 \\
SST24 J143324.2+334239 & 0.96 & 0.17 & 6.67 & 0.47 & 0.53 & 13.8 \\
SST24 J143331.9+352027 & 0.66 & 0.17 & 3.58 & 0.32 & 0.60 & 14.3 \\
SST24 J143349.5+334602 & 0.63 & 0.15 & 4.44 & 0.33 & 0.53 & 14.3 \\
SST24 J143458.8+333437 & 0.49 & 0.20 & 5.14 & 0.51 & 0.57 & 14.0 \\
SST24 J143502.9+342657 & 0.24 & 0.13 & 2.52 & 0.63 & 0.50 & 14.1 \\
SST24 J143503.3+340243 & 0.26 & 0.14 & 3.68 & 0.36 & 0.76 & 14.6 \\
SST24 J143702.0+344631 & 0.10 & 0.11 & 0.17 & 0.33 & 0.33 & 14.2 \\
SST24 J143816.6+333700 & 0.68 & 0.16 & 4.22 & 0.22 & 3.28 & 14.8 \\
\enddata
\label{tab:sources}
\end{deluxetable*}
\begin{figure}[!tbp]
\epsscale{1.0}
\plotone{f4.eps}
\caption{ Color-magnitude diagram for bump DOGs, power-law DOGs , SMGs, and
XFLS ULIRGs at $z > 1.4$ (symbols as in Figure~\ref{fig:samplecmd}). Smaller
symbols indicate objects where the $I$-band measurement has been synthesized
from the $R$-band or $V$-band measurement
\citep{2008ApJ...680..232D,2009ApJ...693..750B}, assuming a power-law of the
form $F_\nu \propto \nu^{-2}$. Arrows indicate 2-$\sigma$ limits. Galaxies
spanning the redshift range $1.5 < z < 2.5$ in the HDF-N (Papovich, personal
communication) and HDF-S \citep{2003AJ....125.1107L} are shown with grey dots.
Power-law DOGs have the reddest $I-H$ colors, followed by bump DOGs, XFLS
ULIRGs, and SMGs, which have colors comparable to high-$z$ HDF galaxies. }
\label{fig:imh}
\end{figure}
\subsection{Morphologies}
\subsubsection{Visual Classification Results} \label{sec:visual}
From the 7 users who entered classifications of the NICMOS images of DOGs, the
main results can be summarized as follows: power-law DOGs were classified as
irregular (43\%) approximately as frequently as they were classified regular
(42\%), with 15\% being too faint to tell. In contrast, bump DOGs were
classified as irregular significantly more often than they were classified as
regular (69\% vs. 26\%, with only 5\% being too faint to tell). These results
can be subdivided into those with very robust classifications (6 or more users
were in agreement), and less robust classifications (fewer than 6 users were in
agreement). The trends quoted earlier become stronger when considering only
the robust classifications, as the ratio of regular:irregular classifications
for this subset is 1.4:1 and 1:3 for power-law and bump DOGs, respectively.
Table~\ref{tab:vismorph} shows the breakdown of visual classifications with
this additional subdivision. In Table~\ref{tab:morph} we provide, for each DOG
in this sample, the mode of the classifications as well as how many users were
in agreement with the mode. Overall, the qualitative morphological assessment
indicates that bump DOGs have irregular, diffuse morphologies more frequently
than power-law DOGs.
\subsubsection{Non-parametric Classification Results} \label{sec:nonpar}
The characterization of galaxy morphologies requires high S/N imaging in order
to provide reliable results. For non-parametric forms of analysis, typical
requirements are ${\rm S/N_{pixel}} > 2$ and $r_{p}({\rm Elliptical}) > 2
\times {\rm FWHM}$ \citep{2004AJ....128..163L} (hereafter, $r_{\rm P}$
indicates the elliptical petrosian radius). In the case of the imaging
presented here, ${\rm FWHM} = 0\farcs16$. None of the 20 bump DOGs in this
study observed with WFPC2 have the per-pixel S/N necessary to compute $r_{\rm
P}$, $G$, $M_{20}$, and $C$. On the other hand, 18 out of 20 sources have
sufficient S/N in the NICMOS imaging. Table~\ref{tab:morph} presents the
visual and non-parametric measures of DOG morphologies, including per-pixel
S/N, $r_{\rm P}$, $G$, $M_{20}$, and $C$ values for the NICMOS images. This
table also includes an estimate of whether the DOG is dominated by a bump or by
a power-law in the mid-IR using IRAC data from \citet{2009ApJ...701..428A} and
the same statistical definition originally used by \citet{2008ApJ...677..943D}.
\begin{deluxetable*}{lcccccc}[!bp]
\tablecolumns{7}
\tablewidth{0in}
\tablecaption{Visual Morphological Classifications\label{tab:vismorph}}
\tablehead{
\colhead{} & \multicolumn{2}{c}{Regular} & \multicolumn{2}{c}{Irregular} &
\multicolumn{2}{c}{Too Faint Too Tell} \\
\colhead{} & \colhead{Agree} & \colhead{Disagree} & \colhead{Agree} &
\colhead{Disagree} & \colhead{Agree} & \colhead{Disagree}
}
\startdata
Power-law DOGs & 34\% & 9\% & 24\% & 18\% & 6\% & 9\% \\
Bump DOGs & 16\% & 10\% & 48\% & 21\% & 5\% & 0\% \\
\enddata
\end{deluxetable*}
Figure~\ref{fig:rpetmag} displays $C$ as a function of $r_{\rm P}$ for
power-law DOGs, bump DOGs, SMGs, and XFLS sources. The error bars indicate the
typical uncertainties in $C$ and $r_{\rm P}$ given the S/N and spatial
resolution associated with the imaging of each galaxy. The left panel of
Figure~\ref{fig:rpetmag}, focusing only on bump and power-law sources that
qualify as DOGs, shows that bump DOGs have larger sizes (median $r_{\rm P} =
8.4$~kpc, $\sigma_{r_{\rm P}} = 2.7$~kpc) than their power-law counterparts
(median $r_{\rm P} = 5.5$~kpc, $\sigma_{r_{\rm P}} = 2.3$~kpc). A two-sided
Kolmogorov-Smirnov (KS) indicates only a 1\% chance the two $r_{\rm P}$
distributions are drawn from the same parent distribution. The right panel of
Figure~\ref{fig:rpetmag} shows SMGs and XFLS sources which are not DOGs. In
this diagram, almost all sources are bumps, and almost all sources have large
sizes (median $r_{\rm P} = 8.5~$kpc, $\sigma_{r_{\rm P}} = 2.9$~kpc).
\begin{figure*}[!tbp]
\epsscale{1.00}
\plotone{f5.eps}
\caption{ $C$ as a function of $r_{\rm P}$ for $z > 1.4$ ULIRGs (symbols are
the same as in Figure~\ref{fig:samplecmd}). {\it Left}: Power-law DOGs, bump
DOGs, SMGs that qualify as DOGs, and XFLS ULIRGs at $z>1.4$ that qualify as
DOGs. Error bars illustrate the typical uncertainty level
given the S/N and spatial resolution associated with the image of each galaxy
\citep{2006ApJ...636..592L}. Bump DOGs have larger sizes than power-law DOGs.
{\it Right}: Same as left panel, but only for $z > 1.4$ ULIRGs (SMGs and XFLS)
that are not DOGs. Regardless of sample selection criteria, power-law $z >
1.4$ ULIRGs are significantly smaller than their bump counterparts (median
$r_{\rm P}$ of 5.6~kpc vs. 8.0~kpc, for the total respective populations).}
\label{fig:rpetmag}
\end{figure*}
When no consideration is given to their $R-[24]$ color, SMGs and XFLS sources
show a similar distinction in their sizes when dividing the samples into bump
(SMG median $r_{\rm P} = 8.6$~kpc, $\sigma_{r_{\rm P}} = 3.3$~kpc; XFLS median
$r_{\rm P} = 7.6$~kpc, $\sigma_{r_{\rm P}} = 2.9$~kpc) and power-law (SMG
median $r_{\rm P} = 4.6$~kpc, $\sigma_{r_{\rm P}} = 4.5$~kpc; XFLS median
$r_{\rm P} = 4.8$~kpc, $\sigma_{r_{\rm P}} = 1.5$~kpc) varieties. Indeed,
considering all $z>1.4$ ULIRGs regardless of whether they are selected at
mid-IR or sub-mm wavelengths, bump sources (median $r_{\rm P} = 8.4$~kpc,
$\sigma_{r_{\rm P}} = 2.9$~kpc) are significantly larger than their power-law
counterparts (median $r_{\rm P} = 5.6$~kpc, $\sigma_{r_{\rm P}} = 1.9$~kpc),
and a two-sided KS test indicates there is only a 1.3\% chance the two
populations could be drawn randomly from the same parent sample. This finding
is consistent with results from Keck $K$-band adaptive optics imaging of DOGs
which shows that power-law DOGs are smaller and more concentrated than bump
DOGs \citep{2009AJ....137.4854M}. One caveat with this result is that the bump
DOG sample is brighter in $H$-band than the power-law DOG sample. Considering
only the DOGs satisfying $H < 22.5$, the bump and power-law DOGs have similar
sizes ($r_{\rm P} \approx 8$~kpc). At the faint end ($H > 22.5$), power-law
DOGs are smaller than bump DOGs (5~kpc vs. 8~kpc, respectively).
The distribution in $G-M_{20}$ space derived from NICMOS imaging of power-law
DOGs, bump DOGs, XFLS sources, and SMGs is shown in Figure~\ref{fig:gm20}. The
error bars indicate the typical uncertainties in $G$ and
$M_{20}$ given the S/N and spatial resolution of the imaging of each galaxy. A sample of 73 local
ULIRGs ($z < 0.2$) is also shown in this diagram \citep{2004AJ....128..163L},
using data from {\it HST} WFPC2/F814W imaging \citep{2000ApJ...529L..77B}. The
dotted line separates major mergers from other types of galaxies and is based
on measurements at roughly the same rest-frame wavelength ($\approx
5000-5500$~\AA) of these 73 local ULIRGs \citep{2004AJ....128..163L}.
The left panel of Figure~\ref{fig:gm20} (including all sources that qualify as
DOGs) shows that bump DOGs appear offset to lower $G$ and higher $M_{20}$
values than power-law DOGs. The median \{$G$, $M_{20}$\} values for bump and
power-law DOGs are \{0.47, -1.08\} and \{0.49, -1.48\}, respectively. A
two-sided KS test indicates that there is only a 0.5\% chance that the two
$M_{20}$ distributions could have been drawn randomly from the same parent
distribution (the two $G$ distributions have a 10\% chance of being drawn from
the same parent distribution). These types of morphologies are consistent with
what is seen in simulations of major mergers during the beginning and end
stages, respectively, of the ``final coalescence'' of the merger when the SFR
peaks and begins to turn over \citep{2008MNRAS.391.1137L}. In the right panel
of Figure~\ref{fig:gm20}, SMGs and XFLS $z>1.4$ ULIRGs that are not DOGs are
shown. Although nearly all of these sources have bump SEDs, their morphologies
bear a greater resemblence to power-law DOGs than bump DOGs. The median \{$G$,
$M_{20}$\} values for the non-DOGs are \{0.52, -1.46\}.
\begin{figure*}[!tbp]
\epsscale{1.00}
\plotone{f6.eps}
\caption{Gini coefficient vs. $M_{20}$ derived from NIC2/F160W images of
high-redshift ULIRGs (symbols same as in Figure~\ref{fig:rpetmag}) and local
ULIRGs \citep[gray plus signs,][]{2004AJ....128..163L}. The evolution of a
typical gas-rich ($f_{\rm gas} = 0.5$) major merger during its peak SFR period
is illustrated by a green vector \citep{2008MNRAS.391.1137L}. The dashed line
is drawn qualitatively to separate ``diffuse'' and ``single-object''
morphologies and bisects the green vector. The dotted line shows the
empirically determined (based on measurements of local ULIRGs) demarcation line
above which objects are obvious major mergers \citep{2004AJ....128..163L}.
{\it Left}: Bump DOGs, power-law DOGs, and SMGs and XFLS ULIRGs qualifying as
DOGs. Within this highly obscured subset of the high redshift ULIRG
population, bump sources are ``diffuse'' (low $G$, high $M_{20}$) more often
than power-law DOGs. In simulations of major mergers, such morphologies occur
during the early half of the peak SFR period of the merger. {\it Right}: Same
as left panel, but for SMGs and XFLS $z>1.4$ ULIRGs that are not DOGs. The
distribution of morphologies for non-DOGs is skewed towards the
``single-object'' region of this diagram. These objects may occur during the
late stage of the peak SFR period of a major merger, or they may be associated
with more secular evolutionary processes.}
\label{fig:gm20}
\end{figure*}
The preceding analysis is largely qualitative in nature. A more quantitative
approach involves the use of contingency tables, which offer a means to
quantify broad-brush distinctions in the properties of two populations of
objects. Three properties are tested here: mid-IR SED shape (bump OR
power-law), extent of obscuration ($R-[24] > 14$ OR $R-[24] < 14$), and
morphology (low $G$, high $M_{20}$ OR high $G$, low $M_{20}$). The division
based on morphology is derived from simulations of major mergers, which
indicate that the high SFR period of a merger is bisected by a line described
by the equation $G = 0.4 M_{20} + 0.9$ \citep{2008MNRAS.391.1137L}.
Table~\ref{tab:contingency} shows the two 2$\times$2 contingency tables that
are needed to account for the three variables used in this analysis.
The first result from this analysis is the paucity of power-law sources in the
non-DOG subset. There are 29 bump DOGs, 23 bump non-DOGs, 31 power-law DOGs,
and only 1 power-law non-DOGs. The $2\times2$ contingency table for this
dataset indicates a negligible probability (Fisher Exact p-value $<0.0001$)
that all four sub-populations are drawn randomly from the same parent sample.
Could this be due to a selection effect? The non-DOG sample comprises ULIRGs
from the XFLS and SMGs. XFLS sources are selected to have high $F_{\rm 24\mu
m} / F_{\rm 8\mu m}$ flux density ratios, which tends to favor the selection of
bump SEDs over power-law ones. On the other hand, the XFLS sources are
selected to be very bright at 24$\mu$m ($F_{\rm 24\mu m} > 0.8$~mJy. At these
24$\mu$m flux densities, power-law sources are more common than bump sources
\citep[e.g.][]{2008ApJ...677..943D}. SMGs are selected at sub-mm wavlengths,
without any knowledge of the mid-IR SED shape. Presently, it is not obvious
that either the XFLS ULIRGs or SMGs are affected by the kind of severe
selection effect necessary to produce the observed trends.
The second result from the contingency table data is that, considering only
bump sources, non-DOGs have a much more skewed distribution of morphologies
than DOGs. Diffuse type morphologies (low $G$, high $M_{20}$) are rare in the
non-DOG population, while in DOGs they occur much more frequently. A
$2\times2$ contingency table here suggests a very low probability (Fisher Exact
p-value $=0.007$) that blue ($R-[24]<14$) and red ($R-[24]>14$) ULIRGs have
morphologies drawn from the same parent distribution. Low $G$ and high
$M_{20}$ values suggest irregular and lumpy (less centrally concentrated)
morphologies that could be caused by a clumpy distribution of stars or
significant dust obscuration \citep{2008MNRAS.391.1137L}. Further discussion
of the implications of this result are deferred to section~\ref{sec:disc}.
Finally, with the highly obscured subset of ULIRGs (DOGs), there is evidence
that bump DOGs have diffuse type morphologies more commonly than power-law
DOGs. A $2\times2$ contingency table indicates an extremely low probability
(Fisher Exact p-value $=0.003$) that bump and power-law DOGs have morphologies
drawn from the same parent distribution. As mentioned earlier, this
distinction is consistent with expectations from simulations of major mergers
during the peak SFR phase of the merger \citep{2008MNRAS.391.1137L}.
\begin{deluxetable}{lcccc}
\tablecolumns{5}
\tablewidth{3.5in}
\tablecaption{NICMOS Morphology Contingency Table Data}
\tablehead{
\colhead{} & \multicolumn{2}{c}{$R-[24]<14$} & \multicolumn{2}{c}{$R-[24]>14$} \\
\colhead{} &
\colhead{Diffuse\tablenotemark{a}} &
\colhead{Single-source\tablenotemark{b}} &
\colhead{Diffuse\tablenotemark{a}} &
\colhead{Single-source\tablenotemark{b}}
}
\startdata
Power-law & 0 & 1 & 7 & 24 \\
Bump & 3 & 20 & 15 & 14 \\
\enddata
\tablenotetext{a}{$G < 0.4 M_{20} + 0.9$}
\tablenotetext{b}{$G > 0.4 M_{20} + 0.9$}
\label{tab:contingency}
\end{deluxetable}
\subsubsection{GALFIT Results}\label{sec:GALFIT}
The results of our GALFIT analysis of the NICMOS images of the Cycle~16 DOGs
are shown in Table~\ref{tab:morph}, along with 1-$\sigma$ uncertainties in the
best-fit parameters. Included in this table are point source fractions (ratio
of flux in the point-source component to the total flux of the source),
effective radius of the S\'ersic component ($R_{\rm eff}$), S\'ersic index
($n$), semi-minor to semi-major axis ratio of the S\'ersic component (Axial
Ratio), number of degrees of freedom ($N_{\rm DOF}$), and reduced chi-squared
($\chi_\nu^2$).
Figure~\ref{fig:reffrpet} shows a comparison of $R_{\rm eff}$ (the radius
within which half the light is enclosed) and $r_{\rm P}$ (the radius at which
the ratio of the surface brightness at $r_{\rm P}$ to the mean surface
brightness within $r_{\rm P}$ is equal to 0.2) for DOGs, SMGs, and XFLS ULIRGs
at $z > 1.4$. For bump DOGs and power-law DOGs, the median $R_{\rm eff}$
values are 3.3~kpc and 2.5~kpc, respectively. Bump sources that are not DOGs
(from the SMG and XFLS samples) have a median effective radius of 3.2~kpc. One
of the bump DOGs (SST24~J143137.1+334500) has the appearance of an edge-on disk
with a semi-major axis of $3\farcs25$, or 27.5~kpc at its redshift of 1.77.
This extremely large $R_{\rm eff}$ value may imply that this object is in fact
a merger viewed edge-on. Spatially resolved dynamical information would be
particularly useful for answering this question.
\begin{figure*}[!tbp]
\epsscale{1.00}
\plotone{f7.eps}
\caption{Comparison of sizes of $z > 1.4$ ULIRGs (symbols same as in
Figure~\ref{fig:rpetmag}) as determined by the effective radius of the
S\'ersic component from GALFIT modeling ($R_{\rm eff}$) and the elliptical
Petrosian radius ($r_{\rm P}$). Error bars represent 1$\sigma$ uncertainty
values from GALFIT. {\it Left}: Bump DOGs, power-law DOGs, and
SMGs and XFLS ULIRGs qualifying as DOGs. {\it Right}: SMGs and XFLS ULIRGs
that are not DOGs. Both size measurements suggest that power-law sources are
on average smaller than bump sources, although a significant population of
compact bump sources exists. }
\label{fig:reffrpet}
\end{figure*}
Our measurements of SMG sizes (median $R_{\rm reff}$ value for the full SMG
population of 3.6~kpc) are in broad agreement, given the different methods
used, with those of \citet{2010MNRAS.405..234S}, who find typical half-light
radii of $2.8\pm0.4$~kpc. For XFLS ULIRGs, \citet{2008ApJ...680..232D} use
GALFIT to find typical effective radii of $2.43\pm0.80$~kpc, consistent with
our results (median $R_{\rm eff}$ of 2.5~kpc). As an additional consistency
check, a strong correlation is evident between $R_{\rm eff}$ and $r_{\rm P}$
for all populations. Note that $r_{\rm P} > R_{\rm eff}$; this is because the
S\'ersic profile is defined such that half of the galaxy's flux is enclosed
within a radius of $r=R_{\rm eff}$, while $r_{\rm P}$ defines the radius at
which the surface brightness is one-fifth the average surface brightness within
$r_{\rm P}$.
Figure~\ref{fig:fpsfn} shows the point source fraction and S\'ersic index for
DOGs, SMGs, and XFLS ULIRGs at $z > 1.4$. The majority of sources have low
point source fractions (point source fraction $<0.3$) and disk-type
morphologies ($n < 2$). Studies have found that when a point source
contributes less than 20\% of the total light, it has an insignificant effect
on the measured morphologies \citep{2010MNRAS.tmp..455P}. Considering only
DOGs with sufficient S/N to be placed on this diagram (left panel of
Figure~\ref{fig:fpsfn}), 6/28 power-law DOGs and 0/17 bump DOGs have either $n
> 3$ or point source fraction $>0.4$. Such sources have compact, centrally
dominated morphologies \citep[$n=1$ corresponds to an exponential profile, and
$n=4$ corresponds to a de Vaucouleurs profile;][]{2002AJ....124..266P}. This
distinction is consistent with the $G$ and $M_{20}$ results in
section~\ref{sec:nonpar}.
\begin{figure*}[!tbp]
\epsscale{1.00}
\plotone{f8.eps}
\caption{S\'ersic index $n$ as a function of point source fraction from GALFIT
modeling (symbols same as in Figure~\ref{fig:reffrpet}). {\it Left}:
Power-law DOGs, bump DOGs, and SMGs and XFLS ULIRGs qualifying as DOGs. Aside
from a handful of power-law DOGs with point source fraction $>0.4$ or $n >
2.5$, there is strong overlap between the bump and power-law DOG populations in
this diagram. {\it Right}: SMGs and XFLS ULIRGs that do not qualify as DOGs.
In contrast to the DOG populations, there are a number of $n>2.5$ bump sources
from the SMG and XFLS samples. As in the analysis of the $G$ and $M_{20}$
values, these could represent objects at the end of the peak SFR period, or
they might not be associated with major merger activity at all. }
\label{fig:fpsfn}
\end{figure*}
On the other hand, the distinction between bump and power-law sources is not as
obvious when considering the SMGs and XFLS sources. For SMGs, 2/3 power-law
and 2/11 bump sources satisfy the compact criteria outlined above, while for
XFLS ULIRGs the respective numbers are 2/6 (power-law sources) and 3/18 (bump
sources). Further discussion of the distinction between the morphological
properties of bump and power-law DOGs is deferred to section~\ref{sec:disc}.
\section{Discussion: Implications for Models of Massive Galaxy
Evolution}\label{sec:disc}
ULIRG activity in the local universe has been known for some time to result
from a major merger of two gas-rich disk galaxies
\citep[e.g.][]{1987AJ.....94..831A, 1988ApJ...325...74S}. Material is funneled
towards the center of the system and drives an intense starburst, producing
large amounts of cold dust, and begins to feed a nascent central black hole.
As the merger evolves, ambient gas and dust particles are heated by feedback
processes. This warm-dust ULIRG stage has been suggested to represent a
transition stage between cold ULIRGs and optically luminous quasars
\citep{1988ApJ...328L..35S}.
Recently, efforts have been made to extend this paradigm to the ultra-luminous
galaxy populations at high-redshift. One possible hypothesis within this
scenario is that SMGs represent the cold-dust ULIRGs created during the early
stage of the merger, whereas {\it Spitzer}-selected sources represent the
warm-dust ULIRGs formed during the later stages of the merger
\citep[e.g.,][]{2008ApJ...677..943D,2009ASPC..408..411D,2009arXiv0910.2234N}.
This basic picture (that SMGs and {\it Spitzer}-selected ULIRGs are related) is
strengthened by the similarity in the measured clustering strengths of $z
\approx 2$ SMGs, DOGs, and QSOs, which suggest that these populations all
reside in similar mass halos at similar epochs
\citep[e.g.,][]{2008ApJ...687L..65B, 2009ASPC..408..411D}.
In this section, we test the viability of this scenario using the morphological evidence presented in
section~\ref{sec:results}. On one hand,
when considering only the most extremely obscured objects (i.e., DOGs), a clear
trend in morphologies emerges. Bump DOGs are larger (i.e., more spatially
extended) than power-law DOGs ($r_{\rm P} \approx 8$~kpc vs. 5~kpc), more
diffuse ($\{G,M_{20}\} \approx \{0.47, -1.08\}$ vs. $\{G,M_{20}\} \approx
\{0.49, -1.48\}$), and more irregular (67\% vs. 50\% visually classified as
irregular). This trend is consistent with expectations from simulations of
major mergers, which indicate that merger morphologies generally evolve from
extended, diffuse, and irregular at the beginning of the peak SFR phase to
compact and regular when star-formation shuts down and the AGN begins to
dominate \citep{2008MNRAS.391.1137L,2009arXiv0910.2234N}.
On the other hand, the less obscured sources (non-DOGs from the SMG and XFLS
sample) show two strong distinctions from their more extreme counterparts.
First, there are very few power-law non-DOGs. If power-law SEDs are more
frequently associated with objects that are more dust reddened, this may imply
a connection between the amount of extinction of the optical light and the
nature of the power source producing the mid-IR emission.
Second, within the bump population of non-DOGs, there are very few diffuse type
morphologies (low $G$, high $M_{20}$). The prevalence of bump sources with
``single-object'' morphologies is difficult to understand within the context of
a major merger scenario in which bump sources evolve into power-law sources.
If the bump phase always precedes the power-law phase, there should be very few
bump sources with compact, single-object morphologies. A number of potential
explanations exist.
Perhaps the most exciting explanation is that high redshift ULIRGs are related
to one another within a single evolutionary scheme driven by major mergers, but
with an additional wrinkle related to the degree of obscuration. During the
highly dust-obscured period of the merger (represented jointly by both bump and
power-law DOGs), the bump phase typically occurs before the power-law phase.
In contrast, the less obscured sources (SMGs and XFLS ULIRGs) sample the merger
over a broader timescale and so the relationship between bump and power-law
sources is not as obvious. For example, there may be a significant population
of blue ULIRGs (non-DOGs) that correspond to the systems near the very end of the high SFR period
of the merger when the obscuring column of dust has decreased and UV light can
escape the galaxy.
An alternative, but potentially equally exciting, way to reconcile the
morphological evidence is by appealing to more quiescent modes of galaxy
assembly for some fraction of the high redshift ULIRG population
\citep[e.g.][]{2008ApJ...687...59G}. Recent theoretical work has suggested
that many SMGs may be produced not by major mergers, but instead by smooth gas
inflow and the accretion of small gas-rich satellites
\citep{2010MNRAS.tmp..360D}. Such an explanation would be surprising, given
the evidence already in place favoring a major merger origin for SMGs largely
based on dynamical and kinematic arguments \citep[e.g.][]{2005MNRAS.359.1165G,
2006MNRAS.371..465S, 2008ApJ...680..246T,2010ApJ...724..233E}. While there is
no definitive evidence in the data presented here that can unambiguously
support this smooth inflow mode of galaxy formation, the relatively normal
morphologies observed in the non-DOGs could suggest that major mergers are not
responsible for driving the prodigious on-going star-formation in these
systems. Given that such intense star-formation bursts can only be sustained
over a short timescale, the morphologies suggest that the fuel may have to be
accreted in less disruptive minor mergers or through some smooth process.
Physical mechanisms explaining how such a process might occur have been
presented recently \citep{2010ApJ...719..229G,2010arXiv1011.0433G}.
Observations of the internal dynamics of these systems \citep[along the lines
of, e.g.,][]{2008ApJ...687...59G,2010arXiv1011.1507F,2010arXiv1011.5360G} are
likely what is needed to continue progress in this area of research.
A third possibility is that the expected trends in morphologies with merger
stage are somewhat sensitive both to the initial conditions of the merger ---
for example, highly radial orbits can have similar $G$ and $M_{20}$ values
throughout the ``final merger'' stage \citep{2008MNRAS.391.1137L} --- as well
as the viewing time and angle. It would be surprising if unusual initial
conditions or viewing times and angles were necessary to explain most high
redshift ULIRGs, particularly since they appear to have fairly typical axial
ratios (see Table~\ref{tab:morph}).
An important consideration related to the XFLS ULIRGs and SMGs analyzed here is
that many of these objects are composite starburst and AGN systems with complex
mid-IR spectral features. \citet{2008ApJ...680..232D} show that the 7.7$\mu$m
PAH feature is usually strong in extended sources, while it varies from strong
to weak in compact sources. The mid-IR spectral analysis of these sources
\citep{2007ApJ...664..713S} indicates that only a few XFLS ULIRGs are clearly
dominated by PAH features or AGN continuum emission. This result is consistent
with the nature of their mid-IR SEDs and underscores the fact that these
objects are composite systems that are not easily classified by either their
mid-IR spectral features or their rest-frame optical morphologies. Only 7 SMGs
in the sample studied here have both high-resolution imaging and mid-IR
spectroscopy \citep{2009ApJ...699..667M}. Of these 7, all are bump sources, 4
have strong PAH emission, and 3 have weak or no PAH emission. It may be the
case that the mid-IR SEDs of the SMG and XFLS ULIRG samples are not
sufficiently distinct to identify significant morphology differences in the
bump vs. power-law sub-samples.
\section{Conclusions} \label{sec:conclusions}
We have used {\it HST} imaging to analyze the morphologies of 22 DOGs at $z
\approx 2$ from the Bo\"{o}tes field selected to show SED features typical of
star-formation dominated systems (bump DOGs). We compare these new data with
similar {\it HST} imaging of DOGs with SED features typical of AGN-dominated
systems (power-law DOGs), sub-millimeter galaxies (SMGs), and a sample of
ULIRGs at high-$z$ selected from the {\it Spitzer} XFLS. Our findings are
summarized below.
\begin{enumerate}
\item Spatially resolved emission is observed in the rest-frame optical imaging
of all bump DOGs. GALFIT modeling indicates that the point source fraction
(ratio of flux in the point-source component to total flux of the source)
in these objects never exceeds 20\% and is typically smaller than that
found in power-law DOGs, suggesting a smaller AGN contribution to the
rest-frame optical light from bump DOGs.
\item Typical S\'ersic indices of the resolved emission of bump DOGs suggest
disk-type rather than bulge-type profiles ($n<2$), similar to power-law
DOGs.
\item At $H < 22.5$, bump and power-law DOGs have similar sizes (median $r_{\rm
P} = 8$~kpc). At $H > 22.5$, bump DOGs are significantly larger than
power-law DOGs (median value of $r_{\rm P} =8$~kpc vs. $r_{\rm P} =
5.4$~kpc, respectively). This distinction is also true for SMGs and XFLS
ULIRGs.
\item In the rest-frame optical, bump DOGs have lower $G$ and higher $M_{20}$
values than power-law DOGs. This difference is consistent with
expectations from simulations of major mergers. On the other hand, less
obscured objects in our sample (SMGs and XFLS ULIRGs that do not qualify as
DOGs) have high $G$ and low $M_{20}$ values that are more typical of
``single-object'' systems.
\end{enumerate}
Overall, our findings highlight the diversity and complexity of high redshift
ULIRG morphologies. Within the highly obscured subset (i.e., DOGs), we find
evidence in support of a major merger paradigm in which bump DOGs evolve into
power-law DOGs. Within the less obscured subset (i.e., SMGs and XFLS ULIRGs),
the picture is not as clear. This may be a result of the timescales over which
obscured and less obscured sources can be observed during a major merger. Alternatively,
that the intense star-formation in these less-obscured ULIRGs is not the result
of a recent major merger, and
may be an indication that more quiescent forms of galaxy assembly are important
for some high redshift ULIRGs.
The work is based primarily on observations made with the {\it Hubble Space
Telescope}. This work also relies in part on observations made with the {\it
Spitzer Space Telescope}, which is operated by the Jet Propulsion Laboratory,
California Institute of Technology under NASA contract 1407. We are grateful to
the expert assistance of the staff Kitt Peak National Observatory where the
Bo\"{o}tes field observations of the NDWFS were obtained. The authors thank
NOAO for supporting the NOAO Deep Wide-Field Survey. In particular, we thank
Jenna Claver, Lindsey Davis, Alyson Ford, Emma Hogan, Tod Lauer, Lissa Miller,
Erin Ryan, Glenn Tiede and Frank Valdes for their able assistance with the
NDWFS data. We also thank the staff of the W.~M.~Keck Observatory, where some
of the galaxy redshifts were obtained.
We gratefully acknowledge the anonymous referee whose helpful suggestions have
resulted in an improved manuscript. RSB gratefully acknowledges financial
assistance from HST grants GO-10890 and GO-11195, without which this research
would not have been possible. Support for Program numbers HST-GO10890 and
HST-GO11195 were provided by NASA through a grant from the Space Telescope
Science Institute, which is operated by the Association of Universities for
Research in Astronomy, Incorporated, under NASA contract NAS5-26555. The
research activities of AD and BTJ are supported by NOAO, which is operated by
the Association of Universities for Research in Astronomy (AURA) under a
cooperative agreement with the National Science Foundation. Support for E. Le
Floc'h was provided by NASA through the Spitzer Space Telescope Fellowship
Program.
|
\section{Introduction}
Galaxy orientations contain important information about the gravitational environment in which it resides. \citet{brown38} pointed out that galaxy orientations may not be isotropic due to the large scale gravitational interaction. \citet{hawley75} reported a weak evidence of anisotropy of the galaxy orientation. Galaxy orientation becomes especially interesting in the vicinity of galaxy clusters, where the strong gravitational field may produce detectable orientation preference for both central galaxies and satellite galaxies. With the advent of modern sky surveys, such as Sloan Digital Sky Survey (SDSS)~\citep{york00}, the shape and orientation of galaxies can be measured to high precision. The large sky coverage substantially increases the sample size to allow statistically significant measurements on cluster galaxy alignments, from which our understanding of the cluster formation process can be greatly improved.
The term galaxy alignment has been used extensively in the literature and refers to alignments in different contexts. In the galaxy cluster environment, there are two types of alignments that are of great interest. The first is the alignment of the major axes of the cluster satellite galaxies towards the cluster center, which we will call ``satellite alignment''. In our case, as an operational definition, we will consider the alignment between the major axes of the satellite galaxies and the BCG\footnote{Though cluster center is well defined as the deepest gravitational potential well, its determination from observational data, especially optical data, is not unambiguous. The central galaxy in a cluster (the one which resides near the bottom of the cluster potential well) is very often the brightest galaxy (BCG) in the cluster. This BCG is then coincident with the region with the deepest potential traditionally identified in theory as the center of a cluster. Using the BCG as the cluster center simplifies precise comparisons between observations and theory.}. The second type of alignment is the alignment of the BCGs major axis towards the distribution of satellite galaxies in the cluster. We will call this alignment ``BCG alignment'' hereafter.
In addition to these two types of alignments, the possible alignments between the major axises of the satallite galaxies and the cluster~\citep{plinois04}, and between the cluster shape and large scale structures~\citep{paz08,flatenbacher09,wangyougang09,paz11} have been studied, but these are beyond the scope of this current paper.
There are extensive studies on these two types of alignments with both simulations and observations. It is argued, based on simulations, that the preferred accretion direction of satellite halos toward the host halo along the filaments is largely responsible for the BCG alignment~\citep{tormen97,vitvitska02,knebe04,zentner05,wang05}. The detections of this alignment from real data has been reported by many teams~\citep{sastry68,austin74,dressler78,carter80,binggeli82,brainerd05,yang06,azzaro07,wangyougang08,siverd09,ostholt10}, though non-detection of this alignment were also reported~\citep{tucker88,ulmer89}. For the satellite alignment, tidal torque is thought to play a major role in its formation~\citep{ciotti94,ciotti98,kuhlen07,pereira08,faltenbacher08,pereira10}. Its detection based on SDSS data has been reported by~\citet{pereira05, agustsson06,faltenbacher07}, while non-detections of this alignment are also reported based on data from both 2dF Galaxy Redshift Survey (2dFGRS)~\citep{colless01,bernstein02} and SDSS~\citep{siverd09}. In Table~\ref{table:summary}, we summarize the previous work that reports the existence and non-existence of these two types of alignments based on real data. On the other hand, the intrinsic alignment of galaxies will contaminate gravitational lensing measurements and therefore needs to be carefully modeled in lensing analysis. Along these lines,~\citet{mandelbaum06,hirata07} reported correlations between intrinsic shear and the density field based on data from SDSS and 2SLAQ~\citep{croom09}.
In general, the cluster galaxy alignment signals are weak, and their measurement requires high quality photometry and well measured galaxy PAs. Moreover, a galaxy cluster catalog with high purity, well-determined BCGs and satellite galaxies is important too. Since measuring the redshift evolution of the alignment is crucial for understanding its origin, the cluster catalog needs to be volume limited and maintain constant purity for a wide redshift range. Most of the existing alignment measurements (see Table~\ref{table:summary}) are based on galaxy clusters/groups selected from spectroscopic data. In SDSS data, due to the high cost of obtaining spectra for a large population of galaxies, the completeness of spectroscopic coverage is limited to $r$ band Petrosian magnitude $r_p \le$17.7, corresponding to a median redshift of 0.1~\citep{strauss02}. This greatly limits the ability to look at the redshift evolution of the alignments.
On the other hand, one can also measure the alignments by using photometrically selected clusters. The advantage of photometrically selected clusters lies in the large data sample as well as relatively deep redshift coverage, allowing a study on the redshift evolution of the alignments. However, there are clear disadvantages too. For example, the satellite galaxies are prone to contamination from the projected field galaxies, which will dilute the alignment signals. The level of this contamination may also vary as redshift changes, complicating the interpretation of the alignment evolution.
In this paper, we show our measurements of the two types of alignments based on a volume limited and highly pure ($\ge$ 90\%) subsample of clusters from the GMBCG cluster catalog for SDSS DR7~\citep{haocat}. The large sample of clusters allows us to examine the dependence of the alignments on cluster richness and redshift with sufficient statistics. With this catalog, we detect a BCG alignment that depends on redshift and the absolute magnitude of BCG, but not on the cluster richness. We also observe that the satellite alignment depends on the apparent brightness of the BCG and the methods the position angles are measured. We can only see a statistically significant satellite alignment at low redshift when we use the isophotal fit PAs (see \S~3.3 for more details). Furthermore, we notice that the satellite alignment based on isophotal fit PAs depends strongly on the apparent magnitude rather than the absolute magnitude of the BCG. This suggests that the measured satellite alignment is more likely due to the isophotal fit PAs, whose measurements are prone to contaminations from the diffuse light of the BCG.
The paper is organized as following: in \S 2, we introduce the two parameters used to quantify the two types of alignments. In \S~3, we introduce the data used in this paper. In \S~4, we present and discuss our measurement results. By convention, we use a $\Lambda$CDM cosmology with $h=1.0$, $\Omega_m=0.3$ and $\Omega_{\Lambda}=0.7$ throughout this paper. All angles measurements that appear in this paper are in units of degrees.
\begin{deluxetable*}{l l|l|l|r}
\tabletypesize{\tiny}
\tablecolumns{5}
\tablewidth{0pt}
\tablecaption{Alignment Measurements Summary\label{table:summary}}
\tablehead{
\multicolumn{1}{c}{Data }&
\multicolumn{2}{c}{BCG Alignment}&
\multicolumn{2}{c}{Satellite Alignment}\\
\colhead{Source}& \colhead{Exist}& \colhead{Non-exist} & \colhead{Exist} & \colhead{Non-exist}
}
\startdata
&\citet{sastry68} &\citet{tucker88} & &\\
&\citet{austin74} &\citet{ulmer89} & &\\
Photometric &\citet{dressler78} & & &\\
Plates &\citet{carter80} & & &\\
&\citet{binggeli82} & & &\\ \tableline
&\citet{brainerd05} & &\citet{pereira05} & \citet{siverd09}\\
SDSS &\citet{yang06} & &\citet{agustsson06} &\\
Spectrosopic &\citet{azzaro07} & &\citet{faltenbacher07} &\\
&\citet{faltenbacher07} & & &\\
&\citet{wangyougang08} & & &\\
&\citet{siverd09} & & &\\ \tableline
SDSS &\citet{ostholt10} & & \citet{pereira05} &\\
Photometric &\color{red}{This Work} & & &\color{red}{This Work}\\ \tableline
2dF & & & &\citet{bernstein02}\\
Spectroscopic & & & &
\enddata
\end{deluxetable*}
\section{Alignment parameters}
In this paper, we consider two types of alignments: (1) Satellite alignment; (2) BCG alignment. Each of them is quantified by a corresponding alignment parameter. For the satellite alignment, we follow~\citet{struble,pereira05} and use the following alignment parameter:
\begin{equation}\label{eq:delta}
\delta = \frac{\sum^{N}_{i=1}{\phi_i}}{N} - 45
\end{equation}
\noindent where $\phi_i$ is the angle between the major axes of the satellite galaxies and the lines connecting their centers to the BCGs, as illustrated in the left panel of Figure~\ref{fig:aangle}. N is the number of satellite galaxies in the cluster. For every cluster, there will be a unique alignment parameter $\delta$ measured, which is the mean angle $\phi$ of all cluster satellite galaxies subtracted by 45. If the major axes of satellite galaxies do not preferentially point to the BCG of the cluster, the $\phi$ will randomly distribute between -45 and 45 (degrees), leading to $\delta = 0$. On the other hand, if the major axes of satellite galaxies preferentially point to the BCG, there will be $\delta < 0$. The standard deviation (error bar) of $\delta$ can be naturally calculated by $\delta_{err} = \sqrt{\sum_{i=1}^N{(\phi_i-\delta-45)^2}}/N$.
For the BCG alignment, we will focus on the angle $\theta$ between the BCGs PA and the lines connecting the BCG to each satellite galaxy, as illustrated in the right panel of Figure~\ref{fig:aangle}. In our cluster sample, each cluster has more than 15 satellite galaxies (see \S 3.2 for more details). Therefore, instead of looking at the full distribution of $\theta$ from all clusters, we will focus on the mean of $\theta$ measured for each cluster. In analogy to the satellite alignment parameter $\delta$, we introduce a BCG alignment parameter $\gamma$ defined as:
\begin{equation}\label{eq:gamma}
\gamma = \frac{\sum^{N}_{i=1}{\theta_i}}{N} - 45
\end{equation}
\noindent where $\theta_i$ is the angle between BCGs PA and the line connecting the BCG to the $i^{th}$ satellite galaxy (see the right panel of Figure~\ref{fig:aangle}). N is the number of satellite galaxies in the cluster. Similarly, each cluster will correspond to a BCG alignment $\gamma$, which is the mean of angle $\theta$ subtracted by 45. If the major axis of BCG preferentially aligns with the distribution of the majority of the satellite galaxies, the $\gamma < 0$. If no such preference exist, $\gamma = 0$ will be expected. The uncertainty of $\gamma$ can be readily calculated as $\gamma_{err} = \sqrt{\sum_{i=1}^N{(\theta_i-\gamma-45)^2}}/N$.
\begin{figure*}
\epsscale{0.9}
\begin{center}
\plottwo{alignment_def_mb_bcg.pdf}{alignment_def_bcg_mb.pdf}
\caption{\textit{left:} Illustration of the angle $\phi$ used in the definition of satellite alignment parameter $\delta$; \textit{right:} Illustration of the angle $\theta$ used in the definition of BCG alignment parameter $\gamma$. }
\label{fig:aangle}
\end{center}
\end{figure*}
These two parameters quantify the two types of alignments and are easy to measure. In the follows, we will focus on these two quantities and their dependencies on various cluster/BCG properties.
\section{Data}
In this section, we describe the details of the galaxies, their PA measurements, and the galaxy cluster sample used in our measurement.
\subsection{Galaxy Catalog}
The galaxies we use are from the Data Release 7 of the Sloan Digital Sky Survey~\citep{york00, abazajian08}. The SDSS is a multi-color digital CCD imaging and spectroscopic sky survey, utilizing a dedicated 2.5-meter telescope at Apache Point Observatory, New Mexico. It has recently completed mapping over one quarter of the sky in $u$,$g$,$r$,$i$ and $z$ filters. DR7 is a mark of the completion of the original goals of the SDSS and the end of the phase known as SDSS-II. It includes a total imaging area of 11663 square degrees with 357 million unique objects identified.
This work focuses on the Legacy Survey area of SDSS DR7, which covers more than 7,500 square degrees of the North Galactic Cap, and three stripes in the South Galactic Cap totaling 740 square degrees~\citep{abazajian08}. The galaxies are selected from the PhotoPrimary view of the SDSS Catalog Archive Server with object type tag set to 3 (galaxy) and $i$-band magnitude less than 21.0. Moreover, we require that the galaxies did not trigger the following error flags: SATURATED, SATUR\_CENTER, BRIGHT, AMOMENT\_MAXITER, AMOMENT\_SHIFT and AMOMENT\_FAINT.
In addition to the above selection criteria, we also reject those galaxies with photometric errors in $r$ and $i$ band greater than 10 percent. Additionally, we require the ellipticity\footnote{The ellipticity is defined as $\mathrm{\sqrt{m_{e1}^2 +m_{e2}^2}}$, where the $\mathrm{m_{e1} = \frac{<col^2> - <row^2>}{<col^2> + <row^2>}}$ and $\mathrm{m_{e2} = \frac{2 <col*row>}{<col^2> + <row^2>}}$. Details of estimating $\mathrm{<...>}$ can be found in \citep{bernsteinjarvis02}} of each galaxy in the $r$-band and $i$-band to be less than 0.8 in order to remove edge-on galaxies whose colors are not well measured. By doing this, we will retain about 95\% of the total galaxies. All the magnitudes used in this paper are dust extinction corrected model magnitudes~\citep{abazajian08}.
\subsection{Galaxy Clusters}
In order to measure the two types of alignments, we need to have a galaxy cluster catalog with well determined member galaxies. The GMBCG cluster catalog for SDSS DR7 is a large catalog of optically selected clusters from SDSS DR7 using the GMBCG algorithm. The catalog is constructed by detecting the BCG plus red sequence feature that exists among most clusters. The cluster satellite galaxies are within 2$\sigma$ of the red sequence mean color detected using a Gaussian Mixture Model. Since the red sequence of each cluster is measured individually, it allows a more accurate satellite galaxy selection than using a universal red sequence model. We count the satellite galaxies down to 0.4L* at the cluster's redshift. In Figure~\ref{fig:gmr_bimodal}, we show the color distribution of galaxies around a cluster in GMBCG catalog. To study the redshift evolution of the alignment, we choose the volume limited sample of clusters with redshift below 0.4, where the satellite galaxies are selected using $g-r$ color. Also, to get high purity, we choose clusters with richness equal to or greater than 15, which leads to a purity above 90\% across the redshift range. As a result, there are about 11,000 clusters with over 260,000 associated satellite galaxies. More details of the cluster catalog can be found in~\citet{haocat}.
\begin{figure*}
\begin{center}
\epsscale{0.8}
\plotone{colorspace.pdf}
\caption{{\em left:} Galaxy $g-r$ color distribution around a cluster
overlaid with a model constructed of a mixture of two Gaussian
distributions. The red curve corresponds to the red sequence
component while the blue one corresponds to the sum of background
galaxies and blue cluster satellites. The green vertical line indicates
the color of the BCG. $\mu$ and $\sigma$ are the means and standard
deviations of the two Gaussian components. {\em right:}
Color-magnitude relation for the same galaxies. Galaxies within the
2$\sigma$ clip of the red sequence component are shown with red
points; the green line indicates the best fit slope and intercept of
this red seqence.}
\label{fig:gmr_bimodal}
\end{center}
\end{figure*}
\subsection{PA Measurements}
In the SDSS data reduction pipeline, the PAs of galaxies are measured with several different methods~\citep{stoughton02}. In this work, we will use three of them: isophotal PA, exponential fit PA~\footnote{Since most satellite galaixes are selected using red sequence, the exponential fit may not be suitable. In this paper, we want to demonstrate how will the PA measurement affect the alignment signal, and therefore include it in our discussion.} and De Vaucouleurs fit PA.
To measure the isophotal PA, the SDSS pipeline measures out to the 25 magnitudes per square arc-second isophote (in all bands). The radius of a particular isophote as a function of angle is measured and Fourier expanded. From the coefficients, PA (isoPhi) together with the centroid (isoRowC,isoColC), the major and minor axes (isoA,isoB) are extracted.
For the exponential fit PA, the SDSS pipeline fit the intensity of galaxy by
\begin{equation}
I(R)=I_0\exp[-1.68 (R/R_{eff})]
\end{equation}
\noindent where the profile is truncated outside of 3$R_{eff}$ and smoothly decreases to zero at 4$R_{eff}$. After correcting for the PSF, the PA is calculated from this fitting and reported as expPhi in the CASJOB database.
The De Vaucouleurs fit PA follows the same procedure as exponential fit PA, except the model in the fitting is
\begin{equation}
I(R)=I_0\exp[-7.67 (R/R_{eff})^{1/4}]
\end{equation}
\noindent where the profile is truncated outside of 7$R_{eff}$ and smoothly decreases to zero at 8$R_{eff}$. The PA from this fitting is reported as devPhi. For more details about these PA measurements, one can refer to~\citet{stoughton02}. All the angles are in degrees East of North by convention. By comparing these different PAs, isophotal PA tends to trace the exterior shape of the galaxy while the two model fit PAs tend to trace the inner profile of the galaxy. At low redshift, the BCG is very bright and its diffuse light may severely affect the measurement of the outer part of the nearby galaxies. Therefore, the isophotal PA is more susceptible to this artifact, leading to an ``artificial'' orientation preference toward BCG. To give a sense of the BCG diffuse light effect, we show a low redshift and a higher redshift cluster in Figure~\ref{fig:diffuse}.
\begin{figure*}
\epsscale{0.75}
\begin{center}
\plottwo{0103.pdf}{0310.pdf}
\caption{Left is a cluster at redshift 0.103 and the right is a cluster at redshift 0.310. The light from BCG will pose a risk on the proper measurements of the PAs of the satellite galaxies, especially at low redshift.}
\label{fig:diffuse}
\end{center}
\end{figure*}
\subsection{Control Sample}
As we talk about the detection of the alignment signals, we need to have a control sample to compare with. Naively, one may directly compare the measured alignment signal to what is expected for non-detection. However, this will implicitly assume that the data as well as the measurements are free from any systematics. When talking about the significance level of the detection, it is not sufficient to consider only the statistical uncertainties. Systematics (intrinsic scatter), if they exist, also need to be considered. Therefore, introducing an appropriate control sample and applying the same measurements we apply to the cluster data will help to eliminate the possible false detection resulted from potential systematics.
For this purpose, we prepare our control sample as follows: we shuffle the BCGs by assigning random positions (RA and DEC) to them, but keep all other information of the BCGs unchanged. Then around each BCG (at new random position), we re-assign ``cluster satellites'' by choosing those galaxies that are falling within the $R_{scale}$ from the BCG. Also, the $i$ band magnitude of the ``satellites'' should be in the range from 14 to 20. The $R_{scale}$, measured in Mpc, plays the role of virial radius~\citep{haocat}. In addition to this non-cluster sample, we also add another random PA to every galaxy in both the true cluster sample and the non-cluster sample by replacing the measured galaxy's PA with a random angle uniformly sampled between 0 and 180 degree. This random PA will serve as another control sample to double check the possible systematics in our measurements.
\section{Results}
\subsection{Satellite Alignment}
There are two basic questions concerning the formation of the satellite alignment: (1) is it a residual feature of the initial condition of the cluster formation? (2) or is it a dynamically evolving effect that varies as the cluster evolves~\citep{pereira05}, for example, due to the tidal torque? The two different scenarios lead to different redshift dependence of $\delta$. If it is left over from the initial alignment, its strength should decrease as redshift decreases. On the other hand, we should see a stronger alignment signal at low redshift if it is a dynamically evolving effect~\citep{ciotti94,catelan00,kuhlen07,pereira08,pereira10}. Therefore, looking at the redshift dependence of the measured $\delta$ is our primary interest. To do this, we first measure the $\delta$ for each cluster using 4 different PAs, i.e., the random PA, exponential fit PA, De Vaucouleurs fit PA and isophotal PA in $r$ band. We bin the clusters into redshift bins of size 0.05. In each bin, we calculate the weighted mean of $\delta$ and the standard deviation of the weighted mean, with the weights specified by $1/\delta_{err}^2$. Then, we perform the same measurement on the control sample, and present the results in Figure~\ref{fig:deltaz} and Figure~\ref{fig:delta_z_random}.
\begin{figure}
\begin{center}
\epsscale{1}
\plotone{delta_z.pdf}
\caption{Satellite alignment for clusters at different redshift bin of size 0.05. The legend random PA indicates using of randomized PAs. exp indicates using exponential fit PAs. dev indicates using De Vaucouleurs fit PAs and iso indicates using isophotal PAs. $r$ indicates the SDSS $r$ filter. This legend convention is also applicable to other figures in the paper.}
\label{fig:deltaz}
\end{center}
\end{figure}
\begin{figure}
\begin{center}
\epsscale{1}
\plotone{delta_z_random.pdf}
\caption{Satellite alignment measured based on the random control sample. }
\label{fig:delta_z_random}
\end{center}
\end{figure}
\begin{figure}
\begin{center}
\epsscale{1}
\plotone{member_ba_z.pdf}
\caption{The axis ratio of the satellite galaxies. Here, the error bar is the standard deviation to the mean in that bin. The legend satellite expAB\_r refers to the $b/a$ ratio is measured by fitting the satellite galaxy with an exponential profile in $r$ band. devAB\_r and isoAB\_r refer to the $b/a$ ratio by fitting De Vaucouleurs and isophotal profiles respectively. This convention is also applicable to other figures in this paper.}
\label{fig:deltambz}
\end{center}
\end{figure}
Based on the results in Figure~\ref{fig:deltaz}, one can see that the $\delta$ measured using isophotal PAs deviated from that based on random PAs in a statistically significant way at low redshift, though approaching zero as redshift increases. While the $\delta$ measured using the exponential fit PA and De Vaucouleurs fit PA are consistent with that measured using random PAs except in the lowest redshift bin. In \citet{pereira05}, the authors used isophotal fit PAs and also considered a cluster sample (sample B) with satellite galaxies selected using red sequence. When limiting the $r$ band magnitude to less than 18, they measured a $\delta = -1.06 \pm 0.37$, which is consistent with our results in the lowest redshift bin using isophotal fit PAs. The results based on the random control sample in Figure~\ref{fig:delta_z_random} show that $\delta$ measured using all types of PAs are consistently zero across the redshift range. In Figure~\ref{fig:deltambz}, we plot the axis ratio of the satellite galaxies in different redshift bins. The axis ratio determines how precisely the PAs are measured. From Figure~\ref{fig:deltambz}, the axis ratio measured using the isophotal method does not vary much as redshift increases, indicating that the diminishing satellite alignment based on isophotal fit PAs is not due to the decreasing S/N at high redshift.
There are two possible explanations for the measured $\delta$ in the low redshift bins. The first one is that the diffuse light from the BCGs affects the measurements of the PAs of the cluster satellite galaxies'. This creates an artificial preference of major axes of the satellite galaxies. This contamination is most severe when the PAs are measured using isophotal fit, but less prominent when the PAs are measured using exponential fit and De Vaucouleurs fit. This is because the isophotal PAs are sensitive to the shape of the outer profile of galaxy while the model fit PAs are more determined by the inner profile of the galaxy.
The second possible explanation to these results is the twisting of galaxy. This leads to different PAs when we use different methods. The outer rim of the galaxy is more susceptible to the tidal torque so that the alignment will show up when we use the isophotal fit PAs. One way to distinguish these two explanations is to look at the way $\delta$ depends on the absolute and apparent magnitudes of the BCG. To see this, we plot the measured $\delta$ for all the clusters with respect to apparent magnitudes and absolute magnitudes of the corresponding BCGs in the Figure~\ref{fig:deltarmaga} and Figure~\ref{fig:deltarmagb} respectively. One can see that $\delta$ shows a strong dependence on the apparent magnitude but not on the absolute magnitude. Therefore, we conclude that the $\delta$ in the low redshift bins in Figure~\ref{fig:deltaz} is more likely resulted from the artifact of the PA measurement. In Figure~\ref{fig:photozramag}, we show the absolute magnitude vs redsfhit for the BCGs.
\begin{figure}
\epsscale{1}
\plotone{delta_lowz_rmag.pdf}
\caption{The dependence of satellite alignment $\delta$ on the r-band apparent magnitudes of the corresponding BCGs. This shows a strong dependence of $\delta$ measured using isophotal PAs on the apparent magnitude of the BCGs.}
\label{fig:deltarmaga}
\end{figure}
\begin{figure}
\epsscale{1}
\plotone{delta_lowz_ramag.pdf}
\caption{The dependence of satellite alignment $\delta$ on the r-band absolute magnitude of the BCGs. This shows that there is very little dependence of $\delta$ measured using isophotal PAs on the $r$ band absolute magnitude of the BCGs}
\label{fig:deltarmagb}
\end{figure}
\begin{figure}
\epsscale{1}
\plotone{photoz_ramag.pdf}
\caption{Photometric redshift vs. r-band absolute magnitude for BCGs. The red overplotted dots are the means in each redshift bin of size 0.05.}
\label{fig:photozramag}
\end{figure}
\subsection{BCG Alignment}
\subsubsection{Redshift Dependence}
Now, we consider the BCG alignment. We first look at the redshift dependence of $\gamma$. Again, we perform our measurements on both the cluster sample and the random control sample. The results are shown in Figure~\ref{fig:gammaz} and Figure~\ref{fig:gammaz_random}
\begin{figure}
\begin{center}
\epsscale{1}
\plotone{gamma_z.pdf}
\caption{BCG alignment at redshift bins of size 0.05 using different PAs measured from the cluster sample.}
\label{fig:gammaz}
\end{center}
\end{figure}
\begin{figure}
\begin{center}
\epsscale{1}
\plotone{gamma_z_random.pdf}
\caption{BCG alignment at redshift bins of size 0.05 using different PAs measured from the random sample.}
\label{fig:gammaz_random}
\end{center}
\end{figure}
The results from the random control sample show that $\gamma$ is consistently zero. From the cluster sample, we detect a clear BCG alignment and its strength decreases as redshift increases. From the figures, we can also see that $\gamma$ is almost the same no matter how the PAs of BCGs are measured. This has two important implications: 1. the PAs of BCGs are well measured by both isophotal fit and model fit; 2. the diffuse light of satellite galaxies does not affect the PA measurement of the BCG. Before we can conclude the redshift dependence of $\gamma$, we still need to do one more test. That is, if the axis ratio ($b/a$) of the BCG and cluster become systematically smaller due to the decreased S/N at higher redshift, the measured strength of $\gamma$ will decrease too.
To calculate the cluster $b/a$, we use the method as described in \citet{kim02,ostholt10}. For clusters, the axis ratio is defined as $b/a = (1-\sqrt{Q^2+U^2})/(1+\sqrt{Q^2+U^2})$, with the stokes parameters $Q=M_{xx}-M_{yy}$ and $U=2 M_{xy}$. The radius-weighted second moments are given by $M_{xx}=\left<x^2/r^2\right>$, $M_{yy}=\left<y^2/r^2\right>$ and $M_{xy}=\left<xy/r^2\right>$ with $r^2=x^2+y^2$. $x$ and $y$ are the distances between the satellite galaxies and BCG in the tangent plane. For BCGs, we use the measured $b/a$ in SDSS pipeline based on isophotal fit, exponential fit and DeVoucular fit. In Figure~\ref{fig:ba}, we plot the $b/a$ in each redshift bin of size 0.05. From the results, the axis ratio does not depend on redshift in a statistically significant way. Furthermore, we also checked that both the BCGs' PAs and cluster PAs are distributed randomly, as show in the right panel of Figure~\ref{fig:pa}. Therefore, the evolution of $\gamma$ in Figure~\ref{fig:gammaz} should not result from the S/N variation of the PA measurements.
\begin{figure}
\epsscale{1}
\plotone{bcg_cluster_ba_z.pdf}
\caption{The axis ratio $b/a$ of BCGs and clusters vs. redshift.}
\label{fig:ba}
\end{figure}
\begin{figure}
\epsscale{1}
\plotone{bcg_cluster_pa_z.pdf}
\caption{The distribution of BCGs' and clusters' PAs.}
\label{fig:pa}
\end{figure}
\subsubsection{Magnitude Dependence}
We measure the BCG alignment vs the $r$ band absolute magnitudes of the BCG. The results are presented in Figure \ref{fig:gamma_ramag}
\begin{figure}
\epsscale{1}
\plotone{gamma_ramag}
\caption{The dependence of BCG alignment on the absolute $r$ band magnitudes of BCGs. }
\label{fig:gamma_ramag}
\end{figure}
From the plot, we see that $\gamma$ strongly depends on the BCGs absolute magnitude. To further show that this is not due to S/N of the measurement of BCGs shape, we plot the axis ratio $b/a$ of BCGs and clusters as a function of the BCGs $r$ band absolute magnitude in Figure~\ref{fig:ba_absmag}. There is no dependence of $b/a$ on the BCGs absolute magnitude.
\begin{figure}
\epsscale{1}
\plotone{bcg_cluster_ba_absmag.pdf}
\caption{The axis ratio $b/a$ of BCGs and clusters vs. the $r$ band absolute magnitude of BCG.}
\label{fig:ba_absmag}
\end{figure}
From the above results, we see that the $\gamma$ depends on both redshift and the BCG absolute magnitude. To show this more clearly, we bin the cluster samples into photoz bins of size 0.05 and absolute magnitude bins of size 0.5. Then, we calculate the mean $\gamma$ in each bin and plot the results in Figure~\ref{fig:gamma_z_magbin}. The color in the plot indicates $\gamma$.
\begin{figure*}
\begin{center}
\epsscale{1.2}
\plotone{gamma_ramag_photoz_contour.pdf}
\caption{Contours of $\gamma$ w.r.t. redshift and BCG absolute magnitude. The crossover of the contour lines are mainly due to the large error bar of each data point, which cannot be expressed in the contour plots.}
\label{fig:gamma_z_magbin}
\end{center}
\end{figure*}
From the plot, we can see that the $\gamma$ increases (i.e. absolute signal decreases) as redshift increases and as the absolute magnitude of the BCG decreases. Since the absolute magnitude of a galaxy is proportionally correlated with its mass, the above results indicate that the more massive BCGs tend to be more aligned with the cluster orientation. A subsample of the BCGs ($\sim$2800 BCGs) has their stellar masses measured in the MPA-JHU value-added catalog for SDSS DR7~\citep{kauffmann03,salim07}. This allows us to directly look at the trend of $\gamma$ with respect to the BCG stellar mass (a proxy of the total mass) and redshift. We choose three stellar mass bins and plot the $\gamma$ vs. redshift in each of them in Figure~\ref{fig:gamma_z_stm}. We can see that the $\gamma$ increases as redshift increases in each stellar mass bin and the massive BCGs shows more negative $\gamma$.
\begin{figure*}
\begin{center}
\epsscale{1.2}
\plotone{gamma_stm_photoz.pdf}
\caption{Redshift dependence of $\gamma$ in three BCG stellar mass bins.}
\label{fig:gamma_z_stm}
\end{center}
\end{figure*}
\subsubsection{Richness Dependence}
Next, we check whether $\gamma$ depends on cluster richness. To do this, we bin the clusters by their richness in bins with edges at 15, 25, 35, 50, 65 and 90. We did not go to the richness bins above 90 due to the smaller number of clusters in that richness range. Then, we look at the mean $\gamma$ in each bin. Note that we did not re-bin them into different redshift bins to keep the number of clusters reasonably large. This will not affect our purpose for detecting richness dependence since clusters from different redshifts are randomly falling into different richness bins. We plot the results in Figure~\ref{fig:gammarich}. Based on the results, we do not see a dependence on cluster richness.
\begin{figure}
\epsscale{1}
\plotone{gamma_ngals.pdf}
\caption{Dependence of BCG alignment on cluster richness. Here, we bin the richness into bins with edges at 15, 25, 35, 50, 65 and 90. The results do not show a statistically significant dependence of BCG alignment on cluster richness.}
\label{fig:gammarich}
\end{figure}
\subsection{Redshift Evolution of BCG Alignment Once Again}
Measuring the redshift evolution of BCG alignment is very important for understanding its origin. However, the redshift evolution of the measured alignment signal needs to be interpreted with great caution, especially for the cluster samples selected using photometric data. There are at least four factors that will introduce systematic redshift dependence and complicate the interpretation. (1) The S/N of the galaxy shape measurements will decrease as redshift increases. (2) When we look at clusters of different redshift, we need to make sure we are comparing the same population of satellite galaxies. That is, the cluster catalog needs to be volume limited in the redshift range. (3) The purity of clusters need to be consistently high across the redshift range. The change of purity will lead to decreased mean alignment signal. (4) The level of contamination from the projected field galaxies are prone to redshift dependence. Higher level of contamination will dilute the alignment signal. We have addressed (1) and (2) in previous sections, where we introduce the results. In the follows, we will focus on the (3) and (4).
The alignment signal from the falsely detected clusters should be consistent with zero, decreasing the mean alignment of the whole sample. For the subsample of GMBCG clusters with richness equal or greater than 15, it has been shown that the purity is consistently above 90\% and does not vary more than 10\% across the redshift range from 0.1 to 0.4~\citep{haocat}. To further show the purity variation will not produce the observed redshift dependence of $\gamma$, we choose another subsample of clusters with even higher purity. We choose clusters with richness greater than 25, which have a purity of above 95\% and vary less than 5\% in the redshift range. In Figure~\ref{fig:gammaz25}, we plot the BCG alignment parameter $\gamma$ from this subsample. Though the overall signal level increases a little at low redshift, the trend of redshift dependence does not differ much from the full sample with a lower richness threshold 15. So, the purity change should not explain the measured strong redshift dependence of $\gamma$.
\begin{figure}
\begin{center}
\epsscale{1}
\plotone{gamma_z_rich25.pdf}
\caption{BCG alignment at redshift bins of size 0.05 based on a subsample with higher purity and lower purity variations across the redshift range. The results indicate that the redshift dependence are not affected by the purity of the cluster sample we are using.}
\label{fig:gammaz25}
\end{center}
\end{figure}
On the other hand, satellite galaxies selected using red sequence colors have different levels of contamination from projected field galaxies as redshift changes. This is mainly caused by the different degree of overlap between the red sequence population and the field galaxy population (e.g. see Figure 14 in \citet{haocat}). There are two competing effects that will increase or decrease the contamination. First, the separation between the red sequence component and the field galaxy component. As redshift increases, the two components separate farther, leading to decreased projection contamination in the red sequence. The second effect is the broadening of the distribution of both red sequence and field galaxy. As redshift increases, the measured width of red sequence increases mainly due to the photometric errors~\citep{haoecgmm}. This will increase the chance of projected field galaxies being identified as satellites when we select the satellite galaxies by color. Therefore, the actual contamination level is the compromise of these two effects.
We can describe the measured alignment parameters, $\gamma$\footnote{Since we did not see a significant $\delta$ signal, we will consider only $\gamma$ in all the discussions hereafter. But the method described can also be applied to the $\delta$ case.}, as a combination of alignment from real cluster satellites and projected field galaxies. If we denote the alignment parameters from our measurements as $\gamma_m$, then we can decompose it into two parts as follows:
\begin{equation}\label{gamma_m}
\gamma_m = \frac{\sum_{i=0}^{N_c}\theta_i+\sum_{j=0}^{N_f}\theta_j}{N_c + N_f} - 45
\end{equation}
\noindent where $N_c$ is the number of true cluster satellite galaxies and $N_f$ is the number of projected field galaxies. We can introduce the fraction of real cluster satellite as $f_c(z) = N_c/(N_c + N_f)$, the BCG alignment from true cluster members as $\gamma_c=\sum_{i=0}^{N_c}\theta_i/N_c - 45$ and the BCG alignment from the projected field galaxies as $\gamma_f=\sum_{j=0}^{N_f}\theta_j/N_f - 45$. Substitute these definitions into Equation~\ref{gamma_m} and take ensemble average of the clusters, we will have:
\begin{equation}
\left<\gamma_m\right> = \left<f_c(z)\right> \left<\gamma_c\right> + \left[1 - \left<f_c(z)\right>\right] \left<\gamma_f\right>
\end{equation}
\noindent where $\left<...\right>$ denotes the average over the cluster ensemble. As the mean alignment signal from the field is consistent with zero, the alignment parameter $\gamma$ from the true cluster satellites is related to the measured one through the redshift dependent fraction $f_c(z)$. To the first order approximation, we can separate $f_c(z)$ into two parts as $f_c(z) = f_{const} \times f(z)$, where $f_{const}$ is a redshift independent component of the fraction, indicating the ``intrinsic'' fraction of true satellite based on color selection. $f(z)$ is the redshift dependent part, corresponding to the effect we described above. Then, the redshift dependence of the measured alignment $\gamma$ will be mainly determined by $f(z)$.
In the GMBCG catalog, we also measured a weighted richness, which takes into account the different degree of overlaps between red sequence and the field galaxies at different redshift~\citep{haocat}. The difference between weighted richness and the direct member count richness is a good estimator of the number of projected galaxies due to the effect described above. The fraction of contamination can therefore be estimated by the ratio of this difference to the direct member count richness. In Figure~\ref{fig:fcz}, we plot the fraction of contamination ($1 - \left<f(z)\right>$) as a function of redshift in bins of size 0.05. The fraction is almost constant except for the lowest redshift bin. Again, this cannot explain away the dependence of $\gamma$ on redshift as shown in Figure~\ref{fig:gammaz} and Figure~\ref{fig:gammaz25}. Therefore, after considering all the possible systematics known to us, the measured redshift dependence of $\gamma$ still cannot be explained. In~\citet{ostholt10}, the authors also reported a different BCG alignment between one low redshift bin (0.08 - 0.26) and another high redshift bin (0.26 - 0.44), which is consistent with the results we find here.
\begin{figure}
\epsscale{1}
\plotone{fcz.pdf}
\caption{Fraction of projected field galaxies at different redshift bin of size 0.05. It maintains constant except in the lowest redshift bin. }
\label{fig:fcz}
\end{figure}
\subsection{Conclusions and Discussions}
We measure the satellite alignment and BCG alignment based on a large sample of photometrically selected galaxy clusters from the SDSS DR7. We detect a satellite alignment only when we use the isphotal PAs. As we noted in \S 3.3, the isophotal PA tends to trace the outer profile of the galaxy while the model fit PAs tend to trace the inner part of the galaxy. A direct interpretation of the measurement results could be that the outer part of the satellite galaxy is more susceptible to the the gravitational torque and thus shows an orientation preference toward the BCG. However the inner part of the galaxy is not affected much by the tidal torque and does not show preference toward the BCG. The measured discrepancy of the satellite alignment from different PAs could be a manifestation of the twisting of galaxy shape from inner part to outer part. However, another possibility of this discrepancy could be that the light from BCG contaminates the measurement of the PA based on the isophote fit to the outer region of the galaxy and lead to a ``artificial'' alignment. By comparing the dependence of $\delta$ on BCG apparent and absolute magnitudes, we favor the latter explanation. This means that, though the tidal torque within the galaxy cluster may induce the satellite alignment, we are not yet able to detect them based on our current SDSS data. It will be definitely an interesting question to address with the forthcoming high quality data such as that from the Dark Energy Survey~\citep{des05}.
For the BCG alignment, by introducing the alignment parameter $\gamma$, we detect a strong redshift and BCG absolute magnitude dependences of the alignment. The redshift dependence cannot be explained by our known systematics. This result implies that the BCGs orientation is a dynamically evolving process and gets stronger as the cluster system evolves. For the dependence of $\gamma$ on the absolute magnitude of BCG, our result is qualitatively consistent with the conclusion that clusters with BCG dominance show stronger BCG alignment in \citep{ostholt10}. Furthermore, based on a subsample of the BCGs whose stellar masses are available, we show that the BCG alignment signal becomes stronger as the BCG stellar mass increases. This result indicates that more massive BCGs (with lower absolute magnitude) are more likely to align with the major axes of clusters.
We must take great caution when interpreting the dependence of $\gamma$ on BCG absolute magnitude and stellar mass since the purity of the cluster sample may also depend on the BCG absolute magnitude and stellar mass. As the cluster purity decreases, the alignment signal will decrease too. The faintest two bins in Figure~\ref{fig:gamma_ramag} show null alignment signal, which may also be due to the significantly decreased cluster purity. Nevertheless, we can still see a trend that $\gamma$ increases as the BCG absolute magnitude increases by looking at the bright end of the sample where we are confident about the cluster purity. Evaluating the cluster purity variation w.r.t BCG absolute magnitude turns out to be difficult because it requires a mock galaxy catalog that has BCG information properly built in. The way the mock catalog is constructed will impact the results significantly. Therefore, we think the best way to check this purity variation w.r.t. magnitude is to perform similar analysis with deeper data in the near future, such as the data from the upcoming Dark Energy Survey~\citep{des05}.
|
\section{Introduction}
Accurate astrometry is crucial in interpreting astronomical observations, allowing the association of sources observed in one region of the electromagnetic spectrum with their counterparts at other wavebands, and thus providing the full spectral energy distribution of an astrophysical object. For Galactic objects such as X-ray binaries, high-precision astrometry can also provide estimates of the proper motion of the source \citep[e.g.][]{Mir01,Mir02,Mir03,Dha07,Mil09a}, and even the source distance via trigonometric parallax \citep[e.g.][]{Bra99,Mil09b}. At present, the highest precision can be achieved in the radio band using the technique of very long baseline interferometry (VLBI). However, the high resolution of VLBI arrays that enables such high-precision astrometry also makes it impractical to use this technique to search for sources without a well-defined error circle from lower-resolution instruments, since the number of pixels to be searched can become unfeasibly large.
Black hole X-ray binaries undergo occasional outbursts in which relativistically-moving jet knots are seen to move away from the central binary system. Accurate astrometry is also required to interpret observations of these jets, since knowledge of the position of the central binary allows us to compare the relative motion of approaching and receding jet components and hence to constrain the product $\beta\cos\theta$, where $\beta$ is the jet speed as a fraction of the speed of light, and $\theta$ is the inclination angle of the jet axis to the line of sight \citep{Mir94}. Knowledge of the Doppler factor in turn helps constrain the energetics of the jets.
As they evolve through their duty cycles, black hole X-ray binaries pass through a range of canonical states, defined by their X-ray spectral and timing characteristics. These X-ray states, representative of specific conditions within the accretion flow, are very well correlated with the behaviour of the associated outflow in the form of radio jets \citep*[for a detailed discussion of the disc-jet coupling, see][]{Fen04}. In the low/hard X-ray spectral state, observed at the beginning and end of an outburst, a steady, compact, flat-spectrum jet exists, which subsequently gives rise to bright, relativistically-moving, discrete ejecta as the X-ray spectrum softens at the peak of the outburst and the source moves through hard and soft intermediate states into a high/soft state. The compact radio jet is then quenched, with no core radio emission being detected until the source moves back from the high/soft state through the intermediate states into the low/hard state once more. The quenching of the core radio emission at the peak of the outburst implies that it is difficult to perform accurate astrometry at this time. To determine the position of the central binary system requires high-resolution VLBI observations during the low/hard state, to detect the compact jet originating from the core of the system. This requires the observations to be triggered sufficiently early in the outburst during the rise phase, or following the reverse transition back to the low/hard state at the end of the outburst, although the hysteresis effect \citep{Mac03} coupled with the radio/X-ray correlation in the low/hard state \citep*{Gal03} implies that the radio emission is less bright in the latter case.
\subsection{XTE J1752-223}
XTE J1752-223 was discovered on 2009 October 23 by the {\it Rossi X-ray Timing Explorer} ({\it RXTE}) during a routine scan of the Galactic bulge region \citep{Mar09a}. The spectrum and lack of pulsations in the X-ray band led \citet{Mar09b} to suggest that the source was a black hole candidate, a conclusion supported by \citet{Mun10a} and \citet{Sha10}, who also claimed a black hole mass of 8--11\,$M_{\odot}$ and a source distance of $3.5\pm0.4$\,kpc from correlations between X-ray spectral and timing properties. A 2-mJy radio counterpart was detected by \citet{Bro09} during the hard state of the system, before its proximity to the sun precluded detailed study by pointed instruments on board X-ray satellites. As the source emerged from this zone of avoidance, a state transition from the hard X-ray spectral state to the intermediate state was detected \citep{Hom10,Sha10a,Neg10}. The outburst was intensively monitored in the X-ray band by the {\it Monitor of All-sky X-ray Image (MAXI)} \citep{Nak10}, {\it RXTE} \citep{Sha10}, and {\it Swift} \citep{Cur11} satellites, which observed behaviour fairly typical for a black hole X-ray binary system.
The initial radio detection was followed up by further monitoring with the Australia Telescope Compact Array (ATCA), showing that the source had brightened by an order of magnitude by 2010 January 21. A VLBI imaging campaign with the European VLBI Network (EVN) and Very Long Baseline Array (VLBA) detected moving, decelerating and expanding jet components \citep{Yan10}, suggesting an initially mildly relativistic jet. While the radio core of the system was not detected, its location was inferred to be between the two components detected in the VLBA image of 2010 February 26. Further target of opportunity VLBA observations were made in 2010 April following the return of the source to the hard X-ray spectral state.
The lack of an accurate core position rendered some of the interpretation of the VLBI images reliant upon various assumptions. In this paper, we use optical astrometry in conjunction with VLBI radio imaging to determine the true core location, demonstrating how these two techniques can be highly complementary when used in parallel. We describe our observations in Section~\ref{sec:observations}, present our results in Section~\ref{sec:results}, and re-interpret the existing VLBI observations in light of our newly-determined core position in Section~\ref{sec:discussion}.
\section{Observations and data reduction}
\label{sec:observations}
\subsection{Optical}
We observed the field containing XTE J1752-223 using the Inamori-Magellan Areal Camera and Spectrograph (IMACS) instrument mounted on the 6.5-m Baade Magellan telescope at Las Campanas Observatory. We obtained four $i^{\prime}$-band and four $g^{\prime}$-band images on 2009 November 3 00:10:26--00:19:28 UTC with exposure times of 5--10\,s. The observing conditions were good with a photometric sky and a seeing of 0.8\arcsec\ and 0.9\arcsec\ in the $i^{\prime}$-band and $g^{\prime}$-bands, respectively. Only a small section of one of the eight CCDs of the IMACS mosaic detector was read to sample a $\sim4^{\prime}\times 4^{\prime}$ field of view at 0.11\arcsec\,pixel$^{-1}$.
We applied aperture photometry on each of the images using DAOPHOT within the Image Reduction and Analysis Facility ({\sc iraf}\footnote{\textsc{iraf} is distributed by the National Optical Astronomy Observatories}) software package to compute the instrumental magnitudes of the detected stars. Flux calibration of the field was performed by observing two Sloan Digital Sky Survey (SDSS) fields, and differential photometry was used to derive the source flux variability as a function of time. The photometric results given here are with respect to the three field stars C1--C3 shown in Fig.~\ref{fig:finder}. The average $i^{\prime}$-band and $g^{\prime}$-band magnitudes of XTE J1752-223 were $16.22\pm0.05$ and $17.79\pm0.05$\,mag, respectively.
\begin{figure*}
\centering
\includegraphics[width=\textwidth]{miller-jones_f1.eps}
\caption{Optical $i^{\prime}$-band finder chart for XTE J1752-223, taken from a 5-s image from 2009 November 3. North is up, east is to the left. The optical counterpart to XTE J1752-223 is indicated with the cross-hairs. The stars labeled C1 to C3 are comparison stars with $i^{\prime}$-band ($g^{\prime}$-band) magnitudes of 13.62 (14.66), 14.53 (16.25) and 14.38 (15.30), respectively. The accuracy in the above magnitudes is of order 1 per cent.}
\label{fig:finder}
\end{figure*}
\subsection{VLBA}
At the end of its 2009--2010 outburst, XTE J1752-223 made a transition back to the hard state late in 2010 March \citep{Mun10b}. As reported by \citet{Yan10}, two VLBA observations were made following this transition, on 2010 April 25 and 29 (MJD 55311.46 and 55315.46), under project code BB291, at an observing frequency of 5\,GHz. The data were taken in dual-polarization mode with a recording rate of 512\,Mbps, corresponding to a total bandwidth of 64\,MHz per polarization. XTE J1752-223 was phase referenced to the nearby calibrator source J1755-2232, from the third VLBA calibrator survey \citep[VCS3;][]{Pet05}. The position assumed for the calibrator source was 17$^{\rm h}$55$^{\rm m}$26\fs285, $-22$\degr32\arcmin10\farcs593 (J\,2000), although the current best position\footnote{http://astrogeo.org/} is 17$^{\rm h}$55$^{\rm m}$26\fs2845, $-22$\degr32\arcmin10\farcs616 (J\,2000), with an uncertainty of 1.4\,mas. Since phase referenced positions are measured relative to the position of the calibrator source, the measured target position must therefore be corrected by the 24-mas difference between the assumed and true positions.
In an attempt to perform accurate astrometry on the fading core of the system, one further epoch of VLBA observations was taken on 2010 June 17 (MJD 55364.31), at an observing frequency of 8.4\,GHz, under project code BM346. The data were split into eight 8-MHz intermediate frequency pairs, corresponding to a total bandwidth of 64\,MHz in each of two independent polarizations. We used the same phase reference source, J1755-2232, but assumed the best known position from the VLBA calibrator manual, as given above. The phase referencing cycle time was 3\,min (2\,min on the target source, 1\,min on the calibrator), and we substituted every eighth scan on the target source for an observation of a nearby check source from the fifth VLBA calibrator survey \citep[VCS-5;][]{Kov07}, J1751-1950. 30\,min at the beginning and end of the observing run were used to observe bright calibrator sources at a wide range of elevations across the entire sky, in order to better calibrate unmodelled clock and tropospheric phase errors using the task DELZN from the Astronomical Image Processing System ({\sc aips}\footnote{{\sc aips} is produced and maintained by the National Radio Astronomy Observatory}) software package \citep{Gre03}, thereby improving the accuracy of the phase transfer and hence the resulting astrometry.
Data reduction was carried out according to standard procedures within {\sc aips}. We corrected for small changes to the Earth orientation parameters used in the initial correlator model, and also for the ionospheric dispersive delay. We used the measured system temperatures to calibrate the amplitude scale, and corrected the instrumental phase using the fringe finder source J1733-1304 from the international celestial reference frame \citep[ICRF;][]{Ma98}. We performed fringe fitting on the phase reference source J1755-2232, which was then subjected to iterative imaging and self-calibration. The final image was used as a model for bandpass calibration before transferring bandpass, amplitude and phase solutions to XTE J1752-223. The strong scattering along the line of sight towards the source led us to use only the shortest baselines (up to 30\,M$\lambda(\nu/5{\rm GHz})$, where $\nu$ is the observing frequency) in making the images, tapering the weights of the visibility data with a Gaussian function of full width at half maximum 22.8\,M$\lambda$ to further downweight the long baselines.
\section{Results}
\label{sec:results}
\subsection{Optical}
We performed astrometry on a 5-second $i^{\prime}$-band exposure. We compared the positions of the stars against entries from the third U.S. Naval Observatory CCD Astrograph Catalog \citep[UCAC3;][]{Zac10}. An astrometric solution was computed by fitting for the reference point position, the scale and the position angle, considering all the sources that are not saturated and appear stellar and unblended. We obtain a solution with 0.031$^{\prime\prime}$\ root-mean-square (rms) residuals from 26 stars (UCAC fit model magnitudes 14--16.5) well distributed on the $\sim$4$^{\prime}$$\times$4$^{\prime}$\ image. The positional accuracy of UCAC3 is estimated to be $\sim0.01$$^{\prime\prime}$\ on stars of such magnitude \citep{Zac10}. In addition, the systematic uncertainty in tying the UCAC3 stars to the International Celestial Reference System (ICRS) is 5\,mas \citep{Zac10}. For the accuracy on our stellar positions we adopt the linear sum of the residuals of the astrometry and the accuracy of the catalogue (as the latter is probably a systematic error): the resulting positional accuracy at 1$\sigma$ is 0.046$^{\prime\prime}$\ on both right ascension and declination. Thus our best optical position is
\begin{equation}
\begin{split}
\notag
{\rm RA} &= 17^{\rm h}52^{\rm m}15\fs093 \pm 0.003\\
{\rm Dec.} &= -22\degr20^{\prime}32\farcs35 \pm 0.05\qquad{\rm (J\,2000)},
\end{split}
\end{equation}
This is 533.9\,mas to the south east of the position of component A (corrected for the error in the assumed calibrator position) as identified by \citet{Yan10} in their VLBA observations of 2010 February 11.
\subsection{Radio}
In the VLBA observations of 2010 April 25, a point-like radio source was detected with the VLBA, at a level of $0.46\pm0.07$\,mJy\,beam$^{-1}$, at a position consistent with the optical position of XTE J1752-223 (Fig.~\ref{fig:image}). After correcting for the error in the assumed calibrator position, the position of the radio source, relative to the VCS-3 calibrator J1755-2232 (with an assumed position of (J\,2000) 17$^{\rm h}$55$^{\rm m}$26\fs2845, $-22$\degr32\arcmin10\farcs616), was
\begin{equation}
\begin{split}
\notag
{\rm RA} &= 17^{\rm h}52^{\rm m}15\fs09509 \pm 0.00002\\
{\rm Dec.} &= -22\degr20\arcmin32\farcs3591 \pm 0.0008\qquad{\rm (J\,2000)},
\end{split}
\end{equation}
where the quoted uncertainties are only the statistical errors from the Gaussian fitting algorithm (JMFIT within {\sc aips}). The systematic error due to the cumulative effect of the error in the assumed calibrator position and the 0.76\degr\ throw between calibrator and target \citep*{Pra06} is estimated to be 0.51\,mas. This VLBA position differs from the optically-derived core position by 30.4\,mas, along a position angle 107$^{\circ}$\ E of N. With the derived uncertainty of 46\,mas in both co-ordinates of the optical position, the radio source is well within the optical error circle. Given the positional agreement, and since XTE J1752-223 was in the hard spectral state at the time of the radio observations, we conclude that this radio source is indeed a compact jet from the core of the system. We note that the $\sim6$-month time offset between the observations from which the optical and radio positions were determined will likely lead to a small positional shift between epochs owing to the parallax and proper motion signatures of the source, of order 1\,mas if the source is at a distance of a few kpc and participates in the Galactic rotation. Furthermore, since we see the radio emission from optical depth 1 at each frequency in a compact jet, the true core could be slightly offset along the jet axis from the position of the radio source.
\begin{figure}
\centering
\includegraphics[width=\columnwidth]{miller-jones_f2.eps}
\caption{5-GHz VLBA image of XTE J1752-223 on 2010 April 25. Contour levels are at $\pm(\sqrt{2})^n$ times the rms noise of 0.066\,mJy\,beam$^{-1}$, where $n=-3,3,4,5...$ The optically-derived core position is marked with a cross, and the $1\sigma$ optical error circle of radius $0.046$$^{\prime\prime}$\ is shown in white. Co-ordinates have been corrected for the error in assumed calibrator position for the phase reference source. The radio position lies well within the optical error circle, giving us confidence that we have correctly identified the radio core of the system.}
\label{fig:image}
\end{figure}
On 2010 April 29, we detected a radio source at this same position, but it had faded to $0.29\pm0.08$\,mJy\,beam$^{-1}$\ at 5\,GHz. On 2010 June 17, the source was only marginally detected, at a level of $0.25\pm0.08$\,mJy\,beam$^{-1}$\ at 8.4\,GHz. The persistent, variable radio emission at this location also supports our conclusion that this radio source does indeed correspond to a compact jet from the core of the system. In none of these three hard state observations was the radio core resolved. Our best constraint on the size scale of the compact jet comes from the first, brightest observation on 2010 April 25, when the source was unresolved down to the beam size of $12.5\times5.2$\,mas$^2$ in position angle 5.5$^{\circ}$\ east of north. This corresponds to a source size of $<12.5(d/{\rm kpc})$\,au, where $d$ is the source distance \citep[i.e.\ $<44$\,au for the distance of 3.5\,kpc claimed by][]{Sha10}.
Following the detection of the radio core, we re-reduced the VLBA data of 2010 February 18, 23 and 26 (program code BB290) presented by \citet{Yan10}, using a larger image size to search for receding ejecta to the south-east of the core. No new components were detected, to $5\sigma$ upper limits of 0.62, 0.74, and 0.45\,mJy\,beam$^{-1}$\ respectively.
\section{Discussion}
\label{sec:discussion}
Using high-precision optical astrometry, we have been able to locate the core of the X-ray binary system XTE J1752-223. From VLBI observations made during the hard spectral state of the system, when the radio emission is dominated by a compact, unresolved core jet, we were able to further refine the core position. In light of this new information, we now reanalyze the VLBI data presented by \citet{Yan10}.
\subsection{A double ejection event}
\label{sec:double}
The VLBA image of 2010 February 26 published by \citet{Yan10} shows two components, labelled A and B. While component A was detected in all four VLBI images from 2010 February, component B is detected in only this one image, and was interpreted by \citet{Yan10} as a receding component, leading the authors to infer that the true core of the system lay between components A and B in this image. Our new determination of the true core position implies that both these components are located to the north-west of the core (Table~\ref{tab:components}). From this, we infer that components A and B must arise from separate ejection events. While we are unable to reliably pinpoint the exact epoch of ejection of each of those events, we can use X-ray spectral and timing information together with constraints from the integrated radio light curves to obtain a rough estimate for component A.
\begin{table}
\begin{center}
\scriptsize
\begin{tabular}{llccc}
\hline\hline
Component & Date & MJD & Separation & P.A.\\
& & (d) & (mas) & ($^{\circ}$)\\
\hline
A & 11-Feb-2010 & 55238.4 & $562.2\pm0.7$ & $-51.3\pm0.1$\\
A & 18-Feb-2010 & 55245.6 & $619.0\pm1.2$ & $-50.3\pm0.1$\\
A & 23-Feb-2010 & 55250.6 & $648.9\pm2.5$ & $-51.0\pm0.2$\\
A & 26-Feb-2010 & 55253.6 & $663.1\pm1.6$ & $-50.9\pm0.1$\\
B & 26-Feb-2010 & 55253.6 & $175.1\pm1.9$ & $-49.9\pm0.6$\\
\hline
\end{tabular}
\end{center}
{\caption{\label{tab:components}Angular separation from our newly-determined core position of the VLBI components detected by \citet{Yan10}.}}
\end{table}
\subsection{The ejection of component A}
\label{sec:ejection}
Radio flares in X-ray binaries have been linked to the transition from a hard intermediate state (HIMS) to a soft intermediate state (SIMS) during a rapid phase of X-ray spectral softening at the peak of the outburst \citep{Fen04}. Also associated with this transition are a sharp drop in the integrated rms variability of the X-ray emission and a reduction in the coherence of the associated quasi-periodic oscillations (QPOs), from high-coherence Type C QPOs associated with flat-topped noise in the power spectrum, to lower-coherence Type A or Type B QPOs associated with weak red noise \citep{Bel05}. However, we note that \citet{Fen09} found that while these changes in the variability properties were closely associated with radio ejection events, the association was not exact, such that one could precede the other by up to a few days.
On MJD\,55215.9, XTE J1752-223 was in a HIMS with a 2.2\,Hz Type C QPO \citep{Sha10}. The integrated rms variability then decreased from 25 to 18 per cent as the QPO frequency rose to 5.3\,Hz by MJD\,55217.9, and by MJD\,55218.8, the observed QPOs had changed from Type C to Type A/B. This suggests that the transition from HIMS to SIMS occurred around MJD\,55218. Supporting this inference is the bright (20\,mJy), flat-spectrum radio emission observed on MJD\,55217 between 1.2 and 19\,GHz \citep{Bro10}. This represents an increase of the radio brightness by an order of magnitude as compared to the initial 2-mJy radio detection in the rising hard state \citep{Bro09}, suggesting the onset of a radio flare. This increase in radio brightness corresponds to an increase of only a factor of $\sim 2$ in the 15--50\,keV {\it Swift}/BAT and 4--10\,keV {\it MAXI}/GSC X-ray count rates \citep{Nak10} over the same period, which is not consistent with the radio/X-ray luminosity correlation \citep{Gal03} found for the compact jets of many hard state black hole candidates \citep[although note an ever-increasing number of outliers; e.g.][]{Gal07}. While this would tend to support the interpretation of a radio flare on MJD\,55217, the flat radio spectrum over more than a decade in frequency instead argues that this radio emission most probably still arises from a compact jet, rather than discrete, optically-thin transient ejecta. The anomalously bright radio emission could then correspond to the period of jet instability known to occur immediately preceding a large radio flare \citep{Fen04}.
Further radio information is available from the ATCA monitoring, which covered the entire outburst from the initial rising hard state through to the decay back to quiescence. Although a full analysis is beyond the scope of this paper and will be presented by Brocksopp et al.\ (in prep.), we summarize the relevant information here. The integrated radio light curve shows at least two large flares, followed by a few smaller events. The last flat-spectrum radio detection was made on MJD\,55217, after which the 9-GHz radio flux density dropped to 3.3\,mJy on MJD\,55220, before peaking at 9.9\,mJy on MJD\,55221. This suggests an initial ejection date between MJD\,55217 and 55220. The second flare was somewhat broader, with the rise phase beginning after MJD\,55226 and the flare peaking at 10.9\,mJy at 9\,GHz on MJD\,55242. The double-peaked light curve supports our conclusion from Section~\ref{sec:double} that the outburst comprised at least two ejection events.
\citet{Neg10} reported a sharp increase of the soft ($<4$\,keV) X-ray flux and a decline of the hard ($>10$\,keV) flux on MJD 55218, with the emergence of a disk blackbody component in the X-ray spectrum. This is consistent with the evidence from the X-ray timing and radio observations, suggesting MJD\,55218 as the likely date of the initial radio ejection event.
\subsection{Deceleration of the ejecta}
\label{sec:deceleration}
\citet{Yan10} found that uniform deceleration of component A fitted their measurements better than a ballistic model with no deceleration. However, their suggested deceleration parameters (Fig.~\ref{fig:angsep}) imply an ejection date of MJD\,55201.8 (2010 January 5). {\it RXTE} was still sun-constrained on this date, so no PCA observations are available to ascertain the X-ray state of the source, but the {\it Swift} and {\it MAXI} observations \citep{Nak10} show that this was significantly prior to the beginning of the X-ray spectral softening, and we deem this unlikely as the true date of ejection.
Assuming an ejection date of MJD\,55218, we are unable to fit the motion of the ejecta with a uniform deceleration model. Our best-fitting model has a reduced-$\chi^2$ value of 230.9. Either our assumed ejection date is wrong or the uniform deceleration model does not describe the data well. A plausible alternative could be the scenario outlined by \citet*{Wan03}, whereby a shock wave propagates into the interstellar medium, sweeping up material as it moves, and decelerating such that the late-time behaviour approaches the Sedov solution, $R\propto t^{2/5}$. Allowing the zero point to float and fitting the measured positions of component A with a simplistic Sedov model $R=R_0+k(t-t_0)^{0.4}$, where $R$ is the angular separation of the component from the core, we find $R_0=406\pm29$\,mas, $k=75.7\pm6.5$\,mas\,d$^{-0.4}$, and $t_0={\rm MJD\,}55232.3\pm1.6$, with a reduced $\chi^2$ value of 0.2. Although the zero time and position ($t_0$ and $R_0$, respectively) do not correspond to our assumed ejection date and derived core position, when coupled with an initial coasting phase where the ejecta travel purely ballistically \citep[as proposed by][]{Hao09}, this model appears to provide a plausible fit to the data (Fig.~\ref{fig:angsep}).
The derived zero time, $t_0$, is just consistent within error bars with an extrapolation of the expansion of component A \citep{Yan10} to zero size, which occurs on MJD\,$55229.7\pm1.0$. It also coincides with the rise phase of the second flare in the integrated radio light curves, suggesting that the breadth of the second flare could be due to the release of energy as component A begins to decelerate, possibly combined with the ejection of component B. However, if component B were ejected during this second radio flare, its non-detection in the three VLBI observations prior to MJD\,55253 is surprising. One explanation could be that the intrinsic jet speed and inclination angle to the line of sight are both high enough for the emission to be Doppler-deboosted until the jets have decelerated by sweeping up the surrounding gas. Alternatively, if the observed jet ejecta are shocks, the delay in the appearance of component B could arise from the time taken for the ejecta to either catch up with the slower-moving material ahead of them (for internal shocks) or to sweep up and interact with the surrounding gas (for external shocks). With only the one VLBI detection of component B and the lack of any signatures in the X-ray light curves that might correspond to a second ejection event, we cannot further constrain the ejection date of component B, and do not discuss it further.
In the absence of a precise ejection date for component A, constraints from the receding components, or from X-ray lightcurves of the ejecta as they decelerate, there are too few constraints to conduct a more meaningful fit to the full model of \citet{Hao09}. However, the measured VLBI angular separations, the integrated radio light curves and the expansion of component A are all consistent with the deceleration, brightening and lateral expansion of that component close to MJD\,55232, as derived from our model fitting. Thus, while we cannot definitively verify the proposed scenario, it is certainly plausible. Should the model be applicable, the angular scale for deceleration ($<0.56$$^{\prime\prime}$) would be significantly smaller for XTE J1752-223 than those derived by \citet{Hao09} for XTE J1550-564 (12--17$^{\prime\prime}$) and H\,1743-322 (3$^{\prime\prime}$). While the distance is not yet well-determined, if XTE J1752-223 is indeed relatively nearby, as implied by the low hydrogen column towards the source \citep{Mar09b,Cur11} and as derived via a more model-dependent method \citep[$3.5\pm0.4$\,kpc;][]{Sha10}, then the discrepancy in the physical scale of the radius at which deceleration begins would be greater still.
\begin{figure}
\centering
\includegraphics[width=\columnwidth]{miller-jones_f3.eps}
\caption{Angular separation of the components detected by \citet{Yan10} from our newly-determined core position, together with their decelerating ejecta model (dashed line) and their ballistic ejecta model (dot-dashed line). Filled points represent component A and the open point component B. Error bars (0.7--2.5\,mas) are smaller than the marker size. The black vertical dotted line indicates the time of the transition between HIMS and SIMS on MJD\,55218, as determined by \citet{Sha10}. Grey line shows a possible model for the data, consisting of a period of pure ballistic motion lasting until $t_0$ (indicated by the grey vertical dotted line), followed by our fitted Sedov phase (Section~\ref{sec:deceleration}). For the assumed ejection date, this appears to provide a better match to the data than either the pure ballistic model or the pure deceleration model.}
\label{fig:angsep}
\end{figure}
\subsection{Quenching of the radio core in the soft state}
Knowing the true core position, we can constrain the quenching factor of the compact jet in the soft state. During the three VLBA observations of 2010 February \citep{Yan10}, the core radio flux was in all cases $<2.4\sigma$, corresponding to a flux density of $<0.35$\,mJy\,beam$^{-1}$. As compared to the 20\,mJy flux density measured on 2010 January 21, this represents a core quenching factor of $>57$. This is consistent with previous lower limits on the quenching factor of the compact core jets in the soft X-ray state \citep[e.g.][]{Fen99}.
\subsection{No receding ejecta}
The reinterpretation of component B of \citet{Yan10} as an approaching component implies that their estimates of jet speed and inclination angle to the line of sight are no longer valid, since they relied on the ratios of sizes and flux densities of approaching and receding components. Since a re-examination of the VLBA data from 2010 February did not show any receding components (Section~\ref{sec:results}), we cannot make revised estimates of the physical parameters of the jets with any accuracy. However, assuming symmetric approaching and receding ejecta, linear expansion of the components \citep[as fitted for component A of XTE J1752-223 by][]{Yan10} and a value for the index of the electron energy spectrum, $p$, \citet*{Mil04} demonstrated that the ratio of the flux densities of approaching and receding components in a single image could be used to determine the product $\beta\cos\theta$. Using the measured flux density of 2.2\,mJy for component A on 2010 February 18, and taking the $5\sigma$ upper limit of 0.62\,mJy for the flux density of the receding component, then assuming a canonical value for optically-thin ejecta of $p=2.2$, we find a lower limit $\beta\cos\theta>0.66$, such that $\beta\geq0.66$ and $\theta\leq49^{\circ}$. As a caveat, we note that if significant deceleration and consequent brightening of the ejecta has occurred prior to this image being taken, as is conceivable (Fig.~\ref{fig:angsep} and Section~\ref{sec:deceleration}), the flux densities of approaching and receding components will no longer be governed by symmetric ejection, adiabatic expansion and Doppler boosting, so this analysis would not be valid. Should the upper limit on $\theta$ be valid however, it would argue that the time delay hypothesis proposed in Section~\ref{sec:deceleration} is more probable than Doppler deboosting as an explanation for the delayed appearance of component B.
\subsection{Future proper motion studies}
The time baseline between the three radio observations in 2010 April and June is not sufficiently large to measure a significant positional shift between epochs, particularly given the low significance of the latter two detections. Also, despite the 6-month time baseline between the observations from which the optical and radio positions were determined, the uncertainties are sufficiently large that we cannot reliably determine the source proper motion between these epochs. However, our high-precision measurement of the core position provides an initial data point for a measurement of the proper motion should the source undergo repeated outbursts in the future, or should it be sufficiently bright in quiescence to be detected with the new generation of sensitive radio instruments such as the Expanded Very Large Array (EVLA) or the High Sensitivity Array (HSA) following the completion of the ongoing bandwidth upgrade at the VLBA. Long time baselines and high-precision astrometry are essential for the measurement of X-ray binary proper motions, and given the relative rarity of outbursts in the majority of sources and the relatively small proper motion signals (typically a few milliarcseconds per year), it is important to take astrometric data at every possible opportunity. We encourage future astrometric observations of this source should it return to a bright hard state. A measurement of the source proper motion would provide information about the formation mechanism of the compact object \citep[e.g.][]{Mir01,Mir03}.
\section{Conclusions}
We have demonstrated how accurate optical astrometry can be used in conjunction with high-resolution VLBI imaging in the hard state of an X-ray binary to locate the compact core of the system to sub-milliarcsecond accuracy. Our determination of the position of the core of XTE J1752-223 mandated a re-interpretation of the published VLBI data from the 2009--2010 outburst of the source. The two components detected in the VLBA image of 2010 February 26 are both on the same side of the core, implying that there were at least two ejection events during the outburst, with the first likely occurring close to the transition from the hard intermediate state to the soft intermediate state on MJD\,55218. With this extra constraint on the motion of component A, its angular separation from the core as a function of time can no longer be fit with a uniform deceleration model. A plausible explanation could be ballistic motion out to some radius after which rapid deceleration occurred as the ejecta swept up the surrounding interstellar medium, and the jets transitioned to a Sedov phase. We constrain the quenching factor of the compact core radio jet in the soft state to be $>57$. No receding ejecta are detected in any of the VLBI observations, and from the upper limit to their flux density, we constrain the product of jet speed and inclination angle, $\beta\cos\theta$, to be $>0.66$.
\section*{Acknowledgments}
PGJ acknowledges support from the Netherlands Organisation for
Scientific Research via a VIDI grant. The VLBA is operated by the
National Radio Astronomy Observatory, a facility of the National
Science Foundation operated under cooperative agreement by Associated
Universities, Inc. This paper includes data gathered with the 6.5-m
Magellan Telescopes located at the Las Campanas Observatory,
Chile. This research has made use of NASA's Astrophysics Data System.
\label{lastpage}
\bibliographystyle{mn2e}
|
\section{Introduction}
This review was not specially written for the {\it Uspekhi Fizicheskikh Nauk
(Physics-Uspekhi)} issue devoted to the memory of V L Ginzburg. Nevertheless,
I would like to hope that the spirit of this review is close to that of other
papers in this issue. I was happy to work closely with Vitaly Lazarevich
for more than 30 years, starting from my graduate student days, and he definitely
played a significant role in my scientific development. So, this review, hopefully,
bears a fraction of the soul of Vitaly Lazarevich.
Vitaly Lazarevich was a passionate person. Astrophysics, undoubtedly, fascinated
him most. But the scale of his personality was such that this passion did not
separate, but instead united, people. So, it is not surprising that at the Lebedev
Institute the astrophysical seminar headed by Vitaly Lazarevich for more
than several decades continues, and scientists from many institutes participate
in its work. The Department of Physics and Astrophysics Problems at Moscow Institute
of Physics and Technology (MIPT), which Vitaly Lazarevich founded in 1968 and headed
until recently, continues to be one of the leading institutions in teaching young
astrophysicists. Numerous pupils of V L Ginzburg and pupils of his pupils working
in the leading astrophysical centers of the world keep his unique trademark in their
studies.
The astrophysical heritage of V L Ginzburg is enormous. He obtained fundamental
results in the theory of propagation of electromagnetic waves in cosmic plasma,
in the theory of the origin of cosmic rays, and in the theory of neutron stars
and black holes. In all cases, a simple model allowing the understanding of the
essence of physical process in observed astrophysical sources laid the basis of
the theory. The present review, hopefully, was written in the same spirit.
Astronomy, as follows from the very appellation, is the science that stemmed
from the observations of stars. During hundreds of years the people observed
stars in the sky and gained insight into the laws of Nature. The stars appeared
to be always unchanged and existing for an infinite amount of time. After the
appearance of spectral analysis, the first astrophysical observations and, later,
the theory of radiation generally confirmed this point of view. The lifetimes of
most stars turned out to be comparable to the age of the Universe. Thus, in the 1950s,
when radio astronomy began, stars emitting thermal radiation seemed to be the main
objects for studies. In radio astronomy, the brightness temperature remains even
now the basic characteristic of radiation intensity.
However, the first radio astronomical observations, and especially observations in
the X-ray and gamma-ray ranges, which started in the middle of the 1970s, discovered
numerous nonthermal sources in the Universe. These objects are sufficiently compact
(i.e., the spatial resolution of the existing detectors is insufficient to determine their
internal structure) and, in addition, are highly variable. In active galactic nuclei
the variability timescale (months or sometimes even days) is small according to the
cosmic timescale, with the variability timescale of radio pulsars and sources of
gamma-ray bursts being the fractions of a second, which is small even to Earth's
measures. The activity, i.e., high variability on timescales $\tau \sim R/c$, as well
as the generation of nonthermal radiation indicates that in most cases we are dealing
with relativistic objects, namely, with objects in which matter moves with velocities
close to the speed of light.
Jet eruptions represent one of the visible appearances of the activity of compact
astrophysical objects. We shall briefly discuss their properties in Section 2. They
are observed in both relativistic objects (such as active galactic nuclei and
microquasars) and in young stars where the motion of matter is definitely nonrelativistic.
This means that we are dealing with some universal and extremely efficient mechanism of
energy release. Therefore, the key theoretical problems include the question of the
energy source of the activity of compact objects, the understanding of their energy
release mechanism, and the collimation of matter outflows. We shall postpone the detailed
discussion of arguments against alternative models until the next section, and here we
only remind the main arguments favoring the magnetohydrody-namic model of activity of
compact sources, which is accepted by most astrophysicists.
The model of the unipolar inductor, i.e., the source of direct current, lies at the
heart of the magnetohydrodynamic approach. As we shall show in Section 2, conditions
for the existence of such a 'central engine' are satisfied in all the compact sources
discussed below. Indeed, all compact sources are assumed to harbor a rapidly spinning
central body (black hole, neutron star, or young star) and some regular magnetic field,
which leads to the emergence of strong induction electric fields. The electric fields,
in turn, lead to the appearance of longitudinal electric currents and effective particle
acceleration. The collimation mechanism in this model is related to the well-known property
of mutual attraction of parallel currents.
The first studies of the electromagnetic model of compact sources (namely, radio pulsars)
were carried out as early as the end of the 1960s [1-4]. It was evidenced that there are
objects in the Universe in which electrodynamical processes can play the decisive role in
the energy release. Then, in 1976 R Blandford [5] and R Lovelace [6] independently suggested
that the same mechanism can also operate in active galactic nuclei. In the same year,
G S Bisnovatyi-Kogan, Yu P Popov, and A A Samokhin proposed a magnetorotational mechanism
of the supernova explosion [7] (i.e., the model of an essentially nonstationary phenomenon),
in which jet eruptions can also be formed [8]. This model has remained the leading one
for nearly 40 years. However, only recently have some key properties become clear. This
is related both to advances in the theory which have at last formulated sufficiently simple
analytical relations, and to the breakthrough in numerical simulations which confirmed
theoretical predictions.
The reader can find the detailed introduction to the analytical theory in the author's
monograph [9] (see also the review in {\it Physics-Uspekhi} in 1997 [10]). However, first, the
monograph was devoted to the basics of the theory, and qualitative predictions for specific
astrophysical sources were discussed only very briefly. Second, the monograph clearly could
not include the results of numerical calculations carried out in the last five years since
its publication. This is the main reason for writing the present review. In addition, here
we shall correct formulas from the monograph in which misprints were found.
Of course, we are still far away from the full understanding of the essence of physical
processes proceeding in compact sources. In fact, now we have only agreement between theory
and numerical modeling. All results have been obtained applying ideal one-liquid
magneto\-hyd\-ro\-dynamics, though by different methods (the theory is based on stationary
equations, while numerically the time relaxation problem is solved). In particular, it
is not yet clear which of the main physical characteristics of the central engine (such as
the mass of the central body or its rotation velocity) should fully determine the observed
energy release. Nevertheless, the progress achieved over recent years raises hopes for test
observations already in the nearest future, which can give insight into physical processes
occurring in active astrophysical sources.
\section{Jets}
\subsection{Active galactic nuclei}
The main properties of the central engine in active galactic nuclei, which are
presently accepted by most astrophysicists, can be summarized as follows [11,12].
In the center of the host galaxy there is a supermassive black hole (its mass reaches
$10^6$--$10^9 \, M_{\odot}$, where \mbox{$M_{\odot} \approx 2 \times 10^{33}$ g}
is the mass of the Sun), onto which accretion of the surrounding matter occurs [13].
Only in this case it is possible to explain the very high efficiency of the energy
release and the compactness of the central engine. The energy source of activity of
galactic nuclei can be related to both the rotational energy of the black hole, viz.
\begin{equation}
E_{\rm tot} = \frac{J_{\rm r}\Omega_{\rm H}^2}{2} \approx
10^{62}\left(\frac{M}{10^9M_{\odot}}\right)\left(\frac{\Omega_{\rm H} r_{\rm g}}{c}\right)^2
\,{\rm erg},
\label{k1}
\end{equation}
and the energy of the accreting matter. Here \mbox{$r_{\rm g} = 2 GM/c^2$}
is the radius of the black hole, $J_{\rm r}$ is the moment of inertia, $\Omega_{\rm H}$ and $M$ are
the angular velocity and the mass of the black hole, respectively, and $c$ is the speed
of light. The existence of supermassive objects is also supported by the fact that the
Eddington luminosity
\begin{equation}
L_{\rm Edd} \approx 10^{47}\left(\frac{M}{10^{9}M_{\odot}}\right)
\, {\rm erg} \, {\rm s}^{-1},
\label{k2}
\end{equation}
(i.e., the luminosity at which the gravitational force acting on the accreting matter is
balanced by the radiation pressure force) is close to the characteristic luminosity of
active galactic nuclei [14]. Moreover, the duration of the active phase
$\tau_{\rm D} = E_{\rm tot}/L_{\rm Edd}$ estimated using formulas (1) and (2) is on the
order of $10^7$ years, which is also in agreement with observations.
Further, it is usually assumed that the accretion of matter proceeds through a disc [15].
Thus, the preferential direction --- the axis of rotation --- emerges naturally in space, along
which the formation of jets is possible. As a black hole itself cannot have the self-magnetic
field (the so-called 'no-hair theorem'), the generation of a large-scale magnetic field in the
vicinity of the black hole is believed to occur in the accretion disc [16-18].
\begin{figure}
\begin{center}
\includegraphics[width=\columnwidth]{fig1_01.ps}
\end{center}
\caption{Radio image (5 GHz) of active regions and jet eruptions
from the nucleus of the Cygnus A galaxy [20]. The distance between
bright spots is about 80 kpc, which is 9-10 orders of magnitude greater
than the size of the central black hole.
}
\label{fig1_01}
\end{figure}
According to the modern concept, massive central objects are present in most galaxies and
remain active only if a sufficient amount of matter falls on them. This restricts their
active lifetime. Unfortunately, as stated above, the angular resolution of modern detectors
does not allow us to directly observe plasma flow on the scales comparable to the black hole
size \mbox{$r_{\rm g} \approx 3 \times 10^{14}(M/10^{9} \, M_{\odot})$ cm.} Therefore, we have to
judge the activity of galactic nuclei only using indirect evidence, by observing flows on
much larger scales.
Let us remember that the diffuse radio emission around active galaxies is observed from
regions located at distances of tens or even a hundred kiloparsecs from their nuclei.
Very shortly after the discovery of these regions at the beginning of the 1960s, this
emission was associated with collimated plasma ejections (jets) flowing out the galactic
nuclei [12]. It is precisely these jets that transport matter and energy from the active
nuclei to those regions (Fig. 1). Observations show that the jets can be accelerated
and collimated very close to a galactic nucleus. For example, in the case of the nearest
active galaxy M87 the formation of the jet occurs within a radius of $60 \, r_{\rm g}$
from the nucleus [19]. In recent years, the internal structure close to the jet base was
resolved in several sources, where the jet tranverse dimension usually does not exceed several
parsecs [21, 22] (Fig. 2).
\begin{figure}
\begin{center}
\includegraphics[width=0.9\columnwidth]{fig1_02.ps}
\caption{Radio image of the jet eruption from the galaxy M87 near the
central engine [22]. The jet transverse size is about 1 pc.
}
\label{fig1_02}
\end{center}
\end{figure}
The matter in jets from active galactic nuclei has a very high energy --- the bulk Lorentz
factor of a jet is at least a few unities. For example, this motion is directly observed
in the M87 galaxy, with the bulk Lorentz factor of the outflow being
$\gamma \approx 6$ [23]. In many cases, the matter continues moving with relativistic
velocities up to huge distances from the nucleus before noticeable braking due to interaction
with the ambient intergalactic medium. Another peculiar feature of jets is their high degree
of collimation within a cone characterized by an opening angle of only several degrees.
Unfortunately, observations do not yet allow reliable estimations of the energy and mass
fluxes in jets from active galactic nuclei, of the magnitude of the magnetic field both close
to the black hole and in the jet itself, or of the composition of jet eruptions. The spectrum
of radiation from galactic jets (in contrast, for example, to the spectrum of jets from young
stars) does not exhibit any spectral features of moving matter, i.e., neither atomic (ionic)
lines nor the electron-positron pair annihilation line are observed. To this regard, there are
arguments both in favor [24] and against [25] the leading role of electron-positron plasma,
so it is now impossible to say exactly which mechanism of energy transfer to the jet actually
operates.
Where a physical nature of the galactic nucleous activity is concerned, several mechanisms
of particle acceleration and jet collimation have been proposed, but so far there is no
definite answer as to which of them are actually realized. It is possible that different
mechanisms operate in different sources, or, just the opposite, all mechanisms are realized
simultaneously.
{\it Gas-dynamic acceleration}. The acceleration and collimation of a jet can be related to the
presence of an ambient medium with high pressure which decreases with distance from the
center [26, 27]. Such a medium could play the role of an external wall collimating the
outflow. The pressure of the external hot medium can, in principle, be estimated from X-ray
observations [28]. This mechanism possibly explains how weak jets in our Galaxy and in some
Seyfert galaxies (i.e., low-active galaxies) are formed. On the other hand, the observed
pressure of the hot matter around the most powerful jets from active galactic nuclei is not
sufficiently high, and there must be an alternative mechanism of plasma confinement.
{\it Acceleration by radiation}. As the photon density near the central source can be very high,
the radiation-driven mechanism of jet matter acceleration by radiation pressure was proposed
[29, 30]. In this model, it is assumed that the inner parts of the disc can serve as a nozzle
directing matter outflows accelerated by the radiation pressure. However, this mechanism also
meets some difficulties. For example, there is no correlation between the jet power and
the luminosity of the source --- many sources with very powerful jets are low-luminous sources
[31]. Another difficulty comes from the fact that, starting from the sufficiently low particle
energies $\gamma \approx 3$, the radiation field more effectively brakes particles than accelerates them [32].
This contradicts observations of 'superluminal' sources in which the energy of plasma particles
is much higher. In addition, if the jet was formed in a system with a thin accretion disc
emitting radiation more or less isotropically, additional mechanisms for the jet collimation
should be invoked. A modification of this model involving the formation of a funnel in a thick
accretion disc can explain the initial jet collimation, but there are indications that such a
structure is unstable [31].
{\it Magnetohydrodynamic mechanism}. As noted above, most researchers favor the
magneto\-hyd\-ro\-dynamic model of jet formation. The magnetohydrodynamic (MHD) model was
successfully utilized to describe many processes in active nuclei, and, in particular, in
connection with the problem of the origin and stability of jets, as well as to explain the
energetics of processes proceeding near the central black hole. The magnetic field here is
the natural link between the central engine and the jet. Moreover, in this model it is easy
to understand why the jet matter can predominantly consist of electron-positron plasma. As
was shown in Refs [33-35], it can be generated on the magnetic field lines threading the
black hole horizon.
In the simplest version, the picture is as follows: the regular magnetic field generated in
the disc links the spinning central engine (the disc and the black hole) with infinity. The
plasma outflow occurs along the magnetic field lines; the electromagnetic energy flux is also
directed along the magnetic field lines. The longitudinal electric current flowing along the
jet forms a toroidal magnetic field, and the magnetic field pressure associated with this
toroidal component can collimate the jet.
It should be noted, however, that in a real astrophysical system the total current flowing
from the central engine should vanish, so the Ampere force in the current closure region will,
on the contrary, decollimate the flow (the antiparallel currents repulse). Therefore, an
external medium (for example, a subrelativistic wind outflow from the accretion disc) is
necessary to collimate the jet. In addition, the question as to whether it is possible to
consider a black hole immersed in the external magnetic field as a unipolar inductor turned
out to be also rather nontrivial. It required almost 30 years of studies after the paper by
Blandford and Znajek [33], which laid the basis of the theory in 1977, before the needed
clarity was reached in this question. We shall discuss these points in more detail in Section 3.
\subsection {Microquasars}
Microquasars comprise galactic objects in which the jet formation is due to accretion onto a
compact relativistic object (neutron star or black hole). In other words, all microquasars
reside in sufficiently close binary systems in which the effective flow of matter from the
star companion occurs. The rate of matter inflow in such systems is larger than can be swallowed
by the central object. As a result, some accreting matter that carries, in particular, an
excessive angular momentum is expelled from the system in the form of jets. Observations of
microquasars show that jets are related to thick accretion discs. In other words, no jets
are known for systems with thin discs. The reason for that is unclear: either a thin disc
insufficiently collimates the outflow, or the magnetic field generated by the thin disc is
not strong enough.
Microquasars represent a small population of objects, including only around ten sources [36],
with only half of them demonstrating noticeable relativistic jets \mbox{($v > 0.9 \, c$).}
The characteristic longitudinal size of jets is usually \mbox{$0.1$ pc}, with the jet spread
angle being within several degrees (Fig. 3). The total energetics are about $10^{37}$ erg s$^{-1}$.
Due to the relativistic velocity of the bulk motion of matter in the jets, some sources
demonstrate the superluminal motion effect, with the apparent angular velocity being several
orders of magnitude larger (due to a relative proximity of these objects) than that observed
in jets from active galactic nuclei.
\begin{figure}
\begin{center}
\includegraphics[width=\columnwidth]{fig1_03.ps}
\caption{Radio image of the jet from the microquasar iE1740.7-2942. The characteristic
length of the jet measures $0.1$ pc [38].
}
\label{fig1_03}
\end{center}
\end{figure}
Historically, the first revealed object of this class was the famous source SS433 [37] in
which, however, the gas ejection velocity in jets is only $0.26 \, c$. Such a velocity can be easily
explained by the radiation pressure from highly heated internal regions of the accretion disc.
As for relativistic jets, the first source was discovered only in 1994 [38]. Since the appearance
of near-light velocities due to radiation or gas pressure is problematic, it has not been ruled
out that, to explain them, an electrodynamic model similar to that used in explaining the origin
and collimation of extragalactic jets should be invoked again. This model is also supported by
the fact that in all but one microquasar (SS433) no emission lines from jets are observed. This
indirectly points to the electron-positron composition of matter in jets [36]. Finally, it
should be noted that in most microquasars the jet is separated in individual blobs at large
distances from the central engine, which is thought to be due to a long duty cycle of the work
of the central engine.
\subsection {Sources of cosmological gamma-ray bursts}
As regards the sources of cosmological gamma-ray bursts, there are indirect, although
sufficiently reliable, arguments in favor of the presence of jets related exactly to
relativistic strongly magnetized outflows, which we shall discuss in this review. It
is well known that the discovery of the optical afterglow [39], as well as afterglows
in other spectral ranges, which allowed the measurement of the distance to these sources
from the observed redshifts of the host galaxies, put serious constraints on their energetics
[40]. If the observed gamma-ray radiation were emitted isotropically, the total energy
release for the typical distance to these sources of several gigaparsecs would reach
$10^{54}$ erg. However, we do not know at present processes with such huge energy liberation.
On the other hand, the small duration of the burst \mbox{($\sim 10 \,$ s)} restricts the size of the
emission region, which, in turn, does not allow us to explain the observed nonthermal
gamma-ray spectra, since the optical depth in the source proves to be very high [41].
If it is assumed that gamma-rays are emitted within a narrow cone angle
$\vartheta \sim 1^{\circ}$, the observed energy can be reduced to $10^{51}$ erg, which
is already to an order of magnitude of the energy release during supernova explosions.
On the other hand, the observed optically thin nonthermal gamma-ray spectra immediately
imply the presence of ultrarelativistic outflows with bulk Lorentz factors of $\sim 100$--$300$.
Only in this case can the compactness problem of the source be resolved, since the
estimated size of the emitting region also increases as the square of the bulk Lorentz
factor (i.e., by $10^{4}$--$10^{5}$ times), and the optical depth, which is proportional
to the density multiplied by the size of the region, decreases respectively by
$10^{8}$--$10^{10}$ times.
However, the ultrarelativistic character of the outflow, in turn, puts constraints on the
particle composition in the jet, since the presence of a significant fraction of baryons
with such energy in the outflow would contradict the total energy release in the gamma-ray
burst. Therefore, the contribution of protons must be smaller than $10^{-2}$ of the total
number of particles, so that only electron-positron jets should be considered. The
existence of jets is also evidenced by the presence of the characteristic bend of the
light curve of the afterglow, when the power law index $\alpha$ in the radiation intensity
dependence on time, $W_{\rm tot} \propto t^{-\alpha}$, changes from $\alpha \approx 1.1$ to
$\alpha \approx 2.0$ after a span of about a few days following the burst. This effect is
related to the cessation of relativistic contraction of the radiation cone in the direction
of particles' motion toward an observer. Incidentally, this model allowed independent
confirmation of the jet spread angle $\vartheta \sim 1^{\circ}$ and the bulk Lorentz factor
$\gamma \sim 100$--$300$ [42].
The nature of the central engine giving rise to strongly magnetized jets can be usually
related to the collision of two neutron stars [43, 44] or of a neutron star and a black
hole [45], or, most likely, to the collapse of the massive core of an unusual supernova
[46, 47]. However, in most models a rapidly spinning solar-mass black hole ultimately
emerges, which loses its rotation energy via the Blandford-Znajek process [45, 48 50].
Indeed, as we have seen, this process easily provides a natural explanation for both the
low baryonic load of the jet and the large bulk Lorentz factors of jet particles. In other
words, the model again is constructed similarly to the scheme proposed for active galactic
nuclei. In particular, the key processes here also include the magnetic field generation
in the plasma around the black hole, the interaction of the black hole with the accretion
disc via magnetic field lines, and the generation of particles in the magnetosphere. To
explain the observed energy release, it is necessary to assume that the magnetic field
near the black hole must be as high as $10^{14}$ or even \mbox{$10^{15}$ G.} The generation of
such a high field is thought to be possible in nonstationary processes like the supernova
core collapse or binary neutron star coalescence [51, 52].
\subsection {Radio pulsars}
The discovery of radio pulsars at the end of the 1960s, which are the sources of pulsating
cosmic radio emission with the characteristic period $P \sim 1$ s [53], is definitely one
of the major astrophysical discoveries of the 20th century. Indeed, for the first time a
new class of cosmic sources related to neutron stars, whose existence was theoretically
predicted away back in the 1930s [54], was discovered. Neutron stars (mass of about
\mbox{$1.2$--$1.4$ $M_{\odot}$,} and radius $R$ of only $10$--$15$ km) must result from the
catastrophic compression (collapse) of usual massive stars at the late stage of their
evolution or, for example, of white dwarfs whose mass exceeds the Chandrasekhar mass
limit of $1.4$ $M_{\odot}$ due to accretion from the companion star. It is this formation
mechanism that provides the simplest explanation for both small spin periods $P$ (the
smallest known spin period $P = 1.39$ ms) and superstrong magnetic fields with
\mbox{$B_0 \sim 10^{12}$ G} [1,2].
Interestingly, the basic physical processes determining the observed activity of radio
pulsars were understood almost immediately after their discovery. For example, it became
clear that highly regular pulsations of observed radio emission are related to the rotation
of neutron stars. Next, radio pulsars are powered by the rotational energy of the neutron
star, and the mechanism of energy release is related to their superstrong magnetic field
with $B_0 \sim 10 ^{12}$ G. Indeed, energy losses estimated using the simple magnetodipole
formula [44] are as follows:
\begin{equation}
W_{\rm tot} = -J_{\rm r}\Omega\dot\Omega \approx
\frac{1}{6}\frac{B_0^2\Omega^4R^6}{c^3}\sin^2\chi,
\label{wmd}
\end{equation}
where $J_{\rm r} \sim MR^2$ is the moment of inertia of the neutron star, $\chi$
is the angle between the magnetic dipole axis and the spin axis, and $\Omega = 2\pi/P$
is the angular velocity of the neutron star rotation. For most pulsars, energy losses range
from $10^{31}$--$10^{34}$ erg s$^{-1}$. These energy losses exactly correspond to the
observed spin-down rate ${\rm d}P/{\rm d}t \sim 10^{-15}$, or to the spin-down time
$\tau_{\rm D} = P/\dot P \sim$ $1$--$10$ mln years. Let us keep in mind that the fraction
of radio emission amounts to only $10^{-4}$--$10^{-6}$ of total energy losses. For most
pulsars this corresponds to \mbox{$10^{26}$--$10^{28}$ erg s$^{-1}$,} which is 5--7 orders of
magnitude less than the luminosity of the Sun.
As shown in Refs [56, 57], the actual energy losses cannot be due to magnetodipole radiation
because the plasma that fills the magnetosphere will fully screen the low-frequency radiation
from the neutron star. However, energy losses can be caused by longitudinal electric currents
circulating in the magnetosphere and looped across the surface of the central engine. As a
result, in this case, too, the main energy release near the neutron star is related to the
electromagnetic energy flux (the Poynting vector flux), and the total energy losses can be
again estimated using formula (3).
Most radio pulsars constitute single neutron stars. Of the 1880 pulsars known by the middle
of 2010, only 140 were members of binary systems. However, in all these cases it is reliably
known that there is no somewhat appreciable mass transfer from the companion star to the
neutron star. Since, as already stressed, the radio luminosities of pulsars are low, the
modern sensitivity of detectors allows observations of pulsars only up to distances of
$3$--$5$ kpc, which is smaller than the distance to the galactic center. Therefore, we can
observe only a small fraction of all 'active' pulsars. The total number of neutron stars
in our Galaxy must be around \mbox{$10^{8}$--$10^{9}$.} Such a big number of extinguished neutron
stars can be naturally related to the small duration of their active life, as discussed above.
The jets are only observed in Crab and Vela radio pulsars [58, 59], which is not surprising,
since, in contrast to the compact objects considered above, the pulsar magnetosphere is not
axisymmetric. On the other hand, only axisymmetric configurations were actually
considered until recently in the theory of pulsar wind. Based on these studies, the main
features of strongly magnetized winds were understood. Nevertheless, even in this approximation
for smooth flows, it has thus far been impossible to construct a self-consistent model which
jointly describes the energy transfer from the neutron star surface to infinity and includes
effective particle acceleration, i.e., an almost complete transformation of the electromagnetic
field energy into the energy of the plasma flowing out. Because of this, different models are
actively being discussed at present, which, to various degrees, propose going beyond the
framework of the 'classical' scheme (see, for example, Refs [60-62]).
Indeed, observations show that most energy far from the neutron star must be carried by
relativistic particles. For example, the analysis of the emission from the Crab Nebula
in the shock region located at a distance of $\sim 10^{17}$ cm from the pulsar in the
region of interaction of the pulsar wind with the supernova remnant definitely shows
that the total flux $W_{\rm em}$ of the electromagnetic energy in this region is no more
than $\sim 10^{-3}$ of the particle energy flux $W_{\rm part}$ [63]. Thus, the Poynting
vector flux in the asymptotically remote region must be completely converted into the
outgoing plasma flux. The presently known axisymmetric numerical models of jets
from radio pulsars [64-66] were constructed exactly under this assumption.
However, the transformation apparently occurs already much closer to the neutron star,
namely at distances comparable to the size of the light cylinder. This is evidenced by
the detection of variable optical emission from companions in some close binary systems
involving radio pulsars [67]. This variable optical emission with a period equal exactly
to the orbital period of the binary can be naturally related to the heating of the companion's
part facing the radio pulsar. It was found that the energy reradiated by the companion star
almost matches the total energy emitted by the radio pulsar into the corresponding solid angle.
Clearly, this fact cannot be understood either in the magnetodipole radiation model or by
assuming a Poynting-dominated strongly magnetized outflow, since the transformation coefficient
of a low-frequency electromagnetic wave cannot be close to unity. Only if a significant fraction
of the energy is related to the relativistic particle flux can the heating of the star's surface
be effective enough. Therefore, the so-called $\sigma$-problem --- the question as to how the energy
is transferred from the electromagnetic field to particles in the pulsar wind --- remains one of
big puzzles in modern astrophysics.
\subsection{Young stars}
Jets from young stars were indirectly discovered at the beginning of the 1950s, when
G Herbig and G Haro [68, 69] discovered a new class of extended diffuse objects usually
existing in pairs and, as became clear later, connected by thin jets with young
rapidly rotating stars [70]. The formation of such jets can naturally be related to the
need of removing most effectively the excessive angular momentum that prevents the formation
of a star. As we see, the situation here is quite similar to that with active galactic nuclei,
where first a diversity of different types of sources (quasars, Seyfert galaxies, and radio
galaxies) were discovered, and only later on did it become clear that the activity of all
these sources has a similar nature. Moreover, the similarity of the observational features
suggests that the physical mechanism of jet formation from young stars can also be similar
to that from active galactic nuclei. And this is despite the fact that physical conditions
near a young star (mass of order $3$--$10$ $M_{\odot}$, and total energy release ranging
from $10^{31}$ to $10^{36}$ erg s$^{-1}$ are dramatically different from those in the centers
of active galactic nuclei. One of the main differences here is the nonrelativistic character
of gas outflow from young stars.
\begin{figure}
\begin{center}
\includegraphics[width=\columnwidth]{fig1_04.ps}
\caption{Optical image of jets from the system HH47 (see, for example, Ref. [70]).
The scale corresponds to 1000 a.u.
}
\label{fig1_04}
\end{center}
\end{figure}
Presently, more than 250 Herbig-Haro objects are known [71]. As shown in Fig. 4, they
represent bright condensations with an angular size of several seconds of arc (linear
size of order $500$--$1000$ a.u.), usually surrounded by a bright diffuse envelope.
Their spectra mainly show emission lines of hydrogen and some other low-excitation
elements. A shock wave propagating with velocities $40$--$200$ km s$^{-1}$ through
a gas with a density of $\sim 10^2$ cm$^{-3}$ is apparently the main source of excitation [70].
As in the case of radio galaxies, the activity of Herbig-Haro objects is dictated by
collimated outflows which are well seen in forbidden lines. Nearly $60\%$ of the objects
demonstrate both jets, while in other cases the receding jet is blocked by the accretion
disc. The extent of the optical jets is of order $0.01$--$2$ pc, and their velocity
reaches $600$ km s$^{-1}$. The gas density in the jets is estimated to be $10$--$100$ cm$^{-3}$,
and the mass outflow rate comes to $10^{-9}$--$10^{-10}$ $M_{\odot}$ yr$^{-1}$. The degree of
collimation of the jets (the ratio of the observed length to the width) can be as high as 30.
The total jet opening angle is in the range of $5$--$10^{\circ}$. In addition to highly
elongated jets, molecular outflows with a much smaller collimation degree are observed near
young stars. Their size may run to \mbox{$0.04$--$4$ pc,} and the velocity of gas motion does not
exceed \mbox{$5$--$100$ km s$^{-1}$.} Here we should stress that this velocity is much higher than
the speed of sound in an outflow with a temperature of only $10$--$90$ K. The total mass of
the ejected gas is estimated to be \mbox{$0.1$--$200$ $M_{\odot}$,} and the total kinetic energy
stored in the molecular outflows can reach $10^{43}$ and even $10^{47}$ erg. The direct
observation of rotation of the jets is the most important recent discovery. The characteristic
velocities at an axial distance of $20$--$30$ a.u. range from 3--10 km s$^{-1}$ [72, 73].
There is also direct evidence of the spiral structure of the magnetic field in the jets [74].
All these facts unambiguously support the MHD model.
As in jets from microquasars, a strong instability frequently develops in collimated outflows
from young stars at large distances from the central engine (see Fig. 4), so that the outflow
is split into separate blobs. On the other hand, as seen from Fig. 5, the flow near the base
of the jet can be considered sufficiently regular.
\begin{figure}[t]
\begin{center}
\includegraphics[width=0.8\columnwidth]{fig1_05.ps}
\caption{Formation of a jet from a young star in the system HH 30 [70].
The accretion disc is clearly seen. Here also the scale corresponds to 1000 a.u.
}
\label{fig1_05}
\end{center}
\end{figure}
As for the physical nature of collimated jet formation, this question is still far from solved.
It is only clear that the power of the central engine is always sufficient to accelerate the
outflowing gas; however, the mechanism of energy transformation remains unclear. We stress that
in contrast to relativistic galactic objects (for example, microquasars), where the formation
of jets is possibly caused by supercritical accretion, the luminosity in young stars never
approaches the Eddington limit. On the other hand, it is clear that the key role in the collimated
outflow formation is just played by accretion discs which undoubtedly exist around young stars.
This is supported by the direct correlation between the power of the gas flux and the mass of the
disc, estimated from its luminosity, as well as some other correlations [75, 76]. The parameters
of the discs can be very different. For example, their masses range from $0.1$--$100$
$M_{\odot}$, while the outer radii can vary from 10 a.u. to 0.1 pc.
It is important that, in contrast to discs around relativistic objects (neutron stars and
black holes), the gas temperature in discs around young stars is only \mbox{$20$--$100$ K.} As a
result, as in the case with active galactic nuclei, neither the radiation pressure force
nor gas pressure can explain the high velocities observed in the collimated outflows [71].
Therefore, to explain the jet formation and particle acceleration, models in which the magnetic
field plays the key role and effectively mediates the interaction between the accretion disc
and the jet were invoked once again. Because the real structure of the magnetic field in the
proximity of a young star is presently unknown, here, too, both models in which the magnetic
field of the star itself has a dominant role [77] and models in which the magnetic field of
the disc plays the decisive role [75, 78] have been proposed. It is seen that here we meet
the same problems regarding the structure of the initial magnetic field as in the study of
the black hole magnetosphere.
\section{Basics of the MHD approach}
\subsection{The key idea --- unipolar inductor}
As already said, the notion of a unipolar inductor is the main physical idea that
underlies the MHD theory of compact objects. Referring to Fig. 6, a rotating magnetized
ball can serve as the battery that determines the energy release from the central engine.
Indeed, assuming the high conductivity of the ball, the freezing-in condition of the magnetic
field, viz.
\begin{equation}
{\bf E}_{\rm in}
+ \frac{{\bf\Omega} \times {\bf r}}{c}\times {\bf B}_{\rm in} = 0
\label{d1}
\end{equation}
(i.e., simply the condition that the electric field in the rotating reference frame
vanishes), leads to the appearance of the potential difference $\delta U$ between points
$a$ and $b$. To an order of magnitude, this potential difference can be
\begin{equation}
\delta U \sim E R_{0} \sim \frac{\Omega R_{0}^2}{c} B,
\label{d1'}
\end{equation}
where $R_{0}$ is the transverse size of the working area. As a result, the total energy
release $W_{\rm tot}$ on the external load ${\cal R}$ will be given as
\begin{equation}
W_{\rm tot} = I \delta U,
\label{IU}
\end{equation}
where the electric current $I = \delta U/{\cal R}$. Here, however, several conditions
should be met. First, the electric circuit must touch the ball at different latitudes,
i.e., at points with different electric potentials. Second, the electric circuit should
rotate with an angular velocity $\Omega$ different from that of the magnetized ball.
The current flowing along a wire tightly welded on the ball will be absent.
\begin{figure}
\begin{center}
\includegraphics[width=0.8\columnwidth]{fig2_01.ps}
\end{center}
\caption{The unipolar inductor as the source of a direct current.
Inside the magnetized ball, the electric current flows against the
electric field direction.
}
\label{fig2_01}
\end{figure}
We stress that the energy source [electromotive force (EMF)] in the unipolar inductor is
due to the kinetic energy of rotation. Indeed, as seen from Fig. 6, charges inside the ball
move against the direction of the electric field. This becomes possible due to the force
by which the lattice acts on charges carried along the wire, which violate the freezing-in
condition inside the ball. Conversely, the Ampere force acting from the side of the surface
electric current on the ball's material brakes its rotation. Therefore, the principle of
work of the unipolar inductor (or, as it is sometimes called, the unipolar Faraday generator)
is not the Faraday effect as such (where the EMF induced in a current loop depends on the
variation of the magnetic flux), since the flux through the circuit remains constant. Notice
that the reverse situation is also possible: if one applies a potential difference to a
magnetized ball (i.e., if one replaces the load in Fig. 6 by the voltage source), the ball
starts rotating. On the site http://fiziks.org.ua/samyj-prostoj-v-mire-elektrodvigatel/,
which is devoted to laboratory studies in secondary school, one can find a video illustrating
the work of such a device.
As we have understood, for the central engine to operate it is necessary to have:
\begin{itemize}
\item
rotating body;
\item
regular magnetic field, and
\item
well-conducting wire.
\end{itemize}
Then the current, and hence the energy losses, will be determined by the value of the
external resistance ${\cal R}$. Let us see now whether these conditions can be met in
compact astrophysical objects.
As we have seen, a central rotating body in active astrophysical sources is undoubtedly
present. For example, the spin periods of young stars are about several days (the inner
parts of accretion discs rotate even faster). The spin periods of most radio pulsars are
close to 1 s; however, they can be as small as a few milliseconds, which is already close
to the limiting speed of rotation ($\Omega R/c \sim 0.1$). The rotational velocities of
black holes in active galactic nuclei, to tell the truth, are unknown, but we can suppose
that due to disc accretion (it is in this way that the millisecond-period pulsars are
thought to have been spun up) their spin parameter $\Omega_{\rm H}R/c = a/2M$ (see the
Appendix) can also be sufficiently large. For example, the estimate of the black hole
rotational velocity in the nucleus of Seyfert galaxy MCG 06-30-15, as inferred from the
iron 6.4-keV line profile distortion, yields $a/M = 0.989_{-0.002}^{+0.009}$ [79] (see
also Ref [80]). As a result, the kinetic energy of rotation ${\cal E}_{\rm kin} = J_r \Omega^2/2$
stored in the central engine turns out to be quite sufficient to explain the energy source
of activity of compact objects.
There are no particular problems with a regular magnetic field, either. In young stars, the
proper magnetic field $B_{0}$ is measured directly and can be as high as \mbox{$10^{3}$ G [71].}
At present, there are no direct observations of magnetic fields in radio pulsars, but they can
be measured in X-ray (accreting) pulsars, which are also neutron stars [11]. Therefore,
nobody now doubts that the magnetic field of a neutron star can reach $10^{12}$ G, and even
extend up to $10^{15}$ G in magnetars [81]. The situation is worse with the magnetic fields
of supermassive black holes. As is well known, a black hole cannot have a proper magnetic
field, but the field can be generated in the surrounding accretion discs [82]. Unfortunately,
so far there is no self-consistent theory of such generation, so we have to apply the estimate
$B_0 \sim B_{\rm Edd}$, where
\begin{equation}
B_{\rm Edd} \approx 10^4 \left(\frac{M}{10^9M_{\odot}}\right)^{-1/2} {\rm G}.
\label{bedd}
\end{equation}
Let us keep in mind that such an estimate comes from the simple assumption that the energy
density of the magnetic field is comparable to the total energy density in the accreting
plasma yielding the Eddington luminosity (2). Clearly, estimate (7) represents rather an
upper limit of the magnetic field near the black hole. In particular, it
does not take into account the contribution from the thermal pressure, which can be significant
in gamma-ray burst sources.
Finally, the problem of the 'electric wiring' can also be easily solved at first glance. Due
to the presence of a strong magnetic field, in all cases the Larmor radius of particles
$r_{\rm L} = m c v/eB$ is always much smaller than the size $R$ of the central engine. Therefore,
one can consider with good accuracy that the electric current flows along the direction of the
regular magnetic field. However, here we meet the problem of current closing, since particles
in the region of the load must move across the magnetic field. We shall necessarily discuss
this point below.
As an example, Fig. 7 illustrates how the braking occurs in an axisymmetric magnetosphere
of radio pulsars. Clearly, the total current flowing out of the pulsar surface must vanish;
thus, there must necessarily be a reverse current in the magnetosphere to compensate for the
loss of charges from the neutron star. As a result, currents ${\bf J}_{\rm s}$ closing the
longitudinal currents in the magnetosphere must flow over the pulsar surface. The ponderomotive
action of these currents must brake the rotation of the radio pulsars [3, 56].
\begin{figure}
\begin{center}
\includegraphics[width=0.95\columnwidth]{fig2_07.ps}
\end{center}
\caption{The structure of electric currents (contour arrows) near the polar
caps of a neutron star. The Ampere force related to the surface current
${\bf J}_{\rm s}$, produces the torque ${\bf K}$ braking the neutron star
rotation. Above the acceleration region, the energy flux is predominantly
transported by the Poynting vector (hatched arrows).
}
\label{fig2_07}
\end{figure}
Thus, the problem of the magnitude of potential difference is solved quite easily. But the
problem of the load that determines the current $I$ and, hence, the energy
losses, proved much more difficult. A long way had to be covered in order to solve it, and
this, essentially, will be discussed in this review. Nevertheless, we shall go somewhat
ahead and give here the preliminary estimates confirming the applicability of the discussed
mechanism. As shown below, a good estimate of the electric current density is given by the
expression
\begin{equation}
j_{\rm GJ} = \rho_{\rm GJ}c,
\label{jGJ}
\end{equation}
where
\begin{equation}
\rho_{\rm GJ} = -\frac{{\bf \Omega} \cdot {\bf B}}{2\pi c}
\label{GJ}
\end{equation}
is the electric charge density that is needed for the electric field in the rotating
reference frame to vanish. Formula (9) can be easily derived from relation (4). It
was first applied to the neutron star magnetosphere in the pioneering paper by P Goldreich
and W H Julian [3], so the charge density (9) is usually called the Goldreich density.
\begin{table*}[ht]
\caption{ Parameters of the central engine: AGN --- active galactic nucleus,
GRB --- gamma-ray burst, $\mu$QSO --- microquasar, PSR --- radio pulsar,
msPSR --- millisecond radio pulsar, and YSO --- young stellar object.}
\vspace{0.3cm}
\centering
\begin{tabular}{|l|c|c|c|c|c|c|}
\hline
& AGN & GRB & $\mu$QSO & PSR & msPSR & YSO \\
\hline
Mass $M$ in $M_{\odot}$& $10^6$--$10^9$ & $\sim 10$ & $\sim 10$ & $\approx 1.4 $ & $\approx 1.4 $ & $\sim 10$ \\
\hline
Radius $R$, cm & $10^{11}$--$10^{14}$ & $\sim 10^6$ & $\sim 10^6$ & $\sim 10^6$ & $\sim 10^6$ & $\sim 10^{11}$ \\
\hline
Working radius $R_0$ & $\sim R$ & $\sim R$ & $\sim R$ & $(\Omega R/c)^{1/2} R$ & $(\Omega R/c)^{1/2} R$ & $\sim R$ \\
\hline
Period $P$ & $10$--$10^3$ \, s & $\sim 1$ \, ms & $\sim 1$ \, ms & $\sim 1$ \, s & $1.39$--$10$ \, ms & $1$--$10$ \, d \\
\hline
$\Omega R/c$ & $\geq 0.1$ & $\geq 0.1$ & $\geq 0.1$ &$ \sim 10^{-4}$ &$\sim 10^{-1}$ & $\sim 10^{-5}$\\
\hline
Magnetic field $B_0$, G & $10^3$--$10^4$ & $\sim 10^{15}$ & $\sim 10^{10}$ & $\sim 10^{12}$& $\sim 10^{8}$ & $\sim 10^3$ \\
\hline
Energy storage ${\cal E}_{\rm kin}$, erg & $10^{58}$--$10^{61}$ & $\sim 10^{52}$ & $\sim 10^{52}$ & $10^{44}$--$10^{46}$ & $\sim 10^{51}$ & $\sim 10^{44}$ \\
\hline
Dimensionless current $i_0$& 1 & 1 & 1 & 1 & 1 & $ \sim c/v_{\rm in}$ \\
\hline
Power $W_{\rm tot}$, erg s$^{-1}$& $10^{42}$--$10^{45}$ & $10^{51}$--$10^{52}$ & $\sim 10^{38}$ & $10^{31}$--$10^{34}$ & $10^{34}$--$10^{35}$ & $\sim 10^{35}$ \\
\hline
Lifetime $\tau_{\rm D}$, yr & $\sim 10^{7}$ & $\sim 10^{-6}$ & $\sim 10^{4}$ & $10^{6}$--$10^{7}$ & $10^{8}$--$10^{9}$& $\sim 10^{4}$ \\
\hline
Current $I$, CGSE & $10^{26}$--$10^{28}$ & $\sim 10^{31}$ & $\sim 10^{25}$ & $10^{21}$--$10^{22}$ & $\sim 10^{22}$& $ \sim 10^{25}$ \\
\hline
\end{tabular}
\label{table2_01}
\end{table*}
Clearly, the total electric current circulating in the magnetosphere of the central
engine can be conveniently written out in the form
\begin{equation}
I_{\rm tot} = i_{0} I_{\rm GJ}.
\label{iq0}
\end{equation}
Here $i_0$ is the dimensionless current, and \mbox{$I_{\rm GJ}= \pi R_0^2 c \rho_{\rm GJ}$,}
i.e., for the case $\rho_{\rm GJ} \approx$ const we obtain
\begin{equation}
I_{\rm GJ}=\frac{\Omega B_{0}R_0^2}{2}.
\label{d46}
\end{equation}
Finally, $R_0$ is again the size of the working area on the central engine surface.
For black holes we can set $R_0 \approx R=r_{\rm g}$, and for neutron stars (radio
pulsars) it must be on the order of the radius of a polar cap from which magnetic field
lines can go beyond the light cylinder $R_{\rm L} = c/\Omega$. Indeed, inside the
closed magnetosphere, by virtue of the remarkable Ferraro isorotation law, the plasma
starts rotating with the star as a solid body, and, hence, this region cannot work as
a unipolar inductor. The working area will include only the region of open field lines,
inside which the plasma rotational velocity can be different from that of the star.
As a result, for the dipole magnetic field we obtain
\begin{equation}
R_{0} \approx R\left(\frac{\Omega R}{c}\right)^{1/2}.
\label{r0}
\end{equation}
For relativistic strongly magnetized wind it is natural to assume that
\begin{equation}
i_0 \approx 1,
\label{i_0r}
\end{equation}
which corresponds to a free plasma outflow with the velocity $c$. As we shall see, this
estimate is indeed correct. Therefore, the total energy losses can be estimated as
\begin{equation}
W_{\rm tot} \approx \left(\frac{\Omega R_0}{c}\right)^{2} B_{0}^{2} R_{0}^{2} c.
\label{i_0rnew}
\end{equation}
In consequence of this, as shown in Table 1, the unipolar inductor model allows us
to explain both the total energy release $W_{\rm tot}$ and the time of activity of compact
sources, $\tau_{\rm D} = {\cal E}_{\rm kin}/W_{\rm tot}$. As mentioned above, for radio
pulsars estimate (14) with account for relation (12) coincides to within an order of
magnitude with the magnetodipole losses (3).
For nonrelativistic outflows, estimate (13) is incorrect, and, as a detailed analysis
shows, $i_0 \gg 1$ [9]. For a sufficiently rapid rotation with $\Omega > \Omega_{\rm cr}$,
where
\begin{equation}
\Omega_{\rm cr} = \frac{v_{\rm in}}{R_{0}}\left(
\frac{4\pi\rho_{\rm in}v_{ \rm in}^2}{B_{0}^2}\right)^{1/2} \sim 10^{-6} \, {\rm s}^{-1},
\label{ocr}
\end{equation}
we have
\begin{equation}
i_0 \approx \frac{c}{v_{\rm in}}\left(
\frac{\Omega_{\rm F}}{\Omega_{\rm cr}}\right)^{-2/3},
\label{i00}
\end{equation}
and the dimensionless current for slow rotation limit takes the form
\begin{equation}
i_0 \approx \frac{c}{v_{\rm in}}.
\label{i_0n}
\end{equation}
Here $\rho_{\rm in}$ is the density of the outflowing matter near the surface of the star,
and $v_{\rm in}$ is the characteristic velocity of the outflow along the jet axis. As a
result, the total energy losses for rapidly rotating stars can be expressed through the
directly observed quantities:
\begin{equation}
W_{\rm tot} \approx \Omega^{4/3} \Psi_{\rm tot}^{4/3} {\dot M}^{1/3},
\label{exta1}
\end{equation}
i.e., through the total magnetic flux $\Psi_{\rm tot} = \pi R_0^2 B_0$, the rotational
angular velocity $\Omega$, and the mass loss rate in the jet ${\dot M}$. For the parameters
typical in young stars we have
\begin{eqnarray}
W_{\rm tot} \sim 10^{36} \,
\left(\frac{P}{10^{6} \, {\rm s}}\right)^{4/3}
\left(\frac{B_{\rm in}}{10^{3} \, {\rm G}}\right)^{4/3} \nonumber \\
\left(\frac{R_{\rm in}}{10^{11} \, {\rm cm}}\right)^{8/3}
\left(\frac{\dot M}{10^{-9}\, M_{\odot} \, {\rm yr}^{-1}}\right)^{1/3} \, {\rm erg} \,{\rm s}^{-1}.
\label{exta3}
\end{eqnarray}
It is seen that this value is indeed close to the energy losses from young stellar objects.
Thus, the unipolar inductor model allows us to explain the main jet characteristics for nonrelativistic
sources, too.
Interestingly, the knowledge of the total energy losses $W_{\rm tot}$ immediately allows the total
longitudinal electric current the circulating in the magnetosphere to be estimated. Indeed, by
comparing expressions (11) and (14), we straightforwardly obtain
\begin{equation}
I \approx i_0 c^{1/2}W_{\rm tot}^{1/2}.
\label{i_0}
\end{equation}
The characteristic amplitudes of currents are also collated in Table 1.
\subsection{Grad-Shafranov equation method}
The Grad-Shafranov equation method lies at the heart of the analytical theory which,
in our opinion, is able to quite successfully describe the main properties of active
compact astrophysical sources. Simply speaking, this approach describes axisymmetric
stationary flows in the framework of ideal magnetohydrodynamics. This approximation
is based on the assumption of a high conductivity of the plasma that fills the
magnetosphere of the central engine (the high energy release guarantees a high degree
of ionization of matter, and the effective production of electron-positron pairs in
the vicinity of black holes). Moreover, most of the sources discussed above (except
for radio pulsars) can be considered to a good approximation as axisymmetric
and stationary.
\begin{figure}
\begin{center}
\includegraphics[width=0.6\columnwidth]{fig2_02.ps}
\end{center}
\caption{Axisymmetric magnetic surfaces $\Psi(r,\theta) = $ const.
}
\label{fig2_02}
\end{figure}
The attractiveness of this approach is related to the fact that there are quite a lot of
integrals of motion in stationary ideal magnetohydrodynamics, i.e., quantities which are
conserved along particle trajectories. This immediately provides us with important
information without complicated calculations. Indeed, to determine the height a throwing
stone reaches it is not necessary to solve equations of its motion: it is sufficient to
apply the energy conservation law.
In the axisymmetric case, as illustrated in Fig. 8, the magnetic field vectors must
lie on the magnetic surfaces which can be easily parametrized using the magnetic flux
function $\Psi(r,\theta)$ that determines the magnetic field
\begin{equation}
{\bf B} = \frac{{\bf\nabla}\Psi \times {\bf e}_{\varphi}}{2\pi\varpi}
-\frac{2I}{c\varpi}{\bf e}_{\varphi}.
\label{d30}
\end{equation}
Here $\varpi = r\sin\theta$ is the distance from the rotational axis, and the
numerical coefficient in the first term is chosen such that the function
$\Psi(r,\theta)$ indeed coincides with the magnetic flux passing through
a circle $r$, $\theta$, \mbox{$0<\varphi<2\pi$}. As for the quantity $I(r,\theta)$,
it represents the total electric current flowing through the same circle.
It is easy to check that the following important properties are satisfied.
\begin{enumerate}
\item
At all times ${\rm d}\Psi = {\bf B} \cdot {\rm d}{\bf S}$ (${\rm d}{\bf S}$
is the surface element). Therefore, the
function $\Psi(r,\theta)$ indeed bears the sense of the magnetic flux.
\item
Since the poloidal part of the magnetic field in formula (21) can
be written out as $(2 \pi)^{-1} \nabla \Psi \times \nabla \varphi$,
the condition $\nabla \cdot {\bf B} = 0$ is automatically
satisfied. Thus, three components of the magnetic field are completely
determined by two scalar functions $\Psi(r,\theta)$ and $I(r,\theta)$.
\item
For the same reason it is clear that the condition ${\bf B} \cdot \nabla \Psi = 0$
will be satisfied for axisymmetric flows. Therefore, the lines $\Psi(r,\theta) =$ const
define the form of magnetic surfaces. As a result, the integrals of motion should
depend on only one scalar function, $\Psi(r,\theta)$.
\end{enumerate}
Let us now understand which integrals of motion appear in the case of axisymmetric
stationary flows. Incidentally, the very structure of the Grad-Shafranov equation
method will become clear. For simplicity, let us consider first a purely hydrodynamic
flow. In this case, in analogy with relation (21), it is necessary to introduce the
function $\Phi(r,\theta)$ of hydrodynamic flux defined as
\begin{equation}
\rho{\bf v}_{\rm p}
= \frac{\nabla \Phi \times {\bf e}_{\varphi}}{2\pi r\sin\theta},
\label{b1}
\end{equation}
where hereinafter the subscript 'p' will correspond to the poloidal [i.e., lying in the
($r, \theta$) plane] components of vectors.
In hydrodynamics, there are five scalar equations (the mass continuity equation,
three components of the momentum Euler equation, and the energy conservation equation)
for five unknown quantities --- three velocity components, and two thermodynamic functions.
However, due to the axial symmetry, stationarity, and ideality of the flow three of five
equations can be represented in the form $({\bf v}\nabla){\cal I}^{(i)} = 0$, which means
that integrals ${\cal I}^{(i)}$ must be constant on the surfaces $\Phi(r,\theta) =$ const.
As is well known, these integrals include the energy (Bernoulli integral) $E_{\rm n}$,
the specific angular momentum $L_{\rm n}$, and the entropy $s$:
\begin{eqnarray}
E_{\rm n} & = & E_{\rm n}(\Phi) = \frac{v^2}{2} + w(\rho, s) + \varphi_{\rm g},
\label{b2} \\
L_{\rm n} & = & L_{\rm n}(\Phi) = v_{\varphi}r\sin\theta,
\label{b6} \\
s & = & s(\Phi).
\label{b3}
\end{eqnarray}
Here, $\varphi_{\rm g}$ is the gravitational potential, $w(\rho,s)$ is the specific enthalpy,
and the subscript 'n' corresponds to nonrelativistic quantities, while Bernoulli integral
$E_{\rm n}(\Phi)$ corresponds to the projection of the Euler equation (i.e., the equation
of the force balance) onto the direction along the poloidal velocity ${\bf v}_{\rm p}$, and
the angular momentum $L_{\rm n}(\Phi)$ represents the projection onto the unit vector
${\bf e}_{\varphi}$. The remaining two first-order equations can be reduced to one second-order
equation for the flux function $\Phi(r,\theta)$. It is clear that this equation will describe
the force balance in the direction perpendicular to surfaces $\Phi(r,\theta) =$ const. In the
compact form, it can be written out as
\begin{eqnarray}
-\varpi^2 \nabla_{k}
\left(\frac{1}{\varpi^{2}\rho}\nabla^{k}\Phi\right)
-4\pi^{2}\rho L_{\rm n}\frac{{\rm d}L_{\rm n}}{{\rm d}\Phi} \nonumber \\
+4\pi^{2}\varpi^{2}\rho \frac{{\rm d}E_{\rm n}}{{\rm d}\Phi}
-4\pi^2\varpi^{2}\rho \frac{T}{m_{\rm p}}\frac{{\rm d}s}{{\rm d}\Phi}=0.
\label{gscomp}
\end{eqnarray}
It should be noted that we deliberately defined the enthalpy $w$ as a function of
density $\rho$ and entropy $s$. The point is that Bernoulli equation with the use
of definition (22) can be recast into the form
\begin{equation}
E_{\rm n} = \frac{(\nabla\Phi)^2}{8\pi^2\varpi^2\rho^2}
+ \frac{1}{2}\,\frac{L_{\rm n}^2}{\varpi^2} + w(\rho,s) + \varphi_{\rm g}.
\label{b17}
\end{equation}
One can see that Bernoulli equation written in this form, in addition to the integrals
of motion and the flux function $\Phi(r,\theta)$, contains only the density $\rho$. Consequently,
it indirectly defines the density $\rho$ though the flux function $\Psi$ and integrals of motion:
\begin{equation}
\rho = \rho(\nabla\Phi;E_{\rm n},L_{\rm n},s;r,\theta).
\label{b18'new}
\end{equation}
This implies that after substituting Eqn (28) into Eqn (26), the latter will contain only
one unknown flux function $\Phi(r,\theta)$ and three integrals of motion depending on it.
Such an equation can also be written out in the framework of ideal magnetohydrodynamics,
too. In this case, however, not three but five integrals of motion exist. Two additional
integrals come from the freezing-in condition
\begin{equation}
{\bf E} + {\bf v} \times {\bf B}/c = 0.
\label{frozen}
\end{equation}
Indeed, from condition (29) follows that the electric field is perpendicular to the
magnetic field. In the axisymmetric case this means that magnetic surfaces
\mbox{$\Psi(r,\theta) =$ const} will be equipotential ones. This condition can be conveniently
rewritten in the form
\begin{equation}
\Omega_{\rm F} = \Omega_{\rm F}(\Psi),
\label{k41'}
\end{equation}
where the scalar quantity $\Omega_{\rm F}$ determines the electric field according
to the definition
\begin{equation}
{\bf E} = -\frac{\Omega_{\rm F}}{2\pi c}{\bf\nabla}\Psi.
\label{d34}
\end{equation}
This is related to the fact that:
\begin{itemize}
\item
in the axisymmetric case ($\partial/\partial t = 0$) Maxwell equation
$\nabla \times {\bf E} = 0$ leads to the condition $E_{\varphi} = 0$,
\item
the freezing-in condition yields $E_{\parallel} = 0$,
\item
the definition [31] together with Maxwell equation $\nabla \times {\bf E} = 0$ leads to the condition
$\nabla\Omega_{\rm F} \times \nabla \Psi = 0$, where relation (30) comes from.
\end{itemize}
Function $\Omega_{\rm F}$ introduced in this way bears the meaning of the angular
velocity of particles (more precisely, the motion of particles is the sum of rotation with
the angular velocity $\Omega_{\rm F}$ and sliding along the magnetic field). Condition (30)
represents the Ferraro isorotation law [83], according to which the angular velocity of
particle motion relative to the magnetic field must be constant on axisymmetric magnetic
surfaces.
On the other hand, the freezing-in plasma condition implies that plasma velocity vectors ${\bf v}$
have also to lie on the magnetic surfaces, i.e., the flux of matter does not intersect the
boundaries of the magnetic surfaces. This means that the particle flux function $\Phi(r,\theta)$
must be a function of the magnetic flux $\Psi(r,\theta)$. This fact allows us to introduce one
more integral of motion
\begin{equation}
\eta_{\rm n}(\Psi) = \frac{{\rm d}\Phi}{{\rm d}\Psi},
\label{p5}
\end{equation}
which, as evidenced by the foregoing, bears the sense of the ratio of the particle flux to the
magnetic field flux. Correspondingly, the poloidal velocity of matter can be written as
\begin{equation}
{\bf v}_{\rm p} = \frac{\eta_{\rm n}}{\rho}{\bf B}_{\rm p}.
\label{p2'extra}
\end{equation}
As for the energy and angular momentum integrals (which in the nonrelativistic
case are usually considered as functions of the particle flux), they now take the form
\begin{eqnarray}
E_{\rm n}(\Phi) & = & \frac{\Omega_{\rm F} I}{2\pi \eta_{\rm n} c}
+\frac{v^2}{2} + w + \varphi_{\rm g},
\label{5a} \\
L_{\rm n}(\Phi) & = & \frac{I}{2\pi \eta_{\rm n} c}+ v_{\varphi}r\sin\theta,
\label{6a}
\end{eqnarray}
respectively. The entropy $s(\Psi)$ is ones again the one more (fifth) invariant. It is clear
that both particles and electromagnetic field contribute to the energy and angular momentum,
and, as can be easily checked, the term $\Omega_{\rm F} I/2\pi \eta_{\rm n} c$
corresponds simply to the Poynting vector flux.
Next, Bernoulli equation (34) can now be rewritten in the form
\begin{eqnarray}
\frac{{\cal M}^4}{64\pi^4\eta_{\rm n}^2}\left(\nabla\Psi\right)^2
= 2\varpi^{2} (E_{\rm n} - w -\varphi_{\rm g}) \nonumber \\
-\frac{(\Omega_{\rm F}\varpi^2- L_{\rm n}{\cal M}^2)^2}
{(1-{\cal M}^2)^2}
-2\varpi^2\Omega_{\rm F}\frac{L_{\rm n}-\Omega_{\rm F}\varpi^2}
{1-{\cal M}^2},
\label{11a}
\end{eqnarray}
where
\begin{equation}
{\cal M}^{2}=\frac{4\pi\eta_{\rm n}^{2}}{\rho}.
\label{10a}
\end{equation}
The quantity ${\cal M}^{2}$ is the square of the Mach number of the poloidal
velocity $v_{\rm p}$ relative to the poloidal component of the Alfv\'en
velocity:
\begin{equation}
v_{\rm Ap} = \frac{B_{\rm p}}{\sqrt{4\pi \rho}},
\label{alfA}
\end{equation}
i.e., ${\cal M}^2=v_{\rm p}^2/v_{\rm Ap}^2$. It should be recalled that the specific
enthalpy $w$ in equation (36) must be considered as a function of entropy $s$, as well
as of the Mach number ${\cal M}^2$ and the integral $\eta_{\rm n}$. The corresponding
relationship has the form
\begin{equation}
\nabla w = c_{\rm s}^2 \left(2\frac{\nabla\eta_{\rm n}}{\eta_{\rm n}}
-\frac{\nabla {\cal M}^{2}}{{\cal M}^{2}}\right)
+\left[\frac{1}{\rho}\left(\frac{\partial P}{\partial s}\right)_{n}
+\frac{T}{m_{\rm p}}\right]\nabla s.
\label{12aa}
\end{equation}
In consequence, as in the hydrodynamic limit, Bernoulli equation allows one to
determine, albeit indirectly, the quantity ${\cal M}^{2}$ via the magnetic flux
$\Psi(r,\theta)$ and five integrals of motion:
\begin{equation}
{\cal M}^2 = {\cal M}^2(\nabla\Psi;E_{\rm n},L_{\rm n},s,\eta_{\rm n},\Omega_{\rm F};r,\theta).
\label{b18'qqq}
\end{equation}
As for the projection of the force balance onto the direction perpendicular to the magnetic
surfaces, it can be written in the form{\footnote{Unfortunately, in monograph [9]
terms $-w$ $- \varphi_{\rm g}$ in Eqn (4.102) were discarded.}}
\begin{eqnarray}
&&\frac{1}{16 \pi^3 \rho}
\nabla_{k}\left(\frac{1-{\cal M}^2}{\varpi^{2}}
\nabla^{k}\Psi\right)
+\frac{{\rm d}E_{\rm n}}{{\rm d}\Psi}
\nonumber \\
&&+\frac{\Omega_{\rm F}\varpi^2 - L_{\rm n}}{1-{\cal M}^2} \,
\frac{{\rm d}\Omega_{\rm F}}{{\rm d}\Psi}
+\frac{1}{\varpi^2} \,
\frac{{\cal M}^2 L_{\rm n} - \Omega_{\rm F}\varpi^2}{1-{\cal M}^2}
\frac{{\rm d}L_{\rm n}}{{\rm d}\Psi}
\nonumber \\
&&+\left[2(E_{\rm n} -w - \varphi_{\rm g}) +
\frac{\Omega_{\rm F}^2\varpi^4 -2\Omega_{\rm F}L_{\rm n}\varpi^2
+ {\cal M}^2 L_{\rm n}^2}{\varpi^2(1-{\cal M}^2)}\right]
\nonumber \\
&&\times\frac{1}{\eta_{\rm n}}\frac{{\rm d}\eta_{\rm n}}{{\rm d}\Psi}
-\frac{T}{m_{\rm p}} \, \frac{{\rm d}s}{{\rm d}\Psi} = 0.
\label{19b}
\end{eqnarray}
Now the structure of the treatment considered here becomes clear. Equation (41)
jointly with Bernoulli equation (36) determines the value of the magnetic flux
$\Psi(r,\theta)$. Then, again using Bernoulli equation, one can determine
the value of the Mach number ${\cal M}$ at each point. After that it turned out
that the number of integrals of motion is enough for determining all other
quantities from simple algebraic equations. For example, one arrives at [84]
\begin{eqnarray}
\frac{I}{2\pi} & = & c\eta_{\rm n}
\frac{L_{\rm n}-\Omega_{\rm F}\varpi^2}{1-{\cal M}^2},
\label{Inrel} \\
v_{\varphi} & = & \frac{1}{\varpi}\frac{\Omega_{\rm F}\varpi^2
-L_{\rm n}{\cal M}^2}{1-{\cal M}^2},
\label{9a}
\end{eqnarray}
and, respectively, $\rho = 4 \pi \eta_{\rm n}^2/{\cal M}^2$. This is essentially the main
attractiveness of the approach we discuss. Sometimes, as we shall see, the key properties
can be directly obtained from algebraic relations. In other words, by making sufficiently
reasonable assumptions about the structure of the flow [i.e., about flux function $\Psi(r,\theta)$],
it is possible not to solve equation (41) at all and to analyze only algebraic, albeit indirect,
relations.
The full version of the nonrelativistic equation containing all five invariants was first
formulated by L S Solov'ev in 1963 in the third volume of the {\it Reviews of Plasma Physics} [85].
Being virtually unknown for astrophysicists, this equation was later reformulated anew several
times [86-88]. For this reason, in particular, to date there has been no unique system of
notations, so sometimes it is difficult to compare the results of different studies. In the
literature, equations of this type are commonly called Grad-Shafranov equations, which were
formulated at the end of the 1950s in relation to controlled thermonuclear fusion [89, 90],
although the hydrodynamic version of this equation was known even earlier (see, for example,
Ref. [91]). Similar equations going back to the classical Tricomi equation, were discussed
as early as the beginning of the twentieth century in the context of transonic hydrodynamic
flows [92, 93].
For simplicity, we have written out above equations only for the nonrelativistic case.
However, it was not too difficult to obtain the corresponding equations both for the
relativistic case [94] and for flows in the vicinities of nonrotating [95] and rotating
[96, 97] black holes, since the Kerr metric is axisymmetric and stationary. It is
these relativistic equations that we shall discuss below. In the main text we shall try
to formulate sufficiently simple asymptotic expressions by focusing on the qualitative
description of the flow properties. Sufficiently lengthy full equations are presented
in the Appendix. Here, we shall restrict ourselves by writing out additionally the
integrals of motion for relativistic flows in flat space.
Clearly, magnetic surfaces remain equipotential in the relativistic case, too. Thus, the
angular frequency $\Omega_{\rm F}$ in definition (31) remains the integral of motion. As
for the integrals of energy $E$ and the $z$-component of the angular momentum $L$, now
they should be written out as{\footnote{To avoid misunderstanding, from now on the electric
field {\bf E} will always be boldfaced.}}
\begin{eqnarray}
& & E = E(\Psi) = \frac{\Omega_{\rm F}I}{2\pi}
+ \gamma \mu \eta c^2,
\label{p31} \\
& & L=L(\Psi) = \frac{I}{2\pi} + \mu\eta \varpi u_{\varphi} c.
\label{p32}
\end{eqnarray}
Here ${\bf u}$ is the spatial part of the four-velocity vector
\mbox{($\gamma = \sqrt{u^2 + 1}$} is the Lorentz factor), and
\begin{equation}
\mu
\approx m_{\rm p}c^2 + m_{\rm p}w + \dots
\label{c17}
\end{equation}
is the relativistic enthalpy including the rest mass of particles. Finally, the relativistic
integral of motion $\eta$ is now determined from the condition
\begin{equation}
{\bf u}_{\rm p} = \frac{\eta}{n}{\bf B}_{\rm p},
\label{p2'}
\end{equation}
where $n$ is the particle number density, and hereinafter all thermodynamic functions
will be defined in the co-moving reference frame. Thus, for relativistic flows where
$|{\bf u}_{\rm p}| \approx \gamma$ we simply have
\begin{equation}
\eta = \frac{n^{({\rm lab})}}{{B}_{\rm p}}.
\label{extraeta}
\end{equation}
It should be noted that the relativistic and nonrelativistic integrals of motion have
different dimensions, since the relativistic integrals are normalized not on the unit
matter flux ${\rm d}\Phi$ but on the unit magnetic flux ${\rm d}\Psi$.
It is evident that again both the energy flux and the angular momentum flux include the
contributions from the electromagnetic field and particles, with the electromagnetic
contribution [accurate up to an additional factor $\eta(\Psi)$] fully coinciding with
that obtained in the nonrelativistic limit. In the general magnetohydrodynamic case,
the total energy and angular momentum losses, $W_{\rm tot}$ and $K_{\rm tot}$, will be
determined by the relationships
\begin{eqnarray}
& & W_{\rm tot} = \frac{1}{c} \int_{0}^{\Psi_{\rm max}}E(\Psi){\rm d}\Psi,
\label{p32a} \\
& & \nonumber \\
& & K_{\rm tot} = \frac{1}{c} \int_{0}^{\Psi_{\rm max}}L(\Psi){\rm d}\Psi,
\label{p32b}
\end{eqnarray}
respectively.
\subsection{Supersonic flows}
\vspace*{.2cm}
{\bf 3.3.1. The model.} In order to clearly understand the main features of the model considered here, it is
convenient from the very beginning to analyze a sufficiently simple geometry of magnetic
surfaces, and to formulate basic parameters characterizing the flow. Figure 9 demonstrates
the simplestsplit monopole model of the magnetized wind [33], which during many years served
as the 'hydrogen atom' for all researchers who studied the nature of the activity of galactic
nuclei, gamma-ray bursts, and microquasars. Notably, most analytical results were obtained
exactly for such flows.
It is assumed in the framework of this model that the 'central engine' involves a compact
object (neutron star or black hole) and an accretion disc which separates the converging
and diverging magnetic field fluxes. The accretion disc here is nedeed both to separate
the oppositely directed magnetic field fluxes and, in the case of a black-hole magnetosphere,
to generate the regular poloidal magnetic field (it will be produced by the toroidal currents
flowing in the disc). In the absence of the accretion disc, a black hole, as is well known,
cannot have the proper magnetic field (the so-called 'no-hair theorem' [18]). In addition,
poloidal currents will also flow in the disc, closing the bulk currents flowing in the
magnetosphere. Notice that a similar configuration with a disc separating magnetic field
fluxes far from the neutron star also emerges in many models of radio pulsars [98 104],
because it is natural to assume that at large distances the flow becomes quasispherical.
\begin{figure}
\begin{center}
\includegraphics[width=0.9\columnwidth]{fig2_03.ps}
\end{center}
\caption{The structure of electromagnetic fields for the split monopole magnetic field near a
slowly rotating black hole [33]. Currents flowing in the highly conducting disc in the
equatorial plane provide both the poloidal magnetic field jump and the closure of bulk
currents flowing out the upper and bottom hemispheres.
}
\label{fig2_03}
\end{figure}
Next, it is important that the crossed fields $E_{\theta}$ and $B_{\varphi}$ form the
electromagnetic energy flux (the Poynting vector flux), which is directed along the
magnetic surfaces. It should be stressed that this energy is transferred at the zero
frequency, so the electromagnetic field that carries the energy is not an electromagnetic
wave in the usual sense. The electromagnetic energy flux therefore appears only due to
the longitudinal current generating the toroidal magnetic field; in the absence of
particles, such energy release becomes impossible. The plasma also moves along the
magnetic surfaces, so the sum of their energy fluxes is the integral of motion. On
the other hand, the longitudinal current $I$ is not the integral of motion, so the MHD
approximation we are considering allows, in principle, describing the current closure
phenomenon. However, the magnetic surfaces here remain equipotential. Therefore, such
flows can carry high electric voltages over large distances from the central engine.
This should always be borne in mind when discussing the interaction of a magnetized wind
with the surrounding medium. For example, this effect must be taken into account in close
binary systems with radio pulsars.
As for the main dimensionless quantities characterizing the flow, they include the
magnetization parameter $\sigma$, the particle production multiplicity $\lambda$, and
the compactness parameter $l_{a}$. The {\it magnetization parameter} $\sigma$ shows by
how much the electromagnetic energy flux near the central engine can exceed the particle
energy flux. Thus, as seen from the definition of the energy integral (44), it can be
more conveniently defined for relativistic flows as
\begin{equation}
\sigma = \left(\frac{E}{\mu\eta c^2}\right)_{\rm max},
\label{sigma}
\end{equation}
where the maximal value is chosen for all magnetic surfaces. As a result, the value
of $\sigma$ corresponds to the maximal Lorentz factor of the plasma that can be
reached in the case where all the electromagnetic field energy is transferred to
the particles. In other words, $\sigma$ is the maximum Lorentz factor that can be
achieved in the magnetized wind. Of course, here the mean hydrodynamic energy of
the plasma flowing out is assumed. In particular, for the split monopole magnetic
field (for which this quantity was first introduced by F C Michel [4] in 1969),
we obtain
\begin{equation}
\sigma = \frac{\Omega^2\Psi_{\rm tot}}{8\pi^2c^2 \mu\eta}.
\label{b12*}
\end{equation}
Correspondingly, for the nonrelativistic flow it is convenient to use the quantity
\begin{equation}
\sigma_{\rm n} = \frac{\Omega^2\Psi_{\rm tot}}{8\pi^2{v_{\rm in}}^3\eta_{\rm n}},
\label{b12*bis}
\end{equation}
where $v_{\rm in}$ is the velocity of matter flow along the jet axis. It is easy to
check that the strong magnetization condition $\Omega > \Omega_{\rm cr}$ (15) coincides
with the condition \mbox{$\sigma_{\rm n} > 1$}.
Recall that we are mainly interested in strongly magnetized flows, i.e., flows in which
the main energy flux near the central engine is due to the Poynting vector flux
$\Omega_{\rm F}I/2\pi$.. In the opposite case, the flow will be only slightly different
from the hydrodynamic outflow. Using definition (51), this condition can be rewritten as
\begin{equation}
\gamma_{\rm in} \ll \sigma,
\label{mdom}
\end{equation}
where $\gamma_{\rm in}$ is the injection Lorentz factor. As we shall see, the magnetization
parameter $\sigma$ is the key parameter determining the basic features of the flow.
Next, to find the ejected plasma density, it is convenient to introduce the dimensionless
particle production {\it multiplicity} $\lambda$
\begin{equation}
\lambda = \frac{n^{({\rm lab})}}{n_{\rm GJ}},
\label{lam}
\end{equation}
where $n_{\rm GJ}=|\rho_{\rm GJ}|/e$. Such a definition is connected with the fact that,
as we shall show below, both in pulsar magnetospheres and in black hole magnetospheres
the densest is the secondary electron-positron plasma generated either due to conversion
of hard gamma-quanta in the magnetic field or due to gamma-ray collisions with thermal
photons [33, 105]. However, hard gamma-quanta must be emitted in both cases by primary
particles whose density is supposed to be close to the Goldreich density. The convenience
of the particle production multiplicity $\lambda$ also stems from the fact that the
magnetization parameter $\sigma$ can be rewritten with its help in the form
\begin{equation}
\sigma = \frac{e\Omega \Psi_{\rm tot}}{4\lambda m_{\rm e}c^3}
\sim \frac{1}{\lambda}\left(\frac{W_{\rm tot}}{W_{\rm A}}\right)^{1/2},
\label{newsigma}
\end{equation}
where $W_{\rm A} = m_{\rm e}^{2}c^{5}/e^{2} \approx 10^{17}$ erg \, s$^{-1}$.
Thus, the knowledge of two of three quantities $W_{\rm tot}$, $\sigma$, and $\lambda$,
allows the determination of the third one.
Finally, the {\it compactness parameter}
\begin{equation}
l_{a} = \frac{\sigma_{\rm T} L_{\rm tot}}{m_{\rm e}c^3 R}
\label{compact}
\end{equation}
is in fact the optical depth for Thomson cross section $\sigma_{\rm T}$ at a distance $R$ from a
source with the total luminosity $L_{\rm tot}$. Below, it will be important for us
that the parameter $l_{a}$ provide an upper limit of particle energy in the acceleration
region. On the other hand, a large $l_{a}$ is necessary for effective particle production.
\vspace*{.2cm}
{\bf 3.3.2. Singular surfaces.}
Singular surfaces represent the most important structural element of flows. As we shall see,
it is the analysis of the conditions of the smooth passing of a flow through singular surfaces
that allows sometimes rather general relationships to be obtained without solving the
Grad-Shafranov equation itself. Notice from the very beginning that, for simplicity, below
we shall only analyze the case of cold flows. The point here is that, at large distances
thermal effects are insignificant for the polytropic index $\Gamma > 1$ (pressure
$P \propto n^{\Gamma}$). This conclusion can readily be obtained from both the Grad-Shafranov
equation itself and Bernoulli equation. Indeed, from the analysis, for instance, of
nonrelativistic equations (36) and (41) it follows that both the enthalpy
$w = c_{\rm s}^2/(\Gamma - 1) \propto n^{\Gamma -1}$ in Eqn (36) and the temperature
$T \propto n^{\Gamma - 1}$ in Eqn (41) decrease with the distance from the compact source,
since for any divergent outflow the particle number density $n \rightarrow 0$ for
$r \rightarrow \infty$. Therefore, the contribution from a final temperature (enthalpy,
entropy) compared to the total energy $E$ and its derivative ${\rm d}E/{\rm d}\Psi$ can
be neglected at large distances. Thus, it becomes clear why in the analysis of relativistic
flows the final temperature effects (and, in particular, critical conditions on the slow
magnetosonic surface) are usually neglected. On the other hand, the pressure can be significant
for cylindrical flows, i.e., for flows in which the density does not decrease with distance
from compact object [9,106].
The first natural scale that emerges in the theory of relativistic winds is the {\it light
cylinder}
\begin{equation}
R_{\rm L} = \frac{c}{\Omega},
\label{d14}
\end{equation}
i.e., the axial distance at which solid-body rotation together with the central object
becomes impossible. It is easy to show that the light cylinder is the scale where:
\begin{enumerate}
\item
the magnitude of the electric field becomes comparable to that of the poloidal
magnetic field;
\item
toroidal electric currents flowing in the magnetosphere start perturbing the poloidal
manetic field of the central engine;
\item
the magnitude of the toroidal magnetic field produced by the longitudinal
Goldreich current becomes comparable with that of the poloidal magnetic field.
\end{enumerate}
It follows from the first statement above and definitions (21) and (31) that beyond the
light cylinder the electric field becomes stronger than the poloidal magnetic field.
In particular, the poloidal magnetic field will decrease as $r^{-2}$ and the electric
field as $r^{-1}$ for the spit monopole outflow shown in Fig. 9. On the other hand,
freezing-in condition (29) requires that the magnetic field be stronger than the electric
field. This can be possible only when a strong enough longitudinal electric current is
flowing in the magnetosphere, because the toroidal magnetic field also decreases as $r^{-1}$
according to Eqn (21).
We can therefore conclude that the question as to whether or not a smooth relativistic
MHD outflow \mbox{($|{\bf E}| < |{\bf B}|$)} exists beyond the light cylinder
is also directly related to the question of the magnitude of the longitudinal current
circulating in the magnetosphere of a compact object. Then, for currents below some
critical value, a so-called {\it light surface} is bound to appear in the magnetosphere,
on which the electric field matches the magnetic field ($|{\bf E}| = |{\bf B}|$), and
hence the approximation considered here itself becomes invalid. Calculations [107, 108]
showed that the closure of currents occurs near this surface in the region with the
thickness $\delta r \sim R_{\rm L}/\lambda$, and particles are effectively accelerated
there up to energies of $\gamma \sim \sigma$.
If the longitudinal currents are sufficiently high, the smooth MHD outflow can exist
beyond the light cylinder as well. The electric field there will be almost equal to the
magnetic field. Indeed, as directly follows from the relativistic Bernoulli equation
($A.12$), in the limit $\varpi \gg R_{\rm L}$ we obtain simply{\footnote{This expression
corrects Eqn (4.144) from monograph [9].}}
\begin{equation}
{\bf B}^2 - |{\bf E}|^2 = \frac{B_{\varphi}^2}{\gamma^2}.
\label{w10k}
\end{equation}
Since, as can be easily checked, $B_{\rm p} \approx B_{\varphi}/x_r$, where
$x_r = \Omega \varpi/c$, we can always apply the estimate
\begin{equation}
B_{\varphi}^2 - |{\bf E}|^2 \leq
\, \frac{B_{\varphi}^2}{\gamma^2}.
\label{w10k'}
\end{equation}
As a result, the radial drift motion in the crossed electromagnetic fields dominates
in a strongly magnetized relativistic outflow beyond the light cylinder. Indeed, as
can be easily verified, the Lorentz factor entering into Eqn (59) satisfies the
condition $\gamma^{-2} = 1 - U_{\rm dr}^2$, where
\begin{equation}
{\bf U}_{\rm dr} = c \frac{{\bf E} \times {\bf B}}{B^2}.
\label{drdr}
\end{equation}
In other words, the velocity parallel to the magnetic field does not contribute at all
to the value of the Lorentz factor [109].
The {\it fast magnetosonic surface} is another important surface of the magnetized flows.
It is fully equivalent to the sonic surface in the ideal hydrodynamics. Indeed,
Bernoulli equation (27) is well known to have a singularity on the sonic surface.
For example, the logarithmic derivative of density determined from Eqn (27) is
written for a spherically symmetric flow in the form
\begin{equation}
\eta_1 = \frac{r}{\rho}\frac{{\rm d}\rho}{{\rm d}r}=
\frac{{\displaystyle{2v^2-\frac{GM}{r}}}}{c_{\rm s}^2-v^2}=
\frac{{\displaystyle{2-\frac{GM}{rv^2}}}}
{{\displaystyle{-1+\frac{c_{\rm s}^2}{v^2}}}}=\frac{N}{D}.
\label{a18}
\end{equation}
It is obvious that derivative (62) has a singularity when the velocity of matter
equals the speed of sound: \mbox{$v = c_{\rm s} = c_{*}$ ($D=0$).} This means that in
order to cross the sonic surface $r=r_{*}$ smoothly, the additional condition
\begin{equation}
N(r_{*}) = 2 - \frac{GM}{r_{*}c_{*}^2} = 0
\label{a19}
\end{equation}
must be satisfied. As a result, the additional critical condition (63) fixes the
accretion (ejection) rate of matter [14].
The fast magnetosonic surface plays a similar role. But now it determines not the
accretion or ejection rate, but the magnitude of the longitudinal current $I$
(more precisely, the integral $L$). It is this critical condition on this surface
that shows us that in the relativistic case the longitudinal current $I$ near the
central engine must be close to the Goldreich current $I_{\rm GJ}$. On the other
hand, as we have already noted, for a nonrelativistic outflow $i_0 \gg 1$, and the
conditions of the smooth crossing of singular surfaces lead to relations (15)--(17)
used above. Their derivation, however, is rather cumbersome, and we shall omit it
here. It should only be emphasized that they can be obtained directly from an analysis
of Bernoulli equation. The point is that the sonic surface is the $X$-point on
the (distance $r$-velocity $v$) plane. That is, it is the point of crossing
the roots of the algebraic Bernoulli equation. The condition of coincidence of roots
of the algebraic equation puts certain bounds on the coefficients of the equation
itself, which enables the magnitude of the longitudinal current to be estimated.
Notice that expressions for the current formulated above were exactly obtained in
Refs [110,111] for the simplest split monopole geometry shown in Fig. 9.
In a similar way, the following theorem can be proved: {\it In a relativistic outflow near
the outer fast magnetosonic surface, the energy of the particles reaches the values
\begin{eqnarray}
\gamma & = & \left(\frac{E}{\mu\eta c^2}\right)^{1/3} \sim \sigma^{1/3}, \quad
\gamma_{\rm in} \ll \sigma^{1/3}, \\
\gamma & = & \gamma_{\rm in}, \hspace*{3cm} \gamma_{\rm in} \gg \sigma^{1/3}.
\end{eqnarray}
Thus, the fraction of energy carried by particles in the vicinity of the fast magnetosonic
surface is a small fraction ($\sim \sigma^{-2/3}$) of the electromagnetic energy flux for
strongly magnetized outflows $\sigma \gg \gamma_{\rm in}$. The surface itself is located
at the distance of
\begin{eqnarray}
r_{\rm F} & \approx & \left(\frac{E}{\mu\eta c^2}\right)^{1/3}R_{\rm L}
\sim \sigma^{1/3}R_{\rm L}, \hspace*{.3cm} \gamma_{\rm in} \ll \sigma^{1/3},
\label{q05} \\
r_{\rm F} & \approx & \left(\frac{E}{\mu\eta\gamma_{\rm in}c^2}\right)^{1/2}R_{\rm L}
\sim \left(\frac{\sigma}{\gamma_{\rm in}}\right)^{1/2}R_{\rm L}, \label{q03} \\
&&\hspace*{4cm} \gamma_{\rm in} \gg \sigma^{1/3}
\nonumber
\end{eqnarray}
(the first relation holds true not too close to the rotational axis)}.
Interestingly, expression (66) is valid for both relativistic and nonrelativistic
flows since it does not include in fact the speed of light $c$.
Finally, the singularity at $A = 1 - {\cal M}^2 = 0$ that appeared in nonrelativistic
equations (42), (43) suggests that the {\it Alfv\'enic surface} must also play an important
role in the structure of magnetized flows. Its location for nonrelativistic flows
can easily be estimated from the numerator of relation (42):
\begin{equation}
\varpi_{\rm A}^2 = \frac{L_{\rm n}}{\Omega_{\rm F}}.
\label{w18}
\end{equation}
In this case, the Alfv\'enic surface turns out to be located close to the fast
magnetosonic surface.
As to a relativistic flow (and a flat space), the corresponding condition should be written differently:
\begin{equation}
A = 1 -\frac{\Omega_{\rm F}^{2}\varpi^{2}}{c^2}-{\cal M}^{2}.
\label{p39}
\end{equation}
On the other hand, it is easy to check that the parameter $q = W_{\rm part}/W_{\rm em}$
(i.e., the particle-to-electromagnetic energy flux ratio) can be presented in the form
\begin{equation}
q = \frac{{\cal M}^2c^2}{\Omega_{\rm F}^2\varpi^2}.
\label{defg}
\end{equation}
Thus, in the region where the energy flux is Poynting-dominated ($q \ll 1$), the following
condition should be satisfied:
\begin{equation}
{\cal M}^2 \ll \frac{\Omega_{\rm F}^2\varpi^2}{c^2}.
\label{w10c}
\end{equation}
Consequently, the Alfv\'enic surface for such flows is located near the light cylinder.
Hence (except for the polar region where both the Alfv\'enic and fast magnetosonic surfaces are
close to each other), the fast magnetosonic surface for strongly magnetized flows is located
$\sigma^{1/3}$ times further from the central engine as compared with the Alfv\'enic surface.
Let us remember that the Alfv\'enic surface in the relativistic case determines the scale on
which the toroidal magnetic field becomes comparable in magnitude to the poloidal field.
It is easy to check that for rapid rotation $\Omega > \Omega_{\rm cr}$ (15) a similar
situation holds for nonrelativistic flows, too. As regards the electric field, it is
always weaker than the magnetic field in the nonrelativistic case. Notably, that is why
the light surface cannot appear in the nonrelativistic case.
The Alfv\'enic surface represents a higher-order singularity as against the fast magnetosonic
surface. Therefore, relations (42) and (43) do not put any constraint on the integrals of
motion and only determine the location of the Alfv\'enic surface and the magnetic field
structure. In this event, however, particles can intersect the Alfv\'enic surface only in
one direction. For example, when the central source loses its rotational energy, crossing
the Alfv\'enic surface is possible only in the direction outward from the compact object.
When the energy flux is directed toward the central engine (for instance, if it is spun
up by the accreting material), the flow must also be directed toward the central engine.
Certainly, this statement is invalid in the region of an accretion disc where viscosity
cannot be neglected.
This statement can easily be proved in the relativistic case by recalling that the motion
of particles is the sum of the drift motion in the crossed fields and the motion along the
magnetic field. The condition $v < c$ that limits the longitudinal velocity puts bounds on
the radial velocity of matter. This is related to the fact that the drift velocity itself
becomes close to the speed of light on the Alfv\'enic surface. However, accretion of matter
with positive energy release cannot be realized in the nonrelativistic case, either. In this
event, the interaction of the supersonic accretion flow with the rotating magnetized central
body would take place. A shock wave is known to be formed in such an interaction [112].
\begin{figure}
\begin{center}
\includegraphics[width=0.9\columnwidth]{fig2_04.ps}
\end{center}
\caption{The location of the Alfv\'enic ($A$) and fast magnetosonic ($F$) surfaces near the
black hole horizon. The dashed line shows the Alfv\'enic surface in the force-free
approximation (i.e., the 'light cylinder'), while the dotted line indicates the
ergosphere surface. Here, $\alpha$ is the gravitational redshift (see the Appendix).
}
\label{fig2_04}
\end{figure}
In conclusion, we should comment on the features of a black hole magnetosphere. As seen
from exact expressions for the Alfv\'enic $A$ ($A.13$) and sonic $D$ ($A.20$) factors presented
in the Appendix, a second family of singular surfaces inevitably emerges near the black hole
horizon. Here, as shown in Fig. 10, the infalling matter, as in the case of the outflow, must
first cross the Alfv\'enic surface and only then intersect the fast magnetosonic surface (recall
that thermal effects are not discussed here). This is related to the fact that the strong
gravitational field of the black hole forces the matter to approach the event horizon.
The appearance of the second family of singular surfaces leads to new important properties.
First of all, the matter can intersect the inner Alfv\'enic surface only in the direction
towards the black hole horizon. But this means that if the central engine loses rotational
energy (and hence particles can cross the outer Alfv\'enic surface only in the direction away
from the compact object), the plasma is bound to be generated in the magnetic field lines
'anchored' to the black hole horizon. Only in this case can electric currents appear in the
black hole magnetosphere, which are necessary, as we have seen, to explain the observed
energy release. In turn, the appearance of one more critical condition on the inner fast
magnetosonic surface proved to be sufficient to determine the angular velocity $\Omega_{\rm F}$.
Here, $\Omega_{\rm F}$ indeed must be close to $\Omega_{\rm H}/2$
[$\Omega_{\rm H} = \omega(r_{\rm g})$ is the angular velocity of the black hole;
see the Appendix], as was understood as early as the Blandford and Znajek paper [33].
This problem has been possible to solve exactly for slow rotation in the split monopole
magnetic field [113].
\vspace*{.2cm}
{\bf 3.3.3. The problem setting.}
Before considering the main results which were obtained using the analytical theory, it is
necessary to discuss the formulation of the problem. The point is that we shall primarily
be interested in transonic flows, i.e., those which are subsonic near the compact object and
supersonic in the wind region. Indeed, as we shall see, the distance from a central
engine to singular surfaces in all compact objects are much smaller than even the transverse
size of the collimated outflows. The difficulty here is that the direct problem setup itself
in the framework of the Grad-Shafranov equation method turns out to be nontrivial. For example,
the second-order equation describing the two-dimensional flow in the hydrodynamic limit, when
only three integrals of motion are available, requires four boundary conditions to be imposed
in the transonic regime. The fifth condition is the critical condition on the sonic surface.
This means that on some surface, for example, two thermodynamic functions and two velocity
components must be specified. We stress that here only flows depending on two variables are
considered. As discussed in detail in monograph [9], the well-known spherically symmetric
flows (the Bondi accretion, the Parker ejection) are degenerate, since the structure of the
flow itself is specified in them. In the general case of spherical accretion, a nonstationary
solution with a shock wave appears (see, for example, Ref. [114]).
However, to determine Bernoulli integral, which is naturally needed in solving the
equilibrium equation, we should specify all three components of the velocity, which is
impossible since the third velocity component itself should be found from the solution.
In the general case, one should set
\begin{equation}
b = 2 + i - s'
\label{a35}
\end{equation}
boundary conditions, where $i$ is the number of invariants, and $s'$ is the number of
singular surfaces (and for the magnetic field lines threading the black hole horizon
this number is doubled, i.e., separately for ejecting and accreting plasma). Such an
internal inconsistency of this approach in the general case does not allow us to solve
direct problems, namely, to determine the structure of the flow in some region using the
given physical parameters on its boundary. Therefore, it is not astonishing that most
researchers primarily interested in astrophysical applications already in the middle
of the 1990s started addressing a totally different class of equations, namely those
covering time relaxation problems, which can only be solved numerically [79, 115-119].
However, only in the last several years has significant progress here been achieved
[103, 109, 120-127], which, among other things, has confirmed many analytical results
obtained before.
It should be noted that this problem does not appear in both subsonic and supersonic
cases. For these flows, all necessary integrals of motion must be determined from the
boundary conditions. In particular, the boundary conditions will determine in the subsonic
case the longitudinal current $I$, too. This exactly corresponds to the
unipolar inductor model, where the current (and hence the energy losses) is determined
by the external load. Unfortunately, this ideology has also spread into the theory of
magnetized winds. This is related to the fact that many results were obtained in the
1970s-1980s using the force-free approximation [i.e., when $\sigma \rightarrow \infty$ and
masses of particles can be neglected]. In this approximation, the Grad-Shafranov equation
becomes elliptical and, hence, the flow structure must depend on conditions at the external
boundary. But the theory of pulsar magnetospheres so far has been constructed
in the force-free approximation. The class of subsonic flows also includes the so-called
'magnetic tower' [128, 129]. As this question is highly important, we shall discuss
it below in more detail.
Correspondingly, the Blandford-Znajek model was also constructed in the force-free
approximation, which required the boundary condition to be set on the black hole horizon.
Since the electromagnetic wave (like other material bodies) can propagate near the horizon
only normal to the horizon toward the black hole, the boundary condition in fact
is equivalent to the Leontovich boundary condition in radio physics [130]. However, it
is usually obtained by requiring the finiteness of fields in the freely falling
observer's frame of reference, which yields $B_{\varphi}(r_{\rm g}) = -E_{\theta}(r_{\rm g})$.
This condition, as is well known, can be rewritten in the form of the Ohm law for the
formally introduced 'surface current' [131]
\begin{equation}
{\bf J}(r_{\rm g}) = \frac{c}{4\pi}{\bf E}(r_{\rm g}),
\label{k50}
\end{equation}
which corresponds to the universal 'internal' resistance of the battery,
${\cal R} = 4\pi/c =$ 377 $\Omega$. In another form, this boundary condition
can be rewritten as
\begin{equation}
4 \pi I(\Psi)
=[\Omega_{\rm H}-\Omega_{\rm F}(\Psi)]\sin\theta
\left(\frac{{\rm d}\Psi}{{\rm d}\theta}\right).
\label{k48}
\end{equation}
Here, we have utilized definitions ($A.7$) and ($A.8$) and, for simplicity, written the
equality $B_{\varphi}(r_{\rm g}) = -E_{\theta}(r_{\rm g})$ for slow rotation. Thus, it
is not surprising that in the framework of this approximation the mechanism of energy
loss by a black hole was connected, in analogy with the unipolar inductor, with the Ampere
force acting on the black hole horizon from the side of the surface current [131].
Only much later was it understood that the force-free approximation gives inaccurate and
sometimes erroneous results. This is related to the fact that in the force-free
approximation, i.e., when particles are assumed to be massless, the flow always remains
subsonic. Under this assumption, the fast magnetosonic surface on which the poloidal
velocity of particles matches the fast magnetosonic wave velocity goes formally to
infinity (and the inner surface, oppositely, coincides with the black hole horizon).
But we have seen that it is the conditions on the fast magnetosonic surface that fix
the longitudinal current circulating in the magnetosphere. In addition, the very
necessity of establishing the boundary condition on the black hole horizon, i.e.,
in the causally disconnected region, shows that the physical interpretation given
above does not relate to the reality [132].
\begin{figure}
\begin{center}
\includegraphics[width=0.6\columnwidth]{fig2_05.ps}
\end{center}
\caption{The structure of a magnetic field behind a switching-on wave propagating with velocity
$c$ from a compact object [120]. Inside the switching-on wave, the flow rapidly becomes
stationary, in correspondence with the analytical solution.
}
\label{fig2_05}
\end{figure}
To make this point more clear and, in particular, to understand the boundedness of stationary
solutions (and hence the analytical method itself) in studies of the current closure, it is
useful to consider the results obtained in paper [120] in which the problem was formulated
as follows. There is a magnetized ball at rest which at the moment $t = 0$ starts
rotating with angular velocity $\Omega$. As a result, the switching-on wave starts propagating
from the ball with velocity $c$, so that the magnetic field remains unperturbed beyond it and
electric currents are absent, while inside the switching-on wave (and this is a very important
result) the solution rapidly approaches the stationary transonic regime, which is in full
agreement with the analytical solution. Thus, the assumption of the stationary solution for
longitudal currents flowing actually along magnetic surfaces is confirmed.
As for the current closure, no current closure as such happens at all in the ideal case
where the outflow occurs in a vacuum. This is related to the fact that in the switching-on
wave the flow is time-dependent (Fig. 11), so there ${\rm div} {\bf j} \neq 0$
(S S Komissarov, private communication). In reality, the current closure will take place on
a shock wave which must necessarily emerge in the region where the supersonic switching-on
wave collides with the surrounding medium. In any case, however, the ambient medium for
transonic flows cannot influence the magnitude of the longitudinal current for $r < r_{\rm F}$ and,
hence, affect the central engine energy release. As soon as the switching-on wave crosses
the singular surfaces (they are also shown in the left lower corner of Fig. 11), the
longitudinal current flowing in the magnetosphere stops depending on time. For this reason,
one can indeed consider in the framework of the stationary approximation that the electric
current closure occurs at infinity, as is usually assumed. Thus, transonic flows are
significantly different from subsonic ones, when the electric current circulating in the
magnetosphere is determined by the conductivity of the boundary of the region occupied by plasma
(see, for example, Ref. [133]).
Accounting for the nonzero mass of particles allows us to clarify the situation with the
'boundary condition on the horizon', and thus with the mechanism of energy release by a
black hole. Indeed, as shown in Fig. 10, the fast magnetosonic surface for nonzero masses
of particles is located above the black hole horizon. Therefore, the critical condition
should also be set here, which is definitely located in the region casually connected to
the outer space. The black hole horizon will be located in the region of the supersonic
flow and, hence, cannot affect the properties of the flow. As a result, the additional
critical condition must also be kept in the force-free limit $m_{\rm p} \rightarrow 0$, when
the fast magnetosonic surface, as noted above, formally coincides with the event horizon.
It is not then surprising that this limit of the critical condition on the fast magnetosonic
surface exactly coincides with condition (74) [10]. Thus, the boundary condition on the
horizon (74), which was necessarily used in the force-free approximation, represents a relict
of the critical condition on the fast magnetosonic surface.
Correspondingly, it also becomes clear how to interpret the infinite time retardation near
the black hole horizon. Indeed, the time it takes for the plasma to reach the horizon must
be infinite from the point of view of the remote observer. As the time of existence of the
black hole (the surrounding plasma, etc.) is finite, the remote observer will register a
'switching-on wave' corresponding to the very initial stages of the existence of the central
engine over the black hole surface. This was, in fact, one more argument in support of the
fact that it is incorrect to set any boundary condition on the black hole horizon [132].
However, as seen by the example of the outflow (see Fig. 11), to form the current system
that fully determines the central engine power it is sufficient to wait a finite time until
the plasma intersects the fast magnetosonic surface. This also implies, inter alia, that
after a finite time the switching-on wave will turn out to be in the supersonic region of
the flow and, hence, cannot affect the magnetosphere structure. This conclusion was
numerically confirmed by Komissarov [134].
If this is indeed the case, however, another source of EMF should be found, since the black
hole itself now can no longer be the source of extraneous forces and, consequently, serve as
a battery. It turned out that the appearance of EMF in the black hole magnetosphere is related
to the Lense-Thirring (the frame-dragging) effect, which appears due to rotation of the
black hole. Indeed, according to general relativity, the space itself in the vicinity of a
Kerr black hole starts rotating with angular velocity $\omega$ ($A.4$). Only in the reference
frame rotating with this angular velocity will an observer not register a precession of
gyroscopes. On the contrary, noninertial forces will appear in the laboratory frame at rest
with respect to remote observers and they can be detected. It is the frame-dragging effect
that leads to the appearance of the electromotive force in the black hole magnetosphere.
\begin{figure}
\begin{center}
\includegraphics[width=\columnwidth]{fig2_06.ps}
\end{center}
\caption{The appearance of the electromotive force in a circuit at rest
relative to a rotating black hole immersed in an external magnetic field.
The electric field directions are shown for observers at rest.
}
\label{fig2_06}
\end{figure}
Indeed, as in the case of any body moving in the magnetic field, the 'motion of space' relative
to the observer at rest will produce the electric field ${\bf E} = -{\bf V} \times {\bf B}/c$,
where now ${\bf V} = {\bf \omega} \times {\bf r}$ is the velocity of the body at rest
relative to the preferred reference frame. Here, it is important that the angular velocity
$\omega$ be different at various distances from the black hole. Therefore,
as shown in Fig. 12, the circulation of the electric field in a circuit will be nonzero
even if the electric circuit is at rest relative to the black hole, i.e., when the magnetic
flux through the contour remains constant (of course, this state is possible only above the
ergosphere). It is the motion of space through the circuit that generates the electromotive
force. We see here that the 'electric battery' will be located above the black hole horizon,
and possibly outside the ergosphere.
Finally, as already stressed, for positive energy flux from the central engine the electric
currents in the black hole magnetosphere can flow only in the case where the plasma generation
mechanism is operating above the horizon. Electron-positron pair production in the collision
of hard gamma-ray quanta emitted from the accretion disc surface was already discussed above.
Here, we should note that, as in the Penrose effect, one particle should fall into the black
hole, and another particle should escape to infinity. It should be recalled that the Penrose
effect has its origins in the remarkable property of rotating black holes --- the relativistic
mass defect can exceed $100\%$ inside the ergosphere
\mbox{$r_{\rm g} < r < r_{\rm e} = M + \sqrt{M^2 - a^2 \cos^2\theta}$}
($G = c = 1$) and, hence, above all the horizon surface [14]. Therefore, it becomes evident
that the Blandford-Znajek effect is in fact the electromagnetic realization of the Penrose
process. The difference is that it concerns not charged particles themselves but the
electromagnetic field they induce. In other words, the spin-down of the black hole is not
due to electric currents flowing over the horizon but due to the negative
electromagnetic energy flux falling onto the black hol. It is this role of
the ergosphere inside which the relativistic energy of any material bodies (including
the electromagnetic field) can become negative that is great. Such interpretation seems
now to be the most likely, and most researchers involved in these studies tend to accept
this interpretation (see, for example, book [135]).
Let us summarize. We have shown that the critical conditions on the fast magnetosonic
surface mostly determine the energy release of the central engine. This surface serves
as a valve that determines the magnitude of the longitudinal current circulating in the
magnetosphere. Like the usual sonic surface in hydrodynamics, it separates subsonic and
supersonic parts of the flow. As a result, the longitudinal current for transonic flows
must be determined not by the external conditions but by the condition of smoothly crossing
the singular surfaces. Even more complicated is the situation in a black hole magnetosphere,
where both the current and the angular velocity $\Omega_{\rm F}$ should be deduced from the
critical conditions on the singular surfaces. In consequence of this, it is these critical
conditions on the singular surfaces that will determine the central engine power.
\section{Theoretical predictions}
\subsection{Collimation}
{\bf 4.1.1 The force balance across the flow.}
Let now analyze the results of the analytical theory. First of all, we should try to
formulate some general properties of a magnetized outflow, which must show their worth
at large distances from the central object. As already noted, we are primarily interested
in transonic flows, in which the flow at large distances is supersonic. Moreover, we shall
unconditionally assume that the solution can be continued to infinity. This is possible,
as we have seen, only if the longitudinal electric current is sufficiently large. As we have
already stressed, thermal effects can almost always be neglected. As a consequence, Bernoulli
equation becomes a fourth-order algebraic equation with respect to ${\cal M}^2$, which in
many cases allows us to write out rather simple analytical asymptotic solutions.
For simplicity, let us start from the nonrelativistic case. By analyzing the leading
terms in the Grad-Shafranov equation (41), it is possible to show that the force balance
equation can be written in the form [136]
\begin{equation}
\frac{\rho v_{\parallel}^2}{R_{\rm c}}
= \frac{1}{c}j_{\parallel}B_{\varphi},
\label{t-f-n}
\end{equation}
where $R_{\rm c}$ is the radius of the magnetic field line in the poloidal plane.
In other words, the equilibrium must be established due to the balance between the
centrifugal force $\rho v_{\parallel}^2/R_{\rm c}$ and the Ampere force related to
the longitudinal electric current $j_{\parallel}$. It is evident that the Ampere
force, depending on the sign of the longitudinal current, can both collimate and
decollimate the flow. In this case, the collimation must occur close to the jet axis,
while the decollimation occurs at its periphery. Exactly this behavior was obtained
both analytically [111] and numerically [120] (Fig. 13). It should be stressed that
here we do not consider how strongly the magnetic surfaces can be bent and only
investigate their form.
\begin{figure}
\begin{center}
\includegraphics[width=0.8\columnwidth]{fig3_01.ps}
\end{center}
\caption{The structure of magnetic surfaces for a nonrelativistic plasma outflow
in the split monopole magnetic field [120]. The decollimation in the region of the
reverse bulk current near the central engine is clearly seen. However, the
redistribution of longitudinal currents occurs at far distances, so that the
current flowing out is concentrated near the rotational axis, and the reverse
current flows near the equatorial plane.
}
\label{fig3_01}
\end{figure}
A similar picture, however, cannot be realized up to very large distances from the
central engine. 'Natura abhorret vacuum', and the diverging magnetic surfaces inevitably
must start collimating. It turned out that this is possible exactly because the
longitudinal current $I$ is not the integral of motion, and hence the electric
current, unlike particles, can intersect the magnetic surfaces. Indeed, it is easy
to check that, for an almost radial flow at large distances, the right-hand side of
equation (75) should decrease as $r^{-3}$, while the numerator on the left-hand side
decreases as $r^{-2}$. As a result, for equality (75) to hold, the radius of curvature
of the magnetic surfaces, $R_{\rm c}$, should increase as $r$. But such a behavior cannot be
realized at mathematical infinity [137], so for $r \rightarrow \infty$ only the right-hand
side of equation (75) is leading.
In the long run, we come at first glance across the paradoxical result that the current
density in the magneto-sphere must vanish at large distances [138]:
\begin{equation}
j_{\parallel} = 0.
\label{asy1}
\end{equation}
In fact, this simply means that at large distances almost all outflowing longitudinal
current must be concentrated near the rotational axis. Indeed, as shown in Ref. [139],
a cylindrical region containing almost all outflowing current must inevitably appear
near the axis.{\footnote{Strictly speaking, this terminology corresponds to the case
of ${\bf \Omega}{\bf B} < 0$, where $\rho_{\rm GJ} > 0$. For the opposite orientation,
the current near the rotational axis will flow towards the central engine.}}
This behavior was later verified by numerical simulations [120]. We shall consider this
point in more detail in Section 4.1.3.
In the relativistic case, the electric force $\rho_{\rm e}{\bf E}$ and the component of the
Ampere force related to the longitudinal current ${\bf j}_{\parallel} \times {\bf B_{\varphi}}$
will mostly act on the outflowing plasma. However, as follows from equation (60), these forces
almost mutually balance each other. Therefore, in addition to the bulk force{\footnote
{The corresponding formula (4.227) from monograph [9] has the incorrect sign}}
\begin{equation}
{\cal F}_{\rm jpol} \approx \rho_{\rm e}{\bf E} -
\nabla \left(\frac{B_{\varphi}^2}{8\pi}\right),
\label{clF2'}
\end{equation}
it is necessary to take into account the bulk centrifugal force
\begin{equation}
{\cal F}_{\rm cent} \approx \frac{nmc^2\gamma+S/c}{R_{\rm c}}
\label{clF1}
\end{equation}
($S \approx c B_{\varphi}^2/4\pi$ is the Poynting vector) and the Ampere force
${\cal F}_{\rm jtor} \approx {\bf j}_{\varphi} \times {\bf B_{\rm p}}/c$ pertaining
to toroidal current. As a consequence, the Grad-Shafranov equation in the limit
$r \gg r_{\rm F}$ can be conveniently rewritten in the form [136, 138, 140]
\begin{eqnarray}
\frac{B_{\varphi}^2 + 4 \pi n m_{\rm p} c^2\gamma}{R_{\rm c}}
+ \frac{1}{2} \, \hat{\bf n} \cdot \nabla ({\bf B}_{\rm p}^2) \nonumber \\
+ \frac{1}{2} \, \hat{\bf n} \cdot \nabla (B_{\varphi}^2 - {\bf E}^2)
- \frac{B_{\varphi}^2 - {\bf E}^2}{\varpi} \, (\hat{\bf n} \cdot {\bf e}_{\varpi}) = 0,
\label{balance}
\end{eqnarray}
where $\hat{\bf n} = \nabla \Psi/|\nabla \Psi|$. Notice that the Poynting vector contributes,
in addition to the contribution from particles, to the centrifugal force. This is related to
the fact that, as noted above, both particles and the electromagnetic energy propagate along
the magnetic surfaces.
\vspace*{0.2cm}
{\bf 4.1.2 The collimation mechanism.}
Thus, the form of the magnetic surfaces close to the rotational axis depends on the balance
between the collimating Ampere force arising from the longitudinal electric current (parallel
currents are attracted) and the decollimating Ampere force related to toroidal currents. Thus,
the question as to whether the collimation will be effective depends on the magnitude of the
longitudinal current. A series of exact solutions [110, 111], which were possible to obtain
by analyzing small deviations from the monopole magnetic field, showed that for nonrelativistic
jets the collimation is large even close to the fast magnetosonic surface. This property is also
confirmed by numerical modeling [120, 141] (see also Fig. 13).
In the relativistic case, where, as we remember, the longitudinal current is close to the
Goldreich current ($i_0 \approx 1$), an almost full compensation of these two forces takes
place. In particular, the balance is met exactly in the force-free approximation and in the
split monopole magnetic field [142], so the vacuum monopole solution remains exact up to
infinity for the magnetosphere filled with the plasma as well (Fig. 14). The current $I$ here
takes the form $I(\theta)=I_{\rm M}^{({\rm A})}\sin^2\theta$, where
\begin{equation}
I_{\rm M}^{({\rm A})} = \frac{\Omega_{\rm F}\Psi_0}{4\pi},
\label{IMA}
\end{equation}
which exactly coincides with the Goldreich current \mbox{($j_{\parallel} = \rho_{\rm GJ} c$).}
\begin{figure}
\begin{center}
\includegraphics[width=0.65\columnwidth]{fig3_02.ps}
\end{center}
\caption{The force-free monopole solution found by Michel [142], in which the electric
field $E_{\theta}$ exactly equals the toroidal magnetic field $B_{\varphi}$. The contour
arrows show the direction of the poloidal current.
}
\label{fig3_02}
\end{figure}
For massive particles (and again for the split monopole magnetic field), the longitudinal
current determined from the condition of smoothly crossing the fast magnetosonic surface
will differ from the Goldreich current only by a factor of the order $\sigma^{-4/3}$ [143].
As a result, the perturbation of the magnetic flux function $\delta\Psi/\Psi$ in the
asymptotically remote region $r \gg r_{\rm F}$ will increase logarithmically slowly [143,144]:
\begin{equation}
\frac{\delta\Psi}{\Psi} \sim \sigma^{-2/3}
\ln^{1/3}\left(\frac{r}{r_{\rm F}}\right).
\label{l119'}
\end{equation}
In other words, the current turns out to be only insignificantly larger than the critical one,
which leads to a vanishingly small collimation. Correspondingly, the particle energy also
increases very slowly:
\begin{equation}
\gamma \approx \sigma^{1/3} \ln^{1/3}\left(\frac{r}{r_{\rm F}}\right).
\label{l119g}
\end{equation}
\begin{figure}
\begin{center}
\includegraphics[width=0.65\columnwidth]{fig3_03.ps}
\end{center}
\caption{The location of the Alfv\'enic (A) and fast magnetosonic (F) surfaces in a parabolic
magnetic field. Such a field can also be generated in the accretion disc [145].
}
\label{fig3_03}
\end{figure}
We stress that above we have considered the proper collimation, i.e., that due to bulk
currents. However, the collimation, generally speaking, can be produced in the source itself.
In Fig. 15, the flow obtained as a small perturbation of the force-free solution is shown,
but for a parabolic field [145]. Such a field may be generated in the accretion disc as well
[5]. For not too small ($\theta \gg \gamma_{\rm in}^2/\sigma$) and not too large
($\theta \ll \sigma^{-1/3}$) angles, the location of the fast magnetosonic surface $r_{\rm F}$
and the value of $\gamma(r_{\rm F})$ are given by the expressions
\begin{eqnarray}
r_{\rm F} & \approx & \left(\frac{\sigma}{\theta}\right)^{1/2} R_{\rm L},
\label{r-f-par} \\
\gamma(r_{\rm F}) & \approx & \sigma^{1/2}\theta^{1/2},
\end{eqnarray}
respectively, which again correspond to the estimates of the particle energy,
$\gamma \sim \sigma^{1/3}$, and the axial distance,
$r_{\rm F}\sin\theta \sim \sigma^{1/3}R_{\rm L}$ obtained above (see Ref. [145]
for more detail). But in this case the account for nonzero particle masses also only
slightly perturbs the force-free solution. On the fast magnetosonic surface,
for example, one finds
\begin{equation}
\left(\frac{\delta\Psi}{\Psi}\right)_{r = r_{\rm F}} \approx \frac{1}{\sigma\theta} \ll 1.
\label{lena2}
\end{equation}
Thus, we can conclude that the proper collimation is impossible in the relativistic case.
That is, the bulk collimation due to longitudinal currents flowing in the magnetosphere is
impossible. Therefore, the magnetic surfaces can be collimated either by a special choice
of currents in the source itself, or due to the interaction of the supersonic wind with the
surrounding medium. Clearly, a sufficiently extended dense disc is required in the first case.
Because of this, for radio pulsars where there is definitely no such disc, one usually assumes
that the flow at distances $r \gg R_{\rm L}$ must be radial.
Unfortunately, it is not clear at present which mechanism is actually responsible for the
collimation of relativistic jets. However, if one assumes that the collimation is indeed
caused by the presence of the surrounding medium, it becomes possible to estimate the transverse
size $r_{\rm jet}$ of the jets. Indeed, by assuming that the pressure of the poloidal magnetic
field in the jet is close to the ambient pressure $P_{\rm ext}$ (this estimate is valid by
assuming the total electric current in the jet to be zero), we obtain from the condition of
the magnetic flux conservation that
\begin{equation}
r_{\rm jet} \sim R\left(\frac{B_{\rm in}^2}{8\pi P_{\rm ext}}\right)^{1/4},
\label{2}
\end{equation}
where $R$ and $B_{\rm in}$ are the radius and the magnetic field of the compact object,
respectively. For example, for active galactic nuclei ($B_{\rm in} \sim 10^{4}$ G,
\mbox{$R \sim 10^{13}$ cm)} we have
\begin{equation}
r_{\rm j} \sim 1 \, {\rm pc},
\label{3}
\end{equation}
which exactly corresponds to the observed transverse sizes of jets [12]. Correspondingly, for
young stellar objects ($B_{\rm in} \sim 10^{3}$ G, $R \sim 10^{11}$ cm) we have
$r_{\rm j} \sim 10^{16}$ cm, which again is in agreement with observations. Thus,
the external medium apparently must significantly affect the collimation of jets.
\vspace*{0.2cm}
{\bf 4.1.3 The dense core.}
As we have seen, near the axis the bulk force ${\cal F}_{\rm jpol}$ related to the
poloidal current is always directed toward the rotational axis, and in the region
of the current closure, away from the axis. Then, for example, close to the cylindrical
core where the curvature of the magnetic surfaces is small and, hence, the centrifugal
force ${\cal F}_{\rm cent}$ (78) can be neglected, the equilibrium can be established
only if the force ${\cal F}_{\rm jtor}$ is directed away from the rotational axis. But
the poloidal magnetic field in this case must decrease with axial distance. Then, the density
of the outflowing plasma will decrease along with the poloidal field as well. Exactly such a
behavior of the solution was illustrated in Fig. 13.
It turned out that the dense core for the nonrelativistic supersonic flow will exist
for both strongly and weakly magnetized flows close to the central engine,
irrespective of the ambient pressure [146]. The radius of the core will correspond
to such a distance from the axis, at which the toroidal field matches the poloidal one.
A straightforward calculation shows that in both cases the longitudinal current in the
core region will be $j = (c/v_{\rm in})j_{\rm GJ}$, i.e., $i_0 = c/v_{\rm in}$, and so
one arrives at
\begin{equation}
r_{\rm core} = \frac{v_{\rm in}}{\Omega}.
\label{core}
\end{equation}
The magnetic flux in such a core must be a significant fraction of the total magnetic flux.
Such a configuration is none other than the $z$-pinch well-known in plasma physics [147]
(the question of stability will be briefly discussed below).
As for the relativistic flow, the appearance of the dense core can be balanced here by the
electric force $\rho_{\rm e}{\bf E}$ which, as we have seen, significantly weakens the
force ${\cal F}_{\rm jpol}$. As a result, the answer is significantly dependent on the
ambient pressure [9]. If the relativistic jet is surrounded by a medium with the total
gas and magnetic pressure $P_{\rm ext}$ above some limiting value $P_{\rm min}$, the dense
core does not form at all. In this case, the poloidal magnetic field in the jet will be
approximately constant: $B_{\rm p}^2/8\pi \approx P_{\rm ext}$. For convenience sake,
we shall express below the limiting value $P_{\rm ext}$ through the equivalent magnetic field
($B_{\rm min}^2/8\pi = P_{\rm min}$). Here, one obtains the relationship
\begin{equation}
B_{\rm min} = \frac{1}{\sigma\gamma_{\rm in}}B(R_{\rm L}),
\label{38}
\end{equation}
where $B(R_{\rm L}) = \Omega^2\Psi_{\rm tot}/\pi c^2$ is the characteristic magnetic field
on the light cylinder. Correspondingly, the density of the outflowing plasma will be constant,
too.
If the ambient pressure is sufficiently low, so that $P_{\rm ext} < P_{\rm min}$, then, as in the
nonrelativistic case, the dense core is formed in the center of the outflow; the radius of
the core must exceed that of the light cylinder:
\begin{equation}
r_{\rm core} = \gamma_{\rm in}R_{\rm L}.
\label{rcore}
\end{equation}
It is easy to check that on the core boundary the energy flux density of the electromagnetic
field matches the particle energy density. The appearance of such a cylindrical jet was
predicted already many years ago in many papers [138, 148, 149], but the magnetic flux in this
core was determined only quite recently [145,146]. The magnetic field near the axis was found
to only very weakly (logarithmically) depend on the ambient pressure, so with good accuracy
one can set $B_{\rm core} = B_{\rm min}$. As a result, the magnetic flux
$\Psi_{\rm core} = \pi r_{\rm core}^2B_{\rm min}$ within the core turns out
to be much smaller than the total flux:
\begin{equation}
\frac{\Psi_{\rm core}}{\Psi_{\rm tot}} \approx \frac{\gamma_{\rm in}}{\sigma} \ll 1.
\label{39new}
\end{equation}
Here, the poloidal magnetic field and the density of matter for $x_r = \varpi/r_{\rm core} > 1$
must behave as power functions
\begin{equation}
B_{\rm p} \propto x_r^{-k_1},\
\label{Bzpower}
\end{equation}
\begin{equation}
\rho^{({\rm lab})} \propto x_r^{-k_2}.
\label{npower}
\end{equation}
respectively. As the ambient pressure decreases, the exponents $k_1$ and $k_2$
gradually increase; however, their difference remains approximately constant, viz.
\begin{equation}
k_1 - k_2 \approx 0.
\end{equation}
As a result, if we have $k_1 \approx k_2 \approx 0$ for $P_{\rm ext} > P_{\rm min}$
(i.e., the poloidal magnetic field and the outflowing plasma density are constant in the
cross section), for ambient pressures corresponding to the magnetic field
$B_{\rm ext} \equiv B_{\rm eq}$, where
\begin{equation}
B_{\rm eq} = \frac{1}{\sigma^2} \, B(R_{\rm L}),
\label{exep1}
\end{equation}
we have, in contrast, $k_1 \approx k_2 \approx 1$. Under such ambient pressures, as we
shall see, the contribution from particles to the energy flux becomes dominant in the
entire volume of the jet.
\begin{figure}
\begin{center}
\includegraphics[width=0.7\columnwidth]{fig3_04.ps}
\end{center}
\caption{Longitudinal magnetic field $B_z$ (G) as a function of the axial
distance $x=\Omega \varpi/c$ for $\sigma=10^3$ and different Mach numbers
${\cal M}$ on the jet axis, which was obtained as a solution of the one-dimensional
problem [146]. The axial magnetic field only slightly deviates from $B_{\rm min}$
when the ambient pressure changes by several orders of magnitude.
}
\label{fig3_04}
\end{figure}
\begin{figure}
\begin{center}
\includegraphics[width=0.8\columnwidth]{fig3_05.ps}
\end{center}
\caption{Longitudinal magnetic field profile as a function of the axial
distance $\xi \propto \varpi$ for different values of the ambient pressure,
which was obtained as a solution of the two-dimensional time relaxation problem
[126]. The power-law dependence corresponds to the analytical estimate (92).
}
\label{fig3_05}
\end{figure}
In Fig. 16, the behavior of the poloidal magnetic field is shown in the double-logarithmic
scale as a function of the distance to the axis [146] (see also Ref. [150]). It was
obtained as the solution to the system of two ordinary differential equations ($A.14$) and
($A.27$) to which, as shown in the Appendix, the Grad-Shafranov equation can be
conveniently reduced in the cylindrical case. The Mach number on the rotational axis served
as the parameter for different curves. It is evident that the poloidal magnetic field at
distances exceeding $r_{\rm core}$ indeed starts decreasing as a power law. At the boundary
of the jet it changes by several orders of magnitude, while the magnetic field on the axis
changes only several-fold times. This behavior was recently confirmed by numerical two-dimensional
calculations, too [126] (Fig. 17). It is important here that in the last case the very
formulation of the problem was principally different (the time relaxation problem was solved
rather than a stationary problem).
In relativistic jets there can be one more regime which is impossible in nonrelativistic
jets [151]. In Fig. 15, it corresponds to cross sections located sufficiently close to the
equator, when the flow in the jet center must be subsonic. By rewriting the condition
$z < (\sigma/\gamma_{\rm in})R_{\rm L}$ in terms of an ambient pressure, we obtain the
inequality $P_{\rm ext} > B_{\rm cr}^2/8\pi$, where
\begin{equation}
B_{\rm cr} = \frac{\gamma_{\rm in}}{\sigma}B(R_{\rm L}).
\label{55a}
\end{equation}
In this case, a subsonic flow region must inevitably be formed in the inner parts
of the jet with $\varpi < r_{\rm s}$, where
\begin{equation}
r_{\rm s} \approx
\sigma\left[\frac{8 \pi P_{\rm ext}}{B(R_{\rm L})^2}\right]^{1/2}R_{\rm L}.
\label{56a}
\end{equation}
At last, the subsonic flow is established in the entire jet volume for higher
ambient pressures corresponding to the magnetic field pressure on the fast
magnetosonic surface.
\vspace*{0.2cm}
{\bf 4.1.4 The current closure.} As we have understood, according to the most
wide-spread point of view, it is the electric current flowing along collimated
outflows, which is responsible for the main jet energy release. Here, the following
theorem can be formulated.
{\it A stationary cylindrical jet with finite magnetic flux
$\Psi_{\rm tot}$ can be formed either with the nonzero full electric current
$I(\Psi_{\rm tot}) \neq 0$ or in the presence of a surrounding medium with nonzero
pressure.}
At first glance, these two variants fully contradict each other. However, this is
not actually the case. As we have seen, a cylindrical core (which contains a significant
part of the outflowing current) must be formed near the jet axis both in the relativistic
and nonrelativistic cases. If the compact object were solitary in the Universe, the reverse
current would indeed return near the equatorial plane. The closing of the current itself
would occur in the switching-on wave. Thus, if one ignores the current closure region and
studies only the structure of the central region, it is indeed possible to assume that the
current closure takes place at infinity.
In fact, when a jet is immersed into a medium with a finite pressure, the reverse current,
as one usually assumes, must flow along the boundary of the cocoon that forms due to the
interaction of the supersonic wind with the external medium. The boundaries of the
corresponding cocoons can be well seen in Figs 1 and 4. It is the force balance at the
cocoon boundary that allowed the magnetic field $B_{\rm min}$ close to the jet axis to be
determined [145].
\vspace*{0.2cm}
{\bf 4.1.5 The stability.} In conclusion, it is necessary to briefly consider the problem
of the stability of jets. The nonrelativistic z-pinches observed in the laboratory are known
to be strongly unstable with respect to constrictions and screw modes [147, 152,153]. Therefore,
the problem of jet stability has been widely discussed in the literature [154-160]. As shown
in Fig. 4, however, jets from young stars at large distances indeed are similar to a sequence
of flying blobs rather than to a regular flow. Only at small distances can the flow be
considered quite regular (see Fig. 5).
In recent years, special laboratory experiments have been carried out under conditions
maximally similar to those in the nonrelativistic jets [161, 162]. In particular, the
plasma velocity was as high as $100$--$400$ km s$^{-1}$ (which is comparable to the plasma
flow velocity in jets from young stars), and the total current was about \mbox{$1$ MA.} In these
experiments, a strong instability leading to rapid fragmentation of the flow into individual
blobs was also observed.
Nevertheless, it is not obvious that the instability of laboratory pinches can be considered
as an undisputable evidence for the immanent strong instability of astrophysical jets. The
point is that astrophysical jets are always 'specially prepared', since they come out from
the quasispherical subsonic flow region. As a result, the plasma density and the longitudinal
magnetic field profiles near the jet base turn out to be close to the equilibrium ones. In
laboratory experiments, in contrast, the initial plasma density is usually very different
from the equilibrium value [153].
As for relativistic jets (which at large distances also frequently show an irregular
structure), they proved to be more stable [163 165]. The recent numerical simulations
[166] (where the jet was found to be stable after more than 1000 rotations of the central
engine) confirmed this conclusion. Thus, there are no doubts now that the nonrelativistic
Kruskal-Shafranov stability criterion [153]
\begin{equation}
\frac{r_{\rm jet}}{L} \, \frac{B_{\rm p}}{B_{\varphi}} > 1,
\label{ksh}
\end{equation}
where $L$ is the length of the jet, cannot be applied to relativistic flows.
Unfortunately, the limits of the present review do not allow us to discuss this most
important point in more detail.
\subsection{Acceleration}
{\bf 4.2.1 The acceleration mechanism.} First of all, let us consider
the acceleration mechanism itself. It is convenient to start
from expression (43) for the toroidal velocity of the non-
relativistic flow. It is evident that in the subsonic region
${\cal M}^2 < 1$ the flow velocity corresponds to the precise corotation:
\begin{equation}
v_{\varphi} \approx \Omega_{\rm F} \varpi.
\label{uphsmnon}
\end{equation}
In other words, particles can be considered as beads on a wire that determines their
angular velocity of rotation. This situation is quite understandable because, as can
be easily checked, the energy density of the magnetic field within the Alfv\'enic surface
exceeds the plasma energy density, so it is the magnetic field that controls the motion
of particles. In fact, the magnetic field plays the role of a slingshot that provides the
constant angular velocity of the plasma rotation. Therefore, the velocity of particles
linearly increases with increasing axial distance. Then, the following theorem can be
formulated.
{\it In the nonrelativistic limit, the smooth crossing of the fast magnetosonic surface
is possible only if the particle energy flux on this surface is at least one-third of the
total energy losses. In other words, the transonic nonrelativistic flow in the supersonic
region must already be effectively accelerated ($q \sim 1$).}
Indeed, we can set $E_{\rm n} \approx \Omega_{\rm F}I/2\pi c \eta_{\rm n}$ for strongly
magnetized flows near the surface of the central engine, which immediately yields
$E_{\rm n} \approx \Omega_{\rm F}^2r_{\rm F}^2$ for $I \approx i_0 I_{\rm GJ}$. Moreover,
it is easy to show that the poloidal velocity $v_{\rm p}$ near the singular surfaces also
becomes on the order of $\Omega_{\rm F}r_{\rm F}$, too.
By this means the particle acceleration mechanism in the strongly magnetized wind is similar
to that by which a slingshot accelerates a stone. This becomes possible exactly due to the
dominant effect of the poloidal magnetic field. However, this acceleration stops at distances
$r_{\rm F} \sim r_{\rm A}$, since there the magnetic field energy density becomes smaller than the
plasma particle energy density. Additionally, the toroidal magnetic field beyond the
Alfv\'enic surface becomes stronger than the poloidal one, so particles start sliding relative
to the magnetic field lines. As a result, the flow passes to another asymptotic behaviour:
$v_{\varphi} \approx \Omega_{\rm F}\varpi_{\rm A}^2/\varpi$.. The plasma kinetic energy in
this region will be mainly due to their poloidal velocity component.
As for the particle acceleration in the ultrarelativistic limit, near the fast magnetosonic
surface, as we have seen, the particle energy flux must be much smaller than the electromagnetic
energy flux. So the question arises as to whether it is possible to effectively accelerate
particles beyond the sonic surface. Surprisingly, the force balance across the flow should be
considered once again to answer this question.
\vspace*{0.2cm}
{\bf 4.2.2 Efficiency. } So, let us come back to equation (79). We have already mentioned
that the particle energy will be completely determined by the drift motion beyond the light
cylinder $\varpi \gg R_{\rm L}$. Therefore, using Eqn (59) we can write out the
expression for the Lorentz factor of particles in the form
\begin{equation}
\frac{1}{\gamma^2} \approx \frac{{\bf B}_{\rm p}^2}{B^2}
+ \frac{B_{\varphi}^2 - {\bf E}^2}{B^2}.
\label{balance1}
\end{equation}
Now, making use of the relation $B_{\varphi} \approx |{\bf E}| \approx x_r B_{\rm p}$, where
again $x_r = \Omega \varpi/c$, we immediately come to the conclusion that when the second term
in expression (100) can be neglected, the particle Lorentz factor must approach the following
asymptotic solution:
\begin{equation}
\gamma = x_r,
\label{gamma!}
\end{equation}
which, as we shall see, is universal enough. If the curvature of the magnetic field lines
is appreciable, then, in contrast, we can neglect in formula (79) the second term corresponding
to the bulk force $j_{\varphi} B_{\rm p}/c$. As a result, by comparing the corresponding terms
in the force balance equation (79) for the strongly magnetized (i.e., Poynting dominated) flow,
we arrive at another asymptotic solution [167]:
\begin{equation}
\gamma \approx {\cal C} \sqrt{\frac{R_{\rm c}}{\varpi}}.
\label{g2rc}
\end{equation}
where ${\cal C} \sim 1$. Moreover, making use of Eqn (59) and the equilibrium condition
(79), we can write in the general case the relationship [109]
\begin{equation}
\frac{1}{\gamma^2} \approx \frac{1}{x_r^2} + \frac{\varpi}{{\cal C}^2 R_{\rm c}}.
\label{gammahar}
\end{equation}
Here, the value of ${\cal C}$ can be exactly determined for strongly collimated flows
and $\Omega_{\rm F} =$ const [109,150]:
\begin{equation}
{\cal C} = \sqrt{3}.
\label{gammahar'}
\end{equation}
Simply speaking, the Lorentz factor will be determined by the least of the two values
giving by expressions (101) and (102).
\begin{figure}
\begin{center}
\includegraphics[width=0.8\columnwidth]{fig3_06.ps}
\end{center}
\caption{The growth of a particle's Lorentz factor $\gamma$ with distance from the equatorial
plane $z$ [168]. The right curve corresponds to strong collimation ($k > 2$). For weaker
collimation (left curves), the particle acceleration at large distances becomes less effective.
}
\label{fig3_06}
\end{figure}
Thus, the choice between asymptotic solutions (101) and (102) must be determined by how bent the
magnetic surfaces are. It is easy to show that the parabolic magnetic field in which the
field line at a large distance from the central source is given by the equation
$z(\varpi) \propto \varpi^2$ corresponds to the terminating case [126,146]. Indeed, as
the curvature radius can be defined as $R_{\rm c} = [(z')^2+1]^{3/2}/z''$, in the limit
$z' = {\rm d}z/{\rm d}\varpi \gg 1$ for the magnetic surfaces specified by the relationship
$z(\varpi) \propto \varpi^k$, Eqn (102) gives the energy of particles moving along the
magnetic field line:
\begin{equation}
\gamma \propto \varpi^{k-1},
\label{asym}
\end{equation}
where $\varpi$ is, in this case, the current axial distance of a particle. At $k = 2$,
the acceleration efficiency determined from expressions (101) and (102) is the same. Thus,
if the magnetic surfaces are collimated more strongly than those for the parabolic field
(i.e., $k > 2$), the curvature of the magnetic surfaces can be neglected at large distances,
and the energy of the particles will be determined by expression (101). If the flow is poorly
collimated (i.e., $1 < k < 2$), the particle acceleration will be less effective and one
should use expression (102). The numerical simulations [124, 168] fully confirm the picture
presented here. As shown in Fig. 18, the acceleration of a particle moving along the magnetic
field line indeed follows the law $\gamma(z) \propto z^{1/k}$ for strongly collimated flows,
in correspondence with asymptotic behaviour (101), while the acceleration for poorly collimated
outflows at large distances, in full agreement with expression (105), becomes less effective.
We thus arrive at the most important conclusion that the acceleration efficiency of particles
in a supersonic ultrarelativistic wind is determined by the degree of collimation of the magnetic
surfaces. The plasma can be effectively accelerated only in the case where the magnetic
surfaces are collimated more strongly than those for the parabolic field. In this event, one
can use asymptotic solution (101) which shows that the acceleration mechanism, in fact, again is due
to the slingshot effect (the larger the axial distance, the higher the energy). But in this
case, too, the total transformation of the electromagnetic energy into the particle flux
energy, $\gamma \approx \sigma$, can take place only if the jet transverse size exceeds the
value
\begin{equation}
r_{\rm eff} = \sigma R_{\rm L}.
\label{rsigma}
\end{equation}
In particular, the fraction of energy carried by particles for $P_{\rm ext} < B_{\rm eq}^2/8\pi$
can be determined from the simple relationship [10,151]
\begin{equation}
\frac{W_{\rm part}}{W_{\rm tot}} \sim
\frac{1}{\sigma}\left[\frac{B^2(R_{\rm L})}{8 \pi P_{\rm ext}}\right]^{1/4}.
\label{46b}
\end{equation}
If the collimation is poor, the acceleration will be ineffective, since particles begin
sliding from the magnetic field lines. In this case, a much higher transverse size of the jet
is required for accelerating particles to limiting energies:
\begin{equation}
r_{\rm eff} = \sigma^{1/(k-1)} R_{\rm L}.
\label{rsigma'}
\end{equation}
This dependence was also numerically confirmed [169]. In particular, the acceleration becomes
actually impossible for the split monopole magnetic field $k = 1$ (which, as we see, is a
special case that requires separate consideration), since, as follows from estimate (82),
the condition $\gamma \approx \sigma$ is reached only at exponentially large distances from
the compact object. On the other hand, it easy to check that for $k > 1$ both for effective
and for ineffective acceleration mechanisms the condition $\gamma \vartheta \sim 1$ will be
satisfied, where $\vartheta \approx \varpi/z$ is the characteristic opening angle of the
acceleration region.
Here, it is important to emphasize that the self-consistent analysis, in which the magnetic
surfaces for the split monopole field were not assumed to be exactly conical, allowed the
determination of the correct structure of the flow [143]. For example, the expression (82)
for the particle's Lorentz factor discussed above is determined precisely by the small
curvature of the magnetic surfaces, and so exactly corresponds to expression (102). It should
be remembered that for a long time the flow in the split monopole magnetic field had been,
in fact, the only example in which analytical results could be obtained. Therefore, a
strong opinion was that the effective particle acceleration beyond the fast magnetosonic
surface is completely impossible. As wee see, this conclusion proved to be incorrect in the
general case.
\subsection{Subsonic flows}
As discussed above, all observed jets must be transonic. It is this property that
allowed us to find the longitudinal current flowing in the magnetosphere and, hence,
to estimate the central engine power. However, until recently many models considered
subsonic flows, and the magnitude of the current was determined from other considerations
[170, 171]. We shall briefly discuss below two the most known examples of such subsonic flows.
\vspace*{0.2cm}
{\bf 4.3.1 Pulsar magnetospheres.} Starting in the early 1970s, pulsar magnetospheres have
been discussed mostly in the force-free approximation [107, 142, 172-174]. This was based
on the fact that the plasma filling the neutron star magnetosphere is secondary with respect
to the magnetic field, and so (at least inside the light cylinder) the particle energy density
can be neglected. The Grad-Shafranov equation ($A.28$) in the force-free approximation (which
in this case is simply called the pulsar equation) becomes elliptical. Therefore, for numerical
modeling of the axisymmetric magnetospheres one has to impose an additional condition at the
external boundary of the integration region [98, 100 104]. Usually, one chooses the condition
of radiality of the magnetic field lines (Fig. 19). In this case it is this additional condition
that fixes the longitudinal current in the magnetosphere. Thus, it is not surprising that this
current turns out to be close to the longitudinal current $I_{\rm M}(\theta)$ (80) obtained by
F C Michel for the monopole solution shown in Fig. 14. If the absence of the reverse current
flowing along the equator (so that the current closing was done by bulk currents only) was
chosen as the additional condition, the magnetic field structure beyond the light cylinder
turned out to be significantly different [175].
\begin{figure}
\begin{center}
\includegraphics[width=0.8\columnwidth]{fig3_07.ps}
\end{center}
\caption{The structure of the axisymmetric radio-pulsar magnetosphere in the model [98].
The flow is assumed to be radial at large distances.
}
\label{fig3_07}
\end{figure}
\begin{figure}
\begin{center}
\includegraphics[width=0.8\columnwidth]{fig3_08.ps}
\end{center}
\caption{The magnetic field structure in the 'rotating split monopole'
model [176]. In the force-free case, particles moving radially at the speed
of light can provide the formation of the current sheet (the wavy curve) which
separates magnetic fluxes in the equatorial region.
}
\label{fig3_08}
\end{figure}
Interestingly, a similar structure also appears in different models of the inclined rotator.
Here, first of all, we should highlight the model of the 'rotating split monopole' (Fig. 20).
In the force-free approximation (when massless particles move with the speed of light), the
monopole magnetic field, namely
\begin{eqnarray}
&&\Psi(r,\theta, \varphi, t) = \Psi_{0}(1-\cos\theta), \nonumber \\
&& \hspace*{1cm} \theta < \pi/2 - \chi \cos(\varphi - \Omega t + \Omega r/c), \\
&&\Psi(r,\theta, \varphi, t) = \Psi_{0}(1+\cos\theta), \nonumber \\
&& \hspace*{1cm} \theta > \pi/2 - \chi \cos(\varphi - \Omega t + \Omega r/c),
\label{e38''}
\end{eqnarray}
was found to be also the solution of the problem [176]. In this case, the electromagnetic
fields inside cones with angles \mbox{$\theta < \pi/2 - \chi$} and
\mbox{$\pi - \theta < \pi/2 - \chi$} near the rotational axis are independent of time and
the angle $\varphi$ and coincide with the fields of an axisymmetric rotator, while
in the equatorial plane the field directions reverse sign at the moment when
the current sheet intersects the given point. A similar structure beyond the light cylinder
was obtained for the rotating dipole, too [177]. True enough, in contrast to the 'rotating
magnetic monopole', the total longitudinal current here depends on the inclination angle
$\chi$, although not very strongly.
The last property can be easily explained. As we have seen, the very existence of the MHD
wind far away from the light cylinder is possible only if the toroidal magnetic field is
comparable in magnitude to the electric field. But this is possible only if a sufficiently
large longitudinal current $I \approx I_{\rm M}^{({\rm A})}$ flows in the magnetosphere.
Let us keep in mind that this value of the current can also be found from the
condition of the smooth crossing of the singular surfaces in the full MHD version. In none
of the numerical calculations mentioned above were bounds on the longitudinal currents
flowing from the neutron star surface set. Thus, it is not surprising that the longitudinal
current obtained from the solution of the problem considered turned out to be on the order
of $I_{\rm M}(\theta)$. As a result, the energy losses in model [176] were found to be
independent of the angle $\chi$. Energy losses in model [177] were also found to be weakly
dependent on the inclination angle:
\begin{equation}
W_{\rm tot} \approx \frac{1}{4} \, \frac{B_0^2\Omega^4R^6}{c^3}\left(1 + \sin^2\chi\right).
\label{Wspit}
\end{equation}
However, the following problem emerges here. All the theories of particle generation near
the neutron star magnetic poles [178-180] unambiguously suggests that the longitudinal
current density cannot be higher than that of the local Goldreich current which, as follows
from definition (9), depends on the inclination angle $\chi$
\begin{equation}
j_{\rm GJ} \approx \frac{ \Omega B}{2\pi}\cos\chi.
\label{GJforj}
\end{equation}
Indeed, the local Goldreich charge density near the magnetic poles for the orthogonal
rotator must be $(\Omega R/c)^{1/2}$ times smaller than in the axisymmetric magnetosphere.
Thus, the longitudinal current flowing along open magnetic field lines should be correspondingly
smaller (for ordinary pulsars this factor can be as high as $10^{2}$). In the 'rotating split
monopole' model considered above, this problem does not arise, because at any inclination angle
in the polar magnetospheric regions the current is always the same as in the axisymmetric case. Just
this current provided the necessary toroidal magnetic field. For the inclined dipole, in contrast,
it is necessary to additionally assume that the longitudinal current in the polar cap regions can
be significantly higher than the local Goldreich current (A Spitkovsky, private communication).
There is one more problem related to the decrease in the longitudinal current density as
$\chi \rightarrow 90^{\circ}$ . The point is that the current losses for the local Goldreich
current must decrease as the inclination angle $\chi$ increases [56,107]:
\begin{equation}
W_{\rm tot} = \frac{f_*^2}{4} \, \frac{B_0^2\Omega^{4}R^6}{c^3}i_0\cos\chi.
\label{e15}
\end{equation}
Here, the coefficient $f_* = 1.59$--$1.96$ depends only on the inclination angle $\chi$.
It is necessary to stress that in addition to the $\cos \chi$ factor (which is related to the
scalar product $W_{\rm tot} = -{\bf \Omega} \cdot {\bf K}$, where ${\bf K}$ is the braking
torque), the significant dependence of the current losses $W_{\rm tot}$ on the inclination
angle is also contained in the quantity $i_0$. The matter is that in the definition of the
dimensionless current $i_0 = I/I_{\rm GJ}$ in expression (113), the denominator contains the
Goldreich current for the axisymmetric case, whereas at nonzero angles $\chi$ the
Goldreich charge density itself depends on the angle $\chi$ near the magnetic poles. It is
logical to expect that the dimensionless current $i_0$ for the inclined rotator will be bounded
from above as $i_0^{(\rm max)}(\chi) \sim \cos\chi$. As a result, the current losses, in
contrast to relationship (111), must decrease with angle $\chi$ at least as $\cos^2\chi$.
In particular, for $\chi = 90^{\circ}$ (when $\cos^2\chi$ must be replaced by its characteristic
value within the polar cap region $\cos^2\chi$, we obtain [56]
\begin{equation}
W_{\rm tot} = c_{\perp}\frac{B_0^2\Omega^{4}R^6}{c^3}
\left(\frac{\Omega R}{c}\right) i_{\rm A}.
\label{e17'}
\end{equation}
Here, $i_{\rm A} = 1$ for the local Goldreich current, and the coefficient $c_{\perp} \sim 1$
now depends not only on the profile of the asymmetric longitudinal current, but also on the
polar cap shape.
Usually, when discussing this point, the following counter-argument against the decrease
in current losses with increasing $\chi$ is invoked. In the expression for the braking
torque, namely
\begin{equation}
{\bf K} =
\frac{1}{c}\int[{\bf r}\times[{\bf J}_{\rm s}\times{\bf B}]]{\rm d}S,
\label{e3}
\end{equation}
The surface current ${\bf J}_{\rm s}$ indeed must decrease as $\cos \chi$ as angle $\chi$
increases. But then the characteristic distance from the axis to the polar cap, in contrast,
will increase as $\sin \chi$, so ultimately the losses will be weakly dependent on the
inclination angle $\chi$ even for the local Goldreich current.
\begin{figure}
\begin{center}
\includegraphics[width=0.7\columnwidth]{fig3_09.ps}
\end{center}
\caption{The structure of electric currents flowing near the magnetic poles of an
orthogonal rotator. Currents flowing along the sepratrix (thin arrows) that separates
the regions of the open and closed magnetic field lines are adjusted to bulk currents
(contour arrows) in such a way that the closing surface current is fully concentrated
within the polar cap.
}
\label{fig3_09}
\end{figure}
As follows from a more precise analysis [107], however, the real structure of the surface
currents in the polar cap region was ignored in this obvious, at first glance, consideration.
Referring to Fig. 21, the closing currents in fact must be such that the current averaged over
the polar cap region is zero. As a result, one needs to consider higher-order effects (like,
for example, the neutron star surface curvature effect), when determining the radio pulsar
spin-down rate. But if the surface current averaged over the polar cap region is indeed zero,
then, as shown in Fig. 21, a surface current comparable in amplitude to the bulk current in
the magnetosphere must flow along the separatrix separating the regions of open and closed
magnetic field lines. Remarkably, in numerical modeling of the inclined rotator [181],
reverse currents flowing along the separatrix were indeed discovered. Finally, it worth
noting that there is no contradiction between relations (111) and (113), either. As we
specially emphasized, the approximation expression (111) was obtained in Ref. [177] for a
flow in which the longitudinal current was larger than the local Goldreich current, which
corresponds to the condition $i_0 > 1$ ($i_{\rm A} > 1$, respectively).
In any case, studies of the last decade, in our opinion, have at last formulated the problem
whose solution will provide significant progress in the understanding of the structure of
radio pulsar magnetospheres. The problem is whether the plasma generation region in the neutron
star magnetosphere can provide a sufficiently large longitudinal current which
is necessary to launch the MHD wind from the inclined rotator. If the necessary current can be
produced, nothing will prohibit the formation of the MHD wind in which the main part of the
energy will be carried by the electromagnetic field. If the generation of a current which is
significantly larger than the local Goldreich current is impossible, then the toroidal magnetic
field near the light cylinder will be smaller than the poloidal magnetic field. In that case, a
light surface will inevitably be formed near the light cylinder, on which the current closing
and particle acceleration up to energies $\gamma \sim \sigma$ will occur [56]. Thus, the problem
of the effective particle acceleration in pulsar winds, which we mentioned in Section 2.4, can be
solved.
Interestingly, the possibility of answering this question has apparently emerged a short
time ago. This test is related to the unusual properties of the radio pulsar
PSR $B$1931$+$24 [182]. It differs from other pulsars in that it stays in the active
state for $5$--$10$ days, and then its radio emission switches off in less than 10 s,
and the source is not observed during the next $25$--$35$ days. It is important that
the spin-down rate in these two states be different:
\begin{eqnarray}
\dot \Omega_{\rm on} & = & -1.02 \times 10^{-14} \, {\rm s}^{-2}, \\
\dot \Omega_{\rm off} & = & -0.68 \times 10^{-14} \, {\rm s}^{-2},
\end{eqnarray}
so that
\begin{equation}
\frac{\dot \Omega_{\rm on}}{\dot \Omega_{\rm off}} \approx 1.5.
\end{equation}
A similar behavior was later observed in the pulsar PSR $J$1832$+$0031
($t_{\rm on} \sim 300$ days, $t_{\rm off} \sim 700$ days), with the ratio
$\dot \Omega_{\rm on}/\dot \Omega_{\rm off} \approx 1.5$ again.
It is logical to assume that the difference in the spin-down rates in these pulsars
is simply associated with the fact that the spin-down in the switch-on state is due
to the current losses, and in the switch-off state, when the magnetosphere is not
filled with plasma, is due to magnetodipole radiation [183, 184]. Then, making use
of equations (3) and (113) we obtain
\begin{equation}
\frac{\dot\Omega_{\rm on}}{\dot\Omega_{\rm off}}
= \frac{3f_{\ast}^2}{2} \, {\rm cot}^2\chi,
\end{equation}
which yields the reasonable inclination angle $\chi \approx 60^{\circ}$. On the other hand,
using expression (111) [177] for the switch-on state, we arrive at the relationship
\begin{equation}
\frac{\dot\Omega_{\rm on}}{\dot\Omega_{\rm off}} = \frac{3}{2} \,
\frac{(1+ \sin^2\chi)}{\sin^2\chi}.
\end{equation}
Clearly, this quantity cannot be equal to 1.5 for any $\chi$. If this interpretation of observations
holds true, this implies that the longitudinal current in the magnetosphere indeed does not exceed
the local Goldreich current.
It should be emphasized that we have assumed above that the magnetospheric plasma
fully screens the magnetodipole radiation of the neutron star. This conclusion,
which was formulated for the first time in paper [107], seems now to be directly
confirmed, since both in the model [176] and in the numerical calculations [107]
there are no alternating electromagnetic fields decaying as $1/r$.
\vspace*{0.2cm}
{\bf 4.3.2 'The magnetic tower'.} 'The magnetic tower' is the model of collimated jets
proposed by D Lynden-Bell in
1996 [128] (Fig. 22) and later on widely discussed in relation to both relativistic
and nonrelativistic sources [171, 185-187]. It is based on the assumption that there
exists an intensive wind outflowing perpendicular to the accretion disc and stretching
the magnetic field lines along the rotational axis. Here, one usually assumes that
the initial magnetic field was quasi-dipole, i.e., it consisted of magnetic field
lines with one end frozen in the central star (or in the inner regions of the accretion
disc) and another end frozen in the outer parts of the disc. Such a cylindric flow cannot
intersect singular surfaces since the magnetic field lines remain at a constant distance
from the axis of rotation. That is why, the longitudinal current in this model will be
determined exactly by the differential rotation that leads to the magnetic field line
twisting. Because of this, as in the case shown in Fig. 11, the magnetic tower top will
propagate upwards by gradually increasing the volume occupied by the twisted magnetic
field, while a stationary configuration restricted by the ambient pressure will be formed
at smaller distances.
\begin{figure}
\begin{center}
\includegraphics[width=0.55\columnwidth]{fig3_10.ps}
\end{center}
\caption{The magnetic field structure in the model of the 'magnetic tower' that can
exist under sufficiently strong ambient pressure $P$ [128]. The energy is transferred
from the central engine along the rotational axis, and back to the accretion disc along
the jet periphery.
}
\label{fig3_10}
\end{figure}
There are two additional important properties that distinguish the magnetic tower model
from the transonic flows we are considering. First, if the magnetic field lines are anchored
in the accretion disc, a configuration with almost zero total magnetic flux will form during
the outflow (see Fig. 22). In other words, the direction of the poloidal magnetic field on the
periphery of the jet will be different from that near the jet axis. Second, when the magnetic
field loops do not extend beyond the light cylinder and, hence, do not open, the energy will
continue being transferred along the magnetic fields lines. But this means that the energy
will be carried away from the central engine only along the rotational axis, while on the
jet periphery the energy flux will be directed back to the accretion disc [186,187].
If such a configuration were stationary, the reverse energy flux would be exactly the same
as the energy flux outgoing from the central engine closely to the rotational axis. However,
we have seen that the equipotenital condition is violated in the switching-on wave, so the
reverse energy flux turns to be smaller than the escaping flux.
Thus, in the very setup the problem of the magnetic tower formation is different from
that describing transonic flows. The longitudinal current determining the energy losses
in no way relates to the critical conditions on the singular surfaces which, as we specially
stressed above, should unavoidably appear in all real compact sources (see also Section 2).
Hence, apparently, the magnetic tower model cannot correspond to the reality. The results
of numerical simulations also support this conclusion. When the flow intersected the singular
surfaces, the magnetic field lines became open and the energy flux was directed outward from
the central engine in the regions of both outgoing and incoming magnetic field lines [188].
On the other hand, in the case where the flow remained subsonic in numerical simulations, the
magnetic tower formation was indeed observed [50,189].
\section{Estimation of parameters}
Thus, we have seen that at present it has turned out to be possible to understand many
key points related to the formation and the internal structure of collimated outflows.
In doing so, we have managed to find several key parameters that determine the basic
physical properties of ejected matter. First and foremost, these include the magnetization
parameter $\sigma$, the multiplicity parameter $\lambda$, and the initial velocity
$v_{\rm in}$ (the Lorentz factor $\gamma_{\rm in}$) of the outflow. We shall try below
to understand how precisely these parameters can be estimated for the observed jets.
It should be emphasized from the very beginning that each time it is necessary to clearly
separate which collimated outflows are being considered. Indeed, the properties of jets
from active galactic nuclei can be quite different on the scale of the host galaxy
(see Fig. 1) and in the formation region (see Fig. 2), where their transverse size is
close to 1 pc. However, as we mentioned above, the relativistic jets can be stable on
all scales. On the other hand, in the majority of cases nonrelativistic jets at large
distances from the star can indeed be unstable (see Fig. 4), so for them the treatment
considered here, strictly speaking, can be applied only in the innermost parts.
\subsection{Active galactic nuclei}
Despite longstanding efforts, we know very little about the internal structure of jets
from active galactic nuclei. In particular, we still do not have an answer to the key
question of whether it is the black hole, and not the inner parts of the accretion disc,
that is the central engine which is responsible for the black hole power [190,191].
In the practical sense, the main uncertainty appears in the determination of the particle
production multiplicity $\lambda$. Indeed, as already noted, the plasma on the magnetic
field lines threading the black hole horizon (which is needed both to screen the
longitudinal electric field and to produce the longitudinal electric current) must
be generated in the magnetosphere itself between two families of singular surfaces.
Some fraction of the plasma will escape the magnetosphere, while another part will
accrete onto the black hole. Correspondingly, it is still unclear which quantities
determine the density of matter flowing out of the accretion disc surface (see, for
example, Refs [192, 193]). Moreover, it cannot be ruled out that the jet at large
distances from the central engine will be additionally 'loaded' due to interaction
with stellar winds from surrounding stars [194], or, for example, due to the 'photon
breeding' effect (creation of the secondary electron-positron plasma by hard gamma-ray
quanta generated in the interaction of the relativistic outflow with the ambient medium)
[195,196]. That is why, the properties of jets on the scales of several kiloparsecs can
be significantly different from those in the jet formation region.
In the region of field lines threading the black hole horizon several plasma generation
mechanisms are currently being discussed, in which, however, plasma is ultimately generated
always due to two-photon pair creation. The one-photon conversion, which plays the leading
role in radio pulsar magnetospheres, here turns out to be ineffective, since the probability
of pair creation in magnetic fields $B \sim B_{\rm Edd} \sim 10^4$ G is vanishingly small.
First and foremost, secondary plasma generation can be related to the direct two-photon
process $\gamma + \gamma \rightarrow e^{+} + e^{-}$ (see, for example, Ref. [105]), where the necessary
gamma-quanta are emitted from the inner regions of an accretion disc. This mechanism, with
a high value of the parameter $\lambda_{\rm AGN_1} \sim 10^{10}$--$10^{12}$, was discussed
in the pioneering paper by Blandford and Znajek [33]. However, sufficiently high temperatures
providing the necessary number of hard gamma-ray photons with energies above the pair creation
threshold ${\cal E}_{\rm min} = m_{\rm e}c^2$, and small free path lengths of photons are required
for this mechanism to be effective. Presently, the accuracy of the compactness parameter estimate
$l_{a,{\rm AGN}} \sim 1$--$100$ does not allow one to make definitive conclusions on the
efficiency of this mechanism of particle creation.
On the other hand, such a high particle density must be typical for a wind outflowing from
the accretion disc surface. Let us keep in mind that even if the energy release related to
such a wind is insignificant, it can play the decisive role in the matter outflow collimation
[197]. Here, the energy of the jet core observed at high radio frequencies and in gamma-rays
will be associated with ultrarelativistic particles extracting energy from the rotating black
hole.
There is another mechanism capable of bringing particles into the region of magnetic field
lines threading the black hole horizon even in the absence of hard gamma-ray quanta. This
mechanism is similar to the particle creation process in the outer gap of the pulsar
magnetosphere [198]. Indeed, the exact relativistic expression for the Goldreich charge density
$\rho_{\rm GJ}$ takes the form [9]
\begin{equation}
\rho_{\rm GJ} = -\frac{1}{8\pi^2}\nabla_k
\left(\frac{\Omega_{\rm F}-\omega}{\alpha}\nabla^k\Psi\right).
\label{k51}
\end{equation}
In particular, near the rotational axis we simply have
\begin{equation}
\rho_{\rm GJ} \approx -\frac{(\Omega_{\rm F} - \omega)B}{2\pi \alpha}.
\label{k52}
\end{equation}
As a result, the general relativity effects cause the Goldreich density to
vanish at $\omega \approx \Omega_{\rm F}$. Therefore, a region quite similar
to the outer gap in pulsar magnetospheres appears in the black hole magnetospheres.
The formation of longitudinal electric fields is also possible in this region, since
the charge-separated plasma flow cannot provide the fullfilment of the condition
$\rho_{\rm e} = \rho_{\rm GJ}$. As a result, it turned out that under real
conditions the size of the acceleration region is much smaller than the system's size,
so that the acceleration region does not affect the global structure of the magnetosphere
[34, 35]. In this model, the particle production multiplicity is rather small:
\begin{equation}
\lambda_{\rm AGN_2} \sim 10 - 100.
\label{lmbagn2}
\end{equation}
Hence, we shall consider below both large and small values of the parameter $\lambda$.
Next, it should be remembered that to explain the high efficiency of the energy release
from the central engine we need to assume that the rotation parameter $\Omega R/c$ must be
not too much smaller than unity. In other words, the light cylinder radius must not exceed
the central engine size too much. As a result, the observed transverse size $r_{\rm jet}$ of
relativistic jets will be three-five orders of magnitude greater than the light cylinder
radius $R_{\rm L}$. That is why, far from the central engine most magnetic field lines
must be far beyond the light cylinder.
However, according to the relationship
\begin{equation}
\frac{B_{\varphi}}{B_{\rm p}} = x_r,
\label{bphi'}
\end{equation}
where again $x_r = \Omega_{\rm F}\varpi/c$, which follows from the definition of the magnetic
field components at $I \approx I_{\rm GJ}$, this implies that the toroidal magnetic field will
be the same three-five orders of magnitude stronger than the poloidal magnetic field. Therefore,
the magnetic field must have a strongly pronounced spiral structure. Correspondingly, the
electric field must also be three-five orders of magnitude larger than the poloidal magnetic
field. At present, VLBI (Very Large Baseline Interferometry) methods provide a lot of data on
the polarization of the innermost parts of jets [199 201]; however, so far it is impossible
to unambiguously determine the magnetic field structure from the observations.
Notice, finally, that strong twisting of magnetic field lines does not imply that the plasma
motion will also occur along strongly twisted trajectories. As stressed above, an almost
poloidal drift motion in the crossed electric and magnetic fields will be the principal motion
of particles beyond the light cylinder. Thus, the toroidal velocity for $x_r > 1$ is given by
the simple relationship
\begin{equation}
v_{\varphi}(x_r) = \frac{c}{x_r}.
\label{44a}
\end{equation}
Further, if the magnetization parameter $\sigma$ exceeds the ratio $r_{\rm jet}/R_{\rm L}$,
the plasma Lorentz factor can again be estimated from the asymptotic solution (101), so that
$\gamma \approx r_{\rm jet}/R_{\rm L}\sim 10^{3}$--$10^{5}$. If $\sigma < r_{\rm jet}/R_{\rm L}$,
the effects of the finite mass of particles will limit their energy growth at large axial
distances. To determine the magnetization parameter $\sigma$, as was shown, it is necessary
to know the particle production multiplicity $\lambda$.
As noted above, it is impossible at present to determine the basic parameters of the
outflowing plasma from observations. Nevertheless, some estimates can still be made.
For example, one of the methods to determine the value of $\lambda$ is based on the
assumption that the synchrotron radiation self-absorption occurs at the base of the
jet [202]. This assumption allows one to estimate the particle number density [203].
On one-parsec scales, the characteristic particle number densities thus were found to
be $10^2$--$10^4$ cm$^{-3}$, which gives
\begin{equation}
\lambda_{\rm AGN_1} \sim 10^{10} - 10^{12}.
\label{lmbagn1}
\end{equation}
If this indeed is the case, intensive secondary particle creation near the black hole
horizon should have occurred. Here, according to relation (56) the value of $\sigma$ cannot
exceed one hundred:
\begin{equation}
\sigma_{\rm AGN_1} \sim 10^{2} - 10^{3},
\label{sgmagn1}
\end{equation}
which is much smaller than the ratio $r_{\rm jet}/R_{\rm L} \sim 10^5$ corresponding
to the maximum possible Lorentz factor derived from the asymptotic solution (101). Therefore,
an almost complete transformation of the electromagnetic energy into particles'
energy must occur in the process of collimation. Notice that in this case, although
$\gamma_{\rm max} = \sigma \sim 10^2$--$10^3$ exceeds the particle energies that are
required to explain the apparent superluminal motion effect, it is still insignificant.
Here, almost all the energy flux in the jet will be related to the flux of accelerated
particles. For $\sigma \sim 10^2$--$10^3$, the radius \mbox{$r_{\rm F} \sim \sigma^{1/3}R_{\rm L}$}
(66) of the fast magnetosonic surface must be smaller than one hundred radii from the
central engine, which is \mbox{$10^{14}$--$10^{16}$ cm.} Thus, the flow in the jet must indeed
be supersonic.
Now making use of expressions (89), (95), and (96), we arrive at the conclusion that
in this case all critical magnetic fields for the reasonable value of $\gamma_{\rm in} \sim 10$
are larger than $B_{\rm ext} \sim 10^{-6}$ G corresponding to the ambient pressure
(for convenience, all parameters discussed here are listed below in Table 2). But this
means that a denser core must exist in the center of the jet, and the subsonic flow near
the axis will not be formed. At last, the ejection rate of electron-positron pairs,
${\dot N} = \pi R_0^2 \cdot \lambda n_{\rm GJ} c$, can be estimated, which, as can be
easily checked, is determined using the simple relation
\begin{equation}
{\dot N} \sim \lambda \left(\frac{W_{\rm tot}c}{e^2}\right)^{1/2}.
\end{equation}
As a result, we have ${\dot N}_{\rm AGN_1} \sim 10^{49}$ particles s$^{-1}$
(hereinafter we set $M = 10^9 M_{\odot}$). In other words, about $10^{63}$
electron-positron pairs will be injected into host galaxy over the time
of the active life of a galactic nucleus, $\tau \sim 10^7$ years. This number,
incidentally, is quite sufficient to explain the intensity of the annihilation
line emitted from the Galactic center, which, as is well known, requires about
$10^{43}$ annihilations per second [204].
If the multiplicity factor of the secondary particle creation is small,
$\lambda_{\rm AGN_2} \sim 10$--$100$, the inner structure of the jet must
be significantly different, since now all critical fields are below the
value of $10^{-6}$ G corresponding to the ambient pressure. Here, one obtains
\begin{equation}
\sigma_{\rm AGN_2} \sim 10^{10} - 10^{12},
\label{sgmagn2}
\end{equation}
so that the plasma Lorentz factor, according to the asymptotic solution (101), can be as high
as approximately $10^4$--$10^5$, and a subsonic flow region must be formed in the center
of the jet. In this case, the energy flux is Poynting-dominated. Correspondingly, the
dense core will not be formed, so that both the poloidal magnetic field and the plasma
density are weakly dependent on the axial distance. The electron-positron pair injection
rate will be much smaller: ${\dot N}_{\rm AGN_2} \sim 10^{39}$ particles s$^{-1}$. However,
in this case, too, the fast magnetosonic surface radius $r_{\rm F} \sim 10^{17}$--$10^{18}$ cm
will be smaller than the jet transverse size. Notice that here there is no direct contradiction
with observations, since, as has been noted, the drift motion of particles will be directed
almost along the poloidal magnetic field. This particle motion does not produce synchrotron
radiation. Hence, one should be cautious in using the standard synchrotron radiation formulas
to estimate the value of the magnetic field and the lifetime of relativistic particles.
We should keep in mind that, when estimating radiation from relativistic jets, one usually
assumes that an approximate equipartition between the energy densities of particles and the magnetic
field takes place ($B^{({\rm com})2}/8\pi \sim \gamma^{({\rm com})} n^{({\rm com})} m c^2$) in
the rest frame of plasma. The parameters we introduced allow us to obtain simple relationships
for all quantities in this reference frame. In particular, it is easy to show that the
characteristic Lorentz factor of particles in the plasma rest frame is simply the following:
\begin{equation}
\gamma^{({\rm com})} \approx \frac{\sigma}{\gamma},
\label{gcom}
\end{equation}
where $\gamma$ is the bulk Lorentz factor of the hydrodynamic flow. On the other hand,
one finds $B^{({\rm com})} \approx (x_r/\gamma)B_{\rm p}$. Consequently, we have
$B^{({\rm com})} \approx B_{\rm p}$ in the effective acceleration region ($\gamma = x_r$),
while for the asymptotic solution (102) we obtain $B^{({\rm com})} \gg B_{\rm p}$.
\subsection {Microquasars}
If the operation of the central engine in microquasars indeed can be described by the
MHD model considered here, it is again possible to assume that the observed subrelativistic
jet velocities must correspond to $\sigma$ on the order of $3$--$10$. Then, according to
relation (56), we should conclude that the particle production multiplicity in the
microquasar magnetosphere must be fairly large:
\begin{equation}
\lambda_{\mu{\rm QSO}} \sim 10^{10}.
\label{lmbmQSO}
\end{equation}
Such a high value is also supported by the compactness parameter $l_{a,\mu{\rm QSO}} \sim 10^4$.
Then, the electron-positron pair ejection rate should be ${\dot N}_{\mu{\rm QSO}} \sim 10^{43}$
particles s$^{-1}$. Finally, the large values of the critical fields \mbox{($B_{\rm min} \sim 10^{4}$ G,}
$B_{\rm eq} \sim 10^{3}$ G, $B_{\rm cr} \sim 10^{6}$ G) indicate that a dense core must exist
in the jet center, and the subsonic flow near the rotational axis does not form.
On the other hand, if $\sigma$ is indeed not too high, the fast magnetosonic surface
($r_{\rm F} \sim \sigma^{1/3}R_{\rm L}$) must lie close to the light cylinder, i.e.,
at distances of order \mbox{$10^{7}$--$10^{8}$ cm.} This scale is much smaller than the distance
from the central engine to the companion star in binaries involving a microquasar. Hence,
we can conclude that the interaction of the stellar wind and matter ejected from the
microquasar occurs in the supersonic regime.
\subsection {Sources of cosmological gamma-ray bursts}
Now let us see which parameters can characterize jets outflowing cosmological gamma-ray
bursts. It should be recalled that, according to one of the most popular models, a rapidly
rotating central engine (magnetar, black hole) is immersed in the progenitor massive star
[40]. That is why, the ambient pressure for the jet in its formation region is provided
not by the surrounding medium with very small pressure, but by the matter of the massive
star itself (the equivalent magnetic field $B_{\rm ext} \sim 10^{6}$--$10^{8}$ G). Notice
also that in cosmological gamma-ray bursts there can be one more mechanism of the
electron-positron pair creation, related to neutrino annihilation. Such neutrinos can be
copiously created during supercritical accretion onto the collapsing stellar core [205].
The starting point that can shed light on the physical conditions inside the central engine
can be the characteristic particle Lorentz factor $\gamma \sim 300$, which in this case can
be naturally related to the magnetization parameter $\sigma$:
\begin{equation}
\sigma_{\rm GRB} \sim 10^{2} - 10^{3}.
\label{sigmaGRB}
\end{equation}
If the condition $\gamma \ll \sigma$ is satisfied, the total energy release from
the central engine would be unrealistically high. Now, using Eqn (56) to estimate the
plasma generation multiplicity, we obtain
\begin{equation}
\lambda_{\rm GRB} \sim 10^{13} - 10^{14}.
\label{lambdaGRB}
\end{equation}
Such a huge value unambiguously evidences that the particle creation efficiency must be high
enough. Indeed, formula (57) gives very large compactness parameter $l_{a} \sim 10^{15}$.
Correspondingly, for the electron-positron pair ejection rate we find
${\dot N}_{\rm GRB} \sim 10^{53}$ particles s$^{-1}$.
Next, the very small sizes of the central engine together with the moderate value of the
magnetization parameter $\sigma$ shows that the fast magnetosonic surface radius $r_{\rm F}$
(66) should not exceed $10^{7}$--$10^{8}$ cm, which is significantly smaller
than the size of the progenitor star. Consequently, the matter outflow in the jet becomes
supersonic before it exits the star. Finally, expressions (89) and (95) for the characteristic
magnetic fields $B_{\rm min}$ and $B_{\rm eq}$ indicate that they are in the range of
\mbox{$10^{6}$--$10^{8}$ G,} i.e., their pressure is comparable with that inside the progenitor star.
Therefore, the jet transverse size will indeed be sufficient to accelerate particles up to
energies $\gamma \sim \sigma$.
Notice at last that the condition $\gamma \vartheta \sim 1$ is certainly not satisfied for
gamma-ray bursts, since in that case the jet spread angles would only be $0.1^{\circ}$, while
observations indicate that $\vartheta \sim 1$--$10^{\circ}$ [206]. Such a flow can also be
realized. For example, it was shown in paper [207] (see also Ref. [208]) that in the model
of an infinitely extended progenitor star, where the ambient pressure decreases graduately
according to a power law, the acceleration turns out to be not very effective in comparison
with a more realistic model in which the ambient pressure beyond the star is assumed to be
low. As it has turned out, in both cases the flow corresponds to a weakly collimated flux
with $1 < k < 2$, where the particle energy follows the asymptotic behaviour
$\gamma \approx (R_{\rm c}/\varpi)^{1/2}$ (102). In the former case, however, the radius
of curvature $R_{\rm c}$ of the magnetic surfaces, which is determined by the pressure
decrease law inside the progenitor star, turns out to be sufficiently small, which precludes
plasma from being effectively accelerated. Beyond the star, magnetic field lines straighten
up (and hence the radius of curvature increases), which leads to a more effective acceleration.
As we see, the simple analytical asymptotic solutions obtained above allow easy interpretation
of the numerical experiment.
\subsection{Radio pulsars}
Radio pulsars, undoubtedly, only indirectly relate to the topic considered here,
since their magnetospheres are certainly not axisymmetric and stationary.
It is not then surprising that jets, as we have noted, are observed only from two
energetic pulsars. Nevertheless, many points, which were possible to clarify in
the theory of the neutron star magnetosphere, undoubtedly allow us to shed light
on the nature of other compact objects, too.
First and foremost, it should be noted that basic parameters characterizing the pulsar
wind are well known inside the light cylinder. This is due to our good knowledge of
the process of plasma creation near the magnetic polar caps of a neutron star. Numerous
calculations have shown that the general properties of the secondary electron-positron
plasma flowing out from the magnetosphere turned out to be only a little sensitive to
the details of the acceleration region structure. For most models [178, 180, 209, 210],
both the density and the energy spectrum of the plasma flowing out are universal enough.
Therefore, with certainty we can say that the plasma flowing along open field lines in
the pulsar magnetosphere includes both the primary particle beam with energy
${\cal E} \approx 10^7$ MeV and density close to the Goldreich density and the secondary
electron-positron component. Its energy spectrum with a good accuracy exhibits a power
law dependence
\begin{equation}
{\cal E} \propto {\cal E}^{-2},
\end{equation}
ranging from ${\cal E}_{\rm min} \sim 10$--$100$ MeV to
${\cal E}_{\rm max} \sim 10^4$ MeV. The total density of the secondary plasma for ordinary
pulsars exceeds the Goldreich density by $10^3$--$10^4$ times. And only for the most energetic
pulsars can the multiplicity factor reach $10^5$.
Thus, the parameter $\lambda$ for radio pulsars is determined quite reliably:
\begin{equation}
\lambda_{\rm PSR} \sim 10^3 - 10^4.
\label{lmbpsr}
\end{equation}
Now, making use again of relation (56) we obtain
\begin{equation}
\sigma_{\rm PSR} \sim 10^3 - 10^4.
\label{sgmpsr}
\end{equation}
And only for the most energetic pulsars do we find $\sigma_{\rm PSR} \sim 10^5$--$10^6$.
Thus, the condition \mbox{$\gamma_{\rm in} \gg \sigma^{1/3}$} is satisfied in
the vast majority of pulsars, which corresponds to a slow rotation of the central object
[99]. Indeed, the equality $\gamma_{\rm in} = \sigma^{1/3}$ can be written out in the
form $P = P_{\rm cr}$, where
\begin{eqnarray}
P_{\rm cr} = \pi \frac{R}{c}\,
\left[ \frac{2}{\lambda \gamma_{\rm in}^3}
\, \left(\frac{\omega_{B}R}{c}\right)\right]^{1/2}
\nonumber \\
\sim 10^{-3} \, \left(\frac{\lambda}{10^4}\right)^{-1/2}
\left(\frac{\gamma}{10^{2}}\right)^{-3/2}
\left(\frac{B_0}{10^{12} \, {\rm G}}\right)^{1/2} \, {\rm s}.
\label{l116'}
\end{eqnarray}
For fast rotation ($P \ll P_{\rm cr}$), the particle energy significantly increases when
particles approach the fast magnetosonic surface, whereas for slowly rotating objects
($P \gg P_{\rm cr}$ the particle energy remains practically unchanged. Their further fate,
as we have shown above, depends on whether the flow intersects the light surface or not.
\subsection{Young stars}
In conclusion, we discuss the main parameters which characterize nonrelativistic jets from
young stars. It should be recalled that in this case the nonrelativistic magnetization
parameter an $\sigma_{\rm n} \approx (\Omega_{\rm F}/\Omega_{\rm cr})^2$ (53) plays the key
role. Under the condition $\sigma_{\rm n} \gg 1$ ($\Omega_{\rm F} \gg \Omega_{\rm cr}$), the
electromagnetic energy flux near the central engine will be much greater than the particle
energy flux; beyond singular surfaces, as stressed above, particles must carry a significant
fraction of the total energy.
As the critical period $P_{\rm cr} = 2 \pi/\Omega_{\rm cr}$, namely
\begin{eqnarray}
P_{\rm cr} \approx 10
\left(\frac{R}{10^{11}{\rm cm}}\right)^{2}
\left(\frac{B_0}{10^{3}\,{\rm G}}\right)
\left(\frac{r_{\rm d}/R}{30}\right)^{-1}
\nonumber \\
\left(\frac{v_{\rm in}}{100 \,{\rm km \, s^{-1}}}\right)^{-3/2}
\left(\frac{\dot M}{10^{-9}M_{\odot} \, {\rm yr}^{-1}}\right)^{-1/2}
\, {\rm d}
\label{pcr}
\end{eqnarray}
($r_{\rm d}$ is the inner radius of the accretion disc) is close to the spin periods
of young stars, in the region of magnetic field lines coming out of the surface of
the star we have $\sigma_{\rm n} \sim 1$. On the other hand, the period $P_{\rm cr}$
is one two orders of magnitude larger than the rotation periods in the inner regions of
accretion discs, $P_{\rm b} = 2 \pi (GM/r^3)^{1/2}$, so that for the corresponding magnetic
field lines one finds $\sigma_{\rm n} \sim 10$--$1000$. That is why, the inner parts of the
accretion disc, not the central star, must play the role of the central engine rotor.
As noted above, there are observational evidences of this being the case [76].
Next, from relation (66) we obtain
\begin{equation}
r_{\rm A} \sim r_{\rm F} \sim \frac{v_{\rm in}}{\Omega} \, \sigma_{\rm n}^{1/3},
\label{aqa}
\end{equation}
so that $r_{\rm F} \sim 10$--$30 \, R$. Thus, the distance to singular surfaces exceeds
the size of the star by $10$--$30$ times, but it is the same $10$--$30$ times smaller
than the transverse size of jets. Consequently, the flow inside the jets must be supersonic
and, hence, the longitudinal current for these objects again must be derived from the
critical conditions on the singular surfaces. Correspondingly, the radius of the jet core
must be on the order of $r_{\rm core} = v_{\rm in}/{\Omega} \sim 0.1$ a.u., the jet magnetic
field should be as high as
$B_{\rm core} \sim 0.1 (\Omega r_{\rm core}/v_{\rm in})^2 B_{\rm in} \sim 10^{-2}$ G, and the
particle number density must range from $10^{8}$ to \mbox{$10^{9}$ cm$^{-3}$.}
Finally, we note that the existence of the integrals of motion allows us to obtain direct
information about the plasma outflow region. For example, if the radial and longitudinal
velocities of the flow in the jet are known at the axial distance $r_{\perp}$ (and such
observations, as noted above, have already been performed for several young stars [72, 73]),
it is possible to estimate the distance $r_{\rm st}$ from the central star at which the
corresponding force line is 'anchored' in the accretion disc [211, 212]:
\begin{eqnarray}
r_{\rm st} \approx 0.7 \, \left(\frac{r_{\perp}}{10 \, {\rm a.u.}}\right)^{2/3}
\left(\frac{{v}_{\varphi}}{10 \, {\rm} \, {\rm km} \, {\rm s}^{-1}}\right)^{2/3} \nonumber \\
\left(\frac{{v}_{\rm p}}{100 \, {\rm km} \, {\rm s}^{-1}}\right)^{-4/3}
\left(\frac{M}{M_{\odot}}\right)^{1/3} \, {\rm a.u.}
\end{eqnarray}
It is obvious that this scale is much larger than the size of the star, so at present it has
been possible to resolve only outermost regions of the outflow.
\section{Conclusions}
Thus, significant progress has indeed been recently achieved in the understanding
of the nature of jets observed in different classes of astrophysical sources. This
became possible because the analytical approach allowed sufficiently simple relationships
between physical parameters characterizing the outflows to be found, and the numerical
modeling (in which, we should keep in mind, the setup of the problem itself has been
significantly different from the analysis of stationary equations) has confirmed these
relationships.
The most important result of the analytical theory includes the understanding of the role
of key dimensionless parameters. For clearness, they are listed in Table 2 (for active
galactic nuclei we set $M = 10^9 \, M_{\odot}$). It turned out that the knowledge of these
parameters allows us to estimate many characteristics of jets, including the fraction of
energy carried by particles, the plasma Lorentz factor, the electron positron pair injection
rate ${\dot N}$, and the compactness parameter $l_{a}$, as well as to determine the main
parameters of the internal structure of jets. The determination of these parameters from
observations would be a significant breakthrough in our understanding of physical processes
which are underway in active astrophysical sources.
Next, we have shown that many properties of relativistic and nonrelativistic jets are
significantly different from each other. For convenience, we also collect them together
in Table 3. As can be seen, one should be very cautious when applying the results, which
were obtained for nonrelativistic jets, to ultrarelativistic flows. Moreover, the asymptotic
relations formulated above clarify many results obtained by means of numerical simulations.
The limited space of the present review did not allow us to discuss in detail many
important issues. In particular, we only briefly discussed the stability of jets.
Finally, here we have no space at all to discuss the nonstationary performance of
the central engine (most papers have recently started focusing exactly on this topic)
or the proper radiation of jets. Nevertheless, we would like to hope that questions
addressed in this review will be useful for future studies of relativistic and
nonrelativistic outflows observed in many astrophysical objects.
\begin{table}[ht]
\caption{
Parameters of jets outflowing from relativistic compact objects. All values are given to
an order of magnitude.
}
\centering
\begin{tabular}{|l|c|c|c|c|c|}
\hline
& AGN$_1$ & AGN$_2$ & $\mu$QSO & GRB & PSR \\
\hline
$\sigma$ & $100$ & $10^{12}$ & $10$ & $10^3$ & $10^{4}$ \\
$\lambda$ & $10^{12}$ & $100$ & $10^{10}$ & $10^{14}$ & $10^{3}$ \\
$\gamma_{\rm in}$ & $10$ & $10$ & $10$ & $10$ & $100$ \\
\hline
$l_{a}$ & $1$--$100$ & $1$--$100$ & $10^{4}$ & $10^{15}$ & $10^{-5}$ \\
$W_{\rm part}/W_{\rm tot}$ & $1$ & $10^{-9}$ & $ 10^{-5}$ & $1$ & $10^{-2}$ \\
$\gamma$ & $10$--$100$ & $10^4$--$10^5$ & $10^3$ & $300$ & $10^3$\\
${\dot N}$, s$^{-1}$ & $10^{49}$ & $10^{39}$ & $10^{33}$ & $10^{53}$ & $10^{32}$ \\
\hline
$B_{\rm min}$, G & $10^{-2}$ & $10^{-12}$ & $10^{4}$ & $10^{8}$ & $10^{-6}$ \\
$B_{\rm eq}$, G & $10^{-4}$ & $10^{-24}$ & $10^{3}$ & $10^{6}$ & $10^{-8}$ \\
$B_{\rm cr}$, G & $10^{-1}$ & $10^{-11}$ & $10^{6}$ & $10^{10}$ & $10^{-2}$ \\
\hline
\end{tabular}
\label{table5_01}
\end{table}
\begin{table*}[ht]
\caption{Main differences between relativistic and nonrelativistic jets.}
\vspace{0.3cm}
\centering
\begin{tabular}{|l|l|}
\hline
{\bf Relativistic flow} & {\bf Nonrelativistic flow} \\
\hline
Longitudinal current is close to the Goldreich current & Longitudinal current is much larger than the\\
& Goldreich current \\
\hline
For strongly magnetized flow ($\sigma \gg 1$), fast & Fast magnetosonic surface is located near Alfv\'enic \\
magnetosonic surface near equatorial plane is located & surface \\
$\sigma^{1/3}$ times farther than Alfv\'enic surface& \\
\hline
On fast magnetosonic surface, the particle energy flux & On fast magnetosonic surface, the particle energy flux\\
is much smaller than the electromagnetic energy flux & is close to the electromagnetic energy flux \\
\hline
Beyond Alfv\'enic surface, the electric field is close & Beyond Alfv\'enic surface, the electric field is much\\
in magnitude to the magnetic field & smaller than the magnetic field \\
\hline
Proper collimation is impossible for both strongly and & Proper collimation becomes effective for strongly \\
weakly magnetized outflows & magnetized outflows\\
\hline
The dense core in jets can be formed only under & The dense core is always formed in jets. Magnetic flux \\
sufficiently small ambient pressures. Magnetic flux & in the core is a significant fraction of the total \\
in the core is much smaller than the total flux & magnetic flux \\
\hline
Cylindrical flow with subsonic core is possible & Cylindrical flow with subsonic core is impossible \\
\hline
\end{tabular}
\label{table5_02}
\end{table*}
\noindent
{\bf Acknowledgments}
\vspace{0.3cm}
I would like to acknowledge M Barkov, S Komissarov, M Romanova, and A D Tchekhovskoy
for carefully reading the manuscript and many notes, and also A V Gurevich, Ya N Istomin,
Yu Yu Kovalev, R Lovelace, Yu Lyubarsky, and A Spitkovsky for useful discussions. The work
was supported by RFBR grant 09-02-00749 and the Federal Agency of Science and Innovations
(contract No. 02.740.11.0250).
\section{Appendix}
In the Appendix, we give for reference the complete set of equations of the Grad-Shafranov
method written out in themost general case, i.e., for axisymmetric stationary flows in the
vicinity of a rotating black hole. First of all, we keep in mind the basic relations for
the Kerr metric of a rotating black hole. In the Boyer-Lindquist coordinates $t$, $r$,
$\theta$, and $\varphi$, it assumes the form
$$
{\rm d}s^{2}=-\alpha^{2}{\rm d}t^{2}
+g_{ik}({\rm d}x^{i}+\beta^{i}{\rm d}t)({\rm d}x^{k}+\beta^{k}{\rm d}t),
\label{c1}
\eqno{(A.1)}
$$
where the quantity
$$
\alpha=\frac{\rho}{\Sigma}\sqrt{\Delta}
\label{c2}
\eqno{(A.2)}
$$
is the gravitational redshift, and the vector ${\bf \beta}$ is toroidal:
$$
\beta^{r} = \beta^{\theta} = 0, \qquad
\beta^{\varphi}=-\omega.
\label{c3}
\eqno{(A.3)}
$$
Here
$$
\omega = \frac{2aMr}{\Sigma^{2}}
\label{c2**'}
\eqno{(A.4)}
$$
is the so-called Lense-Thirring angular velocity. Finally, $M$ and $a$ are the mass and
specific angular momentum of the black hole ($a = J/M$), respectively. In addition, we
introduced the standard notations
$$
\Delta=r^{2}+a^{2}-2Mr, \qquad
\rho^{2}=r^{2}+a^{2}\cos^{2}\theta, \nonumber\\
$$
$$
\Sigma^{2}=(r^{2}+a^{2})^{2}-a^{2}\Delta\sin^{2}\theta, \qquad
\varpi=\frac{\Sigma}{\rho}\sin\theta.
\label{c5}
\eqno{(A.5)}
$$
Here, in all relativistic expressions we use the units in which $c = G = 1$. Finally, it is
important that the three-dimensional metric $g_{ik}$ in formula ($A.1$) be diagonal:
$$
g_{rr}=\frac{\rho^{2}}{\Delta},\qquad
g_{\theta \theta}=\rho^{2}, \qquad
g_{\varphi \varphi}=\varpi^{2}.
\label{c4}
\eqno{(A.6)}
$$
As for the flat space limit, it can be easily obtained by passing to the limit
$\alpha \rightarrow 1$ and $\omega \rightarrow 0$.
As is well known, for calculations it is convenient to introduce a special reference
frame, the so-called ZAMO (Zero Angular Momentum Observers) [131], which has the
following properties:
\begin{itemize}
\item
ZAMO observers are located at constant radius \mbox{$r =$ const,} $\theta =$ const but rotate
with the Lense Thirring angular velocity ${\rm d}\varphi/{\rm d}t=\omega$;
\item
for ZAMO, the four-dimensional metric $g_{\alpha\beta}$ is diagonal, with its three-dimensional
part $g_{ik}$ coinciding with Eqn ($A.6$).
\end{itemize}
Below, all vectors will be written out in this reference frame. In particular,
the operator $\nabla_{i}$, means the covariant derivative in the three-dimensional
metric ($A.6$).
As a result, the electric and magnetic fields can be conveniently written as
$$
{\bf B} = \frac{{\bf\nabla}\Psi \times {\bf e}_{\hat \varphi}}{2\pi\varpi}
-\frac{2I}{\alpha\varpi}{\bf e}_{\hat \varphi},
\label{k21'P}
\eqno{(A.7)}
$$
$$
{\bf E} = -\frac{\Omega_{\rm F}-\omega}{2\pi\alpha}{\bf\nabla}\Psi,
\label{k24''P}
\eqno{(A.8)}
$$
respectively, and the four-velocity of matter is written as
$$
{\bf u} = \frac{\eta}{\alpha n}{\bf B} + \gamma(\Omega_{\rm F}-\omega)\frac{\varpi}
{\alpha}{\bf e}_{\hat\varphi},
\label{p3'}
\eqno{(A.9)}
$$
where $\gamma = 1/\sqrt{1-v^2}$ is the Lorentz factor of matter, and the subscripts with a
cap over them correspond to physical components of vectors. The quantity $\Omega_{\rm F}$
remains the integral of motion. In turn, integrals of motions $E$ and $L$ will be written
out as
$$
E=E(\Psi) = \frac{\Omega_{\rm F}I}{2\pi}
+ \mu\eta(\alpha\gamma + \omega \varpi u_{\hat\varphi}),
\eqno{(A.10)}
\label{p31P}
$$
$$
L=L(\Psi) = \frac{I}{2\pi} + \mu\eta \varpi u_{\hat\varphi}.
\label{p32P}
\eqno{(A.11)}
$$
Next, the relativistic Bernoulli equation $\gamma^{2} - u_{\hat\varphi}^{2} = u_{\rm p}^2 +1$
takes the form
$$
\frac{K}{\varpi^{2}A^{2}}=\frac{1}{64\pi^{4}}\frac{{\cal M}^{4}({\bf\nabla}
\Psi)^{2}}{\varpi^{2}}+\alpha^{2}\eta^{2}\mu^{2},
\label{p38}
\eqno{(A.12)}
$$
where the Alfv\'en factor is
$$
A=\alpha^{2}-(\Omega_{\rm F}-\omega)^{2}\varpi^{2}-{\cal M}^{2},
\label{p39app}
\eqno{(A.13)}
$$
and
$$
K=\alpha^{2}\varpi^{2}(E-\Omega_{\rm F}L)^{2}\left[\alpha^{2}-(\Omega_{\rm F}-
\omega)^{2}\varpi^{2}-2{\cal M}^{2}\right] \nonumber
$$
$$
+{\cal M}^{4}\left[\varpi^{2}(E-\omega L)^{2}-\alpha^{2}L^{2}\right].
\eqno{(A.14)}
\label{p40}
$$
This equation defines the Alfv\'enic Mach number {\cal M}, where
$$
{\cal M}^{2}=\frac{4\pi\eta^{2}\mu}{n}.
\label{p36}
\eqno{(A.15)}
$$
Now, making use of relations ($A. 12$)-($A. 14$), which can be recast in the form
$({\bf\nabla}\Psi)^{2}=F({\cal M}^{2},E,L,\eta,\Omega_{\rm F},\mu)$,
where
$$
F=\frac{64\pi^{4}}{{\cal M}^{4}}\frac{K}{A^{2}}-\frac{64\pi^{4}}{{\cal M}^{4}}\alpha^{2}
\varpi^{2}\eta^{2}\mu^{2},
\label{p45}
\eqno{(A.16)}
$$
we obtain
$$
\nabla_{k}{\cal M}^{2}=\frac{N_k}{D},
\label{p46}
\eqno{(A.17)}
$$
where
$$
N_k = -\frac{A}{({\bf\nabla}\Psi)^{2}}\nabla^{i}\Psi \cdot \nabla_{i}
\nabla_{k}\Psi+\frac{A}{2}\frac{\nabla_{k}'F}{({\bf\nabla}\Psi)^{2}}.
\label{p47}
\eqno{(A.18)}
$$
Here, the operator $\nabla'_k$ acts on all quantities but ${\cal M}^{2}$. The
quantity $\nabla'_k\mu$ must be determined from the relation [97]
$$
\nabla'_k\mu
=\frac{2 c^{2}_{\rm s}}{1-c^{2}_{\rm s}}\mu \, \frac{\nabla_k\eta}{\eta}
+\frac{1}{1-c^{2}_{\rm s}}\left[\frac{1}{n}\left(\frac{\partial
P}{\partial s}\right)_{n}+T\right]\nabla_k s,
\label{p49}
\eqno{(A.19)}
$$
where $c^{2}_{\rm s} = 1/\mu(\partial P/\partial n)_{s}$ is the speed of sound, and $s$
is the entropy. In turn, the denominator $D$ can be rewritten in the form
$$
D=\frac{A}{{\cal M}^{2}}+\frac{\alpha^{2}}{{\cal M}^{2}}\frac{B^{2}_{\hat\varphi}}{B^{2}_
{\rm p}}-\frac{1}{u^{2}_{\rm p}}\frac{A}{{\cal M}^{2}}
\frac{c^{2}_{\rm s}}{1-c^{2}_{\rm s}}.
\label{p48}
\eqno{(A.20)}
$$
As for the Grad-Shafranov equation, in the compact form it can be written out as [96, 97]
$$
\frac{1}{\alpha}\nabla_{k}\left\{\frac{1}{\alpha\varpi^2}
[\alpha^{2}-(\Omega_{\rm F}-\omega)^{2}\varpi^{2}-{\cal M}^{2}]
\nabla^{k}\Psi\right\} \nonumber
$$
$$
+\frac{\Omega_{\rm F}
-\omega}{\alpha^{2}}({\bf\nabla}\Psi)^{2}\frac{{\rm d}
\Omega_{\rm F}}{{\rm d}\Psi}
\nonumber \\
\eqno{(A.21)}
\label{p64}
$$
$$
+\frac{64\pi^{4}}{\alpha^{2}\varpi^{2}}\frac{1}{2{\cal M}^{2}}
\frac{\partial}{\partial\Psi}\left(\frac{G}{A}\right)
-16\pi^{3}\mu n\frac{1}{\eta}\frac{{\rm d}\eta}{{\rm d}\Psi}
-16\pi^{3}nT\frac{{\rm d}s}{{\rm d}\Psi}=0, \nonumber
$$
where
$$
G=\alpha^{2}\varpi^{2}(E-\Omega_{\rm F}L)^{2}+\alpha^{2}{\cal M}^{2}L^{2}
-{\cal M}^{2} \varpi^{2}(E-\omega L)^{2}.
\label{p65}
\eqno{(A.22)}
$$
Now, expanding terms $\nabla_{k}{\cal M}^{2}$ in Eqn ($A.21$) according to definitions
($A.17$)--($A.19$), we finally arrive at
$$
A\left[\frac{1}{\alpha}\nabla_{k}\left(\frac{1}{\alpha\varpi^{2}}
\nabla^{k}\Psi\right)+\frac{1}{\alpha^{2}\varpi^{2}({\bf\nabla}\Psi)^{2}}
\frac{\nabla^{i} \Psi \cdot \nabla^{k} \Psi \cdot \nabla_{i}\nabla_{k}\Psi}{D}\right]
\nonumber
$$
$$
+\frac{1}{\alpha^{2}\varpi^{2}}\nabla_{k}'A \cdot \nabla^{k}\Psi
-\frac{A}{\alpha^{2}\varpi^{2}({\bf\nabla}\Psi)^{2}}
\frac{1}{2D}\nabla_{k}'F \cdot \nabla^{k}\Psi
\nonumber \\
\nonumber
$$
$$
+\frac{\Omega_{\rm F}-\omega}{\alpha^{2}}
({\bf\nabla}\Psi)^{2} \frac{{\rm d}\Omega_{\rm F}}{{\rm d}\Psi}
+\frac{64\pi^{4}}{\alpha^{2}\varpi^{2}}\frac{1}{2{\cal M}^{2}}
\frac{\partial}{\partial\Psi}\left(\frac{G}{A}\right)
\nonumber \\
$$
$$
-16\pi^{3}\mu n\frac{1}{\eta}
\frac{{\rm d}\eta}{{\rm d}\Psi}-16\pi^{3}nT\frac{{\rm d}s}{{\rm d}\Psi}=0,
\label{p66}
\eqno{(A.23)}
$$
where again the gradient $\nabla_{k}'$ acts on all quantities but
${\cal M}^2$, and the derivative $\partial/\partial\Psi$
acts only onh the integrals of motion. Formula ($A27$) determines
in the most general form the equilibrium equation for the magnetic
surfaces. Finally algebraic relations have the form
$$
\frac{I}{2\pi} = \frac{\alpha^{2}L-(\Omega_{\rm F}-\omega)\varpi^{2}
(E-\omega L)}{\alpha^{2}-(\Omega_{\rm F}-\omega)^{2}\varpi^{2}-{\cal M}^{2}},
\label{p33} \\
\eqno{(A.24)}
$$
$$
\gamma = \frac{1}{\alpha\mu\eta}\frac{\alpha^{2}(E-\Omega_{\rm F}L)-{\cal M}^{2}
(E-\omega L)}{\alpha^{2}-(\Omega_{\rm F}-\omega)^{2}\varpi^{2}-{\cal M}^{2}},
\label{p34}
\eqno{(A.25)}
$$
$$
u_{\hat\varphi} = \frac{1}{\varpi\mu\eta}\frac{(E-\Omega_{\rm F}L)
(\Omega_{\rm F}-\omega)\varpi^{2}-L{\cal M}^{2}}{\alpha^{2}-(\Omega_{\rm F}
-\omega)^{2}\varpi^{2}-{\cal M}^{2}}.
\label{p35}
\eqno{(A.26)}
$$
Equations ($A.12$) and ($A.24$)--($A.26$) represent algebraic bounds which allow the
determination, albeit in an indirect form, of all characteristics of the flow from the
given poloidal field ${\bf B}_{\rm p}$ (i.e., from the known potential $\Psi$) and five
integrals of motion. It should be emphasized that for a nonzero temperature they are
extremely lengthy, mainly due to the need to resolve equation ($A.19$). In the case of
cold plasma ($s$ = 0, i.e., \mbox{$\mu = $ const),} Bernoulli equation ($A.12$) becomes a
fourth-order algebraic equation with respect to ${\cal M}^2$. As shown above, this fact
often allows analytical asymptotics to be found.
In the cylindrical case, the second-order Grad-Shafranov equation can be conveniently
reduced to the system of two ordinary differential equations of the first order for
the magnetic flux $\Psi(\varpi)$ and the Mach number ${\cal M}^2$. The equation for
the Mach number has therewith the form [146]
$$
\left[\frac{(e')^2}{\mu^2\eta^2}-1+\frac{\Omega_{\rm F}^2 r^2}{c^2}
-A\frac{c_{\rm s}^2}{c^2}\right]
\frac{{\rm d}{\cal M}^2}{{\rm d} r } =
\frac{{\cal M}^6L^2}{A r ^3 \mu^2\eta^2c^2}
$$
$$
+\frac{\Omega_{\rm F}^2 r {\cal M}^2}{c^{2}}\left[2 - \frac{(e')^2}{A\mu^2\eta^2c^4}\right]
+{\cal M}^2 \frac{e'}{\mu^2\eta^2c^4}\frac{{\rm d}\Psi}{{\rm d}r}\frac{{\rm d}e'}{{\rm d}\Psi}
$$
$$
+{\cal M}^2\frac{ r ^2}{c^2}\Omega_{\rm F}\frac{{\rm d}\Psi}{{\rm d}r}
\frac{{\rm d}\Omega_{\rm F}}{{\rm d}\Psi}
-{\cal M}^2 \left(1-\frac{\Omega_{\rm F}^2 r ^2}{c^2} + 2A\frac{c_{\rm s}^2}{c^2}\right)
\frac{{\rm d}\Psi}{{\rm d} r }\frac{1}{\eta}\frac{{\rm d}\eta}{{\rm d}\Psi}
$$
$$
-\left[\frac{A}{n}\left(\frac{\partial P}{\partial s}\right)_n
+\left(1-\frac{\Omega_{\rm F}^2 r ^2}{c^2}\right)T\right]
\frac{{\cal M}^2}{\mu}\frac{{\rm d}\Psi}{{\rm d}r}
\frac{{\rm d}s}{{\rm d}\Psi},
\label{p36new1}
\eqno{(A.27)}
$$
where $e' = E - \Omega_{\rm F}L$. The equation for the magnetic flux
$\Psi$ will coincide with Bernoulli equation ($A.14$).
Finally, the force-free pulsar equation takes on the form
$$
-\left(1-\frac{\Omega_{\rm F}^2\varpi^2}{c^2}\right)\nabla^2\Psi
+\frac{2}{\varpi}\frac{\partial \Psi}{\partial\varpi}
$$
$$
-\frac{16\pi^{2}}{c^2}I\frac{{\rm d}I}{{\rm d}\Psi}
+\frac{\varpi^{2}}{c^2}\left(\nabla\Psi\right)^{2}
\Omega_{\rm F}\frac{{\rm d}\Omega_{\rm F}}{{\rm d}\Psi}=0,
\label{d39}
\eqno{(A.28)}
$$
where $\nabla^2$ is the Laplace operator. Its generalization to the force-free
black hole magnetosphere is written as [131]
$$
\frac{1}{\alpha}{\bf\nabla}_k\left\{\frac{\alpha}{\varpi^{2}}
\left[1-\frac{(\Omega_{\rm F}-\omega)^{2}\varpi^{2}}{\alpha^{2}}\right]
{\bf\nabla}^k\Psi \right\}
$$
$$
+\frac{\Omega_{\rm F}-\omega}{\alpha^{2}}
({\bf\nabla}\Psi)^{2} \frac{{\rm d} \Omega_{\rm F}} {{\rm d}\Psi}
+\frac{16\pi^{2}}{\alpha^{2} \varpi^{2}} I\frac{{\rm d}I}{{\rm d}\Psi} = 0.
\eqno{(A.29)}
$$
These equations are elliptical in all the space, and so they require boundary conditions on the
integration region boundary or on the black hole horizon.
\newpage
\noindent
{\bf References}
\vspace{0.3cm}
{\small
\noindent
1. Kardashev N S {\it Astron. Zh.} {\bf 41} 807 (1964) [ {\it Sov. Astron.} {\bf 8} 643 (1965)]
\noindent
2. Pacini V {\it Nature} {\bf 216} 567 (1967)
\noindent
3. Goldreich P, Julian W H {\it Astrophys. J.} {\bf 157} 869 (1969)
\noindent
4. Michel F C {\it Astrophys. J.} {\bf 158} 727 (1969)\
\noindent
5. Blandford R D {\it Mon. Not. R. Astron. Soc.} {\bf 176} 465 (1976)
\noindent
6. Lovelace R W E{\it Nature} {\bf 262} 649 (1976)
\noindent
7. Bisnovatyi-Kogan G S, Popov Yu P, Samokhin A A {\it Astrophys. Space Sci.} {\bf 41} 321 (1976)
\noindent
8. Moiseenko S G, Bisnovatyi-Kogan G S, Ardeljan N V {\it Astrophys. J.} {\bf 370} 501 (2006)
\noindent
9. Beskin V S {\it Osesimmetrichnye Statsionamye Techeniya v Astrofizike} ({\it MHD Flows in Compact Astrophysical Objects}) (Moscow: Fizmatlit, 2005) [Translated into English (Berlin: Springer, 2009)]
\noindent
10. Beskin V S {\it Usp. Fiz. Nauk} {\bf 167} 689 (1997) [Phys. Usp. {\bf 40} 659 (1997)]
\noindent
11. Shapiro S, Teukolsky S {\it Black Holes, White Dwarfs, and Neutron Stars} (New York: Wiley, 1983) [Translated into Russian (Moscow: Mir, 1985)]
\noindent
12. Begelman M C, Blandford R D, Rees M J Rev. Mod. Phys. {\bf 56} 255 (1984) [Translated into Russian: in {\it Fizika Vnegalakticheskikh Istochnikov Radioizlucheniya} (Ed. R D Dagkesamanskii) (Moscow: Mir, 1987) p. 1
\noindent
13. Rees M J Annu. Rev. Astron. Astrophys. {\bf 22} 471 (1984)
\noindent
14. Zeldovich Ya B, Novikov I D {\it Relyativistskaya Astrofizika} ({\it Relativistic Astrophysics}) (Moscow: Nauka, 1967) [Translated into English (Chicago: Univ. of Chicago Press, 1971,1983)]
\noindent
15. Lynden-Bell D {\it Nature} {\bf 223} 690 (1969)
\noindent
16. Bisnovatyi-Kogan G S, Ruzmaikin A A {\it Astrophys. Space Sci.} {\bf 28} 45 (1974)
\noindent
17. Bisnovatyi-Kogan G S, Ruzmaikin A A {\it Astrophys. Space Sci.} {\bf 42} 401(1976)
\noindent
18. Novikov I D, Frolov V P {\it Fizika Chernykh Dyr} ({\it Physics of Black Holes}) (Moscow: Nauka, 1986) [Translated into English (Dordrecht: Kluwer Acad., 1989)]
\noindent
19. Junor W, Biretta J A, Livio M {\it Nature} {\bf 401} 891 (1999)
\noindent
20. Perley R A, Dreher J W, Cowan J J {\it Astrophys. J.} {\bf 285} L35 (1984)
\noindent
21. Lobanov A, Hardee P, Eilek J {\it New Astron. Rev.} {\bf 47} 629 (2003)
\noindent
22. Kovalev Y Y et al. {\it Astrophys. J.} {\bf 668} L27 (2007)
\noindent
23. Junor W, Biretta J A Astron. J. {\bf 109} 500 (1995)
\noindent
24. Reynolds C S et al. {\it Astrophys. J.} {\bf 283} 873 (1996)
\noindent
25. Sikora M, Madejski G {\it Astrophys. J.} {\bf 534} 109 (2000)
\noindent
26. Blandford R D, Rees M J {\it Astrophys. J.} {\bf 169} 395 (1974)
\noindent
27. Fabian A C, Rees M J {\it Astrophys. J.} {\bf 277} L55 (1995)
\noindent
28. Feretti L et al. Astron. Astrophys. {\bf 298} 699 (1995)
\noindent
29. Cheng A Y S, O'Dell S L {\it Astrophys. J.} {\bf 251} L49 (1981)
\noindent
30. Proga D, Stone J M, Kallman T R {\it Astrophys. J.} {\bf 543} 686 (2000)
\noindent
31. Ghisellini G et al. {\it Astrophys. J.} {\bf 362} L1 (1990)
\noindent
32. Konigl A, Kartje J F {\it Astrophys. J.} {\bf 434} 446 (1994)
\noindent
33. Blandford R D, Znajek R L {\it Astrophys. J.} {\bf 179} 433 (1977)
\noindent
34. Beskin V S, Istomin Y N, Parev V I {\it Astron. Zh.} {\bf 69} 1258 (1992) [{\it Sov. Astron.} {\bf 36} 642 (1992)]
\noindent
35. Hirotani K, Okamoto I {\it Astrophys. J.} 497 563 (1998)
\noindent
36. Fender R P, in {\it Compact Stellar X-ray Sources} (Eds W Lewin, M van der Klis) (Cambridge: Cambridge Univ. Press, 2006) p. 381
\noindent
37. Spencer R E {\it Nature} {\bf 282} 483 (1979)
\noindent
38. Mirabel I F, Rodriguez L F {\it Nature} {\bf 371} 46 (1994)
\noindent
39. Akerlof C et al. {\it Nature} {\bf 398} 400 (1999)
\noindent
40. Postnov K A {\it Usp. Fiz. Nauk} {\bf 169} 545 (1999) [{\it Phys. Usp.} {\bf 42} 469 (1999)]
\noindent
41. Ruderman M {\it Ann. New York Acad. Sci.} {\bf 262} 164 (1975)
\noindent
42. Panaitescu A, Kumar P {\it Astrophys. J.} {\bf 571} 779 (2002)
\noindent
43. Blinnikov S I et al. {\it Pis'ma Astron. Zh.} {\bf 10} 422 (1984) [{\it Sov. Astron.} Lett. {\bf 10} 177 (1984)]
\noindent
44. Eichler D et al. {\it Nature} {\bf 340} 126 (1989)
\noindent
45. Paczynski B {\it Acta Astron.} {\bf 41} 257 (1991)
\noindent
46. Woosley S E {\it Astrophys. J.} {\bf 405} 273 (1993)
\noindent
47. Paczynski B {\it Astrophys. J.} {\bf 494} L45 (1998)
\noindent
48. Meszaros P, Rees M J {\it Astrophys. J.} {\bf 482} L29 (1997)
\noindent
49. van Putten M H P M, Levinson A {\it Astrophys. J.} {\bf 584} 937 (2003)
\noindent
50. Komissarov S S, Barkov M V {\it Astrophys. J.} {\bf 382} 1029 (2007)
\noindent
51. Usov V V {\it Nature} {\bf 357} 472 (1992)
\noindent
52. Thompson C, Duncan R C {\it Astrophys. J.} {\bf 275} 255 (1995)
\noindent
53. Hewish A et al. {\it Nature} {\bf 217} 708 (1968)
\noindent
54. Baade W, Zwicky F {\it Proc. Natl. Acad. Sci.} {\bf 20} 254 (1934)
\noindent
55. Landau L D, Lifshitz E M {\it Teoriya Polya} ({\it The Classical Theory of Fields}) (Moscow: Nauka, 1973) [Translated into English (Oxford: Pergamon Press, 1975)]
\noindent
56. Beskin V S, Gurevich A V, Istomin Ya N {\it Physics of the Pulsar Magnetosphere} (Cambridge: Cambridge Univ. Press, 1993)
\noindent
57. Mestel L, Panagi P, Shibata S {\it Astrophys. J.} {\bf 309} 388 (1999)
\noindent
58. Weisskopf M C et al. {\it Astrophys. J.} {\bf 536} L81 (2000)
\noindent
59. Helfand D J, Gotthelf E V, Halpern J P {\it Astrophys. J.} {\bf 556} 380 (2001)
\noindent
60. Chiueh T, Li Z-Y, Begelman M C {\it Astrophys. J.} {\bf 505} 835 (1998)
\noindent
61. Lyubarsky Y, Kirk J G {\it Astrophys. J.} {\bf 547} 437 (2001)
\noindent
62. Petri J, Lyubarsky Y {\it Astron. Astrophys.} {\bf 473} 683 (2007)
\noindent
63. Kennel C F, Coroniti F V {\it Astrophys. J.} {\bf 283} 694 (1984)
\noindent
64. Komissarov S S, Lyubarsky Y E {\it Astrophys. J.} {\bf 349} 779 (2004)
\noindent
65. Bogovalov S V et al. {\it Astrophys. J.} {\bf 358} 705 (2005)
\noindent
66. Del Zanna L, Amato E, Bucciantini N {\it Astron. Astrophys.} {\bf 421} 1063 (2004)
\noindent
67. Djorgovsky S, Evans C R {\it Astrophys. J. Lett.} {\bf 335} L61 (1988)
\noindent
68. Herbig G H {\it Astrophys. J.} {\bf 111} 11 (1950)
\noindent
69. Haro G {\it Astron. J.} {\bf 55} 72 (1950)
\noindent
70. Reipurth B, Bally J {\it Annu. Rev. Astron. Astrophys.} {\bf 39} 403 (2001)
\noindent
71. Surdin V G {\it Rozhdenie Zvezd} ({\it The Birth of Stars}) (Moscow: URSS, 2001)
\noindent
72. Bacciotti F et al. {\it Astrophys. J.} {\bf 576} 222 (2002)
\noindent
73. Coffey D et al. {\it Astrophys. J.} {\bf 663} 350 (2007)
\noindent
74. Chrysostomou A, Lucas P W, Hough J H {\it Nature} {\bf 450} 71 (2007)
\noindent
75. Pelletier G, Pudritz R E {\it Astrophys. J.} {\bf 394} 117 (1992)
\noindent
76. Edwards S {\it Proc. Int. Astron. Union} {\bf 3} 171 (2007)
\noindent
77. Shu F H et al. {\it Astrophys. J.} {\bf 429} 797 (1994)
\noindent
78. Pudritz R E, Norman C A {\it Astrophys. J.} {\bf 301} 571 (1986)
\noindent
79. Brenneman L W, Reynolds C S {\it Astrophys. J.} {\bf 652} 1028 (2006)
\noindent
80. Daly R A {\it Astrophys. J. Lett.} {\bf 696} L32 (2009)
\noindent
81. Lyne A G, Graham-Smith F {\it Pulsar Astronomy} 2nd ed. (Cambridge: Cambridge Univ. Press, 1998)
\noindent
82. Bisnovatyi-Kogan G S, Lovelace R V E {\it New Astron. Rev.} {\bf} 45 663 (2001)
\noindent
83. Alfven H, Falthammar C-G {\it Cosmical Electrodynamics} (Oxford: Clarendon Press, 1963) [Translated into Russian (Moscow: Mir, 1967)]
\noindent
84. Weber E J, Davis L (Jr.) {\it Astrophys. J.} {\bf 148} 217 (1967)
\noindent
85. Solov'ev L S, in {\it Voprosy Teorii Plazmy} ({\it Reviews of Plasma Physics}) Vol. 3 (Ed. M A Leontovich) (Moscow: Gosatomizdat, 1963) p. 245 [Translated into English (New York: Consultants Bureau, 1967) p. 277
\noindent
86. Heinemann M, Olbert S J {\it Geophys. Res.} {\bf 83} 2457 (1978)
\noindent
87. Okamoto I {\it Astrophys. J.} {\bf 173} 357 (1975)
\noindent
88. Heyvaerts J, in {\it Plasma Astrophysics} (Lecture Notes Phys., Vol. 468, Eds C Chiuderi, G Einaudi) (Berlin: Springer-Verlag, 1996) p. 31
\noindent
89. Shafranov V D {\it Zh. Eksp. Tear. Fiz.} {\bf 33} 710 (1957) [{\it Sov. Phys. JETP} {\bf 6} 545 (1958)]
\noindent
90. Grad H {\it Rev. Mod. Phys.} {\bf 32} 830 (1960)
\noindent
91. Von Mises R {\it Mathematical Theory of Compressible Fluid Flow} (New York: Academic Press, 1958) [Translated into Russian (Moscow: IL, 1961)]
\noindent
92. Guderley K G {\it Theorie Schallnaher Strbmungen} ({\it Theory of Transonic Flow}) (Berlin: Springer-Verlag, 1957) [Translated into English (Oxford: Pergamon Press, 1962); Translated into Russian (Moscow: IL, I960)]
\noindent
93. Frankl F I {\it Izbrannye Trudy po Gazovoi Dinamike} ({\it Selected Papers on Gasdynamics}) (Moscow: Nauka, 1973)
\noindent
94. Ardavan H {\it Astrophys. J.} {\bf 204} 889 (1976)
\noindent
95. Lovelace R V E et al. {\it Astrophys. J. Suppl.} {\bf 62} 1 (1986)
\noindent
96. Nitta S, Takahashi M, Tomimatsu A {\it Phys. Rev. D} {\bf 44} 2295 (1991)
\noindent
97. Beskin V S, Par'ev V I {\it Usp. Fiz. Nauk} {\bf 163} 95 (1993) [{\it Phys. Usp.} {\bf} 36 529 (1993)]
\noindent
98. Contopoulos I, Kazanas D, Fendt C {\it Astrophys. J.} {\bf 511} 351 (1999)
\noindent
99. Bogovalov S V {\it Astron. Astrophys.} {\bf 371} 1155 (2001)
\noindent
100. Goodwin S P et al. {\it Astrophys. J.} {\bf 349} 213 (2004)
\noindent
101. Gruzinov A {\it Phys. Rev. Lett.} {\bf 94} 021101 (2005)
\noindent
102. Komissarov S S {\it Astrophys. J.} {\bf 367} 19 (2006)
\noindent
103. McKinney J C {\it Astrophys. J.} {\bf 368} L30 (2006)
\noindent
104. Timokhin A N {\it Astrophys. J.} {\bf 368} 1055 (2006)
\noindent
105. Svensson R {\it Astrophys. J.} {\bf 209} 175 (1984)
\noindent
106. Zakamska N L, Begelman M C, Blandford R D {\it Astrophys. J.} {\bf 679} 990 (2008)
\noindent
107. Beskin V S, Gurevich A V, Istomin Ya N {\it Zh. Eksp. Teor. Fiz.} {\bf 85} 401 (1983) [{\it Sov. Phys. JETP} {\bf 58} 235 (1983)]
\noindent
108. Beskin V S, Rafikov R R {\it Astrophys. J.} {\bf 313} 433 (2000)
\noindent
109. Tchekhovskoy A, McKinney J C, Narayan R {\it Astrophys. J.} {\bf 388} 551 (2008)
\noindent
110. Bogovalov S V {\it Pis'ma Astron. Zh.} {\bf 18} 832 (1992) [{\it Sov. Astron.} Lett. {\bf 18} 337 (1992)]
\noindent
111. Beskin V S, Okamoto I {\it Astrophys. J.} {\bf 313} 445 (2000)
\noindent
112. Toropina O D et al. {\it Mem. Soc. Astron. Ital.} {\bf 76} 508 (2005)
\noindent
113. Beskin V S, Kuznetsova I V {\it Nuovo Cimento B} {\bf 115} 795 (2000)
\noindent
114. Kazhdan Y M, Murzina M {\it Astrophys. J.} {\bf 270} 351 (1994)
\noindent
115. Uchida Y, Shibata K {\it Astron. Soc. Jpn.} {\bf 36} 105 (1984)
\noindent
116. Hawley J F, Smarr L L, Wilson J R {\it Astrophys. J.} {\bf 277} 296 (1984)
\noindent
117. Shima E et al. {\it Astrophys. J.} {\bf 217} 367 (1985)
\noindent
118. Petrich L I et al. {\it Astrophys. J.} {\bf 336} 313 (1989)
\noindent
119. Ustyugova G V et al. {\it Astrophys. J.} {\bf 439} L39 (1995)
\noindent
120. Bogovalov S, Tsinganos K {\it Astrophys. J.} {\bf 305} 211 (1999)
\noindent
121. Ustyugova G V et al. {\it Astrophys. J.} {\bf 541} L21 (2000)
\noindent
122. Komissarov S S {\it Astrophys. J.} {\bf 326} L41 (2001)
\noindent
123. Komissarov S S et al. {\it Astrophys. J.} {\bf 380} 51 (2007)
\noindent
124. Barkov M V, Komissarov S S {\it Astrophys. J.} {\bf 385} L28 (2008)
\noindent
125. Tchekhovskoy A, McKinney J C, Narayan R {\it Astrophys. J.} {\bf 699} 1789 (2009)
\noindent
126. Komissarov S S et al. {\it Astrophys. J.} {\bf 394} 1182 (2009)
\noindent
127. Tchekhovskoy A, Narayan R, McKinney J C {\it Astrophys. J.} {\bf 711} 50 (2010)
\noindent
128. Lynden-Bell D {\it Astrophys. J.} {\bf 279} 389 (1996)
\noindent
129. Lynden-Bell D {\it Astrophys. J.} {\bf 341} 1360 (2003)
\noindent
130. Landau L D, Lifshitz E M {\it Elektrodinamika Sploshnykh Sred} ({\it Electrodynamics of Continuous Media}) (Moscow: Nauka, 1982) [Translated into English (Oxford: Pergamon Press, 1984)]
\noindent
131. Thorne K S, Price R H, Macdonald D A {\it Black Holes: the Membrane Paradigm} (New Haven: Yale Univ. Press, 1986) [Translated into Russian (Moscow: Mir, 1988)]
\noindent
132. Punsly B {\it Black Hole Gravitohydromagnetics} (Berlin: Springer, 2001)
\noindent
133. Al'pert Ya L, Gurevich A V, Pitaevskii L P {\it Iskusstvennye Sputniki v Razrezhennoi Plazme} ({\it Space Physics with Artifical Satellites}) (Moscow: Nauka, 1964) [Translated into English (New York: Consultants Bureau, 1965)]
\noindent
134. Komissarov S S {\it Astrophys. J.} {\bf 350} 1431 (2004)
\noindent
135. Camenzind M {\it Compact Objects in Astrophysics} (Heidelberg: Springer, 2007)
\noindent
136. Okamoto I {\it Astrophys. J.} {\bf 307} 253 (1999)
\noindent
137. Heyvaerts J, Norman C {\it Astrophys. J.} {\bf 596} 1240 (2003)
\noindent
138. Heyvaerts J, Norman C {\it Astrophys. J.} {\bf 347} 1055 (1989)
\noindent
139. Bogovalov S V Pis'ma {\it Astron. Zh.} {\bf 21} 633 (1995) [{\it Astron. Lett.} {\bf 21} 565 (1995)]
\noindent
140. Bogovalov S V Pis'ma {\it Astron. Zh.} {\bf 24} 381 (1998) [{\it Astron. Lett.} {\bf 24} 321 (1998)]
\noindent
141. Sakurai T {\it Astron. Astrophys.} {\bf 152} 121 (1985)
\noindent
142. Michel F C {\it Astrophys. J.} {\bf 180} L133 (1973)
\noindent
143. Beskin V S, Kuznetsova IV, Rafikov R R {\it Astrophys. J.} {\bf 299} 341 (1998)
\noindent
144. Tomimatsu A {\it Publ. Astron. Soc. Jpn.} {\bf 46} 123 (1994)
\noindent
145. Beskin V S, Nokhrina E E {\it Astrophys. J.} {\bf 367} 375 (2006)
\noindent
146. Beskin V S, Nokhrina E E {\it Astrophys. J.} {\bf 397} 1486 (2009)
\noindent
147. Trubnikov B A {\it Teoriya Plazmy} ({\it Plasma Theory}) (Moscow: Energoatomizdat, 1996)
\noindent
148. Bogovalov S V {\it Astrophys. J.} {\bf 280} 39 (1996)
\noindent
149. Bogovalov S V Pis'ma {\it Astron. Zh.} {\bf 16} 844 (1990) [{\it Sov. Astron. Lett.} {\bf 16} 362 (1990)]
\noindent
150. Lyubarsky Yu {\it Astrophys. J.} {\bf 698} 1570 (2009)
\noindent
151. Beskin V S, Malyshkin L M {\it Pis'ma Astron. Zh.} {\bf 26} 253 (2000) [{\it Astron. Lett.} {\bf 26} 208 (2000)]
\noindent
152. Kadomtsev B B {\it Kollektivnye Yavleniya v Plazme} ({\it Cooperative Effects in Plasmas}) 2nd ed. (Moscow: Nauka, 1988) [Translateed into English, in Reviews of Plasma Physics Vol. 22 (Ed. V D Shafranov) (New York: Kluwer Acad./Plenum Publ., 2001) p. 1
\noindent
153. Ryutov D D, Derzon M S, Matzen M K {\it Rev. Mod. Phys.} {\bf 72} 167 (2000)
\noindent
154. Bisnovatyi-Kogan G S, Komberg B V, Fridman A M {\it Astron. Zh.} {\bf 46} 465 (1969) [{\it Sov. Astron.} {\bf 13} 369 (1969)]
\noindent
155. Benford G {\it Astrophys. J.} {\bf 247} 792 (1981)
\noindent
156. Hardee P E {\it Astrophys. J.} {\bf 313} 607 (1987)
\noindent
157. Hardee P E, Norman M L {\it Astrophys. J.} {\bf 334} 70 (1988)
\noindent
158. Appl S, Camenzind M {\it Astron. Astrophys.} {\bf 256} 354 (1992)
\noindent
159. Lyubarskii Yu E {\it Astrophys. J.} {\bf 308} 1006 (1999)
\noindent
160. Bisnovatyi-Kogan G {\it Astrophys. Space Sci.} {\bf 311} 287 (2007)
\noindent
161. Ciardi A et al. {\it Astrophys. J.} {\bf 691} L147 (2009)
\noindent
162. Bellan P M et al. {\it Phys. Plasmas} {\bf 16} 041005 (2009)
\noindent
163. Istomin Y N, Pariev VI {\it Astrophys. J.} {\bf 267} 629 (1994)
\noindent
164. Meliani Z, Keppens R {\it Astron. Astrophys.} {\bf 475} 785 (2007)
\noindent
165. Narayan R, Li J, Tchekhovskoy A {\it Astrophys. J.} {\bf 697} 1681 (2009)
\noindent
166. McKinney J C, Blandford R D {\it Astrophys. J.} {\bf 394} L126 (2009)
\noindent
167. Beskin V S, Zakamska N L, Sol H {\it Astrophys. J.} {\bf 347} 587 (2004)
\noindent
168. Narayan R, McKinney J C, Farmer A J {\it Astrophys. J.} {\bf 375} 548 (2006)
\noindent
169. Barkov M V, Komissarov S S {\it Int. J. Mod. Phys. D} {\bf 17} 1669 (2008)
\noindent
170. Bardou A, Heyvaerts J {\it Astron. Astrophys.} {\bf 307} 1009 (1996)
\noindent
171. Lovelace R V E, Romanova M M {\it Astrophys. J.} {\bf 596} L159 (2003)
\noindent
172. Michel F C {\it Astrophys. J.} {\bf 180} 207 (1973)
\noindent
173. Mestel L, Wang Y-M {\it Astrophys. J.} {\bf 188} 799 (1979)
\noindent
174. Lyubarskii Y E Pis'ma {\it Astron. Zh.} {\bf 16} 34 (1990) [{\it Sov. Astron. Lett.} {\bf 16} 16 (1990)]
\noindent
175. Lovelace R V E, Turner L, Romanova M M {\it Astrophys. J.} {\bf 652} 1494 (2006)
\noindent
176. Bogovalov S V {\it Astron. Astrophys.} {\bf 349} 1017 (1999)
\noindent
177. Spitkovsky A {\it Astrophys. J.} {\bf 648} L51 (2006)
\noindent
178. Ruderman M A, Sutherland P G {\it Astrophys. J.} {\bf 196} 51 (1975)
\noindent
179. Arons J {\it Astrophys. J.} {\bf 248} 1099 (1981)
\noindent
180. Gurevich A V, Istomin Ya N {\it Zh. Eksp. Teor. Fiz.} {\bf 89} 3 (1985) [{\it Sov. Phys.JETP} {\bf 62} 1(1985)]
\noindent
181. Bai X-N, Spitkovsky A {\it Astrophys. J.} {\bf 715} 1282 (2010)
\noindent
182. Kramer M et al. {\it Science} {\bf 312} 549 (2006)
\noindent
183. Beskin V S, Nokhrina E E {\it Astrophys. Space Sci.} {\bf 308} 569 (2007)
\noindent
184. Gurevich A V, Istomin Ya N {\it Astrophys. J.} {\bf 377} 1663 (2007)
\noindent
185. Kato Y, Mineshige S, Shibata K {\it Astrophys. J.} {\bf 605} 307 (2004)
\noindent
186. Sherwin B D, Lynden-Bell D {\it Astrophys. J.} {\bf 378} 409 (2007)
\noindent
187. Uzdensky D A, MacFadyen A I {\it Astrophys. J.} {\bf 669} 546 (2007)
\noindent
188. Romanova M M et al. {\it Astrophys. J.} {\bf 399} 1802 (2009)
\noindent
189. Lovelace R V E et al. {\it Astrophys. J.} {\bf 572} 445 (2002)
\noindent
190. Ghosh P, Abramowicz M A {\it Astrophys. J.} {\bf 292} 887 (1997)
\noindent
191. Livio M, Ogilvie G I, Pringle J E {\it Astrophys. J.} {\bf 512} 100 (1999)
\noindent
192. Konigl A {\it Astrophys. J.} {\bf 342} 208 (1989)
\noindent
193. Ferreira J, Pelletier G {\it Astron. Astrophys.} {\bf 295} 807 (1995)
\noindent
194. Krolik J H {\it Active Galactic Nuclei, from the Central Black Hole to the Galactic Environment} (Princeton, NJ: Princeton Univ. Press, 1999)
\noindent
195. Derishev E V et al. {\it Astrophys. Space Sci.} {\bf 297} 21 (2005)
\noindent
196. Stern B E, Poutanen J {\it Astrophys. J.} {\bf 383} 1695 (2008)
\noindent
197. Sol H, Pelletier G, Asseo E {\it Astrophys. J.} {\bf 237} 411 (1989)
\noindent
198. Cheng K S, Ho C, Ruderman M {\it Astrophys. J.} {\bf 300} 500 (1986)
\noindent
199. Zavala R T, Taylor G B {\it Astrophys. J.} {\bf 612} 749 (2004)
\noindent
200. Lister M L, Homan D C {\it Astron. J.} {\bf 130} 1389 (2005)
\noindent
201. Gabuzda D C et al. {\it Astrophys. J.} {\bf 384} 1003 (2008)
\noindent
202. Blandford R D, Konigl A {\it Astrophys. J.} {\bf 232} 34 (1979)
\noindent
203. Lobanov A P {\it Astron. Astrophys.} {\bf 330} 79 (1998)
\noindent
204. Churazov E M et al. {\it Usp. Fiz. Nauk} {\bf 176} 334 (2006) [{\it Phys. Usp.} {\bf 49} 319 (2006)]
\noindent
205. MacFadyen A I, Woosley S E {\it Astrophys. J.} {\bf 524} 262 (1999)
\noindent
206. Piran T {\it Rev. Mod. Phys.} {\bf 76} 1143 (2004)
\noindent
207. Tchekhovskoy A, Narayan R, McKinney J C {\it New Astron.} {\bf 15} 749 (2010)
\noindent
208. Komissarov S S, Vlahakis N, Konigl A {\it Astrophys. J.} {\bf 407} 17 (2010)
\noindent
209. Daugherty J K, Harding A K {\it Astrophys. J.} {\bf 252} 337 (1982)
\noindent
210. Medin Z, Lai D {\it Astrophys. J.} {\bf 406} 1379 (2010)
\noindent
211. Anderson J M et al. {\it Astrophys. J.} {\bf 590} L107 (2003)
\noindent
212. Ferreira J, Dougados K, Cabrit S {\it Astron. Astrophys.} {\bf 453} 785 (2006)
}
\end{document}
|
\section{Introduction}
The measurement of the charge asymmetry of leptons originating from the decay
of singly produced \Wboson~bosons at $pp$, $p\bar{p}$ and $ep$ colliders provides important information
about the proton structure as described by parton distribution
functions (PDFs). The \Wboson~boson charge asymmetry is mainly
sensitive to valence quark distributions~\cite{Berger:1988tu} via the
dominant production process $u\bar{d}(\bar{u}d)\rightarrow W^{+(-)}$
and provides complementary information to that obtained from
measurements of inclusive deep inelastic scattering cross-sections at
the HERA electron-proton collider~\cite{Chekanov:2008aa,Chekanov:2009gm,Adloff:2003uh,:2009wt}.
The HERA data do not strongly constrain the ratio between $u$ and $d$
quarks in the kinematic regime of low $x$, where $x$ is the proton
momentum fraction carried by the parton~\cite{Nakamura:2010zzi}. A precise measurement of the \Wboson~asymmetry
at the Large Hadron Collider (LHC)~\cite{LHC:2008} on the other hand, can contribute
significantly to the understanding of PDFs and quantum chromodynamics
(QCD) in the parton momentum fraction range $10^{-3}\lesssim
x\lesssim 10^{-1}$.
In $pp$ collisions the overall production rate of \Wplus~bosons is
significantly larger than the corresponding \Wminus~rate, since the
proton contains two $u$ and one $d$ valence quarks. The first
measurements of the inclusive $\Wpm$ cross-sections at the LHC by the
ATLAS~\cite{wzpaper} and the CMS~\cite{Khachatryan:2010xn}
Collaborations confirmed the difference predicted by the Standard
Model. The asymmetry in $pp$ collisions is symmetric with respect to
the $\Wboson$ rapidity, whereas in $\antibar{p}$ collisions it is
antisymmetric; the small sensitivity to sea quark contributions is
strongly suppressed in $\antibar{p}$ compared to $pp$
collisions~\cite{Lohwasser:2010sp}. Measurements in $p\bar{p}$
collisions have been performed at the Tevatron by both the
CDF~\cite{Abe:1998rv,Acosta:2005ud} and
D\O~\cite{Abazov:2007pm,Abazov:2008qv} Collaborations, and the data
have been included in global fits of parton
distributions~\cite{Martin:2009iq,Pumplin:2002vw}.
This letter presents a differential measurement of the muon charge
asymmetry from the decay of $W^{\pm}$ bosons in $pp$ collisions at a
centre-of-mass energy of $\sqrt{s}~=~7\TeV$ at the LHC. The asymmetry
varies significantly as a function of the pseudorapidity\footnote{The
nominal $pp$ interaction point at the centre of the detector is
defined as the origin of a right-handed coordinate system. The
positive $x$-axis is defined by the direction from the interaction
point to the centre of the LHC ring, with the positive $y$-axis
pointing upwards. The azimuthal angle $\phi$ is measured around the
beam axis and the polar angle $\theta$ is the angle from the
$z$-axis. The pseudorapidity is defined as $\eta = -\ln
\tan(\theta/2)$.} $\eta_\mu$ of the charged decay lepton owing to
its strong correlation with the momentum fraction $x$ of the partons
producing the \Wboson~boson. It is defined from the cross
sections for $W \to \mu\nu$ production
$d\sigma_{\mathrm{W\mu^\pm}}/d\eta_{\mu}$ as:
\begin{equation}
A_\mu =
\frac{d\sigma_{\mathrm{W\mu^+}}/d\eta_\mu-d\sigma_{\mathrm{W\mu^-}}/d\eta_\mu}
{d\sigma_{\mathrm{W\mu^+}}/d\eta_\mu+d\sigma_{\mathrm{W\mu^-}}/d\eta_\mu}\,,
\label{eq:asymmetry}
\end{equation}
where the cross sections include the event kinematical cuts used to select $W \to \mu\nu$ events. No extrapolation to the full phase space is attempted in order to reduce the dependence on theoretical predictions.
Systematic effects on the $\Wboson$-production cross-section
measurements are typically the same for positive and negative muons,
mostly canceling in the asymmetry. The ATLAS detector measures muons
with two independent detector systems. These two
independent measurements allow systematic uncertainties to be
controlled. The results presented are based on data collected in 2010 with an
integrated luminosity of $31~\ipb$. These results significantly
improve on the previous measurement by the ATLAS
Collaboration~\cite{wzpaper}, which is based on a data set
approximately 100 times smaller.
\section{The ATLAS Detector}
The ATLAS detector~\cite{AtlasDetector,CSCbook} consists of an inner
tracking system (inner detector, or ID) surrounded by a superconducting solenoid providing a 2T magnetic field,
electromagnetic and hadronic calorimeters and a muon spectrometer
(MS). The ID consists of pixel and silicon microstrip (SCT)
detectors, surrounded by a transition radiation tracker (TRT). The
electromagnetic calorimeter is a lead liquid-argon (LAr) detector in the barrel and the endcap, and in the forward region copper LAr technology is used.
Hadron calorimetry is based on two different detector technologies,
with scintillator tiles or LAr as the active media, and with either
steel, copper, or tungsten as the absorber material. There is a poorly instrumented transition region between the barrel and endcap calorimeter, $1.37<|\eta|<1.52$, where electrons cannot be precisely measured. In view of a later combination, this motivates the binning in that region for the present muon analysis. The MS is based
on three large superconducting toroids, and a system of three
stations of chambers for trigger and precise tracking measurements.
There is a transition between the barrel and endcap muon detectors around $|\eta|=1.05$.
\section{Data and Simulated Event Samples}
The data used in this analysis were collected from the end of
September to the end of October 2010. Basic requirements on beam, detector, stable trigger conditions and
data-quality were used in the event selection, resulting
in a total integrated luminosity of $31~\ipb$. Events in this analysis are selected using a single-muon trigger with
a requirement on the momentum transverse to the beam (\pT) of at least
$13\GeV$. The trigger includes three levels of event selection: a
first level hardware-based selection using hit patterns in the MS and
two higher levels of software-based requirements.
Simulated event samples are used for the background estimation, the
acceptance calculation and for comparison of data with theoretical
expectations. The processes considered are the $\Wmn$ signal, and
backgrounds from $\Wtau$, $\Zmm$, $\Ztau$, $\ttbar$ and jet production
via QCD processes (referred to as ``QCD background" in the text). The
signal and background samples (except $\ttbar$) were generated
with PYTHIA 6.421~\cite{pythia} using MRST 2007
$\rm{LO}^*$~\cite{mrst} PDFs. The $\ttbar$ sample was generated with
POWHEG-HVQ v1.01 patch
4~\cite{Powheg}; the PDF set was CTEQ 6.6M~\cite{CTEQ6.6M} for the NLO
matrix element calculations, while CTEQ 6L1 was used for the parton
showering and underlying event via the POWHEG interface to PYTHIA. The
radiation of photons from charged leptons was treated using PHOTOS
v2.15.4~\cite{Photos} and TAUOLA v1.0.2~\cite{Tauola} was used for tau
decays. The underlying and pile-up events were simulated
according to the ATLAS MC09 tune~\cite{MC09tune}.
The generated samples were passed through the GEANT4\\\cite{geant4} simulation of the ATLAS
detector~\cite{AtlasSimulation}, reconstructed and analysed with the
same analysis chain as the data. The cross-section
predictions for $W$ and $Z$ were calculated to
next-to-next-to-leading-order (NNLO) using FEWZ~\cite{FEWZ} with the
MSTW 2008~\cite{MSTW2008} PDFs. The $\ttbar$ cross-section was
obtained at next-to-leading-order (plus next-to-next-to-leading-log,
NNLL) using POWHEG~\cite{Bonciani:1998vc}. The Monte Carlo (MC) were generated with, on average, two soft inelastic collisions overlaid on top of the hard-scattering event.
Events in the MC samples were weighted so that the distribution of the number of inelastic collisions per bunch crossing matched that in data, which has an average of $2.2$.
\section{Event Selection}
The criteria for the event selection and muon identification follow
closely those used for the $W$ boson inclusive cross-section
measurement~\cite{wzpaper}, with an improved muon quality selection~\cite{wjets}.
Events from $pp$ collisions are selected by requiring a collision vertex with
at least three tracks each with transverse momentum greater than
$150\MeV$. A beam-spot constraint has been applied in the collision vertex
reconstruction stage significantly improving the resolution on the
collision vertex position in the transverse plane. To reduce the
contribution of cosmic-ray and beam-halo events, induced by proton
losses from the beam, the analysis
requires the collision vertex position along the beam axis to be
within~$20\cm$ of the nominal interaction point.
Events with a high transverse momentum muon are selected by imposing stringent requirements
to ensure good discrimination of $\Wmn$ events from background. The muon parameters are
first reconstructed separately in the MS and ID. Subsequently, the tracks from the ID and MS are
matched. Their parameters are then combined, weighted by their respective errors, to form a combined
muon. The \Wboson\ candidate events are required to have at least one combined muon track with
$\pT>20\GeV$ and {\pT} measured by the MS alone greater than $\pT^{\mathrm{MS}}>10\GeV$,
within the range $|\eta_{\mu}|<2.4$. The difference between the ID and MS \pT, corrected
for the mean energy loss in the material traversed between the ID and
MS, is required to be less than 0.5 times the ID \pT,
$$
\pT^{\mathrm{MS}} ({\mathrm{energy\;loss\;corrected}})- \pT^{ID}<0.5\phantom{0}\pT^{\mathrm{ID}}.
$$
This requirement increases the robustness against track reconstruction
mismatches, including decays-in-flight of\\
hadrons. In addition, a minimum
number of hits in the ID is required to ensure high quality
tracks~\cite{wjets}. In order to further reduce non-collision
backgrounds, the difference between the $z$ position of the muon track
extrapolated to the beam line and the $z$ coordinate of the collision
vertex is required to be less than $1\cm$.
A track-based isolation for the muon is defined as $\sum \pT^{\mathrm{ID}}/\pT<0.2$, where $\sum\pT^{\mathrm{ID}}$ is the scalar sum of transverse momenta of all other tracks measured in the ID within a cone\footnote{$\Delta R$ is defined as $\Delta R=\sqrt{\Delta \eta^2 + \Delta \phi^2}$.} $\Delta R <0.4$
around the muon direction excluding the ID track associated with the muon, and $\pT$ is
the transverse momentum of the muon combined track.
The reconstruction of the missing transverse energy ($\met$) and the
transverse mass (\mT) follows the prescription in~\cite{wzpaper}. The
\met~is determined from the energy deposits of calibrated calorimeter cells in
three-dimensional clusters and is corrected for the momentum of all muons
reconstructed in the event. Jet-quality requirements are applied to remove a small fraction of events where sporadic
calorimeter noise and non-collision backgrounds can affect the
\met~reconstruction~\cite{CLEANING}.
The transverse mass is defined as
\begin{equation}
\ensuremath{\mT} =
\sqrt{2\ensuremath{\pT}^{\mu}\ensuremath{\pT}^{\nu}(1-\cos(\phi^{\mu}-\phi^{\nu}))},
\end{equation}
where the highest $\pt$ muon is used and the $(x,y)$ components
of the neutrino momentum are inferred
from the corresponding \met~components. Events are required to
have $\met>25\GeV$ and $\mT>40\GeV$, yielding 129572
$\Wboson$ candidates.
\section{$W^{\pm}$ Signal Yield and Background Estimation}
\label{sec:background}
\begin{figure*}[t]
\centering
\subfigure[]{\includegraphics[width=0.40\linewidth]{fig1a}\label{subfig:etapos}}
\subfigure[]{\includegraphics[width=0.40\linewidth]{fig1b}\label{subfig:etaneg}}
\caption{Distribution of the muon pseudorapidity $\eta_\mu$ of
$\Wplus$~\subref{subfig:etapos} and $\Wminus$~\subref{subfig:etaneg}
candidates, after final selection. The data are compared to
MC simulation, broken down into the signal and various
background components. The MC distributions are normalised
to the total number of events in data.}
\label{fig:eta}
\end{figure*}
\begin{table*}[ht]
\begin{center}
\begin{tabular}{l|ccc|ccc}
\hline
\hline
& \multicolumn{3}{|c|}{$\mu^+$} & \multicolumn{3}{c}{$\mu^-$} \\\cline{2-7}
& Observed & Exp. Background & $\CWmp$ & Observed & Exp. Background & $\CWmm$ \\
\hline
$0.00<|\eta_\mu|<0.21$ & $5052$ & $272\pm 51$ & $0.594\pm 0.005$ & $3726$ & $236\pm 55$ & $0.584\pm 0.004$\\
$0.21<|\eta_\mu|<0.42$ & $6519$ & $385\pm 70$ & $0.779\pm 0.009$ & $4757$ & $334\pm 70$ & $0.759\pm 0.008$\\
$0.42<|\eta_\mu|<0.63$ & $6845$ & $481\pm 88$ & $0.808\pm 0.009$ & $4936$ & $357\pm 70$ & $0.800\pm 0.009$\\
$0.63<|\eta_\mu|<0.84$ & $5963$ & $366\pm 76$ & $0.686\pm 0.008$ & $4212$ & $329\pm 64$ & $0.691\pm 0.008$\\
$0.84<|\eta_\mu|<1.05$ & $5933$ & $395\pm 63$ & $0.672\pm 0.007$ & $4207$ & $358\pm 63$ & $0.681\pm 0.008$\\
$1.05<|\eta_\mu|<1.37$ &$10114$ & $627\pm 93$ & $0.735\pm 0.007$ & $6544$ & $585\pm 101$ & $0.752\pm 0.007$\\
$1.37<|\eta_\mu|<1.52$ & $5726$ & $363\pm 57$ & $0.905\pm 0.009$ & $3601$ & $348\pm 59$ & $0.914\pm 0.009$\\
$1.52<|\eta_\mu|<1.74$ & $8228$ & $542\pm 89$ & $0.905\pm 0.008$ & $5043$ & $518\pm 82$ & $0.925\pm 0.008$\\
$1.74<|\eta_\mu|<1.95$ & $7982$ & $605\pm 114$ & $0.896\pm 0.009$ & $4688$ & $456\pm 80$ & $0.898\pm 0.008$\\
$1.95<|\eta_\mu|<2.18$ & $8392$ & $647\pm 100$ & $0.903\pm 0.009$ & $4971$ & $548\pm 91$ & $0.910\pm 0.009$\\
$2.18<|\eta_\mu|<2.40$ & $7562$ & $534\pm 81$ & $0.881\pm 0.010$ & $4571$ & $492\pm 82$ & $0.896\pm 0.010$\\
\hline
\hline
\end{tabular}
\caption{Summary of observed number of events, expected background and
correction factor $\CWm$ for positive and negative muons in bins of $|\eta_{\mu}|$. The
errors given for the background estimates include systematic
uncertainties, including the uncertainty due to the luminosity, used in the normalization of
the electro-weak and $\ttbar$ components. The $\CWm$ factors include trigger
and muon reconstruction scale factors; they include the statistical uncertainty
from the MC sample and the trigger and reconstruction scale factors.
}
\label{tab:Wsummary_muon}
\end{center}
\end{table*}
Many components in the $W$ cross-section measurement, such as the
luminosity or detector efficiencies, are in principle the same for
positive and negative muons and therefore mostly cancel in the asymmetry calculation.
The main experimental biases on the asymmetry measurement come from
possible differences in the reconstruction of positive and negative muons.
Each effect (trigger and reconstruction efficiency and
momentum scale) is examined to check that the two charges
behave in the same way within the systematic uncertainties.
These studies are performed in absolute pseudorapidity in order to reduce
the uncertainty associated with the limited size of the data samples used.
As in past $W$ analyses,
trigger~\cite{wjets} and muon reconstruction~\cite{wzpaper,wjets}
efficiencies as a function of muon $\eta_\mu$ have been measured in data
using a sample of unbiased muons from
\Zmm~decays, which provides a source of muons with small background.
The trigger efficiency is
determined relative to a reconstructed muon satisfying the selection
criteria of the analysis. The average trigger efficiencies after the full
$\Wboson$ selection are $(81\pm 2)$\%
in the central detector region, $|\eta_\mu|<1.05$, and $(94\pm 1)$\% in
the forward detector region, $1.05<|\eta_\mu|<2.4$, where the differences
are due to the geometrical acceptance of the muon trigger chambers.
In the same muon sample, the muon
reconstruction efficiency relative to an ID track is measured to be
$(93\pm 1)$\% overall. The efficiency for reconstructing an ID track
is $(99\pm 1)$\%~\cite{wzpaper}. The quoted uncertainties on these
efficiencies are statistical.
Corrections have been applied to the simulated samples to account for
differences in the trigger and reconstruction efficiencies between
data and simulation. These are based on the ratio of the efficiency in
data and in simulation, and are computed as a function of the muon
$\eta_\mu$ and charge. The corrections for each charge agree within the
statistical uncertainties, so the charge-averaged result is
applied. For the trigger, the corrections are $0.98$ and $1.03$
in the central and forward MS
regions, respectively. For the reconstruction efficiency, the
correction factors are about $0.99$ per $\eta_\mu$ bin except for the central-forward MS
transition region ($|\eta_\mu|$ about $1.05$) where the correction factor is $0.94$.
The muon momentum resolution is affected by the a\-mount of material
traversed by the muon, the spatial resolution of the individual track
points and the degree of internal alignment of the ID and
MS~\cite{muperfcosmics}. This resolution has been measured as a
function of $\eta_\mu$ for the main detector regions
(in $\eta_\mu$ ranges delimited by $1.05, 1.7, 2.0$ and $2.4$) from the
width of the di-muon invariant mass distribution in $\Zmm$ decays and
from the comparison of the momentum measurements in the ID and MS in
$\Zmm$ and $\Wmn$ decays. The measured
resolution is worse than expected from simulation by
$1\;\!$--$\;\!5\%$, with the maximum discrepancy reached in the high-$\eta_\mu$
region of the detector.
The discrepancy is due to residual mis-alignments in the ID and MS,
imperfections in the description of the inert material in simulation and
an imperfect mapping of the magnetic field in the MS transition region where
the field is highly non-uniform. Smearing corrections are therefore applied to
the simulation in order to improve the agreement with data.
If the accuracy of the muon momentum measurement is different for
positive and negative muons, this difference can produce a bias in the
acceptance of $\mu^+$ with respect to $\mu^-$. Differences in the muon
\pT\ measurement between data and simulation have been evaluated
comparing the curvature of muons from \Wboson\ candidates in data and
in templates derived from simulation. A binned likelihood fit for a
momentum-scale correction that yields the best agreement between data
and simulation is performed as a function of $\eta_\mu$ separately for
positive and negative charges. The measured biases in the $\pT$ scale
between the two charges are $<1\%$, but they increase to about
$3\%$ in the transition and high-$\eta_\mu$ regions due to residual
mis-alignments in the ID and MS. These corrections are applied to the
muon momenta in the simulated samples.
Figure~\ref{fig:eta} shows the
pseudorapidity distribution of the selected positive and negative
muons. Data distributions are compared to the MC simulation,
normalised to the total number of events in data. The shape of the
simulation agrees well with the shape of the data after the
corrections for the reconstruction and trigger efficiencies, and the
muon-momentum scale and resolution.
The main backgrounds to $\Wmn$ arise from heavy flavour
decays in multijet events and from the electro-weak background from
\Wtau~where the tau decays to a muon, \Zmm~where
one muon is not reconstructed and \Ztau~where one tau decays to a muon,
as well as semileptonic \ttbar~decays in the muon channel.
Di-boson and single top backgrounds are found to be negligible.
The \Wtau~contribution is treated as a background. While this
contribution presents the same asymmetry as the \Wmn~signal, it is
difficult to include in PDF fits, which assume that the asymmetry is a
function of $\eta_\ell$ for $\Wboson \rightarrow l\nu$.
The background estimates of the electro-weak and $\ttbar$ backgrounds and the QCD
background closely follow the methods used in the $W$ inclusive cross-section
measurement~\cite{wzpaper}. They are determined separately
for positive and negative muons as a function of $\eta_{\mu}$.
The electro-weak and $\ttbar$ backgrounds are estimated using MC simulation.
The QCD background comes primarily from $b$ and $c$ quark decays, with a smaller contribution
from pion and kaon decays in flight. This background is estimated using a data-driven
method similar to the one described
in~\cite{wzpaper}. The sample of events fulfilling the full $\Wboson$
selection criteria with the exception of the muon isolation requirement is
compared before and after the isolation requirement. The
isolation efficiency for non-QCD events is measured in data with
the $\Zmm$ sample. The efficiency for QCD events is estimated in a
control sample of low-$\pt$ muons extrapolated to the high-$\pt$ and
high-$\met$ signal region using the simulated jet sample. Since
the samples before and after isolation can be defined in terms of a
QCD and non-QCD component, the expected number of QCD events can thus
be determined.
\begin{figure*}[tph]
\centering
\subfigure[]{\includegraphics[width=0.4\linewidth]{fig2a}\label{subfig:muptpos}}
\subfigure[]{\includegraphics[width=0.4\linewidth]{fig2b}\label{subfig:muptneg}}
\caption{Distribution of the transverse momentum of positive and negative
muons after the final selection. The data are compared to
MC simulation, broken down into the signal and various
background components. The MC distributions are normalised
to the total number of entries in data.}
\label{fig:mupt}
\end{figure*}
The expected background amounts to $7\%$ of the selected events;
$6\%$ is the electro-weak and $\ttbar$ contribution ($3\%$ $\Zmm$,
$2\%$ $\Wtau$, and $1\%$ for the sum of $\ttbar$ and $\Ztau$) and the
remainder is the QCD background. The cosmic ray
background contamination is estimated to be smaller by a factor of $10^5$
compared to the signal and thus negligible. The $W^{\pm}$ candidate events
and expected background contributions are summarised in
Table~\ref{tab:Wsummary_muon}.
Figure~\ref{fig:mupt} shows the transverse momentum distribution for
positive and negative muons after the full event selection. They
are compared with the distributions predicted by the corrected
MC simulation normalised to the total number of events
in data. The correction factors, $\CWm$,
corresponding to the ratio of reconstructed over generated events in the
simulated $W$ sample, satisfying all kinematic requirements of the event selection,
$p_{\mathrm{T}}^{\mu} > 20\GeV$, $p_{\mathrm{T}}^{\nu} > 25\GeV$, $\mT > 40\GeV$,
are also listed in Table~\ref{tab:Wsummary_muon}. No correction is made
to the full acceptance.
The $\CWm$ factors include trigger and muon reconstruction scale factors to
correct for observed deviations between data and MC efficiencies. Due to a
reduced geometric acceptance in the trigger, the $\CWm$ factors for the lowest
$|\eta_\mu|$ bins are significantly smaller than those for the higher $|\eta_\mu|$
regions.
\section{Systematic Uncertainties}
All systematic uncertainties on the asymmetry measurement are
determined in each $|\eta_\mu|$ bin
accounting for correlations between the charges and are summarised
in Table~\ref{tab:muonsysts_summary_emu_abs}. The dominant sources
of systematic uncertainty on the asymmetry come from the trigger and
reconstruction efficiencies. The determination of these efficiencies is affected by
the statistical uncertainty due to the small available
sample of $\Zmm$ events.
Systematic uncertainties on the
efficiencies are determined from studies of the impact of
the selection criteria and backgrounds, and no significant charge
biases are found.
There is a loss of trigger efficiency in the low
pseudorapidity region due to reduced geometric acceptance,
resulting in a larger statistical error. As a result, the trigger systematic
uncertainty on the asymmetry is largest in the low pseudorapidity
bins (6-7\% for central $|\eta_\mu|$ and 2-3\% for forward $|\eta_\mu|$).
Similarly, the uncertainties associated with the reconstruction
efficiency are larger in the lowest pseudorapidity
bin (about $7\%$), and in the MS central-forward transition region
(about $3\%$), due to geometrical acceptance effects associated with
reduced chamber coverage. In the remaining regions, the uncertainty
is about 1-2\%.
\begin{table*}[hb]
{\small
\centering
\begin{tabular}{l|ccccccc}
\hline
\hline
& \multirow{2}{*}{Trigger} & \multirow{2}{*}{Reconstruction} & $\pt$ Scale and & QCD & Electro-weak and $\ttbar$ & Theoretical\\
& & & Resolution & Normalisation & Normalisation & Modelling\\
\hline
$0.00<|\eta_{\mu}|<0.21$ & $0.011$ & $0.010$ & $0.003$ & $0.003$ & $<0.001$ & $0.007$\\
$0.21<|\eta_{\mu}|<0.42$ & $0.010$ & $0.004$ & $0.003$ & $0.003$ & $<0.001$ & $0.005$\\
$0.42<|\eta_{\mu}|<0.63$ & $0.009$ & $0.004$ & $0.003$ & $0.003$ & $<0.001$ & $0.006$\\
$0.63<|\eta_{\mu}|<0.84$ & $0.012$ & $0.004$ & $0.003$ & $0.002$ & $0.001$ & $0.007$\\
$0.84<|\eta_{\mu}|<1.05$ & $0.013$ & $0.006$ & $0.003$ & $0.003$ & $0.001$ & $0.008$\\
$1.05<|\eta_{\mu}|<1.37$ & $0.006$ & $0.007$ & $0.002$ & $0.002$ & $0.001$ & $0.006$\\
$1.37<|\eta_{\mu}|<1.52$ & $0.006$ & $0.005$ & $0.002$ & $0.003$ & $0.002$ & $0.005$\\
$1.52<|\eta_{\mu}|<1.74$ & $0.005$ & $0.004$ & $0.002$ & $0.003$ & $0.002$ & $0.007$\\
$1.74<|\eta_{\mu}|<1.95$ & $0.006$ & $0.003$ & $0.002$ & $0.002$ & $0.001$ & $0.006$\\
$1.95<|\eta_{\mu}|<2.18$ & $0.006$ & $0.004$ & $0.002$ & $0.003$ & $0.002$ & $0.009$\\
$2.18<|\eta_{\mu}|<2.40$ & $0.007$ & $0.005$ & $0.002$ & $0.003$ & $0.002$ & $0.007$\\
\hline
\hline
\end{tabular}
\caption{Absolute systematic uncertainties on the $\Wboson$ charge
asymmetry from different sources as a function of absolute muon pseudorapidity
that are described in the text.}
\label{tab:muonsysts_summary_emu_abs}
}
\end{table*}
The muon momentum scale and resolution corrections contribute to the
uncertainty primarily due to the limited statistics for the fitting
procedures used to measure the differences between the data and
simulation. An additional source of uncertainty arises from potential
biases in the template shapes. The size of this effect is determined by using different templates created
by shifting the resolution parameters in opposite directions to account
for possible charge biases.
Uncertainties associated with the modelling of the background
contributions to the templates, particularly the QCD background, are
also included. The resulting uncertainty on the asymmetry is in the
1-2\% range, with little dependence on $\eta_\mu$.
The redundant ID and MS momentum
measurements result in a rate of charge mis-identification smaller than
$10^{-4}$ in the $\pt$ range considered, resulting in a
negligible impact on the asymmetry.
The momentum-scale correction procedure is further tested by exploiting the
redundant muon-momentum measurements offered by the ATLAS
detector. The full asymmetry measurement is performed with the ID and
MS components of the combined muon separately, including the scale
corrections. Figure~\ref{fig:idmsasym} compares the two independently
corrected charge-asymmetry distributions, showing good agreement
within the systematic uncertainty associated with the momentum-scale
correction.
\begin{figure}[tbph]
\centering
\includegraphics[width=0.80\linewidth]{fig3}
\caption{\Wboson\ charge asymmetry measured using the ID and MS
separately. The MS measurement is extrapolated to
the collision vertex, and corrected for energy-loss in the
calorimeters. The two measurements are independently corrected for
effects of the muon-momentum scale on the muon acceptance. The two
measurements are statistically correlated to a large extent, since they use
the same muons reconstructed by different subdetectors and algorithms.
The error bar reports therefore only the systematic uncertainty associated
with the momentum-scale
correction.}
\label{fig:idmsasym}
\end{figure}
The systematic uncertainties on the QCD background arise
primarily from the uncertainty on the isolation efficiency for muons in QCD events due
to possible mis-modellings of the extrapolation of the isolation efficiency
to the large $\pt$ and $\MET$ region in the QCD simulation (40\%).
This has been derived from differences in the
efficiency predictions between data and simulation in the low muon $\pt$
control region and in sideband regions where the muon $\pt$ or $\MET$ cuts
are reversed. The electro-weak and $\ttbar$ background and signal contributions are
subtracted from data in these comparisons. Additional uncertainties due to
the non-QCD isolation efficiency and the statistical uncertainty are included
in the total uncertainty on the QCD background estimate.
The corresponding systematic uncertainty on the asymmetry is 1-2\%,
with little dependence on $\eta_\mu$.
For the electro-weak and $\ttbar$ backgrounds, the uncertainty in the
cross-sections includes the PDF uncertainties (3\%), and the uncertainties
estimated from varying the renormalization and factorization scales:
5\% for $W$ and $Z$, and 6\%
for $\ttbar$~\cite{Moch:2008qy,Beneke:2009ye,wzpaper}. An additional
uncertainty from the luminosity of $11\%$ is included, since the backgrounds
are scaled to the luminosity measured in data. The combination
of all these contributions results in an uncertainty on the asymmetry of
less than $1\%$.
The impact of using an NLO MC rather than Pythia in the $\CWm$ factor calculation
has been evaluated and an additional systematic uncertainty of about $3\%$ is
included to account for the small variations observed.
Pythia uses a leading-log calculation for $\Wboson$ production and is expected
to give a reasonably accurate prediction for the low $\Wboson$ transverse
momentum $\Wpt$ region whereas MC@NLO~\cite{Frixione:2002ik} uses higher-order
matrix elements and is therefore expected to be more reliable in the high
$\Wpt$ region. Therefore the differences in the scale factors associated with these
two MC calculations gives a reasonable estimate of the associated systematic error.
\section{Results and Conclusions}
The measured particle-level differential charge asymmetry in eleven
bins of muon absolute pseudorapidity is shown in Table~\ref{tab:asymvals} and
Figure~\ref{fig:asym}. The statistical and systematic uncertainties
per $|\eta_\mu|$ bin are included and contribute comparably to the
total uncertainty. Table~\ref{tab:asymvals} and Figure~\ref{fig:asym}
also show particle-level expectations from $W$ predictions at NLO with
different PDF sets: CTEQ 6.6~\cite{Pumplin:2002vw}, HERA
1.0~\cite{:2009wt} and MSTW 2008~\cite{Martin:2009iq}; all predictions
are presented with 90\% confidence-level error bands. All MC
predictions are calculated using MC@NLO, with
all kinematic selection criteria applied to the truth particles. The PDF
uncertainty bands are obtained by summing in quadrature the deviations
of each of the PDF error sets~\cite{Pumplin:2001ct} from the respective
nominal predictions, according to the specifications of the corresponding
PDF collaborations to get 90\% C.L. bands. These uncertainties for all predictions
include experimental uncertainties as well as model and parametrization
uncertainties. The HERA 1.0~\cite{:2009wt} set also includes the uncertainty in
$\alpha_s$, which, however, is not the dominant source of uncertainty.
While the predictions with different PDF sets differ within their
respective uncertainty bands~\cite{Lohwasser:1265829,Alekhin:2011sk},
they follow the same global trend, rising with $\eta_\mu$. The measured
asymmetry agrees with this expectation. As demonstrated
graphically in Figure~\ref{fig:asym}, the data are roughly
compatible with all the predictions with different PDF sets, though some are
slightly preferred to others. A $\chi^{2}$-comparison using
the measurement uncertainty and the central value of the PDF
predictions yields values per number of degrees of freedom of $9.16/11$
for the CTEQ 6.6 PDF sets, $35.81/11$ for the HERA 1.0 PDF sets and
$27.31/11$ for the MSTW 2008 PDF sets.
In summary, this letter reports a measurement of the $\Wboson$ charge
asymmetry in $pp$ collisions at $\sqrt{s} = 7\TeV$ performed
in the $\Wboson \rightarrow \mu \nu$ decay mode using $31~\ipb$
of data recorded with the ATLAS detector at the LHC. Until the start of the LHC,
it has not been kinematically possible to precisely measure
the valence quark distributions and in particular the ratio of $u/d$ quarks
below $x\lesssim0.05$. Whereas none of the predictions with different PDF sets are inconsistent
with these data, the predictions are not fully consistent with each other since
they are all phenomenological extrapolations in $x$. The input of the data
presented here is therefore expected to contribute to the determination of the next generation
of PDF sets, helping reduce PDF uncertainties, particularly the shapes of the
valence quark distributions in the low-$x$ region.
\renewcommand{\arraystretch}{1.3}
\begin{table*}[tbph]
\centering
\begin{tabular}{r|cccc}
\hline
\hline
& Data & MSTW 2008 & CTEQ 6.6 & HERA 1.0\\
\hline
$0.00<|\eta_\mu|<0.21$ & $0.147 \pm 0.011 \pm 0.017$& $0.142_{-0.014}^{+0.006}$& $0.164_{-0.007}^{+0.006}$& $0.163\pm 0.007$\\
$0.21<|\eta_\mu|<0.42$ & $0.150 \pm 0.010 \pm 0.012$& $0.147_{-0.014}^{+0.007}$& $0.168_{-0.007}^{+0.006}$& $0.167\pm 0.007$\\
$0.42<|\eta_\mu|<0.63$ & $0.158 \pm 0.010 \pm 0.012$& $0.151_{-0.013}^{+0.007}$& $0.173_{-0.007}^{+0.006}$& $0.169\pm 0.007$\\
$0.63<|\eta_\mu|<0.84$ & $0.184 \pm 0.010 \pm 0.015$& $0.163_{-0.012}^{+0.008}$& $0.186_{-0.008}^{+0.007}$& $0.179_{-0.007}^{+0.008}$\\
$0.84<|\eta_\mu|<1.05$ & $0.186 \pm 0.011 \pm 0.017$& $0.176_{-0.012}^{+0.009}$& $0.198_{-0.008}^{+0.007}$& $0.188\pm 0.008$\\
$1.05<|\eta_\mu|<1.37$ & $0.240 \pm 0.008 \pm 0.011$& $0.197\pm 0.010$ & $0.219_{-0.010}^{+0.008}$& $0.203_{-0.008}^{+0.009}$\\
$1.37<|\eta_\mu|<1.52$ & $0.250 \pm 0.011 \pm 0.010$& $0.215_{-0.010}^{+0.011}$& $0.237_{-0.010}^{+0.009}$& $0.214\pm 0.009$\\
$1.52<|\eta_\mu|<1.74$ & $0.269 \pm 0.009 \pm 0.010$& $0.230_{-0.010}^{+0.012}$& $0.251_{-0.011}^{+0.009}$& $0.224\pm 0.009$\\
$1.74<|\eta_\mu|<1.95$ & $0.273 \pm 0.009 \pm 0.010$& $0.251_{-0.009}^{+0.013}$& $0.270_{-0.011}^{+0.010}$& $0.239_{-0.009}^{+0.010}$\\
$1.95<|\eta_\mu|<2.18$ & $0.276 \pm 0.009 \pm 0.012$& $0.266_{-0.010}^{+0.014}$& $0.284_{-0.011}^{+0.010}$& $0.251_{-0.010}^{+0.009}$\\
$2.18<|\eta_\mu|<2.40$ & $0.273 \pm 0.010 \pm 0.012$& $0.272_{-0.011}^{+0.015}$& $0.288_{-0.010}^{+0.009}$& $0.255_{-0.010}^{+0.009}$\\
\hline
\hline
\end{tabular}
\caption{The muon charge asymmetry from \Wboson-boson decays in bins
of absolute pseudorapidity. The data measurements are listed with statistical
and systematic uncertainties respectively. Predicted asymmetries of
the MSTW 2008, CTEQ 6.6, and HERA 1.0 PDF sets are shown for
comparison.}
\label{tab:asymvals}
\end{table*}
\begin{figure}[tbph]
\centering
\includegraphics[width=0.95\linewidth]{fig4}
\caption{The muon charge asymmetry from \Wboson-boson decays in bins
of absolute pseudorapidity.
The kinematic requirements applied are
$p_{\mathrm{T}}^{\mu} > 20\GeV$, $p_{\mathrm{T}}^{\nu} > 25\GeV$
and $\mT > 40\GeV$.
The data points (shown with error
bars including the statistical and systematic
uncertainties) are compared to MC@NLO predictions with different PDF sets.
The PDF uncertainty bands are described in the text and include
experimental uncertainties as well as model and parametrization
uncertainties.}
\label{fig:asym}
\end{figure}
\section{Acknowledgements}
We wish to thank CERN for the efficient commissioning and operation of
the LHC during this initial high-energy data-taking period as well as
the support staff from our institutions without whom ATLAS could not
be operated efficiently.
We acknowledge the support of ANPCyT, Argentina; YerPhI, Armenia; ARC,
Australia; BMWF, Austria; ANAS, Azerbaijan; SSTC, Belarus; CNPq and
FAPESP, Brazil; NSERC, NRC and CFI, Canada; CERN; CONICYT, Chile; CAS,
MOST and NSFC, China; COLCIENCIAS, Colombia; MSMT CR, MPO CR and VSC
CR, Czech Republic; DNRF, DNSRC and Lundbeck Foundation, Denmark;
ARTEMIS, European Union; IN2P3-CNRS,\\ CEA-DSM/IRFU, France; GNAS,
Georgia; BMBF, DFG, HGF, MPG and AvH Foundation, Germany; GSRT,
Greece; ISF, MINERVA, GIF, DIP and Benoziyo Center, Israel; INFN,
Italy; MEXT and JSPS, Japan; CNRST, Morocco; FOM and NWO, Netherlands;
RCN, Norway; MNiSW, Poland; GRICES and FCT, Portugal; MERYS (MECTS),
Romania; MES of Russia and ROSATOM, Russian Federation; JINR; MSTD,
Serbia; MSSR, Slovakia; ARRS and MVZT, Slovenia; DST/NRF, South
Africa; MICINN, Spain; SRC and Wallenberg Foundation, Sweden; SER,
SNSF and Cantons of Bern and Geneva, Switzerland; NSC, Taiwan; TAEK,
Turkey; STFC, the Royal Society and Leverhulme Trust, United Kingdom;
DOE and NSF, United States of America.
The crucial computing support from all WLCG partners is acknowledged
gratefully, in particular from CERN and the ATLAS Tier-1 facilities at
TRIUMF (Canada), NDGF (Denmark, Norway, Sweden), CC-IN2P3 (France),
KIT/GridKA (Germany), INFN-CNAF (Italy), NL-T1\\ (Netherlands), PIC
(Spain), ASGC (Taiwan), RAL (UK) and BNL (USA) and in the Tier-2
facilities worldwide.
\bibliographystyle{model1-num-names}
|
\section{KL divergence and MLE}
The Kullback-Leibler (KL) divergence between two discrete probability distribution $p_1, \ldots, p_K$ and $q_1, \ldots, q_K$ is defined as
\begin{equation}
\KL{p}{q} := \sum_{i=1}^{K} p_i \log_{2} \frac{p_i}{q_i} = \mathbb{E}_p \log_{2} \frac{p_i}{q_i}.
\end{equation}
For the continuous case
\begin{equation}
\KL{p}{q} := \int p(x) \log \frac{p(x)}{q(x)} dx,
\end{equation}
where $p(x)$ and $q(x)$ are probability density functions.
KL divergence measures how far $p$ and $q$ are apart. In particular, note that if $p = q$ then $\KL{p}{q} = 0$; and if $p$ and $q$ are independent (or more precisely their RVs) then $\KL{p}{q} = \infty$. Note however, that in general KL divergence is not symmetric: $\KL{p}{q} \neq \KL{q}{p}$. From the definiton it is intuitive that the second argument is typically a reference/``true'' value, whereas the first argument is the one we want to see how far it is from that reference distribution.
Let $\tilde{p}(x)$ be the empirical distribution of a sequence $(x_1, \ldots, x_N)$
\begin{equation}
\label{eq:empirical_distribution}
\tilde{p}(x) := \sum_{n=1}^{N} \delta(x - x_n),
\end{equation}
and $p(x | \theta)$ be a model distribution.
Typically if we want to estimate the parameter $\theta$ for a given data set we maximize the log-likelihood of the data
\begin{equation}
\ell(x | \theta) := \sum_{n=1}^{N} \log p(x_n | \theta).
\end{equation}
Using KL divergence it is intuitive to select a $\theta$ that minimizes the distance between the model distribution and the empirical distribution. In fact, it turns out that both are equivalent:
\begin{eqnarray}
\KL{\tilde{p}(x)}{p(x | \theta)} &=& \int \tilde{p}(x) \log \frac{\tilde{p}(x)}{p(x | \theta)} dx \\
&=& - H(\tilde{p}(x)) - \int \tilde{p}(x) \log p(x | \theta) dx,
\end{eqnarray}
where $H(\tilde{p}(x)) = \int \tilde{p}(x) \log \tilde{p}(x) dx $ is the entropy of $\tilde{p}(x)$. Note that the entropy is independent of the choice of $\theta$. Thus
\begin{equation}
\arg \min_{\theta} \KL{\tilde{p}(x)}{p(x | \theta)} = \arg \max_{\theta} \mathbb{E}_{\tilde{p}} \log p(x | \theta).
\end{equation}
Plugging \eqref{eq:empirical_distribution} into the right hand side gives
\begin{eqnarray}
\mathbb{E}_{\tilde{p}} \log p(x | \theta) &=& \frac{1}{T} \int \sum_{n=1}^{N} N \delta(x - x_n) \log p(x | \theta) dx \\
&=& \frac{1}{N} \sum_{n=1}^{N} \log p(x_n | \theta) \\
&=& \frac{1}{N} \ell(x | \theta)
\end{eqnarray}
Given this relation we can now evaluate the log-likelihood of $x$ given the KL divergence and entropy of the empirical distribution as
\begin{eqnarray}
\KL{\tilde{p}(x)}{p(x | \theta)} &=& - H(\tilde{p}) - \frac{1}{N} \ell(x | \theta) \\
\Leftrightarrow \ell(x | \theta) &=& -N \cdot \left(\KL{\tilde{p}(x)}{p(x | \theta)} + H(\tilde{p}) \right)
\end{eqnarray}
Since we can view the empirical spectrum as a multinomial probability mass function on the unit circle, the Kullback-Leibler divergence seems to be a natural candidate. The KL divergence between two discrete probability distribution $p_1, \ldots, p_K$ and $q_1, \ldots, q_K$ is defined as.
\begin{equation}
\KL{p}{q} := \sum_{i=1}^{K} p_i \log_{2} \frac{p_i}{q_i} = \mathbb{E}_p \log_{2} \frac{p_i}{q_i}.
\end{equation}
Note however, that in general KL divergence is not symmetric: $\KL{p}{q} \neq \KL{q}{p}$. This violates the conditions needed for a proper application of Laplacian graphs and diffusion maps. A typical ``trick'' is to symmetrize KL divergence as
\begin{equation}
\symKL{p}{q} := \frac{\KL{p}{q} + \KL{q}{p}}{2}.
\end{equation}
Hence for $\mathbf{X} = \left( s_{1}, \ldots, s_{n} \right) \in \mathbb{R}^{T \times n}$ the similarity matrix between spike $s_{i}$ and $s_{j}$ can be defined as
\begin{equation}
\mathbf{S} = \symKL{I(s_i)}{I(s_j)},
\end{equation}
where $I(s_k)$ is the periodogram estimate of spike $s_{k,t}$.
\begin{comment}
\section{Progress report}
For the project proposal I set the following ``milestones'':
\begin{itemize}
\item have an automated procedure that can succesfully extract the actual spikes (only noise remains)
\item characterize spikes theoretically and empirically so well that I can get useful features as a basis for classification.
\item Have a good idea about what a good (parametric) distribution fits the data/features; the raw data is obviously non-Gaussian; it might be possible to approximate the features by (mixtures of) Gaussians.
\item Present preliminary results based on the models I have by then.
\end{itemize}
I think I have accomplished that mostly; maybe in the second point I don't have any highly theoretical characterizations, but the slowness measure seems to be a very reasonable feature that approaches the problem at its heart.
\end{comment}
\section{Clustering spikes}
\section{Clustering in the complex plane}
Doing the FFT of the signal the different wave forms must be visible as ``cones'' on the complex plane, pointing in some particular direction (the important frequency); due to the way the data is collected these cones are very likely pointing towards low frequencies (something like angles greater than $\pi / 16$).
The noise in the measurements will be apparent by a circular bivariate distributed point cloud around $(0,0)$: this is the white noise sequence.
Picking out these cones will give the clusters.
\begin{remark}
The logarithm of the Fourier coefficients converts the ``cones'' to nice horizontal lines; much easire to cluster. Look into the complex logarithm.
\end{remark}
\begin{figure}[!t]
\centering
\makebox{\includegraphics[width=\textwidth]{images/simulated_spike_trains.png}}
\caption{\label{fig:simulated_spike_trains} (top left) the $5$ true signals $S_i(t)$ of $5$ neurons; (top right) $N$ measured signals plus noise $s_{i,t}$, $t = 1, \ldots, 46$; (bottom left) Fourier coefficients in the complex plan of all $N$ signals; (bottom right) logarithm of the Fourier coefficients.}
\end{figure}
\begin{figure}[!t]
\centering
\makebox{\includegraphics[width=\textwidth]{images/true_and_noisy.png}}
\caption{\label{fig:simulated_spike_trains} (top left) the $5$ true signals $S_i(t)$ of $5$ neurons; (top right) $N$ measured signals plus noise $s_{i,t}$, $t = 1, \ldots, 46$; (bottom left) Fourier coefficients in the complex plan of all $N$ signals; (bottom right) logarithm of the Fourier coefficients.}
\end{figure}
\begin{figure}[!t]
\centering
\makebox{\includegraphics[width=\textwidth]{images/joint_density_plot_sim.png}}
\caption{\label{fig:simulated_spike_trains} Joint density of a mixture of $10$ Gamma - von Mises distributions}
\end{figure}
\subsection{Prior knowledge/ distributions}
The higher the variance of the noise, the wider the cones on the plane ($\rightarrow$ linear models theory: $\mathbb{V} \beta = \sigma^2 (X'X)^{-1}$. For Fourier basis, this $X'X = I$ (orthonormal), so the width of the cone should grow linearly with the variance of the noise.)
For an iid sequence all frequencies are equally important (uniform distribition) and the disribution of the radii of the coefficients is exponential (see Brockwell and Davis).
For the signals a von Mises distribution on the frequency and a Gamma distribution for the coefficients for a given frequency would make sense, i.e.\ Fourier coefficients $a_{j} = r \exp(i \varphi)$ are random variables with the frequency $\varphi$ having a von Mises distribution\footnote{\url{https://secure.wikimedia.org/wikipedia/en/wiki/Von_Mises_distribution}}
\begin{equation}
\varphi \sim f(\lambda | \mu, \kappa) = \frac{e^{\kappa\cos(x-\mu)}}{2\pi I_0(\kappa)}
\end{equation}
Then for a given frequency, the radius of the Fourier coefficients should follow a distribution that has the exponential distribution as a special case, since for an iid sequence the distribution is exponential (Brockwell and Davis). For example the Gamma distribution would be useful (also because it's from a conjugate family)
\begin{equation}
r | \varphi \sim \Gamma(k, \theta)
\end{equation}
For a simple model just assume that radius and phase are independent. However, for this data they presumably are not: since we measure the signals only over a certain time frame (that tries to exactly match one complete waveform), very high frequencies will not be important for these series. Thus if $\varphi$ is very small, then also the radius is presumably small. Or in terms of distributions, for low frequencies I expect a more exponential distribution than for low frequencies (that's where Fourier coefficients will show up).
\section{Discussion and outlook}
I introduce a novel technique to detect and classify similar dynamics in signals, where similar dynamics can either mean similar shape for non-stationary signals, or similar auto-correlation for stationary signals. It is an adaptation of the EM algorithm to the power spectra of the signals and thus future research can benefit from the extensive literature in both areas of signal processing. Applications to neural spike sorting and pattern recognition in macro-economic time series demonstrate the usefulness of the presented method.
I also used the recently introduced slowness feature for the classification of neuron spikes. The slowness of signals can separate signals from noise and also distinguish differently shaped signals. Compared to multivariate methods in the literature it is very fast and easily computable, and more robust to outliers than for example the standard approach of a simple threshold method.
\section{Non-parametric frequency domain EM algorithm}
\label{sec:EM_spectral_density}
The fundamental idea of this novel approach to clustering dynamic structures in time series is to identifying each time series with the distribution it induces on the unit circle - and thus on the interval $[-\pi, \pi]$ - by its Fourier transform. These distributions can then be used in a mixture density setting and an adaptation of the EM algorithm yields the classification algorithm.
\begin{definition}[Spectral Density]
The spectral density of a stationary, zero-mean time series $x_t$ with auto-covariance function $\gamma_{x}(\ell) = \mathbb{E} x_t x_{t-\ell}$ is defined as
\begin{equation}
\label{eq:spectral_density}
f_x(\lambda) := \frac{1}{2 \pi} \sum_{\ell=-\infty}^{\infty} \gamma_{x}(\ell) e^{i \lambda \ell}, \quad \lambda \in [-\pi, \pi],
\end{equation}
where the limit is understood point-wise if $\lbrace \gamma_{x}(\ell) \rbrace_{\ell=-\infty}^{\infty}$ is absolutely summable, and in the mean-square sense if $\lbrace \gamma_{x}(\ell) \rbrace_{\ell=-\infty}^{\infty}$ is square summable.
\end{definition}
For real valued processes $\gamma_{x}(\ell) = \gamma(-\ell)$, thus $f_x(\lambda) \geq 0$ for all $\lambda$. Furthermore,
\begin{equation}
\label{eq:spectral_decomposition}
\int_{-\pi}^{\pi} f_x(\lambda) = \sigma_x^2,
\end{equation}
since $\int_{-\pi}^{\pi} e^{i \lambda \ell} d \lambda = 0$ if $\ell \neq 0$ and $\gamma_x(0) = \sigma_x^2$. Equivalence \eqref{eq:spectral_decomposition} is also known as the spectral decomposition of the variance of a time series. Hence, the spectral density is a non-negative function on the interval $[-\pi,\pi]$ and peaks at $\lambda_0$ indicate that this frequency is important for the overall variance of the process, since those peaks contribute a lot to the integral in \eqref{eq:spectral_decomposition}.
An estimate of the spectral density is the power spectrum or \emph{periodogram}.
\begin{definition}[Periodogram]
The periodogram (or power spectrum) of $x_t$ is defined as
\begin{equation}
\label{eq:periodogram}
I_{T,x}(\omega_j) := \left| X(\omega_j) \right|^2 = \Big| \frac{1}{\sqrt{T}} \sum_{k=0}^{T-1} x_t e^{-2 \pi i \omega_j t}\Big|^2, \quad \omega_j= j/T, \quad j=0,1,\ldots, T-1
\end{equation}
where $\omega_j$ are the Fourier frequencies (scaled by $2 \pi$ for easier interpretation).
\end{definition}
For large $T$ a frequent model for $I_{T,x}(\omega_j)$ is \citep[see][]{BrockwellDavis91}
\begin{align}
I_{T,x}(\omega_j) =
\begin{cases}
\chi_1^2, & \text{if $j = 0$ or $T/2$,}
\\
f_x(\omega_j) \, \eta, & \text{otherwise.}
\end{cases}
\end{align}
where $\eta$ is a standard (rate $=$ 1) exponential RV. At each frequency $\omega_j$ (except $0$ and $\pi$) the periodogram is an exponential RV with rate parameter equal to the true spectral density $f_x(\omega_j)$. Therefore $I_{T,x}(\omega_j)$ is asymptotically unbiased ($\mathbb{E} I_{T,x}(\omega_j) = f_x(\omega_j)$), but not consistent ($\mathbb{V} I_{T,x}(\omega_j) = f_x(\omega_j)^2 \stackrel{T \rightarrow \infty}{\nrightarrow} 0)$. This is especially harmful for large values of the true spectral density, as they exactly correspond to those frequencies which are particularly important for the overall variation.\\
There are two main ways to reduce the variance of the raw periodogram. If only one time series is available, then one can reduce the variance of $I_{T,x}(\omega_j)$ by smoothing over neighboring frequencies \citep{OppenheimSchafer89_Discrete-TimeSignalProcessing} - just as in non-parametric density estimation. This works well for series with many observations, but for small samples such as the neuron spikes or many economic time series averaging over neighboring frequencies is not a practical option as it quickly introduces too much bias at each $\omega_j$.
For $M_k$ independent time series $\lbrace x_{m,t} \rbrace_{m=1}^{M_k} $ of the same type (all from sub-system $\mathcal{S}_k$) a variance-reduced estimate of the true $f_{\mathcal{S}_k}(\lambda)$ can be obtained by averaging over all $M_k$ periodograms at each frequency
\begin{equation}
\label{eq:M_periodograms}
\widehat{f}_{\mathcal{S}_k}(\lambda) \mid_{\lambda = \omega_j} = \frac{1}{M_k} \sum_{m = 1}^{M_k} I_{T, x_{m,t}}(\omega_j), \quad j=0, \ldots, T-1.
\end{equation}
Since by assumption all $x_{m,t} \in \mathcal{S}_k$ have the same dynamic structure, $\widehat{f}_{\mathcal{S}_k}(\lambda)$ is also a good estimate of $f_{x_{m,t}}(\lambda)$ for all $m = 1, \ldots, M_k$.
If the sub-series $x_{m,t}$ are far enough apart in a signal $y_t$, then periodograms $I_{T, x_{m,t}}(\omega_j)$ can be considered as independent estimates of the same underlying true spectral density $f_{\mathcal{S}_k}(\lambda)$. Thus \eqref{eq:M_periodograms} is still unbiased but has a much lower variance
\begin{equation}
\mathbb{E} \widehat{f}_{\mathcal{S}_k}(\omega_j) = f_{\mathcal{S}_k}(\omega_j), \quad \mathbb{V} \widehat{f}_{\mathcal{S}_k}(\omega_j) = \frac{ f^2_{\mathcal{S}_k}(\omega_j)}{M_k}.
\end{equation}
\subsection{From spectral density estimation to the EM algorithm}
Equation \eqref{eq:M_periodogram} looks very similar to the M step of an EM algorithm \citep{McLachlanetal08_EM_book}. By averaging over periodograms in \eqref{eq:M_periodograms} we assume we \emph{know} that series $\mathbf{x}_{i,t}$ came from system $\mathcal{S}_k$. This can be a reasonable assumption when repeatedly measuring time series in controlled physical experiments. In many applications, however, it is not \emph{known} where the signal came from. Thus I introduce a non-parametric frequency domain EM algorithm to classify time series. As a general idea this shift from averaging periodograms deterministically to probabilistically is analogous to the shift from hard-thresholding in k-means to soft-thresholding in the EM algorithm.\\
Formally, let $\mathbf{z}_{i}$ be a $K$-dimensional vector indicating from which system series $\mathbf{x}_{i,t}$ comes from; i.e.\ $z_{i k} = 1$ if $\mathbf{x}_{i,t}$ is from $\mathcal{S}_k$, $0$ otherwise. By averaging over periodograms as in \eqref{eq:M_periodograms}, $\mathbf{z}_i$ is treated as a deterministic, known variable. Thus \eqref{eq:M_periodograms} can be rewritten to
\begin{align}
\widehat{f}_{\mathcal{S}_k}(\lambda) \mid_{\lambda = \omega_j} &= \frac{1}{M_k} \sum_{m = 1}^{M_k} I_{x_{m,t}}(\omega_j) \\
\label{eq:M_periodogram_z}
&= \frac{1}{\sum_{i = 1}^{N} z_{ik} }\sum_{i = 1}^{N} z_{ik} I_{\mathbf{x}_{i,t}}(\omega_j), \quad j=0, \ldots, T-1.
\end{align}
For the EM framework we treat $\mathbf{z}_i$ as random variable with marginal distribution $\mathbb{P} \left( z_{ik} = 1 \right) = \pi_k$, also commonly referred to as \emph{mixing weights}. Rather than weighing periodograms with binary weights in \eqref{eq:M_periodogram_z}, the non-parametric frequency domain EM estimator for $\widehat{f}_{\mathcal{S}_k}(\lambda)$ weighs the periodogram of series $\mathbf{x}_{i,t}$ with the probability of coming from system $\mathcal{S}_k$, that is
\begin{align}
\label{eq:gamma_periodogram}
\widehat{f}_{\mathcal{S}_k}(\omega_j) &= \frac{1}{N_k}\sum_{i = 1}^{N} \gamma_{ik} I_{\mathbf{x}_{i,t}}(\omega_j),
\end{align}
where
\begin{equation}
\label{eq:gamma_ik}
\gamma_{ik} := \mathbb{P} \left( z_{ik} = 1 \mid \mathbf{x}_{i,t} \right),
\end{equation}
and $N_k = \sum_{i = 1}^{N} \gamma_{ik}$ is the effective number of time series from sub-system $\mathcal{S}_k$. As a by-result this new method also gives improved spectral density estimates.\\
For the frequency domain EM algorithm we treat the spectral density/periodogram of $\mathbf{x}_{i,t}$ as a pdf/pmf on the on the Fourier frequencies $\mathbf{\lambda}_i = \left( \lambda_{i,0}, \ldots, \lambda_{i,T-1} \right)$. Thus we compute \eqref{eq:gamma_ik} by the probability that ``model'' density $f_{\mathcal{S}_k}(\lambda)$ assigns to the empirical distribution function (edf) of the Fourier frequencies of $\mathbf{x}_{i,t}$ (= periodogram of $\mathbf{x}_{i,t}$), i.e.\
\begin{align}
\mathbb{P}\left( z_{ik} = 1 \mid \mathbf{x}_{i,t} \right) := \mathbb{P}\left( I_{\mathbf{x}_{i,t}}(\lambda) \text{ from } f_{\mathcal{S}_k}(\lambda) \right)
\end{align}
As we do not observe the Fourier frequencies $\mathbf{\lambda}_i$ we cannot compute likelihoods and probabilities such as $ \mathbb{P}\left( I_{\mathbf{x}_{i,t}}(\lambda) \text{ from } f_{\mathcal{S}_k}(\lambda) \right)$ directly. However, eq.\ \eqref{eq:loglik_equals_KL_ent} in the Appendix \ref{sec:KL_MLE} shows how to compute the log-likelihood of $\mathbf{\lambda}_i$ as a linear combination of the Kullback-Leibler (KL) divergence between $I_{\mathbf{x}_{i,t}}(\lambda)$ and $f_{\mathcal{S}_k}(\lambda)$, and the entropy of $I_{\mathbf{x}_{i,t}}(\lambda)$.
Thus the EM algorithm can be implemented as follows:
\begin{enumerate}
\setcounter{enumi}{-1}
\item Initialization: set $\tau = 0$ and randomly assign $\mathbf{x}_{i,t}$ to one of the $K$ sub-systems; set class probabilities $\gamma_{ik}^{(\tau)} := 1$ if $\mathbf{x}_{i,t} \in \mathcal{S}_k$; $0$ otherwise. Compute effective number of time series per cluster $N_k^{(\tau)} = \sum_{i=1}^{N} \gamma_{ik}^{(\tau)}$ and estimate mixing weights by $\widehat{\pi}_k^{(\tau)} = \frac{N_k^{(\tau)}}{N}$.
\item \label{item:start_iteration} Estimate $f_{\mathcal{S}_k}(\lambda)$ by a weighted average of the periodograms of $\mathbf{x}_{i,t}$:
\begin{equation}
\label{eq:est_f_j2}
\widehat{f}_{\mathcal{S}_k}^{(\tau)}(\omega_{j}) = \frac{1}{N_k^{(\tau)}} \sum_{i=1}^{N} \gamma_{i{k}}^{(\tau)} I_{\mathbf{x}_{i,t}}(\omega_{j}), \text{ for each } k =1, \ldots, K.
\end{equation}
This gives $K$ spectral densities $\lbrace \widehat{f}_{\mathcal{S}_1}^{(\tau)}, \ldots, \widehat{f}_{\mathcal{S}_K}^{(\tau)} \rbrace =: \mathcal{F}^{(\tau)}$ at iteration $\tau$. Note that for each $k$, $\mathbb{E} \widehat{f}_{\mathcal{S}_k}^{(\tau)}(\omega_{j}) = f_{\mathcal{S}_k}(\omega_{j})$ and $\mathbb{V} \widehat{f}_{\mathcal{S}_k}^{(\tau)}(\omega_{j})\approx \frac{ f_{\mathcal{S}_k}(\omega_{j}) }{N_k} \ll f_{\mathcal{S}_k}(\omega_{j})$.
\item \label{item:end_iteration} Compute KL divergence between each $I_{\mathbf{x}_{i,t}}(\omega_{j})$ and all $\widehat{f}_{\mathcal{S}_k}^{(\tau)} \in \mathcal{F}^{(\tau)}$:
\begin{equation}
\label{eq:KL_div_periodograms}
\KL{I_{\mathbf{x}_{i,t}}}{\widehat{f}_{\mathcal{S}_k}^{(\tau)}} = \sum_{j=0}^{T-1} I_{\mathbf{x}_{i,t}}(\omega_{j}) \log \frac{I_{\mathbf{x}_{i,t}}(\omega_{j})}{\widehat{f}_{\mathcal{S}_k}^{(\tau)}(\omega_{j})}, \quad \forall i, \forall k,
\end{equation}
and update conditional probability that series $\mathbf{x}_{i,t}$ comes from system $\mathcal{S}_k$
\begin{align}
\gamma_{ik}^{(\tau+1)} &=
\mathbb{P}\left( I_{\mathbf{x}_{i,t}}(\lambda) \text{ from } \widehat{f}_{\mathcal{S}_k}^{(\tau)} \right) \quad \forall i, \forall k
\end{align}
using \eqref{eq:KL_div_periodograms} and \eqref{eq:probability_from_loglik}. Update mixing weights
\begin{equation}
\widehat{\pi}_k^{(\tau+1)} = \frac{N_k^{(\tau+1)}}{N} \text{, where } N_k^{(\tau+1)} = \sum_{i=1}^{N} \gamma_{ik}^{(\tau+1)}.
\end{equation}
Set $\tau = \tau +1$.
\item Repeat steps \ref{item:start_iteration} and \ref{item:end_iteration} until convergence of the overall spectral likelihood
\begin{equation}
\label{eq:loglik_total}
\ell(\mathcal{S}_1, \ldots, \mathcal{S}_K; \pi_1, \ldots, \pi_K \mid \mathbf{x}_{1,t}, \ldots \mathbf{x}_{N,t}) = \sum_{i=1}^{N} \log \left( \sum_{k=1}^{K} \widehat{\pi}_k e^{\ell \left( I_{\mathbf{x}_{i,t}}(\omega) \mid \widehat{f}_{\mathcal{S}_k}(\omega) \right) } \right).
\end{equation}
\end{enumerate}
Since for unit-variance input $x_t$ the spectral density/periodogram are well-defined continuous/discrete probability distributions, this EM algorithm can be applied to both stationary as well as non-stationary signals: in the first case, the spectral density $f_x(\lambda)$ exists as a non-negative, square integrable function and a large part of the time series and econometrics literature is devoted to the spectral analysis of stationary time series \citep{Iacobucci03_spectralanalysiseconomics, Priestly81_SpectralAnalysis,Mathias04_Irregularspaced}; in the second case, the periodogram \eqref{eq:periodogram}, viewed as a purely data-driven method, represents a valid discrete pmf on $\lbrace \omega_j \rbrace_{j=0}^{T-1}$.
Since frequency domain analysis plays a very prominent and successful role in statistics, time series analysis, and signal processing, this frequency domain EM algorithm to detect similar dynamics or shape can be easily implemented and applied to a great variety of problems where the data has a spectral representation. For example, the method can be used for image clustering ($2D$ Fourier transform) as well as classification of a family of positive semi-definite random matrices $\lbrace A_i \rbrace_{i=1}^{N}$, $A_i \in \mathbb{R}^{T \times T}$ considering their normalized eigenvalues $\lbrace \lbrace \tilde{\lambda}_{j} \rbrace_{j=1}^{T} \rbrace_i$ as a discrete distribution on $j = 1, \ldots, T$.\\
It must be noted though that it comes with all the pros and cons of the basic EM algorithm (never decreasing likelihood, but possibly local optima). For a detailed account of convergence results and many other properties see \citet{McLachlanetal08_EM_book} or \citet{Bishop07_ML_book}.
\subsection{Choosing the number of clusters}
So far the number of clusters was fixed a-priori and the algorithm gives the (locally) best $K$-cluster solution. However, since this number is rarely known in practice, we must have a rule to select a good $K$. In some cases there is a ``true'' $K$. For example, the micro-electrode in the brain recorded a certain number of neurons. Thus there is an underlying truth which we try to estimate. In other cases, such as the economic time series example, there may not be a true number of sub-systems but choosing the number of clusters is based on convenience and ease of interpretation. One may choose only two clusters to show vastly contrary situations, and then compare this to a more refined structure by allowing more clusters.\\
While the EM algorithm achieves a (locally) optimal classification by maximizing the expected likelihood function, this criterion cannot be used to choose the optimal number of clusters: the likelihood is non-decreasing in $K$, thus maximizing the likelihood with respect to $K$ will always give $K \equiv N$; that is each time series is its own class. For parametric models one can use the BIC to choose $K$ \citep{BiernackiGovaert98_choosingmodels}, but for non-parametric settings this is not directly applicable. A common heuristic is the ``elbow rule'', where the number of clusters is determined by looking where the likelihood does not show a substantial increase anymore.
\citet{CeleuxSoromenho96_EntropyCriterion} propose an entropy based criterion to assess the optimal number of clusters. The \emph{normalized entropy criterion} (NEC) chooses that $K$ which minimizes
\begin{equation}
\label{eq:NEC}
NEC(K) = \frac{E(K)}{\ell^*(K) - \ell(1)}, K \geq 2,
\end{equation}
where $\ell^*(K)$ is the log-likelihood of the best $K$ cluster solution, and
\begin{equation}
\label{eq:classification_entropy}
E(K) = - \sum_{k=1}^{K} \sum_{i=1}^{N} \gamma_{ik} \log \gamma_{ik} \geq 0,
\end{equation}
is the entropy of the soft classification matrix. Since it is only based on the class probabilities and the log-likelihood, it can be easily computed even for non-parametric classification methods.
The entropy in \eqref{eq:classification_entropy} measures how well the best $K$ cluster partition can separate the data. In the case of perfectly separable classification, $\gamma_{ik} = 1$ for one $k$ and $0$ otherwise (for each $i$); in this case $E(K) = 0$. In practice, classification is not perfect, thus in general $E(K) > 0$. Hence it makes sense to choose that $K$ which minimizes $E(K)$ as this is as close as possible to a perfect separation. Since the baseline value of the likelihood changes for each $K$, the entropy is normalized by the optimal log-likelihood for each $K$. The optimal number of clusters is the one that minimizes \eqref{eq:NEC}. See \citet{Biernacki99_ImprovementNEC, CeleuxSoromenho96_EntropyCriterion} for details and simulation results.
It must be noted though that rather than looking at the global minimum, it is more useful to consider all local minima as possible candidates. Only focusing on the global minimum can lead to an under-estimation of the true order $K$. For example, sometimes a $K = 2$ cluster solution gives binary weights to each class - and thus $E(K) = 0$ - but can be far from representing the true number of clusters, as it averages over several clusters in one region of the space. Thus for simulations and applications I use the NEC rule combined with the ``elbow'' rule in the log-likelihood to choose an appropriate $K$.
\section{Fourier Analysis}
As every bounded, continuous function can be approximated arbitrarily well by Fourier polynomials of sufficiently high order (in the mean square sense), we can write each spike wave as
\begin{equation}
S(t) = \sum_{\ell=-\infty}^{\infty} \alpha_{\ell} e^{i \ell t}, \text{ with } \alpha_{\ell} = \overline{\alpha}_{\ell},
\end{equation}
where $\overline{z} = a - ib$ is the complex conjugate of the complex number $z = a+ib$. This must hold as $S_i(t)$ is a real-valued function; thus the coefficients of the Fourier series must be complex conjugates.
The Fourier coefficients can be computed by
\begin{equation}
\alpha_{\ell} = \frac{1}{2\pi}\int_{-\pi}^{\pi} S(t) e^{-i \ell t}\, dt.
\end{equation}
Since $S_i(t)$ and $\lbrace a_{\ell}^{(i)} \rbrace$ are also in a one-to-one relation we have
\begin{equation}
\text{ Neuron } n_i \leftrightarrow S_i(t) \leftrightarrow \boldsymbol{\alpha}^{(i)}
\end{equation}
Thus if we want to classify the signals we measure to which $n_i$ they are coming from, we can either classify the wave forms $S_i(t)$ or classify Fourier coefficients. The advantage of Fourier coefficients is that in general there will be only very few important coefficients (since the wave forms are already ``nice'' harmonics, and not random signals).
In general the function is not on $[-\pi, \pi]$: for a periodic function $f(t)$ on $[a, a+ \tau]$ the formulas become
\begin{equation}
f(x)=\sum_{\ell =-\infty}^\infty \alpha_{\ell} \cdot e^{i 2\pi \frac{\ell}{\tau} x}.
\end{equation}
and the coefficients are
\begin{equation}
\alpha_{\ell} = \frac{1}{\tau}\int_a^{a+\tau} f(x)\cdot e^{-i 2\pi \frac{\ell}{\tau} x}\, dx.
\end{equation}
For measured data and for ``nice'' wave forms, the infinite sum for spike-signal $S(t)$ can be split into the important frequencies\footnote{Note that the important frequencies are not necessarily the first frequencies in $j$.} $\mathcal{K}_i$ and the noise
\begin{eqnarray}
S_i(t) &=& \sum_{\ell \in \mathcal{K}_i} \alpha_{\ell} \cdot e^{i 2\pi \frac{\ell}{\tau} t} + \sum_{j \notin \mathcal{K}_i} a_n \cdot e^{i 2\pi \frac{\ell}{\tau} t} \\
&=& \sum_{\ell \in \mathcal{K}_i} \alpha_{\ell} \cdot e^{i 2\pi \frac{\ell}{\tau} t} + \varepsilon(t) \\
\end{eqnarray}
This decomposition transfers directly to the statistical model for measured signals $s_{j,t}$
\begin{equation}
s_{j,t} = \sum_{\ell \in \mathcal{K}_i} a_{\ell} \cdot e^{i 2\pi \frac{\ell}{\tau} t} + \varepsilon_t.
\end{equation}
\subsection{Estimating the spectral density}
Clustering the waves in the frequency domain also allows to use the well-developed theory of spectral densities of time series. For example, an for a white noise sequence each frequency is equally important ($\rightarrow$ uniform distribution); for the wave signal only very few frequencies are important. Furthermore, the important frequencies will be the low frequencies, not the high ones. This makes it easy to separate between signal and noise.
In practice the data of a wave form is a time series of length $T$, $(s_{i,1}, \ldots, s_{i,T})$; the Fourier coefficients at the Fourier frequencies $\omega_j = \frac{2 \pi j}{T}$ for $j = 0, \ldots, T-1$ are
\begin{equation}
X_j = \sum_{n=0}^{T-1} S_i(n) e^{-\frac{2 \pi i}{T} j n} \quad \quad j = 0, \dots, T-1
\end{equation}
The fast fourier transform (FFT) can achieve that computation in $\mathcal{T \log T}$ operations.
\begin{table}[!b]
\caption{\label{tab:notation} Notation used in this analysis.}
\begin{center}
\begin{tabular}{ | l | l | l | }
\hline
$n_i$ & Neuron $i$ & $i = 1, \ldots, K$; $K$ unknown\\
$\mathcal{N}$ & collection of all neurons $\mathcal{N} = \lbrace n_1, \ldots, n_K \rbrace$ & \\
\hline
$S_i(t)$ & spike-signal from $n_i$ & $i=1,\ldots, K$, $t \in \mathbb{R}$ \\
$\mathcal{S}$ & collection of all spike-signal wave forms $\mathcal{S} =\lbrace S_1(t), \ldots, S_K(t) \rbrace$ & \\
\hline
$\boldsymbol{\alpha}^{(i)} = \lbrace \alpha_{\ell}^{(i)} \rbrace \in \mathbb{C}$ & Fourier coefficients of $S_i(t)$ & $\ell = 0, \pm 1,\pm2 , \ldots$ \\
$\mathcal{A}$ & collection of all sequences $\mathcal{A} =\lbrace \boldsymbol{\alpha}^{(1)}, \ldots, \boldsymbol{\alpha}^{(K)} \rbrace$ & \\
\hline
$s_{j,t}$ & $N > K$ measured signals & $j = 1, \ldots, N$, $t = 1,\ldots, T$ \\
$\mathbf{a}^{(j)} = \lbrace a_{\ell}^{(j)} \rbrace$ & Fourier coefficients of $s_{j,t}$ & $\ell = 0, 1, \ldots, T-1$ \\
\hline
\end{tabular}
\end{center}
\end{table}
\begin{definition}[Discrete Fourier Transform]
For a sequence $(x_1, \ldots, x_T)$, define the discrete Fourier transform (DFT) as $(X(\omega_0), X(\omega_1), \ldots, X(\omega_{n-1}))$, where
\begin{equation}
X(\omega_k) = \frac{1}{\sqrt{T}} \sum_{t=1}^{T} x_t e^{-2 \pi i \omega_k t},
\end{equation}
and $\omega_k = \frac{k}{T}$ for $k=0, 1, \ldots, T-1$ are the Fourier frequencies.
\end{definition}
Since the Fourier basis $\lbrace e^{2 \pi i \omega_k t} \rbrace$ is an orthonomal sequence, the coordinates of $x_t$ in the Fourier basis are $X(\omega_k)$, i.e.\
\begin{equation}
x_t = \sum_{k=0}^{T-1} \innerprod{x_t, e_k} e_k = \sum_{k=0}^{T-1} X(\omega_k) e^{2 \pi i \omega_k t}.
\end{equation}
\section{Introduction}
Classification of similar signals is a widespread task in signal processing, where similar can either mean similar shape (for non-stationary signals) or similar dynamics (for stationary\footnote{A sequence of random variables (RVs) $\lbrace x_t \rbrace_{t \in \mathbb{Z}}$ is stationary if i) $\mathbb{E} x_t = \mu < \infty$, ii) $\mathbb{V} x_t := \mathbb{E} (x_t - \mu)^2 < \infty$ and iii) the auto-covariance function $\gamma(k):=\mathbb{E} (x_t - \mu) (x_{t-k} - \mu)$ is independent of $t$.} signals). Non-stationary examples are recordings of brain activity (see Section \ref{sec:spike_sorting}) or speech signals; stationary signals can be found in many economic or physical time series. In both cases, researchers want to detect similar dynamics:
\begin{description}
\item Neuro-scientists study the signal shape sent by neurons in order to understand how fast neurons send information across the brain. As a recording can contain signals from many different neurons, it is necessary to cluster them into signals of similar shape \citep{Quiroga04}, which were presumably sent by the same neuron.
\item In economics and public policy one is often interested in similar dynamics of the market/society to characterize, for example, how fast a country recovers from a recession, and how it compares to other countries in the region; or which countries have similar dynamics in their labor market.
\end{description}
Formally, let $\mathcal{X} = \lbrace \mathbf{x}_{1,t}, \ldots, \mathbf{x}_{N,t} \rbrace$ be a family of sequential observations from a dynamical system $\mathcal{S}$, where $\mathbf{x}_{i,t} = (x_{i,1}, \ldots, x_{i,T})$ is the individual time series of entity $i$. For example, $\mathcal{S}$ can be a particular area in the brain or the economic rules in the labor market. Here we consider systems which can be naturally divided into $K$ homogeneous sub-systems $\mathcal{S}_1, \ldots, \mathcal{S}_K$, each one with its own characteristic dynamics. In the neurology context these sub-systems $\mathcal{S}_k$ represent different neurons sending a signal; in economics $\mathcal{S}_k$ could correspond to different dynamics in the market, e.g.\ countries that recover fast from a recession ($\mathcal{S}_1$) versus countries that need more time to catch up again to global economy ($\mathcal{S}_2$).
Many clustering and dimension reduction techniques such as principal component analysis (PCA) \citep{Jolliffe02_PCAbook} focus on the mean and/or variance to make a reduction/classification in similar blocks of data. Yet these two statistics are irrelevant for the correlation over time; \citet{Keogh03_MeaninglessTSclustering} even claim that time series clustering is entirely meaningless. \citet{Simon05_UnfoldingTScluster} show that time series clustering is not meaningless \emph{per se}, but that the similarity measure must be chosen carefully. They embed each time series in a higher dimensional space of lagged variables, $x_t \rightarrow \tp{\left(x_{t-\tau_1}, x_{t-\tau_2} \ldots, x_{t-\tau_s} \right)} \in \mathbb{R}^{s}$, $0 \leq \tau_1 < \tau_2 < \ldots < \tau_s < T$, such that signals with different dynamics can be easily distinguished in the higher dimensional $\mathbb{R}^s$. This method works particularly well for long time series even with non-linear dynamics. If only few observations per series are available ($T \approx 100$ or even only $50$), then time-embeddings are extremely sparse in $\mathbb{R}^{s}$ and thus clustering becomes impractical.
For few observations per series it can be useful to first fit a parametric model $\mathcal{M}_{\theta_j}$ to every series $\mathbf{x}_{j,t}$, and then cluster in the lower-dimensional parameter space $\lbrace \widehat{\theta}_{1}, \ldots, \widehat{\theta}_{n} \rbrace$. For example, for the broad class of auto-regressive integrated moving average (ARIMA) models several approaches have been studied: \citet{Dhiral01_ClusteringARIMA} cluster ARIMA models based on the distance between their estimated coefficients; \citet{Piccolo90_ARIMAdistance} uses the Euclidean distance between their auto-regressive extensions as a metric on the invertible ARIMA model space; \citet{Maharaj00} present a hypothesis test to distinguish between two - not necessarily independent - stationary time series by comparing auto-regressive fits to the data. See \citet{Liao05} for a detailed survey. Although this works well for small $T$, it suffers from a model selection bias: if we pick the wrong model for just some of the series, then the clustering cannot be accurate anymore. Furthermore, if the models are not nested in some sense, then it is hard to compare the parameters of $\mathbf{x}_{j,t}$ to those of $\mathbf{x}_{i,t}$.\\
Here I propose a novel approach to clustering similar dynamics using frequency domain properties of the signals, which avoids the model selection bias and at the same time works even with few observations. Existing frequency domain classification methods are mostly based on defining a metric on the spectrum and then using a clustering algorithm based on the so-obtained distance (or similarity) matrix. \citet{CaiadoNunoPena06} use hierarchical clustering algorithm on the Euclidean distance between the log-spectra; \citet{Savvides08} use a distance measure on cepstral coefficients obtained from the log-spectra. The method proposed here differs from existing techniques as it treats the magnitude of the discrete Fourier transform (DFT) of signal $\mathbf{x}_{j,t}$ as a probability mass function (pmf) on the unit circle and thus leads to a natural classification by an adaptation of the well-known Expectation Maximization (EM) algorithm \citep{Dempster77_EM, Bishop07_ML_book}. Section \ref{sec:EM_spectral_density} describes a non-parametric version which avoids the model selection bias, but it can also be easily adapted to a parametric framework, e.g.\ to cluster time series within the ARIMA model class.
\subsection{Similar dynamics in socio-economic time series}
In macro-economics and public policy researchers are often interested in comparing economies/societies with each other. For example, annual unemployment rates over the course of several decades can show law changes or adaptations of economic interdependencies within a country as well as with the rest of the world.
Here I will consider the annual per-capita income growth rate of the ``lower 48'' in the US from $1958$ to $2008$ compared to the overall US growth rate
\begin{equation}
g_{j,t} := r_{j,t} - r_{US,t}, \quad j \in \lbrace \text{Alabama, $\ldots$, Wyoming} \rbrace,
\end{equation}
where $r_{j,t}$ is the annual growth rate of region $j$ (see Appendix \ref{sec:Data} for details). Clustering states according to similar economic dynamics can help to decide where to provide economic support to overcome a recession faster. For example, if certain states do not show any important dynamics on a 7-8 year period - which is typically considered the ``business cycle'' \citep{Halletetal08_EuropeanBusinessCycle, Iacobucci03_spectralanalysiseconomics} - then it might be more useful for to invest available money in those states that are heavily affect by these global economy swings.
This dataset has also been analyzed in \citet{Dhiral01_ClusteringARIMA}, who fit auto-regressive models of order $1$ ($AR(1)$) to the non-adjusted growth rates $r_{j,t}$ for pre-selected $25$ states, and then cluster them based on the different fits. Although this procedure gives useful results, it is very unlikely that different dynamics for each of the $48$ states only manifest themselves in a different $AR(1)$ coefficient. In particular, simple $AR(1)$ models cannot capture the business cycle dynamics which are clearly visible in the power spectra of the growth rates (even in the adjusted rates) - see Section \ref{sec:marco_econ_results}, Fig.\ \ref{fig:EM_spikes_periodogram}.
The non-parametric EM algorithm introduced in Section \ref{sec:EM_spectral_density} does not face this model selection bias, but can capture different cyclic components in all $48$ time series.
\subsection{Neuron identification - ``spike sorting''}
\label{sec:spike_sorting}
The human brain can be seen as a big information-processing and -storing unit. For example, the information we get from watching our environment must be carried from the eye to the visual cortex. As the visual cortex resides in the back of the brain neurons have to transmit information from the front to the back of the head, just for us to being able to make sense of what we see; set aside the neurons involved in executing our reaction to what we see. Every time a neuron transmits information it emits an electrochemical signal, which can be measured by an electrode put in the brain area of interest. Figure \ref{fig:data_recording_40_50} (top) shows a recorded signal $y_t$ with $73,500$ observations.\footnote{For a detailed description see Appendix \ref{sec:Data}.}.
\begin{figure}[!t]
\centering
\makebox{\includegraphics[width=0.8\textwidth]{images/data_recording_1_50.png}}
\caption{\label{fig:data_recording_40_50} Brain signal recording: (top) entire $49$ seconds; (bottom left) zoom to $[43.2,43.6]$ - the transition between extremely large spikes ($\sim 43.32$) and smaller spikes ($\sim 43.57$); (bottom right) zoom to $[43.26, 43.29]$ where two spikes become visible
\end{figure}
On a macro-level these measurements help to identify active areas of the brain which are involved in performing a particular task; e.g.\ for visual tasks the back of the brain shows up as an active area. Micro-level properties of neuron activity are also important: for example \citet{KassVentura01} analyze how fast neurons can send information - this is characterized by a neuron's ``firing rate''. To do this non-trivial task, however, one implicitly makes an important assumption: it is known which neuron sent which signal. Figure \ref{fig:data_recording_40_50} clearly shows that micro-electrodes cannot single out one neuron, but record a concatenation - and sometimes a superposition - of an unknown number of neurons $n_1, \ldots, n_K$ transmitting information plus a lot of background noise. Hence to successfully analyze the firing rates, it is necessary to
\begin{enumerate}
\item[i)] distinguish actual spikes from background noise, and
\item[ii)] identify and assign each signal to one particular neuron $n_k$, $k = 1, \ldots, K$ where the number of neurons $K$ is unknown: an electrode records as many neurons as there are in its local neighborhood.
\end{enumerate}
Part i) constitutes one of the core problems in signal processing \citep{ZhengmingXiaojun00, DaviesJames07}. Consequently there is an immense literature on signal/noise separation, especially in audio and speech processing \citep{Jangetal03_SingleChannel, Barry05}. For sub-problem ii) we can classify the observed spikes into classes of similarly shaped wave-forms. If these shapes actually correspond to one sole neuron $n_k$ or still to a collection of neurons, depends on whether each neuron has a unique wave form or not. Only if there exists such a one-to-one relation, we can determine the firing rates of each single neuron. Biochemical and physiological findings suggest that each neuron has its own unique wave-form, which can only vary slightly based on the state of the neuron. Thus it should be possible to classify neuron activity according to the form of the signal - the ``spike''. This classification task is commonly known as ``spike sorting'' \citep{Lewicki98,Kim06,Natakanietal01}.
A common and simple approach is performing PCA on the spikes, and then cluster the signals according to the PCA coefficients \citep{Wood04_Automaticspike}. Although generally there are far more spikes than observations per spike ($N \gg T$), still the first $2$-$3$ eigen-vectors of the low-rank correlation matrix capture most of the variation in the data. However, since PCA selects sources by the direction of maximum variance, it will classify low power firings from the same neuron as different neurons.
The frequency domain classification algorithm introduced here, builds on the relation between the shape of the signal and its Fourier coefficients. Similar shapes have similar Fourier coefficients and thus clustering in the frequency domain should reveal these sub-classes.
\section{KL divergence and maximum likelihood}
\label{sec:KL_MLE}
Let $p_k := \mathbb{P}(X = a_k)$ define a probability distribution for the RV $X$ taking values in the finite alphabet $\mathcal{A} := \lbrace a_1, \ldots, a_K \rbrace$. The Kullback-Leibler (KL) divergence between two discrete probability distributions $p = \lbrace p_1, \ldots, p_K \rbrace$ and $q = \lbrace q_1, \ldots, q_K \rbrace$
\begin{equation}
\KL{p}{q} := \sum_{i=1}^{K} p_i \log_{2} \frac{p_i}{q_i} = \mathbb{E}_p \log_{2} \frac{p_i}{q_i}
\end{equation}
measures how far $p$ is from the ``truth'' $q$; in particular, if $p = q$ then $\KL{p}{q} = 0$.
Let $\tilde{p}(x)$ be the empirical distribution function (edf) of a sample $\mathbf{x} = (x_1, \ldots, x_N)$
\begin{equation}
\label{eq:empirical_distribution}
\tilde{p}(\mathbf{x}) := \frac{1}{N} \sum_{n=1}^{N} \delta(x - x_n),
\end{equation}
where $\delta(y)$ is the Dirac delta function, and let $p(\mathbf{x} \mid \theta)$ be a model (distribution) for the RV $X$. The maximum likelihood estimator (MLE) is that $\theta$ which maximizes the log-likelihood of the data (assuming iid)
\begin{equation}
\label{eq:loglik}
\ell( \theta \mid \mathbf{x}) = \sum_{n=1}^{N} \log p(x_n \mid \theta).
\end{equation}
In terms of the KL divergence it is intuitive to select that $\theta$ which minimizes the distance between the empirical distribution of the data, $\tilde{p}(\mathbf{x})$, and the model $p(\mathbf{x} \mid \theta)$. In fact, they are equivalent since
\begin{eqnarray}
\label{eq:KL_entropy_likelihood}
\KL{\tilde{p}(\mathbf{x})}{p(\mathbf{x} \mid \theta)} = \int \tilde{p}(\mathbf{x}) \log \frac{\tilde{p}(\mathbf{x})}{p(\mathbf{x} \mid \theta)} d \mathbf{x} = - H(\tilde{p}(\mathbf{x})) - \int \tilde{p}(\mathbf{x}) \log p(\mathbf{x} \mid \theta) d \mathbf{x},
\end{eqnarray}
where $H(\tilde{p}(\mathbf{x})) = - \int \tilde{p}(\mathbf{x}) \log \tilde{p}(\mathbf{x}) d \mathbf{x} $ is the entropy of $\tilde{p}(\mathbf{x})$, which is independent of $\theta$. Thus
\begin{equation}
\label{eq:KL_equiv_loglik}
\arg \min_{\theta} \KL{\tilde{p}(\mathbf{x})}{p(\mathbf{x} \mid \theta)} = \arg \max_{\theta} \mathbb{E}_{\tilde{p}} \log p(\mathbf{x} \mid \theta).
\end{equation}
Plugging \eqref{eq:empirical_distribution} in the right hand side of \eqref{eq:KL_equiv_loglik} shows the equivalence of KL divergence minimization and log-likelihood maximization as
\begin{eqnarray}
\mathbb{E}_{\tilde{p}} \log p(\mathbf{x} \mid \theta) &=& \frac{1}{N} \int \sum_{n=1}^{N} \delta(x - x_n) \log p(x \mid \theta) dx = \frac{1}{N} \sum_{n=1}^{N} \log p(x_n \mid \theta) \\
\label{eq:1_N_loglik}
&=& \frac{1}{N} \ell( \theta \mid \mathbf{x}).
\end{eqnarray}
Conversely the log-likelihood of $\mathbf{x}$ can be computed by
\begin{eqnarray}
\label{eq:loglik_equals_KL_ent}
\ell( \theta \mid \mathbf{x}) = -N \cdot \left[ \KL{\tilde{p}(\mathbf{x})}{p(\mathbf{x} \mid \theta)} + H(\tilde{p}(\mathbf{x})) \right],
\end{eqnarray}
and consequently
\begin{equation}
\label{eq:probability_from_loglik}
\mathbb{P} \left( \mathbf{x} \mid \theta \right) = e^{\ell( \theta \mid \mathbf{x})}.
\end{equation}
Equations \eqref{eq:loglik_equals_KL_ent} and \eqref{eq:probability_from_loglik} play a key role in the non-parametric EM algorithm defined on the power spectra, as they allow to compute $\ell( \theta \mid \mathbf{x})$ even though $\mathbf{x}$ has not been observed directly, but just its edf $\tilde{p}(\mathbf{x})$ and a model $p(\mathbf{x} \mid \theta)$. In this frequency domain framework, the data $\mathbf{x}$ are the unobserved Fourier frequencies $\omega_0, \ldots, \omega_{T-1}$, the edf $\tilde{p}(\mathbf{x})$ is the periodogram $I_{T,x}(\omega_k)$, and the ``true'' model $p(\mathbf{x} \mid \theta)$ is the EM estimate $\widehat{f}_{\mathcal{S}_k}(\lambda) \mid_{\lambda = \omega_j}$ of the spectral density of sub-system $\mathcal{S}_k$ - see \eqref{eq:est_f_j2}. Thus the conditional probability $\gamma_{ik} = \mathbb{P}(z_{ik} = 1 \mid \mathbf{x}_{i,t})$ can be computed by
\begin{equation}
\gamma_{ik} = \frac{e^{\ell \left( I_{\mathbf{x}_{i,t}}(\lambda) \mid \widehat{f}_{\mathcal{S}_k}(\omega) \right) }}{\sum_{k=1}^{K} e^{\ell \left( I_{\mathbf{x}_{i,t}}(\lambda) \mid \widehat{f}_{\mathcal{S}_k}(\omega) \right) }}, \quad i =1, \ldots, N \text{ and } k = 1, \ldots, K.
\end{equation}
\section{``Spike sorting'' in the time domain}
Let $\mathcal{N}$ be the set of all neurons and assume that each neuron $n_i \in \mathcal{N}$ has a unique characteristic spike $S_i(t) \in \mathcal{C}[a,b]$, where $ \mathcal{C}[a,b]$ is the set of all continuous functions
on $[a,b]$. The spike is unique in the sense that $n_i = n_j$ if and only if $S_i(t) = S_j(t)$, or put in words if we see two different spikes, then we know that two different neurons were active and vice versa.
The micro-electrode only records a subset of neurons $n_k$, $k = 1, \ldots, K$, where $K$ is unknown. In Section \ref{sec:spike_detection} I use a slowness measure to distinguish between signal and noise, and in Section \ref{sec:features} I fit a Gaussian mixture model (GMM) to the slowness to detect different neurons, based on the assumption that a every different spike shape has its characteristic slowness.
\subsection{Spike detection}
\label{sec:spike_detection}
Given the recorded signal $y_t$ it is necessary to extract windows of size $T$ containing a spike.\footnote{Since these extracted windows containing a spike will later on be used as the $N$ time series $\lbrace \mathbf{x}_{1,t}, \ldots, \mathbf{x}_{N,t} \rbrace$ of length $T$ to the classification algorithm, I also use $T$ here to denote the window size.} These signals $s_{j, t}$ of length $T$ represent the family of sequential observations $\mathcal{X} = \lbrace s_{1, t}, \ldots, s_{N, t} \rbrace \in \mathbb{R}^{T \times N}$, where $N$ is the number of detected spikes. As the entire micro-electrode recording is much longer than the length of one single spike there are far more extracted spikes than number of observations per signal ($N \gg T$). Since the electrode only records signals in its local neighborhood, we can also expect a small number of sub-systems (neurons) $\mathcal{S}_1, \ldots, \mathcal{S}_K$ of similar shape ($K \ll N$). The size of the window must match the length of a typical spike: the lower right panel of Fig.\ \ref{fig:data_recording_40_50} suggests that a typical spike lasts for about $0.0035$ seconds $\approx$ $55$ time steps (vertical red lines). Thus for the rest of this section I set $T = 55$.
Since we do not know a-priori where a spike occurs we need a rule that tells us where to look for it. Whereas characterizing spikes visually is easy, designing a quantitative automated rule that can describe spikes is much more difficult. A common approach \citep{Quiroga04} is to set a threshold value $tol$ and a spike is detected if the signal exceeds this threshold. This threshold rule will not only be very sensitive to outliers, but also bias the selected spikes in favor of spikes with large variance (power). Furthermore neurons sometimes fire with lower power than usual, and thus may not exceed such a threshold. Although missing these spikes would not affect the spike sorting algorithm, it will underestimate the firing rate of neuron $n_k$.
Here I characterize ``non-spikes'', i.e.\ noise, in a way that detects spikes according to properties of the entire signal, not of one single observations (such as the threshold rule). One way to characterize noise is that it is moving much faster than any spike - whatever such a spike may look like. \citet{Berkes05} introduced a measure of slowness for a signal $x_t$, defined as the variance of the differenced, unit-variance signal
\begin{equation}
\label{eq:slowness}
\Delta(x_t) = \mathbb{V} \left( x_t - x_{t-1} \right), \quad \mathbb{V} x_t = 1.
\end{equation}
For an independent identically distributed (iid) signal $\varepsilon_t$ the slowness satisfies $\Delta\left(\varepsilon_t\right) = 2$. On the other hand, if $x_t \rightarrow const$ then $\Delta(x_t) \rightarrow 0$. Therefore, the larger $\Delta(x_t)$, the faster $x_t$.
Computing the slowness of the signal in a sliding window over $y_t$ reveals noisy parts (fast) and - complementary - the spikes (slow). The red (right) histogram in Fig.\ \ref{fig:MonteCarlo_slowness} shows simulated $\Delta\left( \varepsilon_t \right)$, where $\varepsilon_t \stackrel{iid}{\sim} \mathcal{N}(0,1)$ with $t = 1, \ldots, T = 55$ for $N = 10,000$ replications. Clearly, the central limit theorem (CLT) comes into play and the simulated values are centered around their true slowness $\Delta\left(\varepsilon_t\right) = 2$.
However, there is no obvious reason to assume that the brain background noise in the neighborhood of the micro-electrode is necessarily iid. In fact, the empirical slowness (blue histogram) of the sliding windows is substantially smaller than $2$, showing that brain background noise is not iid.\footnote{Since $\Delta\left( \varepsilon_t \right) = \Delta\left( const \cdot \varepsilon_t \right)$ by definition ($\mathbb{V} x_t \equiv 1$ in \eqref{eq:slowness}), the lower slowness for the brain signal is not due to a lower variance white noise sequence, but indeed a manifestation of some dependence in the data.}
But even though we do not know how slow it is, we know - and can clearly see in Fig.\ \ref{fig:MonteCarlo_slowness} (bottom) - that noise moves much faster than any of the spikes: $\Delta(\text{background noise}) \gg \Delta\left( \text{any spike} \right)$. Hence, we can learn the boundary value that distinguishes noise and spikes from the data. At this stage we are only concerned with separating spikes from noise, thus we can choose a conservative value for the boundary. If it turns out that this still includes too much noise, then a clustering algorithm will put these falsely extracted ``spikes'' in a \emph{noise} class. On the other hand, a too small boundary will miss spikes and thus bias the analysis of firing rates towards larger firing intervals. The lower panel of Figure \ref{fig:MonteCarlo_slowness} suggests that $tol = 0.25$ provides a good separation between noise on the right and spikes on the left.
\begin{figure}[!t]
\centering
\subfloat[(top): (red) Simulation of $\Delta \left(\varepsilon_t\right)$ with $10,000$ replications for iid $\lbrace \varepsilon_t \rbrace_{t=1}^{T=55}$; (blue) empirical slowness of the rolling window over the data $y_t$. \newline (bottom): zoom into $(0, 0.4)$ with the boundary between spikes ($ < 0.25$) and noise ($> 0.25$).]{\includegraphics[width=.45\textwidth]{images/MonteCarlo_slowness}\label{fig:MonteCarlo_slowness} }
\hspace{1cm}
\subfloat[(top) detected spikes; (below) their $\log$ slowness and a Gaussian mixture fit with $6$ components (chosen according to BIC score)]{\includegraphics[width=.45\textwidth]{images/extracted_spikes}\label{fig:extracted_spikes} }
\caption{How slowly do we think? - the slowness of brain recordings.}
\label{fig:Flatness}
\end{figure}
This rolling window approach gives the so called on-set times (the moment a neuron fires and the spike lasts for $T = 55$ units of time), which are then used to extract possible spikes $s_{j,t}$ from $y_t$. An additional alignment step takes place to avoid slight misplacements of the onset times; following the spike sorting literature, this was done by identifying the maximum of each spike and adjust the window such that all signals have their maximum at the same position.
Figure \ref{fig:MonteCarlo_slowness} shows $n = 1,747$ extracted signals of length $T = 55$ obtained by applying the slowness measures on a sliding window using $tol = 0.25$. The rolling window spike detection could not exclude noise completely, so in the upper left panel of Fig.\ \ref{fig:extracted_spikes} one can visually distinguish $2$-$3$ spikes and some noise. Even though the low-variance signals seem to be noise, they could also be just low-power spikes. Since the slowness measure is invariant to scaling, it does not falsely ignore low power signals.
Before applying a standard classification algorithm on the extracted signals in Section \ref{sec:Results}, I first describe the main contribution of this work.
\section{Applications}
\label{sec:Results}
In this section I demonstrate the usefulness and wide applicability of the presented methods on and income growth (stationary) and neuron spike train (non-stationary) data.
\subsection{States with similar income dynamics in the US}
\label{sec:marco_econ_results}
First, all $48$ (more or less) stationary series $\mathbf{x}_{i,t}$ were transformed via the DFT to get the raw periodograms $I_{\mathbf{x}_{i,t}}(\omega)$, $i=1, \ldots, 48$. Without any further smoothing all $K$-cluster models for $K = 1, \ldots, 6$ were fitted to the data and both the $NEC(K)$ and the log-likelihood suggest that $K = 3$ clusters provide a good fit. The upper row in Fig.\ \ref{fig:US_growth_1960_2008} shows the periodograms of the three classes and the estimate $\widehat{f}_{\mathcal{S}_k}(\lambda)$ (black line) using \eqref{eq:gamma_periodogram}. The x-axis represents the Fourier frequencies $\omega_j$, which have been re-scaled from $[0, \pi]$ to $[0,0.5]$ for easier interpretation. Peaks at frequency $\omega_j$ mean that periods of length $1/\omega_j$ are important for the variation in the data. For example, the blue series show two important low frequencies (long cycles): $\omega \approx 0.04$ and $\omega \approx 0.18$. They correspond to a cycle of $25$ years and $5$-$6$ years -- which represent a generation cycle and a (short) business cycle \citep{Tylecote94_Longcycles}. Note that $AR(1)$ models \citep{Dhiral01_ClusteringARIMA} may be appropriate for the red dynamics ($AR(1)$ coefficient slightly negative), but cannot capture two cycles as shown in the blue and green periodograms.\\
\begin{figure}[!t]
\centering
\makebox{\includegraphics[width=0.9\textwidth]{images/US_growth_1960_2008_3clusters}}
\caption{\label{fig:US_growth_1960_2008} Non-parametric, frequency domain EM detects $3$ dominant dynamics of per-capita income growth (top); (left) normalized entropy from \eqref{eq:NEC} as a function of $K$; (center) spectral log-likelihood \eqref{eq:loglik_total} at the optimum for each $K$; (right) color-coded US map where red/green/blue intensity equals the conditional probability $\gamma_{nk}$ (RGB = ($\widehat{\gamma}_{n1},\widehat{\gamma}_{n2}, \widehat{\gamma}_{n3}$)).}
\end{figure}
The spatial connectivity of the obtained clusters confirms the good model fit as it separates US economy $\mathcal{S}$ in three major sub-economies/regions:\footnote{Any resemblance of the RGB color system to politics is purely coincidental.}
\begin{description}
\item[$\approx$ East \& Rockies \& CA (blue):] economy is highly persistent, changes are slow; business cycle of $\approx 5$-$6$ years is also important.
\item[$\approx$ South-West (green):] also highly persistent and affected by global business cycle of $7-8$ years (peak at $\omega \approx 0.13$).
\item[$\approx$ Mid-West (red):] almost flat spectrum, high frequencies (short cycles) are slightly more important; decoupled from global business cycle.
\end{description}
One possible explanation why the red states have a flat spectrum, is that they are mostly agricultural states, and since people have to eat no matter how the global economy is doing, the red states' income is not affected too much by recession or other market fluctuations. On the contrary, states whose economy - and thus income - relies heavily on industry, production, or technology are more affected by global economy swings, which typically happen every $7$-$8$ years.\\
Hence, the classification map in Fig.\ \ref{fig:US_growth_1960_2008} can provide a basis for more effective policies to boost local economies facing a recession: it might be more effective to allocate main parts of public investments to states that are actually affected by the business cycle, and not put it in states which are decoupled from global economy.
\subsection{Spike sorting}
\label{sec:features}
For the neuron classification we can either try to fit a mixture model directly on the $T$ - dimensional data, or compute ``features'' for each spike that summarize its shape. A good feature selection will reduce the dimensionality of the data, and thus greatly accelerate computations.
Here I will cluster both in the time and frequency domain: for the first I fit a Gaussian mixture model (GMM) on the logarithm of the slowness of each spike, $\log \Delta \left( s_{j,t} \right)$; for the second I use the frequency domain EM algorithm on the power spectra induced by each spike $s_{j,t}$.
\subsubsection{Gaussian mixture model on slowness}
The histogram in Fig.\ \ref{fig:extracted_spikes} of $\lbrace \log \Delta \left( s_{j,t} \right) \rbrace_{j=1}^{1,747}$ shows 5-6 peaks, which presumably correspond to 5-6 differently shaped spikes. Thus I fit GMMs to $\log \Delta \left( s_{j,t} \right)$ and assign each spike $s_{j,t}$ to the cluster with highest a posteriori probability. Table \ref{tab:gaussian.mix} shows parameter estimates of the $6$ component model, which was chosen according to the highest BIC score (Fig.\ \ref{fig:extracted_spikes}) from all GMMs up to order $10$.\footnote{To avoid local maxima, I ran the EM algorithm (package \texttt{mixtools} in R) $100$ times for each $K$ and chose the largest local optimum solution in each run.}
\begin{table}[!t]
\begin{center}
\caption{\label{tab:gaussian.mix} EM estimates of a $6$ component GMM for $\log \Delta(s_{j,t})$}
\begin{tabular}{rrrrrrr}
\hline
& Comp 1 & Comp 2 & Comp 3 & Comp 4 & Comp 5 & Comp 6 \\
\hline
\hline
$\pi_k$ & 0.069 & 0.218 & 0.093 & 0.511 & 0.078 & 0.031 \\
\hline
$\mu$ & -3.285 & -2.766 & -2.331 & -2.171 & -1.671 & -1.442 \\
$\sigma^2$ & 0.155 & 0.171 & 0.037 & 0.156 & 0.125 & 0.042 \\
\hline
\end{tabular}
\end{center}
\end{table}
The corresponding spikes are shown in the upper right panel of Fig.\ \ref{fig:extracted_spikes}. As $tol = 0.25$ was too conservative, two shapes still represent noise, and GMM identifies $K=4$ different neurons.
\subsubsection{Clustering in the frequency domain}
After time-domain techniques, I use the frequency domain EM algorithm described in Section \ref{sec:EM_spectral_density}. An additional advantage of working in the frequency domain compared to the time-domain is that misalignment of the spikes does not affect the clustering.
\begin{figure}[!t]
\centering
\makebox{\includegraphics[width=0.75\textwidth]{images/EM_spikes_periodogram_6clusters_original_data}}
\caption{\label{fig:EM_spikes_periodogram} EM on periodograms of spike signals $s_{j,t}$ with $K = 6$ clusters.}
\end{figure}
Also here I fit all mixture models up to order $K = 9$. In this case $NEC(K)$ achieves a global minimum at $K = 2$ and is monotonically increasing (not shown here). However, the two cluster solution is only optimal in the sense that it separates perfectly between spikes and noise, even though there is a relevant sub-classification within all spikes (similar to the behavior of $NEC(K)$ in the simulations). Hence here I use the ``elbow'' rule in the log-likelihood to determine the number of clusters. The most prominent ``elbow'' occurs at $K = 3$ (Fig.\ \ref{fig:EM_spikes_periodogram}) followed by another level-shift at $K = 6$. Since $K = 6$ was also by the BIC for the time-domain classification, I choose $K = 6$ for easier comparison. The $K = 6$ cluster solution reveals five spikes and one noise class (green shapes). Thus compared to the time-domain technique the frequency domain version detects one more spike.
\section{Simulations}
\label{sec:simulations}
This section shows how the methods perform on simulated data. In particular, I consider $K = 5$ sub-systems consisting of both stationary and non-stationary series: one white noise sequence (flat spectrum), two $AR(1)$ processes with $\phi = 0.5$ and $0.75$ respectively, and two sine waves with frequencies $\omega = 0.1$ and $0.2$ (on the $[0, 0.5]$ scale) corrupted by additive Gaussian noise. For each model I generate $n = 100$ series with $T = 50$ observations each. All series have been scaled to zero mean and unit variance.
\begin{figure}[!t]
\centering
\subfloat[ (left) sample time series from each of the $5$ groups: (1) white noise, (2) $AR(1)$ with $\phi = 0.5$, (3) $AR(1)$ with $\phi = 0.75$, (4) $\sin(1/10 2 \pi t / T)$, (5) $\sin(1/20 2 \pi t / T)$; (right) corresponding spectra.]{\includegraphics[width=.45\textwidth]{images/sample_series5}\label{fig:sample_series5} }
\hspace{1cm}
\subfloat[(top) log-likelihood and NEC as a function of $K$; (bottom-left) conditional class probabilities $\gamma_{nk}$; (bottom-right) logarithm of estimated cluster centers equaling optimal estimates $\widehat{f}_{\mathcal{S}_k}(\lambda)$, $k = 1, \ldots, 5$. ]{\includegraphics[width=.45\textwidth]{images/simulation_5series_results_loglik}\label{fig:simulation_5series_results} }
\caption{Simulation results for the frequency domain EM algorithm.}
\label{fig:simulation_study}
\end{figure}
Figure \ref{fig:sample_series5} shows a representative series of each sub-system. All corresponding non-smoothed periodograms have high variance (not consistent estimate). The nonparametric EM can be directly applied to these raw periodograms to cluster the $500$ time series.
The ``elbow'' rule for the log-likelihood favors a $K = 3$ solution, because separating the signals into the two non-stationary signals plus all stationary signals in the third class provides the largest gain in likelihood. The additional likelihood gain by separating the stationary signals into their sub-systems is negligible and thus is not evident in a plot of the log-likelihood as a function of $K$. The NEC has a global minimum at $K = 2$ (one sine wave plus rest) and a local minimum at $K = 5$. The log-likelihood clearly shows that $K = 2$ can not be an optimal separation, thus we take the $K = 5$ local minimum. Figure \ref{fig:simulation_5series_results} shows a very good separation between all signals, except for some cross-matches between the white noise sequence and the two $AR(1)$. However, as the parameters are close to each other and due to the small sample size ($N = 50$), some overlap between them can not be avoided - even using the true model and an MLE $\widehat{\phi}_{MLE}$ to cluster them (see below).
\subsection{Comparison to model based clustering}
For comparison I also fit $AR(1)$ and $ARMA(1,1)$ models\footnote{The series $x_t$ is an auto-regressive moving average process of order $(1,1)$ if it satisfies $x_t - \phi x_{t-1} = \varepsilon_t - \theta \varepsilon_{t-1}$, where both parameters $\phi$ and $\theta$ must lie in $(-1,1)$ to guarantee stationarity and invertibility.} to each series. Figure \ref{fig:simulation_series5_model_clustering} shows the separation of the series in the parameter space $\phi \in (-1,1)$ and $(\phi, \theta) \in (-1,1) \times (-1,1)$. Using the $AR(1)$ model not only gives a large overlap between the non-stationary signals and the stationary $AR(1)$, $\phi = 0.5$, but also completely fails to distinguish between the two harmonic signals. Even if the true signal is an $AR(1)$, model based clustering still has many falsely classified signals. The overlap in the fitted parameters $\widehat{\phi}_{MLE}$ show that the bad performance of the frequency domain EM for the $AR(1)$ series is not due to the algorithm, but results from the true parameters of distinct $AR(1)$ being very close ($0.5$ and $0.75$). In this case even the maximum likelihood estimator (MLE) provides wrong conclusions.
Extending the $AR(1)$ to $ARMA(1,1)$ models improves the separability between the two harmonic series, but also leads to additional variation in other regions of the parameter space. In particular, the black dots around the 45-degree line show that avoiding the model bias by simply using a larger model class introduces another problem of model based clustering. Here the model class is an $ARMA(1,1)$, but the true process is white noise, which is a special case of an $ARMA(1,1)$ for $\phi = \theta = 0$. However, every $ARMA(1,1)$ with $\phi = \theta$ also describes a white noise process, thus the MLE finds optimal solutions along the $\phi = \theta$ line and thus adds artificial variance - and thus performance loss for the clustering.\\
The exploratory analysis of the AR and ARMA models is an example of how the model selection bias can undermine clustering algorithms. For a good classification we would need to identify the correct model for each series first, and then estimate the parameters on each tuned model. However, even if we had the time and resources to do a model check for all $N$ time series, the $AR(1)$ example shows that even if we found the true model for each series, a large overlap $\widehat{\phi}_{MLE}$ remains (red triangles and green diamonds in the left panel of Fig. \ref{fig:simulation_series5_model_clustering}).
The non-parametric EM approach, on the other hand, does not require any modeling and subsequent checks, and has comparable performance to the model based clustering if we knew the true model (white noise and $AR(1)$) and performs much better if the models are wrong (sine waves versus rest).
\begin{figure}[!t]
\centering
\makebox{\includegraphics[width=0.7\textwidth]{images/simulation_series5_model_clustering}}
\caption{\label{fig:simulation_series5_model_clustering} Model based classification: (left) $\widehat{\phi}$ (y-axis) from fitting an $AR(1)$ to all series; (right) estimate pair $(\widehat{\phi}, \widehat{\theta})$ from fitting an $ARMA(1,1)$ model to each series. Colors and shapes represent the true classes, not estimated clusters from the data.}
\end{figure}
\section{Data}
\label{sec:Data}
\paragraph{Spikes:} The \texttt{PKdata} data set can be obtained from \url{www.biomedicale.univ-paris5.fr/physcerv/C_Pouzat/Data.html}. It contains recordings of the electro-chemical signal in the cerebral slice of a rat. A band pass filter for frequencies between $300$ Hz and $5$ kHz has been applied to the signal $y_t$, which was sampled at a rate of $15$ Khz for $1$ minute.
\paragraph{Income:} The dataset can be obtained from \url{www.bea.gov/regional/spi}. It contains yearly ($1958 - 2008$) average per-capita income of the ``lower $48$'' and the entire US: $I_{j,t}$, $j = 1, \ldots, 49$. As $I_{j,t}$ grew exponentially over time, one typically considers income growth rates $r_{j,t} = \log I_{j,t} - \log I_{j,t-1}$ - also known as log-returns - which are (more or less) stationary. Since we are interested in the individual dynamics of a state compared to the US, I analyze the difference between each state's growth rate to the US, as this is a more refined indicator of the state's dynamics (it removes the overall seasonal dynamics of the US baseline).
|
\section{Introduction}
Despite its deceptively simple definition, percolation provides a wealth of interesting problems that have kept physicists and mathematicians busy for over fifty years \cite{BroadbentHammersley}. In bond percolation, which is the focus of this paper, we declare each bond of a lattice to be open with probability $p$ and closed with probability $1-p$, resulting in a random distribution of connected clusters. In the limit of an infinite lattice, there is a sharp critical probability, $p_c$, at which an infinite cluster first appears. The determination of $p_c$ is an unsolved problem except for a few 2-dimensional cases. There are also many other quantities of interest, and recently, arguments from conformal field theory \cite{Belavin} and the discovery of Schramm-Loewner evolution \cite{Lawler,Gruzberg2006} have facilitated some spectacular calculations in the continuum limit, in which details of the underlying lattice disappear (\cite{Cardy92,Simmons2007,Simmons2009} are just a few examples). However, many new lattice-level results have also appeared in recent years, including exact calculations of percolation thresholds \cite{Scullard06,Ziff06,ScullardZiff06,Spakulova,Bollobas2006,Bollobas} as well as other rigorous \cite{Sedlock,Riordan,Smirnov} and numerical \cite{HajiAkbari,Becker,Ziff2009,Feng} results at the critical point. This paper describes an approximation scheme that allows very accurate determination of bond percolation thresholds for arbitrary periodic lattices. It was originally discussed in \cite{Scullard08}, and expanded in \cite{Scullard10} where thresholds were estimated for all but two of the Archimedean lattices to within $10^{-5}$ of the numerically determined values. I complete this program here by computing the approximate thresholds for the $(4,6,12)$ and $(3^4,6)$ lattices. In addition, I show that the approximations can be refined, leading to a sequence of polynomials that give successively better predictions for the bond thresholds. Finally, I discuss how this refinement provides significant clues about what the eventual solutions of these problems might look like.
\begin{figure}
\includegraphics{4-6-12.eps}
\caption{The $(4,6,12)$ lattice with the assignment of probabilities on the unit cell.} \label{fig:4612}
\end{figure}
\begin{figure}
\includegraphics{3_4-6.eps}
\caption{The $(3^4,6)$ lattice with the assignment of probabilities on the unit cell.} \label{fig:346}
\end{figure}
\section{Critical polynomials}
Every two-dimensional lattice for which the percolation threshold is known exactly is in the category of 3-uniform hyperlattices \cite{Bollobas}. An example of such a hyperlattice is shown in Figure \ref{fig:3uniform}. The unit cell of the graph is contained between three vertices, but the shaded triangle can represent any connected network of sites and bonds. The critical points of all these problems are given by the Ziff criterion \cite{Ziff06},
\begin{equation}
P(A,B,C)=P(\overline{A},\overline{B},\overline{C}), \label{eq:zc}
\end{equation}
where $P(A,B,C)$ is the probability that all vertices $(A,B,C)$ can be connected within a shaded triangle, and $P(\overline{A},\overline{B},\overline{C})$ is the probability that all three are disconnected. Thus, the critical threshold, which marks the appearance of an infinite connection, can be located by comparing probabilities of events on a single unit cell. This is a very special property that is not shared by most lattices. The result is that in this class, every bond threshold, for example, is the root in $[0,1]$ of a polynomial of degree at most $n$ with integer coefficients, where $n$ is the number of bonds in the unit cell. That is, at present, all exactly known thresholds are algebraic numbers.
The criterion (\ref{eq:zc}) provides the opportunity to derive not only critical thresholds, but also critical surfaces for inhomogeneous percolation (also called anisotropic percolation in some applications \cite{Turban}). Here, each bond, $i$, of the unit cell has its own probability, $p_i$, and applying (\ref{eq:zc}) leads to a condition of the form
\begin{equation}
f(p_1,p_2,...,p_n)=0 \label{eq:surface}
\end{equation}
where none of the $p_i$ appears as a power greater than one. For example, for the square lattice, with probability $p_1$ on the vertical bonds and $p_2$ on the horizontal bonds (Figure \ref{fig:SHM}(a)), we have,
\begin{equation}
\mathrm{S}(p_1,p_2)=1-p_1-p_2=0 \label{eq:square}
\end{equation}
and for the 3-probability honeycomb lattice (Figure \ref{fig:SHM}(b)),
\begin{equation}
\mathrm{H}(p_1,p_2,p_3)=1-p_1 p_2 -p_2 p_3 - p_1 p_3 + p_1 p_2 p_3=0 . \label{eq:honeycomb}
\end{equation}
A more complicated example is the martini lattice \cite{Scullard06,Fendley}, which is derived in \cite{Scullard10} and shown in Figure \ref{fig:SHM}(c),
\begin{eqnarray}
\mathrm{M}(p_1,p_2,p_3,r_1,r_2,r_3) &=& 1- p_1 p_2 r_3 - p_2 p_3 r_1 - p_1 p_3 r_2 - p_1 p_2 r_1 r_2 \nonumber \\
&-& p_1 p_3 r_1 r_3 - p_2 p_3 r_2 r_3 + p_1 p_2 p_3 r_1 r_2 \nonumber \\
&+& p_1 p_2 p_3 r_1 r_3 + p_1 p_2 p_3 r_2 r_3 + p_1 p_2 r_1 r_2 r_3 \nonumber \\
&+& p_1 p_3 r_1 r_2 r_3 + p_2 p_3 r_1 r_2 r_3 - 2 p_1 p_2 p_3 r_1 r_2 r_3 = 0 \ .
\label{eq:martini}
\end{eqnarray}
Setting the probabilities equal in these equations gives the corresponding critical polynomial, whose root in $[0,1]$ determines the threshold. A question that arises is whether other lattices, not in the 3-uniform class, also have algebraic critical points. So far, the answer to this question is not known. However, it is possible to find approximations to a range of unsolved bond problems by effectively assuming that the critical threshold is the root of a polynomial of degree $n$, where $n$ is the number of bonds in the unit cell. This method has been used with success in \cite{Scullard08} and \cite{Scullard10}, and a corresponding approximation for the Potts model was used by Wu for lattices of the kagome type \cite{Wu79,Wu2010,Ding}. Here, I use this method to find the critical polynomials approximating the thresholds of the $(4,6,12)$ and $(3^4,6)$ lattices, and I show that this leads to a sequence of approximations that likely approach the exact answer for any periodic lattice.
I will begin by defining some terminology. All the lattices considered here are periodic and, as we have done already, we call the smallest graph that can be copied and translated to form the entire lattice, the unit cell. When we consider inhomogeneous percolation and assign different probabilities to different bonds, we are not confined to remain within a single unit cell; corresponding bonds on neighbouring cells may have different probabilities. In this case, the percolation process defined on the lattice does not have the same periodicity as the lattice itself. I will term the smallest periodic unit of the probabilities the ``base'' of the process. Examples for the $(4,8^2)$ lattice are shown in Figure \ref{fig:FE}. The lattice is in Figure \ref{fig:FE}(a), Figure \ref{fig:FE}(b) depicts a base consisting of a single unit cell, and the base in Figure \ref{fig:FE}(c) has $12$ different probabilities and covers two unit cells.
\begin{figure}
\begin{center}
\includegraphics{3uniform.eps}
\caption{A 3-uniform hyperlattice and unit cell. The shaded area can represent any network of sites or bonds.} \label{fig:3uniform}
\end{center}
\end{figure}
\begin{figure}
\begin{center}
\includegraphics{SqHexMartini.eps}
\caption{a) the square lattice; b) the honeycomb lattice; c) the martini lattice. Inhomogeneous probability assignments are shown below the lattices.} \label{fig:SHM}
\end{center}
\end{figure}
\begin{figure}
\begin{center}
\includegraphics{FE.eps}
\caption{a) the $(4,8^2)$ lattice; b) inhomogeneous assignment of probabilities on the unit cell; c) a base that consists of $12$ probabilities and covers two unit cells.} \label{fig:FE}
\end{center}
\end{figure}
\section{Linearity assumption}
The approximation used here was described fully in \cite{Scullard10}, but I will give a brief summary. Note that in equations (\ref{eq:square}) and (\ref{eq:honeycomb}), all probabilities appear only in first order. This is a natural consequence of (\ref{eq:zc}), and I refer to this property as ``linearity''. As mentioned above, this is a very special property, but it turns out that very good approximations to bond percolation thresholds can be found by assuming that it holds in general. It may not be very clear why such a strategy should produce good approximations. However, the study in \cite{Scullard10} shows that the predicted thresholds are generally within $10^{-5}$ of the best numerical estimates, and that continual refinement is possible.
Consider the $(4,8^2)$ lattice (Figure \ref{fig:FE}(a)), in which each bond on the unit cell has a different probability, and call its critical surface $\mathrm{FE}(p,r,s,t,u,v)$. If we remove the $p-$ bond by setting $p=0$, the result should be the honeycomb lattice, with the trivial difference that two of the bonds are doubled. Contracting the $p-$bond to zero length by setting $p=1$ gives the martini-A lattice (Figure \ref{fig:FEtree}(a)), which is another exactly known threshold, denoted $A(p_1,p_2,r_1,r_2,r_3)$ and given by
\begin{equation}
\mathrm{A}(p_1,p_2,r_1,r_2,r_3)=\mathrm{M}(p_1,p_2,1,r_1,r_2,r_3) . \label{eq:A}
\end{equation}
The only way to satisfy both of these conditions {\it and} keep the function $\mathrm{FE}(p,r,s,t,u,v)$ first-order in its probabilities is to set
\begin{equation}
\mathrm{FE}(p,r,s,t,u,v)=p \mathrm{A}(r,s,t,u,v)+(1-p) \mathrm{H}(s,tv,ru)=0 . \label{eq:FE}
\end{equation}
Expanding gives,
\begin{eqnarray}
\mathrm{FE}(p,r,s,t,u,v)&=&1 - (p r u + s t u + p s v + r t v) - (r s u v + p t u v) +\nonumber \\
p r s u v + p t r u v &+& p s t u v + r s t u v + p r s t u + p r s t v - 2 p r s t u v \nonumber .\label{eq:foureightsquared}
\end{eqnarray}
The prediction for the bond threshold is found by solving the polynomial equation $\mathrm{FE}(p,p,p,p,p,p)=0$,
\begin{equation}
1 - 4 p^3 - 2 p^4 + 6 p^5 - 2 p^6=0,
\end{equation}
with solution on $[0,1]$ $p_c=0.676835...$ . Comparing with the numerical estimate $p_c \approx 0.676802$ with a standard error of $6.3 \times 10^{-7}$, we find that although our solution is very close, it is ruled out numerically. The assumption that $\mathrm{FE}$ is first-order in its arguments is therefore not correct. However, in seemingly all cases, the linearity assumption produces good approximations to bond thresholds.
At this point, one may wonder if the approximation is actually well-defined or if the derived critical surface depends on the method used to find it. It is clear that equation (\ref{eq:FE}) is the only way to satisfy reduction to the honeycomb and A lattices in the appropriate limits of $p$. What is not obvious is whether this choice is consistent with the results one expects by setting bonds other than $p$ to $0$ or $1$. For example, the $v$-bond is not equivalent to the $p$-bond but we may just as well use $v$ to derive the threshold. If we set this bond to $0$, we disconnect the graph and end up with one-dimensional percolation with the decorated bond shown in Figure \ref{fig:FEtree}(b). The probability of crossing this bond is given by $p_D(p,r,s,t,u)=u [1-(1-pt)(1-rs)]$, and, since this is 1-D percolation, our critical surface should be $p_D=1$ or $1-p_D=0$. Setting $v=1$ gives the lattice shown in Figure \ref{fig:FEtree}(c). The critical surface of this lattice, which for the moment I will call $L_1$ (actually, it is the dual of Wierman's bow-tie lattice \cite{Wierman84}), is not known exactly for this configuration of probabilities (however, see \cite{Scullard10}), so we must employ the same procedure to find $L_1(p,r,s,t,u)$. Setting $s=0$ in this lattice gives us the exactly-known martini-B lattice \cite{Scullard06,Ziff06} of Figure \ref{fig:FEtree}(d) \cite{Scullard06,Ziff06}, with critical surface $B(p_1,r_1,r_2,r_3)=M(p_1,1,1,r_1,r_2,r_3)$, while $s=1$ is once again the honeycomb lattice with a doubled bond. Thus we have,
\begin{equation}
L_1(p,r,s,t,u)=s B(u,t,p,r)+(1-s)H(p,r,ut),
\end{equation}
and now
\begin{equation}
\mathrm{FE}(p,r,s,t,u,v)=v L_1(p,r,s,t,u)+(1-v) [1-p_D(p,r,s,t,u)]=0 .
\end{equation}
Expanding gives the same function as in (\ref{eq:FE}). The answer is therefore unchanged whether we choose to impose constraints on the behaviour of $p$ or $v$. We would like to know whether this will be true in general. Given a periodic lattice with a base of $n$ bonds, can we find a critical function, first order in all $p_i$, such that the results of setting each of the $p_i$ alternately to $0$ and $1$ are consistent with the answers required in those limits? To answer this question, we observe that the linearity assumption is equivalent to the requirement that the critical function satisfy the $n-$dimensional Laplace equation in the probabilities. Specification of the limiting functions when $p_i=0$ and $p_i=1$ for all $i$ is equivalent to the imposition of boundary conditions on the faces of an $n-$dimensional hypercube. Provided the boundary conditions are consistent with one another, existence of the function is assured by the standard existence theorem for the $n-$dimensional Laplace equation. This also proves uniqueness, although it was already clear from the derivation of (\ref{eq:FE}).
\begin{figure}
\begin{center}
\includegraphics{FEtree.eps}
\caption{Lattices and probability assignments used to derive the approximate critical surface (\ref{eq:FE}). a) the martini-A lattice; b) one-dimensional percolation with a decorated bond; c) the bow-tie dual; d) the martini-B lattice.} \label{fig:FEtree}
\end{center}
\end{figure}
Thus, we can talk about {\it the} first-order function that serves as the approximate critical surface for a given base, and the method we use to find it is irrelevant. Critical functions were found in \cite{Scullard10} for all but two of the Archimedean lattices by starting with the most general function first-order in its arguments, and imposing symmetry and special cases until all unknown constants were fixed. However, the greater the number of bonds in the base, the more cumbersome this procedure becomes, and it was not possible to find the thresholds of the $(4,6,12)$ and $(3^4,6)$ lattices in this way, which have $18$ and $15$ bonds in their unit cells. In this paper, I will find these two critical surfaces using the technique outlined above for the $(4,8^2)$ lattice; choose a bond, call $L_0$ the lattice resulting from setting $p_i=0$ for some $i$, and $L_1$ the lattice resulting from setting $p_i=1$. If the threshold of either of these lattices, say $L_0$, is unknown then repeat the process on that lattice, finding $L_{00}$ and $L_{01}$. By continuing in this way, the number of bonds will eventually be reduced sufficiently that lattices with exactly known thresholds appear. Although the result can be a fairly large branching process of lattices, the procedure is straightforward and, aside from some trivial computer algebra, requires minimal computational resources.
\section{Archimedean lattices}
\subsection{$(4,6,12)$ lattice}
The $(4,6,12)$ lattice is shown in Figure \ref{fig:4612} and we denote its threshold by
\begin{equation}
\mathrm{FST}(p_1,p_2,p_3,p_4,p_5,p_6,r_1,r_2,r_3,r_4,r_5,r_6,s_1,s_2,t_1,t_2,u_1,u_2) .
\end{equation}
The unit cell contains $18$ bonds, so we have no hope of encountering a known lattice until we have performed many iterations. We begin the journey by setting $p_1=1$, resulting in the graph $L_1$ (Figure \ref{fig:FST1}(a)) with a bond doubled in series, and setting $p_1=0$, giving $L_0$ (Figure \ref{fig:FST0}(a)). Their critical surfaces are given by the functions
\begin{equation}
L_0(p_3,p_4,p_5,r_1,r_2,r_3,r_4,r_5,r_6,s_1,s_2,t_2,u_1,x,y)
\end{equation}
and
\begin{equation}
L_1(p_2,p_3,p_4,p_5,p_6,r_1,r_2,r_3,r_4,r_5,r_6,s_1,s_2,t_1,t_2,u_1,u_2)
\end{equation}
and thus we have
\begin{eqnarray}
\mathrm{FST}&=&(1-p_1) L_0(p_3,p_4,p_5,r_1,r_2,r_3,r_4,r_5,r_6,s_1,s_2,t_2,u_1,t_1 p_2,u_2 p_6) \nonumber \\
&+&p_1 L_1(p_2,p_3,p_4,p_5,p_6,r_1,r_2,r_3,r_4,r_5,r_6,s_1,s_2,t_1,t_2,u_1,u_2) .
\end{eqnarray}
Since neither of these functions are known, we start the procedure over, beginning with $L_0$. The rest of the gory details are described in the Appendix and all the lattices arising in the branching process are shown in Figures \ref{fig:FST0} and \ref{fig:FST1} with probability assignments in Figures \ref{fig:FST0units} and \ref{fig:FST1units}. The final critical function $\mathrm{TFS}$ has $1932$ terms, and is included in a text file in the supplementary material to this submission. The polynomial resulting from setting all probabilities equal is
\begin{eqnarray}
1 &-& 18 p^6 - 6 p^8 + 30 p^9 + 3 p^{10} + 108 p^{11} - 81 p^{12} - 174 p^{13} \nonumber \\
&-& 246 p^{14} + 1090 p^{15} - 1110 p^{16} + 480 p^{17} - 78 p^{18}=0
\end{eqnarray}
with root in $[0,1]$, $p_c=0.69377849...$ . According to the numerical work of Parviainen \cite{Parviainen}, $p_c=0.69373383...$, putting our prediction within $4.5 \times 10^{-5}$ of the numerical value, but outside his standard error of $7.2 \times 10^{-7}$. This level of accuracy is typical for the method \cite{Scullard10}.
Before arriving at this answer we encountered many intermediate lattices (there are 21 in Figures \ref{fig:FST0} and \ref{fig:FST1}), and the procedure may rightly be described as tedious. However, the upside is that we obtain predictions for all these lattices as a by-product, some of which are even exact. Table \ref{table:FST} shows the predictions along with an indication of which are exact.
\begin{table}
\begin{center}
\caption{Predicted thresholds for the lattices in Figures \ref{fig:FST0} and \ref{fig:FST1}.(*) denotes exact threshold.}
\begin{tabular}{c c c|c c c}
\hline
\hline
&Lattice & Bond prediction & & Lattice & Bond prediction\\
\hline
(a) &$L_0$ & $0.736212...$ & (a) &$L_1$ & $0.669513...$\\
(b) &$L_{00}$ & $0.725567...$ & (b) &$L_{10}$ & $0.717320...$\\
(c) &$L_{01}$ & $0.716269...^*$ & (c) &$L_{11}$ & $0.639238...$\\
(d) &$L_{000}$ & $0.704323...$ & (d) &$L_{100}$ & $0.699211...$\\
(e) &$L_{011}$ & $0.693925...^*$ & (e) &$L_{110}$ & $0.659993...$\\
(f) &$L_{0110}$ & $0.695253...^*$ & (f) &$L_{111}$ & $0.612973...$\\
(g) &$L_{0111}$ & $0.666099...^*$ & (g) &$L_{1000}$ & $0.704323...$\\
(h) &rocket & $0.669182...^*$ & (h) &$L_{1001}$ & $0.669182...$\\
(i) &$L_{01111}$ & $0.628312...^*$ & (i) &$L_{1110}$ & $0.610552...$\\
\ &\ & & (j) &$L_{1111}$ & $0.591166...$\\
\ &\ & & (k) &$L_{111110}$ & $0.544637...$\\
\ &\ & & (l) &$(3,4,6,4)$ & $0.524821...$\\
\hline
\hline
\end{tabular}
\label{table:FST}
\end{center}
\end{table}
\subsection{$(3^4,6)$ lattice}
The procedure for the $(3^4,6)$ lattice (Figure \ref{fig:346}) is described in the Appendix. The full critical surface has $12795$ terms and is included in the supplementary material. The polynomial is
\begin{eqnarray}
1 &-& 12 p^3 - 36 p^4 + 21 p^5 + 327 p^6 - 69 p^7 - 2532 p^8 + 6533 p^9 - 8256 p^{10} \nonumber \\
&+& 6255 p^{11} - 2951 p^{12} + 837 p^{13} - 126 p^{14} + 7 p^{15} =0 .
\end{eqnarray}
This predicts $p_c=0.43437077...$, whereas Parviainen gives $p_c=0.43430621...$, so again we are fairly close, differing by $6.5 \times 10^{-5}$ but still outside Parviainen's standard error of $5.0 \times 10^{-7}$. The predicted thresholds for the intermediate lattices are shown in Table \ref{table:TFS}.
\begin{table}
\begin{center}
\caption{Predicted thresholds for the lattices in Figures \ref{fig:TFS0} and \ref{fig:TFS1}.(*) denotes exact threshold. Graphs with threshold $1/2$ are self-dual.}
\begin{tabular}{c c c|c c c}
\hline
\hline
& Lattice & Bond prediction & & Lattice & Bond prediction\\
\hline
(a) & $L_0$ & $0.462592...$ & (a) & $L_1$ & $0.441699...$ \\
(b) & $L_{01}$ & $0.493113...$ & (b) & $L_{10}$ & $0.485729...$ \\
(c) & $L_{00}$ & $0.476682...^*$ & (c) & $L_{11}$ & $0.408991...$ \\
(d) & $L_{001}$ & $1/2^*$ & (d) & $L_{100}$ & $0.560890...$ \\
(e) & $L_{010}$ & $0.529519...$ & (e) & $L_{110}$ & $0.469809...$ \\
(f) & $L_{011}$ & $0.439655...$ & (f) & $L_{111}$ & $0.330818...$ \\
(g) & $L_{0100}$ & $0.634235...^*$ & (g) & $L_{1110}$ & $0.380244...^*$\\
(h) & $L_{0101}$ & $0.439497...^*$ & (h) & $L_{1001}$ & $0.532058...^*$\\
(i) & $L_{0110}$ & $0.532058...^*$ & (i) & $L_{1100}$ & $0.544620...$ \\
(j) & $L_{0111}$ & $0.330818...$ & (j) & $L_{1101}$ & $0.415824...^*$\\
(k) & $L_{0010}$ & $0.483133...$ & (k) & $L_{11001}$ & $1/2^*$ \\
(l) & $L_{0011}$ & $0.516867...$ & (l) & $L_{11000}$ & $0.628312...$ \\
(m) & $L_{00101}$ & $1/2^*$ & & \ & \ \\
(n) & dec. sq. & $1/2^*$ & & & \\
(o) & $(4,8^2)$ dual & $0.323165...$ & & & \\
(p) & dice & $0.475572...$ & & & \\
\hline
\hline
\end{tabular}
\label{table:TFS}
\end{center}
\end{table}
\section{Polynomial sequences}
Every critical bond threshold derived in this way is the root in $[0,1]$ of a polynomial of degree at most $n$, where $n$ is the number of bonds in the base. However, as already mentioned, the base need not consist of only a single unit cell. Considering the base of the $(4,8^2)$ lattice to be two unit cells results in a $12^{\mathrm{th}}$ order polynomial, which I will call the second polynomial, that makes a better prediction (see below) than that found by the first polynomial from the $6-$bond case. It is natural, then, to assume that by considering larger and larger bases, one can get arbitrarily close to the exact threshold. For those lattices to which (\ref{eq:zc}) applies, the first polynomial gives the exact answer. For those situations, all subsequent polynomials must also give the same result even if the full critical surface is not correct.
One possible mathematical question is whether there is always only one root in $[0,1]$ of the polynomial when following the above procedure. I do not know of any argument that establishes this in general, but for all lattices I have encountered, it has always been true.
\subsection{Second polynomials}
The polynomials calculated above represent only the first step in a series of approximations. By extending the base to include a second unit cell, we find a higher-order polynomial for the critical threshold. This generally, but not always, gives a refinement of the estimate and I explore this idea for a few lattices.
\subsection{$(4,8^2)$ lattice}
We already encountered the first polynomial for this lattice, but we can get a better estimate using the base shown in Figure \ref{fig:FE}(c). The full threshold can be found in the supplementary material, but the critical polynomial is
\begin{equation}
1 - 4 p^4 - 16 p^6 + 12 p^7 + 22 p^8 + 16 p^9 - 70 p^{10} + 48 p^{11} - 10 p^{12}=0
\end{equation}
with $p_c=.676787...$ . The difference between this and the numerical value is $1.5 \times 10^{-5}$ as opposed to $3.3 \times 10^{-5}$ for the $6-$bond estimate, so we have cut our error in half. Furthermore, setting corresponding probabilities on each cell equal, $p_1=p_2=p$, $r_1=r_2=r$, etc., we get a prediction for the $6-$bond case that does not reduce to what we found before (equation (\ref{eq:FE})). Note also that the second polynomial is not just the first with a few extra terms, but is completely different.
\subsection{Kagome lattice}
The kagome lattice is shown in Figure \ref{fig:kagome}(a). The threshold using the unit cell as base (Figure \ref{fig:kagome}(b)) was derived in \cite{ScullardZiff06} and \cite{Scullard10}, where it was found that $p_c=0.524430...$ compared with the recent numerical estimate $0.524405...$ \cite{Feng,Ding}. In considering the $12-$bond base (Figure \ref{fig:kagome}(c)), I found the same prediction as in the $6-$bond case, i.e. the first polynomial factors out of the second. Similarly, setting the corresponding probabilities in the neighbouring cells equal, $p_1=p_2=p, r_1=r_2=r$, etc., the $6-$bond critical surface factors out of the larger expression. Presumably, one needs to consider an even larger base to get any refinement.
\begin{figure}
\begin{center}
\includegraphics{kagome.eps}
\caption{a) the kagome lattice; b) the unit cell; c) the base using two unit cells.} \label{fig:kagome}
\end{center}
\end{figure}
\subsection{$(3^3,4^2)$ lattice}
Both the first and second polynomials for this lattice were derived in \cite{Scullard10}, but I will repeat them here because they will be useful later. The $(3^3,4^2)$ lattice is shown in Figure \ref{fig:3cubed4squared}(a). The first polynomial, derived using the base in Figure \ref{fig:3cubed4squared}(b), is
\begin{equation}
1-2p-2p^2+3p^3-p^4=0
\end{equation}
with critical threshold $p_c=0.419308...$ whereas Parviainen gives $p_c=.419642...$ . We refine the estimate by considering the base consisting of two unit cells, shown in Figure \ref{fig:3cubed4squared}(c). In this case, the procedure gives the $10^{\mathrm{th}}$-order polynomial,
\begin{equation}
1-4 p^2-12 p^3+104 p^5-193 p^6+146 p^7-45 p^8+2 p^{10}=0
\end{equation}
and $p_c=0.419615...$, which is closer to the numerical value. Once again, these two polynomials do not appear to have much in common other than a similar root in $[0,1]$. However, that is not to say that no relationship exists, and in fact a recursive method of generating successive polynomials might even be considered an exact solution of the problem.
\begin{figure}
\begin{center}
\includegraphics{3cubed4squared.eps}
\caption{a) the $(3^3,4^2)$ lattice; b) unit cell; c) a base employing two unit cells} \label{fig:3cubed4squared}
\end{center}
\end{figure}
\section{Limiting case}
It is natural, after calculating a few first and second polynomials, to wonder whether the process must really be carried out indefinitely, or if there are some lattices for which an $m^{\mathrm{th}}$ polynomial actually provides the exact answer. Indeed, in the case of the $3-$uniform lattices, the first polynomial is exact and all subsequent polynomials make the same prediction. Could it be that in some cases we just need to consider a large enough base and we will find that the exact critical surface is first-order in the probabilities? I will argue that for many of the lattices considered here, the answer is no; if the first critical surface is not exact, then no finite-sized base will give the exact answer.
Consider the $(3^3,4^2)$ base shown in Figure \ref{fig:3c4scontract}(a), where we have extended it to include $m$ unit cells, and thus we have $5m$ probabilities, denoted by $\{p\}$. Assume also that we have found the threshold by the linearity assumption, $\mathrm{TF^{(m)}}(\{p\})$, and we want to know if it might be exact. To answer this, we may set $p_i=r_i=t_i=0$ for $i>1$ to give the lattice shown in Figure \ref{fig:3c4scontract}(b), which consists of unit cells of the $(3^3,4^2)$ lattice with some long straight paths (shown in red) connecting them. Setting the probabilities of all the bonds in those paths to $1$, we recover the case in which the base is the unit cell (Figure \ref{fig:3cubed4squared}(b)), i.e. we are left with $5$ probabilities and a critical surface that is first order in these. The prediction for this function must agree with the one we found previously, as the linear threshold is unique, but that prediction has been ruled out numerically. Moreover, the critical surface $\mathrm{TF^{(m)}}(\{p\})$ is not internally consistent as it makes two different predictions for the same lattice. Any base consisting of a finite number of units cells will suffer from this defect. We therefore come to the conclusion that although increasing the size of the base of the process on the $(3^3,4^2)$ lattice makes the linearity approximation more accurate, it will never be exact for any finite-sized base.
For some lattices, one may be able to construct a large but finite base that does not suffer from this inconsistency. However, for many problems, such as the kagome and $(4,8^2)$ lattices, this does not seem possible, suggesting that, for these problems, the threshold is not the root of a finite polynomial. However, a few issues would have to be resolved before we could reach that conclusion. Just because the linearity hypothesis fails for any finite base, we cannot conclude that the exact critical surface for a base using only a unit cell does not contain only a finite number of terms. Consider the kagome critical surface for the unit cell (Figure \ref{fig:kagome}(b)), $K(p,r,s,t,u,v)$. Is it possible that there is a maximum order, $m$, of the terms in the exact threshold, so that the highest order term possible is $p^m r^m s^m t^m u^m v^m$ ? It is certainly true that if we had a base of $m$ unit cells, and found the linear threshold, it would predict that $K(p,r,s,t,u,v)$ has maximum order $m$. However, it is not clear if the converse is true. If $K(p,r,s,t,u,v)$ has maximum order $m$, then is it necessarily derived from a critical surface that was first-order on some base of $m$ cells? If we could show that this is the case, then we would know that the exact critical function $K(p,r,s,t,u,v)$ is necessarily an infinite power series in its arguments, unlike any presently-known exact solution. However, it is apparently not trivial to show this, if in fact it is true. Complicating matters is that there are many possibilities for a base of size $m$, as $m$ becomes large, and it is not clear that they will all predict the same polynomial, or reduce to the same function, $K(p,r,s,t,u,v)$. Sorting out these issues would provide crucial insight into the question of whether or not thresholds of these problems are algebraic numbers.
In this paper, I have found the approximate thresholds of the $(4,6,12)$ and $(3^4,6)$ lattices, completing the program begun in \cite{Scullard10}. I also showed how this approximation leads to polynomial sequences for unsolved problems, and calculated the second polynomials for the $(3^3,4^2)$ and $(4,8^2)$ lattices. Finally, I showed how these results give some clues about whether or not generic percolation thresholds are algebraic numbers.
\begin{figure}
\begin{center}
\includegraphics{3_3-4_2contract.eps}
\caption{a) a base for the $(3^3,4^2)$ lattice extending over $m$ unit cells; b) the lattice resulting from setting $p_i=r_i=t_i=0$ for $i>1$. The long red paths between the triangles can be contracted by setting $s_i=t_i=1$ for $i>1$ to recover the process whose base is a single unit cell.} \label{fig:3c4scontract}
\end{center}
\end{figure}
\section{Acknowledgements}
This work was performed under the auspices of the U.S. Department of Energy by Lawrence Livermore National Laboratory under Contract DE-AC52-07NA27344.
\section*{References}
|
\section{Introduction}
In the framework of the perturbative QCD, in the Regge kinematics,
particle interaction can be described by the exchange of reggeized gluons
which emit and absorb real gluons and also may split into several
reggeized gluons. Emission of real gluons from reggeized gluons is
described by effective vertices introduced in ~\cite{bfkl}
and ~\cite{bartels} for non-split and split reggeons.
Originally both type of vertices were calculated directly
from the relevant simple Feynman diagrams in the Regge kinematics.
Later a powerful effective action formalism was proposed in
~\cite{lipatov}, which considers reggeized and normal gluons as independent
entities from the start and thus allows to calculate all QCD diagrams
in the Regge kinematics automatically and in a systematic and self-consistent
way. However the resulting expressions are 4-dimensional and need
reduction to the final 2-dimensional transverse form. This reduction is
trivial for tree diagrams but becomes less trivial for diagrams
with loops.
In the paper of two co-authors of the present paper (M.A.B. and M.I.V.)
~\cite{bravyaz} it was demonstrated that the diffractive
amplitude for the production of a real gluon calculated by means of the
effective action and based on the transition of a Reggeon into one or two
Reggeons and Particle (R$\to$R(R)P transition), after integration over
longitudinal variables, coincides with the results calculated in terms
of the above-mentioned effective vertices
(in the purely tranverse "BFKL-Bartels formalism").
However in the process of reduction to the transverse form a certain
prescription had to be used to give sense to divergent integrals.
In this paper we study a more general case: the gluon production
off two different targets. In this case, in the lowest order, with which we
restrict ourselves, the production amplitude by itself
is a tree diagram, without any internal integrations. Longitudinal
integrals appear only in the inclusive cross-section.
Our purpose is twofold. First we analyze application of the
effective action formalism to the processes where the number of
reggeons can be changed (from one to two in our case). As we shall show
this application requires a careful study of kinematical regions in
which particular diagrams generated by the effective action or
even parts of these diagrams are to be taken into account.
A separate problem
is understanding in which sence singularities which appear in
these diagrams when the "-" components of the momenta transferred
to the two targets ("energies"), $q_{1,2-}$, vanish are to be understood.
Comparison with the standard perturbation approach allows
to solve this problem. Note that this latter point was studied in
~\cite{HBL,hentschinski} for simpler diagrams, without reggeon proliferation.
Our second aim is to obtain convenient expressions for the production
amplitudes which can be used for the calculation of the inclusive
cross-sections. The essential region for the integration over transferred
energies involved in this calculation is $q_{1,2-}>>p_-$, where $p_-$
is the energy of the emitted gluon. The amplitude can be drastically simplfied
in this region and transformed to the expression convenient for the following
integrations.
Our results show that in the general kinematics the production amplitude is
almost completely given by the contribution from the R$\to$RRP effective
vertex
derived in ~\cite{bravyaz}. The pole at $q_{1,2-}=0$ should be taken in the
principal values prescription, in agreement with the assumption made in
~\cite{bravyaz}. However the contribution from the R$\to$RRP vertex
should be supplemented by terms proportional to $\delta(q_{1,2-})$
coming from the double reggeon exchange.
\begin{figure}[h]
\leavevmode \centering{\epsfysize=0.2\textheight\epsfbox{fig1.eps}}
\caption{Production amplitude off two scattering centers}
\label{fig1}
\end{figure}
\section{Kinematics and the target lines}
We study the production amplitude on two centers shown in
Fig. \ref{fig1} in the Regge kinematics. For simplicity we consider
the quark as the projectile and scalar quarks as the two scattering centers.
We consider the central region so that both
$(kp)$ and $(lp)$ are large. This means that $k_+p_-$ and
$l_-p_+$ are much greater than the typical transverse momentum squared
$p_\perp^2$. If $p_+=xk_+$ then our region of $x$ is
\begin{equation}
\frac{|p_\perp^2|}{s}<<x<<1.
\end{equation}
Our longitudinal conservation laws are
\begin{equation}
p_++k'_+=k_+,\ \ p_-+q_{1-}+q_{2_-}=0,
\label{conslaw}
\end{equation}
where we neglected $q_{1+}$, $q_{2+}$ and $k'_-$.
We consider both targets at the same initial momentum $l_1=l_2=l$
in the center of mass system with $l_+=l_\perp=0$.
The transferred momenta in the two collisions are $q_1$ and $q_2$.
The mass-shell conditions give
\begin{equation}
2q_1l+q_1^2=2q_2l+q_2^2=0.
\end{equation}
The initial projectile momentum $k$ has $k_-=k_\perp=0$.
Our Regge kinematics requires that $q_{1-},q_{2-}<<l_-$.
So we put $q_{1-}=\xi_1l_-$ and $q_{2-}=\xi_2l_-$ with
$\xi_{1,2}<<1$. The parameters $\xi_1$ and $\xi_2$ are actually the
conventional Sudakov variables for the transverred momenta $q_1$ and $q_2$.
From
\[q_{1+}=l_{1+}=-\frac{{l'_{1\perp}}^2}{2l'_{1-}}\]
it follows that
\[ q_1^2=2q_{1+}q_{1-}+q_{1\perp}^2=
(1-\xi_1)q_{1\perp}^2\simeq q_{1\perp}^2,\]
so that $q_1$ is almost a purely transverse momentum. as well as $q_2$.
In practical applications we use the axial gauge
in which the gluon field $V$ satisfies $(Vl)=0$
and its propagator is
\begin{equation}
P_{\mu\nu}(q)=\frac{h_{\mu\nu}(q)}{q^2},
\ \ h_{\mu\nu}(q)=g_{\mu\nu}-\frac{l_\mu q_\nu+q_\mu l_\nu}{(ql)}.
\end{equation}
Coupling to the scalar targets generates vectors
\begin{equation}
P_\mu(q)=P_{\mu\nu}(2l^\nu+q^\nu)=P_{\mu\nu}q^\nu=
-l_\mu\frac{q^2}{(ql)}\,\frac{1}{q^2}=2l_\mu\frac{1}{q_\perp^2},
\end{equation}
where $q=q_1$ or $q=q_2$.
So in fact we can study the amplitude with amplutated target lines
and external legs $P_{\mu_1}(q_1)$ and $P_{\mu_2}(q_2)$.
\section {Gluon emission in the effective action formalism}
In the effective action formalism all real and virtual particles
in the direct channels split into groups in correspondence
with their rapidities $y=\frac{1}{2}\ln|p_+/p_-|$.
Gluons with rapidities within some interval $[y-\eta/2, y+\eta/2]$
are described by the usual gluon field $V^{y}_{\mu}=-it^{a}V^{ya}_{\mu}$.
The reggeon field $A^{y}_{{\mu}}=-it^{a}A^{ya}_{\mu}$ with only
non-zero longitudinal components $A_{+}$ and $A_{-}$ corresponds to
virtual gluons in the crossing channels responsible
for the interaction between the groups with essentially different rapidities.
The effective Lagrangian describes the self-interaction of
gluons inside of each group by means of the usual QCD Lagrangian
${\cal L}_{QCD}$ and their interaction with reggeons.
It takes the form \cite{lipatov}:
\begin{equation}
{\cal L}_{eff}={\cal L}_{QCD}(V^{y}_\mu+A^{y}_\mu)
+ 2 {\rm Tr}\,\Big(({\cal A}_+(V^{y}_+ +A^{y}_+)-A^{y}_+)\partial^2_{\perp} A^{y}_-
+ ({\cal A}_-(V^{y}_- +A^{y}_-)-A^{y}_-)\partial^2_{\perp} A^{y}_+\Big),
\label{e1}
\end{equation}
where
$$
{\cal A}_{\pm}(V_{\pm})=-\frac{1}{g}\partial_{\pm}\frac{1}{D_{\pm}}\partial_{\pm}*1=
\sum_{n=0}^{\infty}(-g)^nV_{\pm}(\partial_\pm^{-1}V_\pm)^n
$$
\begin{equation}
=V_{\pm}-gV_{\pm}\partial_\pm^{-1}V_\pm+g^2V_{\pm}\partial_\pm^{-1}V_\pm
\partial_\pm^{-1}V_\pm+ -...
\label{e2}
\end{equation}
It is local in rapidity, so the rapidity index $y$ can be omitted.
The shift $V_\mu\to V_\mu+A_\mu$ with $A_\perp=0$ is done
to exclude direct gluon-reggeon transitions.
The reggeon propagator in momentum representation
\begin{equation}
<A_+^{y'a}A_-^{yb}>=-i\frac{\delta_{ab}}{q_\perp^2}
\,\theta(y'-y-\eta)
\label{e3}
\end{equation}
is to be contracted
with field $A_-$ interacting with a group of a higher rapidity $y'$
and field $A_+$ interacting with a group of a smaller rapidity $y$.
From the kinematical constraints it easily follows
that the momentum $q_-$ of field $A_+$ is small compared to ``-''
components of momenta flowing in the group with a higher rapidity,
so the kinematical condition is implied
\begin{equation}
\partial_- A_+ =0.
\label{e4}
\end{equation}
Analogously, the kinematical condition
\begin{equation}
\partial_+ A_- =0
\label{e5}
\end{equation}
reflects the comparative smallness of the momentum $q_+$
of the field $A_-$.
Formally, in the framework of the effective action approach
one can introduce quite a number of diagrams which describe
interaction with two centers. But one has to take into account
that the rules assumed in the derivation of effective action
put certain restrictions on the kinematical regions appropriate
for particular contributions, so that many of the diagrams which
can formally be introduced have in fact to be dropped once
a particular kinematics is condidered. For some other diagrams
these restrictions may lead to neglecting some terms.
In short these conditions require that real particles should be emitted
in the multiregge kinematics described above. The emitted particle
should carry nearly all the "+" component of the momentum of the
higher rapidity reggeons and nearly all the "$-$" component of the
lower rapidity reggeons from which it is emitted.
\begin{figure}[h]
\leavevmode \centering{\epsfysize=0.15\textheight\epsfbox{fig2.eps}}
\caption{Production amplitude in the effective action formalism}
\label{fig2}
\end{figure}
The essential contribution to the production amplitude
in the effective action formalism is
given by the three diagrams shown in Fig. \ref{fig2}.
Also for the two last diagrams the analogous diagrams with
the interchanged targets ($1\leftrightarrow 2$) have to be added.
The diagram 1 in this figure contains the R$\to$RRP vertex $V$ derived
in ~\cite{bravyaz} in a general kinematics $q_{1,2-} \sim p_-$,
$q_{1,2+} << p_+$:
\begin{equation}
V=V_1+V_2,
\end{equation}
where
\begin{eqnarray}
&& V_1=\frac{i}{2}\,\frac{f^{db_1c}f^{cb_2a}}{(q-q_1)^2+i0}\times
\nonumber \\
&& \Big\{ q_+(4q_1+p)_\mu
-\left((q+q_1)(p-q_2)+q_2^2-q_1^2+(q-q_1)^2
+q_+ q_{1-}\right)n^+_{\mu}
\nonumber \\
&& +\frac{q^2 (q-q_1)^2}{p_- q_{1-}}n^-_{\mu}
+\Big(2q_+ -\frac{q^2}{q_{1-}} \Big)
\Big[-2q_+n^-_{\mu}+(p+2q_2)_\mu+
(p_- -q_{2-}+\frac{q_2^2}{q_+})n^+_{\mu}\Big]\Big\}
\cdot{e^{\mu}}
\label{12a}
\end{eqnarray}
and $V_2=V_1(1\leftrightarrow 2)$. Here $q=p+q_1+q_2$ with $q^2=q_{\perp}^2$
is the momentum transferred from the projectile,
$e^{\mu}$ is the gluon polarization vector and
$n_{\mu}^{\pm}=(1,0,0,\mp 1)/\sqrt{2}$ are the light-cone unit
\footnote{
Note that the normalization
$a_{\mu}b^{\mu}=a_{+}b_{-}+a_{-}b_{+}+a_{\perp}b_{\perp}$
for the longitudinal components is used here,
whereas it was taken as
$a_{\mu}b^{\mu}=\frac{1}{2}a_{+}b_{-}+\frac{1}{2}a_{-}b_{+}
+a_{\perp}b_{\perp}$ in \cite{bravyaz,braun2}.
}
vectors.
This vertex is found to be transversal with respect to
the gluon momentum $p$.
In the chosen gauge $(Vl)=0$ which is equivalent to $V_+ =0$,
the vertex crucially simplifies:
\begin{equation}
V_1=i\frac{f^{db_1c}f^{cb_2a}}{(q-q_1)^2+i0}
\Big\{2q_+(eq)_{\perp} -\frac{q_\perp^2}{q_{1-}}
\Big[(e,q-q_1)_{\perp}
-\frac{(q-q_1)^2}{p_\perp^2}(ep)_{\perp}\Big]\Big\}\ .
\label{12}
\end{equation}
It is possible to present the contribution (\ref{12}) in
terms of effective R$\to$RP vertices.
In the second term in the brackets
we present the product $1/q_{1-}[(q-q_1)^2+i0]$ as
\[
\frac{1}{q_{1-}[(q-q_1)^2+i0]}
=\frac{2q_+}{(q-q_1)_\perp^2[(q-q_1)^2+i0]}
+\frac{1}{q_{1-}(q-q_1)_\perp^2}
\]
to obtain
\begin{equation}
V_1=W_1+R_1 ,
\end{equation}
where
\begin{equation}
W_{1}=-i\frac{2q_+q_{\perp}^2}{(q-q_1)^2+i0}f^{db_1c}f^{cb_2a}
B(p,q_2,q_1),
\end{equation}
\begin{equation}
R_{1}=i\frac{q_{\perp}^2}{q_{1-}}f^{db_1c}f^{cb_2a}
L(p,q_2)
\end{equation}
and vertices $L$ and $B$ are
\begin{equation}
L(p,q_1)=\frac{(p\epsilon_\perp)}{p_\perp^2}-
\frac{(p+q_1,\epsilon_\perp)}{(p+q_1)_\perp^2},\ \
B(p,q_1,q_2)=L(p+q_1,q_2).
\label{lev}
\end{equation}
The contribution from $R_1$ contains the
singularity at $q_{1}=0$. In ~\cite{bravyaz} this singularity
was understood in the principal value prescription.
This rule will be proven when we compare the contribution from
the vertex (\ref{12}) to the emission amplitude with the expression
found by the standard perturbation technique.
Coupling to the projectile and taking into account
the upper reggeon propagator
we obtain the part of the amplitude generated by the effective R$\to$RRP
vertex
\begin{equation}
{\cal A}^{ef}_1= 32(kl)^2\frac{1}{2k_+q_\perp^2}(V_1+V_2) t^d\, .
\end{equation}
For the following comparison it is convenient to separate
contributions with different
polarization factors
\[
{\cal A}^{ef}_1=i 32(kl)^2t^d
\Big\{\frac{p_+}{k_+}\frac{(eq)_\perp}{q_\perp^2}
\Big(\frac{f^{db_1c}f^{cb_2a}}{(q-q_1)^2+i0}+
\frac{f^{db_2c}f^{cb_1a}}{(q-q_2)^2+i0}\Big)\]
\[+
\frac{(ep)_\perp}{p_\perp^2}\Big(\frac{f^{db_1c}f^{cb_2a}}{2k_+q_{1-}}+
\frac{f^{db_2c}f^{cb_1a}}{2k_+q_{2-}}\Big)\]
\begin{equation}
-(e,q-q_1)_\perp \frac{f^{db_1c}f^{cb_2a}}{2k_+q_{1-}}\frac{1}{(q-q_1)^2+i0}-
(e,q-q_2)_\perp \frac{f^{db_2c}f^{cb_1a}}{2k_+q_{2-}}\frac{1}{(q-q_2)^2+i0}\Big\}
\ .
\label{aeff}
\end{equation}
For the calculation of the diagrams 2,3 in Fig. \ref{fig2}
one must take into account that with the kinematical conditions
$q_{1,2+}<<p_+ <<k_+$ and $k_+ q_{1,2-} >> q_{1,2\perp}^2,p_{\perp}^2$
the following approximation for the denominators
of the quark propagators can be done
\[
(k-q_1)^2+i0=-2k_+ q_{1-} + q_{1\perp}^2 +i0
\approx -2k_+ q_{1-} +i0\ ,
\]
\begin{equation}
(k-p-q_1)^2+i0=2k_+ (-p_{-}-q_{1-}) + (p+q_1)_{\perp}^2 +i0
\approx 2k_+ q_{2-} +i0\ .
\label{prop2}
\end{equation}
The total contribution from these two diagrams
together with terms $(1\leftrightarrow 2)$ is given by
\begin{equation}
{\cal A}^{ef}_2+{\cal A}^{ef}_3+(1\leftrightarrow 2)=
\frac{32(kl)^2}{2k_+q_{2-}+i0}L(p,q_1)f^{b_1ac}t^{b_2}t^c+
\frac{32(kl)^2}{-2k_+q_{1-}+i0}L(p,q_2)f^{b_2ac}t^ct^{b_1}+
\Big(1\leftrightarrow 2\Big).
\label{ampp}
\end{equation}
Using
$$
\frac{1}{\pm 2k_+q_{1,2-} +i0}
=\pm\frac{1}{2k_+}\cdot P\frac{1}{q_{1,2-}}
-i\pi\delta(2k_+ q_{1,2-})
$$
we separate parts with or without $\delta$-function, which we denote
by upper indeces 0 and 1 respectively.
We find (suppressing factor $32(kl)^2$)
\begin{equation}
\Big({\cal A}^{ef}_2+{\cal A}^{ef}_3+(1\leftrightarrow 2)\Big)^{(0)}=
-i\pi\delta(2k_+q_{2-})L(p,q_1)f^{b_1ac}\{t^{b_2},t^c\}
-i\pi\delta(2k_+q_{1-})L(p,q_2)f^{b_2ac}\{ t^{b_1},t^c\}
\label{ampp0}
\end{equation}
and
\begin{equation}
\Big({\cal A}^{ef}_2+{\cal A}^{ef}_3+(1\leftrightarrow 2)\Big)^{(1)}=
\frac{L(p,q_1)}{2k_+}f^{b_1ac} [t^{b_2},t^c] P\frac{1}{q_{2-}}+
\frac{L(p,q_2)}{2k_+}f^{b_2ac}[t^{b_1},t^c] P\frac{1}{q_{1-}}\ .
\label{ampp1}
\end{equation}
However one should take into account the mentioned restriction
to the multiregge character of particle emission.
Then one immediately finds that diagrams 2 and 3 can only contribute in the
region of $q_{1-}$ or $q_{2-}$ close to zero, since otherwise the "$-$' component
of the lower reggeon from which the gluon is emitted will not be totally
transferred to this gluon. This implies that only terms with
$\delta(q_{1.2-})$ should be retained in the contributions from the
diagrams 2 and 3.
This circumstance can be explained in more detail by
the above-mentioned condition of locality in rapidity.
For example, in the diagram 2 the difference between
the rapidity of the projectile
$\frac{1}{2}\ln\frac{s}{M^2}$,
where the small quark mass $M$ is introduced for finiteness,
and the rapidity of the virtual quark with the momentum
$k-q_1$:
\begin{equation}
\frac{1}{2}\ln\left|
\frac{k_+ -q_{1+}}{k_- -q_{1-}}
\right|
\label{e6}
\end{equation}
cannot be more than the cut-off parameter $\eta$.
This condition is equivalent to
\begin{equation}
\frac{M^2}{{s}} e^{-\eta} <
\frac{|k_{-}-q_{1-}|}{|k_{+}-q_{1+}|}
< \frac{M^2}{{s}} e^{+\eta} \ .
\label{e7}
\end{equation}
As pointed out in \cite{lipatov},
the parameter $\eta$ have to be chosen numerically large
but significantly smaller than the relative rapidities
of colliding particles:
$$
1<< \eta << \ln\frac{s}{M^2}\ .
$$
For our kinematical conditions $q_{1+}\to 0$, $k_{-}\to 0$
and the appropriate $\eta$
the inequality (\ref{e7}) leads to the restriction
\begin{equation}
|q_{1-}| < \frac{M^2}{\sqrt{s}}e^{+\eta}\ ,
\label{e8}
\end{equation}
where the limit tends to zero for large $s$.
It follows that the result of the calculation of the diagram
has to be multipied by some $\theta$-function
which allows the values of $q_{1-}$ to be
only within a very small interval around the point
of the pole at $q_{1-}=0$.
For the diagram 3 the same condition reads for $q_{2-}$.
This means that if we understand (\ref{ampp})
in the sense of generalized function
then terms (\ref{ampp1}) should be dropped
and only the $\delta$-function terms (\ref{ampp0})
remain.
It can be shown that for the diagram 1 in Fig. \ref{fig2}
from the same analysis of rapidity restrictions
it does not follow such a narrow limit for values of $q_{1,2-}$.
Qualitatively, one can say that
the multiperipheral diagram 1 describes emission
for arbitrary relations between $q_{1-}$, $q_{2-}$ and $p_{-}$
excluding the regions close to $q_{1,2-}=0$
but the diagrams 2,3 describe emission from this quasi-elastic
scattering region when $q_{1-}=0$ or $q_{2-}=0$.
As a result, in the effective action formalism the total amplitude
${\cal A}^{ef}$ is given
by the sum of contribution from the R$\to$RRP vertex (\ref{aeff}) and
the part (\ref{ampp0}) of the double reggeon exchange containing
$\delta(q_{1,2-})$. In the following section we shall see that this
identically coincides with the amplitude calculated by the standard
perturbative QCD provided the singularities at $q_{1,2-}=0$ in
the part with R$\to$RRP vertex are understood in the principal value
sense.
\begin{figure}[h]
\leavevmode \centering{\epsfysize=0.15\textheight\epsfbox{fig3.eps}}
\caption{Diagrams containing the three-reggeon vertex}
\label{fig3}
\end{figure}
In the effective vertex formalism
the three diagrams in Fig. \ref{fig3} containing the R$\to$RR
transition (three-reggeon) vertex can also be drawn.
The found contribution from the diagram in Fig. \ref{fig2},1
actually transforms into the contribution of one
of these diagrams depending on the correct multiregge character
of emission in a given kinematics. Therefore
these diagrams should not be added to the amplitude ${\cal A}^{ef}$
to avoid double counting.
They rather describe emisssion for particular relations
between $p_-,\,q_{1-}$ and $q_{2-}$.
For example the diagram in Fig. \ref{fig3},2
corresponds to the situation when $p_-\simeq -q_{1-}>>q_{2-}$.
Of special interest is the diagram
in Fig. \ref{fig3},3, which corresponds to the situation when
$|q_{1,2-}|>>p_-$ considered in Section 5.
\begin{figure}[h]
\leavevmode \centering{\epsfysize=0.15\textheight\epsfbox{fig4.eps}}
\caption{The vertex R$\to$RR}
\label{fig4}
\end{figure}
The three-reggeon vertex is determined by the
contribution $S_3$ to the action ~\cite{lipatov}. The relevant term
in the Lagrangian is
\begin{equation}
{\cal L}_3=
-2g{\rm Tr}\,\Big( A_-\partial^{-1}_-A_-\partial^2_\perp A_+\Big)
=-{g}{\rm Tr}\,\Big(\Big[ A_-,\partial^{-1}_-A_-\Big]
\partial^2_\perp A_+\Big).
\end{equation}
It corresponds to the vertex shown in Fig. \ref{fig4}:
\[
{\cal L}_3=-{g}(-i)^3A_-^{b_1}\partial_-^{-1}A_-^{b_2}
\partial^2_\perp A_+^a{\rm Tr}\Big([t^{b_1},t^{b_2}] t^a\Big) \, .
\]
We have
\[
{\rm Tr}\Big([t^{b_1},t^{b_2}] t^a\Big)=\frac{i}{2}f^{ab_1b_2}
\]
and also
\[\partial_-^{-1}A_-^{b_2}=-i\frac{1}{q_{2-}}A_-^{b_2},\ \
\partial^2_\perp A_+^a=-q_{\perp}^2A_+^a \ . \]
It leads to the vertex
\begin{equation}
\Gamma_{R\to RR}=\frac{q_\perp^2}{2q_{1-}}f^{ab_1b_2}=
-\frac{q_\perp^2}{2q_{2-}}f^{ab_1b_2} \ .
\label{rrrv}
\end{equation}
Note that the kinematical condition $\partial_- A_+ =0$
applied to ${\cal L}_3$ is equivalent to
the momentum relation $q_{1-}+q_{2-}=0$
which is fully correspondent to
the considered case $|q_{1,2-}|>>p_-=-(q_{1-}+q_{2-})$.
Using the vertex we find that the contribution of diagram 3 in
Fig. \ref{fig3} to the amplitude is
\begin{equation}
{\cal A}^{ef}_{R\to RR}= 32(kl)^2\frac{1}{2k_+q^2}t^d V_3 \, ,
\end{equation}
where
\begin{equation}
V_3=
i\frac{q_\perp^2}{q_{1-}}L(p,q_1+q_2)f^{adc}f^{cb_1b_2} \ .
\label{add2}
\end{equation}
As it will be calculated in Section 5, this contribution
is exactly equal to the expression
(\ref{aa}) obtained for the antisymmetric part
of the amplitude in the kinematics $q_{1,2-}>>p_-$.
Clearly, it also coincides with the contribution of the diagram
of Fig. ~\ref{fig2},1 with the R$\to$RRP vertex in this limit,
since the latter gives the part of the amplitude
free from $\delta(q_{1,2})$ terms in the general case.
It assumes the principal value prescription also
for the pole at $q_{1,2-}=0$ in the vertex (\ref{rrrv}).
So in this particular kinematics the part
of the amplitude without $\delta$-functions
can be fully described by the diagram 3 in Fig. \ref{fig3}.
\section{Production amplitude in the lowest order of the perturbative QCD}
We study production of a gluon with momentum $p$ in collision
of the projectile with momentum $k$ and two targets with
their intial momenta $l$ each.
In the standard Feynman formalism the production amplitude
is given by 6 diagrams $A,...,F$ in Fig. \ref{fig5}
summed with the same diagrams with the transposed targets
$1\leftrightarrow 2$.
\begin{figure}[h]
\leavevmode \centering{\epsfysize=0.35\textheight\epsfbox{fig5.eps}}
\caption{Production amplitude off two scattering centers in the
lowest order}
\label{fig5}
\end{figure}
\subsection{Diagrams $A,B,C$}
We start from three similar diagrams $A,B$ and $C$.
In all the three the quark line generates a momentum factor
\begin{equation}
M=-32i(kl)^2\frac{k_+}{p_+}(ep)_\perp.
\end{equation}
Here
the factor $-i$ comes from 2 vertices, 2 propagators and the minus sign
which originates from the relation
\begin{equation}
(ek)=e_-k_+=-\frac{k_+}{p_+}(ep)_\perp.
\end{equation}
The colour factors are
\begin{equation}
C_A=t^{b_2}t^{b_1}t^a,\ \ C_B=t^{b_2}t^{a}t^{b_1},\ \
C_C=t^at^{b_2}t^{b_1}.
\end{equation}
The propagators contribute factors
\[
P_A=\frac{1}{[(k-p)^2+i0][(k'+q_2)^2+i0]},\ \
P_B=\frac{1}{[(k-q_1)^2+i0][(k'+q_2)^2+i0]},\]\begin{equation}
P_C=\frac{1}{[(k-q_1)^2+i0][(k'+p)^2+i0]}.
\end{equation}
Using $k_+p_->>|p_\perp^2|$ we can write
\[(k-p)^2=-(k'+p)^2=\frac{k_+}{p_+}p_\perp^2\]
to finally obtain
\begin{equation}
A=-32i(kl)^2\frac{(ep)_\perp}{p_\perp^2}
\frac{1}{(k'+q_2)^2+i0}t^{b_2}t^{b_1}t^a
\end{equation}
and
\begin{equation}
C=32i(kl)^2\frac{(ep)_\perp}{p_\perp^2}
\frac{1}{(k-q_1)^2+i0}t^at^{b_2}t^{b_1}.
\end{equation}
As to the contribution $B$ we can formally write it in a similar
fashion
\begin{equation}
B=-32i(kl)^2\frac{(ep)_\perp}{p_\perp^2}\frac{k_+}{p_+}
\frac{p_\perp^2}{[(k-q_1)^2+i0][(k'+q_2)^2+i0]}t^{b_2}t^at^{b_1}.
\end{equation}
Its relative weight depends on the values of $q_{1-}$ and $q_{2-}$.
They are constrained by the conservation law (\ref{conslaw}), so that their
sum is of the order $p_-$.
\subsection{Diagrams $D$ and $E$}
To study the diagrams $D$, $E$ and $F$ we shall use some properties
of the gluon propagator in the axial gauge, derived in ~\cite{bra4}.
Namely, interaction with the projectile introduces into the gluon line
the vertex (see Fig. \ref{fig6})
\begin{equation}
2(pl)g_{\mu\nu}f^{abc}.
\end{equation}
So the two gluon propagators connected by this vertex, apart from the
standard denominators, contains the momentum factor
\begin{equation}
H_{\mu\nu}(p_1,p_2)=g_{\mu\nu}-\frac{l_\mu p_{1\nu}+p_{2\mu}l_\nu}{(pl)}
+\frac{(p_1p_2)l_\mu l_\nu}{(pl)^2}.
\label{hh}
\end{equation}
Here it is used that the "$+$" component of the gluon momentum does not change
in its interaction with the target.
\begin{figure}[h]
\leavevmode \centering{\epsfysize=0.12\textheight\epsfbox{fig6.eps}}
\caption{Insertion of an interaction with the target into the gluon line.
All lines are assumed outgoing; $(p_1l)=(p_2l)\equiv(pl)$.}
\label{fig6}
\end{figure}
Coupling to the gluon polarization vector gives
\begin{equation}
h_{\mu\nu}(p)e^\nu=e_\mu-l\mu\frac{(ep)}{(pl)}\equiv E_\mu(p),
\end{equation}
\begin{equation}
H_{\mu\nu}(p_1,p_2)e^\nu=E_\mu(p_2).
\label{hhe}
\end{equation}
Note that we have
\begin{equation}
E_-(p)=-\frac{1}{p_+}(ep)_\perp.
\end{equation}
Armed with these relations we start from diagram $D$.
The momentum factor from the quark and gluon lines is
\begin{equation}
M_D=32(pl)(kl)\frac{k_+}{p_+}(e,p+q_1)_\perp
\end{equation}
(we have $i^3$ from two vertices and the propagator in the quark line,
an $(-i)$ from the gluon line and a $(-1)$ from $E$).
The color factor is
\begin{equation}
C_D=f^{ab_1c}t^{b_2}t^c.
\end{equation}
The two propagators give
\begin{equation}
P=\frac{1}{[(k'+q_2)^2+i0][(p+q_1)^2+i0]}.
\end{equation}
So diagram $D$ gives a contribution
\begin{equation}
D=32(kl)^2\frac{(e,p+q_1)_\perp}{(p+q_1)_\perp^2}
\frac{(p+q_1)_\perp^2}{[(k'+q_2)^2+i0][(p+q_1)^2+i0]}
f^{ab_1c}t^{b_2}t^c,
\label{diad}
\end{equation}
where we used $(pl)=(kl)p_+/k_+$.
Analogous calculations give for diagram $E$
\begin{equation}
E=32(kl)^2\frac{(e,p+q_2)_\perp}{(p+q_2)_\perp^2}
\frac{(p+q_2)_\perp^2}{[(k-q_1)^2+i0][(p+q_2)^2+i0]}
f^{ab_2c}t^ct^{b_1}.
\label{eiad}
\end{equation}
\subsection{Diagram $F$}
Using our formulas (\ref{hh}) and (\ref{hhe}) we get the
momentum factor
\begin{equation}
M_F=-32i(pl)^2[kE(p+q_1+q_2)]=
32i(kl)^2\frac{k_+}{p_+}(e,p+q_1+q_2)_{\perp}.
\end{equation}
The colour factor is
\begin{equation}
C_F=f^{ab_2c}f^{cb_1d}t^d.
\end{equation}
The propagators give
\begin{equation}
P_F=\frac{1}{[(p+q_2)^2+i0][(p+q_2+q_1)^2+i0]}.
\end{equation}
In fact $p_-+q_{2-}+q_{1-}\simeq 0$ so in the second denominator
only the transversal part remains.
We get our final result
\begin{equation}
F=32i(kl)^2\frac{p_+}{k_+}\frac{(e,p+q_1+q_2)_{\perp}}{(p+q_1+q_2)_{\perp}^2}
\frac{1}{[(p+q_2)^2+i0]}f^{ab_2c}f^{cb_1d}t^d\ .
\end{equation}
\subsection{Parts with and without $\delta_(q_{1,2-})$}
To compare with the results of the effective action formalism,
we split the total perturbative contribution into parts with and
without $\delta_(q_{1,2-})$.
We start by rewriting the propagator
in the contribution of the diagram $B$ in Fig. \ref{fig5}
as
\[
P_B=\frac{1}{(-2k_+q_{1-}+i0)(2k_+q_{2-}+i0)}=
\frac{1}{2k_+(q_{1-}+q_{2-})}\Big(\frac{1}{-2k_+q_{1-}+i0}-
\frac{1}{2k_+q_{2-}+i0}\Big)\]\begin{equation}=
\frac{1}{2k_+p_-}\Big(\frac{1}{2k_+q_{1-}-i0}+
\frac{1}{2k_+q_{2-}+i0}\Big).
\end{equation}
So we find
\begin{equation}
B=32i(kl)^2\frac{(ep)_\perp}{p_\perp^2}
\Big(\frac{1}{2k_+q_{1-}-i0}+\frac{1}{2k_+q_{2-}+i0}\Big)
t^{b_2}t^at^{b_1}.
\end{equation}
Combining this contribution with terms $A$ and $C$
we obtain
\begin{equation}
A+B+C=-i32(kl)^2\frac{(ep)_\perp}{p_\perp^2}
\Big\{\frac{t^{b_2}[t^{b_1},t^a]}{2k_+q_{2-}+i0}+
\frac{[t^a,t^{b_2}]t^{b_1}}{2k_+q_{1-}-i0}\Big\}.
\label{abc}
\end{equation}
The contribution with 1 and 2 interchanged is
\begin{equation}
(A+B+C)(1\leftrightarrow 2)=
-i32(kl)^2\frac{(ep)_\perp}{p_\perp^2}
\Big\{\frac{t^{b_1}[t^{b_2},t^a]}{2k_+q_{1-}+i0}+
\frac{[t^a,t^{b_1}]t^{b_2}}{2k_+q_{2-}-i0}\Big\}.
\label{abc1}
\end{equation}
Separating parts with and without the $\delta$ function,
which we again denote by upper indices $0$ and $1$ respectively we obtain
\[
\Big(A+B+C+(1\leftrightarrow 2)\Big)^{(0)}=
-i32(kl)^2\pi \frac{(ep)_\perp}{p_\perp^2}
\Big(\delta(2k_+q_{2-})f^{b_1ac}\{t^{b_2},t^c\}+
\delta(2k_+q_{1-})f^{b_2ac}\{t^{b_1},t^c\}\Big)
\]
and
\[
\Big(A+B+C+(1\leftrightarrow 2)\Big)^{(1)}=
i32(kl)^2\frac{(ep)_\perp}{p_\perp^2}t^d\Big(
\frac{P}{2k_+q_{2-}}f^{b_2cd}f^{b_1ac}+
\frac{P}{2k_+q_{1-}}f^{b_1cd}f^{b_2ac}\Big).
\]
We pass to our diagrams Fig. \ref{fig5} $D$ and $E$.
We rewrite our formulas (\ref{diad}) and (\ref{eiad}) as
\[
D=32(kl)^2(e,p+q_1)_\perp\Big(\frac{P}{2k_+q_{2-}}-
i\pi\delta(2k_+q_{2-})\Big)
\frac{1}{(p+q_1)^2+i0}
f^{ab_1c}t^{b_2}t^c
\]
and
\[
E=-32(kl)^2(e,p+q_2)_\perp\Big(\frac{P}{2k_+q_{1-}}+i\pi
\delta(2k_+q_{1-})\Big)\frac{1}{(p+q_2)^2+i0}
f^{ab_1c}t^{b_2}t^c.
\]
Adding the terms with $1\leftrightarrow 2$ we find the principal value and
$\delta$-function parts of the sum as
\[
\Big(D+E+(1\leftrightarrow 2)\Big)^{(0)}\]\[=
-i32(kl)^2\pi\frac{(e,p+q_1)_\perp}{(p+q_1)_\perp^2}
\delta(2k_+q_{2-})f^{ab_1c}\{t^{b_2},t^c\}
-i32(kl)^2\pi\frac{(e,p+q_2)_\perp}{(p+q_2)_\perp^2}
\delta(2k_+q_{1-})f^{ab_2c}\{t^{b_1},t^c\}
\]
and
\[
\Big(D+E+(1\leftrightarrow 2)\Big)^{(1)}\]\[=
i32(kl)^2 \frac{P}{2k_+q_{2-}}\frac{(e,p+q_1)_\perp}{(p+q_1)^2+i0}
f^{b_2cd}f^{ab_1c}t^d
+i32(kl)^2\frac{P}{2k_+q_{1-}}\frac{(e,p+q_2)_\perp }{(p+q_2)^2+i0}
f^{b_1cd}f^{ab_2c}t^d.
\]
In the derivation of the part with the $\delta$-function we used that
\[\delta(2k_+q_{2-})\frac{1}{(p+q_1)^2+i0}=
\delta(2k_+q_{2-})\frac{1}{(p+q_1)^2_\perp},\]
since in this relation $(p+q_1)_-=-q_{2-}=0$.
In the general kinematics diagram $F$ does not contain denominators
singular in $q_{1,2-}$. So the only contribution is
\[\Big(F+(1\leftrightarrow 2)\Big)^{(1)}=
i 32(kl)^2\frac{p_+}{k_+}\frac{(e,p+q_1+q_2)_\perp}{(p+q_1+q_2)_\perp^2}t^d
\Big(\frac{f^{ab_2c}f^{cb_1d}}{(p+q_2)^2+i0}+
\frac{f^{ab_1c}f^{cb_2d}}{(p+q_1)^2+i0}\Big).
\]
Comparison of the total perturbative amplitude demonstrates that
the terms without the $\delta(q_{1,2-})$ sum into exacttly the
expression (\ref{aeff}) coming from the R$\to$RRP vertex, provided we
interprete
in the latter the singularities at $q_{1,2-}=0$ in the principal value
prescription. The terms
containing $\delta(q_{1,2-})$ are exactly reproduced by the corresponding
part (\ref{ampp0})
of the contribution from the double gluon exchange. Thus comparison
of the amplitudes
obtained in the effective action
formalism and perturbative QCD demonstrates that they are completely
identical. As a byproduct of this comparison we find that the singularties
at $q_{1,2-}=0$ in the effective R$\to$RRP vertex should be understood
in the principal value prescription.
\section{Production amplitude in the region $q_{1-},q_{2-}>>p_-$}
The "-"-components of momenta $q_{1,2}$ transferred to the targets
can be taken arbitrary. In fact in the inclusive cross-sections one
integrates over all values of $q_{1-}$ and $q_{2-}$ related by the
conservation law (\ref{conslaw}). The two production amplitudes enter
into the inclusive cross-section at different momenta of target quarks,
shifted by the momentum $\lambda$ transferred to the nucleus.
Its value is determined by the properties of the nucleus. In the rest
system of the target $\lambda_-\sim\sqrt{m\epsilon}$ where
$m$ is the nucleon mass and $\epsilon$ is the binding energy. In the same
system in the central emission region $p_-\sim |p_\perp|m/\sqrt{s}$, so that
$p_-<<\lambda_-$. As a result, in the integration over $q_{1,2-}$
the essential values of $q_{1,2-}$ are determined by $\lambda_-$ and so
are much larger than $p_-$
In this kinematical region
\begin{equation}
q_{1-},q_{2-}>>p_-,\ \ q_{-}+q_{2-}=0
\label{cond}
\end{equation}
our expressions for the amplitude can be substantially simplified.
With (\ref{cond}) we have
\[
\Big|\frac{k_+}{p_+}\frac{1}{(k-q_1)^2}\Big|
=\Big|\frac{1}{p_+q_{1-}}\Big|<<\Big|\frac{1}{p_+p_{-}}\Big|=
\Big|\frac{1}{p_\perp^2}\Big|
\]
and the contribution $B$ is much smaller than $A$ and $C$. So in the
kinematics (\ref{cond}) the diagram $B$ can be neglected.
Also with (\ref{cond}) we have
\begin{equation}
(k-q_1)^2=(k'+q_2)^2=2k_+q_{2-}
\end{equation}
and for the sum $A+C$ we have
\[
A+C=-32i(kl)^2\frac{(ep)_\perp}{p_\perp^2}
\frac{1}{2k_+q_{2-}+i0}(t^{b_2}t^{b_1}t^a-t^at^{b_2}t^{b_1})\]\begin{equation}=
32(kl)^2\frac{(ep)_\perp}{p_\perp^2}
\frac{1}{2k_+q_{2-}+i0}(f^{b_1ac}t^{b_2}t^c+f^{b_2ac}t^ct^{b_1}).
\label{tac1}
\end{equation}
Next in the kinematics (\ref{cond}) we find that the contributions of diagrams
$D$ and $E$ are small, since in them $q_{2-}$ appears squared
in the denominators. However
the contribution from diagram $D$ contains a $\delta$-function, since the
poles in, say, $q_{2-}$ are located on different sides of the real axis.
In our kinematics $q_{1-}=-q_{2-}$ so that
\begin{equation}
D=32(pl)^2\frac{(e,p+q_1)_\perp}{(p+q_1)_\perp^2}
\frac{(p+q_1)_\perp^2}{[2k_+q_{2-}+i0][-2p_+q_{2-}+(p+q_1)_\perp^2+i0]}
f^{ab_1c}t^{b_2}t^c.
\label{diad1}
\end{equation}
Integration over $q_{2-}$ shows that
that $D$ contains a $\delta$-contribution
\begin{equation}
\Delta D=-32(kl)^22\pi i\delta(2k_+q_{2-})
\frac{(e,p+q_1)_\perp}{(p+q_1)_\perp^2}
f^{ab_1c}t^{b_2}t^c.
\end{equation}
In $E$ the two poles in $q_{2-}$ are located on the same side of the real
axis, so that it does not contain $\delta$-like contributions.
So the only contribution which remains from $D$ and $E$ in our
kinematics is
\[
D+D(1\leftrightarrow 2)=\Delta D+\Delta D(1\leftrightarrow 2)\]\begin{equation}=
32(kl)^22\pi i\delta(2k_+q_{1-})\Big(\frac{(e,p+q_2)_\perp}{(p+q_2)_\perp^2}
f^{ab_2c}t^ct^{b_1}+\frac{(e,p+q_1)_\perp}{(p+q_1)_\perp^2}
f^{ab_1c}t^ct^{b_2}\Big).
\end{equation}
Finally the expression for $E$ can be simplified to
\begin{equation}
F=32i(kl)^2\frac{(e,p+q_1+q_2)_{\perp}}{(p+q_1+q_2)_{\perp}^2}
\frac{1}{2k_+q_{2-}+i0}f^{ab_2c}f^{cb_1d}t^d.
\end{equation}
The total contribution to the amplitude
\begin{equation}
{\cal A}=A+C+D+F+(1\leftrightarrow 2)
\end{equation}
can be rewritten in the form which follows the rules
of the BFKL-Bartels formalism. In the following we suppress the
common factor $32(kl)^2$. Having this in mind
we add to our
contribution a term
\begin{equation}
T=T^{(1)}+T^{(2)},
\end{equation}
where
\begin{equation}
T^{(1)}=-\frac{1}{2k_+q_{2-}+i0}
\Big(\frac{(e,p+q_2)_\perp}{(p+q_2)_\perp^2}f^{acb_2}t^{b_1}t^c+
\frac{(e,p+q_1)_\perp}{(p+q_1)_\perp^2}f^{acb_1}t^{b_2}t^c\Big)
+\Big(1\leftrightarrow 2\Big)
\label {term1}
\end{equation}
and
\begin{equation}
T^{(2)}=-2\pi i\delta(2k_+q_{2-})
\Big(\frac{(e,p+q_2)_\perp}{(p+q_2)_\perp^2}f^{acb_2}t^{b_1}t^c+
\frac{(e,p+q_1)_\perp}{(p+q_1)_\perp^2}f^{acb_1}t^{b_2}t^c\Big),
\label{term2}
\end{equation}
which is zero, since with $q_{1-}=-q_{2-}$
\[\frac{1}{2k_+q_{2-}+i0}+\frac{1}{2k_+q_{1-}+i0}
=-2\pi i\delta(2k_+q_{2-}).\]
We see that term $T^{(2)}$ cancels the contribution
from $D$:
\begin{equation}
D+D(1\leftrightarrow 2)+T^{(2)}=0.
\end{equation}
We present the remaining expressions as:
\begin{equation}
A+C+(1\lra2)=\frac{(ep)_\perp}{p_\perp^2}\frac{1}{2k_+q_{2-}+i0}
\Big(f^{b_1ac}t^{b_2}t^c+
f^{b_2ac}t^ct^{b_1}\Big)+\Big(1\leftrightarrow 2\Big)
\label{tac}
\end{equation}
and
\begin{equation}
F+(1\leftrightarrow 2)=i\frac{(e,p+q_1+q_2)_{\perp}}{(p+q_1+q_2)_{\perp}^2}
\frac{1}{2k_+q_{2-}+i0}
f^{ab_2c}f^{cb_1d}t^d+\Big(1\leftrightarrow 2\Big)
\label{tf}
\end{equation}
and shall transform the explicitly shown expressions.
We combine the second term in (\ref{term1}) with the first term in
(\ref{tac}) to get
\begin{equation}
T_1=\frac{1}{2k_+q_{2-}+i0}\Big(\frac{(ep)_\perp}{p_\perp^2}-
\frac{(e,p+q_1)_\perp}{(p+q_1)_\perp^2}\Big)f^{b_1ac}t^{b_2}t^c.
\label{t1}
\end{equation}
We transform the colour factor in the first term in (\ref{term1})
as
\begin{equation}
f^{acb_2}t^{b_1}t^c= f^{acb_2}t^ct^{b_1}+if^{acb_2}f^{b_1cd}t^d
\label{rel}
\end{equation}
and the contribution from the first term in this relation
combine with the second term in (\ref{tac}) to get
\begin{equation}
T_2=\frac{1}{2k_+q_{2-}+i0}\Big(\frac{(ep)_\perp}{p_\perp^2}-
\frac{(e,p+q_2)_\perp}{(p+q_2)_\perp^2}\Big)f^{b_2ac}t^ct^{b_1}.
\label{t2}
\end{equation}
Finally we combine the contribution from the second term in (\ref{rel})
with $F$ and obtain
\begin{equation}
T_3=-i\frac{1}{2k_+q_{2-}+i0}\Big(\frac{(e,p+q_2)_\perp}{(p+q_2)_\perp^2}-
\frac{(e,p+q_1+q_2)_{\perp}}{(p+q_1+q_2)_{\perp}^2}\Big)
f^{ab_2c}f^{cb_1d}t^d.
\label{t3}
\end{equation}
As we observe that all the contributions nicely arrange into three
terms with R$\to$RP effective vertices.
We find that the total amplitude is a sum
\begin{equation}
{\cal A}=\Big({\cal A}_1+{\cal A}_2+{\cal A}_3\Big)+\Big(1\leftrightarrow 2\Big),
\end{equation}
where
\begin{equation}
{\cal A}_1=\frac{1}{(k'+q_2)^2+i0}L(p,q_1)f^{b_1ac}t^{b_2}t^c,
\label{amp1}
\end{equation}
\begin{equation}
{\cal A}_2=\frac{1}{(k-q_1)^2+i0}L(p,q_2)f^{b_2ac}t^ct^{b_1},
\label{amp2}
\end{equation}
\begin{equation}
{\cal A}_3=-i\frac{1}{2k_+q_{2-}+i0}
B(p,q_2,q_1)f^{ab_2c}f^{cb_1d}t^d=
-i\frac{(ql)}{(kl)}\frac{1}{(p+q_2)^2+i0}B(p,q_2,q_1)f^{ab_2c}f^{cb_1d}t^d.
\label{amp3}
\end{equation}
The three terms ${\cal A}_{1,2,3}$ correspond to the diagrams shown in Fig.
\ref{fig7} with the vertices $L$ and $B$ and expected dependence
on $q_{1-}$ and $q_{2-}$.
\begin{figure}[h]
\leavevmode \centering{\epsfysize=0.15\textheight\epsfbox{fig7.eps}}
\caption{Production amplitude off two scattering centers in terms of the
$L$ and $B$ vertices}
\label{fig7}
\end{figure}
The denominator $2k_+q_{2-}+i0$ splits into parts symmetric and antisymmetric
with repect to the change $q_{2-}\to q_{1-}$, that is $u\to s$
\begin{equation}
\frac{1}{2k_+q_{2-}+i0}=P\frac{1}{2k_+q_{2-}}-\pi i\delta(2k_+q_{2-}).
\end{equation}
The first term corresponds to the antisymmetric and the second to
symmetric parts. Note that the $\delta$ term does not literally mean
that $q_{2-}=0$, which seems to violate the adopted kinematics
$q_{2-}>>p_{2-}$. Rather it means that in the integration over $q_{2-}$
in the complex plane one can neglect terms of the order $p_-$ in the integrand
and take the integral around the resulting singularity at $q_{2-}=0$.
It is instructive to find the final expressions for the
parts of the amplitude symmetric and
antisymetric under the interchange $s\leftrightarrow u$,
$\cal{A}^{(+)}$ and
$\cal{A}^{(-)}$ respectively. The antisymmetric part is
\[
{\cal{A}}^{(-)}=P\frac{1}{2k_+q_{2-}}\Big[L(p,q_1)f^{b_1ac}t^{b_2}t^c
+L(p.q_2)f^{b_2ac}t^ct^{b_1}-iB(p,q_2,q_1)f^{ab_2c}f^{cb_1d}t^d\]\[-
L(p,q_2)f^{b_2ac}t^{b_1}t^c
-L(p.q_1)f^{b_1ac}t^ct^{b_2}+iB(p,q_1,q_2)f^{ab_1c}f^{cb_2d}t^d\Big].
\]
Combining terms with the same R$\to$RP vertices
and using
\begin{equation}
L(p,q_1)+B(p,q_1,q_2)=L(p,q_2)+B(p,q_2,q_1)=L(p,q_1+q_2)
\end{equation}
we find that the antisymmetric part amplitude is given by a
simple expresssion
\begin{equation}
{\cal{A}}^{(-)}=P\frac{1}{2k_+q_{2-}}L(p,q_1+q_2)it^d
(f^{b_1ac}f^{b_2cd}-f^{b_2ac}f^{b_1cd}).
\end{equation}
Using further the Jacobi identity
\[f^{b_1ac}f^{b_2cd}-f^{b_2ac}f^{b_1cd}=f^{dac}f^{cb_1b_2}\]
we finally obtain
\begin{equation}
{\cal{A}}^{(-)}=P\frac{1}{2k_+q_{2-}}L(p,q_1+q_2)it^d
f^{dac}f^{cb_1b_2}.
\label{aa}
\end{equation}
As expected, in the adopted kinematics
the antisymmetric part gives a contribution
which corresponds to the reggeon diagram 3 (\ref{add2})
in Fig. \ref{fig3}.
From its structure one concludes that it corresponds to the
interaction with the target quarks having the $t$-channel with the
gluon colour and gives no contribution to to the interaction with the
vacuum channel.
The symmetric of the amplitude is given by
\[
{\cal{A}}^{(+)}=-\pi i\delta(2k_+q_{2-})\Big[L(p,q_1)f^{b_1ac}\{t^{b_2},t^c\}
+L(p.q_2)f^{b_2ac}\{t^{b_1},t^c\}\]\[-B(p,q_2,q_1)f^{b_2ac}[t^{b_1},t^c]
-B(p,q_1,q_2)f^{b_1ac}[t^{b_2},t^c]\Big],\]
or, in terms of R$\to$RP vertices,
\[
{\cal A}^{(+)}=-\pi i\delta(2k_+q_{2-})\Big[2L(p,q_1)f^{b_1ac}t^ct^{b_2}
+2L(p.q_2)f^{b_2ac}t^ct^{b_1}-iL(p,q_1+q_2)\Big(f^{b_2ac}f^{cb_1d}+
f^{b_1ac}f^{cb_2d}\Big)t^d\Big]\ .
\]
\section{Conclusions}
We have found that in the general kinematics, out of all possible
diagrams which one can formally draw in the effective action approach,
only quite a few are to be taken into account in accordance with the
requirement of the multiregge kinematics. The contribution from the
R$\to$RRP effective vertex gives all the terms in the amplitude which
do not contain $\delta(q_{1,2-})$. The latter are supplied by the
double reggeon exchange. All the rest diagrams just reproduce the limiting
cases of the R$\to$RRP contribution in different kinematical regions,
that is relations between $q_{1,2-}$ and $p_-$.
In the kinematics
$q_{1,2-}>>p_-$ appropriate for the calculation of the inclusive
cross-section of gluon production on two centers, the production amplitude
is reproduced by a set of reggeon diagrams with
effective vertices multiplied by energetic factors $1/s$ and $1/u$.
The set is the same as used in the calculation of the inclusive cross-sections
directly in the purely transversal technique. So one expects that after
integration over the intermediate target momenta one will obtain the same
results for the inclusive cross-section as currently obtained in the
purely transversal approaches (BFKL-Bartels or dipole).
A byproduct of our study is justification of the rule of
integration over the singularities at $q_{1,2-}=0$ in the principal value
prescription imposed {\it ad hoc} in ~\cite{bravyaz}. As follows from our
study, terms in the effective vertex with this singularity contribute
only to the part of the amplitude without $\delta$ functions in
the transferred energies
and so should be integrated in this prescription.
\section{Acknowledgement}
The authors are most thankful to J.Bartels and G.P.Vacca for helpful and
constructive discussions. M.A.B. is indebted to the 2nd Institute
of theoretical Physics at Hamburg university
for hospitality and financial support.
|
\section{Introduction}
The chromatic number of graphs with bounded degrees has been studied
for many years. Brooks' theorem perhaps is one of the most
fundamental results; it is included by many textbooks on
graph theory. Given a simple connected graph $G$, let $\Delta(G)$ be
the maximum degree, $\omega(G)$ be the clique number,
and $\chi(G)$ be the chromatic number. Brooks'
theorem states that $\chi(G)\leq \Delta(G)$ unless $G$ is a complete
graph or an odd cycle.
Reed \cite{reed1} proved that $\chi(G)\leq
\Delta(G)-1$ if $\omega(G)\leq \Delta(G)-1$ and $\Delta(G)\geq \Delta_0$
for some large constant $\Delta_0$. This excellent result was proved by
probabilistic methods, and $\Delta_0$ is at least
hundreds. Before this result, Borodin and Kostochka \cite{bk} made
the following conjecture.
{\bf Conjecture \cite{bk}:}
Suppose that $G$ is a connected graph.
If $\omega(G)\leq \Delta(G)-1$
and $\Delta(G)\geq 9$, then we have
$$\chi(G)\leq \Delta(G)-1.$$
If the conjecture is true, then it is best possible since there is a
$K_8$-free graph $G=C_5\boxtimes K_3$ (actually $K_7$-free, see Figure \ref{fig:1}) with
$\Delta(G)=8$ and $\chi(G)=8$.
\begin{figure}[htbp]
\label{fig:1}
\centerline{\psfig{figure=image/c53.eps, width=0.4\textwidth} }
\caption{The graph $C_5 \boxtimes K_3$.}
\end{figure}
Here we use the following notation of the strong product. Given two graphs $G$
and $H$, the {\it strong product} $G\boxtimes H$ is the graph with vertex
set $V(G)\times V(H)$, and $(a,x)$ is connected to $(b,y)$ if one of the following holds
\begin{itemize}
\item $a=b$ and $xy\in E(H)$,
\item $ab\in E(G)$ and $x=y$,
\item $ab\in E(G)$ and $xy\in E(H)$.
\end{itemize}
Reed's result \cite{reed1} settled Borodin and Kostochka's conjecture for sufficiently large $\Delta(G)$,
but the cases with small $\Delta(G)$ are hard to cover using the probabilistic method.
In this paper we consider a fractional analogue of this problem.
The fractional chromatic number $\chi_f(G)$ can be defined as
follows. A $b$-{\it fold coloring} of $G$ assigns a set of $b$
colors to each vertex such that any two adjacent vertices receive
disjoint sets of colors. We say a graph $G$ is $a$:$b$-{\it
colorable} if there is a $b$-fold coloring of $G$ in which each
color is drawn from a palette of $a$ colors. We refer to such a
coloring as an $a$:$b$-coloring. The $b$-{\it fold coloring number},
denoted by $\chi_b(G)$, is the smallest integer $a$ such that $G$
has an $a$:$b$-coloring. Note that $\chi_1(G)=\chi(G)$. It was shown
that $\chi_{a+b}(G) \leq \chi_a(G) + \chi_b(G) $. The {\it
fractional chromatic number} $\chi_f(G)$ is
$ \underset{b \rightarrow \infty} \lim \frac{\chi_b(G)}{b}.$
By the definition, we have $\chi_f(G)\leq \chi(G)$.
The fractional chromatic number can be viewed as a relaxation of the
chromatic number. Many problems involving the chromatic number can be asked
again using the fractional chromatic number. The fractional analogue often
has a simpler solution than the original
problem. For example, the famous $\omega-\Delta-\chi$ conjecture of
Reed \cite{reed} states that for any simple graph $G$, we have
$$\chi(G)\leq \left\lceil \frac{\omega(G)+\Delta(G)+1}{2}\right
\rceil.$$
The fractional analogue of $\omega-\Delta-\chi$ conjecture was
proved by Molloy and Reed \cite{mr}; they actually proved a stronger
result with ceiling removed, i.e.,
\begin{equation} \label{eq:1}
\chi_f(G)\leq \frac{\omega(G)+\Delta(G)+1}{2}.
\end{equation}
In this paper, we classify all connected graphs $G$ with
$\chi_f(G)\geq \Delta(G)$.
\begin{theorem}\label{main}
A connected graph $G$ satisfies $\chi_f(G)\geq \Delta(G)$ if and only if
$G$ is one of the following
\begin{enumerate}
\item a complete graph,
\item an odd cycle,
\item a graph with $\omega(G)=\Delta(G)$,
\item $C^2_8$,
\item $C_5\boxtimes K_2$.
\end{enumerate}
\end{theorem}
For the complete graph $K_n$, we have $\chi_f(K_n)=n$ and
$\Delta(K_n)=n-1$. For the odd cycle $C_{2k+1}$, we have
$\chi_f(C_{2k+1})=2+\frac{1}{k}$ and $\Delta(C_{2k+1})=2$. If $G$ is
neither a complete graph nor an odd cycle but contains a clique of
size $\Delta(G)$, then we have
\begin{equation} \label{less}
\Delta(G)\leq \omega(G)\leq \chi_f(G)\leq \chi(G)\leq \Delta(G).
\end{equation}
The last inequality is from Brooks' theorem. The sequence of
inequalities above implies $\chi_f(G)=\Delta(G)$.
If $G$ is a vertex-transitive graph,
then we have \cite{su}
$$\chi_f(G)=\frac{|V(G)|}{\alpha(G)},$$
where $\alpha(G)$ is the independence number of $G$.
Note that both graphs $C^2_8$ and $C_5\boxtimes K_2$
are vertex-transitive and have the independence number $2$.
Thus we have
$$\chi_f(C^2_8)=4=\Delta(C^2_8)\quad \mbox { and }\quad
\chi_f(C_5\boxtimes K_2)=5= \Delta(C_5\boxtimes K_2).$$
\begin{figure}[htbp]
\centerline{ \psfig{figure=image/c8square.eps, width=0.24\textwidth} \hspace{4cm} \psfig{figure=image/c52.eps,
width=0.24\textwidth} }
\centerline{$C_8^2$ \hspace{6cm} $C_5 \boxtimes K_2$}
\caption{The graph $C_8^2$ and $C_5 \boxtimes K_2$ .}
\end{figure}
Actually, Theorem \ref{main} is a corollary of the following stronger
result.
\begin{theorem}\label{submain}
Assume that a connected graph $G$ is neither $C^2_8$ nor
$C_5\boxtimes K_2$. If $\Delta(G)\geq 4$ and $\omega(G)\leq
\Delta(G)-1$, then we have
$$\chi_f(G)\leq \Delta(G) -\frac{2}{67}.$$
\end{theorem}
\noindent {\bf Remark:} In the case $\Delta(G)=3$, Heckman and
Thomas \cite{ht} conjectured that $\chi_f(G) \leq 14/5$ if $G$ is
triangle-free. Hatami and Zhu \cite{hz} proved $\chi_f(G) \leq 3-
\frac{3}{64}$ for any triangle-free graph $G$ with $\Delta(G) \leq
3$. The second and third authors showed an improved
result $\chi_f(G) \leq 3 - \frac{3}{43}$ in the previous paper \cite{lup}. Thus we need only consider
the cases $\Delta(G) \geq 4$.
For any
connected graph $G$ with sufficiently large $\Delta(G)$ and
$\omega(G) \leq \Delta(G)-1$, Reed's result \cite{reed1}
$\chi(G)\leq \Delta(G)-1$ implies $\chi_f(G) \leq \Delta(G)-1$.
The method introduced in \cite{hz} and strengthened in \cite{lup},
has a strong influence on this paper. The readers are encouraged to
read these two papers \cite{hz, lup}.
Let $f(k)=\inf_G\{\Delta(G)-\chi_f(G)\}$, where the infimum is taken
over all connected graphs $G$ with $\Delta(G)=k$ and not one of the
graphs listed in Theorem \ref{main}. Since $\chi_f(G)\geq \omega(G)$,
by taking a graph with $\omega(G)=\Delta(G)-1$, we have $f(k)\leq 1$.
Theorem \ref{submain} says
$f(k)\geq \frac{2}{67}$ for any $k\geq 4$.
Reed's result \cite{reed1}
implies $f(k)= 1$ for sufficiently large $k$. Heckman and Thomas
\cite{ht} conjectured $f(3)=1/5$. It is an interesting problem to
determine the value of $f(k)$ for small $k$. Here we conjecture
$f(4)=f(5)=\frac{1}{3}$. If Borodin and Kostochka's conjecture is
true, then $f(k)= 1$ for $k\geq 9$.
Theorem 2 is proved by induction on $k$.
Because the proof is quite long, we split the proof into
the following two lemmas.
\begin{lemma} \label{f4}
We have $f(4) \geq \frac{2}{67}$.
\end{lemma}
\begin{lemma}\label{increase}
For each $k \geq 6$, we have $f(k)\geq \min \left\{f(k-1),
\frac{1}{2}\right\}.$ We also have $f(5)\geq \min\left\{f(4),
\frac{1}{3}\right\}$.
\end{lemma}
It is easy to see the combination of Lemma \ref{f4} and Lemma
\ref{increase} implies Theorem \ref{submain}. The idea of reduction
comes from the first author, who pointed out $f(k)\geq \min \left\{f(k-1),
\frac{1}{2}\right\}$ for $k\geq 7$ based on his recent results
\cite{king}. The second and third authors orginally proved
$f(k)\geq \frac{C}{k^5}$ (for some $C>0$) using different method in the first version;
they also prove the reductions
at $k=5,6$, which are much harder than
the case $k\geq 7$. We do not know
whether a similar reduction exists for $k=4$.
The rest of this paper is organized as follows. In section 2, we
will introduce some notation and prove Lemma \ref{increase}. In
section 3 and section 4, we will prove $f(4) \geq \frac{2}{67}$.
\section{Proof of Lemma \ref{increase}}
In this paper, we use the following notation. Let $G$ be a simple
graph with vertex set $V(G)$ and edge set $E(G)$. The {\it
neighborhood} of a vertex $v$ in $G$, denoted by $\Gamma_G(v)$, is
the set $\{u \colon uv \in E(G)\}$. The {\it degree} $d_G(v)$ of $v$
is the value of $|\Gamma_G(v)|$. The {\it independent
set} (or {\it stable set}) is a set $S$ such that no edge with both
ends in $S$. The {\it independence number} $\alpha(G)$ is the largest
size of $S$ among all the independent sets $S$ in $G$. When $T
\subset V(G)$, we use $\alpha_G(T)$ to denote the independence number
of the induced subgraph of $G$ on $T$. Let $\Delta(G)$ be the maximum
degree of $G$. For any two vertex-sets $S$ and $T$, we define
$E_G(S,T)$ as $\{uv \in E(G): u \in S \ \textrm{and} \ v \in T\}$.
Whenever $G$ is clear under context, we will drop the subscript $G$
for simplicity.
If $S$ is a subset of vertices in $G$, then {\it contracting} $S$
means replacing vertices in $S$ by a single fat vertex, denoted by $\underline{S}$,
whose incident
edges are all edges that were incident to at least one vertex in
$S$, except edges with both ends in $S$. The new graph obtained by
contracting $S$ is denoted by $G/S$. This operation is also known
as {\it identifying vertices of $S$} in the literature. For
completeness, we allow $S$ to be a single vertex or even the empty
set. If $S$ only consists of a single vertex, then $G/S=G$; if
$S=\emptyset$, then $G/S$ is the union of $G$ and an isolated
vertex. When $S$ consists of $2$ or $3$ vertices, for convenience,
we write $G/uv$ for $G/\{u,v\}$ and $G/uvw$ for $G/\{u,v,w\}$; the
fat vertex will be denoted by $\underline{uv}$ and
$\underline{uvw}$, respectively. Given two disjoint subsets $S_1$
and $S_2$, we can contract $S_1$ and $S_2$ sequentially. The order
of contractions does not matter; let $G/S_1/S_2$ be the resulted
graph. We use $G-S$ to denote the subgraph of $G$ induced by
$V(G)-S$.
In order to prove Lemma \ref{increase}, we need use the following theorems
due to King \cite{king}.
\begin{theorem}[King \cite{king}]
\label{king0}
If a graph $G$ satisfies $\omega(G) > \frac{2} {3} (\Delta(G) + 1)$,
then $G$ contains a stable set $S$ meeting every maximum clique.
\end{theorem}
\begin{theorem}[King \cite{king}]
\label{king}
For a positive integer $k$, let $G$ be a graph with vertices
partitioned into cliques $V_1,\ldots,V_r$. If for every $i$ and
every $v \in V_i$, $v$ has at most $\min\{k, |V_i|-k\}$ neighbors
outside $V_i$, then $G$ contains a stable set of size $r$.
\end{theorem}
\begin{lemma} \label{6to5}
Suppose that $G$ is a connected graph with $\Delta(G)\leq 6$ and $\omega(G)\leq 5$.
Then there exists an independent set meeting all induced copies
of $K_5$ and $C_5\boxtimes K_2$.
\end{lemma}
\noindent
{\bf Proof:} We first show that there exists an independent set
meeting all copies of $K_5$. If $G$ contains no $K_5$, then this is
trivial. Otherwise, we can apply Theorem \ref{king0} to get the
desired independent set since $\omega(G)>\frac{2}{3}(\Delta(G)+1)$
is satisfied.
Now we prove the Lemma by contradiction. Suppose the Lemma is false.
Let $G$ be a minimum counterexample (with the smallest number of
vertices). For any independent set $I$, let $C(I)$ be the number of
induced copies of $C_5\boxtimes K_2$ in $G-I$. Among all independent
sets which meet all copies of $K_5$, there exists one such
independent set $I$ such that $C(I)$ is minimized.
Since $C(I)>0$, there is an induced copy of $C_5\boxtimes K_2$ in
$G-I$; we use $H$ to denote it. In $C_5\boxtimes K_2$, there is a
unique perfect matching such that identifying
the two ends of each edge in this matching results a $C_5$.
An edge in this unique matching is called a {\em canonical } edge.
We define a new graph $G'$ as follows: First we contract all
canonical edges in $H$ to get a $C_5$, where its vertices are called {\em fat}
vertices.
Second we add five edges
turning the $C_5$ into a $K_5$. Observe that each vertex in this
$C_5$ can have at most two neighbors in $G-H$ and $\Delta(G') \leq
6$. We will consider the following four cases.
\noindent{\bf Case 1:} There is a $K_6$ in the new graph $G'$. Since
the original graph $G$ is $K_6$-free, the $K_6$ is formed by the
following two possible ways.
{\bf Subcase 1a:} This $K_6$ contains $5$ fat vertices. By the symmetry of $H$,
there is an induced $C_5$ in $H$
such that the vertices in $C_5$ contain a common neighbor vertex $v$ in $G \setminus V(H)$, see Figure \ref{fig:1a}.
Since $H$ is $K_5$-free, we can find $x, y$ in this $C_5$ such that $x,y$ is a non-edge.
Let $I':=(I\setminus\{v\})\cup\{x, y\}$; $I'$ is also an
independent set. Observe that $v$ is not in any $K_5$ in $G-I'$. Thus the set $I'$ is also an independent set and meets every
$K_5$ in $G$. Since $C_5\boxtimes K_2$ is a $5$-regular graph, any
copy of $C_5\boxtimes K_2$ containing $v$ must contain at least one
of $x$ and $y$. Thus, $C(I')<C(I)$. Contradiction!
\begin{figure}[htbp]
\centerline{ {\psfig{figure=image/case1a.eps, width=0.32\textwidth}}
\hfil \psfig{figure=image/case1b.eps, width=0.32\textwidth}}
\begin{multicols}{2}
\caption{Subcase 1a.}
\label{fig:1a}
\newpage
\caption{Subcase 1b.}
\label{fig:1b}
\end{multicols}
\end{figure}
{\bf Subcase 1b:} This $K_6$ contains $4$ fat vertices.
Let $u,v$ be the other two vertices.
By the symmetry of $H$, there is a unique way to connect $u$ and $v$ to $H$
as shown by Figure \ref{fig:1b}.
Since $uv$ is an edge, one of $u$ and $v$ is not in $I$. We assume
$u \not \in I$. Let $\{x,y\} \subset \Gamma_G(v) \cap V(H)$ as shown
in Figure \ref{fig:1b} and $I'=I \setminus \{v\} \cup \{x,y\} $.
Observe that $I'$ is an independent set and $v$ is not in a $K_5$ in $G-I'$. Thus $I'$ is an independent set meeting each $K_5$ in $G$.
Since each $C_5 \boxtimes K_2$ containing $v$ must contain one of
$x$ and $y$. Thus $C(I')<C(I)$. Contradiction!
\noindent {\bf Case 2:}
There is a $K_5$ intersecting $H$ with $4$
vertices. Let $v$ be the vertex of this $K_5$ but not in $H$, see Figure \ref{fig:2}.
We have two subcases.
{\bf Subcase 2a:} The vertex $v$ has another neighbor $y$ in $H$ but
not in this $K_5$. Since $H$ is $K_5$-free, we can select a vertex
$x$ in this $K_5$ such that $xy$ is not an edge of $G$. Let
$I':=I\setminus \{v\}\cup \{x,y\}$. Note that $v$ is not in a $K_5$ in $G-I'$, and $I'$ is an independent set. Thus $I'$ is an
independent set meeting each $K_5$ in $ G$. Since any $C_5
\boxtimes K_2$ containing $v$ must contain one of $x$ and $y$, we
have $C(I')<C(I)$. Contradiction!
{\bf Subcase 2b:} All neighbors of $v$ in $H$ are in this $K_5$.
Let $x$ be any vertex in this $K_5$ other than $v$, and
$I':=I\setminus \{v\}\cup \{x\}$. In this case, there is only one
$K_5$ containing $v$. Thus, $I'$ is also an independent set meeting
every copy of $K_5$ in $G$. Observe that $\Gamma_G(v) \setminus
\{x\}$ is disconnected. If $v \in H'=C_5 \boxtimes K_2$, then
$\Gamma_G(v) \cap H'$ is connected. Thus $v$ is not in a $C_5\boxtimes
K_2$ in $G-I'$ and $C(I')< C(I)$. Contradiction!
\begin{figure}[htbp]
\centerline{ {\psfig{figure=image/case2a.eps, width=0.32\textwidth}}
\hfil \psfig{figure=image/case3.eps, width=0.32\textwidth}}
\begin{multicols}{2}
\caption{Case 2.}
\label{fig:2}
\newpage
\caption{Case 3.}
\label{fig:3}
\end{multicols}
\end{figure}
\noindent {\bf Case 3:} There is an induced subgraph $H'$
isomorphic to $C_5\boxtimes K_2$ such that $H'$ and $H$ are
intersecting, see Figure \ref{fig:3}.
Since $V(H)\cap V(H')\not =\emptyset$ and $H\not=H'$, we can find a
canonical edge $uv$ of $H$ and a canonical edge $uv'$ of $H'$ such
that $v\not\in V(H')$ and $v'\not\in V(H)$.
If $vv'$ is a non-edge, then let $I':=I\setminus \{v'\} \cup \{u\}$.
It is easy to check $I'$ is still an independent set. We also
observe that any possible $K_5$ containing $v'$ must also contain
$u$. Thus, $I'$ meets every copy of $K_5$ in $G$. We have $v'$ in no $C_5 \boxtimes K_2$ in $G-I'$ since $vv'$ is not an edge. We therefore
get $C(I')<C(I)$. Contradiction! If $vv'$ is an edge, then locally
there are two $K_5$ intersecting at $u$, $v$, and $v'$; say the
other four vertices are $x_1, x_2, y_1, y_2$, where two cliques are
$\{x_1,x_2,u,v,v'\}$ and $\{y_1,y_2,u,v,v'\}$, see Figure
\ref{fig:3}. Let $I'=I\cup\{x_1,y_1\}\setminus \{v'\}$. Note that
$I'$ is an independent set and $v'$ is not in a $K_5$ in $G-I'$. Thus
$I'$ is an independent set meeting each $K_5$ in $G$. Observe that
any copies of $C_5\boxtimes K_2$ containing $v'$ must contain one
of $x_1$ and $y_1$; we have $C(I')<C(I)$. Contradiction!.
\noindent
{\bf Case 4:} This is the remaining case, $G'$ is $K_6$-free. We
have $\omega(G')\leq 5$ and $|V(G')| < |V(G)|$. By the minimality of
$G$, there is an independent set $I'$ of $G'$ meeting every copy of
$K_5$ and $C_5\boxtimes K_2$. In $I'$, there is a unique vertex $x$
of the $K_5$ obtained from contracting canonical edges of $H$. Let
$uv$ be the canonical edge corresponding to $x$. Let $I''=I'
\setminus \{x\} \cup \{u\}$, we get an independent set $I''$ of $G$.
Note that any $v \in H \setminus \{u\}$ is not in any $K_5$ of
$G-I''$ by Case 2 as well as not in any $C_5 \boxtimes K_2$ of
$G-I''$ by Case 3. Thus $I''$ hits each $K_5$ in $G$ and $C(I'')=0$.
Contradiction! \hfill $\square$
The following lemma extends Theorem \ref{king0} when $\omega(G)=4$; a similar result was proved independently in \cite{cek}.
\begin{lemma} \label{5to41}
Let $G$ be a connected graph with $\Delta(G)\leq 5$ and $\omega(G)\leq 4$.
If $G \not = C_{2l+1} \boxtimes K_2$ for some $l \geq 2$,
then there is an independent set $I$ hitting all copies of $K_4$ in $G$.
\end{lemma}
{\bf Proof:} We will prove it by contradiction. If the lemma is
false, then let $G$ be a minimum counterexample. If $G$ is
$K_4$-free, then there is nothing to prove. Otherwise, we consider
the clique graph ${\cal C}(G)$, whose edge set is the set of all edges
appearing in some copy of $K_4$. Because of $\Delta(G)=5$, here are
all possible connected component of ${\cal C}(G)$.
\begin{enumerate}
\item $C_{t}\boxtimes K_2$ for $t\geq 4$. If this type occurs, then every vertex
in $C_{t}\boxtimes K_2$ has degree $5$; thus, this is the entire
graph $G$. If $t$ is even, then we can find an independent set $I$
meeting every $K_4$. If $t$ is odd, then it is impossible to find
such an independent set. However, this graph is excluded from the
assumption of the Lemma.
\item $P_t\boxtimes K_2$ for $t\geq 3$. In this case, all internal vertices
have degree $5$ while the four end vertices have degree $4$.
Consider a new graph $G'$ which is obtained by deleting all internal
vertices and adding four edges to make the four end vertices as a
$K_4$. It is easy to check $\Delta(G')\leq 5$ and $\omega(G')\leq
4$. Since $|G'|<|G|$, there is an independent set $I$ of $G'$
meeting every copy of $K_4$ in $G'$. Note that there is exactly one
end vertex in $I$. Observe that any one end vertex can be extended
into a maximal independent set meeting every copy of $K_4$ in
$P_t\boxtimes K_2$. Thus, we can extend $I$ to an independent set
$I'$ of $G$ such that $I'$ meets every copy of $K_4$ in $G$. Hence,
this type of component does not occur in ${\cal C}(G)$.
\item There are four other types listed in Figure \ref{k4}.
\begin{figure}[htbp]
\centerline{ {\psfig{figure=image/k4.eps, width=0.2\textwidth}}
\hfil
\psfig{figure=image/k422.eps, width=0.2\textwidth}\hfil
{\psfig{figure=image/k431.eps, width=0.2\textwidth}}
\hfil \psfig{figure=image/k432.eps, width=0.2\textwidth}}
\caption{All types of components in the clique graph $C(G)$.}
\label{k4}
\end{figure}
For each component $C_i$ in ${\cal C}(G)$, let $V_i$ be the set of common
vertices in all $K_4$'s of $C_i$; for the leftmost figure in Figure
\ref{k4}, $V_i$ is the set of all 4 vertices; for the middle two
figures, $V_i$ is the set of bottom three vertices; for the
rightmost figure, $V_i$ consists of the left-bottom vertex and the
middle-bottom vertex. Note that all $V_i$'s are pairwise disjoint.
Let $G'$ be the induce subgraph of $G$ on $\cup_i V_i$. Note that
$G'$ does not contains any vertex in $C_i\setminus V_i$. By checking
each type, we find out that
for each $i$ and each $v\in V_i$, $v$ has at most $\min\{2,|V_i|-2\}$
neighbors outside $V_i$ in $G'$ (not in $G$!). Applying Theorem
\ref{king} to $G'$, we conclude that there exists an independent set
$I$ of $G'$ meeting every $V_i$; thus $I$ meets every $K_4$ in $G$.
Contradiction!
\end{enumerate}
\hfill $\square$
\begin{lemma} \label{5to42}
Let $G$ be a connected graph with $\Delta(G)\leq 5$ and
$\omega(G)\leq 4$. If $G \not = C_{2l+1} \boxtimes K_2$ for some $l
\geq 2$, then there exists an independent set meeting all induced
copies of $K_4$ and $C_8^2$.
\end{lemma}
\noindent
{\bf Proof:} We will use proof by contradiction. Suppose the Lemma
is false. Let $G$ be a minimum counterexample (with the smallest
number of vertices). For any independent set $I$, let $C(I)$ be the
number of induced copies of $C_8^2$ in $G-I$. Among all independent
sets which meet all copies of $K_4$, there exists an independent set
$I$ such that $C(I)$ is minimized. Since $C(I)>0$, let $H$ be a copy
of $C_8^2$ in $G-I$. The vertices of $H$ are listed by $u_i$ for
$i\in {\mathbb Z}_8$ anticlockwise such that $u_iu_j$ is an edge of
$H$ if and only if $|i-j| \leq 2$. The vertex $v_{i+4}$ is the {\em
antipode} of $v_i$ for any $i\in {\mathbb Z}_8$.
\noindent
{\bf Case 1:} There exists a vertex $v\not\in V(H)$ such that $v$ has
five neighbors in $H$. By the Pigeonhole Principle, $\Gamma(v)$ contains
a pair of antipodes. Without loss of generality, say $u_0, u_4\in
\Gamma(v)$. If the other three neighbors of $v$ do not form a
triangle, then we let $I':=I\setminus \{v\} \cup \{u_0,u_4\}$; note
that $v$ is not in any $K_4$ of $G-I'$. Thus $I'$ is an independent
set meeting every copy of $K_4$. Since every copy of $C_8^2$
containing $v$ must contain one of $u_0$ and $u_4$, we have
$C(I')<C(I)$. Contradiction! Hence, the other three neighbors of $v$
must form a triangle. Without loss of generality, we can assume
that the three neighbors are $u_1$, $u_2$, and $u_3$. Now we let
$I'':=I\setminus \{v\} \cup \{u_0,u_3\}$; note that $v \not \in K_4
\subset G-I'$. Thus $I''$ is also an independent set meeting every
copy of $K_4$ of $G$. Since every copy of $C_8^2$ containing $v$
must contain one of $u_0$ and $u_3$, we have $C(I'')<C(I)$.
Contradiction!
\noindent
{\bf Case 2:} There exists a vertex $v\not\in V(H)$ such that $v$
has exactly four neighbors in $H$. Since $H$ is $K_4$-free, we can
find $u_i, u_j\in \Gamma(v)\cap V(H)$ such that $u_iu_j$ is a
non-edge. Let $I':=I\setminus \{v\} \cup \{u_i,u_j\}$; $I'$ is also
an independent set. Note that $\Gamma(v)\setminus \{u_i, u_j\}$
can not be a triangle, $v$ is not in any $K_4 \subset G-I'$. Thus
$I'$ meets every copy of $K_4$. Since every copy of $C_8^2$
containing $v$ must contain one of $u_i$ and $u_j$, we have
$C(I')<C(I)$. Contradiction!
\noindent
{\bf Case 3:} There exists a vertex $v\not\in V(H)$ such that $v$
has exactly three neighbors in $H$. If the $3$ neighbors do not form
a triangle, then choose $u_i, u_j\in \Gamma(v)\cap V(H)$ such that
$u_iu_j$ is a non-edge. Note that $\Gamma(v)\setminus \{u_i, u_j\}$
can not be a triangle; $v$ is not in any $K_4 \subset G-I'$. Let
$I':=I\setminus \{v\} \cup \{u_i,u_j\}$; $I'$ is also an independent
set meeting every copy of $K_4$. Since every copy of $C_8^2$
containing $v$ must contain one of $u_i$ and $u_j$, we have
$C(I')<C(I)$. Contradiction! Else, the three neighbors form a
triangle; let $u_i$ be one of them and $I':=I\setminus \{v\} \cup
\{u_i\}$; $v$ is not in any $K_4 \subset G-I'$. Thus $I'$ is an
independent set meeting every copy of $K_4$. Note that
$\Gamma(v)\setminus \{u_i\}$ has only two vertices in $H$. The
induced graph on $\Gamma(v)\setminus \{u_i\}$ is disconnected.
However, for any vertex $v$ in $H'=C_8^2$, the subgraph induced by
$\Gamma_G(v) \cap H'$ is a $P_4$. There is no $C_8^2$ in $G-I'$
containing $v$. Thus, $C(I')<C(I)$. Contradiction!
\noindent
{\bf Case 4:} Every vertex outside $H$ can have at most $2$
neighbors in $H$. We identify each pair of antipodes of $H$ to get a
new graph $G'$ from $G$. After identifying, $H$ is turned into a
$K_4$; where the vertices of this $K_4$ are referred as fat
vertices.
{\bf Subcase 4a:} $G'\not=C_{2l+1}\boxtimes K_2$.
Observe $\Delta(G')\leq 5$. We claim $G'$ is $K_5$-free.
Suppose not. Since every vertex in $H$ has at most one neighbor
outside $H$, then each fat vertex can have at most two neighbors
outside $H$. Recall that the original graph $G$ is $K_5$-free. If
$G'$ has some $K_5$, then this $K_5$ contains either $3$ or $4$ fat
vertices. Let $w$ be one of the other vertices in this $K_5$.
We get $w$ has at least three neighbors in
$H$. However, this is covered by Case 1, Case 2, or Case 3. Thus,
$G'$ is $K_5$-free.
Since $|G'|<|G|$, by the minimality of $G$, $G'$ has an independent set $I'$
meeting every copy of $K_4$ and $C_8^2$ in $G'$. There is exactly
one fat vertex in $I'$. Now replacing this fat vertex by its
corresponding pair of antipodal vertices, we get an independent set
$I''$; we assume the pair of antipodal vertices are $u_2$ and $u_6$.
It is easy to check that $I''$ is an independent set of $G$. Next we
claim any $v \in V(H) \setminus \{u_2,u_6\}$ is neither in a $K_4
\subset V(G)-I''$ nor in a $C_8^2 \subset V(G)-I''$. Suppose there is
some $v$ such that $v \in K_4 \subset G-I''$. Recall each $v \in
V(H)$ has at most one neighbor outside $H$ and $H$ is $K_4$-free;
there is some $w \not \in V(H)$ such that $w$ has at least three
neighbors in $H$. This is already considered by Case 1, Case 2, or
Case 3. We are left to show that $v \not \in C_8^2 \subset G-I''$
for each $v \in V(H) \setminus \{u_2,u_6\}$. If not, there exists
a copy $H'$ of $C_8^2$ in $G-I''$ containing $v$. Note $H'$ is $4$-regular,
any vertex in $H'$ can have at most one neighbor in $I''$; in particular,
$v\not=u_0, u_4$. Without loss of generality, we assume $v=u_3$.
Then there is a vertex $w \not \in V(H)$ such
that $u_3w$ is an edge, see Figure \ref{subcase4a}. Observe that the
neighborhood of each vertex of an induced $C_8^2$ is is a $P_4$.
Since $u_1u_4$ and $u_1u_5$ are two non-edges, we have $wu_1$ being
an edge. Observe $\Gamma_G(u_1)=\{u_7,u_0,u_2,u_3,w\}$. Since $u_2
\not \in H'$, we have $u_0 \in H'$; $u_0$ has two neighbors ($u_2$ and $u_6$)
outside $H'$, contradiction!
Therefore, $I''$ meets every copy of $K_4$ and $C_8^2$ in $G$.
Contradiction!
\begin{figure}[htbp]
\centerline{ \psfig{figure=image/2c82.eps, width=0.3\textwidth}}
\caption{Subcase 4a.}
\label{subcase4a}
\end{figure}
{\bf Subcase 4b:} $G'=C_{2l+1}\boxtimes K_2$. The graph $G$ can be
recovered from $G'$. It consists of an induced subgraph $H=C_8^2$
and an induced subgraph $P_{2l-1}\boxtimes K_2$. For each vertex $u$
in $H$, there is exactly one edge connecting it to one of the four
end vertices of $P_{2l-1}\boxtimes K_2$; for each end vertex $v$ of
$P_{2l-1}\boxtimes K_2$, there are exactly two edges connecting $v$
to the vertices in $H$. First, we take any maximum independent set
$I'$ of $P_{2l-1}\boxtimes K_2$. Observe that $I'$ has exactly two
end points of $P_{2l-1} \boxtimes K_2$; so $I'$ has exactly four
neighbors in $H$. In the remaining four vertices of $H$, there
exists a non-edge $u_iu_j$ since $H$ is $K_4$-free. Let
$I:=I'\cup\{u_i,u_j\}$. Clearly $I$ is an independent set of $G$
meeting every copy of $K_4$ and $C_8^2$. Contradiction! \hfill
$\square$
We are ready to prove Lemma \ref{increase}.
\noindent
{\bf Proof of Lemma \ref{increase}:} We need prove
for $k\geq 5$ and any connected graph $G$ with $\Delta(G)=k$
and $\omega(G)\leq k-1$ satisfies
\begin{equation}
\label{eq:rec}
\chi_f(G)\leq k- \min\left\{f(k-1),\frac{1}{2}\right\}.
\end{equation}
If $\omega(G)\leq k-2$,
then by inequality (\ref{eq:1}), we have
$$\chi_f(G)\leq \frac{\Delta(G)+\omega(G)+1}{2}\leq k-\frac{1}{2}.$$
Thus, inequality (\ref{eq:rec}) is satisfied. From now on, we assume
$\omega(G)=\Delta(G)-1$.
For $\Delta(G)=k\geq 6$ and $\omega(G)=k-1$, the condition
$\omega(G)>\frac{2}{3}(\Delta(G)+1)$ is satisfied. By Theorem
\ref{king0}, $G$ contains an independent set meeting every maximum
clique. Extend this independent set to a maximal independent set and
denote it by $I$.
Note that $\Delta(G-I) \leq k-1$ and $\omega(G-I)\leq k-2$.
\noindent {\bf Case 1:} $k\geq 7$. From the definition of $f(k-1)$,
we have $\chi_f(G-I) \leq \Delta(G-I)-f(k-1)$. Thus,
$$
\chi_f(G) \leq \chi_f(G-I)+1 \leq k-1-f(k-1)+1=k-f(k-1).
$$
Thus, we have $f(k) \geq \min\{f(k-1),1/2\}$.
\noindent {\bf Case 2:} $k=6$. By Lemma \ref{6to5}, we can find an
independent set meets every copy of $K_5$ and $C_5\boxtimes K_2$;
we extend this independent set as a maximal independent set $I$.
Note that
$G-I$ contains no induced subgraph isomorphic $C_5\boxtimes K_2$.
We have $\chi_f(G-I) \leq 5-f(5)$; it implies $\chi_f(G)\leq
6-f(5)$.
Thus, $f(6) \geq \min\{f(5),1/2\}$ and we are done.
\noindent
{\bf Case 3:} $k=5$. If
$G=C_{2l+1}\boxtimes K_2$ for some $l\geq 3$; then $G$ is
vertex-transitive and $\alpha(G)=l$. It implies that
$$\chi_f(G)=\frac{|V(G)|}{\alpha(G)}=4+\frac{2}{l}\leq 5-\frac{1}{3}.$$
If $G\not =C_{2l+1}\boxtimes K_2$, then by Lemma \ref{5to42}, we can
find an independent set meeting every copy of $K_4$ and $C_8^2$; we
extend it as a maximal independent set $I$. Note that
$G-I$ contains no induced subgraph isomorphic $C_8^2$.
We have $\chi_f(G-I) \leq 4-f(4)$; it implies $\chi_f(G)\leq 5-f(4)$.
Thus, $f(3) \geq \min\{f(4),1/3\}$ and we are finished.
\hfill $\square$
\section{The case $\Delta(G)=4$}
To prove $f(4)\geq \frac{2}{67}$, we will use an approach which is
similar to those in \cite{hz, lup}. We will construct 133 4-colorable auxiliary graphs, and from these colorings we will construct a 134-fold coloring of $G$ using 532 colors.
It suffices to prove that the
minimum counterexample does not exist.
Let $G$ be a graph with the smallest number of vertices and
satisfying
\begin{enumerate}
\item $\Delta(G)=4$ and $\omega(G)\leq 3$;
\item $\chi_f(G)> 4-\frac{2}{67}$;
\item $G\not=C_8^2$.
\end{enumerate}
By the minimality of $G$, each vertex in $G$ has degree either $4$
or $3$.
To prove Lemma \ref{f4},
we
will show $\chi_f (G) \leq 4 - \frac{2}{67}$, which gives us the
desired contradiction.
For a given vertex $x$ in $V(G)$, it is easy to color its
neighborhood $\Gamma_G(x)$ using $2$ colors. If $d_G(x)=3$, then we
pick a non-edge $S$ from $\Gamma_G(x)$ and color the two vertices in
$S$ using color 1. If $d_G(x)=4$ and $\alpha(\Gamma_G(x))\geq 3$,
then we pick an independent set $S$ in $\Gamma_G(x)$ of size 3 and
assign the color 1 to each vertex in $S$. If $d_G(x)=4$ and
$\alpha(\Gamma_G(x))=2$, then we pick two
disjoint non-edges $S_1$ and $S_2$ from $\Gamma_G(x)$ ; we assign color 1 to each vertex in $S_1$
and color 2 to each vertex in $S_2$.
The following Lemma shows that $G$ has a key property, which
eventually implies that this local coloring scheme works
simultaneously for $x$ in a large subset of $V(G)$.
\begin{lemma}\label{D41}
For each $x \in V(G)$
with $d_G(x)=4$ and $\alpha(\Gamma_G(x))=2$, there exist two
vertex-disjoint non-edges $S_1(x), S_2(x) \subset \Gamma_G(x)$
satisfying the following property. If we contract $S_1(x)$ and
$S_2(x)$, then the resulting graph $G/S_1(x) /S_2(x)$ contains
neither $K^-_{5}$ nor $G_0$. Here $K_5^-$ is the graph obtained
from $K_5$ by removing one edge and $G_0$ is the graph shown in
Figure \ref{fig:H13}.
\begin{figure}[htbp]
\centering
\psfig{figure=image/H122.eps,
width=0.4\textwidth}
\caption{The graph $G_0$.}
\label{fig:H13}
\end{figure}
\end{lemma}
The proof of this lemma is quite long and we will present its proof
in section 4.
For each vertex $x$ in $G$, we associate
a small set of vertices $S(x)$ selected from $\Gamma_G(x)$ as follows. If
$d_G(x)=3$, then let $S(x)$ be the endpoints of a non-edge in
$\Gamma_G(x)$ and label the vertices in $S(x)$ as 1; if $d_G(x)=4$
and $\alpha(\Gamma_G(x))\geq 3$, then let $S(x)$ be any independent
set of size $3$ in $\Gamma_G(x)$ and label all vertices in $S(x)$ as
1; if $d_G(x)=4$ and $\alpha(\Gamma_G(x))=2$, then let
$S(x)=S_1(x)\cup S_2(x)$, where $S_1(x)$ and $S_2(x)$ are guaranteed
by Lemma \ref{D41}; we label the vertices in $S_1(x)$ as 1 and the
vertices in $S_2(x)$ as 2. For any $x\in V(G)$, we have
$|S(x)|=2$, $3$, or $4$.
The following definitions depend on the choice of $S(\ast)$, which
is assumed to be fixed through this section. For $v \in G$ and $j
\in \{1,2,3\}$, we define
$$
N_G^j(v) = \{u| \ \mbox{there is a path} \ vv_0 \ldots v_{j-2}u \
\mbox{in}\ G \ \mbox{of length} \ j \ \mbox{such that} \ v_0
\in S(v) \ \mbox{and} \ v_{j-2} \in S(u)\}.
$$
We now define $N_G^j(u)$ for $j\in \{4,5,7\}$; each $N_G^j(u)$ is a subset of the $j$th neighborhood of $u$. For $j=4$, $v \in N_G^4(u)$ if $d_G(u)=4$,
$\alpha(\Gamma_G(u))=2$, $u$ and $v$ are connected as shown in
Figure \ref{fig:4n}; otherwise $N_G^4(u)=\emptyset$.
In Figure \ref{fig:4n}, $w$ is connected to one of the two vertices in $S_2(u)$.
Similarly, in Figure \ref{fig:5n} and \ref{fig:5n}, a vertex is connected to a group
of vertices means it is connected to any vertex in this group.
For $j=5$, $v \in N_G^5(u)$ if $d_G(w)=4$, $\alpha(\Gamma_G(w))=2$ for
$w \in \{u,v\}$ and $u$ and $v$ are connected as shown in Figure
\ref{fig:5n}; otherwise $N_G^5(u)=\emptyset$.
For $j=7$, $v \in N_G^7(u)$ if $d_G(w)=4$, $\alpha(\Gamma_G(w))=2$ for
$w \in \{u,v\}$ and $u$ and $v$ are connected as shown in Figure
\ref{fig:7n}; otherwise $N_G^7(u)=\emptyset$.
\begin{figure}[htbp]
\centering{
\psfig{figure=image/N4.eps, width=0.25\textwidth} \hfil
\psfig{figure=image/N5.eps, width=0.34\textwidth} \hfil
\psfig{figure=image/N7.eps, width=0.32\textwidth}}
\begin{multicols}{3}
\caption{\! 4-th neighborhood.\!\!\!}
\label{fig:4n}
\newpage
\caption{\! 5-th neighborhood.\!\!\!}
\label{fig:5n}
\newpage
\caption{\! 7-th neighborhood.\!\!\!}
\label{fig:7n}
\end{multicols}
\end{figure}
Note that for $j\in\{1,2,3,5,7\}$, $v\in N_G^j(u)$ if and only if
$u\in N_G^j(v)$; but this does not hold for $j=4$.
We have the following lemma.
\begin{lemma} \label{N5}
For $u \in V(G)$ such that $d_G(u)=4$ and $\alpha(\Gamma_G(u))=2$,
we have $|N_G^1(u) \cup N_G^2(u) \cup N_G^3(u) \cup N_G^4(u) \cup
N_G^5(u) \cup N_G^7(u)| \leq 96$.
\end{lemma}
{\bf Proof:} It is clear that $|N_G^1(u) \cup N_G^2(u) \cup
N_G^3(u)| \leq 4+8+8 \times 3=36$. We next estimate $|N_G^4(u)|$.
In Figure \ref{fig:4n}, observe that $w$ is
connected to one vertex of $S_2(u)$ and $w \not \in \Gamma_G(u)$.
For a fixed $u$, there are at most four choices for $w$, at most
three choices for $z$, and at most three choices for $v$. Therefore,
we have $|N_G^4(u)| \leq 4 \times 3 \times 3=36$.
Let us estimate $|N_G^5(u)|$. In Figure
\ref{fig:5n}, for a fixed $u$, we have four choices for $w$ and
two choices for $z$. Fix a $z$. Assume $\Gamma_G(z) \setminus
\{w\}=\{a,b,c\}$. Let $T_1=\{a,b\}$, $T_2=\{b,c\}$, and
$T_3=\{a,c\}$. We have the following claim.
\noindent {\bf Claim} There are at most three $v \in N_G^5(u)$ such
that for each $v$ we have $ \Gamma_G(z) \cap \Gamma_G(v)=T_i$ for
some $1 \leq i\leq 3$ as shown in Figure \ref{fig:5n}.
\noindent {\bf Proof of the claim:} For each $1 \leq i \leq 3$,
there are at most three $v \in N_G^5(u)$ such that $ \Gamma_G(z)
\cap \Gamma_G(v)=T_i$ as shown in Figure \ref{fig:5n} since each
vertex in $T_i$ has at most three neighbors other than $z$. If
the claim is false, then there is $1 \leq i \not =j \leq 3$ such
that $\Gamma_G(z) \cap \Gamma_G(v_{i})=T_i$ and $\Gamma_G(z) \cap
\Gamma_G(v_{i}')=T_i$ for some $v_{i}, v_{i}' \in N_G^5(u)$, and
$\Gamma_G(z) \cap \Gamma_G(v_j)=T_j$ for some $v_j \in N_G^5(u)$,
where $v_i,v_i', v_j$ are distinct. Without loss of generality, we
assume $\Gamma_G(z) \cap \Gamma_G(v_1)=\Gamma_G(z) \cap
\Gamma_G(v_1')=T_1$ for $v_1, v_1' \in N_G^5(u)$, and $\Gamma_G(z)
\cap \Gamma_G(v_2)=T_2$ for some $v_2 \in N_G^5(u)$, see Figure
\ref{N51}. Observe that $\Gamma_G(b)=\{v_1,v_1',v_2,z\}$. Since
$\Gamma_G(z) \cap \Gamma_G(v_1)=T_1$ as shown in Figure
\ref{fig:5n}, $a$ and one of $b$'s neighbors form $S_i(v_1)$ for
some $i \in \{1,2\}$; we assume it is $S_1(v_1)$. Note
$\{z,v_1,v_1'\} \subset \Gamma_G(a)$. Thus $S_1(v_1)=\{a,v_2\}$ and
$v_2 \in \Gamma_G(v_1)$. Similarly, we can show
$S_1(v_1')=\{a,v_2\}$ and $v_2 \in \Gamma_G(v_1')$. Now, observe
that $\Gamma_G(v_2)=\{v_1,v_1',b,c\}$. Since $\Gamma_G(z) \cap
\Gamma_G(v_2)=T_2$ as shown in Figure \ref{fig:5n}, $b$ and one of
neighbors of $v_2$ form $S_i(v_2)$ for some $i \in \{1,2\}$; we
assume $i=1$. Because $\{v_1,v_1'\} \subset \Gamma_G(b)$, then
$S_1(v_2)=\{b,c\}$. However, $b$ and $c$ are not is in the same
independent set in the definition of $N_G^5(u)$, see Figure
\ref{fig:5n}. This is a contradiction and this case can not
happen. The claim follows.
\begin{figure}[htbp]
\centering{
\psfig{figure=image/N51.eps, width=0.35\textwidth}}
\caption{The picture for the claim.}\label{N51}
\end{figure}
Therefore, $|N_G^5(u)| \leq 4 \times 2 \times 3=24.$
In Figure \ref{fig:7n}, for a fixed $u$, we
have two choices for the edge $e$, one choice for $w$, two choices
for $z$, and three choices for the edge $f$. Fix a $z$. By
considering the degrees of the endpoints of $f$, there is at most
one $f$ and at most one $v \in N_G^7(u)$ such that $|\Gamma_G(f)
\cap \Gamma_G(v)|=4 $ as shown in Figure \ref{fig:7n}. Therefore,
we have $|N_G^7(u)| \leq 2 \times 2 \times 1=4$.
Last, we estimate $|N_G^5(u) \cup N_G^7(u)|$. If there is some $v
\in N_G^7(u)$, then we observe that there are at most five $z$'s
(see Figure \ref{fig:5n}). We get the
number of $v \in N_G^5(u)$ is at most $5 \times 3=15$. In this case,
we have
$$
|N_G^5(u) \cup N_G^7(u)| \leq 4+15 < 24.
$$
If $N_G^7(u)=\emptyset$, then also we have
$$
|N_G^5(u) \cup N_G^7(u) \leq 24.$$
Therefore $$|N_G^1(u) \cup N_G^2(u) \cup N_G^3(u) \cup N_G^4(u)
\cup N_G^5(u) \cup N_G^7(u)| \leq 36+36+24=96.$$ \hfill $\square$
Based on the graph $G$, we define an auxiliary graph $G^{\ast}$ on
vertex set $V(G)$. The edge set is defined as follows: $uv \in E(G^{\ast})$ precisely if
either $u \in N_G^1(v) \cup N_G^2(v) \cup N_G^3(v) \cup N_G^4(v) \cup N_G^5(v)
\cup N_G^7(v) $, or $v \in N_G^4(u)$.
We have the following lemma.
\begin{lemma} \label{121color}
The graph $G^{\ast}$ is 133-colorable.
\end{lemma}
\noindent
{\bf Proof:} Let $\sigma$ be an increasing order of $V(G^{\ast})$
satisfying the following conditions.
\begin{description}
\item [1:] For $u$ and $v$ such that $d_G(u)=3$ and $d_G(v)=4$, we have $\sigma(u) <
\sigma(v)$.
\item [2:] For $u$ and $v$ such that $d_G(u)=d_G(v)=4$, $\alpha(\Gamma_G(u)) \geq
3$, and $\alpha(\Gamma_G(v))=2$, we have $\sigma(u) < \sigma(v)$.
\end{description}
We will color $V(G^{\ast})$ according to the order $\sigma$. For
each $v$, we have the following estimate on the number of colors
forbidden to use for $v$.
\begin{description}
\item [1:] For $v$ such that $d_G(v)=3$, the number of colors
forbidden to use for $v$ is at most $|N_G^1(v) \cup N_G^2(v) \cup
N_G^3(v)| \leq 3+9+27=39$.
\item [2:] For $v$ such that $d_G(v)=4$ and $\alpha(\Gamma_G(v)) \geq
3$, the number of colors forbidden to use for $v$ is at most
$|N_G^1(v) \cup N_G^2(v) \cup N_G^3(v)| \leq 3+9+27=39$.
\item [3:] For $v$ such that $d_G(v)=4$ and
$\alpha(\Gamma_G(v))=2$, the number of colors forbidden to use for
$v$ is at most $|N_G^1(v) \cup N_G^2(v) \cup N_G^3(v) \cup N_G^4(v)
\cup N_G^5(v) \cup N_G^7(v)|+|N_G^4(v)| \leq 96+36=132$ by Lemma
\ref{N5}.
\end{description}
Therefore, the greedy algorithm shows $G^{\ast}$ is 133-colorable.
\hfill $\square$
Let $X$ be a color class of $G^{\ast}$. We define a new graph
$G(X)$ by the following process.
\begin{enumerate}
\item For each $x \in X$, if $|S(x)|=2$ or $|S(x)|=3$, then we contract $S(x)$ as a single vertex,
delete the vertices in $\Gamma_G(v) \setminus S(v)$, and keep
label 1 on the new vertex; if $|S(x)|=4$, i.e., $S(x)=S_1(x)
\cup S_2(x)$, then we contract $S_1(x)$ and $S_2(x)$ as single
vertices and keep their labels. After that, we delete $X$. Let $H$
be the resulting graph.
\item Note that $\Gamma_H(x) \cap \Gamma_H(y) = \emptyset$ and there is no edge from
$\Gamma_H(x)$ to $\Gamma_H(y)$ for any $x,y \in X$ as $X$ is a color class.
\item We identify all vertices with label $i$ as a single vertex $w_i$ for $i \in \{1,2\}$. Let $G(X)$ be the resulted graph.
\end{enumerate}
We have the following lemma on the chromatic number of $G(X)$.
\begin{lemma} \label{4colorable}
The graph $G(X)$ is $4$-colorable for each color class.
\end{lemma}
We postpone the proof of this lemma until the end of this section
and prove Lemma \ref{f4} first.
\noindent
{\bf Proof of Lemma \ref{f4}:} By Lemma \ref{121color}, there is a
proper 133-coloring of $G^{\ast}$. We assume
$V(G^{\ast})=V(G)=\cup_{i=1}^{133} X_i $, where $X_i$ is the $i$-th
color class.
For each $i \in \{1,\ldots, 133\}$, Lemma \ref{4colorable} shows
$G(X_i)$ is $4$-colorable; let $c_i\colon V(G(X_i)) \to T_i$ be a
proper $4$-coloring of the graph $G(X_i)$. Here $T_1,T_2,\ldots,
T_{133}$ are pairwise disjoint; each of them consists of $4$ colors.
For $i \in \{1,\ldots, 133\}$, the $4$-coloring $c_i$ can be viewed
as a $4$-coloring of $G \setminus X_i$ since each vertex with label
$j$ receives the color $c_i(w_i)$ for $j=1,2$ and each removed
vertex has at most three neighbors in $G \setminus X_i$.
Now we reuse the notation $c_i$ to denote this $4$-coloring of $G
\setminus X_i$. For each $v \in X_i$, we have $|\cup_{u \in
\Gamma_G(v)}c_i(u)| \leq 2$. We can assign two unused colors,
denoted by the set $Y(v)$, to $v$. We define $f_i: V(G) \rightarrow
\mathcal P(T_i)$ (the power set of $T_i$) satisfying
$$
f_i(v)= \left\{
\begin{array}{ll}
\{c_i(v)\} \ & \textrm{if} \ v \in V(G) \setminus X_i, \\
Y(v) & \textrm{if} \ v \in X_i.\\
\end{array}
\right.
$$
Observe that each vertex in $X_i$ receives two colors from $f_i$ and
every other vertex receives one color. Let $\sigma: V(G)
\rightarrow \mathcal P(\cup_{i=1}^{k} T_i)$ be
a mapping such that $\sigma(v)=\cup_{i=1}^{m} f_i(v)$. It is easy
to verify
$\sigma$ is a $134$-fold coloring of $G$ such that each color is drawn
from a palette of $532$ colors; namely we have
$$\chi_f(G) \leq
\frac{532}{134} =4-\frac{2}{67}.$$
The proof of Lemma \ref{f4} is finished. \hfill $\square$
Before we prove Lemma \ref{4colorable}, we need the following definitions.
A {\it block} of a graph is a maximal $2$-connected induced subgraph.
A {\it Gallai tree} is a connected graph in which all blocks are
either complete graphs or odd cycles. A {\it Gallai forest} is a
graph all of whose components are Gallai trees. A $k$-{\it Gallai
tree (forest)} is a Gallai tree (forest) such that the degree of all
vertices are at most $k-1$. A $k$-critical graph is a graph $G$
whose chromatic number is $k$ and deleting any vertex can decrease
the chromatic number. Gallai showed the following Lemma.
\begin{lemma}\label{gallai} \cite{gallai}
If $G$ is a $k$-critical graph, then the subgraph of $G$ induced on
the vertices of degree $k-1$ is a k-Gallai forest.
\end{lemma}
\noindent
{\bf Proof of Lemma \ref{4colorable}:} We use proof by
contradiction. Suppose that $G(X)$ is not $4$-colorable.
The only possible vertices in $G(X)$ with degree greater than $4$ are the
vertices $w_1$ and $w_2$, which are
obtained by contracting the vertices with label 1 and 2 in the intermediate graph $H$.
The simple greedy algorithm shows that $G(X)$ is always
$5$-colorable. Let
$G'(X)$ be a $5$-critical subgraph of $G(X)$. Applying
Lemma \ref{gallai} to $G'(X)$, the subgraph of $G'(X)$ induced on the
vertices of degree $4$ is a $5$-Gallai forest $F$. The vertex set of $F$
may contain $w_1$ or $w_2$. Delete $w_1$ and $w_1$ from $F$ if $F$
contains one of them. Let $F'$ be the resulting Gallai forest. (Any
induced subgraph of a Gallai forest is still a Gallai forest.) The Gallai forest $F'$ is not empty. Let $T$ be a
connected component of $F'$ and $B$ be a leaf block of $T$. The
block $B$ is either a clique or an odd cycle from the definition of a
Gallai tree.
Let $v$ be a vertex in $B$. As $v$ has at most two neighbors ($w_1$
and $w_2$) outside $F'$ in $G(X)$, we have $d_{F'}(v)\geq 2$. If
$v$ is not in other blocks of $F'$, then we have $d_B(v)\geq 2$.
It follows that $|B|\geq 3$. Since $B$ is a subgraph of $G$ and
$G$ is $K_4$-free, the block $B$ is an
odd cycle.
Let $v_1v_2$ be an edge in $B$ such that $v_1$ and $v_2$ are not in
other blocks. The degree requirement implies $v_iw_j$ are edges in
$G(X)$ for all $i, j\in\{1,2\}$. For $i=1,2$, there are vertices
$x_i, y_i\in X$ satisfying $S(x_i)\cap \Gamma_G(v_i)\not =\emptyset$
and $S(y_i)\cap \Gamma_G(v_i)\not =\emptyset$; moreover either
$|S(x_i)| =4$ or $|S(y_i)|=4$ since one of its neighborhood has
label 2. Without loss of generality, we assume $|S(x_i)|=4$ for $i
\in \{1,2\}$. If $x_i\not= y_i$, then $y_i \in N_G^4(x_i)$, i.e.,
$y_i \in \Gamma_{G^{\ast}}(x_i)$; this contradicts $X$ being a
color class. Thus we have $x_i=y_i$
and $|S(x_i)|=4$ for $i \in \{1,2\}$. For
$\{i,j\}=\{1,2\}$, if $x_i\not= y_j$, then $y_i \in N_G^5(x_i)$,
i.e., $y_i \in \Gamma_{G^{\ast}}(x_i)$; this is a contradiction of $X$ being a color class.
Thus we
have
$$x_1=x_2 = y_1=y_2.$$
Let $x$ denote this common vertex above. Then
$d_G(x)=4$ and $\alpha(\Gamma_G(x))=2$.
Let $v_0$ be the only vertex in $B$ shared by other blocks. Since
$B-v_0$ is connected, the argument above shows there is a common $x$
for all edges in $B-v_0$. If $\Gamma_{G(X)}(v_0)\cap
\{w_1,w_2\}\not=\emptyset$, the there is some vertex $x_0 \in X$
such that $S(x_0) \cup \Gamma_G(v_0) \not =\emptyset$. By the
similar argument, we also have $x_0=x$.
Therefore, $x$ depends only on $B$. In the sense that for any $y\in X$ and any $v\in B$, if $S(y)\cap \Gamma_G(v)\not=\emptyset$,
then $y=x$.
The block $B$ is an odd cycle as we mentioned above. Suppose $|B|=2r+1$. Let $v_0,v_1,\ldots,v_{2r}$ be the vertices
of $B$ in cyclic order and $v_0$ be the only vertex which may be shared by other block.
Let $x\in X$ be the vertex determined by $B$. Recall
$d_G(x)=4$ and $\alpha(\Gamma_G(x))=2$.
Each vertex in $\Gamma(x)$ can have at most $2$ edges to $B$. We get
\begin{equation}
\label{eq:12}
4r\leq |E(B,\Gamma(x))|\leq 8.
\end{equation}
We have $r\leq 2$. The block $B$ is either a $C_5$ or a $K_3$.
We claim both $v_0w_1$ and $v_0w_2$ are non-edges of
$G(X)$.
If $B=C_5$, then inequality (\ref{eq:12}) implies
that $v_0$ has no neighbor in
$\Gamma(x)$ and the claim holds.
If $B=K_3$, then the claim also holds; otherwise $B\cup
\{\underline{S_1(x)}, \underline{S_2(x)}\}$ forms a $K_5^-$ in
$G/S_1(x)/S_2(x)$, which is a contradiction to Lemma \ref{D41}.
Let $u_1$ and $u_2$ be the two neighbors
of $v_0$ in other blocks of $F'$. If $u_1$ and $u_2$ are in the same block,
then this block is an odd cycle; otherwise, $v_0u_1$ and $v_0u_2$ are in two different blocks.
The union of non-leaf blocks of $T$ is a Gallai-tree, denoted by
$T'$. The argument above shows every leaf block of $T'$ must be an
odd cycle. Let $C$ be such a leaf block of $T'$. Now $C$ is an odd
cycle, and $C$ is connected to $|C|-1$ leaf blocks of $T$. Let $B$
and $B'$ be two leaf blocks of $T$ such that $B\cap C$ is adjacent
to $B'\cap C$. Without loss of generality, we may assume $B$ is the
one we considered before. By the same argument, $B'$ is an odd
cycle of size $2r'+1$ with $r'\in \{1,2\}$. Let $v'_0,v'_1,\ldots,
v'_{2r'}$ be the vertices of $B'$ and $v'_0$ be the only vertex in
$B'\cap C$. For $i$ in $\{1,2,\ldots, 2r'\}$ and $j$ in $\{1,2\}$,
$v'_iw_j$ are edges in $G'(X)$. Similarly, there exists a vertex
$x'\in X$ with $d_G(x')=4$ and $\alpha(\Gamma_G(x'))=2$ such that
$|E(v_i, S_1(x'))|\geq 1$ and $|E(v_i, S_2(x'))|\geq 1$. We must
have $x=x'$; otherwise $x'\in N_G^7(x)$, i.e., $x' \in
\Gamma_{G^{\ast}}(x)$, and this contradicts the fact that $X$ is a
color class in $D$. Now we have $|E(\Gamma(x), B)|\geq 4r$ and
$E(\Gamma(x), B')|\geq 4r'$. By counting the degrees of vertices in
$\Gamma(x)$ in $G$, we have
$$4r+4r'+4+4\leq 16.$$
We get $r=r'=1$. Both $B$ and $B'$ are $K_3$'s. In this case,
$G/S_1(x)/S_2(x)$ contains the graph $G_0$, see figure
\ref{fig:H13}. This contradicts Lemma \ref{D41}.
We can find the desired contradiction, so the lemma
follows. \hfill $\square$
\section{Proof of Lemma \ref{D41}}
In this section, we will prove Lemma \ref{D41}. We first review a Lemma from \cite{lup}.
\begin{lemma}\label{cut2}
Let $G$ be a graph. Suppose that $G_1$ and $G_2$ are two subgraphs
such that $G_1 \cup G_2=G$ and $V(G_1) \cap V(G_2) =\{u,v\}$.
\begin{enumerate}
\item If $uv$ is an edge of $G$, then we have
$$\chi_f(G) = \max\{\chi_f(G_1), \chi_f(G_2)\}.$$
\item If $uv$ is not an edge of $G$, then we have
$$\chi_f(G) \leq \max\{\chi_f(G_1), \chi_f(G_2+uv), \chi_f(G_2/uv)\},$$
\end{enumerate}
where $G_2+uv$ is the graph obtained from $G_2$ by adding edge $uv$ and $G_2/uv$ is the graph obtained from $G_2$ by contracting $\{u, v\}$.
\end{lemma}
\noindent {\bf Proof of Lemma \ref{D41}:} Recall that $G$ is a
connected $K_4$-free graph with minimum number of vertices such that
$G \not = C_8^2$ and $ \chi_f(G) > 4-\frac{2}{67} $. Note that $G$
is 2-connected. We will prove it by contradiction.
Suppose Lemma \ref{D41} fails for some vertex $x$ in $G$. Observe
$\Gamma_G(x)$ is one of the graphs in Figure \ref{Delta=4}. Here we
assume $\Gamma_G(x)=\{a,b,c,d\}$. Through the proof of the lemma, let
$S_1$ and $S_2$ be two vertex-disjoint independent sets in
$\Gamma_G(x)$, $H$ be a triangle in $V(G) \setminus ( \{x\} \cup
\Gamma_G(x))$, then say $(S_1,S_2,H)$ is a bad triple if
$\{\underline {S_1}, \underline {S_2}, H\}$ contains a $K_5^-$ in
$G/S_1/S_2$.
\begin{figure}[htbp]
\centerline{\psfig{figure=image/c42.eps,width=0.18\textwidth} \hfil
\psfig{figure=image/c41.eps, width=0.18\textwidth} \hfil \psfig{figure=image/c43.eps, width=0.18\textwidth}
}
\centerline{$\Gamma_G(x)=P_4$ \hfil $\Gamma_G(x)=2~e$ \hfil $\Gamma_G(x)=C_4$}
\caption{Three possible cases of $\Gamma_G(x)$. } \label{Delta=4}
\end{figure}
If $\Gamma_G(x)=P_4$, then $\{a,d\}$ and $\{b,c\}$ is the only pair
of disjoint non-edges. There is a triangle $H$ with
$V(H)=\{y,z,w\}$ such that $(\{a,d\}, \{b,c\}, H)$ is a bad triple.
Note that $|E(\{a,b,c,d\},\{y,z,w\})|=5$ or $6$.
By an exhaustive search, the induced subgraph of $G$ on $\{x, a,b,c,d,y,z,w\}$
is one of the following six graphs (see Figure \ref{fig:P4}).
\begin{figure}[htbp]
\centerline{\psfig{figure=image/N25.eps, width=0.25\textwidth} \hfil
\psfig{figure=image/H2.eps, width=0.25\textwidth} \hfil
\psfig{figure=image/N22.eps, width=0.25\textwidth} }
\centerline{ \hspace*{0.05\textwidth} $H_{1}$ \hspace*{0.15\textwidth} \hfil $H_{2}$ \hspace*{0.15\textwidth} \hfil $H_{3}$ }
\vspace*{5mm}
\centerline{\psfig{figure=image/N23.eps,width=0.25\textwidth} \hfil
\psfig{figure=image/N24.eps, width=0.25\textwidth}
\hfil
\psfig{figure=image/N21.eps,
width=0.25\textwidth}}
\centerline{ \hspace*{0.05\textwidth}$H_{4}$ \hfil \hspace*{0.15\textwidth} $H_{5}$ \hfil \hspace*{0.15\textwidth} $H_{6}$ }
\caption{If $\Gamma_G(x)=P_4$, then there are six possible induced subgraphs.}
\label{fig:P4}
\end{figure}
If $\Gamma_G(x)=2~e$, then $(\{a,c\}, \{b,d\})$ and $(\{a,d\},
\{b,c\})$ are two pairs of disjoint non-edges. By considering the
degrees of vertices in $\Gamma_G(x)$, there is only one triangle
$H$ with $V(H)=\{y,z,w\}$ such that $(\{a,c\}, \{b,d\}, H)$ and
$(\{a,d\}, \{b,c\}, H)$ are two bad triples. By an exhaustive
search, the induced subgraph of $G$ on $\{x, a,b,c,d,y,z,w\}$ is one
of the following three graphs (see Figure \ref{fig:C4}).
\begin{figure}[htbp]
\centerline{ \psfig{figure=image/G1.eps, width=0.25\textwidth} \hfil \hfil \psfig{figure=image/G2.eps,
width=0.25\textwidth} \hfil \psfig{figure=image/G3.eps,width=0.25\textwidth} }
\centerline{ \hspace*{0.05\textwidth} $H_7$ \hspace*{0.15\textwidth} \hfil $H_8$ \hspace*{0.15\textwidth} \hfil $H_9$ }
\caption{If $\Gamma_G(x)=2\ e$, then there are three possible induced subgraphs.}
\label{fig:C4}
\end{figure}
It suffices to show that $G$ cannot contain $H_i$ for $1\leq i\leq
9$. Since all vertices in $H_1$ (and $H_2$) have degree 4, $H_1$
(and $H_2$) is the entire graph $G$. Observe that $H_1$ is
isomorphic to $C_8^2$ and $H_2$ is 11:3-colorable (see Figure
\ref{fig:H2H7}). Contradiction!
\begin{figure}[htbp]
\centering
\psfig{figure=image/N2224.eps, width=0.24\textwidth} \hfil
\psfig{figure=image/G11.eps, width=0.24\textwidth}
\centerline{$H_2$ \hspace*{1cm} \hfil $H_7$ }
\caption{The graph $H_2$ and $H_7$ are 11:3-colorable.}
\label{fig:H2H7}
\end{figure}
In $H_7$, the vertex $d$ is the only vertex with degree less than
$4$. If $H_7$ is not the entire graph $G$, then $d$ is a cut vertex
of $G$. This contradicts the fact that $G$ is $2$-connected. Thus
$G=H_7$. The graph $H_7$ is 11:3-colorable as shown by Figure
\ref{fig:H2H7}. Contradiction!
Now we consider the case $H_3$. Note $H_3+bz$ is the graph $H_2$. We
have $\chi_f(H_3)\leq \chi_f(H_2)\leq 11/3$. The graph $H_3$ must be
a proper induced subgraph of $G$, and the pair $\{b,z\}$ is a vertex
cut of $G$. Let $G'$ be the induced subgraph of $G$ by deleting all
vertices in $H_3$ but $b,z$. We apply Lemma \ref{cut2} to $G+bz$
with $G_1=H_3+bz=H_2$ and $G_2=G'+bz$. We have
$$\chi_f(G+bz)\leq \max\{\chi_f(H_2), \chi_f(G'+bz)\}.$$
Note $\chi_f(H_2)\leq 11/3$ and $11/3<\chi_f(G)\leq \chi_f(G+bz)$.
We have $\chi_f(G)\leq \chi_f(G'+bz)$. Both $b$ and $z$ have at most $2$ neighbors
in $G'+bz$. Thus $G'+bz$ is $K_4$-free; $G'+bz\not=C_8^2$
and has fewer vertices than $G$.
This contradicts to the minimality of $G$.
Note $H_5+cy=H_2$. The case $H_5$ is similar to the case $H_3$.
Note that $H_4$, $H_6$, and $H_8$ are isomorphic to each other.
It suffices to show $G$ does not contain $H_4$.
Suppose that $H_4$ is a proper induced subgraph of $G$.
Let $G_1$ be the induced subgraph of $G$ by deleting
all vertices in $H_4$. Note $C_8^2$ is not a proper subgraph of any graph in ${\cal G}_4$.
We have $G_1\not=C_8^2$.
Note that $c$ and $z$ have degree $3$ while other vertices in $H_4$ have degree $4$.
Since $G$ is 2-connected, $c$ has a unique neighbor, denoted by $u$,
in $V(G_1)$. Similarly, $z$ has a unique neighbor, denoted by $v$,
in $V(G_1)$. Observe that the pair $\{u,v\}$ forms a vertex cut of $G$.
Let $G_2$ be the induced graph of $G$ on $V(H_4)\cup \{u,v\}$.
Applying Lemma \ref{cut2} to $G$ with $G_1$ and $G_2$, we have
$$\chi_f(G) \leq \max\{\chi_f(G_1), \chi_f(G_2+uv), \chi_f(G_2/uv)\}.$$
Figure \ref{fig:H4} shows $\chi_f(G_2+uv)$ and $\chi_f(G_2/uv)$ are at most $11/3$.
\begin{figure}[htbp]
\centering
\psfig{figure=image/N23+uv.eps, width=0.24\textwidth} \hfil
\psfig{figure=image/N23-uv.eps, width=0.24\textwidth}
\centerline{$G_2+uv$ \hspace*{1cm} \hfil $G_2/uv$ }
\caption{Case $H_4$: both graph $G_2+uv$ and $G_2/uv$ are 11:3-colorable.}
\label{fig:H4}
\end{figure}
Since $\chi_f(G)> 11/3$, we have $\chi_f(G)\leq \chi_f(G_1)$.
Now $G_1$ is $K_4$-free and has maximum degree at most $4$; $G_1$ has
fewer vertices than $G$. This contradicts the minimality of $G$.
If $G=H_4$, then $\chi_f(H_4)\leq 11/3$, since $H_4$ is a subgraph of $G_2+uv$
in Figure \ref{fig:H4}.
Now we consider the last case $H_9$. First, we contract $b,c,z$ into a
fat vertex denoted by $\underline{bcz}$. We write $G/bcz$ for the graph after
this contraction. Observe that $\{\underline{bcz},d\}$ is a vertex-cut of
$G/bcz$. Let $G_4$ and $G_4'$ be two connected subgraphs of $G/bcz$
such that $G_4 \cup G_4'=G/bcz$, $G_4 \cap G_4'=\{\underline{bcz},d\}$, and
$\{u,v\} \subset G_4'$. Note that $G_4$ is $11$:$3$ colorable, see
Figure \ref{fig:H9}. Now by Lemma \ref{cut2}, we have
$$
\chi_f(G/bcz) \leq \max \{\chi_f(G_4), \chi_f(G_4')\}.
$$
As $\{b,c,z\}$ is an independent set, each $a$:$b$-coloring of
$G/bcz$ gives an $a$:$b$-coloring of $G$, that is $\chi_f(G/bcz)\geq
\chi_f(G)>11/3$. The graph $G_4$ is 11:3-colorable; see Figure
\ref{fig:H9}. Thus we have $\chi_f(G_4')\geq \chi_f(G/bcz)\geq
\chi_f(G)$. It is easy to check that $G_4'$ has maximum degree $4$,
$K_4$-free, and it is not $C_8^2$. Hence $G_4'$ must contain a
$K_4$. Otherwise, it contradicts the minimality of
$G$.
Second, we contract $\{b,d,z\}$ into a fat vertex $\underline{bdz}$
and denote the graph by $G/bcz$. Let $G_5$ and $G_5'$ be two
connected subgraphs of $G/bdz$ such that $G_5 \cup G_5'=G/bzd$, $G_5
\cap G_5'=\{\underline{bzd},c\}$, and $\{u,v\} \subset G_5'$. Note
that $G_5$ is 11:3-colorable; see Figure \ref{fig:H9}. By a similar
argument, $G_5'$ must contain a $K_4$.
\begin{figure}[htbp]
\centerline{ \psfig{figure=image/G31bcz.eps, width=0.33\textwidth} \hfill \psfig{figure=image/G31bdz.eps,
width=0.33\textwidth} \hfill \psfig{figure=image/G5.eps,width=0.33\textwidth}}
\centerline{$G_4$ \hspace{3.4cm} $G_5$ \hspace{3.4cm} $G_6$}
\caption{Case $H_9$: the graphs $G_4$, $G_5$, and $G_6$ are 11:3-colorable.}
\label{fig:H9}
\end{figure}
The remaining case is that both $G_4'$ and $G_5'$ have a $K_4$ when we contract $b$ and $z$.
Since the original graph $G$ is $K_4$-free, the $K_4$ in $G_4'$ (and in
$G_5'$) must contain the fat vertex $\underline{bcz}$ (or
$\underline{bdz}$), respectively. Note that each of the four
vertices $b,c,d,z$ has at most one edge leaving $H_9$. There must be
a triangle $uvp$ in $G$ and these four outward edges are connected
to some element of $\{u,v,p\}$. The graph $G/{bz}$ must contain the
subgraph $G_6$ as drawn in Figure \ref{fig:H9}.
Note that $\{u,v\}$ is a vertex-cut in $G/ bz$. Let $G_6$ and $G_6'$
be two connected subgraphs of $G/bz,$ which satisfy $G_6 \cup
G_6'=G$, $G_6 \cap G_6'=\{u,v\}$, and $\underline{bz} \in G_6$. By
Lemma \ref{cut2}, we have
$$
\chi_f(G/bz) \leq \max\{\chi_f(G_6), \chi_f(G_6')\}.
$$
Note that $G_6$ is 11:3-colorable; see Figure \ref{fig:H9}. We also
have $\chi_f(G/bz) \geq \chi_f(G) >\frac{11}{3}$. We obtain
$\chi_f(G_6') \geq \chi_f(G/bz) \geq \chi_f(G)$. Observe that $G_6'$
is a subgraph of $G$. We arrive at a contradiction of the minimality
of $G$.
If $\Gamma_G(x)=C_4$, then the only possible choice for the two
independent sets are $\{a,c\}$ and $\{b,d\}$. If there is some
triangle $H$ such that $(\{a,c\},\{b,d\},H)$ is a bad triple, then
we have
$$
|E(\Gamma_G(x),H)| \geq 5.
$$
However, $|E(\Gamma_G(x),H)| \leq 4$. This is a contradiction. Thus
the lemma follows in this case.
We can select two
vertex disjoint non-edges $S_1$ and $S_2$ such that the graph
$G/S_1/S_2$ contains no $K_5^-$. For these particular $S_1$ and
$S_2$, if $G/S_1/S_2$ contains no $G_0$, then Lemma \ref{D41} holds.
Without loss of generality, we assume that $G/S_1/S_2$ does contain
$G_0$. Let $s_i=\underline{S_i}$ for $i=1,2$.
Observe that both $s_1$ and $s_2$ have
four neighbors $u,v,p,q$ other than $x$ in $G_0$. It follows
that $$|E(S_1\cup S_2, \{u,v,p,q\})|\geq 8.$$
On the one hand, we have
\begin{eqnarray*}
|E(G\mid_{S_1\cup S_2})| &=&\frac{1}{2}\left(\sum_{v\in S_1\cup S_2}d(v)
-|E(S_1\cup S_2, \{u,v,p,q\})| - 4\right)\\
&\leq& \frac{1}{2}(16- 8 - 4)\\
&=&2.
\end{eqnarray*}
On the other hand, $\alpha(\Gamma(x))=2$ implies $G\mid_{S_1\cup
S_2}$ contains at least two edges. Thus, we have $\Gamma_G(x)=2 ~
e$. Label the vertices in $\Gamma_G(x)$ by $a,b,c,d$ as in Figure
\ref{Delta=4}. We assume $ab$ and $cd$ are edges while $ac, bd, ad,
bc$ are non-edges. Observe that each vertex in $\{u,v,p,q\}$ has
exactly two neighbors in $\{a,b,c,d\}$.
If one vertex, say $u$, has two neighbors forming a non-edge, say $ac$, then
we can choose $S'_1=\{a,c\}$ and $S_2'=\{b,d\}$. It is easy to check that
$G/S_1'/S_2'$ contains neither $G_0$ nor $K_5^-$. We are done in this case.
In the remaining case, we can assume that for each vertex $y$ in
$\{u,v,p,q\}$, the neighbors of $y$ in $\{a,b,c,d\}$ always form an
edge. Up to relabeling vertices, there is only one arrangement for
edges between $\{u,v,p,q\}$ and $\{a,b,c,d\}$; see the graph
$H_{10}$ defined in Figure \ref{fig:H14}.
\begin{figure}[htbp]
\centerline{\psfig{figure=image/H14.eps, width=0.28\textwidth} \hfil
\psfig{figure=image/H141.eps, width=0.32\textwidth}}
\caption{$H_{10}$ and an 11:3-coloring of $H_{10}$.} \label{fig:H14}
\end{figure}
The graph $H_{10}$ is 11:3-colorable as shown in Figure
\ref{fig:H14}. Since $\chi_f(G)>11/3$, $H_{10}$ is a proper subgraph
of $G$.
Note in $H_{10}$, every vertices except $w$ and $r$
has degree $4$; both $w$ and $r$ have degree $3$. Thus, $\{w,r\}$ is
a vertex cut of $G$. Let $G_1=H_{10}$ and $G_2$ be the subgraph of
$G$ by deleting vertices in $\{x,a,b,c,d,p,q,u,v\}$. Applying Lemma
\ref{cut2} with $G_1$ and $G_2$ defined above, we have
$$
\chi_f(G) \leq \max \{\chi_f(G_1),\chi_f(G_2)\}.
$$
Since $\chi_f(G)> 11/3$ and $\chi_f(G_1) \leq 11/3 $ (see Figure
\ref{fig:H14}), we must have $\chi_f(G_2) \geq \chi_f(G)$. Note
that $G_2$ has fewer number of vertices than $G$. This contradicts
the minimality of $G$. Therefore, the lemma follows. \hfill
$\square$
\section{Introduction}
The chromatic number of graphs with bounded degrees has been studied
for many years. Brooks' theorem perhaps is one of the most
fundamental results; it is included by many textbooks on
graph theory. Given a simple connected graph $G$, let $\Delta(G)$ be
the maximum degree, $\omega(G)$ be the clique number,
and $\chi(G)$ be the chromatic number. Brooks'
theorem states that $\chi(G)\leq \Delta(G)$ unless $G$ is a complete
graph or an odd cycle.
Reed \cite{reed1} proved that $\chi(G)\leq
\Delta(G)-1$ if $\omega(G)\leq \Delta(G)-1$ and $\Delta(G)\geq \Delta_0$
for some large constant $\Delta_0$. This excellent result was proved by
probabilistic methods, and $\Delta_0$ is at least
hundreds. Before this result, Borodin and Kostochka \cite{bk} made
the following conjecture.
{\bf Conjecture \cite{bk}:}
Suppose that $G$ is a connected graph.
If $\omega(G)\leq \Delta(G)-1$
and $\Delta(G)\geq 9$, then we have
$$\chi(G)\leq \Delta(G)-1.$$
If the conjecture is true, then it is best possible since there is a
$K_8$-free graph $G=C_5\boxtimes K_3$ (actually $K_7$-free, see Figure \ref{fig:1}) with
$\Delta(G)=8$ and $\chi(G)=8$.
\begin{figure}[htbp]
\label{fig:1}
\centerline{\psfig{figure=image/c53.eps, width=0.4\textwidth} }
\caption{The graph $C_5 \boxtimes K_3$.}
\end{figure}
Here we use the following notation of the strong product. Given two graphs $G$
and $H$, the {\it strong product} $G\boxtimes H$ is the graph with vertex
set $V(G)\times V(H)$, and $(a,x)$ is connected to $(b,y)$ if one of the following holds
\begin{itemize}
\item $a=b$ and $xy\in E(H)$,
\item $ab\in E(G)$ and $x=y$,
\item $ab\in E(G)$ and $xy\in E(H)$.
\end{itemize}
Reed's result \cite{reed1} settled Borodin and Kostochka's conjecture for sufficiently large $\Delta(G)$,
but the cases with small $\Delta(G)$ are hard to cover using the probabilistic method.
In this paper we consider a fractional analogue of this problem.
The fractional chromatic number $\chi_f(G)$ can be defined as
follows. A $b$-{\it fold coloring} of $G$ assigns a set of $b$
colors to each vertex such that any two adjacent vertices receive
disjoint sets of colors. We say a graph $G$ is $a$:$b$-{\it
colorable} if there is a $b$-fold coloring of $G$ in which each
color is drawn from a palette of $a$ colors. We refer to such a
coloring as an $a$:$b$-coloring. The $b$-{\it fold coloring number},
denoted by $\chi_b(G)$, is the smallest integer $a$ such that $G$
has an $a$:$b$-coloring. Note that $\chi_1(G)=\chi(G)$. It was shown
that $\chi_{a+b}(G) \leq \chi_a(G) + \chi_b(G) $. The {\it
fractional chromatic number} $\chi_f(G)$ is
$ \underset{b \rightarrow \infty} \lim \frac{\chi_b(G)}{b}.$
By the definition, we have $\chi_f(G)\leq \chi(G)$.
The fractional chromatic number can be viewed as a relaxation of the
chromatic number. Many problems involving the chromatic number can be asked
again using the fractional chromatic number. The fractional analogue often
has a simpler solution than the original
problem. For example, the famous $\omega-\Delta-\chi$ conjecture of
Reed \cite{reed} states that for any simple graph $G$, we have
$$\chi(G)\leq \left\lceil \frac{\omega(G)+\Delta(G)+1}{2}\right
\rceil.$$
The fractional analogue of $\omega-\Delta-\chi$ conjecture was
proved by Molloy and Reed \cite{mr}; they actually proved a stronger
result with ceiling removed, i.e.,
\begin{equation} \label{eq:1}
\chi_f(G)\leq \frac{\omega(G)+\Delta(G)+1}{2}.
\end{equation}
In this paper, we classify all connected graphs $G$ with
$\chi_f(G)\geq \Delta(G)$.
\begin{theorem}\label{main}
A connected graph $G$ satisfies $\chi_f(G)\geq \Delta(G)$ if and only if
$G$ is one of the following
\begin{enumerate}
\item a complete graph,
\item an odd cycle,
\item a graph with $\omega(G)=\Delta(G)$,
\item $C^2_8$,
\item $C_5\boxtimes K_2$.
\end{enumerate}
\end{theorem}
For the complete graph $K_n$, we have $\chi_f(K_n)=n$ and
$\Delta(K_n)=n-1$. For the odd cycle $C_{2k+1}$, we have
$\chi_f(C_{2k+1})=2+\frac{1}{k}$ and $\Delta(C_{2k+1})=2$. If $G$ is
neither a complete graph nor an odd cycle but contains a clique of
size $\Delta(G)$, then we have
\begin{equation} \label{less}
\Delta(G)\leq \omega(G)\leq \chi_f(G)\leq \chi(G)\leq \Delta(G).
\end{equation}
The last inequality is from Brooks' theorem. The sequence of
inequalities above implies $\chi_f(G)=\Delta(G)$.
If $G$ is a vertex-transitive graph,
then we have \cite{su}
$$\chi_f(G)=\frac{|V(G)|}{\alpha(G)},$$
where $\alpha(G)$ is the independence number of $G$.
Note that both graphs $C^2_8$ and $C_5\boxtimes K_2$
are vertex-transitive and have the independence number $2$.
Thus we have
$$\chi_f(C^2_8)=4=\Delta(C^2_8)\quad \mbox { and }\quad
\chi_f(C_5\boxtimes K_2)=5= \Delta(C_5\boxtimes K_2).$$
\begin{figure}[htbp]
\centerline{ \psfig{figure=image/c8square.eps, width=0.24\textwidth} \hspace{4cm} \psfig{figure=image/c52.eps,
width=0.24\textwidth} }
\centerline{$C_8^2$ \hspace{6cm} $C_5 \boxtimes K_2$}
\caption{The graph $C_8^2$ and $C_5 \boxtimes K_2$ .}
\end{figure}
Actually, Theorem \ref{main} is a corollary of the following stronger
result.
\begin{theorem}\label{submain}
Assume that a connected graph $G$ is neither $C^2_8$ nor
$C_5\boxtimes K_2$. If $\Delta(G)\geq 4$ and $\omega(G)\leq
\Delta(G)-1$, then we have
$$\chi_f(G)\leq \Delta(G) -\frac{2}{67}.$$
\end{theorem}
\noindent {\bf Remark:} In the case $\Delta(G)=3$, Heckman and
Thomas \cite{ht} conjectured that $\chi_f(G) \leq 14/5$ if $G$ is
triangle-free. Hatami and Zhu \cite{hz} proved $\chi_f(G) \leq 3-
\frac{3}{64}$ for any triangle-free graph $G$ with $\Delta(G) \leq
3$. The authors showed an improved
result $\chi_f(G) \leq 3 - \frac{3}{43}$ in the previous paper \cite{lup}. Thus we need only consider
the cases $\Delta(G) \geq 4$.
For any
connected graph $G$ with sufficiently large $\Delta(G)$ and
$\omega(G) \leq \Delta(G)-1$, Reed's result \cite{reed1}
$\chi(G)\leq \Delta(G)-1$ implies $\chi_f(G) \leq \Delta(G)-1$.
The method introduced in \cite{hz} and strengthened in \cite{lup},
has a strong influence on this paper. The readers are encouraged to
read these two papers \cite{hz, lup}.
Let $f(k)=\inf_G\{\Delta(G)-\chi_f(G)\}$, where the infimum is taken
over all connected graphs $G$ with $\Delta(G)=k$ and not one of the
graphs listed in Theorem \ref{main}. Since $\chi_f(G)\geq \omega(G)$,
by taking a graph with $\omega(G)=\Delta(G)-1$, we have $f(k)\leq 1$.
Theorem \ref{submain} says
$f(k)\geq \frac{2}{67}$ for any $k\geq 4$.
Reed's result \cite{reed1}
implies $f(k)= 1$ for sufficiently large $k$. Heckman and Thomas
\cite{ht} conjectured $f(3)=1/5$. It is an interesting problem to
determine the value of $f(k)$ for small $k$. Here we conjecture
$f(4)=f(5)=\frac{1}{3}$. If Borodin and Kostochka's conjecture is
true, then $f(k)= 1$ for $k\geq 9$.
Theorem 2 is proved by induction on $k$.
Because the proof is quite long, we split the proof into
the following two lemmas.
\begin{lemma} \label{f4}
We have $f(4) \geq \frac{2}{67}$.
\end{lemma}
\begin{lemma}\label{increase}
For each $k \geq 6$, we have $f(k)\geq \min \left\{f(k-1),
\frac{1}{2}\right\}.$ We also have $f(5)\geq \min\left\{f(4),
\frac{1}{3}\right\}$.
\end{lemma}
It is easy to see the combination of Lemma \ref{f4} and Lemma
\ref{increase} implies Theorem \ref{submain}. The idea of reduction
comes from the first author, who pointed out $f(k)\geq \min \left\{f(k-1),
\frac{1}{2}\right\}$ for $k\geq 7$ based on his recent results
\cite{king}. The second and third authors orginally proved
$f(k)\geq \frac{C}{k^5}$ (for some $C>0$) using different method in the first version;
they also prove the reductions
at $k=5,6$, which are much harder than
the case $k\geq 7$. We do not know
whether a similar reduction exists for $k=4$.
The rest of this paper is organized as follows. In section 2, we
will introduce some notation and prove Lemma \ref{increase}. In
section 3 and section 4, we will prove $f(4) \geq \frac{2}{67}$.
\section{Proof of Lemma \ref{increase}}
In this paper, we use the following notation. Let $G$ be a simple
graph with vertex set $V(G)$ and edge set $E(G)$. The {\it
neighborhood} of a vertex $v$ in $G$, denoted by $\Gamma_G(v)$, is
the set $\{u \colon uv \in E(G)\}$. The {\it degree} $d_G(v)$ of $v$
is the value of $|\Gamma_G(v)|$. The {\it independent
set} (or {\it stable set}) is a set $S$ such that no edge with both
ends in $S$. The {\it independence number} $\alpha(G)$ is the largest
size of $S$ among all the independent sets $S$ in $G$. When $T
\subset V(G)$, we use $\alpha_G(T)$ to denote the independence number
of the induced subgraph of $G$ on $T$. Let $\Delta(G)$ be the maximum
degree of $G$. For any two vertex-sets $S$ and $T$, we define
$E_G(S,T)$ as $\{uv \in E(G): u \in S \ \textrm{and} \ v \in T\}$.
Whenever $G$ is clear under context, we will drop the subscript $G$
for simplicity.
If $S$ is a subset of vertices in $G$, then {\it contracting} $S$
means replacing vertices in $S$ by a single fat vertex, denoted by $\underline{S}$,
whose incident
edges are all edges that were incident to at least one vertex in
$S$, except edges with both ends in $S$. The new graph obtained by
contracting $S$ is denoted by $G/S$. This operation is also known
as {\it identifying vertices of $S$} in the literature. For
completeness, we allow $S$ to be a single vertex or even the empty
set. If $S$ only consists of a single vertex, then $G/S=G$; if
$S=\emptyset$, then $G/S$ is the union of $G$ and an isolated
vertex. When $S$ consists of $2$ or $3$ vertices, for convenience,
we write $G/uv$ for $G/\{u,v\}$ and $G/uvw$ for $G/\{u,v,w\}$; the
fat vertex will be denoted by $\underline{uv}$ and
$\underline{uvw}$, respectively. Given two disjoint subsets $S_1$
and $S_2$, we can contract $S_1$ and $S_2$ sequentially. The order
of contractions does not matter; let $G/S_1/S_2$ be the resulted
graph. We use $G-S$ to denote the subgraph of $G$ induced by
$V(G)-S$.
In order to prove Lemma \ref{increase}, we need use the following theorems
due to King \cite{king}.
\begin{theorem}[King \cite{king}]
\label{king0}
If a graph $G$ satisfies $\omega(G) > \frac{2} {3} (\Delta(G) + 1)$,
then $G$ contains a stable set $S$ meeting every maximum clique.
\end{theorem}
\begin{theorem}[King \cite{king}]
\label{king}
For a positive integer $k$, let $G$ be a graph with vertices
partitioned into cliques $V_1,\ldots,V_r$. If for every $i$ and
every $v \in V_i$, $v$ has at most $\min\{k, |V_i|-k\}$ neighbors
outside $V_i$, then $G$ contains a stable set of size $r$.
\end{theorem}
\begin{lemma} \label{6to5}
Suppose that $G$ is a connected graph with $\Delta(G)\leq 6$ and $\omega(G)\leq 5$.
Then there exists an independent set meeting all induced copies
of $K_5$ and $C_5\boxtimes K_2$.
\end{lemma}
\noindent
{\bf Proof:} We first show that there exists an independent set
meeting all copies of $K_5$. If $G$ contains no $K_5$, then this is
trivial. Otherwise, we can apply Theorem \ref{king0} to get the
desired independent set since $\omega(G)>\frac{2}{3}(\Delta(G)+1)$
is satisfied.
Now we prove the Lemma by contradiction. Suppose the Lemma is false.
Let $G$ be a minimum counterexample (with the smallest number of
vertices). For any independent set $I$, let $C(I)$ be the number of
induced copies of $C_5\boxtimes K_2$ in $G-I$. Among all independent
sets which meet all copies of $K_5$, there exists one such
independent set $I$ such that $C(I)$ is minimized.
Since $C(I)>0$, there is an induced copy of $C_5\boxtimes K_2$ in
$G-I$; we use $H$ to denote it. In $C_5\boxtimes K_2$, there is a
unique perfect matching such that identifying
the two ends of each edge in this matching results a $C_5$.
An edge in this unique matching is called a {\em canonical } edge.
We define a new graph $G'$ as follows: First we contract all
canonical edges in $H$ to get a $C_5$, where its vertices are called {\em fat}
vertices.
Second we add five edges
turning the $C_5$ into a $K_5$. Observe that each vertex in this
$C_5$ can have at most two neighbors in $G-H$ and $\Delta(G') \leq
6$. We will consider the following four cases.
\noindent{\bf Case 1:} There is a $K_6$ in the new graph $G'$. Since
the original graph $G$ is $K_6$-free, the $K_6$ is formed by the
following two possible ways.
{\bf Subcase 1a:} This $K_6$ contains $5$ fat vertices. By the symmetry of $H$,
there is an induced $C_5$ in $H$
such that the vertices in $C_5$ contain a common neighbor vertex $v$ in $G \setminus V(H)$, see Figure \ref{fig:1a}.
Since $H$ is $K_5$-free, we can find $x, y$ in this $C_5$ such that $x,y$ is a non-edge.
Let $I':=(I\setminus\{v\})\cup\{x, y\}$; $I'$ is also an
independent set. Observe that $v$ is not in any $K_5 \subset
G-I'$. Thus the set $I'$ is also an independent set and meets every
$K_5$ in $G$. Since $C_5\boxtimes K_2$ is a $5$-regular graph, any
copy of $C_5\boxtimes K_2$ containing $v$ must contain at least one
of $x$ and $y$. Thus, $C(I')<C(I)$. Contradiction!
\begin{figure}[htbp]
\centerline{ {\psfig{figure=image/case1a.eps, width=0.32\textwidth}}
\hfil \psfig{figure=image/case1b.eps, width=0.32\textwidth}}
\begin{multicols}{2}
\caption{Subcase 1a.}
\label{fig:1a}
\newpage
\caption{Subcase 1b.}
\label{fig:1b}
\end{multicols}
\end{figure}
{\bf Subcase 1b:} This $K_6$ contains $4$ fat vertices.
Let $u,v$ be the other two vertices.
By the symmetry of $H$, there is a unique way to connect $u$ and $v$ to $H$
as shown by Figure \ref{fig:1b}.
Since $uv$ is an edge, one of $u$ and $v$ is not in $I$. We assume
$u \not \in I$. Let $\{x,y\} \subset \Gamma_G(v) \cap V(H)$ as shown
in Figure \ref{fig:1b} and $I'=I \setminus \{v\} \cup \{x,y\} $.
Observe that $I'$ is an independent set and $v \not \in K_5 \subset
G-I'$. Thus $I'$ is an independent set meeting each $K_5$ in $G$.
Since each $C_5 \boxtimes K_2$ containing $v$ must contain one of
$x$ and $y$. Thus $C(I')<C(I)$. Contradiction!
\noindent {\bf Case 2:}
There is a $K_5$ intersecting $H$ with $4$
vertices. Let $v$ be the vertex of this $K_5$ but not in $H$, see Figure \ref{fig:2}.
We have two subcases.
{\bf Subcase 2a:} The vertex $v$ has another neighbor $y$ in $H$ but
not in this $K_5$. Since $H$ is $K_5$-free, we can select a vertex
$x$ in this $K_5$ such that $xy$ is not an edge of $G$. Let
$I':=I\setminus \{v\}\cup \{x,y\}$. Note that $v \not \in K_5
\subset G-I'$ and $I'$ is an independent set. Thus $I'$ is an
independent set meeting each $K_5$ in $ G$. Since any $C_5
\boxtimes K_2$ containing $v$ must contain one of $x$ and $y$, we
have $C(I')<C(I)$. Contradiction!
{\bf Subcase 2b:} All neighbors of $v$ in $H$ are in this $K_5$.
Let $x$ be any vertex in this $K_5$ other than $v$, and
$I':=I\setminus \{v\}\cup \{x\}$. In this case, there is only one
$K_5$ containing $v$. Thus, $I'$ is also an independent set meeting
every copy of $K_5$ in $G$. Observe that $\Gamma_G(v) \setminus
\{x\}$ is disconnected. If $v \in H'=C_5 \boxtimes K_2$, then
$\Gamma_G(v) \cap H'$ is connected. Thus $v \not \in C_5\boxtimes
K_2 \subset G-I'$ and $C(I')< C(I)$. Contradiction!
\begin{figure}[htbp]
\centerline{ {\psfig{figure=image/case2a.eps, width=0.32\textwidth}}
\hfil \psfig{figure=image/case3.eps, width=0.32\textwidth}}
\begin{multicols}{2}
\caption{Case 2.}
\label{fig:2}
\newpage
\caption{Case 3.}
\label{fig:3}
\end{multicols}
\end{figure}
\noindent {\bf Case 3:} There is an induced subgraph $H'$
isomorphic to $C_5\boxtimes K_2$ such that $H'$ and $H$ are
intersecting, see Figure \ref{fig:3}.
Since $V(H)\cap V(H')\not =\emptyset$ and $H\not=H'$, we can find a
canonical edge $uv$ of $H$ and a canonical edge $uv'$ of $H'$ such
that $v\not\in V(H')$ and $v'\not\in V(H)$.
If $vv'$ is a non-edge, then let $I':=I\setminus \{v'\} \cup \{u\}$.
It is easy to check $I'$ is still an independent set. We also
observe that any possible $K_5$ containing $v'$ must also contain
$u$. Thus, $I'$ meets every copy of $K_5$ in $G$. We have $v' \not
\in C_5 \boxtimes K_2 \subset G-I'$ since $vv'$ is not an edge. We
get $C(I')<C(I)$. Contradiction! If $vv'$ is an edge, then locally
there are two $K_5$ intersecting at $u$, $v$, and $v'$; say the
other four vertices are $x_1, x_2, y_1, y_2$, where two cliques are
$\{x_1,x_2,u,v,v'\}$ and $\{y_1,y_2,u,v,v'\}$, see Figure
\ref{fig:3}. Let $I'=I\cup\{x_1,y_1\}\setminus \{v'\}$. Note that
$I'$ is an independent set and $v' \not \in K_5 \subset G-I'$. Thus
$I'$ is an independent set meeting each $K_5$ in $G$. Observe that
any copies of $C_5\boxtimes K_2$ containing $v'$ must contain one
of $x_1$ and $y_1$; we have $C(I')<C(I)$. Contradiction!.
\noindent
{\bf Case 4:} This is the remaining case, $G'$ is $K_6$-free. We
have $\omega(G')\leq 5$ and $|V(G')| < |V(G)|$. By the minimality of
$G$, there is an independent set $I'$ of $G'$ meeting every copy of
$K_5$ and $C_5\boxtimes K_2$. In $I'$, there is a unique vertex $x$
of the $K_5$ obtained from contracting canonical edges of $H$. Let
$uv$ be the canonical edge corresponding to $x$. Let $I''=I'
\setminus \{x\} \cup \{u\}$, we get an independent set $I''$ of $G$.
Note that any $v \in H \setminus \{u\}$ is not in any $K_5$ of
$G-I''$ by Case 2 as well as not in any $C_5 \boxtimes K_2$ of
$G-I''$ by Case 3. Thus $I''$ hits each $K_5$ in $G$ and $C(I'')=0$.
Contradiction! \hfill $\square$
\begin{lemma} \label{5to41}
Let $G$ be a connected graph with $\Delta(G)\leq 5$ and $\omega(G)\leq 4$.
If $G \not = C_{2l+1} \boxtimes K_2$ for some $l \geq 2$,
then there is an independent set $I$ hitting all copies of $K_4$ in $G$.
\end{lemma}
{\bf Proof:} We will prove it by contradiction. If the lemma is
false, then let $G$ be a minimum counterexample. If $G$ is
$K_4$-free, then there is nothing to prove. Otherwise, we consider
the clique graph ${\cal C}(G)$, whose edge set is the set of all edges
appearing in some copy of $K_4$. Because of $\Delta(G)=5$, here are
all possible connected component of ${\cal C}(G)$.
\begin{enumerate}
\item $C_{t}\boxtimes K_2$ for $t\geq 4$. If this type occurs, then every vertex
in $C_{t}\boxtimes K_2$ has degree $5$; thus, this is the entire
graph $G$. If $t$ is even, then we can find an independent set $I$
meeting every $K_4$. If $t$ is odd, then it is impossible to find
such an independent set. However, this graph is excluded from the
assumption of the Lemma.
\item $P_t\boxtimes K_2$ for $t\geq 3$. In this case, all internal vertices
have degree $5$ while the four end vertices have degree $4$.
Consider a new graph $G'$ which is obtained by deleting all internal
vertices and adding four edges to make the four end vertices as a
$K_4$. It is easy to check $\Delta(G')\leq 5$ and $\omega(G')\leq
4$. Since $|G'|<|G|$, there is an independent set $I$ of $G'$
meeting every copy of $K_4$ in $G'$. Note that there is exactly one
end vertex in $I$. Observe that any one end vertex can be extended
into a maximal independent set meeting every copy of $K_4$ in
$P_t\boxtimes K_2$. Thus, we can extend $I$ to an independent set
$I'$ of $G$ such that $I'$ meets every copy of $K_4$ in $G$. Hence,
this type of component does not occur in ${\cal C}(G)$.
\item There are four other types listed in Figure \ref{k4}.
\begin{figure}[htbp]
\centerline{ {\psfig{figure=image/k4.eps, width=0.2\textwidth}}
\hfil
\psfig{figure=image/k422.eps, width=0.2\textwidth}\hfil
{\psfig{figure=image/k431.eps, width=0.2\textwidth}}
\hfil \psfig{figure=image/k432.eps, width=0.2\textwidth}}
\caption{All types of components in the clique graph $C(G)$.}
\label{k4}
\end{figure}
For each component $C_i$ in ${\cal C}(G)$, let $V_i$ be the set of common
vertices in all $K_4$'s of $C_i$; for the leftmost figure in Figure
\ref{k4}, $V_i$ is the set of all 4 vertices; for the middle two
figures, $V_i$ is the set of bottom three vertices; for the
rightmost figure, $V_i$ consists of the left-bottom vertex and the
middle-bottom vertex. Note that all $V_i$'s are pairwise disjoint.
Let $G'$ be the induce subgraph of $G$ on $\cup_i V_i$. Note that
$G'$ does not contains any vertex in $C_i\setminus V_i$. By checking
each type, we find out that
for each $i$ and each $v\in V_i$, $v$ has at most $\min\{2,|V_i|-2\}$
neighbors outside $V_i$ in $G'$ (not in $G$!). Applying Theorem
\ref{king} to $G'$, we conclude that there exists an independent set
$I$ of $G'$ meeting every $V_i$; thus $I$ meets every $K_4$ in $G$.
Contradiction!
\end{enumerate}
\hfill $\square$
\begin{lemma} \label{5to42}
Let $G$ be a connected graph with $\Delta(G)\leq 5$ and
$\omega(G)\leq 4$. If $G \not = C_{2l+1} \boxtimes K_2$ for some $l
\geq 2$, then there exists an independent set meeting all induced
copies of $K_4$ and $C_8^2$.
\end{lemma}
\noindent
{\bf Proof:} We will use proof by contradiction. Suppose the Lemma
is false. Let $G$ be a minimum counterexample (with the smallest
number of vertices). For any independent set $I$, let $C(I)$ be the
number of induced copies of $C_8^2$ in $G-I$. Among all independent
sets which meet all copies of $K_4$, there exists an independent set
$I$ such that $C(I)$ is minimized. Since $C(I)>0$, let $H$ be a copy
of $C_8^2$ in $G-I$. The vertices of $H$ are listed by $u_i$ for
$i\in {\mathbb Z}_8$ anticlockwise such that $u_iu_j$ is an edge of
$H$ if and only if $|i-j| \leq 2$. The vertex $v_{i+4}$ is the {\em
antipode} of $v_i$ for any $i\in {\mathbb Z}_8$.
\noindent
{\bf Case 1:} There exists a vertex $v\not\in V(H)$ such that $v$ has
five neighbors in $H$. By the Pigeonhole Principle, $\Gamma(v)$ contains
a pair of antipodes. Without loss of generality, say $u_0, u_4\in
\Gamma(v)$. If the other three neighbors of $v$ do not form a
triangle, then we let $I':=I\setminus \{v\} \cup \{u_0,u_4\}$; note
that $v$ is not in any $K_4$ of $G-I'$. Thus $I'$ is an independent
set meeting every copy of $K_4$. Since every copy of $C_8^2$
containing $v$ must contain one of $u_0$ and $u_4$, we have
$C(I')<C(I)$. Contradiction! Hence, the other three neighbors of $v$
must form a triangle. Without loss of generality, we can assume
that the three neighbors are $u_1$, $u_2$, and $u_3$. Now we let
$I'':=I\setminus \{v\} \cup \{u_0,u_3\}$; note that $v \not \in K_4
\subset G-I'$. Thus $I''$ is also an independent set meeting every
copy of $K_4$ of $G$. Since every copy of $C_8^2$ containing $v$
must contain one of $u_0$ and $u_3$, we have $C(I'')<C(I)$.
Contradiction!
\noindent
{\bf Case 2:} There exists a vertex $v\not\in V(H)$ such that $v$
has exactly four neighbors in $H$. Since $H$ is $K_4$-free, we can
find $u_i, u_j\in \Gamma(v)\cap V(H)$ such that $u_iu_j$ is a
non-edge. Let $I':=I\setminus \{v\} \cup \{u_i,u_j\}$; $I'$ is also
an independent set. Note that $\Gamma(v)\setminus \{u_i, u_j\}$
can not be a triangle, $v$ is not in any $K_4 \subset G-I'$. Thus
$I'$ meets every copy of $K_4$. Since every copy of $C_8^2$
containing $v$ must contain one of $u_i$ and $u_j$, we have
$C(I')<C(I)$. Contradiction!
\noindent
{\bf Case 3:} There exists a vertex $v\not\in V(H)$ such that $v$
has exactly three neighbors in $H$. If the $3$ neighbors do not form
a triangle, then choose $u_i, u_j\in \Gamma(v)\cap V(H)$ such that
$u_iu_j$ is a non-edge. Note that $\Gamma(v)\setminus \{u_i, u_j\}$
can not be a triangle; $v$ is not in any $K_4 \subset G-I'$. Let
$I':=I\setminus \{v\} \cup \{u_i,u_j\}$; $I'$ is also an independent
set meeting every copy of $K_4$. Since every copy of $C_8^2$
containing $v$ must contain one of $u_i$ and $u_j$, we have
$C(I')<C(I)$. Contradiction! Else, the three neighbors form a
triangle; let $u_i$ be one of them and $I':=I\setminus \{v\} \cup
\{u_i\}$; $v$ is not in any $K_4 \subset G-I'$. Thus $I'$ is an
independent set meeting every copy of $K_4$. Note that
$\Gamma(v)\setminus \{u_i\}$ has only two vertices in $H$. The
induced graph on $\Gamma(v)\setminus \{u_i\}$ is disconnected.
However, for any vertex $v$ in $H'=C_8^2$, the subgraph induced by
$\Gamma_G(v) \cap H'$ is a $P_4$. There is no $C_8^2$ in $G-I'$
containing $v$. Thus, $C(I')<C(I)$. Contradiction!
\noindent
{\bf Case 4:} Every vertex outside $H$ can have at most $2$
neighbors in $H$. We identify each pair of antipodes of $H$ to get a
new graph $G'$ from $G$. After identifying, $H$ is turned into a
$K_4$; where the vertices of this $K_4$ are referred as fat
vertices.
{\bf Subcase 4a:} $G'\not=C_{2l+1}\boxtimes K_2$.
Observe $\Delta(G')\leq 5$. We claim $G'$ is $K_5$-free.
Suppose not. Since every vertex in $H$ has at most one neighbor
outside $H$, then each fat vertex can have at most two neighbors
outside $H$. Recall that the original graph $G$ is $K_5$-free. If
$G'$ has some $K_5$, then this $K_5$ contains either $3$ or $4$ fat
vertices. Let $w$ be one of the other vertices in this $K_5$.
We get $w$ has at least three neighbors in
$H$. However, this is covered by Case 1, Case 2, or Case 3. Thus,
$G'$ is $K_5$-free.
Since $|G'|<|G|$, by the minimality of $G$, $G'$ has an independent set $I$
meeting every copy of $K_4$ and $C_8^2$ in $G'$. There is exactly
one fat vertex in $I$. Now replacing this fat vertex by its
corresponding pair of antipodal vertices, we get an independent set
$I'$; we assume the pair of antipodal vertices are $u_2$ and $u_6$.
It is easy to check that $I'$ is an independent set of $G$. Next we
claim any $v \in V(H) \setminus \{u_2,u_6\}$ is neither in a $K_4
\subset V(G)-I'$ nor in a $C_8^2 \subset V(G)-I'$. Suppose there is
some $v$ such that $v \in K_4 \subset G-I'$. Recall each $v \in
V(H)$ has at most one neighbor outside $H$ and $H$ is $K_4$-free;
there is some $w \not \in V(H)$ such that $w$ has at least three
neighbors in $H$. This is already considered by Case 1, Case 2, or
Case 3. We are left to show that $v \not \in C_8^2 \subset G-I'$
for each $v \in V(H) \setminus \{u_2,u_6\}$. If not, there exists
a copy $H'$ of $C_8^2$ in $G-I'$ containing $v$. Note $H'$ is $4$-regular,
any vertex in $H'$ can have at most one neighbor in $I'$; in particular,
$v\not=u_0, u_4$. Without loss of generality, we assume $v=u_3$.
Then there is a vertex $w \not \in V(H)$ such
that $u_3w$ is an edge, see Figure \ref{subcase4a}. Observe that the
neighborhood of each vertex of an induced $C_8^2$ is is a $P_4$.
Since $u_1u_4$ and $u_1u_5$ are two non-edges, we have $wu_1$ being
an edge. Observe $\Gamma_G(u_1)=\{u_7,u_0,u_2,u_3,w\}$. Since $u_2
\not \in H'$, we have $u_0 \in H'$; $u_0$ has two neighbors ($u_2$ and $u_6$)
outside $H'$, contradiction!
Therefore, $I'$ meets every copy of $K_4$ and $C_8^2$ in $G$.
Contradiction!
\begin{figure}[htbp]
\centerline{ \psfig{figure=image/2c82.eps, width=0.3\textwidth}}
\caption{Subcase 4a.}
\label{subcase4a}
\end{figure}
{\bf Subcase 4b:} $G'=C_{2l+1}\boxtimes K_2$. The graph $G$ can be
recovered from $G'$. It consists of an induced subgraph $H=C_8^2$
and an induced subgraph $P_{2l-1}\boxtimes K_2$. For each vertex $u$
in $H$, there is exactly one edge connecting it to one of the four
end vertices of $P_{2l-1}\boxtimes K_2$; for each end vertex $v$ of
$P_{2l-1}\boxtimes K_2$, there are exactly two edges connecting $v$
to the vertices in $H$. First, we take any maximum independent set
$I'$ of $P_{2l-1}\boxtimes K_2$. Observe that $I'$ has exactly two
end points of $P_{2l-1} \boxtimes K_2$; so $I'$ has exactly four
neighbors in $H$. In the remaining four vertices of $H$, there
exists a non-edge $u_iu_j$ since $H$ is $K_4$-free. Let
$I:=I'\cup\{u_i,u_j\}$. Clearly $I$ is an independent set of $G$
meeting every copy of $K_4$ and $C_8^2$. Contradiction! \hfill
$\square$
We are ready to prove Lemma \ref{increase}.
\noindent
{\bf Proof of Lemma \ref{increase}:} We need prove
for $k\geq 5$ and any connected graph $G$ with $\Delta(G)=k$
and $\omega(G)\leq k-1$ satisfies
\begin{equation}
\label{eq:rec}
\chi_f(G)\leq k- \min\left\{f(k-1),\frac{1}{2}\right\}.
\end{equation}
If $\omega(G)\leq k-2$,
then by inequality (\ref{eq:1}), we have
$$\chi_f(G)\leq \frac{\Delta(G)+\omega(G)+1}{2}\leq k-\frac{1}{2}.$$
Thus, inequality (\ref{eq:rec}) is satisfied. From now on, we assume
$\omega(G)=\Delta(G)-1$.
For $\Delta(G)=k\geq 6$ and $\omega(G)=k-1$, the condition
$\omega(G)>\frac{2}{3}(\Delta(G)+1)$ is satisfied. By Theorem
\ref{king0}, $G$ contains an independent set meeting every maximum
clique. Extend this independent set to a maximal independent set and
denote it by $I$.
Note that $\Delta(G-I) \leq k-1$ and $\omega(G-I)\leq k-2$.
\noindent {\bf Case 1:} $k\geq 7$. From the definition of $f(k-1)$,
we have $\chi_f(G-I) \leq \Delta(G-I)-f(k-1)$. Thus,
$$
\chi_f(G) \leq \chi_f(G-I)+1 \leq k-1-f(k-1)+1=k-f(k-1).
$$
Thus, we have $f(k) \geq \min\{f(k-1),1/2\}$.
\noindent {\bf Case 2:} $k=6$. By Lemma \ref{6to5}, we can find an
independent set meets every copy of $K_5$ and $C_5\boxtimes K_2$;
we extend this independent set as a maximal independent set $I$.
Note that
$G-I$ contains no induced subgraph isomorphic $C_5\boxtimes K_2$.
We have $\chi_f(G-I) \leq 5-f(5)$; it implies $\chi_f(G)\leq
6-f(5)$.
Thus, $f(6) \geq \min\{f(5),1/2\}$ and we are done.
\noindent
{\bf Case 3:} $k=5$. If
$G=C_{2l+1}\boxtimes K_2$ for some $l\geq 3$; then $G$ is
vertex-transitive and $\alpha(G)=l$. It implies that
$$\chi_f(G)=\frac{|V(G)|}{\alpha(G)}=4+\frac{2}{l}\leq 5-\frac{1}{3}.$$
If $G\not =C_{2l+1}\boxtimes K_2$, then by Lemma \ref{5to42}, we can
find an independent set meeting every copy of $K_4$ and $C_8^2$; we
extend it as a maximal independent set $I$. Note that
$G-I$ contains no induced subgraph isomorphic $C_8^2$.
We have $\chi_f(G-I) \leq 4-f(4)$; it implies $\chi_f(G)\leq 5-f(4)$.
Thus, $f(3) \geq \min\{f(4),1/3\}$ and we are finished.
\hfill $\square$
\section{The case $\Delta(G)=4$}
To prove $f(4)\geq \frac{2}{67}$, we will use an approach which is
similar to those in \cite{hz, lup}. It suffices to prove that the
minimum counterexample does not exist.
Let $G$ be a graph with the smallest number of vertices and
satisfying
\begin{enumerate}
\item $\Delta(G)=4$ and $\omega(G)\leq 3$;
\item $\chi_f(G)> 4-\frac{2}{67}$;
\item $G\not=C_8^2$.
\end{enumerate}
By the minimality of $G$, each vertex in $G$ has degree either $4$
or $3$.
To prove Lemma \ref{f4},
we
will show $\chi_f (G) \leq 4 - \frac{2}{67}$, which gives us the
desired contradiction.
For a given vertex $x$ in $V(G)$, it is easy to color its
neighborhood $\Gamma_G(x)$ using $2$ colors. If $d_G(x)=3$, then we
pick a non-edge $S$ from $\Gamma_G(x)$ and color the two vertices in
$S$ using color 1. If $d_G(x)=4$ and $\alpha(\Gamma_G(x))\geq 3$,
then we pick an independent set $S$ in $\Gamma_G(x)$ of size 3 and
assign the color 1 to each vertex in $S$. If $d_G(x)=4$ and
$\alpha(\Gamma_G(x))=2$, then we pick two
disjoint non-edges $S_1$ and $S_2$ from $\Gamma_G(x)$ ; we assign color 1 to each vertex in $S_1$
and color 2 to each vertex in $S_2$.
The following Lemma shows that $G$ has a key property, which
eventually implies that this local coloring scheme works
simultaneously for $x$ in a large subset of $V(G)$.
\begin{lemma}\label{D41}
For each $x \in V(G)$
with $d_G(x)=4$ and $\alpha(\Gamma_G(x))=2$, there exist two
vertex-disjoint non-edges $S_1(x), S_2(x) \subset \Gamma_G(x)$
satisfying the following property. If we contract $S_1(x)$ and
$S_2(x)$, then the resulted graph $G/S_1(x) /S_2(x)$ contains
neither $K^-_{5}$ nor $G_0$. Here $K_5^-$ is the graph obtained
from $K_5$ by removing one edge and $G_0$ is a graph defined in
Figure \ref{fig:H13}.
\begin{figure}[htbp]
\centering
\psfig{figure=image/H122.eps,
width=0.4\textwidth}
\caption{The graph $G_0$.}
\label{fig:H13}
\end{figure}
\end{lemma}
The proof of this lemma is quite long and we will present its proof
in section 4.
For each vertex $x$ in $G$, we associate
a small set of vertices $S(x)$ selected from $\Gamma_G(x)$ as follows. If
$d_G(x)=3$, then let $S(x)$ be the endpoints of a non-edge in
$\Gamma_G(x)$ and label the vertices in $S(x)$ as 1; if $d_G(x)=4$
and $\alpha(\Gamma_G(x))\geq 3$, then let $S(x)$ be any independent
set of size $3$ in $\Gamma_G(x)$ and label all vertices in $S(x)$ as
1; if $d_G(x)=4$ and $\alpha(\Gamma_G(x))=2$, then let
$S(x)=S_1(x)\cup S_2(x)$, where $S_1(x)$ and $S_2(x)$ are guaranteed
by Lemma \ref{D41}; we label the vertices in $S_1(x)$ as 1 and the
vertices in $S_2(x)$ as 2. For any $x\in V(G)$, we have
$|S(x)|=2$, $3$, or $4$.
The following definitions depend on the choice of $S(\ast)$, which
is assumed to be fixed through this section. For $v \in G$ and $j
\in \{1,2,3\}$, we define
$$
N_G^j(v) = \{u| \ \mbox{there is a path} \ vv_0 \ldots v_{j-2}u \
\mbox{in}\ G \ \mbox{of length} \ j \ \mbox{such that} \ v_0
\in S(v) \ \mbox{and} \ v_{j-2} \in S(u)\}.
$$
For $j=4$, $v \in N_G^4(u)$ if $d_G(u)=4$,
$\alpha(\Gamma_G(u))=2$, $u$ and $v$ are connected as shown in
Figure \ref{fig:4n}; otherwise $N_G^4(u)=\emptyset$.
In Figure \ref{fig:4n}, $w$ is connected to one of the two vertices in $S_2(u)$.
Similarly, in Figure \ref{fig:5n} and \ref{fig:5n}, a vertex is connected to a group
of vertices means it is connected to any vertex in this group.
For $j=5$, $v \in N_G^5(u)$ if $d_G(w)=4$, $\alpha(\Gamma_G(w))=2$ for
$w \in \{u,v\}$ and $u$ and $v$ are connected as shown in Figure
\ref{fig:5n}; otherwise $N_G^5(u)=\emptyset$.
For $j=7$, $v \in N_G^7(u)$ if $d_G(w)=4$, $\alpha(\Gamma_G(w))=2$ for
$w \in \{u,v\}$ and $u$ and $v$ are connected as shown in Figure
\ref{fig:7n}; otherwise $N_G^7(u)=\emptyset$.
\begin{figure}[htbp]
\centering{
\psfig{figure=image/N4.eps, width=0.25\textwidth} \hfil
\psfig{figure=image/N5.eps, width=0.34\textwidth} \hfil
\psfig{figure=image/N7.eps, width=0.32\textwidth}}
\begin{multicols}{3}
\caption{\! 4-th neighborhood.\!\!\!}
\label{fig:4n}
\newpage
\caption{\! 5-th neighborhood.\!\!\!}
\label{fig:5n}
\newpage
\caption{\! 7-th neighborhood.\!\!\!}
\label{fig:7n}
\end{multicols}
\end{figure}
Note that for $j\in\{1,2,3,5,7\}$, $v\in N_G^j(u)$ if and only if
$u\in N_G^j(v)$; but this does not hold for $j=4$.
We have the following lemma.
\begin{lemma} \label{N5}
For $u \in V(G)$ such that $d_G(u)=4$ and $\alpha(\Gamma_G(u))=2$,
we have $|N_G^1(u) \cup N_G^2(u) \cup N_G^3(u) \cup N_G^4(u) \cup
N_G^5(u) \cup N_G^7(u)| \leq 96$.
\end{lemma}
{\bf Proof:} It is clear that $|N_G^1(u) \cup N_G^2(u) \cup
N_G^3(u)| \leq 4+8+8 \times 3=36$. We next estimate $|N_G^4(u)|$.
In Figure \ref{fig:4n}, observe that $w$ is
connected to one vertices of $S_2(u)$ and $w \not \in \Gamma_G(u)$.
For a fixed $u$, there are at most four choices for $w$, at most
three choices for $z$, and at most three choices for $v$. Therefore,
we have $|N_G^4(u)| \leq 4 \times 3 \times 3=36$.
Let us estimate $|N_G^5(u)|$. In Figure
\ref{fig:5n}, for a fixed $u$, we have four choices for $w$ and
two choices for $z$. Fix a $z$. Assume $\Gamma_G(z) \setminus
\{w\}=\{a,b,c\}$. Let $T_1=\{a,b\}$, $T_2=\{b,c\}$, and
$T_3=\{a,c\}$. We have the following claim.
\noindent {\bf Claim} There are at most three $v \in N_G^5(u)$ such
that for each $v$ we have $ \Gamma_G(z) \cap \Gamma_G(v)=T_i$ for
some $1 \leq i\leq 3$ as shown in Figure \ref{fig:5n}.
\noindent {\bf Proof of the claim:} For each $1 \leq i \leq 3$,
there are at most three $v \in N_G^5(u)$ such that $ \Gamma_G(z)
\cap \Gamma_G(v)=T_i$ as shown in Figure \ref{fig:5n} since each
vertices in $T_i$ has at most three neighbors other than $z$. If
the claim is false, then there is $1 \leq i \not =j \leq 3$ such
that $\Gamma_G(z) \cap \Gamma_G(v_{i})=T_i$ and $\Gamma_G(z) \cap
\Gamma_G(v_{i}')=T_i$ for some $v_{i}, v_{i}' \in N_G^5(u)$, and
$\Gamma_G(z) \cap \Gamma_G(v_j)=T_j$ for some $v_j \in N_G^5(u)$,
where $v_i,v_i', v_j$ are distinct. Without loss of generality, we
assume $\Gamma_G(z) \cap \Gamma_G(v_1)=\Gamma_G(z) \cap
\Gamma_G(v_1')=T_1$ for $v_1, v_1' \in N_G^5(u)$, and $\Gamma_G(z)
\cap \Gamma_G(v_2)=T_2$ for some $v_2 \in N_G^5(u)$, see Figure
\ref{N51}. Observe that $\Gamma_G(b)=\{v_1,v_1',v_2,z\}$. Since
$\Gamma_G(z) \cap \Gamma_G(v_1)=T_1$ as shown in Figure
\ref{fig:5n}, $a$ and one of $b$'s neighbors form $S_i(v_1)$ for
some $i \in \{1,2\}$; we assume it is $S_1(v_1)$. Note
$\{z,v_1,v_1'\} \subset \Gamma_G(a)$. Thus $S_1(v_1)=\{a,v_2\}$ and
$v_2 \in \Gamma_G(v_1)$. Similarly, we can show
$S_1(v_1')=\{a,v_2\}$ and $v_2 \in \Gamma_G(v_1')$. Now, observe
that $\Gamma_G(v_2)=\{v_1,v_1',b,c\}$. Since $\Gamma_G(z) \cap
\Gamma_G(v_2)=T_2$ as shown in Figure \ref{fig:5n}, $b$ and one of
neighbors of $v_2$ form $S_i(v_2)$ for some $i \in \{1,2\}$; we
assume $i=1$. Because $\{v_1,v_1'\} \subset \Gamma_G(b)$, then
$S_1(v_2)=\{b,c\}$. However, $b$ and $c$ are not is in the same
independent set in the definition of $N_G^5(u)$, see Figure
\ref{fig:5n}. This is a contradiction and this case can not
happen. The claim follows.
\begin{figure}[htbp]
\centering{
\psfig{figure=image/N51.eps, width=0.35\textwidth}}
\caption{The picture for the claim.}\label{N51}
\end{figure}
Therefore, $|N_G^5(u)| \leq 4 \times 2 \times 3=24.$
In Figure \ref{fig:7n}, for a fixed $u$, we
have two choices for the edge $e$, one choice for $w$, two choices
for $z$, and three choices for the edge $f$. Fix a $z$. By
considering the degrees of the endpoints of $f$, there is at most
one $f$ and at most one $v \in N_G^7(u)$ such that $|\Gamma_G(f)
\cap \Gamma_G(v)|=4 $ as shown in Figure \ref{fig:7n}. Therefore,
we have $|N_G^7(u)| \leq 2 \times 2 \times 1=4$.
Last, we estimate $|N_G^5(u) \cup N_G^7(u)|$. If there is some $v
\in N_G^7(u)$, then we observe that there are at most five $z$'s
(see Figure \ref{fig:5n}). We get the
number of $v \in N_G^5(u)$ is at most $5 \times 3=15$. In this case,
we have
$$
|N_G^5(u) \cup N_G^7(u)| \leq 4+15 < 24.
$$
If $N_G^7(u)=\emptyset$, then also we have
$$
|N_G^5(u) \cup N_G^7(u) \leq 24.$$
Therefore $$|N_G^1(u) \cup N_G^2(u) \cup N_G^3(u) \cup N_G^4(u)
\cup N_G^5(u) \cup N_G^7(u)| \leq 36+36+24=96.$$ \hfill $\square$
Based on the graph $G$, we define an auxiliary graph $G^{\ast}$ on
vertex set $V(G)$. The edge set is following: $uv \in E(G^{\ast})$
if either $u \in N_G^1(v) \cup N_G^2(u) \cup N_G^3(u) \cup N_G^5(u)
\cup N_G^7(u) $, or $v \in N_G^4(u)$, or $u \in N_G^4(v)$.
We have the following lemma.
\begin{lemma} \label{121color}
The graph $G^{\ast}$ is 133-colorable.
\end{lemma}
\noindent
{\bf Proof:} Let $\sigma$ be an increasing order of $V(G^{\ast})$
satisfying the following conditions.
\begin{description}
\item [1:] For $u$ and $v$ such that $d_G(u)=3$ and $d_G(v)=4$, we have $\sigma(u) <
\sigma(v)$.
\item [2:] For $u$ and $v$ such that $d_G(u)=d_G(v)=4$, $\alpha(\Gamma_G(u)) \geq
3$, and $\alpha(\Gamma_G(v))=2$, we have $\sigma(u) < \sigma(v)$.
\end{description}
We will color $V(G^{\ast})$ according to the order $\sigma$. For
each $v$, we have the following estimate on the number of colors
forbidden to use for $v$.
\begin{description}
\item [1:] For $v$ such that $d_G(v)=3$, the number of colors
forbidden to use for $v$ is at most $|N_G^1(v) \cup N_G^2(v) \cup
N_G^3(v)| \leq 4+8+24=36$.
\item [2:] For $v$ such that $d_G(v)=4$ and $\alpha(\Gamma_G(v)) \geq
3$, the number of colors forbidden to use for $v$ is at most
$|N_G^1(v) \cup N_G^2(v) \cup N_G^3(v)| \leq 4+8+24=36$.
\item [3:] For $v$ such that $d_G(v)=4$ and
$\alpha(\Gamma_G(v))=2$, the number of colors forbidden to use for
$v$ is at most $|N_G^1(v) \cup N_G^2(v) \cup N_G^3(v) \cup N_G^4(v)
\cup N_G^5(v) \cup N_G^7(v)|+|N_G^4(v)| \leq 96+36=132$ by Lemma
\ref{N5}.
\end{description}
Therefore, The greedy algorithm shows $G^{\ast}$ is 133-colorable.
\hfill $\square$
Let $X$ be a coloring class of $G^{\ast}$. We define a new graph
$G(X)$ by the following process.
\begin{enumerate}
\item For each $x \in X$, if $|S(x)|=2$ or $|S(x)|=3$, then we contract $S(x)$ as a single vertex,
delete the vertices in $\Gamma_G(v) \setminus S(v)$, and keep
labeling the new vertex as 1; if $|S(x)|=4$, i.e., $S(x)=S_1(x)
\cup S_2(x)$, then we contract $S_1(x)$ and $S_2(x)$ as single
vertices and keep their labels. After that, we delete $X$. Let $H$
be the resulted graph.
\item Note that $\Gamma_H(x) \cap \Gamma_H(y) = \emptyset$ and there is no edge from
$\Gamma_H(x)$ to $\Gamma_H(y)$ for any $x,y \in X$ as $X$ is a coloring class.
\item We identify all vertices with label $i$ as a single vertex $w_i$ for $i \in \{1,2\}$. Let $G(X)$ be the resulted graph.
\end{enumerate}
We have the following lemma on the chromatic number of $G(X)$.
\begin{lemma} \label{4colorable}
The graph $G(X)$ is $4$-colorable for each coloring class.
\end{lemma}
We postpone the proof of this lemma until the end of this section
and prove the Lemma \ref{f4} first.
\noindent
{\bf Proof of Lemma \ref{f4}:} By Lemma \ref{121color}, there is a
proper 133-coloring of $G^{\ast}$. We assume
$V(G^{\ast})=V(G)=\cup_{i=1}^{133} X_i $, where $X_i$ is the $i$-th
coloring class.
For each $i \in \{1,\ldots, 133\}$, Lemma \ref{4colorable} shows
$G(X_i)$ is $4$-colorable; let $c_i\colon V(G(X_i)) \to T_i$ be a
proper $4$-coloring of the graph $G(X_i)$. Here $T_1,T_2,\ldots,
T_{133}$ are pairwise disjoint; each of them consists of $4$ colors.
For $i \in \{1,\ldots, 133\}$, the $4$-coloring $c_i$ can be viewed
as a $4$-coloring of $G \setminus X_i$ since each vertex with label
$j$ receives the color $c_i(w_i)$ for $j=1,2$ and each removed
vertex has at most three neighbors in $G \setminus X_i$.
Now we reuse the notation $c_i$ to denote this $4$-coloring of $G
\setminus X_i$. For each $v \in X_i$, we have $|\cup_{u \in
\Gamma_G(v)}c_i(u)| \leq 2$. We can assign two unused colors,
denoted by the set $Y(v)$, to $v$. We define $f_i: V(G) \rightarrow
\mathcal P(T_i)$ (the power set of $T_i$) satisfying
$$
f_i(v)= \left\{
\begin{array}{ll}
\{c_i(v)\} \ & \textrm{if} \ v \in V(G) \setminus X_i, \\
Y(v) & \textrm{if} \ v \in X_i.\\
\end{array}
\right.
$$
Observe that each vertex in $X_i$ receives two colors from $f_i$ and
every other vertex receives one color. Let $\sigma: V(G)
\rightarrow \mathcal P(\cup_{i=1}^{k} T_i)$ be
a mapping such that $\sigma(v)=\cup_{i=1}^{m} f_i(v)$. It is easy
to verify
$\sigma$ is a $134$-fold coloring of $G$ such that each color is drawn
from a palette of $532$ colors; namely we have
$$\chi_f(G) \leq
\frac{532}{134} =4-\frac{2}{67}.$$
The proof of Lemma \ref{f4} is finished. \hfill $\square$
Before we prove Lemma \ref{4colorable}, we need the following definitions.
A {\it block} of a graph is a maximal $2$-connected induced subgraph.
A {\it Gallai tree} is a connected graph in which all blocks are
either complete graphs or odd cycles. A {\it Gallai forest} is a
graph all of whose components are Gallai trees. A $k$-{\it Gallai
tree (forest)} is a Gallai tree (forest) such that the degree of all
vertices are at most $k-1$. A $k$-critical graph is a graph $G$
whose chromatic number is $k$ and deleting any vertex can decrease
the chromatic number. Gallai showed the following Lemma.
\begin{lemma}\label{gallai} \cite{gallai}
If $G$ is a $k$-critical graph, then the subgraph of $G$ induced on
the vertices of degree $k-1$ is a k-Gallai forest.
\end{lemma}
\noindent
{\bf Proof of Lemma \ref{4colorable}:} We use proof by
contradiction. Suppose that $G(X)$ is not $4$-colorable.
The only possible vertices in $G(X)$ with degree greater than $4$ are the
vertices $w_1$ and $w_2$, which are
obtained by contracting the vertices with label 1 and 2 in the intermediate graph $H$.
The simple greedy algorithm shows that $G(X)$ is always
$5$-colorable. Let
$G'(X)$ be a $5$-critical subgraph of $G(X)$. Applying
Lemma \ref{gallai} to $G'(X)$, the subgraph of $G'(X)$ induced on the
vertices of degree $4$ is a $5$-Gallai forest $F$. The vertex set of $F$
may contain $w_1$ or $w_2$. Delete $w_1$ and $w_1$ from $F$ if $F$
contains one of them. Let $F'$ be the resulted Gallai forest. (Any
induced subgraph of a Gallai forest is still a Gallai forest.) The Gallai forest $F'$ is not empty. Let $T$ be a
connected component of $F'$ and $B$ be a leaf block of $T$. The
block $B$ is either a clique or an odd cycle from the definition of a
Gallai tree.
Let $v$ be a vertex in $B$. As $v$ has at most two neighbors ($w_1$
and $w_2$) outside $F'$ in $G(X)$, we have $d_{F'}(v)\geq 2$. If
$v$ is not in other blocks of $F'$, then we have $d_B(v)\geq 2$.
It follows that $|B|\geq 3$. Since $B$ is a subgraph of $G$ and
$G$ is $K_4$-free, the block $B$ is an
odd cycle.
Let $v_1v_2$ be an edge in $B$ such that $v_1$ and $v_2$ are not in
other blocks. The degree requirement implies $v_iw_j$ are edges in
$G(X)$ for all $i, j\in\{1,2\}$. For $i=1,2$, there are vertices
$x_i, y_i\in X$ satisfying $S(x_i)\cap \Gamma_G(v_i)\not =\emptyset$
and $S(y_i)\cap \Gamma_G(v_i)\not =\emptyset$; moreover either
$|S(x_i)| =4$ or $|S(y_i)|=4$ since one of its neighborhood has
label 2. Without loss of generality, we assume $|S(x_i)|=4$ for $i
\in \{1,2\}$. If $x_i\not= y_i$, then $y_i \in N_G^4(x_i)$, i.e.,
$y_i \in \Gamma_{G^{\ast}}(x_i)$; this contradicts $X$ being a
coloring class. Thus we have $x_i=y_i$
and $|S(x_i)|=4$ for $i \in \{1,2\}$. For
$\{i,j\}=\{1,2\}$, if $x_i\not= y_j$, then $y_i \in N_G^5(x_i)$,
i.e., $y_i \in \Gamma_{G^{\ast}}(x_i)$; this is a contradiction of $X$ being a coloring class.
Thus we
have
$$x_1=x_2 = y_1=y_2.$$
Let $x$ denote this common vertex above. Then
$d_G(x)=4$ and $\alpha(\Gamma_G(x))=2$.
Let $v_0$ be the only vertex in $B$ shared by other blocks. Since
$B-v_0$ is connected, the argument above shows there is a common $x$
for all edges in $B-v_0$. If $\Gamma_{G(X)}(v_0)\cap
\{w_1,w_2\}\not=\emptyset$, the there is some vertex $x_0 \in X$
such that $S(x_0) \cup \Gamma_G(v_0) \not =\emptyset$. By the
similar argument, we also have $x_0=x$.
Therefore, $x$ depends only on $B$. In the sense that for any $y\in X$ and any $v\in B$, if $S(y)\cap \Gamma_G(v)\not=\emptyset$,
then $y=x$.
The block $B$ is an odd cycle as we mentioned above. Suppose $|B|=2r+1$. Let $v_0,v_1,\ldots,v_{2r}$ be the vertices
of $B$ in cyclic order and $v_0$ be the only vertex which may be shared by other block.
Let $x\in X$ be the vertex determined by $B$. Recall
$d_G(x)=4$ and $\alpha(\Gamma_G(x))=2$.
Each vertex in $\Gamma(x)$ can have at most $2$ edges to $B$. We get
\begin{equation}
\label{eq:12}
4r\leq |E(B,\Gamma(x))|\leq 8.
\end{equation}
We have $r\leq 2$. The block $B$ is either a $C_5$ or a $K_3$.
We claim both $v_0w_1$ and $v_0w_2$ are non-edges of
$G(X)$.
If $B=C_5$, then inequality (\ref{eq:12}) implies
that $v_0$ has no neighbor in
$\Gamma(x)$ and the claim holds.
If $B=K_3$, then the claim also holds; otherwise $B\cup
\{\underline{S_1(x)}, \underline{S_2(x)}\}$ forms a $K_5^-$ in
$G/S_1(x)/S_2(x)$, which is a contradiction to Lemma \ref{D41}.
Let $u_1$ and $u_2$ be the two neighbors
of $v_0$ in other blocks of $F'$. If $u_1$ and $u_2$ are in the same block,
then this block is an odd cycle; otherwise, $v_0u_1$ and $v_0u_2$ are in two different blocks.
The union of non-leaf blocks of $T$ is a Gallai-tree, denoted by
$T'$. The argument above shows every leaf block of $T'$ must be an
odd cycle. Let $C$ be such a leaf block of $T'$. Now $C$ is an odd
cycle, and $C$ is connected to $|C|-1$ leaf blocks of $T$. Let $B$
and $B'$ be two leaf blocks of $T$ such that $B\cap C$ is adjacent
to $B'\cap C$. Without loss of generality, we may assume $B$ is the
one we considered before. By the same argument, $B'$ is an odd
cycle of size $2r'+1$ with $r'\in \{1,2\}$. Let $v'_0,v'_1,\ldots,
v'_{2r'}$ be the vertices of $B'$ and $v'_0$ be the only vertex in
$B'\cap C$. For $i$ in $\{1,2,\ldots, 2r'\}$ and $j$ in $\{1,2\}$,
$v'_iw_j$ are edges in $G'(X)$. Similarly, there exists a vertex
$x'\in X$ with $d_G(x')=4$ and $\alpha(\Gamma_G(x'))=2$ such that
$|E(v_i, S_1(x'))|\geq 1$ and $|E(v_i, S_2(x'))|\geq 1$. We must
have $x=x'$; otherwise $x'\in N_G^7(x)$, i.e., $x' \in
\Gamma_{G^{\ast}}(x)$, and this contradicts the fact that $X$ is a
coloring class in $D$. Now we have $|E(\Gamma(x), B)|\geq 4r$ and
$E(\Gamma(x), B')|\geq 4r'$. By counting the degrees of vertices in
$\Gamma(x)$ in $G$, we have
$$4r+4r'+4+4\leq 16.$$
We get $r=r'=1$. Both $B$ and $B'$ are $K_3$'s. In this case,
$G/S_1(x)/S_2(x)$ contains the graph $G_0$, see figure
\ref{fig:H13}. This contradicts Lemma \ref{D41}.
We can find the desired contradiction; so the lemma
follows. \hfill $\square$
\section{Proof of Lemma \ref{D41}}
In this section, we will prove Lemma \ref{D41}. We first review a Lemma in \cite{lup}.
\begin{lemma}\label{cut2}
Let $G$ be a graph. Suppose that $G_1$ and $G_2$ are two subgraphs
such that $G_1 \cup G_2=G$ and $V(G_1) \cap V(G_2) =\{u,v\}$.
\begin{enumerate}
\item If $uv$ is an edge of $G$, then we have
$$\chi_f(G) = \max\{\chi_f(G_1), \chi_f(G_2)\}.$$
\item If $uv$ is not an edge of $G$, then we have
$$\chi_f(G) \leq \max\{\chi_f(G_1), \chi_f(G_2+uv), \chi_f(G_2/uv)\},$$
\end{enumerate}
where $G_2+uv$ is the graph obtained from $G_2$ by adding edge $uv$ and $G_2/uv$ is the graph obtained from $G_2$ by contracting $\{u, v\}$.
\end{lemma}
\noindent {\bf Proof of Lemma \ref{D41}:} Recall that $G$ is a
connected $K_4$-free graph with minimum number of vertices such that
$G \not = C_8^2$ and $ \chi_f(G) > 4-\frac{2}{67} $. Note that $G$
is 2-connected. We will prove it by contradiction.
Suppose Lemma \ref{D41} fails for some vertex $x$ in $G$. Observe
$\Gamma_G(x)$ is one of the graphs in Figure \ref{Delta=4}. Here we
assume $\Gamma_G(x)=\{a,b,c,d\}$. Through the proof of the lemma, le
$S_1$ and $S_2$ be two vertex-disjoint independent sets in
$\Gamma_G(x)$, $H$ be a triangle in $V(G) \setminus ( \{x\} \cup
\Gamma_G(x))$, then say $(S_1,S_2,H)$ is a bad triple if
$\{\underline {S_1}, \underline {S_2}, H\}$ contains a $K_5^-$ in
$G/S_1/S_2$.
\begin{figure}[htbp]
\centerline{\psfig{figure=image/c42.eps,width=0.18\textwidth} \hfil
\psfig{figure=image/c41.eps, width=0.18\textwidth} \hfil \psfig{figure=image/c43.eps, width=0.18\textwidth}
}
\centerline{$\Gamma_G(x)=P_4$ \hfil $\Gamma_G(x)=2~e$ \hfil $\Gamma_G(x)=C_4$}
\caption{Three possible cases of $\Gamma_G(x)$. } \label{Delta=4}
\end{figure}
If $\Gamma_G(x)=P_4$, then $\{a,d\}$ and $\{b,c\}$ is the only pair
of disjoint non-edges. There is a triangle $H$ with
$V(H)=\{y,z,w\}$ such that $(\{a,d\}, \{b,c\}, H)$ is a bad triple.
Note that $|E(\{a,b,c,d\},\{y,z,w\})|=5$ or $6$.
By an exhaustive search, the induced subgraph of $G$ on $\{x, a,b,c,d,y,z,w\}$
is one of the following six graphs (see Figure \ref{fig:P4}).
\begin{figure}[htbp]
\centerline{\psfig{figure=image/N25.eps, width=0.25\textwidth} \hfil
\psfig{figure=image/H2.eps, width=0.25\textwidth} \hfil
\psfig{figure=image/N22.eps, width=0.25\textwidth} }
\centerline{ \hspace*{0.05\textwidth} $H_{1}$ \hspace*{0.15\textwidth} \hfil $H_{2}$ \hspace*{0.15\textwidth} \hfil $H_{3}$ }
\vspace*{5mm}
\centerline{\psfig{figure=image/N23.eps,width=0.25\textwidth} \hfil
\psfig{figure=image/N24.eps, width=0.25\textwidth}
\hfil
\psfig{figure=image/N21.eps,
width=0.25\textwidth}}
\centerline{ \hspace*{0.05\textwidth}$H_{4}$ \hfil \hspace*{0.15\textwidth} $H_{5}$ \hfil \hspace*{0.15\textwidth} $H_{6}$ }
\caption{If $\Gamma_G(x)=P_4$, then there are six possible induced subgraphs.}
\label{fig:P4}
\end{figure}
If $\Gamma_G(x)=2~e$, then $(\{a,c\}, \{b,d\})$ and $(\{a,d\},
\{b,c\})$ are two pairs of disjoint non-edges. By considering the
degrees of vertices in $\Gamma_G(x)$, there is only one triangle
$H$ with $V(H)=\{y,z,w\}$ such that $(\{a,c\}, \{b,d\}, H)$ and
$(\{a,d\}, \{b,c\}, H)$ are two bad triples. By an exhaustive
search, the induced subgraph of $G$ on $\{x, a,b,c,d,y,z,w\}$ is one
of the following three graphs (see Figure \ref{fig:C4}).
\begin{figure}[htbp]
\centerline{ \psfig{figure=image/G1.eps, width=0.25\textwidth} \hfil \hfil \psfig{figure=image/G2.eps,
width=0.25\textwidth} \hfil \psfig{figure=image/G3.eps,width=0.25\textwidth} }
\centerline{ \hspace*{0.05\textwidth} $H_7$ \hspace*{0.15\textwidth} \hfil $H_8$ \hspace*{0.15\textwidth} \hfil $H_9$ }
\caption{If $\Gamma_G(x)=2\ e$, then there are three possible induced subgraphs.}
\label{fig:C4}
\end{figure}
It suffices to show that $G$ cannot contain $H_i$ for $1\leq i\leq
9$. Since all vertices in $H_1$ (and $H_2$) have degree 4, $H_1$
(and $H_2$) is the entire graph $G$. Observe that $H_1$ is
isomorphic to $C_8^2$ and $H_2$ is 11:3-colorable (see Figure
\ref{fig:H2H7}). Contradiction!
\begin{figure}[htbp]
\centering
\psfig{figure=image/N2224.eps, width=0.24\textwidth} \hfil
\psfig{figure=image/G11.eps, width=0.24\textwidth}
\centerline{$H_2$ \hspace*{1cm} \hfil $H_7$ }
\caption{The graph $H_2$ and $H_7$ are 11:3-colorable.}
\label{fig:H2H7}
\end{figure}
In $H_7$, the vertex $d$ is the only vertex with degree less than
$4$. If $H_7$ is not the entire graph $G$, then $d$ is a cut vertex
of $G$. This contradicts the fact that $G$ is $2$-connected. Thus
$G=H_7$. The graph $H_7$ is 11:3-colorable as shown by Figure
\ref{fig:H2H7}. Contradiction!
Now we consider the case $H_3$. Note $H_3+bz$ is the graph $H_2$. We
have $\chi_f(H_3)\leq \chi_f(H_2)\leq 11/3$. The graph $H_3$ must be
a proper induced subgraph of $G$, and the pair $\{b,z\}$ is a vertex
cut of $G$. Let $G'$ be the induced subgraph of $G$ by deleting all
vertices in $H_3$ but $b,z$. We apply Lemma \ref{cut2} to $G+bz$
with $G_1=H_3+bz=H_2$ and $G_2=G'+bz$. We have
$$\chi_f(G+bz)\leq \max\{\chi_f(H_2), \chi_f(G'+bz)\}.$$
Note $\chi_f(H_2)\leq 11/3$ and $11/3<\chi_f(G)\leq \chi_f(G+bz)$.
We have $\chi_f(G)\leq \chi_f(G'+bz)$. Both $b$ and $z$ have at most $2$ neighbors
in $G'+bz$. Thus $G'+bz$ is $K_4$-free; $G'+bz\not=C_8^2$
and has fewer vertices than $G$.
This contradicts to the minimality of $G$.
Note $H_5+cy=H_2$. The case $H_5$ is similar to the case $H_3$.
Note that $H_4$, $H_6$, and $H_8$ are isomorphic to each other.
It suffices to show $G$ does not contain $H_4$.
Suppose that $H_4$ is a proper induced subgraph of $G$.
Let $G_1$ be the induced subgraph of $G$ by deleting
all vertices in $H_4$. Note $C_8^2$ is not a proper subgraph of any graph in ${\cal G}_4$.
We have $G_1\not=C_8^2$.
Note that $c$ and $z$ have degree $3$ while other vertices in $H_4$ have degree $4$.
Since $G$ is 2-connected, $c$ has a unique neighbor, denoted by $u$,
in $V(G_1)$. Similarly, $z$ has a unique neighbor, denoted by $v$,
in $V(G_1)$. Observe that the pair $\{u,v\}$ forms a vertex cut of $G$.
Let $G_2$ be the induced graph of $G$ on $V(H_4)\cup \{u,v\}$.
Applying Lemma \ref{cut2} to $G$ with $G_1$ and $G_2$, we have
$$\chi_f(G) \leq \max\{\chi_f(G_1), \chi_f(G_2+uv), \chi_f(G_2/uv)\}.$$
Figure \ref{fig:H4} shows $\chi_f(G_2+uv)$ and $\chi_f(G_2/uv)$ are at most $11/3$.
\begin{figure}[htbp]
\centering
\psfig{figure=image/N23+uv.eps, width=0.24\textwidth} \hfil
\psfig{figure=image/N23-uv.eps, width=0.24\textwidth}
\centerline{$G_2+uv$ \hspace*{1cm} \hfil $G_2/uv$ }
\caption{Case $H_4$: both graph $G_2+uv$ and $G_2/uv$ are 11:3-colorable.}
\label{fig:H4}
\end{figure}
Since $\chi_f(G)> 11/3$, we have $\chi_f(G)\leq \chi_f(G_1)$.
Now $G_1$ is $K_4$-free and has maximum degree at most $4$; $G_1$ has
fewer vertices than $G$. This contradicts the minimality of $G$.
If $G=H_4$, then $\chi_f(H_4)\leq 11/3$, since $H_4$ is a subgraph of $G_2+uv$
in Figure \ref{fig:H4}.
Now we consider the last case $H_9$. First, we contract $b,c,z$ into a
fat vertex denoted by $\underline{bcz}$. We write $G/bcz$ for the graph after
this contraction. Observe that $\{\underline{bcz},d\}$ is a vertex-cut of
$G/bcz$. Let $G_4$ and $G_4'$ be two connected subgraphs of $G/bcz$
such that $G_4 \cup G_4'=G/bcz$, $G_4 \cap G_4'=\{\underline{bcz},d\}$, and
$\{u,v\} \subset G_4'$. Note that $G_4$ is $11$:$3$ colorable, see
Figure \ref{fig:H9}. Now by Lemma \ref{cut2}, we have
$$
\chi_f(G/bcz) \leq \max \{\chi_f(G_4), \chi_f(G_4')\}.
$$
As $\{b,c,z\}$ is an independent set, each $a$:$b$-coloring of
$G/bcz$ gives an $a$:$b$-coloring of $G$, that is $\chi_f(G/bcz)\geq
\chi_f(G)>11/3$. The graph $G_4$ is 11:3-colorable; see Figure
\ref{fig:H9}. Thus we have $\chi_f(G_4')\geq \chi_f(G/bcz)\geq
\chi_f(G)$. It is easy to check that $G_4'$ has maximum degree $4$,
$K_4$-free, and it is not $C_8^2$. Hence $G_4'$ must contain a
$K_4$. Otherwise, it contradicts the minimality of
$G$.
Second, we contract $\{b,d,z\}$ into a fat vertex $\underline{bdz}$
and denote the graph by $G/bcz$. Let $G_5$ and $G_5'$ be two
connected subgraphs of $G/bdz$ such that $G_5 \cup G_5'=G/bzd$, $G_5
\cap G_5'=\{\underline{bzd},c\}$, and $\{u,v\} \subset G_5'$. Note
that $G_5$ is 11:3-colorable; see Figure \ref{fig:H9}. By a similar
argument, $G_5'$ must contain a $K_4$.
\begin{figure}[htbp]
\centerline{ \psfig{figure=image/G31bcz.eps, width=0.33\textwidth} \hfill \psfig{figure=image/G31bdz.eps,
width=0.33\textwidth} \hfill \psfig{figure=image/G5.eps,width=0.33\textwidth}}
\centerline{$G_4$ \hspace{3.4cm} $G_5$ \hspace{3.4cm} $G_6$}
\caption{Case $H_9$: the graphs $G_4$, $G_5$, and $G_6$ are 11:3-colorable.}
\label{fig:H9}
\end{figure}
The remaining case is that both $G_4'$ and $G_5'$ have a $K_4$ when we contract $b$ and $z$.
Since the origin graph $G$ is $K_4$-free, the $K_4$ in $G_4'$ (and in
$G_5'$) must contain the fat vertex $\underline{bcz}$ (or
$\underline{bdz}$), respectively. Note that each of the four
vertices $b,c,d,z$ has at most one edge leaving $H_9$. There must be
a triangle $uvp$ in $G$ and these four outward edges are connected
to some element of $\{u,v,p\}$. The graph $G/{bz}$ must contain the
subgraph $G_6$ as drawn in Figure \ref{fig:H9}.
Note that $\{u,v\}$ is a vertex-cut in $G/ bz$. Let $G_6$ and $G_6'$
be two connected subgraphs of $G/bz,$ which satisfy $G_6 \cup
G_6'=G$, $G_6 \cap G_6'=\{u,v\}$, and $\underline{bz} \in G_6$. By
Lemma \ref{cut2}, we have
$$
\chi_f(G/bz) \leq \max\{\chi_f(G_6), \chi_f(G_6')\}.
$$
Note that $G_6$ is 11:3-colorable; see Figure \ref{fig:H9}. We also
have $\chi_f(G/bz) \geq \chi_f(G) >\frac{11}{3}$. We obtain
$\chi_f(G_6') \geq \chi_f(G/bz) \geq \chi_f(G)$. Observe that $G_6'$
is a subgraph of $G$. We arrive at a contradiction of the minimality
of $G$.
If $\Gamma_G(x)=C_4$, then the only possible choice for the two
independent sets are $\{a,c\}$ and $\{b,d\}$. If there is some
triangle $H$ such that $(\{a,c\},\{b,d\},H)$ is a bad triple, then
we have
$$
|E(\Gamma_G(x),H)| \geq 5.
$$
However, $|E(\Gamma_G(x),H)| \leq 4$. This is a contradiction. Thus
the lemma follows in this case.
We can select two
vertex disjoint non-edges $S_1$ and $S_2$ such that the graph
$G/S_1/S_2$ contains no $K_5^-$. For these particular $S_1$ and
$S_2$, if $G/S_1/S_2$ contains no $G_0$, then Lemma \ref{D41} holds.
Without loss of generality, we assume that $G/S_1/S_2$ does contain
$G_0$. Let $s_i=\underline{S_i}$ for $i=1,2$.
Observe that both $s_1$ and $s_2$ have
four neighbors $u,v,p,q$ other than $x$ in $G_0$. It follows
that $$|E(S_1\cup S_2, \{u,v,p,q\})|\geq 8.$$
On the one hand, we have
\begin{eqnarray*}
|E(G\mid_{S_1\cup S_2})| &=&\frac{1}{2}\left(\sum_{v\in S_1\cup S_2}d(v)
-|E(S_1\cup S_2, \{u,v,p,q\})| - 4\right)\\
&\leq& \frac{1}{2}(16- 8 - 4)\\
&=&2.
\end{eqnarray*}
On the other hand, $\alpha(\Gamma(x))=2$ implies $G\mid_{S_1\cup
S_2}$ contains at least two edges. Thus, we have $\Gamma_G(x)=2 ~
e$. Label the vertices in $\Gamma_G(x)$ by $a,b,c,d$ as in Figure
\ref{Delta=4}. We assume $ab$ and $cd$ are edges while $ac, bd, ad,
bc$ are non-edges. Observe that each vertex in $\{u,v,p,q\}$ has
exactly two neighbors in $\{a,b,c,d\}$.
If one vertex, say $u$, has two neighbors forming a non-edge, say $ac$, then
we can choose $S'_1=\{a,c\}$ and $S_2'=\{b,d\}$. It is easy to check that
$G/S_1'/S_2'$ contains neither $G_0$ nor $K_5^-$. We are done in this case.
In the remaining case, we can assume that for each vertex $y$ in
$\{u,v,p,q\}$, the neighbors of $y$ in $\{a,b,c,d\}$ always form an
edge. Up to relabeling vertices, there is only one arrangement for
edges between $\{u,v,p,q\}$ and $\{a,b,c,d\}$; see the graph
$H_{10}$ defined in Figure \ref{fig:H14}.
\begin{figure}[htbp]
\centerline{\psfig{figure=image/H14.eps, width=0.28\textwidth} \hfil
\psfig{figure=image/H141.eps, width=0.32\textwidth}}
\caption{$H_{10}$ and an 11:3-coloring of $H_{10}$.} \label{fig:H14}
\end{figure}
The graph $H_{10}$ is 11:3-colorable as shown in Figure
\ref{fig:H14}. Since $\chi_f(G)>11/3$, $H_{10}$ is a proper subgraph
of $G$.
Note in $H_{10}$, every vertices except $w$ and $r$
has degree $4$; both $w$ and $r$ have degree $3$. Thus, $\{w,r\}$ is
a vertex cut of $G$. Let $G_1=H_{10}$ and $G_2$ be the subgraph of
$G$ by deleting vertices in $\{x,a,b,c,d,p,q,u,v\}$. Applying Lemma
\ref{cut2} with $G_1$ and $G_2$ defined above, we have
$$
\chi_f(G) \leq \max \{\chi_f(G_1),\chi_f(G_2)\}.
$$
Since $\chi_f(G)> 11/3$ and $\chi_f(G_1) \leq 11/3 $ (see Figure
\ref{fig:H14}), we must have $\chi_f(G_2) \geq \chi_f(G)$. Note
that $G_2$ has fewer number of vertices than $G$. This contradicts
the minimality of $G$. Therefore, the lemma follows. \hfill
$\square$
|
\section{Introduction}
A wide literature is dedicated to development of numerical
algorithms for solving the global optimization problem (see, for
example, references given in \cite{Horst&Pardalos(1995)}). The
problem may be formulated as
\be
f^*=f(x^*)=\min_{x\in \mathcal{F} } f(x), \ \ \mathcal{F} \subset \mathbb{R}
^N,\label{glop}
\ee
where $f(x)$ is a multiextremal and possibly non-differentiable
function and $\mathcal{F}$ is a compact set.
One of the approaches to studying and verifying validity of
numerical algorithms is their comparison on test problems (see,
e.g., \cite{Ali:et:al.(2003)}, \cite{Dixon&Szego(1978)},
\cite{Facchinei:et:al.(1997)}, \cite{Floudas&Pardalos(1990)},
\cite{Floudas:et:al(1999)}, \cite{Gaviano&Lera(1998)},
\cite{Horst&Pardalos(1995)}, \cite{Kalantari&Rosen(1986)},
\cite{Khoury:et:al.(1993)}, \cite{Li&Pardalos(1992)},
\cite{Locatelli(2003)}, \cite{More:et:al.(1981)},
\cite{Moshirvaziri(1994)}, \cite{Moshirvaziri:et:al(1996)},
\cite{Pardalos(1987)}; \cite{Pardalos(1991)},
\cite{Pinter(2002)}, \cite{Schittkowski(1980)};
\cite{Schittkowski(1987)}, \cite{Schoen(1993)},
\cite{Sung&Rosen(1982)}). Many global optimization tests were
taken from real-life problems and for this reason comprehensive
information about them is not available. The number of local
minima may be unknown, as well as their locations, regions of
attraction, and even values (including that of the global
minimum).
Recently \cite{Gaviano&Lera(1998)} introduced two types of
functions with a priori known local minima and their regions of
attraction. The tests proposed take a convex quadratic function
(called hereafter `paraboloid') systematically distorted by cubic
polynomials and by quintic polynomials to introduce local minima
and to construct test functions that are continuously
differentiable in some region $\Omega \supseteq \mathcal{F}$
(called hereafter `D-type' test functions) and twice continuously
differentiable in $\Omega \supseteq \mathcal{F}$ (called hereafter
`D2-type' test functions), where $\mathcal{F}$ is
from~(\ref{glop}) and $\Omega$ is a hyperrectangle.
To define a function of one of these types it is necessary to
determine a number of correlated parameters. Unfortunately, the
correlations do not allow simple and fast generation of the test
functions. Additionally, generation of different functions having
similar properties becomes difficult and non-intuitive when
dimension and/or number of local minima increase.
In this paper, in addition to the two types of test functions from
\cite{Gaviano&Lera(1998)}, the third type of non-differentiable
test functions (called hereafter `ND-type') is presented and a
generator for these three types of test functions is proposed. The
software to be introduced generates classes of test functions and
provides procedures for calculating the first order derivatives of
the D-type test functions and the first and second order
derivatives of the D2-type test functions.
Each class contains 100 functions and is defined by the following
parameters (the only ones to be determined by the user):
\begin{enumerate}
\item problem dimension;
\item number of local minima;
\item value of the global minimum;
\item radius of the attraction region of the global minimizer;
\item distance from the global minimizer to the vertex of the
paraboloid.
\end{enumerate}
The other necessary parameters (i.e., locations of all minimizers,
their regions of attraction, and values of minima) are chosen
randomly by the generator. After generation a special notebook
containing a complete description of all the functions from the
generated class is supplied to the user.
The rest of the paper is structured as follows. In
Section~\ref{sectionMatDescr}, a mathematical description of the
three types of test functions is given.
Section~\ref{sectionDescr}~introduces the generator and details of
its implementation. Section~\ref{sectionUsing}~is devoted to usage
of the generator.
\section{Mathematical description} \label{sectionMatDescr}
In this section, the three types of test functions are briefly
described. Let us start with the D-type and D2-type functions
(see~\cite{Gaviano&Lera(1998)}). A function $f(x)$ of the D-type
is determined over an admissible region $\Omega \supseteq
\mathcal{F}$, where $\mathcal{F}$ is from~(\ref{glop}) and
\be
\Omega=[a,b\,]=\{x \in \mathbb{R} ^N \, : \, a \leq x \leq
b \}, \hspace{3mm} a < b, \hspace{2mm} a,b \in \mathbb{R}^N.
\label{D}
\ee
The function is constructed by modifying a paraboloid $Z$:
\be
Z:\ g(x)= \| x-T \| ^2 + \, t,\ \ x \in \Omega, \label{Z}
\ee
(hereafter $\| \cdot \|$ denotes the Euclidean norm) with the
minimum $t$ at a point \mbox{$T \in {\rm int}(\Omega)$} in such a
way that the resulting function $f(x)$ has $m$, $m \geq 2$, local
minimizers: point~$T$ from (\ref{Z}) (we denote it by $M_1 := T$)
and points
\be
M_i \in {\rm int}(\Omega), \ \ M_i \neq T, \ M_i \neq M_j, \ \ i,j=2,\ldots,m,\ i \neq
j. \label{M(i)}
\ee
The paraboloid $Z$ from~(\ref{Z}) is modified by a function
$C_i(x)$, which is constructed by using cubic polynomials within
balls $S_i \subset \Omega$ around each point $M_i$,
$i=2,\ldots,m$, where
\be
S_i = \{x \in \mathbb{R} ^N : \, \| x - M_i \| \ \leq
\rho_i,\, \rho_i > 0\}, \ \ i= 1, \ldots, m. \label{S(i)}
\ee
Functions $Q_i(x)$, $i=2,\ldots,m$, use quintic polynomials to
determine the D2-type test functions.
Selection of radii $\rho_i$, $i=1,\ldots,m$, is carried out in
such a manner that sets $S_i$ from~(\ref{S(i)}) do not overlap:
\be
S_i \cap S_j = \varnothing,\ \ i,j=1,\ldots,m,\ i \neq j. \label{S(i)xS(j)}
\ee
It is not required that each attraction region $S_i$, $i=1,
\ldots, m$, be entirely contained in $\Omega$. Note that we use
the notation ``attraction region'' with respect to the balls
$S_i$, $i=1,\ldots, m$, just for simplicity. Naturally, definition
of the real attraction region for each local minimizer will depend
on the method used for optimization and will change from one
algorithm to another.
Formally, D-type functions \cite{Gaviano&Lera(1998)} are described
as follows:
\be
f(x)=\left\{
\begin{array}{ll}
C_i(x), & x\in S_i, \, i \in \{2,\ldots,m\} , \\
g(x), & x \notin S_2 \cup \ldots \cup S_m
\, ,
\end{array}
\right. \label{f_cubic}
\ee
where $g(x)$ is from~(\ref{Z}), sets $S_i$, $i=2,\ldots,m$,
from~(\ref{S(i)}) satisfy~(\ref{S(i)xS(j)}), and
$$
C_i(x) =\left ( \frac{2}{\rho_i^2} \frac{<\!x-M_i,\, T-M_i\!>}{\| x-M_i \| } - \frac{2}{\rho_i^3}A_i \right) \| x-M_i\| ^3 +
$$
\be
+ \left( 1- \frac{4}{\rho_i}\frac{<\!x-M_i,\, T-M_i\!>}{\| x-M_i \| }+\frac{3}{\rho_i^2}A_i \right) \| x-M_i\| ^2 + f_i. \label{C_cubic}
\ee
In~(\ref{C_cubic}) radii $\rho_i$, $i=2,\ldots,m$, determine the
sets $S_i$ from~(\ref{S(i)}), $<\! \cdot,\cdot \!>$ denotes the
usual scalar product, and the values $A_i$, $i=2,\ldots,m$, are
found as
\be
A_i= \| T-M_i\| ^2 + t - f_i, \label{A}
\ee
where $f_1 = t$ and $f_i$, $i=2,\ldots,m$, are the function values
at local minimizers $M_i$:
\be
f_i = \min \{g(x) \, : x \in B_i\} - \gamma_i,\ \ \gamma_i > 0,
\label{f(i)}
\ee
where $B_i$ is the boundary of the ball $S_i$:
\be
B_i=\{x \in \mathbb{R} ^N : \, \| x - M_i \| \ = \rho_i, \, \rho_i > 0\}, \ \ i= 2, \ldots,
m, \label{B(i)}
\ee
and $\gamma_i$ is a parameter ensuring that the value $f_i$ is
less than the minimum of the paraboloid $Z$ from~(\ref{Z}) over
$B_i$.
Analogously, D2-type functions \cite{Gaviano&Lera(1998)} are
defined by
\be
f(x)=\left\{
\begin{array}{ll}
Q_i(x), & x\in S_i, \, i \in \{2,\ldots,m\} , \\
g(x), & x \notin S_2 \cup \ldots \cup S_m
\, ,
\end{array}
\right. \label{f_quintic}
\ee
where
\begin{center}
$$
Q_i(x) =\left [-\frac{6}{\rho_i^4} \frac{<\!x-M_i,\, T-M_i\!>}{\| x-M_i \| } + \frac{6}{\rho_i^5}A_i +
\frac{1}{\rho_i^3}(1-\frac{\delta}{2})\right]
\| x-M_i\| ^5 +
$$
$$
\left [\frac{16}{\rho_i^3} \frac{<\!x-M_i,\, T-M_i\!>}{\| x-M_i \| } - \frac{15}{\rho_i^4}A_i -
\frac{3}{\rho_i^2}(1-\frac{\delta}{2})\right]
\| x-M_i\| ^4 +
$$
$$
\left [-\frac{12}{\rho_i^2} \frac{<\!x-M_i,\, T-M_i\!>}{\| x-M_i \| } + \frac{10}{\rho_i^3}A_i +
\frac{3}{\rho_i}(1-\frac{\delta}{2})\right]
\| x-M_i\| ^3 +
$$
\be
\frac{1}{2}\delta \| x-M_i\| ^2 + f_i \label{Q_quintic}
\ee
\end{center}
with $A_i$ and $f_i$, $i=2,\ldots,m$, from~(\ref{A})
and~(\ref{f(i)}), and $\delta$ is an arbitrary positive real
number (see \cite[Lemma 3.1]{Gaviano&Lera(1998)}).
The properties of these functions have been studied by
\cite{Gaviano&Lera(1998)}. In particular, the following results
can be proved:
i. D-type functions~(\ref{f_cubic})--(\ref{C_cubic}) are
continuously differentiable in~$\Omega$~\cite[Lemma
2.1]{Gaviano&Lera(1998)}.
ii. D2-type functions~(\ref{f_quintic})--(\ref{Q_quintic}) are
twice continuously differentiable in~$\Omega$ \cite[Lemma
3.1]{Gaviano&Lera(1998)}.
Let us now describe the ND-type test functions, which are
continuous in $\Omega$ but non-differentiable in the whole region
$\Omega$. An analogous procedure is considered: the paraboloid $Z$
from~(\ref{Z}) is modified by a function $P_i(x)$ constructed from
second degree polynomials within each region $S_i \subset \Omega$
from~(\ref{S(i)}) in such a way that the resulting function $f(x)$
is continuous in the feasible region $\Omega$ from~(\ref{D}),
differentiable at each local minimizer $M_i$, $i=2,\ldots, m$,
from~(\ref{M(i)}), but generally non-differentiable at the points
of the boundaries $B_i$ of the balls $S_i$, $i=2, \ldots, m$,
determined by~(\ref{B(i)}). That is,
\be
f(x)=\left\{
\begin{array}{ll}
P_i(x), & x\in S_i, \, i \in \{2,\ldots,m\} , \\
g(x), & x \notin S_2 \cup \ldots \cup S_m
\, ,
\end{array}
\right. \label{f_nonsmooth}
\ee
where $g(x)$ is from~(\ref{Z}), sets $S_i$, $i=2,\ldots,m$,
from~(\ref{S(i)}) satisfy~(\ref{S(i)xS(j)}), and
\be
P_i(x) =\left ( 1 -\frac{2}{\rho_i} \frac{<\!x-M_i,\, T-M_i\!>}{\| x-M_i \| } + \frac{1}{\rho_i^2}A_i \right) \| x-M_i\| ^2 + f_i.
\label{P_quadratic}
\ee
In~(\ref{P_quadratic}) the values $\rho_i$, $A_i$, and $f_i$
($i=2,\ldots,m$) are determined in the same way as for the D- and
D2-type functions by formulae~(\ref{S(i)})--(\ref{S(i)xS(j)}),
(\ref{A}), and~(\ref{f(i)}), respectively.
\section{Generation of tests classes} \label{sectionDescr}
As one can see from the previous section, all three function types
have many parameters to be coordinated. Moreover, their
characteristics (for example, the mutual positions of the local
minimizers, the global minimizer, and the paraboloid vertex; the
size of the attraction regions of local minimizers; the function
values at local minima) influence the properties of the test
functions significantly from the point of view of global
optimization algorithms. For example, coincidence of the global
minimizer with the paraboloid vertex leads to generation of too
simple functions. Existence of many deep minima having narrow
regions of attraction can lead to the impossibility of global
minimizer location even by the most ``intelligent'' global
optimization algorithms. All these features should be added to the
general scheme from Section~\ref{sectionMatDescr} in order to
obtain well-structured test classes.
In the generator, the user sets just a few parameters defining a
desirable class while all the other parameters are chosen
randomly. The generator is also employed in maintaining conditions
distinguishing each class -- for example, the distance of the
global minimizer from the minimizer of the paraboloid, dependence
of the local minima values on the attraction regions sizes, etc.
Thus, the generator gives the researcher the ability to construct
classes of 100 test functions of arbitrary dimension with
arbitrary number of local minima.
This section describes how a class consisting of D-type test
functions is generated. Classes consisting of D2-type and ND-type
functions are constructed analogously.
Each test class generated by the introduced software contains 100
test functions $f(x)$ and is defined by the following parameters
to be fixed by the user:
\begin{enumerate}
\item the problem dimension $N$, $N \geq 2$;
\item the number of local minimizers $m$, $m \geq 2$, including the
minimizer $T$ for the paraboloid~(\ref{Z}) (all the minimizers
are chosen randomly);
\item the global minimum value $f^*$, the same for all the
functions of the class;
\item the radius $\rho ^*$ of the attraction region of the global
minimizer $x^*$;
\item the distance $r^*$ from the paraboloid vertex $T$ to the global
minimizer \mbox{$x^*\in \Omega$} (whose coordinates are also chosen
randomly).
\end{enumerate}
By changing these parameters the user can create classes with
different properties.
Each function of a test class is specified by its number~$n$, $1
\leq n \leq 100$. The other parameters of the functions
from~(\ref{Z})--(\ref{P_quadratic}) are chosen randomly by means
of the random number generator proposed in~\cite{Knuth(1997)}.
The input parameters $f^*$, $r^*$, and $\rho ^*$ must be chosen
in such a way that the following simple conditions are satisfied:
\be
f^* < t \label{f*}
\ee
(which means that the global minimizer is not a vertex of the
paraboloid; this requirement allows us to avoid too simple
functions with a global minimum at the vertex of the paraboloid
$Z$ from~(\ref{Z})),
\be
0 < r ^* < 0.5\min_{1 \leq j \leq N} | b(j) - a(j) | \label{r_global_cond}
\ee
(i.e., the global minimizer $x^*$ belongs to the admissible region
$\Omega$ even in the case when the paraboloid vertex $T$ is at the
center of $\Omega$), and
\be
0 < \rho ^* \leq 0.5r^*. \label{rho_global}
\ee
Note that it is not required that each attraction region $S_i$,
$i=1, \ldots, m$, from~(\ref{S(i)}) entirely belongs to $\Omega$.
The admissible region $\Omega$ is taken as $\Omega=[-1,1]^N$ and
the minimal value of the paraboloid~(\ref{Z}) is fixed at $t=0$ by
default (naturally, these parameters can be changed by the user).
Let us discuss in more detail the random procedure generating
parameters for test functions. (The unique difference for the
D2-type is that the parameter $\delta$ from~(\ref{Q_quintic}) is
required; this parameter is chosen randomly from the open interval
$(0,\Delta)$, where $\Delta$ is a positive number taken by default
$\Delta=10$.) Hereafter the vertex $T$ from~(\ref{Z}) in the
set~(\ref{M(i)}) of local minimizers has the index 1, $M_1 := T$,
and the global minimizer $x^*$ has the index 2, $M_2 := x^*$.
Naturally, among the minimizers $M_i$, $i=3, \ldots, m$, another
global minimizer $y^* \neq x^*$ can be generated.
First, coordinates of the paraboloid vertex $T$, coordinates of
the global minimizer~$x^*$, and coordinates of the remaining local
minimizers $M_i$ (controlling the satisfaction of~(\ref{M(i)}))
are chosen randomly. Then, the attraction regions radii $\rho_i$,
$i \neq 2$, from~(\ref{S(i)}) are determined: to do this the
attraction regions of each local minimizer from~(\ref{M(i)}) ($i
\neq 2$ because the attraction region of the global minimizer is
fixed: $\rho_2 = \rho^*$) are expanded until
condition~(\ref{S(i)xS(j)}) is not violated. Finally, values of
the function $f(x)$ at local minima $M_i$, $i=3, \ldots, m$, are
fixed by choosing random values $\gamma_i$, $i=3, \ldots, m$,
from~(\ref{f(i)}) (recall that $f_1=t$ and $f_2 = f^*$).
Let us consider these three principal operations in detail.
Coordinates of the local minimizers $M_i$, $i=3, \ldots, m$,
from~(\ref{M(i)}), coordinates of the vertex $T$ of the
paraboloid~(\ref{Z}), and location of the global minimizer $x^*$
are chosen randomly at the intersection of $\Omega$ and the sphere
of radius $r^*$ with a center at~$T$ so that~(\ref{M(i)}) is
satisfied. For the positioning of $x^*$ we use generalized
spherical coordinates
$$
x^*_j := T_j + r^* \cos \phi _j \prod_{k=1}^{j-1}\sin \phi _k, \ \
j=1,\ldots,N-1,
$$
\be
x^*_N := T_N + r^* \prod_{k=1}^{N-1}\sin \phi _k, \label{x_polar}
\ee
where the components of the vector
$$
\phi = (\phi _1, \ldots, \phi _N) \in \Phi = \{0 \leq \phi_1 \leq \pi;
\ 0 \leq \phi_j \leq 2\pi, \ j=2,\ldots,N \}
$$
are chosen randomly. In this case, if some $x^*_k \notin \Omega$,
$1 \leq k \leq N$, this coordinate is redefined as
$$
x^*_k := 2 T_k - x^*_k.
$$
After selection of coordinates of the paraboloid vertex $T$ and of
the global minimizer $x^*$, coordinates of the points $M_i$, $i=3,
\ldots, m$, are generated in such a way that beside
condition~(\ref{M(i)}) the condition
\be
\| M_i - x^* \| - \rho ^* = \zeta, \ \ \zeta > 0
\label{gap}
\ee
is satisfied with some positive parameter $\zeta$. This condition
follows from~(\ref{S(i)xS(j)}) and does not allow the local
minimizers to be very close to the attraction region of the global
minimizer $x^*$. Thus, in~(\ref{gap}) the parameter $\zeta$ should
not be too small. The value $\zeta = \rho^*$ is chosen by default.
The next step of the test function construction sets attraction
regions. Each value $\rho_i$, $i \neq 2$, from~(\ref{S(i)}) is
initially calculated as half of the minimum distance between the
minimizer $M_i$ and the remaining local minimizers
$$
\rho_i := 0.5 \min_{1 \leq j \leq m,\ j \neq i} \| M_i-M_j\| , \ \ i=1,\ldots,m, \ i \neq 2,
$$
\be \label{rho(i)init}
\rho_2 := \rho ^*
\ee
(in such a way that the attraction regions from~(\ref{S(i)}) do
not overlap). Then, an attempt to increase the values $\rho_i$,
$i=1,\ldots,m$, $i \neq 2$ (i.e., an attempt to enlarge the
attraction regions) is made:
\be
\rho _i := \max \left( \rho_i,\min_{1 \leq j \leq m, \ j \neq
i} \{ \| M_i-M_j\| - \rho_j \}\right), \ \
i=1,\ldots,m, \ i \neq 2. \label{rho(i)expand}
\ee
Because of the recursive character of
formulae~(\ref{rho(i)expand}), an expansion of the attraction
regions depends on the order in which these regions are selected
(an ascending order of the indices is chosen).
Finally, the values of the radii $\rho_i$ are corrected by the
weight coefficients $w_i$:
$$
\rho_i := w_i\,\rho_i, \ \ i=1, \ldots, m,
$$
where $0 < w_i \leq 1$, $i=1,\ldots, m$, and the values $w_i$ are
chosen by default as
\be
w_i = 0.99, \ \ i=1, \ldots, m, \ i \neq 2, \hspace{7mm} {\rm
and} \hspace{7mm} w_2 = 1. \label{rho(i)weight}
\ee
At the last step the function values $f_i$, $i=3, \ldots, m$, at
the local minima are generated by using formula~(\ref{f(i)}),
where $\gamma_i$ must be specified. Each value $\gamma_i$, $i=3,
\ldots, m$, is chosen (note that the values $\gamma_1$ and
$\gamma_2$ are not considered because the function values $f_1=t$
at the paraboloid vertex and $f_2=f^*$ at the global minimizer
have been fixed by the user without using~(\ref{f(i)})) as the
minimum of two values generated randomly from the open intervals
$(\rho_i, 2\rho_i)$ and $(0, Z_{B_i} - f^*)$, where $Z_{B_i}$ is
the minimum of the paraboloid $Z$ from~(\ref{Z}) over $B_i$
from~(\ref{B(i)}). In such a way, the values $f_i$ in~(\ref{f(i)})
depend on radii $\rho_i$ of the attraction regions $S_i$, $i=3,
\ldots, m$, and at the same time the following condition is
satisfied:
$$
f^* \leq f_i,\ \ i=3, \ldots, m.
$$
Note that dependence of the function values at local minima on the
radii of the attraction regions is not respected by the global
optimum value $f_2=f^*$ because the user defines the function
value at the global minimizer and the radius $\rho ^*$ of its
region of attraction directly when choosing the corresponding test
class.
\begin{figure}[t]
\centerline{\psfig{file=Fig1.eps,width=110mm,height=90mm,angle=0,silent=}}
\caption{The function number~9 from a class of two-dimensional
D-type test functions with~10 local minima} \label{fig1}
\end{figure}
Figure~\ref{fig1} shows an example of the D-type test function.
This function is defined in the region $\Omega=[-1,1]^2$ and is
number~9 in the class of D-type functions with the following
parameters:
\begin{enumerate}
\item dimension $N=2$;
\item number of local minima $m=10$;
\item value of the global minimum $f^*=-1$;
\item radius of the attraction region of the global minimizer $\rho^*=\frac{1}{3}$;
\item distance from the global minimizer $x^*$ to the vertex $T$ of the
paraboloid from~(\ref{Z}) is $r^*=\frac{2}{3}$.
\end{enumerate}
The generated global minimizer of this function is $x^*=(-0.911,
0.989)$ and the paraboloid minimizer is $T=(-0.711, 0.353)$.
\section{Usage of the test classes generator} \label{sectionUsing}
The generator package has been written in ANSI Standard C and
successfully tested on Windows and UNIX platforms. Our
implementation follows the procedure described in
Section~\ref{sectionDescr}. First, the general structure of the
package is described, then instructions for using the test classes
generator (called hereafter GKLS-generator) are given.
\subsection{Structure of the package}
The package includes the following files:
\begin{description}
\item[\bf gkls.c] -- the main file;
\item[\bf gkls.h] -- the header file that users should include in
their application projects in order to call subroutines from
the file {\bf gkls.c};
\item[\bf rnd\_gen.c] -- the file containing the uniform random number
generator proposed in Knuth~[\cite{Knuth(1997)};
\cite{Knuth:HomePage}];
\item[\bf rnd\_gen.h] -- the header file for linkage to the file {\bf
rnd\_gen.c};
\item[\bf example.c] -- an example of the GKLS-generator usage;
\item[\bf Makefile] -- an example of a UNIX makefile provided to
UNIX users for a simple compilation and linkage of separate files
of the application project.
\end{description}
For implementation details the user can consult the C codes. Note
that the random number generator in {\bf rnd\_gen.c} uses the
logical-and operation `\&' for efficiency, so it is not strictly
portable unless the computer uses two's complement representation
for integer. It does not limit portability of the package because
almost all modern computers are based on two's complement
arithmetic.
\subsection{Calling sequence for generation and usage of the tests
classes}
Here we describe how to generate and use classes of the ND-, D-,
and D2-type test functions. Again, we concentrate on the D-type
functions. The operations for the remaining two types are
analogous.
To utilize the GKLS-generator the user must perform the following
steps:
\begin{description}
\item[\bf Step 1.] Input of the parameters defining a specific test class.
\item[\bf Step 2.] Generating a specific test function of the defined test
class.
\item[\bf Step 3.] Evaluation of the generated test function and, if
necessary, its partial derivatives.
\item[\bf Step 4.] Memory deallocating.
\end{description}
Let us consider these steps in turn.
\subsubsection{Input of the parameters defining a specific test
class} \label{sectionInput}
This step is subdivided into: (a) defining the parameters of the
test class, (b) defining the admissible region $\Omega$, and (c)
checking (if necessary).
-- {\it (a) Defining the parameters of the test class}. The
parameters to be defined by the user determine a specific class
(of the ND-, D- or D2-type) of 100 test functions (a specific
function is retrieved by its number). There are the following
parameters:
\begin{description}
\item[{\it GKLS\_dim}] -- ({\bf unsigned int}) dimension $N$
(from~(\ref{glop})) of test functions; $N \geq 2$ (since
multidimensional problems are considered in~(\ref{glop})) and
$N<$ NUM\_RND in {\bf rnd\_gen.h}; this value is limited by the
power of \mbox{{\bf unsigned int}}-representation; default
$N=2$;
\item[{\it GKLS\_num\_minima}] -- ({\bf unsigned int}) number $m$
(from~(\ref{M(i)})) of local minima including the paraboloid $Z$
minimum (from~(\ref{Z})) and the global minimum; $m \geq 2$; the
upper bound of this parameter is limited by the power of
\mbox{{\bf unsigned int}}-representation;
default $m=10$;
\item[{\it GKLS\_global\_value}] -- ({\bf double}) global minimum value $f^*$ of
$f(x)$; condition~(\ref{f*}) must be satisfied; the default
value is $-1.0$ (defined in the file {\bf gkls.h} as a constant~GKLS\_GLOBAL\_MIN\_VALUE);
\item[{\it GKLS\_global\_dist}] -- ({\bf double}) distance $r^*$ from the paraboloid
vertex $T$ in~(\ref{Z}) to the global minimizer $x^* \in \Omega$
of $f(x)$; condition~(\ref{r_global_cond}) must be
satisfied; the default value is
$$
{\rm GKLS\_global\_dist}\ \stackrel{\rm def}{=} \ \min_{1 \leq j
\leq N} | b(j) - a(j) | /\, 3,
$$
where the vectors $a$ and $b$ determine the admissible region
$\Omega$ in~(\ref{D});
\item[{\it GKLS\_global\_radius}] -- ({\bf double}) radius $\rho ^*$ of the attraction
region of the global
minimizer $x^*\in \Omega$ of $f(x)$; condition~(\ref{rho_global}) must be
satisfied; the default value is
$$
{\rm GKLS\_global\_radius}\ \stackrel{\rm def}{=} \ \min_{1 \leq
j \leq N} | b(j) - a(j) | /\, 6.
$$
\end{description}
The user may call subroutine \mbox{{\it GKLS\_set\_default}()} to
set the default values of these five variables.
-- {\it (b) Defining the admissible region} $\Omega$. With $N$
determined, the user must allocate dynamic arrays {\it
GKLS\_domain\_left} and {\it GKLS\_domain\_right} to define the
boundary of the hyperrectangle $\Omega$. This is done by calling
subroutine
{\bf int} {\it GKLS\_domain\_alloc} ();\\
which has no parameters and returns the following error codes
defined in~{\bf gkls.h}:
\begin{description}
\item[\bf GKLS\_OK] -- no errors;
\item[\bf GKLS\_DIM\_ERROR] -- the problem dimension is out of range;
it must be greater than or equal to 2 and less than NUM\_RND defined
in {\bf rnd\_gen.h};
\item[\bf GKLS\_MEMORY\_ERROR] -- there is not enough memory to
allocate.
\end{description}
The same subroutine defines the admissible region $\Omega$. The
default value $\Omega =[-1,1]^N$ is set by \mbox{{\it
GKLS\_set\_default}()}.
-- {\it (c) Checking}. The following subroutine allows the user to
check validity of the input parameters:
{\bf int} {\it GKLS\_parameters\_check} ().\\
It has no parameters and returns the following error codes (see
{\bf gkls.h}):
\begin{description}
\item[\bf GKLS\_OK] -- no errors;
\item[\bf GKLS\_DIM\_ERROR] -- problem dimension error;
\item[\bf GKLS\_NUM\_MINIMA\_ERROR] -- number of local minima error;
\item[\bf GKLS\_BOUNDARY\_ERROR] -- the admissible region boundary
vectors are ill-defined;
\item[\bf GKLS\_GLOBAL\_MIN\_VALUE\_ERROR] -- the global minimum
value is not less than the paraboloid~(\ref{Z}) minimum value $t$
defined in {\bf gkls.h} as a constant GKLS\_PARABOLOID\_MIN;
\item[\bf GKLS\_GLOBAL\_DIST\_ERROR] -- the parameter $r^*$ does not
satisfy~(\ref{r_global_cond});
\item[\bf GKLS\_GLOBAL\_RADIUS\_ERROR] -- the parameter $\rho^*$ does not
satisfy~(\ref{rho_global}).
\end{description}
\subsubsection{Generating a specific test function of the defined
test class}
After a specific test class has been chosen (i.e., the input
parameters have been determined) the user can generate a specific
function that belongs to the chosen class of 100 test functions.
This is done by calling subroutine
{\bf int} {\it GKLS\_arg\_generate} ({\bf unsigned int} {\it nf}); \\
where
\begin{description}
\item[\it nf] -- the number
of a function from the test class (from~1~to~100).
\end{description}
This subroutine initializes the random number generator, checks
the input parameters, allocates dynamic arrays, and generates a
test function following the procedure of
Section~\ref{sectionDescr}. It returns an error code that can be
the same as for subroutines {\it GKLS\_parameters\_check}() and
{\it GKLS\_domain\_alloc}(), or additionally:
\begin{description}
\item[\bf GKLS\_FUNC\_NUMBER\_ERROR] -- the number of a test function to
generate exceeds 100 or it is less than 1.
\end{description}
{\it GKLS\_arg\_generate}() generates the list of all local minima
and the list of the global minima as parts of the structures {\it
GKLS\_minima} and {\it GKLS\_glob}, respectively. The first
structure gathers the following information about all local minima
(including the paraboloid minimum and the global one): coordinates
of local minimizers, local minima values, and attraction regions
radii. The second structure contains information about the number
of global minimizers and their indices in the set of local
minimizers. It has the following fields:
\begin{description}
\item[\it num\_global\_minima] -- ({\bf unsigned int}) total number of
global minima;
\item[\it gm\_index] -- ({\bf unsigned int *}) list of indices of generated
minimizers, which are the global ones (elements 0 to ($\mbox{\it
num\_global\_minima} - 1$) of the list) and the local ones
(the remaining elements of the list).
\end{description}
The elements of the list {\it GKLS\_glob}.{\it gm\_index} are
indices to a specific minimizer in the first structure {\it
GKLS\_minima} characterized by the following fields:
\begin{description}
\item[\it local\_min] -- ({\bf double **}) list of local minimizers
coordinates;
\item[\it f] -- ({\bf double *}) list of local minima values;
\item[\it rho] -- ({\bf double *}) list of attraction regions radii;
\item[\it peak] -- ({\bf double *}) list of parameters $\gamma_i$ values
from~(\ref{f(i)});
\item[\it w\_rho] -- ({\bf double *}) list of parameters $w_i$ values
from~(\ref{rho(i)weight}).
\end{description}
The fields of these structures can be useful if one needs to study
properties of a specific generated test function more deeply.
\subsubsection {Evaluation of a generated test function or its partial
derivatives}
While there exists a structure {\it GKLS\_minima} of local minima,
the user can evaluate a test function (or partial derivatives of
D- and D2-type functions) that is determined by its number (a
parameter to the subroutine {\it GKLS\_arg\_generate}()) within
the chosen test class. If the user wishes to evaluate another
function within the same class he should deallocate dynamic arrays
(see the next subsection) and recall the generator {\it
GKLS\_arg\_generate}() (passing it the corresponding function
number) without resetting the input class parameters (see
subsection~\ref{sectionInput}). If the user wishes to change the
test class properties he should reset also the input class
parameters.
Evaluation of an ND-type function is done by calling subroutine
{\bf double} {\it GKLS\_ND\_func} ({\it x}).\\
Evaluation of a D-type function is done by calling subroutine
{\bf double} {\it GKLS\_D\_func} ({\it x}).\\
Evaluation of a D2-type function is done by calling subroutine
{\bf double} {\it GKLS\_D2\_func} ({\it x}).\\
All these subroutines have only one input parameter
\begin{description}
\item[\it x] -- ({\bf double *}) a point $x \in \mathbb{R} ^N$
where the function must be evaluated.
\end{description}
All the subroutines return a test function value corresponding to
the point $x$. They return the value GKLS\_MAX\_VALUE (defined in
{\bf gkls.h}) in two cases: (a) vector~$x$ does not belong to the
admissible region $\Omega$ and (b) the user tries to call the
subroutines without generating a test function.
The following subroutines are provided for calculating the partial
derivatives of the test functions (see Appendix).
Evaluation of the first order partial derivative of the D-type
test functions with respect to the variable $x_j$
(see~(\ref{df_cubic})--(\ref{dC_cubic}) in Appendix) is done by
calling subroutine
{\bf double} {\it GKLS\_D\_deriv} ({\it j}, {\it x}). \\
Evaluation of the first order partial derivative of the D2-type
test functions with respect to the variable $x_j$
(see~(\ref{df_quintic})--(\ref{dQ_quintic}) in Appendix) is done
by calling subroutine
{\bf double} {\it GKLS\_D2\_deriv1} ({\it j}, {\it x}). \\
Evaluation of the second order partial derivative of the D2-type
test functions with respect to the variables $x_j$ and $x_k$ (see
in Appendix the
formulae~(\ref{d2(jk)f_quintic})--(\ref{d2(jk)Q_quintic}) for the
case $j \neq k$
and~(\ref{d2(jj)f_quintic})--(\ref{d2(jj)Q_quintic}) for the case
$j = k$) is done by calling subroutine
{\bf double} {\it GKLS\_D2\_deriv2} ({\it j}, {\it k}, {\it x}). \\
Input parameters for these three subroutines are:
\begin{description}
\item[\it j, k] -- ({\bf unsigned int}) indices of the variables
(that must be in the range from 1 to {\it GKLS\_dim}) with respect to
which the partial derivative is evaluated;
\item[\it x] -- ({\bf double *}) a point $x \in \mathbb{R} ^N$ where
the derivative must be evaluated.
\end{description}
All subroutines return the value of a specific partial derivative
corresponding to the point $x$ and to the given direction. They
return the value GKLS\_MAX\_VALUE (defined in {\bf gkls.h}) in
three cases: (a) index ($j$ or $k$) of a variable is out of the
range [1,{\it GKLS\_dim}]; (b) vector $x$ does not belong to the
admissible region~$\Omega$; (c) the user tries to call the
subroutines without generating a test function.
Subroutines for calculating the gradients of the D- and D2-type
test functions and for calculating the Hessian matrix of the
D2-type test functions at a given feasible point are also
provided. These are
{\bf int} {\it GKLS\_D\_gradient} ({\it x}, {\it g}), \\
{\bf int} {\it GKLS\_D2\_gradient} ({\it x}, {\it g}), \\
{\bf int} {\it GKLS\_D2\_hessian} ({\it x}, {\it h}). \\
Here
\begin{description}
\item[\it x] -- ({\bf double *}) a point $x \in \mathbb{R} ^N$
where the gradient or Hessian matrix must be evaluated;
\item[\it g] -- ({\bf double *}) a pointer to the gradient vector
calculated at {\it x};
\item[\it h] -- ({\bf double **}) a pointer to the Hessian matrix
calculated at {\it x}.
\end{description}
Note that before calling these subroutines the user must allocate
dynamic memory for the gradient vector {\it g} or the Hessian
matrix {\it h} and pass the pointers {\it g} or {\it h} as
parameters of the subroutines.
These subroutines call the subroutines described above for
calculating the partial derivatives and return an error code ({\bf
GKLS\_DERIV\_EVAL\_ERROR} in the case of an error during
evaluation of a particular component of the gradient or the
Hessian matrix, or {\bf GKLS\_OK} if there are no errors).
\subsubsection {Memory deallocating}
When the user concludes his work with a test function he should
deallocate dynamic arrays allocated by the generator. This is done
by calling subroutine
{\bf void} {\it GKLS\_free} ({\bf void});\\
with no parameters.
When the user abandons the test class he should deallocate dynamic
boundaries vectors {\it GKLS\_domain\_left} and {\it
GKLS\_domain\_right} by calling subroutine
{\bf void} {\it GKLS\_domain\_free} ({\bf void});\\
again with no parameters.
It should be finally highlighted that if the user, after
deallocating memory, wishes to return to the same class,
generation of the class with the same parameters produces the same
100 test functions.
An example of the generation and use of some of the test classes
can be found in the file {\bf example.c}.
\vspace{5mm}
\begin{Acknowledgement}
This research was partially supported by the following projects:
FIRB RBAU01JYPN, FIRB RBNE01WBBB, and RFBR 01-01-00587. The
authors thank Associate Editor Michael Saunders and an anonymous
referee for their subtle suggestions.
\end{Acknowledgement}
|
\section{The setting}
We consider actions of higher rank abelian groups
$\mathbb{Z}^k\times \mathbb{R}^\ell, \,\, k+l\ge 2$ that come
from the following general algebraic construction:
Let $G$ be a
connected Lie group, $A\subseteq G$ a closed Abelian subgroup which
is isomorphic to $\mathbb{Z}^k\times \mathbb{R}^\ell$, $L$ a compact subgroup of
the centralizer $Z(A)$ of $A$, and $\Gamma$ a cocompact lattice in
$G$. Then $A$ acts by left translation on the compact space
$M=L\backslash G/\Gamma$.
Throughout this paper $G$ will always denote a semisimple connected Lie group of $\mathbb{R}$- rank $\ge 2$ without compact factors and with finite center.
\subsection{Symmetric space examples} Let $D$ be the
connected component of a split Cartan subgroup of $G$. $D$ is
isomorphic to $\mathbb{R}^{\text{rank}_\mathbb{R} G}$. Suppose $\Gamma$ is an
irreducible torsion-free cocompact lattice in $G$. The centralizer
of $Z(D)$ of $D$ splits as a product $Z(D)=KD$ where $K$ is compact
and commutes with $D$. Let $D_+$ be a closed subgroup of $D$
isomorphic to $\mathbb{R}^\ell\times\mathbb{Z}^k$, $k+\ell\geq 2$. Let $\Phi$
denote the restricted root system of $G$. Then the Lie algebra
$\mathfrak{G}$ of $G$ decomposes
\begin{align*}
\mathfrak{G}=\mathfrak{K}+\mathfrak{D}+\sum_{\phi\in\Phi}\mathfrak{g}^\phi
\end{align*}
where $\mathfrak{g}^\phi$ is the root space of $\phi$ and
$\mathfrak{K}$ and $\mathfrak{D}$ are the Lie algebras of $K$ and
$D$ respectively. Elements of
$\mathfrak{D}\backslash\bigcup_{\phi\in\Phi}\ker(\phi)$ are
\emph{regular} elements. Connected components of the set of regular
elements are \emph{Weyl chambers}.
We will consider actions of higher rank subgroups of $D$ by
left translations on double coset spaces
$L\backslash G/\Gamma$ where $L\subset K$ is a connected subgroup.
The action of $D$ on $ G/\Gamma$ is sometimes referred to as {\em full Cartan action}. The action of on $K\backslash G/\Gamma$ is called {\em Weyl chamber flow}. It is Anosov (normally hyperbolic with respect to the orbit foliation). Regular elements of the Weyl chamber flow are normally hyperbolic.
Full Cartan action coincides with the Weyl chamber flow if and only if the group $G$
is $\mathbb{R}$ split.
\subsection{Twisted symmetric space examples}Let $\rho:\Gamma\rightarrow
SL(m,\mathbb{Z})$ be a representation of $\Gamma$ which admits no invariant
subspace with eigenvalue $1$. Then $\Gamma$ acts on the $N$-torus
$\mathbb T^N$ via $\rho$. By Margulis' superrigidity theorem
\cite{Margulis}, semisimplicity of the algebraic hull $H$ of
$\rho(\Gamma)$ and the non-compactness of $\rho(\Gamma)$ the
representation $\rho$ of $\Gamma$ extends to a rational homomorphism
$G\rightarrow H_{ad}$ over $\mathbb{R}$ where $H_{ad}$ is the adjoint group
of $H$. Note that $\rho(\Gamma)$ has finite center $Z$ as follows
for example from Margulis' finiteness theorem \cite{Margulis}, then
$G$ acts on orbifold $\mathbb{R}^N/Z$ via $\rho$, which can be lifted to a
representation of $G$ on $\mathbb{R}^N$. We still denote by
$\rho$.
Now assume notations of the previous section.
Then we have semi-direct product Lie group $G\ltimes \mathbb{R}^N$ twisted by $\rho$. The
multiplication of elements in $G\ltimes \mathbb{R}^N$ is defined by
\begin{align*}
(g_1,r_1)\cdot(g_2,r_2)=(g_1g_2,\rho(g_2^{-1})r_1+r_2).
\end{align*}
Then $\Gamma\ltimes\mathbb{Z}^N$ is a lattice in $G\ltimes \mathbb{R}^N$. We can
view $G\ltimes\mathbb{R}^N/\Gamma\ltimes\mathbb{Z}^N$ as a torus bundle over
$G/\Gamma$. So
we may assume without loss of generality that $\rho$ is irreducible
over $\mathbb{R}$ on $\mathbb{R}^N$.
Let $\Phi_1$ be the set of weights of the representation $\rho$.
Then the Lie algebra $\mathfrak{G}+\mathbb{R}^N$ of $G\ltimes\mathbb{R}^N$
decomposes
\begin{align*}
\mathfrak{G}+\mathbb{R}^N=\mathfrak{K}+\mathfrak{D}+\sum_{\phi\in\Phi}\mathfrak{g}^\phi+\sum_{\mu\in\Phi_1}\mathfrak{v}^\mu
\end{align*}
where $\mathfrak{v}^\mu$ is the weight space of $\mu$. We call the
representation $\rho$ on $\mathbb{R}^{N}$ {\em Anosov} or {\em
hyperbolic}, if there is no $0$ weight in $\Phi_1$ and {\em
genuinely partially hyperbolic} on $\mathbb{R}^{N}$ if $0\in\Phi_1$.
Elements of
$\mathfrak{D}\backslash\bigcup_{\phi\in\Phi\cup\Phi_1\backslash
0}\ker(\phi)$ are \emph{regular} elements. Connected components of
the set of regular elements are \emph{Weyl chambers}.
Similarly to the symmetric space setting we will consider actions of
higher rank subgroups of $D$ by left
translations on double coset spaces
\\ $L\backslash G\ltimes\mathbb{R}^N/\Gamma\ltimes\mathbb{Z}^N$ where $L\subset
K$ is a connected subgroup.
The action of the whole group $D$ on $K\backslash G\ltimes\mathbb{R}^N/\Gamma\ltimes\mathbb{Z}^N$ is called {\em twisted Weyl chamber flow}. It is Anosov if and only if the representation $\rho$ is Anosov.
\begin{remark}
The requirement ``$\Gamma$ which admits no invariant subspace with
eigenvalue $1$'' is necessary for rigidity. Otherwise,
there is a factor $\mathbb{R}^{N_i}$ on which the $G$ acts trivially.
\end{remark}
\subsection{Regular restrictions and coarse Lyapunov distributions}\label{regularrestrictions}
Let $X$ be a double coset space $L\backslash G/\Gamma$ as in
symmetric space examples or $L\backslash
G\ltimes\mathbb{R}^N/\Gamma\ltimes\mathbb{Z}^N$ as in twisted symmetric space
examples.
\begin{definition}
A two-dimensional plane in $\mathbb{P}\subset\mathbb{D}$, the Lie
algebra of $D$, is {\em in regular position} if it contains at least
one regular element. \end{definition}
Let $\mathbb{D}_+\subset \mathbb{D}$ be a closed
subgroup which contains a lattice $\mathbb{L}$ in a plane in regular
position and let $D_+=\exp\mathbb{D}_+$.
\begin{definition}
The action $\alpha_{D_+}$ of $D_+$ by left translations on $X$ will be referred to as a {\em higher-rank regular restriction of
split Cartan actions} or just a {\em regular restriction} for short.
\end{definition}
Throughout this paper terms ``{\em symmetric space examples}'' and ``{\em twisted symmetric space examples}'' are used as synonyms for ``higher rank regular restrictions of split Cartan actions'' in the corresponding cases.
If a plane contains a lattice in regular position, then there exists
a linearly independent basis of the plane consisting of two regular
elements in the lattice which we denote by $a$ and $b$. $a$ and $b$ will be
referred to as {\em regular generators}.
\begin{definition} {\em Coarse Lyapunov distributions} are defined as minimal
non-trivial intersections of stable distributions of various action
elements.\end{definition}
In the setting of present paper those are homogeneous distributions
or their perturbations, that integrate to homogeneous foliations
called {\em coarse Lyapunov foliations}; see \cite[Section
2]{Damjanovic1} and \cite{Kalinin} for detailed discussion in
greater generality.
The standard root system comes from the decomposition of
$\mathfrak{G}$, the Lie algebra of $G$, into the eigenspaces of
adjoint representation of $\mathfrak{D}$ whose elements are
simultaneously diagonalizable. For any connected subgroup $P$ of $D$
with Lie algebra $\mathbb{P}$, we can also consider the
decomposition of $\mathfrak{G}$ with respect to the adjoint
representation of $\mathbb{P}$ and the resulting root system is
called the {\em root system with respect to $P$}. Then the coarse
Lyapunov distributions of regular restrictions both these classes of
examples are the positively proportional root spaces obtained from
root systems with respect to $D_+$.
For the symmetric space examples let $r$ be the smallest integer such
that the coarse Lyapunov distributions of full Cartan actions as
well as their commutators of length $r$ span the tangent space at
any $x\in M$, then $r$ also has the same property for any regular
restriction.
For the twisted symmetric space examples, notice that in
$\mathfrak{G}+\mathbb{R}^N$
\begin{align}\label{for:79}
[(X_1,0),(X_2,0)]=[X_1,X_2]_\mathfrak{G}\quad\text{ and
}\quad[(X,0),(0,t)]=d\rho(X)t
\end{align}
where $X_1\,,X_2\,,X\in\mathfrak{G}$, $t\in\mathbb{R}^N$;
$[\cdot,\cdot]_\mathfrak{G}$ denotes the Lie bracket in
$\mathfrak{G}$ and $d\rho$ is the induced Lie algebra representation
on $\mathbb{R}^N$ from $\rho$. By complete irreducibility of
semisimple groups, there is a decomposition
of $\mathbb{R}^N=\bigoplus_i\mathbb{R}^{N_i}$ such that $\rho$ is irreducible over
$\mathbb{R}$ on each $\mathbb{R}^{N_i}$ and hence $\rho$ is irreducible over $\mathbb{R}$
on each $\mathbb{R}^{N_i}$. It follows that there exists an integer $r_1$
such that the coarse Lyapunov distributions of full Cartan actions
as well as their commutators of length $r_1$ also span the tangent
space at any $x\in M$. Similarly $r_1$ has the same property for
any regular restriction.
Let $\mathfrak L$ be the Lie algebra of the group $L$. Regular
restrictions are partially hyperbolic:
\begin{itemize}
\item
For the symmetric space examples the neutral distribution is
$\mathfrak L\backslash\mathfrak{N}$ where $\mathfrak{N}=
\mathfrak{K}+\mathfrak{D}$;
\item for
the Anosov twisted symmetric space examples the neutral distribution
is $\mathfrak L\backslash\mathfrak{N}$ where $\mathfrak{N}=
\mathfrak{K}+\mathfrak{D}$;
\medskip
Fix a positive definite inner product
$\langle\cdot\rangle$ on $\mathfrak{K}+\mathfrak{D}$ that are
invariant under $\text{Ad}_K$. Then $\text{Ad}_K\subset
O(\mathfrak{N},\langle\cdot\rangle)$. Let $\mathfrak{L}^\bot$ be the
orthogonal complement of $\mathfrak{L}$. Then
$\mathfrak{L}^\bot=L\backslash\mathfrak{N}$ and both $\mathfrak{L}$
and $\mathfrak{L}^\bot$ are invariant under $\text{Ad}_{Z(L)}$.
\medskip
\item for the genuinely partially hyperbolic twisted symmetric
space examples the neutral distribution is
$L\backslash\mathfrak{N}$ where $\mathfrak{N}=
\mathfrak{K}+\mathfrak{D}+\sum \mathfrak{v}^0$ and $\mathfrak{v}^0$
are weight spaces of $0$ weight.
\medskip
Fix positive definite inner products $\langle\cdot\rangle_1$ on
$\mathfrak{K}+\mathfrak{D}$ and $\langle\cdot\rangle_2$ on
$\mathfrak{v}^0$ that are invariant under $\text{Ad}_{K}$ and
$\rho(K)$ respectively. Then $\text{Ad}_{K,0}\subset
O(\mathfrak{N},\langle\cdot\rangle)$. Let
$\langle\cdot\rangle=\langle\cdot\rangle_1+\langle\cdot\rangle_2$
and $\mathfrak{L}^\bot$ be the orthogonal complement of
$\mathfrak{L}$. Then $\mathfrak{L}^\bot=L\backslash\mathfrak{N}$
and
$\mathfrak{L}^\bot=\mathfrak{L}^\bot\bigcap(\mathfrak{K}+\mathfrak{D})
+\mathfrak{v}^0$. Notice
\begin{align*}
\text{Ad}_{l,t_0}(\mathfrak{l},t)=\text{Ad}_{l,0}\text{Ad}_{e,t}(\mathfrak{l},t)\subseteq\text{Ad}_{l,0}(\mathfrak{l},\mathfrak{v}^0)
\subseteq \mathfrak{L}^\bot
\end{align*}
where $l\in Z(L)\bigcap\exp(\mathfrak{K}+\mathfrak{D})$, $t_0$ and
$t\in \mathfrak{v}^0$ and $\mathfrak{l}\in
\mathfrak{L}^\bot\bigcap(\mathfrak{K}+\mathfrak{D})$. Hence it
follows $\mathfrak{L}^\bot$ are invariant under $\mathfrak{L}$
and $\text{Ad}_{Z(L)}$.
\end{itemize}
\medskip
\begin{remark} Notice that the neutral distribution for any regular restriction coincides with the
homogeneous distribution into cosets of the centralizer of the full Cartan action, or its factor by $L$ in the case of actions on double coset spaces.
\end{remark}
We
call the following simple fact {\em the uniqueness decomposition
property} which will be often used in the future: there exist
bounded, open, connected neighborhoods $U_{\mathfrak{L}}$ and
$U_{\mathfrak{L}^\bot}$ of $0$ in $\mathfrak{L}$ and
$\mathfrak{L}^\bot$ respectively, such that the mapping
$\mathcal{T}:(U_{\mathfrak{L}},U_{\mathfrak{L}^\bot})\rightarrow
\exp(U_{\mathfrak{L}})\exp(U_{\mathfrak{L}^\bot})$ is a
diffeomorphism of $U_{\mathfrak{L}}\times U_{\mathfrak{L}^\bot}$ onto an open neighborhood of $e$ in $Z(D)$.
\subsection{Rigidity} One can naturally think of
$D_+$ as the image of an embedding $i_0 :
\mathbb{Z}^k\times\mathbb{R}^{\ell}\rightarrow D_+$. Then one can consider
the action $\alpha_{i_0}$ of $A=\mathbb{Z}^k\times\mathbb{R}^l$ on $M$
given by \begin{align}\label{for:15} \alpha_{i_0}(a,x)=i_0(a)\cdot x \end{align}
Thus $\alpha_{i_0}$ is $\alpha_{D_+}$ with a fixed system of generators. We will say that $A$ action $\alpha_{i_0}$ generates $D_+$ action $\alpha_{D_+}$.
Of course, $D_+$ can be obtained as the image of different embeddings; corresponding actions of $A$ differ by a time change.
A {\em standard perturbation} of the action $\alpha_{i_0}$ is an action $\alpha_i$ where $i: A \to D$ is a homomorphism close to $i_0$. Since a small perturbation of an embedding is still an embedding a standard perturbation also generates a regular restriction
$\alpha_{D'_+}$ where $D'_+\subset D$ is the image of $i$.
\begin{remark}
In the hyperbolic situation, i.e. for $D_+=D$ any standard perturbation is simply a time change
corresponding to an automorphism of the acting group but in the partially hyperbolic cases standard perturbations are usually essentially different from each other.
\end{remark}
A proper notion of rigidity for algebraic actions, i.e homogeneous actions, their factors, as well as affine actions, states that any perturbation of the action in a properly defined regularity class is conjugate to an algebraic action obtained by perturbing acting subgroup. In our setting this translates into the following definition.
\begin{definition}An action $\alpha_{i_0}$ of $A=\mathbb{Z}^k\times\mathbb{R}^l$ on $M$ is $C^{k,r,\ell}$ {\em locally
rigid} if any $C^k$ perturbation $\tilde{\alpha}$ which is
sufficiently $C^r$ close to $\alpha_{i_0}$ on a compact generating
set is $C^\ell$ conjugate to a standard perturbation
$\alpha_i$.
An action $\alpha_{D_+}$ is $C^{k,r,\ell}$ {\em locally
rigid} if there exists a homomorphism $i_0: \mathbb{Z}^k\times\mathbb{R}^l\to D$ whose image equals $D_+$
such that $\alpha_{i_0}$ is $C^{k,r,\ell}$ locally
rigid.
\end{definition}
\begin{remark} It is immediately obvious that if $\alpha_{i_0}$ is locally rigid then the same is true for any time change obtained by an automorphism of $A$; hence the notion of local rigidity for $\alpha_{D_+}$ depends only of the subgroup $D_+$.
\end{remark}
\section{History of the rigidity problem}
\subsection{Rigidity of hyperbolic actions}
Differentiable rigidity of higher rank algebraic Anosov actions including Weyl chamber flows and twisted Weyl chamber flows was proved in the mid-1990s \cite{Spatzier}.
The proof consists of two major parts:
(i) An {\em a priori regularity} argument that shows smoothness of the Hirsch-Pugh-Shub orbit equivalence
\cite{shub}. The key part of the argument is the theory of non-statinary normal forms developed in \cite{Guysinsky}.
(ii) {\em Cocycle rigidity} used to ``straighten out'' a time change; it is proved by a harmonic analysis method in \cite{Spatzier1}.
Both these ingredients also appear in the present paper.
\subsection{Difference between hyperbolic and partially hyperbolic actions}\label{sec:2.2}
The next step is to consider algebraic {\em partially hyperbolic
actions}. Unlike the hyperbolic case, the a priori regularity method is not directly
applicable here since individual elements of such actions are not
even structurally stable. In addition to their work on hyperbolic
rigidity, the first author and R. Spatzier also considered cocycle
problem for certain partially hyperbolic actions in
\cite{Spatzier2} and proved essential cocycle rigidity results.
Before proceeding with the chronological account let us explain an essential point that also plays a role in the present work.
In the hyperbolic case smooth orbit rigidity reduces the local differentiable rigidity problem to rigidity of vector valued cocycles. For, the expression of the {\em old time}, i.e. that of the unperturbed action through the {\em new time}, i.e. that of the perturbed action is a cocycle over the {\em unperturbed action} with values in the acting group that is a vector space or its subgroup. This is the scheme of \cite{Spatzier}. In other words, smooth orbit rigidity reduces the differentiable conjugacy problem to a time change problem.
In the partially hyperbolic cases instead of smooth orbit rigidity one may hope at best to have
{\em smooth rigidity of neutral foliation} when the scheme of \cite{Spatzier} is applicable. For that one needs the following property:\medskip
\noindent ($\mathfrak B'$) {\em The stable directions of various
action elements or, equivalently, coarse Lyapunov
distributions,(see above
Section~\ref{regularrestrictions}),
together with the orbit direction, generate the tangent space as a
Lie algebra, i.e. those distributions and their brackets of all
orders generate the Lie algebra linearly.} \medskip
In this paper we deal with cases, namely regular restrictions of split Cartan actions, where an even stronger property ($\mathfrak B$) holds:\medskip
\noindent ($\mathfrak B$) {\em The stable directions of various action elements generate the tangent space as a Lie algebra.}
\medskip
Still even after smooth rigidity of the orbit foliation has been established, the problem of differentiable conjugacy is not reduced to a cocycle problem over the unperturbed action; rather it reduces to the cocycle problem over the
perturbed action; furthermore, the values of the cocycle may be in a non-abelian group: it is actually the exponential of the central (neutral) distribution.
\subsection{Previous work on partially hyperbolic rigidity}
\subsubsection{KAM method }
The work on the differentiable rigidity of partially hyperbolic but not hyperbolic actions started in earnest in 2004 \cite{DD-AK-ERA} (complete proofs appeared in print in \cite{Damjanovic4}). The situation considered in that work, actions by commuting partially hyperbolic automorphisms of a torus, is algebraically more amenable than
the one of \cite{Spatzier2} or the present paper, but is geometrically more subtle because stable directions for different action elements commute as any homogeneous distributions on the torus do. To handle this problem a
new method was introduced: a KAM type iteration scheme formally similar to that employed by J.Moser \cite{Moser} in the higher-rank version of a conventional setting for application of KAM scheme -- diffeomorphisms of the circle.
While causing geometric difficulties, abelian nature of the torus situation helps both with algebra and with analysis involved in carrying out
the iterative scheme. It allows for an explicit calculation of solutions of linearized conjugacy equation, as well
as the splitting for ``almost cocycles'' that transforms producing needed tame estimates
into manageable problems of Fourier analysis. At the time applications of the KAM scheme even to the most basic semisimple situations looked very problematic. It seemed that specific information from representation theory would be needed that may be available for some semisiple Lie groups and not for others. Only in the present work new algebraic and analytic insights and tools are developed that made possible applications of the KAM type iterative scheme to all semisimple and various other cases.
\subsubsection{Geometric/Algebraic $K$-theory method }
In \cite{Damjanovic1,Damjanovic2} a different method has been developed that bypasses subtleties of analysis and representation theory altogether. It builds upon the already mentioned observation that the problem of differentiable conjugacy reduces to the cocycle problem over the perturbed action. Solution of coboundary equations for actions satisfying condition ($\mathfrak B$)
are built along broken paths consisting of pieces of stable foliations for different elements of the action. It was first observed in \cite{Damjanovic1} that consistency of such a construction follows in certain cases from description of generators and relations in the ambient Lie group. In \cite{Damjanovic2} another key ingredient was introduced: under certain circumstances the web of Lyapunov foliations is so robust that the construction of solutions of the coboundary equation carries out to the {\em perturbed action} thus providing a solution of the conjugacy problem in the H\"older category. After that smoothness of the conjugacy is established by a priori regularity method as in \cite{Spatzier}.
A great strength of this method is that it requires only $C^2$ closeness for the perturbation (possibly even less), unlike the KAM scheme where number of derivatives is large and dependent on the data.
But this comes with a price. In order to carry out this scheme: \smallskip
\noindent (i) very detailed information about specific generators and relations in the ambient group is required, and
\noindent (ii) coarse Lyapunov distributions for the partially hyperbolic restriction should be the same as for the ambient Cartan action. \smallskip
The latter requirement leads to an extra general position (generic) assumption
on the acting subgroup of $D$ that, while open and dense in the spaces of actions,
(i.e. of embeddings $i_0: \mathbb{Z}^k\times\mathbb{R}^l\to D$) is more restrictive than ($\mathfrak B$) and is clearly far from necessary for rigidity.
In \cite{Damjanovic1,Damjanovic2} the special case $G=SL(n,\mathbb{R})(n\geq
3)$ is considered. In this case necessary algebraic information can be extracted from the classical work of Steinberg, Matsumoto and Milnor, see \cite{Steinberg, Matsumoto, Milnor}.
The approach of \cite{Damjanovic1,Damjanovic2} was further employed
in \cite{Damjanovic3}, \cite{Zhenqi1} and \cite{Zhenqi2} for
extending cocycle rigidity and differentiable
rigidity from
$SL(n,\mathbb{R})/\Gamma$ and $SL(n,\mathbb{C})/\Gamma$ to compact homogeneous
spaces obtained from simple split Lie groups and some non-split Lie
groups. All actions considered in \cite{Damjanovic1, Damjanovic2, Damjanovic3, Zhenqi1} are ``generic restrictions'' and in
\cite{Zhenqi2} only a special example of non-generic restrictions is
considered.
There is no way to deal with more general situations like regular
restrictions of Weyl chamber flows treated in the present
paper. Unlike the case
of generic restrictions, in both symmetric space examples and
twisted symmetric space examples coarse Lyapunov distributions
of regular restrictions may be different from those of full Cartan
actions.
Another problem with the geometric method is that it requires
consideration of simple Lie groups case-by-case. While in
\cite{Damjanovic3, Zhenqi2} algebraic results needed for the
treatment of actions on homogeneous spaces of split Lie groups are
still deduced from the literature with a moderate effort, treatment
of non-split series $SO^o(m,n), |m-n|\ge 2$ and $SU(m,n)$ in
\cite{Zhenqi1} requires an algebraic {\em tour de force}. The
remaining classical cases that are quaternionic groups, and
non-split exceptional groups, would require even heavier algebraic
calculations although results of the present work indicate that
rigidity is likely to hold with lower regularity requirements on
the closeness of perturbations. It is unlikely that this method can
be adapted to the case of semisimple but not simple groups.
The geometric method can
be applied to some hyperbolic twisted symmetric spaces examples,
which will appear in another paper by the second author, but it
can't deal with genuinely partially hyperbolic cases. For these
cases there are no good invariant layers isomorphic to simple
groups under perturbations as in the hyperbolic twisted symmetric
spaces examples and due to the abelian nature of torus bundles geometric structures are not robust under perturbations.
\subsubsection{Comparison} To summarize, two methods developed to prove rigidity for partially hyperbolic actions can be viewed as somewhat complementary: the method of the present paper requires less specific information and hence applies in a greater variety of settings to a broader class of actions, while the geometric/algebraic $K$-theory method, whose applicability is more limited, produces stronger conclusions by requiring
lower regularity for perturbations. \medskip
\section{Statement of results}
In this paper we develop an approach for proving local
differentiable rigidity of higher rank abelian groups, i.e.
$\mathbb{Z}^k\times \mathbb{R}^\ell$, $k + \ell\geq 2$, based on KAM-type
iteration scheme, that was first introduced in \cite{DD-AK-ERA, Damjanovic4}
where local differentiable rigidity was proved for $\mathbb{Z}^k$,
$k\geq 2$ actions by partially hyperbolic automorphisms of a torus.
Here we deal with actions of $\mathbb{Z}^k\times \mathbb{R}^\ell$, $k
+ \ell\geq 2$ on $L\backslash G/\Gamma$ and $L\backslash G\ltimes\mathbb{R}^n/\Gamma\ltimes \mathbb{Z}^n$ that are regular restrictions of split Cartan actions. \medskip
The work on commuting toral automorphisms together with the present work covers all significant instances of higher rank partially hyperbolic algebraic actions with an exception of actions by automorphisms of nilmanifolds that is the subject of a work in progress by Damjanovic and collaborators.
\subsection{Some Lie groups preliminaries}
Lojasiewicz inequality \cite[Theorem 4.1]{Malgrange} implies the
following statement: for any subalgebra $\mathfrak{S}\subseteq
\mathfrak{N}$ there exist constants $d\,,q\,,\delta>0$ such that
$\mathfrak{n}_1\,,\mathfrak{n}_2\in\mathfrak L$ any $1\geq\gamma>0$,
if $\norm{[\mathfrak{n}_1,\mathfrak{n}_2]}\leq \gamma$ then there
exist $\mathfrak{n}'_1\,,\mathfrak{n}'_2\in\mathfrak{S}$ such that
\begin{align}\label{for:112}
\norm{[\mathfrak{n}_1,\mathfrak{n}_2]}=0,
\norm{\mathfrak{n}_1-\mathfrak{n}_1'}\leq d\gamma^\delta,
\norm{\mathfrak{n}_2-\mathfrak{n}_2'}\leq d\gamma^\delta, \text{ if
}\norm{\mathfrak{n}_1}+\norm{\mathfrak{n}_2}\leq q.
\end{align}
\begin{definition}\label{def:1}
Let $\delta(q)$ be the maximum of all $\delta$ satisfying
\eqref{for:112} and let $\delta_0(\mathfrak{S})=\max\{\delta(t),
0<t\leq q\}$.
\end{definition}
As we will see now, for the symmetric
space examples and Anosov twisted symmetric space examples for any
subalgebra $\mathfrak{S}\subseteq \mathfrak{N}$ we have
$\delta_0(\mathfrak{S})\geq\frac{1}{2}$.
For the genuinely partially
hyperbolic examples, if $G$ is split over $\mathbb{R}$, then $\mathfrak{N}$
is abelian and hence $\delta_0(\mathfrak{S})=\infty$ for any
$\mathfrak{S}\subseteq \mathfrak{N}$. If $G$ is non-split over
$\mathbb{R}$, that is the compact part $\mathfrak{K}$ is nontrivial,
$\delta_0$ depends on the the representation $\rho$. Even if $G$
is quasi-split over $\mathbb{R}$, that is $\mathfrak{K}$ is nontrivial but
abelian, there is no definite answer about $\delta_0$.
\begin{lemma}\label{le:15}
For the symmetric space examples and the Anosov twisted symmetric
space examples, for any subalgebra $\mathfrak{S}\subseteq
\mathfrak{N}$, $\delta_0(\mathfrak{S})\geq\frac{1}{2}$.
\end{lemma}
\begin{proof}
For these cases
$\mathfrak{S}=\mathfrak{K}\bigcap\mathfrak{S}+\mathfrak{D}\bigcap\mathfrak{S}$.
Suppose $\mathfrak{n}_1\,,\mathfrak{n}_2\in\mathfrak{S}$ and
$\norm{[\mathfrak{n}_1,\mathfrak{n}_2]}\leq \gamma$. Since
$\mathfrak{K}\bigcap\mathfrak{S}$ is compact it can be written
as a direct sum
$\mathfrak{K}\bigcap\mathfrak{S}=\mathfrak{K}_0+\mathfrak{K}_s$
where the ideals $\mathfrak{K}_s$ and $\mathfrak{K}_0$ are
semisimple and abelian, respectively. Let $p$ be the projection from
$\mathfrak{S}$ to $\mathfrak{K}_s$. Let $K'$ be the connected
subgroup in $G$ with Lie algebra $\mathfrak{K}_s$. Then $K'$ is
compact. Let $U$ be the maximal torus in $K'$ containing
$p(\mathfrak{n}_1)$ with Lie algebra $\mathfrak{U}$. Let
$(\mathfrak{K}_s)_{\mathbb{C}}$ be the complexification of $\mathfrak{K}_s$
and let $\mathfrak{U}'$ denote the subalgebra of
$(\mathfrak{K}_s)_{\mathbb{C}}$ generated by $\mathfrak{U}$. Let $\Delta^*$
denote the set of nonzero roots of $(\mathfrak{K}_s)_{\mathbb{C}}$ with
respect to $\mathfrak{U}'$. Then there exists $E_\psi\,,F_\psi$,
$(\psi\in (\Delta^*)_+)$ is a base of $\mathfrak{K}_s$ (mod
$\mathfrak{U}$) and
\begin{align*}
&[H,E_\psi]=-\textrm{i}\psi(H)F_\psi\notag\\
&[H,F_\psi]=\textrm{i}\psi(H)E_\psi
\end{align*}
for all $H\in \mathfrak{U}$.
Let
$p(\mathfrak{n}_2)=\sum_{\psi\in (\Delta^*)_+ }(x_\psi E_\psi+y_\psi
F_\psi)+u$ where $u\in \mathfrak{U}$. Let
$$O=\{\psi\in (\Delta^*)_+| \text{ either
}\abs{x_\psi}>\sqrt{\gamma}\text{ or }\abs{y_\psi}>\sqrt{\gamma}\}$$
and let $p'(\mathfrak{n}_2)=\sum_{\psi\in O }(x_\psi E_\psi+y_\psi
F_\psi)+u$, then
$\norm{p'(\mathfrak{n}_2)-p(\mathfrak{n}_2)}<C\sqrt{\gamma}$. By
assumption $\psi\bigl(p(\mathfrak{n}_1)\bigl)<\sqrt{\gamma}$ if
$\psi\in O$. Since the maximum cardinality of the set of linearly
independent elements in $O$ is not greater than $\dim \mathfrak{U}$ then
there exists $p'(\mathfrak{n}_1)\in \mathfrak{U}$ such that
$\norm{p'(\mathfrak{n}_1)-p(\mathfrak{n}_1)}<C\sqrt{\gamma}$ and
$\psi\bigl(p'(\mathfrak{n}_1)\bigl)=0$ for any $\psi\in O$. Then we
also have $[p'(\mathfrak{n}_1),p'(\mathfrak{n}_2)]=0$. Hence we let
$\mathfrak{n}_1'=\mathfrak{n}_1-p(\mathfrak{n}_1)+p'(\mathfrak{n}_1)$
and
$\mathfrak{n}_2'=\mathfrak{n}_2-p(\mathfrak{n}_2)+p'(\mathfrak{n}_2)$.
Then $\mathfrak{n}_1'\,,\mathfrak{n}_2'$ satisfy \eqref{for:112} for
$\delta=\frac{1}{2}$.
\end{proof}
\subsection{Main theorems}In what follows $G$ always denotes a connected semisimple Lie group of real rank greater than one.
In the symmetric space case we prove rigidity for regular restrictions.
\begin{theorem}\label{th:3}
Let $\alpha_{D_+}$ be a restriction of the split Cartan action
$\alpha$ on a factor $L\backslash G/\Gamma$ of $M= G/\Gamma$ to a $D_+\subset D$ in
regular position. Then there exists a constant $\ell=
\ell(\alpha_{D_+},M)$ such that $\alpha_{D_+}$ is
$C^{\infty,\ell,\infty}$-locally rigid.
\end{theorem}
Recall that for the twisted symmetric space examples, if $\rho$ is irreducible on
$\mathbb{R}^N$ and the weights do not contains $0$, then we call the resulted
twisted symmetric space {\em hyperbolic} and if they contain $0$ we call them \emph{irreducible genuinely partially
hyperbolic}.
\begin{theorem}\label{th:6}
Let $\alpha_{D_+}$ be the restriction of the split Cartan action
on a factor of a hyperbolic twisted symmetric space $M=G\ltimes\mathbb{R}^N/\Gamma\ltimes\mathbb{Z}^N$ to
$D_+\subset D$ in regular position. Then
there exists a constant $\ell= \ell(\alpha_{D_+},M)$ such that
$\alpha_{D_+}$ is $C^{\infty,\ell,\infty}$-locally rigid.
\end{theorem}
\begin{theorem}\label{th:4} Let
$\alpha_{D_+}$ be the restriction of the split Cartan action of the
irreducible genuinely partially hyperbolic twisted symmetric space
on $X=L\backslash G\ltimes\mathbb{R}^N/\Gamma\ltimes\mathbb{Z}^N$ to $D_+\subset
D$ in regular position. Let $\mathfrak{Z}$ be the Lie algebra of
$Z(D)\bigcap Z(L)$. If $\delta_0(\mathfrak{Z})>\frac{1}{4}$ then
there exists a constant $\ell= \ell(\alpha_{D_+},M)$ such that
$\alpha_{D_+}$ is $C^{\infty,\ell,\infty}$-locally rigid.
\end{theorem}
\begin{remark}
The condition $\delta_0(\mathfrak{Z})>\frac{1}{4}$ can not be
weakened in the framework of our approach. This condition will be
used in finding two commuting regular generators in small
neighborhoods of two in general not commuting elements in the
iterative step. The norm of the commutator of these two not commuting elements can at
best be estimated by the fourth power of the error, so
$\delta_0>\frac{1}{4}$ is necessary. In fact, this condition also
applies to Theorem \ref{th:3} and \ref{th:6}. We omit it since in
these cases $\delta_0\geq \frac{1}{2}$ by Lemma~\ref{le:15}.
If $G$ is split over $\mathbb{R}$, then $\delta_0=\infty$, hence for all
irreducible genuinely partially hyperbolic representations, we have
rigidity of higher rank regular restrictions.
\end{remark}
Thus what is not fully covered by our theorems is rigidity of
general higher rank irreducible genuinely partially hyperbolic
symmetric space examples for non-split groups. as well as higher rank restrictions that contain no regular elements. The following result
demonstrates the existence of irreducible genuinely partially
hyperbolic symmetric space examples for non-split groups for which
$\delta_0(\mathfrak{Z})>\frac{1}{4}$ for any $L$, and hence
Theorem~\ref{th:4} applies.
\begin{theorem}\label{th:7} For $G$ and $\Gamma$ with assumed
notations irreducible genuinely partially hyperbolic twisted
symmetric space examples exist. For these particular examples
$\delta_0(\mathfrak{Z})\geq\frac{1}{3}$ if $G$ is
quasi-split or the semisimple part of $K$ is $SO(3)$.
\end{theorem}
As for higher rank non-regular restriction of Cartan actions the there three issues here.
\begin{enumerate}
\item First thing is property ($\mathfrak B$). It is necessary for our method and it is satisfied in many cases.
\item Next issue is that the neutral foliation is not any more a combination of an abelian and compact parts. This changes somewhat the character of linearized conjugacy equations since adjoint action is not any more by isometries. In many situations this problem is not particularly serious.
\item And finally there is a control over $\delta_0$ that even in the symmetric space cases need not be greater than $1/4$ any more. However, again in many situations a proper estimate can be carried out.
\end{enumerate}
To summarize, one can define a reasonably broad class of non-regular restrictions to which our results extend with a moderate amount of technical changes.
To put our results into perspective let us mention that a scheme similar to that of the present paper applies to certain
{\em parabolic} cases, i.e. homogeneous actions of unipotent abelian groups. In those situations there are no stable manifolds altogether and hence geometric considerations cannot even get started. While the first paper on the subject \cite{DK-parabolic} uses lots of specific information
about the actions at hand and specific constructions of splitting, insights and constructions introduced in the present paper allow to treat parabolic situations of considerably greater generality
\cite{DFK-parabolic}.
\subsection{Structure of the proof of Theorems~\ref{th:3}, \ref{th:6} and \ref{th:4}}
All three theorems are proved simultaneously; along the way we
sometimes have to consider three cases separately. Typically
distinctions between the symmetric space and twisted hyperbolic
symmetric space cases are minimal, and appear at the level of
notations. However, the genuinely partially hyperbolic twisted
symmetric space case sometimes requires a separate argument, see
proofs of Lemmas \ref{le:11}, \ref{le:16} and \ref{le:1}.
In the next section we formulate without proofs essential ingredients used in the construction of solutions of conjugacy equations. Those include:
\begin{itemize}
\item Decay estimates for matrix coefficients for irreducible unitary representations in the form
presented in \cite{Kleinbock}. Those are essential in the first
reduction step and construction of distribution solutions for
coboundary equations.
\item Elliptic regularity results used previously in the work on cocycle rigidity in the form presented in \cite{Spatzier2}
that allow to deduce global regularity of solutions.
\end{itemize}
In Sections~\ref{sec:prelim} and \ref{section:1gen} we carry out two essential reductions: to the special class of perturbations along the neutral foliation of the original action, and to the conjugacy problem for one generator of the action. After that we derive the linearized conjugacy equation and consider linearized commutativity condition in Section~\ref{sec:1}.
In Section~\ref{section:linear} we solve the linearized conjugacy equation in appropriately chosen Sobolev spaces and produce tame estimates for the solutions. This is done in two steps: we first solve
linearized equation for a single generator with tame estimates assuming vanishing of naturally defined obstructions (Lemma~\ref{le:11}) and then show vanishing of the obstructions in the case of higher rank actions due to the linearized commutativity condition thus producing a common solution for the linearized conjugacy equation (Lemma~\ref{le:1}).
The next and more delicate step is producing a tame splitting for ``almost cocycles'' in Section~\ref{section:splitting}. We first do that in auxiliary ``incomplete'' or ``partial'' Sobolev spaces where the norms include differentiation only along hyperbolic directions and hence no loss of regularity appears in the solution of linearized conjugacy equations (Section~\ref{splitting-prelim}), and then use that
to obtain tame estimates for the splitting in the ``real'' Sobolev spaces (Section~\ref{splitting-final}). There are several essential new ingredients in constructions of the solutions of linearized equations and the splitting that are explained at the beginning of Section~\ref{sec:cohstability} and in Section~\ref{splitting-intro}.
Finally, in Section~\ref{section-KAM} the iteration process is carried out. It consists of two parts: The first part (Sections~\ref{sec:8.2}--\ref{sec:8.4}) that follows the general scheme developed in \cite{Damjanovic4}, and the parameter adjustment (Section~\ref{sec:8.5}) that is specific for our situation since we prove the conjugacy not with the original action but with its algebraic perturbation. It is only in the second part that condition $\delta_0>1/4$ appears, see derivation of inequality \eqref{for:113} and subsequent calculation leading to \eqref{delta14}.
\subsection{Elliptic regularity and decay of matrix coefficients} Now we formulate two important facts that form the analysis that are used in the proofs.
The first one summarizes elliptic regularity results that we need, and the second contains essential information on the decay of matrix coefficients.
Let $X_1,\cdots,X_\ell$ be smooth vector
fields on $M$ whose commutators of length at most $r$ span the
tangent space at each
point on $M$ and satisfy the following technical condition: \\
(*) for each $j$, the dimension of the space spanned by the
commutators of length at most $j$ at each point is constant in a
neighborhood.
\begin{theorem}\cite{Spatzier2}\label{th:5}
Suppose $f$ is in $L^2(M)$ or a distribution on $M$ and $m$ is a
positive integer greater than $r$. Denote by $\mathcal{H}_s$ the
$s'$th Sobolev space of $M$ with Sobolev norm $\norm{\cdot}_s$. If
the $m$'th partial derivative $X_j^m(f)$ exists as a continuous or a
$L^2(M)$ function, then $f\in\mathcal{H}_{\frac{m}{r}-1}$ with
estimate
\begin{align}\label{eq: modSobolev}
\norm{f}_{\frac{m}{r}-1}\leq
C_m(\sum_{j=1}^{\ell}\sum_{i=1}^m\norm{X_j^if}_0+\norm{f}_0)
\end{align}
where $\norm{\cdot}_0$ denotes the $L^2$-norm.
\end{theorem}
Recall that $G$ is a semisimple Lie group of real rank $\ge 2$.
Let us fix a Riemannian metric dist$(\cdot,\cdot)$ on $G$
bi-invariant with respect to $K$. Let $\pi$ be a unitary
representation of $G$ on a Hilbert space $\mathcal{H}$. Say that a
vector $v\in\mathcal{H}$ is $\delta$-Lipschitz if
\begin{align*}
\delta=\sup_{g\in
G-\{e\}}\frac{\norm{\pi(g)v-v}}{\text{dist}(e,g)}<\infty;
\end{align*}
we will refer to the number $\delta$ as to the $\delta$-Lipschitz
coefficient of $v$, and say that the vector $v$ is
$\delta$-Lipschitz.
\begin{theorem}\cite[appendix]{Kleinbock}\label{th:2} Let $\pi$ be an irreducible unitary representation of
$G$ with discrete kernel. There exist constants $\gamma,E>0$,
dependent only on $G$ such that if $v_i$, $i=1,2$, be
$\delta_i$-Lipschitz vectors in the representation space of $\pi$
then for any $g\in G$
\begin{align*}
\abs{\langle\pi(g)v_1,v_2\rangle}&\leq
(E\norm{v_1}\norm{v_2}+\delta_1\norm{v_2}+\delta_2\norm{v_1}+\delta_1\delta_2)e^{-\gamma
\emph{dist}(e,g)}.
\end{align*}
\end{theorem}
\begin{corollary}\label{cor:1}
There exist constants $\gamma,E>0$, dependent only on $G$ such that
if $f_i$, $i=1,2$, be $C^1$-functions on $M$ where $M=G/\Gamma$ or
$M=G\ltimes\mathbb{R}^N/\Gamma\ltimes\mathbb{Z}^N$ orthogonal to the constants
then for any $g\in G$
\begin{align*}
\abs{\langle f_1(g),f_2\rangle}&\leq
E(\norm{f_1}_{0}\norm{f_2}_{0}+\norm{f_1}_{C^1}\norm{f_2}_0\\
&+\norm{f_2}_{C^1}\norm{f_1}_0+\norm{f_1}_{C^1}\norm{f_2}_{C^1})e^{-\gamma
\emph{dist}(e,g)}
\end{align*}
where $\langle \cdot,\cdot\rangle$ the inner product in $L^2(M)$
with respect to the Haar measure.
\end{corollary}\begin{proof}
Denote by $\rho_0$ the regular representation of $G$ on the space
$L^2(M)$ and let $V$ be a non-trivial irreducible component of
$L^2(M)$. If we can show that the kernel of $\rho_0$ is discrete,
then Theorem \ref{th:2} applies since every $C^1$ function $f$ on
$M$ is $\delta$-Lipschitz with $\delta\leq C\norm{f}_{C^1}$ where
$C$ is a constant only dependent on $M$.
If $g_0$ is in the kernel of $\rho_0$ then we have
$f(gg_0g^{-1})=f$ for every $f\in V$ and $g\in G$. Let $H$ be the
subgroup generated by $\{gg_0g^{-1},g\in G\}$. Obviously $H$ is a
normal subgroup of $G$. $H$ is discrete if and only if $g_0$ is in
the center of $G$. If $H$ is not discrete then the non-compactness
follows from the fact there is no nontrivial compact normal
subgroups in $G$. Then there exists a non-compact one parameter
subgroup $\{h_t\}$ in $H$.
For the symmetric space examples, since $\Gamma$ is irreducible,
there are no $L^2$-functions on $G/\Gamma$ orthogonal to the
constants which are invariant under a non-compact element in $G$
by Moore's ergodicity theorem \cite{Zimmer}. Hence every non-trivial
irreducible component of $L^2(G/\Gamma)$ has discrete kernel.
For the twisted symmetric space examples we want to show ergodicity
of noncompact one parameter subgroups of $G$ on
$G\ltimes\mathbb{R}^N/\Gamma\ltimes\mathbb{Z}^N$. Brezin and Moore \cite{Brezin}
show that a one parameter subgroup of $G$ acts ergodically on
$G\ltimes\mathbb{R}^N/\Gamma\ltimes\mathbb{Z}^N$ if and only if the quotient flows
on the maximal Euclidean quotient and the maximal semisimple
quotient are ergodic. The Euclidean quotient is $G\ltimes\mathbb{R}^N/E$
where $[G\ltimes\mathbb{R}^N,G\ltimes\mathbb{R}^N]\subseteq E$. Notice that the derived
group is $G\ltimes\mathbb{R}^N$, as follows from
of \eqref{for:79}. Hence the maximal Euclidean quotient
is a point. The latter quotient is obtained by factoring
$G\ltimes\mathbb{R}^N$ by the closure of $\Gamma\ltimes\mathbb{R}^N$ which is
isomorphic to $G/\Gamma$. Since $\Gamma$ is irreducible, all
noncompact one parameter subgroups of $G$ are ergodic on $G/\Gamma$.
Then we get ergodicity of noncompact one parameter subgroups of $G$
on $G\ltimes\mathbb{R}^N/\Gamma\ltimes\mathbb{Z}^N$, especially for $\{h_t\}$.
Hence every non-trivial irreducible component of
$L^2(G\ltimes\mathbb{R}^N/\Gamma\ltimes\mathbb{Z}^N)$ has discrete kernel.
\end{proof}
\section{Preparatory steps and notations}\label{section:prep}
We begin by describing two important steps that reduce the proof of our theorems
to a more specialized situation.
\subsection{Smooth conjugacy for neutral foliations}\label{sec:prelim}
Let $\tilde\alpha$ be a $C^\infty$ action that is $C^l$ close to
$\alpha_{D_+}$. The first key step in the proof of our results is
finding a $C^\infty$ diffeomorphism $H_1$ (neutral foliations
conjugacy) that maps the neutral foliation of the perturbed
action $\tilde\alpha$ onto that of the unperturbed action
$\alpha_{D_+}$. This is done by the application of \cite[Theorem
1]{Spatzier}. A short comment is in order. The quoted theorem has
words ``Anosov action'' in its assumptions and its conclusion is
existence of smooth orbit equivalence for such an action satisfying
some technical assumptions, and its perturbation. However, the proof
applies verbatim to our situation replacing the orbit foliation by
the neutral foliation. This is transparent already from the outline
at the beginning of Section 2.2 in \cite{Spatzier} and from the fact
any one parameter subgroup in $D_{+}$ acts ergodicly (see Corollary
\ref{cor:1}). Furthermore, the foliation
conjugacy $H_1$ is $C^{\frac{\ell}{r}}$ close to identity where $r$
is the least integer such that the coarse Lyapunov distributions of
$\alpha_{D_+}$ as well as their commutators of length $r$ span the
tangent space at any $x\in M$. Let
$\tilde{\alpha}=H_1^{-1}\circ\alpha'\circ H_1$. The proof consists
of showing that the Hirsch-Pugh-Shub neutral foliation conjugacy
that is transversally unique can be chosen to be $C^\infty$ along
coarse Lyapunov foliations without loss of regularity. The
principal tool here is the theory of non-stationary normal forms,
see \cite{Guysinsky}. After that global smoothness follows from
general elliptic regularity results, i.e. Theorem~\ref{th:5}, as in
\cite{Spatzier}, \cite{Damjanovic2} or \cite{Margulis2}.
Conjugating $\tilde\alpha$ by this diffeomorphism $H_1$ produces an action whose neutral foliation is the same as for $\alpha_{D_+}$. Thus general local conjugacy problem for perturbations of $\alpha_{D_+}$ is reduced to considering a special class of perturbations along the neutral foliation. As we will see later the analytic machinery we use to construct the conjugacy inductively depends crucially of the special form of the linearized conjugacy equation
for such perturbations. At this step there is a crucial difference with \cite{Damjanovic4}. In the torus situation considered there the Hirsch-Pugh-Shub neutral foliation conjugacy can {\em a priori} only be shown to be smooth along coarse Lyapunov directions. But because of the lack of condition ($\mathfrak B$) elliptic regularity cannot be used and global smoothness does not follow. However, on the torus this is not too high price to pay for considering general linearized conjugacy equation: only stronger requirement for the regularity of the perturbation.
This equation still is solved with tame estimates and tame splitting is produced. This is due to the fact that smooth functions on the torus have super-polynomial decay of Fourier coefficients that result to super-exponential decay of correlations for such functions under the action of any irreducible partially hyperbolic automorphism. By contrast, in the semisimple and other cases at hand there is a particular speed of exponential decay of matrix coefficients, however smooth the functions are, and it is not sufficient to construct distribution solutions for the twisted coboundary equations that appear in the linearization of the conjugacy problem for general perturbations. We comment on this point more specifically later, see Remark~\ref{rem: exponential}.
\subsection{Reduction to a conjugacy for a single generator}\label{section:1gen}
We will say that a homeomorphism $\theta: X\rightarrow X$ {\em preserves a
foliation $\mathcal{F}$ everywhere} if
$\theta(\mathcal{F}(x))\subseteq \mathcal{F}(x)$ for any $x\in X$.
The following Lemma shows that obtaining a $C^\infty$-conjugacy
preserving the neutral foliation everywhere for one regular
generator suffices for the proofs of Theorems \ref{th:3}, \ref{th:6}
and \ref{th:4}:
\begin{lemma}\label{le:9}
Let $a\in D_+$ be a regular element for $\alpha_{D_+}$. Suppose
$\tilde{\alpha}$ is a sufficiently small $C^1$ perturbation of
$\alpha_{D_+}$ and $H$ is a $C^1$ map of $X=L\backslash M$ that is $C^1$ close to identity and preserves the neutral
foliation everywhere. If $H$ satisfies:
\begin{align}\label{for:116}
&\tilde{\alpha}(a)\circ H=H\circ \alpha(a')
\end{align}
for some element $a'\in Z(D)\bigcap Z(L)$, then $H$
conjugates the corresponding maps for all the elements of the action
i.e. there exists a homomorphism $i_0:D_+\rightarrow Z(D)\bigcap
Z(L)$ such that for all $d\in D_+$ we have
\begin{align*}
&\tilde{\alpha}(d)\circ H=H\circ \alpha(i_0(d)\cdot d).
\end{align*}
\end{lemma}
\begin{proof}
Let $d$ be any element $D_+$, other than $a$. It follows from
\eqref{for:116} and commutativity that
\begin{align*}
\tilde{\alpha}(a)\circ F=F\circ \alpha(a')
\end{align*}
where
\begin{align*}
F=\tilde{\alpha}(d)^{-1}\circ H\circ\alpha(d).
\end{align*}
Therefore
\begin{align}\label{for:63}
\alpha(a')\circ F^{-1}\circ H=F^{-1}\circ H\circ\alpha(a').
\end{align}
We can lift $\theta=F^{-1}\circ H$ to a map $\tilde{\theta}$ on $M$.
Since $\theta$ is $C^1$ close to identity and preserves the neutral
foliation everywhere, there exist a $C^1$ small map $R:M\rightarrow
\mathfrak{L}^\bot$ such that $\tilde{\theta}=\exp(R)\cdot I$. Since
$\tilde{\theta}$ is $L$-fiber preserving $R$ satisfies:
\begin{align*}
l'\cdot\exp\bigl(\text{Ad}_{l}R(l^{-1}x)\bigl)=\exp(R(x))\qquad
\text{ for any }l\in L.
\end{align*}
where $l'\in L$ is dependent on $l$. If $\norm{R}_{C^0}$ is small
enough by unique decomposition property near identity then it
follows
\begin{align}\label{for:62}
\text{Ad}_{l}R(l^{-1}x)=R(x)\qquad \text{ for any }l\in L.
\end{align}
Let $p$ be the natural projection $p: M\rightarrow L\backslash M$.
Then $$\theta(\bar{x})=p\bigl(\exp(R(x))\cdot p(x)\bigl)$$ for any
$x\in M$ satisfying $p(x)=\bar{x}$. We write $\theta=\exp(R)\cdot I$
without confusion. By \eqref{for:63} we have:
\begin{align*}
l(x)\cdot a'\cdot\exp(R(x))\cdot x=\exp(R(a'x))\cdot a'\cdot
x\qquad\text{for any }x\in M
\end{align*}
where $l:M\rightarrow L$ is a $C^1$ map, or equivalently
\begin{align*}
l\cdot \exp(\textrm{Ad}_{a'}R)=\exp(R\circ\alpha(a'))
\end{align*}
By assumption $a'\in Z(L)$ $\textrm{Ad}_{a'}R$ takes values in
$\mathfrak{L}^\bot$, then use unique decomposition property again
the above equation is reduced to
\begin{align}\label{for:117}
\textrm{Ad}_{a'}\circ R=R\circ\alpha(a').
\end{align}
If $\textrm{Ad}_{a'}$ is diagonalizable, the equation
\eqref{for:117} reduces to several equations of the following type:
\begin{align}
\lambda\omega=\omega\circ\alpha(a')
\end{align}
where $\omega$ is a $C^1$ function and $\lambda$ is an eigenvalue of
$\textrm{Ad}_{a'}$ with $\abs{\lambda}=1$.
Consider integrals on $M$ with respect to the Haar measure
$\mu$:
\begin{align*}
\int_{M} \lambda\omega d\mu=\int_{M}
\omega\circ\alpha(a')d\mu=\int_{M} \omega d\mu.
\end{align*}
Hence $\int_{M} \omega d\mu=0$ or $\lambda=1$.
If $\int_{M} \omega d\mu=0$ by Corollary \ref{cor:1} for any $f\in
C^1(M)$ there exist $C\,,\gamma>0$ such that
\begin{align}\label{for:118}
\abs{\langle\omega((a')^n),f\rangle}&\leq Ce^{-\abs{n}\gamma }\qquad
\text{ for all } n\in \mathbb{Z}.
\end{align}
While on the other hand
\begin{align}\label{for:119}
\abs{\langle\omega,f\rangle}=\abs{\langle\lambda^n\omega,f\rangle}=\abs{\langle\omega((a')^n),f\rangle},\qquad
\forall n\in\mathbb{Z},
\end{align}
so $\omega$ is a $0$ distribution. Since $\omega$ is also $C^1$,
then $\omega=0$.
if $\lambda=1$, let $\omega'=\omega-\int_{M} \omega d\mu$, then
$\omega'=\omega'\circ\alpha(a')$. Similarly to the above argument
$\omega'=0$, hence $\omega=\int_{M} \omega d\mu$. This implies $R$
is constant.
If there are Jordan blocks for $\textrm{Ad}_{a'}$, the argument
above is still sufficient to deduce that $R$ has to be constant.
Namely, if there is, say, a $3$-Jordan block, then equation
\eqref{for:117} reduces to
$$ \begin{pmatrix}\lambda & 1 & 0 \\
0 & \lambda & 1 \\
0 & 0 & \lambda \\
\end{pmatrix}\begin{pmatrix}\omega_1 \\
\omega_2 \\
\omega_3 \\
\end{pmatrix}=\begin{pmatrix}\omega_1\circ\alpha(a') \\
\omega_2\circ\alpha(a') \\
\omega_3\circ\alpha(a') \\
\end{pmatrix}$$
This implies
\begin{align*}
\lambda\omega_1+\omega_2&=\omega_1\circ\alpha(a')\\
\lambda\omega_2+\omega_3&=\omega_2\circ\alpha(a')\\
\lambda\omega_3&=\omega_3\circ\alpha(a').
\end{align*}
From the third equation above, if $\lambda\neq 1$, using deday of matrix coefficients as in the case of a simple
eigenvalue, we deduce that $\omega_3=0$. Substituting into the
second equation, we obtain $\omega_2=0$ and finally using this fact
in the first equation above, we obtain $\omega_1=0$. If $\lambda= 1$
then by the same argument as in the case of a simple eigenvalue
$\omega_3=c_3$ for some $c_3\in\mathbb{C}$. Substituting into the second
equation, we get $c_3=0$ by integrating each side and then it
follows $\omega_2=c_2$ for some $c_2\in\mathbb{C}$. Using a similar argument
again by substituting into the first equation we obtain $c_2=0$ and
then $\omega_1=c_1$ for some $c_1\in\mathbb{C}$. One can obviously by
induction obtain that $R$ is constant for Jordan blocks of any dimension;
hence it follows $R$ is constant. Since $R$ satisfies \eqref{for:62}
$\exp(R)\in Z(L)$ and therefore $F=H\circ \alpha(d')$ for some
constant $d'\in Z(L)$ and
\begin{align*}
H=\tilde{\alpha}(d)^{-1}\circ H\circ\alpha(i_0(d)\cdot d)
\end{align*}
for an arbitrary $d\in D_+$ where $i_0(d)=d'^{-1}$.
As has been explained in the previous section, the smoothness of
the diffeomorphism $H$ follows from the non-stationary normal forms
theory and Theorem~\ref{th:5}.
\end{proof}
\begin{remark} Alternatively Lemma~\ref{le:9} can be proved using
the ``geometric'' construction of solutions
for cohomological equations first introduced in \cite{Kononenko} and
developed in the setting of abelian group actions in
\cite{Damjanovic1}. We prefer the more analytic proof to demonstrate that
the machinery of periodic cycle functionals that is central for the geometric approach can be avoided
in the present setting.
\end{remark}
\subsection{Conjugacy problem and linearization}\label{sec:1}
Now we proceed to the main part of the proof of our theorem. Since
it is carried out via a KAM-type iteration scheme we first need to
deduce linearized conjugacy equation in a convenient form on
$X=L\backslash M$.
Due to Lemma~ \ref{le:9} it is sufficient to prove the existence
of a $C^1$ conjugacy $H$ preserving neutral foliation
$\mathfrak{L}^\bot$ everywhere such that
\begin{align}
\tilde{\alpha}(d)\circ H=H\circ \alpha(i_0(d))\qquad \text{ for
}\forall d\in D_+
\end{align}
for some homomorphism $i_0:D_+\rightarrow \exp(\mathfrak{N})\bigcap
Z(L)$.
Any map on $X$ can be lifted to a map on $M$. For any $d\in D_+$ the
lifts of $\tilde{\alpha}(d)$ and $H$ are
$\exp(R(d))\cdot\alpha(i_0(d))$ and $\exp(\Omega)\cdot I$
respectively. Since we are considering only perturbations along the
neutral foliation and correspondingly looking for a conjugacy that
preserve leaves of that foliation, this means that $\exp(R)$ and
$\exp(\Omega)$ take values in $\exp(\mathfrak{L}^\bot)$ and
hence
$$R,\Omega:M\rightarrow \mathfrak{L}^\bot\quad
\text{are} \quad C^\infty\quad \text{maps close to identity.}$$
Let $\bar{e}$ be the image of the
identity $e$ of $G$ or $G\ltimes\mathbb{R}^N$ on $M$. For a small
neighborhood $V$ of the tangent bundle $T_{e}G$ or
$T_{e}G\ltimes\mathbb{R}^N$, we can identify the metrics on $V$ and $\exp
V\cdot x$ for any $x\in M$.
Since $\tilde{\alpha}$ is $L$-fiber preserving $R
:M\rightarrow\mathfrak{L}^\bot$ satisfies the \emph{adjoint
$L$-invariant condition}
\begin{align}\label{for:64}
\text{Ad}_{l}R(l^{-1}x)=R(x)\qquad \text{ for any }l\in L.
\end{align}
Then $\Omega :M\rightarrow\mathfrak{L}^\bot$ is also adjoint
$L$-invariant and for any $\bar x\in X$ and any $x\in M$ such that $p(x)=\bar{x}$, where $p$ is the natural
projection $M\rightarrow L\backslash M$
\begin{align*}
H(\bar{x})&=p\bigl(\exp(\Omega(x))\cdot p(x)\bigl)\\
\tilde{\alpha}(d)(\bar{x})&=p\bigl(\exp(R(x))\cdot
\alpha(i_0(d))\cdot p(x)\bigl)\qquad \forall d\in D_+
\end{align*}
We write without confusion $H=\exp(\Omega)\cdot I$ and
$\tilde{\alpha}(d)=\exp(R(d))\cdot\alpha(i_0(d))$ for any $d\in
D_+$.
One can consider the problem of finding a conjugacy as a problem of
solving the following non-linear functional equation:
\begin{align}\label{for:96}
\exp(R\circ \tilde{H})\cdot
(i_0(d)\exp(\Omega)i_0(d)^{-1})=l\cdot\exp(\Omega\circ\alpha\circ
i_0(d)).
\end{align}
where $l:M\rightarrow L$ and $\tilde{H}$ is the lift
of $H$. Let $\mathcal{S}(R,\Omega)=R\circ
\tilde{H}+\textrm{Ad}_{i_0(d)}\Omega-\Omega\circ\alpha( i_0(d))$,
then
\begin{align*}
\exp(\mathcal{S}(R,\Omega))\exp(\textrm{Res}_1(R,\Omega))&=l
\end{align*}
where $\textrm{Res}_1(R,\Omega)$ is quadratically small with respect
to $R$ and $\Omega$. By assumption the inage of the homomorphism $i_0$ is in $Z(L)$; hence
$\mathcal{S}(R,\Omega)$ takes values in $\mathfrak{L}^\bot$. By the
unique decomposition property
\begin{align}\label{for:94}
\exp(\mathcal{S}(R,\Omega))=\exp(\textrm{Res}(R,\Omega))
\end{align}
where $\textrm{Res}(R,\Omega)$ is quadratically small with respect to
$R$ and $\Omega$.
Thus the operator $\mathcal{S}$
has the following form:
\begin{align}\label{for:95}
\mathcal{S}(R,\Omega)&=\mathcal{S}(0,0)+D_1\mathcal{S}(R,0)R\notag\\
&+D_2\mathcal{S}(0,\Omega)\Omega+\textrm{Res}_2(R,\Omega)\notag\\
&=R+\textrm{Ad}_{i_0(d)}\Omega-\Omega\circ\alpha(
i_0(d))+\textrm{Res}_2(R,\Omega)
\end{align}
where $\textrm{Res}_2(R,\Omega)$ is quadratically small with respect
to $R$ and $\Omega$. Here $D_1\mathcal{S}(R,0)$ and
$D_2\mathcal{S}(0,\Omega)$ denote Frechet derivatives of the map
$\mathcal{S}$ in the first and second variable, respectively, at the
point $(0,0)$ so that the linearization of $\mathcal{S}$ at $(0,0)$ is equal to
$ R+\textrm{Ad}_{i_0(d)}\Omega-\Omega\circ\alpha(
i_0(d))$.
Combining \eqref{for:94} and \eqref{for:95} we obtain the
linearized equation of \eqref{for:96}:
\begin{align}\label{for:97}
\Omega\circ\alpha( i_0(d))-\textrm{Ad}_{i_0(d)}\Omega=R.
\end{align}
\begin{remark}\label{rem: exponential} In deriving the algebraic form of the linearized equation we have not used specific form of our perturbation. For a general
perturbation $R$ takes values in the whole Lie algebra and hence
$\Omega$ also must take values there but the linearized equation
has the same form \eqref{for:97}. There is however a crucial
difference: for our special perturbations $\text{Ad}$ have
eigenvalues of absolute value one and, as we have seen in Lemma
\ref{le:9} and will soon see, weak
exponential estimates on the decay of matrix coefficients given by
Theorem~\ref{th:2} are sufficient to show vanishing of
functions, see \eqref{for:118} and to construct special distribution
solutions $\Omega_{\binom{+}{-}}$, see \eqref{for:28}. In general
however the eigenvalues of the $\text{Ad}$ beat the exponential
decay of matrix coefficients and the right-hand part in
\eqref{for:119} or \eqref{for:28} diverges too fast to produce even
a distribution.
\end{remark}
The equation \eqref{for:97} actually consists of infinitely many
equations corresponding to different elements $d\in D_+$ of the
action and we need to find a common approximate solution $\Omega$ to all those
equations.
Lemma \ref{le:9} shows that it is enough to produce a conjugacy for
one regular element of the action. It is clear however that, in
general, it is not possible to produce a $C^\infty$ conjugacy for a
perturbation of a single element of the action, since a single
element of a genuinely partially hyperbolic action is not even
structurally stable. Indeed, as we will soon see in Lemma
\ref{le:11} that there are infinitely many obstructions to solving
linearized equation for one generator. Therefore, we consider two
regular elements of the action, and reduce the problem of solving
the linearized equation \eqref{for:97} to solving simultaneously the
following system:
\begin{align}\label{for:101}
&\Omega\circ\alpha(a_1)-\textrm{Ad}_{a_1}\Omega=R_{a},\notag\\
&\Omega\circ\alpha(a_2)-\textrm{Ad}_{a_2}\Omega=R_{b}
\end{align}
where $a_1=i_0(a)$ and $a_2=i_0(b)$ are regular commuting elements of $Z(D)$ close
to $a$ and $b$, respectively and
\begin{align}\label{for:98}
R_{a}:=R(a)\qquad R_{b}:=R(b).
\end{align}
If the solution of the system \eqref{for:101} on $M$ has an adjoint
$L$-invariant solution then it implies that the solution can descend
to a map on $X$ and then we obtain a conjugacy on $X$. Hence throughout the paper, we only
solve the system \eqref{for:101} on $M$ and show that if $R$
satisfies adjoint $L$-invariant condition then the solution also
does.
Notice that since $\Omega $ has values in the Lie algebra
$\mathfrak{N}$ of the centralizer of the
unperturbed action
$\textrm{Ad}_{a}\Omega=\textrm{Ad}_{b}\Omega=\textrm{Id}$. Moreover,
the centralizer acts on itself by isometries. Hence
$\textrm{Ad}_{a_1}$ and $\textrm{Ad}_{a_2}$ act on $\Omega$ as
isometries close to identity.
Assume now that $\exp(R)\cdot(\alpha\circ i_0)$ is a commutative
action. Linearized form of commutativity relation gives the following {\em twisted cocycle condition}:
\begin{align}\label{twistedcocyclefirst}
R_{b}\circ
a_1-\textrm{Ad}_{a_1}R_{b}-R_{a}\circ
a_2+\textrm{Ad}_{a_2}R_{a}=0.
\end{align}
We will now justify this formal linearization by showing that \\ $L(R_{b}, R_{a})^{(a_1,a_2)}\stackrel{\text{def}}{=}R_{b}\circ
a_1-\textrm{Ad}_{a_1}R_{b}-R_{a}\circ
a_2+\textrm{Ad}_{a_2}R_{a}$
is quadratically small with respect to $R$.
\begin{lemma}\label{le:5}
Let $\tilde{\alpha}=\exp(R)\cdot(\alpha\circ i_0)$ be a commutative
$C^s$ action on $X=L\backslash M$ where $R$ has values in
$\mathfrak{L}^\bot$ and the image of the homomorphism $i_0$ is in $Z(L)\bigcap Z(D)$. If
$d(i_0(a),a)+d(i_0(b),b)\leq\eta$ and $\norm{R_{a},R_{b}}_{C^s}\leq
1$, then for $0\leq m\leq s-1$ the following inequalities hold
\begin{align*}
\norm{L(R_{b},R_{a})^{(a_1,a_2)}}_{C^{m}}\leq
C_{m,\eta}\norm{R_{a},R_{b}}_{C^{m}}\norm{R_{b},R_{a}}_{C^{m+1}}
\end{align*}
where $C_{m,\eta}$ is a constant only dependent on $m\,,\eta$.
\end{lemma}
\begin{proof}
Re-writing the commutativity condition $\tilde{a}\circ\tilde{b}=\tilde{b}\circ\tilde{a}$
in the Lie algebra terms we obtain
\begin{align*}
&\exp\bigl(R_{a}\bigl(\exp(R_{b}(x))\cdot a_2\cdot
x\bigl)\bigl)\cdot
a_1\cdot\exp(R_{b}(x))\cdot a_2\cdot x\\
&=l(x)\exp\bigl(R_{b}\bigl(\exp(R_{a}(x))\cdot a_1\cdot
x\bigl)\bigl)\cdot a_2\cdot\exp(R_{a}(x))\cdot a_1\cdot x\qquad
\forall x\in M
\end{align*}
where $l:M\rightarrow L$. Equivalently
$$\exp\bigl(R_{a}(\exp(R_{b})\cdot
a_2)\bigl)\cdot\exp(\text{Ad}_{a_1}R_{b})=l\cdot\exp\bigl(R_{b}(\exp(R_{a})\cdot
a_1)\bigl)\cdot\exp(\text{Ad}_{a_2}R_{a}).
$$
Let $\mathcal{S}=R_{a}(\exp(R_{b})\cdot
a_2)+\text{Ad}_{a_1}R_{b}-R_{b}(\exp(R_{a})\cdot
a_1)-\text{Ad}_{a_2}R_{a}$, then
\begin{align}
\exp (\mathcal{S}+\textrm{Res}(R_a,R_b))=l
\end{align}
By the Baker-Campbell-Hausdorff formula
there exists $C_{m,\eta}>0$ only dependent on $m,\eta$ such that
\begin{align*}
&\norm{\textrm{Res}(R_a,R_b)}_{C^{m}}\leq
C_{\eta,m}\norm{R_{a}}_{C^{m}}\norm{R_{b}}_{C^{m}}.
\end{align*}
By assumption $i_0$ valued in $Z(L)$ then $\mathcal{S}$ takes
values in $\mathfrak{L}^\bot$. By unique decomposition property
\begin{align}\label{for:99}
&\norm{\mathcal{S}}_{C^{m}}\leq
C_{\eta,m}\norm{R_{a}}_{C^{m}}\norm{R_{b}}_{C^{m}}.
\end{align}
Furthermore
\begin{align*}
L(R_{b},R_{a})^{(a_1,a_2)}&=R_{b}\circ
a_1-\text{Ad}_{a_1}R_{b}-R_{a}\circ
a_2+\text{Ad}_{a_2}R_{a_1}\\
&=(R_{a}(\exp(R_{b})\cdot a_2)-R_{a}\circ
a_2)-(R_{b}(\exp(R_{a})\cdot a_1)-R_{b}\circ
a_1)\\
&-(R_{a}(\exp(R_{b})\cdot
a_2)+\text{Ad}_{a_1}R_{b}-R_{b}(\exp(R_{a})\cdot
a_1)-\text{Ad}_{a_2}R_{a}),
\end{align*}
then by \cite[Appendix II]{Lazutkin} it follows that there exists
$C_{m}>0$ only dependent on $m$ such that
\begin{align}\label{for:100}
&\norm{R_{a}(\exp(R_{b})\cdot a_2)-R_{a}\circ a_2}_{C^{m}}\leq
C_{m}\norm{R_{a}}_{C^m}\norm{R_{b}}_{C^{m+1}}\notag\\
&\norm{R_{b}(\exp(R_{a})\cdot a_1)-R_{b}\circ a_1}_{C^{m}}\leq
C_{m}\norm{R_{b}}_{C^m}\norm{R_{a}}_{C^{m+1}}.
\end{align}
Then combining \eqref{for:99} and \eqref{for:100} we obtain desired estimate
\begin{align*}
\norm{L(R_{b},R_{a})^{(a_1,a_2)}}_{C^m}&\leq
2C_{m}\norm{R_{b}}_{C^m}\norm{R_{a}}_{C^{m+1}}+C_{\eta,m}\norm{R_{a}}_{C^{m}}\norm{R_{b}}_{C^{m}}\\
&\leq
(2C_{m}+C_{m,\eta})\norm{R_{a},R_{b}}_{C^{m}}\norm{R_{b},R_{a}}_{C^{m+1}}.
\end{align*}
\end{proof}
\subsection{Some notation}\label{section:notation} We try as much as possible to develop a unified systems of notations
for symmetric space examples and twisted symmetric space examples. We will use notations from this section throughout subsequent sections. So the reader
should consult this section if an unfamiliar symbol appears.
\begin{enumerate}
\item Let $U=Z(D)\times S^1_{\mathbb{C}}$, where $S^1_{\mathbb{C}}$ is the set of complex numbers of absolute value one.
For any $x\in U$, $B_\eta(x)\stackrel{\text{def}}{=}\{y\in
U|d(x,y)\leq\eta\}$; for any $y\in Z(D)$,
$B_\eta(y)\stackrel{\text{def}}{=}\{x\in Z(D)|d(x,y)\leq\eta\}$. \\
\item Fix positive definite inner products $\langle\cdot\rangle_1$ on $\mathfrak{G}$ and
$\langle\cdot\rangle_2$ on $\mathbb{R}^N$ that are invariant under
$\text{Ad}_K$ and $\rho(K)$ respectively. Let
$\langle\cdot\rangle=\langle\cdot\rangle_1+\langle\cdot\rangle_2$
and $\mathfrak{L}^\bot$ be the orthogonal complement of
$\mathfrak{L}$ in $\mathfrak{N}$.
In the setting of twisted symmetric space examples for $X+v\in \mathfrak{G}+\mathbb{R}^N$ where $X\in
\mathfrak{G}$ and $v\in\mathbb{R}^N$, let
$\norm{X+v}\stackrel{\text{def}}{=}\max\{\norm{X}_1,\norm{v}_2\}$,
where $\norm{\cdot}_1$ and $\norm{\cdot}_2$ are induced by
$\langle\cdot\rangle_1$ and $\langle\cdot\rangle_2$ respectively. \\
\item For symmetric space examples,
let $X_1,\cdots,X_{p}$ be a base of
$\{\mathfrak{g}^\phi\}_{\phi\in\Phi}$, the root spaces of the Cartan
action of $D$, such that their commutators of length at most $r$
span $\mathfrak{K}+\mathfrak{D}$. Let $Y_1,\cdots,Y_{q}$ be base of
$\mathfrak{N}$ generated by these commutators.
For twisted symmetric space examples let
$v_1,\cdots,v_{N-N_0}$ be a base of
$\{\mathfrak{v}^\mu\}_{\mu\in\Phi_1\backslash 0}$ the non-zero
weight spaces of the Cartan action of $D$, such that together with
$X_1,\cdots,X_{p}$ their commutators of length at most $r$ span
$\mathfrak{N}=\mathfrak{K}+\mathfrak{D}+\mathbb{R}^{N_0}$. Let
$u_1\cdots,u_{N_0},Y_1,\cdots,Y_{q}$ be a base of $\mathfrak{N}$
generated by these
commutators.\\
\item For any
function $f$ on $M$, $v\in \mathfrak{G}$ (correspondingly $v\in
\mathfrak{G}+\mathbb{R}^{N}$), $v^m (f)$ denotes the the $m$'th partial
derivative along direction $v$ if it exists.\\
\item\label{item: incompleteSobolev} Let
$m> r$ be an integer. For symmetric space examples, let
$\mathcal{L}^m$ to be the subspace of $L^2(M)$ such that $f$ and
$X_j^k (f)$ exist as $L^2$ functions for $1\leq j\leq p\,,k\leq
m$; for twisted symmetric space examples, let $\mathcal{L}^m$ be
the subspace of $L^2(M)$ such that $f$ $X_j^k (f)$ and $v_i^k (f)$
exist as $L^2$ functions for $1\leq j\leq p\,,1\leq i\leq
N-N_0\,,k\leq m$.
By Theorem \ref{th:5}, $\mathcal{L}^m$ for $m>r$ can be made into a Hilbert space
with the norm
\begin{align}\label{eq: normmodSobolev}
\norm{f}'_m\stackrel{\text{def}}=(\sum_{j=1}^{p}\sum_{i=1}^m\norm{X_j^if}_0^2+\norm{f}_0^2)^{1/2}
\end{align}
for symmetric space examples and
\begin{align}\label{eq: normmodSobolev1}
\norm{f}'_m\stackrel{\text{def}}=\bigl(\sum_{n=1}^{N-N_0}\sum_{j=1}^{p}\sum_{i=1}^m\norm{X_j^if}_0^2+\norm{v_n^i
(f)}_0^2+\norm{f}_0^2\bigl)^{1/2}
\end{align}
for twisted symmetric space examples.
Denote by $\mathcal{H}^s$ the $s'$th Sobolev
space of $M$ with Sobolev norm $\norm{\cdot}_s$ and let
$\norm{\cdot}_{C^r}$ stand for $C^r$ norm for functions on $M$. Let
$\mathcal{L}_0^r\stackrel{\text{def}}{=}\{f\in\mathcal{L}^r\mid\int_{M}f=0\}$
and
$\mathcal{H}_0^r\stackrel{\text{def}}{=}\{f\in\mathcal{H}^r\mid\int_{M}f=0\}$.
Assume $f\in \mathcal{H}^{s+\sigma}$ where $\sigma>\frac{\dim
M}{2}+1$ and $s\in\mathbb{N}$. The following relations hold by Sobolev
embedding theorem
\begin{align*}
\norm{f}_{s}\leq C\norm{f}_{C^s}\qquad\text{and
}\qquad\norm{f}_{C^{s}}\leq C_s\norm{f}_{s+\sigma}
\end{align*}
In particular, one
may take $\sigma= \frac{\dim M}{2}+1+\delta$ with small $\delta>
0$.\\
\item For a map $\mathcal{F}$ with coordinate functions $f_i$,
$1\leq i\leq n_0$ define: $\norm{\mathcal{F}}_r=\max_{1\leq i\leq
n_0}\norm{f_i}_r$ and $\int_M \mathcal{F}=(\int_M
f_1d\mu,\cdots,\int_M f_{n_0}d\mu)$ where $\mu$ is the Haar measure.
For two maps $\mathcal{F}$, $\mathcal{G}$ define
$\norm{\mathcal{F},\mathcal{G}}_r=\max\{\norm{\mathcal{F}}_r,\norm{\mathcal{G}}_r\}$.
$\norm{\mathcal{F}}_{C^r}$, $\norm{\mathcal{F}}'_{r}$ and
$\norm{\mathcal{F},\mathcal{G}}_{C^r}$ are defined similarly. We
write $\mathcal{F}\in C^r$ if $f_i\in C^r$ for $1\leq i\leq n_0$.
$\mathcal{F}\in\mathcal{H}^s$, $\mathcal{F}\in\mathcal{H}_0^s$,
$\mathcal{F}\in\mathcal{L}^s$ and $\mathcal{F}\in\mathcal{L}_0^s$
are defined similarly. We say that $\mathcal{F}$ is a distribution
if coordinates $f_i$ are distribution, $1\leq i\leq n_0$.
For any connected subgroup $L\subseteq K$, let $\mathcal{S}_L$ be
the set of maps from $M$ to $\mathfrak{L}^\bot$.
$$\mathcal{L}_{0,L}^r\stackrel{\text{def}}{=}\{\mathcal{F}\in\mathcal{L}_0^r\bigcap \mathcal{S}_L\mid\text{Ad}_{l^{-1}}\mathcal{F}(lx)=\mathcal{F}(x)\}, \forall l\in L$$
and
$$\mathcal{H}_{0,L}^r\stackrel{\text{def}}{=}\{\mathcal{F}\in\mathcal{H}_0^r\bigcap \mathcal{S}_L\mid\text{Ad}_{l^{-1}}\mathcal{F}(lx)=\mathcal{F}(x)\}, \forall l\in
L.$$\\
\item For the symmetric space examples, any $z\in Z(D)$ can be
decomposed as $z=dk$ where $\log d\in\mathfrak{D}$ and $\log
k\in\mathfrak{K}$; for the twisted symmetric space examples, any
$z\in Z(D)$ can be decomposed as $z=dkv$ where $\log
d\in\mathfrak{D}$, $\log k\in\mathfrak{K}$ and $v\in \mathfrak{v}^0$. In either
case we call $d$ {\em the split part} of $z$ and $k$ {\em the compact part} of
$z$. We call $a_1,a_2\in Z(D)$ linearly independent if their splits
are linearly independent over $\mathbb{R}$. If $a_1,a_2\in Z(D)$ are linearly independent with split parts $d_1$ and
$d_2$ respectively then $\sum_{\phi\in\Phi}\abs{\phi(j_1\log
d_1+j_2\log d_2)}>0$ for any $j_1,j_2\in\mathbb{R}$ with $j_1^2+j_2^2\neq
0$. \\
\item For $(z,\lambda)\in U$ and a function $f$ on $M$,
define the twisted coboundary operators:
\begin{align*}
(z,\lambda)^\tau f=f(z)-\lambda f
\end{align*}
In what follows $\lambda$ will be either equal to $1$ or to an eigenvalue of
$\textrm{Ad}_k$ or $\rho(k)$, where $k$ is the compact part of $z$.
Recall that $|\lambda|=1$.
For any $s>0$, $(z,\lambda)^\tau$
is a bounded linear operator on $\mathcal{L}_0^s$. Denote the operator norm by $\norm{(z,\lambda)^\tau}'_s$.\\
\item\label{item(10)} Let $f$ be a function or a distribution. We introduce notations for the following formal sums:
\begin{align*}
&\sum_{+}^{(z,\lambda)}f\stackrel{\text{def}}{=}-\sum_{j\geq
0}\lambda^{-(j+1)}f\circ z^j\\
&\sum_{-}^{(z,\lambda)}f\stackrel{\text{def}}{=}\sum_{j\leq -1
}\lambda^{-(j+1)}f\circ z^j\\
&\sum^{(z,\lambda)}f\stackrel{\text{def}}{=}\sum_{j=-\infty
}^{+\infty}\lambda^{-(j+1)}f\circ z^j
\end{align*}\\
\item For $z\in Z(D)$ and a map $\mathcal{F}:M\rightarrow \mathfrak{N}$,
define the twisted coboundary operator:
\begin{align*}
\mathcal{T}_{z}\mathcal{F}=\mathcal{F}(z)-\text{Ad}_{z}\mathcal{F}.
\end{align*}
Similarly to the scalar case we introduce notations for the following formal sums:
\begin{align*}
\sum^{z}\mathcal{F}\stackrel{\text{def}}{=}\sum_{j=-\infty
}^{+\infty}\text{Ad}_{z}^{-(j+1)}\mathcal{F}\circ z^j.
\end{align*}
\item For two functions $\theta\,,\varphi$ and maps
$\mathcal{F}\,,\mathcal{G}:M\rightarrow \mathfrak{N}$,\\ for
$(a_1,\lambda_1)\,,(a_2,\lambda_2)\in U$ define the following
operators:
\begin{align*}
&L(\theta,\varphi)^{(a_1,a_2)}_{(\lambda_1,\lambda_2)}\stackrel{\text{def}}{=}(a_2,\lambda_2)^\tau
\theta-(a_1,\lambda_1)^\tau\varphi\\
&L(\mathcal{F},\mathcal{G})^{(a_1,a_2)}\stackrel{\text{def}}{=}\mathcal{F}\circ
a_2-\textrm{Ad}_{a_2}\mathcal{F}-\mathcal{G}\circ
a_1+\textrm{Ad}_{a_1}\mathcal{G}=\mathcal{T}_{a_2}\mathcal{F}-\mathcal{T}_{a_1}\mathcal{G}.
\end{align*}
We will sometimes use generic notation $L(\mathcal{F},\mathcal{G})$ when the generators are either not specified or clear from the context.
\\
\item In what follows, $C$ will denote any constant that depends only
on the given action $\alpha_{D_+}$ with two linearly independent
regular generators and on the dimension of group $G$.
$C_{x,y,z,\cdots}$ will denote any constant that in addition to the
above depends also on parameters $x, y, z,\cdots$.\\
\item Smoothing operators and some norm inequalities\\ The space $M$
is compact, then there exists a collection of smoothing operators
$S_t:C^\infty(M)\rightarrow C^\infty(M)$, $t > 0$, such that the
following holds:
\begin{align*}
\norm{S_tf}_{C^{s+s'}}&\leq C_{s,s'}t^{s'}\norm{f}_{C^s}\notag\\
\norm{(I-S_t)f}_{C^{s-s'}}&\leq C_{s,s'}t^{-s'}\norm{f}_{C^s}\notag\\
\norm{S_tf}_{s+s'}&\leq C_{s,s'}t^{s'}\norm{f}_{s}\notag\\
\norm{(I-S_t)f}_{s-s'}&\leq C_{s,s'}t^{-s'}\norm{f}_{s}.
\end{align*}
\end{enumerate}
\section{Solution of the linearized equation}\label{section:linear}
We consider scalar equations that appear as projections of the linearized conjugacy equations on common eigenspaces for two commuting $\textrm{Ad}$ operators.
\subsection{Cohomological stability}\label{sec:cohstability}
First we define obstructions to solvability of the linearized conjugacy equation, i.e. twisted coboundary equation, for a single element and show that
vanishing of those obstructions implies solvability of the equation
with tame estimates wrt. Sobolev norms. The latter property is an
instance of {\em cohomological stability}, the notion first
defined in \cite{Katok-constructions}.
Scheme of the proof is as follows:
\begin{enumerate}
\item Decay estimates for matrix coefficients imply existence of two distribution solutions obtained
by iteration in positive and negative directions: one of those solutions is differentiable along stable directions and the other along unstable directions.
\item Vanishing of the obstructions implies that those distribution solutions coincide.
\item Condition $(\mathfrak B)$ allows to apply elliptic regularity and deduce that solution is really a smooth function. At his stage however there is a large loss of regularity, roughly from
$m$th norm to $\frac{m}{r}$th norm.
\item Since solution along the stable and unstable directions is given by explicit exponentially converging ``telescoping sums'' they can be differentiated without loss of regularity. Up to this point the proof follows the same general scheme as in \cite{Spatzier1} although we obtain more elaborate information about estimates in different norms.
\item Remaining directions are those of the centralizer of the acting element; in particular, the adjoint representation acts by isometries and hence derivatives of all orders in those direction are bounded.
Tame estimates follow form that and from the fact that those vector-fields can be expressed as polynomial of hyperbolic ones, i.e. from condition $(\mathfrak B)$.
\end{enumerate}
\begin{lemma}\label{le:11}
Let $a$ and $b$ be two regular generators for the unperturbed
action $\alpha_{D_+}$. Let $(z,\lambda)\in B_\eta(a,1)\bigcup
B_\eta(b,1)$ with $\eta$ small enough and
$\theta\in\mathcal{H}_0^m$, $m\in\mathbb{N}$ and $m\geq m_i$, where
\begin{equation}\label{eqm1} m_1\stackrel{\rm{def}}{=}\max\{r\dim M/2+r^2+4r,r\dim
M/2+rr_0+4r,r_0r+4r\}\end{equation} where $r_0$ is a positive constant defined
in \eqref{for:31} and \eqref{for:102} below respectively,
depending on the eigenvalues of $\textrm{Ad}_a$ and $\textrm{Ad}_b$.
If
\begin{align}\label{for:61}
\sum_{j=-\infty}^{+\infty}\lambda^{-(j+1)}\theta\circ z^j=0
\end{align}
as a
distribution, then the equation
\begin{align}\label{eq:10}
\Omega\circ z-\lambda\Omega=\theta
\end{align}
have a solution $\Omega\in\mathcal{H}_0^{m-r-1}$ and the following
estimate holds
\begin{align}\label{for:60}
\norm{\Omega}_{m-r-1}\leq C_{m,\eta} \norm{\theta}_{m}.
\end{align}
\end{lemma}
We present a detailed proof for the symmetric space examples showing in particular how the
above scheme is implemented, and then describe additional ingredients that appear in the twisted cases.
\begin{proof}[Proof for symmetric space examples]\*
\noindent{\small{\bf Distribution solution.}} To find the solution $\Omega$ let us first show that the formal
solutions
\begin{align}\label{for:28}
\Omega_{\binom{+}{-}}=\binom{-}{+}\sum_{\binom{j\geq0}{j\leq
-1}}\lambda^{-(j+1)}\theta\circ z^j
\end{align}
are distributions.
Denote $z=dk$ where $d$ is the split part and $k$ is the compact
part. Since $m>\frac{\dim M}{2}+2$ then $\theta\in C^{1}(M)$.
Let $g\in C^\infty(G/\Gamma)$. By Corollary \ref{cor:1}, there exist
constants $\gamma,E>0$ only dependent on $G$ such that
\begin{align}\label{for:38}
\abs{\langle \lambda^{-(j+1)}\theta(z^j),g\rangle}\leq
E\norm{\theta}_{C^1}\norm{g}_{C^1}e^{-\gamma \abs{j}l(z)}
\end{align}
where $l(z)=\frac{1}{2}\sum_{\phi\in\Phi}\abs{\phi(d)}$. Hence
$\sum_{j=0}^{+\infty}\langle \lambda^{-(j+1)}\theta(z^j),g\rangle$
converges absolutely, and there is a constant $C> 0$ such that
$\abs{\sum_{j=0}^{+\infty}\langle
\lambda^{-(j+1)}\theta(z^j),g\rangle}\leq C\norm{g}_{C^1}$. Thus
$\Omega_{+}$ and similarly $\Omega_{-}$ are distributions. By
assumption $\sum^{(z,\lambda)}\theta=0$ hence
$\Omega\stackrel{\text{def}}{=}\Omega_{-}=\Omega_{+}$. This gives a
formal solution $\Omega$.
\medskip
\noindent{\small{\bf Smoothness in stable and unstable directions.}} Next we will show differentiability of $\Omega$ along the
stable and unstable foliations of $z$ by using
both of its forms.
Let $\phi$ be the root corresponding to $X_i$. If $\phi(\log d)<0$, we may use the $\Omega_{+}$ form for
the solution to obtain the following bound on $s$'th derivative
\begin{align}\label{for:30}
\sum_{j=0}^\infty X_i^s(\lambda^{-(j+1)}\theta\circ
z^j)=\sum_{j=0}^\infty e^{sj\phi(\log d)}\lambda^{-(j+1)}
Z^s_j(\theta)\circ z^j
\end{align}
where $Z_j=\textrm{Ad}^j_k(X_i)$ for all $1\leq s\leq m$. Notice that
$\norm{X_i}=\norm{\textrm{Ad}_k(X_i)}$ hence the left-hand side of
\eqref{for:30} converges absolutely in $L^2$ norm on $M$ and we get
\begin{align}\label{for:37}
\norm{X_i^s(\Omega)}_0\leq C_{s,\eta} \norm{\theta}_s.
\end{align}
Similarly, if $\phi(\log d)>0$ using the form $\Omega=\Omega_{-}$,
the estimate $\norm{X_i^s(\Omega)}_0\leq C_{s,\eta} \norm{\theta}_s$
holds if $1\leq s\leq m$.\medskip
\noindent{\small{\bf Smoothness of the solution.}}
By Theorem \ref{th:5}, $\Omega\in \mathcal{H}_0^{\frac{m}{r}-1}$ and
it follows that
\begin{align}\label{for:5}
\norm{\Omega}_{\frac{m}{r}-1}\leq
C_{m,\eta}(\norm{\theta}_{m}+\norm{\Omega}_0).
\end{align}
By assumption $\frac{m}{r}-1-\frac{\dim M}{2}-2\geq 1$, then
$\Omega\in C^1(M)$. Using Sobolev embedding theorem together with
\eqref{for:5} we have
\begin{align}\label{for:14}
\norm{\Omega}_{C^1}\leq C\norm{\Omega}_{\frac{\dim M}{2}+3}\leq
C\norm{\Omega}_{\frac{m}{r}-1}\leq
C_{m,\eta}(\norm{\theta}_{m}+\norm{\Omega}_0).
\end{align}
By Corollary \ref{cor:1}, we have
\begin{align*}
\norm{\Omega}_0^2=\abs{\langle
\sum_{j=0}^{+\infty}\lambda^{-(j+1)}\theta(z^j),\Omega\rangle}\leq
\sum_{j=0}^{+\infty}\abs{\langle \theta(z^j),\Omega\rangle}\leq
A_0(\norm{\Omega}_0+\norm{\Omega}_{C^1})
\end{align*}
where $0< A_0\leq C_{\eta}\norm{\theta}_{C^1}$. Together with
\eqref{for:14} we have
\begin{align}\label{for:120}
\norm{\Omega}_0\leq C_{m,\eta}\norm{\theta}_{m}.
\end{align}
Substituting \eqref{for:120} into \eqref{for:5} it follows
\begin{align}\label{for:11}
\norm{\Omega}_{\frac{m}{r}-1}\leq C_{m,\eta}\norm{\theta}_{m}.
\end{align}
\noindent{\small{\bf Tame estimates in stable and unstable directions.}} Let $x$ stand for either $a$ or $b$. Define $r_0$ to be the minimal
positive integer satisfying:
\begin{align}\label{for:31}
&\chi(\log x)+r_0\phi(\log x)>0 \qquad\text{ if }\phi(\log x)>0\notag\\
&\chi(\log x)+r_0\phi(\log x)<0 \qquad\text{ if }\phi(\log x)<0
\end{align}
for any $\chi,\phi\in\Phi$.
Let $\eta$ be small enough such that
\begin{align}\label{for:102}
&\chi(\log g)+r_0\phi(\log g)>0 \qquad\text{ if }\phi(\log g)>0\notag\\
&\chi(\log g)+r_0\phi(\log g)<0 \qquad\text{ if }\phi(\log g)<0
\end{align}
for any $\chi,\phi\in\Phi$ and any $g\in Z(D)$ with either
$d(g,a)\leq \eta$ or $d(g,b)\leq \eta$.
By assumption $\frac{m}{r}-1>\frac{\dim M}{2}+r_0+2$, then
$\Omega\in C^{r_0+1}(M)$ with bound
\begin{align}\label{for:26}
\norm{\Omega}_{C^{r_0+1}}\leq C\norm{\Omega}_{\frac{m}{r}-1}\leq
C_{m,\eta}\norm{\theta}_{m}.
\end{align}
Fix $X_i$, $1\leq i\leq p$ with corresponding root $\phi$. If
$\phi(\log d)<0$ then
\begin{align*}
X_i\Omega=-\sum_{j\geq0}e^{j\phi(\log
d)}\lambda^{-(j+1)}(Z_j\theta)\circ z^j
\end{align*}
where $Z_j=\textrm{Ad}^j_k(X_i)$.
Note that $\norm{X_i}=\norm{Z_j}$ for all $j$, then for any $1\leq
j\leq q$ we have
\begin{align}\label{for:20}
\norm{Y_j^s(X_i\Omega)}_0\leq C_{m,\eta}\norm{\theta}_{m}\qquad
\text{if } s\leq m-1.
\end{align}
For any $X_j$, $1\leq j\leq p$ with the corresponding root $\chi$,
if $\chi(\log d)<0$ we have
\begin{align}\label{for:3}
\norm{X_j^s(X_i\Omega)}_0\leq
C_{m,\eta}\norm{\theta}_{m},\qquad\text{ if }s\leq m-1.
\end{align}
For any $X_l$, $1\leq l\leq p$ with the corresponding root $\nu$, if
$\nu(\log d)>0$, let $P_j=\textrm{Ad}^j_k(X_l)$ and let
$P_j^{r_0}Z_j$ act on each side of equation \eqref{eq:10} we get
\begin{align}
&e^{\phi(\log d)+r_0\nu(\log d)}(P_{j+1}^{r_0}Z_{j+1}\Omega)\circ
z-\lambda P_j^{r_0}Z_j\Omega= P_j^{r_0}Z_j\theta
\end{align}
where each side is continuous function for any $j\in\mathbb{Z}$ by
\eqref{for:26}.
By assumption of $r_0$, $e^{\phi(\log d)+r_0\nu(\log d)}>1$,
then
\begin{align}
X_l^{r_0}(X_i\Omega)&=P_0^{r_0}Z_0(X_i\Omega)\notag\\
&=\sum_{j\leq -1}e^{j\phi(\log d)+jr_0\nu(\log
d)}\lambda^{-(j+1)}(P_j^{r_0}Z_j\theta)\circ z^j.
\end{align}
Then if $0\leq s\leq m-1-r_0$ we have
\begin{align}\label{for:1}
\norm{X_l^{r_0+s}(X_i\Omega)}_0\leq C_{m,\eta}\norm{\theta}_{m}.
\end{align}
By assumption $\frac{m}{r}-1>r_0+2$ and \eqref{for:11} we also have
\begin{align}\label{for:4}
\norm{X_l^{s}(X_i\Omega)}_0\leq C_{m,\eta}\norm{\theta}_{m}\qquad
\text{if } s\leq r_0.
\end{align}
Combine \eqref{for:20}, \eqref{for:3}, \eqref{for:1} and
\eqref{for:4} by Theorem \ref{th:5} $X_i\Omega\in
\mathcal{H}_0^{m-2}$ with estimate
\begin{align}
\norm{X_i\Omega}_{m-2}\leq C_{m,\eta}\norm{\theta}_{m}.
\end{align}
If $\phi(d)>0$, the arguments follows in a similar way. Similarly we
get
\begin{align}\label{for:29}
\norm{X_j\Omega}_{m-2}\leq C_{m,\eta}\norm{\theta}_{m}
\end{align}
for all $1\leq j\leq p$.
\medskip
\noindent{\small{\bf Tame estimates in neutral directions.}}
In the universal enveloping algebra
$U(\mathfrak{G})$ let $p_j$ be a polynomial with degree no greater than
$r$ such that $p_j(X_{j(1)},\cdots,X_{j(i)})=Y_j$, $1\leq j\leq
q$. Such a polynomial exists due to the condition $(\mathfrak B)$. Note that
\begin{align}
Y_j^s(\Omega)=Y_j^{s-1}p_j(X_{j(1)},\cdots,X_{j(i)})\Omega
\end{align}
Recall that $C^0$ norms of all powers of $Y_j$ are uniformly bounded since $Y_j$ is the generates action by isometries.
By \eqref{for:29} if $s-1+r-1\leq m-2$ we have
\begin{align*}
\norm{Y_j^s(\Omega)}_0\leq
C\sum_{j(i)}\norm{X_{j(i)}\Omega}_{r-1+s-1}\leq
C_{m,\eta}\norm{\theta}_{m}
\end{align*}
combine with \eqref{for:37} by Theorem \ref{th:5} $\Omega\in
\mathcal{H}_0^{m-r-1}$ with estimate
\begin{align}
\norm{\Omega}_{m-r-1}\leq C_{m,\eta}\norm{\theta}_{m}.
\end{align}
It follows that there exists a solution $\Omega$ to equation
\eqref{eq:10} stratifying estimates \eqref{for:60} providing the
condition \eqref{for:61} is satisfied. \end{proof}
\begin{proof}[Proof for twisted symmetric space examples] We assume
notations in the previous part if there is no confusion.
(i) If $\rho$ is Anosov then the neutral distribution is still
$\mathfrak{D}+\mathfrak{K}$. Then $z=(dk,0)$ where $d$ is the split
part and $k$ is the compact part. Notice for any $(g,t)\in G\times
\mathbb{R}^N$, $z\cdot(g,t)=(dkg,t)$, then the statement follows
essentially verbatim as in the symmetric space examples.
(ii) If $\rho$ is genuinely partially hyperbolic then the neutral
distribution is $\mathfrak{D}+\mathfrak{K}+\mathfrak{v}^0$. Let $z=(dk,t_0)$
where $d$ is the split part, $k$ is the compact part and $t_0\in
\mathfrak{v}^0$ then $z^j=(d^jk^j,t_j)$ where
$t_j=\sum_{i=0}^{j-1}\rho(k)^{-i}t_0$ for any $j$.
First let us show that two formal solutions $\Omega_{\binom{+}{-}}$ in
\eqref{for:28} are distributions. By Corollary
\ref{cor:1}, there exist constants $\gamma,E>0$ only dependent on
$G$ such that for any $f\in
C^\infty(G\ltimes\mathbb{R}^N/\Gamma\ltimes\mathbb{Z}^N)\,,j\in\mathbb{Z}$ we have
\begin{align}\label{for:69}
&\abs{\langle \lambda^{-(j+1)}\theta(z^j
x),f(x)\rangle}\notag\\
&=\abs{\langle
\theta((dk)^j,0)x),f((0,-t_j)x)\rangle}\notag\\
&\leq E\norm{\theta}_{C^1}\norm{f((0,-t_j)x)}_{C^1}e^{-\gamma
\abs{j}l(z)}\notag\\
&\leq CE\norm{t_0}\norm{\theta}_{C^1}\norm{f}_{C^1}\abs{j}e^{-\gamma
\abs{j}l(z)}
\end{align}
where
$l(z)=\frac{1}{2}\sum_{\phi\in\Phi}\abs{\phi(d)}$. Hence
$\sum_{j=0}^{+\infty}\langle \lambda^{-(j+1)}\theta(z^j),f\rangle$
converges absolutely, and thus $\Omega_{+}$ and $\Omega_{-}$ are
distributions. By the assumption $\sum^{(z,\lambda)}\theta=0$ we have
$\Omega\stackrel{\text{def}}{=}\Omega_{-}=\Omega_{+}$.
Note that for any $X_i$, $1\leq i\leq p$ with corresponding root
$\phi$ we have
\begin{align}\label{for:51}
\textrm{Ad}_z(X_i)&=e^{\phi(\log d)}\textrm{Ad}_{(k,t_0)}(X_i)\notag\\
&=(e^{\phi(\log d)}\textrm{Ad}_k(X_i),-e^{\phi(\log
d)}\rho(k)d\rho(X_i)t_0)
\end{align}
and for any $v_i$, $1\leq i\leq N-N_0$ with corresponding weight
$\mu$ we have
\begin{align}\label{for:42}
\textrm{Ad}_z(v_i)=e^{\mu(\log d)}\rho(k)v_i=e^{\mu(\log
d)}\textrm{Ad}_{(k,t_0)}(v_i),
\end{align}
the remaining part follows almost the same way as in the
symmetric space examples if we can show
$\norm{\textrm{Ad}_{(k,t_0)}(X_i)}$ and
$\norm{\textrm{Ad}_{(k,t_0)}(v_i)}$ are bounded for any $j\in\mathbb{Z}$.
It is obvious for the latter one since
$\norm{\textrm{Ad}_{(k,t_0)}(v_i)}=\norm{\rho(k)v_i}=\norm{v_i}$ by
assumption.
For the first one if $\eta$ is small enough then
\begin{align*}
\norm{\rho(k)d\rho(X_i)t_0}=\norm{d\rho(X_i)t_0}\leq
\norm{X_i}=\textrm{Ad}_k(X_i)
\end{align*}
then we have
\begin{align*}
\norm{\textrm{Ad}_{(k,t_0)}(X_i)}= \norm{\textrm{Ad}_k(X_i)}.
\end{align*}
\end{proof}
\begin{corollary}\label{cor:2}
If the equation
\begin{align}\label{eq:13}
\Omega\circ z-\lambda\Omega=\theta
\end{align}
has a solution in $C^1(M)\bigcap \mathcal{H}_0^0$ then it is unique.
\end{corollary}
\begin{proof}
If $\Omega'\in C^1(M)\bigcap \mathcal{H}_0^0$ is also a solution of
equation \eqref{eq:13}, then $\lambda^{-j}(\Omega-\Omega')\circ
z^j=\Omega-\Omega'$ for any $j\in\mathbb{Z}$. Using Corollary \ref{cor:1}
as in \eqref{for:38} and \eqref{for:69} the left side is a $0$
distribution. Since $\Omega-\Omega'\in C^1(M)$ then
$\Omega=\Omega'$.
\end{proof}
\begin{lemma}\label{le:16}
Let $\eta$ be sufficiently small and $m\in\mathbb{N}$ with $m\geq m_1+r+1$
where $m_1$ is defined as in Lemma \ref{le:11}. For any $z\in
B_\eta(a)\bigcup B_\eta(b)$ and any $\mathcal{H}_0^m$ map
$\mathcal{F}:M\rightarrow \mathfrak{N}$. If
\begin{align}\label{for:122}
\sum_{j=-\infty}^{+\infty}\Ad_z^{-(j+1)}(\mathcal{F}\circ
z^j)=0
\end{align}
as a
distribution, then the equation
\begin{align}\label{eq:2}
\Lambda\circ z-\Ad_z\Lambda=\mathcal{F}
\end{align}
have a solution $\Lambda\in\mathcal{H}_0^{m-2r-2}$ and the following
estimate holds
\begin{align}\label{for:123}
\norm{\Lambda}_{m-2r-2}\leq C_{m,\eta} \norm{\theta}_{m}.
\end{align}
Furthermore if $z\in Z(L)$ and $\mathcal{F}\in
\mathcal{H}_{0,L}^{m}$, then $\Lambda\in\mathcal{H}_{0,L}^{m-2r-2}$.
\end{lemma}
\begin{proof}[Proof for symmetric space examples and hyperbolic twisted
examples] Let \\ $\mathfrak N_{\mathbb C}$ be the complexification of
the subalgebra $\mathfrak N=\mathfrak{K}+\mathfrak{D}$. There exists an orthonormal basis that diagonalizes
$\textrm{Ad}_{z}$. As usual, this basis may be chosen to consists
of several real vectors and several pairs of complex conjugate
vectors. Equations \eqref{for:122} and \eqref{eq:2} split into
finitely many equations of the form
\begin{align}\label{for:124}
\sum_{j=-\infty}^{+\infty}\lambda^{-(j+1)}\varphi\circ z^j=0
\end{align}
and
\begin{align}\label{eq:7}
\omega\circ z-\lambda\omega=\varphi
\end{align}
where $\varphi$ is a $\mathcal{H}_0^m$ function and $\lambda$ in
$S^1_\mathbb{C}$ is corresponding eigenvalue of $\textrm{Ad}_{z}$. Notice
that since the coefficients and the right-hand part of the equation
\eqref{eq:2} are real-valued, the unique solution in $C^1(M)\bigcap
\mathcal{H}_0^0$ (see Corollary~\ref{cor:2}) is real-valued as well.
Then the conclusion follows directly from Lemma
\ref{le:11}.\qed
\smallskip
\noindent{\em Proof for genuinely partially hyperbolic twisted
examples}. Now
the neutral distribution is
$\mathfrak{N}=\mathfrak{D}+\mathfrak{K}+\mathfrak{v}^0$. Let
$z=(dk,t)$ where $d$ is the split part, $k$ is compact part and
$t\in \mathfrak{v}^0$. Consider as before the complexification
$\mathfrak{N}_\mathbb{C}$ of $\mathfrak{N}$. By \eqref{for:51} and
\eqref{for:42}
there
exists an orthonormal basis in $\mathfrak{N}_\mathbb{C}$ such that $\textrm{Ad}_{z}$
and $\textrm{Ad}_{z}^{-1}$ are represented in that basis by matrices $J_1$ and $J_2$ where
$$ J_i=\begin{pmatrix}A_i & 0 & 0\\
B_i & D_i & 0\\
0 & 0 & E_i
\end{pmatrix}.$$
Here $E_i=I_{\dim(\mathfrak{D})}$, $A_i$ are
$\dim\mathfrak{K}\times \dim\mathfrak{K}$
diagonal matrices,
$D_i$ are $\dim\mathfrak{v}^0\times \dim\mathfrak{v}^0$
diagonal matrices all of
whose eigenvalues are of absolute value $1$ all elements of
$B_i$ have absolute value smaller than $1$ if $\eta$ is small
enough. This basis can be chosen as in the symmetric space case to
include real vectors and pairs of complex conjugate vectors.
Then equations \eqref{eq:2} have the form:
\begin{align}\label{eq:3}
&\Lambda\circ z-J_1\Lambda=\Theta
\end{align}
and condition \eqref{for:122} can be written as
\begin{align}\label{for:125}
\sum_{j=-\infty}^{+\infty}J_1^{-(j+1)}\Theta\circ z^j=0.
\end{align}
We will show that the formal solutions
\begin{align}\label{for:132}
\Lambda_{\binom{+}{-}}=\binom{-}{+}\sum_{\binom{j\geq0}{j\leq
-1}}J_1^{-(j+1)}\Theta\circ z^j
\end{align}
are in fact $\mathcal{H}_0^{m-2r-2}$ solutions. Denote by $\Lambda_-^i$ and
$\Lambda_+^i$ the $i$-th coordinates of $\Lambda_-$ and $\Lambda_+$
respectively.
Denote entries of the matrices $J_1$ and $J_2$ by $q_1^{ij}$ and $q_2^{ij}$ correspondingly and let
$q_1^{ii}=\lambda_i$ for $1\leq i,j\leq t_0$ where $t_0=\dim
\mathfrak{D}+\dim \mathfrak{v}^0+\dim \mathfrak{K}$. By simply
comparing coefficients, it is easy to obtain the following relation
between the coefficients of $\textrm{Ad}_{z}$ and
$\textrm{Ad}_{z}^{-1}$:
\begin{align}\label{for:126}
q_1^{ij}+\lambda_i\lambda_jq_2^{ij}=0
\end{align}
for $1\leq j \leq\dim(\mathfrak{K})$, $\dim(\mathfrak{K})+1\leq i\leq\dim(\mathfrak{K})+\dim \mathfrak{v}^0$.
Let the coordinate functions of $\Theta$ be $\vartheta_i$, $1\leq
i\leq t_0$. For $1\leq i\leq\dim(\mathfrak{K})$ or
$\dim(\mathfrak{K})+\dim \mathfrak{v}^0+1\leq i\leq t_0$, the
$i$-th equation in \eqref{eq:3} becomes:
\begin{align}\label{eq:4}
&\omega_i\circ z-\lambda_i\omega_i=\vartheta_i
\end{align}
and the condition \eqref{for:125} splits as
\begin{align}\label{for:127}
\sum_{j=-\infty}^{+\infty}\lambda_i^{-(j+1)}\vartheta_i\circ z^j=0.
\end{align}
Then the existence of a solution follows the same way as in Lemma
\ref{le:11}. Moreover, the estimate:
\begin{align}\label{for:131}
\norm{\omega_i}_{m-r-1}\leq C_{m,\eta}\norm{\vartheta_i}_m\leq
C_{m,\eta}\norm{\Theta}_m
\end{align}
holds for $1\leq i\leq\dim(\mathfrak{K})$ or
$\dim(\mathfrak{K})+\dim \mathfrak{v}^0+1\leq i\leq t_0$. Equality
$\omega_i=\Lambda_-^i=-\Lambda_+^i$ follows directly from Lemma
\ref{le:11}.
For $n_0\geq i\geq\dim(\mathfrak{K})+1$ the $i$-th equation in
\eqref{eq:3} is:
\begin{align}\label{for:128}
&\omega_{i}\circ
a_1-\lambda_i\omega_{i}=\vartheta_{i}+\sum_{j=1}^{k_0}q_1^{ij}\omega_{j}
\end{align}
where $n_0=\dim(\mathfrak{K})+\dim \mathfrak{v}^0$ and
$k_0=\dim(\mathfrak{K})$. By \eqref{for:126} the $i$-th coordinate
function of $J_1^{-(j+1)}\Theta\circ z^j$ is
\begin{align}
&\lambda_i^{-(j+1)}\vartheta_i\circ
z^j+\sum_{k=1}^{k_0}\sum_{n=0}^{-j-2}
q_1^{ik}\lambda^{n}_k\lambda^{-j-2-n}_i\vartheta_k\circ
z^j\qquad\text{ for
any }j\leq -2\label{for:135}\\
&\lambda_i^{-(j+1)}\vartheta_i\circ
z^j-\sum_{k=1}^{k_0}\sum_{n=-1}^{-j-1}
q_1^{ik}\lambda^{n}_k\lambda^{-j-n-2}_i\vartheta_k\circ
z^j\qquad\text{ for any }j\geq 0\label{for:136}
\end{align}
and it follows:
\begin{align*}
\Lambda_-^i&=\vartheta_i\circ z^{-1}+\sum_{j\leq
-2}\bigl(\lambda_i^{-(j+1)}\vartheta_i\circ
z^j+\sum_{k=1}^{k_0}\sum_{n=0}^{-j-2}
q_1^{ik}\lambda^{n}_k\lambda^{-j-2-n}_i\vartheta_k\circ z^j\bigl)\notag\\
\Lambda_+^i&=-\sum_{j\geq 0}\bigl(\lambda_i^{-(j+1)}\vartheta_i\circ
z^j+\sum_{k=1}^{k_0}\sum_{n=-1}^{-j-1}
q_1^{ik}\lambda^{n}_k\lambda^{-j-n-2}_i\vartheta_k\circ z^j\bigl).
\end{align*}
Using Corollary \ref{cor:1} and the fact that $e^{-\delta\abs{j}}$
decreases faster than any negative power of $\abs{j}$ for any
$\delta>0$ and $j\in\mathbb{Z}$, similar to \eqref{for:69} we can show that both
$\Lambda_-^i$ and $\Lambda_+^i$ are distributions. Now we use the
fact that all the subsequent equations are solved i.e. we substitute
all $\vartheta_{j}$ for all $1\leq j\leq k_0$ into above functions
using their expression as in \eqref{eq:4}. This implies:
\begin{align*}
\Lambda_-^i&=\sum_{j\leq
-1}\lambda_i^{-(j+1)}\bigl(\vartheta_i(z^j)+\sum_{j=1}^{k_0}q_1^{ik}\omega_{j}(z^j)\bigl)
+\lim_{j\rightarrow -\infty}\sum_{n=1}^{-j-1}\sum_{k=1}^{k_0}q_1^{ik}\lambda^n_k\lambda_i^{-n-j-1}\omega_{k}(z^{j})\notag\\
\Lambda_+^i&=-\sum_{j\geq
0}\lambda_i^{-(j+1)}\bigl(\vartheta_i(z^j)+\sum_{j=1}^{k_0}q_1^{ik}\omega_{j}(z^j)\bigl)
+\lim_{j\rightarrow
+\infty}\sum_{n=1}^{j+1}\sum_{k=1}^{k_0}q_1^{ik}\lambda_i^{-j+n-2}\lambda_k^{-n}\omega_{k}(z^{j+1}).
\end{align*}
By Corollary \ref{cor:1} and similar to \eqref{for:69} we can show
without difficulty that
\begin{align*}
\lim_{j\rightarrow
-\infty}\sum_{n=1}^{-j-1}\sum_{k=1}^{k_0}q_1^{ik}\lambda^n_k\lambda_i^{-n-j-1}\omega_{k}(z^{j})\text{
and }\lim_{j\rightarrow
+\infty}\sum_{n=1}^{j+1}\sum_{k=1}^{k_0}q_1^{ik}\lambda_i^{-j+n-2}\lambda_k^{-n}\omega_{k}(z^{j+1})
\end{align*}
are $0$ distributions. Hence it follows
\begin{align}\label{for:130}
\Lambda_-^i-\Lambda_+^i=\sum_{j=-\infty}^{+\infty}\lambda_i^{-(j+1)}\bigl(\vartheta_i(z^j)+\sum_{j=1}^{k_0}q_1^{ik}\omega_{j}(z^j)\bigl)
\end{align}
moreover by assumption \eqref{for:125} the right side is $0$
distribution.
Thus the equation \eqref{for:128} satisfies the solvability
condition \eqref{for:61} and notice that its right-hand side is a
$\mathcal{H}_0^{m-r-1}$ function
by \eqref{for:131} therefore we may use Lemma \ref{le:11} again to
conclude that the equation \eqref{for:128} has a solution
$\omega_i\in\mathcal{H}_0^{m-2r-2}$ and
\begin{align}
\omega_i=\Lambda_-^i=-\Lambda_+^i\qquad
\end{align}
for any $ n_0\geq i\geq\dim(\mathfrak{K})+1$.
As a consequence of assumption \eqref{for:131} this solution
satisfies the estimate
\begin{align}\label{for:133}
\norm{\omega_i}_{m-2r-2}\leq
C_{m,\eta}\norm{\vartheta_i+\sum_{j=1}^{k_0}q_1^{ik}\omega_{j}}_{m-r-1}\leq
C_{m,\eta}\norm{\Theta}_m
\end{align}
for $n_0\geq i\geq\dim(\mathfrak{K})+1$ if $m-r-1\geq m_1$.
Combine \eqref{for:131} and \eqref{for:133} we obtain a
$\mathcal{H}_0^{m-2r-2}$ solution $\Lambda$ of the equation
\eqref{eq:2} with estimate:
\begin{align*}
\norm{\Lambda}_{m-2r-2}\leq C_{m,\eta}\norm{\mathcal{F}}_{m}.
\end{align*}
If $z\in Z(L)$ and $\mathcal{F}\in \mathcal{H}_{0,L}^{m}$, then
$\Lambda$ takes values in $\mathfrak{L}^\bot$ and
\begin{align}\label{for:134}
\textrm{Ad}_{l^{-1}}\Lambda(lx)&=\sum_{j\leq
-1}\textrm{Ad}_{l^{-1}}\textrm{Ad}_z^{-(j+1)}\mathcal{F}(
z^jlx)=\sum_{j\leq
-1}\textrm{Ad}_z^{-(j+1)}\textrm{Ad}_{l^{-1}}\mathcal{F}( lz^jx)\notag\\
&=\sum_{j\leq -1}\textrm{Ad}_z^{-(j+1)}\mathcal{F}(
z^jx)=\Lambda(x).
\end{align}
Hence $\Lambda\in\mathcal{H}_{0,L}^{m-2r-2}$.
\end{proof}
Now we derive a similar result for ``partial'' Sobolev norms
$\norm{\cdot}_m'$ defined in Section~\ref{section:notation}
(\ref{item: incompleteSobolev}). The argument is actually simpler
since it only involves differentiability along stable and unstable
direction and no loss of regularity appears.
\begin{corollary}\label{cor:3}
For any $z\in B_\eta(a)\bigcup B_\eta(b)$ with $\eta$ sufficiently
small and for any $\mathcal{L}_0^m$ map $\mathcal{F}:M\rightarrow
\mathfrak{N}$, $m\in \mathbb{N}$ with $m>\frac{r\dim M}{2}+3r$ if
\begin{align}
\sum_{j=-\infty}^{+\infty}\Ad_z^{-(j+1)}(\mathcal{F}\circ
z^j)=0
\end{align}
as a
distribution, then the equation
\begin{align}
\Omega\circ z-\Ad_z\Omega=\mathcal{F}
\end{align}
has a solution $\Omega\in\mathcal{L}_0^{m}$ such that
\begin{align}
\norm{\Omega}'_{m}\leq C_{m,\eta} \norm{\theta}'_{m}.
\end{align}
Furthermore if $z\in Z(L)$ and $\mathcal{F}\in
\mathcal{L}_{0,L}^{m}$, then $\Omega\in\mathcal{L}_{0,L}^{m}$.
\end{corollary}
\begin{proof}
If $\frac{m}{r}-1-\frac{\dim M}{2}-1>1$ then $\mathcal{F}\in
C^1(M)$. Using the same method as in Lemma \ref{le:11} we show
the two formal solutions $\Omega_{\binom{+}{-}}$ for any coordinate
of $\mathcal{F}$ in \eqref{for:28} are distributions. The
differentiability of $\Omega$ along all stable and unstable
foliations and the estimates follows from \eqref{for:30}.
Using the same method as in Lemma \ref{le:16} we show the two
formal solutions $\Lambda_{\binom{+}{-}}$ in \eqref{for:132} are
also distributions. The estimates of $\Lambda$ follows similarly
from \eqref{for:131} and \eqref{for:133} just by substituting norm
$\norm{\cdot}$ with $\norm{\cdot}'$ and by noticing that there is
no loss of regularity by above discussion in the first part. Finally
the invariance of $\Omega$ under $L$ is the same as \eqref{for:134}.
\end{proof}
Similarly to Corollary \ref{cor:2} we also have the following
``uniqueness'' property:
\begin{corollary}\label{cor:6} If the equation
\begin{align}\label{eq:14}
\Omega\circ z-\Ad_z\Omega=\mathcal{F}
\end{align}
has a solution in $C^1(M)\bigcap \mathcal{H}_0^0$ then it is unique.
\end{corollary}
\begin{proof}
[Proof for symmetric space examples and hyperbolic twisted examples] For\\ both these cases $\textrm{Ad}_z$ has diagonal
form in $\mathfrak{N}_\mathbb{C}$ and then the equation \eqref{eq:14}
splits into finitely many equations of the form
\begin{align}\label{eq:7}
\omega\circ z-\lambda\omega=\theta
\end{align}
where $\lambda$ is an eigenvalue of $\textrm{Ad}_z$ and $\theta\in
C^1(M)\bigcap \mathcal{H}_0^0$ is the coordinate function of
$\mathcal{F}$ under new basis. By Corollary \ref{cor:2} the solution
of the above equation is unique hence the solution of \eqref{eq:16}
is unique. \end{proof}
\begin{proof}[Proof for genuinely partially hyperbolic twisted
examples] We assume notations in Lemma \ref{le:16}
if there is no confusion. If $\Omega'\in C^1(M)\bigcap
\mathcal{H}_0^0$ is also a solution of equation \eqref{eq:14}, then
\begin{align}\label{eq:16}
J_1^{-j}(\Omega-\Omega')\circ z^j=\Omega-\Omega'\qquad\text{ for any
}j\in\mathbb{Z}.
\end{align}
For $1\leq i\leq\dim(\mathfrak{K})$ or $\dim(\mathfrak{K})+\dim
\mathfrak{v}^0+1\leq i\leq t_0$, the trivialization of the $i$-th
equation in \eqref{eq:16} follows the same ways as in the previous
part.
For $n_0\geq i\geq\dim(\mathfrak{K})+1$ the left side of $i$-th
equation in \eqref{eq:16} has the same form either as
\eqref{for:135} or as \eqref{for:136} depending on $j$ by
substituting $\vartheta_i$ with $\theta_i$ where $\theta_i\in
C^1(M)\bigcap \mathcal{H}_0^0$ are coordinate functions of
$\Omega-\Omega'$ under new basis. Using Corollary \ref{cor:1} as in
\eqref{for:69} we can show without difficulty that both
\eqref{for:135} and \eqref{for:136} are $0$ distributions as
$j\rightarrow\infty$. By above argument $\theta_i=0$ for $n_0\geq
i\geq\dim(\mathfrak{K})+1$. Hence $\Omega=\Omega'$.
\end{proof}
\subsection{Higher rank trick and trivialization of cohomology}
Now we will show that in the higher rank case obstructions to the solution of the linearized conjugacy equation vanish.
The reason for that is the linearized form of the commutation relation \eqref{twistedcocyclefirst}
$L(\mathcal{F},\mathcal{G})^{(a_1,a_2)}=0$ that mean that the pair $\mathcal{F},\mathcal{G}$
form a twisted cocycle over the homogeneous action generated by $a_1$ and $a_2$. Joint solvability of the linearized conjugacy equations for commuting elements means that this cocycles is a coboundary, hence corresponding twisted first cohomology is trivial.
The proof consists of three parts:
\begin{enumerate}
\item Reduction of the vector values linearized conjugacy equation \eqref{for:101} and linearized commutativity condition \ref{twistedcocyclefirst} to scalar equations.\footnote{ Somewhat similar arguments already appeared in the proof of Lemma~\ref{le:16}.} This is straightforward for the symmetric space and
Anosov twisted symmetric space cases but for genuinely partially hyperbolic twisted symmetric space examples requires certain algebraic manipulations somewhat similar to those that appear in the case of commuting toral automorphisms in the presence of Jordan blocks, see \cite[Section 3.2]{Damjanovic4}.
\item The ``higher rank trick'' that proves vanishing of the obstructions \eqref{for:61}. It
appears in virtually identical form in all proofs of cocycle and differentiable rigidity for actions of higher rank abelian groups that use some form of dual, i.e. harmonic analysis arguments.
For its earliest appearance see Lemmas 4.3, 4.6 and 4.7 in \cite{Spatzier1}.
\item Application of Lemma~\ref{le:11} in the case of Sobolev norms and Corollary~\ref{cor:3} for partial Sobolev norms $\norm{\cdot}'_m$.
\end{enumerate}
\begin{lemma}\label{le:1}
Let $\eta$ be sufficiently small and $m\in\mathbb{N}$ with $m\geq m_1+r+1$
where $m_1$ is defined as in Lemma \ref{le:11}. For any $a_1\in
B_\eta(a)$ and $a_2\in B_\eta(b)$ where $a_1\,,a_2$ commute and any
two $\mathcal{H}_0^m$ maps $\mathcal{F}\,,\mathcal{G}:M\rightarrow
\mathfrak{N}$ satisfying $L(\mathcal{F},\mathcal{G})^{(a_1,a_2)}=0$, or, equivalently, $\mathcal{T}_{a_2}\mathcal{F}=\mathcal{T}_{a_1}\mathcal{G}$
then the equations
\begin{align}\label{eq:6}
&\Omega\circ a_1-\emph{Ad}_{a_1}\Omega=\mathcal{F}\notag\\
&\Omega\circ a_2-\emph{Ad}_{a_2}\Omega=\mathcal{G}
\end{align}
or equivalently $$\mathcal{F}=\mathcal{T}_{a_1}\Omega ,\,\, \mathcal{G}=\mathcal{T}_{a_2}\Omega$$
have a common solution $\Omega\in\mathcal{H}_0^{m-2r-2}$ with the
following estimate
\begin{align}\label{for:49}
\norm{\Omega}_{m-2r-2}\leq C_{m,\eta}
\norm{\mathcal{F},\mathcal{G}}_{m}.
\end{align}
Furthermore if $a_1,a_2\in Z(L)$ and $\mathcal{F}, \mathcal{G}\in
\mathcal{H}_{0,L}^{m}$, then $\Omega\in\mathcal{H}_{0,L}^{m-2r-2}$.
\end{lemma}
\begin{proof}[Proof for symmetric space examples]
We consider symmetric space examples at first. Let $a_i=d_ik_i$,
$i=1,2$ where $d_i$ is the split part and $k_i$ is compact part.
Consider the complexification $\mathfrak N_{\mathbb C}$ of the
subalgebra $\mathfrak N=\mathfrak{K}+\mathfrak{D}$. There
exists an orthonormal basis that diagonalizes
$\textrm{Ad}_{a_1}$ and
$\textrm{Ad}_{a_2}$. As usual, this basis may be chosen to consists of several real vectors and
several pairs of complex conjugate vectors.
The equations $L(\mathcal{F},\mathcal{G})^{(a_1,a_2)}=0$ and
\eqref{eq:6} split into finitely many equations of the form
\begin{align}\label{for:48}
L(\varphi,\vartheta)^{(a_1,a_2)}_{(\lambda_1,\lambda_2)}=(a_1,\lambda_1)^\tau
\varphi-(a_2,\lambda_2)^\tau\vartheta=0
\end{align}
and
\begin{align}\label{eq:7}
&\omega\circ a_1-\lambda_1\omega=\vartheta \notag\\
&\omega\circ a_2-\lambda_2\omega=\varphi
\end{align}
where $\vartheta,\,\varphi$ are $\mathcal{H}_0^m$ functions and
$\lambda_1$ and $\lambda_2$ in $S^1_\mathbb{C}$ are corresponding
eigenvalues of $\textrm{Ad}_{a_1}$ and $\textrm{Ad}_{a_2}$,
respectively. Notice that since the coefficients and right-hand parts of the equations \eqref{eq:6}
are real-valued, the unique solution in $\mathcal{H}_0^0$ (see Corollary~\ref{cor:2}) is real-valued as well.
By the assumption \eqref{for:48} we get
\begin{align*}
\sum_{j=-n}^{j=n}\lambda_1^{-(j+1)}\vartheta(a_2a_1^j)-\sum_{j=-n}^{j=n}\lambda_2\lambda_1^{-(j+1)}\vartheta(a_1^j)
=\lambda_1^{-(n+1)}\varphi(a_1^{n+1})-\lambda_1^{n}\varphi(a_1^{-n}).
\end{align*}
By Corollary \ref{cor:1} the right-hand converges to $0$ as a distribution when
$n\rightarrow \infty$. Hence using notation from Section~\ref{section:notation} (\ref{item(10)})
\begin{align}\label{for:39}
\sum^{(a_1,\lambda_1)}\lambda_2^{-1}\vartheta(a_2)=\sum^{(a_1,\lambda_1)}\vartheta
\end{align}
as distributions.
Let $\phi\in \Phi$. For any $j,n\in\mathbb{Z}$ write
\begin{align*}
\sum_{\phi\in\Phi} \abs{\phi(j\log d_1+n\log
d_2)}=(\abs{j}+\abs{n})\sum_{\phi\in\Phi}\abs{\phi(j_1\log
d_1+j_2\log d_2)}
\end{align*}
where $j_1=\frac{j}{\abs{j}+\abs{n}}$ and
$j_2=\frac{n}{\abs{j}+\abs{n}}$. If $\eta$ is small enough then
$a_1$ and $a_2$ are also linearly independent elements. Hence
$$c_0=\min_{\substack{\abs{r_1}+\abs{r_2}=1\\(r_1,r_2)\in\mathbb{R}^2}}\frac{1}{2}(\sum_{\phi\in\Phi}\abs{\phi(r_1\log d_1+r_2\log d_2)})>0.$$
For any $f\in C^\infty(G/\Gamma)$ by Corollary \ref{cor:1} there
exist constants $\gamma,E>0$ only dependent on $G$ such that
\begin{align*}
&\sum_{n=-\infty}^{\infty}\sum_{j=-\infty}^{\infty}\abs{\langle
\lambda_1^{-(j+1)}\lambda_2^{-(n+1)}\vartheta(a_2^na_1^j),f\rangle}\\
&\leq \sum_{n=-\infty}^{\infty}\sum_{j=-\infty}^{\infty}\abs{\langle
\vartheta(a_2^na_1^j),f\rangle}\\
&\leq
\sum_{n=-\infty}^{\infty}\sum_{j=-\infty}^{\infty}E\norm{\vartheta}_{C^1}\norm{f}_{C^1}e^{-\gamma
c_0 (\abs{n}+\abs{j})}<\infty.
\end{align*}
Hence the sum
$\sum_{n=-\infty}^{\infty}\sum_{j=-\infty}^{\infty}\abs{\langle
\lambda_1^{-(j+1)}\lambda_2^{-(n+1)}\vartheta(a_2^na_1^j),f\rangle}$
converges absolutely and thus
$\sum^{(a_2,\lambda_2)}\sum^{(a_1,\lambda_1)}\vartheta$ is a
distribution. On the other hand by iterating equation \eqref{for:39}
we obtain:
\begin{align*}
\sum^{(a_1,\lambda_1)}\lambda_2^{-k}\vartheta(a_2^k)=\sum^{(a_1,\lambda_1)}\vartheta
\end{align*}
as distributions for any $j\in\mathbb{Z}$. Therefore
\begin{align}\label{for:40}
\sum^{(a_2,\lambda_2)}\sum^{(a_1,\lambda_1)}\vartheta=\sum_k\sum^{(a_1,\lambda_1)}\lambda_2^{-(k+1)}\vartheta(a_2^k)=
\sum_k\lambda_2^{-1}\sum^{(a_1,\lambda_1)}\vartheta
\end{align}
The series in the left hand side of \eqref{for:40} is not a
distribution unless $\sum^{(a_1,\lambda_1)}\vartheta$ is a $0$
distribution. Similarly $\sum^{(a_2,\lambda_2)}\varphi$ is also a $0$
distribution.
This is the ``higher rank trick''!
By Lemma \ref{le:11} each equation of \eqref{eq:7} has a
$\mathcal{H}_0^{m-r-1}$ solution. Moreover, they coincide. Indeed,
if $\omega$ solves the first equation, i.e.
$(a_1,\lambda_1)^\tau\omega=\vartheta$ then by \eqref{for:48} we
have
\begin{align*}
(a_2,\lambda_2)^\tau(a_1,\lambda_1)^\tau\omega=(a_2,\lambda_2)^\tau\vartheta=(a_1,\lambda_1)^\tau\varphi
\end{align*}
Since operators $(a_2,\lambda_2)^\tau$ and $(a_2,\lambda_2)^\tau$
commute this implies
\begin{align*}
(a_1,\lambda_1)^\tau((a_2,\lambda_2)^\tau\omega-\varphi)=0
\end{align*}
By Corollary \ref{cor:2} $(a_1,\lambda_1)^\tau$ is an injective
operator if $\omega,\varphi\in C^1(M)$, that is $m-r-1>\frac{\dim
M}{2}+2$. Therefore $(a_2,\lambda_2)^\tau\omega-\varphi=0$ i.e.
$\omega$ solves the second equation as well.
\end{proof}
\begin{proof}[Proof for twisted symmetric space examples]
We assume
notations from the previous part if there is no confusion.
(1) If $\rho$ is Anosov then the neutral distribution is still
$\mathfrak{D}+\mathfrak{K}$. Let $a_i=d_ik_i$, $i=1,2$ where $d_i$
is the split part and $k_i$ is compact part. Notice for any
$(g,t)\in G\times \mathbb{R}^N$, $a_i\cdot(g,t)=(d_ik_ig,t)$, $i=1,2$ then
the statement follows essentially verbatim as in the case of symmetric space
examples.
(2) If $\rho$ is genuinely partially hyperbolic, then the neutral
distribution is $\mathfrak{N}=\mathfrak{D}+\mathfrak{K}+\mathfrak{v}^0$. Let
$a_i=(d_ik_i,t_i)$, $i=1,2$ where $d_i$ is the split part, $k_i$ is
compact part and $t_i\in \mathfrak{v}^0$. We repeat with appropriate modifications construction from the proof of Lemma~\ref{le:16}. Consider as before the complexification $\mathfrak{N}_\mathbb{C}$
of $\mathfrak{N}$. By \eqref{for:51} and \eqref{for:42}
there
exists an orthonormal basis in $\mathfrak{N}_\mathbb{C}$ such that $\textrm{Ad}_{a_i}$,
$i=1,2$ has the following form in this basis
$$ J_i=\begin{pmatrix}A_i & 0 & 0\\
B_i & D_i & 0\\
0 & 0 & E_i
\end{pmatrix}$$
where $E_i=I_{\dim(\mathfrak{D})}$, $A_i$ are
$\dim\mathfrak{K}\times \dim\mathfrak{K}$
diagonal matrices and
$D_i$ are $ \dim\mathfrak{v}^0\times \dim\mathfrak{v}^0$ diagonal
matrices all of whose eigenvalues are of absolute value $1$ and
every element of $B_i$ with absolute value smaller than $1$ if
$\eta$ is small enough. This basis can be chosen as in the symmetric
space case to include real vectors and pairs of complex conjugate
vectors.
Then
equations \eqref{eq:6} have the form:
\begin{align}\label{eq:8}
&\Omega\circ a_1-J_1\Omega=\Theta\notag\\
&\Omega\circ a_2-J_2\Omega=\Psi
\end{align}
and the condition $L(\Psi,\Theta)^{(a_1,a_2)}=0$ can be written as
\begin{align}\label{for:52}
J_2\Theta-\Theta\circ a_2=J_1\Psi-\Psi\circ a_1.
\end{align}
Denote $J_1=(q_1^{ij})$ and $J_2=(q_2^{ij})$ and let
$q_1^{ii}=\lambda_i$ and $q_2^{ii}=\mu_i$ for $1\leq i,j\leq t_0$
where $t_0=\dim \mathfrak{D}+\dim \mathfrak{v}^0+\dim \mathfrak{K}$.
Since $\textrm{Ad}_{a_1}$ and $\textrm{Ad}_{a_2}$
commute, by comparing coefficients, one obtains the
following relation between the coefficients of $\textrm{Ad}_{a_1}$ and
$\textrm{Ad}_{a_2}$:
\begin{align}\label{for:58}
\lambda_iq_2^{ij}+\mu_jq_1^{ij}=\mu_iq_1^{ij}+\lambda_jq_2^{ij}
\end{align}
for $1\leq j \leq\dim(\mathfrak{K})$, $\dim(\mathfrak{K})+1\leq i\leq\dim(\mathfrak{K})+\dim \mathfrak{v}^0$.
Let the coordinate functions of $\Theta$ and $\Psi$ be $\vartheta_i$
and $\varphi_i$ $1\leq i\leq t_0$, respectively. For $1\leq
i\leq\dim(\mathfrak{K})$ or $\dim(\mathfrak{K})+\dim
\mathfrak{v}^0+1\leq i\leq t_0$, the $i$-th pair of equations in
\eqref{eq:8} is:
\begin{align}\label{eq:15}
&\omega_i\circ a_1-\lambda_i\omega_i=\vartheta_i\notag\\
&\omega_i\circ a_2-\mu_i\omega_i=\varphi_i
\end{align}
and the condition $L(\Psi,\Theta)^{(a_1,a_2)}=0$ splits as
\begin{align}\label{for:53}
L(\varphi_i,\vartheta_i)^{(a_1,a_2)}_{(\lambda_i,\mu_i)}=(a_1,\lambda_i)^\tau
\varphi_i-(a_2,\mu_i)^\tau\vartheta_i=0
\end{align}
If we can show that $\sum^{(a_2,\mu_i)}\sum^{(a_1,\lambda_i)}\vartheta_i$
is a distribution then $\sum^{(a_1,\lambda_i)}\vartheta=0$ and
hence
the existence of a common solution follows the same way as in the previous part.
As before for integers $j,n$ we define
$j_1=\frac{j}{\abs{j}+\abs{n}}$ and
$j_2=\frac{n}{\abs{j}+\abs{n}}$. If $\eta$ is small enough then
$a_1$ and $a_2$ are also linearly independent elements hence
$$c_0=\min_{\substack{\abs{r_1}+\abs{r_2}=1\\(r_1,r_2)\in\mathbb{R}^2}}\frac{1}{2}(\sum_{\phi\in\Phi}\abs{\phi(r_1\log d_1+r_2\log d_2)})>0.$$
For any $f\in C^\infty(M)$, let $c_i^j=d_i^jk_i^j$,
$t_{n,j}=\sum_{l=0}^{n-1}\rho(k_1)^{-j}\rho(k_2)^{-l}t_2+\sum_{l=0}^{j-1}\rho(k_1)^{-l}t_1$,
$i=1,2$ for any $n\,,j\in\mathbb{Z}$, by Corollary \ref{cor:1} there exist
constants $\gamma,E>0$ only dependent on $G$ satisfying
\begin{align*}
&\sum_{n=-\infty}^{\infty}\sum_{j=-\infty}^{\infty}\abs{\langle
\mu_i^{-(n+1)}\lambda_i^{-(j+1)}\vartheta_i(a_2^na_1^k),f\rangle}\\
&\leq \sum_{n=-\infty}^{\infty}\sum_{j=-\infty}^{\infty}\abs{\langle
\vartheta_i(a_2^na_1^j),f\rangle}\\
&\leq\sum_{n=-\infty}^{\infty}\sum_{j=-\infty}^{\infty}\abs{\langle
\vartheta_i\bigl((c_2^nc_1^j,0)x\bigl),f((0,-t_{n,j})x)\rangle}\\
&\leq
\sum_{n=-\infty}^{\infty}\sum_{j=-\infty}^{\infty}E\norm{\vartheta}_{C^1}\norm{f}_{C^1}(\norm{t_1}+\norm{t_2})(\abs{j}+\abs{n})e^{-\gamma
c_0 (\abs{n}+\abs{j})}<\infty
\end{align*}
then the sum
$\sum_{n=-\infty}^{\infty}\sum_{j=-\infty}^{\infty}\abs{\langle
\mu_i^{-(n+1)}\lambda_i^{-(j+1)}\vartheta(a_2^na_1^k),f\rangle}$
converges absolutely. Hence
$\sum^{(a_2,\mu_i)}\sum^{(a_1,\lambda_i)}\vartheta_i$ is a
distribution.
Therefore, using \eqref{for:53} the same way as in the previous part, we deduce that there
exist $\omega_i\in\mathcal{H}_0^{m-r-1}$ which solve
simultaneously the equations in \eqref{eq:15}. Moreover, for $1\leq
i\leq\dim(\mathfrak{K})$ or $\dim(\mathfrak{K})+\dim
\mathfrak{v}^0+1\leq i\leq t_0$, the estimate:
\begin{align}\label{for:43}
\norm{\omega_i}_{m-r-1}\leq
C_{m,\eta}\norm{\vartheta_i,\varphi_i}_m\leq
C_{m,\eta}\norm{\Theta,\Phi}_m
\end{align}
follows from Lemma \ref{le:11}.
For $\dim(\mathfrak{K})+\dim \mathfrak{v}^0\geq
i\geq\dim(\mathfrak{K})+1$ the $i$-th pair of equations in
\eqref{eq:8} is:
\begin{align}\label{for:56}
&\omega_{i}\circ a_1-\lambda_i\omega_{i}-\sum_{j=1}^{k_0}q_1^{ij}\omega_{j}=\vartheta_{i}\notag\\
&\omega_{i}\circ
a_2-\mu_i\omega_{i}-\sum_{j=1}^{k_0}q_2^{ij}\omega_{j}=\varphi_{i}
\end{align}
where $n_0=\dim(\mathfrak{K})+\dim \mathfrak{v}^0$ and
$k_0=\dim(\mathfrak{K})$ and the assumption \eqref{for:52} for
$\vartheta_{i}$ and $\varphi_{i}$ splits as:
\begin{align}\label{for:54}
\mu_i\vartheta_{i}+\sum_{j=1}^{k_0}q_2^{ij}\vartheta_{j}-\vartheta_{i}\circ
a_2=\lambda_i\varphi_{i}+\sum_{j=1}^{k_0}q_1^{ij}\varphi_{j}-\varphi_{i}\circ
a_1
\end{align}
Now we use the fact that all the subsequent pairs of equations are
solved i.e. we substitute all $\vartheta_{j}$ and $\varphi_j$ for all
$1\leq j\leq k_0$ into \eqref{for:54} using their expression as in
\eqref{eq:15}. This implies:
\begin{align*}
\sum_{j=1}^{k_0}q_2^{ij}(a_1,\lambda_j)^\tau\omega_{j}-(a_2,\mu_i)^\tau\vartheta_{i}
=\sum_{j=1}^{k_0}q_1^{ij}(a_2,\mu_j)^\tau\omega_{j}-(a_1,\lambda_i)^\tau\varphi_{i}
\end{align*}
Since $\textrm{Ad}_{a_1}$ and $\textrm{Ad}_{a_2}$ commute, we can
use the equation \eqref{for:58} for the coefficients and the
linearity of operators $(a_2,\mu_j)^\tau$ and
$(a_1,\lambda_j)^\tau$, to simplify the above expression to:
\begin{align*}
(a_2,\mu_i)^\tau(\vartheta_{i}+\sum_{j=1}^{k_0}q_1^{ij}\omega_{j})=(a_1,\lambda_i)^\tau(\varphi_{i}+\sum_{j=1}^{k_0}q_2^{ij}\omega_{j})
\end{align*}
Thus the functions
$\vartheta_{i}+\sum_{j=1}^{k_0}q_1^{ij}\omega_{j}$ and
$\varphi_{i}+\sum_{j=1}^{k_0}q_2^{ij}\omega_{j}$ satisfy the
solvability condition
\begin{align}\label{for:138}
L(\vartheta_{i}+\sum_{j=1}^{k_0}q_1^{ij}\omega_{j},\varphi_{i}+\sum_{j=1}^{k_0}q_2^{ij}\omega_{j})^{(a_2,a_1)}_{(\mu_i,\lambda_i)}=0
\end{align}
they are in $\mathcal{H}_0^{m-r-1}$ by \eqref{for:43} therefore we
may use previous part again to conclude that the pair of equations
\eqref{for:56} has a common solution
$\omega_i\in\mathcal{H}_0^{m-2r-2}$. As a consequence of assumptions
\eqref{for:43} this solution satisfies the estimate
\begin{align}
\norm{\omega_i}_{m-2r-2}\leq
C_{m,\eta}\norm{\vartheta_i,\varphi_i}_{m}\leq
C_{m,\eta}\norm{\Theta,\Phi}_m
\end{align}
for $n_0\geq i\geq\dim(\mathfrak{K})+1$ if $m-r-1\geq m_1$.
We obtain the following estimate for the norm of the
$\mathcal{H}_0^{m-2r-2}$ solution $\Omega$ of the system
\eqref{eq:6}:
\begin{align}
\norm{\Omega}_{m-2r-2}\leq
C_{m,\eta}\norm{\mathcal{F},\mathcal{G}}_{m}.
\end{align}
Both for the symmetric space examples and hyperbolic twisted
symmetric space examples it is obvious that
\begin{align}\label{for:139}
\Omega=\Lambda_{\binom{+}{-}}=\binom{-}{+}\sum_{\binom{j\geq0}{j\leq
-1}}J_{1}^{-(j+1)}\Theta\circ a_1^j.
\end{align}
For the genuinely partially hyperbolic twisted symmetric space
examples, similar to the former cases for the diagonal blocks $A$
and $E$ of $\text{Ad}_{a_1}$ and $\text{Ad}_{a_2}$, that is for
$1\leq i\leq\dim(\mathfrak{K})$ or $\dim(\mathfrak{K})+\dim
\mathfrak{v}^0+1\leq i\leq t_0$, the $i$-th coordinate $\Omega_i$ of
$\Omega$ is the $i$-th coordinate of $\Lambda_{\binom{+}{-}}$.
For $n_0\geq i\geq\dim(\mathfrak{K})+1$, by \eqref{for:138}
\begin{align*}
\sum^{(a_1,\lambda_i)}(\vartheta_{i}+\sum_{j=1}^{k_0}q_1^{ij}\omega_{j})=0\qquad\text{
as a distribution}
\end{align*}
combined with \eqref{for:130} the $i$-th coordinate of
$\Lambda_--\Lambda_+$ is also $0$. Then we get
$\Lambda_-=\Lambda_+$. By Lemma \ref{le:16} and Corollary
\ref{cor:6} $\Omega=\Lambda_-=\Lambda_+$. Invariance of $\Omega$
under $\text{Ad}_L$ follows the same way as in \eqref{for:134}.
\end{proof}
Now we consider the case of partial Sobolev norms $\norm{\cdot}'_m$.
\begin{corollary}\label{cor:4}
Let $\eta$ be sufficiently small and $m\in \mathbb{N}$ with $m>\frac{r\dim
M}{2}+3r$. For any $a_1\in B_\eta(a)$ and $a_2\in B_\eta(b)$ where
$a_1\,,a_2$ commute and for any two
$\mathcal{F}\,,\mathcal{G}:M\rightarrow \mathfrak{N}$ satisfying
$L(\mathcal{F},\mathcal{G})^{(a_2,a_1)}=0$, the equations
\begin{align}\label{eq:1}
&\Omega\circ a_1-\emph{Ad}_{a_1}\Omega=\mathcal{F}\notag\\
&\Omega\circ a_2-\emph{Ad}_{a_2}\Omega=\mathcal{G}
\end{align}
or, using our compact notation, $ \mathcal{T}_{a_1}\Omega=\mathcal{F},\,\,\, \mathcal{T}_{a_2}\Omega=\mathcal{G}$
have a common solution $\omega\in\mathcal{L}_0^{m}$ with the
following estimate
\begin{align}\label{for:65}
\norm{\Omega}'_{m}\leq C_{m,\eta}
\norm{\mathcal{F},\mathcal{G}}'_{m}.
\end{align}
Furthermore if $a_1\,,a_2\in Z(L)$ and $\mathcal{F}\,,\mathcal{G}\in
\mathcal{L}_{0,L}^{m}$, then $\Omega\in\mathcal{L}_{0,L}^{m}$.
\end{corollary}
\begin{proof}
If
$\frac{m}{r}-1-\frac{\dim M}{2}-1>1$ then
$\mathcal{F}\,,\mathcal{G}\in C^1(M)$. Using the same method as in
Lemma \ref{le:1} we show that both
$\sum_{j=-\infty}^{+\infty}\textrm{Ad}_{a_1}^{-(j+1)}(\mathcal{F}\circ
a_1^j)$ and
$\sum_{j=-\infty}^{+\infty}\textrm{Ad}_{a_2}^{-(j+1)}(\mathcal{G}\circ
a_2^j)$ are $0$ distributions. Then by Corollary \ref{cor:3} each
equation of \eqref{eq:1} has a $\mathcal{L}_{0}^{m}$ solution.
Moreover, they coincide. If $\Omega$ solves the first equation, i.e.
$\mathcal{T}_{a_1}\Omega=\mathcal{F}$, then by assumption
$L(\mathcal{F},\mathcal{G})^{(a_1,a_2)}=0$ we have
$\mathcal{T}_{a_2}\mathcal{F}=\mathcal{T}_{a_1}\mathcal{G}$ and thus
\begin{align*}
\mathcal{T}_{a_2} \mathcal{T}_{a_1}\Omega=\mathcal{T}_{a_2}\mathcal{F}=\mathcal{T}_{a_1}\mathcal{G}.
\end{align*}
Since operators $\mathcal{T}_{a_1}$ and $\mathcal{T}_{a_2}$ commute this implies
\begin{align*}
\mathcal{T}_{a_1}(\mathcal{T}_{a_2}\Omega-\mathcal{G})=0
\end{align*}
By Corollary \ref{cor:6} $\mathcal{T}_{a_1}$ is an injective operator if
$\Omega,\mathcal{G}\in C^1(M)$ which is satisfied by our assumption.
Therefore $\mathcal{T}_{a_2}\Omega-\mathcal{G}=0$ i.e. $\Omega$ solves
$\Omega\circ a_2-\textrm{Ad}_{a_2}\Omega=\mathcal{G}$ as well. Then
estimate \eqref{for:65} follows by Corollary \ref{cor:3}.
\end{proof}
\section{Approximate solution of linearized equation}\label{section:splitting}
\subsection{The splitting problem}\label{splitting-intro}
As we mentioned in Section~\ref{sec:2.2}, conjugacy problem cannot be reduced to a cohomology problem for the unperturbed action. In other words, perturbations do not satisfy cocycle equations exaclty. However they satisfy those equations approximately. The method of proof of our main theorems is based on the iteration procedure. At each step we have an almost cocycle and show that it is an almost coboundary. To achieve that we need to show that an almost cocycle $\mathcal F, \mathcal G$ can be split into a real cocycle and an error term that can be estimated tamely through the values of the coboundary operator $L(\mathcal F, \mathcal G)$.
>From the general functional analysis point of view the problem does not look very hopeful. We have a bounded operator with infinite-dimensional co-kernel and without a spectral gap.
What may help of course is that we are content with a finite loss of regularity but still in general tame splitting is not likely. So one needs to use special features of the operators to construct desired splittings.
The first thing that comes to mind is to use the fact that our action is a part of the actions of the whole group $G$ (we ignore additional factorization for the sake of this discussion). One can split
the unitary representation of $G$ in $L^2(M)$ into irreducibles that are orthogonal not only with respect to the $L^2$ norm itself but also with respect to Sobolev norms, and try to construct splitting in each irreducible representation space. A similar approach works for the actions of $\mathbb{Z}^k$ by automorphisms of the torus \cite{Damjanovic4}. In the semisimple case it has been successfully applied to the unipotent action on homogeneous spaces of $SL(2,\mathbb{R})\times SL(2,\mathbb{R})$ in \cite{DK-parabolic}. There are some other cases where this approach may (and should) work, such as actions by automorphisms of nilmanifolds, partially hyperbolic actions on factors of $SL(2,\mathbb{R})\times SL(2,\mathbb{R})$, or unipotent actions on homogeneous spaces of $SL(2,\mathbb{C})$. In all these cases one uses specific explicit constructions of the splittings that do not depend on regularity of functions involved that can be loosely described as ``pushing the obstructions to the lowest possible level''. However, even for simple Lie groups of rank $\ge 2$ the structure of irreducible representations is too complicated to carry out similar specific constructions.
We solve the problem in a fairly general way by using the algebraic structure of the coboundary operators that allows to reduce the problem of splitting to orthogonal projections in sufficiently high Sobolev spaces.\footnote{This can also be used instead of specific constructions in the situations considered in \cite{Damjanovic4, DK-parabolic}} While algebra is transparent, the analysis part is involved and subtle. In order to carry out our method for a parametric family of operators one needs to work with operators in a fixed Hilbert space; in other words
not to allow any loss of regularity. This is the reason we consider Sobolev spaces
$\mathcal{L}_{0,L}^m$ where derivatives are only considered in the hyperbolic directions
and no loss of regularity appears in the solution of the coboundary equations, see Corollaries~\ref{cor:3}
and \ref{cor:4}. Splittings for those spaces are constructed in the next section.
However spaces $\mathcal{L}_{0,L}^m$ play an auxiliary role: they cannot be used in the iteration scheme. We use this splitting in Section~\ref{splitting-final} to get tame estimates uniform in parameters for the splitting in the ``real'' Sobolev spaces $\mathcal{H}_{0,L}^m$ (see \eqref{crucial} and \eqref{for:89}
in the proof of Lemma~\ref{le:2}). This is the only place in the main line of arguments where
auxiliary norms $\norm{\cdot}'_m$ appears but it seems to be crucial.
\subsection{Construction of splittings in $\mathcal{L}_{0,L}^m$}\label{splitting-prelim}
We begin with a preparatory lemma about properties of coboundaries.
The main part here is the last statement that asserts that on the
intersection of images of two commuting coboundary operators the
product of those operators can be inverted in the space
$\mathcal{L}_{0,L}^m$.
\begin{lemma}\label{le:12}
If $m\in\mathbb{N}$ with $m>\frac{r\dim M}{2}+3r$ and if $\eta$ is small
enough the following properties hold for any $z\in B_\eta(a)\bigcup
B_\eta(b)$ with $z\in Z(L)$:
\begin{enumerate}
\item if $f=\mathcal{T}_z \Omega$ and $\Omega\in \mathcal{L}_{0,L}^m$ then
$\sum^{z}f=0$ as a distribution. \label{for:2}
\item if $\mathcal{T}_z\Omega=0$ with $\Omega\in\mathcal{L}_{0,L}^{m}$
then $\Omega=0$.\label{for:9}
\item if $\mathcal{T}_z \Omega=f\in \mathcal{L}_{0,L}^{m}$
and $\Omega\in C^1\bigcap \mathcal{L}_{0,L}^{0}$ then
$\Omega\in\mathcal{L}_{0,L}^{m}$ and\\ $\norm{\Omega}'_{m}\leq
C_{m,\eta} \norm{f}'_{m}$.\label{for:10}
\item if $z_1\in B_{\eta}(a)$ and $z_2\in
B_{\eta}(b)$ and $z_1\,,z_2$ commute with $z_1,z_2\in Z(L)$ and
$f\in \mathcal{L}_{0,L}^{m}$ satisfying $\sum^{z_1}f=0$ and
$\sum^{z_2}f=0$ as distributions then there exists $\Omega\in
\mathcal{L}_{0,L}^{m}$ satisfying $\mathcal{T}_{z_1}\mathcal{T}_{z_2}\Omega=f$ and\\
$\norm{\Omega}'_{m}\leq C_{m,\eta} \norm{f}'_{m}$.
\label{for:103}
\end{enumerate}
\end{lemma}
\begin{proof}We assume notations of Lemma \ref{le:16} if there is no
confusion.
(1) By assumption we get
\begin{align}\label{for:129}
\sum_{j=-n}^{k}\Ad_{z}^{-(j+1)}f(z^j)=\Ad_{a_1}^{-(k+1)}\Omega(z^{k+1})-\Ad_{a_1}^n\Omega(z^{-n}).
\end{align}
First consider symmetric space examples and hyperbolic twisted
symmetric space examples. For both these cases $\textrm{Ad}_z$ has
diagonal form in $\mathfrak{N}_\mathbb{C}$ and then any coordinate function
of \eqref{for:129} has the form
\begin{align*}
\lambda^{-(k+1)}\vartheta(z^{k+1})-\lambda^n\vartheta(z^{-n})
\end{align*}
where $\lambda$ is a eigenvalue of $\textrm{Ad}_z$ and $\theta\in
\mathcal{L}_{0,L}^{0}$.
By \eqref{for:38} and \eqref{for:69}
$\lambda^{-(k+1)}\vartheta(z^{k+1})$ and
$\lambda^n\vartheta(z^{-n})$ converge to $0$ as distributions as
$k\rightarrow\infty$ and $n\rightarrow\infty$ respectively, then
$\sum^{(z,\lambda)}f$ is a $0$ distribution.
Now consider genuinely partially hyperbolic twisted symmetric space
examples.
For $1\leq i\leq\dim(\mathfrak{K})$ or $\dim(\mathfrak{K})+\dim
\mathfrak{v}^0+1\leq i\leq t_0$, the $i$-th coordinate in
\eqref{for:129} has the form:
\begin{align}\label{for:137}
\lambda_i^{-(k+1)}\vartheta_i(z^{k+1})-\lambda_i^n\vartheta_i(z^{-n})
\end{align}
where $\theta_i\in \mathcal{L}_{0,L}^{m}$. Arguing as above we see that expression
\eqref{for:137} converges to a $0$ distribution as as $k\rightarrow\infty$ and $n\rightarrow\infty$.
For $n_0\geq i\geq\dim(\mathfrak{K})+1$ by \eqref{for:135} and
\eqref{for:136} the $i$-th coordinate of
$\emph{Ad}_{a_1}^{-(k+1)}\Omega(z^{k+1})$ and
$\emph{Ad}_{a_1}^n\Omega(z^{-n})$ in \eqref{for:129}
are equal to
\begin{align*}
&-\lambda_i^{-(k+1)}\vartheta_i(z^{k+1})+\sum_{j=1}^{k_0}\sum_{n=-1}^{-k-2}
q_1^{ij}\lambda^{n}_j\lambda^{-k-n-2}_i\vartheta_j(z^{k+1})\\
\text{and}\\
&+\lambda_i^{n}\vartheta_i(z^{-n})+\sum_{k=1}^{k_0}\sum_{j=0}^{n-2}
q_1^{ik}\lambda^{j}_k\lambda^{n-1-j}_i\vartheta_k(z^{-n})
\end{align*}
where $\vartheta_i\in \mathcal{L}_{0,L}^{m}$.
Using Corollary \ref{cor:1} as in \eqref{for:69} we see that those expressions converge to $0$ distributions as
$k\rightarrow\infty$ and $n\rightarrow\infty$ respectively.
Combining both these cases we deduce that $\sum^{(z,\lambda)}f$ is a $0$
distribution.
(2) A direct consequence of Corollary \ref{cor:6} since if
$\Omega\in \mathcal{L}_{0,L}^{m}$ then $\Omega\in C^1$.
(3) By \eqref{for:2}, we get $\sum^{z}f=0$ as a distribution. Then
by Corollary \ref{cor:3} there exist $\Omega_1\in
\mathcal{L}_{0,L}^{m}$ satisfying: $\mathcal{T}_z\Omega_1=f$ such that
$\norm{\Omega_1}'_m\leq C_{m,\eta}\norm{f}'_m$. By (2) $\Omega$ and
$\Omega_1$ coincide.
(4) By assumption $\sum^{z_1}f=0$ and $\sum^{z_2}f=0$ it follows
from Corollary \ref{cor:3} that there exist $\Omega_1\,,\Omega_2\in
\mathcal{L}_{0,L}^{m}$ satisfying: $\mathcal{T}_{z_1}\Omega_1=f$ and
$\mathcal{T}_{z_2}\Omega_2=f$ with estimate $\norm{\Omega_1,\Omega_2}'_m\leq
C_{m,\eta}\norm{f}'_m$. Since $\mathcal{T}_{z_1}\Omega_1=\mathcal{T}_{z_2}\Omega_2$,
then by Corollary \ref{cor:4} there exists $\Omega\in
\mathcal{L}_{0,L}^{m}$ satisfying: $\mathcal{T}_{z_2}\Omega=\Omega_1$ and
$\mathcal{T}_{z_1}\Omega=\Omega_2$ with estimate $\norm{\Omega}'_m\leq
C_{m,\eta}\norm{\Omega_1,\Omega_2}'_m\leq C^2_{m,\eta}\norm{f}'_m$
and $\mathcal{T}_{z_1}\mathcal{T}_{z_2}\Omega=\mathcal{T}_{z_1}\Omega_1=f$.
\end{proof}
For any $z\in B_\eta(a)\bigcup B_\eta(b)$ with $z\in Z(L)$ let
$$U_{z}^{m}=\{f\in\mathcal{L}_{0,L}^m|\sum^{z}f\text{ is a $0$ distribution}\} $$
By Lemma \ref{le:12} if $m>\frac{r\dim M}{2}+3r$, for a` sufficiently
small $\eta$ $U_{z}^m$ is a closed subspace of
$\mathcal{L}_{0,L}^{m}$.
For any pair $z_1,z_2$ where $z_1\in B_{\eta}(a)$ and $z_2\in
B_{\eta}(b)$
let the orthogonal complement of
$U^{(3,m)}_{z_1,z_2}=U_{z_1}^{m}\bigcap U_{z_2}^{m}$ in
$U_{z_1}^{m}$ be $U^{(1,m)}_{z_1,z_2}$ and in $U_{z_2}^{m}$ be
$U^{(2,m)}_{z_1,z_2}$. Then we have a decomposition
\begin{align}\label{cobound-orthogonal}
U_{z_1}^{m}+U_{z_2}^{m}=U^{(1,m)}_{z_1,z_2}\bigoplus
U^{(2,m)}_{z_1,z_2}\bigoplus U^{(3,m)}_{z_1,z_2}.
\end{align}
For any $f\in
U_{z_1}^{m}+U_{z_2}^{m}$, write $f=\sum_{i=1}^3 f_i$ where $f_i\in
U^{(i,m)}_{z_1,z_2}$. By the Open Mapping Theorem there exists $C>0$
such that
\begin{align}\label{for:47}
\norm{f}_m'\leq\sum_{i=1}^3\norm{f_i}'_m\leq
C_{m,z_1,z_2}\norm{f}_m'
\end{align}
where $C_{m,z_1,z_2}$ is dependent on $m,z_1,z_2$.
Now we are ready to construct the advertised splitting. It can be
described as follows: given a pair of functions that represent an
``almost cocycle'', i.e $\psi,\theta\in \mathcal{L}_{0,L}^m$,
$a_1,a_2\in Z(L)$ and $L(\psi,\theta)^{(a_1,a_2)}=\omega$, project
the coboundary $\mathcal{T}_{a_2}\theta$ orthogonally to the space
$U_{z_1}^{m}\bigcap U_{z_2}^{m}$ of joint coboundaries. By
Lemma~\ref{le:12} (4) this projection is in the image of the
product of the coboundary operators that can be inverted in the
space $\mathcal{L}_{0,L}^m$ producing a function $\Omega\in
\mathcal{L}_{0,L}^m$. Coboundaries $\mathcal{T}_{a_1}\Omega$ and
$\mathcal{T}_{a_2}\Omega$ form a cocycle that approximates our pair
$\psi,\theta$ with an error of the order of $\omega$.
\begin{lemma}\label{le:6}
Let $\eta$ be sufficiently small and $m\in\mathbb{N}$ with $m>\frac{r\dim
M}{2}+3r$, $a_1\in B_\eta(a)$ and $a_2\in B_\eta(b)$ where
$a_1\,,a_2$ commute and $a_1\,,a_2\in Z(L)$. Suppose that
$\psi,\theta\in \mathcal{L}_{0,L}^m$ and
$L(\psi,\theta)^{(a_1,a_2)}=\omega$.
Then there exists $\Omega\in \mathcal{L}_{0,L}^m$ such that
\begin{align}\label{L7-2-1}
&\norm{\theta-\mathcal{T}_{a_1}\Omega}'_{m}\leq
C_{m,a_1,a_2}\norm{\omega}'_{m}\\\label{L7-2-2}
&\norm{\psi-\mathcal{T}_{a_2}\Omega}'_{m}\leq
C_{m,a_1,a_2}\norm{\omega}'_{m}\\\label{L7-2-3}
&\norm{\Omega}'_{m}\leq C_{m,\eta}\norm{\theta,\psi}'_{m}.
\end{align}
\end{lemma}
\begin{proof}
Let $p_i$ be the projection from $U_{a_1}^{m}+U_{a_2}^{m}$ to
$U^{(i,m)}_{(a_1,a_2)}$, $1\leq i\leq 3$. By \eqref{for:2} of Lemma
\ref{le:12} $\mathcal{T}_{a_2}\theta\in U_{a_2}^{m}$ and $\mathcal{T}_{a_1}\psi\in
U_{a_1}^{m}$ then we write using the decomposition
\eqref{cobound-orthogonal}:
\begin{align}
\mathcal{T}_{a_2}\theta&=p_2(\mathcal{T}_{a_2}\theta)+p_3(\mathcal{T}_{a_2}\theta),\label{for:44}\\
\mathcal{T}_{a_1}\psi&=p_1(\mathcal{T}_{a_1}\psi)+p_3(\mathcal{T}_{a_1}\psi)\label{for:57}.
\end{align}
Since $p_3(\mathcal{T}_{a_2}\theta)\in U_{a_1}^{m}\bigcap U_{a_2}^{m}$ and
$p_3(\mathcal{T}_{a_1}\psi)\in U_{a_1}^{m}\bigcap U_{a_2}^{m}$ then by (4) of
Lemma \ref{le:12} there exist $\Omega,\Omega'\in
\mathcal{L}_{0,L}^{m}$ such that
\begin{align}
&\mathcal{T}_{a_1}\mathcal{T}_{a_2}\Omega=p_3(\mathcal{T}_{a_2}\theta),\label{for:45}\\
&\mathcal{T}_{a_1}\mathcal{T}_{a_2}\Omega'=p_3(\mathcal{T}_{a_1}\psi)\label{for:46}
\end{align}
satisfying
\begin{align*}
&\norm{\Omega}'_{m}\leq
C_{m,\eta}\norm{p_3(\mathcal{T}_{a_2}\theta)}'_{m}\notag\\
&\leq C_{m,\eta}\norm{\mathcal{T}_{a_2}\theta}'_{m}\leq C_{m,\eta}
C_{m,\eta}\norm{\theta}'_{m}.
\end{align*}
Hence \eqref{L7-2-3} holds. Now we are going to show that $\Omega$ satisfies \eqref{L7-2-1} and \eqref{L7-2-2}.
(Naturally, by symmetry all three inequalities will also hold for $\Omega'$).
Substituting \eqref{for:45} into \eqref{for:44} we get
\begin{align}\label{for:71}
\mathcal{T}_{a_2}(\theta-\mathcal{T}_{a_1}\Omega)=p_2(\mathcal{T}_{a_2}\theta).
\end{align}
Since $L(\psi,\theta)^{(a_1,a_2)}=\mathcal{T}_{a_1}\psi-\mathcal{T}_{a_2}\theta=\omega$ and\\
$\mathcal{T}_{a_1}\psi\in U^{(1,m)}_{(a_1,a_2)}\bigoplus
U^{(3,m)}_{(a_1,a_2)}$, we have $-p_2(\mathcal{T}_{a_2}\theta)=p_2\omega$.
Then by \eqref{for:47}
\begin{align}\label{for:55}
\norm{p_2(\mathcal{T}_{a_2}\theta)}'_m=\norm{p_2\omega}'_m\leq
C_{m,a_1,a_2}\norm{\omega}'_m.
\end{align}
Combining \eqref{for:71} and \eqref{for:55} and using \eqref{for:10}
of Lemma \ref{le:12}
we have
\begin{align*
\norm{\theta-\mathcal{T}_{a_1}\Omega}'_{m}&\leq
C_{m,\eta}\norm{p_2(\mathcal{T}_{a_2}\theta)}'_{m}\leq C_{m,\eta}C_{m,a_1,
a_2}\norm{\omega}'_m.
\end{align*}
This gives \eqref{L7-2-1}.
Substituting \eqref{for:45} and \eqref{for:46} into
$p_3(\mathcal{T}_{a_1}\psi)-p_3(\mathcal{T}_{a_2}\theta)=p_3\omega$ gives
\begin{align}\label{for:72}
\mathcal{T}_{a_1}\mathcal{T}_{a_2}\Omega'-\mathcal{T}_{a_1}\mathcal{T}_{a_2}\Omega_2=p_3\omega.
\end{align}
Since $\mathcal{T}_{a_2}\theta\in U^{(2,m)}_{(a_1,a_2)}\bigoplus
U^{(3,m)}_{(a_1,a_2)}$, $p_1(\mathcal{T}_{a_1}\psi)=p_1\omega$. Then by
\eqref{for:47} we have
\begin{align}\label{for:59}
\norm{p_1(\mathcal{T}_{a_1}\psi)}'_m=\norm{p_1\omega}'_m\leq
C_{m,a_1,a_2}\norm{\omega}'_m.
\end{align}
Substituting \eqref{for:46} into \eqref{for:57} gives
\begin{align}\label{for:70}
\mathcal{T}_{a_1}(\psi-\mathcal{T}_{a_2}\Omega')=p_1(\mathcal{T}_{a_1}\psi).
\end{align}
Combining \eqref{for:72} and \eqref{for:70} gives
\begin{align*}
&\mathcal{T}_{a_1}(\psi-\mathcal{T}_{a_2}\Omega)\\
&=\mathcal{T}_{a_1}(\psi-\mathcal{T}_{a_2}\Omega')+\bigl(\mathcal{T}_{a_1}\mathcal{T}_{a_2}\Omega'-
\mathcal{T}_{a_1}\mathcal{T}_{a_2}\Omega\bigl)\\
&=p_1(\mathcal{T}_{a_1}\psi)+p_3\omega,
\end{align*}
then use \eqref{for:10} of Lemma \ref{le:12} again and combine with
\eqref{for:59}:
\begin{align*
&\norm{\psi-\mathcal{T}_{a_2}\Omega}'_{m}\notag\\
&\leq
C_{m,\eta}\norm{p_1(\mathcal{T}_{a_1}\psi)+p_3\omega}'_{m}\notag\\
&\leq
C_{m,\eta}\norm{p_1(\mathcal{T}_{a_1}\psi)}'_m+C_{m,\eta}\norm{p_3\omega}'_m\notag\\
&\leq
C_{m,\eta}C_{m,a_1,a_2}\norm{\omega}'_m+C_{m,\eta}\norm{\omega}'_m,
\end{align*}
i.e. \eqref{L7-2-2} also holds.
\end{proof}
Next we will show there exists a upper bound for $C_{m,a_1,a_2}$ in
$B_\eta(a)\bigcup B_\eta(b)$ for sufficiently small $\eta$ if $a_1$
and $a_2$ commute and $a_1,a_2\in Z(L)$. The proof uses a fairly
straightforward compactness argument that works because we operate
in a fixed Hilbert space.
\begin{lemma}\label{le:4} In the notation of the previous lemma the constant $C_{m,a_1,a_2}$ can be chosen to depend only on $m$ and $\eta$.
\end{lemma}
\begin{proof}
Let $U'=\{(a_1,a_2)\in B_\eta(a)\times B_\eta(b)| a_1\text{ and
}a_2\text{ commute and } a_1,a_2\in Z(L)\}.$ Since $U'$ is a compact
set we just need to show: for any pair $a_1,a_2\in U'$ there exits
$\delta\,,C'>0$ such that for any pair $z_1,z_2\in U'$ satisfying
$z_1\in B_\delta(a_1)$ and $z_2\in B_\delta(a_2)$,
$C_{m,z_1,z_2}<C'$.
If not, for any $n\in\mathbb{N}$ there exist $z_n\in B_{\frac{1}{n}}(a_1)$,
$b_n\in B_{\frac{1}{n}}(a_2)$ where $z_n,b_n\in U'$ and $\theta_n\in
U^{(1,m)}_{(z_n,b_n)}$, $\psi_n\in U^{(2,m)}_{(z_n,b_n)}$ with
$\norm{\theta_n}'_m+\norm{\psi_n}'_m=1$ while $\norm{\omega_n}'_m<
\frac{1}{n}$ where $\omega_n=\theta_n-\psi_n$. By assumption there
exists a sequence $c_n\rightarrow 0$ as $n\rightarrow \infty$ such
that
\begin{align}\label{for:74}
\norm{\mathcal{T}_{z_n}-\mathcal{T}_{a_1}}'_m+\norm{\mathcal{T}_{b_n}-\mathcal{T}_{a_2}}'_m\leq c_n.
\end{align}
Since $\theta_n\in U_{z_n}^{m}$ and $\psi_n\in U_{b_n}^{m}$ by
Corollary \ref{cor:3} there exists $\Omega_n\,,\Omega'_n\in
\mathcal{L}_{0,L}^m$ such that
\begin{align}\label{for:73}
\mathcal{T}_{z_n}\Omega_n=\theta_n\qquad \mathcal{T}_{b_n}\Omega'_n=\psi_n
\end{align}
with estimate
\begin{align}\label{for:77}
\norm{\Omega_n,\Omega'_n}'_m\leq
C_{m,\eta}\norm{\theta_n,\psi_n}'_m.
\end{align}
Let $\mathcal{T}_{a_1} \Omega_n-\mathcal{T}_{a_2} \Omega'_n=\omega_n'$ then combining
\eqref{for:74} and \eqref{for:73} we obtain
\begin{align}\label{for:76}
\norm{\omega_n'}'_{m}&\leq
\norm{\omega_n}'_{m}+\norm{\mathcal{T}_{z_n}\Omega_n-\mathcal{T}_{a_1}\Omega_n}'_m\notag\\
&+
\norm{\mathcal{T}_{b_n}\Omega'_n-\mathcal{T}_{a_2}\Omega'_n}'_m\notag\\
&\leq
\frac{1}{n}+c_n(\norm{\theta_n}'_m+\norm{\psi_n}'_m)=\frac{1}{n}+c_n.
\end{align}
By Lemma \ref{le:6}, there exists $\phi_n\in \mathcal{L}_0^{m}$
satisfying
\begin{align}\label{for:75}
&\norm{\Omega_n-\mathcal{T}_{a_2} \phi_n}'_{m}\leq
C_{m,a_1,a_2}\norm{\omega_n'}'_{m}\notag\\
&\norm{\Omega'_n-\mathcal{T}_{a_1} \phi_n}'_{m}\leq
C_{m,a_1,a_2}\norm{\omega_n'}'_{m}\notag\\
&\norm{\phi_n}'_{m}\leq C_{m,\eta}\norm{\Omega_n,\Omega'_n}'_{m}.
\end{align}
Let $c=\max\{\norm{\mathcal{T}_z}'_m,\forall z\in U'\}$ then by
\eqref{for:74}, \eqref{for:77}, \eqref{for:76} and \eqref{for:75} we
have
\begin{align*}
&\norm{\mathcal{T}_{z_n}(\Omega_n-\mathcal{T}_{b_n} \phi_n)}'_{m}+\norm{\mathcal{T}_{b_n} (\Omega'_n-\mathcal{T}_{z_n}\phi_n)}'_{m}\\
&\leq c\norm{\Omega_n-\mathcal{T}_{b_n} \phi_n}'_{m}+c\norm{\Omega'_n-\mathcal{T}_{z_n}\phi_n}'_{m}\\
&\leq c\norm{\Omega_n-\mathcal{T}_{a_2} \phi_n}'_{m}+c\norm{\mathcal{T}_{a_2}\phi_n-\mathcal{T}_{b_n} \phi_n}'_{m}\\
&+c\norm{\Omega'_n-\mathcal{T}_{a_1}\phi_n}'_{m}+c\norm{\mathcal{T}_{a_1} \phi_n-\mathcal{T}_{z_n} \phi_n}'_{m}\\
&\leq 2cC_{m,a_1,a_2}\norm{\omega_n'}'_{m}+2cc_n\norm{\phi_n}'_{m}\\
&\leq 2cC_{m,a_1,a_2}(\frac{1}{n}+c_n)+2cc_nC^2_{m,\eta}.
\end{align*}
Hence
\begin{align}\label{for:78}
\norm{\mathcal{T}_{z_n} (\Omega_n-\mathcal{T}_{b_n} \phi_n)}'_{m}+\norm{\mathcal{T}_{b_n}
(\Omega'_n-\mathcal{T}_{z_n}\phi_n)}'_{m}\rightarrow 0
\end{align}
as $n\rightarrow \infty$.
On the other hand notice
\begin{align*}
\mathcal{T}_{z_n}\mathcal{T}_{b_n} \phi_n=\mathcal{T}_{b_n} \mathcal{T}_{z_n}\phi_n\in U^{(3,m)}_{(z_n,b_n)}
\end{align*}
then by \eqref{for:73} for any $n\in\mathbb{N}$ we have
\begin{align*}
&\norm{\mathcal{T}_{z_n} (\Omega_n-\mathcal{T}_{b_n} \phi_n)}'_{m}+\norm{\mathcal{T}_{b_n}
(\Omega'_n-\mathcal{T}_{z_n}\phi_n)}'_{m}\\
&=\norm{\theta_n-\mathcal{T}_{z_n}\mathcal{T}_{b_n} \phi_n}'_{m}+\norm{\psi_n-\mathcal{T}_{z_n}
\mathcal{T}_{b_n}
\phi_n}'_{m}\\
&\geq \norm{\theta_n}'_{m}+\norm{\psi_n}'_{m}\\
&=1
\end{align*}
which contradicts \eqref{for:78}. \end{proof}
\subsection{Construction of splittings in $\mathcal{H}_{0,L}^m$}\label{splitting-final}
The following lemma is proved exactly the same way as Lemma
\ref{le:12} using Lemma \ref{le:16}, instead of
Corollary~\ref{cor:3} and Lemma \ref{le:1} instead of
Corollary~\ref{cor:4}.
\begin{lemma}\label{le:13} Let $m_1$ be defined as in Lemma \ref{le:11}.
If $\eta$ is small enough the following properties hold for any
$z\in B_\eta(a)\bigcup B_\eta(b)$ and $z\in Z(L)$:
\begin{enumerate}
\item if $f=\mathcal{T}_z \Omega$ and $\Omega\in \mathcal{H}_{0,L}^m$ then
$\sum^{z}f=0$ if $m>\frac{r\dim M}{2}+3r$. \label{for:7}
\item if $\mathcal{T}_z \Omega=0$ with $\Omega\in\mathcal{H}_{0,L}^{m}$ then
$\Omega=0$ if $m>\frac{r\dim M}{2}+3r$.\label{for:8}
\item if $\mathcal{T}_z \Omega=f\in\mathcal{H}_{0,L}^{m}$ and $\Omega\in C^1\bigcap \mathcal{H}_{0,L}^{0}$
then $\Omega\in\mathcal{H}_{0,L}^{m-2r-2}$ and
$\norm{\Omega}_{m-2r-2}\leq C_{m,\eta} \norm{f}_{m}$ if $m\geq
m_1+r+1$.\label{for:6}
\item if $z_1\in B_{\eta}(a)$ and $z_2\in
B_{\eta}(b)$ and $z_1\,,z_2$ commute with $z_1\,,z_2\in Z(L)$ and
$f\in \mathcal{H}_{0,L}^{m}$ satisfying $\sum^{z_1}f=0$ and
$\sum^{z_2}f=0$ as distributions then there exists $\Omega\in
\mathcal{H}_0^{m-4r-4}$ satisfying $\mathcal{T}_{z_1}\mathcal{T}_{z_2}\Omega=f$ and
$\norm{\Omega}_{m-4r-4}\leq C_{m,\eta} \norm{f}_{m}$ if $m\geq
m_1+2r+2$.
\label{for:107}
\end{enumerate}
\end{lemma}
Notice finite loss of regularity in statements (3) and (4). Let
$$V_{z}^{m}=\{f\in\mathcal{H}_{0,L}^m|\sum^{z}f\text{ is a $0$ distribution}\}. $$
\begin{lemma}
If $m\geq
m_1+r+1$ then
$V_{z}^m$ is a closed subspace of $\mathcal{H}_{0,L}^{m}$.
\end{lemma}
\begin{proof}
If $f_n\in V_{z}^{m}$ and
$f_n\rightarrow f$ in $\mathcal{H}_{0,L}^{m}$, by \eqref{for:6} of
Lemma \ref{le:13}, there exist
$\Omega_n\in\mathcal{H}_{0,L}^{m-2r-2}$ such that $\mathcal{T}_z\Omega_n=f_n$
with estimates $\norm{\Omega_n}_{m-2r-2}\leq
C_{m,\eta}\norm{f_n}_m$. Then there exists a subsequence
$\Omega_{k_n}\rightarrow \Omega$ for some
$\Omega\in\mathcal{H}_{0}^{m-2r-2-\epsilon}$ in
$\mathcal{H}_{0}^{m-2r-2-\epsilon}$ by Rellich's Lemma where
$\epsilon$ is sufficiently small. By continuity of the operator $\mathcal{T}_z$ we
get $\mathcal{T}_z\Omega=f$. Then by \eqref{for:7} of Lemma \ref{le:13} $f\in
V_{z}^m$. Hence $V_{z}^m$ is a closed subspace.\end{proof}
Now we proceed similarly to the previous section. Let $a_1,a_2$ be commuting elements in $Z(L)$ and $a_1\in B_{\eta}(a), \,\,a_2\in B_{\eta}(b)$. Denote
the orthogonal complement of
$V^{(3,m)}_{(a_1,a_2)}=V_{a_1}^{m}\bigcap V_{a_2}^{m}$ in
$V_{a_1}^{m}$ by $V^{(1,m)}_{(a_1,a_2)}$ and in $V_{a_2}^{m}$ by
$V^{(2,m)}_{(a_1,a_2)}$. Then we have a decomposition
\begin{align*}
V_{a_1}^{m}+V_{a_2}^{m}=V^{(1,m)}_{(a_1,a_2)}\bigoplus
V^{(2,m)}_{(a_1,a_2)}\bigoplus V^{(3,m)}_{(a_1,a_2)}.
\end{align*}
For any $f\in
V_{a_1}^{m}+V_{a_2}^{m}$, write $f=\sum_{i=1}^3 f_i$ where $f_i\in
V^{(i,m)}_{(a_1,a_2)}$.
Next lemma is the central part of the splitting argument. Its
conclusion is similar to that Lemma~\ref{le:4} but since we deal
with ordinary Sobolev norms estimates, they are weaker in three respects: (i) initial data have high
regularity $s>r(m+1) $ compared to the regularity $m$ appearing in
the estimates; (ii) there is fixed $(2r+2)$ loss of regularity in
the estimates; (iii) quality of approximation is not linear in
the norm error but is estimated by a certain power of that norm.
\begin{lemma}\label{le:2}
Let $\eta$ be sufficiently small, $a_1, a_2$ as before, and
$m\in\mathbb{N}$ such that $m\geq m_1+2r+2$, $s>r(m+1)$, Suppose that
$\theta\,,\psi \in \mathcal{H}_{0,L}^s$, and
$L(\psi,\theta)^{(a_1,a_2)}=\omega$ with $\norm{\omega}_s=R_0\leq
1$.
Then there exists $\Omega\in \mathcal{H}_{0,L}^{m-2r-2}$ such that
\begin{align}
&\label{L7-5-1}\norm{\theta-\mathcal{T}_{a_1}\Omega}_{m-4r-4}\leq
C_{m,\eta,s}\norm{\omega}_{m}^{\frac{s-2rm+m-2r}{s-rm-r}},\\
&\label{L7-5-2}\norm{\psi-\mathcal{T}_{a_2}\Omega}_{m-4r-4}\leq
C_{m,\eta,s}\norm{\omega}_{m}^{\frac{s-2rm+m-2r}{s-rm-r}},\\
&\label{L7-5-3}\norm{\Omega}_{m-4r-4}\leq
C_{m,\eta}\norm{\theta,\psi}_{m}
\end{align}
\end{lemma}
\begin{proof} The structure of the proof is similar to that of Lemma~\ref{le:6}.
Let $\mathcal{P}_i$, $1\leq i\leq 3$ be the orthogonal projection
from $V_{a_1}^{m}+V_{a_2}^{m}$ to $V^{(i,m)}_{(a_1,a_2)}$. By (1) of
Lemma \ref{le:13}, $\mathcal{T}_{a_2}\theta\in V_{a_2}^{m}$ and
$\mathcal{T}_{a_1}\psi\in V_{a_1}^{m}$ so that
\begin{align}
\mathcal{T}_{a_2}\theta&=\mathcal{P}_2(\mathcal{T}_{a_2}\theta)+\mathcal{P}_3(\mathcal{T}_{a_2}\theta)\label{for:83},\\
\mathcal{T}_{a_1}\psi&=\mathcal{P}_1(\mathcal{T}_{a_1}\psi)+\mathcal{P}_3(\mathcal{T}_{a_1}\psi)\label{for:84}.
\end{align}
Notice that since $s>r(m+1)$ then
$p_3(T_{a_2}\theta)\in \mathcal{H}^m_{0,L}$. Using Theorem
\ref{th:5}, Lemma \ref{le:4} and properties of smoothing operators,
we have for any $t>0$:
\begin{align}\label{crucial}
&\norm{\mathcal{P}_2(\mathcal{T}_{a_2}\theta)}_{m}\\
&= \norm{\mathcal{T}_{a_2}\theta- \mathcal{P}_3(\mathcal{T}_{a_2}\theta)}_{m}\leq \norm{\mathcal{T}_{a_2}\theta- p_3(\mathcal{T}_{a_2}\theta)}_{m}\notag\\
& \leq
C_{rm+r}\norm{\mathcal{T}_{a_2}\theta-p_3(\mathcal{T}_{a_2}\theta)}'_{r(m+1)}\notag\\
& \leq C_{rm+r}C_{rm+r,\eta}\norm{\omega}'_{r(m+1)}\leq
C_{rm+r}C_{rm+r,\eta}\norm{\omega}_{r(m+1)}\notag\\& \leq
C_{m,\eta}\norm{(I-S_t)\omega}_{r(m+1)}+C_{m,\eta}\norm{S_t\omega}_{r(m+1)}\notag\\&\leq
C_{m,\eta}C_{s,s-r(m+1)}t^{r(m+1)-s}\norm{\omega}_{s}+C_{m,\eta}C_{r(m+1)-m,m}t^{r(m+1)-m}\norm{\omega}_{m}.\notag
\end{align}
If we let
$t=R_0^{\frac{1}{s-rm-r}}\norm{\omega}_{m}^{\frac{1}{rm+r-s}}$, then
we have
\begin{align}
&\norm{\mathcal{P}_2(\mathcal{T}_{a_2}\theta)}_{m}\notag\\
&\leq
C_{m,\eta}C_{s,s-r(m+1)}\norm{\omega}_{m}+C_{m,\eta}C_{r(m+1)-m,m}R_0^{\frac{rm-m}{s-rm}}\norm{\omega}_{m}^{\frac{s-2rm+m-2r}{s-rm-r}}\notag\\
&\leq
C_{m,\eta,s}\norm{\omega}_{m}^{\frac{s-2rm+m-2r}{s-rm-r}}\label{for:85}.
\end{align}
Similarly we have
\begin{align}\label{for:89}
\norm{\mathcal{P}_1(\mathcal{T}_{a_1}\psi)}_{m}\leq
C_{m,\eta,s}\norm{\omega}_{m}^{\frac{s-2rm+m-2r}{s-rm-r}}.
\end{align}
\begin{remark}As we mentioned before, derivation of \eqref{for:85} and \eqref{for:89} (more precisely, the second inequality in\eqref{crucial} and similar inequality for $\psi$)
is the only but crucial place where the auxiliary norms $\norm{\cdot}'$ appear
in the main line of proof of our main theorems. Those norms can be used due to condition ($\mathfrak B$). \footnote{Variations of our method work in other situations where that condition does not hold or even where there are no stable directions altogether, \cite{DFK-parabolic}. However in those situations the ``target'' action is unique. Hence there is no problem of uniformity of estimates in the parametric family of standard perturbations that we handle in Lemma~\ref{le:4}.} \end{remark}
Now we repeat the arguments from the proof of Lemma~\ref{le:6} with appropriate modifications for the norms. Algebra is identical.
By \eqref{for:107} of Lemma \ref{le:13} there exist
$\Omega,\Omega'\in \mathcal{H}_0^{m-4r-4}$ such that
\begin{align}
&\mathcal{T}_{a_1}\mathcal{T}_{a_2}\Omega=\mathcal{P}_3(\mathcal{T}_{a_2}\theta),\label{for:81}\\
&\mathcal{T}_{a_1}\mathcal{T}_{a_2}\Omega'=\mathcal{P}_3(\mathcal{T}_{a_1}\psi)\label{for:82}
\end{align}
satisfying
\begin{align}\label{for:80}
&\norm{\Omega}_{m-4r-4}\notag\\
&\leq
C_{m,\eta}\norm{\mathcal{P}_3(\mathcal{T}_{a_2}\theta)}_{m}\leq C_{m,\eta}\norm{\mathcal{T}_{a_2}\theta}_{m}\notag\\
&\leq C_{m,\eta} C_{m,\eta}\norm{\theta}_{m}.
\end{align}
This gives \eqref{L7-5-3}. Similar inequality holds for $\Omega'$ although we do not use it.
\noindent Substituting \eqref{for:81} into \eqref{for:83} we obtain
\begin{align}\label{for:50}
\mathcal{T}_{a_2}(\theta-\mathcal{T}_{a_1}\Omega)=\mathcal{P}_2(\mathcal{T}_{a_2}\theta).
\end{align}
By \eqref{for:85} and \eqref{for:6} of Lemma \ref{le:13} the
following inequalities hold:
\begin{align}\label{for:86}
&\norm{\theta-\mathcal{T}_{a_1}\Omega}_{m-4r-4}\notag\\
&\leq
C_{m,\eta}\norm{\mathcal{P}_2(\mathcal{T}_{a_2}\theta)}_m\notag\\
&\leq
C_{m,\eta}C_{m,\eta,s}\norm{\omega}_{m}^{\frac{s-2rm+m-2r}{s-rm-r}}.
\end{align}
Thus \eqref{L7-5-1} holds.
Substituting \eqref{for:81} and \eqref{for:82} into the identity
$$\mathcal{P}_3(\mathcal{T}_{a_1}\psi)-\mathcal{P}_3(\mathcal{T}_{a_2}\theta)=\mathcal{P}_3\omega$$
we have
\begin{align}\label{for:87}
\mathcal{T}_{a_1}\mathcal{T}_{a_2}\Omega'-\mathcal{T}_{a_1}\mathcal{T}_{a_2}\Omega=\mathcal{P}_3\omega.
\end{align}
Substituting \eqref{for:82} into \eqref{for:84} we have
\begin{align}\label{for:88}
\mathcal{T}_{a_1}(\psi-\mathcal{T}_{a_2}\Omega')=\mathcal{P}_1(\mathcal{T}_{a_1}\psi).
\end{align}
Combine \eqref{for:87} and \eqref{for:88} it follows
\begin{align*}
&\mathcal{T}_{a_1}(\psi-\mathcal{T}_{a_2}\Omega)\\
&=\mathcal{T}_{a_1}(\psi-\mathcal{T}_{a_2}\Omega')+\bigl(\mathcal{T}_{a_1}\mathcal{T}_{a_2}\Omega'-
\mathcal{T}_{a_1}\mathcal{T}_{a_2}\Omega\bigl)\\
&=\mathcal{P}_1(\mathcal{T}_{a_1}\psi)+\mathcal{P}_3\omega,
\end{align*}
then use \eqref{for:10} of Lemma \ref{le:13} again and combine
\eqref{for:89} we have
\begin{align*}
&\norm{\psi-\mathcal{T}_{a_2}\Omega}_{m-4r-4}\\
&\leq
C_{m,\eta}\norm{\mathcal{P}_1(\mathcal{T}_{a_1}\psi)+\mathcal{P}_3\omega}_{m}\\
&\leq
C_{m,\eta}\norm{\mathcal{P}_1(\mathcal{T}_{a_1}\psi)}_m+C_{m,\eta}\norm{\mathcal{P}_3\omega}_m\\
&\leq
C_{m,\eta}C_{m,\eta,s}\norm{\omega}_{m}^{\frac{s-2rm+m-2r}{s-rm-r}}+C_{m,\eta}\norm{\omega}_m\\
&\leq C_{m,\eta,s}\norm{\omega}_{m}^{\frac{s-2rm+m-2r}{s-rm-r}}.
\end{align*}
This gives \eqref{L7-5-2} and completes the proof.
\end{proof}
Using the Sobolev embedding theorem we translate estimates for the previous lemma to those in $C^m$ norms that are used in our iteration process. Recall that regularity threshold $m_1$ has been defined in \eqref{eqm1},\eqref{for:31} and \eqref{for:102}. Thus we obtain
\begin{corollary}[\bf{Main Estimate}]\label{le:3}
Let $\eta$ be sufficiently small and $m\in\mathbb{N}$ with $m\geq m_0=
m_1-\frac{\dim M}{2}-2r-3$. For any $a_1\in B_\eta(a)$ and $a_2\in
B_\eta(b)$ where $a_1\,,a_2$ commute and $a_1\,,a_2\in Z(L)$ and any
two $C^{s}$ $L$-invariant maps
$\mathcal{F}\,,\mathcal{G}:M\rightarrow \mathfrak{N}$ with
$L(\mathcal{G},\mathcal{F})^{(a_1,a_2)}=\Psi$ satisfying $\int_M
\mathcal{F}(x)dx=\int_M \mathcal{G}(x)dx=0$ and
$\norm{\Psi}_{C^s}\leq 1$, then there exists $\Omega\in C^{m}$ such
that
\begin{align*}
&\norm{\mathcal{F}-(\Omega\circ
a_1-\textrm{Ad}_{a_1}\Omega)}_{C^{m}}\leq
C_{m,\eta,s}\norm{\Psi}^{\gamma(m+\sigma,s)}_{C^{m+\sigma}}\\
&\norm{\mathcal{G}-(\Omega\circ
a_2-\textrm{Ad}_{a_2}\Omega)}_{C^{m}}\leq
C_{m,\eta,s}\norm{\Psi}^{\gamma(m+\sigma,s)}_{C^{m+\sigma}}\\
&\norm{\Omega}_{C^{m}}\leq
C_{m,\eta}\norm{\mathcal{F},\mathcal{G}}_{C^{m+\sigma}}.
\end{align*}
where $s\ge s(m)$ and
\begin{equation}\label{eqsm}
s(m)=r(m+\dim M/2+4r+6),\end{equation}
$\sigma=[\frac{\dim M}{2}+4r+6]$ and
$\gamma(m,s)=\frac{s-2rm+m-2r}{s-rm-r}$.
\end{corollary}
\section{Iteration procedure and completion of proof}\label{section-KAM}
\subsection{Scheme of proof}
Let $\bar{e}$ be the image of the identity $e$ of $G$
(corr. $G\ltimes\mathbb{R}^N$) on $M$ and $d(\,,)$ a right invariant metric
on $G$ (corr. $G\ltimes\mathbb{R}^N$). For a small neighborhood $V$ of the
tangent bundle $T_{e}G$ or $T_{e}G\ltimes\mathbb{R}^N$, we can identify the
metrics on $V$, $\exp V$ and $\exp V\cdot \bar{e}$, that is, for any
$x\in V\cdot \bar{e}$, by writing $x=\exp(v)\cdot \bar{e}$ where
$v\in V$ then $d(x,\bar{e})=d(\exp v,e)=d(v,0)$. For simplicity we
denote $\norm{x}=\norm{\exp(v)}=\norm{v}=d(x,\bar{e})$.
For any continuous map $f$ valued on $\exp V$ (corr. $\exp V\cdot
\bar{e}$), we can define $\norm{f}_{C^0}=\max_{y\in D(f)}d(f(y),e)$
(corr. $\norm{f}_{C^0}\stackrel{\text{def}}{=}\max_{y\in
D(f)}d(f(y),\bar{e})$) where $D(f)$ is the domain of $f$.
Assuming $\alpha'$ is
a $C^\infty$ action that is $C^{\ell r}$ close to $\alpha_{D_+}$
(where $\ell$ is fixed and will be determined in the proof, see
\eqref{for:115}), there is a $C^\infty$ orbit conjugacy $H_1$
between $\alpha'$ and $\alpha_{D_+}$ which is $C^{\ell}$ close to
identity, see Section~\ref{sec:prelim}. We define a linear
$L$-averaging operator ``$-$'' on the set of smooth maps from $M$ to
$\mathfrak{N}$:
$$\overline{f}(x)\stackrel{\text{def}}=\int_L\text{Ad}_{l^{-1}}f(l\cdot
x)dl.$$
Let $\tilde{\alpha}=H_1^{-1}\circ\alpha'\circ H_1$.
We can represent $\tilde{\alpha}=\exp(R) \cdot\alpha_{D_+}$ where
$R$ is close to $0$
in $C^{\ell}$ norm and we will show that
$\tilde{\alpha}$ is smoothly conjugate to $\alpha$. The conjugacy is
produced for two regular generators and by Lemma \ref{le:9} it
works for all elements of the action. This proof is similar to the
iterative proof in \cite{Damjanovic4} with essential additions of the balancing of the norms explained below and the
parameter adjustment argument in Section~\ref{sec:8.5}.
Following the scheme described in Section \ref{sec:1} $R$ is an
$L$-invariant map on $M$ (i.e, $\overline{R}=R$) valued on
$\mathfrak{L}^\bot$, then at each step of the iterative procedure we
solve the linearized equation \eqref{for:101}:
\begin{align*}
&\Omega\circ\alpha(a_1)-\textrm{Ad}_{a_1}\Omega=R_{a},\notag\\
&\Omega\circ\alpha(a_2)-\textrm{Ad}_{a_2}\Omega=R_{b}
\end{align*}
approximately where $a_1\in B_\eta(a)\bigcap Z(L)$, $a_2\in
B_\eta(b)\bigcap Z(L)$ and $a_1$ and $a_2$ commute. By the Main
Estimate (Corollary~\ref{le:3}) this linearized equation has an approximate
solution $\Omega$ which is $C^\infty$ although we can only compare
its $C^m$ norm to the in $C^{m+\sigma}$ of $R$ where $\sigma$ is
large but fixed. The norm of the error by construction in the Main
Estimate
is comparable to a certain power of the norm of
$L(R_{b}, R_{a})^{(a_1,a_2)}$. Thus it is small with respect to $R$
by Lemma \ref{le:5}, but comparison again comes with fixed loss of
derivatives.
Very essential part of the argument is the proper balancing of various norms.
Notice that in the Main Estimate the estimate with fixed loss of regularity is obtained for the $C^m$ norm
for an $m$ above the threshold $m_0$ determined by the data (see \eqref{eqm1},\eqref{for:31} and \eqref{for:102})
under the additional assumption that much higher $C^s$ norm is bounded, where $s=s(m)$ (see \eqref{eqsm}), in particular $s>rm$.
Thus we cannot simply make closeness assumption for a fixed norm and offset the finite loss of regularity by straightforward application of smoothing. Instead we assume closeness of the perturbation to the original action in a high $C^\ell$ norm where $\ell$ is
defied in \eqref{for:115}. Essential requirement is $\ell\ge 3\ell_0>3s(m_0)$, see \eqref{for:106}. In particular $l> 3rm_0$. In the iterative step
we use smoothing both for the data and for the solution. By making very strong requirements on the decrease of the $C^0$ norms of the successive errors that are allowed by the quadratic convergence and controlling the growth of $C^\ell$ norms (see \eqref{for:32}), we guarantee via interpolation inequalities fast decrease of still high $C^{\ell_0}$ norms. Since $l_0> s(m_0)$ this allows for successive applications of the Main Estimate.
At the end we guarantee convergence of conjugacies in $C^1$ and appeal to the a priory regularity
to conclude that is is $C^\infty$. Of course our argument also gives convergence in $C^{l_0}$ but since we
only work in finite regularity we do not produce $C^\infty$ conjugacy directly.
\subsection{Smoothing operators and some norm inequalities} To overcome this fixed loss of derivatives at each step of
the iteration process, it is standard (see for example
\cite{Zehnder}) to introduce the family of smoothing operators:
$\{S_t,t\in\mathbb{R}\}$. Since $S_tR$ is not necessarily $L$-invariant we combine smoothing with $L$-averaging and
will solve approximately the following system:
\begin{align}
&\Omega\circ\alpha(a_1)-\textrm{Ad}_{a_1}\Omega=\overline{S_tR_{a}},\notag\\
&\Omega\circ\alpha(a_2)-\textrm{Ad}_{a_2}\Omega=\overline{S_tR_{b}}.
\end{align}
Denote the averages
$\int_{M}\overline{S_tR_{a}}d\mu=\mathcal{A}(\overline{S_tR_{a}})$
and
$\int_{M}\overline{S_tR_{b}}d\mu=\mathcal{A}(\overline{S_tR_{b}})$. Recall that $L(\mathcal{F},\mathcal{G})^{(a_1,a_2)}=\mathcal{F}\circ
a_2-\textrm{Ad}_{a_2}\mathcal{F}-\mathcal{G}\circ
a_1+\textrm{Ad}_{a_1}\mathcal{G}=\mathcal{T}_{a_2}\mathcal{F}-\mathcal{T}_{a_1}\mathcal{G}.$
By Lemma \ref{le:5} and $L$-invariance of the data i.e. $\overline{R_{b}}=R_{b}$ and
$\overline{R_{a}}=R_{a}$, we have
\begin{align}\label{for:16}
&\norm{L(\overline{S_tR_{b}}-\mathcal{A}(\overline{S_tR_{b}}),\overline{S_tR_{a}}-\mathcal{A}(\overline{S_tR_{a}}))^{(a_1,a_2)}}_{C^m}\\
&=\norm{L(\overline{S_tR_{b}},\overline{S_tR_{a}})^{(a_1,a_2)}-\int_{M}L(\overline{S_tR_{b}},\overline{S_tR_{a}})^{(a_1,a_2)}d\mu}_{C^m}\notag\\
&\leq C\norm{L(\overline{S_tR_{b}},\overline{S_tR_{a}})^{(a_1,a_2)}}_{C^m}\notag\\
&\leq C\norm{L(\overline{R_{b}},\overline{R_{a}})^{(a_1,a_2)}}_{C^m}+C\norm{L(\overline{(I-S_t)R_{b}},\overline{(I-S_t)R_{a}})^{(a_1,a_2)}}_{C^m}\notag\\
&\leq C\norm{L(R_{b},R_{a})^{(a_1,a_2)}}_{C^m}+C_{m,\eta}\norm{(I-S_t)R_{a},(I-S_t)R_{b}}_{C^m}\notag\\
&\leq
C\norm{L(R_{b},R_{a})^{(a_1,a_2)}}_{C^m}+C_{m,j,\eta}t^{-j}\norm{R_{a},R_{b}}_{C^{j+m}}\notag\\
&\leq
C_{m,\eta}\norm{R_{a},R_{b}}_{C^{m}}\norm{R_{b},R_{a}}_{C^{m+1}}+C_{m,j,\eta}t^{-j}\norm{R_{a},R_{b}}_{C^{j+m}}\notag.
\end{align}
for any $j\in \mathbb{N}$.
\subsection{Iterative step and the error estimate}\label{sec:8.2}
At each step of the iterative scheme we first choose a smoothing
operator $S_t$ with an appropriately chosen $t$. In order to solve
approximately
\begin{align}
&\Omega\circ a_1-\textrm{Ad}_{a_1}\Omega=\overline{S_tR_{a}}-\mathcal{A}(\overline{S_tR_{a}})\notag\\
&\Omega\circ
a_2-\textrm{Ad}_{a_2}\Omega=\overline{S_tR_{b}}-\mathcal{A}(\overline{S_tR_{b}})
\end{align}
we use the Main Estimate (Corollary~\ref{le:3}) to obtain an
approximate solution $\Omega\in C^{m_0}$:
\begin{align}
&\Omega\circ
a_1-\textrm{Ad}_{a_1}\Omega=\overline{S_tR_{a}}-\mathcal{A}(\overline{S_tR_{a}})-\mathcal{R}
\bigl(\overline{S_tR_{a}}-\mathcal{A}(\overline{S_tR_{a}})\bigl)\notag\\
&\Omega\circ
a_2-\textrm{Ad}_{a_2}\Omega=\overline{S_tR_{b}}-\mathcal{A}(\overline{S_tR_{b}})-\mathcal{R}\bigl(\overline{S_tR_{b}}-\mathcal{A}(\overline{S_tR_{b}})\bigl)
\end{align}
such that for any $m\leq m_0$
\begin{align}\label{for:19}
\norm{\Omega}_{C^{m}}&\leq
C_{m_0,\eta}\norm{\overline{S_tR_{a}}-\mathcal{A}(\overline{S_tR_{a}}),\overline{S_tR_{b}}-\mathcal{A}(\overline{S_tR_{b}})}_{C^{m_0+\sigma}}\notag\\
&\leq CC_{m_0,\eta}\norm{S_tR_{a},S_tR_{b}}_{C^{m_0+\sigma}}\notag\\
&\leq C_{m_0,\eta}\norm{R_{a},R_{b}}_{C^{m_0+\sigma}}
\end{align}
and also
\begin{align}\label{for:121}
\norm{\Omega}_{C^{m}}&\leq
C_{m_0,\eta}\norm{\overline{S_tR_{a}}-\mathcal{A}(\overline{S_tR_{a}}),\overline{S_tR_{b}}-\mathcal{A}(\overline{S_tR_{b}})}_{C^{m_0+\sigma}}\notag\\
&\leq CC_{m_0,\eta}\norm{S_tR_{a},S_tR_{b}}_{C^{m_0+\sigma}}\notag\\
&\leq C_{m_0,\eta}t^{-j}\norm{R_{a},R_{b}}_{C^{m_0+\sigma+j}}
\end{align}
for any $j\in\mathbb{N}$.
Here we used the properties of smoothing operators. As was explained above, applicability of the Main Estimate will be guaranteed by uniform boundedness of the data in $C^{\ell_0}$ for a sufficiently large $\ell_0$.
The error terms $\mathcal{R}
\bigl(\overline{S_tR_{a}}-\mathcal{A}(\overline{S_tR_{a}}))$ and $\mathcal{R}
\bigl(\overline{S_tR_{b}}-\mathcal{A}(\overline{S_tR_{b}}))$ will be estimated by the Main Estimate.
Then we define $H=\exp(S_s\Omega)\cdot I$ for a certain
$s>0$. Notice that by properties of smoothing operators $\norm{S_s\Omega}_{C^1}\leq C\norm{\Omega}_{C^1}$,
and since $\Omega$ is small in $C^1$
throughout the iteration, then $H$ is invertible and $H$ projects
into
an invertible map $h$ on $X$. This is the conjugacy at the iterative step. Let
$\tilde{\alpha}^{(1)}=h^{-1}\circ\tilde{\alpha}\circ h$. This is the new action and we need to estimate is distance form a certain standard algebraic perturbation of the unperturbed action $\alpha_{D_+}$.
For any
$d\in D_+$, $\tilde{\alpha}^{(1)}(d)$ can be lifted to a map on $M$
which we still denote by $\tilde{\alpha}^{(1)}(d)$ without
confusion. Hence there exists $l_d:M\rightarrow L$ such that
$$H\circ\tilde{\alpha}^{(1)}(d)=l_d\cdot (\tilde{\alpha}(d)\circ
H)
$$
so that
\begin{align*}
\tilde{\alpha}^{(1)}&=H^{-1}(l\cdot(\tilde{\alpha}\circ
H))\\
&=\exp(-S_s\Omega\circ\tilde{\alpha}^{(1)})\cdot l\cdot\exp(R\circ
H)\cdot\exp(\text{Ad}_{\alpha_{D_+}}S_s\Omega)\cdot
\alpha_{D_+}\\
&=l\cdot\exp(-S_s\Omega\circ (l^{-1}\cdot\tilde{\alpha}^{(1)}))\cdot\exp\bigl(S_s\Omega(\exp(\mathcal{A}R)\cdot\alpha_{D_+})\bigl)\\
&\cdot\exp\bigl(-S_s\Omega(\exp(\mathcal{A}R)\cdot\alpha_{D_+})\bigl)\cdot\exp(S_s\Omega\circ\alpha_{D_+})\\
&\cdot\exp(-S_s\Omega\circ\alpha_{D_+})\cdot\exp(R\circ H)\cdot\exp(-R)\cdot\exp(S_s\Omega\circ\alpha_{D_+})\\
&\cdot\exp(-S_s\Omega\circ\alpha_{D_+})\cdot\exp(R)\cdot\exp(\text{Ad}_{\alpha_{D_+}}S_s\Omega)\cdot\exp(-\mathcal{A}R)\\
&\cdot\exp(S_s\Omega\circ\alpha_{D_+}-R-\text{Ad}_{\alpha_{D_+}}S_s\Omega+\mathcal{A}R)\\
&\cdot\exp(-S_s\Omega\circ\alpha_{D_+}+R+\text{Ad}_{\alpha_{D_+}}S_s\Omega-\mathcal{A}R)\\
&\cdot\exp(\mathcal{A}R)\cdot\alpha_{D_+}.
\end{align*}
The new error is:
\begin{align*}
\exp(R'_{(1)})=l^{-1}\cdot\tilde{\alpha}^{(1)}\cdot\bigl(\exp(\mathcal{A}R)\cdot\alpha_{D_+}\bigl)^{-1}
\end{align*}
and it can be decomposed as $\exp(R'_{(1)})=E_2\cdot\exp(E_1)$ where:
\begin{enumerate}
\item $\exp(E_1)$ is the error coming from solving the linearized equation only
approximately:
\begin{align*}
E_1&=\mathcal{R}(\overline{S_tR}-\mathcal{A}(\overline{S_tR}))+\overline{(I-S_t)R}+\mathcal{A}\bigl(\overline{(I-S_t)R}\bigl)\\
&+(I-\text{Ad}_{\alpha_{D_+}})\circ(I-S_s)\Omega
\end{align*}
\item $E_2$ is the standard error coming from the linearization:
\begin{align*}
E_2&=\exp(-S_s\Omega\circ(l^{-1}\cdot\tilde{\alpha}^{(1)}))\cdot\exp\bigl(S_s\Omega(\exp(\mathcal{A}R)\cdot\alpha_{D_+})\bigl)\\
&\cdot\exp\bigl(-S_s\Omega(\exp(\mathcal{A}R)\cdot\alpha_{D_+})\bigl)\cdot\exp(S_s\Omega\circ\alpha_{D_+})\\
&\cdot\exp(-S_s\Omega\circ\alpha_{D_+})\cdot\exp(R\circ H)\cdot\exp(-R)\cdot\exp(S_s\Omega\circ\alpha_{D_+})\\
&\cdot\exp(-S_s\Omega\circ\alpha_{D_+})\cdot\exp(R)\cdot\exp(\text{Ad}_{\alpha_{D_+}}S_s\Omega)\cdot\exp(-\mathcal{A}R)\\
&\cdot\exp(S_s\Omega\circ\alpha_{D_+}-R-\text{Ad}_{\alpha_{D_+}}S_s\Omega+\mathcal{A}R)
\end{align*}
\end{enumerate}
{\em Estimate of $E_1$ in $C^0$.} Using Main Estimate,
properties of
smoothing operators and inequality \eqref{for:16} for $j= \ell-m_0-\sigma$ we have for any
$m\leq m_0$
\begin{align}\label{for:34}
&\norm{\mathcal{R}\bigl(\overline{S_tR}-\mathcal{A}(\overline{S_tR})\bigl)}_{C^{m}}\\
&\leq
C_{m_0,\eta,\ell_0}\norm{L(\overline{S_tR_{b}}-\mathcal{A}(\overline{S_tR_{b}}),\overline{S_tR_{a}}-
\mathcal{A}(\overline{S_tR_{a}}))^{(a_1,a_2)}}_{C^{m_0+\sigma}}^{\gamma}\notag\\
&\leq
C_{m_0,\eta,\ell_0}\norm{R_{b},R_{a}}^\gamma_{C^{m_0+\sigma}}\norm{R_{b},R_{a}}^\gamma_{C^{m_0+\sigma+1}}\notag\\
&+C_{m_0,\eta,j,\ell_0}t^{-\gamma
(\ell-m_0-\sigma)}\norm{R_{a},R_{b}}^\gamma_{C^{\ell}}\notag
\end{align}
where $\gamma=\gamma(m_0+\sigma,\ell_0)$ providing
$\norm{L(\overline{S_tR_{b}}-\mathcal{A}(\overline{S_tR_{b}}),\overline{S_tR_{a}}-\mathcal{A}(\overline{S_tR_{a}}))^{(a_1,a_2)}}_{C^{\ell_0}}$
is bounded throughout the procedure for well chosen $\ell_0$ and $t$
where $a_1=i_0(a)$, $a_2=i_0(b)$ and $\gamma$ is as in Main Estimate
determined by $m_0\,,\sigma\,,\ell_0$. Also, using properties
of smoothing operators:
\begin{align}\label{for:35}
\norm{\overline{(I-S_t)R}}_{C^{0}}&\leq
C\norm{(I-S_t)R}_{C^{0}}\notag\\
&\leq C_{\ell}t^{-\ell}\norm{R}_{C^{\ell}}
\end{align}
and thus
\begin{align}\label{for:66}
\norm{\mathcal{A}\bigl(\overline{(I-S_t)R}\bigl)}_{C^0}\leq
C_{\ell}t^{-\ell}\norm{R}_{C^{\ell}}
\end{align}
and by \eqref{for:121}
\begin{align}\label{for:105}
&\norm{(I-\text{Ad}_{\alpha_{D_+}})\circ(I-S_s)\Omega}_{C^0}\notag\\
&\leq
C\norm{(I-S_s)\Omega}_{C^0}\leq C_{m_0}s^{-m_0}\norm{\Omega}_{C^{m_0}}\notag\\
&\leq
C_{m_0,\eta,\ell}s^{-m_0}t^{-(\ell-m_0-\sigma)}\norm{R_{a},R_{b}}_{C^{\ell}}
\end{align}
Thus, combining \eqref{for:34}, \eqref{for:35},
\eqref{for:66} and \eqref{for:105} we have for
$\gamma=\gamma(m_0+\sigma,\ell_0)$ (since $R$ is kept bounded in
$C^{\ell_0}$ throughout the iteration):
\begin{align}\label{for:25}
\norm{E_1}_{C^{0}}&\leq
C_{m_0,\eta,s}\norm{R}^\gamma_{C^{m_0+\sigma}}\norm{R}^\gamma_{C^{m_0+\sigma+1}}
+C_{\ell}t^{-\ell}\norm{R}_{C^{\ell}}\notag\\
&+C_{m_0,\eta,\ell_0,\ell}t^{-\gamma
(\ell-m_0-\sigma)}\norm{R}^\gamma_{C^{\ell}}+C_{m_0,\eta,\ell}s^{-m_0}t^{-(\ell-m_0-\sigma)}\norm{R}_{C^{\ell}}.
\end{align}
\begin{align*}
E_2^1&=\exp(-S_s\Omega\circ(l_d^{-1}\cdot\tilde{\alpha}^{(1)}))\cdot\exp\bigl(S_s\Omega(\exp(\mathcal{A}R)\cdot\alpha_{D_+})\bigl)\\
E_2^2&=\exp\bigl(-S_s\Omega(\exp(\mathcal{A}R)\cdot\alpha_{D_+})\bigl)\cdot\exp(S_s\Omega\circ\alpha_{D_+})\\
E_2^3&=\exp(-S_s\Omega\circ\alpha_{D_+})\cdot\exp(R\circ
H)\cdot\exp(-R)\cdot\exp(S_s\Omega\circ\alpha_{D_+})\\
E_2^4&=\exp(-S_s\Omega\circ\alpha_{D_+})\cdot\exp(R)\cdot\exp(\text{Ad}_{\alpha_{D_+}}S_s\Omega)\cdot\exp(-\mathcal{A}R)\\
&\cdot\exp(S_s\Omega\circ\alpha_{D_+}-R-\text{Ad}_{\alpha_{D_+}}S_s\Omega+\mathcal{A}R)
\end{align*}
then we have
\begin{align*}
\norm{E_2}_{C^0}&\leq
\norm{E_2^1}_{C^0}+\norm{E_2^2}_{C^0}+\norm{E_2^3}_{C^0}+\norm{E_2^4}_{C^0}
\end{align*}
which follows from the fact that for any small enough
$x_1\,,x_2\,,x_3\,,x_4\in \exp(V)$ by right invariance of metric $d$
we have
\begin{align*}
&d(x_1x_2x_3x_4,e)\\
&\leq
d(x_1x_2x_3x_4,x_2x_3x_4)+d(x_2x_3x_4,x_3x_4)+d(x_3x_4,x_4)+d(x_4,e)\\
&=d(x_1,e)+d(x_2,e)+d(x_3,e)+d(x_4,e).
\end{align*}
Thus there are four terms that we estimate as
follows:\smallskip
\noindent {\small{\bf First term.}} Using \eqref{for:19} we obtain
\begin{align}\label{for:21}
&\norm{\exp\bigl(-S_s\Omega\circ(l_d^{-1}\cdot\tilde{\alpha}^{(1)})\bigl)\cdot\exp\bigl(S_s\Omega(\exp(\mathcal{A}R)\cdot\alpha_{D_+})\bigl)}_{C^0}\notag\\
&\leq\norm{S_s\Omega(\exp(\mathcal{A}R)\cdot\alpha_{D_+})-S_s\Omega\circ(l_d^{-1}\cdot\tilde{\alpha}^{(1)})}_{C^0}+
C\norm{S_s\Omega}_{C^0}^2\notag\\
&\leq
C\norm{S_s\Omega}_{C^1}\norm{R'_{(1)}}_{C^0}+C\norm{\Omega}_{C^{0}}^{2}\notag\\
&\leq
C\norm{\Omega}_{C^1}\norm{R'_{(1)}}_{C^0}+C\norm{\Omega}_{C^{0}}^{2}\notag\\
&\leq
\frac{1}{4}\norm{R'_{(1)}}_{C^0}+C_{m_0,\eta}\norm{R}^2_{C^{m_0+\sigma}}
\end{align}
thus this term is absorbed into
$\norm{R'_{(1)}}_{C^0}$ providing $\norm{\Omega}_{C^1}$ remains sufficiently small
throughout the procedure.\smallskip
\noindent{\small{\bf Second term.}} We estimate similarly to $E_1$:
\begin{align}\label{for:22}
&\norm{\exp\bigl(-S_s\Omega(\exp(\mathcal{A}R)\cdot\alpha_{D_+})\bigl)\cdot\exp(S_s\Omega\circ\alpha_{D_+})}_{C^0}\notag\\
&\leq\norm{S_s\Omega(\exp(\mathcal{A}R)\cdot\alpha_{D_+})-S_s\Omega\circ\alpha_{D_+}}_{C^0}+C\norm{S_s\Omega}_{C^0}^2
\notag\\
&\leq
C\norm{S_s\Omega}_{C^{1}}\norm{R}_{C^0}+C\norm{S_s\Omega}_{C^{0}}^2\notag\\
&\leq
C\norm{\Omega}_{C^{1}}\norm{R}_{C^0}+C\norm{\Omega}_{C^{0}}^2\notag\\
&\leq
C_{m_0,\eta}\norm{R}_{C^0}\norm{R}_{C^{m_0+\sigma}}+C_{m_0,\eta}\norm{R}^2_{C^{m_0+\sigma}}.
\end{align}
\noindent{\small{\bf Third term.}} We first notice:
\begin{align}\label{thirdterm}
&\norm{\exp(R\circ H)\exp(-R)}_{C^0}\notag \\
&\leq
C\norm{R\circ H-R}_{C^0}+C\norm{R}^2_{C^0} \notag\\
&\leq C\norm{R}_{C^1}\norm{S_s\Omega}_{C^0} +C\norm{R}^2_{C^0}\notag\\
&\leq C\norm{R}_{C^1}\norm{\Omega}_{C^0} +C\norm{R}^2_{C^0},
\end{align}
then use \eqref{for:19}, the fact that $\norm{\Omega}_{C^1}$ and
$\norm{R}_{C^0}$ are bounded throughout the procedure and \eqref{thirdterm} we get
\begin{align}\label{for:23}
&\norm{\exp(-S_s\Omega\circ\alpha_{D_+})\cdot\exp(R\circ
H)\cdot\exp(-R)\cdot\exp(S_s\Omega\circ\alpha_{D_+})}_{C^0}\notag\\
&\leq \norm{\text{Ad}_{S_s\Omega\circ\alpha_{D_+}}}_{C^0}\norm{\exp(R\circ H)\exp(-R)}_{C^0}\notag\\
&\leq C(\norm{S_s\Omega\circ\alpha_{D_+}}_{C^0}+1)\norm{\exp(R\circ H)\exp(-R)}_{C^0}\notag\\
&\leq C(\norm{\Omega}_{C^0}+1)\norm{\exp(R\circ H)\exp(-R)}_{C^0}\notag\\
&\leq C\norm{R}_{C^1}\norm{\Omega}_{C^0}
+C\norm{R}^2_{C^0}\notag\\
&\leq
C_{m_0,\eta}\norm{R}_{C^{m_0+\sigma}}\norm{R}_{C^1}+C\norm{R}^2_{C^0}.
\end{align}
\noindent{\small{\bf Fourth term.}} Use \eqref{for:19} and the fact
$C^0$-norm of Lie brackets between $S_s\Omega$, $R$, $\mathcal{A}R$
and $\text{Ad}_{\alpha_{D_+}}S_s\Omega$ are uniformly bounded by
$C_{m_0,\eta}\norm{R}^2_{C^{m_0+\sigma}}$, hence we have
\begin{align}\label{for:24}
&\norm{\exp(-S_s\Omega\circ\alpha_{D_+})\cdot\exp(R)\cdot\exp(\text{Ad}_{\alpha_{D_+}}S_s\Omega)\cdot\exp(-\mathcal{A}R)\notag\\
&\cdot\exp(S_s\Omega\circ\alpha_{D_+}-R-\text{Ad}_{\alpha_{D_+}}S_s\Omega+\mathcal{A}R)}_{C^0}\notag\\
&\leq C_{m_0,\eta}\norm{R}^2_{C^{m_0+\sigma}}.
\end{align}\smallskip
By combining \eqref{for:21}, \eqref{for:22}, \eqref{for:23},
\eqref{for:24} we obtain the following estimate for $E_2$:
\begin{align}\label{for:27}
&\norm{E_2}_{C^0}\leq C_{m_0,\eta}\norm{R}^2_{C^{m_0+\sigma}}
\end{align}
By combining \eqref{for:25} and \eqref{for:27} we obtain
an estimate for the new error for
$\gamma=\gamma(m_0+\sigma,\ell_0)$:
\begin{align}\label{for:33}
\norm{R'_{(1)}}_{C^0}&\leq \norm{E_1}_{C^0}+\norm{E_2}_{C^0}\notag\\
&\leq
C_{m_0,\eta,s}\norm{R}^\gamma_{C^{m_0+\sigma}}\norm{R}^\gamma_{C^{m_0+\sigma+1}}
+C_{\ell}t^{-\ell}\norm{R}_{C^{\ell}}\notag\\
&+C_{m_0,\eta,\ell_0,\ell}t^{-\gamma
(\ell-m_0-\sigma)}\norm{R}^\gamma_{C^{\ell}}+C_{m_0,\eta,\ell}s^{-m_0}t^{-(\ell-m_0-\sigma)}\norm{R}_{C^{\ell}}\notag\\
&+C_{m_0,\eta}\norm{R}^2_{C^{m_0+\sigma}}.
\end{align}
This completes the $C^0$ estimate.
Now we need to prepare for the coordinate change that will make the constant term quadratically small with respect to $\norm{R'_{(1)}}_{C^0}$.
Since $R$ is $L$-invariant then $\mathcal{A}R\in Z(D)\bigcap Z(L)$,
then $\exp(\mathcal{A}R)\cdot\alpha_{D_+}$ can descend to $X$, then
there exists smooth map $R':M\rightarrow \mathfrak{L}^\bot$ such
that
\begin{align}\label{for:68}
\tilde{\alpha}^{(1)}(p(x))=p\bigl(\exp(R'(x))\cdot\exp(\mathcal{A}R)\cdot\alpha_{D_+}(x)\bigl)\qquad
\forall x\in M
\end{align}
where $p$ is the natural projection from $M$ to $X=L\backslash
M$. Notice $l^{-1}\cdot\tilde{\alpha}^{(1)}$ descends to the same
map on $X$ as $\tilde{\alpha}^{(1)}$ does then combine
\eqref{for:33} we have
\begin{align*}
\norm{R'}_{C^n}\leq C\norm{R'_{(1)}}_{C^n}\qquad \forall n\in\mathbb{N}.
\end{align*}
Since $R'$ is $L$-invariant then $\mathcal{A}R'\in Z(D)\bigcap Z(L)$
and $\exp(\mathcal{A}R')\cdot\exp(\mathcal{A}R)\cdot\alpha_{D_+}$
can descend to $X$, hence there exists a smooth map
$R_{(1)}:M\rightarrow \mathfrak{L}^\bot$ such that
\begin{align}\label{for:91}
\tilde{\alpha}^{(1)}(p(x))=p\bigl(\exp(R_{(1)}(x))\cdot\exp(\mathcal{A}R')\cdot\exp(\mathcal{A}R)\cdot\alpha_{D_+}(x)\bigl)\qquad
\forall x\in M
\end{align}
and hence
\begin{align*}
\norm{R_{(1)}}_{C^n}\leq C\norm{R'}_{C^n}\qquad \forall n\in\mathbb{N}.
\end{align*}
Combine \eqref{for:68} and \eqref{for:91} we have
\begin{align*}
l\cdot\exp(R')= \exp(R_{(1)})\cdot\exp(\mathcal{A}R').
\end{align*}
where $l:M\rightarrow L$. Notice $\log(\mathcal{A}R')\in
\mathfrak{L}^\bot$ and thus it follows
\begin{align*}
R'=R_{(1)}+\mathcal{A}R'+\text{Res}
\end{align*}
where $\norm{\text{Res}}_{C^0}\leq C\norm{R',R_{(1)}}^2_{C^0}\leq
C\norm{R'_{(1)}}^2_{C^0} $. Then integrate each side of above
equation we have
\begin{align}\label{for:92}
\norm{\int_MR_{(1)}d\mu}=\norm{\int_M\text{Res} d\mu}\leq
C\norm{R'_{(1)}}^2_{C^0}
\end{align}
and
\begin{align}\label{for:67}
&\norm{R_{(1)}}_{C^0}\leq C\norm{R'_{(1)}}_{C^0}\notag\\
&\leq
C_{m_0,\eta,s}\norm{R}^\gamma_{C^{m_0+\sigma}}\norm{R}^\gamma_{C^{m_0+\sigma+1}}
+C_{\ell}t^{-\ell}\norm{R}_{C^{\ell}}\notag\\
&+C_{m_0,\eta,\ell_0,\ell}t^{-\gamma
(\ell-m_0-\sigma)}\norm{R}^\gamma_{C^{\ell}}+C_{m_0,\eta,\ell}s^{-m_0}t^{-(\ell-m_0-\sigma)}\norm{R}_{C^{\ell}}\notag\\
&+C_{m_0,\eta}\norm{R}^2_{C^{m_0+\sigma}}.
\end{align}
{\em Estimate of the new error in $C^\ell$.}
>From $\tilde{\alpha}^{(1)}=h^{-1}\tilde{\alpha}h$, using the fact
that $\Omega$ satisfies the estimate \eqref{for:121} we have:
\begin{align}\label{Clerror}
\norm{R_{(1)}}_{C^\ell}&\leq C_{\ell,\eta}
(\norm{S_s\Omega}_{C^\ell}+1+\norm{R}_{\ell})\leq C_{\ell,\eta}
(s^{\ell-m_0}\norm{\Omega}_{C^{m_0}}+1+\norm{R}_{\ell})\notag \\
&\leq C_{\ell,\eta}
(s^{\ell-m_0}t^{-(\ell-m_0-\sigma)}\norm{R}_{C^{\ell}}+1+\norm{R}_{\ell})
\end{align}
\subsection{The iteration scheme}
To set up the iterative process we first let:
\begin{align*}
R^{(0)}=R;\qquad \tilde{\alpha}^{(0)}=\tilde{\alpha};\qquad
\alpha^{(0)}=\alpha_{D_+};\qquad H^{(0)}=I
\end{align*}
Now construct $R^{(n)}$ inductively for every $n$: for $R^{(n)}$
choose an appropriate number $t_n$ to obtain $S_{t_n}R^{(n)}$ which
produces, after solving approximately the linearized equation, new
$\Omega^{(n)}$. Then we construct new abelian action
$\alpha^{(n+1)}$ as follows: At first
define
\begin{align*}
H^{(n)}&=\exp(\Omega^{(n)})\cdot I\\
\tilde{\alpha}^{(n+1)}&=(H^{(n)})^{-1}\circ\tilde{\alpha}^{(n)}\circ
H^{(n)}\\
\exp(R'_{(n+1)})&=\tilde{\alpha}^{(n+1)}\cdot\bigl(\exp(\mathcal{A}R^{(n)})\cdot\alpha^{(n)}\bigl)^{-1}
\end{align*}
Now let
\begin{align*}
\alpha_{(n+1)}&=\exp(\mathcal{A}R'_{(n+1)})\cdot\exp(\mathcal{A}R^{(n)})\cdot\alpha^{(n)}\\
\exp(R_{(n+1)})&=\tilde{\alpha}^{(n+1)}\cdot(\alpha_{(n+1)})^{-1}\\
\end{align*}
Notice $\alpha_{(n+1)}$ is not necessarily
an abelian action, so we need to do some
coordinate change: find a new $\exp(\mathcal{A}R^{(n)})'$ close
enough to $\exp(\mathcal{A}R'_{(n+1)})\cdot\exp(\mathcal{A}R^{(n)})$
such that
$\alpha^{(n+1)}=\exp(\mathcal{A}R^{(n)})'\cdot\alpha^{(n)}$ is
abelian and thus define
$R^{(n+1)}=\tilde{\alpha}^{(n+1)}\cdot\bigl(\alpha^{(n+1)}\bigl)^{-1}$
(see Section \ref{sec:8.5}). Consequently:
\begin{align*}
\tilde{\alpha}^{(n+1)}&=(H^{(n)})^{-1}\circ
(H^{(n-1)})^{-1}\circ\cdots\circ(H^{(0)})^{-1}\circ\tilde{\alpha}\circ
H^{(0)}\circ\cdots\circ H^{(n)}\\
&=H_n^{-1}\circ\tilde{\alpha}\circ H_n
\end{align*}
where $H_n=H^{(0)}\circ\cdots\circ H^{(n)}$.
Let $\delta_1=\frac{\delta_0}{2}+\frac{1}{8}$ where $\delta_0$ is
defined in Definition \ref{def:1}. To ensure $C^1$ convergence of the
process set:
\begin{align}\label{for:32}
\norm{R^{(n)}}_{C^0}&\leq \varepsilon_n=\varepsilon^{(k^n)}\notag\\
\norm{R^{(n)}}_{C^\ell}&\leq \varepsilon_n^{-1}\notag\\
s_n=t_n&=\varepsilon_n^{-\frac{1}{k_0\sigma+2}}
\end{align}
where $k=\frac{k_0+1}{k_0}$ and
$k_0>\max\{\frac{\frac{7\delta_0}{4}+\frac{1}{16}}
{\frac{\delta_0}{4}-\frac{1}{16}},4\delta_1,\frac{1}{1-\frac{1}{\delta_1+\delta_0}}\}$.
Let $\ell_0\in\mathbb{N}$ large enough
such that
\begin{align}\label{for:106}
&\ell_0>r(m_0+2r+\frac{\dim
M}{2}+6),\notag\\
&1>\gamma(m_0+\sigma,\ell_0)> \frac{1}{2}+\frac{1}{8\delta_1}
\end{align}
where $\tau=\max\{\frac{k_0+1}{k_0},\frac{1}{2\delta_1}\}$. Notice that $\tau>1$.
>From interpolation inequalities it follows that:
\begin{align}\label{intCl0}
&\norm{R^{(n)}}_{C^{\ell_0}}\leq
C_\ell\norm{R^{(n)}}_{C^0}^{1-\frac{\ell_0}{\ell}}\norm{R^{(n)}}_{C^\ell}^{\frac{\ell_0}{\ell}}\leq
C_\ell \varepsilon_n^{1-\frac{2\ell_0}{\ell}}<\varepsilon_n^{\frac{1}{6}}\notag\\
&\norm{R^{(n)}}_{C^{\ell_0+1}}\leq
C_\ell\norm{R^{(n)}}_{C^0}^{1-\frac{\ell_0+1}{\ell}}\norm{R^{(n)}}_{C^\ell}^{\frac{\ell_0+1}{\ell}}\leq
C_\ell
\varepsilon_n^{1-\frac{2\ell_0+2}{\ell}}<\varepsilon_n^{\frac{1}{6}}
\end{align}
then by \eqref{for:16}
\begin{align}\label{intconclusion}
&\norm{L\bigl(\overline{S_{t_n}R^{(n)}_{b}}-\mathcal{A}(\overline{S_{t_n}R^{(n)}_{b}}),\overline{S_{t_n}R^{(n)}}_{a}-
\mathcal{A}(\overline{S_{t_n}R^{(n)}_{a})}\bigl)^{(a^1_n,a^2_n)}}_{C^{\ell_0}}\notag\\
&\leq
C_{\ell,\eta}\norm{R^{(n)}}_{C^{\ell_0}}\norm{R^{(n)}}_{C^{\ell_0+1}}+C_{\ell,\eta}t_n^{-2\ell_0}\norm{R^{(n)}}_{C^{\ell}}\notag\\
&\leq
C_{\ell,\eta}\varepsilon_n^{\frac{1}{3}}+C_{\ell,\eta}t_n^{-2\ell_0}\norm{R^{(n)}}_{C^{\ell}}<\varepsilon_n^{\frac{1}{6}}<1.
\end{align}
At this point fix $\ell$:
\begin{align}\label{for:115}
\ell\geq \max\{3\ell_0,\frac{2m_0+2\sigma}{1-\frac{1}{k_0}},
\frac{2m_0+2\sigma+2}{2-\frac{\tau}{\frac{1}{2}+\frac{1}{8\delta_1}}},\frac{2\tau}{\tau-\frac{1}{\delta_1+\delta_0}},
(\tau+2+m_0)(k_0\sigma+2)\}
\end{align}
This seemingly cumbersome condition will be needed to estimate $\norm{R^{(n)}}_{C^0}$.
\subsection{Convergence}\label{sec:8.4}
By induction it is proved that all the bounds \eqref{for:32} hold
for every $n\in\mathbb{N}$. By \eqref{Clerror} and by the inductive assumption we obtain
\begin{align*}
\norm{R_{(n+1)}}_{C^\ell}&\leq C_{\ell,\eta}
(s^{\ell-m_0}t^{-(\ell-m_0-\sigma)}\norm{R}_{C^{\ell}}+1+\norm{R}_{\ell})\\
&\leq C_{\ell,\eta}
(\varepsilon_n^{-\frac{\ell-m_0}{k_0\sigma+2}}\varepsilon_n^{\frac{\ell-m_0-\sigma}{k_0\sigma+2}}\varepsilon_n^{-1}+1+\varepsilon_n^{-1})\\
&\leq C_{\ell,\eta}
(\varepsilon_n^{-\frac{\sigma}{k_0\sigma+2}-1}+1+\varepsilon_n^{-1})\leq
2C_{\ell,\eta}
\varepsilon_n^{-\frac{\sigma}{k_0\sigma+2}-1}\\
&<\varepsilon_n^{-\frac{1}{k_0+1}}\varepsilon_n^{-1}
=\varepsilon_n^{-\frac{k_0+2}{k_0+1}}\\
&<\varepsilon_{n+1}^{-1}.
\end{align*}
Now by \eqref{intCl0} and \eqref{intconclusion} Main Estimate
applies to get \eqref{for:34}.
From interpolation
inequalities it follows that:
\begin{align}\label{for:114}
\norm{R^{(n)}}_{C^{m_0+\sigma}}&\leq
C_\ell\norm{R^{(n)}}_{C^0}^{1-\frac{m_0+\sigma}{\ell}}\norm{R^{(n)}}_{C^\ell}^{\frac{m_0+\sigma}{\ell}}\notag\\
\norm{R^{(n)}}_{C^{m_0+\sigma+1}}&\leq
C_\ell\norm{R^{(n)}}_{C^0}^{1-\frac{m_0+\sigma+1}{\ell}}\norm{R^{(n)}}_{C^\ell}^{\frac{m_0+\sigma+1}{\ell}}\notag\\
\norm{R^{(n)}}_{C^{1}}&\leq
C_\ell\norm{R^{(n)}}_{C^0}^{1-\frac{1}{\ell}}\norm{R^{(n)}}_{C^\ell}^{\frac{1}{\ell}}
\end{align}
Along with \eqref{for:67} and this implies for
$\gamma=\gamma(m_0+\sigma,\ell_0)$:
\begin{align*}
\norm{R_{(n+1)}}_{C^0}&\leq
C_{m_0,\eta,s}\norm{R}^\gamma_{C^{m_0+\sigma}}\norm{R}^\gamma_{C^{m_0+\sigma+1}}
+C_{\ell}t^{-\ell}\norm{R}_{C^{\ell}}\notag\\
&+C_{m_0,\eta,\ell_0,\ell}t^{-\gamma
(\ell-m_0-\sigma)}\norm{R}^\gamma_{C^{\ell}}+C_{m_0,\eta,\ell}s^{-m_0}t^{-(\ell-m_0-\sigma)}\norm{R}_{C^{\ell}}\notag\\
&+C_{m_0,\eta}\norm{R}^2_{C^{m_0+\sigma}}\notag\\
&\leq
C_{m_0,\eta,\ell_0}\varepsilon_n^{\gamma(2-\frac{2m_0+2\sigma+2}{\ell})}
+C_{\ell}\varepsilon_n^{-1}\varepsilon_n^{\frac{\ell}{k_0\sigma+2}}\notag\\
&+C_{m_0,\eta,\ell_0,\ell}\varepsilon_n^{-\gamma}\varepsilon_n^{\gamma(\ell-\frac{m_0-\sigma}{k_0\sigma+2})}
+C_{m_0,\eta,\ell}\varepsilon_n^{-1}\varepsilon_n^{\frac{m_0}{k_0\sigma+2}}\varepsilon_n^{\frac{\ell-m_0-\sigma}{k_0\sigma+2}}\\
&+C_{m_0,\eta}\varepsilon_n^{2-\frac{2m_0+2\sigma}{\ell}}\\
&=C(\varepsilon_n^x+\varepsilon_n^y+\varepsilon_n^z+\varepsilon_n^u+\varepsilon_n^v)\leq
\varepsilon_n^{\tau}
\end{align*}
where $\tau=\max\{\frac{k_0+1}{k_0},\frac{1}{2\delta_1}\}$ providing
\begin{align*}
\gamma(2-\frac{2m_0+2\sigma+2}{\ell})&>\tau,\qquad
\frac{\ell}{k_0\sigma+2}-1
>\tau,\\
\gamma(\ell-\frac{m_0-\sigma}{k_0\sigma+2})-\gamma &>\tau,\qquad
\frac{\ell-m_0-\sigma}{k_0\sigma+2}+\frac{m_0}{k_0\sigma+2}-1
>\tau\\
2-\frac{2m_0+2\sigma}{\ell}&>\tau
\end{align*}
all inequalities above are satisfied for
\begin{align*}
& \ell>\max\{(\tau+2+m_0)(k_0\sigma+2),
\frac{2m_0+2\sigma+2}{2-\frac{\tau}{\frac{1}{2}+\frac{1}{8\delta_1}}}\},\quad
k_0>4\delta_1.
\end{align*}
Using interpolation inequalities we can get the $C^1$ bound for
$R_{(n+1)}$:
\begin{align*}
\norm{R_{(n+1)}}_{C^1}&\leq
C_\ell\norm{R_{(n+1)}}_{C^0}^{1-\frac{1}{\ell}}\norm{R_{(n+1)}}_{C^\ell}^{\frac{1}{\ell}}\leq
C_\ell\varepsilon_n^{\tau(1-\frac{1}{\ell})-\frac{\tau}{\ell}}\leq
\varepsilon_n^{\frac{1}{\delta_1+\delta_0}}
\end{align*}
providing
$\tau(1-\frac{1}{\ell})-\frac{\tau}{\ell}>\frac{1}{\delta_1+\delta_0}$
which is satisfied for
$\ell>\frac{2\tau}{\tau-\frac{1}{\delta_1+\delta_0}}$ and
$k_0>\frac{1}{1-\frac{1}{\delta_1+\delta_0}}$
\subsection{Construction of coordinate changes}\label{sec:8.5} By induction we just formed:
\begin{align}
&\tilde{a}^{(n+1)}=\exp(R^a_{(n+1)})\cdot \alpha(a_n)\notag\\
&\tilde{b}^{(n+1)}=\exp(R^b_{(n+1)})\cdot\alpha(b_n)
\end{align}
where $\alpha(a_n)=\alpha_{(n+1)}(a)$ and
$\alpha(b_n)=\alpha_{(n+1)}(a)$ with $a_n\,,b_n\in Z(L)$ and
\begin{align}\label{for:108}
&\norm{R^a_{(n+1)},R^b_{(n+1)}}_{C^0}\leq
\varepsilon_n^{\tau},\qquad \norm{R^a_{(n+1)},R^b_{(n+1)}}_{C^1}\leq
\varepsilon_n^{\frac{1}{\delta_0+\delta_1}}\notag\\
&\norm{R^a_{(n+1)},R^b_{(n+1)}}_{C^\ell}\leq
\varepsilon_n^{-\frac{k_0+2}{k_0+1}}.
\end{align}
Using \eqref{for:92} we have
\begin{align}\label{for:93}
&\norm{\mathcal{A}R^a_{(n+1)},\mathcal{A}R^b_{(n+1)}}\leq
\varepsilon_n^{2\tau}
\end{align}
Since $\tilde{a}^{(n+1)}$ and $\tilde{b}^{(n+1)}$
commute, we can argue similarly to the proof of Lemma \ref{le:5}. Let
$\exp(X)=\log(a_nb_n(a_nb_n)^{-1})$, then
\begin{align}\label{for:110}
X&=-R^a_{(n+1)}(\exp(R^b_{(n+1)})\cdot
b_n)-\text{Ad}_{a_n}R^b_{(n+1)}\notag\\
&+R^b_{(n+1)}(\exp(R^a_{(n+1)})\cdot
a_n)+\text{Ad}_{b_n}R^a_{(n+1)}+\text{Res}\notag\\
&=R^b_{(n+1)}\circ a_n-\text{Ad}_{a_n}R^b_{(n+1)}-R^a_{(n+1)}\circ b_n+\text{Ad}_{b_n}R^a_{(n+1)}\notag\\
&+R^b_{(n+1)}(\exp(R^a_{(n+1)})\cdot a_n)-R^b_{(n+1)}\circ
a_n\notag\\
& +R^a_{(n+1)}\circ b_n-R^a_{(n+1)}(\exp(R^b_{(n+1)})\cdot
b_n)+\text{Res}
\end{align}
where
\begin{align*}
\norm{\text{Res}}_{C^0}\leq
C\norm{R^a_{(n+1)},R^b_{(n+1)}}^2_{C^0}\leq
C\varepsilon_n^{\frac{1}{\delta_1}}<\varepsilon_n^{\frac{2}{\delta_0+\delta_1}}.
\end{align*}
By \eqref{for:108} and along proof line in Lemma \ref{le:5}, we have
\begin{align}\label{for:109}
&\norm{R^b_{(n+1)}(\exp(R^a_{(n+1)})\cdot a_n)-R^b_{(n+1)}\circ
a_n}_{C^0}\notag\\
& \leq C\norm{R^b_{(n+1)}}_{C^1}\norm{R^a_{(n+1)}}_{C^0}\leq
C\varepsilon_n^{\frac{1}{2\delta_1}+\frac{1}{\delta_0+\delta_1}}<\varepsilon_n^{\frac{2}{\delta_0+\delta_1}},
\end{align}
similarly we also have
\begin{align}\label{for:111}
\norm{R^a_{(n+1)}(\exp(R^b_{(n+1)})\cdot b_n)-R^a_{(n+1)}\circ
b_n}_{C^0}<\varepsilon_n^{\frac{2}{\delta_0+\delta_1}}.
\end{align}
And thus if we integrate each side of \eqref{for:110}, since the
left side is constant and norm of integral of
\begin{align*}
R^b_{(n+1)}\circ a_n-\text{Ad}_{a_n}R^b_{(n+1)}-R^a_{(n+1)}\circ
b_n+\text{Ad}_{b_n}R^a_{(n+1)}
\end{align*}
is bounded by $\varepsilon_n^{\frac{1}{\delta_1}}$ by
\eqref{for:93}, combine \eqref{for:109} and \eqref{for:111} it
follows that
\begin{align*}
\norm{X}\leq\varepsilon_n^{\frac{2}{\delta_0+\delta_1}}+\varepsilon_n^{\frac{2}{\delta_0+\delta_1}}
+\varepsilon_n^{\frac{2}{\delta_0+\delta_1}}+\varepsilon_n^{\frac{1}{\delta_1}}<4\varepsilon_n^{\frac{2}{\delta_0+\delta_1}}
\end{align*}
hence we have
\begin{align}\label{for:36}
\norm{[\log(a_n),\log(b_n)]}\leq C\norm{X}\leq
4C\varepsilon_n^{\frac{2}{\delta_0+\delta_1}}<
\varepsilon_n^{\frac{2}{\frac{3\delta_0}{2}+\frac{\delta_1}{2}}}.
\end{align}
By assumption of $\delta_0$ and using \eqref{for:36}
there exist $\mathfrak{n}'_1\,,\mathfrak{n}'_2\in \mathfrak{Z}$
where $\mathfrak{Z}$ is the Lie algebra of $Z(D)\bigcap Z(L)$ such
that $[\mathfrak{n}'_1,\mathfrak{n}'_2]=0$ and
\begin{align}\label{for:113}
\norm{\mathfrak{n}'_1-\log(a_n),\mathfrak{n}'_2-\log(b_n)}\leq\norm{CX}^{\delta_0-\epsilon}\leq
\varepsilon_n^{\frac{2(\delta_0-\epsilon)}{\frac{3\delta_0}{2}+\frac{\delta_1}{2}}}.
\end{align}
In order for the process to converge we need the power in the
right-hand part of \eqref{for:113} to be greater than
$\frac{k_0+1}{k_0}$ for some $\epsilon>0$. Since
$\delta_1=\frac{\delta_0}{2}+\frac{1}{8}$ this is equivalent to
$$\frac{2\delta_0}{\frac{3\delta_0}{2}+\frac{\delta_1}{2}}=\frac{2\delta_0}{\frac{7\delta_0}{4} +\frac{1}{16}}> \frac{k_0+1}{k_0}$$ or
\begin{align}\label{delta14}\delta_0>\frac{\frac{1}{16}}{\frac{2k_0}{k_0+1}-\frac{7}{4}}\end{align}
which is satisfied if $k_0>\frac{\frac{7\delta_0}{4}+\frac{1}{16}}
{\frac{\delta_0}{4}-\frac{1}{16}}$. Notice that the minimum of the right-hand
side of above inequality is $\frac{1}{4}$. That is the key
assumption needed to carry out the proof.
Let $\tilde{a}_n=\exp(\mathfrak{n}'_1)$ and
$\tilde{b}_n=\exp(\mathfrak{n}'_2)$ and also let
\begin{align}
&R_a^{(n+1)}=\log\bigl(\exp(R^a_{(n+1)})a_n(\tilde{a}_n)^{-1}\bigl),\notag\\
&R_b^{(n+1)}=\log\bigl(\exp(R^b_{(n+1)})b_n(\tilde{b}_n)^{-1}\bigl)
\end{align}
combine \eqref{for:108} and \eqref{for:113} it follows
\begin{align*}
\norm{R_a^{(n+1)},R_b^{(n+1)}}_{C^0}&\leq
C\norm{R^a_{(n+1)},\mathfrak{n}_1'-\log(a_n),\mathfrak{n}'_2-\log(b_n)}_{C^0}\leq
\varepsilon_n^{\frac{k_0+1}{k_0}}\\
\norm{R_a^{(n+1)},R_b^{(n+1)}}_{C^\ell}&\leq
C\norm{R^a_{(n+1)},R^b_{(n+1)}}_{C^\ell}\leq
C\varepsilon_n^{-\frac{k_0+2}{k_0+1}}\leq \varepsilon_n^{-1}.
\end{align*}
Then we get
\begin{align}
&\tilde{a}^{(n+1)}=\exp\bigl(R_a^{(n+1)}\bigl)\cdot\alpha^{(n+1)}(a)\notag\\
&\tilde{b}^{(n+1)}=\exp\bigl(R_a^{(n+1)})\bigl)\cdot\alpha^{(n+1)}(b)
\end{align}
where $\alpha^{(n+1)}(a)=\alpha(\tilde{a}_n)$ and
$\alpha^{(n+1)}(b)=\alpha(\tilde{b}_n)$.
Using \eqref{for:19} and \eqref{for:114} we may check the $C^1$
bound for $\Omega$:
\begin{align*}
\norm{\Omega^{(n+1)}}_{C^1}&\leq C
\norm{R_a^{(n)},R_b^{(n)}}_{C^{a_0m_0+\sigma}}\leq
C\varepsilon_{n}^{1-\frac{2a_0m_0+2\sigma}{\ell}}<C\varepsilon_{n}^{\frac{1}{k_0}}
<\varepsilon_{n}^{\frac{1}{2k_0}}
\end{align*}
providing $1-\frac{2a_0m_0+2\sigma}{\ell}>\frac{1}{k_0}$ which is
satisfied for $\ell>\frac{2a_0m_0+2\sigma}{1-\frac{1}{k_0}}$.
Thus for sufficiently small $\norm{R}_{C^0}$ and $\norm{R}_{C^\ell}$
the process converges to a solution $\Omega\in C^1$ with
$\norm{\Omega}_{C^1}<\frac{1}{4}$. Hence $\exp{\Omega}$ conjugates $\tilde\alpha$ with a standard perturbation of $\alpha_{D_+}$.
Since the $C^1$ conjugacy thus constructed is also an orbit conjugacy between the neutral foliations, it is $C^\infty$ by \cite[Theorem
1]{Spatzier} it is $C^\infty$. This completes the proof of our theorems.
\begin{remark} The last argument is a shortcut that is available due to condition $(\mathfrak B)$. However the general method of the present paper is applicable to certain cases where this condition either does not hold as in the setting of \cite{Damjanovic4} or where there is no hyperbolicity to begin with as in \cite{DK-parabolic, DFK-parabolic}. A direct application of our scheme with orthogonal projection in a fixed Sobolev space and iteration procedure described above
would only produce conjugacy of finite regularity and a priory estimates are not available in those situations.
However, the iterative scheme may be modified by increasing $\ell$ along the way. The ideas is to achieve closeness of conjugacy constructed on an iteration step to identity in a high norm that will allow to raise the threshold
$m_0$ in the Main Estimate.
\end{remark}
\section{Rigid genuinely partially hyperbolic twisted examples}
\subsection{Preliminaries on arithmetic groups} To prove the statement of
Theorem \ref{th:7} we first recall some algebraic notations and
theorems.
Let $H$ be an algebraic group defined over $\mathbb{Q}$, and we can specify
an embedding $H\hookrightarrow GL(n,\mathbb{C})$. Then we can define
$H(\mathbb{Z}) = H(\mathbb{C})\bigcap GL(n,\mathbb{Z})$ and $H(\mathbb{Q}) = H(\mathbb{C})\bigcap
GL(n,\mathbb{Q})$. Obviously $H(\mathbb{Z})$ and $H(\mathbb{Q})$ depend on the embedding
$H\hookrightarrow GL(n,\mathbb{C})$.
\begin{definition}A subgroup $S$ of $H(\mathbb{Q})$ is called an {\em arithmetic subgroup} if it
is commensurable with $H(\mathbb{Z})$, i.e., the intersection $S\bigcap
H(\mathbb{Z})$ has finite index in both $S$ and $H(\mathbb{Z})$. In particular,
$H(\mathbb{Z})$ and its subgroups of finite index are arithmetic groups.
\end{definition}
Now we give the definition of arithmetic subgroups of Lie groups.
Since a Lie group may not be equal to the real locus of an algebraic
group, we need a more general definition of arithmetic subgroups.
\begin{definition}
Let $L$ be a Lie group, $\Lambda\subseteq L$ be a discrete subgroup.
$\Lambda$ is called an arithmetic subgroup of $L$ if there exists an
algebraic group $H$ defined over $\mathbb{Q}$ and a Lie group homomorphism
$\varphi : L\rightarrow H(\mathbb{R})$ whose kernel is compact such that
the image $\varphi(\Lambda)$ is an arithmetic subgroup of $H(\mathbb{Q})$.
\end{definition}
One has the following fundamental theorem due to Margulis:
\begin{theorem}\emph{(Margulis' arithmeticity theorem
\cite{Margulis})} Suppose $L$ is a connected semisimple Lie group
without compact factors. If the rank of $L$ is at least $2$, then
every irreducible lattice in $L$ is arithmetic.
\end{theorem}
\subsection{Proof of Theorem \ref{th:7}}
By Margulis arithmeticity theorem $\Gamma$ is
arithmetic. Hence there exists $\varphi: G\rightarrow H(\mathbb{R})$ whose
kernel is compact such that the image $\varphi(\Gamma)$ is an
arithmetic subgroup of $H(\mathbb{Q})$. Since $G$ has no compact factor,
the kernel of $\varphi$ is finite. Consider the adjoint
representation of $\varphi(\Gamma)$ on the Lie algebra of
$\varphi(G)$ which is isomorphic to $\mathfrak{G}$. Since
$\text{Ad}_{\varphi(\Gamma)}\in SL(m,\mathbb{Q})$ ($m=\dim \mathfrak{G}$)
with bounded denominators, there exists a base of $\mathfrak{G}$
such that $\text{Ad}_{\varphi(\Gamma)}\in SL(m,\mathbb{Z})$. Also notice the adjoint
representation admits no invariant subspace of eigenvalue $1$ hence
$\text{Ad}\circ \varphi$ is an genuinely partially hyperbolic
representation of $G$ on $\mathbb{R}^m$.
If $G$ is quasi-split then obviously $\mathfrak{N}$ is abelian hence
$\delta_0=\infty$. If $\mathfrak{K}=\mathfrak{K}_0+\mathfrak{so}(3)$
where the ideal $\mathfrak{K}_0$ is abelian, next we will show
$\delta_0\geq\frac{1}{3}$ in $\mathfrak{N}$. In fact we just need to
show for any $\mathfrak{n}_1\,,\mathfrak{n_2}\,,t_1\,,t_2\in
\mathfrak{so}(3)$, if
$\norm{[(\mathfrak{n}_1,t_1),(\mathfrak{n}_2,t_2)]}<\gamma$ then
there exist $\mathfrak{n}_1'\,,\mathfrak{n}_2'\,,t_1'\,,t_2'\in
\mathfrak{so}(3)$ such that
$\norm{[(\mathfrak{n}'_1,t'_1),(\mathfrak{n}'_2,t'_2)]}=0$ with
$d(\mathfrak{n}_i',\mathfrak{n}_i)<C\gamma^{\frac{1}{3}}$ and
$d(t_i',t_i)<C\gamma^{\frac{1}{3}}$ for $i=1,2$.
There exist bases of $\mathfrak{so}(3)$ such that
$\text{ad}_{\mathfrak{n}_1}$ and $\text{ad}_{\mathfrak{n}_2}$
have the forms $\text{ad}_{\mathfrak{n}_i}=\begin{pmatrix}0 & \theta_i & 0 \\
-\theta_i & 0 & 0\\
0 & 0& 0
\end{pmatrix}, i=1,2
$ respectively. We can assume either $\abs{\theta}_1>\gamma^{\frac{1}{3}}$ or
$\abs{\theta}_2>\gamma^{\frac{1}{3}}$. Otherwise there exists $C>0$
such that $\norm{\mathfrak{n}_i}<C\gamma^{\frac{1}{3}}$, $i=1,2$.
Then let $\mathfrak{n}_i'=0$ and $t_i'=t_i$, $i=1,2$.
Assume that $\abs{\theta}_1>\gamma^{\frac{1}{3}}$. Denote $\text{ad}_{t_1}=\begin{pmatrix}0 & q & b \\
-q & 0 & c\\
-b & -c & 0
\end{pmatrix}
$, $\mathfrak{n}_2=\begin{pmatrix}x_1 \\
y_1 \\
z_1
\end{pmatrix}$ and $t_2=\begin{pmatrix}x_2 \\
y_2 \\
z_2
\end{pmatrix}$ under the basis. Then by assumption we have
\begin{align*}
&\abs{\theta_1x_1}<\gamma, \quad\abs{\theta_1y_1}<\gamma,\quad
\abs{\theta_1y_2-qy_1-bz_1}<\gamma\\
&\abs{-\theta_1x_2+qx_1-cz_1}<\gamma,\quad \abs{bx_1+cy_1}<\gamma.
\end{align*}
Since $\abs{\theta}_1>\gamma^{\frac{1}{3}}$ then
$\abs{x}_1<\gamma^{\frac{2}{3}}$ and
$\abs{y}_1<\gamma^{\frac{2}{3}}$. Then it follows:
\begin{align*}
\abs{d_1=\theta_1y_2-bz_1}<C\gamma^{\frac{2}{3}}\text{ and
}\abs{d_2=\theta_1x_2+cz_1}<C\gamma^{\frac{2}{3}}.
\end{align*}
Then let $\mathfrak{n}_1'=\mathfrak{n}_1$,
$\mathfrak{n}'_2=(0,0,z_1)$, $t_1'=t_1$ and
$t_2'=(x_2-d_2\theta_1^{-1},y_2-d_1\theta_1^{-1},z_2)$. Notice
$\norm{d_2\theta_1^{-1},d_2\theta_1^{-1}}\leq C\gamma^{\frac{1}{3}}$
then it is easy to check they satisfy the requirement.
Hence $\delta_0> \frac{1}{4}$ works for all the following examples:
$G=SO(m+3,m)$, $SU(m+2,m)$ or $Sp(m+1,m)$ or it's product with any
quasi-split semisimple Lie groups.
|
\section{Introduction}
\label{intro}
Quantum walks, developed as the quantum analog of the classical random walks\,\cite{Ria58, Fey86, Par88, ADZ93} first emerged as a powerful tool in the development of quantum algorithms\,\cite{Amb03, CCD03, SKW03, AJR05}. Subsequently, its rich dynamics is constituting as a framework to understand and simulate the dynamics in various systems. For example, they have been used to explain phenomena such as the breakdown of an electric-field driven system\,\cite{OKA05} and mechanism of wavelike energy transfer within photosynthetic systems\,\cite{ECR07, MRL08}, to demonstrate the coherent quantum control over atoms\,\cite{CL08} and localization of Bose-Einstein condensates in optical lattice\,\cite{Cha11a} and to explore topological phases\,\cite{KRB10}. The quantum walk evolution are widely studied in two forms: continuous-time quantum walk (CTQW)\,\cite{FG98} and discrete-time quantum walk (DTQW)\,\cite{ADZ93, DM96, ABN01, NV01, BCG04, Ken07, CSL08}. In this letter we focus on the DTQW evolution which is defined on the {\it position} Hilbert space, ${\cal H}_p$ and the {\it coin} (particle) Hilbert space, ${\cal H}_c$.
During last few years, experimental implementation of the DTQW has been demonstrated with energy levels in NMR\,\cite{RLB05},
ions\,\cite{SMS09, ZKG10}, photons\,\cite{PLP08, PLM10, SCP10, BFL10}, and atoms\,\cite{KFC09}. These experimental implementations using one-dimensional (1D) DTQW model on a two-level system using a degree two coin operation have opened up a new dimension to simulate quantum dynamics in physical systems like the recent demonstration of localization of photon's wavepacket\,\cite{SCP11}. Now the immediate interest would be to extend the implementation to two-dimensional (2D) and three-dimensional (3D) lattice structure with the available resources. This will give way to simulate and explore the possibility of mimicking the dynamics in various naturally occurring physical systems. One of the extension to the 2D is the Grover walk which is defined on a four-level particle with a specific initial state\,\cite{MBS02, TFM03}. An alternative extension to higher ($d$) dimensions is to use a $d$ coupled qubits to describe the internal states\,\cite{ EMB05, OPD06}. This is extremely challenging with the available resources to implement it experimentally. To overcome this challenge, an alternative 2D DTQW using a two-state particle was very recently proposed. By evolving the particle in superposition of position space in one dimension followed by the evolution in the other direction using Hadamard coin operation was show to be equivalent to four-state Grover walk\,\cite{FGB11}.
\par
In this letter we present a new scheme using different Pauli basis as translational eigenstate in different axis and show that the two-state particle walk can be implemented on a physically relevant 2D, square, triangular, kagome, and 3D, cubic lattice structures. Using basis vectors of the three Pauli matrices is very common in quantum optics experiments and various other physical systems making this scheme implementable in the present experimental setups. We also present Hamiltonian form of the evolution which can serve as a general framework to simulate, control, and study the dynamics in different 2D and 3D physical systems. We then show that the coin operation which is required for 1D DTQW is not a necessary requirement for 2D and 3D DTQW but can be used as an additional degree of freedom to control the dynamics. In particular, we demonstrate that the probability distribution obtained using Grover walk and alternative two-state walk on a square lattice \cite{FGB11} is effectively reproduced without a coin operation in our scheme. This further reduces the resource required for the experimental implementation.
\section{One-dimensional DTQW and its Hamiltonian form}
\label{DTQW}
The standard form of DTQW evolution on a two-state particle in 1D lattice is defined on a coin (c) and the position (p) Hilbert space ${\cal H} = {\cal H}_c\otimes {\cal H}_p$. The basis states of ${\cal H}_c$ are the internal states of the particle, $|\downarrow \rangle = \begin{bmatrix} 1 \\ 0 \end{bmatrix}$ and $|\uparrow \rangle =\begin{bmatrix} 0 \\ 1 \end{bmatrix}$ and they are also the eigenstates of the Pauli matrix $\sigma_3= \begin{bmatrix} 1 & ~~0\\ 0 & -1 \end{bmatrix}$. The basis states of ${\cal H}_p$ is described in terms of $|\psi_z\rangle$, where $z \in {\mathbbm I}$, the set of integers associated with the lattice sites. Each step evolution of 1D DTQW is described using a quantum coin operation $B_{\sigma_3} (\theta) \equiv \begin{bmatrix}\cos(\theta) & \sin(\theta) \\
-\sin(\theta) & \cos(\theta)
\end{bmatrix}$\footnote{Though $B_{\sigma_3}(\xi, \theta, \zeta)\in {\rm SU(2)}$ describes the general form of evolution of each step of the DTQW, its main essence can be completely captured by assigning $\xi = \zeta = 0$.}
which evolves the particle (coin) into the superposition of the basis states followed by the unitary shift operator
$S_{\sigma_3}\equiv \sum_z \left [ |\downarrow \rangle\langle
\downarrow|\otimes|\psi_{z-1}\rangle\langle \psi_z| + | \uparrow \rangle\langle
\uparrow|\otimes |\psi_{z+1}\rangle\langle \psi_z| \right ]$,
which shift the state of the particle in superposition of the position space. Therefore, the effective operation for each step of the DTQW
is written in the form:
\begin{equation}
\label{Wop}
W_{\sigma_3} (\theta)\equiv S_{\sigma_3}[B_{\sigma_3}(\theta) \otimes {\mathbbm 1}].
\end{equation}
The state after the $t$ step evolution of the DTQW is given by,
$|\Psi_t\rangle=[W_{\sigma_3}(\theta)]^t|\Psi_{\rm in}\rangle$,
where $|\Psi_{\rm in}\rangle= \left ( \cos(\delta/2)| \downarrow \rangle + e^{i\eta}
\sin(\delta/2)|\uparrow \rangle \right )\otimes |\psi_{0}\rangle$,
is the initial state of the particle at position $z=0$. The coin parameter $\theta$ controls the variance of the probability distribution of the walk and $\theta \neq 0, \pi/2$ is required to spread the amplitude in superposition of position space.
\par
The net evolution of each step of the DTQW in the form of Eq.\,(\ref{Wop}), can also be generated by the time independent effective Hamiltonian $H_{\sigma_3}(\theta)$. That is,
$W_{\sigma_3}(\theta) \equiv e^{-iH_{\sigma_3}(\theta)\tau}$ where $\hbar = 1$ and $\tau$ is the time required to implement one step of the walk.
To obtain an expression for $H(\theta)$ we will expand Eq.\,(\ref{Wop}) and rewrite to the form:
\begin{eqnarray}
\label{eq:combop}
W_{\sigma_3}(\theta)
= \begin{bmatrix}
\mbox{~~}\cos(\theta)e^{-i\hat{P}_Zl\tau} & & \sin(\theta)e^{-i\hat{P}_Zl\tau}
\\ - \sin(\theta)e^{+i\hat{P}_Zl\tau} & & \cos(\theta)e^{+i\hat{P}_Zl\tau}
\end{bmatrix}
\end{eqnarray}
where $\hat{P}_Z$ is the momentum operator whose action on all position space in the $Z$ axis is local such that $e^{\pm i \hat{P}_Zl \tau} |\psi_z \rangle = |\psi_{z \pm l} \rangle$ \cite{ADZ93, Kem03}, $\tau$ is the time required to implement each step of walk length $l$ (lattice separation). Hereafter, we will consider the time required to implement the unit length of each step of walk to be one ($\tau = l = 1$). To obtain the Hamiltonian form, $W_{\sigma_3}(\theta) = e^{-i H_{\sigma_3}}(\theta)$ is written as
\begin{eqnarray}
\label{Ham}
-iH_{\sigma_3}(\theta) &=& \ln \begin{bmatrix}
\mbox{~~}\cos(\theta) e^{-i\hat{P}_Z} & & \sin(\theta)e^{-i\hat{P}_Z}
\\ - \sin(\theta)e^{+i\hat{P}_Z} & & \cos(\theta)e^{+i\hat{P}_Z}
\end{bmatrix} \nonumber \\
&=& \ln \,(A ) = V \,\ln(\,\lambda_{\sigma_3}\,)\, V^{-1}.
\end{eqnarray}
\begin{equation}
\lambda_{\sigma_3} =\begin{bmatrix}
\lambda^-_Z & &
0 \\
0 & & \lambda^+_Z \end{bmatrix}
\end{equation}
is the diagonal matrix composed of the eigenvalues
\begin{equation}
\label{l1}
\lambda^{\mp}_Z = \cos(\theta) \cos(\hat{P}_Z) \mp \sqrt {\cos^2(\theta)\cos^2(\hat{P}_Z) - 1}
\end{equation}
of matrix $A$ where $\lambda^{+}_Z \lambda^{-}_Z = \lambda^{-}_Z \lambda^{+}_Z = 1$.
\begin{eqnarray}
V =\begin{bmatrix}
\frac{\cos(\theta)e^{i \hat{P}_Z} - \lambda_{-}}{\sin(\theta) e^{i \hat{P}_Z}} & & \frac{\cos(\theta)e^{i \hat{P}_Z} - \lambda_{+}}{\sin(\theta) e^{i \hat{P}_Z}} \\
1 & & 1 \end{bmatrix}
\end{eqnarray}
and
\begin{eqnarray}
V^{-1} =\begin{bmatrix}
\frac{\sin(\theta) e^{i\hat{P}_Z}}{2 \sqrt{\cos^2(\theta)\cos^2(\hat{P}_Z) -1}} & &
\frac{\lambda_{+} - \cos(\theta) e^{i\hat{P}_Z}} { 2 \sqrt{\cos^2(\theta)\cos^2(\hat{P}_Z) -1}} \\
\frac{-\sin(\theta)e^{+i\hat{P}_Z}}{2 \sqrt{\cos^2(\theta)\cos^2(\hat{P}_Z) -1}} & &
\frac{-\lambda_{-} + \cos(\theta) e^{i\hat{P}_Z}} { 2 \sqrt{\cos^2(\theta)\cos^2(\hat{P}_Z) -1} }\end{bmatrix}
\end{eqnarray}
are the matrix composed of eigenvectors of $A$ and its inverse. By substituting these elements into Eq.\,(\ref{Ham}),
\begin{widetext}
\begin{eqnarray}
\label{hamil2}
H_{\sigma_{3}}(\theta)& = &
\frac{ \ln \left (\frac{\lambda^{+}_Z}{\lambda^{-}_Z} \right )}{2 \sqrt{\cos^2(\theta)\cos^2(\hat{P}_Z) -1}}
\begin{bmatrix}
\cos(\theta) \sin(\hat{P}_Z) &
-\sin(\theta)[\sin(\hat{P}_Z) + i \cos(\hat{P}_Z)]\\
\sin(\theta)[\sin(\hat{P}_Z) - i \cos(\hat{P}_Z)]
&
\cos(\theta) \sin(\hat{P}_Z)
\end{bmatrix}
\cdot \sigma_3
\end{eqnarray}
\end{widetext}
where $\sin(\hat{P}_Z) |\psi_{z} \rangle = \frac{i}{2}(|\psi_{z-1} \rangle - |\psi_{z+1} \rangle)$ and $\cos(\hat{P}_Z) |\psi_{z} \rangle = \frac{1}{2}(|\psi_{z-1} \rangle + |\psi_{z+1} \rangle)$.
\section{Two-state particle walk on different lattices}
\label{2sqw}
\subsection{Square Lattice and Cubic Lattice}
\label{square}
A simple 2D lattice structure is a square lattice with four direction of propagation and two quantization axis, $X$ and $Z$ [Fig.\,\ref{fig:1a}]. Similarly, a simple 3D lattice is a cubic lattice with six direction of propagation and three quantization axis, $X$, $Y$, and $Z$ [Fig.\,\ref{fig:1b}].
\begin{figure}[ht]
\begin{center}
\subfigure[]{\includegraphics[width=4.0cm]{square.png}
\label{fig:1a}}
\subfigure[]{\includegraphics[width=4.0cm]{cubea.png}
\label{fig:1b}}
\caption{\footnotesize{(a) Square lattice with two direction of propagation in each $X$ and $Y$ axis. (b) Cubic lattice with two direction of propagation in each $X$, $Y$ and $Z$ axis. Evolution in $X$, $Y$ and $Z$ axis are quantized by the basis states of the Pauli operators $\sigma_{1}$, $\sigma_{2}$ and $\sigma_{3}$, respectively.}}
\end{center}
\end{figure}
\par
Two-state particle DTQW on a square and cubic lattice can be realized by quantizing the evolution using different Pauli basis states as translational eigenstate for each axis in the lattice structure.
For a walk in 1D ($Z$ axis) we used basis states, $|\downarrow\rangle \equiv |+\rangle_{\sigma_{3}}$ and $|\uparrow\rangle \equiv |-\rangle_{\sigma_{3}}$ of Pauli operator $\sigma_{3}$ as translational state. Similarly, for $X$ and $Y$ axis we will use the basis states,
\begin{eqnarray}
|+\rangle_{\sigma_{1}} = \frac{1}{\sqrt 2} \begin{bmatrix} 1 \\ 1 \end{bmatrix} ~;~ |-\rangle_{\sigma_{1}}= \frac{1}{\sqrt 2} \begin{bmatrix} 1 \\ -1 \end{bmatrix} \\
|+\rangle_{\sigma_{2}} = \frac{1}{\sqrt 2} \begin{bmatrix} 1 \\ i \end{bmatrix} ~;~ |-\rangle_{\sigma_{2}}= \frac{1}{\sqrt 2} \begin{bmatrix} 1 \\ -i \end{bmatrix}
\end{eqnarray}
of Pauli operators $\sigma_{1} = \begin{bmatrix} 0 & 1\\ 1 & 0 \end{bmatrix}$ and $\sigma_{2}= \begin{bmatrix} 0 & -i\\ i & ~0 \end{bmatrix}$ as translational states. The choice of a particular Pauli basis for particular axis is purely conventional and all the matrices hereafter are represented in the basis formed by the eigenvectors of $\sigma_{3}$.
The general form of the coin operation in any of the basis states $|\pm \rangle_{\sigma_{\alpha}}$ of the Pauli operators $\sigma_{\alpha}$ with $\alpha = 1, 2, 3$ will be
\begin{eqnarray}
\label{coinP}
B_{\sigma_{\alpha}}(\theta) = \cos (\theta) |+ \rangle_{\sigma_{_{\alpha}}} \langle + |
+ \sin (\theta) |+ \rangle_{\sigma_{\alpha}} \langle -| \nonumber \\
- \sin (\theta) |- \rangle_{\sigma_{\alpha}} \langle + |
+ \cos (\theta) |- \rangle_{\sigma_{\alpha}} \langle - |.
\end{eqnarray}
The shift operator for each axis of the square lattice ($Z$ and $X$) and the cubic lattice ($Z$, $X$, $Y$) will be
\begin{subequations}
\begin{eqnarray}
\label{shiftS}
S_{\sigma_3}^{sq} \equiv \sum_{x, z} [ |+ \rangle_{\sigma_{3}}\langle
+|\otimes|\psi_{x, z-1}\rangle\langle \psi_{x, z}| \nonumber \\
+ | - \rangle_{\sigma_{3}}\langle
-|\otimes |\psi_{x, z+1}\rangle\langle \psi_{x, z}|], \\
S_{\sigma_1}^{sq} \equiv \sum_{x, z} [ |+ \rangle_{\sigma_{1}}\langle
+|\otimes|\psi_{x-1, z}\rangle\langle \psi_{x, z}| \nonumber \\
+ | - \rangle_{\sigma_{1}}\langle
-|\otimes |\psi_{x+1, z}\rangle\langle \psi_{x, z}|], \\
S_{\sigma_3}^{cub} \equiv \sum_{x,y,z} [ |+ \rangle_{\sigma_{3}}\langle
+|\otimes|\psi_{x, y,z-1}\rangle\langle \psi_{x, y, z}| \nonumber \\
+ | - \rangle_{\sigma_{3}}\langle
-|\otimes |\psi_{x,y, z+1}\rangle\langle \psi_{x, y, z}|], \\
S_{\sigma_1}^{cub} \equiv \sum_{x, y, z} [ |+ \rangle_{\sigma_{1}}\langle
+|\otimes|\psi_{x-1, y, z}\rangle\langle \psi_{x,y, z}| \nonumber \\
+ | - \rangle_{\sigma_{1}}\langle
-|\otimes |\psi_{x+1,y, z}\rangle\langle \psi_{x,y, z}|], \\
S_{\sigma_2}^{cub} \equiv \sum_{x, y, z} [ |+ \rangle_{\sigma_{2}}\langle
+|\otimes|\psi_{x, y-1, z}\rangle\langle \psi_{x,y, z}| \nonumber \\
+ | - \rangle_{\sigma_{1}}\langle
-|\otimes |\psi_{x,y+1, z}\rangle\langle \psi_{x,y, z}|],
\end{eqnarray}
\end{subequations}
where position state $|\psi_{x, y, z}\rangle = |\psi_{x}\rangle \otimes |\psi_{y}\rangle \otimes |\psi_{z}\rangle$. One complete step of the DTQW on a square and cubic lattice using two-state particle composes of the evolution in one axis followed by the evolution in the other, that is,
\begin{subequations}
\begin{eqnarray}
\label{w2d}
W^{sq}(\theta) = W_{\sigma_1}^{sq} (\theta) W_{\sigma_3}^{sq} (\theta),\\
\label{w3d}
W^{cub}(\theta) = W_{\sigma_2}^{cub} (\theta)W_{\sigma_1}^{cub} (\theta) W_{\sigma_3}^{cub} (\theta),
\end{eqnarray}
\end{subequations}
where $W_{\sigma_\alpha}^{sq} (\theta)= S_{\sigma_\alpha}^{sq} [ B_{\sigma_{\alpha}}(\theta) \otimes {\mathbbm 1}_X\otimes {\mathbbm 1}_Z]$ and $W_{\sigma_\alpha}^{cub} (\theta)=S_{\sigma_\alpha}^{cub} [ B_{\sigma_{\alpha}}(\theta) \otimes {\mathbbm 1}_X\otimes {\mathbbm 1}_Y \otimes {\mathbbm 1}_Z]$, respectively. Equivalent form of the evolution operators are,
\begin{subequations}
\begin{eqnarray}
W_{\sigma_3}^{sq} (\theta) \equiv{\mathbbm 1}_X \otimes W_{\sigma_3}(\theta) = {\mathbbm 1}_X \otimes e^{-iH_{\sigma_3}(\theta)} ~\\
W_{\sigma_1}^{sq} (\theta)\equiv W_{\sigma_1} (\theta)\otimes {\mathbbm 1}_Z \equiv e^{-iH_{\sigma_1}(\theta)} \otimes {\mathbbm 1}_Z~\\
W_{\sigma_3}^{cub} (\theta) \equiv {\mathbbm 1}_X\otimes {\mathbbm 1}_Y \otimes W_{\sigma_{3}}(\theta)
= {\mathbbm 1}_X\otimes {\mathbbm 1}_Y \otimes e^{-iH_{\sigma_3}(\theta)}~\\
W_{\sigma_1}^{cub} (\theta) \equiv W_{\sigma_{1}}(\theta) \otimes {\mathbbm 1}_Y \otimes {\mathbbm 1}_Z
= e^{-iH_{\sigma_1}(\theta)}\otimes {\mathbbm 1}_Y \otimes {\mathbbm 1}_Z ~\\
W_{\sigma_2}^{cub} (\theta) \equiv {\mathbbm 1}_X \otimes W_{\sigma_{2}}(\theta) \otimes {\mathbbm 1}_Z
= {\mathbbm 1}_X \otimes e^{-iH_{\sigma_2}(\theta)} \otimes {\mathbbm 1}_Z~.
\end{eqnarray}
\end{subequations}
The operator $W_{\sigma_3}(\theta)$ and $H_{\sigma_3}(\theta)$ are given by Eq.\,(\ref{eq:combop}) and Eq.\,(\ref{hamil2}). Similarly, the operator for $X$ and $Y$ axis,
\begin{subequations}
\begin{eqnarray}
&W_{\sigma_1} (\theta) = S_{\sigma_1} [ B_{\sigma_{1}}(\theta) \otimes {\mathbbm 1}] \equiv e^{-iH_{\sigma_1}(\theta)} \\
\label{eq:combop33}
&= \frac{e^{-i\hat{P}_X}}{2} \begin{bmatrix}
\cos(\theta)+ \sin(\theta) & \cos(\theta)-\sin(\theta)\\
\cos(\theta)+\sin(\theta) & \cos(\theta)-\sin(\theta)
\end{bmatrix} + \nonumber \\
& \frac{e^{+i\hat{P}_X}}{2} \begin{bmatrix}
\cos(\theta)-\sin(\theta) & -\cos(\theta)-\sin(\theta)\\
-\cos(\theta)+\sin(\theta) & \cos(\theta)+\sin(\theta)
\end{bmatrix};\\
& W_{\sigma_2} (\theta) = S_{\sigma_2} [ B_{\sigma_{2}}(\theta) \otimes {\mathbbm 1}] \equiv e^{-iH_{\sigma_2}(\theta)} \\
\label{eq:combop4}
&=\frac{e^{-i\hat{P}_Y}}{2}\begin{bmatrix}
\cos(\theta)+ \sin(\theta) & -i\cos(\theta)+i\sin(\theta)
\\ i\cos(\theta)+i\sin(\theta) & \cos(\theta)-\sin(\theta)
\end{bmatrix} \nonumber \\
&+\frac{e^{i\hat{P}_Y}}{2} \begin{bmatrix}
\cos(\theta)-\sin(\theta) & i\cos(\theta)+i\sin(\theta)
\\ -i\cos(\theta)+\sin(\theta) & \cos(\theta)+\sin(\theta)
\end{bmatrix},
\end{eqnarray}
\end{subequations}
where $\hat{P}_X$ and $\hat{P}_Y$ are the momentum operator on $X$ and $Y$ axis.
The eigenvalues, $\lambda^\mp_X$ of $W_{\sigma_1}(\theta)$ and $\lambda^\mp_Y$ of $W_{\sigma_2}(\theta)$ are same as the eigenvalues $\lambda^\mp_Z$ with only a replacement of $\hat{P}_X$ and $\hat{P}_Y$ in place of $\hat{P}_Z$ in Eq.\,(\ref{l1}). Therefore, using the eigenvalues, eigenvectors and inverse of eigenvector, the Hamiltonian form for the evolution in $X$ and $Y$ axis are
\begin{widetext}
\begin{subequations}
\begin{eqnarray}
\label{hamil2b}
H_{\sigma_{1}}(\theta)=
\frac{\ln \left (\frac{\lambda^{+}_X}{\lambda^{-}_X} \right )}{2 \sqrt{\cos^2(\theta)\cos^2(\hat{P}_X) -1}}
\begin{bmatrix}
\cos(\theta)\sin(\hat{P}_X) - i \cos(\hat{P}_X)\sin(\theta) \, &
\sin(\theta) \sin(\hat{P}_X)\, \\
-\sin(\theta) \sin(\hat{P}_X) \,
&
\cos(\theta)\sin(\hat{P}_X) + i \cos(\hat{P}_X)\sin(\theta)
\end{bmatrix} \cdot \sigma_1 \\
\label{hamil2d}
H_{\sigma_{2}}(\theta) =
\frac{\ln \left (\frac{\lambda^{+}_Y}{\lambda^{-}_Y} \right )}{2 \sqrt{\cos^2(\theta)\cos^2(\hat{P}_Y) -1}}
\begin{bmatrix}
\cos(\theta)\sin(\hat{P}_Y)-i \cos(\hat{P}_Y)\sin(\theta) &
-i\sin(\theta) \sin(\hat{P}_Y)\\
-i\sin(\hat{P}_Y) \sin(\theta)
&
\cos(\theta)\sin(\hat{P}_Y)+i\cos(\hat{P}_Y)\sin(\theta)
\end{bmatrix} \cdot \sigma_2
\end{eqnarray}
\end{subequations}
\end{widetext}
When $\theta = 0$, that is, in absence of a coin operation $\lambda^{\mp}_Z = e^{\mp i\hat{P}_Z}$, $\lambda^{\mp}_X = e^{\mp i\hat{P}_X}$, and $\lambda^{\mp}_Y = e^{\mp i\hat{P}_Y}$. This reduces the Hamiltonian form to, $H_{\sigma_{3}}(0) =
\begin{bmatrix}
\hat{P}_Z & 0 \\
0 & \hat{P}_Z
\end{bmatrix} \cdot \sigma_3$, $H_{\sigma_{1}}(0) =
\begin{bmatrix}
\hat{P}_X & 0 \\
0 & \hat{P}_X
\end{bmatrix} \cdot \sigma_1$, and $H_{\sigma_{2}}(0) =
\begin{bmatrix}
\hat{P}_Y & 0 \\
0 & \hat{P}_Y
\end{bmatrix} \cdot \sigma_2$, respectively.
\par
If the initial state of the particle on a square lattice $|C \rangle=\frac{1}{\sqrt{2}}[|\downarrow\rangle + i |\uparrow \rangle]$,
$|\Psi_{\rm in} \rangle = \frac{1}{\sqrt 2}\begin{bmatrix}
[|\psi_{x_0}\rangle \otimes |\psi_{z_0}\rangle] \\
i[|\psi_{x_0}\rangle \otimes |\psi_{z_0}\rangle]
\end{bmatrix}$ and the state after $t$ step [$(W^{sq}(\theta))^t$],
\begin{eqnarray}
|\Psi_t\rangle = \sum_{x =-t}^t \sum_{z =-t}^t \Big [ \alpha^{(1)}_{(x, z, t)}|\downarrow
\rangle + \alpha^{(2)}_{(x, z, t)}|\uparrow \rangle \Big ]
\otimes|\psi_{x, z}\rangle.
\label{eq:lr2}
\end{eqnarray}
When $\theta =0$, $\alpha^{(1)}_{(x, y,t)}$ and $\alpha^{(2)}_{(x, y, t)}$ are given by the
coupled iterative relations
\begin{subequations}
\label{eq:iter}
\begin{eqnarray}
\alpha^{(1)}_{(x, y,t)} = \frac{1}{2}\Big [ \alpha^{(1)}_{(x+1, y+1,t-1)} + \alpha^{(1)}_{(x+1, y-1,t-1)} \nonumber \\
+ \alpha^{(2)}_{(x-1, y+1,t-1)} - \alpha^{(2)}_{(x-1, y-1,t-1)}\Big ] \\
\alpha^{(2)}_{(x, y,t)} = \frac{1}{2}\Big [ \alpha^{(1)}_{(x+1, y+1,t-1)} - \alpha^{(1)}_{(x+1, y-1,t-1)} \nonumber \\
+ \alpha^{(2)}_{(x-1, y+1,t-1)} + \alpha^{(2)}_{(x-1, y-1,t-1)}\Big ].
\end{eqnarray}
\end{subequations}
\begin{figure}[ht]
\begin{center}
\subfigure[]{\includegraphics[width=63mm]{squareL00.png}
\label{fig:3a}}
\subfigure[]{\includegraphics[width=63mm]{squareL1515.png}\label{fig:3b}}
\caption{\footnotesize{Probability distribution of a 50 step two-state particle DTQW with the initial state $|\Psi_{\rm in} \rangle = \frac{1}{\sqrt 2}[|\downarrow \rangle + i|\uparrow \rangle]\otimes|\psi_{x_0}\rangle \otimes |\psi_{z_0}\rangle$ on a square lattice using basis state of different Pauli operator for each axis ($\sigma_1$ for $X$ axis and $\sigma_3$ for $Z$ axis). (a) The distribution is after the evolution without the coin operation in both the axis ($\theta =0$) and same distribution is obtained for Grover walk. (b) The distribution is after the evolution with the coin operation $\theta = \pi/12$ in both the axis.}}
\end{center}
\end{figure}
In Fig.\,\ref{fig:3a}, the probability distribution of the 50 step DTQW on a square lattice using the Pauli basis scheme without the coin operation ($\theta = 0$) is shown. This probability distribution obtained is identical to the ones reported for Grover walk on a four-state particle \cite{TFM03} and for the alternative walk on a two-state particle with initial state, $\frac{1}{\sqrt{2}}[|\downarrow \rangle + i |\uparrow \rangle]$ using Hadamard operator as the coin operation \cite{FGB11}. Each step of Grover walk is on a 2D is realized using the Grover diffusion operator
$G = \frac{1}{2}\begin{bmatrix}-1 & ~~1 & ~~1 & ~~1 \\
~~1 & -1 & ~~1 & ~~1 \\
~~1 & ~~1 & -1 & ~~1 \\
~~1 & ~~1 & ~~1 & -1
\end{bmatrix}$ as coin operation Followed by $
S^G \equiv \sum_{x, z} \Big[|\downarrow \rangle\langle \downarrow|\otimes|
\psi_{x-1, z-1}\rangle\langle \psi_{x, z}|
+ |\uparrow \rangle\langle \uparrow|\otimes |
\psi_{x-1, z+1}\rangle\langle \psi_{x, z}|
+ |\leftarrow \rangle\langle \leftarrow|\otimes|
\psi_{x+1, z-1}\rangle\langle \psi_{x, z}|
+ |\rightarrow \rangle\langle \rightarrow|\otimes |
\psi_{x+1, z+1}\rangle\langle \psi_{x, z}| \Big]$ on a particle in a specific initial state, $|\Psi_{\rm in}^{G}\rangle=\frac{1}{2}[|\downarrow \rangle -|\uparrow \rangle -|\leftarrow \rangle +|\rightarrow \rangle]$.
The state after $t$ step of the Grover walk [$S^G(G \otimes {\mathbbm 1})^t]$,
\begin{eqnarray}
|\Psi^G_t\rangle = \sum_{x =-t}^t \sum_{z =-t}^t \Big [ \beta^{(1)}_{(x, z, t)}|\downarrow
\rangle + \beta^{(2)}_{(x, z, t)}|\uparrow \rangle + \beta^{(3)}_{(x, z, t)}|\leftarrow \rangle \nonumber \\
+ \beta^{(4)}_{(x, z, t)}|\rightarrow \rangle \Big ]
\otimes|\psi_{x, z}\rangle,
\label{eq:lr2}
\end{eqnarray}
where $\beta(x, y,t)$'s are given by the
quadrupled iterative relation coupling the $X$ and $Z$ axis
\begin{subequations}
\label{eq:4s_iter}
\begin{eqnarray}
\beta^{(1)}(x, z, t) &= \frac{1}{2} \Big[-\beta^{(1)}_{(x+1, z+1,t-1)} + \beta^{(2)}_{(x+1, z+1,t-1)} \nonumber \\
& + \beta^{(3)}_{(x+1, z+1,t-1)} + \beta^{(4)}_{(x+1, z+1,t-1)}\Big ] \\
\beta^{(2)}(x, z, t) &= \frac{1}{2} \Big[\beta^{(1)}_{(x+1, z-1,t-1)} - \beta^{(2)}_{(x+1, z-1,t-1)} \nonumber \\
&+ \beta^{(3)}_{(x-1, z-1,t-1)} + \beta^{(4)}_{(x+1, z-1,t-1)} \Big] \\
\beta^{(3)}(x, z, t) &= \frac{1}{2} \Big[\beta^{(1)}_{(x-1, z+1,t-1)} + \beta^{(2)}_{(x-1, z+1,t-1)} \nonumber \\
& - \beta^{(3)}_{(x-1, z+1,t-1)} + \beta^{(4)}_{(x-1, z+1,t-1)}\Big ]\\
\beta^{(4)}(x, z, t) &= \frac{1}{2} \Big [\beta^{(1)}_{(x-1, z-1,t-1)} + \beta^{(2)}_{(x-1, z-1,t-1)} \nonumber \\
& + \beta^{(3)}_{(x-1, z-1,t-1)} - \beta^{(4)}_{(x-1, z-1,t-1)} \Big ].
\end{eqnarray}
\end{subequations}
From Eqs.\,(\ref{eq:iter}) and Eqs.\,(\ref{eq:4s_iter}) we can note that for both, two-state walk and the Grover walk, the amplitude at any position $(x, z)$ for a given time $t$ is dependent on the amplitude at the four diagonally opposite sites at time $t-1$. Therefore, starting from a specific initial state of two-state and four-state particle as discussed in this section, these amplitudes returns the same probability distribution.
\par
Unlike the Grover walk which is very specific to the initial state and the coin operation, probability distribution with the two-state walk using the Pauli basis can be controlled by introducing the coin operation ($\theta \neq 0$) and/or using different initial state of the particle. In Fig.\,\ref{fig:3b} the probability distribution of the 50 step DTQW with coin operation, $\theta = \pi/12$ is show to squeeze the distribution towards the diagonal of the square lattice.
\par
The basis states of $\sigma_3$ can be written as a superposition of a basis states of the $\sigma_1$. Therefore, even in absence of a coin operation, $[W^{sq}_{\sigma_{1}}(0)W^{sq}_{\sigma_{3}}(0)]^t$ evolves the particle in superposition of position space and implement a DTQW on a square lattice. Similarly, for a two-state walk on a cubic lattice using different basis states, due to the relationship between the basis states of the Pauli operators, the particle evolves in superposition of position space even in absence of coin operation. However, the coin operation can be effectively used to control the dynamics and the probability distribution of the walk.
\subsection{Triangular lattice}
The triangular lattice structures shown in Fig.~\ref{fig:4} also has a three axis of propagation, $X, Y$, and $Z$. Therefore, the walk can be quantized using the eigenstates, $|+\rangle_{\sigma_{\alpha}}$ and $|-\rangle_{\sigma_{\alpha}}$ of the Pauli operators $\sigma_{\alpha}$ where $\alpha =$ 1, 2, and 3, as translational basis states.
\begin{figure}[ht]
\begin{center}
\includegraphics[width=7.6cm]{triangularb.png}
\caption{\footnotesize{ Triangular lattice structure with labeling of lattice position in all the three axis, $X$, $Y$, and $Z$ of propagation quantized by the basis states of the Pauli operators $\sigma_{1}$, $\sigma_{2}$ and $\sigma_{3}$, respectively.}}
\label{fig:4}
\end{center}
\end{figure}
\begin{figure}[ht]
\subfigure[]{
\includegraphics[width=5.5cm]{triangularc.png}
\label{fig:5}}
\subfigure[]{\includegraphics[width=6.2cm]{kagome.png}
\label{fig:6}}
\caption{\footnotesize{ Scheme for evolution of DTQW on: (a) The triangular lattice, starting from the middle, the arrow marks show the shift in position space during one step of DTQW evolution (evolution in $Z$ axis followed by the evolution in the $Y$ and $X$ axis). (b) Kagome lattice structure with two axis of propagation at each lattice site. From lattice sites {\bf o}, {\bf p}, and {\bf q}, we can see that they are associated with different combination of quantization axis. Starting from position {\bf p}, the arrow marks show the shift in position space during one step of DTQW evolution. The final positions are encircled.}}
\end{figure}
Unlike the cubic lattice the three axis for propagation in triangular lattice are not orthogonal to each other and hence, the evolution in one axis alters the evolution in other two axis as well.
Therefore, the shift operator has to be defined according to the lattice structure. Shift operator in each axis can be defined such that the unit shift in the main axis is accompanied by half of the unit shift in the other two axis. In Fig.\,\ref{fig:4}, the labeling of the position is shown and for convenience we choose the two unit position shift in the main axis and one unit position shift in the other two axis. Therefore, the shift operation for $\sigma_3$ basis states is,
\begin{eqnarray}
S_{\sigma_3} \equiv \sum_{x, y, z} |\downarrow \rangle\langle
\downarrow |\otimes|\psi(x+1, y-1, z-2)\rangle\langle \psi(x, y, z)| \nonumber \\
+ \sum_{x, y, z} | \uparrow \rangle\langle
\uparrow |\otimes |\psi(x-1, y+1, z+2)\rangle\langle \psi(x, y, z)|.
\end{eqnarray}
and the effective evolution operator
\begin{eqnarray}
\label{eq:comboptri}
W^{\prime}_{\sigma_3}(\theta) = \begin{bmatrix}
\mbox{~~}\cos(\theta) e^{-i\hat{P}_3} & & \sin(\theta)e^{-i\hat{P}_3}
\\ - \sin(\theta)e^{+i\hat{P}_3} & & \cos(\theta)e^{+i\hat{P}_3}
\end{bmatrix},
\end{eqnarray}
where $\hat{P}_3 = -\hat{P}_{X} \otimes \hat{P}_{Y} \otimes 2\hat{P}_{Z}$. Therefore, the Hamiltonian $H^{\prime}_{\sigma_{3}}(\theta)$ will be in the same form of Eq. (\ref{hamil2}) with a replacement of $\hat{P}_{Z}$ by $\hat{P}_3$. Similarly, the evolution operator $W^{\prime}_{\sigma_1}(\theta)$ and $W^{\prime}_{\sigma_2}(\theta)$, and the Hamiltonian form in the $X$ and $Y$ axis, $H^{\prime}_{\sigma_{1}}(\theta)$ and $H^{\prime}_{\sigma_{2}}(\theta)$, will be in the same form of Eqs. (~\ref{eq:combop33}),~(\ref{eq:combop4}) and Eqs. (~\ref{hamil2b}),~(\ref{hamil2d}), respectively with the replacement of $\hat{P}_{X}$ by $\hat{P}_1 = 2\hat{P}_{X} \otimes \hat{P}_{Y} \otimes -\hat{P}_{Z}$ and $\hat{P}_{Y}$ by $\hat{P}_2 = -\hat{P}_{X} \otimes 2\hat{P}_{Y} \otimes \hat{P}_{Z}$.
\begin{figure}[ht]
\begin{center}
\subfigure[]{\includegraphics[width=70.5mm]{triagularL0.png}
\label{fig:5a}}
\hskip -0.17in
\subfigure[]{\includegraphics[width=70.5mm]{triagularL1.png}
\label{fig:5b}}
\vskip -0.25cm
\subfigure[]{\includegraphics[width=70mm]{triagular0450.png}
\label{fig:5c}}
\caption{\footnotesize{Probability distribution of 20 step two-state particle DTQW on triangular lattice.
(a) The initial state is $|\Psi_{\rm in} \rangle = |\downarrow \rangle \otimes|\psi_{x_0}\rangle \otimes |\psi_{y_0}\rangle \otimes |\psi_{z_0}\rangle$ and the walk is evolved without a coin operation. (b) The initial state is $|\Psi_{\rm in} \rangle = |\uparrow \rangle \otimes|\psi_{x_0}\rangle \otimes |\psi_{y_0}\rangle \otimes |\psi_{z_0}\rangle$ and the walk is evolved without a coin operation. (c) The initial state is $|\Psi_{\rm in} \rangle = |\downarrow \rangle \otimes|\psi_{x_0}\rangle \otimes |\psi_{y_0}\rangle \otimes |\psi_{z_0}\rangle$ and the walk is evolved with one coin operation [$H^{\prime}_{\sigma_{1}}(0)+ H^{\prime}_{\sigma_{2}}(\pi/4)+ H^{\prime}_{\sigma_{3}}(0)$] to obtain a symmetric probability distribution in position space.}}
\end{center}
\end{figure}
Each Hamiltonian, $H^{\prime}_{\sigma_{1}}(\theta)$, $H^{\prime}_{\sigma_{2}}(\theta)$, and $H^{\prime}_{\sigma_{3}}(\theta)$ evolve the state enabling the interaction between the three quantization axis.
Therefore, each step of DTQW can be realized by the evolution using one Pauli basis followed by the other, $W^{tri}(\theta) = W^{\prime}_{\sigma_1}(\theta)W^{\prime}_{\sigma_2}(\theta)W^{\prime}_{\sigma_3}(\theta)$ as shown in Fig.~\ref{fig:5}. We should note that the three effective Hamiltonian for the operators, $H^{\prime}_{\sigma_{1}}(\theta)$, $H^{\prime}_{\sigma_{2}}(\theta)$, and $H^{\prime}_{\sigma_{3}}(\theta)$ together commute and complete Hamiltonian for each step of DTQW on triangular lattice can be written as
\begin{eqnarray}
\label{hamil55}
H^{\prime}(\theta) &=& H^{\prime}_{\sigma_{1}}(\theta) + H^{\prime}_{\sigma_{2}}(\theta) + H^{\prime}_{\sigma_{3}}(\theta).
\end{eqnarray}
The choice of the order of the basis in which the particle is evolved is purely conventional for the triangular lattice. Even when $\theta =0$, due to the interplay between different Pauli basis for translation in each axis, a two-state particle evolve in superposition of position space resulting in a diffused probability distribution. However, a coin operation with different $\theta$ for each axis can be extensively used for the evolution to get addition freedom to control the evolution and obtained the desired probability distribution. In Fig.\,\ref{fig:5a} and \,\ref{fig:5b}, we show the probability distribution of a 20 step DTQW without a coin operation [$H^{\prime}(0) = H^{\prime}_{\sigma_{1}}(0) + H^{\prime}_{\sigma_{2}}(0) + H^{\prime}_{\sigma_{3}}(0)$] on a two-state particle initially in state $|\downarrow \rangle$ and $|\uparrow \rangle$, respectively. We can see that the probability distribution in Fig.\,\ref{fig:5a} and Fig.\,\ref{fig:5b} are not symmetric distribution in position space but are symmetric to each other. In Fig.\,\ref{fig:5c} we show the symmetric probability distribution obtained by introducing a coin operation with $\theta = \pi/4$ for only one operation during each step evolution [$H^{\prime}_{\sigma_{1}}(0) + H^{\prime}_{\sigma_{2}}(\pi/4) + H^{\prime}_{\sigma_{3}}(0)$].
\subsection{Kagome Lattice}
Kagome lattice structure can also be labeled the same way as the triangular lattice. The evolution operator and its Hamiltonian form in each basis ($H^{\prime}_{\sigma_{1}}(\theta)$, $H^{\prime}_{\sigma_{2}}(\theta)$, and $H^{\prime}_{\sigma_{3}}(\theta)$) will be in the same form as presented for triangular lattice. But, unlike triangular lattice which has three quantization axis at each lattice site, kagome lattice shown in Fig. \ref{fig:6} has only two quantization axis with four direction of propagation for the walk at each lattice site. The two quantization axis at each lattice site is not the same for all lattice sites. In Fig. \ref{fig:6}, lattice sites {\bf o}, {\bf p}, and {\bf q} have axis $X$ and $Z$ ($\sigma_1$ and $\sigma_3$), $X$ and $Y$ ($\sigma_1$ and $\sigma_2$), and $Y$ and $Z$ ($\sigma_2$ and $\sigma_3$) as quantization axis, respectively. Therefore, to implement each step of DTQW in kagome lattice certain simple order for using evolution operators in different axis has to be followed. For example, if the initial position is {\bf p} (as marked in Fig.~\ref{fig:6}), each step of DTQW can be realized by $W^{kag}(\theta) = W^{\prime}_{\sigma_1}(\theta)W^{\prime}_{\sigma_3}(\theta)W^{\prime}_{\sigma_2}(\theta)$ and the effective Hamiltonian form for each step of DTQW will be
$H^{\prime}(\theta) = H^{\prime}_{\sigma_{1}}(\theta) + H^{\prime}_{\sigma_{3}}(\theta) + H^{\prime}_{\sigma_{2}}(\theta)$.
The main consideration in choosing the first axis is, not to pick the evolution in the axis which is nonexistent in that initial position.
\section{Conclusion}
\label{conc}
In this paper we presented a new scheme for the evolution of DTQW on 2D and 3D lattices using a two-state particle. Our scheme used different Pauli basis states as translational eigestate in different axis and showed that the coin operation is not a necessary requirement to implement a walk on 2D and 3D systems but can be used as an additional degree of freedom to control the dynamics. We also discussed the Hamiltonian form of evolution for the walk which can serve as a general framework to simulate, control, and study the dynamics in different physical systems. The Pauli basis states for translation is commonly used to describe the dynamics in various physical systems, in particular, in quantum optics and optical lattice \cite{DDL03}. Therefore, use of Pauli basis state for translation without the use of coin operation and the Hamiltonian form can serve as a frame work for experimental implementation of DTQW with a minimum resource in various 2D and 3D physical structures. Our scheme for evolution on square, cubic, triangular, and kagome lattice can be straight away extended to other 2D and 3D Bravais lattice and the extension to other higher dimensions is also possible by permuting the three Pauli basis states for each translational axis. This description of dynamics in Hamiltonian form helps to further explore topological phase, establish connection between physical process in nature which are generally not 1D and does not involve larger internal (more than two) dimension of the particle.
|
\section{Supplemental material for `` Klein Tunneling and Berry Phase $\pi$ in Bilayer Graphene with a Band Gap '' }
In this material, we discuss the details (the calculation method and the properties) of the Veselago lens. We also discuss how to detect the Klein effect in a bilayer graphene nanoribbon.
\section{I.~ Electronic Veselago lens in bilayer graphene}
In this section, we provide the method of calculating the properties of the Veselago lens, and also more information about the Veselago lens.
The wave intensity in the Veselago setup (a bilayer graphene $np$ junction) is obtained,
using the plane-wave expansion method \cite{Cincotti}.
The setup is formed by the potential energy $V(x)$ and band gap $\Delta (x)$,
\begin{equation}
(V(x),~\Delta (x)) = \left\lbrace
\begin{array}{ll}
(V_1 ,~\Delta_1), \,\,\, & \hbox{$x<0$},\\
(V_2,~\Delta_2), \,\,\, & \hbox{$x>0$}.
\end{array}
\right. \nonumber
\end{equation}
The electron point source is located at $(x, y) = (x_{{\rm src}},0) = (-100 l_0, 0)$ in the $n$ region, where
$l_0 = \hbar v / \gamma \approx 1.7 \textrm{nm}$.
The wave diverging from the source, $\Psi^{{\rm K}({\rm K}')}_{{\rm src}}$,
is described by the out-going wave of angular momentum zero,
\begin{equation}\label{SrcWave}
\Psi^{{\rm K}({\rm K}')}_{\rm{src}}(x,y)
=
\frac{H^{(1)}_{1} ( \rho k_1 )}{\sqrt{N_1}}
\left( \begin{array}{c}
\sqrt{V^{2}_1 - (\Delta^{2}_1 /4)}~ e^{\mp i \phi} \\
(V_1 + \Delta_1 /2)~ e^{\pm i \phi}
\end{array}
\right). \nonumber
\end{equation}
Here, the upper (lower) sign corresponds to K (K$'$) valley,
we used the cylindrical coordinate $(\rho, \phi)$ whose origin is at $(x_{{\rm src}},0)$,
$(\rho~{\rm cos}~\phi, \rho~{\rm sin}~\phi) = (x - x_{{\rm src}}, y)$,
$H^{(1)}_{1}$ is the Hankel function of the first kind, $k_1 = \sqrt{(\gamma/ \hbar^2 v^2) \sqrt{V^{2}_{1} - \Delta^{2}_{1}/4}}$ is the wave vector of the wave, and $N_1 = 2 V^{2}_1 + V_1 \Delta_1$ is the normalization constant of the pseudospin part.
The energy $E$ of the wave, related to $k_1$ by $(E-V_1)^2 = \Delta_1^2 / 4 + (\hbar^2 v^2 k_1^2/\gamma)^2$, is set to be $E=0$ as in the main text.
In the calculation of wave refraction, we apply the following approximation.
In the spatial region where $\rho k_1 \gg 1$ is satisfied, $\Psi^{{\rm K}({\rm K}')}_{{\rm src}}$
can be approximately written \cite{Cincotti} in terms of plane waves,
\begin{widetext}
\begin{equation}
\Psi^{{\rm K}({\rm K}')}_{{\rm src}}(x,y)
\approx
\frac{-i}{\pi \sqrt{N_1}} \int^{\pi/2}_{-\pi/2} d \theta_1~ e^{i k_{1_x}(x - x_{{\rm src}}) + i k_{y} y}
\left( \begin{array}{c}
\sqrt{V^{2}_1 - (\Delta^{2}_1 /4)} ~e^{\mp i \theta_1} \\
(V_1 + \Delta_1 /2) ~e^{\pm i \theta_1}
\end{array}
\right), \nonumber
\end{equation}
\end{widetext}
where $k_{1_x} = s_1 k_1 \cos \theta_1$,
$k_{y} = s_1 k_1 \sin \theta_1$, and
$s_1 = -\textrm{sgn} (V_1)$.
Here, $\theta_1$ can be regarded as the propagation angle of the plane wave in the $n$ region.
The setup in Fig.~4 of the main text satisfies the condition of $x_\textrm{src} k_1 \gg 1$, hence the above approximation is well applicable to it;
we have checked that the error due to the approximation does not alter the main features discussed in the main text.
The approximation allows us to express
the refracted waves $\Psi^{{\rm K}({\rm K}')}_{{\rm rfr}}$ in terms of the transmission coefficient $t_{np}(\theta_1)$ through the junction, which is mentioned in Eq.~(6) of the main text,
\begin{widetext}
\begin{equation}\label{RfrWave}
\Psi^{{\rm K}({\rm K}')}_{{\rm rfr}}(x,y)
\approx
\frac{-i}{\pi \sqrt{N_2}} \int^{\pi/2}_{-\pi/2} d \theta_1~
t_{np}(\theta_1)~e^{-i k_{1_x} x_{{\rm src}} + i k_{2_x} x + i k_{y} y}
\left( \begin{array}{c}
\sqrt{V^{2}_2 - (\Delta^{2}_2 /4)} ~e^{\mp i \theta_2} \\
(V_2 + \Delta_2 /2) ~e^{\pm i \theta_2}
\end{array}
\right). \nonumber
\end{equation}
\end{widetext}
$k_{2_x}, N_2,$ and $s_2$, are defined in the same way as $k_{1_x}, N_1,$ and $s_1$, except
for the subscript $2$.
The propagation angle $\theta_2$ of the refracted wave is governed by the conservation
of $p_y$, $s_1 \sqrt{\gamma \sqrt{V^{2}_{1} - \Delta^{2}_{1}/4}}~ \textrm{sin} \theta_1
= s_2 \sqrt{\gamma \sqrt{V^{2}_{2} - \Delta^{2}_{2}/4}}~ \textrm{sin} \theta_2$.
We calculate $t_{np}$ by using the boundary matching method of plane waves at $x = 0$, and then obtain the wave intensity from $I_{{\rm K}({\rm K}')}(x,y) = |\Psi^{{\rm K}({\rm K}')}_{{\rm rfr}}(x,y)|^2$. As we are considering the regime where the intervalley scattering is absent, the total intensity $I$ is the sum of $I_{\rm K}$ and $I_{\rm K'}$, i.e., there is no interference between the waves of K and K$'$ valleys.
Hereafter, we give additional information about the Veselago lens. We first mention the Klein effect in the smooth $np$ junction where $V(x)$ and $\Delta(x)$ vary smoothly around $x=0$.
In the main text, for the sharp junction in which the values of $V(x)$ and $\Delta(x)$ jump at $x=0$, we find the feature that $I(x,0)=0$ along the line of $y=0$ in the $p$ region of $x > 0$, when the condition of $\alpha_2 /\alpha_1 =1$ is achieved.
This feature robustly appears in the smooth $np$ junction.
The robustness comes from the facts that the symmetry in Eq.~(5) of the main text is still valid, and that the value of $\Delta(x) / V(x)$ is almost constant~\cite{McCann_smooth} when $V(x)$ varies sufficiently smoothly.
Next we mention about electron propagation in the Veselago lens setup shown in Fig.~4 of the main text. In the $p$ region, the interference between three different refracted waves can appear at $(x \ge x_\textrm{cusp}, y=0)$; in contrast, there appears only a single refracted wave at $0 < x < x_\textrm{cusp}$. The three different waves propagate from the source to $(x,y=0)$ as follows.
One of them propagates from the source to the junction interface with incident angle $\theta_1 = \theta_x$ to the interface, and then it arrives at $(x,y=0)$ after its refraction at the interface. Here, the relation between $\theta_x$ and $x$ is found to be $x = x_{\rm cusp} \sqrt{(1-|V_1/V_2| {\rm sin}^2 \theta_x)/{\rm cos}^2 \theta_x}$, where $x_{\rm cusp} = |x_{\rm src}| \sqrt{|V_2/V_1|}$. Another wave follows the path with $\theta_1 = - \theta_x$, which is the mirror-reflection (about $y=0$ axis) path of the wave with $\theta_1 = \theta_x$. The other wave follows the path with normal incidence of $\theta_1 = 0$; the transmission amplitude of this wave through the interface of $x=0$ is in fact zero, when the condition of $\alpha_1 / \alpha_2 = 1$ for the Klein effect is achieved.
The interference between the three waves shows the clear signature of the Klein effect and the Berry phase $\pi$, as shown in the main text.
\section{II. ~ Klein effect in armchair nanoribbon}
In the main text, we find that the Klein effect and the Berry phase $\pi$ can be detected in the Veselago lens.
In this section, we discuss another way of experimentally detecting the Klein effect
in a bilayer graphene armchair nanoribbon.
The nanoribbon has armchair edges along $\hat{x}$ axis and shows metallic behavior in the absence of external potential and band gap. Its stacking configuration is shown in Supplementary Figure 1(a).
We consider an $npn$ junction in the nanoribbon, formed by the potential energy $V(x)$ and band gap $\Delta(x)$,
\begin{equation} \label{Ribbon_potential}
(V(x),~\Delta (x)) = \left\lbrace
\begin{array}{ll}
(V_1 ,~\Delta_1), \,\,\, & \hbox{$x<0$ and $x > d$},\\
(V_2,~\Delta_2), \,\,\, & \hbox{$0<x<d$}.
\end{array}
\right. \nonumber
\end{equation}
To see the signature of the Klein effect that the transmission probability of an electron through the junction is zero at normal incidence of $\theta_1 = 0$, one needs to study the energy regime having only one transverse mode, which corresponds to the state with $k_y=0$ in the bulk limit~\cite{Gonzalez}.
Focusing on this energy regime,
we numerically calculate the transmission probability $T = |t|^2$ through the $npn$ junction, by combining the tight-binding method and Green's function~\cite{Datta, Sim2}; as mentioned in the main text, $t$ is the transmission coefficient, and the energy of the incident electron is set by $E=0$.
\begin{figure}
\includegraphics[width=0.45\textwidth]{Supp_Ribbon1.eps}
\caption{Supplemantary figure 1.
(a) Stacking structure of a bilayer graphene armchair ribbon. Unfilled (filled) circles connected by dashed (solid) lines represent the lower (upper) layer of the bilayer ribbon. Electrons propagate along $\hat{x}$ direction.
(b) Transmission probability $T$ (with log scale plot on the right side) through an $npn$ junction
as a function of $\alpha_2 / \alpha_1$. The Klein effect of $T=0$ is shown at $\alpha_2 / \alpha_1 \approx 1$.
}
\label{fig:ribbon}
\end{figure}
In Supplementary figure 1(b), we plot $T$ as a function of $\alpha_2 / \alpha_1$, where $\alpha_{1(2)}= \Delta_{1(2)}/ V_{1(2)}$.
We choose $\alpha_1 = -0.3$, $V_1 = -0.05 \gamma$, $V_2 = 0.26 \gamma$,
$d = 60 l_0$, and the width of the ribbon $W = 6 a$, where $a$ is the lattice constant of graphene.
Here, the subindex $1$ ($2$) refers to the region of $x < 0$ and $x > d$ ($0 < x < d$).
We choose $V$ and $\Delta$ spatially varying over the length scale of $27 l_0$ around $x=0$ and $d$; in this case the intervalley mixing is prevented.
The transmission probability $T$ shows the perfect reflection of $T=0$ at $\alpha_2 / \alpha_1 \approx 1$. This shows the Klein effect, the transmission zero at normal incidence of $\theta_1=0$ in the bulk limit, hence, one can observe the Klein effect with tuning $\alpha_j$ by gate voltages.
This signature of the Klein effect (the perfect reflection) is distinguishable from the transmission valleys between resonance peaks. It is because the perfect reflection is independent of parameters such as $d/\lambda$, while the resonance peaks do depend on those parameters; $\lambda = 2 \pi \hbar v / \sqrt{\gamma \sqrt{V^{2}_{2} - \Delta^{2}_{2}/4}}$ is an electron wavelength in the $p$ region.
The Klein effect in the armchair nanoribbon is robust against the details of the shapes
of $V(x)$ and $\Delta (x)$ around the interfaces of the $npn$ junction, provided that the valley mixing is negligible.
The robustness implies that the Klein effect
will survive in the presence of screened-Coulomb interaction, because the interaction
may induce small change of the spatial shape of $V(x)$ and $\Delta (x)$ at the interfaces \cite{Zhang}.
The above finding appears in a metallic armchair ribbon with arbitrary width, provided that the energy
regime has one transverse mode corresponding to the states with $k_y = 0$ in the bulk limit.
Note that it is difficult to observe the effect in ribbons with non-metallic armchair edges or with zigzag edges.
This is because there is no state corresponding to the $k_y = 0$ state injecting normal to the junction, i.e., because the band dispersion of the ribbon in the low-energy regime is different
from that in bulk limit due to the edges.
The perfect reflection may disappear in the presence of valley-symmetry-breaking disorders, e.g.,
short range disorders and edge disorders.
Such disorders induce valley mixing, and hence the Berry phase argument based on the time reversal and
electron-hole symmetries defined in a single valley is not applicable.
|
\section{#1}}
\newcommand{\conclusio{section}{1} \setcounter{equation}{0}{\conclusio{section}{1} \setcounter{equation}{0}
\section*{Appendix \Alph{section}}}
\setlength{\parindent}{0.22in} \setlength{\textheight}{9.2in}
\setlength{\textwidth}{15.5cm} \setlength{\topmargin}{-.3in}
\setlength{\evensidemargin}{-1cm}
\setlength{\oddsidemargin}{-.2cm}
\newsavebox{\PSLASH}
\sbox{\PSLASH}{$p$\hspace{-1.8mm}/}
\newcommand{\usebox{\PSLASH}}{\usebox{\PSLASH}}
\begin{document}
\title{Conformal symmetry in non-local field theories}
\author{ M. A. Rajabpour$^{a}$\footnote{e-mail: <EMAIL>} \\ \\
$^{a}$SISSA and INFN,
\textit{Sezione di Trieste}, via Bonomea 265, 34136 Trieste, Italy}
\maketitle
\begin{abstract}
We have shown that a particular class of
non-local free field theory has conformal symmetry in arbitrary dimensions. Using the local field theory counterpart of this class, we have found the Noether currents and Ward identities of the translation, rotation and scale symmetries. The operator product expansion of the energy-momentum tensor with quasi-primary fields is also investigated.
\end{abstract}
\section{Introduction}\
Non-local field theories are well-known as a method to describe the scaling limit of the long-range interacting systems and they are much studied in statistical physics and string theory. Long-range spin systems \cite{FNM}, rough surfaces \cite{rough} and diffusion processes \cite{Gornflo} are just few examples among many in statistical physics.
Although many studies have dealt with non-local field theories, mostly calculating the renormalization group equations and scaling properties, we still know very little about the role of the symmetries in these systems.
One of the very powerful symmetries is conformal symmetry, which is the Poincare symmetry plus scale and special conformal symmetry. It is powerful because it puts several restrictions on the form of the correlation functions. For an extensive study of conformal field theory (CFT), see \cite{DMS,Cardy1}. For a recent tutorial on conformal symmetry in diverse dimensions, see \cite{JP}. Since special conformal symmetry is a symmetry made of translation and inversion, it could be quite surprising if we find conformal symmetry in non-local field theories. The rule for conformal symmetry in non-local field theories is already discussed in the string theory community by usually defining the theory by its correlation functions\footnote{There are also some comments about the conformal symmetry at the level of the non-local Lagrangian \cite{RV}.} \cite{HPPS, EP}. Some deformed versions of conformal symmetry in non-local field theories are also discussed in ageing systems, see \cite{Malte1,Malte2} and reference therein. In this paper we will study systematically the fractional Laplacian field theory (at the level of the lagrangian) which is a generalization of the field theory with different powers of the Laplacian. We first show that this field theory has conformal symmetry in any dimension. In two dimensions it has just global conformal symmetry and not full conformal symmetry. Then we will study different consequences of conformal symmetry in these field theories.
The paper is organized as follows. In the section ~2 we define the fractional Laplacian and we show that the fractional Laplacian field theory has conformal symmetry in arbitrary dimensions. In the section ~3 we will introduce a method to find the Noether currents of the symmetries in non-local field theories by mapping the non-local field theory to a local field theory. We also study the Ward identities of the translational, rotational and scale invariance. In the section ~4 we investigate the operator product expansion structure in the non-local field theories. Finally in the last section we will summarize our findings and give some comments about possible future directions.
\section{Conformal symmetry of the fractional Laplacian field theory}\
\setcounter{equation}{0}
In this section we will discuss the conformal symmetry of the fractional Laplacian field theory in general dimensions. The action of the fractional Laplacian field theory is as follows:
\begin{eqnarray}\label{fractional laplacian field theory}
S=\frac{1}{2}\int \Phi(x) (-\bigtriangleup)^{\frac{\alpha}{2}}\Phi(x)d^dx,
\end{eqnarray}
where $(-\bigtriangleup)^{\frac{\alpha}{2}}$ is the fractional laplacian defined by its Fourier transform
\begin{eqnarray}\label{fractional laplacian definition1}
\widehat{(-\bigtriangleup)^{\frac{\alpha}{2}}}\Phi(\mathbf{p})=|\mathbf{p}|^\alpha \widehat{\Phi(\mathbf{p})}.
\end{eqnarray}
Here we define the Fourier transform as
\begin{eqnarray}\label{Fourier transform}
F[f(x)]=\widehat{f(\mathbf{p})}=\int f(\mathbf{x})e^{i\mathbf{p}.\mathbf{x}}d^d\mathbf{x}.
\end{eqnarray}
The non-local nature of the fractional Laplacian can be seen in its representation in the real space, which is
\begin{eqnarray}\label{fractional laplacian definition2}
(-\bigtriangleup)^{\frac{\alpha}{2}}\Phi(x)=C(d,\alpha)\int \frac{\Phi(x)-\Phi(y)}{|x-y|^{d+\alpha}}d^dy,\hspace{1cm}0 < \alpha < 2,
\end{eqnarray}
where $C(d,\alpha)$ is a constant \cite{Samko}. The inverse of the fractional Laplacian is the Riesz potential and has the following form
\begin{eqnarray}\label{Riesz potential}
(-\bigtriangleup)^{-\frac{\alpha}{2}}\Psi(x)=D_1(d,\alpha)\int \frac{\Psi(y)}{|x-y|^{d-\alpha}}d^dy,\hspace{1cm}\hspace{1cm}0 < \alpha < d,
\end{eqnarray}
where $\Psi(x)=(-\bigtriangleup)^{\frac{\alpha}{2}}\Phi(x)$ and $D_1(d,\alpha)=\frac{\Gamma(\frac{d-\alpha}{2})}{2^\alpha \pi^{\frac{d}{2}}\Gamma(\frac{\alpha}{2})}$.
The conformal algebra in $d$ dimensions is made of $d$ generators of translation, $\frac{d(d-1)}{2}$ generators of rotation, $1$ generator of the scale transformation and $d$ generators of the special conformal transformation. It is quite simple to see that the action (\ref{fractional laplacian field theory}) has translational, rotational and scale symmetry for $0 < \alpha < 2$. The invariance of the action with respect to the special conformal transformation, which is composed of translation plus inversion, is not trivial. The transformation corresponding to inversion is $\mathbf{x}\rightarrow{\frac{\mathbf{x}}{|\mathbf{x}|^2}}$. To see that the action (\ref{fractional laplacian field theory}) is conformally invariant, we need to show that under the transformation
\begin{eqnarray}\label{conformal transform}
\mathbf{x}\rightarrow{\mathbf{x}'=g(\mathbf{x})},\hspace{1cm}\Phi'(\mathbf{x}')=|J_g|^{\frac{-d+\alpha}{2}}\Phi(\mathbf{x}),
\end{eqnarray}
where $g(\mathbf{x})$ is the conformal transformation and $|J_g|$ is the $d$th root of the Jacobian of the transformation, the action does not change. To have conformal symmetry, we need to show that
\begin{equation}\label{conformal transform1}
\frac{1}{2}\int \Phi(x) (-\bigtriangleup)^{\frac{\alpha}{2}}\Phi(x)d^dx=\frac{1}{2}\int \Phi'(x') (-\bigtriangleup')^{\frac{\alpha}{2}}\Phi'(x')d^dx',
\end{equation}
or
\begin{equation}\label{conformal transform2}
|J_g|^{\frac{\alpha+d}{2}}(-\bigtriangleup)^{\frac{\alpha}{2}} \Phi'(x)|_{x=x'} =(-\bigtriangleup)^{\frac{\alpha}{2}}(|J_g|^{\frac{-\alpha+d}{2}}\Phi'(x')).
\end{equation}
Using the Jacobian of the the inversion transformation, i.e. $|J|=\frac{1}{|\mathbf{x}|^2}$, and using the equality
\begin{eqnarray}\label{equality}
|\mathbf{x}|^{-\alpha-d}\int |\frac{\mathbf{x}}{|\mathbf{x}|^2}-\mathbf{y}|^{-\alpha-d}f(\mathbf{y})d^dy=\int |\mathbf{x}-\mathbf{y}|^{-\alpha-d}|\mathbf{y}|^{-d+\alpha}f(\frac{\mathbf{y}}{|\mathbf{y}|^2})d^dy,
\end{eqnarray}
one can easily show that equation (\ref{conformal transform2}) is valid for the inversion transformation. Checking the validity of equality (\ref{conformal transform1}) for translation, rotation and scale transformations is almost a trivial task.
The above calculation shows that the action (\ref{fractional laplacian field theory}) is conformally invariant for $0 < \alpha < d$. The definition of the case $d=\alpha$ is very similar to the previous one, we just need to consider logarithm in the definition of the Riesz potential
\begin{eqnarray}\label{Riesz potential2}
(-\bigtriangleup)^{-\frac{d}{2}}\Psi(x)=D_2(d,\alpha)\int \log|\mathbf{x}-\mathbf{y}|^{-1}\Psi(y)d^dy,
\end{eqnarray}
where $D_2(d,\alpha)=\frac{1}{(4\pi)^{\frac{d}{2}}\Gamma(\frac{d}{2})}$. The ordinary Laplacian is just $\alpha=d=2$, and it is well-known that it is fully conformally invariant. In the appendix we show that $\alpha=d=4$ is also conformally invariant. We conjecture that the action (\ref{fractional laplacian field theory}) is conformally invariant for all positive real $\alpha$'s. The field $\Phi(\mathbf{x})$ is a quasi-primary field with conformal weight $x_{\phi}=\frac{d-\alpha}{2}$; in other words, the two-point correlation function of the $\Phi(\mathbf{x})$ operator is
\begin{eqnarray}\label{correlation function}
<\Phi(x)\Phi(0)> =
\left\{
\begin{array}{lr}
D_1(d,\alpha)\frac{1}{|x|^{d-\alpha}},&0 < \alpha < d,\\
D_2(d,\alpha) \log|x|^{-1},&\alpha = d.
\end{array}
\right.
\end{eqnarray}
The operator $\Phi(x)$ is not the only operator with power law correlations, for $\alpha = d$ it is easy to see that the operator $e^{i\beta\Phi(x)}$ is also quasi-primary and
\begin{eqnarray}\label{correlation function vertex oparator}
<e^{i\beta\Phi(x)}e^{-i\beta\Phi(0)}>=\frac{1}{|x|^{\beta^2 D_2(d,\alpha)}}.
\end{eqnarray}
It is not difficult to see that to get a non-trivial correlation function one needs to fix the charges to zero, in other words
\begin{eqnarray}\label{correlation function vertex oparators}
<e^{i\alpha_1\Phi(x_1)}e^{i\alpha_2\Phi(x_2)}...e^{i\alpha_n\Phi(x_n)}>=0,
\end{eqnarray}
if $\alpha_1+\alpha_2+...+\alpha_n \neq 0$. This is a consequence of the symmetry of the action with respect to $\Phi(x)\rightarrow{\Phi(x)+a}$ ( this can be seen after partial integration). It is worth mentioning that the following action is also conformally invariant
\begin{eqnarray}\label{conformally invariant action}
S=\frac{1}{2}\int \Big{(}\Phi(x) (-\bigtriangleup)^{\frac{\alpha}{2}}\Phi(x)+(\Phi^2(x))^{\frac{d}{d-\alpha}}\Big{)}d^dx.
\end{eqnarray}
The scale invariance of the above action can be checked by simple dimensional argument.
\section{Ward identities}\
\setcounter{equation}{0}
The field theory defined by (\ref{fractional laplacian field theory}) is a non-local field theory. The non-locality makes the study of the variations of the field theory difficult. One way to overcome this difficulty is to use the local counterpart of the fractional Laplacian. This can be done by following the Caffarelli and Silvestre trick \cite{CS}, see also \cite{Chang}. The correspondence is based on the equivalence of the fractional field theory in $d$ dimensions with an ordinary Laplacian in $n=d+1$ dimensions, in other words
\begin{eqnarray}\label{local-field-theory}
S=\frac{1}{2}\int_{y>0} (\partial_\mu \tilde{\Phi}(x,y))^2y^{1-\alpha}d^dxdy=C\frac{1}{2}\int \Big{(}\Phi(x) (-\bigtriangleup)^{\frac{\alpha}{2}}\Phi(x)\Big{)}d^dx,
\end{eqnarray}
where $C$ is a constant and $0<\alpha<2$ (for an extension to the range $0<\alpha<d$ as long as $\alpha$ is not an integer see \cite{Chang}) and $\tilde{\Phi}(x,0)=\Phi(x)$. The above equality can be easily shown in momentum space \cite{CS} . It is also easy to show that $\tilde{\Phi}(x,y)$ and $\Phi(x)$ have the same two point functions in $\mathbf{x}$ space. Using the Euler-Lagrange equation of the local field theory, i.e. $\partial^\mu(y^{1-\alpha}\partial_\mu\tilde{\Phi}(x,y))=0$, one can show that
\begin{eqnarray}\label{local-nonlocal equality}
C(-\bigtriangleup)^{\frac{\alpha}{2}}\Phi(x)=\lim_{y\rightarrow{0}}y^{1-\alpha}\tilde{\Phi}_{y}(x,y)=\delta^{d}(x).
\end{eqnarray}
Motivated by the above correspondence, we study the symmetries of the local field theory instead of the non-local fractional Laplacian. First of all we notice that our field theory is translational invariant in $\mathbf{x}$ space so we will have $d$ generators of translation. In addition we have $\frac{d(d-1)}{2}$ generators of rotation in $\mathbf{x}$ space and also $d$ generators of special conformal transformations and a generator of the scale transformation. The space of the symmetries of the two field theories in the equation (\ref{local-field-theory}) is the same. The very tricky symmetry is the scale invariance which is highly dependent on the $y$ coordinate.
The conserved currents $J^{\mu}_{a}$ of the local field theory are defined by the response of the action to the infinitesimal transformation of the coordinates $x'^{\mu}=x^{\mu}+\omega_{a}\frac{\delta x'^{\mu}}{\delta\omega_{a}}(x)$ and fields $\tilde{\Phi'}(x',y')=\tilde{\Phi}(x,y)+\omega^{a}\frac{\delta \mathcal{F}}{\delta\omega_{a}}(x,y)$. For the generic field theory $S=\int \mathcal{L}[\tilde{\Phi}(x,y),\partial_{\mu}\tilde{\Phi}(x,y)]y^{1-\alpha}d^dxdy$ the transformation of the action is
\begin{eqnarray}\label{action-change}
S'-S=\int_{y>0} \partial_{\mu}J^{\mu}_{a}\omega^{a}d^dxdy,
\end{eqnarray}
where
\begin{eqnarray}\label{current}
J^{\mu}_{a}=y^{1-\alpha}\Big{(}(\delta_{\nu}^{\mu}\mathcal{L}-\frac{\partial\mathcal{L}}{\partial(\partial_{\mu}\tilde{\Phi}(x,y))}
\partial_{\nu}\tilde{\Phi}(x,y))\frac{\delta x^{\nu}}{\delta\omega_{a}} + \frac{\partial\mathcal{L}}{\partial(\partial_{\mu}\tilde{\Phi}(x,y))}\frac{\delta \mathcal{F}}{\delta\omega_{a}}\Big{)},
\end{eqnarray}
and $\mu=1,2,...,n$. From now on we will use italic letters for the quantities in the $\mathbf{x}$ space and the Greek letters for the quantities in the $(\mathbf{x},y)$ space. It is worth mentioning that to get the equation (\ref{action-change}) we do not need to do any partial integration. The conserved current associated with the translational invariance is the canonical energy-momentum tensor and has the following form:
\begin{eqnarray}\label{energy-momentum}
T^{c}_{\mu i}=y^{1-\alpha}\Big{(}\partial_{\mu}\tilde{\Phi}(x,y)\partial_{i}\tilde{\Phi}(x,y)-\frac{1}{2}\delta_{\mu i}(\partial\tilde{\Phi}(x,y))^2\Big{)},
\end{eqnarray}
where $i=1,2,...,d$. Using the equations of motion it is easy to show that $\partial^{\mu}T^{c}_{\mu i}=0$. Formally one can also define $T^{c}_{\mu y}$ by the same formula, but one should keep in mind that it is not a conserved current. Since the above energy-momentum tensor is already symmetric, one can write the associated conserved current of the rotational invariance as
\begin{eqnarray}\label{current-rotation}
j_{\mu ij}=T^{c}_{\mu i}x_{j}-T^{c}_{\mu j}x_{i}.
\end{eqnarray}
Using the equations of motion, one can again easily show that $\partial^{\mu}j_{\mu ij}=0$. The last and the most tricky conserved current is the dilation current, it can be written as
\begin{eqnarray}\label{current-dilation}
D_{\mu}=-x_{\nu}T_{\mu}^{c\nu}+y^{1-\alpha}x_{\phi} \tilde{\Phi}(x,y)\partial_{\mu}\tilde{\Phi}(x,y),
\end{eqnarray}
where $x_{\phi}=\frac{n-1-\alpha}{2}=\frac{d-\alpha}{2}$ and $\partial^{\mu}D_{\mu}=0$. Notice that the energy-momentum tensor is living in $d+1$ dimensions, so we expect that the weight of this energy-momentum tensor will be $d+1$. On the other hand, the non-local field theory is living in $d$ dimensions so to get a sensible energy-momentum tensor for our system it is reasonable to integrate $T(x,y)$ over $y$ from zero to infinity. We define
\begin{eqnarray}\label{energy-momentum tensor non-local}
t_{\mu\nu}=\int_{0}^{\infty}T^{c}_{\mu\nu}dy.
\end{eqnarray}
Using the above definition, one can write the Ward identity of the translational invariance in $\mathbf{x}$ space as
\begin{eqnarray}\label{ward identity-translation}
\partial^{j}<t_{ij}\textbf{X}>=-\sum_k \delta(x-x_k)\frac{\partial}{\partial x_{k}^{i}}<\textbf{X}>,
\end{eqnarray}
where $\textbf{X}$ is any operator living in the $\textbf{x}$ space and without any derivative over $y$. The Ward identity of the rotational symmetry is
\begin{eqnarray}\label{ward identity-rotation}
<(t_{ij}-t_{ji})\textbf{X}>=-i\sum_k \delta(x-x_k)S_{ij}^{k}<\textbf{X}>,
\end{eqnarray}
where $S_{ij}^{k}$ is the spin generator appropriate for the $k$-th field of the set $X$. The Ward identity of the scale symmetry has a form that is a bit disappointing;
\begin{eqnarray}\label{ward identity-scale}
<t_{i}^iX>&+&\int dy y<\partial^iT_{iy}\textbf{X}>+x_{\phi_k}\int dy y^{1-\alpha}<\partial^i\Big{(}\tilde{\Phi}(x,y)\partial_i\tilde{\Phi}(x,y)\Big{)}\textbf{X} >\nonumber\\&=&-\sum_k \delta(x-x_k)x_{\phi_k}<\textbf{X}>,
\end{eqnarray}
where $x_{\phi_k}$ is the weight appropriate for the $k$-th field of the set $\textbf{X}$. In usual CFT studies the second and the third terms in the left hand side disappear because there one starts with a traceless energy-momentum tensor which helps to write the conserved dilation current as $D_{\mu}=x^\nu T_{\nu\mu}$. It seems that this is not a trivial task to do in our model because making $T_{\mu\nu}$ traceless in $n$ dimensions does not mean that we can write $D_{\mu}$ fully with respect to the energy-momentum tensor. The trouble comes from the fact that the energy-momentum tensor is not conserved in the $y$ direction while $D_{\mu}$ is fully affected by also the $y$ space. In other words $\partial^\mu D_\mu=T^{\mu}_{\mu}+x^{\nu}\partial^\mu T_{\nu\mu}=T_\mu^\mu+y\partial^\mu T_{\mu y}$, where the second term is zero in the usual CFT due to the conservation of the energy-momentum tensor in the whole space. Another way to view this problem is to look at the form of the virial in the equation (\ref{ward identity-scale}), $V_i=\int dy yT_{iy}+x_{\phi}\int dy y^{1-\alpha}\Big{(}\tilde{\Phi}(x,y)\partial_i\tilde{\Phi}(x,y)\Big{)}$. Although it is possible to write the second term as the gradient of a tensor, it is impossible to do the same with the first term. This is in contrast with what one expects from the previous arguments \cite{JP,Polchinski} which state that the virial should be a gradiant of a tensor for conformally invariant models (for an example see the appendix). This means that although $t_{ij}$ carries some properties of the energy-momentum tensor in the $\textbf{x}$ space it is not exactly the same as the conventional energy-momentum tensor.
In two dimensions, it is probably impossible to find a traceless energy-momentum tensor because tracelessness means that the action is invariant under full conformal symmetry \cite{Polchinski} which we already know is not true in our case.
\section{Operator product expansion}\
\setcounter{equation}{0}
In this section, we study the leading terms of the operator product expansion of the energy-momentum tensor with the primary fields. For simplicity we first focus on the case $\alpha=1$. Although the explicit dependence on $y$ in the action and the energy-momentum tensor disappears, the original field theory is still a non-local field theory. For this case one can simply use the traceless energy-momentum tensor of the bosonic field theory in $n$ dimensions. To bring out the connection to well-known facts in bosonic field theory, we write the action as
\begin{eqnarray}\label{bosonic action}
S=\frac{1}{2}\int (\partial_\mu \tilde{\Phi}(x,y))^2d^dxdy,
\end{eqnarray}
with the following two point function
\begin{eqnarray}\label{two point function}
<\tilde{\Phi}(x,y)\tilde{\Phi}(0,0)>=\frac{1}{S_n(n-2)r^{n-2}},
\end{eqnarray}
where $r=\sqrt{\mathbf{x}^2+y^2}$ and $S_n=\frac{2\pi^{n/2}}{\Gamma[\frac{n}{2}]}$. We define the energy-momentum tensor with the following equation\footnote{The normalization is chosen to be coherent with the well-known results for $\alpha=1$, see for example \cite{Capelli}. As we will see for $\alpha\neq 1$ different normalization leads to the known OPEs.}:
\begin{eqnarray}\label{renormalized energy-momentum tensor}
\delta S=-\frac{1}{S_n}\int \partial^\mu T_{\mu\nu}\omega^\nu d^dxdy.
\end{eqnarray}
Then the traceless energy-momentum tensor is
\begin{eqnarray}\label{energy-momentum2}
\frac{T_{\mu \nu}}{S_n}=-\frac{1}{2(n-1)}\Big{(}n\partial_{\mu}\tilde{\Phi}(x,y)\partial_{\nu}\tilde{\Phi}(x,y)-(n-2)\tilde{\Phi}(x,y)\partial_{\mu}\partial_{\nu}\tilde{\Phi}(x,y)-\nonumber\\\delta_{\mu \nu}\Big{(}(\partial\tilde{\Phi}(x,y))^2-\frac{n-2}{n}\tilde{\Phi}(x,y)\partial^2\tilde{\Phi}(x,y)\Big{)}\Big{)}.
\end{eqnarray}
The operator product expansion of the above energy-momentum tensor with a primary operator is\footnote{This form is valid for generic CFT with generic primary operators \cite{Cardy2}.}
\begin{eqnarray}\label{OPEbosonic-alpha=1}
T_{\mu\nu}\tilde{\Phi}(0,0)=A_{\mu\nu}^n\tilde{\Phi}(0,0)+B_{\mu\nu\lambda}^n\partial^{\lambda}\tilde{\Phi}(0,0),
\end{eqnarray}
where
\begin{eqnarray}\label{A and B}
A_{\mu\nu}^n&=&\frac{\frac{nx_{\tilde{\phi}}}{n-1} (r_\mu r_\nu-\frac{1}{n}r^2\delta_{\mu\nu})}{r^{n+2}},\\
B_{\mu\nu\lambda}^n&=&B_{\mu\nu\lambda}^{1n}+B_{\mu\nu\lambda}^{2n},\\
B_{\mu\nu\lambda}^{1n}&=&\frac{n}{2(n-1)}\Big{(}\frac{r_\mu\delta_{\nu\lambda}+r_\nu\delta_{\mu\lambda}-(2/n)r_\lambda \delta_{\mu\nu}}{r^n}\Big{)},\\
B_{\mu\nu\lambda}^{2n}&=&
\frac{n(n-2)}{2(n-1)}\Big{(}\frac{(r_\mu r_\nu-\frac{1}{n}r^2\delta_{\mu\nu})r_\lambda}{r^{n+2}}\Big{)},
\end{eqnarray}
with $x_{\tilde{\phi}}=\frac{n-2}{2}$. The integration over $y$ gives
\begin{eqnarray}\label{OPE-alpha=1 }
t_{ij}\Phi(x)=\frac{\sqrt{\pi}d\Gamma(\frac{d}{2})}{4\Gamma(\frac{d+3}{2})}A_{ij}^d\Phi(0)+\Big{(}\frac{\sqrt{\pi}d\Gamma(\frac{d}{2})}{2\Gamma(\frac{d+1}{2})}B_{ijk}^{1d}+d(d-1)\frac{\sqrt{\pi}\Gamma(\frac{d}{2})}{4\Gamma(\frac{d+3}{2})}B_{ijk}^{2d}\Big{)}\partial^{k}\Phi(0)+...,
\end{eqnarray}
where
\begin{eqnarray}\label{A and B after integration}
A_{ij}^d&=&\frac{\frac{(d+1)x_{\phi}}{d} (r_i r_j-\frac{1}{d}r^2\delta_{ij})}{r^{d+2}},\\
B_{ijk}^{1d}&=&\frac{d+1}{2d}\Big{(}\frac{r_i\delta_{jk}+r_j\delta_{ik}-(2/d)r_k \delta_{ij}}{r^d}\Big{)}\\
B_{ijk}^{2d}&=&
\frac{(d-1)(d+1)}{2d}\Big{(}\frac{(r_i r_j-\frac{1}{d}r^2\delta_{ij})r_k}{r^{d+2}}\Big{)}.
\end{eqnarray}
To get an idea about the above results we now investigate the most interesting dimension $d=2$. Define the following energy-momentum tensors in the complex plane
\begin{eqnarray}\label{energy-momentum complex plane}
t_{zz}&=&t_{00}-2it_{10}-t_{11},\\
t_{\bar{z}\bar{z}}&=&t_{00}+2it_{10}-t_{11},\\
t_{\bar{z}z}&=&\frac{1}{4}(t_{00}+t_{11}).
\end{eqnarray}
It is easy to show that
\begin{eqnarray}\label{OPE energy-momentum complex plane}
t_{zz}\Phi(0)&=&\frac{x_\phi}{z^2}\Phi(0)+\frac{5\partial_z\Phi(0,0)}{2z}+\frac{\bar{z}\partial_{\bar{z}}\Phi(0,0)}{2z^2}+...,\\
t_{\bar{z}\bar{z}}\Phi(0)&=&\frac{x_{\phi}}{\bar{z}^2}\Phi(0)+\frac{5\partial_{\bar{z}}\Phi(0,0)}{2\bar{z}}+\frac{z\partial_z\Phi(0,0)}{2\bar{z}^2}+...,
\end{eqnarray}
where $x_\phi=\frac{d-1}{2}$. All of the terms, except the most singular one, do not have the usual CFT form (in usual CFT we have $t_{zz}\Phi(0)=\frac{x_\phi}{z^2}\Phi(0)+\frac{\partial_z\Phi(0)}{z}$). This is not surprising because as we already noticed the Ward identity of the dilation symmetry does not have the usual form. The first term is very similar to the usual CFT term, however, one should be careful that the above equation still needs to be renormalized by $C$.
The equation (\ref{OPE energy-momentum complex plane}) shows that $t_{ij}$ has some properties that we expect for the conformal field theories. $t_{ij}$ carries some important features of the energy-momentum tensor; for example, (\ref{OPE energy-momentum complex plane}) suggests that one can consider $t_{zz}(t_{\bar{z}\bar{z}})$ as a holomorphic (anti-holomorphic) operator with spin $2(-2)$. Getting the OPE of the energy-momentum tensor with itself is a non-trivial task. Doing blindly the two integrations over the $y$ space of the OPE of the two energy-momentum tensors in the (\textbf{x},y) space will give the infinity. This seems consistent with the observation in \cite{EP}, however, we think that more study in this direction is necessary.
For generic $\alpha$ one can consider the following conserved energy-momentum tensor
\begin{eqnarray}\label{energy-momentum3}
\frac{T_{\mu i}}{S_n}=-\frac{y^{1-\alpha}}{2(n-1)}\Big{(}n\partial_{\mu}\tilde{\Phi}(x,y)\partial_{i}\tilde{\Phi}(x,y)-(n-2)\tilde{\Phi}(x,y)\partial_{\mu}\partial_{i}\tilde{\Phi}(x,y)-\delta_{\mu i}(\partial\tilde{\Phi}(x,y))^2\Big{)},
\end{eqnarray}
where $\partial^\mu T_{\mu i}=0$. Note that the above energy-momentum tensor is symmetric but not traceless, i.e. $T^i_i\neq 0$. Then the OPE of $T$ with $\tilde{\Phi}$ after using (\ref{two point function}) and Wick's theorem is
\begin{eqnarray}\label{OPE energy-momentum 2}
T_{\mu \nu}\tilde{\Phi}(0,0)=\frac{(n+1-\alpha)y^{1-\alpha}}{2(n-1)}\frac{ (r_\mu r_\nu-\frac{1}{n+1-\alpha}r^2\delta_{\mu\nu})}{r^{n+3-\alpha}}\tilde{\Phi}(0,0)+....
\end{eqnarray}
After integration over $y$ one can simply write
\begin{eqnarray}\label{OPE energy-momentum 2 after integration}
t_{ij}\Phi(0)=\frac{(d-1)\Gamma(\frac{d}{2})\Gamma(1-\frac{\alpha}{2})}{4\Gamma(\frac{2+d-\alpha}{2})}\frac{ (r_i r_j-\frac{1}{d}r^2\delta_{ij})}{r^{d+2}}\Phi(0)+....
\end{eqnarray}
Renormalization of $t_{ij}$ will lead us to the well-known form of the first term of the OPE of the energy-momentum tensor and a primary field.
\section{Conclusion}\
In this paper we showed that non-local field theories can have conformal symmetry in arbitrary dimensions. This gives an example of a field theory in two dimensions which is globally conformally invariant but not fully conformally
symmetric \cite{Polchinski,Riva,slava2}. Using the local counterpart of the non-local field theory, we proposed a way to extract the Noether currents of the non-local field theories. The Ward identities of the translation, rotation and scale transformations were derived. The Ward identity of the scale symmetry does not have the usual form that we expect in CFT. In other words we have not been able to write all the conserved currents with respect to the energy-momentum tensor. We do not have a particular argument to rule out the possibility of writing all the conserved currents with respect to an improved energy-momentum tensor except in two dimensions. Finding such an improved energy-momentum-tensor is a subject for further studies in this direction. We should point out here that since our starting point was the fractional Laplacian field theory, we encountered the possibility of using the fractional derivative in the construction of the operator structure of our model. This is the reason that we have energy-momentum tensor in our construction which is absent in the earlier discussions \cite{HPPS}.
Using the energy-momentum tensor defined in the third section, we also calculated the different operator product expansions. Some terms in the operator product expansions (the most singular terms) have forms very similar to the cases in the usual CFTs, the less singular terms are quite different from the usual CFT. We have not been able to find a consistent OPE for the two energy-momentum tensors.
\textbf{Acknowledgments:}
The author is indebted to Sebastian Guttenberg for many useful comments and discussions. I also thank Malte Henkel for discussions and introducing me to the book \cite{Malte2} and Slava Rychkov for drawing my attention to the references \cite{RV,HPPS,EP,slava2}. I also thank R. Jackiw, S. Rouhani and R. T. Requist for reading the manuscript.
\section*{Appendix :Symmetric traceless energy-momentum tensor of the fourth derivative
free theory
}
In this appendix we study the symmetries of the fourth derivative free theory. The purpose of this appendix is to give an idea about the difference between the conformal symmetry in the two dimensions and higher dimensions. We show that it is possible to find a symmetric traceless energy-momentum tensor in $d>2$ but it is impossible to find a traceless-energy-momentum tensor in $d=2$ except for $\alpha=2$. The action of the fourth derivative free theory is
\begin{eqnarray}\label{fourth derivative scalar theory}
S=\int d^dx (\bigtriangleup\psi(x))^2,
\end{eqnarray}
where $\bigtriangleup\psi(x)=\partial_\mu\partial^\mu\psi(x)$. Consider the transformation (\ref{conformal transform}) with $\alpha=4$. The canonical energy-momentum tensor which is the generator of translation has the following form
\begin{eqnarray}\label{canonical energy-momentum fourth derivative scalar theory}
T^c_{\mu\nu}=-2\partial_\mu\partial_\nu\psi(x)\bigtriangleup\psi(x)+2\partial_\nu\psi\partial_\mu(\bigtriangleup\psi(x))+\delta_{\mu\nu}(\bigtriangleup\psi(x))^2.
\end{eqnarray}
The above energy-momentum tensor is not symmetric. To get a symmetric energy-momentum tensor one can use the following Belinfante tensor defined for arbitrary dimension:
\begin{eqnarray}\label{Beleinfante}
B_{\rho\mu\nu}=-2\delta_{\mu\nu}\bigtriangleup\psi(x)\partial_{\rho}\psi+2\delta_{\rho\nu}\bigtriangleup\psi(x)\partial_{\mu}\psi,
\end{eqnarray}
which is antisymmetric with respect to the first two indices. Using the above tensor one can write the Belinfante energy-momentum tensor as
\begin{eqnarray}\label{Beleinfante energy-momentum}
T^B_{\mu\nu}=T^c_{\mu\nu}+\partial^\rho B_{\mu\rho\nu}=2\partial_\nu\psi\partial_\mu(\bigtriangleup\psi(x))+\hspace{3cm}\nonumber\\\hspace{2cm}2\partial_\mu\psi\partial_\nu(\bigtriangleup\psi(x))-\delta_{\mu\nu}\Big{(}(\bigtriangleup\psi(x))^2+2\partial_\rho\psi\partial_{\rho}(\bigtriangleup\psi(x))\Big{)}.
\end{eqnarray}
Using the above energy-momentum tensor, one can simply write the generator of the rotation as
\begin{eqnarray}\label{rotation}
j_{\mu\nu\rho}= T^B_{\mu\nu}x_\rho-T^B_{\mu\rho}x_\nu.
\end{eqnarray}
The generator of the scale invariance is
\begin{eqnarray}\label{scale}
D^\mu=x^\nu T^{c\mu}_\nu-\frac{d-2}{2}\bigtriangleup\psi(x)\partial^\mu\psi+(\frac{d}{2}-3)\psi\partial^\mu \bigtriangleup\psi(x).
\end{eqnarray}
We define now the following tensor
\begin{eqnarray}\label{sigma}
\sigma^{\alpha\mu}=2\partial^\alpha\psi\partial^\mu\psi-2\psi\partial^\alpha\partial^\mu\psi+(d-2)\delta^{\alpha\mu}\psi \bigtriangleup\psi(x),
\end{eqnarray}
then the virial of the field $\psi$ will have the following form
\begin{eqnarray}\label{virial}
V^\mu=\partial_\alpha\sigma^{\alpha\mu}=d\bigtriangleup\psi(x)\partial^\mu\psi+(d-4)\psi \partial^\mu\bigtriangleup\psi(x).
\end{eqnarray}
One can easily show that
\begin{eqnarray}\label{virial}
\partial_\mu V^\mu=d(\bigtriangleup\psi(x))^2+(2d-4)\partial_\mu\psi\partial^\mu\bigtriangleup\psi(x)
\end{eqnarray}
Using the above definitions one can define the following traceless improved energy-momentum tensor \cite{CCJ} :
\begin{eqnarray}\label{traceless}
T^t_{\mu\nu}=T^B_{\mu\nu}+\frac{1}{2}\partial_{\lambda}\partial_{\rho}X^{\lambda\rho\mu\nu},
\end{eqnarray}
where
\begin{eqnarray}\label{X}
X^{\lambda\rho\mu\nu}=\frac{2}{d-2}\Big{(}\delta^{\lambda\rho}\sigma^{\mu\nu}-\delta^{\lambda\mu}\sigma^{\rho\nu}-\delta^{\lambda\nu}\sigma^{\rho\mu}+\delta^{\mu\nu}\sigma^{\lambda\rho}+\frac{1}{d-1}(\delta^{\lambda\rho}\delta^{\mu\nu}-\delta^{\lambda\mu}\delta^{\rho\nu})\sigma^\alpha_\alpha\Big{)}.
\end{eqnarray}
Then one can also write
\begin{eqnarray}\label{D}
D^\mu=x^\nu T^{t\mu}_\nu.
\end{eqnarray}
The above procedure is not repeatable for $d = 2$ because $X^{\lambda\rho\mu\nu}$ in (\ref{X})
becomes singular. This is related to the fact that the action (\ref{fourth derivative scalar theory})
does not have full conformal symmetry, as we already know. It is only invariant under M\"obius transformations. The situation is very similar to the non-local field theories that we have discussed in sections ~2 and ~3.
|
\section{Introduction}\label{sec:01}
Neutrino physics is a very active area of research which involves
some of the most intriguing problems in particle physics. The
nature of neutrinos and the origin of the small mass of neutrinos
are two examples of these kind of problems. Since neutrinos are
electrically neutral, the nature of these elementary particles can
be Majorana or Dirac fermions. The first possibility, i. e.
neutrinos being Majorana fermions was introduced by Etore Majorana
\cite{M37} when he suggested that massive neutral fermions with
specific momenta have associated only two helicity states implying
that neutrinos and anti-neutrinos are the same particles. The
second possibility implies that Dirac neutrinos are described by
four-component spinorial fields which are different from spinorial
fields describing anti-neutrinos. In this work, we will consider
neutrinos as Majorana fermions which is favored by simplicity
because they have only two degrees of freedom \cite{RM02,GK07}.
Direct and indirect experimental evidences show that neutrinos are
massive fermions with masses smaller than $1$ eV \cite{PDG}. The
most accepted way to generate neutrino masses is by mean of the
seesaw mechanism \cite{Bot}. Mass for neutrinos is a necessary
ingredient to understand the oscillations between neutrino flavor
states which have been observed experimentally \cite{PDG}.
Neutrino oscillations are originated by the interference between
mass states whose mixing generates flavor states. This phenomenon
means that a neutrino created in a weak interaction process with a
specific flavor can be detected with a different flavor. Neutrino
oscillations were first described by Pontecorvo \cite{Ponteco} as
an extension for the leptonic sector of the strange oscillations
observed in the neutral Kaon system. Neutrino oscillations can be
described in context of Quantum Mechanics
\cite{Ponteco1}-\cite{Zralec} as an application of the two level
system \cite{Sassa0}.
Description of neutrino oscillations in the context of Quantum
Field Theory (QFT) is a very well studied topic
\cite{Blasone}-\cite{Aknme}. In the literature it is possible to
find two kinds of QFT models describing neutrino oscillations:
intermediate models and external models \cite{Beuthe}. In the
framework of intermediate models Sassaroli developed a model based
in an interacting Lagrangian density which includes the coupling
between two flavor fields \cite{Sassa1,Sassa2}. This model was
framed by Beuthe as a hybrid model owing to it goes half-a-way to
QFT \cite{Beuthe}. Sassaroli model was first developed for a
coupled system of two Dirac equations \cite{Sassa1} and then it
was extended for a coupled system of two Majorana ones
\cite{Sassa2}. The probability amplitude of transition between two
neutrino flavor states for these two systems \cite{Sassa1,Sassa2}
was obtained starting from flavor states which are used on the
standard treatment of neutrino oscillations.
The standard definition of flavor states can originate some
possible limitations in the description of neutrino oscillations
as was observed by Giunti et al. \cite{Giunti2}. Specifically, in
reference \cite{Giunti2} it was shown that flavor states can
define an approximate Fock space of weak states in the following
two cases: (i) In the extremely relativistic limit, i. e. if
neutrino momentum is much larger than the maximum mass eigenvalue
of a neutrino mass state; (ii) for almost degenerated neutrino
mass eigenvalues, i. e. if the differences between neutrino mass
eigenvalues are much smaller than the neutrino momentum. The first
case leads to the standard definition of flavor states. The second
case has associated a real mixing matrix which is restricted to a
specific interaction process. Additionally these authors have
proposed that oscillations can be described appropriately for
ultra-relativistic and non-relativistic neutrinos by defining
appropriate flavor states which are superpositions of mass states
weighted by their transitions amplitudes in the process under
consideration \cite{Giunti2}.
By considering the limitations mentioned in the last paragraph it
was set down by Beuthe in \cite{Beuthe} that Sassaroli hybrid
model can only be applied consistently if lepton flavor wave
functions are considered as observable and the ultra-relativistic
limit is taken into account. On the other hand, the Sassaroli
model describing Majorana neutrino oscillations \cite{Sassa1} was
developed without considering the four-momentum conservation for
neutrinos which implies the existence of a specific momentum for
every neutrino mass state.
The main goal of this work is to study neutrino oscillations in
vacuum between two flavor states considering neutrinos as Majorana
fermions and to obtain the probability densities of transition for
left-handed neutrinos (ultra-relativistic limit) and for
right-handed neutrinos (non-relativistic limit). This work is
developed in the context of a type I seesaw scenario which leads
to get light left-handed and heavy right-handed Majorana
neutrinos. In this context, we perform an extension of the model
developed by Sassaroli in which the Majorana neutrino oscillations
are obtained for the case of flavor states constructed as
superpositions of mass states \cite{Sassa1}. Our extension
consists in considering neutrino mass states as plane waves with
specific momenta. The model that we consider in this work, which
is developed in the canonical formalism of Quantum Field Theory,
has the advantage that in the same theoretical treatment it is
possible to study neutrino oscillations for light neutrinos and
for heavy neutrinos. To do this, we first perform the canonical
quantization procedure for Majorana neutrino fields of definite
masses and then we write the neutrino flavor states as
superpositions of mass states using quantum field operators. Next
we calculate the probability amplitude of transition between two
different neutrino flavor states for the light and heavy neutrino
cases and we establish normalization and boundary conditions for
the probability density. These probability densities for the
left-handed neutrino case after the ultra-relativistic limit is
taking lead to the standard probability densities which describe
light neutrino oscillations. For the right-handed neutrino case,
the expressions describing heavy neutrino oscillations in the
non-relativistic limit are different respect to the ones of the
standard neutrino oscillations. However, the right-handed neutrino
oscillations are phenomenologically restricted as is shown when
the propagation of heavy neutrinos is considered as superpositions
of mass-eigenstate wave packets \cite{Kim}. The oscillations do
not take place in this case because the coherence is not
preserved: in other words, the oscillation length is comparable or
larger than the coherence length of the neutrino system
\cite{Kim}.
The content of this work has been organized as follows: In section
2, after establishing the Majorana condition, we obtain and solve
the two-component Majorana equation for a free fermion; in section
3, we consider a type I seesaw scenario which leads to get light
left-handed neutrinos and heavy right-handed neutrinos; in section
4, we obtain the Majorana neutrino fields with definite masses,
then we carry out the canonical quantization procedure of these
Majorana neutrino fields and we obtain relation between neutrino
flavor states and neutrino mass states using operator fields; in
section 5, we determine the probability density of transition
between two left-handed neutrino flavor states, additionally we
establish normalization and boundary conditions and then we obtain
left-handed neutrino oscillations for ultra-relativistic light
neutrinos; in section 6, we study the right-handed neutrino
oscillations for non-relativistic heavy neutrinos; finally, in
section 7 we present some conclusions.
\section{Two-component Majorana equation} \label{Majeq}
In 1937 Ettore Majorana proposed a symmetric theory for electron
and positron through a generalization of a variational principle
for fields which obey Fermi-Dirac statistics \cite{M37}. When this
theory is applied to a neutral fermion which has a specific
momentum then there exist only two helicity states. The Majorana
theory implies that it does not exist antiparticles associated to
these fermions, i. e. Majorana fermions are their own
antiparticles. For convenience we study the equation of motion for
neutral fermions but using the two-component theory developed by
Case in \cite{Case}.
In contrast with a Dirac fermion, a Majorana fermion can only be
described by a two-component spinor. To show it we consider a free
relativistic fermionic field $\psi$ of mass $m$ described by the
Dirac equation $(i\gamma^{\mu}\partial_{\mu}-m)\psi=0$, where Dirac matrixes $\gamma^\mu$
obey the anticonmutation relations
$\llav{\gamma^\mu,\gamma^\nu}=2g^{\mu\nu}$ and metric tensor
satisfies $g_{\mu\nu}=g^{\mu\nu}=\text{diag}(1,-1,-1,-1)$. Using the
chirality matrix given by $\gamma^5\equiv
i\gamma^0\gamma^1\gamma^2\gamma^3$, the left- and right-handed
chiral projections of the fermion field $\psi$ are
$\psi_{R,L}=\frac{1}{2}\corc{1\pm\gamma^5}\psi$, respectively. If we
write the Dirac matrixes projected on the chiral subspace as
$\gamma^\mu_{\pm}=\frac{1}{2}\corc{1\pm\gamma^5}\gamma^\mu$, we
obtain that the coupled equations for the chiral components of the
fermionic field $\psi$ are given by
\begin{subequations}\label{eq:EQ}
\begin{align}
i\gamma^\mu_+\partial_\mu\psi_L=m\psi_R,\label{eq:EQ1}\\
i\gamma^\mu_-\partial_\mu\psi_R=m\psi_L.\label{eq:EQ2}
\end{align}
\end{subequations}
We introduce the charge conjugation operation that will allow us to
describe Majorana fermions. The charged conjugated field (or
conjugated field) $\psi^c$ is defined as
\begin{equation}
\psi^c=\mathcal{\hat C}\bar\psi^T,
\end{equation}
where the charge conjugation operator $\mathcal{\hat C}$ satisfies
the properties $\mathcal{\hat C}(\gamma^\mu)^T\mathcal{\hat
C}^{-1}=-\gamma^\mu$, $\mathcal{\hat C}^{-1}=\mathcal{\hat C}^\dag$,
$\mathcal{\hat C}^T=-\mathcal{\hat C}$ \cite{GK07}. Using these
properties we find that the conjugated field $\psi^c$ obeys the
Dirac equation $(i\gamma^{\mu}\partial_{\mu}-m)\psi^c=0$. As $\psi$ describes a fermion with
a specific charge, its conjugated field $\psi^c$ represents a
fermion with an opposite charge and with the same mass, i. e.
$\psi^c$ describes the antifermion of $\psi$. The Dirac equation for
$\psi^c$ should be projected on the chiral subspace and for this
reason it is necessary to remember that $\mathcal{\hat
C}(\gamma^5)^T=\gamma^5\mathcal{\hat C}$ \cite{GK07}. So the coupled
equations for the chiral components of the conjugated field $\psi^c$
are
\begin{subequations}\label{eq:ECQ}
\begin{align}
i\gamma^\mu_+\partial_\mu(\psi_R)^c=m(\psi_L)^c,\label{eq:ECQ1}\\
i\gamma^\mu_-\partial_\mu(\psi_L)^c=m(\psi_R)^c.\label{eq:ECQ2}
\end{align}
\end{subequations}
We observe that the chiral components of the fermionic field $\psi$
under charge conjugation $(\psi^c)_L$, $(\psi^c)_R$ and the chiral
components of the conjugated field $(\psi_R)^c$, $(\psi_L)^c$ are
related by $(\psi_L)^c=(\psi^c)_R$, $(\psi_R)^c=(\psi^c)_L$, showing
how the charge-conjugation operation changes the chirality of
fields.
We define the Majorana condition by taking the fermionic field as
proportional to the conjugated field
\begin{align}\label{eq:CM}
\psi \equiv \xi\psi^c,
\end{align}
where the proportional constant is a complex phase factor of the
form $\xi\equiv e^{i\alpha}$ which plays an important role on
applications of Majorana theory. The equality (\ref{eq:CM}) implies
that Majorana fermions are their own antiparticles. Now the chiral
components of the Majorana field satisfy
\begin{align}\label{eq:CMQ}
\psi_L=\xi\ \mathcal{\hat C}\bar \psi_R^T,\quad \psi_R=\xi\
\mathcal{\hat C}\bar \psi_L^T.
\end{align}
So we can write equations (\ref{eq:EQ1}) and (\ref{eq:EQ2}) in the
form
\begin{subequations}\label{eq:EM}
\begin{align}
i\gamma^\mu_+\partial_\mu\psi_L&=m\xi\mathcal{\hat C}\bar \psi_L^T,\label{eq:EM1}\\
i\gamma^\mu_-\partial_\mu\mathcal{\hat C}\bar
\psi_L^T&=m\xi^*\psi_L.\label{eq:EM2}
\end{align}
\end{subequations}
If we apply the Majorana condition (\ref{eq:CMQ}) into the
equations (\ref{eq:ECQ1}) and (\ref{eq:ECQ2}), we obtain equations
(\ref{eq:EM1}) and (\ref{eq:EM2}). Additionally we can observe
that equations (\ref{eq:EM1}) and (\ref{eq:EM2}) are related to
themselves by means of a complex conjugation. In this way, we have
gone from four coupled equations describing a fermion and its
antifermion to two decoupled equations describing a left-handed
chiral field $\psi_L$ and a right-handed chiral field $\psi_R$.
Due to the fact that the right-handed chiral field can be
constructed from the left-handed chiral field \cite{Case}, as it
is shown in (\ref{eq:CMQ}), now we have only an independent field
given by $\psi_L$. For the last fact, we will be able to describe
Majorana fermion by means of field $\psi_L$ which now has two
components. To verify this sentence we rewrite equation
(\ref{eq:EM1}) as
\begin{align}\label{eq:EQM1}
i\mathcal{\hat
C}\gamma^0\gamma^\mu_+\partial_\mu\psi_L=m\xi\psi_L^*.
\end{align}
If we define
\begin{equation}\label{eq:MM}
\eta^\mu \equiv\mathcal{\hat C}\gamma^0\gamma^\mu_+,
\end{equation}
and if we take $\phi\equiv\psi_L$, then equation (\ref{eq:EQM1}) can
be written as
\begin{align}\label{eq:EMDC1}
i\eta^\mu\partial_\mu\phi=m\xi\phi^*,
\end{align}
which is known as the Majorana equation. This equation in which a
particle is indistinguishable from its antiparticle has two
components because the matrixes $\gamma^\mu_\pm$ are projected on
the chiral subspaces of two components. The matrixes $\eta^\mu$ are
called Majorana matrixes and these should not be confused with the
Dirac matrixes written in Majorana representation.
Now we are interested in knowing the kind of relations that
Majorana matrixes $\eta^\mu$ obey. So we first apply definition
(\ref{eq:MM}) into equation (\ref{eq:EM2}) and we obtain
$-i\eta^{\mu*}\partial_\mu\phi^*=m\xi^*\phi$, with
$\eta^{\mu*}=\gamma^\mu_-\gamma^0\mathcal{\hat C}$. Then we apply
$-i\eta^{\nu*}\partial_\nu$ into (\ref{eq:EMDC1}) and we have
$\eta^{\nu*}\eta^\mu\partial_\nu\partial_\mu\phi-m^2\phi=0 $ or
its equivalent
$\frac{1}{2}\llav{\eta^{\mu*}\eta^\nu+\eta^{\nu*}\eta^\mu}\partial_
\nu\partial_\mu\phi-m^2\phi=0$, where we have used
$\abs{\xi}^2=1$. Accordingly, Majorana matrixes should satisfy
relations $\eta^{\mu*}\eta^\nu+\eta^{\nu*}\eta^\mu=-2g^{\mu\nu}$
and then the field $\phi$ is satisfying the Klein-Gordon equation
given by $\corc{\partial^\mu\partial_\mu+m^2}\phi=0$.
In this work we have taken a particular representation of matrixes
$\eta^\mu$ which has permitted us to write the two-component
Majorana equation in the form given by (\ref{eq:EMDC1}). Now we
can consider a matrix $A$ which satisfies the following relations
\cite{Case}
\begin{align}\label{eq:PA}
A\sigma^{i*}A^{-1}=-\sigma^i,\quad A=A^\dag=A^{-1}=-A^T,
\end{align}
where $\sigma^i$ represents Pauli matrixes in a given
representation. We take $\eta^\mu=i\sigma_2\bar{\sigma}^\mu$,
where $\bar{\sigma}^\mu\equiv(\mathbb{I},-\vec{\mathbf{\sigma}})$
being $\mathbb{I}$ the unit matrix $2\times 2$ and
$\vec{\mathbf{\sigma}}=(\sigma_1,\sigma_2,\sigma_3)$ Pauli
matrixes. Since $\sigma_2$ satisfies properties (\ref{eq:PA}), we
have taken $A\equiv\sigma_2$. So the equation (\ref{eq:EMDC1}) can
be written as
\begin{align}\label{eq:EMDC}
i\bar{\sigma}^\mu\partial_\mu\phi+im\xi\sigma_2\phi^*=0.
\end{align}
This equation is the well known two-component Majorana equation
\cite{Pal,Marsch}, which will be solved in next subsection.
\subsection{Canonical quantization for Majorana field}
With the purpose of studying the canonical quantization for the
Majorana field we will obtain the free-particle solution of
equation (\ref{eq:EMDC}). On the outset, we consider bi-spinors
$\chi$ which obey the following relations
\begin{subequations}
\begin{align}
\frac{\vec{\mathbf{\sigma}}\cdot\vec{p}}{\abs{\vec{p}}}
\chi^h(\vec{p})&=h\chi^h(\vec{p}),\\
-i\sigma_2(\chi^h(\vec{p}))^*&=h\chi^{-h}(\vec{p}),
\end{align}
\end{subequations}
where these bi-spinors correspond to helicity eigenstates. If we
take the momentum in spherical coordinates
$\vec{p}=\abs{\vec{p}}(\sin\theta\cos\varphi, \sin\theta\sin\varphi,
\cos\theta)$, then the helicity operator has the form
\begin{align}
\frac{\vec{\mathbf{\sigma}}\cdot\vec{p}}{\abs{\vec{p}}}=\begin{pmatrix}
\cos\theta & \sin\theta e^{-i\varphi}\\
\sin\theta e^{i\varphi} & -\cos\theta
\end{pmatrix}.
\end{align}
We choose the following representation for these bi-spinors
\begin{align}
\chi^+(\vec{p})=\begin{pmatrix}
\cos\frac{\theta}{2}\\
\sin\frac{\theta}{2}e^{i\varphi}
\end{pmatrix},\quad
\chi^-(\vec{p})=\begin{pmatrix}
-\sin\frac{\theta}{2}e^{-i\varphi}\\
\cos\frac{\theta}{2}
\end{pmatrix}.
\end{align}
We can prove that the following solution satisfies the two-component
Majorana equation (\ref{eq:EMDC})
\begin{align} \label{eq:SeM}
\phi^h_{\vec{p}}(x)=\sqrt{\frac{E-h\abs{\vec{p}}}{2E}}\chi^h(\vec{p})
e^{-ipx}-h\sqrt{\frac{E+h\abs{\vec{p}}}{2E}}\chi^{-h}(\vec{p})e^{ipx},
\end{align}
with $px\equiv p_\mu x^\mu=Et-\vec{p}\cdot\vec{x}$. We observe that
Majorana field can be written as superposition of positive and
negative energy states.
The Lagrangian density which describes a free two-component
Majorana field is given by
\begin{align} \label{eq:Mld}
\mathcal{L}_M=\phi^{\dag}i\bar\sigma^{\mu}\partial_{\mu}\phi-
\frac{m\xi}{2}\corc{\phi^{T}i\sigma_2\phi-\phi^{\dag}i\sigma_2\phi^*},
\end{align}
where the two-component Majorana field $\phi$ and its conjugated
field $\phi^{\dag}$ behave as Grassmann variables. It is very easy
to prove that the two-component Majorana equation (\ref{eq:EMDC})
can be obtained from the Lagrangian density (\ref{eq:Mld}) using
the Euler-Lagrange equation. Additionally, we can obtain the
following energy-momentum tensor from (\ref{eq:Mld}), $
T_{\mu\nu}=\phi^{\dag}i\bar\sigma_{\mu}\partial_{\nu}\phi\notag
-g_{\mu\nu}\esp{\phi^{\dag}i\bar\sigma^{\lambda}\partial_{\lambda}
\phi-\frac{m}{2}\corc{\phi^{T}i\sigma_2\phi-\phi^{\dag}i\sigma_2\phi^*}}
$. Following the standard canonical quantization procedure, we now
consider the Majorana field $\phi$ and its conjugated field
$\phi^{\dag}$ as operators which satisfy the usual canonical
anticonmutation relations given by $
\llav{\hat\phi_{\alpha}(\vec{r},t),\hat\phi_{\beta}(\vec{r}\,',t)}=
\llav{\hat\phi^{\dag}_{\alpha}(\vec{r},t),\hat\phi^{\dag}_{\beta}(\vec{r}\,',t)}=0$,
$
\llav{\hat\phi_{\alpha}(\vec{r},t),\hat\phi^{\dag}_{\beta}(\vec{r}\,',t)}=
\delta_{\alpha\beta}\delta^3(\vec{r}-\vec{r}\,')$, where
$\alpha,\beta=1,2$. Using the Heisenberg equation for the Majorana
field $
i\partial_t\hat\phi_\alpha(\vec{r},t)=\esp{\hat\phi_\alpha(\vec{r},t),\hat{
H}}$, we can obtain its corresponding Majorana equation
(\ref{eq:EMDC}). By means of the energy-momentum tensor it is
possible to prove that the Hamiltonian operator can be written as
$\hat{H}=\int\
d^3x\corc{\hat{\phi}^{\dag}i\vec{\mathbf{\sigma}}\cdot
\nabla\hat{\phi}+\frac{m\xi}{2}\corc{\hat{\phi}^{T}i\sigma_2\hat{\phi}-
\hat{\phi}^{\dag}i\sigma_2\hat{\phi}^*}}$. The expansion in a
Fourier series for the Majorana field operator is
\cite{Giunti2}-\cite{Case}
\begin{align}\label{eq:EONM}
\hat\phi(x)=\int\frac{d^3p}{(2\pi)^{3/2}(2E)^{1/2}}&\sum_{h=\pm1}
\left[\sqrt{E-h\abs{\vec{p}}}\,\des{a}{}(\vec{p},h)\chi^{h}(\vec{p})
e^{-ip\cdot x}\right.\notag\\ &\qquad\quad\left.-h\sqrt{E+h\abs{\vec{p}}}\,
\crea{a}{}(\vec{p},h)\chi^{-h}(\vec{p})e^{ip\cdot x}\right],
\end{align}
where we have used the free-particle solution (\ref{eq:SeM}) and
operators $\des{a}{}(\vec{p},h)$, $\crea{a}{}(\vec{p},h)$ which
satisfy the anticonmutation relations
$\llav{\des{a}{}(\vec{p},h),\des{a}{}(\vec{p}\,',h')}=
\llav{\crea{a}{}(\vec{p},h),\crea{a}{}(\vec{p}\,',h')}=0$,
$\llav{\des{a}{}(\vec{p},h),\crea{a}{}(\vec{p}\,',h')}=
\delta_{h,h'}\delta^3(\vec{p}-\vec{p}\,')$. Then we can identify
$\des{a}{}(\vec{p},h)$ as the annihilation operator and
$\crea{a}{}(\vec{p},h)$ as the creation operator of a Majorana
fermion with momentum $\vec{p}$ and helicity $h$.
\section{Masses for Majorana neutrino fields}
The most accepted way to generate neutrino masses is through the
seesaw mechanism. In this section we consider a type I seesaw
scenario which leads to get light left-handed neutrinos and heavy
right-handed neutrinos. For the case of two neutrino generations,
a Dirac-Majorana mass term is given by \cite{Bilpet}
\begin{equation}\label{eq:DMT}
\mathcal{L}^{M+D}_Y=\frac{1}{2} \bar N_L \mathcal{\hat C}
M^{M+D}_\nu N_L + H.c.,
\end{equation}
where $H.c.$ represents the hermitic conjugate term, $N_L$ is the
vector of flavor neutrino fields written as
\begin{align}\label{eq:NL1}
N_L=\begin{pmatrix}
\nu_L\\
\mathcal{\hat C} \bar \nu^T_R
\end{pmatrix}
=\begin{pmatrix}
\nu_L\\
\nu^c_R
\end{pmatrix},
\end{align}
where $\nu_L$ represents a doublet of left-handed neutrino fields
active under the weak interaction and $\nu^c_R$ represents a
doublet of right-handed Majorana neutrino fields non active
(sterile) under the weak interaction. These doublets are given by
\begin{align}\label{eq:NL2}
\nu_L=\begin{pmatrix}
\nu_{e_L}\\
\nu_{\mu_L}
\end{pmatrix};
\,\,\,\nu^c_R=\begin{pmatrix}
\nu^c_{e_R}\\
\nu^c_{\mu_R}
\end{pmatrix}.
\end{align}
In the Dirac-Majorana term (\ref{eq:DMT}), $M^{M+D}_\nu$ is a
$4\times4$ non-diagonal matrix of the form
\begin{align}\label{eq:MMMnu}
M^{M+D}_\nu=\begin{pmatrix}
M^{\prime}_L & (M_D)^T\\
M_D & M^{\prime}_R
\end{pmatrix},
\end{align}
where $M_L$, $M_R$ and $M_D$ are $2\times2$ matrixes. The vector
of neutrino fields with definite masses $n_L$ can be written by
mean of a unitary matrix $U^\nu_L$ as follows
\begin{equation}\label{eq:FTUM}
N_L=U^\nu_L n_L,
\end{equation}
where $n_L$ has the form
\begin{align}\label{eq:NL3}
n_L=\begin{pmatrix}
n_{1}\\
n_{2}\\
\end{pmatrix}=\begin{pmatrix}
\nu_{1}\\
\nu_{2}\\
\nu_{3}\\
\nu_{4}
\end{pmatrix}.
\end{align}
The unitary matrix $U^\nu_L$ is chosen in such a way that the
non-diagonal matrix $M^{M+D}_\nu$ can be diagonalized through the
similarity transformation
\begin{equation}\label{eq:MMdiag}
(U^\nu_L)^{-1} M^{M+D}_\nu U^\nu_L = M_\nu,
\end{equation}
where $M_\nu$ is a diagonal matrix which is defined by
$(M_\nu)_{ab} = m_a \delta_{ab}$, where $a,b=1,2,3,4$. The masses
of the neutrino fields of definite masses $\nu_{a}$ are $m_a$,
with $a=1,2,3,4$.
The seesaw scenario is established imposing the following
conditions into the matrix (\ref{eq:MMMnu}): $M^{\prime}_L = 0$,
$(M_D)_{kj} \ll (M^{\prime}_R)_{kj}$, thus the matrix
$M^{M+D}_\nu$ is diagonalized as
\begin{equation}\label{eq:MMdiag}
(U^\nu_L)^{-1} M^{M+D}_\nu U^\nu_L = \begin{pmatrix}
M_{l}& 0\\
0 & M_{h}
\end{pmatrix},
\end{equation}
where $M_{l}$ is the light neutrino mass matrix and $M_{h}$ is the
heavy neutrino mass matrix. If the unitary matrix $U^\nu_L$ is
expanding considering terms until of the order $(M^\prime_R)^{-1}
M_D$, the light and heavy neutrino mass matrixes can be written as
\begin{equation}\label{eq:MMlh}
M_{l}=\begin{pmatrix}
m_1 & 0\\
0 & m_2
\end{pmatrix};
\,\,\,M_{h}=\begin{pmatrix}
m_3 & 0\\
0 & m_4
\end{pmatrix}.
\end{equation}
The Dirac-Majorana mass term (\ref{eq:DMT}) can be written in
terms of the neutrino fields of definite masses $\nu_{a}$ as
\begin{equation}\label{eq:DMT1}
\mathcal{L}^{M+D}_Y=\frac{1}{2} \bar n_{1} \mathcal{\hat C} M_{l}
n_{1}+\frac{1}{2} \bar n_{2} \mathcal{\hat C} M_{h} n_{2}+ H.c.,
\end{equation}
where the matrixes $M_l$ and $M_h$ are given by (\ref{eq:MMlh})
and the doublets $n_{1_L}$ and $n_{2_L}$ are written as
\begin{align}\label{eq:NL4}
n_{1}=\begin{pmatrix}
\nu_{1}\\
\nu_{2}
\end{pmatrix};
\,\,\,n_{2}=\begin{pmatrix}
\nu_{3}\\
\nu_{4}
\end{pmatrix}.
\end{align}
The neutrino fields of definite masses $\nu_{1}$ and $\nu_{2}$
have associate respectively the light masses $m_1 \sim
m_e^2/(f_{11} v_R)$ and $m_2 \sim m_\mu^2/(f_{22} v_R)$ and the
neutrino fields of definite masses $\nu_{3}$ and $\nu_{4}$ have
associate respectively the heavy masses $m_3 = f_{33} v_R$ and
$m_4 = f_{44} v_R$, where $v_R \rightarrow \infty$, $f_{ab}$ are
Yukawa couplings, $m_e$ is the electron mass and $m_\mu$ is the
muon mass.
As it will be shown in the next section, starting from the
Dirac-Majorana mass term
\begin{equation}\label{eq:DMT2}
\mathcal{L}^{M+D}_Y=\frac{1}{2} \bar \nu_L \mathcal{\hat C} M_{L}
\nu_L+\frac{1}{2} \bar \nu^c_R \mathcal{\hat C} M_{R} \nu^c_R+
H.c.,
\end{equation}
where $\nu_L$ and $\nu^c_R$ are the flavor doublets of
non-definite masses given by (\ref{eq:NL2}), while $M_L$ and $M_R$
are $2\times2$ non-diagonal matrixes, it will be possible to
obtain the Dirac-Majorana mass term (\ref{eq:DMT1}) after the
diagonalization of the matrixes $M_L$ and $M_R$.
\section{Mass and flavor neutrino states}
In the next we suppose that the Majorana fields $\nu_{e_L}$ and
$\nu_{\mu_L}$ describe the active light left-handed neutrinos that
are produced and detected in the laboratory, while the Majorana
fields $\nu^c_{e_R}$ and $\nu^c_{\mu_R}$ describe the sterile
heavy right-handed neutrinos which there exist in a type I seesaw
scenario.
In the section (\ref{Majeq}) we have presented a lagrangian
density (\ref{eq:Mld}) which describes a free Majorana fermion.
This lagrangian density can be extended to describe a system of
two flavor left-handed neutrinos and two flavor right-handed
neutrinos with non-definite masses. Using the Dirac-Majorana mass
term given by (\ref{eq:DMT2}), the lagrangian density describing
this system is given by
\begin{align}\label{eq:MAC}
\mathcal{L}=\bar \nu_L i\bar\sigma^{\mu}\partial_{\mu}\nu_L+ \bar
\nu^c_R i\bar\sigma^{\mu}\partial_{\mu}\nu^c_R-\frac{1}{2} \bar
\nu_L M_L i\sigma_2 \nu_L -\frac{1}{2} \bar \nu^c_R M_R i\sigma_2
\nu^c_R + H.c.
\end{align}
where the non-diagonal mass matrixes $M_L$ and $M_R$ are written
as
\begin{align}\label{eq:MMML}
M_L=\begin{pmatrix}
m_{\nu_{e_L}}e^{i\delta_1} & m_{\nu_{e_L}\nu_{\mu_L}}e^{i(\delta_1+\delta_2)}\\
m_{\nu_{e_L}\nu_{\mu_L}}e^{i(\delta_1+\delta_2)} & m_{\nu_{\mu_L}} e^{i\delta_2}
\end{pmatrix},
\end{align}
\begin{align}\label{eq:MMMR}
M_R=\begin{pmatrix}
m_{\nu_{e_R}}e^{i\delta_3} & m_{\nu_{e_R}\nu_{\mu_R}}e^{i(\delta_3+\delta_4)}\\
m_{\nu_{e_R}\nu_{\mu_R}}e^{i(\delta_3+\delta_4)} & m_{\nu_{\mu_R}} e^{i\delta_4}
\end{pmatrix},
\end{align}
We observe that the form of the matrixes $M_L$ and $M_R$ is the
same. In the next, we will restrict to the left-handed Majorana
neutrinos, but the results are directly extended to the
right-handed Majorana neutrinos. From the Euler-Lagrange equations
we obtain that the coupled equation of motion for the flavor
left-handed neutrino fields $\nu_{e_L}$ and $\nu_{\mu_L}$ are
\begin{subequations}\label{eq:MAC2}
\begin{align}
i\bar\sigma^{\lambda}\partial_{\lambda}
\nu_{e_L}&=-im_{\nu_{e_L}}\sigma_{2}e^{i\delta_1}\nu_{e_L}^{*}-im_{\nu_{e_L}\nu_{\mu_L}}
\sigma_{2}e^{i(\delta_1+\delta_2)}\nu_{\mu_L}^{*},\\
i\bar\sigma^{\lambda}\partial_{\lambda}\nu_{\mu_L}&=-im_{\mu_L}
\sigma_{2}e^{i\delta_2}\nu_{\mu}^{*}-im_{\nu_{e_L}\nu_{\mu_L}}\sigma_{2}e^{i(\delta_1+\delta_2)}
\nu_{e_L}^{*},
\end{align}
\end{subequations}
respectively. We observe that flavor neutrino fields are coupled
by means of the parameter $m_{\nu_{e_L}\nu_{\mu_L}}$. With the
purpose of decoupling the equations of motion for the flavor
left-handed neutrino fields, now we consider the most general
unitary matrix $U_L$ given by
\begin{align}
U_L=\frac{1}{\sqrt{1+\Lambda_L^2}}\begin{pmatrix}
\Lambda_L e^{-i\frac{\delta_1}{2}} &
e^{-\frac{i}{2}(\delta_1+\alpha_L)}\\
-e^{-i\frac{\delta_2}{2}} & \Lambda_L e^{-\frac{i}{2}
(\delta_2+\alpha_L)}
\end{pmatrix},
\end{align}
where the phases $e^{-i\delta_1}$ and $e^{-i\delta_2}$ appear as a
consequence of the Majorana condition (\ref{eq:CMQ}). The
definite-mass neutrino field doublet $n_1$ given by (\ref{eq:NL4})
is related to the flavor left-handed neutrino doublet $\nu_L$
given by (\ref{eq:NL2}) by mean of
\begin{equation}\label{eq:CMN}
\nu_L=U_L n_1.
\end{equation}
Without a lost of generality, we can change the phases of the
flavor left-handed neutrino fields by means of
$\nu_{e_L}\rightarrow \exp\{-\frac{i\delta_1}{2}\}\nu_{e_L}$ and
$\nu_{\mu_L}\rightarrow \exp\{-\frac{i\delta_2}{2}\}\nu_{\mu_L}$.
Thus there is just a phase $\exp\{-\frac{i\alpha_L}{2}\}$ that can
not be eliminated. So the matrix $U$ can be rewritten as
\begin{align}
U_L=\frac{1}{\sqrt{1+\Lambda_L^2}}\begin{pmatrix}
\Lambda_L & e^{-\frac{i\alpha_L}{2}}\\
-1 & \Lambda_L e^{-\frac{i\alpha_L}{2}}
\end{pmatrix}.
\end{align}
Now the diagonalization of the mass matrix (\ref{eq:MMML}) given
by
\begin{align}\label{eq:DONM}
M_{D_L}=U_L^\dag M_L U_L=\text{diag}(m_1,m_2),
\end{align}
is valid for
\begin{equation} \label{eq:Lambda}
\Lambda_L =\frac{m_{\nu_{\mu_L}}-m_{\nu_{e_L}}+R_L}{2m_{\nu_{e_L}\nu_{\mu_L}}},\quad
{\rm with}\quad R_L^2=(m_{\nu_{e_L}}-m_{\nu_{\mu_L}})^2+4m_{\nu_{e_L}\nu_{\mu_L}}^2,
\end{equation}
and thus the neutrino fields with definite masses $\nu_1$ and
$\nu_2$ have respectively the following masses
\begin{align}
m_{1}&=\frac{1}{2}(m_{\nu_{e_L}}+m_{\nu_{\mu_L}}-R_L),\\
m_{2}&=\frac{1}{2}(m_{\nu_{e_L}}+m_{\nu_{\mu_L}}+R_L)e^{-i\alpha_L}.
\end{align}
We observe that in the expression for $m_2$ appears the factor
$\exp\{-i\alpha_L\}$ which suggest that this mass could be
complex. However the diagonalization given by (\ref{eq:DONM}) is
not completely right because $M_L$ is a symmetric matrix. So from
(\ref{eq:DONM}) the diagonalization should be of the form
\begin{align}
&M_{D_L}^2=U_L^\dag M_L^\dag M_L U_L,
\end{align}
where we have considered that this matrix is hermitic, i. e.
$M_L^2\equiv M_L^\dag M_L$. So the values $m_1$ and $m_2$ are the
quadratic roots of the eigenvalues of $M_L^2$. This last result
implies that these eigenvalues can be multiplied by a complex
phase.
The expression (\ref{eq:CMN}) gives the mixing of the flavor
neutrino fields in terms of the neutrino fields with definite
masses. The neutrino fields with definite masses $\nu_1$ and
$\nu_2$ obey Majorana field equations of the form
\begin{subequations}
\begin{align}
i\bar\sigma^{\mu}\partial_{\mu}\nu_{1}&=-im_{1}\sigma_{2}\nu_{1}^{*},\\
i\bar\sigma^{\mu}\partial_{\mu}\nu_{2}&=-im_{2}e^{-i\alpha}\sigma_{2}\nu_{2}^{*}.
\end{align}
\end{subequations}
With the purpose of eliminating the phase $\alpha_L$ from the last
equation of motion, we can make the following phase transformation
$\nu_2\rightarrow \exp\{-\frac{i\alpha_L}{2}\}\nu_2$. Now the
unitary matrix can be written as
\begin{align}\label{eq:MUD}
U_L=\frac{1}{\sqrt{1+\Lambda_L^2}}\begin{pmatrix}
\Lambda_L & e^{-i\alpha_L}\\
-1 & \Lambda_L e^{-i\alpha_L}
\end{pmatrix}.
\end{align}
We observe that the phase $\exp\{-i\alpha_L\}$ was eliminated from
the last equation of motion but not from the unitarian matrix
$U_l$. So it proves that this phase is physical and should be
involved in some processes. This phase could play an important
role in the case of doublet beta decay process.
Following a similar procedure for the right-handed Majorana
neutrinos, we find that the definite-mass neutrino field doublet
$n_2$ given by (\ref{eq:NL4}) is related to the flavor
right-handed neutrino doublet $\nu^c_R$ given by (\ref{eq:NL2}) by
mean of
\begin{equation}\label{eq:CMNR}
\nu^c_R=U_R n_2,
\end{equation}
where the unitary matrix $U_R$ is given by
\begin{align}\label{eq:MUDR}
U_R=\frac{1}{\sqrt{1+\Lambda_R^2}}\begin{pmatrix}
\Lambda_R & e^{-i\alpha_R}\\
-1 & \Lambda_R e^{-i\alpha_R}
\end{pmatrix}.
\end{align}
where $\Lambda_R$ is given by
\begin{equation} \label{eq:LambdaR}
\Lambda_R =\frac{m_{\nu^c_{\mu_R}}-m_{\nu^c_{e_R}}+R_R}{2m_{\nu^c_{e_R}\nu^c_{\mu_R}}},\quad
{\rm with}\quad
R_R^2=(m_{\nu^c_{e_R}}-m_{\nu^c_{\mu_R}})^2+4m_{\nu^c_{e_R}\nu^c_{\mu_R}}^2.
\end{equation}
Next we consider the canonical quantization of the
neutrino fields with definite mass by stating the anticonmutation
relations given by
$\llav{\des{\nu}{a}(\vec{r},h),\crea{\nu}{b}(\vec{r}',h')}=
\delta_{ab}\delta^3(\vec{r}-\vec{r}')$,
$\llav{\des{\nu}{a}(\vec{r},h),\des{\nu}{b}(\vec{r}',h')}= 0$ and
$\llav{\crea{\nu}{a}(\vec{r},h),\crea{\nu}{b}(\vec{r}',h')}= 0$,
where $a,b=1,2,3,4$ represent neutrino mass states. Each one of
the definite-mass neutrino field operators $\hat\nu_a(x)$ obeys a
Majorana equation. It is possible to expand each one of these
field operators on a plane-wave basis set as was shown in
(\ref{eq:EONM})
\begin{align}\label{eq:FEOFDM}
\hat\nu_a(x)=\int
\frac{d^3p}{(2\pi)^{3/2}(2E_a)^{1/2}}&\sum_{h=\pm1}\left[\sqrt{E_a-h
\abs{\vec{p}}}\,\des{a}{a}(\vec{p},h)\chi^{h}(\vec{p})e^{-ip\cdot
x}\right.\notag\\
&\qquad\quad\left.-h\sqrt{E_a+h\abs{\vec{p}}}\,\crea{a}{a}(\vec{p},h)
\chi^{-h}(\vec{p})e^{ip\cdot
x}\right],
\end{align}
where $E_a^2=\abs{\vec{p}}^2+m_a^2$ is the energy of the neutrino
field with definite mass which is tagged by $a=1,2,3,4$.
The flavor neutrino field operators tagged by $\hat \nu_\alpha$
are defined as superposition of the definite-mass neutrino field
operators $\hat \nu_a$ given by (\ref{eq:FEOFDM}) through the
expression
\begin{align}
\des{\nu}{\alpha}(x)=\sum_aU_{\alpha a}\des{\nu}{a}(x),
\end{align}
where $U$ is the unitarian matrix defined by (\ref{eq:MUD}) for
left-handed neutrinos and by (\ref{eq:MUDR}) for right-handed
neutrinos, meanwhile neutrino flavor states $\ket{\nu_\alpha}$ are
defined in terms of the neutrino mass states $\ket{\nu_a}$ as
\begin{align}\label{eq:DefEstSab}
\ket{\nu_\alpha}=\sum_aU^*_{\alpha a}\ket{\nu_a}.
\end{align}
Thus we have found a relation between neutrino flavor states and
neutrino mass states using operator fields. As flavor states are
physical states since they could be detected in interaction
processes, flavor states are non-stationary. So their temporal
evolution gives the probability of transition between them.
Therefore, this probability describes Majorana neutrino
oscillations studied as follows.
\section{Left-handed neutrino oscillations}
Now we will focuss our interest in the description of left-handed
neutrino oscillations in vacuum from a cinematical point of view.
For this reason we will not consider in detail the weak
interaction processes involved in the creation and detection of
left-handed neutrinos. However, these processes are manifested
when boundary conditions are imposed in the probability amplitude
of transition between two neutrino flavor states. We suppose that
a neutrino with a specific flavor is created in a point of
space-time $x_0^\mu=(t_0,\vec{r}_0)$ as a result of a certain weak
interaction process. We will determine the probability amplitude
to find out the neutrino with another flavor in a different point
of space-time $x^\mu=(t,\vec{r})$. We assume that neutrinos are
created under the same production process with different values of
energy and momentum. These dynamical quantities are related among
themselves under the specific production process.\\
The initial left-handed neutrino flavor state in the production
time ($t_0$) corresponds to the following superposition of
neutrino mass states
\begin{align}
\ket{\nu_L(t_0)}=A\ket{\nu_1}+B\ket{\nu_2}.
\end{align}
where $\abs{A}^2+\abs{B}^2=1$. Each of these neutrino mass states
has associated a specific four-momentum. We assume that in the
production point it was created a left-handed electronic neutrino
with each massive field having a four-momentum given by
$p_a^\mu=(E_a,\vec{p}_a)$, with $a=1,2$. The initial left-handed
electronic neutrino state satisfying the condition
$\abs{A}^2+\abs{B}^2=1$ is written as
\begin{align}\label{eq:EI}
\ket{\nu_L(t_0)}=\sum_{h=\pm 1}\frac{\Lambda_L
}{\sqrt{1+\Lambda_L^2}}\ket{\nu_1}+\frac{e^{-i\alpha_L}}{\sqrt{1+
\Lambda_L^2}}\ket{\nu_2},
\end{align}
where the sum over helicities is taken over the neutrino mass
states. This sum over helicities must be considered to describe
appropriately the initial left-handed neutrino flavor state
because the helicity is a property which is not directly measured
in the experiments. The manner as the electronic left-handed
neutrino state has been built in the production point is in
agreement with the experimental fact that left-handed neutrinos
are ultra-relativistic.
The neutrino mass states involve in the superposition given by
(\ref{eq:EI}) are obtained from the vacuum state as
$\ket{\nu_a}=e^{ip_ax_0}\crea{a}{a}(\vec{p}_a,h)\ket{0}$, where we
have included the phase factor $\exp\{ip_ax_0\}$. This phase
factor gives us information about the four-space time where the
left-handed neutrino was created. The probability amplitudes for
transitions to electronic and muonic left-handed neutrinos are
respectively given by
\begin{align} \label{eq:nuapt}
\nu_{e_L}(X)&\equiv\bra{0}\des{\nu}{e_L}(x)\ket{\nu_L(t_0)}\notag\\
&=\frac{1}{(2\pi)^{3/2}}\sum_{h}\left\{
\frac{\Lambda_L^2}{1+\Lambda_L^2}\sqrt{\frac{E_1-h\abs{\vec
p}_1}{2E_1}} e^{-ip_1X}\chi^h(\vec{p}_1)
+\frac{1}{1+\Lambda_L^2}\sqrt{\frac{E_2-h \abs{\vec
p}_2}{2E_2}}e^{-ip_2X}\chi^h(\vec{p}_2)\right\},
\end{align}
\begin{align} \label{eq:eapt}
\nu_{\mu_L}(X)&\equiv\bra{0}\hat\nu_{\mu_L}(x)\ket{\nu_L(t_0)}\notag\\
&=\frac{1}{(2\pi)^{3/2}}\sum_{h}\left\{
-\frac{\Lambda_L}{1+\Lambda_L^2}\sqrt{\frac{E_1-h\abs{\vec
p}_1}{2E_1}} e^{-ip_1X}\chi^h(\vec{p}_1)
+\frac{\Lambda_L}{1+\Lambda_L^2} \sqrt{\frac{E_2-h\abs{\vec
p}_2}{2E_2}}e^{-ip_2X}\chi^h(\vec{p}_2)\right\},
\end{align}
where we have used some expansions over the Majorana fields and we
have taken $X\equiv x-x_0$ which corresponds to a four-vector
associated to the distance and time of neutrino propagation. The
probability densities $\rho_{\nu_\alpha
L}(X)=\abs{\nu_{\alpha_L}(X)}^2$ respectively are
\begin{align} \label{eq:nudpt}
\rho_{\nu_{eL}}(X)&=\frac{1}{(2\pi)^{3}}\frac{1}{(1+\Lambda_L^2)^2}\times\notag\\
&\quad\left\{1+\Lambda_L^4+\frac{\Lambda_L^2}{(E_1E_2)^{1/2}}
\sum_h\sqrt{(E_1-h\abs{\vec p}_1)(E_2-h\abs{\vec p}_2)}\right.
\notag\\&\quad \left.-\frac{(2\Lambda_L)^2}{2(E_1E_2)^{1/2}}
\sum_h\sqrt{(E_1-h\abs{\vec p}_1)(E_2-h\abs{\vec p}_2)}
\sin^2\corc{\frac{p_1-p_2}{2}X}\right\},
\end{align}
\begin{align}\label{eq:edpt}
\rho_{\nu_{\mu L}}(X)&=\frac{1}{(2\pi)^{3}}\frac{1}{(1+\Lambda_L^2)^2}\times\notag\\
&\quad\left\{2\Lambda_L^2-\frac{\Lambda_L^2}{(E_1E_2)^{1/2}}\sum_h
\sqrt{(E_1-h\abs{\vec p}_1)(E_2-h\abs{\vec p}_2)}\right.
\notag\\&\quad
\left.+\frac{(2\Lambda_L)^2}{2(E_1E_2)^{1/2}}\sum_h\sqrt{(E_1-h\abs{\vec
p}_1)(E_2-h\abs{\vec
p}_2)}\sin^2\corc{\frac{p_1-p_2}{2}X}\right\},
\end{align}
where we have taken into account that spinors $\chi^h(\vec{p}_1)$
and $\chi^h(\vec{p}_2)$ are the same because vectors $\vec{p}_1$ and
$\vec{p}_2$ are co-linear. The probability densities
(\ref{eq:nudpt}) and (\ref{eq:edpt}) that we have found present a
serious problem. If we fix $X=0$ into (\ref{eq:nudpt}) and
(\ref{eq:edpt}) we find that
\begin{align} \label{eq:nudptxo}
\rho_{\nu_{eL}}(X=0)=\frac{1}{(2\pi)^{3}}\frac{1}{(1+\Lambda_L^2)^2}
\left\{1+\Lambda_L^4+\frac{\Lambda_L^2}{(E_1E_2)^{1/2}}\sum_h
\sqrt{(E_1-h\abs{\vec p}_1)(E_2-h\abs{\vec p}_2)}\right\},
\end{align}
\begin{align}\label{eq:edptxo}
\rho_{\nu_{\mu L}}(X=0)=\frac{1}{(2\pi)^{3}}\frac{1}{(1+\Lambda_L^2)^2}
\left\{2\Lambda_L^2-\frac{\Lambda_L^2}{(E_1E_2)^{1/2}}\sum_h
\sqrt{(E_1-h\abs{\vec p}_1)(E_2-h\abs{\vec p}_2)}\right\},
\end{align}
and we observe that the probability density (\ref{eq:edptxo}) can
be different from zero, i. e. it can exist a muonic neutrino in
the production point which disagrees with the initial conditions.
The origin of this problem is related to the weak state definition
\eqref{eq:DefEstSab} that we have used before. As it was
previously mentioned into the introduction, the flavor definition
\eqref{eq:DefEstSab} is not complectly consistent and it is
necessary to define appropriate flavor states \cite{Giunti2}.
\subsection{Ultra-relativistic limit: Left-handed neutrino oscillations}
This problem can be solved by taking an approximation in the
probability densities (\ref{eq:nudpt}) and (\ref{eq:edpt}) based
on the fact that left-handed neutrinos are ultra-relativistic
particles because their masses are very small. Here we consider
energy and momentum different for every mass state. In general we
can write
\begin{align}\label{eq:CEM1}
\abs{\vec p}_a^2&=E^2-\xi m_a^2+\zeta m_a^4,
\end{align}
\begin{align}\label{eq:CEM2}
E_a^2&=E^2+(1-\xi) m_a^2+\zeta m_a^4,
\end{align}
where the parameters $\xi$ and $\zeta$ are determined in the
production process and $E$ is the energy for the case in which
neutrinos were massless. For instance, for the pion decay process we
have
\begin{align}
E&=\frac{m_\pi}{2}\corc{1-\frac{{m'}_{\mu}^2}{m_\pi^2}},\\
\xi&=\frac{1}{2}\corc{1+\frac{{m'}_{\mu}^2}{m_\pi^2}},\qquad
\zeta=\frac{1}{4m_\pi^2},
\end{align}
where ${m'}_\mu$ is the muon mass and $m_\pi$ is the pion mass.
Because for the ultra-relativistic limit $m_a\rightarrow 0$, we can
approximate the expressions (\ref{eq:CEM1}) and (\ref{eq:CEM2}) to
\begin{align}\label{eq:AUR}
\abs{\vec p}_a&\approx E-\xi\frac{m_a^2}{2E},\\
E_a&\approx E+(1-\xi)\frac{m_a^2}{2E}.
\end{align}
Now it is possible to prove that the right side of the relation
\begin{align}\label{eq:AFS}
\frac{1}{2(E_1E_2)^{1/2}}\sum_{h=\pm1}\sqrt{(E_1-h\abs{\vec p}_1)
(E_2-h\abs{\vec p}_2)}\approx 1-\frac{(m_1-m_2)^2}{8E^2},
\end{align}
can be approximated to the unit because $(\delta
m_{12}/E)^2\approx 0$, where $\delta m_{12}= m_1-m_2$. On the
other hand, neutrino propagation time $T$ is not measured in
neutrino experiments \cite{GK07,Beuthe,Giunti1}. In this kind of
experiments is measured the distance $L$ between the neutrino
source and the detector. By this reason, it can be possible to
find a analytical expression that establishes a relation between
$T$ and the propagation distance $L=\abs{\vec{L}}$. In our
approach using plane waves, for the ultra-relativistic limit we
can write
\begin{align}\label{eq:RTL}
L\approx T.
\end{align}
This relation implies that the propagation distance and the
propagation time for neutrinos are approximately equal because in
the ultra-relativistic limit a neutrino mass state has a mass too
small and its velocity of propagation $v_k$ is approximately equal
to speed velocity $c=1$, i. e. $v_k \approx 1$. However, a most
precise relation between $L$ and $T$ must be described by an
expression that should include explicitly the velocities of the
two neutrino mass states involved in such a way that this
expression for the ultra-relativistic limit should lead to
\eqref{eq:RTL}.\\
So for the ultra-relativistic limit the probability densities
(\ref{eq:nudpt}) and (\ref{eq:edpt}) can be written as
\begin{align}\label{eq:nudpt2}
\rho_{\nu_{e L}}(L)&=\frac{1}{(2\pi)^{3}}\left\{1-\corc{\frac{2\Lambda_L}
{1+\Lambda_L^2}}^2\sin^2\corc{\frac{\Delta m_{12}^2}{4E}L}\right\},
\end{align}
\begin{align}\label{eq:edpt2}
\rho_{\nu_{\mu L}}(L)&=\frac{1}{(2\pi)^{3}}\corc{\frac{2\Lambda_L}
{1+\Lambda_L^2}}^2\sin^2\corc{\frac{\Delta m_{12}^2}{4E}L},
\end{align}
where we have used $L\approx T$ and $\Delta m_{12}^2\equiv
m_1^2-m_2^2$. Under this approximation it is clear that these
probability density does not depend from the production process
due to that there is no dependence from $\xi$. Thus these
probability densities satisfy the boundary conditions that we have
imposed.
In the next we will prove that the probability densities
(\ref{eq:nudpt2}) and (\ref{eq:edpt2}) have the form of the
standard probability densities for neutrino oscillations. In the
context of the standard formalism of neutrino oscillations
(assuming CP conservation), for the two generation case
considering here, the representation of the unitary matrix $U_L$
that appears into the expression (\ref{eq:CMN}) is given by
\cite{CK93}
\begin{align}\label{eq:MUDSR}
U_L=\begin{pmatrix}
\cos\theta_L & \sin\theta_L\\
-\sin\theta_L & \cos\theta_L
\end{pmatrix},
\end{align}
where $\theta_L$ is the mixing angle. If we compare the unitary
matrix given by (\ref{eq:MUD}) with the one given by
(\ref{eq:MUDSR}), we observe that $\cos\theta_L = \Lambda_L
/(\sqrt{1+\Lambda_L^2})$ and then it is very easy to obtain that
\begin{equation}\label{eq:thlamb}
\sin 2\theta_L=\frac{2\Lambda_L}{1+\Lambda_L^2}.
\end{equation}
Substituting (\ref{eq:thlamb}) into (\ref{eq:nudpt2}) and
(\ref{eq:edpt2}), we obtain the expressions
\begin{align}\label{eq:nudpt3}
\rho_{\nu_{e L}}(L)&=\frac{1}{(2\pi)^{3}}\left\{1-\sin^2(2\theta_L)
\sin^2\corc{\frac{\Delta m_{12}^2}{4E}L}\right\},
\end{align}
\begin{align}\label{eq:edpt3}
\rho_{\nu_{\mu L}}(L)&=\frac{1}{(2\pi)^{3}}\sin^2(2\theta_L)
\sin^2\corc{\frac{\Delta m_{12}^2}{4E}L},
\end{align}
which are the standard probability densities for left-handed
neutrino oscillations in the two flavor case \cite{CK93}.
\section{Right-handed neutrino oscillations}
The initial right-handed neutrino flavor state in the production
time ($t_0$) corresponds to the following superposition of
neutrino mass states
\begin{align}
\ket{\nu^c(t_0)}=C\ket{\nu_3}+D\ket{\nu_4}.
\end{align}
where $\abs{C}^2+\abs{D}^2=1$. Each of these neutrino mass states
has associated a specific four-momentum. We assume that in the
production point it was created a right-handed electronic neutrino
with each massive field having a four-momentum given by
$p_a^\mu=(E_a,\vec{p}_a)$, with $a=3,4$. The initial right-handed
electronic neutrino state satisfying the condition
$\abs{C}^2+\abs{D}^2=1$ is written as
\begin{align}\label{eq:EIR}
\ket{\nu^c_R(t_0)}=\sum_{h=\pm 1}\frac{\Lambda_R
}{\sqrt{1+\Lambda_R^2}}\ket{\nu_3}+\frac{e^{-i\alpha_R}}{\sqrt{1+
\Lambda_R^2}}\ket{\nu_4},
\end{align}
where the sum over helicities is taken over the neutrino mass
states.
The probability amplitudes for transitions to electronic and
muonic right-handed neutrinos are respectively given by
\begin{align} \label{eq:nuaptR}
\nu^c_{e_R}(X)&\equiv\bra{0}\des{\nu^c}{e_R}(x)\ket{\nu^c_R(t_0)}\notag\\
&=\frac{1}{(2\pi)^{3/2}}\sum_{h}\left\{
\frac{\Lambda_R^2}{1+\Lambda_R^2}\sqrt{\frac{E_3-h\abs{\vec
p}_3}{2E_3}} e^{-ip_3X}\chi^h(\vec{p}_3)
+\frac{1}{1+\Lambda_R^2}\sqrt{\frac{E_4-h \abs{\vec
p}_4}{2E_4}}e^{-ip_4X}\chi^h(\vec{p}_4)\right\},
\end{align}
\begin{align} \label{eq:eaptR}
\nu^c_{\mu_R}(X)&\equiv\bra{0}\hat\nu^c_{\mu_R}(x)\ket{\nu^c_R(t_0)}\notag\\
&=\frac{1}{(2\pi)^{3/2}}\sum_{h}\left\{
-\frac{\Lambda_R}{1+\Lambda_R^2}\sqrt{\frac{E_3-h\abs{\vec
p}_3}{2E_3}} e^{-ip_3X}\chi^h(\vec{p}_3)
+\frac{\Lambda_R}{1+\Lambda_R^2} \sqrt{\frac{E_4-h\abs{\vec
p}_4}{2E_4}}e^{-ip_4X}\chi^h(\vec{p}_4)\right\}.
\end{align}
The probability densities $\rho_{\nu_\alpha
R}(X)=\abs{\nu^c_{\alpha_R}(X)}^2$ respectively are
\begin{align} \label{eq:nudptR}
\rho_{\nu^c_{eR}}(X)&=\frac{1}{(2\pi)^{3}}\frac{1}{(1+\Lambda_R^2)^2}\times\notag\\
&\quad\left\{1+\Lambda_R^4+\frac{\Lambda_R^2}{(E_3E_4)^{1/2}}
\sum_h\sqrt{(E_3-h\abs{\vec p}_3)(E_4-h\abs{\vec p}_4)}\right.
\notag\\&\quad \left.-\frac{(2\Lambda_R)^2}{2(E_3E_4)^{1/2}}
\sum_h\sqrt{(E_3-h\abs{\vec p}_3)(E_4-h\abs{\vec p}_4)}
\sin^2\corc{\frac{p_3-p_4}{2}X}\right\},
\end{align}
\begin{align}\label{eq:edptR}
\rho_{\nu^c_{\mu R}}(X)&=\frac{1}{(2\pi)^{3}}\frac{1}{(1+\Lambda_R^2)^2}\times\notag\\
&\quad\left\{2\Lambda_R^2-\frac{\Lambda_R^2}{(E_3E_4)^{1/2}}\sum_h
\sqrt{(E_3-h\abs{\vec p}_3)(E_4-h\abs{\vec p}_4)}\right.
\notag\\&\quad
\left.+\frac{(2\Lambda_R)^2}{2(E_3E_4)^{1/2}}\sum_h\sqrt{(E_3-h\abs{\vec
p}_3)(E_4-h\abs{\vec
p}_4)}\sin^2\corc{\frac{p_3-p_4}{2}X}\right\},
\end{align}
where we have taken into account that spinors $\chi^h(\vec{p}_3)$
and $\chi^h(\vec{p}_4)$ are the same because vectors $\vec{p}_3$
and $\vec{p}_4$ are co-linear. The probability densities
(\ref{eq:nudptR}) and (\ref{eq:edptR}) that we have found present
a serious problem. If we fix $X=0$ into (\ref{eq:nudptR}) and
(\ref{eq:edptR}) we find that
\begin{align} \label{eq:nudptxoR}
\rho_{\nu^c_{e
R}}(X=0)=\frac{1}{(2\pi)^{3}}\frac{1}{(1+\Lambda_R^2)^2}
\left\{1+\Lambda_R^4+\frac{\Lambda_R^2}{(E_3E_4)^{1/2}}\sum_h
\sqrt{(E_3-h\abs{\vec p}_3)(E_4-h\abs{\vec p}_4)}\right\},
\end{align}
\begin{align}\label{eq:edptxoR}
\rho_{\nu_{\mu R}}(X=0)=\frac{1}{(2\pi)^{3}}\frac{1}{(1+\Lambda_R^2)^2}
\left\{2\Lambda_R^2-\frac{\Lambda_R^2}{(E_3E_4)^{1/2}}\sum_h
\sqrt{(E_3-h\abs{\vec p}_3)(E_4-h\abs{\vec p}_4)}\right\},
\end{align}
and we observe that the probability density (\ref{eq:edptxoR}) can
be different from zero.
\subsection{Non-relativistic limit: Right-handed neutrino oscillations}
This problem can be solved by taking an approximation in the
probability densities (\ref{eq:nudptR}) and (\ref{eq:edptR}) based
on the fact that right-handed neutrinos are non-relativistic
particles because their masses are very large. By this reason, we
take the non-relativistica approximation, i. e. $m_a\gg p_a$. So
we have
\begin{align}
E_a\approx m_a+\frac{\abs{\vec p}_a^2}{2m_a}.
\end{align}
Therefore, we suppose that heavy right-handed Majorana neutrinos
obey simply the relativistic dispersion relation. So we obtain the
following approximation
\begin{align}
\frac{1}{2(E_3E_4)^{1/2}}\sum_{h=\pm1}\sqrt{(E_3-h\abs{\vec
p}_3)(E_4-h\abs{\vec p}_4)}=
1+\frac{v_3v_4}{2}-\frac{1}{8}(v_3^2+v_4^2)(v_3+v_4)^2+\cdots,
\end{align}
where the non-relativistic velocity of the neutrino is $v_i\equiv
\frac{\abs{\vec p}_i}{m_i}\approx 0$, meanwhile the phase is
approximated to
\begin{align}
(E_3-E_4)T-(p_3-p_4)L\approx\Delta m_{34}T,
\end{align}
with $\Delta m_{34}\equiv m_3-m_4$. So the probability densities
of transition are given by
\begin{align}
\rho_{\nu^c_{e R}}(T)&=\frac{1}{(2\pi)^{3}}\left\{1-\corc{\frac{2\Lambda_R}
{1+\Lambda_R^2}}^2\sin^2\corc{\frac{\Delta m_{34}}{2}T}\right\},
\label{eq:rhMnpde}\\
\rho_{\nu^c_{\mu R}}(T)&=\frac{1}{(2\pi)^{3}}\corc{\frac{2\Lambda_R}{1+
\Lambda_R^2}}^2\sin^2\corc{\frac{\Delta m_{34}}{2}T}\label{eq:rhMnpdmu},
\end{align}
where $\Lambda_R$ is given by (\ref{eq:LambdaR}). The last
probability densities satisfy the normalization and boundary
conditions. Unlikely to the case of left-handed neutrino
oscillations described by (\ref{eq:nudpt}) and (\ref{eq:edpt}),
the argument of the periodic function for the right-handed
neutrino oscillations depends on the linear mass difference
$\Delta m_{34}$ and the propagation time $T$. The description of
heavy right-handed neutrino oscillations that we present here
could be of interest in cosmological problems \cite{Gershtein}. As
it has been proposed in the literature, heavy-heavy neutrino
oscillations could be responsible for the baryon asymmetry of the
universe through a leptogenesis mechanism
\cite{AkhRubaSmir,Volkas}. But it should be noted that if the
propagation of heavy right-handed neutrinos is considered as
superpositions of mass-eigenstate wave packets \cite{Kim}, then
the oscillations do not take place because the coherence is not
preserved: in other words, the oscillation length is comparable or
larger than the coherence length of the right-handed neutrino
system \cite{Kim}.
\section{Conclusions}
In this work we have studied neutrino oscillations in vacuum
between two flavor states considering neutrinos as Majorana
fermions. We have performed this study for the case of flavor
states constructed as superpositions of mass states extending the
Sassaroli model which describes Majorana neutrino oscillations by
considering neutrino mass states as plane waves with specific
momenta. In the context of a type I seesaw scenario which leads to
get light left-handed and heavy right-handed Majorana neutrinos,
the main contribution of this work has been to obtain in a same
formalism the probability densities which describe the
oscillations for light left-handed neutrinos (ultrarelativistic
limit) and for heavy right-handed neutrinos (non-relativistic
limit). In this work we have performed the canonical quantization
procedure for Majorana neutrino fields of definite masses and then
we have written the neutrino flavor states as superpositions of
mass states using quantum field operators. We have calculated the
probability amplitude of transition between two different neutrino
flavor states for the light and heavy neutrino cases and we have
established normalization and boundary conditions for the
probability density. After the ultra-relativistic limit was taken
in the probability densities for the left-handed neutrino case
lead to the standard probability densities which describe light
neutrino oscillations. For the right-handed neutrino case, the
expressions describing heavy neutrino oscillations in the
non-relativistic limit were different respect to the ones of the
standard neutrino oscillations. However, the right-handed neutrino
oscillations are phenomenologically restricted as is shown when
the propagation of heavy neutrinos is considered as superpositions
of mass-eigenstate wave packets \cite{Kim}.
This work establish a framework to study Majorana neutrino
oscillations for the case where mass states are described by
Gaussian wave packets as will be presented in a forthcoming work
\cite{PQ1}. The wave packet treatment is necessary owing to the
neutrinos are produced in weak interaction processes without a
specific momenta. Additionally the plane wave treatment can not
describe production and detection localized processes as occur in
neutrino oscillations.
\section*{Acknowledgments}
C. J. Quimbay thanks DIB by the financial support received through
the research project "Propiedades electromagn\'eticas y de
oscilaci\'on de neutrinos de Majorana y de Dirac". C. J. Quimbay
thanks also Vicerrector\'ia de Investigaciones of Universidad
Nacional de Colombia by the financial support received through the
research grant "Teor\'ia de Campos Cu\'anticos aplicada a sistemas
de la F\'isica de Part\'iculas, de la F\'isica de la Materia
Condensada y a la descripci\'on de propiedades del grafeno".
|
\section{Introduction}
``How do massive stars form?'' is still a debated question (e.g.,
\citealt[]{Beuther07}). One basic problem is massive protostars become
so luminous that radiation pressure may stop the accretion and growth
of the star. One possible mechanism to form a massive star can be
considered as a scaled-up version of low-mass star formation. If the
gas is turbulent and threaded by sufficiently strong magnetic fields
then fragmentation may be suppressed for cores much more massive than
the Jeans mass. The conditions leading to this suppression should be
relatively rare since massive stars are rare and make up only a small
mass fraction of the final star cluster. These massive cores are
expected to form from highly pressurized clumps of gas, in which case
they start with high densities, short free-fall times and therefore
high accretion rates. This is the basic scenario of the Turbulent Core
Model (\citealt[]{MT03}, hereafter MT03).
Other radically different possibilities are that massive stars form
through stellar mergers (\citealt[]{Bonnell98}) or accrete most of
their mass from initially unbound material (competitive accretion
model, \citealt[]{Bonnell01}, \citealt[]{Bonnell04},
\citealt{Bate09a,Bate09b}).
Answering this question is difficult observationally because massive
star formation occurs in distant and highly obscured regions. However,
with the improving sensitivity and spatial resolution provided by the
instruments such as the {\it Herschel Space Telescope}, {\it
Stratospheric Observatory for Infrared Astronomy (SOFIA)}, {\it Gran
Telescopio Canarias (GTC) - CanariCam}, the {\it James Webb Space
Telescope} and Thirty-Meter class telescopes, we expect to see a
faster advance in the research of massive star formation and hope this
question can be finally solved.
In order to interpret observations such as spectral energy
distributions (SEDs) or images, and then have a better understanding
of the properties of a massive protostar and its evolution, a number
of models have already been developed to fit or compare with
observations. For example, \citet[]{Robitaille06} have developed a
very impressive model grid containing 200,000 SEDs covering a large
parameter space, which is publicly available and now widely used
(referred to below as the ``Robitaille model''). However, this grid
mainly covers low-mass young stellar objects (YSOs) and when massive
protostars are considered, they do not have the properties expected
for fiducial parameters of the turbulent core model. Also these models
do not consider the presence of an optically thick gaseous disk inside
the dust destruction front. \citet[]{Molinari08} used the same methods
employed by the Robitaille model but now for protostellar parameters
based on the turbulent core model, to developed a SED model grid for
massive YSOs. They found that the SED can be a diagnostic tool to
determine the evolutionary stage of a massive YSO. However, in this
work the representation of the turbulent core model is quite
approximate and limitations of the Robitaille model framework are
still present. \citet[]{CM05} developed an analytic solution for
far-IR SED of a protostar embedded in a spherically symmetric
molecular cloud, but this method could not allow for presence of
accretion disks and protostellar outflow cavities.
\citet[]{Indebetouw06} investigated the effects of clumpy structures
in the molecular envelope around massive protostars. However, again,
their models were not tuned to the parameters of the turbulent core
model: for example, they considered much more massive structures
(e.g. $\sim 5\times 10^4\;M_\odot$ contained in a sphere of radius
2.5 pc), more representative of star forming clumps that form entire
star clusters. They did not consider protostellar disks.
Our aim is thus to develop a new model of massive protostars, based on
the Turbulent Core Model of \citet{MT02} and MT03, including all
the important components self-consistently, and then perform
simulations of the radiation transfer to see whether different
components or evolutionary stages are represented in the SEDs and
images. Starting with a fiducial hydrostatic core of $60\;M_\odot$
bounded by the pressure of a self-gravitating clump of mean mass
surface density $\Sigma_{\rm cl}=1\:{\rm g\:cm^{-2}}$, we then
consider its appearance once an $8\;M_\odot$ protostar has formed at its
center. We develop a series of protostellar models of increasing
realism: starting with a simple hydrostatic core, we then apply the
inside-out expansion wave solution (\citealt[]{Shu77}), but
generalized to singular polytropic spheres (\citealt[]{MP97}) and the
free-fall rotating collapse solution by \citet[]{Ulrich}. A
circumstellar disk is expected to form around the protostar and this
is important to transfer angular momentum and to solve the radiation
pressure problem (e.g. \citealt[]{jijina96}, \citealt[]{Krumholz07}).
Presently, there is evidence for rotating toroids around massive
protostars (e.g. \citealt[]{Beltran05}) but little direct evidence for
Keplerian protostellar disks, which will probably require the angular
resolution of ALMA. We include the disk with an $\alpha$-disk model
(\citealt[]{SS73}). Unlike \citet[]{Robitaille06} and
\citet[]{Molinari08}, we include an optically thick inner disk with
gas opacities inside the dust destruction radius, which is important
to conserve the accretion energy naturally. Accretion is expected to
drive strong bipolar outflows and sweep up the material in the core
and form cavities. These outflow cavities have been observed around
massive protostars and may determine the mid-IR morphology (e.g.,
\citealt[]{deBuizer06}).
The assumptions of each component of our model and the model series
are introduced in detail in the next section. In
Section. \ref{sec:simulation}, we discuss our simulations, including
the Monte Carlo radiation transfer code, and the dust and gas
opacities we use. In Section. \ref{sec:results}, we present the SED
and image results of our models. In Section. \ref{sec:conclusions}, we
summarize our main results, including a comparison of our model with
other works. In future papers we will examine additional refinements,
especially the development and material content of the outflow
cavities, and present results of the evolutionary sequence of massive
star formation based on the fiducial model we have started to develop
here.
\section{Massive Protostar Model}
\subsection{Envelope}
\subsubsection{Hydrostatic Outer Envelope}
\label{sec:core}
Following MT03 and \citet[]{Tan08}, we define a
``star-forming core'' as a region of a molecular cloud that will form
a single star or close binary, and assume it is self-similar,
self-gravitating in near virial equilibrium and spherical. The density
and pressure each have power-law dependencies on radius, $\rho\propto
r^{-k_\rho}$ and $P\propto r^{-k_p}$. This smooth power-law density
distribution is only an approximation, especially given that massive
cores in virial equilibrium but with gas temperatures $\sim 10\;{\rm
K}$ must be supported by non-thermal forms of pressure support, such
as turbulence and/or magnetic fields. Clumpy substructures are likely
to form inside a turbulent core and this will affect radiative
transfer through the core - basically making it somewhat easier for
shorter wavelength photons to propagate through the
core. \citet[]{Indebetouw06} performed radiative transfer models of
clumpy cores. However, given the uncertain nature of the clumping, we
defer such considerations to a future paper, and first calculate the
properties of smoothly distributed gas and dust.
From the above power law distributions, it follows that the core is
polytropic with $P\propto \rho^{\gamma_p}$. The case with $\gamma_p=1$
is a singular isothermal sphere (\citealt[]{Shu77}) and in the other
limit $\gamma_p=0$ corresponds to a logotropic sphere
(\citealt{MP96,MP97}). For the models presented here,
we follow MT03 and adopt $k_\rho=1.5$, thus $\gamma_p=\frac{2}{3}$ and
$k_p=1$. The equation of hydrostatic equilibrium gives
\begin{equation}
M(r)=\frac{k_pc^2r}{G},
\end{equation}
and
\begin{equation}
\rho(r)=\frac{(3-k_\rho)(k_\rho-1)c^2}{2\pi G r^2}, \label{den}
\end{equation}
where $c=(P/\rho)^{1/2}$ is the effective sound speed.
We assume the total mass in the core is 60 $M_\odot$. If the efficiency
$\epsilon_{*f}$ is 0.5, which is estimated from low-mass cases
(\citealt[]{MM00}), a star with mass
$m_{*f}=\epsilon_{*f}
M_\mathrm{core}=30\;M_\odot$ can finally form out of the core. For a
core with such mass, the core radius is (eq. (20) in MT03)
\begin{equation}
R_\mathrm{core}=0.057\Big(\frac{M_\mathrm{core}}{60\;M_\odot}\Big)^{1/2}\Sigma_\mathrm{cl}^{-1/2}\;\mathrm{pc},
\end{equation}
where $\Sigma_\mathrm{cl}$ is the mean surface density in the
molecular clump in which the core is embedded, and $1\;
\mathrm{g/cm}^{-2}$ is used as a fiducial value. We also consider
models with higher and lower values of $\Sigma_{\rm cl}$.
\subsubsection{Expansion Wave}
\label{sec:expansionwave}
\begin{figure*}
\begin{center}
\includegraphics[width=\columnwidth]{radialprofile.eps}
\includegraphics[width=\columnwidth]{polarprofile.eps}\\
\caption{Density distribution with radius $r$ (left panel) and polar angle
$\theta$ (right panel). $n_\mathrm{He}=0.1n_\mathrm{H}$ is assumed here.
In the left panel, different curves correspond to
different polar angles, which is $0^\circ$ when along the rotation axis. The thick black
curve is without any rotation effect for comparison. The right panel shows the density dependence on $\theta$ at $r=2r_d$ and $r=4r_d$.}
\label{fig:rotation}
\end{center}
\end{figure*}
\citet[]{Shu77} developed the inside-out expansion-wave solution for
the problem of the gravitational collapse of an isothermal
sphere. \citet[]{MP97} applied this solution to the collapse of a
logotropic sphere and also gave the general formulae for a polytropic
sphere. Following their work, we calculate the density profile in the
core. A core with initial density profile of eq. (\ref{den}) is not
stable and a perturbation at the center can trigger collapse of the
innermost region. The position where the material begins to fall
progresses outward, which is called the expansion wave. While the
material collapses inside the expansion wave front, the region outside
it is still hydrostatic. The similarity is always kept during the
whole collapse in this solution.
\citet[]{MP97} did not include the effect of magnetic fields which
can help to support more mass in the core and increase the accretion
rate after collapse starts. We estimate the effect of magnetic fields
from the work of \citet[]{LS97}, who found that the equilibrium
surface density is increased by a factor of $(1+H_0)$ when magnetic
fields are considered, where $H_0$ is a parameter. So the real mass
and density profile should increased by the same factor,
\begin{equation}
M(r,t)=M_{\mathrm{non}}(r,t)(1+H_0)=m(x)(1+H_0)a_t^3t/G,
\end{equation}
and
\begin{equation}
\rho(r,t)=\rho_{\mathrm{non}}(r,t)(1+H_0)=\frac{\alpha(x)(1+H_0)}{4\pi Gt^2},
\end{equation}
where $m(x)$ and $\alpha(x)$ are the similarity variables for mass and
density, and $a_t=[K\gamma_p(4\pi Gt^2)^{1-\gamma_p}]^{1/2}$ has
dimension of velocity. In our case, $H_0$ is set to 1 following the
assumption by MT03.
The collapsed mass at the center now is
\begin{equation}
M(0)=m(0)(1+H_0)a_t^3t/G,
\end{equation}
therefore,
\begin{equation}
M(0)\propto a_t^3t \propto t^{4-3\gamma_p},
\end{equation}
and
\begin{equation}
\dot{M}(0)\propto t^{3-3\gamma_p}\propto t
\end{equation}
in our case, which is consistent with the analysis of
MT03. So, given a certain collapsed mass $M(0,t)$, with the
star formation time (the time that the whole core takes to collapse,
eq. (44) in MT03)
\begin{equation}
t_{*f}=1.29\times 10^5
\Big(\frac{M_\mathrm{core}}{60\;M_\odot}\Big)^{1/4}\Sigma_{\mathrm{cl}}^{-3/4}\;\mathrm{yr}
\end{equation}
and the final collapsed mass $M(0,t=t_{*f})=M_\mathrm{core}=60\;M_\odot$
(The whole core collapses at the end, either into the star-disk system or into the outflow
and escapes), we can
calculate $t$ and further the density $\rho(r,t)$, velocity
$u(r,t)$ and mass $M(r,t)$ at that moment. At some time the
expansion wave will reach the boundary of the core and lead to a backward wave, thus,
for simplicity we choose such a
collapsed mass that the expansion wave has not reached the core boundary yet.
(We will discuss this in detail in Section \ref{sec:star}.)
\subsubsection{Rotating Infall}
\label{sec:rotating}
We consider a slowly rotating core. For simplicity, we only include
the effect of rotation inside the sonic point, where the infall
becomes supersonic. Once such infall starts, we expect it is difficult
to transfer angular momentum from inside the sonic point to outside
(\citealt[]{TM04}). So we assume that the angular momentum is
conserved inside the sonic point until gas accretes onto the star or
disk.
We use the solution of \citet[]{Ulrich} (referred to below as the
``Ulrich solution'') to describe the velocity field and density profile
inside the sonic point. This solution assumes that a particle with an
initial distance $r_\infty$ from the center, an initial polar angle
$\theta_0$ and angular velocity $\Omega_\infty$ about the axis of
rotation, due to a point mass $M$ at the center, moves in a parabolic
path described by
\begin{equation}
r=\frac{r_d\cos\theta_0\sin^2\theta_0}{\cos\theta_0-\cos\theta}, \label{streamline}
\end{equation}
where
\begin{equation}
r_d=\frac{\Omega_\infty^2r_\infty^4}{GM}=\frac{j_\infty^2}{GM}.
\end{equation}
All particles starting from a spherical shell of radius $r_\infty$
will hit the equatorial plane at radius $r<r_d$, so naturally $r_d$
can be thought of as the radius of the accretion disk.
$r_d$ is also the centrifugal radius which marks the radial extent of
centrifugal balance, inside which the flow will become disk-like
(e.g. \citealt[]{jijina96}). It can be estimated as below:
\begin{equation}
r_d=\frac{6a_g}{5a_i}\beta r=\frac{6a_g}{5a_i}\beta
\Big(\frac{M}{M_\mathrm{core}}\Big)^\frac{1}{3-k_\rho}, \label{rd}
\end{equation}
where
$a_g=5|E_\mathrm{grav}|r/(3GM^2)=(5/3)(3-k_\rho)/(5-2k_\rho)\rightarrow1.25$,
$a_i=2E_\mathrm{rot}/(Mr^2\Omega^2)=(2/3)(3-k_\rho)/(5-k_\rho)\rightarrow
0.286$, and $\beta =E_\mathrm{rot}/|E_\mathrm{grav}|$ is set to have a
fiducial value of 0.02 for the slowly rotating core, based on
observations of low-mass cores (\citealt[]{Goodman93}). For any
spherical shell at radius $r$ inside the sonic point, it infalls
following eq. (\ref{streamline}), with $r_d$ to be the outer radius of
the disk determined by the position of the sonic point.
\begin{equation}
r_d=\frac{6a_g}{5a_i}\beta
\Big(\frac{M_\mathrm{sp}}{M_\mathrm{core}}\Big)^\frac{1}{3-k_\rho},
\end{equation}
where $M_\mathrm{sp}$ is the mass inside the sonic point. In our
fiducial case, when the collapsed mass is 10.67 $M_\odot$ (8 $M_\odot$
star and 2.67 $M_\odot$ disk, discussed in Section \ref{sec:model}),
the disk radius is 449 AU.
The rotation changes the density distribution from spherical symmetry to
axissymmetry as below
\begin{equation}
\rho(r,\theta)\,\propto\,\Big(1+\frac{\cos\theta}{\cos\theta_0}\Big)^{-1/2}\Big(\frac{\cos\theta}{2\cos\theta_0}+\frac{r_d}{r}\cos^2\theta_0\Big)^{-1}. \label{rotateden}
\end{equation}
This result is valid for an infall rate that is constant with radius,
which is not exactly right for the expansion wave solution of a
polytropic sphere. Thus we only scale the density distribution with
the angular dependence of eq. (\ref{rotateden}) and keep the total
mass in each shell unchanged, so that the infall rate is the same as
the non-rotating case. Fig. \ref{fig:rotation} shows the radial
density profiles at different polar angles (upper panel) and the
density profiles with polar angle at different radii (lower panel),
for a core with 10.67 $M_\odot$ collapsed at the center. In the upper
panel, the thick black line shows the density profile in a
non-rotating core. The thin curves correspond to different polar
angles. There is a discontinuity at $r\sim 2500$ AU, which marks the
position of the sonic point, only inside of which we consider
rotation. For a real core, this discontinuity would not exist. The
peaks and dips mark the disk radius $r_d$. The lower panel of
Fig. \ref{fig:rotation} shows the dependence of density on $\theta$ at
$2r_d$ and $4r_d$.
\subsection{Accretion Disk}
\label{sec:disk}
Inside the centrifugal radius $r_d$, the infalling flow is
circularized and forms an accretion disk around the protostar. The disk
can provide an efficient way to transfer angular momentum outwards and
enable high accretion rates to form a massive star. The high accretion
rate indicates that the disk is relatively massive (comparable to the
star). But gravitational instability can become very efficient when
the mass of the disk is high enough and lead it to fragment. We assume
the disk mass is always a constant fraction of the stellar mass
(\citealt[]{TM04}):
\begin{equation}
m_{*d}=m_*+m_d=(1+f_d)m_*,
\end{equation}
where $f_d$ is assumed to have a fiducial value of 1/3. Recently,
\citet[]{Kratter10} performed a numerical parameter study on accretion disks
of massive stellar systems and found that a disk can be stable and does not
fragment with an even higher disk-to-star mass ratio ($f_d\sim 1$).
We follow the description for the disk structure by
\citet[]{Whitney03b}, which is a standard $\alpha$-disk
(\citealt[]{SS73}, \citealt[]{LP74}, \citealt[]{Pringle81},
\citealt[]{Bjorkman97}, \citealt[]{Hartmann98}) with density
distribution
\begin{equation}
\rho=\rho_0 \Big(1-\sqrt\frac{r_*}{r}\Big)
\Big(\frac{r_*}{r}\Big)^\alpha \exp \Big\{-\frac{1}{2}
\Big(\frac{z}{H(r)}\Big)^2 \Big\}, \label{rho}
\end{equation}
where $r$ is the radial coordinate in the disk midplane, $z$ is the
distance from the midplane, and $r_*$ is the stellar radius. Basically
the density has a power-law profile with radius (smoothed at the
innermost region) and a Gaussian profile with the height. $H$ is the
scale height of the disk,
\begin{equation}
H=H_0\Big(\frac{r}{r_*}\Big)^\beta. \label{height}
\end{equation}
Thus, the disk structure is specified by following parameters: disk
mass $m_d$, disk inner radius $r_{d,\mathrm{in}}$, disk outer radius
$r_d$ (the centrifugal radius), disk scale height $H_0$ at the stellar
surface, and the power indices $\alpha$ and $\beta$.
If the dust is the only source of opacity in the disk, then it equals
to a disk truncated at dust destruction radius $r_\mathrm{sub}$,
inside of which the temperature is too high for dust to
exist. However, in reality, the gas opacity at the innermost region of
the disk can be significant, so the disk should be optically thick
down to the surface of the protostar. Thus, in our fiducial model, the
inner radius of the disk is set to be the stellar radius.
We consider both geometrically thin ($H_0/r_*= 0.01$) and moderately
thick ($H_0/r_*= 0.1$) disks to see the effects of the disk scale
height. $H_0/r_*=0.01$ is the value used by
\citep{Whitney03a,Whitney03b}. $H_0/r_*=0.1$ is closer to the aspect
ratio expected for a disk composed of dust only and is a typical value
for a moderately thick disk. For these cases, $\alpha$ and $\beta$ are
chosen to be 1.875 and 1.125, respectively, as in the standard
$\alpha$ disk models considered by Whitney. However, in our final
fiducial model, we calculate the disk scale height, $\alpha$ and
$\beta$ self-consistently from the gas and dust opacities, in which
case, fitting with eq. (\ref{rho}) and (\ref{height}) gives
$H_0/r_*=0.06$ and $\alpha=1.75$, $\beta=1.08$.
In the limit of a slowly rotating protostar, the accretion luminosity
from the system is
\begin{equation}
L_{\mathrm{acc}}=\frac{Gm_*\dot{m}}{r_*},\label{totalacc}
\end{equation}
half of which is emitted from the viscous disk (disk accretion
luminosity $L_\mathrm{disk}$) and the other half is emitted when the
material hits the surface of the protostar (hot-spot luminosity
$L_\mathrm{hotspot}$),
\begin{equation}
L_\mathrm{disk}=L_\mathrm{hotspot}=\frac{1}{2}L_\mathrm{acc}.
\end{equation}
For simplicity, we add the hot-spot luminosity to the star's
homogeneously. The star then has a single, enhanced temperature to
describe its black body spectrum. If the disk is truncated at
$r_{d,\mathrm{in}}$, the disk luminosity will be
\begin{equation}
L_{\mathrm{disk}}(r_{d,\mathrm{in}})=\frac{Gm_*\dot{m}}{2r_{d,\mathrm{in}}}\Big(3-2\left(r_*/r_{d,\mathrm{in}}\right)^{1/2}\Big),\label{ldisk}
\end{equation}
which indicates that only when the disk extends to the stellar radius
can the total energy be conserved. In fact, the disk accretion
luminosity is very sensitive to the inner radius of the disk. The
accretion rate $\dot{m}$ of a massive protostar can be very high,
reaching $\sim 10^{-4} - 10^{-3} \;M_\odot$/yr. Here we adopt the
protostar evolution model by MT03, which gives an accretion
rate of $2.40\times 10^{-4}\;M_\odot$/yr for an 8 $M_\odot$ protostar
in a 60 $M_\odot$ core. This accretion rate gives a disk luminosity and a
hotspot luminosity which are comparable to the stellar luminosity.
The disk accretion rate is related to the $\alpha$ parameter of the disk by
\begin{equation}
\dot{m}=\sqrt{18\pi^3}\alpha_\mathrm{disk}V_c\rho_0 H_0^3/r_*
\end{equation}
with $V_c=(Gm_*/r_*)^{1/2}$. Parameters used in our fiducial model
correspond to a case with $\alpha_\mathrm{disk}=1.43$, which is consistent
with the results of numerical simulations by \citet[]{Krumholz07},
in which they found an effective $\alpha_\mathrm{disk}\;\sim1.0-1.6$.
Also, part of the energy may
be used to drive the outflow (mechanical luminosity $L_w$,
MT03). But for now we only
consider the radiative luminosity from the disk and the hot-spot.
\subsection{Outflow Cavity}
\label{sec:outflow}
Powerful bipolar outflows are ubiquitous phenomena around protostars
(e.g. \citealt[]{KP2000}).
The prevalent interpretation is that outflows are powered by accretion
activity, being driven by spinning magnetic fields that thread the
disk. There are several theoretical models to describe this process
such as the X-wind from the innermost region of the disk
(\citealt[]{Shu94}) and disk wind model (\citealt[]{KP2000}).
A common feature of these models is the production of a bipolar
outflow with momentum distribution $p_w\propto (\sin\theta_w)^{-2}$ for
$\theta_w>\theta_{w0}$, where $\theta_w$ is measured from the outflow axis
and $\theta_{w0}\sim 10^{-2}$ is a small angle (\citealt[]{MM99}). On scales large
compared to the source,
\begin{equation}
\frac{\mathrm{d} \dot{p}_w}{\mathrm{d}
\Omega}=\frac{\dot{p}_w}{4\pi\ln(2/\theta_{w0})(1+\theta_{w0}^2-\cos^2\theta_w)}. \label{outflow}
\end{equation}
We follow the discussion of \citet[]{TM02}. Assuming an opening angle
of $\theta_{w,\mathrm{esc}}$ for the outflow cavity, the outflow material
with $\theta_w>\theta_{w,\mathrm{esc}}$ cannot escape from the core and will
go back to the infalling flow. We parameterize the fraction of the
outflow momentum that escapes from the core with
$f_{w,\mathrm{esc}}$. \citet[]{NS94} showed that the velocity of the
wind is approximately independent of the polar angle so that
$f_{w,\mathrm{esc}}$ can also describe the ratio between the mass loss
rates. We assume that part of the material accreted to the disk will
be transferred to the star at a rate $\dot{m}_*$, and another part will
leave to the outflow at a rate $\dot{m}_w$, while the rest is left in
the disk so the disk grows in a rate $\dot{m}_d$. We also assume that
$\dot{m}_d = f_d \dot{m}_*$, $f_w=\dot{m}_w/\dot{m}_*$ and
$f_{w,\mathrm{esc}}$ are all constant.
Generally, we assume the outflow starts at time $t_1$ when the stellar
mass is $m_*(t_1)$ and the disk mass is $m_d(t_1)$ ($t_1=0$
corresponds to situation where the outflow cavity forms as soon as the
collapse begins.) When $t<t_1$ there is no outflow and we have
\begin{equation}
m_{*d,0}(t_1)=m_{*d}(t_1)=m_*(t_1)+m_d(t_1)=(1+f_d)m_*(t_1),
\end{equation}
where $m_{*d,0}$ is the collapsed mass of the polytropic core
(a hypothetical star-disk mass if the feedback is absent). After $t_1$ until
the time when the whole core has collapsed $t_f$, we have
\begin{equation}
\dot{m}_*+\dot{m}_d+f_{w,\mathrm{esc}}\dot{m}_w=\cos\theta_{w,\mathrm{esc}}\dot{m}_{*d,0},
\end{equation}
where $\cos\theta_{w,\mathrm{esc}}$ means only part of the infalling core
accretes to the disk because of the existence of the outflow
cavity. Therefore, the instantaneous star formation efficiency is
\begin{equation}
\epsilon_{*d}\equiv\frac{\dot{m}_{*d}}{\dot{m}_{*d,0}}=\frac{m_{*d}(t_f)-m_{*d}(t_1)}{M_\mathrm{core}-m_{*d,0}(t_1)}=\frac{(1+f_d)\cos\theta_{w,\mathrm{esc}}}{1+f_d+f_wf_{w,\mathrm{esc}}},
\end{equation}
and the mean star formation efficiency is
\begin{equation}
\bar{\epsilon}_{*d}\equiv\frac{m_{*d}(t_f)}{m_{*d,0}(t_f)}=\frac{m_{*d}(t_f)}{M_\mathrm{core}}
\end{equation}
After the whole core has collapsed, part of material in the disk will accrete
onto the star and rest of them will leave the star-disk system to the outflow
wind, which gives us
\begin{equation}
m_{*f}=m_*(t_f)+\frac{m_d(t_f)}{1+f_w}=\frac{1+f_d+f_w}{(1+f_w)(1+f_d)}m_{*d}(t_f),
\end{equation}
where $m_{*f}$ is the mass of the star finally born out of the core, which is
assumed to be half of the initial total mass of the core $M_\mathrm{core}$,
i.e, the final star formation efficiency is
\begin{equation}
\epsilon_{*f}\equiv\frac{m_{*f}}{M_\mathrm{core}}=\frac{1+f_d+f_w}{(1+f_w)(1+f_d)}\bar{\epsilon}_{*d}=\frac{1}{2}. \label{epsilon}
\end{equation}
For the models with outflow cavities that we investigate in this
paper, we make the assumption that the outflow cavity has only now
formed when $m_*(t_1)=8\;M_\odot$, so $m_{*d}(t_1)=10.67\;M_\odot$. The
validity of this assumption will be examined in a future paper.
Combining eq. (\ref{outflow}) to (\ref{epsilon}) we solve for the
opening angle $\theta_{w,\mathrm{esc}}$ and $f_{w,\mathrm{esc}}$
simultaneously. For $\epsilon_{*f}=0.5$ and $f_d=1/3$, we find that
when $f_w=0.1\sim 0.8$, the opening angle varies from 64$^\circ$ to
45$^\circ$ and $f_{w,\mathrm{esc}}$ varies from 0.91 to 0.84. So the
fraction of the outflow coming back to the infalling core is always
small, and the opening angle is typically large except when $f_w$ is
close to 1. We choose $f_w=0.6$ as our fiducial value, making
$f_{w,\mathrm{esc}}=0.86$ and the opening angle of the outflow cavity
to be 51$^\circ$. In our models, the cavity wall follows the
streamline of the rotating infalling material
(eq. \ref{streamline}). For now we set the outflow cavity to be empty
(i.e. the optically thin limiting case). One would expect it to be
relatively free of dust if most of the outflow is launched from the
region of the disk inside the dust destruction front and assuming
there is little time for new dust grains to form in the rapidly
expanding outflow. We will improve upon this assumption by studying
the detailed density distribution in the outflow cavity in our
next paper. Since the cavity does not change the density distribution
in rest of envelope, the total mass of the material in the envelope
after outflow cavities are carved out is about half its original
value.
\begin{figure*}
\begin{center}
\includegraphics[width=0.9\columnwidth]{figure1.1.ps}
\includegraphics[width=0.9\columnwidth]{figure1.2.ps}\\
\caption{Structure of the massive star forming core in the fiducial
model. The right panel is a zoom-in view of the central region of
the left panel. The positions of core boundary, expansion wave
front, sonic point (inside which we consider rotation), disk scale
height, dust destruction radius and the outer radius of the disk are
marked. The outer boundary of the disk is chosen to be a surface
with constant density joining the outflow cavity wall at $r_d$.}
\label{fig:sketch}
\end{center}
\end{figure*}
\subsection{Protostar}
\label{sec:star}
Assuming that the core mass is a constant, the mass of the central
protostar indicates its age and evolutionary stage. At some moment,
the outgoing expansion wave front will reach the boundary of the core
and, depending on the properties of the boundary, may lead to a
backward wave and a breakdown of self-similarity. In our case, this
happens when the collapsed mass reaches $m_{*d,0}\sim 12\; M_\odot$.
For the present paper we have chosen to consider a series of models
with $m_*=8\;M_\odot$, and thus a maximum central collapsed mass of
10.67 $M_\odot$ for those cases with rotating infall to a disk. Thus
the central object is on the verge of becoming a ``massive
protostar'', following the definitions of \citet[]{Beuther07} and
\citet[]{Tan08}. The full evolutionary sequence, including a treatment for
greater masses, will be considered in a future paper.
MT03 studied the evolution of massive protostars, and for
$m_*=8\;M_\odot$ and $\Sigma_{\rm cl}=1\:{\rm g\:cm^{-2}}$, the
protostar is not expected to have yet contracted to the main sequence.
They calculated the values of radius and luminosity of the protostar:
$r_*=12.05\;R_\odot$ and $L_*=2.81 \times 10^3\;L_\odot$. For simplicity,
we assume a black body spectrum for the star, with surface temperature
$T_*=1.22\times 10^4$ K in this condition. For models with an active
disk, we also add the hot-spot luminosity to the stellar spectrum,
assuming it is emitted homogeneously from the stellar surface, which
means the temperature now is $T_{*,\mathrm{hotspot}}=1.43\times
10^4$ K.
For the cases with $\Sigma_{\rm cl}$ = 0.316 g cm$^{-2}$, we use the
following parameters: $r_*=11.3\;R_\odot$, $T_*=1.25\times10^4$ K, and
$T_{*,\mathrm{hotspot}}=1.37\times10^4$ K. For the cases with
$\Sigma_{\rm cl}$ = 3.16 g cm$^{-2}$, these parameters are:
$r_*=5.93\;R_\odot$, $T_*=1.74\times10^4$ K, and
$T_{*,\mathrm{hotspot}}=2.62\times10^4$ K.
\subsection{Model Series}
\label{sec:model}
\begin{table*}
\begin{center}
\caption{Properties of features included in the model series.\label{table1}}
\begin{tabular}{c|c|c|c|c}
\hline
\bf Models & \bf Star & \bf Disk & \bf Envelope & \bf Outflow \\
\hline
1 & 8$M_\odot$ & no & 52$M_\odot$, $\rho\propto r^{-1.5}$ & no \\
\hline
2 & 8$M_\odot$ & no & 52$M_\odot$, expansion wave & no\\
\hline
\multirow{2}{*}{3} & \multirow{2}{*}{8$M_\odot$} & $1/3m_*$, thin ($H_0/r_* = 0.01$), &
$49.333M_\odot$, expansion & \multirow{2}{*}{no} \\
& & passive, $r_{d,\mathrm{in}}=r_\mathrm{sub}$ & wave, rotation & \\
\hline
\multirow{2}{*}{4} & \multirow{2}{*}{8$M_\odot$} & \multirow{2}{*}{as above} &
only $\sim 29 M_\odot$ left, & \multirow{2}{*}{yes} \\
& & & expansion wave, rotation & \\
\hline
\multirow{2}{*}{5} & \multirow{2}{*}{8$M_\odot$} & $1/3m_*$, thin, & \multirow{2}{*}{ as above
} & \multirow{2}{*}{yes} \\
& & $r_{d,\mathrm{in}}=r_\mathrm{sub}$, active & & \\
\hline
\multirow{2}{*}{6} & \multirow{2}{*}{8$M_\odot$} & $1/3m_*$, active, $r_{d,\mathrm{in}}=r_\mathrm{sub}$& \multirow{2}{*}{ as above
} & \multirow{2}{*}{yes} \\
& & thick ($H_0/r_* = 0.1$) & & \\
\hline
\multirow{2}{*}{7} & \multirow{2}{*}{8$M_\odot$} & same as Model 8, & \multirow{2}{*}{ as above
} & \multirow{2}{*}{yes} \\
& & except $r_{d,\mathrm{in}}=r_\mathrm{sub}$ & & \\
\hline
8 & \multirow{2}{*}{8$M_\odot$} & 1/3$m_*$, active, & \multirow{2}{*}{as above} & \multirow{2}{*}{yes}\\
(fiducial) & & $H_0/r_*=0.06$, $r_{d,\mathrm{in}}=r_*$ && \\
\hline
\end{tabular}
\end{center}
\end{table*}
\begin{table*}
\begin{center}
\caption{Parameters of the models comparing different column density. \label{table2}}
\begin{tabular}{l|c|c|c}
\hline
& Model 8 (fiducial) & Model 8l & Model 8h\\
\hline
mean surface density $\Sigma_\mathrm{cl}$ (g/cm$^2$) & 1 & 0.316 & 3.16\\
\hline
core radius $R_\mathrm{core}$ (pc) & 0.057 & 0.10 & 0.032\\
\hline
position of expansion wave front $r_\mathrm{ew}$ (pc) & 0.049 & 0.088 & 0.028\\
\hline
formation time $t_f$ (yr) & $1.29 \times 10^5$ & $3.06 \times 10^5$ & $5.44 \times 10^4$ \\
\hline
position of sonic point $r_\mathrm{sp}$ (AU) & $2.57\times 10^3$ & $4.57\times 10^3$ & $1.45\times 10^3$\\
\hline
outer radius of disk $r_d$ (AU) & 449.4 & 801.4 & 253.4\\
\hline
disk accretion rate $\dot{m}$ ($M_\odot$/yr) & $2.398\times 10^{-4}$ & $1.035\times 10^{-4}$ & $5.667\times 10^{-4}$ \\
\hline
stellar radius $r_*$ ($R_\odot$) & 12.0 & 11.3 & 5.93\\
\hline
stellar surface temperature $T_*$ (K) & $1.22\times 10^4$ & $1.25\times 10^4$ & $1.74\times 10^4$\\
\hline
stellar + hotspot temperature $T_{*,\mathrm{hotspot}}$ (K) & $1.42\times 10^4$ & $1.37\times 10^4$ & $2.62\times 10^4$\\
\hline
stellar luminosity $L_*$ ($L_\odot$) & $2.81\times 10^3$ & $2.82 \times 10^3$ & $2.84\times 10^3$\\
\hline
total accretion luminosity $L_\mathrm{acc}$ ( $L_\odot$) & $4.90\times 10^3$ & $2.25\times 10^3$ & $2.36\times 10^4$ \\
\hline
disk scale height at stellar surface $H_0/r_*$ & 0.06 & 0.053 & 0.078\\
\hline
\end{tabular}
\end{center}
\end{table*}
In order to study the effects of all the features discussed above on
SEDs and images, we construct a model series starting from the
simplest to the one containing all these features. All the models
assume an initial core mass of 60 $M_\odot$ from which an 8 $M_\odot$
protostar has formed at the center.
In Model 1, we assume a spherical symmetric density distribution in
the core with a power-law dependence on the radius, $\rho\propto
r^{k_\rho}$. The total mass in the envelope is $52\;M_\odot$
since an $8 M_\odot$ protostar has already formed at the center.
In Model 2, We change the radial density distribution in the envelope
to the expansion wave solution (similar to the thick line in the upper
panel of Fig. \ref{fig:rotation}, but here the expansion wave front is at
$r=0.0408$ pc and the envelope mass is again fixed at $52\:M_\odot$).
We begin to consider rotating infall inside the sonic point and thus a
disk around the star in Model 3. The disk is geometrically thin
($H_0/r_* = 0.01$) and passive (no accretion luminosity) with the
inner radius set to be the dust destruction radius $r_\mathrm{sub}$,
which is empirically determined to be (\citealt[]{Whitney04})
\begin{equation}
r_\mathrm{sub}=r_*(T_\mathrm{sub}/T_*)^{-2.1}
\end{equation}
where $T_\mathrm{sub}$ is the dust sublimation temperature and we adopt
$T_\mathrm{sub}=1600$ K.
The expansion wave
front now reaches $r=0.0494$ pc for the collapsed mass now is
$m_*+m_d=10.67\;M_\odot$. The envelope mass is now $49.33\;M_\odot$.
Outflow cavities are added in from Model 4. We keep the density
profile in the envelope unchanged, which corresponds to a case that
the outflow has just swept up the material to form the bipolar
cavities. The envelope mass is now $\sim 29\:M_\odot$.
The accretion luminosities (both from disk and hot-spot) are turned on
in Model 5. However, since the disk is truncated at the dust
destruction radius which in this case is $\sim 98.4\; r_*$, most of the
disk accretion luminosity is lost and the rest of it
($L_\mathrm{disk}=25.29\;L_\odot$) is much lower than
the hotspot luminosity ($L_\mathrm{hotspot}=2.45\times10^3\;L_\odot$).
Note that in Robitaille model,
part of this missing disk luminosity is included
in the stellar $+$ hotspot luminosity.
In Model 6 we adjust the disk to be a geometrically thick one
($H_0/r_* = 0.1$) to see the effects of the height of the disk.
In our fiducial model, the disk is extended to the stellar radius so
that the accretion luminosity can be conserved. Inside the dust
destruction front ($T >1600$ K), gas opacities are used. Here we assume
that the $\alpha$-disk model is still valid.
The scale height, $H_0$, and radial and vertical scaling parameters
$\alpha$ and $\beta$ are calculated self-consistently from the assumed
opacities and other stellar and envelope parameters: $H_0/r_*=0.06$,
$\alpha=1.75$, and $\beta=1.08$. Fig. \ref{fig:sketch} is a schematic
of the structure in the fiducial model. The positions of core
boundary, expansion wave front, sonic point, outer radius of the disk,
dust destruction radius, and the scale height of the disk are all
marked. Unlike the previous models, in which disk material can exist
at any height (though the density is very low at large $z$), we
truncate the disk at a fixed density of $2\times10^{-15}\;\mathrm{g}\;\mathrm{cm}^{-3}$
so that the disk atmosphere joins smoothly
with the infalling envelope. (The input density profile can be seen in
the lower panel of Fig. \ref{fig:tdinit}.) We expect that the
existence of optically thick gas in the innermost region of the disk
can make a significant difference. In order to investigate this
effect, we construct Model 7 to be the same as the fiducial model
except without gas disk ($r_{d,\mathrm{in}}=r_\mathrm{sub}$). And the
fiducial model is labeled with Model 8. Table \ref{table1} lists the
differences of all these models.
The column density of the clump where the core is embedded in can
affect the surface pressure of the core, therefore its size, outer
radius of the disk, accretion rate and also the scale height of the
disk. In order to investigate this effect, we consider other two
values for the mean surface density of the clump,
$\Sigma_\mathrm{cl}=0.316$ and 3.16 g/cm$^2$. These correspond to a
change in column density of the core by a factor of 10 and the surface
pressure by a factor of 100. These two models are labeled with Model
8l (low surface density) and 8h (high surface density). Table
\ref{table2} lists the differences of these three models.
\section{Simulations}
\label{sec:simulation}
\begin{figure*}
\begin{center}
\includegraphics[width=0.8\textwidth]{trho.ps}\\
\caption{Input temperature and density for our fiducial model.
$n_\mathrm{He}=0.1n_\mathrm{H}$ is assumed here. The
black contours in the upper panels correspond to the dust destruction
front (1600K), cells inside which are assigned with gas
opacities. The white contours show the photon-diffusive region calculated in the fiducial model,
a photon emitted inside which will move to the surface through a path with constant $r$ and emitted again from the surface. The temperatures of the cells on that path are calculated with grey atmosphere approximation, which is why the the dust destruction fronts have these vertical structures. Also, since the frequency of the photon, when it finally leave the photon-diffusive region, is calculated from the surface temperature, these vertical structures inside it should not affect the results.}
\label{fig:tdinit}
\end{center}
\end{figure*}
\begin{figure}
\begin{center}
\includegraphics[width=\columnwidth]{sede_series_60.ps}\\
\caption{SEDs for the model series, assuming a distance of 1 kpc. A
typical inclination of $60^\circ$ is chosen. The upper panel shows
the SEDs and the lower panel shows the differences of each model
from the fiducial one. The black solid curve is for the fiducial model,
the red thin curves are for Model 1 - 3 (models without outflow
cavities), and the blue thick curves are for Models 4 - 7 (models with
outflow cavities) The black dotted line is the input stellar
spectrum (black-body) which is used for all the models.}
\label{fig:sed_series_60}
\end{center}
\end{figure}
\begin{figure}
\begin{center}
\includegraphics[width=\columnwidth]{sede_series_30.ps}\\
\caption{The same as Fig. \ref{fig:sed_series_60}, except for an inclination of $30^\circ$.}
\label{fig:sed_series_30}
\end{center}
\end{figure}
\begin{figure*}
\begin{center}
\includegraphics[width=0.8\textwidth]{tdiffus_model78.ps}\\
\caption{Output temperature distribution of Model 7 and 8, showing the
difference the gaseous innermost disk makes. The black contours are
dust destruction front ($T$=1600K).
}
\label{fig:tdiffus_78}
\end{center}
\end{figure*}
\begin{figure}
\begin{center}
\includegraphics[width=\columnwidth]{sede_fiducial_new.ps}\\
\caption{SEDs of the fiducial model (Model 8) at 4 inclinations,
assuming a distance of 1 kpc ($F_\nu$ in upper panel and $\nu F_\nu$ in lower panel).
The thicker lines are total fluxes
and the thiner red lines are only the scattered light. The lower dotted blue
line is the input stellar spectrum (black-body). The upper dotted
line is the total luminosity of star and hot-spot emitted as a
black-body.}
\label{fig:sed_fiducial}
\end{center}
\end{figure}
We use the Monte Carlo radiation transfer code by
\citet[]{Whitney03b} to perform our calculations. This code includes
thermal emission, non-isotropic scattering and polarization due to
scattering from the dust in a spherical-polar grid, using the method
of Monte Carlo radiative equilibrium by \citet[]{BW01}. This method
requires that the opacities are set up at the beginning of each run
and kept unchanged. During the run, the temperature in each cell is
always being corrected by calculating the balance of absorption and
emission of new photon packets. This becomes a problem when gas
opacities are included because they are highly dependent on the
temperature and density. So we have to iterate to obtain the correct
temperature profile for the disk (especially for the inner region) and
then use it to set up opacities for each cell.
We choose the analytical temperature profile from $\alpha$-disk model
in our case as an initial condition. After several iterations, the
temperature of the outer disk $\gtrsim$ 4 AU becomes quite
stable. However, because of the discontinuity of the opacities at the
transition region between gas and dust, the temperature profile in the
inner region oscillates between two phases. Even though the difference
can be large between results from two adjacent steps
(i.e. $T_n(r,\theta)$ and $T_{n-1}(r, \theta)$), the averaged
temperature profile of two adjacent step should keep similar
(i.e. $\log(T'_n(r,\theta))=(\log(T_n(r,\theta))+\log(T_{n-1}(r,\theta)))/2$).
So we stop the iteration when $T'_n\sim T'_{n-1}$, more specifically,
when standard deviation of the distribution of $({\rm log} T'_n - {\rm
log} T'_{n-1})/{\rm log} T'_n$ for the cells inside 4 AU becomes
consistently $<0.1$, which corresponds to a variation of
$\sim$ 25\%. Then we use the temperature profile which is higher in the
midplane to be the input for the final run with large number of
photons to generate SEDs and images. We performed simulations with
both input temperature profiles and saw no significant differences on the SEDs
and images. We also doubled the photon number for the iteration but
found no significant dependence of the standard deviation on the
photon number. It should be noted here that 0.1 standard deviation
inside 4 AU is only an arbitrary standard.
The upper panel of Fig. \ref{fig:tdinit} shows the input temperature
profile. The black contours correspond to the dust destruction front
($T=1600$ K), inside which gas opacities are assigned depending on the
temperature and density. We can also see that the dust destruction
front extends to $\sim 10$ AU in the midplane and $\sim 3$ AU on the
surface of the disk, which agrees with the estimate of the dust
destruction radius of 5.5 AU in Model 3 - 7.
We note that temperature iteration is only used in Model 8, while in
other models, the dust opacities are assigned only depending on the
region and the density. Also, this input temperature is only used to
assign opacities. It is not the initial temperature condition for each
run.
The inner region of the disk around the midplane is very optically
thick and so detailed radiative transfer simulation becomes very
time-consuming. Thus in the code the grey atmosphere approximation is
used to describe this region (photon-diffusive region). A photon
generated inside this photon-diffusive region will move to the surface
of the region by following a path with same $r$, and is then emitted
with a frequency calculated based on the temperature of the surface
cell. The temperatures of the cells on that path are calculated
accordingly with the grey atmosphere approximation. In models except
the fiducial one (Model 3 - 7), as in the original code, a cell is set
to be in the photon-diffusive region if the optical depth from
$z=\infty$ to it is larger than 10. In the fiducial model, we change
this to a more restrictive local definition, that the photon-diffusive
region is where the mean free path is smaller than 0.1 $H(r)$. Note
that as long as this photon-diffusive region is set small enough, it
should not affect the final results.
A 3000 $\times$ 1499 grid is used to resolve the $r \times \theta$
space. In $r$ space, $\sim600$ cells are used to resolve the disk with
a finer grid in inner region, $\sim150$ cells are used to cover the
region inside $\sim 6$ AU in the fiducial model. In $\theta$ space,
the grid is finer in the disk (especially around the midplane) and
around the opening angle of the outflow cavity
$\theta_{w,\mathrm{esc}}$. $\sim 700$ cells are used between
$20^\circ$ above and below the disk midplane and $\sim 300$ cells are
used between $50^\circ$ and $53^\circ$
($\theta_{w,\mathrm{esc}}=51^\circ)$.
For each run, SEDs at ten inclinations (evenly distributed in cosine
space) can be produced simultaneously, while if the ``peeling-off''
mode is used, images and SEDs with higher signal-to-noise ratio are
produced for the particular ``peeling-off'' angle. The ``peeling-off''
mode is very time-consuming. For most of our models, $10^8$ photons
are used in one run, and it typically takes several days to one week
running on a single processor. This number of photon packets is still
not perfect for an image of the whole core, especially for those at
wavelengths with low fluxes. Because this code does not enable
parallel computing now, in order to save time, for each model we run
10 times with different random seeds simultaneously on different
processors, and superposed their results, making our final images
contain $10^9$ photons.
Images are produce at several wavelengths and convolved with filter
functions for comparison to observations. The code has already
included filters such as Spitzer IRAC filters at 3.6, 4.5, 5.8 and 8.0
{\micron}, Spitzer MIPS filters at 24, 70 and 160 {\micron}, and 2MASS
J, H, K bands. We also add in the filters of GTC-CanariCam, Herschel
PACS and SPIRE,
and SOFIA FORCAST. We will show images both before and after
convolution with resolution of these particular instruments.
\subsection{Dust Opacity}
\label{sec:dust}
We use three dust grain models for different regions: (1) the
envelope; (2) lower density regions of the disk, and (3) the densest
regions in the disk ($n_{\mathrm{H}_2}>10^{10}\;\mathrm{cm}^{-3}$, the
criterion used by \citealt[]{Robitaille06}). For our present models
there is no dust (or gas) in the outflow cavities. The default dust
models in the code are used without any change
(\citealt[]{Whitney03a}).
The dust model used in the envelope is based on that derived by
\citet[]{KMH94} for the diffuse interstellar medium (ISM). The size
distribution is modified by \citet[]{Whitney03a} to fit an extinction
curve typical of the more dense regions of the Taurus molecular cloud
with $R_V=4$. These grains also include a water ice mantle covering
5\% of the radius. For the lower density regions of the disk, the
grain model of \citet[]{Cotera01} is used. It has grains larger than
ISM grains, but not as large as the disk midplane grains.
For the densest regions of the disk, we use the dust model with large
grains ($\sim$1 mm; model 1 in \citealt[]{Wood02}), which fit the
HH 30 disk SED well. Compared to ISM grains, the larger dust grain
model has a shallower wavelength-dependent opacity: lower at short
wavelengths and larger at long wavelengths.
\subsection{Gas Opacity}
\label{sec:gas}
For most regions of the disk, dust grains dominate the opacity, even if
the mass ratio between dust and gas is low (0.01 is used in our
models). However, in the innermost region of the disk, where the dust
cannot exist ($T>1600$ K), opacity is dominated by gas. Especially when
the temperature is high ($\sim 10^4$ K), the mean opacities of the gas
can be comparable or higher than dust opacities. As discussed in
Section \ref{sec:disk}, most of the disk accretion luminosity is from
this innermost region. Thus, gas opacity should be included in a
realistic model.
In order to simulate the frequency of each photon packet emitted from
the gas-dominated region correctly, not only Rosseland or Planck mean
opacities, but also the frequency-dependent opacities are
needed. Besides, the gas opacity is highly dependent on the
temperature and the density, which make our problem much more
complicated. Since our aim is only to simulate the disk emission
correctly, especially in near and mid-IR, rather than to simulate the
details of line profiles, we smooth the monochromatic opacities for
simplicity and smaller memory demand. Also, we assign opacities to a
grid in temperature and density, rather than to interpolate to obtain
opacities at exact temperature and density.
For temperatures higher than $\sim 3000$ K, we adopt gas opacities from
OP project (\citealt[]{Seaton94}, \citealt[]{Badnell05} and
\citealt[]{Seaton05}). They provide monochromatic gas opacities of
hydrogen, helium and other 15 elements, in large ranges of temperature
($\sim 3000$ K to $10^8$ K, we only use those of temperature up to
$\sim 10^6$ K in our present model since the maximum temperature
we can reach in the models is less than this) and density.
The opacities are mainly due to the line absorption, ionization, H$^-$
bound-free absorption, electron-hydrogen/helium free-free
absorption. For scattering, Thompson scattering with a collective
effect is considered (\citealt[]{Boercker87}).
For regions with $T<3000$ K, the opacity model of \citet[]{AF94} is
adopted. However, they do not provide monochromatic opacities for a
range of temperatures and densities. So we use the monochromatic
opacities shown in Fig. 4 of \citet[]{AF94} which is for a temperature
of 2000 K and a density of $8\times 10^{-12}\; \mathrm{g/cm}^3$, and
rescale it for other temperatures and densities based on their
Rosseland mean opacities. Fortunately, the temperature range for this
model is quite narrow (1600K - 3000K). Alexander and Ferguson's model
includes both gas and molecules. At $\sim 2000$K, the total opacity is
dominated by molecules, especially H$_2$O and TiO. Atomic lines and CO
lines are important at some wavelengths. Other continuous sources
include H$^-$ absorption and Rayleigh scattering of H and H$_2$.
In this way we construct an opacity grid of temperature and
density. The temperature varies from 1600 K to 10$^6$ K with an
interval of 0.1 in logarithmic space, and the density varies from
$\sim 10^{-8}$ to $\sim 10^{-15}$ g/cm$^3$ with an interval of 0.5 in
logarithmic space. For each $T$ and $\rho$, a monochromatic opacity
file is assigned, so that we have totally 148 gas opacity files in our
present model. At the beginning of each run, a temperature profile is
read in, and in each cell the gas opacity file with closest $T$ and
$\rho$ to the read-in values is chosen. The opacities are not changed
during the calculation.
Since here we are not concerned with line absorption and emission, it
is better to smooth the monochromatic opacities to save computing
memory. In the code, three important values are calculated: 1)
Rosseland mean opacity
\begin{equation}
\frac{1}{\kappa_\mathrm{R}}=\frac{\int_0^\infty
(\partial B_\nu / \partial T) / \kappa_\mathrm{ext}(\nu)
\mathrm{d} \nu}{\int_0^\infty (\partial B_\nu / \partial T) \mathrm{d} \nu},
\end{equation}
used to determine the photon-diffusive region inside the disk; 2) Planck
mean opacity
\begin{equation}
\kappa_\mathrm{P}=\frac{\int_0^\infty \kappa_\mathrm{abs}(\nu)
B_\nu \mathrm{d} \nu}{ \int_0^\infty B_\nu \mathrm{d} \nu},
\end{equation}
used to calculate the energy equilibrium in a cell; and 3)
\begin{equation}
P_\nu=\int_0^\nu
\frac{\kappa_\mathrm{abs}(\nu)}{ \kappa_\mathrm{P}} \frac{\partial B_\nu}{\partial T} \mathrm{d} \nu,
\end{equation}
where $P_\nu$ is the probability that a photon packet is emitted from a cell with a frequency
between 0 and $\nu$ (\citealt[]{Bjorkman97}).
For $\kappa_P$ and $P_\nu$ it is best to smooth opacities
with linear averaging, thus giving better estimates of equilibrium
temperatures and photon frequencies. However, this method tends to
increase $\kappa_R$ by $\sim$10\% or more. Averaging ${\rm log}\kappa$
reduces the importance of line absorption and yields more accurate
Rosseland mean opacities. For our problem, the precise location of the
boundary of the photon-diffusive region is is not very important, so
we smooth opacities with linear averages. The original gas opacity files contain $10^4$ frequencies between $h\nu/kT=0.001$ and 20. We smooth them by averaging 50 adjacent frequencies.
\section{Results}
\label{sec:results}
\subsection{SEDs}
\begin{figure*}
\begin{center}
\includegraphics[width=0.8\textwidth]{sede_sigma_new.ps}\\
\caption{SEDs of Model 8 ($\Sigma_\mathrm{cl}$= 1 g/cm$^2$), Model 8l
($\Sigma_\mathrm{cl}$= 0.316 g/cm$^2$) and Model 8h
($\Sigma_\mathrm{cl}$= 3.16 g/cm$^2$) at inclinations of $60^\circ$
(left panels) and $30^\circ$ (right panels), assuming a distance of
1 kpc. For each inclination, the upper panel shows the total fluxes
(black), scattered fluxes (red) and stellar input (blue), while the
lower panel shows fluxes from the disk only.}
\label{fig:sed_sigma}
\end{center}
\end{figure*}
\subsubsection{SEDs of the Model Series}
Figures \ref{fig:sed_series_60} and \ref{fig:sed_series_30} show the
SEDs of the model series at inclination angle between our line of
sight and the protostar rotation axis of $60^\circ$ (i.e. a more
equatorial view) and $30^\circ$ (i.e. a more polar view)
respectively. In each figure, the upper panel shows the total fluxes
while in the lower panel these SEDs are compared to the fiducial model
(Model 8). The $\theta=60^\circ$ line of sight goes through the
envelope material, in which case the photons from the protostar and
the disk are all reprocessed by the dust before they can escape the
core. The $\theta=30^\circ$ line of sight from the central star passes
through the outflow cavity (for Model 4 - 8), in which case we can see
the stellar black-body spectra in the short wavelength region of the
SEDs. Some important features in the SEDs include the water ice
feature at 3 {\micron} and silicate features at 10 {\micron} and 18
{\micron}. The ice feature is only present for the higher inclination
view for which the lines of sight pass through the envelope material,
which uses a dust model with ice mantles.
Models 1 to 3 show very similar SEDs, where we do not see any
radiation at short wavelengths. The 10 {\micron} silicate features are
all very deep. The occurrence of the expansion wave (Model 2) and
rotation/disk (Model 3) shift the far-IR peak a little to shorter
wavelengths, and increases the mid-IR emission, making the 18
{\micron} silicate feature deeper. This is because the expansion wave
and the Ulrich solution decrease the density of the inner region of
the core, and thus reduce the extinction. For Model 3, which is not
spherically symmetric, the SEDs do not show much difference between
the different inclinations. It should be noted that in Model 3
rotation is only considered inside the sonic point. In a more
realistic solution, the material in the outer region of the envelope
would also be redistributed by the effect of the rotation (like the
solution of \citet[]{TSC84} for an isothermal core) so that one might
see larger differences.
The outflow cavities change the shape of SEDs significantly. With
outflow cavities (Model 4 to 8), the SEDs show a large excess at
wavelengths shorter than 10 {\micron}. The position and height of the
far-IR peak and the 20 {\micron} $\sim$ 70 {\micron} slopes are
affected by the cavity as well. Especially for a low inclination, the
star and disk can be seen directly. The fluxes at wavelengths shorter
than 10 {\micron} are larger than those at a high inclination by about
two orders of magnitude.
Note that the short wavelength emission seen in the $\theta=60^\circ$
view is essentially all due to scattered emission from the outflow
cavity walls.
Compared to a passive disk, an active disk with accretion luminosity
increases the fluxes at all wavelengths without many changes in the
shape of the SEDs (comparing Model 4 and 5). The accretion luminosity
in Model 5 is mainly due to the hot-spot, while most of the disk
accretion luminosity is lost here because of the absence of the
innermost disk. The total energy is conserved only in Model 8, which
the disk is extended to the stellar surface.
The effect of the thickness of the disk is distinct on the SEDs
(comparing Model 5 and 6). A thicker disk tends to obscure more
photons at high inclinations and emit or scatter them to low
inclination directions - i.e. more flux escapes via the outflow
cavities. Therefore, with a thicker disk,
Model 6 at $\theta=30^\circ$ shows a rise between 1 and 10 {\micron}
which even smooths out the far-IR peak, while at $\theta=60^\circ$ the
SED shows a decrease in near-IR and shorter wavelengths, and a deeper
silicate features.
Model 8 shows how the SED changes by including the flux from the
innermost disk region. Compared to previous models, the optical and
near-IR emission is significantly increased, in both high and low
inclinations. Model 7 has exactly the same geometry and density setup
as Model 8, except that it has a disk truncated at
$r_\mathrm{sub}$. The opacities are chosen depending on the input
temperature and density in Model 8 while in Model 7 they only depend
on the density, therefore, even outside $r_\mathrm{sub}$ the opacity
setup of these two models can be different. Because of the truncated
disk, the majority of the disk accretion luminosity is lost in Model
7. This missing luminosity shows up in Model 8 mainly as optical
radiation. However, the near-IR flux in Model 8 is not so bright as
Model 7. The far-IR SEDs of these two models do not show much
difference.
Fig. \ref{fig:tdiffus_78} shows the final output temperature
distributions of these two models. The temperature of the disk reaches
$\sim10^5$ K at the innermost region in Model 8, while the highest
temperature in Model 7 is only $\sim 3000$ K. This explains the optical
excess in the SED of Model 8. However, the existence of the innermost
disk shields the flux from the star, therefore in Model 7 the
temperature at the surface of the disk and the base of the outflow
cavity becomes higher, leading to the higher near-IR flux on the
SED. For the models presented here, we do not have any material in the outflow
cavity. If any dust grain exists there, the optical emission in Model
8 will be suppressed and the disk luminosity may appear as more near-
and mid-IR radiation. This will be examined in a future paper.
\subsubsection{SEDs of the Fiducial Model}
Fig. \ref{fig:sed_fiducial} shows the SEDs of our fiducial model
(Model 8) at four inclinations. Even after we smoothed the
monochromatic opacity curves of gas, the original output SEDs still
show some significant emission features, mainly H$\alpha$. Since the
line profile is not exactly correct in our models and it is not the
interest of our present study, we subtract the H$\alpha$ feature
and smooth out other line features with wavelengths $<2$
{\micron}. The energy contained in the H$\alpha$ line is
typically $\lesssim$ few\%.
The SEDs at 4 inclinations are shown here, from a view along the
rotation/outflow axis to that through the disk plane. The inclination
of the viewing angle changes the observed flux at short wavelengths
significantly. It also affects the height and position of the peak in
the far-IR region, and the mid-IR spectral slope. The SED at
wavelengths longer than $\sim$100 {\micron} is not affected by the
inclination. Inclinations of $30^\circ$ and $0^\circ$ have very
similar SEDs ($0^\circ$ inclination contains 8\% more energy than
$30^\circ$ inclination). Recall, the outflow cavity has an opening
angle of $51^\circ$, which means we can directly see the star in both
these cases. The stellar spectrum and the black-body spectrum
containing both stellar and hot-spot luminosity are shown by the
dotted lines. Their difference shows the luminosity from the hot-spot
region. From high inclinations to low inclinations, because of the
change of the optical depth, the silicate feature change from a big
absorption feature to a weak emission feature.
The dashed lines show the SEDs of only scattered light. At high
inclination angles, the observed light at short-wavelengths has always
been scattered, while at low inclinations, we can see stellar
radiation and thermal emission from the disk. In the far-IR, the flux is
dominated by the thermal emission of the envelope. Such significant
difference between low inclinations and high inclinations is partly
because we have an empty outflow cavity.
\subsubsection{Effect of Different Mass Surface Density}
\begin{figure*}
\begin{center}
\includegraphics[width=0.8\textwidth]{img_60_new.eps}\\
\caption{Images for the fiducial model at inclination of $60^\circ$,
assuming a distance of 1 kpc, at different wavelengths. Each image
has 149 $\times$ 149 pixels, and field of view of 30" $\times$ 30".
Each image is normalized to its maximum surface brightness which is
labeled in the bottom-left corner. The values on the top-right
corners are the total fluxes. The blue contour is 1\% of the maximum
surface brightness}
\label{fig:img_60}
\end{center}
\end{figure*}
\begin{figure*}
\begin{center}
\includegraphics[width=0.8\textwidth]{img_60_conv_new.eps}\\
\caption{Same as Fig. \ref{fig:img_60}, except convolved with the
resolution of each particular instrument, labeled in the top-right
corner and shown as red circles of
radius equal to the HWHM (or line segments with lengths equal to the
HWHMs for Herschel SPIRE) on the bottom right corners. The maximum
surface brightness (which is dependent on the resolution) is labeled
in the bottom-left corner. The contours for most of the images have
intervals of half an order of magnitude from the maximum
brightness. Because some images (1.23, 10.3 and 20 {\micron})
are relatively noisy at faint fluxes, only five contours with
intervals of half an order of magnitude are shown here. For images at
250 and 500 {\micron}, dashed contours have intervals of 0.1 dex.}
\label{fig:img_60_conv}
\end{center}
\end{figure*}
\begin{figure*}
\begin{center}
\includegraphics[width=0.8\textwidth]{img_30_new.eps}\\
\caption{Same as Fig. \ref{fig:img_60}, except at inclination of $30^\circ$.}
\label{fig:img_30}
\end{center}
\end{figure*}
\begin{figure*}
\begin{center}
\includegraphics[width=0.8\textwidth]{img_30_conv_new.eps}\\
\caption{Same as Fig. \ref{fig:img_60_conv}, except at inclination of $30^\circ$ and that for images at 1.23, 10.3 and 20 {\micron}, only five contours with intervals of an order of magnitude are shown.}
\label{fig:img_30_conv}
\end{center}
\end{figure*}
\begin{figure}
\begin{center}
\includegraphics[width=\columnwidth]{fluxdist.ps}\\
\caption{Flux distribution along the projected outflow axis. Fluxes
are normalized by the mean value of a narrow strip across the core
along the axis. Different curves show different wavelengths. The
upper panel is at inclination of $60^\circ$ and the lower panel is
at $30^\circ$.}
\label{fig:fluxdist}
\end{center}
\end{figure}
Model 8, 8l and 8h compare the effect of different mean surface
densities of clump in which the core is embedded, which affects the
surface pressure of the core, and therefore the size of the core, the
accretion rate, the disk structure and the protostar evolution. The
size of the core, the radius of the expansion wave front, the position
of the sonic point and the disk radius are all proportional to
$\Sigma_\mathrm{cl}^{-1/2}$, so with higher clump surface density, the
core is more compact and accordingly the accretion rate is higher. The
scale hight of the disk is also larger in the high
$\Sigma_\mathrm{cl}$ case. The stellar luminosities of these three
models are similar but Model 8h has a much bluer stellar spectrum,
because of the higher hot-spot accretion luminosity.
Some important parameters of these three models are listed in
Table. \ref{table2}
Fig. \ref{fig:sed_sigma} compares SEDs of these three models at both
$60^\circ$ and $30^\circ$. As discussed above, higher surface density
leads to higher bolometric luminosities, which can be seen from the
SEDs: At inclination of $30^\circ$, with a higher surface density, the
flux is higher at all wavelengths; And at $60^\circ$ inclination we
can see the same effect in optical, near-IR and far-IR
emissions. However, in mid-IR, Model 8l shows a rise of the flux and a
very significant 10 {\micron} silicate feature, while the other two
models do not. To explain this, we also show the disk flux in the
lower panels. Here, disk flux contains photons which have their last
emission in the disk, and then either escape the core directly or are
scattered before they reach the observer. At the inclination of
$60^\circ$, the disk is blocked by the envelope. The short-wavelength
fluxes from the disk should all have been scattered. They keep the
trend that the model with higher surface density has higher
fluxes. However, in mid- and far-IR, the fluxes should have suffered
the extinction of the envelope, making the model with higher surface
density have lower fluxes because of the higher extinction. Thus, even
though Model 8 and 8h have very strong silicate features in their disk
SEDs, they are buried in the envelope fluxes in the total SEDs, while
Model 8l shows the high disk flux level in mid-IR and a significant
silicate absorption feature because of its lower extinction and lower
envelope flux.
\subsection{Images}
\subsubsection{Images of the Fiducial Model}
Fig. \ref{fig:img_60}, \ref{fig:img_60_conv}, \ref{fig:img_30} and
\ref{fig:img_30_conv} show the images for the fiducial model (Model 8)
at inclinations of $60^\circ$ and $30^\circ$, assuming a distance of 1
kpc and no foreground extinction. The instruments filters chosen here
are: 2MASS J, H, K bands, Spitzer IRAC 3.5 {\micron}, 4.5 {\micron},
5.8 {\micron} and 8 {\micron}, MIPS 24 {\micron}, GTC-CanariCam N band
and Q wide band, SOFIA FORCAST 37.1 {\micron}, Herschel PACS 70
{\micron}, 100 {\micron}, 160 {\micron}, and SPIRE 250 {\micron} and
500 {\micron}. We show both resolved images and those after
convolving with PSFs of the particular instruments. Images showed here
are all normalized to their maximum surface brightnesses which are
labeled on each image. $10^9$ photon packets are used. However they
are still not enough to produce a smooth image for the narrow filters
or at the wavelengths with low fluxes. Besides, the simulation grid
also contributes to this problem, like the strip patterns in the
images at some wavelengths (e.g. 20 {\micron}). Around the opening
angle of the outflow cavity, the grid with very small change of the
polar angle intersects with the cavity wall at quite different
radii. Even we have made the grid much finer in polar angle at the
region around the cavity wall, these patterns can still be seen,
especially at outer regions of the envelope. A finer grid would demand
more memory and computing time.
On the images at inclination of $60^\circ$ before convolution with the
instrument PSF, the most significant features are the outflow
cavities. They can be seen in any wavelength, though they become not
so obvious in far-IR wavelengths where the thermal emission of the
envelope dominates. At this inclination, the line of sight intersects
with the envelope, thus in near-IR the only emission we can see is
scattered by the cavity wall and escapes from the opening region,
especially from the side facing us. Deeper regions appear as the
wavelength increases. The central protostar and disk begins to show up
in mid-IR images. The dark lane around it shows the size of the
disk. In the mid-IR, the cavity walls dominates the emissions, as has
been discussed by \citet[]{deBuizer06}. Both sides can be clearly seen
and the brightest region is the base of the cavity. The emission of
the envelope takes over at longer wavelengths, making the image
symmetric and the outflow cavity begins to fade. The central star and
disk can still be seen as the brightest region in those wavelengths.
At the inclination of $30^\circ$, the central object can always be
seen through the empty outflow cavity. At shorter wavelengths, only
the side facing us of the cavity is significant and the opposite side
is very dim. Far-IR emission is dominated by the envelope and it is
hard to tell the features of the cavities. It should be noted here,
especially when comparing the 30$^\circ$ and 60$^\circ$ images, that
the images are all scaled to their maximum surface brightness, which
is generally much greater for the 30$^\circ$ viewing angles.
After convolving with PSF of particular instruments, in the far-IR,
the contours become very symmetric and we cannot tell the inclination
or the opposite outflow cavities. It is possible to see the opposite
outflow in MIPS 24 {\micron} and FORCAST 37.1 {\micron} at high
inclinations, if the S/N is large enough. The two sides of the
outflow cavities are clear in IRAC images and in near-IR
images. GTC-CanariCam has very high resolution so it may enable us to
see the central region in mid-IR.
Fig. \ref{fig:fluxdist} shows the flux profile along the system axis
at inclinations of both $60^\circ$ and $30^\circ$. Each profile is
normalized by the mean flux of a very narrow strip along the axis
across the whole core. Thus, the general shape and level of the
profiles are independent on the resolution of the image. But the curve
would be smoothed out if a larger PSF-FWHM is used. At inclination of
$60^\circ$, as we discussed above, at short wavelengths, the central
region is totally blocked by the envelope, most of the emission comes
from the outflow cavity. In the mid-IR, the maximum surface brightness
comes from the base of the outflow cavity rather than central star and
disk, on the side the outflow cavity is opposite to us, the flux drops
very fast by almost four orders of magnitude, while on the side the
outflow is facing us the flux drops much more gradually. In the
far-IR, the profile is symmetric on both sides, and central region has
the maximum surface brightness. At lower inclination, the maximum flux
always comes from the center. In the far-IR, the profile is very
similar to the case in higher inclination. At shorter wavelengths, the
contrast between outer regions and the center point is much higher
than the previous case, so that the profile drops very fast. At 1.66
{\micron} and 10.3 {\micron} we can only see the flux due to the
outflow cavity facing us, while the other side only shows up at 20
{\micron}. Because in our models it is empty outside the core, the
profiles all cut off at the core radius, which is not true in reality
since clump material can extend to larger radii.
\subsubsection{Effect of Different Surface Density}
\begin{figure*}
\begin{center}
\includegraphics[width=0.8\textwidth]{img_sigma_60_new.ps}\\
\caption{Images for Model 8, Model 8l and Model 8h at the inclination
of $60^\circ$, assuming a distance of 1 kpc. The size of the image
is 50" $\times$ 50". The resolution here is set to be 0.5", so the
surface brightnesses can be compared. Each images are scaled to the
maximum brightness of Model 8h at that band, but their own maximum
brightness is also labelled in the bottom-left corners. The total
flux in each image is also labelled in the top-right corner. The
blue contour is 1\% of the maximum surface brightness in that
image.}
\label{fig:img_sigma_60}
\end{center}
\end{figure*}
Fig. \ref{fig:img_sigma_60} compares the images of Model 8, 8l and 8h
at 6 wavelengths. The size of the images are all 50"$\times$50". We
can see that the core is smaller when the surface pressure is
higher. The surface brightness is dependent on the resolution, so here
we convolved all images with same PSF with FWHM of 0.5". The images
are scaled to the maximum surface brightness of Model 8h at each
wavelengths, so the brightnesses of the three images at each
wavelength can be directly compared.
The total fluxes and the maximum surface brightness in images of Model
8h are generally higher than those of the other two models. This is
natural since the total luminosity is much higher in this model. The
optical depth is larger when $\Sigma_\mathrm{cl}$ is
higher.
In the near-IR, we can see the base of the outflow
cavity facing us and most of the opposite side in Model 8l, while we
can only see the light from the opening region of the cavities in
Model 8h. The central region shows up in Model 8h only in
far-IR, while it can be seen at mid-IR in Model 8 and 8l. With lower
surface density, the contrast between two sides of the outflow
cavities becomes smaller. In Model 8h the side towards us is much
brighter than the other side, even at 70 {\micron}, while the other
two models show almost symmetric images.
We can also see that the mid-IR emission is dominated by the cavity walls,
especially in the cases with higher extinction, as opposed to that with lower extinction
the emissions are more concentrated to the central region.
\section{Discussions and Conclusions}
\label{sec:conclusions}
We have constructed a model for individual massive star formation,
with a $60\;M_\odot$ initial core that forms an $8M_\odot$ protostar. We included the
inside-out expansion wave in the core, free-fall rotating collapse in
the inner region, an accretion disk of 1/3 the mass of the star, and
bipolar outflow cavities with large opening angle, parameters of which
are all calculated self-consistently. For the first time, we
considered an optically thick inner disk with gas opacities which were
assigned depending on the pre-calculated temperature profile. This
inner disk enabled us to calculate the emission from the active disk
with correct total luminosity and spectrum.
Compared to the Robitaille model, our parameter space covers higher
accretion rates, higher disk mass, denser envelopes and larger outflow
cavities. Also, in the Robitaille model the disk accretion luminosity
is much lower than that in our model since the disk is truncated at
the dust destruction radius, while the hot-spot luminosity on the
stellar surface is set to be $Gm_*\dot{m}(1/r_*-1/r_\mathrm{co})$
where $r_\mathrm{co}$ is the magnetic co-rotation radius and set to be
$5r_*$, making the hot-spot luminosity is 4/5 of the total accretion
luminosity $L_\mathrm{acc}$. So energy of
$\lessapprox0.2\;L_\mathrm{acc}$ is lost. In our model, the total
accretion luminosity is divided equally into disk accretion luminosity
(emitted from the disk extended to stellar radius with gas and dust
opacities) and hot-spot luminosity (added on the stellar
spectrum). Also, the Ulrich solution is used for the whole envelope in
Robitaille model, which means the whole envelope is assumed to be
free-falling and the infalling rate is constant at all radii. As
discussed in Section \ref{sec:rotating}, the core undergoes free-fall
only in the inner region. Therefore, for a given core mass and
accretion rate, our model has a more compact core with higher
extinction, causing the far-IR peak of the SED to be at lower fluxes
and at longer wavelengths.
We have presented SEDs of the model series, the fiducial model, and
the models with higher and lower surface pressure, at typical
inclinations. We also have presented images for the fiducial model at
JHK bands, IRAC and MIPS bands, GTC-CanariCam bands, SOFIA bands and
Herschel bands, both resolved and convolved with the resolutions of
each instrument. The main conclusions can be summarized as:
1. Outflow cavities affect the SEDs significantly and cause a large
difference between low and high inclinations of our viewing
angle. However, the present modeling assumes these cavities are
optically thin, which may not be valid, especially at the shorter
wavelengths. This issue will be investigated in a follow-up paper.
The height of the disk also affects the SEDs. With a thicker
disk, the near- and mid-IR fluxes at low inclinations become higher,
while at high inclinations it suppresses the fluxes (especially at
short wavelengths) and creates deep silicate features. Also, the
density distribution in the core (especially the inner region) can
affect the mid-IR flux levels, the silicate features, and the
far-IR peaks.
2. The temperature of the innermost region of the disk can reach $\sim
10^5$ K. The disk becomes optically thick in such conditions even if
no dust can exist there. SEDs show the rise of optical emission due to
this hot inner disk. This optically thick inner disk also shields flux
from the protostar, leading to lower temperatures on the surface of
the disk and the base of the cavity wall, and therefore lower fluxes in
near-IR and mid-IR part of the SEDs.
3. The SEDs of the fiducial model show that the inclination can affect
SEDs at wavelengths shorter than 100 {\micron}, including the far-IR
peak. The mid-IR spectral slope changes significantly with
inclination. The silicate feature changes from a deep absorption
feature to an emission feature from high inclinations to low
inclinations, due to the change of the optical depth.
4. With higher surface pressure, the core becomes more compact and the
accretion rate and luminosity become higher, leading to higher fluxes
at all wavelengths (except in the mid-IR for high inclination cases
which suffer higher extinction).
High extinction can also cause the mid-IR flux to be dominated by the
envelope, and thus hide the silicate absorption feature. Thus, only in
the model with low extinction does the silicate absorption feature
show up.
5. Outflow cavities are the most significant features on the images,
except at wavelengths longer than 70 {\micron}. At inclinations of
$60^\circ$, from short wavelengths to long wavelengths, the brightest
point moves from the outer region of the cavity to the base of the
cavity wall, and to the center of the core in far-IR, while at
inclination of $30^\circ$, it is always the central region. At
inclination of $60^\circ$, the opposite outflow cavity can be seen
n the mid-IR if fluxes $\sim 10^2 - 10^3$ times fainter than the peak
can be probed. It is very difficult to see the opposite cavity at an
inclination of $30^\circ$. GTC-CanariCam (and other 10 m class
telescopes with mid-IR cameras) has very high angular resolution so it
may enable us to resolve the central disk system. The flux distribution
along the outflow axis can help constrain model assumptions and
inclination angles and so will useful to measure.
In the mid-IR, the cavity walls seem to dominate the emission, but for
lower density cores with lower extinction, the central region becomes
brighter. The contrast between the two sides of the outflows in mid-
and far-IR increases as the extinction becomes higher.
The model we have presented will be improved in future papers by a
detailed consideration of the effect of including gas and dust in the
outflow cavity. Currently models for the density distribution here are
quite uncertain, which is why we defer this study to a future date. An
evolutionary sequence of protostellar models with and without outflow
opacity will then be presented. To gauge the degree of inhomogeneity
we will consider the results of numerical simulations of core
accretion and outflow interaction (e.g. \citealt[]{Krumholz07},
\citealt[]{Staff10}), and then incorporate these inhomogeneities into
the models in a parametrized form.
\acknowledgements
We thank James De Buizer and Barbara Whitney for helpful discussions.
YZ acknowledges support from an Alumni Fellowship of the University of
Florida. JCT acknowledges support from NSF CAREER grant AST-0645412;
NASA Astrophysics Theory and Fundamental Physics grant ATP09-0094; and
NASA Astrophysics Data Analysis Program ADAP10-0110.
|
\section{Introduction}
To calculate the band structure of a crystal, some initial guess of the
electron wave-functions is required. One of the rare cases where analytic
solutions to the Schr\"odinger equation are known for a semiconductor or an
insulator are separable potentials that can be written as the sum of three
one-dimensional square-wave potentials. Such separable potentials were first
considered by Kronig and Penney\cite{de_l._kronig_quantum_1931}. There is a
simple analytic expression for the energy -- wave-number dispersion relation
for the well-known problem of an electron moving in a one-dimensional
square-wave potential. Here we show that there is also an analytic expression for
the density of states for the one-dimensional problem. These results can be
simply combined to produce the dispersion relations and density of states of
separable three dimensional potentials. A limiting case of this class of
potentials is the widely used free electron model which occurs when the
amplitude of the square-wave is zero. As the amplitude of the potential
increases, the potentials can correspond to metals, semiconductors, or
insulators. Normally determining the electronic density of states from the
band structure is a computationally intensive process that requires sampling
many points of the Brillouin zone. For the separable square-wave potentials,
the electronic density of states is however easily calculable as a
convolution of three analytically known functions. This makes it easy to
calculate the thermodynamic properties corresponding to this class of
potentials. From the density of states we can calculate the temperature
dependence of thermodynamic quantities such as the chemical potential, the
internal energy, the specific heat, the entropy, and the Helmholtz free
energy. Half a million plots of the thermodynamic properties of separable
potentials are available as online supplementary material\cite{supp}. The
collective data set allows us to
consider how an increase in the amplitude of the potential causes a
transition from a metal to an insulator. There are cases where increasing the
amplitude of the potential causes a transition from a metal to an insulator,
then back to a metal and then finally back to an insulator.
While band gaps appear in the one-dimensional problem for an arbitrarily small
amplitude of the periodic potential, a finite amplitude is needed for the creation
of band gaps in three dimensions. In section \ref{sec:3} we quantify how large
the amplitude of the potential must be for bands to be formed for the specific
class of three-dimensional potentials we have considered. This minimum amplitude is,
\begin{equation}
V = 0.9\frac{{\pi ^2 \hbar ^2 }}{{ma^2 }}.
\label{eq:1}
\end{equation}
Here $V$ is the amplitude of the potential, $\hbar$ is the reduced Planck's constant, $m$
is the mass of an electron, and $a$ is the lattice constant. The amplitude of the potential
necessary for the formation of bands depends strongly on the lattice constant.
\section{Solutions of the Schr\"odinger equation for the Kronig-Penney potential}
This section reviews the solutions to the Kronig-Penney
model\cite{mcquarrie_1996,szumlowicz_1997} and derives an expression for
the one-dimensional density of states which will be used in section \ref{sec:3} to construct
the dispersion relations and densities of states for three-dimensional separable potentials.
Figure \ref{fig:1} shows the one-dimensional potential that was first considered by Kronig
and Penney\cite{de_l._kronig_quantum_1931}.
\begin{figure}
\includegraphics{Figure1}
\caption{\label{fig:1} One-dimensional square-wave potential.}
\end{figure}
Because of the translational symmetry of the potential, the eigenfunctions of the Hamiltonian
are simultaneously eigenfunctions of the translation operator. The eigenfunctions of the
translation operator $T$ can be readily constructed from any two linearly independent solutions
of the one-dimensional Schr\"odinger equation. A convenient choice is,
\begin{equation}
\psi _1 (0) = 1,{\rm{ }}\frac{{d\psi _1 }}{{dx}}(0) = 0,{\rm{ }}\psi _2 (0) = 0,{\rm{ }}\frac{{d\psi _2 }}{{dx}}(0) = 1.
\label{eq:2}
\end{equation}
The solutions in region 1 ($0 < x < b$) are,
\begin{equation}
\psi _1 (x) = \cos (k_1 x),{\rm{ }}\psi _2 (x) = \frac{{\sin (k_1 x)}}{{k_1 }},
\label{eq:3}
\end{equation}
while the solutions in region 2 ($b < x < a$) are\cite{merzbacher_1961},
\begin{subequations}
\label{eq:4}
\begin{align}
\psi _1 (x) = \cos (k_2 (x - b))\cos (k_1 b) \nonumber\\
- \frac{k_1}{k_2} \sin (k_2 (x - b))\sin (k_1 b),
\end{align}
\begin{align}
\psi _2 (x) = \frac{{1}}{{k_1 }} \cos (k_2 (x - b))\sin (k_1 b)\nonumber\\
+ \frac{{1}}{{k_2 }} \sin (k_2 (x - b))\cos (k_1 b).
\end{align}
\end{subequations}
Here $k_r = \sqrt {2m\left( {E - V_r } \right)/\hbar ^2 }$; $r = 1,2$ is the label of the
region, and $E$ is the energy. For energies where $k_r$ is imaginary, the solutions are still real
since $\cos(i\theta) = \cosh(\theta)$ and $\sin(i\theta) = i \sinh(\theta)$.
Any other solution can be written as a linear combination of $\psi_1(x)$ and $\psi_2(x)$. In
particular, $\psi_1(x + a)$ and $\psi_2(x + a)$ can be written in terms of $\psi_1(x)$ and
$\psi_2(x)$. These solutions are related to each other by the matrix representation of the
translation operator\cite{magnus_hills_1966}.
\begin{equation}
\left[ {\begin{array}{*{20}c}
{\psi _1 \left( {x + a} \right)} \\
{\psi _2 \left( {x + a} \right)} \\
\end{array}} \right] = \left[ {\begin{array}{*{20}c}
{T_{11} } & {T_{12} } \\
{T_{21} } & {T_{22} } \\
\end{array}} \right]\left[ {\begin{array}{*{20}c}
{\psi _1 \left( x \right)} \\
{\psi _2 \left( x \right)} \\
\end{array}} \right].
\label{eq:5}
\end{equation}
The elements of the translation matrix can be determined by evaluating Eq. \eqref{eq:5} and
its derivative at $x = 0$.
\begin{equation}
\left[ {\begin{array}{*{20}c}
{\psi _1 \left( {x + a} \right)} \\
{\psi _2 \left( {x + a} \right)} \\
\end{array}} \right] = \left[ {\begin{array}{*{20}c}
{\psi _1 \left( a \right)} & {\frac{{d\psi _1 }}{{dx}}\left( a \right)} \\
{\psi _2 \left( a \right)} & {\frac{{d\psi _2 }}{{dx}}\left( a \right)} \\
\end{array}} \right]\left[ {\begin{array}{*{20}c}
{\psi _1 \left( x \right)} \\
{\psi _2 \left( x \right)} \\
\end{array}} \right].
\label{eq:6}
\end{equation}
The eigenfunctions and eigenvalues $\lambda$ of the translation operator are,
\begin{align}
\psi _ \pm \left( x \right) = \frac{{2\psi _2 (a)}}{{\frac{{d\psi _2 (a)}}{{dx}} - \psi _1 (a) \pm \delta }}\psi _1 (x) + \psi _2 (x),\nonumber\\
{\rm{ }}\lambda _ \pm = \frac{1}{2}\left( {\alpha \pm \delta } \right),
\label{eq:7}
\end{align}
where $\delta = \sqrt {\alpha ^2 - 4} {\rm{ }}$ and
\begin{align}
\alpha = \psi _1 (a) + \frac{{d\psi _2 (a)}}{{dx}} = 2\cos \left( {k_2 \left( {a - b} \right)} \right)\cos \left( {k_1 b} \right) \nonumber\\
- \left( {\frac{{k_2 }}{{k_1 }} + \frac{{k_1 }}{{k_2 }}} \right)\sin \left( {k_2 \left( {a - b} \right)} \right)\sin \left( {k_1 b} \right).
\label{eq:8}
\end{align}
If periodic boundary conditions are used for a potential with $N$ unit cells, then
applying the translation operator $N$ times brings the function back to its original
position
\begin{equation}
T^N \psi \left( x \right) = \psi \left( {x + Na} \right) = \lambda ^N \psi \left( x \right) = \psi \left( x \right).
\label{eq:9}
\end{equation}
The eigenvalues of the translation operator are therefore the solutions to the equation
$\lambda^N = 1$. These solutions are,
\begin{equation}
\lambda _j = \exp \left( \frac{i2\pi j}{N} \right) = \exp \left( \frac{i2\pi aj}{L} \right) = \exp \left( {ik_j a} \right),
\label{eq:10}
\end{equation}
where $j$ is an integer between $-N/2$ and $N/2$, $L = Na$ is the length of the crystal,
and $k_j = 2\pi j/L$ are the allowed $k$ values in the first Brillouin zone. The dispersion
relation can be determined by first calculating $\alpha$ for a specific energy and then
solving Eqs. \eqref{eq:7} and \eqref{eq:10} for the wave-number\cite{bender_1999},
\begin{equation}
k = \pm \frac{1}{a}\tan ^{ - 1} \left( {\frac{{\sqrt {4 - \alpha ^2 } }}{\alpha }} \right).
\label{eq:11}
\end{equation}
\begin{figure*}
\includegraphics{Figure2}
\caption{\label{fig:2}(a) The energy -- wave-number dispersion relation. The dashed
line is the Fermi energy. (b) The density of states. (c) The internal energy
density (solid line) and Helmholtz free energy density (dashed line). (d) The
chemical potential (solid line) and the specific heat (dashed line). All of the
plots were drawn for a square-wave potential with the parameters: $V_1 = \unit[0]{eV}$,
$V_2 = \unit[12.5]{eV}$, $a = \unit[2 \times 10^{-10}]{m}$, $b = \unit[5 \times 10^{-11}]{m}$,
and an electron density of $n = \unit[3]{}${electrons/primitive cell}.}
\end{figure*}
The dispersion relation can be used to determine the density of states which is
needed to calculate the thermodynamic properties of a system of noninteracting
electrons\cite{lin_2003}. The one-dimensional density of states in $k$ space is $D(k) = 2/\pi$
and thus the density of states in energy is,
\begin{equation}
D(E) =
\begin{cases}
\frac{2}{\pi }\frac{{dk}}{{dE}} = \frac{2}{\pi }\frac{{dk}}{{d\alpha }}\frac{{d\alpha }}{{dE}} & \left| \alpha \right| < 2 \\
0 & \left| \alpha \right| > 2
\end{cases}
\label{eq:12}
\end{equation}
where
\begin{equation*}
\frac{{dk}}{{d\alpha }} = \frac{1}{{a\sqrt {4 - \alpha ^2 } }},
\end{equation*}
and
\begin{widetext}
\begin{align}
\frac{{d\alpha }}{{dE}} = - \frac{m}{{\hbar ^2 }}\left( {\frac{{2b}}{{k_1 }} + \left( {\frac{{k_2 }}{{k_1 }} + \frac{{k_1 }}{{k_2 }}} \right)\frac{{\left( {a - b} \right)}}{{k_2 }}} \right)\cos \left( {k_2 \left( {a - b} \right)} \right) \sin \left( {k_1 b} \right) - \frac{m}{{\hbar ^2 }}\left( {2\frac{{\left( {a - b} \right)}}{{k_2 }} + \left( {\frac{{k_2 }}{{k_1 }} + \frac{{k_1 }}{{k_2 }}} \right)\frac{b}{{k_1 }}} \right) \nonumber\\
\times \sin \left( {k_2 \left( {a - b} \right)} \right) \cos \left( {k_1 b} \right) + \frac{m}{{\hbar ^2 }}\left( {\left( {\frac{{k_2 }}{{k_1^2 }} - \frac{1}{{k_2 }}} \right)\frac{1}{{k_1 }} + \left( {\frac{{k_1 }}{{k_2^2 }} - \frac{1}{{k_1 }}} \right)\frac{1}{{k_2 }}} \right) \sin \left( {k_2 \left( {a - b} \right)} \right)\sin \left( {k_1 b} \right).
\label{eq:13}
\end{align}
\end{widetext}
Thermodynamic properties such as the chemical potential, the internal energy,
the specific heat, the entropy, or the Helmholtz free energy of a system of
noninteracting fermions can be calculated from the electron density $n$, the density
of states $D(E)$, and the temperature $T$.
The chemical potential $\mu$ is implicitly defined by the relation,
\begin{equation}
n = \int\limits_{ - \infty }^\infty {D(E)F(E)dE} ,
\label{eq:14}
\end{equation}
where $F(E)$ is the Fermi function,
\begin{equation}
F(E) = \frac{1}{{\exp \left( {\frac{{E - \mu }}{{k_B T}}} \right) + 1}}.
\label{eq:15}
\end{equation}
Here $k_B$ is Boltzmann's constant. Once the chemical potential has been determined,
it can be used to calculate the internal energy density $u$ and the Helmholtz free
energy density $f$.
\begin{equation}
u = \int\limits_{ - \infty }^\infty {ED(E)F(E)dE}
\label{eq:16}
\end{equation}
\begin{equation}
f = \mu n - k_B T\int\limits_{ - \infty }^\infty {D(E)\ln \left( {1 + \exp \left( {\frac{{\mu - E}}{{k_B T}}} \right)} \right)dE}
\label{eq:17}
\end{equation}
Finally, the specific heat and the entropy are given by the partial derivatives,
\begin{equation}
c_v = \left. {\frac{{\partial u}}{{\partial T}}} \right|_{N,V} \text{ and }
s = - \left. {\frac{{\partial f}}{{\partial T}}} \right|_{N,V}.
\label{eq:18}
\end{equation}
The band structure of a one dimensional potential and the corresponding thermodynamic
properties are plotted in Fig. \ref{fig:2}.
\section{Three-dimensional separable potentials} \label{sec:3}
In this section, 5 of the calculated band structures that are available in
the supplementary material are presented. This illustrates the variety of results
that can be obtained with this simple model. The five examples are a free electron
gas, a metal where the charge carriers at the Fermi surface are electron-like, a
metal where the charge carriers at the Fermi surface are hole-like, a direct band
gap semiconductor, and an indirect band gap semiconductor.
The dispersion relation and density of states for any three-dimensional potential
of the form
\begin{equation}
U_{3d}(x,y,z) = U_x(x) + U_y(y) + U_z(z)
\label{eq:19}
\end{equation}
are easily calculated from the one-dimensional results\cite{berezin_1986}.
The energy of an electron in a three-dimensional separable potential is the sum of the
energies of the constituent one-dimensional potentials. The three-dimensional density of
states is the convolution of the three one-dimensional densities of states. Once the
three-dimensional density of states is known, the thermodynamic quantities can be
calculated as outlined above.
The one-dimensional bands can be indexed by integers, 1 corresponding to the band
with the lowest energy, 2 to the band with the next lowest energy, etc. The three-dimensional
bands can then be indexed by the three integers that correspond to the
one-dimensional bands that make up the three-dimensional band. The three dimensional
band with the lowest energy is the 111 band. For cubic crystals, the next three
bands 211, 121, and 112 are degenerate in energy.
Figure \ref{fig:3} shows the band structure, the density of states, and corresponding
thermodynamic properties for a constant potential that corresponds to a free electron
gas. The band structure in Fig. \ref{fig:3}a is plotted along a path in the Brillouin
zone going from the M point (0.5, 0.5, 0) through the points $\Gamma$ (0, 0, 0),
X(0.5, 0, 0), M, R (0.5, 0.5, 0.5), X, $\Gamma$, ending at the point R. All of the
standard results for a free electron gas are reproduced by the numerical calculation.
The dispersion relation is parabolic; the density of states increases with the square
root of the energy and the specific heat is a linear function of the temperature.
\begin{figure*}
\includegraphics{Figure3}
\caption{\label{fig:3}The band structure and thermodynamic properties of a free electron
gas. (a) The energy-momentum dispersion relation. The dashed line is the Fermi energy.
(b) The density of states. (c) The internal energy density (solid line) and the Helmholtz
free energy density (dashed line). (d) The chemical potential (solid line) and the
specific heat (dashed line). All plots are for a cubic potential with the parameters:
$V_1 = \unit[0]{eV}$, $V_2 = \unit[0]{eV}$, $a = \unit[0.2]{nm}$, $b = \unit[0.09]{nm}$,
and an electron density of $n = \unit[2]{}${electrons/primitive cell}.}
\end{figure*}
Figures \ref{fig:4} and \ref{fig:5} show the results for a potential with an amplitude
of $\unit[4.2]{eV}$, $a = \unit[0.2]{nm}$, and $b = \unit[0.09]{nm}$. Gaps open in the
dispersion relation at the Brillouin zone boundaries and some kinks appear in the density
of states. These kinks make both the analytical and numerical evaluation of the
thermodynamic properties difficult. Both Figs. \ref{fig:4} and \ref{fig:5} correspond
to metals but in Fig. \ref{fig:4} the electron density is $\unit[2]{}$electrons/primitive
cell and the density of states is increasing at the Fermi energy while in Fig. \ref{fig:5} the
electron density is $\unit[1]{}$electron/unit cell and the density of states is decreasing
at the Fermi energy. The material in Fig. \ref{fig:4} has a Fermi surface that consists
of electron-like states and the chemical potential decreases with increasing temperature
as it does for the free electron gas of Fig. \ref{fig:3}. The material in Fig.
\ref{fig:5} has a Fermi surface that consists of hole-like states and consequently the
chemical potential increases with increasing temperature.
\begin{figure*}
\includegraphics{Figure4}
\caption{\label{fig:4}The band structure and thermodynamic properties for a cubic potential
$V_1 = \unit[0]{eV}$, $V_2 = \unit[4.2]{eV}$, $a = \unit[0.2]{nm}$, $b = \unit[0.09]{nm}$
and $n = \unit[2]{}$electrons/primitive cell. (a)-(d) as in Fig. \ref{fig:3}.}
\end{figure*}
\begin{figure*}
\includegraphics{Figure5}
\caption{\label{fig:5}The band structure and thermodynamic properties for a cubic potential
$V_1 = \unit[0]{eV}$, $V_2 = \unit[4.2]{eV}$, $a = \unit[0.2]{nm}$, $b = \unit[0.09]{nm}$
and $n = \unit[1]{}$electrons/primitive cell. (a)-(d) as in Fig. \ref{fig:3}.}
\end{figure*}
Figure \ref{fig:6} shows the band structure for a material with two band gaps.
The lower band gap is indirect with a band gap energy of $\unit[3.9]{eV}$. The
upper band is direct with a band gap energy of $\unit[0.4]{eV}$. If the material
shown in Fig. \ref{fig:6} had $\unit[2]{}$electrons/primitive cell, the lowest
band would be completely filled and it would be an insulator. For an electron
density of $\unit[8]{}$electrons/primitive cell, the Fermi energy lies in the
second band gap (illustrated by the dashed line in Fig. \ref{fig:6}a. The
thermodynamic properties in Figs. \ref{fig:6}c and \ref{fig:6}d were calculated
assuming an electron density of $\unit[8]{}$electrons/primitive cell. Since this
material is a semiconductor, the electronic contribution to the specific heat is
negligible at low temperature and increases exponentially at high temperatures.
The chemical potential of a semiconductor is a linear function of the temperature.
\begin{figure*}
\includegraphics{Figure6}
\caption{\label{fig:6}The band structure and thermodynamic properties for a cubic potential
$V_1 = \unit[0]{eV}$, $V_2 = \unit[8.8]{eV}$, $a = \unit[0.4]{nm}$, $b = \unit[0.21]{nm}$
and $n = \unit[8]{}$electrons/primitive cell. This band structure corresponds to a direct
band gap semiconductor. (a)-(d) as in Fig. \ref{fig:3}.}
\end{figure*}
Figure \ref{fig:7} shows the band structure for a potential where the 311 band has
moved down lower than the 221 band. In this case the lower band is indirect with
a band gap of $\unit[4.7]{eV}$ and the upper band gap is also indirect with a
band gap energy of $\unit[0.5]{eV}$. For an electron density of
$\unit[8]{}$electrons/primitive cell, the Fermi energy lies in the second band
gap where the dashed line is drawn and the material is an indirect band gap
semiconductor.
\begin{figure*}
\includegraphics{Figure7}
\caption{\label{fig:7}The band structure and thermodynamic properties for a cubic potential
$V_1 = \unit[0]{eV}$, $V_2 = \unit[9.5]{eV}$, $a = \unit[0.4]{nm}$, $b = \unit[0.19]{nm}$
and $n = \unit[8]{}$electrons/primitive cell. This band structure corresponds to an indirect
band gap semiconductor. (a)-(d) as in Fig. \ref{fig:3}.}
\end{figure*}
The dispersion relations, densities of states, and thermodynamic properties like
those shown in Fig. \ref{fig:3}-\ref{fig:7} were calculated for 10000 cubic potentials.
For these calculations, it is convenient to normalize the Schr\"odinger equation
so that length is measured in terms of the lattice constant $a$. When this is done,
energies are measured in units of $\pi^2\hbar^2 / 2ma^2$ and there are only two
independent parameters in the problem, which can be taken to be $\tilde{b} = b/a$
and $\tilde{V} = 2ma^2V / \pi^2\hbar^2$. Here $V = V_2 - V_1$. The calculations
were performed for normalized parameters in the range $\tilde{b} = \left( 0,1 \right)$
and $\tilde{V} = \left( 0,27 \right)$. It becomes more difficult to perform the
numerical calculations of the thermodynamic properties for large values of $\tilde{V}$.
The upper limit of $\tilde{V} = 27$ corresponds to the amplitude where our
calculations of the thermodynamic properties become unreliable.
Figure \ref{fig:8} shows the sizes of the first two band gaps that occur. The
three-dimensional potential must have a finite amplitude for there to be a band gap.
\begin{figure*}
\centering
\includegraphics{Figure8a}\hfill
\includegraphics{Figure8b}
\caption{\label{fig:8} (a) The band gap of the lowest gap between band 111 and the
degenerate bands 112, 121, and 211 in units of $\pi^2\hbar^2 / 2ma^2$. This band
gap is always indirect. (b) The second band gap. The dashed lines in (b) indicate
the transitions from direct to indirect band gaps. The indirect band gaps are in
between the two dashed lines. Note that there is a reentrant regime at $\tilde{b} = 0.3$
where an increase in the amplitude $V$ causes a transition from metal to insulator,
then back to a metal and finally to an insulator.}
\end{figure*}
\section{Thermodynamic properties of the metals} \label{sec:4}
The band structures of the potentials that were studied always correspond to
metals when there is no band gap. This region includes the free electron model
at $\tilde{V} = 0$ as well as $\delta$-function potentials in the regions
near $\tilde{b} = 0$ and $\tilde{b} = 1$. When band gaps occur, the situation
is more complicated. The potentials can correspond to metals, semiconductors,
or insulators depending on the electron density. We compared the thermodynamic
properties of these metals to results obtained by using the Sommerfeld expansion.
Sommerfeld showed that the electronic contribution to the thermodynamic
properties of metals can be approximated in terms of just two quantities: the
density of states at the Fermi energy $D(E_F)$ and the derivative of the density
of states at the Fermi energy\cite{sommerfeld_zur_1928}
\begin{equation}
D'(E_F ) = \left. {\frac{{dD(E)}}{{dE}}} \right|_{E_F }.
\label{eq:20}
\end{equation}
To lowest order in the temperature $T$, the Sommerfeld expression for the
chemical potential, the internal energy, specific heat, entropy, and Helmholtz
free energy are,
\begin{equation}
\mu \approx E_F - \frac{{\pi ^2 }}{6}\left( {k_B T} \right)^2 \frac{{D'(E_F )}}{{D(E_F )}},
\label{eq:21}
\end{equation}
\begin{equation}
u \approx u(T = 0) + \frac{{\pi ^2 D(E_F )}}{6}\left( {k_B T} \right)^2,
\label{eq:22}
\end{equation}
\begin{equation}
c_v \approx \frac{{\pi ^2 D(E_F )}}{3}k_B^2 T,
\label{eq:23}
\end{equation}
\begin{equation}
s \approx \frac{{\pi ^2 D(E_F )}}{3}k_B^2 T,
\label{eq:24}
\end{equation}
\begin{equation}
f \approx u(T = 0) - \frac{{\pi ^2 D(E_F )}}{6}\left( {k_B T} \right)^2.
\label{eq:25}
\end{equation}
Figure \ref{fig:9} shows the value of the density of states at the Fermi energy
and its derivative for an electron density of $n = \unit[1]{}$electron/primitive
cell. This figure makes it possible to estimate the thermodynamic properties of
the metals by substituting the values for $D(E_F)$ and $D'(E_F)$ into
Eqs. \eqref{eq:21}--\eqref{eq:25}.
\begin{figure*}
\includegraphics{Figure9}
\caption{\label{fig:9} The density of states at the Fermi energy and the
derivative of the density of states at the Fermi energy for an electron density
of $n = \unit[1]{}$electrons/primitive cell. It was not possible to determine
the density of states at the Fermi energy reliably in the crosshatched regions.
Similar plots for electron densities $\unit[2\text{--}10]{}$electrons/primitive cell are
available in the supplementary information online.}
\end{figure*}
The Sommerfeld expansion assumes that the density of states is a smoothly varying
function in an energy range about $k_BT$ wide in the vicinity of the Fermi energy.
However, the density of states must contain kinks known as Van Hove
singularities\cite{ziman_principles_1969}. If the Fermi energy is within an energy $k_BT$ of one
of the kinks, then the assumptions of the Sommerfeld theory are invalid. The numerical
simulations in the supplementary material show that the Sommerfeld approximation is
valid in the majority of cases where there are no band gaps. When gaps are present,
the bands are narrow and it is more likely that the Sommerfeld theory fails. In the
case of extremely narrow bands it becomes difficult to determine the temperature
dependence of the thermodynamic properties reliably because of the discontinuities
in the density of states.
\section{Thermodynamic properties of the semiconductors and insulators} \label{sec:5}
If the Fermi energy falls in a band gap so that the material is a semiconductor or an
insulator, the density of states at the Fermi energy and its derivative are not
suitable for describing the thermodynamic quantities. Instead, it is
common to specify the band gap and the effective masses of the electrons and holes.
For cubic crystals, the first gap appears above the 111 band and beneath the next
three bands (211, 121, and 112) which are degenerate due to symmetry reasons. Above
these three degenerate bands, a second energy gap sometimes appears. The Fermi
energy lies in the first gap for an electron density of 2 electrons per unit cell
and it can only be in the second gap for an electron density of 8 electrons per unit
cell.
The first gap is always indirect with the maximum of the valence band occurring at
R. The minimum in the conduction band for the first gap occurs at X. The second
band is sometimes direct and sometimes indirect. The valence band maximum of the
second gap is always at M. When the second band gap is direct, the conduction band
minimum is also at M (see Fig. \ref{fig:6}) but when the second band gap is
indirect, the conduction band minimum is at $\Gamma$ (see Fig. \ref{fig:7}).
The effective masses near the top of the valence band and bottom of the conduction
band can be found by linearizing $\alpha$ given by Eq. \eqref{eq:8} near the
band edges and inserting this into Eq. \eqref{eq:11}. The effective masses of
electrons and holes are,
\begin{subequations}
\label{eq:26}
\begin{equation}
m_e^* = \frac{{\hbar ^2 }}{{2a^2 }}\left( {\left. { - \frac{{d\alpha }}{{dE}}} \right|_{E = E_c } } \right) \text{ and}
\label{eq:26a}
\end{equation}
\begin{equation}
m_h^* = \frac{{\hbar ^2 }}{{2a^2 }}\left( {\left. {\frac{{d\alpha }}{{dE}}} \right|_{E = E_v } } \right).
\label{eq:26b}
\end{equation}
\end{subequations}
Here $E_c$ is the energy at the bottom of the conduction band and $E_v$ is the energy
at the top of the valence band. The effective masses are plotted in Fig. \ref{fig:10}.
\begin{figure*}
\centering
\includegraphics{Figure10a}\hfill
\includegraphics{Figure10b}\hfill
\includegraphics{Figure10c}\hfill
\includegraphics{Figure10d}
\caption{\label{fig:10} (a) Effective masses of electrons in the first band gap. (b)
Effective masses of holes in the first band gap. (c) The effective masses of electron
in the second band gap. The dashed lines indicate the transitions from direct to
indirect band gaps. The indirect band gaps are in between the two dashed lines.
(d) The effective masses of holes in the second band gap.}
\end{figure*}
The thermodynamic properties of semiconductors or insulators are typically calculated
using the Boltzmann approximation. In this approximation, the density of states
near the Fermi energy is described by the function,
\begin{equation}
D(E) =
\begin{cases}
\left( 2m_h^* \right)^{3/2} \frac{\sqrt{E_v - E}}{2 \pi^2 \hbar^3} & {E < E_v } \\
0 & {E_v < E < E_c } \\
\left( 2m_e^* \right)^{3/2} \frac{\sqrt{E - E_c}}{2 \pi^2 \hbar^3} & {E_c < E }
\end{cases}
\label{eq:27}
\end{equation}
In the Boltzmann approximation it is further assumed that in the conduction band
the Fermi function in Eqs. \eqref{eq:14},\eqref{eq:16}--\eqref{eq:18} can be replaced by
a Boltzmann factor $F(E) \approx exp((\mu-E)/k_BT)$ while in the valence band
$F(E) \approx 1 - exp((E-\mu)/k_BT)$. The result for the chemical potential of a
semiconductor in this approximation is found in many textbooks\cite{kittel_introduction_2005,ashcroft_solid_1976}.
The chemical potential of a semiconductor or insulator is a linear function of the
temperature.
\begin{equation}
\mu \approx \frac{{E_v + E_c }}{2} - \frac{3}{4}k_B T\ln \left( {\frac{{m_e^* }}{{m_h^* }}} \right)
\label{eq:28}
\end{equation}
The Boltzmann approximation can also be used to calculate the electronic contribution
to other thermodynamic quantities in terms of the band gap and the effective masses.
\begin{align}
u \approx u(T = 0) + \frac{{\sqrt {2\pi } }}{{2\pi ^2 \hbar ^3 }}\left( {m_e^* m_h^* } \right)^{3/4} \exp \left( {\frac{{ - E_g }}{{2k_B T}}} \right) \nonumber \\
\times \left( {k_B T} \right)^{3/2} \left( {3k_B T + E_g } \right),
\label{eq:29}
\end{align}
\begin{align}
c_v \approx \frac{{\sqrt {2\pi } }}{{2\pi ^2 \hbar ^3 }}\left( {m_e^* m_h^* } \right)^{3/4} \exp \left( {\frac{{ - E_g }}{{2k_B T}}} \right)\left( {k_B T} \right)^{3/2} \nonumber\\
\times \left( {\frac{{15}}{2}k_B + \frac{{3E_g }}{T} + \frac{{E_g^2 }}{{2k_B T^2 }}} \right),
\label{eq:30}
\end{align}
\begin{align}
s \approx \frac{{\sqrt {2\pi } }}{{2\pi ^2 \hbar ^3 }}\left( {m_e^* m_h^* } \right)^{3/4} \exp \left( {\frac{{ - E_g }}{{2k_B T}}} \right)\left( {k_B T} \right)^{3/2} \nonumber\\
\times \left( {5k_B + \frac{{E_g }}{T}} \right),
\label{eq:31}
\end{align}
\begin{align}
f \approx u(T = 0) - \frac{{\sqrt {2\pi } }}{{\pi ^2 \hbar ^3 }}\left( {m_e^* m_h^* } \right)^{3/4} \left( {k_B T} \right)^{5/2} \nonumber\\
\times \exp \left( {\frac{{ - E_g }}{{2k_B T}}} \right).
\label{eq:32}
\end{align}
Here $E_g$ is the band gap. The Boltzmann approximation assumes that there are no
other kinks in the density of states besides the square root behavior described by
Eq. \eqref{eq:27} within $k_BT$ of the band edge. This approximation works best
for small band gaps. For large band gaps the bands are narrow and there are Van
Hove singularities near the band edges. The electronic contributions to the thermodynamic
quantities of semiconductors and insulators are exponentially suppressed at low
temperatures by the factor $\exp(-E_g/2k_BT)$ and are often simply ignored. Equations
\eqref{eq:29}--\eqref{eq:32} can be used to estimate the temperatures where it is no longer
reasonable to ignore the electronic contribution to the thermodynamic properties of
semiconductors.
\section{Conclusions}
The classic problem of electrons moving in a 1-D square-wave potential was considered
and expressions were derived for the density of states and the effective masses
of electrons and holes. These expressions allowed us to efficiently calculate the
band structure and electronic contribution to the thermodynamic properties in 3-D
separable square-wave potentials. This relatively simple model produces a wide
range of band structures including metals, semiconductors, and insulators. Plots
were presented showing the parameters for which band gaps appear. It was observed that
for this class of potentials, band gaps only appear when $V > 0.9 \pi^2\hbar^2/ma^2$.
Thus the condition for the existence of band gaps depends strongly on the lattice
constant. The density of states at the Fermi energy and its derivative as well as the
effective masses of the electrons and holes were also calculated. This makes it
possible to estimate the thermodynamic properties using the standard approximations
of the Sommerfeld expansion for metals and the Boltzmann approximation for semiconductors.
Expressions were derived for the thermodynamic properties of semiconductors in the
Boltzmann approximation. These simple models assume that the density of states is a
smooth function in an energy range $k_BT$ wide. For those cases where a Van Hove
singularity causes the density of states not to be a smooth function, the thermodynamic
properties were calculated numerically and the results are available in the supplementary
material.
\newpage
|
\section{Introduction}
Many researchers investigate optimal control problems with uncertainties. In
\cite{GKP10a}, Gabasov et al. consider optimal preposterous observation and
optimal control problems for dynamic systems under uncertainty with use of a
priori and current information about the controlled object behavior and
uncertainty. In \cite{GKP09}, Gabasov et al. investigate for an optimal
control problem under uncertainty the positional solutions, which are based
on the results of inexact measurements of input and output signals of
controlled object. In \cite{GKP10b}, Gabasov et al. study a problem of
optimal control of a linear dynamical system under set-membership
uncertainty.
Fuzzy time-optimal control problem is investigated in different forms in
\cite{Plot00}-\cite{MP09}. In \cite{Plot00}, Plotnikov proves necessary
maximin and maximax conditions for a control problem, when behavior of the
object is described by a controllable differential inclusion with
multivalued performance criterion. In \cite{SIKI96}, Sakawa et al. propose a
fuzzy satisficing method for multiobjective linear optimal control problems.
To solve these problems, they discretize the time and replace the system of
differential equations by system of difference equations. In \cite{MP09},
Molchanyuk and Plotnikov study the problem of high-speed operation for
linear control systems with fuzzy right-hand sides. For this problem, they
introduce the notion of optimal solution and establish necessary and
sufficient conditions of optimality in the form of the maximum principle.
In this paper, we consider a time-optimal control problem with crisp
dynamics and with fuzzy start and target states. We interpret the optimal
time as a fuzzy variable and propose a numerical method to calculate it.
The paper consists of 5 sections. In Section 2, we describe the classical
time-optimal control problem. In Section 3, we define the fuzzy time-optimal
control problem and propose a method for calculation of fuzzy optimal time.
In Section 4, we show the proposed approach by an example. Finally, we give
concluding remarks in Section 5.
\section{Classical linear time-optimal control problem}
Let the behavior of a controlled object is definite (crisp) and described by
the following linear system of differential equations:
\begin{equation}
\dot{x}=Ax+u \label{Motion}
\end{equation}
Here $x$ is $n$-dimensional vector-function that describes the phase state
of the object, $A$ is an $n\times n$ matrix, $u$ is $n$-dimensional control
vector-function.
Let $U\subseteq R^{n}$ be a nonempty compact set. If measurable function $u$%
, defined on the interval $I=\left[ t_{0},t_{1}\right] $, satisfies the
condition $u(t)\in U$ for each $t\in I$, then $u$ is called as admissible
control. It is known that for any admissible function $u$ and for any
initial state $p$ the initial value problem
\begin{equation*}
\dot{x}=Ax+u
\end{equation*}
\begin{equation*}
x(t_{0})=p
\end{equation*}
has a unique solution \cite{Blag01}. This solution $x$ describes how the
phase state changes under the influence of admissible control $u$.
Assume that the start time $t_{0}$ and the start state $p$ are given. If we
want to transfer the object to a given state $q$ in the shortest time by
choosing an appropriate admissible control $u$, we have the following
Classical time-optimal control problem of 1st type:
\begin{equation}
t_{1}-t_{0}\rightarrow \underset{u}{\min } \label{eqA1}
\end{equation}
Subject to
\begin{equation}
\dot{x}=Ax+u \label{eqA2}
\end{equation}
\begin{equation}
x(t_{0})=p \label{eqA3}
\end{equation}
\begin{equation}
x(t_{1})=q \label{eqA4}
\end{equation}
Note, that the finish time $t_{1}$ is not known beforehand and is determined
as a result of solving the problem. Summarizing, 1st type Classical
time-optimal problem (\ref{eqA1})-(\ref{eqA4}) is a problem of finding an
admissible control $u$, which transfers the system from the initial phase
state $p$ to the final phase state $q$ in the shortest time.
Now, let nonempty compact sets $M_{0}$ and $M_{1}$ from $R^{n}$, an interval
$I=\left[ t_{0},t_{1}\right] ,$ and an admissible function $u$ on this
interval are given. If the system (\ref{Motion}) has a solution $x(t)$ such
that $x(t_{0})\in M_{0}$ and $x(t_{1})\in M_{1}$, then it is said that the
control function $u$ transfers the object from the initial phase set $M_{0}$
to the final phase set $M_{1}$ on the interval $\left[ t_{0},t_{1}\right] $.
If we want to transfer the object from the set $M_{0}$ to the set $M_{1}$ in
the shortest time, we have the following Classical time-optimal control
problem of 2nd type:
\begin{equation}
t_{1}-t_{0}\rightarrow \underset{u}{\min } \label{eqB1}
\end{equation}
Subject to
\begin{equation}
\dot{x}=Ax+u \label{eqB2}
\end{equation}
\begin{equation}
x(t_{0})\in M_{0} \label{eqB3}
\end{equation}
\begin{equation}
x(t_{1})\in M_{1} \label{eqB4}
\end{equation}
where $M_{0}$ and $M_{1}$ are given start and target sets. The solution $u$
of the problem (\ref{eqB1})-(\ref{eqB4}) is called optimal control. The
solution $x$ of the system (\ref{eqB2})-(\ref{eqB4}), corresponding to the
optimal control $u$, is called optimal trajectory. If $u(t)$ is an optimal
control and $x(t)$ is a corresponding optimal trajectory, then $(u(t),x(t))$
is called to be an optimal pair.
We note that the Classical problem of 2nd type can also be reformulated as
follows:
\begin{equation}
t_{1}-t_{0}\rightarrow \underset{u;\ p\in M_{0};\ q\in M_{1}}{\min }
\label{eqD1}
\end{equation}
\begin{equation}
\dot{x}=Ax+u \label{eqD2}
\end{equation}
\begin{equation}
x(t_{0})=p \label{eqD3}
\end{equation}
\begin{equation}
x(t_{1})=q \label{eqD4}
\end{equation}
2nd type Classical time-optimal problem (\ref{eqB1})-(\ref{eqB4}) (or (\ref%
{eqD1})-(\ref{eqD4})) is well studied \cite{Blag01}. Below\ we\ give\
necessary\ conditions\ of\ optimality for this problem \cite{Blag01}.
\begin{definition}
(Maximum principle). Let $u$ be an admissible control defined on an interval
$\left[ t_{0},t_{1}\right] $ and let $x$ be a solution of the system (\ref%
{eqB2})-(\ref{eqB4}). We say that the pair $(u(t),x(t))$ satisfies maximum
principle on the interval $\left[ t_{0},t_{1}\right] $ if the conjugate
system
\begin{equation*}
\dot{\psi}=-A^{*}\psi
\end{equation*}
has such a nontrivial solution $\psi =(\psi _{1},\psi _{2},...,\psi _{n})$
that the following conditions hold:
1) maximum condition: $\left\langle u(t),\psi (t)\right\rangle =c(U,\psi
(t)) $ for almost any $t\in \left[ t_{0},t_{1}\right] $;
2) transversality condition on $M_{0}$: $\left\langle x(t_{0}),\psi
(t_{0})\right\rangle =c(M_{0},\psi (t_{0}))$;
3) transversality condition on $M_{1}$: $\left\langle x(t_{1}),-\psi
(t_{1})\right\rangle =c(M_{1},-\psi (t_{1}))$.
\end{definition}
Here $\left\langle u,\psi \right\rangle =u_{1}\psi _{1}+u_{2}\psi
_{2}+...+u_{n}\psi _{n}$ denotes the inner product of vectors $u$ and $\psi $
from $R^{n}$ and $c(S,\psi )=\underset{s\in S}{\max }\left\langle s,\psi
\right\rangle $ denotes the support function of the compact set $S$ from $%
R^{n}$.
\begin{theorem}
\cite{Blag01} (Necessary conditions of optimality for the time-optimal
control problem). Let $M_{0}$ and $M_{1}$ be nonempty convex compact sets.
Also let the function $u$ defined on $\left[ t_{0},t_{1}\right] $ be an
optimal control for the problem (\ref{eqB1})-(\ref{eqB4}) and $x$ be a
corresponding optimal trajectory. Then the pair $(u(t),x(t))$ satisfies
maximum principle on the interval $\left[ t_{0},t_{1}\right] $.
\end{theorem}
\section{Fuzzy linear time-optimal control problem}
If\ the start and target values in Classical problem of 1st or 2nd type are
fuzzy, we obtain the following Fuzzy time-optimal control problem:
\begin{equation}
t_{1}-t_{0}\rightarrow \underset{u}{\min } \label{eqF1}
\end{equation}
Subject to
\begin{equation}
\dot{x}=Ax+u \label{eqF2}
\end{equation}
\begin{equation}
x(t_{0})=\widetilde{\xi } \label{eqF3}
\end{equation}
\begin{equation}
x(t_{1})=\widetilde{\zeta } \label{eqF4}
\end{equation}
where $\widetilde{\xi }$ and $\widetilde{\zeta }$ are given fuzzy initial
and final vectors (or sets).
Depending on different definitions for derivative of fuzzy function or
different definitions for solution of system of differential equations, the
problem (\ref{eqF1})-(\ref{eqF4}) can be interpreted by different ways. We
will interpret the problem (\ref{eqF1})-(\ref{eqF4}) as a set of 1st type
Classical problems (\ref{eqA1})-(\ref{eqA4}). Each problem is obtained by
taking the initial value $p$ from $\xi $ and the final value $q$ from $\zeta
$. We denote by $t_{1,pq}$, $u_{pq}$ and $x_{pq}$ the solutions of the
problem (\ref{eqA1})-(\ref{eqA4}). Let $\alpha =\min \left\{ \mu _{\xi
}(p),\mu _{\zeta }(q)\right\} $ (where $\mu _{\xi }(p)$ denotes the
membership of $p$ in $\xi $). We call $(t_{1,pq},u_{pq},x_{pq})$ to be a
solution of the problem (\ref{eqF1})-(\ref{eqF4}) with possibility $\alpha $.
Set of all $t_{1,pq}$, defined above, determines a fuzzy number $\widetilde{%
t_{1}}$. We will investigate how to calculate $\widetilde{t_{1}}$. Functions
\underline{$t_{1}$}$(\alpha )$ and $\overline{t_{1}}(\alpha ),$ which
indicate the left and right boundaries of $\alpha $-cuts, determine the
number $\widetilde{t_{1}}$ fully. Thus, the problem of calculation of fuzzy
optimal time is reduced to calculation of the functions \underline{$t_{1}$}$%
(\alpha )$ and $\overline{t_{1}}(\alpha )$.
As it is known, the initial and final values of the optimal solution $x(t)$
of the problem (\ref{eqB1})-(\ref{eqB4}) are achieved on boundaries of the
sets $M_{0}$ and $M_{1}$ \cite{Blag01}. Thus, value of \underline{$t_{1}$}$%
(\alpha )$ can be obtained by solving the problem (\ref{eqB1})-(\ref{eqB4})
with taking $M_{0}=\xi _{\alpha }$ and $M_{1}=\zeta _{\alpha }$ ($\xi
_{\alpha }$ and $\zeta _{\alpha }$ denote $\alpha $-cuts of $\xi $ and $%
\zeta $, respectively), namely the problem:
\begin{equation}
t_{1}-t_{0}\rightarrow \underset{u}{\min } \label{eqC1}
\end{equation}
\begin{equation}
\dot{x}=Ax+u \label{eqC2}
\end{equation}
\begin{equation}
x(t_{0})\in \xi _{\alpha } \label{eqC3}
\end{equation}
\begin{equation}
x(t_{1})\in \zeta _{\alpha } \label{eqC4}
\end{equation}
The problem (\ref{eqC1})-(\ref{eqC4}) is a classical problem of 2nd type.
Taking into account (\ref{eqD1}), it can be seen that \underline{$t_{1}$}$%
(\alpha )=\underset{p\in \xi _{\alpha };\ \ q\in \zeta _{\alpha }}{\min }%
t_{1,pq}$. Note that, the value \underline{$t_{1}$}$(\alpha )$ means the
shortest time between two points, one of them is from the set $\xi _{\alpha
} $ and another is from $\zeta _{\alpha }$, in the best case. Similarly, $%
\overline{t_{1}}(\alpha )$ means the shortest time in the worst case:
\begin{equation}
\overline{t_{1}}(\alpha )=\underset{p\in \xi _{\alpha };\ \ q\in \zeta
_{\alpha }}{\max }t_{1,pq} \label{t1_up}
\end{equation}
Calculation of $\overline{t_{1}}(\alpha )$ can be performed similarly to
calculation of \underline{$t_{1}$}$(\alpha )$.
\section{Example}
In this section, we apply the proposed approach to a fuzzy time-optimal
control problem. The problem is a fuzzified version of the crisp problem of
damping of mathematical pendulum, presented in \cite{Blag01}.
\begin{example}
Solve the fuzzy time-optimal control problem (Note that below $t_{0}=0$):
$t_{1}\rightarrow \underset{u}{\min }$
$\dot{x}_{1}=x_{2}$
$\dot{x}_{2}=-x_{1}+u_{2}$
$U=\left\{ u=(u_{1},u_{2})|u_{1}=0,\left| u_{2}\right| \leq 1\right\}
\subseteq R^{2}$
$x_{1}(0)=\xi _{1}=(-6,-5,-4);\ \ \ \ x_{2}(0)=\xi _{2}=(2,3,4)$
$x_{1}(t_{1})=\zeta _{1}=(-0.5,0,0.5);\ \ \ x_{2}(t_{1})=\zeta
_{2}=(-0.5,0,0.5)$
\end{example}
Here $\xi _{1},\xi _{2},\zeta _{1}$ and $\zeta _{2}$ are triangular fuzzy
numbers.
From above $\xi _{\alpha }=\left\{ (x_{1},x_{2})|\alpha -6\leq x_{1}\leq
-4-\alpha ,\ \alpha +2\leq x_{2}\leq 4-\alpha \right\} \subseteq R^{2}$ and $%
\zeta _{\alpha }=\left\{ (x_{1},x_{2})|\alpha -1\leq x_{1}\leq 1-\alpha ,\
\alpha -1\leq x_{2}\leq 1-\alpha \right\} \subseteq R^{2}$.
System's matrix is $A=\left[
\begin{array}{ll}
0 & 1 \\
-1 & 0%
\end{array}
\right] $. Since, $-A^{*}=A$ (here $A^{*}$ denotes the conjugate matrix of $%
A $), the conjugate system is:
$\dot{\psi} _{1}=\psi _{2}$
$\dot{\psi}_{2}=-\psi _{1}$
Support function of $U$ is $c(U,\psi )=\left\vert \psi _{2}\right\vert $.
Then, from the maximum condition $\left\langle u(t),\psi (t)\right\rangle
=c(U,\psi (t))$ we have $u_{2}(t)\psi _{2}(t)=\left\vert \psi
_{2}(t)\right\vert $. Consequently,
$u_{2}(t)=1,$ if $\psi _{2}(t)>0$;
$u_{2}(t)=-1,$ if $\psi _{2}(t)<0$;
$-1\leq u_{2}(t)\leq 1,$ if $\psi _{2}(t)=0$.
Let us find the solution of the conjugate system corresponding to an initial
condition $\psi (0)\in C$, where $C$ is the unit circle. Initial point can
be represented in the form of $\psi (0)=(\cos \alpha ,\sin \alpha )$ with $%
\alpha \in \left[ 0,2\pi \right) $. Then the solution of the conjugate
system is $\psi _{1}(t)=\cos (\alpha -t),\psi _{2}(t)=\sin (\alpha -t)$. The
function $\psi _{2}(t)=\sin (\alpha -t)$ changes its sign for first time at $%
\tau \leq \pi $ $(\tau =\alpha $, if $0<\alpha \leq \pi $ and $\tau =\alpha
-\pi $, if $\pi <\alpha <2\pi )$ and then after each $\pi $ time period.
Depending on $\alpha $, the sign of the function $\psi _{2}(t)=\sin (\alpha
-t)$ in the interval $\left[ 0,\tau \right] $ is either positive or
negative. Thus, according to the maximum condition, the initial value of the
optimal control $u_{2}(t)$ is either $1$ or $-1$. After $\tau \leq \pi $
time units, it switches from $1$ to $-1$ or vice versa. Then, it repeatedly
changes its sign after each $\pi $ time period.
Below we interpret the behavior of the object as a motion of the object in
the phase plane $x_{1}x_{2}$.
Solutions of dynamic system corresponding to $u_{2}(t)=1$ are in the form $%
x(t)=(1+c\cos (\varphi -t),c\sin (\varphi -t))$. In the phase plane $R^{2}$
these solutions constitute concentric circles with center at $L(1,0)$ (Fig.
1). The motion on these circles is clockwise with constant speed and whole
turn takes $2\pi $ time units.
\begin{figure}
\begin{center}
\includegraphics[width=2.6818in,height=1.7521in,keepaspectratio=false,angle=0]{fig1.eps}
\end{center}
\caption{An optimal trajectory can be realized by combining of clockwise motions on circles with centers $K$ and $L$.}
\label{fig1}
\end{figure}
Similarly, solutions of dynamic system corresponding to $u_{2}(t)=-1$ are in
the form $x(t)=(-1+c\cos (\varphi -t),c\sin (\varphi -t))$. In $R^{2}$ these
solutions constitute circles with center at $K(-1,0)$ (Fig. 1). The motion
on these circles is clockwise with constant speed and whole turn takes $2\pi
$ time units.
Note that angular speed is $\omega =1$ for both motions mentioned above. So,
the angle formed by the object during its motion and passed time are equal
in value.
Let us emphasize two facts which will be used in arguments below. 1) In
circular motion with $\omega =1$ after $\pi $ time period the object will be
in the position which is central symmetric point of the previous position.
2) The symmetric point of $(a,b)$ is $(-a-2,-b)$ with respect to center
point $K$. If center is $L$, then the symmetry of point $(c,d)$ is $%
(-c+2,-d) $.
Now we investigate how is a motion of the object corresponding to an optimal
control in the phase plane for a start point $S$ and a target point $T$. Let
us consider the case when the object starts with control $u=-1$ (The case
with start control $u=1$ can be investigated similarly). Let $k$ denote the
number of control switches. We consider the cases $k=0$ (motion without
switch) and $k\geq 1$ separately.
In the case $k=0$, running from the start position $S$ and moving along a
circle with center $K$ the object reaches the target position $T$. This case
occurs, only if $\left\vert KS\right\vert =\left\vert KT\right\vert $ (Here $%
\left\vert AB\right\vert $ denotes the length of the segment $AB$). The
motion time is $t_{1}=\theta =\angle SKT$ (Here $\angle SKT$ denotes the
value of the angle $SKT$).
Now let $k\geq 1$. We differ the cases when $k$ is odd and when $k$ is even.
Let us consider the case that $k$ is odd number and take $k=3$ for clarity.
The object runs from the point $S$ along a circle with center $K$ and after $%
\tau $ time period arrives a point $X_{1}(x,y)$ (Fig. 2). The points $S$ and
$X_{1}$ are on the same circle. Consequently:
\begin{equation}
\left| KX_{1}\right| =\left| KS\right| \label{C1}
\end{equation}
\begin{figure}
\begin{center}
\includegraphics[width=3.9496in,height=3.0234in,keepaspectratio=false,angle=0]{fig2.eps}
\end{center}
\caption{A sample of optimal trajectory with 3 switches.}
\label{fig2}
\end{figure}
At the
point $X_{1}$ the control switches for the first time and becomes $u=1$.
Under this control, the object moves along a circle with center $L$. After $%
\pi $ time units it arrives a point $X_{2}(-x+2,-y)$. Here the control
switches for the second time and under new control $u=-1$ (moving on circle
with center $K$) after $\pi $ time the object reaches a point $%
X_{k}=X_{3}(x-4,y)$. At the point $X_{k}$ the control switches for last time
and becomes $u=1$. The object continues its motion on a circle with center $%
L $ up to the target point $T$. For the aforementioned motion, the points $%
X_{k}$ and $T$ must be on the same circle with center $L$, i.e.,
\begin{equation}
\left| LX_{k}\right| =\left| LT\right| \label{C2}
\end{equation}
It can be seen from Table 1 that for an odd $k$ (including $k=1$) the last
point of control switch is
\begin{equation}
X_{k}=(x_{k},y_{k})=(x-2(k-1),y) \label{Xk}
\end{equation}
\begin{table}[t]
\caption{Point of $k$-th control switch for optimal motion}
\label{table1}\centering
\begin{tabular}{cccc}
\hline
$k$ (odd) & $X_{k}$ & $k$ (even) & $X_{k}$ \\ \hline
1 & $(x,y)$ & 2 & $(-x+2,-y)$ \\
3 & $(x-4,y)$ & 4 & $(-x+6,-y)$ \\
5 & $(x-8,y)$ & 6 & $(-x+10,-y)$ \\
7 & $(x-12,y)$ & 8 & $(-x+14,-y)$%
\end{tabular}%
\end{table}
Let $S=(p_{x},p_{y})$ and $T=(q_{x},q_{y})$. To calculate unknown
coordinates $x$ and $y$ we use equations (\ref{C1}) and (\ref{C2}). Using (%
\ref{Xk}), these equations can be rewritten in coordinates as follows:
\begin{eqnarray}
(x+1)^{2}+y^{2} &=&r_{1}^{2}=(p_{x}+1)^{2}+p_{y}^{2} \label{C1n} \\
(x+1-2k)^{2}+y^{2} &=&r_{2}^{2}=(q_{x}+1)^{2}+q_{y}^{2} \label{C2n}
\end{eqnarray}
Subtracting (\ref{C2n}) from (\ref{C1n}) we have: $%
4k(x+1)-4k^{2}=r_{1}^{2}-r_{2}^{2}$. Then we can determine $x$ and $y$ as
follows:
\begin{eqnarray}
x &=&\frac{r_{1}^{2}-r_{2}^{2}}{4k}+k-1 \label{x} \\
y &=&\pm \sqrt{r_{1}^{2}-(x+1)^{2}} \label{y}
\end{eqnarray}
If $x$ and $y$ have been determined we can calculate the passed time:
\begin{equation}
t_{1}=\angle SKX_{1}+(k-1)\pi +\angle X_{k}LT \label{t1}
\end{equation}
Let us find an evaluation for $k$. From (\ref{x}) and (\ref{y}) we have
\begin{equation*}
y^{2}=r_{1}^{2}-\left( \frac{r_{1}^{2}-r_{2}^{2}}{4k}+k\right) ^{2}\geq
0\Longleftrightarrow k^{4}-\frac{r_{1}^{2}+r_{2}^{2}}{2}\ k^{2}+\left( \frac{%
r_{1}^{2}-r_{2}^{2}}{4}\right) ^{2}\leq 0\Longleftrightarrow
\end{equation*}%
\begin{equation*}
\frac{(r_{1}-r_{2})^{2}}{4}\leq k^{2}\leq \frac{(r_{1}+r_{2})^{2}}{4}
\end{equation*}%
Hence, we obtain the following evaluation
\begin{equation}
k_{\min }=\left\lceil \left\vert r_{1}-r_{2}\right\vert /2\right\rceil \leq
k\leq \left\lfloor (r_{1}+r_{2})/2\right\rfloor =\widehat{k} \label{k_eval}
\end{equation}%
where $\left\lceil x\right\rceil $ and $\left\lfloor x\right\rfloor $ denote
ceiling and floor of $x$, respectively. By taking $k=k_{\min }$, we have a
feasible motion. Hence, using formula (\ref{t1}), we get:
\begin{equation*}
t_{1,opt}\leq t_{1}=\angle SKX_{1}+(k_{\min }-1)\pi +\angle X_{k}LT<2\pi
+(k_{\min }-1)\pi +2\pi
\end{equation*}%
Then, we have $k_{\max }<k_{\min }+4\Longleftrightarrow k_{\max }\leq
k_{\min }+3$. Consequently, we obtain the following upper evaluation for $k$%
, by using (\ref{k_eval}):
\begin{equation*}
k_{\max }=\min \left\{ k_{\min }+3,\ \widehat{k}\right\}
\end{equation*}
The case when $k\geq 1$ and $k$ is even can be investigated by similar way.
In this case the last point of the control switch is (Table 1):
\begin{equation}
X_{k}=(x_{k},y_{k})=(-x+2(k-1),-y) \label{X_even k}
\end{equation}
The last control is $u=-1$ and, consequently, the object finishes its motion
on a circle with center $K$. Hence, $r_{2}^{2}=(q_{x}-1)^{2}+q_{y}^{2}$.
Except this value, the formulas for $x$ and $y$ become the same as (\ref{x})
and (\ref{y}). The motion time is:
\begin{equation}
t_{1}=\angle SKX_{1}+(k-1)\pi +\angle X_{k}KT \label{t1_even k}
\end{equation}
Above we have investigated the case when the start control $u$ equals to $-1$%
. In the case where $u$ is $1$ we have the following final formulas:
\begin{eqnarray}
r_{1}^{2} &=&(p_{x}-1)^{2}+p_{y}^{2} \\
r_{2}^{2} &=&\left\{
\begin{array}{c}
(q_{x}+1)^{2}+q_{y}^{2},\ \ \ \ \text{if }k\text{ is odd} \\
(q_{x}-1)^{2}+q_{y}^{2},\ \ \ \ \ \text{if }k\text{ is even}%
\end{array}
\right. \\
X_{k} &=&(x_{k},y_{k})=\left\{
\begin{array}{c}
(x+2(k-1),y),\ \ \ \ \ \ \ \text{if }k\text{ is odd} \\
(-x-2(k-1),-y),\ \ \ \ \ \text{if }k\text{ is even}%
\end{array}
\right. \\
x &=&-\left( \frac{r_{1}^{2}-r_{2}^{2}}{4k}+k-1\right) \\
y &=&\pm \sqrt{r_{1}^{2}-(x-1)^{2}} \\
t_{1} &=&\angle SLX_{1}+(k-1)\pi +\left\{
\begin{array}{c}
\angle X_{k}KT,\ \ \ \ \text{if }k\text{ is odd} \\
\angle X_{k}LT,\ \ \ \ \ \text{if }k\text{ is even}%
\end{array}
\right.
\end{eqnarray}
The above formulas, given for different situations, were obtained on the
base of the necessary conditions for optimality. Therefore, every solution
constructed on these formulas may not be optimal. However, the optimal
solution is among all solutions, constructed for different start controls
and for different values of $k$.
Based on the above arguments and formulas a computer program is implemented
to calculate the optimal control for a given pair of start point $S$ and
target point $T$. Firstly, by taking start control $u=-1$, after taking $u=1$
and in both cases by changing the value of $k$ from $k_{\min }$ to $k_{\max
} $ a solution is constructed (if there is any). The solution with the
shortest time is the optimal solution, transferring the object from $S$ to $%
T $.
Now, let us describe how we calculate the fuzzy optimal time $\widetilde{%
t_{1}}$ numerically. To calculate the value \underline{$t_{1}$}$(\alpha )$,
we place equally spaced nodes on the boundaries of the regions $\xi _{\alpha
}$\ and\ $\zeta _{\alpha }$. The shortest time among all possible
start-destination node pairs $(p,q)$ gives the approximate value of
\underline{$t_{1}$}$(\alpha )$.
To calculate the function $\overline{t_{1}}(\alpha )$ we discretize the
problem (\ref{t1_up}) and solve it numerically.
The membership function of fuzzy optimal time $t_{1}$, obtained from
calculations, is depicted in Fig. 3. The value $t_{1}\approx 8.78$ with
possibility $1$ corresponds to the solution of the crisp problem ($p=(-5,3)$
and $q=(0,0)$). The least value $t_{1}\approx 5.97$ with possibility $0$
occurs when $p=(-4,2)$ and $q=(-0.5,0.5)$. The largest value $t_{1}\approx
11.76$ with possibility $0$ corresponds to the pair $p=(-6,4)$ and $%
q=(0.5,0.5)$.
\begin{figure}
\begin{center}
\includegraphics[width=5.0704in,height=2.5754in,keepaspectratio=false,angle=0]{fig3.eps}
\end{center}
\caption{The membership function of fuzzy optimal time $t_{1}$. }
\label{fig3}
\end{figure}
\section{Conclusion}
In this paper, we investigate the problem of time-optimal control with fuzzy
initial and final states. We interpret the problem as a set of crisp
problems. We perform the calculation of fuzzy optimal time by solving crisp
optimal control problems of two types. We exhibit the proposed approach on a
numerical example.
\smallskip
|
\section{Introduction}
For independent random variables, the Rosenthal inequalities relate higher
moments of partial sums of random variables with the variance of partial sums.
One variant of this inequality is the following (see Rosenthal (1970), p.
279): let $(X_{k})_{k}$ be independent and centered real valued random
variables with finite moments of order $p$, $p\geq2$. Then for every positive
integer $n$,
\begin{equation}
{\mathbf{E}}\big(\max_{1\leq j\leq n}\ |S_{j}|^{p}\big)\ll\sum_{k=1}%
^{n}{\mathbf{E}}(|X_{k}|^{p})+\big (\sum_{k=1}^{n}{\mathbf{E}}(X_{k}%
^{2})\big )^{p/2}\,, \label{Rosenind}%
\end{equation}
where $S_{j}=\sum_{k=1}^{j}X_{k}$; the notation $a_{n}\ll b_{n}$ will often
replace the Vinogradov symbol $O$ and means that there is a positive constant
$C$ (that may depend on $p$ in this paper) such that $a_{n}\leq Cb_{n}$ for
all positive integers $n$.
Besides of being useful to compare the norms ${\mathbf{L}}^{p}$ and
${\mathbf{L}}^{2}$ of partial sums, these inequalities are important tools to
obtain a variety of results, including convergence rates with respect to the
strong law of large numbers (see for instance Wittmann (1985)) or almost sure
invariance principles (see Wu (2007) and Gou\"{e}zel (2010) for recent
results). Since the 70's, there has been a great amount of works which
extended the inequality (\ref{Rosenind}) to dependent sequences. See, for
instance among many others: Peligrad (1985) and Shao (1995) for the case of
$\rho$-mixing sequences; Shao (1988), Peligrad (1989) and Utev (1991) for the
case of $\phi$-mixing sequences; Peligrad and Gut (1999) and Utev and Peligrad
(2003) for interlaced mixing; Theorem 2.2 in Viennet (1997) for $\beta$-mixing
processes; Theorem 6.3 in Rio (2000) for the strongly mixing case; Rio (2009)
for projective criteria.
The main goal of the paper is to generalize the Rosenthal inequality from
sequences of independent variables to stationary dependent sequences including
martingales, allowing then to consider examples that are not necessarily
dependent in the sense of the dependence structures mentioned above.
In order to present our results, let us first introduce some notations and
definitions used all along the paper.
\begin{notation}
\label{notation1} Let $(\Omega,{\mathcal{A}},{\mathbf{P}})$ be a probability
space and let $T:\Omega\mapsto\Omega$ be a bijective bi-measurable
transformation preserving the probability ${\mathbf{P}}$. Let $\mathcal{F}%
_{0}$ be a $\sigma$-algebra of $\mathcal{A}$ satisfying $\mathcal{F}%
_{0}\subseteq T^{-1}(\mathcal{F}_{0})$. We then define the nondecreasing
filtration $(\mathcal{F}_{i})_{i\in\mathbb{Z}}$ by $\mathcal{F}_{i}%
=T^{-i}(\mathcal{F}_{0})$ and the stationary sequence $(X_{i})_{i\in
\mathbb{Z}}$ by $X_{i}=X_{0}\circ T^{i}$ where $X_{0}$ is a real-valued random
variable. The sequence will be called adapted to the filtration $(\mathcal{F}%
_{i})_{i\in\mathbb{Z}}$ if $X_{0}$ is ${}\mathcal{F}_{0}$-measurable. The
following notations will also be used: ${\mathbf{E}}_{k}(X)={\mathbf{E}%
}(X|\mathcal{F}_{k})$ and the norm in ${\mathbf{L}}^{p}$ of $X$ is denoted by
$||X||_{p}$. Let $S_{n}=\sum_{j=1}^{n}X_{j}$.
\end{notation}
In the rest of this section the sequence $(X_{i})_{i\in\mathbb{Z}}$ $\ $is
assumed to be stationary and adapted to $(\mathcal{F}_{i})_{i\in\mathbb{Z}}$
and the variables are in ${\mathbf{L}}^{p}$.
If $(X_{k})_{k}$ are stationary martingale differences, the martingale form of
the inequality (\ref{Rosenind}) is
\begin{equation}
\mathbf{||}\max_{1\leq j\leq n}\ |S_{j}|\text{ }||_{p}\ll\ n^{1/p}%
\mathbf{||}X_{1}||_{p}+\big \Vert\sum_{k=1}^{n}{\mathbf{E}}_{k-1}(X_{k}%
^{2})\big \Vert_{p/2}^{1/2}\,\text{\ for any }p\geq2, \label{Rosenmart}%
\end{equation}
(see Burkholder (1973)). One of our goals is to replace the last term in this
inequality with a new one containing terms of the form $\Vert{\mathbf{E}}%
_{0}(S_{n}^{2})\Vert_{p/2}.$ The reason for introducing this term comes from
the fact that for many stationary sequences $\Vert{\mathbf{E}}_{0}(S_{n}%
^{2})\Vert_{p/2}$ is closer to the variance of partial sums. In addition, we
are interested to point out a Rosenthal-type inequality for a larger class of stationary adapted
sequences that includes the martingales.
Two recent results by Peligrad and Utev (2005) and Wu and Zhao (2008) show
that
\[
\mathbf{||}\max_{1\leq j\leq n}\ |S_{j}|\text{ }||_{p}\ll n^{1/p}\left(
\mathbf{||}X_{1}||_{p}+\sum_{k=1}^{n}\frac{1}{k^{1+1/p\ }}||{\mathbf{E}}%
_{0}(S_{k})||_{p}\right) \,\text{\ for any }1\leq p\leq2.
\]
To find a suitable extension of this inequality for $p>2$, the first step in
our approach is to establish the following maximal inequality that has
interest in itself:%
\begin{equation}
\Vert\max_{1\leq j\leq n}|S_{j}|\Vert_{p}\ll n^{1/p}\left( \max_{1\leq j\leq
n}\Vert S_{j}\Vert_{p}/n^{1/p}+\sum_{k=1}^{n}\frac{1}{k^{1+1/p\ }%
}||{\mathbf{E}}_{0}(S_{k})||_{p}\right) \,\text{\ for any }p>1. \label{max}%
\end{equation}
This inequality can be viewed as generalization of the well-known Doob's
maximal inequality for martingales.
Then, we combine the inequality (\ref{max}) with several inequalities for
$\Vert S_{n}\Vert_{p}$ that will further be established in this paper$.$
As we shall see in Section \ref{sectiongeneralpowers}, by a direct approach
using dyadic induction combined with the maximal inequality (\ref{max}), we
shall prove that, for any $p>2$,
\begin{equation}
\Vert\max_{1\leq j\leq n}\ |S_{j}|\Vert_{p}\ll n^{1/p}\left( \mathbf{||}%
X_{1}||_{p}+\sum_{k=1}^{n}\frac{1}{k^{1+1/p\ }}||{\mathbf{E}}_{0}(S_{k}%
)||_{p}+\Big (\sum_{k=1}^{n}\frac{1}{k^{1+2\delta/p}}\Vert{\mathbf{E}}%
_{0}(S_{k}^{2})\Vert_{p/2}^{\delta}\Big )^{1/(2\delta)}\right) \,,
\label{inedirect}%
\end{equation}
where $\delta=\min(1,1/(p-2))$. For $2<p\leq3$ our inequality provides a
maximal form for Theorem 3.1 in Rio (2009). When $p\geq 4$, we shall see that
the last term in the right hand side dominates the second term that can be then
omitted in this case. Inequality (\ref{inedirect}) shows that in order to
relate $\Vert\max_{1\leq j\leq n}\ |S_{j}|\Vert_{p}$ to the vector $(\Vert
S_{j}\Vert_{2})_{1\leq j\leq n}$ we have to control $\sum_{k=1}^{n}%
k^{-(p+1)/p}||{\mathbf{E}}_{0}(S_{k})||_{p}\ $ and $\ \sum_{k=1}%
^{n}k^{-(p+2\delta)/p}\Vert{\mathbf{E}}_{0}(S_{k}^{2})-{\mathbf{E}}(S_{k}%
^{2})\Vert_{p/2}^{\delta}$.
In Section \ref{martingalecase}, we study the case of stationary martingale
difference sequences showing that for all even powers $p\geq4$, the inequality
(\ref{inedirect}) holds with $\delta=2/(p-2)$. This result is possible for
stationary martingale differences with the help of a special symmetrization
for martingales initiated by Kwapie\'{n} and Woyczynski (1991). In addition,
by using martingale approximation techniques, we obtain, for any even integer
$p$, another form of Rosenthal-type inequality than (\ref{inedirect}) for
stationary adapted processes (see the section \ref{sectionmartappro}), that
gives, for instance, better results for functionals of linear processes with
independent innovations.
We also investigate the situation when the conditional expectation with
respect to both the past and the future of the process is used. For instance,
when $p\geq4$ is an even integer, and the process is reversible, then the
inequality (\ref{inedirect}) holds (see Theorem \ref{stateven} and Corollary
\ref{Rev}) with $\delta=1$.
In Section \ref{sectionBurk} we show that our inequalities imply the
Burkholder-type inequality as stated in Theorem 1 of Peligrad, Utev and Wu
(1997). For the sake of applications in Section \ref{sectionindsum} we express
the terms that appear in our Rosenthal inequalities in terms of individual summands.
Our paper is organized as follows. In Section \ref{sectionmaxine}, we prove a
new maximal inequality allowing to relate the moments of the maximum of
partial sums of an adapted sequence, that is not necessarily stationary, to
the moments of its partial sums. The maximal inequality (\ref{max}) combined
with moment estimates allows us to obtain the Rosenthal-type inequalities
stated in Theorems \ref{directprop} and \ref{stateven} of Section
\ref{sectiongeneralpowers}. Section \ref{sectionMA} is devoted to
Rosenthal-type inequalities for even powers for the special case of stationary
martingale differences and to an application to stationary processes via a
martingale approximation technique. In Section \ref{sectionappliexamples}, we
give others applications of the maximal inequalities stated in Section
\ref{sectionmaxine} and provide examples for which we compute the quantities
involved in the Rosenthal-type inequalities of Section \ref{sectionmomentine}.
One of the applications presented in this section is a Bernstein inequality
for the maximum of partial sums for strongly mixing sequences, that extends
the inequality in Merlev\`{e}de \textit{et al.} (2009). The applications are
given to Arch models, to functions of linear processes and reversible Markov
chains. In Section \ref{densitysection}, we apply the inequality
(\ref{inedirect}) to estimate the random term of the ${\mathbf{L}}^{p}%
$-integrated risk of kernel estimators of the unknown marginal density of a
stationary sequence that is assumed to be $\beta$-mixing in the weak sense
(see the definition \ref{beta}). Some technical results are postponed to the appendix.
\section{Maximal inequalities for adapted sequences}
\label{sectionmaxine}
The next proposition is a generalization of the well-known Doob's maximal
inequality for martingales to adapted sequences. It states that the moment of
order $p$ of the maximum of the partial sums of an adapted process can be
compared to the corresponding moment of the partial sum plus a correction term
which is zero for martingale differences sequences. The proof is based on
convexity and chaining arguments.
\begin{proposition}
\label{maxinequality} Let $p>1$ and let $q$ be its conjugate. Let $Y_{i}$,
$1\leq i\leq2^{r}$ be real random variables in ${\mathbf{L}}^{p}$, where $r$
is a positive integer. Assume that the random variables are adapted to an
increasing filtration $({\mathcal{F}}_{i})_{i}$. Let $S_{r}=Y_{1}+\cdots
+Y_{r}$ and $S_{r}^{\ast}=\max_{1\leq i\leq r}|S_{i}|$. We have that
\begin{equation}
\Vert\max_{1\leq m\leq2^{r}}|S_{m}|\Vert_{p}\leq q\Vert S_{2^{r}}\Vert
_{p}+q\sum_{l=0}^{r-1}\Big (\sum_{k=1}^{2^{r-l}-1}||{\mathbf{E}}%
(S_{(k+1)2^{l}}-S_{k2^{l}}|{\mathcal{F}}_{k2^{l}})||_{p}^{p}\Big )^{1/p}\,.
\label{inequality}%
\end{equation}
\end{proposition}
\begin{corollary}
\label{cormaxstat} In the stationary case, we get that for any integer
$r\geq1$,
\[
\Vert\max_{1\leq m\leq2^{r}}|S_{m}|\Vert_{p}\leq q\Vert S_{2^{r}}\Vert
_{p}+q2^{r/p}\sum_{l=0}^{r-1}2^{-l/p}||{\mathbf{E}}(S_{2^{l}}|{\mathcal{F}%
}_{0})||_{p}\,\text{\ }.
\]
\end{corollary}
\begin{remark}
\label{Rem1inegalite} The inequality in Corollary \ref{cormaxstat} easily
implies that%
\begin{equation}
\Vert\max_{1\leq m\leq n}|S_{m}|\Vert_{p}\leq2q\max_{1\leq m\leq n}\Vert
S_{m}\Vert_{p}+(2^{1/p}q)n^{1/p}\sum_{l=0}^{r-1}2^{-l/p}||{\mathbf{E}%
}(S_{2^{l}}|{\mathcal{F}}_{0})||_{p}\,, \label{maxdyatic}%
\end{equation}
for any integer $n\in\lbrack2^{r-1},2^{r}[$, where $r$ is a positive integer.
Moreover, due to the subadditivity of the sequence $\big (||{\mathbf{E}}%
(S_{n}|{\mathcal{F}}_{0})||_{p}\big)_{n\geq1}$, according to Lemma
\ref{lmasubadd}, we also have that for any positive integer $n$,
\begin{equation}
\Vert\max_{1\leq m\leq n}|S_{m}|\Vert_{p}\leq2q\max_{1\leq m\leq n}\Vert
S_{m}\Vert_{p}+(q\frac{2^{2+2/p}}{2^{1+1/p}-1})n^{1/p}\sum_{j=1}^{n}%
j^{-1-1/p}||{\mathbf{E}}(S_{j}|{\mathcal{F}}_{0})||_{p}\,. \label{cons1sub}%
\end{equation}
\end{remark}
For several applications involving exponential bounds we point out the
following proposition.
\begin{proposition}
\label{propmaxineproba} Let $p>1$ and let $q$ be its conjugate. Let
$(Y_{i})_{i\geq1}$, be real random variables in ${\mathbf{L}}^{p}$. Assume
that the random variables are adapted to an increasing filtration
$({\mathcal{F}}_{i})_{i}$. Let $S_{n}=Y_{1}+\cdots+Y_{n}$ and $S_{n}^{\ast
}=\max_{1\leq i\leq n}|S_{i}|$. Let $\varphi$ be a nondecreasing, non
negative, convex and even function. Then for any positive real $x$, and any
positive integer $r$, the following inequality holds
\begin{equation}
{\mathbf{P}}(S_{2^{r}}^{\ast}\geq2x)\leq\frac{1}{\varphi(x)}{\mathbf{E}%
}(\varphi(S_{2^{r}}))+q^{p}x^{-p}\Big(\sum_{l=0}^{r-1}\Big (\sum
_{k=1}^{2^{r-l}-1}||{\mathbf{E}}(S_{(k+1)2^{l}}-S_{k2^{l}}|{\mathcal{F}%
}_{k2^{l}})||_{p}^{p}\Big )^{1/p}\Big )^{p}\,. \label{inegalitemax}%
\end{equation}
Assume in addition that there exists a positive real $M$ such that $\sup
_{i}\Vert Y_{i}\Vert_{\infty}\leq M$. Then for any positive real $x$, and any
positive integer $r$, the following inequality holds
\begin{equation}
{\mathbf{P}}(S_{2^{r}}^{\ast}\geq4x)\leq\frac{1}{\varphi(x)}{\mathbf{E}%
}(\varphi(S_{2^{r}}))+q^{p}x^{-p}\Big(\sum_{l=0}^{r-1}\Big (\sum
_{k=1}^{2^{r-l}-1}||{\mathbf{E}}(S_{v+(k+1)2^{l}}-S_{v+k2^{l}}|{\mathcal{F}%
}_{k2^{l}})||_{p}^{p}\Big )^{1/p}\Big )^{p}\,, \label{inegalitemax2proba}%
\end{equation}
with $v=[x/M]$ (where $[\,.\,]$ denotes the integer part).
\end{proposition}
\bigskip
\noindent\textbf{Proof of Propositions \ref{maxinequality} and
\ref{propmaxineproba}.} For any $m\in\lbrack0,2^{r}-1]$, we have that
\[
S_{2^{r}-m}={\mathbf{E}}(S_{2^{r}}|{\mathcal{F}}_{2^{r}-m})-{\mathbf{E}%
}(S_{2^{r}}-S_{2^{r}-m}|{\mathcal{F}}_{2^{r}-m})\,.
\]
So,%
\begin{equation}
S_{2^{r}}^{\ast}\leq\max_{0\leq m\leq2^{r}-1}|{\mathbf{E}}(S_{2^{r}%
}|{\mathcal{F}}_{2^{r}-m})|+\max_{0\leq m\leq2^{r}-1}|{\mathbf{E}}(S_{2^{r}%
}-S_{2^{r}-m}|{\mathcal{F}}_{2^{r}-m})|\,. \label{dec1max}%
\end{equation}
Since $({\mathbf{E}}(S_{2^{r}}|{\mathcal{F}}_{k}))_{k\geq1}$ is a martingale,
we shall use Doob's maximal inequality to deal with the first term in the
right-hand side of (\ref{dec1max}). Hence, since $\varphi$ is a nondecreasing,
non negative, convex and even function, we get that
\begin{equation}
{\mathbf{P}}\big (\max_{0\leq m\leq2^{r}-1}|{\mathbf{E}}(S_{2^{r}%
}|{\mathcal{F}}_{2^{r}-m})|\geq x\big )\leq\frac{1}{\varphi(x)}{\mathbf{E}%
}(\varphi(S_{2^{r}}))\,, \label{b1max}%
\end{equation}
and also that
\begin{equation}
\Vert\max_{0\leq m\leq2^{r}-1}|{\mathbf{E}}(S_{2^{r}}|{\mathcal{F}}_{2^{r}%
-m})|\Vert_{p}\leq q\Vert S_{2^{r}}\Vert_{p}\,. \label{momentb1max}%
\end{equation}
Write now $m$ in basis $2$ as follows:
\[
m=\sum_{i=0}^{r-1}b_{i}(m)2^{i},\ \text{ where $b_{i}(m)=0$ or $b_{i}(m)=1$%
}\,.
\]
Set $m_{l}=\sum_{i=l}^{r-1}b_{i}(m)2^{i}$ and notice that for any $p\geq1$, we
have
\[
|{\mathbf{E}}(S_{2^{r}}-S_{2^{r}-m}|{\mathcal{F}}_{2^{r}-m})|^{p}%
\leq\Big (\sum_{l=0}^{r-1}|{\mathbf{E}}(S_{2^{r}-m_{l+1}}-S_{2^{r}-m_{l}%
}|{\mathcal{F}}_{2^{r}-m})|\Big )^{p}\,.
\]
Hence setting
\[
\alpha_{l}=\Big (\sum_{k=1}^{2^{r-l}-1}||{\mathbf{E}}(S_{(k+1)2^{l}}%
-S_{k2^{l}}|{\mathcal{F}}_{k2^{l}})||_{p}^{p}\Big )^{1/p}\text{ and }%
\lambda_{l}=\frac{\alpha_{l}}{\sum_{l=0}^{r-1}\alpha_{l}}\,,
\]
we get by convexity
\begin{align*}
|{\mathbf{E}}(S_{2^{r}}-S_{2^{r}-m}|{\mathcal{F}}_{2^{r}-m})|^{p} & \leq
\sum_{l=0}^{r-1}\lambda_{l}^{1-p}|{\mathbf{E}}(S_{2^{r}-m_{l+1}}%
-S_{2^{r}-m_{l}}|{\mathcal{F}}_{2^{r}-m})|^{p}\\
& \leq\sum_{l=0}^{r-1}\lambda_{l}^{1-p}|{\mathbf{E}}(|{\mathbf{E}}%
(S_{2^{r}-m_{l+1}}-S_{2^{r}-m_{l}}|{\mathcal{F}}_{2^{r}-m_{l}}%
)|\big |{\mathcal{F}}_{2^{r}-m})|^{p}\,.
\end{align*}
Now $m_{l}\neq m_{l+1}$ only if $b_{l}(m)=1$, and in that case $m_{l}%
=k_{m}2^{l}$ with $k_{m}$ odd. It follows that
\begin{align*}
|{\mathbf{E}}(S_{2^{r}-m_{l+1}}-S_{2^{r}-m_{l}}|{\mathcal{F}}_{2^{r}-m_{l}})|
& \leq\max_{1\leq k\leq2^{r-l},k\text{ odd}}|{\mathbf{E}}(S_{2^{r}%
-(k-1)2^{l}}-S_{2^{r}-k2^{l}}|{\mathcal{F}}_{2^{r}-k2^{l}})|\\
& :=A_{r,l}\,.
\end{align*}
Hence, using the fact that if $|X|\leq|Y|$ then ${\mathbf{E}}(|X||{\mathcal{F}%
})\leq{\mathbf{E}}(|Y||{\mathcal{F}})$, we get that
\[
\Vert\max_{0\leq m\leq2^{r}-1}|{\mathbf{E}}(S_{2^{r}}-S_{2^{r}-m}%
|{\mathcal{F}}_{2^{r}-m})|\Vert_{p}^{p}\leq\sum_{l=0}^{r-1}\lambda_{l}%
^{1-p}{\mathbf{E}}(\max_{0\leq m\leq2^{r}-1}|{\mathbf{E}}(A_{r,l}%
|{\mathcal{F}}_{2^{r}-m})|)^{p}\,.
\]
Notice that $({\mathbf{E}}(A_{r,l}|{\mathcal{F}}_{k}))_{k\geq1}$ is a
martingale and by Doob's maximal inequality, we get that
\[
{\mathbf{E}}(\max_{0\leq m\leq2^{r}-1}{\mathbf{E}}(A_{r,l}|{\mathcal{F}%
}_{2^{r}-m}))^{p}\leq q^{p}\Vert A_{r,l}\Vert_{p}^{p}\leq q^{p}\alpha_{l}%
^{p}\,.
\]
Using the definition of $\lambda_{l}$, we then get that
\begin{gather}
\Vert\max_{0\leq m\leq2^{r}-1}|{\mathbf{E}}(S_{2^{r}}-S_{2^{r}-m}%
|{\mathcal{F}}_{2^{r}-m})|\Vert_{p}^{p}\leq q^{p}\big (\sum_{l=0}^{r-1}%
\alpha_{l}\big )^{p}\label{b2maxmomentproba}\\
\leq q^{p}\Big(\sum_{l=0}^{r-1}\Big (\sum_{k=1}^{2^{r-l}-1}||{\mathbf{E}%
}(S_{(k+1)2^{l}}-S_{k2^{l}}|{\mathcal{F}}_{k2^{l}})||_{p}^{p}\Big )^{1/p}%
\Big )^{p}\,.\nonumber
\end{gather}
Starting from (\ref{dec1max}) and using (\ref{b1max}) and
({\ref{b2maxmomentproba}) combined with Markov's inequality, Inequality
(\ref{inegalitemax}) of Proposition \ref{propmaxineproba} follows. To end the
proof of Proposition \ref{maxinequality}, we start from (\ref{dec1max}) and
consider the bounds (\ref{momentb1max}) and (\ref{b2maxmomentproba}). }
We turn now to the proof of Inequality (\ref{inegalitemax2proba}). We start
from (\ref{dec1max}) and we write that
\begin{align*}
S_{2^{r}}^{\ast} & \leq\max_{0\leq m\leq2^{r}-1}|{\mathbf{E}}(S_{2^{r}%
}|{\mathcal{F}}_{2^{r}-m})|+\max_{0\leq m\leq2^{r}-1}|{\mathbf{E}}(S_{2^{r}%
+v}-S_{2^{r}+v-m}|{\mathcal{F}}_{2^{r}-m})|\\
& +\max_{0\leq m\leq2^{r}-1}|{\mathbf{E}}(S_{2^{r}+v}-S_{2^{r}}|{\mathcal{F}%
}_{2^{r}-m})|+\max_{0\leq m\leq2^{r}-1}|{\mathbf{E}}(S_{2^{r}+v-m}-S_{2^{r}%
-m}|{\mathcal{F}}_{2^{r}-m})|\,.
\end{align*}
By the fact that the variables are uniformly bounded by $M$, we then derive
that
\[
S_{2^{r}}^{\ast}\leq\max_{0\leq m\leq2^{r}-1}|{\mathbf{E}}(S_{2^{r}%
}|{\mathcal{F}}_{2^{r}-m})|+\max_{0\leq m\leq2^{r}-1}|{\mathbf{E}}(S_{2^{r}%
+v}-S_{2^{r}+v-m}|{\mathcal{F}}_{2^{r}-m})|+2vM\,.
\]
Since $vM\leq x$, it follows that
\begin{gather}
{\mathbf{P}}(S_{2^{r}}^{\ast}\geq4x)\leq{\mathbf{P}}(\max_{0\leq m\leq2^{r}%
-1}|{\mathbf{E}}(S_{2^{r}}|{\mathcal{F}}_{2^{r}-m})|\geq
x)\label{b3maxmomentproba}\\
+{\mathbf{P}}(\max_{0\leq m\leq2^{r}-1}|{\mathbf{E}}(S_{2^{r}+v}-S_{2^{r}%
+v-m}|{\mathcal{F}}_{2^{r}-m})|\geq x)\,.\nonumber
\end{gather}
Using chaining arguments, convexity and the Doob's maximal inequality, as
above, we infer that for any $p>1$,
\begin{align}
& \Vert\max_{0\leq m\leq2^{r}-1}|{\mathbf{E}}(S_{2^{r}+v}-S_{2^{r}%
+v-m}|{\mathcal{F}}_{2^{r}-m})|\Vert_{p}^{p}\label{b4maxmomentproba}\\
& \leq q^{p}\Big(\sum_{l=0}^{r-1}\Big (\sum_{k=1}^{2^{r-l}-1}||{\mathbf{E}%
}(S_{v+(k+1)2^{l}}-S_{v+k2^{l}}|{\mathcal{F}}_{k2^{l}})||_{p}^{p}%
\Big )^{1/p}\Big )^{p}\,.\nonumber
\end{align}
Starting from (\ref{b3maxmomentproba}) and using (\ref{b1max}) and
({\ref{b4maxmomentproba}) combined with Markov's inequality, Inequality
(\ref{inegalitemax2proba}) of Proposition \ref{propmaxineproba} follows.
}$\diamond$
\section{Moment inequalities for the maximum of partial sums under projective
conditions}
\label{sectionmomentine}
\subsection{Rosenthal-type inequalities for stationary processes}
\label{sectiongeneralpowers}
Using a direct approach that combines the maximal inequality (\ref{cons1sub})
and the Lemma \ref{cross}, we obtain the following Rosenthal inequality for
the maximum of the partial sums of a stationary process for all powers $p>2$.
\begin{theorem}
\label{directprop} Let $p>2$ be a real number and let $(X_{i})_{i\in
\mathbb{Z}}$ be an adapted stationary sequence in the sense of Notation
\ref{notation1}. Then, for any positive integer $n$, the following inequality
holds:
\[
{\mathbf{E}}\big (\max_{1\leq j\leq n}|S_{j}|^{p}\big )\ll n{\mathbf{E}%
}(|X_{1}|)^{p}+cn\left( \sum_{k=1}^{n}\frac{1}{k^{1+1/p\ }}||{\mathbf{E}}%
_{0}(S_{k})||_{p}\right) ^{p}+n\left( \sum_{k=1}^{n}\frac{1}{k^{1+2\delta
/p}}\Vert{\mathbf{E}}_{0}(S_{k}^{2})\Vert_{p/2}^{\delta}\right)
^{p/(2\delta)}\,,
\]
for any positive integer $r$ with $\delta=\min(1,1/(p-2))$ and $c=1.$ When
$p\geq4$ we can take $c=0$ by enlarging the constant involved.
\end{theorem}
\begin{comment}
\label{commentgreaterdelta} 1. Notice that for $2<p\leq3$ the inequality holds
with $\delta=1$ and therefore it provides a maximal form for Theorem 3.1 in
Rio (2009). \newline2. It is interesting to indicate the monotonicity of the
right-hand side of the inequality in $\delta$. To be more precise, for any
$0<\delta\leq\gamma\leq1$, the following inequality holds:
\[
\Big (\sum_{k=1}^{n}k^{-1-2\gamma/p}\Vert{\mathbf{E}}_{0}(S_{k}^{2}%
)\Vert_{p/2}^{\gamma}\Big )^{1/\gamma}\leq2^{(1+\gamma)(\gamma-\delta
)/(\delta\gamma)}\Big (\sum_{k=1}^{n}k^{-1-2\delta/p}\Vert{\mathbf{E}}%
_{0}(S_{k}^{2})\Vert_{p/2}^{\delta}\Big )^{1/\delta}\,.
\]
To see this, we notice the subadditivity property $\Vert{\mathbf{E}}%
_{0}(S_{i+j}^{2})\Vert_{p/2}\leq2\Vert{\mathbf{E}}_{0}(S_{i}^{2})\Vert
_{p/2}+2\Vert{\mathbf{E}}_{0}(S_{j}^{2})\Vert_{p/2}$ and apply then the item 3
of Lemma \ref{lmasubadd} with $C=2$. \smallskip\newline3. On the other hand
for any $0<\delta<1$ and any $\gamma>1/\delta-1$, by H\"{o}lder's inequality,
there exists a positive constant $C$ depending on $p$, $\gamma$ and $\delta$,
such that
\begin{equation}
\Big (\sum_{k=1}^{n}k^{-1-2\delta/p}\Vert{\mathbf{E}}_{0}(S_{k}^{2}%
)\Vert_{p/2}^{\delta}\Big )^{p/(2\delta)}\leq C\Big (\sum_{k=1}^{n}%
k^{-1-2/p}(\log k)^{\gamma}\ \Vert{\mathbf{E}}_{0}(S_{k}^{2})\Vert
_{p/2}\Big )^{p/2}\,. \label{delta1}%
\end{equation}
\newline4. As a matter of fact we shall prove first the inequality from
Theorem \ref{directprop} in a slightly different form which is equivalent up
to multiplicative constants: for any positive integer $r$ and any integer $n$
such that $2^{r-1}\leq n<2^{r}$, ($\delta$ and $c$ as above$)$
\begin{equation}
{\mathbf{E}}\big (\max_{1\leq j\leq n}|S_{j}|^{p}\big )\ll n{\mathbf{E}%
}(|X_{1}|)^{p}+cn\Big (\sum_{k=0}^{r-1}2^{-k/p}\Vert{\mathbf{E}}_{0}(S_{2^{k}%
})\Vert_{p}\Big )^{p}+n\Big (\sum_{k=0}^{r-1}2^{-2k\delta/p}\Vert{\mathbf{E}%
}_{0}(S_{2^{k}}^{2})\Vert_{p/2}^{\delta}\Big )^{p/(2\delta)}\,.
\label{dyaticform}%
\end{equation}
\end{comment}
With applications to Markov processes in mind, by conditioning with respect to
both the future and to the past of the process, our next result gives an
alternative inequality than the one given in Theorem \ref{directprop} when $p$
is an even integer. For this case, the power $\delta$ appearing in Theorem
\ref{directprop} is always equal to one. Before stating the result, we first
introduce the following notation to define the additional nonincreasing
filtration that we consider.
\begin{notation}
\label{notation2} Let $\mathcal{\bar{F}}_{0}$ be a $\sigma$-algebra of
$\mathcal{A}$ satisfying $T^{-1}(\mathcal{\bar{F}}_{0})\subseteq
\mathcal{\bar{F}}_{0}$. We then define the nonincreasing filtration
$(\mathcal{\bar{F}}_{i})_{i\in\mathbb{Z}}$ by $\mathcal{\bar{F}}_{i}%
=T^{-i}(\mathcal{\bar{F}}_{0})$. In what follows, we use the notation
${\mathbf{\bar{E}}}_{k}(Y)={\mathbf{E}}(Y|\mathcal{\bar{F}}_{k})$.
\end{notation}
\begin{theorem}
\label{stateven} Let $p\geq4$ be an even integer and let $X_{0}$ be a real
valued random variable such that $\Vert X_{0}\Vert_{p}<\infty$ and measurable
with respect to ${\mathcal{F}}_{0}$ and to ${\mathcal{\bar{F}}}_{0}$. We
construct the stationary sequence $(X_{i})_{i\in\mathbb{Z}}$ as in Notation
\ref{notation1}. Then for any integer $n$,
\begin{align*}
{\mathbf{E}}\big (\ \max_{1\leq j\leq n}\ |S_{j}|^{p}\big ) & \ll
n{\mathbf{E}}(|X_{1}|^{p})+n\Big (\sum_{k=1}^{n}\frac{1}{k^{1+1/p}}%
\big (\Vert{\mathbf{E}}_{0}(S_{k})\Vert_{p}+\Vert{\mathbf{\bar{E}}}%
_{k+1}(S_{k})\Vert_{p}\big )\Big )^{p}\\
& +n\Big (\sum_{k=1}^{n}\frac{1}{k^{1+2/p}}\big (\Vert{\mathbf{E}}_{0}%
(S_{k}^{2})\Vert_{p/2}+\Vert{\mathbf{\bar{E}}}_{k+1}(S_{k}^{2})\Vert
_{p/2}\big )\Big )^{p/2}\,.
\end{align*}
\end{theorem}
As a corollary to the proof of Theorem \ref{stateven} we obtain:
\begin{theorem}
\label{pinteger}Let $p\geq4$ be a real number and let $X_{0}$ be a real valued
random variables such that $\Vert X_{0}\Vert_{p}<\infty$ and measurable with
respect to ${\mathcal{F}}_{0}$ and to ${\mathcal{\bar{F}}}_{0}$. We construct
the stationary sequence $(X_{i})_{i\in\mathbb{Z}}$ as in Notation
\ref{notation1}. Then for any integer $n$,
\[
{\mathbf{E}}\big (\ \max_{1\leq j\leq n}\ |S_{j}|^{p}\big )\ll n{\mathbf{E}%
}(|X_{1}|^{p})+n\Big (\sum_{k=1}^{n}\frac{1}{k^{1+1/p}}\big (\Vert{\mathbf{E}%
}_{0}(S_{k}^{2})\Vert_{p/2}^{1/2}+\Vert{\mathbf{\bar{E}}}_{k+1}(S_{k}%
^{2})\Vert_{p/2}^{1/2}\big )\Big )^{p}\,.
\]
\end{theorem}
This theorem is also valid for $2<p<4$. In this range however, according to
the comment \ref{commentgreaterdelta}, Theorem \ref{directprop} gives better bounds.
\medskip
\noindent\textbf{Proof of Theorem \ref{directprop}.} The proof of this theorem
is based on dyadic induction and involves several steps. With the notation
$a_{n}=||S_{n}||_{p}$ we shall establish a recurrence formula: for any
positive integer $r$,
\begin{equation}
a_{2^{r}}^{p}\leq2^{r}\Big (a_{2^{0}}^{p}+c_{1}\sum_{k=0}^{r-1}2^{-k}a_{2^{k}%
}^{p-1}\Vert{\mathbf{E}}_{0}(S_{2^{k}})\Vert_{p}+c_{2}\sum_{k=0}^{r-1}%
2^{-k}a_{2^{k}}^{p-2\delta}\Vert{\mathbf{E}}_{0}(S_{2^{k}}^{2})\Vert
_{p/2}^{\delta}\Big ), \label{recurrence}%
\end{equation}
where $c_{1}$ and $c_{2}$ are positive constants depending only on $p$. Before
proving it, let us show that (\ref{recurrence}) implies our result that we
state as a lemma.
\begin{lemma}
\label{reclemma}Assume that for some $0<\delta\leq1$ the recurrence formula
(\ref{recurrence}) holds. Then the inequalities (\ref{dyaticform}) and
(\ref{inedirect}) hold with the same $\delta.$
\end{lemma}
Let us prove the lemma. We shall establish first the inequality
(\ref{dyaticform}). Due to the maximal inequality (\ref{maxdyatic}), it
suffices to prove that the inequality is satisfied for $\max_{1\leq j\leq
n}{\mathbf{E}}(|S_{j}|^{p})$ instead of ${\mathbf{E}}\big (\max_{1\leq j\leq
n}|S_{j}|^{p}\big )$. Denote $\bar{S}_{n}=X_{n+1}+\dots+X_{2n}$.
The proof is divided in several steps. The goal is to establish that for any
positive integer $r$ and any integer $n$ such that $2^{r-1}\leq n<2^{r}$,
\begin{equation}
\max_{1\leq j\leq n}{\mathbf{E}}\big (|S_{j}|^{p}\big )\ll n{\mathbf{E}%
}(|X_{1}|)^{p}+cn\Big (\sum_{k=0}^{r-1}2^{-k/p}\Vert{\mathbf{E}}_{0}(S_{2^{k}%
})\Vert_{p}\Big )^{p}+n\Big (\sum_{k=0}^{r-1}2^{-2k\delta/p}\Vert{\mathbf{E}%
}_{0}(S_{2^{k}}^{2})\Vert_{p/2}^{\delta}\Big )^{p/2\delta}\,.
\label{ineqtoshow}%
\end{equation}
With the notation $\displaystyle B_{r}=\max_{0\leq k\leq r}(a_{2^{k}}%
^{p}/2^{k}),$ starting from (\ref{recurrence}), we get
\[
B_{r}\leq a_{2^{0}}^{p}+c_{1}B_{r}^{1-1/p}\sum_{k=0}^{r-1}2^{-k/p}%
\Vert{\mathbf{E}}_{0}(S_{2^{k}})\Vert_{p}+c_{2}B_{r}^{1-2\delta/p}\sum
_{k=0}^{r-1}2^{-2k\delta/p\ }\Vert{\mathbf{E}}_{0}(S_{2^{k}}^{2})\Vert
_{p/2}^{\delta}\,.
\]
Therefore, taking into account that either $B_{r}\leq3a_{2^{0}}^{p}$ or
$B_{r}^{1/p}\leq3c_{1}\sum_{k=0}^{r-1}2^{-k/p}\Vert{\mathbf{E}}_{0}(S_{2^{k}%
})\Vert_{p}$ or $B_{r}^{2\delta/p}\leq3c_{2}\sum_{k=0}^{r-1}2^{-2k\delta
/p\ }\Vert{\mathbf{E}}_{0}(S_{2^{k}}^{2})\Vert_{p/2}^{\delta}$, we derive
that
\begin{equation}
a_{2^{r}}^{p}\leq2^{r}\left( 3a_{2^{0}}^{p}+\Big (3c_{1}\sum_{k=0}%
^{r-1}2^{-k/p}\Vert{\mathbf{E}}_{0}(S_{2^{k}})\Vert_{p}\Big )^{p}%
+\Big (3c_{2}\sum_{k=0}^{r-1}2^{-2k\delta/p\ }\Vert{\mathbf{E}}_{0}(S_{2^{k}%
}^{2})\Vert_{p/2}^{\delta}\Big )^{p/2\delta}\right) \,.
\label{boundpropdyatic2}%
\end{equation}
Let now $2^{r-1}\leq n<2^{r}$ and write its binary expansion:
\begin{equation}
n=\sum_{k=0}^{r-1}2^{k}b_{k}\text{ where }b_{r-1}=1\text{ and }b_{k}%
\in\{0,1\}\text{ for }k=0,\dots,r-2\,. \label{binaryexpansion}%
\end{equation}
Notice that by stationarity
\[
\Vert S_{n}\Vert_{p}\leq\sum_{k=0}^{r-1}b_{k}\Vert S_{2^{k}}\Vert_{p}\,.
\]
Then, by using (\ref{boundpropdyatic2}) and the fact that $\sum_{k=0}%
^{r-1}b_{k}2^{k/p}\leq2^{r/p}/(1-2^{-1/p})$, we derive the inequality
(\ref{ineqtoshow}) for ${\mathbf{E}}(|S_{n}|^{p})$ and also for $\max_{1\leq
j\leq n}{\mathbf{E}}\big (|S_{j}|^{p}\big )$. The inequality (\ref{dyaticform}%
) follows now by the maximal inequality (\ref{maxdyatic}).
We indicate now how to derive from (\ref{dyaticform}), the inequality stated
in Theorem \ref{directprop}. Notice that, by stationarity, for any integers
$i$ and $j$,
\[
\Vert{\mathbf{E}}_{0}(S_{i+j})\Vert_{p}\leq\Vert{\mathbf{{E}}}_{0}(S_{i}%
)\Vert_{p}+\Vert{\mathbf{{E}}}_{0}(S_{j})\Vert_{p}\,,
\]
and also that for any $0<\delta\leq1$,
\[
\Vert{\mathbf{E}}_{0}(S_{i+j}^{2})\Vert_{p/2}^{\delta}\leq2^{\delta}%
\Vert{\mathbf{E}}_{0}(S_{i}^{2})\Vert_{p/2}^{\delta}+2^{\delta}\Vert
{\mathbf{E}}_{0}(S_{j}^{2})\Vert_{p/2}^{\delta}\,.
\]
Using Item 1 of Lemma \ref{lmasubadd}, it follows that
\[
\sum_{k=0}^{r-1}2^{-k/p}\Vert{\mathbf{E}}_{0}(S_{2^{k}})\Vert_{p}\ll\sum
_{k=1}^{n}k^{-1-1/p\ }\Vert{\mathbf{E}}_{0}(S_{k})\Vert_{p}\,,
\]
and%
\begin{equation}
\sum_{k=0}^{r-1}2^{-2k\delta/p}\Vert{\mathbf{E}}_{0}(S_{2^{k}}^{2})\Vert
_{p/2}^{\delta}\ll\sum_{k=1}^{n}k^{-1-2\delta/p}\Vert{\mathbf{E}}_{0}%
(S_{k}^{2})\Vert_{p/2}^{\delta}\,. \label{majdyadelta}%
\end{equation}
The results follows by the above considerations via the inequality
(\ref{dyaticform}). $\diamond$
\bigskip
\textbf{End of the proof of Theorem \ref{directprop}.} It remains to establish
the recurrence formula (\ref{recurrence}). We divide the proof in three cases
according to the values of $p$.
The case $2<p\leq3$ was discussed in Rio (2009). We give here a shorter
alternative proof. We apply inequality (\ref{Rio}) of Lemma \ref{cross} with
$x=S_{2^{r-1}}$ and $y=\bar{S}_{2^{r-1}}.$ Then by taking the expectation and
using stationarity and properties of conditional expectation, we obtain
\[
{\mathbf{E}}(|S_{2^{r}}|)^{p}\leq2{\mathbf{E}}(|S_{2^{r-1}}|)^{p}%
+p{\mathbf{E}}(|S_{2^{r-1}}|^{p-1}\mathrm{sign}(S_{2^{r-1}}){\mathbf{E}%
}_{2^{r-1}}(\bar{S}_{2^{r-1}}))+C_{p}^{2}{\mathbf{E}}(|S_{2^{r-1}}%
|^{p-2}{\mathbf{E}}_{2^{r-1}}(\bar{S}_{2^{r-1}}^{2}))\,,
\]
where $C_{p}^{2}=p(p-1)/2$. This inequality combined with the H\"{o}lder's
inequality gives
\[
a_{2^{r}}^{p}\leq2a_{2^{r-1}}^{p}+pa_{2^{r-1}}^{p-1}\Vert{\mathbf{E}}%
_{0}(S_{2^{r-1}})\Vert_{p}+C_{p}^{2}a_{2^{r-1}}^{p(1-2/p)}\Vert{\mathbf{E}%
}_{0}(S_{2^{r-1}}^{2})\Vert_{p/2}\,.
\]
By recurrence according to the first term in the right hand side, it follows
that for any integer $r$,
\[
a_{2^{r}}^{p}\leq2^{r}\Big (a_{2^{0}}^{p}+2^{-1}p\sum_{k=0}^{r-1}%
2^{-k}a_{2^{k}}^{p-1}\Vert{\mathbf{E}}_{0}(S_{2^{k}})\Vert_{p}+2^{-1}C_{p}%
^{2}\sum_{k=0}^{r-1}2^{-k}a_{2^{k}}^{p(1-2/p)}\Vert{\mathbf{E}}_{0}(S_{2^{k}%
}^{2})\Vert_{p/2}\Big )\,,
\]
and therefore (\ref{recurrence}) holds with $\delta=1$, $c_{1}=2^{-1}p$, and
$c_{2}=2^{-1}C_{p}^{2}$.
Assume now that $p\in]3,4[$. Using the inequality (\ref{Rio34}) of Lemma
\ref{cross} with $x=S_{2^{r-1}}$ and $y=\bar{S}_{2^{r-1}}$, taking the
expectation, and using Lemma \ref{basic}, we get by stationarity that for any
positive integer $r$,
\[
a_{2^{r}}^{p}\leq2a_{2^{r-1}}^{p}+pa_{2^{r-1}}^{p-1}\Vert{\mathbf{E}}%
_{0}(S_{2^{r-1}})\Vert_{p}+C_{p}^{2}a_{2^{r-1}}^{p(1-2/p)}\Vert{\mathbf{E}%
}_{0}(S_{2^{r-1}}^{2})\Vert_{p/2}+2p(p-2)^{-1}a_{2^{r-1}}^{p-2/(p-2)}%
||{\mathbf{E}}_{0}(S_{2^{r-1}}^{2})||_{p/2}^{1/(p-2)}\,.
\]
Since $\Vert{\mathbf{E}}_{0}(S_{2^{r-1}}^{2})\Vert_{p/2}\leq a_{2^{r-1}%
}^{2(p-3)/(p-2)}\Vert{\mathbf{E}}_{0}(S_{2^{r-1}}^{2})\Vert_{p/2}^{1/(p-2)}$,
it follows that
\[
\label{boundarp34}a_{2^{r}}^{p}\leq2a_{2^{r-1}}^{p}+pa_{2^{r-1}}^{p-1}%
\Vert{\mathbf{E}}_{0}(S_{2^{r-1}})\Vert_{p}+4pa_{2^{r-1}}^{p-2/(p-2)}%
||{\mathbf{E}}_{0}(S_{2^{r-1}}^{2})||_{p/2}^{1/(p-2)}\,.
\]
By recurrence, we derive that for any integer $r$,
\[
a_{2^{r}}^{p}\leq2^{r}\Big (a_{2^{0}}^{p}+2^{-1}p\sum_{k=0}^{r-1}%
2^{-k}a_{2^{k}}^{p-1}\Vert{\mathbf{E}}_{0}(S_{2^{k}})\Vert_{p}+2p\sum
_{k=0}^{r-1}2^{-k}a_{2^{k}}^{p-2/(p-2)}\Vert{\mathbf{E}}_{0}(S_{2^{k}}%
^{2})\Vert_{p/2}^{1/(p-2)}\Big )\,.
\]
It follows that (\ref{recurrence}) holds with $\delta=1/(p-2)$, $c_{1}%
=2^{-1}p$, and $c_{2}=2p$.
It remains to prove the inequality (\ref{inedirect}) for $p\geq4$. Using the
inequality (\ref{int}) of Lemma \ref{cross} with $x=S_{2^{r-1}}$ and
$y=\bar{S}_{2^{r-1}}$, and taking the expectation, we get by stationarity
that
\[
a_{2^{r}}^{p}\leq2a_{2^{r-1}}^{p}+4^{p}{\mathbf{E}}\big (|S_{2^{r-1}}%
|^{p-1}\mathrm{|}\bar{S}_{2^{r-1}}|+\ \mathrm{|}\bar{S}_{2^{r-1}}%
|^{p-1}|S_{2^{r-1}}|\big )\,.
\]
Using Lemma \ref{basic} together with stationarity, it follows that
\[
{\mathbf{E}}(|S_{2^{r-1}}||\bar{S}_{2^{r-1}}|^{p-1})\leq a_{2^{r-1}%
}^{p-2/(p-2)}||{\mathbf{E}}_{0}(S_{2^{r-1}}^{2})||_{p/2}^{1/(p-2)}\,,
\]
and that
\[
{\mathbf{E}}(\ |S_{2^{r-1}}|^{p-1}|\bar{S}_{2^{r-1}}|)\leq a_{2^{r-1}}%
^{p-1}||{\mathbf{E}}_{0}(S_{2^{r-1}}^{2})||_{p/2}^{1/2}\leq a_{2^{r-1}%
}^{p-2/(p-2)}||{\mathbf{E}}_{0}(S_{2^{r-1}}^{2})||_{p/2}^{1/(p-2)}\,.
\]
From these estimates we deduce
\[
a_{2^{r}}^{p}\leq2a_{2^{r-1}}^{p}+2(4^{p})a_{2^{r-1}}^{p-2/(p-2)}%
\Vert{\mathbf{E}}_{0}(S_{2^{r-1}}^{2})\Vert_{p/2}^{1/(p-2)}\,.
\]
By recurrence, it follows that for any integer $r$,
\[
a_{2^{r}}^{p}\leq2^{r}\Big (a_{2^{0}}^{p}+4^{p}\sum_{k=0}^{r-1}2^{-k}a_{2^{k}%
}^{p-2/(p-2)}\Vert{\mathbf{E}}_{0}(S_{2^{k}}^{2})\Vert_{p/2}^{1/(p-2)}%
\Big )\,.
\]
It follows that (\ref{recurrence}) holds with $\delta=1/(p-2)$, $c_{1}=0$, and
$c_{2}=4^{p}$. Therefore, in this case (\ref{ineqtoshow}) holds with $c=0$.
Then by the maximal inequality (\ref{maxdyatic}), the inequality
(\ref{dyaticform}) holds with $c=1$. We show now that, in this case, the
second term in the inequality can be bounded up to a multiplicative constant
by the third term. By Jensen inequality and since in this case $\delta<1/2,$
we have
\[
\Big (\sum_{k=0}^{r-1}2^{-k/p}||{\mathbf{E}}_{0}(S_{2^{k}})||_{p}%
\leq\Big (\sum_{k=0}^{r-1}2^{-(2k/p)(1/2)}||{\mathbf{E}}_{0}(S_{2^{k}}%
^{2})||_{p/2}^{1/2}\Big )^{2}\leq\Big (\sum_{k=0}^{r-1}2^{-(2k\ /p)\delta
}\Vert{\mathbf{E}}_{0}(S_{2^{k}}^{2})\Vert_{p/2}^{\delta}\Big )^{1/\delta}.
\]
We then finish the proof by using (\ref{majdyadelta}). $\diamond$
\bigskip
\medskip\noindent\textbf{Proof of Theorem \ref{stateven}.} Denote $\bar{S}%
_{n}=X_{n+1}+\dots+X_{2n}$. Starting from the inequality (\ref{evenint}) of
Lemma \ref{cross} applied with $x=S_{n}$ and $y=\bar{S}_{n}$ and using the
notation $a_{n}=||S_{n}||_{p}$, by stationarity, we get that
\begin{equation}
a_{2n}^{p}\leq2a_{n}^{p}+p\big ({\mathbf{E}}(S_{n}^{p-1}{\bar{S}}%
_{n})+{\mathbf{E}}(S_{n}{\bar{S}}_{n}^{p-1})\big )+2^{p}\big ({\mathbf{E}%
}(S_{n}^{p-2}\bar{S}_{n}^{2})+{\mathbf{E}}(\bar{S}_{n}^{p-2}S_{n}%
^{2})\big )\,. \label{b1a2n}%
\end{equation}
By using the H\"{o}lder's inequality and recurrence, we then derive that for
any positive integer $r$,
\begin{align*}
a_{2^{r}}^{p} & \leq2^{r}a_{2^{0}}^{p}+2^{-1}p\sum_{k=0}^{r-1}%
2^{r-k}a_{2^{k}}^{p-1}\big (\Vert{\mathbf{E}}_{0}(S_{2^{k}})\Vert_{p}%
+\Vert{\mathbf{\bar{E}}}_{2^{k}+1}(S_{2^{k}})\Vert_{p}\big )\\
& +2^{p-1}\sum_{k=0}^{r-1}2^{r-k}a_{2^{k}}^{p-2}\big (\Vert{\mathbf{E}}%
_{0}(S_{2^{k}}^{2})\Vert_{p/2}+\Vert{\mathbf{\bar{E}}}_{2^{k}+1}(S_{2^{k}}%
^{2})\Vert_{p/2}\big )\,.
\end{align*}
By using the arguments of the proof of Lemma \ref{reclemma}, we get for
$2^{r-1}\leq n<2^{r},$ \
\begin{align*}
{\mathbf{E}}\big (\ \max_{1\leq j\leq n}\ |S_{j}|^{p}\big ) & \ll
n{\mathbf{E}}(|X_{1}|^{p})+n\Big (\sum_{k=1}^{r-1}2^{-k/p}\big (\Vert
{\mathbf{E}}_{0}(S_{2^{k}})\Vert_{p}+\Vert{\mathbf{\bar{E}}}_{k+1}(S_{2^{k}%
})\Vert_{p}\big )\Big )^{p}\\
& +n\Big (\sum_{k=1}^{n}2^{-2k/p}\big (\Vert{\mathbf{E}}_{0}(S_{2^{k}}%
^{2})\Vert_{p/2}+\Vert{\mathbf{\bar{E}}}_{k+1}(S_{2^{k}}^{2})\Vert
_{p/2}\big )\Big )^{p/2}\,.
\end{align*}
Noticing in addition that, by stationarity, for any integer $i$ and $j$
\[
\Vert{\mathbf{E}}_{0}(S_{i+j})\Vert_{p}\leq\Vert{\mathbf{E}}_{0}(S_{i}%
)\Vert_{p}+\Vert{\mathbf{E}}_{0}(S_{j})\Vert_{p}\text{ , }\Vert{\mathbf{\bar
{E}}}_{i+j+1}(S_{i+j})\Vert_{p}\leq\Vert{\mathbf{\bar{E}}}_{i+1}(S_{i}%
)\Vert_{p}+\Vert{\mathbf{\bar{E}}}_{j+1}(S_{j})\Vert_{p}\,,
\]%
\begin{equation}
\Vert{\mathbf{\bar{E}}}_{i+j+1}(S_{i+j}^{2})\Vert_{p/2}\leq2\Vert
{\mathbf{\bar{E}}}_{i+1}(S_{i}^{2})\Vert_{p/2}+2\Vert{\mathbf{\bar{E}}}%
_{j+1}(S_{j}^{2})\Vert_{p/2}\,, \label{*}%
\end{equation}
and that
\begin{equation}
\Vert{\mathbf{E}}_{0}(S_{i+j}^{2})\Vert_{p/2}\leq2\Vert{\mathbf{E}}_{0}%
(S_{i}^{2})\Vert_{p/2}+2\Vert{\mathbf{E}}_{0}(S_{j}^{2})\Vert_{p/2}\,.
\label{**}%
\end{equation}
We obtain the desired result by using Lemma \ref{lmasubadd}. $\diamond$
\bigskip
\medskip\noindent\textbf{Proof of Theorem \ref{pinteger}.} To proof this
theorem we apply the inequality (\ref{int}) of Lemma \ref{cross} with
$x=S_{n}$ and $y=\bar{S}_{n}$, where $\bar{S}_{n}=X_{n+1}+\dots+X_{2n}$. With
the notation $a_{n}=||S_{n}||_{p}$, we then have by stationarity that
\[
a_{2n}^{p}\leq2a_{n}^{p}+4^{p}\big ({\mathbf{E}}(|S_{n}|^{p-1}|{\bar{S}}%
_{n}|)+{\mathbf{E}}(|S_{n}||{\bar{S}}_{n}|^{p-1})\big )\,.
\]
By conditioning and then applying the Jensen's inequality followed by the
H\"{o}lder's inequality we obtain
\begin{align*}
a_{2n}^{p} & \leq2a_{n}^{p}+4^{p}\big ({\mathbf{E}}(|S_{n}|^{p-1}%
{\mathbf{E}}_{n}^{1/2}({\bar{S}}_{n}^{2})|+{\mathbf{E}}(|{\bar{S}}_{n}%
|^{p-1}{\mathbf{\bar{E}}}_{n+1}^{1/2}(S_{n}^{2})\big )\,\\
& \leq2a_{n}^{p}+4^{p}\ a_{n}^{p-1}\big (\Vert{\mathbf{E}}_{0}(S_{n}%
^{2})\Vert_{p/2}^{1/2}+\Vert{\mathbf{\bar{E}}}_{n+1}(S_{n}^{2})\Vert
_{p/2}^{1/2}\big )\,.
\end{align*}
By recurrence, we then derive that for any positive integer $r$,
\[
a_{2^{r}}^{p}\leq2^{r}\Big (a_{0}^{p}+2^{2p-1}\sum_{k=0}^{r-1}2^{-k}a_{2^{k}%
}^{p-1}\big (\Vert{\mathbf{E}}_{0}(S_{2^{k}}^{2})\Vert_{p/2}^{1/2}%
+\Vert{\mathbf{\bar{E}}}_{2^{k}+1}(S_{2^{k}}^{2})\Vert_{p/2}^{1/2}%
\big )\Big )\,.
\]
The proof is completed by the arguments developed in the proof of Lemma
\ref{reclemma} and by using Lemma \ref{lmasubadd} by taking into account the
inequalities (\ref{*}) and (\ref{**}) .$\diamond$
\subsection{Relation with the Burkholder-type Inequality.}
\label{sectionBurk}
Next lemma shows how to compare $\Vert{\mathbf{E}}_{0}(S_{n}^{2})\Vert_{p/2}$
with quantities involving only $\Vert{\mathbf{E}}_{0}(S_{n})\Vert_{p}$.
\begin{lemma}
\label{comparison} Let $p\geq2$ be a real number and let $(X_{n})$ be an
adapted stationary sequence in the sense of Notation \ref{notation1}. Then,
for any positive integer $n$,
\begin{equation}
\Vert{\mathbf{E}}_{0}(S_{n}^{2})\Vert_{p/2}\ll n\Vert{\mathbf{E}}_{0}%
(X_{1}^{2})\Vert_{p/2}+n\left( \sum_{j=1}^{n}\frac{\Vert{\mathbf{E}}%
_{0}(S_{j})\Vert_{p}}{j^{3/2}}\right) ^{2}\,. \label{comp2}%
\end{equation}
\end{lemma}
As a consequence of the above lemma, we get that for any $0<\delta\leq1$ and
any real $p>2$,
\[
n\left( \sum_{j=1}^{n}\frac{\Vert{\mathbf{E}}_{0}(S_{j}^{2})\Vert
_{p/2}^{\delta}}{j^{1+2\delta/p}}\right) ^{p/(2\delta)}\ll n^{p/2}%
\Vert{\mathbf{E}}_{0}(X_{1}^{2})\Vert_{p/2}^{p/2}+n^{p/2}\left( \sum
_{j=1}^{n}\frac{\Vert{\mathbf{E}}_{0}(S_{j})\Vert_{p}}{j^{3/2}}\right)
^{p}\,.
\]
Theorem \ref{directprop} then implies the following Burkholder-type inequality
that was established by Peligrad, Utev and Wu (2007, Theorem 1):
\begin{corollary}
Let $p>2$ be a real number and let $(X_{n})$ be an adapted stationary sequence
in the sense of Notation \ref{notation1}. Then, for any integer $n$,
\[
{\mathbf{E}}\big (\max_{1\leq j\leq n}|S_{j}|^{p}\big )\ll n^{p/2}{\mathbf{E}%
}(|X_{1}|^{p})+n^{p/2}\Big(\sum_{j=1}^{n}\frac{\Vert{\mathbf{E}}_{0}%
(S_{j})\Vert_{p}}{j^{3/2}}\Big)^{p}\,.
\]
\end{corollary}
\noindent\textbf{Proof of Lemma \ref{comparison}.} We shall first prove that
for any positive integer $k$,
\begin{equation}
\Vert{\mathbf{E}}_{0}(S_{2^{k}}^{2})\Vert_{p/2}\leq2^{k+1}\Vert{\mathbf{E}%
}_{0}(X_{1}^{2})\Vert_{p/2}+2^{k+2}\left( \sum_{j=0}^{k-1}\frac
{\Vert{\mathbf{E}}_{0}(S_{2^{j}})\Vert_{p}}{2^{j/2}}\right) ^{2}\,.
\label{comp1}%
\end{equation}
By using the notation $\bar{S}_{2^{k-1}}=X_{2^{k-1}+1}+\dots+X_{2^{k}}$ and
the fact that $S_{2^{k}}^{2}=S_{2^{k-1}}^{2}+\bar{S}_{2^{k-1}}^{2}%
+2S_{2^{k-1}}\bar{S}_{2^{k-1}}$, we get, by stationarity, that
\[
\Vert{\mathbf{E}}_{0}(S_{2^{k}}^{2})\Vert_{p/2}\leq2\Vert{\mathbf{E}}%
_{0}(S_{2^{k-1}}^{2})\Vert_{p/2}+2\Vert{\mathbf{E}}_{0}\big (S_{2^{k-1}%
}{\mathbf{E}}_{2^{k-1}}(\bar{S}_{2^{k-1}})\big)\Vert_{p/2}\,.
\]
Now Cauchy-Schwartz inequality applied first to the conditional expectation
gives
\begin{align*}
\Vert{\mathbf{E}}_{0}\big (S_{2^{k-1}}{\mathbf{E}}_{2^{k-1}}(\bar{S}_{2^{k-1}%
})\big)\Vert_{p/2} & \leq\Vert{\mathbf{E}}_{0}^{1/2}(S_{2^{k-1}}%
^{2}){\mathbf{E}}_{0}^{1/2}({\mathbf{E}}_{2^{k-1}}^{2}(\bar{S}_{2^{k-1}%
}))\Vert_{p/2}\\
& \leq\Vert{\mathbf{E}}_{0}(S_{2^{k-1}}^{2})\Vert_{p/2}^{1/2}\Vert
{\mathbf{E}}_{0}(S_{2^{k-1}})\Vert_{p}\,.
\end{align*}
Hence setting $b_{2^{k}}=\Vert{\mathbf{E}}_{0}(S_{2^{k}}^{2})\Vert_{p/2}$, it
follows that
\[
b_{2^{k}}\leq2b_{2^{k-1}}+2b_{2^{k-1}}^{1/2}\Vert{\mathbf{E}}_{0}(S_{2^{k-1}%
})\Vert_{p}\,.
\]
By induction, this gives that
\[
b_{2^{k}}\leq2^{k}b_{0}+\sum_{j=0}^{k-1}2^{k-j}b_{2^{j}}^{1/2}\Vert
{\mathbf{E}}_{0}(S_{2^{j}})\Vert_{p}\,.
\]
With the notation $B_{k}=\max_{0\leq j\leq k}2^{-j}b_{2^{j}}$, we derive that
\[
B_{k}\leq2\max\big(b_{0},B_{k}^{1/2}\sum_{j=0}^{k-1}2^{-j/2}\Vert{\mathbf{E}%
}_{0}(S_{2^{j}})\Vert_{p}\big )\,,
\]
implying that
\[
2^{-k}b_{2^{k}}\leq B_{k}\leq2b_{0}+2^{2}\Big (\sum_{j=0}^{k-1}2^{-j/2}%
\Vert{\mathbf{E}}_{0}(S_{2^{j}})\Vert_{p}\Big)^{2}\,.
\]
This ends the proof of the inequality (\ref{comp1}).
We turn now to the proof of (\ref{comp2}). Let $r$ be the positive integer
such that $2^{r-1}\leq n<2^{r}$. Starting with the binary expansion
(\ref{binaryexpansion}), and using Minkowski's inequality twice, first with
respect to the conditional expectation, and second with respect to the norm in
${\mathbf{L}}^{p}$, we get by stationarity that
\[
\Vert{\mathbf{E}}_{0}(S_{n}^{2})\Vert_{p/2}\leq\Big (\sum_{k=0}^{r-1}%
b_{k}\Vert({\mathbf{E}}_{0}(S_{2^{k}}^{2}))^{1/2}\Vert_{p}\Big )^{2}%
\leq\Big (\sum_{k=0}^{r-1}\Vert{\mathbf{E}}_{0}(S_{2^{k}}^{2})\Vert
_{p/2}^{1/2}\Big )^{2}\,.
\]
Using then, the inequality (\ref{comp1}), we derive that
\begin{equation}
\Vert{\mathbf{E}}_{0}(S_{n}^{2})\Vert_{p/2}\ll n\Vert{\mathbf{E}}_{0}%
(X_{1}^{2})\Vert_{p/2}+n\left( \sum_{j=0}^{r-1}\frac{\Vert{\mathbf{E}}%
_{0}(S_{2^{j}})\Vert_{p}}{2^{j/2}}\right) ^{2}\,. \label{binary}%
\end{equation}
Since $(\Vert{\mathbf{E}}_{0}(S_{n})\Vert_{p})_{n\geq1}$ is subadditive, using
Item 1 of Lemma \ref{lmasubadd}, Inequality (\ref{comp2}) follows from
(\ref{binary}). $\diamond$
\subsection{Rosenthal inequalities for martingales and the case of even
powers.}
\label{sectionMA}
\subsubsection{The martingale case}
\label{martingalecase}
For any real $p>2$, Theorem \ref{directprop} applied to stationary martingale
differences gives the following inequality:
\[
{\mathbf{E}}\big (\ \max_{1\leq j\leq n}\ |S_{j}|^{p}\big )\ll n{\mathbf{E}%
}(|X_{1}|^{p})+n\Big (\sum_{k=1}^{n}\frac{1}{k^{1+2\delta/p}}\Vert{\mathbf{E}%
}_{0}(S_{k}^{2})\Vert_{p/2}^{\delta}\Big )^{p/(2\delta)}\,,
\]
where $\delta=\min(1,1/(p-2))$.
Since for stationary martingale differences~we have ${\mathbf{E}}(S_{n}%
^{2})=n{\mathbf{E}}(X_{1}^{2})$, we can express the inequality in the
following form useful for applications: \
\begin{equation}
{\mathbf{E}}\big (\ \max_{1\leq j\leq n}\ |S_{j}|^{p}\big )\ll n^{p/2}%
({\mathbf{E}}(X_{1}^{2}))^{p/2}+n{\mathbf{E}}(|X_{1}|^{p})+n\Big (\sum
_{k=1}^{n}\frac{1}{k^{1+2\delta/p}}\Vert{\mathbf{E}}_{0}(S_{k}^{2}%
)-{\mathbf{E}}(S_{k}^{2})\Vert_{p/2}^{\delta}\Big )^{p/(2\delta)}\,.
\label{consdirectmart}%
\end{equation}
As we shall see in the next result, for stationary sequence $(d_{i}%
)_{i\in{\mathbf{Z}}}$ of martingale differences in ${\mathbf{L}}^{p}$ for
$p\geq4$ an even integer, this inequality can be sharpened since it holds with
$\delta=2/(p-2)$ (see Comment \ref{commentgreaterdelta}). As a consequence, we
recover, in case $p=4$, the inequality (1.6) stated in Rio (2009) that was
obtained using the classical Burkholder's inequality combined with the one
given in Theorem 3 in Wu and Zhao (2008) for variables in ${\mathbf{L}}^{q}$
with $q=p/2$. Notice that the inequality (1.6) stated in Rio (2009) cannot be
generalized for $p>4$ since Theorem 3 in Wu and Zhao (2008) is only valid for
variables in ${\mathbf{L}}^{q}$ with $1<q\leq2$.
\begin{theorem}
\label{mart2}Let $p\geq4$ be an even integer and let $d_{0}$ be a real random
variable in ${\mathbf{L}}^{p}$, measurable with respect to $\mathcal{F}_{0}$
and such that ${\mathbf{E}}(d_{0}|{\mathcal{F}}_{-1})=0$. Let $d_{i}%
=d_{0}\circ T^{i}$ and $S_{n}=\sum_{i=1}^{n}d_{i}$. Then for any integer $n$,
\[
{\mathbf{E}}\big (\max_{1\leq j\leq n}\ |S_{j}|^{p}\big )\ll n\ {\mathbf{E}%
}(|d_{1}|^{p})+n\Big (\sum_{k=1}^{n}\frac{1}{k^{1+4/p(p-2)}}\Vert{\mathbf{E}%
}_{0}(S_{k}^{2})\Vert_{p/2}^{2/(p-2)}\Big )^{p(p-2)/4}\,.
\]
\end{theorem}
The technique that makes this result possible is a special symmetrization for
martingales initiated by Kwapie\'n and Woyczynski (1991).
\begin{proposition}
\label{tangent}Assume that $e_{k}$ are stationary martingale differences
adapted to an increasing filtration $(\mathcal{F}_{k})_{k}$ that are
conditionally symmetric (the distribution of $e_{k}$ given $\mathcal{F}_{k-1}$
is equal to the distribution of $-e_{k}$ given $\mathcal{F}_{k-1}$). Assume,
in addition, that the $e_{k}$'s are conditionally independent given a sigma
algebra $\mathcal{G}$ and such that the law of $e_{k}$ given $\mathcal{G}$ is
the same as the law of $e_{k}$ given $\mathcal{F}_{k-1}.$ Let $S_{n}%
=\sum_{i=1}^{n}e_{i}$. Then for any even integer $p\geq4$ and any integer
$n\geq1$,
\begin{equation}
{\mathbf{E}}\big (\max_{1\leq j\leq n}\ |S_{j}|^{p}\big )\ll n\ {\mathbf{E}%
}(|e_{1}|^{p})+n\Big (\sum_{k=1}^{n}\frac{1}{k^{1+4/p(p-2)}}\Vert{\mathbf{E}%
}_{0}(S_{k}^{2})\Vert_{p/2}^{2/(p-2)}\Big )^{p(p-2)/4}\,. \label{ine2tangent}%
\end{equation}
\end{proposition}
\noindent\textbf{Proof of Proposition \ref{tangent}}. Due to the Doob's
maximal inequality, $\Vert\max_{1\leq j\leq n}|S_{j}|\Vert_{p}\leq q\Vert
S_{n}\Vert_{p}$ where $q=p(p-1)^{-1}$. Then, it suffices to show that the
inequality (\ref{ine2tangent}) holds for ${\mathbf{E}}(|S_{n}|^{p})$. We shall
base this proof again on dyadic induction. Denote $\bar{S}_{n}=$
$e_{n+1}+...+e_{2n}$ and $a_{n}=\Vert S_{n}\Vert_{p}$.
We start from the inequality (\ref{b1a2n}). Since the sequence of martingale
differences $(e_{k})$ is conditionally symmetric and conditionally independent
given a master sigma algebra $\mathcal{G}$, we have ${\mathbf{E}}(S_{n}%
^{p-1}{\bar{S}}_{n})+{\mathbf{E}}(S_{n}{\bar{S}}_{n}^{p-1})=0$ and therefore
\[
a_{2n}^{p}\leq2a_{n}^{p}+2^{p}\big ({\mathbf{E}}(S_{n}^{p-2}\bar{S}_{n}%
^{2})+{\mathbf{E}}(\bar{S}_{n}^{p-2}S_{n}^{2})\big )\,.
\]
Using Lemma \ref{basic}, we have that
\[
{\mathbf{E}}(S_{n}^{p-2}\bar{S}_{n}^{2})\leq a_{n}^{p-2}\Vert{\mathbf{E}}%
_{0}(S_{n}^{2})\Vert_{p/2}\text{ and }{\mathbf{E}}(S_{n}^{2}\bar{S}_{n}%
^{p-2})\leq a_{n}^{p-4/(p-2)}\Vert{\mathbf{E}}_{0}(S_{n}^{2})\Vert
_{p/2}^{2/(p-2)}\,.
\]
Therefore, by combining all these bounds, we obtain for every even integer
$p\geq4$,
\begin{align*}
a_{2n}^{p} & \leq2a_{n}^{p}+2^{p}\big (a_{n}^{p-2}\Vert{\mathbf{E}}%
_{0}(S_{n}^{2})\Vert_{p/2}+a_{n}^{p-4/(p-2)}\Vert{\mathbf{E}}_{0}(S_{n}%
^{2})\Vert_{p/2}^{2/(p-2)}\big )\\
& \leq2a_{n}^{p}+2^{p+1}a_{n}^{p-4/(p-2)}\Vert{\mathbf{E}}_{0}(S_{n}%
^{2})\Vert_{p/2}^{2/(p-2)}\,.
\end{align*}
By induction we easily get that for any integer $r$,
\[
a_{2^{r}}^{p}\leq2^{r}\Big (a_{2^{0}}^{p}+2^{p}\sum_{k=0}^{r-1}2^{-k}a_{2^{k}%
}^{p-4/(p-2)}\Vert{\mathbf{E}}_{0}(S_{2^{k}}^{2})\Vert_{p/2}^{2/(p-2)}%
\Big )\,.
\]
We end the proof by Lemma \ref{reclemma}. $\diamond$
\medskip
\noindent\textbf{Proof of Theorem \ref{mart2}.}
We consider our general martingale differences sequence $(d_{k})_{k}$ and we
construct two decoupled tangent versions $(e_{k})_{k}$ and $(\tilde{e}%
_{k})_{k}$ that are $\mathcal{G-}$conditionally independent between them.
These are martingale differences as in Proposition \ref{tangent} with the
additional property that the conditional distribution of $d_{k}$ given
$\mathcal{F}_{k-1}$ is equal to the distribution of $e_{k}$ given
$\mathcal{F}_{k-1}$ and also to the distribution of $\tilde{e}_{k}$ given
$\mathcal{F}_{k-1}$ (see Proposition 6.1.5. in de la Pe\~{n}a and Gin\'{e}
(1999)). Therefore, for any even integer $p$,
\[
{\mathbf{E}}\Big (\sum_{i=1}^{n}d_{i}\Big )^{p}={\mathbf{E}}\Big (\sum
_{i=1}^{n}(d_{i}-{\mathbf{E}}(e_{i}|\mathcal{G}) \Big )^{p}\leq{\mathbf{E}%
}\Big (\sum_{i=1}^{n}(d_{i}-e_{i})\Big )^{p}\,.
\]
Now we use corollary 6.6.8. in de la Pe\~{n}a and Gin\'{e} (1999) (see also
Zinn (1985)). Since $(e_{i}-\tilde{e}_{i})_{i}$ is a decoupled tangent
sequence of $(d_{i}-e_{i})_{i}$, it follows that
\[
{\mathbf{E}}\Big (\sum_{i=1}^{n}d_{i}\Big )^{p}\ll{\mathbf{E}}\Big (\sum
_{i=1}^{n}(e_{i}-\tilde{e}_{i})\Big )^{p}\,.
\]
Notice that the distribution of $e_{i}-\tilde{e}_{i}$ is conditionally
symmetric given $\mathcal{G}$. Therefore, using the Doob's maximal inequality
and applying Proposition \ref{tangent}, we obtain that for every even integer
$p\geq4$ and any integer $n$,
\begin{gather}
{\mathbf{E}}\Big (\max_{1\leq k\leq n}\big (\sum_{i=1}^{k}d_{i}\big )^{p}%
\Big )\ll{\mathbf{E}}\Big (\sum_{i=1}^{n}(e_{i}-\tilde{e}_{i})\Big )^{p}%
\label{b1thmmart2}\\
\ll n\ {\mathbf{E}}(|e_{1}-\tilde{e}_{1}|^{p})+n\Big (\sum_{k=1}^{n}\frac
{1}{k^{1+4/p(p-2)}}\Big \|{\mathbf{E}}_{0}\Big (\Big (\sum_{i=1}^{k}%
(e_{i}-\tilde{e}_{i})\Big )^{2}\Big )\Big \|_{p/2}^{2/(p-2)}\Big )^{p(p-2)/4}%
\,.\nonumber
\end{gather}
Notice now that $\sum_{i=1}^{k}{\mathbf{E}}(e_{i}^{2}|{\mathcal{F}}%
_{i-1})=\sum_{i=1}^{k}{\mathbf{E}}(d_{i}^{2}|{\mathcal{F}}_{i-1})$ since both
quantities are obtained using only the conditional distributions of the
$d_{i}$'s and $e_{i}$'s respectively, and these two sequences are tangent.
Tangency also implies that $d_{i}$ and $e_{i}$ have the same distributions.
Hence $\Vert d_{1}\Vert_{p}=\Vert e_{1}\Vert_{p}$. For the same reasons, we
also have $\sum_{i=1}^{k}{\mathbf{E}}(\tilde{e}_{i}^{2}|{\mathcal{F}}%
_{i-1})=\sum_{i=1}^{k}{\mathbf{E}}(d_{i}^{2}|{\mathcal{F}}_{i-1})$ and $\Vert
d_{1}\Vert_{p}=\Vert\tilde{e}_{1}\Vert_{p}$. Therefore, $\Vert e_{1}-\tilde
{e}_{1}\Vert_{p}\leq2\Vert d_{1}\Vert_{p}$ and
\begin{align*}
\Big \|{\mathbf{E}}_{0}\Big (\Big (\sum_{i=1}^{k}(e_{i}-\tilde{e}%
_{i})\Big )^{2}\Big )\Big \|_{p/2} & \leq2\Big \|{\mathbf{E}}_{0}%
\Big (\sum_{i=1}^{k}{\mathbf{E}}(e_{i}^{2}|{\mathcal{F}}_{i-1}%
)\Big )\Big \|_{p/2}+2\Big \|{\mathbf{E}}_{0}\Big (\sum_{i=1}^{k}{\mathbf{E}%
}(\tilde{e}_{i}^{2}|{\mathcal{F}}_{i-1})\Big )\Big \|_{p/2}\\
& =4\Big \|{\mathbf{E}}_{0}\Big (\sum_{i=1}^{k}{\mathbf{E}}(d_{i}%
^{2}|{\mathcal{F}}_{i-1})\Big )\Big \|_{p/2}=4\Big \|\sum_{i=1}^{k}%
{\mathbf{E}}_{0}(d_{i}^{2})\Big \|_{p/2}\,.
\end{align*}
Theorem \ref{mart2} follows by introducing these bounds in the inequality
(\ref{b1thmmart2}). $\diamond$
\subsubsection{Application to stationary processes via martingale
approximation}
\label{sectionmartappro} Theorem \ref{mart2} together with the martingale
approximation provide an alternative Rosenthal-type inequality involving the
projection operator, very useful for analyzing linear processes.
Next lemma is a slight reformulation of the martingale approximation result
that can be found in the paper by Wu and Woodroofe (Theorem 1, 2004). See also
Zhao and Woodroofe (2008) and Gordin and Peligrad (2010).
\begin{lemma}
\label{mdec}Let $p\geq1$ and let $(X_{n})$ be an adapted stationary sequence
in the sense of Notation \ref{notation1}. Then there is a triangular array of
row-wise stationary martingale differences satisfying
\begin{equation}
D_{0}^{n}=\frac{1}{n}\sum_{i=1}^{n}({\mathbf{E}}_{1}(S_{i})-{\mathbf{E}}%
_{0}(S_{i}))\text{ \ ; }D_{k}^{n}=D_{0}^{n}\circ T^{k} \label{defdiffmart}%
\end{equation}
such that
\[
S_{k}=M_{k}^{n}+R_{k}^{n}\text{ \ \ where }M_{k}^{n}=\sum_{i=1}^{k}D_{i}%
^{n}\,,
\]
and
\[
\max_{1\leq k\leq n}\Vert R_{k}^{n}\Vert_{p}\leq2\Vert X_{0}\Vert_{p}+\frac
{3}{n}\sum_{i=1}^{n}\Vert{\mathbf{E}}_{0}(S_{i})\Vert_{p}\,.
\]
\end{lemma}
We state now the Rosenthal-type inequality that can be provided with the help
of the approximation result above.
\begin{theorem}
\label{general}Let $p\geq4$ be an even integer and let $(X_{i})_{i\in
\mathbb{Z}}$ be as in Theorem \ref{directprop}. Then the following inequality
is valid: for any integer $n$,
\begin{align*}
{\mathbf{E}}(\max_{1\leq k\leq n} & |S_{k}|^{p})\ll n\Vert D_{0}^{n}%
\Vert_{p}^{p}+n\Vert X_{0}\Vert_{p}^{p}+n^{1-p}\Big (\sum_{i=1}^{n}%
\Vert{\mathbf{E}}_{0}(S_{i})\Vert_{p}\Big )^{p}\\
& +n\Big (\sum_{k=1}^{n-1}\frac{1}{k^{1+4/p(p-2)}}\Vert{\mathbf{E}}_{0}%
(S_{k}^{2})\Vert_{p/2}^{2/(p-2)}\Big )^{p(p-2)/4}\,,
\end{align*}
where $D_{0}^{n}$ is defined by (\ref{defdiffmart}).
\end{theorem}
\begin{remark}
\label{rmkcompmartge} Theorems \ref{directprop} and \ref{general} are in
general not comparable. Indeed, for $p\geq4$, Theorem \ref{directprop} applies
with $\delta=1/(p-2)$ so the last term of the inequality stated in Theorem
\ref{general} can be bounded by the last term in the inequality from Theorem
\ref{directprop} (see Comment \ref{commentgreaterdelta}). However the term
involving the quantity $\Vert{\mathbf{E}}_{0}(S_{n})\Vert_{p}$ in Theorem
\ref{general} gives additional contribution. When, for instance,
$\Vert{\mathbf{E}}_{0}(S_{n})\Vert_{p}=O(1)$, (which is the case under the
assumptions of Corollary \ref{Thlin}; see also the remark \ref{rmklinear}),
Theorem \ref{general} might provide a sharper bound.
\end{remark}
\begin{remark}
\label{projest}According to Remark 3.3 in Dedecker, Merlev\`{e}de and Peligrad
(2009), since ${\mathbf{||}}D_{0}^{n}||_{p}\leq\sum_{k=1}^{n}||P_{0}%
(X_{k})||_{p}$, where $P_{0}(X_{i})={\mathbf{E}}_{0}(X_{i})-{\mathbf{E}}%
_{-1}(X_{i})$, we notice that the following bound is valid:
\[
{\mathbf{||}}D_{0}^{n}||_{p}\ll\sum_{k=1}^{n}\frac{1}{k^{1/p}}||{\mathbf{E}%
}_{0}(X_{k})||_{p}\,.
\]
\end{remark}
\noindent\textbf{Proof of Theorem \ref{general}.} By the martingale
approximation of Lemma \ref{mdec} combined with Theorem \ref{mart2}, we get
that
\begin{align*}
& {\mathbf{E}}(\max_{1\leq k\leq n}|S_{k}|^{p})\ll{\mathbf{E}}(\max_{1\leq
k\leq n}|M_{k}^{n}|^{p})+{\mathbf{E}}(\max_{1\leq k\leq n}|R_{k}^{n}|^{p})\\
& \ll n\ {\mathbf{E}}(|D_{0}^{n}|^{p})+\Vert X_{0}\Vert_{p}^{p}%
+\Big (\frac{1}{n}\sum_{i=1}^{n}\Vert{\mathbf{E}}_{0}(S_{i})\Vert
_{p}\Big )^{p}+n\Big (\sum_{k=1}^{n}\frac{1}{k^{1+4/p(p-2)}}\Vert{\mathbf{E}%
}_{0}((M_{k}^{n})^{2})\Vert_{p/2}^{2/(p-2)}\Big )^{p(p-2)/4}\,.
\end{align*}
Since
\[
{\mathbf{E}}_{0}((M_{k}^{n})^{2})\leq2{\mathbf{E}}_{0}(S_{k}^{2}%
)+2{\mathbf{E}}_{0}((R_{k}^{n})^{2})\,,
\]
by using Lemma \ref{mdec}
\[
\Vert{\mathbf{E}}_{0}((M_{k}^{n})^{2})\Vert_{p/2}\ll\Vert{\mathbf{E}}%
_{0}(S_{k}^{2})\Vert_{p/2}+\Vert X_{0}\Vert_{p}^{2}+n^{-2}\Big (\sum_{i=1}%
^{n}\Vert{\mathbf{E}}_{0}(S_{i})\Vert_{p}\Big )^{2}\, ,
\]
and Theorem \ref{general} follows. $\ \diamond$
\subsection{Rosenthal inequality in terms of individual summands.}
\label{sectionindsum}For the sake of applications in this section we indicate
how to estimate the terms that appear in our Rosenthal inequalities in terms
of individual summands and formulate some specific inequalities. By
substracting ${\mathbf{\ E}}(S_{k}^{2})$ and applying the triangle inequality
we can reformulate all the inequalities in terms of the quantities
${\mathbf{E}}(S_{k}^{2}),$ $||{\mathbf{E}}_{0}(S_{k})||_{p}$ and
$\Vert{\mathbf{E}}_{0}(S_{k}^{2})-{\mathbf{E}}(S_{k}^{2})\Vert_{p/2}$. Next
lemma proposes a simple way to estimate these quantities in terms of
coefficients in the spirit of Gordin (1969).
\begin{lemma}
\label{mixing}Under the stationary setting assumptions in Notation
\ref{notation1}, we have the following estimates:
\begin{equation}
{\mathbf{E}}(S_{k}^{2})\leq2k\sum_{j=0}^{k-1}|{\mathbf{E}}(X_{0}X_{j})|\,,
\label{var}%
\end{equation}%
\begin{equation}
||{\mathbf{E}}_{0}(S_{k})||_{p}\leq\sum_{\ell=1}^{n}||{\mathbf{E}}_{0}%
(X_{\ell})||_{p}\,, \label{condterm}%
\end{equation}
and
\begin{gather}
\Vert{\mathbf{E}}_{0}(S_{k}^{2})-{\mathbf{E}}(S_{k}^{2})\Vert_{p/2}\leq
2\sum_{i=1}^{k}\sum_{j=0}^{k-i}\Vert{\mathbf{E}}_{0}(X_{i}X_{i+j}%
)-{\mathbf{E}}(X_{i}X_{i+j})\Vert_{p/2}\label{varsquare}\\
\leq2\sum_{i=1}^{k}\sum_{j=0}^{k-i}\sup_{\ell\geq0}\Vert{\mathbf{E}}_{0}%
(X_{i}X_{i+\ell})-{\mathbf{E}}(X_{i}X_{i+\ell})\Vert_{p/2}\wedge(2\Vert
X_{0}{\mathbf{E}}_{0}(X_{j})\Vert_{p/2})\text{ }\nonumber\\
\leq4\sum_{j=1}^{k}j\Vert X_{0}{\mathbf{E}}_{0}(X_{j})\Vert_{p/2}+2\sum
_{i=1}^{k}i\sup_{j\geq i}\Vert{\mathbf{E}}_{0}(X_{i}X_{j})-{\mathbf{E}}%
(X_{i}X_{j})\Vert_{p/2}\text{.}\nonumber
\end{gather}
\end{lemma}
Mixing coefficients are useful to continue the estimates from Lemma
\ref{mixing}. We refer to the books by Bradley (2007, Theorem 4.13 via Remark
4.7, VI), Rio (2000, Theorem 2.5 and Appendix, Section C) and Dedecker
\textit{et al.} (2007, Remark 2.5 and Ch 3) for various estimates of the
coefficients involved in Lemma \ref{mixing} and examples. We shall also
provide applications and explicit computations of the quantities involved.
We formulate the following proposition:
\begin{proposition}
\label{consdirect} Let $p>2$ be a real number and let $(X_{i})_{i\in
\mathbf{Z}}$ be a stationary sequence of real-valued random variables in
${\mathbf{L}}_{p}$ adapted to an increasing filtration $({\mathcal{F}}_{i})$.
For any $j\geq1$, let
\begin{equation}
\lambda(j)=\max\big (\Vert X_{0}{\mathbf{E}}_{0}(X_{j})\Vert_{p/2},\sup_{i\geq
j}\Vert{\mathbf{E}}_{0}(X_{i}X_{j})-{\mathbf{E}}(X_{i}X_{j})\Vert
_{p/2}\big )\,. \label{notalambda}%
\end{equation}
Then for every positive integer $n$,
\begin{align*}
\mathbf{||}\max_{1\leq j\leq n}|S_{j}|\text{ }||_{p} & \ll n^{1/2}%
\Big (\text{ }\sum_{k=0}^{n-1}|{\mathbf{E}}(X_{0}X_{k})|\Big )^{1/2}%
+n^{1/p}{\mathbf{|}}|X_{1}||_{p}\\
& +cn^{1/p}\sum_{k=1}^{n}\frac{1}{k^{1/p}}||{\mathbf{E}}_{0}(X_{k}%
)||_{p}+n^{1/p}\Big (\sum_{k=1}^{n}k^{1-2/p}(\log k)^{\gamma}\lambda
(k)\Big )^{1/2}\,.
\end{align*}
where $\gamma$ can be taken $\gamma=0$ for $2<p\leq3$ and $\gamma>p-3$ for
$p>3;$ $c=1$ for $2<p<4$ $\ $and $c=0$ for $p\geq4.$ The constant that is
implicitly involved in the notation $\ll$ depends on $p$ and $\gamma$ but it
does not depend on $n$.
\end{proposition}
\noindent\textbf{Proof of Proposition \ref{consdirect}. }The proof of this
proposition is basically a combination of Theorem \ref{directprop} and Lemma
\ref{mixing}. By the triangle inequality
\[
||{\mathbf{E}}_{0}(S_{k}^{2})||_{p/2}\leq\Vert{\mathbf{E}}_{0}(S_{k}%
^{2})-{\mathbf{E}}(S_{k}^{2})\Vert_{p/2}+{\mathbf{E}}_{0}(S_{k}^{2})\,.
\]
By (\ref{var}), for any $p>2$ and any $\delta>0$, we easily obtain
\[
\left( \sum_{k=1}^{n}\frac{\ 1}{k^{1+2\delta/p}}({\mathbf{E}}(S_{k}%
^{2}))^{\delta}\right) ^{1/(2\delta)}\ll n^{1/2-1/p}\Big (\sum_{j=0}%
^{n-1}|{\mathbf{E}}(X_{0}X_{j})|\Big )^{1/2}\,.
\]
Then, we use inequality (\ref{condterm}) and changing the order of summation
\[
\sum_{k=1}^{n}\frac{1}{k^{1+1/p}}||{\mathbf{E}}_{0}(S_{k})||_{p}\ll\sum
_{k=1}^{n}\frac{1}{k^{1/p}}||{\mathbf{E}}_{0}(X_{k})||_{p}\,.
\]
Now for the situation $0<\delta<1$, \ by H\"{o}lder's inequality,
\[
\left( \sum_{k=1}^{n}\frac{1}{k^{1+2\delta/p}}\Vert{\mathbf{E}}_{0}(S_{k}%
^{2})-{\mathbf{E}}(S_{k}^{2})\Vert_{p/2}^{\delta}\right) ^{p/(2\delta)}%
\ll\left( \sum_{k=1}^{n}\frac{(\log k)^{\gamma}\ }{k^{1+2/p}}\Vert
{\mathbf{E}}_{0}(S_{k}^{2})-{\mathbf{E}}(S_{k}^{2})\Vert_{p/2}\right)
^{p/2}\,.
\]
where $\gamma>1/\delta-1$. We continue the estimate by using (\ref{varsquare})
and get
\[
\sum_{k=1}^{n}\frac{(\log k)^{\gamma}}{k^{1+2/p}}\Vert{\mathbf{E}}_{0}%
(S_{k}^{2})-{\mathbf{E}}(S_{k}^{2})\Vert_{p/2}\ll\sum_{k=1}^{n}(\log
k)^{\gamma}k^{1-2/p}\lambda(k)\text{ }\,.
\]
Proposition \ref{consdirect} follows by using Theorem \ref{directprop}
combined with all the above estimates. $\diamond$
\medskip
We give now a consequence of Theorem \ref{directprop} that will be used in one
of our applications. The proof is omitted since it follows the spirit of the
proof Proposition \ref{consdirect}; namely, the use of Lemma \ref{mixing}
combined with the H\"{o}lder's inequality, and the fact that $||{\mathbf{E}%
}_{0}(S_{k})||_{p}\leq||{\mathbf{E}}_{0}(S_{k}^{2})-{\mathbf{E}}(S_{k}%
^{2})||_{p/2}^{1/2}+\big ({\mathbf{E}}(S_{k}^{2})\big)^{1/2}$.
\begin{proposition}
\label{consdirect2} Let $p>2$ be a real number and let $(X_{i})_{i\in
\mathbf{Z}}$ be a stationary sequence of real-valued random variables in
${\mathbf{L}}_{p}$ adapted to an increasing filtration $({\mathcal{F}}_{i})$.
Let $(\lambda(j))_{j \geq1}$ be defined by (\ref{notalambda}). For every
positive integer $n$, the following inequality holds: for any $\varepsilon
>0$,
\begin{align*}
{\mathbf{E}}\big (\max_{1\leq j\leq n}|S_{j}|^{p}\big ) & \ll n^{p/2}%
\Big (\text{ }\sum_{k=0}^{n-1}|{\mathbf{E}}(X_{0}X_{k})|\Big )^{p/2}%
+n{\mathbf{E}} ( |X_{1}|^{p})+n \sum_{k=1}^{n}k^{p-2+\varepsilon}\lambda
^{p/2}(k)\,.
\end{align*}
The constant that is implicitly involved in the notation $\ll$ depends on $p$
and $\varepsilon$ but it does not depend on $n$.
\end{proposition}
\section{Applications and examples}
\label{sectionappliexamples}
As we have seen Propositions \ref{maxinequality} and \ref{propmaxineproba}
give a direct approach to compare the moments of order $p$ of the maximum of
the partial sums to the corresponding ones of the partial sum. We start this
section by presenting two additional applications of these propositions to the
convergence of maximum of partial sums and to the maximal Bernstein inequality
for dependent structures. In the last three examples, we apply our results on
the Rosenthal type inequalities to different classes of processes.
\subsection{Convergence of the maximum of partial sums in ${\mathbf{L}}^{p}$.}
\begin{corollary}
\label{cortightlp} Let $p\geq2$ and let $(X_{i})_{i\in{\mathbf{Z}}}$ be a
strictly stationary sequence of centered real-valued random variables in
${\mathbf{L}}^{p}$ adapted to an increasing and stationary filtration
$({\mathcal{F}}_{i})_{i\in{\mathbf{Z}}}$. Assume that
\begin{equation}
\lim_{n\rightarrow\infty}n^{-1/p}\Vert S_{n}\Vert_{p}=0\,. \label{hypoappmart}%
\end{equation}
Assume in addition that
\begin{equation}
\sum_{n\geq1}\frac{\Vert{\mathbf{E}}(S_{n}|{\mathcal{F}}_{0})\Vert_{p}%
}{n^{1+1/p}}<\infty\,. \label{maxwoodLp}%
\end{equation}
Then
\begin{equation}
\lim_{n\rightarrow\infty}n^{-1/p}\Vert\max_{1\leq k\leq n}|S_{k}|\Vert
_{p}=0\,. \label{resmaxtightlp}%
\end{equation}
\end{corollary}
\begin{remark}
This corollary is particularly useful for studying the asymptotic behavior of
a partial sum via a martingale approximation. Assume there exists a strictly
stationary sequence $(d_{i})_{i\in{\mathbf{Z}}}$ of martingale differences
with respect to $({\mathcal{F}}_{i})_{i\in{\mathbf{Z}}}$ that are in
${\mathbf{L}}^{p}$, such that $\lim_{n\rightarrow\infty}n^{-1/p}\Vert
S_{n}-\sum_{i=1}^{n}d_{i}\Vert_{p}=0\,.$ Then, if condition (\ref{maxwoodLp})
holds for the sequence $(X_{i})_{i\in{\mathbf{Z}}}$, then by a construction in
Woodroofe and Zhao (2008) and by the uniqueness of the martingale
approximation, the sequence $(X_{i}-d_{i})_{i\in{\mathbf{Z}}}$ is still a
strictly stationary sequence and by our theorem $\lim_{n\rightarrow\infty
}n^{-1/p}\Vert\max_{1\leq k\leq n}|S_{k}-\sum_{i=1}^{k}d_{i}|\Vert_{p}=0\,.$
As a matter of fact, for $p=2$, our corollary leads to the functional form of
the central limit theorem for $\{n^{-1/2}S_{[nt]},t\in\lbrack0,1]\}$ (see also
Theorem 1.1 in Peligrad and Utev, 2005).
\end{remark}
\noindent\textbf{Proof of Corollary \ref{cortightlp}.}
Let $m$ be an integer and $k=k_{n,m}=[n/m]$ (where $[x]$ denotes the integer
part of $x$).
The initial step of the proof is to divide the variables in blocks of size $m$
and to make the sums in each block. Let
\[
X_{i,m}=\sum_{j=(i-1)m+1}^{im}X_{j}\,,\text{ }i\geq1.
\]
Notice first that
\[
\Vert\sup_{t\in\lbrack0,1]}\big |\sum_{j=1}^{[nt]}X_{j}-\sum_{i=1}%
^{[kt]}X_{i,m}\big |\Vert_{p}\leq\Vert\sup_{t\in\lbrack0,1]}\big |\sum
_{i=[kt]m+1}^{[nt]}X_{i}\big |\Vert_{p}\leq m\Vert\max_{1\leq i\leq n}%
X_{i}\Vert_{p}\,.
\]
Since for every $\varepsilon>0$,
\[
{\mathbf{E}}(\max_{1\leq i\leq n}|X_{i}|^{p})\leq\varepsilon^{p}+\sum
_{i=1}^{n}{\mathbf{E}}\big (|X_{i}|^{p}\mathbf{1}_{\{|X_{i}|>\varepsilon
\}}\big )\,,
\]
and since $\Vert X_{i}\Vert_{p}<\infty$ for all $i$, we derive that for any
fixed $m$, $\lim_{n\rightarrow\infty}m\Vert\max_{1\leq i\leq n}X_{i}\Vert
_{p}/n^{1/p}=0$. Hence to prove (\ref{resmaxtightlp}) it remains to show that
\begin{equation}
\lim_{m\rightarrow\infty}\limsup_{n\rightarrow\infty}{\mathbf{\ }}%
n^{-1/p}\Vert\sup_{t\in\lbrack0,1]}\big |\sum_{i=1}^{[kt]}X_{i,m}%
\big |\Vert_{p}=0\,. \label{butresmaxtightlp}%
\end{equation}
Applying Proposition \ref{maxinequality} to the variables $(X_{i,m})_{1\leq
i\leq k}$ which are adapted with respect to ${\mathcal{F}}_{im}$, and taking
into account the remark \ref{Rem1inegalite}, we get that
\[
\Vert\sup_{t\in\lbrack0,1]}\big |\sum_{i=1}^{[kt]}X_{i,m}\big |\Vert_{p}%
\ll\max_{1\leq j\leq k}\Vert\sum_{\ell=1}^{jm}X_{\ell}\Vert_{p}+k^{1/p}%
\sum_{j=1}^{k}\frac{||{\mathbf{E}}(S_{jm}|{\mathcal{F}}_{0})||_{p}}{j^{1+1/p}%
}\,,
\]
where for the last term we used the fact that for any positive integer $u$,
$\Vert{\mathbf{E}}\big (\sum_{j=um2^{\ell}+1}^{(u+1)m2^{\ell}}X_{j}%
|{\mathcal{F}}_{um2^{\ell}}\big )\Vert_{p}=\Vert{\mathbf{E}}\big (\sum
_{j=1}^{m2^{\ell}}X_{j}|{\mathcal{F}}_{0}\big )\Vert_{p}$. Condition
(\ref{hypoappmart}) implies that
\[
\max_{1\leq j\leq k}\Vert\sum_{\ell=1}^{jm}X_{\ell}\Vert_{p}=o((km)^{1/p}%
)=o(n^{1/p})\,.
\]
Now, by subadditivity of the sequence $\big (||{\mathbf{E}}(S_{n}%
|{\mathcal{F}}_{0})||_{p}\big)_{n\geq1}$ and \ applying Lemma \ref{lmasubadd}
we have
\begin{equation}
n^{-1/p}k^{1/p}\sum_{j=1}^{k}\frac{||{\mathbf{E}}(S_{jm}|{\mathcal{F}}%
_{0})||_{p}}{j^{1+1/p}}\ll\sum_{\ell=1}^{m}\frac{||{\mathbf{E}}(S_{\ell
}|{\mathcal{F}}_{0})||_{p}}{(\ell+m)^{1+1/p}}+\sum_{j\geq m}\frac
{||{\mathbf{E}}(S_{j}|{\mathcal{F}}_{0})||_{p}}{j^{1+1/p}}\,. \label{fatou}%
\end{equation}
Hence, under (\ref{maxwoodLp}) and using Fatou's lemma for the first term in
the right-hand side of (\ref{fatou}), we get that
\[
\lim_{m\rightarrow\infty}\limsup_{n\rightarrow\infty}{\mathbf{\ }}%
n^{-1/p}k^{1/p}\sum_{j=1}^{k}\frac{||{\mathbf{E}}(S_{jm}|{\mathcal{F}}%
_{0})||_{p}}{j^{1+1/p}}=0\,,
\]
which ends the proof of the corollary. $\diamond$
\subsection{\medskip Maximal exponential inequalities for strongly mixing.}
Let us first recall the definition of strongly mixing sequences, introduced by
Rosenblatt (1956): For any two $\sigma$ algebras $\mathcal{A}$ and
$\mathcal{B}$, we define the $\alpha$-mixing coefficient by
\[
\alpha(\mathcal{A},\mathcal{B})=\sup_{A\in\mathcal{A},B\in\mathcal{B}%
}|\mathbf{P}(A\cap B)-\mathbf{P}(A)\mathbf{P}(B)|\text{ }.
\]
Let $(X_{k},k\geq1)$ be a sequence of real-valued random variables defined on
$\left( \Omega,\mathcal{A},\mathbf{P}\right) $. This sequence will be called
strongly mixing if
\begin{equation}
\alpha(n):=\sup_{k\geq1}\alpha\left( \mathcal{F}_{k},\mathcal{G}%
_{k+n}\right) \rightarrow0\mbox{ as }n\rightarrow\infty\,, \label{defalpha}%
\end{equation}
where $\mathcal{F}_{j}:=\sigma(X_{i},i\leq j)$ and $\mathcal{G}_{j}%
:=\sigma(X_{i},i\geq j)$ for $j\geq1$.
In 2009, Merlev\`{e}de, Peligrad and Rio have proved (see their Theorem 2)
that for a strongly mixing sequence of centered random variables satisfying
$\sup_{i\geq1}\Vert X_{i}\Vert_{\infty}\leq M$ and for a certain $c>0$
\begin{equation}
\alpha(n)\leq\exp(-cn)\,, \label{alphacond}%
\end{equation}
the following Bernstein-type inequality is valid: there is a constant $C$
depending only on $c$ such that for all $n\geq2$,
\begin{equation}
\mathbf{P}(|S_{n}|\geq x)\leq\exp\big (-{\frac{Cx^{2}}{v^{2}n+M^{2}+xM(\log
n)^{2}}}\big )\,, \label{bernSn}%
\end{equation}
where
\begin{equation}
v^{2}=\sup_{i>0}\Bigl (\mathrm{Var}(X_{i})+2\sum_{j>i}|\mathrm{Cov}%
(X_{i},X_{j})|\Bigr)\,. \label{defv2}%
\end{equation}
Proving the maximal version of the inequality (\ref{bernSn}) cannot be handled
directly neither using Theorem 2.2 in M\'{o}ricz, Serfling and Stout (1982),
nor using Theorem 1 in Kevei and Mason (2010) since the left-hand side of
(\ref{bernSn}) does not satisfy the assumptions of both these papers. However,
an application of Proposition \ref{propmaxineproba} \ leads to the maximal
version of Theorem 2 in Merlev\`{e}de, Peligrad and Rio (2009).
\begin{corollary}
\label{cormaxbernstein} Let $(X_{j})_{j\geq1}$ be a sequence of centered
real-valued random variables. Suppose that there exists a positive $M$ such
that $\sup_{i\geq1}\Vert X_{i}\Vert_{\infty}\leq M$ and that the strongly
mixing coefficients $(\alpha(n))_{n\geq1}$ of the sequence satisfy
(\ref{alphacond}). Then there exists constants $C=C(c)$ and $K=K(M,c)$ such
that for all integer $n\geq2$ and all real $x>K\log n$,
\begin{equation}
\mathbf{P}(\max_{1\leq k\leq n}|S_{k}|\geq x)\leq\exp\big (-{\frac{Cx^{2}%
}{v^{2}n+M^{2}+xM(\log n)^{2}}}\big )\,. \label{bern2}%
\end{equation}
\end{corollary}
\noindent\textbf{Proof of Corollary \ref{cormaxbernstein}.} We first apply
Inequality (\ref{inegalitemax2proba}) of Proposition \ref{propmaxineproba}
with $p=2$ and $\varphi(x)=e^{t|x|}$ where $t$ is a positive integer.
According to Theorem 2 in Merlev\`{e}de, Peligrad and Rio (2009), there exist
positive constants $C_{1}$ and $C_{2}$ depending only on $c$ such that for all
$n\geq2$ and any positive $t$ such that $t<\frac{1}{C_{1}M(\log n)^{2}}$, the
following inequality holds:
\[
\log\mathbf{E}\big (\exp(tS_{n})\big )\leq\frac{C_{2}t^{2}(nv^{2}+M^{2}%
)}{1-C_{1}tM(\log n)^{2}}\, .
\]
Then an optimization on $t$ gives that there is a constant $C_{3}$ depending
only on $c$ such that for all $n\geq2$ and any positive real $x$,
\begin{align}
\mathbf{P}(\max_{1 \leq k \leq n} & |S_{k}|\geq4x)\leq\exp\big (-{\frac
{C_{3}x^{2}}{v^{2}n+M^{2}+xM(\log n)^{2}}}\big )\nonumber\\
& +4x^{-2}\Big(\sum_{l=0}^{r-1}\Big (\sum_{k=1}^{2^{r-l}-1}||{\mathbf{E}%
}(S_{v+(k+1)2^{l}}-S_{v+k2^{l}}|{\mathcal{F}}_{k2^{l}})||_{2}^{2}%
\Big )^{1/2}\Big )^{2}\, , \label{bern2*}%
\end{align}
where $r$ is the positive integer satisfying $2^{r-1} < n \leq2^{r}$ and $v =
[x/M]$.
It remains to bound up the second term in the right-hand side of the above
inequality. Notice first that for any centered variable $Z$ such that $\Vert
Z\Vert_{\infty}\leq B$,
\[
\Vert{\mathbf{E}}(Z|{\mathcal{F}})\Vert_{2}^{2}\leq B\Vert{\mathbf{E}%
}(Z|{\mathcal{F}})\Vert_{1}=B\mathrm{Cov}(\mathrm{sign}({\mathbf{E}%
}(Z|{\mathcal{F}})),Z)\,.
\]
Hence by the Ibragimov's covariance inequality (see Theorem 1.11 in Bradley,
2007),
\[
\Vert{\mathbf{E}}(Z|{\mathcal{F}})\Vert_{2}^{2}\leq4B^{2}\alpha({\mathcal{F}%
},\sigma(Z))\,.
\]
Therefore, applying this last estimate with $Z=S_{v+(k+1)2^{l}}-S_{v+k2^{l}}$
and ${\mathcal{F}}={\mathcal{F}}_{k2^{l}}$, we get that
\[
||{\mathbf{E}}(S_{v+(k+1)2^{l}}-S_{v+k2^{l}}|{\mathcal{F}}_{k2^{l}})||_{2}%
^{2}\leq4M^{2}2^{2l}\alpha(v)\,.
\]
implying that
\[
\Big(\sum_{l=0}^{r-1}\Big (\sum_{k=1}^{2^{r-l}-1}||{\mathbf{E}}%
(S_{v+(k+1)2^{l}}-S_{v+k2^{l}}|{\mathcal{F}}_{k2^{l}})||_{2}^{2}%
\Big )^{1/2}\Big )^{2}\leq4M^{2}2^{2r}(\sqrt{2}+1)^{2}\alpha(v)\,.
\]
Since $2^{2r}\leq4n^{2}$ and $[x/M]\geq x/(2M)$, for $x\geq2M$, by using
(\ref{alphacond}), we get that, for any $x\geq2M$,
\begin{equation}
\Big(\sum_{l=0}^{r-1}\Big (\sum_{k=1}^{2^{r-l}-1}||{\mathbf{E}}%
(S_{v+(k+1)2^{l}}-S_{v+k2^{l}}|{\mathcal{F}}_{k2^{l}})||_{2}^{2}%
\Big )^{1/2}\Big )^{2}\leq3\times2^{5}M^{2}n^{2}\exp(-cx/(2M))\,.
\label{bern2*b2}%
\end{equation}
Starting from (\ref{bern2*}) and using (\ref{bern2*b2}), we then derive that
for any $x\geq2M\max(1,4c^{-1}\log n)$,
\[
{\mathbf{P}}(S_{2^{r}}^{\ast}\geq4x)\leq\exp\Big (-{\frac{C_{3}x^{2}}%
{v^{2}n+M^{2}+xM(\log n)^{2}}}\Big )+96\exp\Big (-\frac{xc}{4M}\Big )\,,
\]
proving the inequality (\ref{bern2}). $\diamond$
\subsection{Application to Arch models.}
Theorem \ref{directprop} applies to the case where $(X_{i})_{i\in{\mathbf{Z}}%
}$ has an ARCH($\infty$) structure as described by Giraitis \textit{et al.}
(2000), that is
\begin{equation}
X_{n}=\sigma_{n}\eta_{n},\ \text{with}\ \sigma_{n}^{2}=c+\sum_{j=1}^{\infty
}c_{j}X_{n-j}^{2}\,, \label{defARCH}%
\end{equation}
where $(\eta_{n})_{n\in\mathbf{Z}}$ is a sequence of i.i.d. centered random
variables such that ${\mathbf{E}}(\eta_{0}^{2})=1$, and where $c\geq0$,
$c_{j}\geq0$, and $\sum_{j\geq1}c_{j}<1$. Notice that $(X_{i})_{i
\in{\mathbf{Z}}}$ is a stationary sequence of martingale differences adapted
to the filtration $({\mathcal{F}}_{i})$ where ${\mathcal{F}}_{i}= \sigma(
\eta_{k} , k \leq i)$.
Let $p>2$ and assume that $\Vert\eta_{0}\Vert_{p}<\infty$. Notice first that
\begin{equation}
\Vert{\mathbf{E}}(X_{j}^{2}|{\mathcal{F}}_{0})-{\mathbf{E}}(X_{0}^{2}%
)\Vert_{p/2}=\Vert{\mathbf{E}}(\sigma_{j}^{2}|{\mathcal{F}}_{0})-{\mathbf{E}%
}(\sigma_{j}^{2})\Vert_{p/2}\,. \label{ARCHfirstequality}%
\end{equation}
In addition, since ${\mathbb{E}}(\eta_{0}^{2})=1$ and $\sum_{j\geq1}c_{j}<1$,
the unique stationary solution to (\ref{defARCH}) is given by Giraitis
\textit{et al.} (2000):
\begin{equation}
\sigma_{n}^{2}=c+c\sum_{\ell=1}^{\infty}\sum_{j_{1},\dots,j_{\ell}=1}^{\infty
}c_{j_{1}}\dots c_{j_{\ell}}\eta_{n-j_{1}}^{2}\dots\eta_{n-(j_{1}%
+\dots+j_{\ell})}^{2}\,. \label{solARCH}%
\end{equation}
Starting from (\ref{ARCHfirstequality}) and using (\ref{solARCH}), one can
prove that
\[
\Vert{\mathbf{E}}(X_{j}^{2}|{\mathcal{F}}_{0})-{\mathbf{E}}(X_{0}^{2}%
)\Vert_{p/2}\leq2c\Vert\eta_{0}\Vert_{p}^{2}\sum_{\ell=1}^{\infty}\ell
\kappa^{\ell-1}\sum_{i=[j/\ell]}^{\infty}c_{i}\,,
\]
where $\kappa=\Vert\eta_{0}\Vert_{p}^{2}\sum_{j\geq1}c_{j}$ (see Section 6.6
in Dedecker and Merlev\`{e}de (2010) for more detailed computations).
Consequently if
\begin{equation}
\Vert\eta_{0}\Vert_{p}^{2}\sum_{j\geq1}c_{j}<1\text{ and }\sum_{j\geq n}%
c_{j}=O(n^{-b})\ \text{for}\ b>1-2/p\,, \label{condARCH}%
\end{equation}
we get that for any $\delta\in]0,1]$,
\[
\sum_{k=1}^{\infty}\frac{1}{k^{1+2\delta/p}}||{\mathbf{E}}_{0}(S_{k}%
^{2})-{\mathbf{E}}(S_{k}^{2})||_{p/2}^{\delta}<\infty\,.
\]
Applying Theorem \ref{directprop} for the martingale case, we then get the
following corollary:
\begin{corollary}
\label{corARCH} Let $(X_{i})_{i \in{\mathbf{Z}}}$ be defined by (\ref{defARCH}%
) and $S_{n} = \sum_{i=1}^{n} X_{i}$. Let $p >2$ and assume that
(\ref{condARCH}) is satisfied. Then for any integer $n$,
\[
{\mathbf{E}}(\max_{1 \leq k \leq n}|S_{k}|^{p}) \ll\big (n{\mathbf{E}}%
(X_{0}^{2})\big )^{p/2}+ n \, (1+{\mathbf{E}}(|X_{0}|^{p}))\, .
\]
\end{corollary}
\subsection{\noindent\ Application to functions of linear processes.}
Let
\begin{equation}
X_{k}=h\Big(\sum_{i\in{\mathbf{Z}}}a_{i}\varepsilon_{k-i}\Big)-{\mathbf{E}%
}\Big(h\Big(\sum_{i\in{\mathbf{Z}}}a_{i}\varepsilon_{k-i}\Big)\Big)\,,
\label{def2suite}%
\end{equation}
where $(\varepsilon_{i})_{i\in{\mathbf{Z}}}$ is a sequence of i.i.d. random
variables. Denote by $w_{h}(.,M)$ the modulus of continuity of the function
$h$ on the interval $[-M,M]$, that is
\[
w_{h}(t,M)=\sup\{|h(x)-h(y)|,|x-y|\leq t,|x|\leq M,|y|\leq M\}\,.
\]
Applying Theorem \ref{general}, the following result holds:
\begin{corollary}
\label{Thlin} Let $(a_{i})_{i\in{{\mathbf{Z}}}}$ be a sequence of real numbers
in $\ell^{2}$ and $(\varepsilon_{i})_{i\in\mathbf{Z}}$ be a sequence of i.i.d.
random variables in ${\mathbf{L}}^{2}$. Let $X_{k}$ be defined as in
(\ref{def2suite}). Assume that $h$ is $\gamma$-H\"{o}lder on any compact set,
with $w_{h}(t,M)\leq Ct^{\gamma}M^{\alpha}$, for some $C>0$, $\gamma\in]0,1]$
and $\alpha\geq0$. Let $p\geq4$ be an even integer. Assume that for
$\lambda>p/2-2$,
\begin{equation}
{\mathbf{E}}(|\varepsilon_{0}|^{2\vee(\alpha+\gamma)p})<\infty\quad
\text{and}\quad\sum_{i\geq1}i^{1-2/p}(\log i)^{\lambda}\Big(\sum_{j\geq
i}a_{j}^{2}\Big)^{\gamma/2}<\infty, \label{gammaholder}%
\end{equation}
and that
\begin{equation}
\text{there exists $\beta>2/p$ such that }n^{- \beta}{\mathbf{E}}(S_{n}^{2})
\text{ is increasing.} \label{condvar}%
\end{equation}
Then for any integer $n$,
\[
{\mathbf{E}}(\max_{1 \leq k \leq n }|S_{k}|^{p})\ll n \, (1+{\mathbf{E}%
}(|X_{0}|^{p})) + \big ( {\mathbf{E}}(S_{n}^{2}) \big )^{p/2} \, .
\]
\end{corollary}
\begin{remark}
\label{rmklinear} The proof of this result is based on Theorem \ref{general}.
Our proof reveals that an application of Theorem \ref{directprop} would
involve a more restrictive condition on $\lambda$, namely $\lambda>p-3$.
\end{remark}
\noindent\textbf{Proof of Corollary \ref{Thlin}.}
Applying Theorem \ref{general}, for any positive integer $n$, we get that
\begin{align*}
{\mathbf{E}} & \big (\max_{1\leq j\leq n}\ |S_{j}|^{p}\big )\ll
n\Big (\sum_{k=1}^{n}\frac{1}{k^{1+4/p(p-2)}}({\mathbf{E}}(S_{k}%
^{2}))^{2/(p-2)}\Big )^{p(p-2)/4}+n\Big (\sum_{k=1}^{n}||{\mathbf{E}}%
_{0}(X_{k})||_{p}\Big )^{p}+n\Vert X_{0}\Vert_{p}^{p}\\
& +n\Big (\sum_{k=1}^{n}\frac{1}{k^{1+4/p(p-2)}}||{\mathbf{E}}_{0}(S_{k}%
^{2})-{\mathbf{E}}(S_{k}^{2})\text{ }||_{p/2}^{2/(p-2)})\Big )^{p(p-2)/4}\,.
\end{align*}
Using the condition (\ref{condvar}), it follows that
\[
n\Big (\sum_{k=1}^{n}\frac{1}{k^{1+4/p(p-2)}}({\mathbf{E}}(S_{k}%
^{2}))^{2/(p-2)}\Big )^{p(p-2)/4}\ll({\mathbf{E}}(S_{n}^{2}))^{p/2}\,.
\]
Therefore the theorem follows if we prove that (\ref{gammaholder}) implies
that
\begin{equation}
\sum_{k\geq1}||{\mathbf{E}}_{0}(X_{k})||_{p}<\infty\text{ and }\sum_{k\geq
1}\frac{1}{k^{1+4/p(p-2)}}||{\mathbf{E}}_{0}(S_{k}^{2})-{\mathbf{E}}(S_{k}%
^{2})\text{ }||_{p/2}^{2/(p-2)}<\infty\,. \label{cons1gammaholder*}%
\end{equation}
Notice now that by H\"{o}lder's inequality and (\ref{varsquare}) to prove the
second part of (\ref{cons1gammaholder*}), it suffices to prove that for
$\lambda>p/2-2$,
\[
\sum_{k\geq1}\frac{(\log k)^{\lambda}}{k^{1+2/p}}\sum_{i=1}^{k}\sum
_{j=0}^{k-i}\Vert{\mathbf{E}}_{0}(X_{i}X_{j+i})-{\mathbf{E}}(X_{i}%
X_{j+i})\Vert_{p/2}<\infty\,.
\]
Corollary \ref{Thlin} is then a consequence of the following proposition
applied with $b=\lambda$ and $c=2/p$ (see the proof of Theorem 4.2 in
Dedecker, Merlev\`{e}de and Rio 2009, page 988).
\begin{proposition}
\label{modulo} Let $(a_{i})_{i \in{\mathbf{Z}}}$, $(\varepsilon_{i})_{i
\in{\mathbf{Z}}}$ and $(X_{i})_{i \in{\mathbf{Z}}}$ be as in Corollary
\ref{Thlin}. Let $(\varepsilon^{\prime}_{i})_{i \in{\mathbf{Z}}}$ be an
independent copy of $(\varepsilon_{i})_{i \in{\mathbf{Z}}}$. Let $V_{0}=
\sum_{i \geq0} a_{i} \varepsilon_{-i}$ and
\[
M_{1,i}= |V_{0}| \vee\Big|\sum_{0 \leq j<i}a_{j} \varepsilon_{-j}+\sum_{j \geq
i \geq0} a_{j} \varepsilon^{\prime}_{-j}\Big|.
\]
Let $p \geq2$. If for $b \geq0$ and $0 < c \leq1$,
\begin{equation}
\label{mod1}\sum_{i \geq1} i^{1-c}(\log i)^{b} \Big \| w_{h}\Big(\Big|\sum_{j
\geq i} a_{j} \varepsilon_{-j}\Big |, M_{1,i}\Big) \Big \|_{p} < \infty\, ,
\end{equation}
then $\sum_{k\geq1}||{\mathbf{E}}_{0}(X_{k})||_{p} < \infty$ and
\[
\sum_{ k \geq1} \frac{(\log k)^{b}}{k^{1+c}}\sum_{i=1}^{k}\sum_{j=0}%
^{k-i}\Vert{\mathbf{E}}_{0}(X_{i}X_{j+i}) -{\mathbf{E}}(X_{i}X_{j+i})
\Vert_{p/2} < \infty\, .
\]
\end{proposition}
The proof of the above proposition is direct following the lines of the proof
of Proposition 4.2 Dedecker, Merlev\`ede and Rio (2009). $\diamond$
\subsection{Application to a stationary reversible Markov chain.}
First we want to mention that all our results can be formulated in the Markov
chain setting. We assume that $(\zeta_{n})_{n\in\mathbb{Z}}$ denotes a
stationary Markov chain defined on a probability space $(\Omega,\mathcal{A}%
,{\mathbf{P}})$ with values in a measurable space $(E,\mathcal{E})$. The
marginal distribution and the transition kernel are denoted by $\pi
(A)={\mathbf{P}}(\zeta_{0}\in A)$ and $Q(\zeta_{0},A)={\mathbf{P}}(\zeta
_{1}\in A|\,\zeta_{0})$. In addition $Q$ denotes the operator {acting via
$(Qf)(\zeta)=\int_{E}f(s)Q(\zeta,ds).$ Next, let $f$ be a function on $E$ such
that $\int_{E}|f|^{p}d\pi<\infty$ and $\int_{E}fd\pi=0.$ Denote by
$\mathcal{F}_{k}$ the $\sigma$--field generated by }$\zeta${$_{i}$ with $i\leq
k,$ $X_{i}=f(\zeta_{i})$, and $S_{n}(f)=\sum\limits_{i=1}^{n}X_{i}$. Notice
that any stationary sequence $(Y_{k})_{k\in\mathbb{Z}}$ can be viewed as a
function of a Markov process $\zeta_{k}=(Y_{i};i\leq k),$ for the function
$g(\zeta_{k})=Y_{k}$. }
\medskip
The Markov chain is called reversible if $Q=Q^{\ast},$ where $Q^{\ast}$ is the
adjoint operator of $Q$. In this setting, an application of Theorem
\ref{stateven} gives the following estimate:
\begin{corollary}
\label{Rev}Let $(\zeta_{n})$ be a reversible Markov chain. For any even
integer $p\geq4$ and any positive integer $n$,
\begin{gather*}
{\mathbf{E}}\big (\max_{1\leq k\leq n}|S_{k}(f)|^{p}\big )\ll n\mathbf{E}%
(|f(\zeta_{1})|^{p})+n\Big (\sum_{k=1}^{n}\frac{1}{k^{1+1/p}}\Vert{\mathbf{E}%
}_{0}(S_{k}(f))\Vert_{p}\Big )^{p}\\
+n\Big (\sum_{k=1}^{n}\frac{1}{k^{1+2/p}}\Vert{\mathbf{E}}_{0}(S_{k}%
^{2}(f))-{\mathbf{E}}(S_{k}^{2}(f))\Vert_{p/2}\Big )^{p/2}\,+n\Big (\sum
_{k=1}^{n}\frac{1}{k^{1+2/p}}{\mathbf{E}}(S_{k}^{2}(f))\Big )^{p/2}.
\end{gather*}
\end{corollary}
Moreover using Theorem \ref{pinteger} we obtain:
\begin{corollary}
\label{Revgen} Let $(\zeta_{n})$ be a reversible Markov chain. For any real
number $p> 4$ and any positive integer $n$,
\begin{gather*}
{\mathbf{E}}\big (\max_{1\leq k\leq n}|S_{k}(f)|^{p}\big )\ll n\mathbf{E}%
(|f(\zeta_{1})|^{p})\\
+n\Big (\sum_{k=1}^{n}\frac{1}{k^{1+1/p}}\Vert{\mathbf{E}}_{0}(S_{k}%
^{2}(f))-{\mathbf{E}}(S_{k}^{2}(f))\Vert_{p/2}^{1/2}\Big )^{p}\,+n\Big (\sum
_{k=1}^{n}\frac{1}{k^{1+1/p}}{\mathbf{||}}S_{k}(f)||_{2}\Big )^{p} \, .
\end{gather*}
\end{corollary}
This corollary is also valid for any real $2 < p \leq4$. For this range
however, according to the comment \ref{commentgreaterdelta}, Theorem
\ref{directprop} (respectively Corollary \ref{Rev} ) gives a better bound for
$p\in]2,4[$ (respectively for $p=4$).
\medskip
For a particular example let $E=[-1,1]$ and let $\upsilon$ be a symmetric
atomless law on $E$. The transition probabilities are defined by
\[
Q(x,A)=(1-|x|)\delta_{x}(A)+|x|\upsilon(A)\,,
\]
where $\delta_{x}$ denotes the Dirac measure. Assume that $\theta=\int
_{E}|x|^{-1}\upsilon(dx)<\infty$. Then there is a unique invariant measure
\[
\pi(dx)=\theta^{-1}|x|^{-1}\upsilon\,(dx)
\]
and the stationary Markov chain $(\zeta_{i})_{i}$ is reversible and positively recurrent.
Assume the following assumption on the measure $\upsilon$: there exists a
positive constant $c$ such that for any $x\in\lbrack0,1]$,
\begin{equation}
\frac{d\upsilon}{dx}(x)\leq cx^{p/2-1}(\log(1+1/x))^{-\lambda}\text{ for some
$\lambda>0$.} \label{hyp2upsilon}%
\end{equation}
As an application of Corollary \ref{Revgen} we shall establish:
\begin{corollary}
\label{cor2MarkovChain} Let $p>2$ be a real number and let $f(-x)=-f(x)$ for
any $x\in E$. Assume that $|f(x)|<C|x|^{1/2}$ for any $x$ in $E$ and a
positive constant $C$. Assume in addition that (\ref{hyp2upsilon}) is
satisfied for $\lambda> p$. Then for any integer $n$,
\begin{equation}
{\mathbf{E}}\big (\max_{1\leq k\leq n}|S_{k}(f)|^{p}\big )\ll n+n^{p/2}%
\Big (\int_{0}^{1}f^{2}(x)x^{-2}\upsilon(dx)\Big )^{p/2}\,. \label{revineq}%
\end{equation}
\end{corollary}
Notice that this example of reversible Markov chain has been considered by Rio
(2009, Section 4) under a slightly more stringent condition on the measure
than (\ref{hyp2upsilon}). Corollary \ref{cor2MarkovChain} then extends
Proposition 4.1 (b) in Rio (2009) to all real $p>2$.
\medskip
\noindent\textbf{Proof of Corollary \ref{cor2MarkovChain}.}
To get this result we shall apply Corollary \ref{Revgen} along with Lemma
\ref{mixing}. We start by noticing that $f$ being an odd function we have
\begin{equation}
{\mathbf{E}}(f(\zeta_{k})|\zeta_{0})=(1-|\zeta_{0}|)^{k}f(\zeta_{0})\text{
a.s.} \label{operator}%
\end{equation}
Therefore, for any $j \geq0$,
\[
{\mathbf{E}}(X_{0}X_{j})={\mathbf{E}}(f(\zeta_{0}){\mathbf{E}}(f(\zeta
_{j})|\zeta_{0}))=\theta^{-1}\int_{E}f^{2}(x)(1-|x|)^{j}|x|^{-1}%
\upsilon(dx)\,.
\]
Then
\begin{align}
{\mathbf{E}}(S_{k}^{2}(f)) & \leq2k\theta^{-1}\Big (\int_{0}^{1}%
f^{2}(x)x^{-1}\upsilon(dx)+2\sum_{j=1}^{k-1}\int_{0}^{1}f^{2}(x)(1-x)^{j}%
x^{-1}\upsilon(dx)\Big )\label{estimatevar}\\
& \leq2k\theta^{-1}\Big (\int_{0}^{1}f^{2}(x)x^{-1}\upsilon(dx)+2\int_{0}%
^{1}f^{2}(x)x^{-2}\upsilon(dx)\Big )\,.\nonumber
\end{align}
We estimate next the quantity $\Vert{\mathbf{E}}_{0}(S_{n}^{2}(f))-{\mathbf{E}%
}(S_{n}^{2}(f))\Vert_{p/2}$. Notice first that
\[
\Vert{\mathbf{E}}_{0}(S_{n}^{2}(f))-{\mathbf{E}}(S_{n}^{2}(f))\Vert_{p/2}
\leq\sum_{k=1}^{n} \Big ( \int_{E} \Big | (\delta_{x} Q^{k} - \pi)
\big ( f^{2} + 2 f \sum_{k=1}^{n-k} Q^{k} f \big ) \Big |^{p/2}\pi(dx)
\Big )^{2/p}\, ,
\]
by using the fact that for any positive $k$, $\pi Q^{k} = \pi$ (see also the
inequality (4.12) in Rio (2009)). Now, by using the relation (\ref{operator}),
one can prove that for any $x \in E$,
\[
\Big ( f^{2} + 2 f \sum_{k=1}^{n} Q^{k} f\Big ) (x) = f^{2}(x) \big ( 1 + 2
(1-(1-|x|)^{n}) ( |x|^{-1}-1) \big )\, ,
\]
(see the computations in Rio (2009) leading to his relation (4.13)). Then,
since $|f(x)|\leq C|x|^{1/2}$, it follows that
\[
\label{marjoration1varSn}\sup_{x \in E} \big | f^{2}(x) + 2 f(x) \sum
_{k=1}^{n} Q^{k} f(x) \big | \leq2 C^{2} \, .
\]
Therefore, for any $p >2$,
\begin{align}
\label{maje0sn2}\Vert{\mathbf{E}}_{0}(S_{n}^{2}(f)) & -{\mathbf{E}}%
(S_{n}^{2}(f))\Vert_{p/2} \leq4 C^{2} \sum_{k=1}^{n}\Big (\int_{ E}\Vert
Q^{k}(x,\cdot)-\pi(\cdot)\Vert\pi(dx)\Big )^{2/p}\, ,
\end{align}
where $\Vert\mu(\cdot)\Vert$ denotes the total variation of the signed measure
$\mu$ (see also the inequality (4.15) in Rio (2009)). We estimate next the
coefficients of absolute regularity $\beta_{n}$ as defined in
(\ref{boundbetacor2}). Let $a=p/2-1$. We shall prove now that under
(\ref{hyp2upsilon}), there exists a positive constant $K$ such that
\begin{equation}
2\beta_{n}:=\int_{E}\Vert Q^{n}(x,\cdot)-\pi(\cdot)\Vert\pi(dx)\leq
Kn^{-a}(\log n)^{-\lambda}\, . \label{boundbetacor2}%
\end{equation}
Notice first that by Lemma 2, page 75, in Doukhan, Massart and Rio (1994), we
have that
\begin{equation}
\beta_{n}\leq3\int_{E}(1-|x|)^{[n/2]}\pi(dx)\, . \label{relbetatau}%
\end{equation}
Let $k\geq2$ be an integer. Clearly, for any $\alpha\in]0,1[$,
\begin{equation}
\int_{0}^{1}(1-x)^{k}\pi(dx)\leq c\int_{0}^{k^{-\alpha}}(1-x)^{k}x^{a-1}%
(\log(1+1/x))^{-\lambda}dx+c\int_{k^{-\alpha}}^{1}(1-x)^{k}x^{a-1}%
(\log(1+1/x))^{-\lambda}dx\,. \label{calculmomentT}%
\end{equation}
Notice now that
\[
\int_{0}^{k^{-\alpha}}(1-x)^{k}x^{a-1}(\log(1+1/x))^{-\lambda}dx\leq
(\alpha\log k)^{-\lambda}\int_{0}^{1}(1-x)^{k}x^{a-1}dx\,.
\]
Hence, by the properties of the Beta and Gamma functions,
\begin{equation}
\lim_{k\rightarrow\infty}k^{a}(\log k)^{\lambda}\int_{0}^{k^{-\alpha}%
}(1-x)^{k}x^{a-1}(\log(1+1/x))^{-\lambda}dx\leq\alpha^{-\lambda}a\Gamma(a)\,.
\label{p1boundbeta}%
\end{equation}
On the other hand, we have that
\[
\int_{k^{-\alpha}}^{1}(1-x)^{k}x^{a-1}(\log(1+1/x))^{-\lambda}dx\leq
(\log2)^{-\lambda}(1-k^{-\alpha})^{k}\int_{0}^{1}x^{a-1}dx\,,
\]
and then, since $\alpha<1$, we easily obtain
\begin{equation}
\lim_{k\rightarrow\infty}k^{a}(\log k)^{\lambda}\int_{k^{-\alpha}}%
^{1}(1-x)^{k}x^{a-1}(\log(1+1/x))^{-\lambda}dx=0\,. \label{p2boundbeta}%
\end{equation}
Starting from (\ref{relbetatau}) and taking into account (\ref{calculmomentT}%
), (\ref{p1boundbeta}) and (\ref{p2boundbeta}), (\ref{boundbetacor2}) follows.
Then by using the inequality (\ref{maje0sn2}) combined with
(\ref{boundbetacor2}), we derive that
\[
\Vert{\mathbf{E}}_{0}(S_{n}^{2}(f))-{\mathbf{E}}(S_{n}^{2}(f))\Vert_{p/2}\ll
n^{2/p}(\log n)^{-2\lambda/p}\,.
\]
Therefore, for any $\delta\in]0,1]$,
\begin{equation}
\sum_{k=1}^{n}\frac{1}{k^{1+1/p}}\Vert{\mathbf{E}}_{0}(S_{k}^{2}%
(f))-{\mathbf{E}}(S_{k}^{2}(f))\Vert_{p/2}^{1/2}=O(1)\text{ for $\lambda>p$%
}\,. \label{convergencelog}%
\end{equation}
Considering the estimates (\ref{estimatevar}) and (\ref{convergencelog}),
Corollary \ref{cor2MarkovChain} follows from an application of Corollary
\ref{Revgen} (taking also into account the comment after its statement for
$2<p\leq4$). $\diamond$
\section{Application to density estimation}
\label{densitysection}
In this section, we estimate the ${\mathbf{L}}^{p}$-integrated risk for
$p\geq4$, for the kernel estimator of the unknown marginal density $f$ of a
stationary sequence $(Y_{i})_{i\geq0}$.
Applying our theorem \ref{directprop}, we shall show that if the coefficients
of dependence $((\beta_{2,Y}(k))_{k\geq1}$ (see the definition \ref{beta}
below) of the sequence $(Y_{i})_{i\in{\mathbf{Z}}}$ satisfy $\beta
_{2,Y}(k)=O(n^{-a})$ for $a>p-1$, then the bound of the ${\mathbf{L}}^{p}%
$-norm of the random term of the risk is of the same order of magnitude as the
one obtained in Bretagnolle and Huber (1979) in the independence setting (see
their corollary 2), provided that the density is bounded and the kernel $K$
satisfies the assumption \textbf{A$_{p}$} below. \medskip
\noindent\textit{\textbf{Assumption A}}$_{p}.$ \textit{ }$K$ \textit{is a BV
(bounded variation) function such that }%
\[
\int_{\mathbf{R}}|K(u)|du<\infty\text{ and }\int_{\mathbf{R}}|K(u)|^{p}%
du<\infty\,.
\]
\begin{definition}
\label{beta} Let $(Y_{i})_{i\in{\mathbf{Z}}}$ be a stationary sequence of real
valued random variables, and let ${\mathcal{F}}_{0}=\sigma(Y_{i},i\leq0)$. For
any positive $i$ and $j$, define the random variables
\begin{align*}
b({\mathcal{F}}_{0},i,j) & =\ \sup_{(s,t)\in{\mathbf{R}}^{2}}|\mathbf{P}%
(Y_{i}\,\leq t,Y_{j}\,\leq s|{\mathcal{F}}_{0})-\mathbf{P}(Y_{i}\,\leq
t,Y_{j}\,\leq s)|\,.
\end{align*}
Define now the coefficient
\[
\beta_{2,Y}(k)=\sup_{i\geq j\geq k}{\mathbf{E}}(b({\mathcal{F}}_{0},i,j))\,.
\]
\end{definition}
\begin{proposition}
\label{applidensity} Let $p\geq4$ and $K$ be any real function satisfying
assumption \textbf{A$_{p}$}. Let $(Y_{i})_{i\geq0}$ be a stationary sequence
with unknown marginal density $f$ such that $\Vert f\Vert_{\infty}<\infty$.
Define
\[
X_{k,n}(x)=K(h_{n}^{-1}(x-Y_{k}))\text{ and }f_{n}(x)=\frac{1}{nh_{n}}%
\sum_{k=1}^{n}X_{k,n}(x)\,,
\]
where $(h_{n})_{n\geq1}$ is a sequence of positive real numbers. Assume that
for some $\eta>0$,
\begin{equation}
\beta_{2,Y}(n)=O(n^{-(p-1+\eta)})\,. \label{condbetathm}%
\end{equation}
Then there exists positive constants $C_{1}$ and $C_{2}$ depending on $\eta$
and $p$ such that for any positive integer $n$,
\begin{align}
& {\mathbf{E}}\int_{\mathbf{R}}|f_{n}(x)-{\mathbf{E}}(f_{n}(x))|^{p}dx\leq
C_{1}(nh_{n})^{-p/2}||dK||^{p/2}||f||_{\infty}^{p/2-1}\Big (\int_{\mathbf{R}%
}|K(u)|du\Big )^{p/2}\label{resdensity}\\
& +C_{2}(nh_{n})^{1-p}\Big (\int_{\mathbf{R}}|K(u)|^{p}du+||dK||\,\int
_{\mathbf{R}}|K(u)|^{p-1}du+||dK||^{2}\int_{\mathbf{R}}|K(u)|^{p-2}%
du\Big )\,,\nonumber
\end{align}
where $||dK||$ is the total variation norm of the measure $dK$.
\end{proposition}
The bound obtained in Proposition \ref{applidensity} can be also compared to
the one obtained in Theorem 3.3 in Viennet (1997) under the assumption that
the strong $\beta$-mixing coefficients in the sense of Rozanov and Volkonskii
(1959) of the sequence $(Y_{i})_{i\in{\mathbf{Z}}}$ , denoted by
$\beta_{\infty}(k)$, satisfy: $\sum_{k\geq1}k^{p-2}\beta_{\infty}(k)<\infty$.
Our condition is then comparable to the one imposed by Viennet (1997) but less
restrictive in the sense that many processes are such that the sequence
$\beta_{2,Y}(n)$ tends to zero as $n\rightarrow\infty$ which is not the case
for $\beta_{\infty}(n)$ (see the examples given in Dedecker and Prieur (2007)).
If we assume that $f$ has a derivative of order $s$, where $s\geq1$ is an
integer and that the following bound holds for the bias term:
\begin{equation}
\int_{\mathbf{R}}|f(x)-{\mathbf{E}}(f_{n}(x))|^{p}dx\leq Mh_{n}^{sp}\Vert
f^{(s)}\Vert_{p}\,, \label{biais}%
\end{equation}
where $M$ is a constant depending on the Kernel $K$, then the choice of
$(nh_{n})^{p/2}h_{n}^{sp}=O(1)$ leads to the following estimate:
\begin{equation}
{\mathbf{E}}\int_{\mathbf{R}}|f_{n}(x)-f(x)|^{p}%
dx=O(n^{-sp/(2s+1)})\,. \label{erreurtotale}%
\end{equation}
We mention that (\ref{biais}) holds for any Parzen Kernel of order $s$ (see
Section 4 in Bretagnolle and Huber (1979)). We also mention that if we only
assume that $\sum_{k\geq1}k^{p-2}\beta_{2,Y}(k)<\infty$ instead of
(\ref{condbetathm}) in Proposition \ref{applidensity} then the inequality
(\ref{resdensity}) is valid with $(nh_{n})^{1-p}n^{\varepsilon}$ (for any
$\varepsilon>0$) replacing $(nh_{n})^{1-p}$ in the second term of the
right-hand side. In this situation, the bound (\ref{biais}) combined with a
choice of $h_{n}$ of order $n^{-1/(1+2s)}$ still leads to the estimate
(\ref{erreurtotale}).
\bigskip
\noindent\textbf{Proof of Proposition \ref{applidensity}.}
Setting $X_{i,n}(x)=K((x-Y_{i})/h_{n})-{\mathbf{E}}(K((x-Y_{i})/h_{n}))$, we
have that
\begin{equation}
{\mathbf{E}}\int_{\mathbf{R}}|f_{n}(x)-{\mathbf{E}}(f_{n}(x))|^{p}%
dx\leq(nh_{n})^{-p}\int_{\mathbf{R}}{\mathbf{E}}\Big |\sum_{i=1}^{n}%
X_{i,n}(x)\Big |^{p}dx\,. \label{trivialbounddensity}%
\end{equation}
Starting from (\ref{trivialbounddensity}) and applying Proposition
\ref{consdirect2} to the stationary sequence $(X_{i,n}(x))_{i\in\mathbf{Z}}$,
Proposition \ref{applidensity} follows provided we establish the following
bounds (in what follows $C$ is a positive constant not depending on $n$ which
may vary from line to line):
\begin{equation}
\int_{\mathbf{R}}{\mathbf{E}}|X_{1,n}(x)|^{p}dx\leq2^{p+1}h_{n}\int
_{\mathbf{R}}|K(u)|^{p}du\,, \label{estdensityb0}%
\end{equation}%
\begin{equation}
\int_{\mathbf{R}}\Big(\sum_{j=0}^{n-1}|{\mathbf{E}}(X_{0,n}(x)X_{j,n}%
(x))|\Big )^{p/2}dx\leq Ch_{n}^{p/2}||dK||^{p/2}||f||_{\infty}^{p/2-1}%
\Big (\int_{\mathbf{R}}|K(u)|du\Big )^{p/2}\,, \label{estdensityb1}%
\end{equation}
and that for $\varepsilon>0$ small enough,
\begin{equation}
\sum_{j=1}^{n}j^{p-2+\varepsilon}\int_{\mathbf{R}}\Vert X_{0,n}(x){\mathbf{E}%
}_{0}(X_{j,n}(x))\Vert_{p/2}^{p/2}dx\leq Ch_{n}||dK||\,\int_{\mathbf{R}%
}|K(u)|^{p-1}du\,, \label{estdensityb3}%
\end{equation}
and
\begin{align}
\sum_{j=1}^{n}j^{p-2+\varepsilon}\sup_{i\geq j} & \int_{\mathbf{R}}%
\Vert{\mathbf{E}}_{0}(X_{i,n}(x)X_{j,n}(x))-{\mathbf{E}}(X_{i,n}%
(x)X_{j,n}(x))\Vert_{p/2}^{p/2}dx\label{estdensityb4}\\
& \leq Ch_{n}||dK||^{2}\,\int_{\mathbf{R}}|K(u)|^{p-2}du\,.\nonumber
\end{align}
In what follows, we shall prove these bounds. Notice first that
\[
\int_{\mathbf{R}}{\mathbf{E}}|X_{1,n}(x)|^{p}dx\leq2^{p+1}\int_{\mathbf{R}%
}\int_{\mathbf{R}}|K((x-y)h_{n}^{-1})|^{p}f(y)dxdy\,,
\]
proving (\ref{estdensityb0}) by the change of variables $u=(x-y)h_{n}^{-1}$.
To prove (\ref{estdensityb1}), we first apply Item 1 of Lemma \ref{covbeta}
implying that
\[
\sum_{j=0}^{n-1}|{\mathbf{E}}(X_{0,n}(x)X_{j,n}(x))|\leq||dK||\,{\mathbf{E}%
}\big (\widetilde{b}({\mathcal{F}}_{0},n)\,|K((x-Y_{0})/h_{n})|\big )\,,
\]
where $\widetilde{b}({\mathcal{F}}_{0},n)=\sum_{j=0}^{n-1}b({\mathcal{F}}%
_{0},j,j)$. An application of H\"{o}lder's inequality as done in Viennet
(1997) at the bottom of page 474, then gives
\[
\int_{\mathbf{R}}\Big(\sum_{j=0}^{n-1}|{\mathbf{E}}(X_{0,n}(x)X_{j,n}%
(x))|\Big )^{p/2}dx\leq h_{n}^{p/2}\Vert f\Vert_{\infty}^{p/2-1}{\mathbf{E}%
}\big (\widetilde{b}({\mathcal{F}}_{0},n))^{p/2}\Big (\int_{\mathbf{R}%
}|K(u)|du\Big )^{p/2}\,.
\]
This proves (\ref{estdensityb1}) since ${\mathbf{E}}\big (\widetilde
{b}({\mathcal{F}}_{0},n))^{p/2}\leq C\sum_{k=1}^{n}k^{p-2}\beta_{2,Y}(k)$ and
$\sum_{k=1}^{n}k^{p-2}\beta_{2,Y}(k)=O(1)$ by condition (\ref{condbetathm}).
\noindent We turn now to the proof of (\ref{estdensityb3}). With this aim we
notice that
\[
\Vert X_{0,n}(x){\mathbf{E}}_{0}(X_{j,n}(x))\Vert_{p/2}^{p/2}={\mathbf{E}%
}(Z_{0}(x){\mathbf{E}}_{0}(X_{j,n}(x)))={\mathbf{E}}(Z_{0}(x)X_{j,n}(x))\,,
\]
where $Z_{0}(x)=|X_{0,n}|^{p/2}|{\mathbf{E}}_{0}(X_{j,n}(x))|^{p/2-1}%
\mathrm{sign}\big ({\mathbf{E}}_{0}(X_{j,n}(x))\big )$. Consequently, by using
Item 1 of Lemma \ref{covbeta} , we derive that
\begin{equation}
\Vert X_{0,n}(x){\mathbf{E}}_{0}(X_{j,n}(x))\Vert_{p/2}^{p/2}=\mathrm{Cov}%
\Big(Z_{0}(x),K((x-Y_{j})/h_{n})\Big )\leq||dK||\,{\mathbf{E}}%
\big (b({\mathcal{F}}_{0},j,j)\,|Z_{0}(x)|\big )\,. \label{used1lemmacovbeta}%
\end{equation}
Notice now that by using the elementary inequality: $x^{\alpha}y^{1-\alpha
}\leq x+y$ valid for $\alpha\in\lbrack0,1]$ and nonnegative $x$ and $y$, we
get that $|Z_{0}(x)|\leq\big (|X_{0,n}(x)|+|{\mathbf{E}}_{0}(X_{j,n}%
(x))|\big )^{p-1}$. Therefore, some computations involving the Jensen's
inequality lead to
\begin{equation}
\int_{\mathbf{R}}|Z_{0}(x)|dx\leq4^{p}h_{n}\int_{\mathbf{R}}|K(u)|^{p-1}du\,.
\label{ine1Z0}%
\end{equation}
Starting from (\ref{used1lemmacovbeta}), we end the proof of
(\ref{estdensityb3}) by taking into account (\ref{ine1Z0}) and the fact that
\[
\sum_{j=1}^{n}j^{p-2+\varepsilon}{\mathbf{E}}(b({\mathcal{F}}_{0}%
,j,j))\leq\sum_{j=1}^{n}j^{p-2+\varepsilon}\beta_{2,Y}(j)
\]
is convergent by condition (\ref{condbetathm}) for any $\varepsilon<\eta$.
It remains to prove (\ref{estdensityb4}). We first write that
\[
\Vert{\mathbf{E}}_{0}(X_{i,n}(x)X_{j,n}(x))-{\mathbf{E}}(X_{i,n}%
(x)X_{j,n}(x))\Vert_{p/2}^{p/2}={\mathbf{E}}(Z_{0}^{(0)}(x)X_{i,n}%
(x)X_{j,n}(x))\,,
\]
where the notation $X^{(0)}$ stands for $X^{(0)}=X-{\mathbf{E}}(X)$ and
\[
Z_{0}(x)=|{\mathbf{E}}_{0}(B_{i,j}(x))|^{p/2-1}\mathrm{sign}\big ({\mathbf{E}%
}_{0}(B_{i,j}(x))\big )\,,
\]
with $B_{i,j}(x)=X_{i,n}(x)X_{j,n}(x)-{\mathbf{E}}(X_{i,n}(x)X_{j,n}(x))$.
Since the variables $X_{i,n}(x)$ and $X_{j,n}(x)$ are centered, an application
of Item 2 of Lemma \ref{covbeta} then gives
\begin{align*}
\Vert{\mathbf{E}}_{0}(X_{i,n}(x)X_{j,n}(x)) & -{\mathbf{E}}(X_{i,n}%
(x)X_{j,n}(x))\Vert_{p/2}^{p/2}\\
& \leq||dK||^{2}\,{\mathbf{E}}\big (|Z_{0}(x)|(b({\mathcal{F}}_{0}%
,i,i)+b({\mathcal{F}}_{0},j,j)+b({\mathcal{F}}_{0},i,j))\big )\,.
\end{align*}
Notice now that since $p/2-1 \geq1,$ we can easily get
\[
\int_{\mathbf{R}}|Z_{0}(x)|dx\leq c_{p}h_{n}\int_{\mathbf{R}}|K(u)|^{p-2}du\,
,
\]
where $c_{p}$ is a positive constant depending on $p$. In addition
\[
\sum_{j=1}^{n}j^{p-2+\varepsilon}\sup_{i\geq j}{\mathbf{E}}(b({\mathcal{F}%
}_{0},i,i)+b({\mathcal{F}}_{0},j,j)+b({\mathcal{F}}_{0},i,j))\leq3\sum
_{j=1}^{n}j^{p-2+\varepsilon}\beta_{2,Y}(j)\,,
\]
which is convergent by condition (\ref{condbetathm}) for any $\varepsilon
<\eta$. Then (\ref{estdensityb4}) holds and so does the proposition.
$\diamond$
\section{Appendix}
This section is devoted to some technical lemmas. Next lemma gives estimates
for terms of the type ${\mathbf{E}}(X_{0}^{u}X_{1}^{p-u})$ .
\begin{lemma}
\label{basic} Let $p$ and $u$ be real numbers such that $0\leq u\leq p-2$. Let
$X_{0}$ and $X_{1}$ be two positive identically distributed random variables,
With the notation $a^{p}={\mathbf{E}}(X_{0}^{p})$, ${\mathbf{E}}_{0}%
(X_{1})={\mathbf{E}}(X_{1}|X_{0})$ the following estimates hold
\begin{equation}
{\mathbf{E}}(X_{0}^{u}X_{1}^{p-u})\leq a^{p-2u/(p-2)}||{\mathbf{E}}_{0}%
(X_{1}^{2})||_{p/2}^{u/(p-2)}\,, \label{b1basic}%
\end{equation}
and
\begin{equation}
{\mathbf{E}}(\ X_{0}^{p-1}X_{1})\leq a^{p-1}||{\mathbf{E}}_{0}(X_{1}%
^{2})||_{p/2}^{1/2}\,. \label{b2basic}%
\end{equation}
\end{lemma}
\noindent\textbf{Proof of Lemma \ref{basic}.} The inequality (\ref{b1basic})
is trivial for $u=0$. To prove it for $u=p-2$, it suffices to write that
${\mathbf{E}}(X_{0}^{p-2}X_{1}^{2})={\mathbf{E}}(X_{0}^{p-2}{\mathbf{E}}%
_{0}(X_{1}^{2}))$, and then to use the H\"{o}lder's inequality.
We prove now the inequality (\ref{b1basic}) for $0<u<p-2$. Select
$x=(p/2-1)/u=(p-2)/2u$. Notice that $2x>1$ and $p-u-1/x>0$ since $u<p-2$.
Then, since the variables are identically distributed,
\begin{gather*}
{\mathbf{E}}(X_{0}^{u}X_{1}^{p-u})={\mathbf{E}}(X_{0}^{u}X_{1}^{1/x}%
X_{1}^{p-u-1/x})\leq||X_{0}^{u}X_{1}^{1/x}|||_{2x}||X_{1}^{p-u-1/x}%
||_{2x/(2x-1)}\\
\leq\big ({\mathbf{E}}(X_{0}^{p-2}X_{1}^{2})\big )^{u/(p-2)}(a^{p}%
)^{1-u/(p-2)}\,.
\end{gather*}
Now, again by H\"{o}lder's inequality applied with $x=p/(p-2)$ and
$1-1/x=2/p$,
\[
{\mathbf{E}}(X_{0}^{p-2}X_{1}^{2}\,)={\mathbf{E}}(X_{0}^{p-2}{\mathbf{E}}%
_{0}(X_{1}^{2}))\leq\big ({\mathbf{E}}(X_{0}^{p})\big )^{(p-2)/p}%
\big ({\mathbf{E}}({\mathbf{E}}_{0}(X_{1}^{2}))^{p/2}\big )^{2/p}%
=(a^{p})^{(p-2)/p}||{\mathbf{E}}_{0}(X_{1}^{2})^{\ }||_{p/2}\,.
\]
Overall
\begin{align*}
{\mathbf{E}}(X_{0}^{u}X_{1}^{p-u}) & \leq a^{u}||{\mathbf{E}}_{0}(X_{1}%
^{2})||_{p/2}^{u/(p-2)}(a^{p})^{1-u/(p-2)}\\
& =a^{p-2u/(p-2)}||{\mathbf{E}}_{0}(X_{1}^{2})||_{p/2}^{u/(p-2)}\,,
\end{align*}
ending the proof of the inequality (\ref{b1basic}).
To prove the inequality (\ref{b2basic}), we use the H\"{o}lder's inequality
which entails that
\[
{\mathbf{E}}(\ X_{0}^{p-1}X_{1})\leq{\mathbf{E}}(\ X_{0}^{p-1}{\mathbf{E}}%
_{0}^{1/2}(X_{1}^{2}))\leq a^{p-1}||{\mathbf{E}}_{0}(X_{1}^{2})||_{p/2}%
^{1/2}\,.
\]
$\diamond$
\bigskip Next lemma gives covariance-type inequalities in terms of beta
coefficients as defined in Definition \ref{beta}.
\begin{lemma}
\label{covbeta} Let $Z$ be a ${\mathcal{F}}_{0}$-measurable real valued random
variable and let $h$ and $g$ be two BV functions (denote by $|| dh ||$ (resp.
$|| dg ||$) the total variation norm of the measure $dh$ (resp. $dg$)). Denote
$Z^{(0)} = Z - {\mathbf{E}} (Z)$, $h^{(0)}(Y_{i}) = h(Y_{i})- {\mathbf{E}}
(h(Y_{i}))$ and $g^{(0)}(Y_{j}) = g(Y_{j})- {\mathbf{E}} (g(Y_{j}))$. Define
the random variables $b({\mathcal{F}}_{0}, i,j)$ as in Definition \ref{beta}. Then
\begin{enumerate}
\item $\big | {\mathbf{E}} \big ( Z^{(0)} h^{(0)}(Y_{i}) \big ) \big | =
\big | \mathrm{Cov} (Z, h(Y_{i}) \big | \leq|| dh || \, {\mathbf{E}}
\big ( |Z| b({\mathcal{F}}_{0}, i,i) \big ) \, . $
\item $\big | {\mathbf{E}} \big ( Z^{(0)} h^{(0)}(Y_{i}) g^{(0)}(Y_{j})
\big ) \big | \leq|| dh || \times|| dg || \, {\mathbf{E}} \big ( |Z|
(b({\mathcal{F}}_{0}, i,i) +b({\mathcal{F}}_{0}, j,j) + b({\mathcal{F}}_{0},
i,j) \big ) \, . $
\end{enumerate}
\end{lemma}
\noindent\textbf{Proof of Lemma \ref{covbeta}.} Item 1 has been proven by
Dedecker and Prieur (2005) (see Item 2 of their proposition 1). Item 2 needs a
proof. We first notice that
\[
h^{(0)}(X)g^{(0)}(Y)=\int\!\!\!\!\int(\mathbf{1}_{X\leq t}-F_{X}%
(t))(\mathbf{1}_{Y\leq s}-F_{Y}(s))dh(t)dg(s)\,.
\]
Therefore
\begin{align*}
& {\mathbf{E}}\big (Z^{(0)}h^{(0)}(Y_{i})g^{(0)}(Y_{j})\big )\\
& \quad={\mathbf{E}}\Big (Z\int\!\!\!\!\int\Big (\mathbf{1}_{Y_{i}\leq
t}^{(0)}\mathbf{1}_{Y_{j}\leq s}^{(0)}-{\mathbf{E}}\big (\mathbf{1}_{Y_{i}\leq
t}^{(0)}\mathbf{1}_{Y_{j}\leq s}^{(0)}\big )\Big )dh(t)dg(s)\Big )\\
& \quad={\mathbf{E}}\Big (Z\int\!\!\!\!\int{\mathbf{E}}\Big (\mathbf{1}%
_{Y_{i}\leq t}^{(0)}\mathbf{1}_{Y_{j}\leq s}^{(0)}-{\mathbf{E}}%
\big (\mathbf{1}_{Y_{i}\leq t}^{(0)}\mathbf{1}_{Y_{j}\leq s}^{(0)}%
\big )|{\mathcal{F}}_{0}\Big )dh(t)dg(s)\Big )\,,
\end{align*}
which proves Item 2 by noticing that
\[
|{\mathbf{E}}\Big (\mathbf{1}_{Y_{i}\leq t}^{(0)}\mathbf{1}_{Y_{j}\leq
s}^{(0)}-{\mathbf{E}}\big (\mathbf{1}_{Y_{i}\leq t}^{(0)}\mathbf{1}_{Y_{j}\leq
s}^{(0)}\big )|{\mathcal{F}}_{0}\Big )|\leq b({\mathcal{F}}_{0}, i,i)
+b({\mathcal{F}}_{0}, j,j) + b({\mathcal{F}}_{0}, i,j) \, .
\]
$\diamond$
\bigskip
Next lemma gives different ranges of inequalities for $|x+y|^{p}$ where $p
\geq2$ is a real number.
\begin{lemma}
\label{cross}
\begin{enumerate}
\item Let $x$ and $y$ be two real numbers and $2\leq p \leq3.$ Then%
\begin{equation}
|x+y|^{p}\leq|x|^{p}+|y|^{p}+p|x|^{p-1}\mathrm{sign}(x)y+ \frac{p(p-1)}{2}
|x|^{p-2}y^{2}. \label{Rio}%
\end{equation}
\item Let $x$ and $y$ be two real numbers and $3< p\leq4.$ Then%
\begin{equation}
|x+y|^{p}\leq|x|^{p}+|y|^{p}+p|x|^{p-1}\mathrm{sign}(x)y+ \frac{p(p-1)}{2}
|x|^{p-2}y^{2} +\frac{2p}{(p-2)} |x| |y|^{p-1}. \label{Rio34}%
\end{equation}
\item Let $x$ and $y$ be two positive real numbers and $p\geq1$ any real
number. Then%
\begin{equation}
(x+y)^{p}\leq x^{p}+y^{p}+4^{p}(x^{p-1}y+xy^{p-1}). \label{int}%
\end{equation}
\item Let $x$ and $y$ be two real numbers and $p$ an even positive integer.
Then%
\begin{equation}
(x+y)^{p}\leq x^{p}+y^{p}+p(x^{p-1}y+xy^{p-1})+2^{p}(x^{2}y^{p-2}+x^{p-2}%
y^{2}). \label{evenint}%
\end{equation}
\end{enumerate}
\end{lemma}
\noindent\textbf{Proof of Lemma \ref{cross}.} The inequality (\ref{Rio}) was
established in Rio (2007, Relation (3.3)) by using Taylor expansion with
integral rest for evaluating the difference $|x+y|^{p}-|x|^{p}$. To prove the
inequality (\ref{Rio34}), we also use the Taylor integral formula of order $2$
that gives
\begin{equation}
|x+y|^{p}-|x|^{p}=p|x|^{p-1}\mathrm{sign}(x)y+C_{p}^{2}|x|^{p-2}y^{2}%
+2C_{p}^{2}y^{2}\int_{0}^{1}(1-t)(|x+ty|^{p-2}-|x|^{p-2})dt\,, \label{order2}%
\end{equation}
where $C_{p}^{2}=p(p-1)/2$. Notice now that, for $3<p\leq4$,
\[
|x+ty|^{p-2}\leq\frac{x^{2}+2|x||ty|+y^{2}}{(|x|+|ty|)^{4-p}}\leq
|x|^{p-2}+2|x||ty|^{p-3}+|ty|^{p-2}%
\]
Hence
\begin{align}
2C_{p}^{2}y^{2} & \int_{0}^{1}(1-t)(|x+ty|^{p-2}-|x|^{p-2})dt\nonumber\\
& \leq2C_{p}^{2}|y|^{p}\int_{0}^{1}(1-t)t^{p-2}dt+4C_{p}^{2}|x||y|^{p-1}%
\int_{0}^{1}(1-t)t^{p-3}dt\nonumber\\
& =2|y|^{p}C_{p}^{2}\frac{\Gamma(p-1)}{\Gamma(p+1)}+4|x||y|^{p-1}C_{p}%
^{2}\frac{\Gamma(p-2)}{\Gamma(p)}\nonumber\\
& =|y|^{p}+\frac{2p}{(p-2)}|x||y|^{p-1}\,. \label{betafunction}%
\end{align}
Starting from (\ref{order2}) and using (\ref{betafunction}), the inequality
(\ref{Rio34}) follows.
The inequality (\ref{int}) was observed by Shao (1995, page 957). We shall
establish now (\ref{evenint}). We start by noticing that for any $a$, $b$ two
positive real numbers and $\ 2\leq k\leq p-2\ $\ we have
\begin{equation}
a^{p-k}b^{k}\leq a^{2}b^{p-2}+a^{p-2}b^{2} \label{ab}%
\end{equation}
\noindent Now for $p$ an even positive integer and $x$ and $y$ two real
numbers, the Newton binomial formula gives
\begin{align*}
(x+y)^{p} & =x^{p}+y^{p}+p(x^{p-1}y+xy^{p-1})+\sum_{k=2}^{p-2}C_{p}%
^{k}x^{p-k}y^{k}\\
& \leq x^{p}+y^{p}+p(x^{p-1}y+xy^{p-1})+\sum_{k=2}^{p-2}C_{p}^{k}%
|x|^{p-k}|y|^{k}%
\end{align*}
Whence, by (\ref{ab}) and the fact that $\sum_{k=0}^{p}C_{p}^{k}=2^{p}$
inequality (\ref{evenint}) follows. $\diamond$
\begin{lemma}
\label{lmasubadd} Let $(V_{i})_{i \geq0}$ be a sequence of non negative
numbers such that $V_{0} = 0$ and for all $i,j \geq0$,
\begin{equation}
\label{condsubadd}V_{i+j} \leq C ( V_{i} + V_{j}) \, ,
\end{equation}
where $C \geq1$ is a constant not depending on $i$ and $j$. Then
\begin{enumerate}
\item For any integer $r \geq1$, any integer $n$ satisfying $2^{r-1} \leq n <
2^{r}$ and any real $q \geq0$
\[
\sum_{i=0}^{r-1} \frac{1}{2^{iq}} V_{2^{i}} \leq C 2^{q+2} ( 2^{q+1} -1)^{-1}
\sum_{\ell= 1}^{n} \frac{V_{k}}{k^{1+q}} \, .
\]
\item For any positive integers $k$ and $m$ and any real $q>0$,
\[
\sum_{j=1}^{k}\frac{1}{j^{q}}V_{jm}\leq2^{q+1}Cq^{-1}m^{q-1}\sum_{\ell=1}%
^{m}\frac{1}{(\ell+m)^{q}}V_{\ell}+2Cq^{-1}m^{q-1}\sum_{\ell=m+1}^{km}\frac
{1}{\ell^{q}}V_{\ell}\,.
\]
\item Let $0< \delta\leq\gamma\leq1.$ Then for any real $q \geq 0$,
\[
\left( \sum_{k=1}^{n}\frac{1}{k^{1+q\gamma}}V_{k}^{\gamma}\right)
^{1/\gamma}\leq2^{1/\delta- 1/ \gamma} C^{(\gamma-\delta)/\delta}\left(
\sum_{k=1}^{n}\frac{1}{k^{1+q\delta}}V_{k}^{\delta}\right) ^{1/\delta} \, .
\]
\end{enumerate}
\end{lemma}
\begin{remark}
If $(V_{i})_{i \geq0}$ satisfies (\ref{condsubadd}) with $C=1$, then the
sequence is said to be subadditive.
\end{remark}
\noindent\textbf{Proof of Lemma \ref{lmasubadd}.}
The condition (\ref{condsubadd}) implies that for any integer $k$ and any
integer $0\leq j\leq k$,
\begin{equation}
V_{k}\leq C(V_{j}+V_{k-j})\text{ and then that }(k+1)V_{k}\leq2C\sum_{j=1}%
^{k}V_{j}\,. \label{condsubadd2}%
\end{equation}
Therefore for $2^{r-1}\leq n <2^{r}$,
\[
\sum_{i=0}^{r-1}\frac{1}{2^{iq}}V_{2^{i}}\leq2C\sum_{j=1}^{2^{r-1}}V_{j}%
\sum_{i:2^{i}\geq j}\frac{1}{2^{i(q+1)}}\,,
\]
proving Item 1. To prove Item 2, using again (\ref{condsubadd2}), it suffices
to notice that
\begin{align*}
\sum_{j=1}^{k}\frac{1}{j^{q}}V_{jm} & \leq2C\sum_{j=1}^{k}\frac
{1}{(1+jm)j^{q}}\sum_{\ell=1}^{jm}V_{\ell}\\
& \leq2Cm^{-1}(2m)^{q}\sum_{j=1}^{k}j^{-q-1}\sum_{\ell=1}^{m}\frac{1}%
{(\ell+m)^{q}}V_{\ell}+2Cq^{-1}m^{q-1}\sum_{\ell=m+1}^{km}\frac{1}{\ell^{q}%
}V_{\ell}\, .
\end{align*}
To prove Item 3, we first notice that (\ref{condsubadd}) entails that
\[
V^{\gamma}_{i+j} \leq C^{\gamma} ( V^{\gamma}_{i} + V^{\gamma}_{j}) \text{ and
then that }(k+1)V^{\gamma}_{k}\leq2C^{\gamma}\sum_{j=1}^{k}V^{\gamma}_{j}\, .
\]
Then for any real $q \geq 0$,
\begin{equation}
\label{ineitem3}k^{-q(\gamma- \delta)} V_{k}^{\gamma- \delta} \leq2^{1-
\delta/\gamma} C^{\gamma- \delta} \Big (\sum_{j=1}^{k}j^{-(1+q\gamma)}%
V_{j}^{\gamma}\ \Big )^{1 - \delta/ \gamma}\, .
\end{equation}
Writing that $k^{-(1+q\gamma)}V_{k}^{\gamma}= k^{-(1+q\delta)}V_{k}^{\delta
}\times k^{-q(\gamma-\delta)}V_{k}^{\gamma-\delta}$ and using (\ref{ineitem3}%
), the following inequality holds
\[
\sum_{k=1}^{n}\frac{1}{k^{1+q\gamma}}V_{k}^{\gamma} \leq2^{1- \delta/\gamma}
C^{\gamma- \delta} \Big (\sum_{k=1}^{n}k^{-(1+q\delta)}V_{k}^{\delta
}\ \Big ) \Big (\sum_{j=1}^{n}j^{-(1+q\gamma)}V_{j}^{\gamma}\ \Big )^{1 -
\delta/ \gamma} \, ,
\]
proving Item 3. $\diamond$
\medskip
|
\section{Introduction\label{sec1}}
\indent\indent
Extensions of the standard model (SM) with a warped dimension\cite{RS,Chang,Csaki,
Gherghetta1}, where all SM fields are propagating in the bulk, provide a naturally
geometrical solution to the hierarchy problem regarding the huge difference between
the Planck scale and the electroweak one. The small mixing between zero modes
and heavy Kaluza-Klein (KK) excitations can induce the observed fermion masses
and corresponding weak mixing angles\cite{Grossman,Gherghetta2}, and suppress
flavor changing neutral current (FCNC) couplings\cite{Huber1,Agashe1}. Additional,
realistic models of electroweak symmetry breaking in warped extra dimension are constructed
in literature\cite{Agashe2,Csaki2,Agashe3,Csaki3,Contino,Carena}, and the gauge coupling
unification with a warped extra dimension is also analyzed in Ref.\cite{Agashe4,Agashe5}.
If the SM gauge group $SU(2)_L\times U(1)_Y$ is chosen in the bulk for extensions
of the SM with a warped extra dimension, the electroweak precision observables, for example
the experimental data on the $S,\;T$ parameters and the well-measured
$Z\bar{b}_Lb_L$ coupling\cite{Agashe6,Casagrande,Bauer1,Bauer2}, generally require
that the exciting KK modes are heavier than 10 TeV and exceed the reach of
colliders running now. In order to accomodate light exciting KK modes with
${\cal O}(1{\rm TeV})$ masses in warped extra dimension,
literature\cite{Agashe7,Santiago1} enlarges the gauge group in the bulk to
$SU(3)_c\times SU(2)_L\times SU(2)_R\times U(1)_X\times P_{LR}$. With an appropriate
choice of quark bulk masses, one indeed obtains the agreement with
the electroweak precision data in the presence of light KK excitations\cite{Djouadi,Bouchart}.
Actually, the electroweak precision observables are consistent with the light
fermion KK modes with masses even below $1{\rm Tev}$ while the masses of the KK
gauge bosons are forced to be at least $(2-3){\rm TeV}$ to be consistent with
the data on the parameter $S$.
However, the large FCNC transitions are aroused by light exciting KK modes if we assume
that the hierarchy of fermion masses together with corresponding weak mixings solely originate
from geometry and the fundamental Yukawa couplings in five dimension are anarchic\cite{Csaki4,
Buras1,Agashe8,Davidson,Iltan,Agashe9}. To suppress the large FCNC
processes mediated by light exciting KK modes, literature\cite{Cacciapaglia} introduces
the bulk and brane flavor symmetries where the naturally geometric explanation
of fermion hierarchies is abandoned. The author of Ref.\cite{Santiago2} proposes
the minimal flavor protection where a global $U(3)$ bulk flavor symmetry is imposed
on the triplet reprsentations of the local gauge symmetry $SU(2)_L\times SU(2)_R$
in the quark sector. In a similar way, literature\cite{Chen1} chooses the global flavor
symmetry $U(3)_L\times U(3)_R$ on lepton sector to control relevant
FCNC transitions. Another approach to suppress FCNC processes in warped
extra dimension introduces two horizontal $U(1)$ symmetries which guarantee
bulk masses in alignment with Yukawa couplings for charge $-1/3$ quarks and charge
$-1$ leptons respectively\cite{Csaki5}.
An analogous strategy to solve the FCNC problems in warped geometry is based on
$A_4$ flavor symmetry\cite{Csaki6,Kadosh}. A very different approach has been
presented in Ref.\cite{Perez1,Perez2} where the bulk mass matrices are expressed
in terms of five dimensional Yukawa couplings, thus flavor violation at low energy can be suppressed
rationally. Comparing with the choices mentioned above,
the warped extra dimension with a soft-wall\cite{Falkowski1,Gherghetta3,
Delgado,Santiago3,Gherghetta4,Cabrer1} perhaps provides a natural solution
to accomodate FCNC transitions at acceptable level and the lightest KK excitations
with ${\cal O}(1{\rm TeV})$ masses simultaneously\cite{Huber2}.
In a warped extra dimension with the custodial symmetry
$SU(3)_c\times SU(2)_L\times SU(2)_R\times U(1)_X\times P_{LR}$, a meticulous
analysis on the electroweak and flavor structure is provided in Ref.\cite{Buras2},
a complete study of rare $K$ and $B$ meson decays is presented in Ref.\cite{Buras3},
the impact from KK excitations of fermions on the couplings among
the SM particles is given in Ref.\cite{Buras4}. Assuming all fields are
propagating in the bulk, the authors of literature\cite{Azatov} analyze the FCNC processes
mediated by a light Higgs, the authors of literature\cite{Agashe10}
present an analysis on production and decay of KK gravitons at the LHC,
and the authors of literature\cite{Agashe11} study signals for KK
excitations of electroweak and strong gauge bosons in the LHC. Additional, an analysis of
loop-induced rare lepton FCNC transition $\mu\rightarrow e\gamma$ in a warped extra
dimension with anarchic Yukawa couplings and IR brane localized Higgs is presented
in literature\cite{Csaki7} through the five dimensional mixed position/momentum space
formalism\cite{Carena1} and mass insertion approach.
It is well known that all virtual KK excitations contribute their corrections
to theoretical predictions on the physical quantities at electroweak scale,
and those theoretical corrections should be summed over infinite KK modes in principle\cite{Hirn,Azatov}.
In this work, we verify some lemmas on the eigenvalues of exciting KK modes in
extension of the SM with a warped extra dimension and the custodial symmetry
$SU(3)_c\times SU(2)_L\times SU(2)_R\times U(1)_X\times P_{LR}$\cite{Agashe1,Agashe12}.
Performing properly analytic extensions of the bulk profiles, we sum over the infinite series of
KK modes through the residue theorem. Additional, we also present the sufficient
condition for a convergent series of infinite KK modes in extensions of the SM
with a warped extra dimension. The method can also be applied to sum over
the infinite series of KK modes in unified extra dimension.
The emphasized point here is that the authors of literature\cite{Hirn} also propose
to sum over infinite series of KK modes applying the equations of motion and
corresponding completeness relations of bulk profiles. Nevertheless, the method
proposed there can only be applied to sum over the infinite series of KK modes
for the five dimensional fields with zero bulk mass and zero modes in extensions
of the SM with a warped extra dimension.
Our presentation is organized as follows. In section \ref{sec2}, the main ingredients
of a warped extra dimension with custodial symmetry are summarized briefly, the KK decompositions
of all five dimensional fields and relevant bulk profiles are given here also.
In section \ref{sec3}, we present the verification of relevant lemmas on the eigenvalues
of exciting KK modes in detail. In order to sum over the infinite series of KK modes properly, we also
discuss how to extend bulk profiles analytically. We discuss the possible relation
between the perturbative expansions in four dimensional effective theory and five
dimensional full theory in section \ref{sec4}. Furthermore, we also recover the equations
presented in Ref.\cite{Casagrande} through the residue theorem. In section \ref{sec5}, we
show how to sum over the infinite series of KK modes in unified extra dimension using residue theorem
through recover an important relation applied extensively in literature.
As an example, we demonstrate in section \ref{sec6} that the corrections from neutral Higgs to the
Wilson coefficients of relevant operators for $B\rightarrow X_s\gamma$ contain the suppression factor
$m_b^3m_s/m_{_{\rm w}}^4$ in comparison with that
from other sectors, thus can be neglected safely. Our conclusions are summarized
in Section \ref{sec7}.
\section{A warped extra dimension with custodial symmetry\label{sec2}}
\indent\indent
In the Randall-Sundrum (RS) scenario, four dimensional Minkowskian space-time
is embedded into a slice of five dimensional anti de-Sitter (ADS$_5$)
space with curvature $k$. The fifth dimension is a $S^1/Z_2\times Z_2^\prime$
orbifold of size r labeled by a coordinate $\phi\in[-\pi,\pi]$, thus
the points $(x^\mu,\phi)$, $(x^\mu,\pi-\phi)$, $(x^\mu,\pi+\phi)$ and $(x^\mu,-\phi)$ are
identified all. The corresponding metric of non-factorizable RS geometry is written as
\begin{eqnarray}
&&ds^2=e^{-2\sigma(\phi)}\eta_{\mu\nu}dx^\mu dx^\nu-r^2d\phi^2,
\;\sigma(\phi)=kr|\phi|,
\label{metric}
\end{eqnarray}
where $x^\mu\;(\mu=0,\;1,\;2,\;3)$ are the coordinates on the four dimensional
hyper-surfaces of constant $\phi$ with metric $\eta_{\mu\nu}=(1,-1,-1,-1)$,
and $e^\sigma$ is called the warp factor. Two branes are located on the
orbifold fixed points $\phi=0$ and $\phi=\pi/2$, respectively. The brane
on $\phi=0$ is called Planck or ultra-violet (UV) brane, and the brane on
$\phi=\pi/2$ is called TeV or infra-red (IR) brane. Assuming the parameters
$k$ and $1/r$ to be of order the fundamental Planck scale $M_{\rm PI}$ and choosing
the product $kr\simeq24$, one gets the inverse warp factor
\begin{eqnarray}
&&\epsilon={\Lambda_{\rm IR}\over\Lambda_{\rm UV}}\equiv e^{-kr\pi/2}\simeq10^{-16},
\label{warped-factor}
\end{eqnarray}
which explains the hierarchy between the electroweak and Planck scale naturally.
Meanwhile, the mass scale of low-lying KK excitations is set as
\begin{eqnarray}
&&\Lambda_{_{KK}}\equiv k\epsilon=ke^{-kr\pi/2}={\cal O}(1{\rm TeV}).
\label{KK-scale}
\end{eqnarray}
In the gauge symmetry $SU(2)_L\times SU(2)_R\times U(1)_X\times P_{LR}$,
the discrete symmetry $P_{LR}$ interchanging the local groups $SU(2)_L$ and
$SU(2)_R$ implies that the five dimensional gauge couplings satisfy the relation
$g_{5L}=g_{5R}=g_5$. The local gauge group $SU(2)_L\times SU(2)_R\times U(1)_X$
is broken to the SM gauge group by the boundary conditions (BCs) on the UV brane:
\begin{eqnarray}
&&W_{L,\mu}^{1,2,3}(++),\;B_\mu(++),\;W_{R,\mu}^{1,2}(-+),\;Z_{X,\mu}(-+),\;\;
(\mu=0,\;1,\;2,\;3),
\nonumber\\
&&W_{L,5}^{1,2,3}(--),\;B_5(--),\;W_{R,5}^{1,2}(+-),\;Z_{X,5}(+-).
\label{BCs-gauge}
\end{eqnarray}
Here, the first (second) sign is the BC on the UV (IR) brane: $+$ denotes a Neumann
BC and $-$ denotes a Dirichlet BC. The zero modes of fields with (++)BCs are massless
before the electroweak symmetry is spontaneously broken, and are
identified as the SM gauge bosons. The fields with other BCs only contain
massive KK modes. The third component of $SU(2)_R$ gauge fields
$W_{R,M}^3$ and the $U(1)_X$ gauge field $\tilde{B}_M$ are expressed in
terms of the neutral gauge fields $Z_{X,M}$ and $B_M$ as
\begin{eqnarray}
&&W_{R,M}^3={g_5Z_{X,M}+g_{5X}B_M\over\sqrt{g_5^2+g_{5X}^2}},\;
\tilde{B}_M=-{g_{5X}Z_{X,M}-g_5B_M\over\sqrt{g_5^2+g_{5X}^2}},\;(M=0,\;1,\;2,\;3,\;5),
\label{gauge1}
\end{eqnarray}
where $g_{5X}$ is the five dimensional gauge coupling of $U(1)_X$. The Lagrangian
for gauge sector is written as
\begin{eqnarray}
&&{\cal L}_{gauge}={\sqrt{\cal G}\over r}{\cal G}^{KM}{\cal G}^{LN}\Big(-{1\over4}W_{KL}^iW_{MN}^i
-{1\over4}\tilde{W}_{KL}^i\tilde{W}_{MN}^i-{1\over4}\tilde{B}_{KL}\tilde{B}_{MN}
-{1\over4}G_{KL}^aG_{MN}^a\Big)\;.
\label{L_gauge}
\end{eqnarray}
For convenience in our discussion further, we define the following five dimensional
fields\cite{Agashe11}
\begin{eqnarray}
&&A_A={\sqrt{g_5^2+g_{5X}^2}B_A+g_{5X}W_{L,A}^3\over\sqrt{g_5^2+2g_{5X}^2}},
\nonumber\\
&&Z_{A}={-g_{5X}B_A+\sqrt{g_5^2+g_{5X}^2}W_{L,A}^3\over\sqrt{g_5^2+2g_{5X}^2}},
\nonumber\\
&&W_{L,A}^{\pm}={1\over\sqrt{2}}\Big(W_{L,A}^1\mp iW_{L,A}^2\Big)\;,
\nonumber\\
&&W_{R,A}^{\pm}={1\over\sqrt{2}}\Big(W_{R,A}^1\mp iW_{R,A}^2\Big)\;.
\label{gauge2}
\end{eqnarray}
As regards the matter fields, the quarks of one generation are embedded into the multiplets\cite{Buras5}:
\begin{eqnarray}
&&Q_L^i=\left(\begin{array}{ll}\chi_{u_L}^i(-+)_{5/3}&q_{u_L}^i(++)_{2/3}\\
\chi_{d_L}^i(-+)_{2/3}&q_{d_L}^i(++)_{-1/3}\end{array}\right)\;,\;\;
Q_{u_R}^i=u_R^i(++)_{2/3}
\nonumber\\
&&\tilde{Q}_{d_R}^i=\left(\begin{array}{l}X_R^i(-+)_{5/3}\\U_R^i(-+)_{2/3}
\\D_R^i(-+)_{-1/3}\end{array}\right)\;,\;\;
Q_{d_R}^i=\left(\begin{array}{l}\tilde{X}_R^i(-+)_{5/3}\\ \tilde{U}_R^i(-+)_{2/3}
\\d_R^i(++)_{-1/3}\end{array}\right)\;,
\label{SM-quarks}
\end{eqnarray}
and the states with opposite chirality are written as
\begin{eqnarray}
&&Q_R^i=\left(\begin{array}{ll}\chi_{u_R}^i(+-)_{5/3}&q_{u_R}^i(--)_{2/3}\\
\chi_{d_R}^i(+-)_{2/3}&q_{d_R}^i(--)_{-1/3}\end{array}\right)\;,\;\;
Q_{u_L}^i=u_L^i(--)_{2/3}
\nonumber\\
&&\tilde{Q}_{d_L}^i=\left(\begin{array}{l}X_L^i(+-)_{5/3}\\U_L^i(+-)_{2/3}
\\D_L^i(+-)_{-1/3}\end{array}\right)\;,\;\;
Q_{d_L}^i=\left(\begin{array}{l}\tilde{X}_L^i(+-)_{5/3}\\ \tilde{U}_L^i(+-)_{2/3}
\\d_L^i(--)_{-1/3}\end{array}\right)\;.
\label{opposite-quarks}
\end{eqnarray}
Here, $i=1,\;2,\;3$ denote the indices of generation, the $U(1)_X$ charges are all assigned as
\begin{eqnarray}
&&Y_{Q^i}=Y_{u^i}=Y_{Q_d^i}={2\over3}\;.
\label{U1_charges}
\end{eqnarray}
In order to give the kinetic terms of triplets, we redefine the quarks in
triplet as
\begin{eqnarray}
&&\tilde{T}_{Q_R}^i=\left(\begin{array}{c}{1\over\sqrt{2}}(X_R^i+D_R^i)
\\{i\over\sqrt{2}}(X_R^i-D_R^i)\\U_R^i\end{array}\right)\;,\;\;
T_{Q_R}^i=\left(\begin{array}{c}{1\over\sqrt{2}}(\tilde{X}_R^i+d_R^i)
\\{i\over\sqrt{2}}(\tilde{X}_R^i-d_R^i)\\ \tilde{U}_R^i\end{array}\right)\;,
\nonumber\\
&&\tilde{T}_{Q_L}^i=\left(\begin{array}{c}{1\over\sqrt{2}}(X_L^i+D_L^i)
\\{i\over\sqrt{2}}(X_L^i-D_L^i)\\U_L^i\end{array}\right)\;,\;\;
T_{Q_L}^i=\left(\begin{array}{c}{1\over\sqrt{2}}(\tilde{X}_L^i+d_L^i)
\\{i\over\sqrt{2}}(\tilde{X}_L^i-d_L^i)\\ \tilde{U}_L^i\end{array}\right)\;.
\label{redefine-triplet}
\end{eqnarray}
Correspondingly, the Lagrangian for kinetic terms of quarks can be written as
\begin{eqnarray}
&&{\cal L}_{Q}
={\sqrt{\cal G}\over2r}\sum\limits_{i=1}^3
\Big\{(\overline{Q}^i)_{a_1a_2}iE_A^M\gamma^A\Big[\Big({1\over2}(\partial_M
-\overleftarrow{\partial}_M)+ig_{5s}T^aG^a_M+ig_{5X}Y_{Q^i}\tilde{B}_M\Big)\delta_{a_1b_1}
\delta_{a_2b_2}
\nonumber\\
&&\hspace{1.2cm}
+ig_5({\sigma^{c_1}\over2})_{a_1b_1}W_{L,M}^{c_1}\delta_{a_2b_2}
+ig_5({\sigma^{c_2}\over2})_{a_2b_2}W_{R,M}^{c_2}\delta_{a_1b_1}\Big](Q^i)_{b_1b_2}
\nonumber\\
&&\hspace{1.2cm}
+(\overline{Q}^i)_{a_1a_2}\Big[iE_A^M\gamma^A\omega_M-{\rm sgn}(\phi)(c_{_B})_{ij}\Big]
(Q^j)_{a_1a_2}
\nonumber\\
&&\hspace{1.2cm}
+\overline{u}^i\Big[iE_A^M\gamma^A\Big({1\over2}(\partial_M
-\overleftarrow{\partial}_M)+ig_{5s}T^aG^a_M+ig_{5X}Y_{u^i}\tilde{B}_M\Big)\delta_{ij}
\nonumber\\
&&\hspace{1.2cm}
+iE_A^M\gamma^A\omega_M-{\rm sgn}(\phi)(c_{_S})_{ij}\Big]u^j
\nonumber\\
&&\hspace{1.2cm}
+(\overline{\tilde{T}}_Q^i)_{a1}iE_A^M\gamma^A\Big[\Big({1\over2}(\partial_M
-\overleftarrow{\partial}_M)+ig_{5s}T^aG^a_M+ig_{5X}Y_{Q_d^i}\tilde{B}_M\Big)
\delta_{a_1b_1}
\nonumber\\
&&\hspace{1.2cm}
+g_5\varepsilon_{a_1b_1c_1}W_{L,M}^{c_1}\Big](\tilde{T}_Q^i)_{b_1}
+(\overline{\tilde{T}}_Q^i)_{a_1}\Big[iE_A^M\gamma^A\omega_M-{\rm sgn}(\phi)(\eta_3)_{ij}\Big]
(\tilde{T}_Q^j)_{a_1}
\nonumber\\
&&\hspace{1.2cm}
+(\overline{T}_Q^i)_{a_1}iE_A^M\gamma^A\Big[\Big({1\over2}(\partial_M
-\overleftarrow{\partial}_M)+ig_{5s}T^aG^a_M+ig_{5X}Y_{Q_d^i}\tilde{B}_M\Big)
\delta_{a_1b_1}
\nonumber\\
&&\hspace{1.2cm}
+g_5\varepsilon_{a_1b_1c_1}W_{R,M}^{c_1}\Big](T_Q^i)_{b_1}
+(\overline{T}_Q^i)_{a_1}\Big[iE_A^M\gamma^A\omega_M-{\rm sgn}(\phi)(c_{_T})_{ij}
\Big](T_Q^j)_{a_1}+h.c.\Big\}\;,
\label{kinetic_quark}
\end{eqnarray}
with $\gamma^A=(\gamma^\mu,\;-i\gamma^5)$, the inverse vielbein
$E_B^A={\it diag}(e^{\sigma(\phi)},\;e^{\sigma(\phi)},\;e^{\sigma(\phi)},\;
e^{\sigma(\phi)},{1\over r})$, and the spin connection
$\omega_A=({\rm sgn}(\phi){i\over2}ke^{-\sigma(\phi)}\gamma_\mu\gamma^5,0)$.
Generally, three bulk mass matrices
$c_{_B},\;c_{_S},\;c_{_T}$ are arbitrarily hermitian $3\times3$ matrices.
To break down the electroweak symmetry, we introduce an IR brane located
Higgs which transforms as a self-dual bidoublet under the gauge group
$SU(2)_L\times SU(2)_R$, and transforms as a singlet with charge $Y_H=0$
under the gauge group $U(1)_X$:
\begin{eqnarray}
&&H=\left(\begin{array}{cc}-i\pi^+/\sqrt{2}&-(h^0-i\pi^0)/2\\
(h^0+i\pi^0)/2&i\pi^-/\sqrt{2}
\end{array}\right)\;.
\label{Higgs-bidoublet}
\end{eqnarray}
After normalizing the kinetic term of Higgs in five dimension, we write the
corresponding Lagrangian as
\begin{eqnarray}
&&{\cal L}_{H}={\rm Tr}\Big[\Big(D_\mu
\Phi(x)\Big)^\dagger\Big(D^\mu\Phi(x)\Big)\Big]
-\mu^2{\rm Tr}\Big(\Phi^\dagger(x)\Phi(x)\Big)
+{\lambda\over2}\Big[{\rm Tr}\Big(\Phi^\dagger(x)\Phi(x)\Big)\Big]^2\;.
\label{IR-Brane-Higgs0}
\end{eqnarray}
Accordingly, the Yukawa couplings between quarks and Higgs can be formulated as
\begin{eqnarray}
&&{\cal L}_{Y}^Q=
e^{kr\pi/2}\sqrt{-{\cal G}_{_{IR}}}\sum\limits_{i,j=1}^3\Big\{\sqrt{2}
\lambda_{ij}^{u}\overline{Q}_{a\alpha}^iH_{a\alpha}u^j-2\lambda_{ij}^{d}\Big[
\overline{Q}_{a\alpha}^i(\tau^c)_{ab}(\tilde{T}_d^j)_cH_{b\alpha}
\nonumber\\
&&\hspace{1.2cm}
+\overline{Q}_{a\alpha}^i(\tau^c)_{\alpha\beta}(T_d^j)_cH_{a\beta}\Big]+h.c.\Big\}\;,
\label{Yukawa-quarks}
\end{eqnarray}
here the metric on the IR brane ${\cal G}_{_{IR}}^{\mu\nu}=e^{kr\pi/2}\eta^{\mu\nu}$.
In the following we choose to work in the background gauge
with $\xi=1$\cite{Randall1}. Furthermore,
we also assume that three bulk mass matrices $c_{_B},\;c_{_S},\;c_{_T}$ are real and
diagonal, i.e. each of them is described by three real parameters. This can
always be obtained through some appropriate field redefinitions.
The KK decompositions of five dimensional gauge fields are extensively studied
in literature, and can be written in our notations as
\begin{eqnarray}
&&A_\mu(x,\phi)={1\over\sqrt{r}}\sum\limits_{n=0}^\infty A_\mu^{(n)}(x)
\chi_{_{(++)}}^A(y_{_{(++)}}^{A(n)},\phi),
\nonumber\\
&&A_\phi(x,\phi)={1\over \Lambda_{_{KK}}\sqrt{r}}\sum\limits_{n=1}^\infty{1\over y_{_A}^{(n)}}
\varphi_{_A}^{(n)}(x)\partial_\phi\chi_{_{(++)}}^A(y_{_{(++)}}^{A(n)},\phi),
\nonumber\\
&&Z_\mu(x,\phi)={1\over\sqrt{r}}\sum\limits_{n=0}^\infty Z_\mu^{(n)}(x)
\chi_{_{(++)}}^Z(y_{_{(++)}}^{Z(n)},\phi),
\nonumber\\
&&Z_\phi(x,\phi)={1\over \Lambda_{_{KK}}\sqrt{r}}\sum\limits_{n=1}^\infty{1\over y_{_Z}^{(n)}}\varphi_{_Z}^{(n)}(x)
\partial_\phi\chi_{_{(++)}}^Z(y_{_{(++)}}^{Z(n)},\phi),
\nonumber\\
&&Z_{X,\mu}(x,\phi)={1\over\sqrt{r}}\sum\limits_{n=1}^\infty Z_{X,\mu}^{(n)}(x)
\chi_{_{(-+)}}^{Z_X}(y_{_{(-+)}}^{Z_X(n)},\phi),
\nonumber\\
&&Z_{X,\phi}(x,\phi)={1\over \Lambda_{_{KK}}\sqrt{r}}\sum\limits_{n=1}^\infty
{1\over y_{_{Z_X}}^{(n)}}\varphi_{_{Z_X}}^{(n)}(x)\partial_\phi
\chi_{_{(-+)}}^{Z_X}(y_{_{(-+)}}^{Z_X(n)},\phi),
\nonumber\\
&&W_{L,\mu}^\pm(x,\phi)={1\over\sqrt{r}}\sum\limits_{n=0}^\infty W_{L,\mu}^{\pm(n)}(x)
\chi_{_{(++)}}^{W_L}(y_{_{(++)}}^{W_L(n)},\phi),
\nonumber\\
&&W_{L,\phi}^\pm(x,\phi)={1\over \Lambda_{_{KK}}\sqrt{r}}\sum\limits_{n=1}^\infty{1\over y_{_{W_L}}^{(n)}}
\varphi_{_{W_L}}^{\pm(n)}(x)\partial_\phi\chi_{_{(++)}}^{W_L}(y_{_{(++)}}^{W_L(n)},\phi),
\nonumber\\
&&W_{R,\mu}^\pm(x,\phi)={1\over\sqrt{r}}\sum\limits_{n=1}^\infty W_{R,\mu}^{\pm(n)}(x)
\chi_{_{(-+)}}^{W_R}(y_{_{(-+)}}^{W_R(n)},\phi),
\nonumber\\
&&W_{R,\phi}^\pm(x,\phi)={1\over \Lambda_{_{KK}}\sqrt{r}}\sum\limits_{n=1}^\infty{1\over y_{_{W_R}}^{(n)}}
\varphi_{_{W_R}}^{\pm(n)}(x)\partial_\phi\chi_{_{(-+)}}^{W_R}(y_{_{(-+)}}^{W_R(n)},\phi),
\nonumber\\
&&G^\pm(x,\phi)=\sum\limits_n\Big\{b_n^{W_L}(\phi)\varphi_{W_L}^{\pm,(n)}(x)
+b_n^{W_R}(\phi)\varphi_{W_R}^{\pm,(n)}(x)\Big\},
\nonumber\\
&&G^0(x,\phi)=\sum\limits_n\Big\{b_n^Z(\phi)\varphi_Z^{(n)}(x)
+b_n^{Z_X}(\phi)\varphi_{Z_X}^{(n)}(x)\Big\},
\nonumber\\
&&G_\mu^a(x,\phi)={1\over\sqrt{r}}\sum\limits_{n=0}^\infty G_\mu^{a,(n)}(x)
\chi_{_{(++)}}^g(y_{_{(++)}}^{g(n)},\phi),
\nonumber\\
&&G_\phi^a(x,\phi)={1\over \Lambda_{_{KK}}\sqrt{r}}\sum\limits_{n=1}^\infty{1\over y_{_g}^{(n)}}
\varphi_{_g}^{a,(n)}(x)\partial_\phi\chi_{_{(++)}}^g(y_{_{(++)}}^{g(n)},\phi).
\label{KK-decomposition-Gauge-Higgs}
\end{eqnarray}
As $G=A,\;Z,\;W_L^\pm,\;g$, $y_{_{(++)}}^{G(n)}\;(n=0,\;1,\;\cdots,\;\infty)$ denote
the roots of equation $z^2R_{_{(++)}}^{G,\epsilon}(z)\equiv0$ with
\begin{eqnarray}
&&R_{_{(++)}}^{G,\epsilon}(z)=Y_0(z)J_0(z\epsilon)-J_0(z)Y_0(z\epsilon)\;.
\label{root1}
\end{eqnarray}
When $G=Z_{_X},\;W_R^\pm$, $y_{_{(-+)}}^{G(n)}\;(n=1,\;2,\;\cdots,\;\infty)$ denote
the roots of equation $R_{_{(-+)}}^{G,\epsilon}(z)\equiv0$ with
\begin{eqnarray}
R_{_{(-+)}}^{G,\epsilon}(z)=Y_0(z)J_1(z\epsilon)-J_0(z)Y_1(z\epsilon)\;.
\label{root2}
\end{eqnarray}
In order to give the bulk profiles of those five dimensional fields conveniently, we introduce
a coordinate $t=\epsilon\exp(\sigma(\phi))$ which takes values between $t=\epsilon$
(UV brane) and $t=1$ (IR brane). In terms of the variable $t$, we write the bulk files
for those gauge fields as
\begin{eqnarray}
&&\chi_{_{(++)}}^G(y_{_{(++)}}^{G(n)},\phi)={t\Phi_{_{(uv)}}^{G(+)}(y_{_{(++)}}^{G(n)},t)\over
N_{_{(uv)}}^{G(+)}(y_{_{(++)}}^{G(n)})}={t\Phi_{_{(ir)}}^{G(+)}(y_{_{(++)}}^{G(n)},t)\over
N_{_{(ir)}}^{G(+)}(y_{_{(++)}}^{G(n)})}\;,
\nonumber\\
&&\chi_{_{(-+)}}^G(y_{_{(-+)}}^{G(n)},\phi)={t\Phi_{_{(uv)}}^{G(-)}(y_{_{(-+)}}^{G(n)},t)\over
N_{_{(uv)}}^{G(-)}(y_{_{(-+)}}^{G(n)})}={t\Phi_{_{(ir)}}^{G(+)}(y_{_{(-+)}}^{G(n)},t)\over
N_{_{(ir)}}^{G(+)}(y_{_{(-+)}}^{G(n)})}\;,
\label{gauge-5th}
\end{eqnarray}
with
\begin{eqnarray}
&&\Phi_{_{(uv)}}^{G(+)}(y,t)=Y_0(y\epsilon)J_1(yt)-J_0(y\epsilon)Y_1(yt)\;,
\nonumber\\
&&\Phi_{_{(ir)}}^{G(+)}(y,t)=Y_0(y)J_1(yt)-J_0(y)Y_1(yt)\;,
\nonumber\\
&&\Phi_{_{(uv)}}^{G(-)}(y,t)=Y_1(y\epsilon)J_1(yt)-J_1(y\epsilon)Y_1(yt)\;,
\nonumber\\
&&\Psi_{_{(uv)}}^{G(-)}(y,t)=Y_0(y\epsilon)J_0(yt)-J_0(y\epsilon)Y_0(yt)\;,
\nonumber\\
&&\Psi_{_{(ir)}}^{G(-)}(y,t)=Y_0(y)J_0(yt)-J_0(y)Y_0(yt)\;,
\nonumber\\
&&\Psi_{_{(uv)}}^{G(+)}(y,t)=Y_1(y\epsilon)J_0(yt)-J_1(y\epsilon)Y_0(yt)\;.
\label{Phi}
\end{eqnarray}
It is easy to check Eq.(\ref{Phi}) satisfying the equation of motions
\begin{eqnarray}
&&t{\partial\Phi_{_{(uv)}}^{G(+)}\over\partial t}(y,t)+\Phi_{_{(uv)}}^{G(+)}(y,t)
=yt\Psi_{_{(uv)}}^{G(-)}(y,t)\;,
\nonumber\\
&&t{\partial\Phi_{_{(ir)}}^{G(+)}\over\partial t}(y,t)+\Phi_{_{(ir)}}^{G(+)}(y,t)
=yt\Psi_{_{(ir)}}^{G(-)}(y,t)\;,
\nonumber\\
&&t{\partial\Phi_{_{(uv)}}^{G(-)}\over\partial t}(y,t)+\Phi_{_{(uv)}}^{G(-)}(y,t)
=yt\Psi_{_{(uv)}}^{G(+)}(y,t)\;,
\nonumber\\
&&t{\partial\Psi_{_{(uv)}}^{G(-)}\over\partial t}(y,t)
=-yt\Phi_{_{(uv)}}^{G(+)}(y,t)\;,
\nonumber\\
&&t{\partial\Psi_{_{(ir)}}^{G(-)}\over\partial t}(y,t)
=-yt\Phi_{_{(ir)}}^{G(+)}(y,t)\;,
\nonumber\\
&&t{\partial\Psi_{_{(uv)}}^{G(+)}\over\partial t}(y,t)
=-yt\Phi_{_{(uv)}}^{G(-)}(y,t)\;.
\label{EM-gauge}
\end{eqnarray}
The corresponding normalization constants are formulated as
\begin{eqnarray}
&&\Big[N_{_{(uv)}}^{G(+)}(y)\Big]^2={2\over kr}\Big\{\Big(\Phi_{_{(uv)}}^{G(+)}(y,1)\Big)^2
-\epsilon^2\Big(\Phi_{_{(uv)}}^{G(+)}(y,\epsilon)\Big)^2+\Big(\Psi_{_{(uv)}}^{G(-)}(y,1)\Big)^2
\nonumber\\
&&\hspace{2.6cm}
-{2\over y}\Phi_{_{(uv)}}^{G(+)}(y,1)\Psi_{_{(uv)}}^{G(-)}(y,1)\Big\}\;,
\nonumber\\
&&\Big[N_{_{(ir)}}^{G(+)}(y)\Big]^2={2\over kr}\Big\{\Big(\Phi_{_{(ir)}}^{G(+)}(y,1)\Big)^2
-\epsilon^2\Big(\Phi_{_{(ir)}}^{G(+)}(y,\epsilon)\Big)^2-\epsilon^2\Big(\Psi_{_{(ir)}}^{G(-)}(y,\epsilon)\Big)^2
\nonumber\\
&&\hspace{2.6cm}
+{2\epsilon\over y}\Phi_{_{(ir)}}^{G(+)}(y,\epsilon)\Psi_{_{(ir)}}^{G(-)}(y,\epsilon)\Big\}\;,
\nonumber\\
&&\Big[N_{_{(uv)}}^{G(-)}(y)\Big]^2={2\over kr}\Big\{\Big(\Psi_{_{(uv)}}^{G(+)}(y,1)\Big)^2
-\epsilon^2\Big(\Psi_{_{(uv)}}^{G(+)}(y,\epsilon)\Big)^2+\Big(\Phi_{_{(uv)}}^{G(-)}(y,1)\Big)^2
\nonumber\\
&&\hspace{2.6cm}
-{2\over y}\Psi_{_{(uv)}}^{G(+)}(y,1)\Phi_{_{(uv)}}^{G(-)}(y,1)\Big\}\;.
\label{normalization-gauge}
\end{eqnarray}
Similarly, the KK decompositions of five dimensional quark fields are written as
\begin{eqnarray}
&&\chi_{u_L}^i(x,\phi)={e^{2\sigma(\phi)}\over\sqrt{r}}\sum\limits_n
\chi_{u_L}^{i,(n)}(x)f_{_{(-+)}}^{L,c_B^i}(y_{_{(\mp\pm)}}^{c_B^i(n)},\phi),
\nonumber\\
&&\chi_{d_L}^i(x,\phi)={e^{2\sigma(\phi)}\over\sqrt{r}}\sum\limits_n
\chi_{d_L}^{i,(n)}(x)f_{_{(-+)}}^{L,c_B^i}(y_{_{(\mp\pm)}}^{c_B^i(n)},\phi),
\nonumber\\
&&q_{u_L}^i(x,\phi)={e^{2\sigma(\phi)}\over\sqrt{r}}\sum\limits_n
q_{u_L}^{i,(n)}(x)f_{_{(++)}}^{L,c_B^i}(y_{_{(\pm\pm)}}^{c_B^i(n)},\phi),
\nonumber\\
&&q_{d_L}^i(x,\phi)={e^{2\sigma(\phi)}\over\sqrt{r}}\sum\limits_n
q_{d_L}^{i,(n)}(x)f_{_{(++)}}^{L,c_B^i}(y_{_{(\pm\pm)}}^{c_B^i(n)},\phi),
\nonumber\\
&&u_R^i(x,\phi)={e^{2\sigma(\phi)}\over\sqrt{r}}\sum\limits_n
u_R^{i,(n)}(x)f_{_{(++)}}^{R,c_S^i}(y_{_{(\mp\mp)}}^{c_S^i(n)},\phi),
\nonumber\\
&&X_R^i(x,\phi)={e^{2\sigma(\phi)}\over\sqrt{r}}\sum\limits_n
X_R^{i,(n)}(x)f_{_{(-+)}}^{R,c_T^i}(y_{_{(\pm\mp)}}^{c_T^i(n)},\phi),
\nonumber\\
&&U_R^i(x,\phi)={e^{2\sigma(\phi)}\over\sqrt{r}}\sum\limits_n
U_R^{i,(n)}(x)f_{_{(-+)}}^{R,c_T^i}(y_{_{(\pm\mp)}}^{c_T^i(n)},\phi),
\nonumber\\
&&D_R^i(x,\phi)={e^{2\sigma(\phi)}\over\sqrt{r}}\sum\limits_n
D_R^{i,(n)}(x)f_{_{(-+)}}^{R,c_T^i}(y_{_{(\pm\mp)}}^{c_T^i(n)},\phi),
\nonumber\\
&&\tilde{X}_R^i(x,\phi)={e^{2\sigma(\phi)}\over\sqrt{r}}\sum\limits_n
\tilde{X}_R^{i,(n)}(x)f_{_{(-+)}}^{R,c_T^i}(y_{_{(\pm\mp)}}^{c_T^i(n)},\phi),
\nonumber\\
&&\tilde{U}_R^i(x,\phi)={e^{2\sigma(\phi)}\over\sqrt{r}}\sum\limits_n
\tilde{U}_R^{i,(n)}(x)f_{_{(-+)}}^{R,c_T^i}(y_{_{(\pm\mp)}}^{c_T^i(n)},\phi),
\nonumber\\
&&d_R^i(x,\phi)={e^{2\sigma(\phi)}\over\sqrt{r}}\sum\limits_n
d_R^{i,(n)}(x)f_{_{(++)}}^{R,c_T^i}(y_{_{(\mp\mp)}}^{c_T^i(n)},\phi),
\nonumber\\
&&X_L^i(x,\phi)={e^{2\sigma(\phi)}\over\sqrt{r}}\sum\limits_n
X_L^{i,(n)}(x)f_{_{(+-)}}^{L,c_T^i}(y_{_{(\pm\mp)}}^{c_T^i(n)},\phi),
\nonumber\\
&&U_L^i(x,\phi)={e^{2\sigma(\phi)}\over\sqrt{r}}\sum\limits_n
U_L^{i,(n)}(x)f_{_{(+-)}}^{L,c_T^i}(y_{_{(\pm\mp)}}^{c_T^i(n)},\phi),
\nonumber\\
&&D_L^i(x,\phi)={e^{2\sigma(\phi)}\over\sqrt{r}}\sum\limits_n
D_L^{i,(n)}(x)f_{_{(+-)}}^{L,c_T^i}(y_{_{(\pm\mp)}}^{c_T^i(n)},\phi),
\nonumber\\
&&\tilde{X}_L^i(x,\phi)={e^{2\sigma(\phi)}\over\sqrt{r}}\sum\limits_n
\tilde{X}_L^{i,(n)}(x)f_{_{(+-)}}^{L,c_T^i}(y_{_{(\pm\mp)}}^{c_T^i(n)},\phi),
\nonumber\\
&&\tilde{U}_L^i(x,\phi)={e^{2\sigma(\phi)}\over\sqrt{r}}\sum\limits_n
\tilde{U}_L^{i,(n)}(x)f_{_{(+-)}}^{L,c_T^i}(y_{_{(\pm\mp)}}^{c_T^i(n)},\phi),
\nonumber\\
&&d_L^i(x,\phi)={e^{2\sigma(\phi)}\over\sqrt{r}}\sum\limits_n
d_L^{i,(n)}(x)f_{_{(--)}}^{L,c_T^i}(y_{_{(\mp\mp)}}^{c_T^i(n)},\phi),
\nonumber\\
&&u_L^i(x,\phi)={e^{2\sigma(\phi)}\over\sqrt{r}}\sum\limits_n
u_L^{i,(n)}(x)f_{_{(--)}}^{L,c_S^i}(y_{_{(\mp\mp)}}^{c_S^i(n)},\phi),
\nonumber\\
&&\chi_{u_R}^i(x,\phi)={e^{2\sigma(\phi)}\over\sqrt{r}}\sum\limits_n
\chi_{u_R}^{i,(n)}(x)f_{_{(+-)}}^{R,c_B^i}(y_{_{(\mp\pm)}}^{c_B^i(n)},\phi),
\nonumber\\
&&\chi_{d_R}^i(x,\phi)={e^{2\sigma(\phi)}\over\sqrt{r}}\sum\limits_n
\chi_{d_R}^{i,(n)}(x)f_{_{(+-)}}^{R,c_B^i}(y_{_{(\mp\pm)}}^{c_B^i(n)},\phi),
\nonumber\\
&&q_{u_R}^i(x,\phi)={e^{2\sigma(\phi)}\over\sqrt{r}}\sum\limits_n
q_{u_R}^{i,(n)}(x)f_{_{(--)}}^{R,c_B^i}(y_{_{(\pm\pm)}}^{c_B^i(n)},\phi),
\nonumber\\
&&q_{d_R}^i(x,\phi)={e^{2\sigma(\phi)}\over\sqrt{r}}\sum\limits_n
q_{d_R}^{i,(n)}(x)f_{_{(--)}}^{R,c_B^i}(y_{_{(\pm\pm)}}^{c_B^i(n)},\phi).
\label{KK-decomposition-quark}
\end{eqnarray}
In Eq.(\ref{KK-decomposition-quark}), the eigenvalues $y_{_{(\pm\pm)}}^{c(n)}\;(n\ge1)$
satisfy the equation $R_{_{(\pm\pm)}}^{c,\epsilon}(z)\equiv0$,
$y_{_{(\mp\mp)}}^{c(n)}\;(n\ge1)$ satisfy the equation $R_{_{(\mp\mp)}}^{c,\epsilon}(z)\equiv0$,
$y_{_{(\pm\mp)}}^{c(n)}\;(n\ge1)$ satisfy the equation
$R_{_{(\pm\mp)}}^{c,\epsilon}(z)\equiv0$, and the eigenvalues $y_{_{(\mp\pm)}}^{c(n)}\;(n\ge1)$
satisfy the equation $R_{_{(\mp\pm)}}^{c,\epsilon}(z)\equiv0$, respectively.
Here, the concrete expressions of $R_{_{(\pm\pm)}}^{c,\epsilon}(z),\;R_{_{(\pm\mp)}}^{c,\epsilon}(z)
\;,R_{_{(\mp\pm)}}^{c,\epsilon}(z),\;R_{_{(\mp\mp)}}^{c,\epsilon}(z)$ are
\begin{eqnarray}
&&R_{_{(\pm\pm)}}^{c,\epsilon}(z)=\left\{\begin{array}{ll}Y_{N}(z)J_{N}(z\epsilon)-J_{N}(z)Y_{N}(z\epsilon),
&c=N+{1\over2}\\
J_{-c+{1\over2}}(z)J_{c-{1\over2}}(z\epsilon)-J_{c-{1\over2}}(z)J_{-c+{1\over2}}(z\epsilon),
&c\neq N+{1\over2}\end{array}\right.\;,
\nonumber\\
&&R_{_{(\pm\mp)}}^{c,\epsilon}(z)=\left\{\begin{array}{ll}J_{N+1}(z)Y_{N}(z\epsilon)-Y_{N+1}(z)J_{N}(z\epsilon),
&c=N+{1\over2}\\
J_{c+{1\over2}}(z)J_{-c+{1\over2}}(z\epsilon)+J_{-c-{1\over2}}(z)J_{c-{1\over2}}(z\epsilon),
&c\neq N+{1\over2}\end{array}\right.\;,
\nonumber\\
&&R_{_{(\mp\pm)}}^{c,\epsilon}(z)=\left\{\begin{array}{ll}Y_{N}(z)J_{N+1}(z\epsilon)-J_{N}(z)Y_{N+1}(z\epsilon),
&c=N+{1\over2}\\
J_{-c+{1\over2}}(z)J_{c+{1\over2}}(z\epsilon)+J_{c-{1\over2}}(z)J_{-c-{1\over2}}(z\epsilon),
&c\neq N+{1\over2}\end{array}\right.\;,
\nonumber\\
&&R_{_{(\mp\mp)}}^{c,\epsilon}(z)=\left\{\begin{array}{ll}J_{N+1}(z)Y_{N+1}(z\epsilon)
-Y_{N+1}(z)J_{N+1}(z\epsilon),&c=N+{1\over2}\\
J_{c+{1\over2}}(z)J_{-c-{1\over2}}(z\epsilon)-J_{-c-{1\over2}}(z)J_{c+{1\over2}}(z\epsilon),
&c\neq N+{1\over2}\end{array}\right.\;.
\label{root-quark}
\end{eqnarray}
In terms of the variable $t$, those bulk profiles for five dimensional fermions are formulated as
\begin{eqnarray}
&&f_{_{(++)}}^{L,c}(y_{_{(\pm\pm)}}^{c(n)},\phi)={\sqrt{t}\;\varphi_{_{L_{(uv)}}}^{c(+)}
(y_{_{(\pm\pm)}}^{c(n)},t)\over N_{_{L_{(uv)}}}^{c(+)}(y_{_{(\pm\pm)}}^{c(n)})}
={\sqrt{t}\;\varphi_{_{L_{(ir)}}}^{c(+)}(y_{_{(\pm\pm)}}^{c(n)},t)
\over N_{_{L_{(ir)}}}^{c(+)}(y_{_{(\pm\pm)}}^{c(n)})}\;,
\nonumber\\
&&f_{_{(+-)}}^{L,c}(y_{_{(\pm\mp)}}^{c(n)},\phi)={\sqrt{t}\;\varphi_{_{L_{(uv)}}}^{c(+)}
(y_{_{(\pm\mp)}}^{c(n)},t)\over N_{_{L_{(uv)}}}^{c(+)}(y_{_{(\pm\mp)}}^{c(n)})}
={\sqrt{t}\;\varphi_{_{L_{(ir)}}}^{c(-)}(y_{_{(\pm\mp)}}^{c(n)},t)\over
N_{_{L_{(ir)}}}^{c(-)}(y_{_{(\pm\mp)}}^{c(n)})}\;,
\nonumber\\
&&f_{_{(-+)}}^{L,c}(y_{_{(\mp\pm)}}^{c(n)},\phi)={\sqrt{t}\;\varphi_{_{L_{(uv)}}}^{c(-)}
(y_{_{(\mp\pm)}}^{c(n)},t)\over N_{_{L_{(uv)}}}^{c(-)}(y_{_{(\mp\pm)}}^{c(n)})}
={\sqrt{t}\;\varphi_{_{L_{(ir)}}}^{c(+)}(y_{_{(\mp\pm)}}^{c(n)},t)\over
N_{_{L_{(ir)}}}^{c(+)}(y_{_{(\mp\pm)}}^{c(n)})}\;,
\nonumber\\
&&f_{_{(--)}}^{L,c}(y_{_{(\mp\mp)}}^{c(n)},\phi)={\sqrt{t}\;\varphi_{_{L_{(uv)}}}^{c(-)}
(y_{_{(\mp\mp)}}^{c(n)},t)\over N_{_{L_{(uv)}}}^{c(-)}(y_{_{(\mp\mp)}}^{c(n)})}
={\sqrt{t}\;\varphi_{_{L_{(ir)}}}^{c(-)}(y_{_{(\mp\mp)}}^{c(n)},t)\over
N_{_{L_{(ir)}}}^{c(-)}(y_{_{(\mp\mp)}}^{c(n)})}\;,
\nonumber\\
&&f_{_{(++)}}^{R,c}(y_{_{(\mp\mp)}}^{c(n)},\phi)={\sqrt{t}\;\varphi_{_{R_{(uv)}}}^{c(+)}
(y_{_{(\mp\mp)}}^{c(n)},t)\over N_{_{R_{(uv)}}}^{c(+)}(y_{_{(\mp\mp)}}^{c(n)})}
={\sqrt{t}\;\varphi_{_{R_{(ir)}}}^{c(+)}(y_{_{(\mp\mp)}}^{c(n)},t)\over
N_{_{R_{(ir)}}}^{c(+)}(y_{_{(\mp\mp)}}^{c(n)})}\;,
\nonumber\\
&&f_{_{(+-)}}^{R,c}(y_{_{(\mp\pm)}}^{c(n)},\phi)={\sqrt{t}\;\varphi_{_{R_{(uv)}}}^{c(+)}
(y_{_{(\mp\pm)}}^{c(n)},t)\over N_{_{R_{(uv)}}}^{c(+)}(y_{_{(\mp\pm)}}^{c(n)})}
={\sqrt{t}\;\varphi_{_{R_{(ir)}}}^{c(-)}(y_{_{(\mp\pm)}}^{c(n)},t)\over
N_{_{R_{(ir)}}}^{c(-)}(y_{_{(\mp\pm)}}^{c(n)})}\;,
\nonumber\\
&&f_{_{(-+)}}^{R,c}(y_{_{(\pm\mp)}}^{c(n)},\phi)={\sqrt{t}\;\varphi_{_{R_{(uv)}}}^{c(-)}
(y_{_{(\pm\mp)}}^{c(n)},t)\over N_{_{R_{(uv)}}}^{c(-)}(y_{_{(\pm\mp)}}^{c(n)})}
={\sqrt{t}\;\varphi_{_{R_{(ir)}}}^{c(+)}(y_{_{(\pm\mp)}}^{c(n)},t)\over
N_{_{R_{(ir)}}}^{c(+)}(y_{_{(\pm\mp)}}^{c(n)})}\;,
\nonumber\\
&&f_{_{(--)}}^{R,c}(y_{_{(\pm\pm)}}^{c(n)},\phi)={\sqrt{t}\;\varphi_{_{R_{(uv)}}}^{c(-)}
(y_{_{(\pm\pm)}}^{c(n)},t)\over N_{_{R_{(uv)}}}^{c(-)}(y_{_{(\pm\pm)}}^{c(n)})}
={\sqrt{t}\;\varphi_{_{R_{(ir)}}}^{c(-)}(y_{_{(\pm\pm)}}^{c(n)},t)\over
N_{_{R_{(ir)}}}^{c(-)}(y_{_{(\pm\pm)}}^{c(n)})}
\label{quark-5th}
\end{eqnarray}
with
\begin{eqnarray}
&&\varphi_{_{L_{(ir)}}}^{c(+)}(y,t)=\left\{\begin{array}{ll}Y_{N}(y)J_{N+1}(yt)-J_{N}(y)Y_{N+1}(yt),
&c=N+{1\over2}\\
J_{-c+{1\over2}}(y)J_{c+{1\over2}}(yt)+J_{c-{1\over2}}(y)J_{-c-{1\over2}}(yt),
&c\neq N+{1\over2}\end{array}\right.\;,
\nonumber\\
&&\varphi_{_{L_{(uv)}}}^{c(+)}(y,t)=\left\{\begin{array}{ll}Y_{N}(y\epsilon)J_{N+1}(yt)-J_{N}(y\epsilon)Y_{N+1}(yt),
&c=N+{1\over2}\\
J_{-c+{1\over2}}(y\epsilon)J_{c+{1\over2}}(yt)+J_{c-{1\over2}}(y\epsilon)J_{-c-{1\over2}}(yt),
&c\neq N+{1\over2}\end{array}\right.\;,
\nonumber\\
&&\varphi_{_{L_{(ir)}}}^{c(-)}(y,t)=\left\{\begin{array}{ll}Y_{N+1}(y)J_{N+1}(yt)-J_{N+1}(y)Y_{N+1}(yt),
&c=N+{1\over2}\\
J_{-c-{1\over2}}(y)J_{c+{1\over2}}(yt)-J_{c+{1\over2}}(y)J_{-c-{1\over2}}(yt),
&c\neq N+{1\over2}\end{array}\right.\;,
\nonumber\\
&&\varphi_{_{L_{(uv)}}}^{c(-)}(y,t)=\left\{\begin{array}{ll}Y_{N+1}(y\epsilon)J_{N+1}(yt)-J_{N+1}(y\epsilon)Y_{N+1}(yt),
&c=N+{1\over2}\\
J_{-c-{1\over2}}(y\epsilon)J_{c+{1\over2}}(yt)-J_{c+{1\over2}}(y\epsilon)J_{-c-{1\over2}}(yt),
&c\neq N+{1\over2}\end{array}\right.\;,
\nonumber\\
&&\varphi_{_{R_{(ir)}}}^{c(-)}(y,t)=\left\{\begin{array}{ll}Y_{N}(y)J_{N}(yt)-J_{N}(y)Y_{N}(yt),
&c=N+{1\over2}\\
J_{-c+{1\over2}}(y)J_{c-{1\over2}}(yt)-J_{c-{1\over2}}(y)J_{-c+{1\over2}}(yt),
&c\neq N+{1\over2}\end{array}\right.\;,
\nonumber\\
&&\varphi_{_{R_{(uv)}}}^{c(-)}(y,t)=\left\{\begin{array}{ll}Y_{N}(y\epsilon)J_{N}(yt)-J_{N}(y\epsilon)Y_{N}(yt),
&c=N+{1\over2}\\
J_{-c+{1\over2}}(y\epsilon)J_{c-{1\over2}}(yt)-J_{c-{1\over2}}(y\epsilon)J_{-c+{1\over2}}(yt),
&c\neq N+{1\over2}\end{array}\right.\;,
\nonumber\\
&&\varphi_{_{R_{(ir)}}}^{c(+)}(y,t)=\left\{\begin{array}{ll}Y_{N+1}(y)J_{N}(yt)-J_{N+1}(y)Y_{N}(yt),
&c=N+{1\over2}\\
J_{-c-{1\over2}}(y)J_{c-{1\over2}}(yt)+J_{c+{1\over2}}(y)J_{-c+{1\over2}}(yt),
&c\neq N+{1\over2}\end{array}\right.\;,
\nonumber\\
&&\varphi_{_{R_{(uv)}}}^{c(+)}(y,t)=\left\{\begin{array}{ll}Y_{N+1}(y\epsilon)J_{N}(yt)-J_{N+1}(y\epsilon)Y_{N}(yt),
&c=N+{1\over2}\\
J_{-c-{1\over2}}(y\epsilon)J_{c-{1\over2}}(yt)+J_{c+{1\over2}}(y\epsilon)J_{-c+{1\over2}}(yt),
&c\neq N+{1\over2}\end{array}\right.\;.
\label{varphi-i}
\end{eqnarray}
It should pointed out that those bulk profiles satisfy the following equations of motion
\begin{eqnarray}
&&t{\partial\varphi_{_{L_{(ir)}}}^{c(\pm)}\over\partial t}(y,t)+\Big(c+{1\over2}\Big)
\varphi_{_{L_{(ir)}}}^{c(\pm)}(y,t)=yt\varphi_{_{R_{(ir)}}}^{c(\mp)}(y,t)\;,
\nonumber\\
&&t{\partial\varphi_{_{R_{(ir)}}}^{c(\pm)}\over\partial t}(z,t)-\Big(c-{1\over2}\Big)
\varphi_{_{R_{(ir)}}}^{c(\pm)}(y,t)=-yt\varphi_{_{L_{(ir)}}}^{c(\mp)}(y,t)\;,
\nonumber\\
&&t{\partial\varphi_{_{L_{(uv)}}}^{c(\pm)}\over\partial t}(y,t)+\Big(c+{1\over2}\Big)
\varphi_{_{L_{(uv)}}}^{c(\pm)}(y,t)=yt\varphi_{_{R_{(uv)}}}^{c(\mp)}(y,t)\;,
\nonumber\\
&&t{\partial\varphi_{_{R_{(uv)}}}^{c(\pm)}\over\partial t}(y,t)-\Big(c-{1\over2}\Big)
\varphi_{_{R_{(uv)}}}^{c(\pm)}(y,t)=-yt\varphi_{_{L_{(uv)}}}^{c(\mp)}(y,t)\;.
\label{EM-fermion}
\end{eqnarray}
Meanwhile, the normalization factors can be written as
\begin{eqnarray}
&&\Big[N_{_{L_{(ir)}}}^{c(+)}(y)\Big]^2={2\over kr\epsilon}\Big\{\Big(\varphi_{_{L_{(ir)}}}^{c(+)}(y,1)\Big)^2
-\epsilon^2\Big(\varphi_{_{L_{(ir)}}}^{c(+)}(y,\epsilon)\Big)^2-\epsilon^2\Big(\varphi_{_{R_{(ir)}}}^{c(-)}(y,\epsilon)\Big)^2
\nonumber\\
&&\hspace{2.5cm}
+(2c+1){\epsilon\over y}\varphi_{_{L_{(ir)}}}^{c(+)}(y,\epsilon)\varphi_{_{R_{(ir)}}}^{c(-)}(y,\epsilon)\Big\}\;,
\nonumber\\
&&\Big[N_{_{R_{(ir)}}}^{c(-)}(y)\Big]^2={2\over kr\epsilon}\Big\{\Big(\varphi_{_{L_{(ir)}}}^{c(+)}(y,1)\Big)^2
-\epsilon^2\Big(\varphi_{_{L_{(ir)}}}^{c(+)}(y,\epsilon)\Big)^2-\epsilon^2\Big(\varphi_{_{R_{(ir)}}}^{c(-)}(y,\epsilon)\Big)^2
\nonumber\\
&&\hspace{2.5cm}
+(2c-1){\epsilon\over y}\varphi_{_{L_{(ir)}}}^{c(+)}(y,\epsilon)\varphi_{_{R_{(ir)}}}^{c(-)}(y,\epsilon)\Big\}\;,
\nonumber\\
&&\Big[N_{_{L_{(ir)}}}^{c(-)}(y)\Big]^2={2\over kr\epsilon}\Big\{
-\epsilon^2\Big(\varphi_{_{L_{(ir)}}}^{c(-)}(y,\epsilon)\Big)^2
+\Big(\varphi_{_{R_{(ir)}}}^{c(+)}(y,1)\Big)^2-\epsilon^2\Big(\varphi_{_{R_{(ir)}}}^{c(+)}(y,\epsilon)\Big)^2
\nonumber\\
&&\hspace{2.5cm}
+(2c+1){\epsilon\over y}\varphi_{_{L_{(ir)}}}^{c(-)}(y,\epsilon)\varphi_{_{R_{(ir)}}}^{c(+)}(y,\epsilon)\Big\}\;,
\nonumber\\
&&\Big[N_{_{R_{(ir)}}}^{c(+)}(y)\Big]^2={2\over kr\epsilon}\Big\{
-\epsilon^2\Big(\varphi_{_{L_{(ir)}}}^{c(-)}(y,\epsilon)\Big)^2
+\Big(\varphi_{_{R_{(ir)}}}^{c(+)}(y,1)\Big)^2-\epsilon^2\Big(\varphi_{_{R_{(ir)}}}^{c(+)}(y,\epsilon)\Big)^2
\nonumber\\
&&\hspace{2.5cm}
+(2c-1){\epsilon\over y}\varphi_{_{L_{(ir)}}}^{c(-)}(y,\epsilon)\varphi_{_{R_{(ir)}}}^{c(+)}(y,\epsilon)\Big\}\;,
\nonumber\\
&&\Big[N_{_{L_{(uv)}}}^{c(+)}(y)\Big]^2={2\over kr\epsilon}\Big\{\Big(\varphi_{_{L_{(uv)}}}^{c(+)}(y,1)\Big)^2
-\epsilon^2\Big(\varphi_{_{L_{(uv)}}}^{c(+)}(y,\epsilon)\Big)^2+\Big(\varphi_{_{R_{(uv)}}}^{c(-)}(y,1)\Big)^2
\nonumber\\
&&\hspace{2.5cm}
-{2c+1\over y}\varphi_{_{L_{(uv)}}}^{c(+)}(y,1)\varphi_{_{R_{(uv)}}}^{c(-)}(y,1)\Big\}\;,
\nonumber\\
&&\Big[N_{_{R_{(uv)}}}^{c(-)}(y)\Big]^2={2\over kr\epsilon}\Big\{\Big(\varphi_{_{L_{(uv)}}}^{c(+)}(y,1)\Big)^2
-\epsilon^2\Big(\varphi_{_{L_{(uv)}}}^{c(+)}(y,\epsilon)\Big)^2+\Big(\varphi_{_{R_{(uv)}}}^{c(-)}(y,1)\Big)^2
\nonumber\\
&&\hspace{2.5cm}
-{2c-1\over y}\varphi_{_{L_{(uv)}}}^{c(+)}(y,1)\varphi_{_{R_{(uv)}}}^{c(-)}(y,1)\Big\}\;,
\nonumber\\
&&\Big[N_{_{L_{(uv)}}}^{c(-)}(y)\Big]^2={2\over kr\epsilon}\Big\{
\Big(\varphi_{_{L_{(uv)}}}^{c(-)}(y,1)\Big)^2
+\Big(\varphi_{_{R_{(uv)}}}^{c(+)}(y,1)\Big)^2-\epsilon^2\Big(\varphi_{_{R_{(uv)}}}^{c(+)}(y,\epsilon)\Big)^2
\nonumber\\
&&\hspace{2.5cm}
-{2c+1\over y}\varphi_{_{L_{(uv)}}}^{c(-)}(y,1)\varphi_{_{R_{(uv)}}}^{c(+)}(y,1)\Big\}\;,
\nonumber\\
&&\Big[N_{_{R_{(uv)}}}^{c(+)}(y)\Big]^2={2\over kr\epsilon}\Big\{
\Big(\varphi_{_{L_{(uv)}}}^{c(-)}(y,1)\Big)^2
+\Big(\varphi_{_{R_{(uv)}}}^{c(+)}(y,1)\Big)^2-\epsilon^2\Big(\varphi_{_{R_{(uv)}}}^{c(+)}(y,\epsilon)\Big)^2
\nonumber\\
&&\hspace{2.5cm}
-{2c-1\over y}\varphi_{_{L_{(uv)}}}^{c(-)}(y,1)\varphi_{_{R_{(uv)}}}^{c(+)}(y,1)\Big\}\;.
\label{normalization-quark}
\end{eqnarray}
With the preparations above, we verify some lemmas on the
eigenvalues of KK modes in the framework of warped extra dimension,
then show how to sum over the infinite series of KK modes using the residue theorem.
\section{Summing over infinite series of KK modes\label{sec3}}
\indent\indent
As mentioned above, the radiative corrections from all virtual KK modes to the physics
quantities at electroweak scale should be summed over in principle in order to obtain the
theoretical predictions in extensions of the SM with a warped or unified extra dimension.
To sum over the infinite series $\sum\limits_{n=1}^\infty(f(y_{_{(BCs)}}^{c(n)})+f(y_{_{(BCs)}}^{c(-n)}))$
where $f(z)$ is analytic function except some limited in number nonzero poles
$z_1,\;z_2,\;\cdots,\;z_{n_0}$ and the possible pole $z=0$, we construct another function $G(z)$ which
is analytic except its poles of order one $y_{_{(BCs)}}^{c(\pm1)},\;\cdots,\;
y_{_{(BCs)}}^{c(\pm n)},\cdots$, and the possible pole of order $m$: $z=0\;\;(m\ge1)$. Additional,
the residues of $G(z)$ are uniform one at nonzero poles $z=\;y_{_{(BCs)}}^{c(\pm1)},\;\cdots,\;
y_{_{(BCs)}}^{c(\pm n)},\cdots$. If the closed rectifiable contour
$C_{_{r_{(n)}}}$ does not pass through any of the points $y_{_{(BCs)}}^{c(\pm k)}\;(k=1,\;2,\;\cdots,\;n)$,
and the region with the boundary $C_{_{r_{(n)}}}$ contains the points
$0,\;y_{_{(BCs)}}^{c(\pm1)},\;\cdots,\;y_{_{(BCs)}}^{c(\pm n)}$ and $z_1,\;z_2,\;\cdots,\;z_{n_0}$.
According the residue theorem\cite{textbook}, we obtain
\begin{eqnarray}
&&\oint_{C_{_{r_{(n)}}}}G(z)f(z)dz=i2\pi\Big\{\sum\limits_{i=1}^n\Big[
f(y_{_{(BCs)}}^{c(i)})+f(y_{_{(BCs)}}^{c(-i)})\Big]
\nonumber\\
&&\hspace{3.6cm}
+{\rm Res}\Big(G(z)f(z),z=0\Big)+\sum\limits_{i=1}^{n_0}{\rm Res}\Big(G(z)f(z),z=z_i\Big)\Big\}\;.
\label{residue1}
\end{eqnarray}
If the limit $$\lim\limits_{n\rightarrow\infty}\oint_{C_{_{r_{(n)}}}}G(z)f(z)dz=0\;,$$
then we can sum over the infinite series
\begin{eqnarray}
\sum\limits_{i=1}^\infty\Big[f(y_{_{(BCs)}}^{c(i)})
+f(y_{_{(BCs)}}^{c(-i)})\Big]=-{\rm Res}\Big(G(z)f(z),z=0\Big)-\sum\limits_{i=1}^{n_0}{\rm Res}\Big(G(z)f(z),z=z_i\Big)\;.
\label{residue1aa}
\end{eqnarray}
To construct the function $G(z)$ and find the suitable contour $C_{_{r_{(n)}}}$,
we verify some lemmas on the eigenvalues of KK modes firstly.
{\it Lemma1: If $y_{_{\rm(BCs)}}^{c(n)}(n=1,\;2,\;\cdots,\;\infty)$ satisfy the equation
$R_{_{\rm(BCs)}}^{c,\epsilon}(y_{_{\rm(BCs)}}^{c(n)})=0$ where $c$ denotes the bulk mass
of five dimensional fermions, then $y_{_{\rm(BCs)}}^{c(n)}$ is real.}
Proof. Taking ${\rm(BCs)}=(\pm\pm)$ as an example, we show how to demonstrate
that the roots of $R_{_{\pm\pm}}^{c,\epsilon}(y_{_{\pm\pm}}^{c(n)})=0$ are real.
If the left-handed five dimensional fermion satisfies the BCs $(++)$ and the dual
right-handed five dimensional fermion satisfies the BCs $(--)$, the corresponding bulk files
are written as
\begin{eqnarray}
&&f_{_{(++)}}^{L,c}(y_{_{(\pm\pm)}}^{c(n)},\phi)={\sqrt{t}\varphi_{_{L_{(ir)}}}^{c(+)}
(y_{_{(\pm\pm)}}^{c(n)},t)\over N_{_{L_{(ir)}}}^{c(+)}(y_{_{(\pm\pm)}}^{c(n)})}
={\sqrt{t}\varphi_{_{L_{(uv)}}}^{c(+)}
(y_{_{(\pm\pm)}}^{c(n)},t)\over N_{_{L_{(uv)}}}^{c(+)}(y_{_{(\pm\pm)}}^{c(n)})}\;,
\nonumber\\
&&f_{_{(--)}}^{R,c}(y_{_{(\pm\pm)}}^{c(n)},\phi)={\sqrt{t}\varphi_{_{R_{(ir)}}}^{c(-)}
(y_{_{(\pm\pm)}}^{c(n)},t)\over N_{_{R_{(ir)}}}^{c(-)}(y_{_{(\pm\pm)}}^{c(n)})}
={\sqrt{t}\varphi_{_{R_{(uv)}}}^{c(-)}
(y_{_{(\pm\pm)}}^{c(n)},t)\over N_{_{R_{(uv)}}}^{c(-)}(y_{_{(\pm\pm)}}^{c(n)})}\;,
\label{lemma1a}
\end{eqnarray}
where $\varphi_{_{L_{(ir)}}}^{c(+)}(y_{_{(\pm\pm)}}^{c(n)},t),\;\varphi_{_{L_{(uv)}}}^{c(+)}(y_{_{(\pm\pm)}}^{c(n)},t),
\;\varphi_{_{R_{(ir)}}}^{c(-)}(y_{_{(\pm\pm)}}^{c(n)},t),\;\varphi_{_{R_{(uv)}}}^{c(-)}(y_{_{(\pm\pm)}}^{c(n)},t)$
satisfy the equations of motion in Eq.(\ref{EM-fermion}).
Using Eq.(\ref{EM-fermion}), we can derive easily the following equations
\begin{eqnarray}
&&{1\over t}{\partial\over\partial t}\Big(t{\partial\varphi_{_{L_{(ir)}}}^{c(+)}\over\partial t}\Big)
(y_{_{(\pm\pm)}}^{c(n)},t)+\bigg(\Big[y_{_{(\pm\pm)}}^{c(n)}\Big]^2
-{(c+{1\over2})^2\over t^2}\bigg)\varphi_{_{L_{(ir)}}}^{c(+)}(y_{_{(\pm\pm)}}^{c(n)},t)=0
\;,\nonumber\\
&&{1\over t}{\partial\over\partial t}\Big(t{\partial\varphi_{_{R_{(ir)}}}^{c(-)}\over\partial t}\Big)
(y_{_{(\pm\pm)}}^{c(n)},t)+\bigg(\Big[y_{_{(\pm\pm)}}^{c(n)}\Big]^2
-{(c-{1\over2})^2\over t^2}\bigg)\varphi_{_{R_{(ir)}}}^{c(-)}(y_{_{(\pm\pm)}}^{c(n)},t)=0\;.
\label{lemma1c}
\end{eqnarray}
If $y_{_{(\pm\pm)}}^{c(n)}=r+is$ ($r,s$ are real and $s\neq0$) satisfies
$R_{_{(\pm\pm)}}^{c,\epsilon}(y_{_{(\pm\pm)}}^{c(n)})=0$, then its conjugate
$\overline{y}_{_{(\pm\pm)}}^{c(n)}=r-is$ also satisfies $R_{_{(\pm\pm)}}^{c,\epsilon}
(\overline{y}_{_{(\pm\pm)}}^{c(n)})=0$. This implies
\begin{eqnarray}
&&{1\over t}{\partial\over\partial t}\Big(t{\partial\varphi_{_{L_{(ir)}}}^{c(+)}\over\partial t}\Big)
(\overline{y}_{_{(\pm\pm)}}^{c(n)},t)+\bigg(\Big[\overline{y}_{_{(\pm\pm)}}^{c(n)}\Big]^2
-{(c+{1\over2})^2\over t^2}\bigg)\varphi_{_{L_{(ir)}}}^{c(+)}(\overline{y}_{_{(\pm\pm)}}^{c(n)},t)=0\;.
\label{lemma1d}
\end{eqnarray}
Using the first equation in Eq(\ref{lemma1c}) and that in Eq(\ref{lemma1d}), we find
\begin{eqnarray}
&&\bigg(\Big[y_{_{(\pm\pm)}}^{c(n)}\Big]^2-\Big[\overline{y}_{_{(\pm\pm)}}^{c(n)}\Big]^2\bigg)
\int_\epsilon^1dtt\varphi_{_{L_{(ir)}}}^{c(+)}(y_{_{(\pm\pm)}}^{c(n)},t)
\varphi_{_{L_{(ir)}}}^{c(+)}(\overline{y}_{_{(\pm\pm)}}^{c(n)},t)
\nonumber\\
&&\hspace{-0.6cm}=
\bigg\{t\varphi_{_{L_{(ir)}}}^{c(+)}(y_{_{(\pm\pm)}}^{c(n)},t){\partial\varphi_{_{L_{(ir)}}}^{c(+)}\over\partial t}
(\overline{y}_{_{(\pm\pm)}}^{c(n)},t)-t\varphi_{_{L_{(ir)}}}^{c(+)}(\overline{y}_{_{(\pm\pm)}}^{c(n)},t)
{\partial\varphi_{_{L_{(ir)}}}^{c(+)}\over\partial t}(y_{_{(\pm\pm)}}^{c(n)},t)\bigg\}_\epsilon^1
\nonumber\\
&&\hspace{-0.6cm}=
\bigg\{t\overline{y}_{_{(\pm\pm)}}^{c(n)}\varphi_{_{L_{(ir)}}}^{c(+)}(y_{_{(\pm\pm)}}^{c(n)},t)
\varphi_{_{R_{(ir)}}}^{c(-)}
(\overline{y}_{_{(\pm\pm)}}^{c(n)},t)-t\overline{y}_{_{(\pm\pm)}}^{c(n)}
\varphi_{_{L_{(ir)}}}^{c(+)}(\overline{y}_{_{(\pm\pm)}}^{c(n)},t)
\varphi_{_{R_{(ir)}}}^{c(-)}(y_{_{(\pm\pm)}}^{c(n)},t)\bigg\}_\epsilon^1\;.
\label{lemma1e}
\end{eqnarray}
Similarly, we also obtain
\begin{eqnarray}
&&\bigg(\Big[y_{_{(\pm\pm)}}^{c(n)}\Big]^2-\Big[\overline{y}_{_{(\pm\pm)}}^{c(n)}\Big]^2\bigg)
\int_\epsilon^1dtt\varphi_{_{R_{(ir)}}}^{c(-)}(y_{_{(\pm\pm)}}^{c(n)},t)
\varphi_{_{R_{(ir)}}}^{c(-)}(\overline{y}_{_{(\pm\pm)}}^{c(n)},t)
\nonumber\\
&&\hspace{-0.6cm}=
\bigg\{-t\overline{y}_{_{(\pm\pm)}}^{c(n)}\varphi_{_{R_{(ir)}}}^{c(-)}(y_{_{(\pm\pm)}}^{c(n)},t)
\varphi_{_{L_{(ir)}}}^{c(+)}(\overline{y}_{_{(\pm\pm)}}^{c(n)},t)+t\overline{y}_{_{(\pm\pm)}}^{c(n)}
\varphi_{_{R_{(ir)}}}^{c(-)}(\overline{y}_{_{(\pm\pm)}}^{c(n)},t)
\varphi_{_{L_{(ir)}}}^{c(+)}(y_{_{(\pm\pm)}}^{c(n)},t)\bigg\}_\epsilon^1\;.
\label{lemma1f}
\end{eqnarray}
Since $t\in(\epsilon,1)$ is real, $\varphi_{_{L_{(ir)}}}^{c(+)}(\overline{y}_{_{(\pm\pm)}}^{c(n)},t)
=\Big[\varphi_{_{L_{(ir)}}}^{c(+)}(y_{_{(\pm\pm)}}^{c(n)},t)\Big]^*$,
$\varphi_{_{R_{(ir)}}}^{c(-)}(\overline{y}_{_{(\pm\pm)}}^{c(n)},t)
=\Big[\varphi_{_{R_{(ir)}}}^{c(-)}(y_{_{(\pm\pm)}}^{c(n)},t)\Big]^*$, then
\begin{eqnarray}
&&\int_\epsilon^1dtt\varphi_{_{L_{(ir)}}}^{c(+)}(y_{_{(\pm\pm)}}^{c(n)},t)
\varphi_{_{L_{(ir)}}}^{c(+)}(\overline{y}_{_{(\pm\pm)}}^{c(n)},t)
=\int_\epsilon^1dtt\Big|\varphi_{_{L_{(ir)}}}^{c(+)}(y_{_{(\pm\pm)}}^{c(n)},t)\Big|^2>0\;,
\nonumber\\
&&\int_\epsilon^1dtt\varphi_{_{R_{(ir)}}}^{c(-)}(y_{_{(\pm\pm)}}^{c(n)},t)
\varphi_{_{R_{(ir)}}}^{c(-)}(\overline{y}_{_{(\pm\pm)}}^{c(n)},t)
=\int_\epsilon^1dtt\Big|\varphi_{_{R_{(ir)}}}^{c(-)}(y_{_{(\pm\pm)}}^{c(n)},t)\Big|^2>0
\label{lemma1g}
\end{eqnarray}
for the nontrivial functions $\varphi_{_{L_{(ir)}}}^{c(+)}(y_{_{(\pm\pm)}}^{c(n)},t),
\varphi_{_{R_{(ir)}}}^{c(-)}(y_{_{(\pm\pm)}}^{c(n)},t)$. If $r\neq0$, then
$\Big[y_{_{(\pm\pm)}}^{c(n)}\Big]^2-\Big[\overline{y}_{_{(\pm\pm)}}^{c(n)}\Big]^2=i4rs\neq0$.
Applying $(--)$ BCs satisfied by the bulk profiles of right-handed fermions, we
derive the following equations from Eq(\ref{lemma1e}) and Eq(\ref{lemma1f}):
\begin{eqnarray}
&&\int_\epsilon^1dtt\Big|\varphi_{_{L_{(ir)}}}^{c(+)}(y_{_{(\pm\pm)}}^{c(n)},t)\Big|^2=0\;,
\nonumber\\
&&\int_\epsilon^1dtt\Big|\varphi_{_{R_{(ir)}}}^{c(-)}(y_{_{(\pm\pm)}}^{c(n)},t)\Big|^2=0\;,
\label{lemma1h}
\end{eqnarray}
which are contrary to the inequalities in Eq(\ref{lemma1g}). For $r=0$,
\begin{eqnarray}
&&\int_\epsilon^1dtt\Big|\varphi_{_{L_{(ir)}}}^{c(+)}(y_{_{(\pm\pm)}}^{c(n)},t)\Big|^2
\nonumber\\
&&\hspace{-0.6cm}=
\lim\limits_{r\rightarrow0}{1\over i4rs}
\bigg\{t\overline{y}_{_{(\pm\pm)}}^{c(n)}\varphi_{_{L_{(ir)}}}^{c(+)}(y_{_{(\pm\pm)}}^{c(n)},t)
\varphi_{_{R_{(ir)}}}^{c(-)}(\overline{y}_{_{(\pm\pm)}}^{c(n)},t)-t\overline{y}_{_{(\pm\pm)}}^{c(n)}
\varphi_{_{L_{(ir)}}}^{c(+)}(\overline{y}_{_{(\pm\pm)}}^{c(n)},t)
\varphi_{_{R_{(ir)}}}^{c(-)}(y_{_{(\pm\pm)}}^{c(n)},t)\bigg\}_\epsilon^1
\nonumber\\
&&\hspace{-0.6cm}=
{1\over i4s}\bigg\{t\varphi_{_{L_{(ir)}}}^{c(+)}(y_{_{(\pm\pm)}}^{c(n)},t)\varphi_{_{R_{(ir)}}}^{c(-)}
(\overline{y}_{_{(\pm\pm)}}^{c(n)},t)+{\overline{y}_{_{(\pm\pm)}}^{c(n)}\over y_{_{(\pm\pm)}}^{c(n)}}
t^2{\partial\varphi_{_{L_{(ir)}}}^{c(+)}\over\partial t}(y_{_{(\pm\pm)}}^{c(n)},t)\varphi_{_{R_{(ir)}}}^{c(-)}
(\overline{y}_{_{(\pm\pm)}}^{c(n)},t)
\nonumber\\&&
+t^2\varphi_{_{L_{(ir)}}}^{c(+)}(y_{_{(\pm\pm)}}^{c(n)},t){\partial\varphi_{_{R_{(ir)}}}^{c(-)}\over\partial t}
(\overline{y}_{_{(\pm\pm)}}^{c(n)},t)-t\varphi_{_{L_{(ir)}}}^{c(+)}(\overline{y}_{_{(\pm\pm)}}^{c(n)},t)
\varphi_{_{R_{(ir)}}}^{c(-)}(y_{_{(\pm\pm)}}^{c(n)},t)
\nonumber\\&&
-{y_{_{(\pm\pm)}}^{c(n)}\over \overline{y}_{_{(\pm\pm)}}^{c(n)}}
t^2{\partial\varphi_{_{L_{(ir)}}}^{c(+)}\over\partial t}(\overline{y}_{_{(\pm\pm)}}^{c(n)},t)\varphi_{_{R_{(ir)}}}^{c(-)}
(y_{_{(\pm\pm)}}^{c(n)},t)-t^2\varphi_{_{L_{(ir)}}}^{c(+)}(\overline{y}_{_{(\pm\pm)}}^{c(n)},t)
{\partial\varphi_{_{R_{(ir)}}}^{c(-)}\over\partial t}(y_{_{(\pm\pm)}}^{c(n)},t)
\bigg\}_\epsilon^1
\nonumber\\
&&\hspace{-0.6cm}=
{1\over2}\Big\{\Big|\varphi_{_{L_{(ir)}}}^{c(+)}(y_{_{(\pm\pm)}}^{c(n)},1)\Big|^2
-\epsilon^2\Big|\varphi_{_{L_{(ir)}}}^{c(+)}(y_{_{(\pm\pm)}}^{c(n)},\epsilon)\Big|^2\Big\}\;.
\label{lemma1i}
\end{eqnarray}
In the last step, we apply the equation of motion for the bulk profiles of fermions
in Eq(\ref{EM-fermion}) and $(--)$ BCs satisfied by the bulk files of
right-handed fermions. Similarly, we can derive
\begin{eqnarray}
&&\int_\epsilon^1dtt\Big|\varphi_{_{R_{(ir)}}}^{c(-)}(y_{_{(\pm\pm)}}^{c(n)},t)\Big|^2
=-{1\over2}\Big\{\Big|\varphi_{_{L_{(ir)}}}^{c(+)}(y_{_{(\pm\pm)}}^{c(n)},1)\Big|^2
-\epsilon^2\Big|\varphi_{_{L_{(ir)}}}^{c(+)}(y_{_{(\pm\pm)}}^{c(n)},\epsilon)\Big|^2\Big\}\;.
\label{lemma1j}
\end{eqnarray}
Eq(\ref{lemma1i}) and Eq(\ref{lemma1j}) are also contrary to the inequalities in Eq(\ref{lemma1g}).
In other words, $R_{_{(\pm\pm)}}^{c,\epsilon}(z)=0$ only has the real roots.
Analogously, we can verify that the equations $R_{_{(\pm\mp)}}^{c,\epsilon}(z)=0$,
$R_{_{(\mp\pm)}}^{c,\epsilon}(z)=0$, $R_{_{(\mp\mp)}}^{c,\epsilon}(z)=0$,
$R_{_{(++)}}^{G,\epsilon}(z)=0$ and $R_{_{(-+)}}^{G,\epsilon}(z)=0$ only have
real roots.
{\it Lemma2: If $y_{_{\rm(BCs)}}^{c(n)}(n=1,\;2,\;\cdots,\;\infty)$ satisfy
$R_{_{\rm(BCs)}}^{c,\epsilon}(y_{_{\rm(BCs)}}^{c(n)})=0$, then
$y_{_{\rm(BCs)}}^{c(-n)}=-y_{_{\rm(BCs)}}^{c(n)}$ and $\pm y_{_{\rm(BCs)}}^{c(1)},\;
\cdots,\;\pm y_{_{\rm(BCs)}}^{c(n)},\;\cdots$ are the zeros of order one of the function
$R_{_{\rm(BCs)}}^{c,\epsilon}(z)$.}
Proof. Assuming ${\rm(BCs)}=(\pm\pm)$, we firstly demonstrate
that $y_{_{(\pm\pm)}}^{c(1)},\;\cdots,\;y_{_{(\pm\pm)}}^{c(n)},\;\cdots$ are
the zeros of order one of the function $R_{_{(\pm\pm)}}^{c,\epsilon}(z)$.
For $c\neq N+{1\over2}$
\begin{eqnarray}
&&{dR_{_{(\pm\pm)}}^{c,\epsilon}\over dz}(z)=\Big[J_{-c-{1\over2}}(z)J_{c-{1\over2}}(z\epsilon)
+J_{c+{1\over2}}(z)J_{-c+{1\over2}}(z\epsilon)\Big]
\nonumber\\
&&\hspace{2.3cm}
-\epsilon\Big[J_{-c+{1\over2}}(z)J_{c+{1\over2}}(z\epsilon)+J_{c-{1\over2}}(z)
J_{-c-{1\over2}}(z\epsilon)\Big]
\nonumber\\
&&\hspace{2.3cm}
+{2c-1\over z}\Big[J_{-c+{1\over2}}(z)J_{c-{1\over2}}(z\epsilon)
-J_{c-{1\over2}}(z)J_{-c+{1\over2}}(z\epsilon)\Big]
\nonumber\\
&&\hspace{2.0cm}=
\varphi_{_{R_{(ir)}}}^{c(+)}(z,\epsilon)-\epsilon\varphi_{_{L_{(ir)}}}^{c(+)}(z,\epsilon)
+{2c-1\over z}\varphi_{_{R_{(ir)}}}^{c(-)}(z,\epsilon)\;.
\label{lemma2e}
\end{eqnarray}
As $c=N+1/2$,
\begin{eqnarray}
&&{dR_{_{(\pm\pm)}}^{c,\epsilon}\over dz}(z)=-\Big[
Y_{N+1}(z)J_{N}(z\epsilon)-J_{N+1}(z)Y_{N}(z\epsilon)\Big]
\nonumber\\
&&\hspace{2.3cm}
-\epsilon\Big[
Y_{N}(z)J_{N+1}(z\epsilon)-J_{N}(z)Y_{N+1}(z\epsilon)\Big]
\nonumber\\
&&\hspace{2.3cm}
+{2N\over z}\Big[Y_{N}(z)J_{N}(z\epsilon)-J_{N}(z)Y_{N}(z\epsilon)\Big]
\nonumber\\
&&\hspace{2.0cm}=
\varphi_{_{R_{(ir)}}}^{c(+)}(z,\epsilon)-\epsilon\varphi_{_{L_{(ir)}}}^{c(+)}(z,\epsilon)
+{2c-1\over z}\varphi_{_{R_{(ir)}}}^{c(-)}(z,\epsilon)\;.
\label{lemma2f}
\end{eqnarray}
When $y_{_{(\pm\pm)}}^{c(n)}(n=1,\;2,\;\cdots,\;\infty)$ satisfy the equation
$R_{_{(\pm\pm)}}^{c,\epsilon}(y_{_{(\pm\pm)}}^{c(n)})=0$, then
\begin{eqnarray}
&&{dR_{_{(\pm\pm)}}^{c,\epsilon}\over dz}(z)\bigg|_{z=y_{_{(\pm\pm)}}^{c(n)}}=
\varphi_{_{R_{(ir)}}}^{c(+)}(y_{_{(\pm\pm)}}^{c(n)},\epsilon)-\epsilon\varphi_{_{L_{(ir)}}}^{c(+)}
(y_{_{(\pm\pm)}}^{c(n)},\epsilon)\neq0\;.
\label{lemma2g}
\end{eqnarray}
In other words, $y_{_{(\pm\pm)}}^{c(n)}(n=1,\;2,\;\cdots,\;\infty)$ are the zeros of order one
for the function $R_{_{(\pm\pm)}}^{c,\epsilon}(z)$. Using concrete expressions
of the Bessel functions $J_{\nu}$ and $Y_{_\nu}$, we can verify directly $R_{_{(\pm\pm)}}^{c,\epsilon}(
-y_{_{(\pm\pm)}}^{c(n)})=0$ if $R_{_{(\pm\pm)}}^{c,\epsilon}(y_{_{(\pm\pm)}}^{c(n)})=0$.
Furthermore, we can obtain those similar results on the zeros of the functions
$R_{_{(\pm\mp)}}^{c,\epsilon}(z)$, $R_{_{(\mp\pm)}}^{c,\epsilon}(z)$, $R_{_{(\mp\mp)}}^{c,\epsilon}(z)$,
$R_{_{(++)}}^{G,\epsilon}(z)$ and $R_{_{(-+)}}^{G,\epsilon}(z)$.
When $z\rightarrow0$ and $c\neq N+{1\over2}$, we expand the function $R_{_{(\pm\pm)}}^{c,\epsilon}(z)$
according $z$ as
\begin{eqnarray}
&&R_{_{(\pm\pm)}}^{c,\epsilon}(z)={2(\epsilon^{c-{1\over2}}-\epsilon^{{1\over2}-c})\over(1-2c)
\Gamma({1\over2}-c)\Gamma({1\over2}+c)}\Big\{1+{\cal O}(z^2)\Big\}\;.
\label{lemma2b}
\end{eqnarray}
When $z\rightarrow0$ and $c=N+{1\over2}$, the function $R_{_{(\pm\pm)}}^{c,\epsilon}(z)$
is approximated as
\begin{eqnarray}
&&R_{_{(\pm\pm)}}^{c,\epsilon}(z)=\left\{\begin{array}{ll}
{1-\epsilon^{2N}\over N\pi\epsilon^N}\Big\{1+{\cal O}(z^2)\Big\}, & N\neq 0\\
-{2\ln\epsilon\over\pi}\Big\{1+{\cal O}(z^2)\Big\}, & N=0\end{array}\right.\;.
\label{lemma2c}
\end{eqnarray}
In other words, $z=0$ is not the zero of $R_{_{(\pm\pm)}}^{c,\epsilon}(z)$. Similarly,
we find that $z=0$ is not the zero of the functions $R_{_{(\mp\mp)}}^{c,\epsilon}(z)$ as well as
$R_{_{(++)}}^{G,\epsilon}(z)$ also, and is the pole of order one of the functions
$R_{_{(\pm\mp)}}^{c,\epsilon}(z)$, $R_{_{(\mp\pm)}}^{c,\epsilon}(z)$ together with
$R_{_{(-+)}}^{G,\epsilon}(z)$.
{\it Lemma3: Let function $f(z)$ be analytic except for limited in number isolated
singularities on the complex plane. If there are two constants ${\cal M}>0$ and ${\cal R}>0$,
we have $|zf(z)|\leq{\cal M}$ when $|z|>{\cal R}$. Then
\begin{eqnarray}
&&\lim\limits_{n\rightarrow\infty}\oint_{C_{_{r_{(n)}}}}\Big\{{2\over z}
+{1\over R_{_{\rm(BCs)}}^{c,\epsilon}(z)}{dR_{_{\rm(BCs)}}^{c,\epsilon}\over dz}(z)
\Big\}f(z)dz=0\;,
\label{lemma3a}
\end{eqnarray}
where the path $C_{_{r_{(n)}}}$ is the rectangular contour with four vertices
$(1\pm i)r_{(n)}$ and $(-1\pm i)r_{(n)}$ with
$y_{_{\rm(BCs)}}^{c(n)}<r_{(n)}<y_{_{\rm(BCs)}}^{c(n+1)}$.
Proof. Firstly, we illustrate how to demonstrate the lemma for the case ${\rm(BCs)}=(\pm\pm)$.
Since $|zf(z)|\leq{\cal M}$ when $|z|>{\cal R}$, all singularities of the function $f(z)$
all distribute within the region $|z|\le{\cal R}$. This implies that $zf(z)$ is analytic at
$z=\infty$:
\begin{eqnarray}
&&zf(z)=a_0+{a_1\over z}+{a_2\over z^2}+\cdots\;,\;\;|z|>{\cal R}\;,
\label{lemma3b}
\end{eqnarray}
or
\begin{eqnarray}
&&f(z)={a_0\over z}+{a_1\over z^2}+{a_2\over z^3}+\cdots\;,\;\;|z|>{\cal R}\;.
\label{lemma3c}
\end{eqnarray}
Applying residue theorem, we have
\begin{eqnarray}
&&\oint_{C_{_{r_{(n)}}}}\Big\{{2\over z^2}+{1\over zR_{_{(\pm\pm)}}^{c,\epsilon}(z)}
{dR_{_{(\pm\pm)}}^{c,\epsilon}\over dz}(z)\Big\}dz
\nonumber\\
&&\hspace{-0.6cm}=
i2\pi\Big\{{\rm Res}\Big({2\over z^2}+{1\over zR_{_{(\pm\pm)}}^{c,\epsilon}(z)}
{dR_{_{(\pm\pm)}}^{c,\epsilon}\over dz}(z),z=0\Big)
\nonumber\\&&
+\sum\limits_{i=1}^n\Big[{\rm Res}\Big({1\over zR_{_{(\pm\pm)}}^{c,\epsilon}(z)}
{dR_{_{(\pm\pm)}}^{c,\epsilon}\over dz}(z),z=y_{_{(\pm\pm)}}^{c(i)}\Big)
\nonumber\\&&
+{\rm Res}\Big({1\over zR_{_{(\pm\pm)}}^{c,\epsilon}(z)}
{dR_{_{(\pm\pm)}}^{c,\epsilon}\over dz}(z),z=-y_{_{(\pm\pm)}}^{c(i)}\Big)\Big]\Big\}
\label{lemma3d}
\end{eqnarray}
Because $2/z^2+(dR_{_{(\pm\pm)}}^{c,\epsilon}(z)/dz)/(zR_{_{(\pm\pm)}}^{c,\epsilon}(z))$
is even function of $z$, its Laurent series at the point $z=0$ does not contain
the term which is proportional to $1/z$. One directly has
\begin{eqnarray}
&&\oint_{C_{_{r_{(n)}}}}\Big\{{2\over z^2}+{1\over zR_{_{(\pm\pm)}}^{c,\epsilon}(z)}
{dR_{_{(\pm\pm)}}^{c,\epsilon}\over dz}(z)\Big\}dz
\nonumber\\
&&\hspace{-0.6cm}=
i2\pi\sum\limits_{i=1}^n\Big[{\rm Res}\Big({1\over zR_{_{(\pm\pm)}}^{c,\epsilon}(z)}
{dR_{_{(\pm\pm)}}^{c,\epsilon}\over dz}(z),z=y_{_{(\pm\pm)}}^{c(i)}\Big)
\nonumber\\&&
+{\rm Res}\Big({1\over zR_{_{(\pm\pm)}}^{c,\epsilon}(z)}
{dR_{_{(\pm\pm)}}^{c,\epsilon}\over dz}(z),z=-y_{_{(\pm\pm)}}^{c(i)}\Big)\Big]
\nonumber\\
&&\hspace{-0.6cm}=
i2\pi\sum\limits_{i=1}^n\Big[{1\over y_{_{(\pm\pm)}}^{c(i)}}
+{1\over -y_{_{(\pm\pm)}}^{c(i)}}\Big]=0\;,
\label{lemma3e}
\end{eqnarray}
then gets
\begin{eqnarray}
&&\oint_{C_{_{r_{(n)}}}}\Big\{{2\over z}+{1\over R_{_{(\pm\pm)}}^{c,\epsilon}(z)}
{dR_{_{(\pm\pm)}}^{c,\epsilon}\over dz}(z)\Big\}f(z)dz
\nonumber\\
&&\hspace{-0.6cm}=
\oint_{C_{_{r_{(n)}}}}\Big\{{2\over z}+{1\over R_{_{(\pm\pm)}}^{c,\epsilon}(z)}
{dR_{_{(\pm\pm)}}^{c,\epsilon}\over dz}(z)\Big\}\Big\{f(z)-{a_0\over z}\Big\}dz\;.
\label{lemma3f}
\end{eqnarray}
For $f(z)-a_0/z=1/z^2\Big\{a_1+a_2/z+\cdots+a_{n+1}/z^n+\cdots\Big\}$,
$a_1+a_2/z+\cdots+a_{n+1}/z^n+\cdots$ is the analytic function and its absolute
value has an upper limit in the region $|z|\ge{\cal R}^\prime,\;{\cal R}^\prime>{\cal R}$.
Assuming
\begin{eqnarray}
&&\Big|a_1+{a_2\over z}+\cdots+{a_{n+1}\over z^n}+\cdots\Big|\le {\cal M}^\prime,\;
as\;|z|\ge{\cal R}^\prime\;,
\label{lemma3g}
\end{eqnarray}
then one gets
\begin{eqnarray}
&&\Big|f(z)-{a_0\over z}\Big|\le {{\cal M}^\prime\over|z|^2},\;
as\;|z|\ge{\cal R}^\prime\;.
\label{lemma3h}
\end{eqnarray}
As $n$ is sufficiently large, we have $|z|\ge{\cal R}^\prime$ for $z\in
C_{_{r_{(n)}}}$, then get
\begin{eqnarray}
&&\bigg|\oint_{C_{_{r_{(n)}}}}\Big\{{2\over z}+{1\over R_{_{(\pm\pm)}}^{c,\epsilon}(z)}
{dR_{_{(\pm\pm)}}^{c,\epsilon}\over dz}(z)\Big\}\Big\{f(z)-{a_0\over z}\Big\}dz\bigg|
\nonumber\\
&&\hspace{-0.6cm}
\le{4r_{(n)}{\cal M}^\prime\over r_{(n)}^2}\times\Big\{the\;upper\;bound\;of\;
\Big|{1\over R_{_{(\pm\pm)}}^{c,\epsilon}(z)}{dR_{_{(\pm\pm)}}^{c,\epsilon}\over dz}(z)\Big|
\;for\;z\in C_{_{r_{(n)}}}\Big\}\;.
\label{lemma3i}
\end{eqnarray}
In order to obtain the upper bound of $\Big|(dR_{_{(\pm\pm)}}^{c,\epsilon}(z)/dz)
/R_{_{(\pm\pm)}}^{c,\epsilon}(z)\Big|$ for $z\in C_{_{r_{(n)}}}$, we express the
Bessel functions as \cite{Wangzx}
\begin{eqnarray}
&&J_\nu(z)={1\over\sqrt{2\pi z}}\Big\{\Big[e^{i(z-{\nu\pi\over2}-{\pi\over4})}
+e^{-i(z-{\nu\pi\over2}-{\pi\over4})}\Big]\Big(1+{\cal O}({1\over z^2})\Big)
\nonumber\\
&&\hspace{1.6cm}
+{i\over2\nu}(\nu^2-{1\over4})\Big[e^{i(z-{\nu\pi\over2}-{\pi\over4})}
-e^{-i(z-{\nu\pi\over2}-{\pi\over4})}\Big]\Big(1+{\cal O}({1\over z^2})\Big)\Big\}
\;,\nonumber\\
&&Y_\nu(z)={1\over\sqrt{2\pi z}}\Big\{-i\Big[e^{i(z-{\nu\pi\over2}-{\pi\over4})}
-e^{-i(z-{\nu\pi\over2}-{\pi\over4})}\Big]\Big(1+{\cal O}({1\over z^2})\Big)
\nonumber\\
&&\hspace{1.6cm}
+{1\over2\nu}(\nu^2-{1\over4})\Big[e^{i(z-{\nu\pi\over2}-{\pi\over4})}
+e^{-i(z-{\nu\pi\over2}-{\pi\over4})}\Big]\Big(1+{\cal O}({1\over z^2})\Big)\Big\}
\label{lemma3j}
\end{eqnarray}
for $|z|\rightarrow\infty$. Using the above equations, we approximate
$R_{_{(\pm\pm)}}^{c,\epsilon}(z)$ and $dR_{_{(\pm\pm)}}^{c,\epsilon}(z)/dz$ as
\begin{eqnarray}
&&R_{_{(\pm\pm)}}^{c,\epsilon}(z)=-{i\cos c\pi\over\pi\sqrt{\epsilon}z}
\Big\{\Big[e^{i(1-\epsilon)z}-e^{-i(1-\epsilon)z}\Big]\Big(1+{\cal O}({1\over z^2})\Big)
\nonumber\\
&&\hspace{2.1cm}
+{ic(c-1)\over2z}\Big({1\over\epsilon}-1\Big)\Big[e^{i(1-\epsilon)z}
+e^{-i(1-\epsilon)z}\Big]\Big(1+{\cal O}({1\over z^2})\Big)\Big\}
\;,\nonumber\\
&&{dR_{_{(\pm\pm)}}^{c,\epsilon}\over dz}(z)={\cos c\pi\over\pi\sqrt{\epsilon}z}
\Big\{(1-\epsilon)\Big[e^{i(1-\epsilon)z}+e^{-i(1-\epsilon)z}\Big]\Big(1+{\cal O}({1\over z^2})\Big)
+{i\over z}\Big[1
\nonumber\\
&&\hspace{2.3cm}
+\Big(1-{(1+\epsilon^2)\over2\epsilon}\Big)c(c-1)\Big]\Big[e^{i(1-\epsilon)z}
-e^{-i(1-\epsilon)z}\Big]\Big(1+{\cal O}({1\over z^2})\Big)\Big\}
\label{lemma3k}
\end{eqnarray}
for $c\neq N+1/2$ and $|z|\rightarrow\infty$. When $c=N+1/2$, the functions
$R_{_{(\pm\pm)}}^{c,\epsilon}(z)$ and $dR_{_{(\pm\pm)}}^{c,\epsilon}(z)/dz$ at
$|z|\rightarrow\infty$ can be similarly approximated as
\begin{eqnarray}
&&R_{_{(\pm\pm)}}^{c,\epsilon}(z)=-{i\over\pi\sqrt{\epsilon}z}
\Big\{\Big[e^{i(1-\epsilon)z}-e^{-i(1-\epsilon)z}\Big]\Big(1+{\cal O}({1\over z^2})\Big)
\nonumber\\
&&\hspace{2.1cm}
+{i\over2z}\Big(N^2-{1\over4}\Big)\Big({1\over\epsilon}-1\Big)\Big[e^{i(1-\epsilon)z}
+e^{-i(1-\epsilon)z}\Big]\Big(1+{\cal O}({1\over z^2})\Big)\Big\}
\;,\nonumber\\
&&{dR_{_{(\pm\pm)}}^{c,\epsilon}\over dz}(z)={1\over\pi\sqrt{\epsilon}z}
\Big\{(1-\epsilon)\Big[e^{i(1-\epsilon)z}+e^{-i(1-\epsilon)z}\Big]\Big(1+{\cal O}({1\over z^2})\Big)
+{i\over z}\Big[1
\nonumber\\
&&\hspace{2.3cm}
+\Big(1-{(1+\epsilon^2)\over2\epsilon}\Big)(N^2-{1\over4})\Big]\Big[e^{i(1-\epsilon)z}
-e^{-i(1-\epsilon)z}\Big]\Big(1+{\cal O}({1\over z^2})\Big)\Big\}\;.
\label{lemma3l}
\end{eqnarray}
Using Eq(\ref{lemma3k}) and Eq(\ref{lemma3l}), one obtains
\begin{eqnarray}
&&{1\over R_{_{(\pm\pm)}}^{c,\epsilon}(z)}{dR_{_{(\pm\pm)}}^{c,\epsilon}\over dz}(z)
\nonumber\\
&&\hspace{-0.6cm}=
i(1-\epsilon){e^{i(1-\epsilon)z}+e^{-i(1-\epsilon)z}
+{i\over z}\Big[{1\over1-\epsilon}-{(1-\epsilon)c(c-1)\over2\epsilon}\Big]\Big[e^{i(1-\epsilon)z}
-e^{-i(1-\epsilon)z}\Big]\over e^{i(1-\epsilon)z}-e^{-i(1-\epsilon)z}
+{ic(c-1)\over2z}\Big({1\over\epsilon}-1\Big)\Big[e^{i(1-\epsilon)z}
+e^{-i(1-\epsilon)z}\Big]}\Big[1+{\cal O}({1\over z^2})\Big]\;.
\label{lemma3m}
\end{eqnarray}
If $n$ is sufficiently large, $y_{_{(\pm\pm)}}^{c(n)}$ is approximately
given by \cite{Buras5}
\begin{eqnarray}
&&y_{_{(\pm\pm)}}^{c(n)}\simeq\Big[n+{1\over2}\Big(\Big|c+{1\over2}\Big|-1\Big)-{1\over4}\Big]\pi\;.
\label{lemma3n}
\end{eqnarray}
The fact implies that the interval between $y_{_{(\pm\pm)}}^{c(n)}$ and $y_{_{(\pm\pm)}}^{c(n+1)}$
is about $\pi$ as $n\gg1$. When
\begin{eqnarray}
&&y_{_{(\pm\pm)}}^{c(n)}<{\pi\over1-\epsilon}\Big(N_0+{1\over4}\Big)
<{\pi\over1-\epsilon}\Big(N_0+{3\over4}\Big)\leq y_{_{(\pm\pm)}}^{c(n+1)}
\label{lemma3o}
\end{eqnarray}
or
\begin{eqnarray}
&&y_{_{(\pm\pm)}}^{c(n)}<{\pi\over1-\epsilon}\Big(N_0+{1\over4}\Big)
< y_{_{(\pm\pm)}}^{c(n+1)}\leq{\pi\over1-\epsilon}\Big(N_0+{3\over4}\Big)
\label{lemma3o1}
\end{eqnarray}
where the positive integer $N_0$ obviously turns large along with increasing of the number $n$,
one can choose
\begin{eqnarray}
&&r_{(n)}={\pi\over1-\epsilon}\Big(N_0+{1\over4}\Big)\;.
\label{lemma3p}
\end{eqnarray}
When the point $z$ belongs to the left- and right- borders of $C_{_{r_{(n)}}}$, i.e.
$z=\mp\Big(N_0+1/4\Big)\pi/(1-\epsilon)+iy$ with
$-\Big(N_0+1/4\Big)\pi/(1-\epsilon)\leq y\leq\Big(N_0+1/4\Big)\pi/(1-\epsilon)$,
we have
\begin{eqnarray}
&&\bigg|{1\over R_{_{(\pm\pm)}}^{c,\epsilon}(z)}{dR_{_{(\pm\pm)}}^{c,\epsilon}\over dz}(z)\bigg|
\nonumber\\
&&\hspace{-0.6cm}\leq
(1-\epsilon){\Big|e^{i(1-\epsilon)z}+e^{-i(1-\epsilon)z}\Big|
+{1\over|z|}\Big|{1\over1-\epsilon}-{(1-\epsilon)c(c-1)\over2\epsilon}\Big|\Big|e^{i(1-\epsilon)z}
-e^{-i(1-\epsilon)z}\Big|\over\Big|e^{i(1-\epsilon)z}-e^{-i(1-\epsilon)z}\Big|
-{|c(c-1)|\over2|z|}\Big({1\over\epsilon}-1\Big)\Big|e^{i(1-\epsilon)z}
+e^{-i(1-\epsilon)z}\Big|}\Big[1+{\cal O}({1\over|z|^2})\Big|
\nonumber\\
&&\hspace{-0.6cm}
\leq(1-\epsilon)\bigg\{1+{1\over r_{(n)}}\Big[{1\over1-\epsilon}
+|c(c-1)|\Big({1\over\epsilon}-1\Big)\Big]\Big\}
\Big[1+{\cal O}({1\over r_{(n)}^2})\Big]\;.
\label{lemma3x}
\end{eqnarray}
When the point $z$ belongs to the upper- and down- borders of $C_{_{r_{(n)}}}$, i.e.
$z=x\pm i\Big(N_0+1/4\Big)\pi/(1-\epsilon)$ with
$-\Big(N_0+1/4\Big)\pi/(1-\epsilon)\leq x\leq\Big(N_0+1/4\Big)\pi/(1-\epsilon)$,
we similarly obtain
\begin{eqnarray}
&&\bigg|{1\over R_{_{(\pm\pm)}}^{c,\epsilon}(z)}{dR_{_{(\pm\pm)}}^{c,\epsilon}\over dz}(z)\bigg|
\nonumber\\
&&\hspace{-0.6cm}\leq
(1-\epsilon){\Big|e^{i2(1-\epsilon)x}+e^{(2N_0+{1\over2})\pi}\Big|
+{1\over|z|}\Big|{1\over1-\epsilon}-{(1-\epsilon)c(c-1)\over2\epsilon}\Big|\Big|e^{i2(1-\epsilon)x}
-e^{(2N_0+{1\over2})\pi}\Big|\over\Big|e^{i2(1-\epsilon)x}-e^{(2N_0+{1\over2})\pi}\Big|
-{|c(c-1)|\over2|z|}\Big({1\over\epsilon}-1\Big)\Big|e^{i2(1-\epsilon)x}
+e^{(2N_0+{1\over2})\pi}\Big|}\Big[1+{\cal O}({1\over|z|^2})\Big|
\nonumber\\
&&\hspace{-0.6cm}\leq
(1-\epsilon)\Big\{1+{1\over|z|}\Big|{1\over1-\epsilon}-{(1-\epsilon)c(c-1)\over2\epsilon}\Big|
+{|c(c-1)|\over2|z|}\Big({1\over\epsilon}-1\Big)\Big\}
\nonumber\\
&&\hspace{0.0cm}\times
{e^{(2N_0+{1\over2})\pi}+1\over e^{(2N_0+{1\over2})\pi}
-1-{|c(c-1)|\over|z|}\Big({1\over\epsilon}-1\Big)}
\Big[1+{\cal O}({1\over |z|^2})\Big]
\nonumber\\
&&\hspace{-0.6cm}\leq
2(1-\epsilon)\bigg\{1+{1\over r_{(n)}}\Big[{1\over1-\epsilon}
+|c(c-1)|\Big({1\over\epsilon}-1\Big)\Big]\Big\}
\Big[1+{\cal O}({1\over r_{(n)}^2})\Big]\;.
\label{lemma3y}
\end{eqnarray}
As
\begin{eqnarray}
&&{\pi\over1-\epsilon}\Big(N_0+{1\over4}\Big)\leq y_{_{(\pm\pm)}}^{c(n)}
<{\pi\over1-\epsilon}\Big(N_0+{3\over4}\Big)< y_{_{(\pm\pm)}}^{c(n+1)}\;,
\label{lemma3z}
\end{eqnarray}
we choose
\begin{eqnarray}
&&r_{(n)}={\pi\over1-\epsilon}\Big(N_0+{3\over4}\Big)\;,
\label{lemma3A}
\end{eqnarray}
and similarly get the upper bounds in Eq(\ref{lemma3x}) and Eq(\ref{lemma3y}).
Applying Eq(\ref{lemma3i}), we have
\begin{eqnarray}
&&\lim\limits_{r_{(n)}\rightarrow\infty}
\bigg|\oint_{C_{_{r_{(n)}}}}\Big\{{2\over z}+{1\over R_{_{(\pm\pm)}}^{c,\epsilon}(z)}
{dR_{_{(\pm\pm)}}^{c,\epsilon}\over dz}(z)\Big\}\Big\{f(z)-{a_0\over z}\Big\}dz\bigg|
\nonumber\\
&&\hspace{-0.6cm}
\le\lim\limits_{r_{(n)}\rightarrow\infty}{4(1-\epsilon){\cal M}^\prime\over r_{(n)}}
\bigg\{1+{1\over r_{(n)}}\Big[{1\over1-\epsilon}+|c(c-1)|
\Big({1\over\epsilon}-1\Big)\Big]\bigg\}\Big[1+{\cal O}({1\over r_{(n)}^2})\Big]
\nonumber\\
&&\hspace{-0.6cm}=0\;.
\label{lemma3B}
\end{eqnarray}
In other words, the integral
\begin{eqnarray}
&&\lim\limits_{r_{(n)}\rightarrow\infty}
\oint_{C_{_{r_{(n)}}}}\Big\{{2\over z}+{1\over R_{_{(\pm\pm)}}^{c,\epsilon}(z)}
{dR_{_{(\pm\pm)}}^{c,\epsilon}\over dz}(z)\Big\}f(z)=0\;.
\label{lemma3C}
\end{eqnarray}
Using the equation in Eq(\ref{root-quark}) and Eq(\ref{lemma3j}), we have
\begin{eqnarray}
&&{1\over R_{_{(\pm\mp)}}^{c,\epsilon}(z)}{dR_{_{(\pm\mp)}}^{c,\epsilon}\over dz}(z)=
i(1-\epsilon){e^{i(1-\epsilon)z}-e^{-i(1-\epsilon)z}\over e^{i(1-\epsilon)z}+e^{-i(1-\epsilon)z}}
\Big\{1-{i\over2z}\Big[{c(c-1)\over\epsilon}
\nonumber\\
&&\hspace{3.9cm}
-c-1+{(c^2-1)\epsilon^2\over1-\epsilon}\Big]{e^{i(1-\epsilon)z}+e^{-i(1-\epsilon)z}\over
e^{i(1-\epsilon)z}-e^{-i(1-\epsilon)z}}
\nonumber\\
&&\hspace{3.9cm}
+{ic\over2z}\Big({c-1\over\epsilon}-c-1\Big){e^{i(1-\epsilon)z}-e^{-i(1-\epsilon)z}
\over e^{i(1-\epsilon)z}+e^{-i(1-\epsilon)z}}\Big\}\Big[1+{\cal O}({1\over z^2})\Big]\;,
\label{lemma3D}
\end{eqnarray}
for the case ${\rm BCs}=(\pm\mp)$. If $n$ is sufficiently large, $y_{_{(\pm\mp)}}^{c(n)}$
is approximately given by \cite{Buras5}
\begin{eqnarray}
&&y_{_{(\pm\mp)}}^{c(n)}\simeq\Big[n+{1\over2}\Big(\Big|c+{1\over2}\Big|-1\Big)+{1\over4}\Big]\pi\;.
\label{lemma3E}
\end{eqnarray}
The fact also implies that the interval between $y_{_{(\pm\mp)}}^{c(n)}$ and $y_{_{(\pm\mp)}}^{c(n+1)}$
is about $\pi$ as $n\gg1$. When
\begin{eqnarray}
&&y_{_{(\pm\pm)}}^{c(n)}<{\pi\over1-\epsilon}\Big(N_0-{1\over4}\Big)
<{\pi\over1-\epsilon}\Big(N_0+{1\over4}\Big)\leq y_{_{(\pm\pm)}}^{c(n+1)}
\label{lemma3F}
\end{eqnarray}
or
\begin{eqnarray}
&&y_{_{(\pm\pm)}}^{c(n)}<{\pi\over1-\epsilon}\Big(N_0-{1\over4}\Big)
< y_{_{(\pm\pm)}}^{c(n+1)}\leq{\pi\over1-\epsilon}\Big(N_0+{1\over4}\Big)
\label{lemma3F1}
\end{eqnarray}
where the positive integer $N_0$ obviously turns large along with increasing of the number $n$.
One can obviously choose
\begin{eqnarray}
&&r_{(n)}={\pi\over1-\epsilon}\Big(N_0-{1\over4}\Big)\;.
\label{lemma3G}
\end{eqnarray}
As
\begin{eqnarray}
&&{\pi\over1-\epsilon}\Big(N_0-{1\over4}\Big)\leq y_{_{(\pm\pm)}}^{c(n)}
<{\pi\over1-\epsilon}\Big(N_0+{1\over4}\Big)< y_{_{(\pm\pm)}}^{c(n+1)}\;,
\label{lemma3G1}
\end{eqnarray}
one can choose $r_{(n)}=(N_0+1/4)\pi/(1-\epsilon)$.
Then performing the similar analysis above, we finally get
\begin{eqnarray}
&&\lim\limits_{r_{(n)}\rightarrow\infty}
\oint_{C_{_{r_{(n)}}}}\Big\{{2\over z}+{1\over R_{_{(\pm\mp)}}^{c,\epsilon}(z)}
{dR_{_{(\pm\mp)}}^{c,\epsilon}\over dz}(z)\Big\}f(z)=0\;.
\label{lemma3H}
\end{eqnarray}
As for the function $R_{_{(\mp\pm)}}^{c,\epsilon}(z),\;R_{_{(\mp\mp)}}^{c,\epsilon}(z),\;
R_{_{(++)}}^{G,\epsilon}(z)$ and $R_{_{(-+)}}^{G,\epsilon}(z)$, we derive the similar equations.
Using the lemmas verified above and Eq(\ref{residue1}), we summarize the summing over
infinite series of KK modes as
\begin{eqnarray}
&&\sum\limits_{i=1}^\infty\Big[f(y_{_{\rm(BCs)}}^{c(i)})+f(-y_{_{\rm(BCs)}}^{c(i)})\Big]
=-{\rm Res}\Big[\Big({2\over z}+{1\over R_{_{\rm(BCs)}}^{c,\epsilon}(z)}
{dR_{_{\rm(BCs)}}^{c,\epsilon}\over dz}(z)\Big)f(z),z=0\Big]
\nonumber\\
&&\hspace{5.2cm}
-\sum\limits_{i=1}^{n_0}{\rm Res}\Big[\Big({2\over z}
+{1\over R_{_{\rm(BCs)}}^{c,\epsilon}(z)}
{dR_{_{\rm(BCs)}}^{c,\epsilon}\over dz}(z)\Big)f(z),z=z_i\Big]\;,
\nonumber\\
&&\sum\limits_{i=1}^\infty\Big[f(y_{_{\rm(BCs)}}^{G(i)})+f(-y_{_{\rm(BCs)}}^{G(i)})\Big]
=-{\rm Res}\Big[\Big({2\over z}+{1\over R_{_{\rm(BCs)}}^{G,\epsilon}(z)}
{dR_{_{\rm(BCs)}}^{G,\epsilon}\over dz}(z)\Big)f(z),z=0\Big]
\nonumber\\
&&\hspace{5.2cm}
-\sum\limits_{i=1}^{n_0}{\rm Res}\Big[\Big({2\over z}
+{1\over R_{_{\rm(BCs)}}^{G,\epsilon}(z)}
{dR_{_{\rm(BCs)}}^{G,\epsilon}\over dz}(z)\Big)f(z),z=z_i\Big]
\label{lemma3I}
\end{eqnarray}
where
\begin{eqnarray}
&&\lim\limits_{|z|\rightarrow\infty}|zf(z)|\le{\cal M},\;\;0<{\cal M}<\infty\;.
\label{lemma3J}
\end{eqnarray}
Actually, Eq(\ref{lemma3J}) is the sufficient condition to judge if
the infinite series $\sum\limits_{i=1}^\infty\Big[f(y_{_{(BCs)}}^{c(i)})+f(-y_{_{(BCs)}}^{c(i)})\Big]$
is convergent.
In extensions of the SM with a warped extra dimension, the bulk profiles in Eq.(\ref{quark-5th}) affect
amplitudes for relevant processes in terms of $\Big[\chi_{_{(BCs)}}^G(y_{_{(BCs)}}^{G(n)},\phi)\Big]
\Big[\chi_{_{(BCs)}}^G(y_{_{(BCs)}}^{G(n)},\phi^\prime)\Big]$, $\Big[f_{_{(BCs)}}^{L,c}(y_{_{(BCs)}}^{c(n)},\phi)\Big]
\Big[f_{_{(BCs)}}^{L,c}(y_{_{(BCs)}}^{c(n)},\phi^\prime)\Big]$ and $\Big[f_{_{(BCs)}}^{R,c}(y_{_{(BCs)}}^{c(n)},\phi)\Big]
\Big[f_{_{(BCs)}}^{R,c}(y_{_{(BCs)}}^{c(n)},\phi^\prime)\Big]$. In order to sum over the
infinite series of KK modes properly, one should analytically extend the above combinations
of bulk profiles to the complex plane. Here, we illustrate how to extend analytically the
combinations of bulk profiles for gauge fields satisfying $(++)$ ($(-+)$) BCs
in the complex plane. When $y=y_{_{(++)}}^{G(n)}$ satisfies the equation
$R_{_{(++)}}^{G,\epsilon}(y_{_{(++)}}^{G(n)})=0$, the combination of bulk profiles for
gauge fields satisfying $(++)$ BCs can be formulated as
\begin{eqnarray}
&&\Big[\chi_{_{(++)}}^G(y,\phi)\Big]\Big[\chi_{_{(++)}}^G(y,\phi^\prime)\Big]
={tt^\prime\Phi_{_{(uv)}}^{G(+)}(y,t)\Phi_{_{(uv)}}^{G(+)}
(y,t^\prime)\over\Big[N_{_{(uv)}}^{G(+)}(y)\Big]^2}
\nonumber\\
&&\hspace{4.6cm}
={tt^\prime\Phi_{_{(ir)}}^{G(+)}(y,t)\Phi_{_{(ir)}}^{G(+)}
(y,t^\prime)\over\Big[N_{_{(ir)}}^{G(+)}(y)\Big]^2}
\nonumber\\
&&\hspace{4.6cm}
={tt^\prime\Phi_{_{(uv)}}^{G(+)}(y,t)\Phi_{_{(ir)}}^{G(+)}
(y,t^\prime)\over\Big[N_{_{(uv)}}^{G(+)}(y)\Big]\Big[N_{_{(ir)}}^{G(+)}(y)\Big]}
\nonumber\\
&&\hspace{4.6cm}
={tt^\prime\Phi_{_{(ir)}}^{G(+)}(y,t)\Phi_{_{(uv)}}^{G(+)}
(y,t^\prime)\over\Big[N_{_{(uv)}}^{G(+)}(y)\Big]\Big[N_{_{(ir)}}^{G(+)}(y)\Big]}\;,
\label{bulkfiles-extension1}
\end{eqnarray}
with $t=\epsilon e^{\sigma(\phi)},\;t^\prime=\epsilon e^{\sigma(\phi^\prime)}$.
When $y=y_{_{(-+)}}^{G(n)}$ satisfies the equation
$R_{_{(-+)}}^{G,\epsilon}(y_{_{(-+)}}^{G(n)})=0$, the combination of bulk profiles for
gauge fields satisfying the BCs $(-+)$ can be written as
\begin{eqnarray}
&&\Big[\chi_{_{(-+)}}^G(y,\phi)\Big]\Big[\chi_{_{(-+)}}^G(y,\phi^\prime)\Big]
={tt^\prime\Phi_{_{(uv)}}^{G(-)}(y,t)\Phi_{_{(uv)}}^{G(-)}
(y,t^\prime)\over\Big[N_{_{(uv)}}^{G(-)}(y)\Big]^2}
\nonumber\\
&&\hspace{4.6cm}
={tt^\prime\Phi_{_{(ir)}}^{G(+)}(y,t)\Phi_{_{(ir)}}^{G(+)}
(y,t^\prime)\over\Big[N_{_{(ir)}}^{G(+)}(y)\Big]^2}
\nonumber\\
&&\hspace{4.6cm}
={tt^\prime\Phi_{_{(uv)}}^{G(-)}(y,t)\Phi_{_{(ir)}}^{G(+)}
(y,t^\prime)\over\Big[N_{_{(uv)}}^{G(-)}(y)\Big]\Big[N_{_{(ir)}}^{G(+)}(y)\Big]}
\nonumber\\
&&\hspace{4.6cm}
={tt^\prime\Phi_{_{(ir)}}^{G(+)}(y,t)\Phi_{_{(uv)}}^{G(-)}
(y,t^\prime)\over\Big[N_{_{(uv)}}^{G(-)}(y)\Big]\Big[N_{_{(ir)}}^{G(+)}(y)\Big]}\;.
\label{bulkfiles-extension1a}
\end{eqnarray}
The combinations of bulk profiles for gauge fields certainly satisfy the corresponding BCs:
\begin{eqnarray}
&&{\partial\over\partial\phi_{_{(uv)}}}
\Big[\chi_{_{(++)}}^G(y,\phi)\Big]\Big[\chi_{_{(++)}}^G(y,\phi^\prime)\Big]
\bigg|_{\phi_{_{(uv)}}=0}=0\;,
\nonumber\\
&&{\partial\over\partial\phi_{_{(ir)}}}
\Big[\chi_{_{(++)}}^G(y,\phi)\Big]\Big[\chi_{_{(++)}}^G(y,\phi^\prime)\Big]
\bigg|_{\phi_{_{(ir)}}=\pi/2}=0\;,
\nonumber\\
&&\Big[\chi_{_{(-+)}}^G(y,\phi)\Big]\Big[\chi_{_{(-+)}}^G(y,\phi^\prime)\Big]
\bigg|_{\phi_{_{(uv)}}=0}=0\;,
\nonumber\\
&&{\partial\over\partial\phi_{_{(ir)}}}
\Big[\chi_{_{(-+)}}^G(y,\phi)\Big]\Big[\chi_{_{(-+)}}^G(y,\phi^\prime)\Big]
\bigg|_{\phi_{_{(ir)}}=\pi/2}=0\;,
\label{bulkfiles-extension2}
\end{eqnarray}
with $\phi_{_{(uv)}}=\min(\phi,\;\phi^\prime)$, $\phi_{_{(ir)}}=\max(\phi,\;\phi^\prime)$.
Considering Eq.(\ref{bulkfiles-extension2}), we analytically extend
the combinations of bulk profiles from Eq.(\ref{bulkfiles-extension1}) and Eq.(\ref{bulkfiles-extension1a})
in the complex plane as
\begin{eqnarray}
&&\Big[\chi_{_{(++)}}^G(z,\phi)\Big]\Big[\chi_{_{(++)}}^G(z,\phi^\prime)\Big]
={tt^\prime\over\Big[N_{_{(uv)}}^{G(+)}(z)\Big]\Big[N_{_{(ir)}}^{G(+)}(z)\Big]}
\Big\{\theta(t-t^\prime)\Phi_{_{(uv)}}^{G(+)}(z,t^\prime)\Phi_{_{(ir)}}^{G(+)}(z,t)
\nonumber\\
&&\hspace{4.8cm}
+\theta(t^\prime-t)\Phi_{_{(uv)}}^{G(+)}(z,t)\Phi_{_{(ir)}}^{G(+)}(z,t^\prime)\Big\}\;,
\nonumber\\
&&\Big[\chi_{_{(-+)}}^G(z,\phi)\Big]\Big[\chi_{_{(-+)}}^G(z,\phi^\prime)\Big]
={tt^\prime\over\Big[N_{_{(uv)}}^{G(-)}(z)\Big]\Big[N_{_{(ir)}}^{G(+)}(z)\Big]}
\Big\{\theta(t-t^\prime)\Phi_{_{(uv)}}^{G(-)}(z,t^\prime)\Phi_{_{(ir)}}^{G(+)}(z,t)
\nonumber\\
&&\hspace{4.8cm}
+\theta(t^\prime-t)\Phi_{_{(uv)}}^{G(-)}(z,t)\Phi_{_{(ir)}}^{G(+)}(z,t^\prime)\Big\}\;.
\label{bulkfiles-extension3}
\end{eqnarray}
Here, the step function $\theta(x)$ is defined as
\begin{eqnarray}
&&\theta(x)=\left\{\begin{array}{ll}1,&x>0\;;\\{1\over2},&x=0\;;\\0,&x<0\;.\end{array}\right.
\label{step-function}
\end{eqnarray}
To guarantee that the combinations of bulk profiles are
uniformly bounded in the complex plane, we analytically extend the corresponding
normalization factors in Eq.(\ref{bulkfiles-extension3}) as
\begin{eqnarray}
&&\Big|N_{_{(uv)}}^{G(+)}(z)\Big|^2={2\over kr(z^2-\bar{z}^2)}
\Big\{\bar{z}\Phi_{_{(uv)}}^{G(+)}(z,1)\Psi_{_{(uv)}}^{G(-)}(\bar{z},1)
-z\Phi_{_{(uv)}}^{G(+)}(\bar{z},1)\Psi_{_{(uv)}}^{G(-)}(z,1)\Big\}
\nonumber\\
&&\hspace{2.5cm}
+{2\over kr}\Upsilon(z)\;,
\nonumber\\
&&\Big|N_{_{(uv)}}^{G(-)}(z)\Big|^2={2\over kr(z^2-\bar{z}^2)}
\Big\{\bar{z}\Phi_{_{(uv)}}^{G(-)}(z,1)\Psi_{_{(uv)}}^{G(+)}(\bar{z},1)
-z\Phi_{_{(uv)}}^{G(-)}(\bar{z},1)\Psi_{_{(uv)}}^{G(+)}(z,1)\Big\}
\nonumber\\
&&\hspace{2.5cm}
+{2\over kr}\Upsilon(z)\;,
\nonumber\\
&&\Big|N_{_{(ir)}}^{G(+)}(z)\Big|^2={2\epsilon\over kr(z^2-\bar{z}^2)}
\Big\{z\Phi_{_{(ir)}}^{G(+)}(\bar{z},\epsilon)\Psi_{_{(ir)}}^{G(-)}(z,\epsilon)
-\bar{z}\Phi_{_{(ir)}}^{G(+)}(z,\epsilon)\Psi_{_{(ir)}}^{G(-)}(\bar{z},\epsilon)\Big\}
\nonumber\\
&&\hspace{2.5cm}
+{2\over kr}\Upsilon(z)\;.
\label{normalization-z-plane}
\end{eqnarray}
Here, $\bar{z}$ represents the conjugate of $z$, and the non-negative function $\Upsilon(z)$
is defined as
\begin{eqnarray}
&&\Upsilon(z)={1\over\pi^2|z|^2}\bigg\{(1-\epsilon)\Big(e^{-i(1-\epsilon)(z-\bar{z})}
+e^{i(1-\epsilon)(z-\bar{z})}\Big)
+{e^{-i(1-\epsilon)(z-\bar{z})}-e^{i(1-\epsilon)(z-\bar{z})}\over i(z-\bar{z})}\bigg\}.
\label{upsilon}
\end{eqnarray}
In the limit of $\bar{z}=z$ (i.e. $z$ is real), one easily gets
\begin{eqnarray}
&&\lim\limits_{\bar{z}\rightarrow z}\Upsilon(z)=0
\label{upsilon-limit}
\end{eqnarray}
and the normalization factors in Eq.(\ref{normalization-z-plane}) recover the corresponding
expressions in Eq.(\ref{normalization-gauge}). Similarly, we can analytically generalize
the normalization constants of bulk profiles for fermions as
\begin{eqnarray}
&&\Big|N_{_{L_{(ir)}}}^{c(+)}(z)\Big|^2={2\over kr(z^2-\bar{z}^2)}
\Big\{z\varphi_{_{L_{(ir)}}}^{c(+)}(\bar{z},\epsilon)\varphi_{_{R_{(ir)}}}^{c(-)}(z,\epsilon)
-\bar{z}\varphi_{_{L_{(ir)}}}^{c(+)}(z,\epsilon)\varphi_{_{R_{(ir)}}}^{c(-)}(\bar{z},\epsilon)\Big\}
\nonumber\\
&&\hspace{2.5cm}
+{2\cos^2c\pi\over kr\epsilon}\Upsilon(z)\;,
\nonumber\\
&&\Big|N_{_{L_{(uv)}}}^{c(+)}(z)\Big|^2={2\over kr\epsilon(z^2-\bar{z}^2)}
\Big\{\bar{z}\varphi_{_{L_{(uv)}}}^{c(+)}(z,1)\varphi_{_{R_{(uv)}}}^{c(-)}(\bar{z},1)
-z\varphi_{_{L_{(uv)}}}^{c(+)}(\bar{z},1)\varphi_{_{R_{(uv)}}}^{c(-)}(z,1)\Big\}
\nonumber\\
&&\hspace{2.5cm}
+{2\cos^2c\pi\over kr\epsilon}\Upsilon(z)\;,
\nonumber\\
&&\Big|N_{_{R_{(ir)}}}^{c(-)}(z)\Big|^2={2\over kr(z^2-\bar{z}^2)}
\Big\{\bar{z}\varphi_{_{L_{(ir)}}}^{c(+)}(\bar{z},\epsilon)\varphi_{_{R_{(ir)}}}^{c(-)}(z,\epsilon)
-z\varphi_{_{L_{(ir)}}}^{c(+)}(z,\epsilon)\varphi_{_{R_{(ir)}}}^{c(-)}(\bar{z},\epsilon)\Big\}
\nonumber\\
&&\hspace{2.5cm}
+{2\cos^2c\pi\over kr\epsilon}\Upsilon(z)\;,
\nonumber\\
&&\Big|N_{_{R_{(uv)}}}^{c(-)}(z)\Big|^2={2\over kr\epsilon(z^2-\bar{z}^2)}
\Big\{z\varphi_{_{L_{(uv)}}}^{c(+)}(z,1)\varphi_{_{R_{(uv)}}}^{c(-)}(\bar{z},1)
-\bar{z}\varphi_{_{L_{(uv)}}}^{c(+)}(\bar{z},1)\varphi_{_{R_{(uv)}}}^{c(-)}(z,1)\Big\}
\nonumber\\
&&\hspace{2.5cm}
+{2\cos^2c\pi\over kr\epsilon}\Upsilon(z)\;,
\nonumber\\
&&\Big|N_{_{L_{(ir)}}}^{c(-)}(z)\Big|^2={2\over kr(z^2-\bar{z}^2)}
\Big\{z\varphi_{_{L_{(ir)}}}^{c(-)}(\bar{z},\epsilon)\varphi_{_{R_{(ir)}}}^{c(+)}(z,\epsilon)
-\bar{z}\varphi_{_{L_{(ir)}}}^{c(-)}(z,\epsilon)\varphi_{_{R_{(ir)}}}^{c(+)}(\bar{z},\epsilon)\Big\}
\nonumber\\
&&\hspace{2.5cm}
+{2\cos^2c\pi\over kr\epsilon}\Upsilon(z)\;,
\nonumber\\
&&\Big|N_{_{L_{(uv)}}}^{c(-)}(z)\Big|^2={2\over kr\epsilon(z^2-\bar{z}^2)}
\Big\{\bar{z}\varphi_{_{L_{(uv)}}}^{c(-)}(z,1)\varphi_{_{R_{(uv)}}}^{c(+)}(\bar{z},1)
-z\varphi_{_{L_{(uv)}}}^{c(-)}(\bar{z},1)\varphi_{_{R_{(uv)}}}^{c(+)}(z,1)\Big\}
\nonumber\\
&&\hspace{2.5cm}
+{2\cos^2c\pi\over kr\epsilon}\Upsilon(z)\;,
\nonumber\\
&&\Big|N_{_{R_{(ir)}}}^{c(+)}(z)\Big|^2={2\over kr(z^2-\bar{z}^2)}
\Big\{\bar{z}\varphi_{_{L_{(ir)}}}^{c(-)}(\bar{z},\epsilon)\varphi_{_{R_{(ir)}}}^{c(+)}(z,\epsilon)
-z\varphi_{_{L_{(ir)}}}^{c(-)}(z,\epsilon)\varphi_{_{R_{(ir)}}}^{c(+)}(\bar{z},\epsilon)\Big\}
\nonumber\\
&&\hspace{2.5cm}
+{2\cos^2c\pi\over kr\epsilon}\Upsilon(z)\;,
\nonumber\\
&&\Big|N_{_{R_{(uv)}}}^{c(+)}(z)\Big|^2={2\over kr\epsilon(z^2-\bar{z}^2)}
\Big\{z\varphi_{_{L_{(uv)}}}^{c(-)}(z,1)\varphi_{_{R_{(uv)}}}^{c(+)}(\bar{z},1)
-\bar{z}\varphi_{_{L_{(uv)}}}^{c(-)}(\bar{z},1)\varphi_{_{R_{(uv)}}}^{c(+)}(z,1)\Big\}
\nonumber\\
&&\hspace{2.5cm}
+{2\cos^2c\pi\over kr\epsilon}\Upsilon(z)\;.
\label{normalization-z-plane1}
\end{eqnarray}
Using the above normalization constant defined in
Eq.(\ref{normalization-z-plane1}), we write the uniformly bounded combinations
of bulk profiles for fermion fields in the complex plane as
\begin{eqnarray}
&&\Big[f_{_{(++)}}^{L,c}(z,\phi)\Big]\Big[f_{_{(++)}}^{L,c}(z,\phi^\prime)\Big]
={\sqrt{tt^\prime}\over N_{_{L_{(uv)}}}^{c(+)}(z)N_{_{L_{(ir)}}}^{c(+)}(z)}
\Big\{\theta(t-t^\prime)\varphi_{_{L_{(uv)}}}^{c(+)}(z,t^\prime)\varphi_{_{L_{(ir)}}}^{c(+)}(z,t)
\nonumber\\
&&\hspace{4.8cm}
+\theta(t^\prime-t)\varphi_{_{L_{(uv)}}}^{c(+)}(z,t)\varphi_{_{L_{(ir)}}}^{c(+)}(z,t^\prime)\Big\}\;,
\nonumber\\
&&\Big[f_{_{(--)}}^{R,c}(z,\phi)\Big]\Big[f_{_{(--)}}^{R,c}(z,\phi^\prime)\Big]
={\sqrt{tt^\prime}\over N_{_{R_{(uv)}}}^{c(-)}(z)N_{_{R_{(ir)}}}^{c(-)}(z)}
\Big\{\theta(t-t^\prime)\varphi_{_{R_{(uv)}}}^{c(-)}(z,t^\prime)\varphi_{_{R_{(ir)}}}^{c(-)}(z,t)
\nonumber\\
&&\hspace{4.8cm}
+\theta(t^\prime-t)\varphi_{_{R_{(uv)}}}^{c(-)}(z,t)\varphi_{_{R_{(ir)}}}^{c(-)}(z,t^\prime)\Big\}\;.
\nonumber\\
&&\Big[f_{_{(+-)}}^{L,c}(z,\phi)\Big]\Big[f_{_{(+-)}}^{L,c}(z,\phi^\prime)\Big]
={\sqrt{tt^\prime}\over N_{_{L_{(uv)}}}^{c(+)}(z)N_{_{L_{(ir)}}}^{c(-)}(z)}
\Big\{\theta(t-t^\prime)\varphi_{_{L_{(uv)}}}^{c(+)}(z,t^\prime)\varphi_{_{L_{(ir)}}}^{c(-)}(z,t)
\nonumber\\
&&\hspace{4.8cm}
+\theta(t^\prime-t)\varphi_{_{L_{(uv)}}}^{c(+)}(z,t)\varphi_{_{L_{(ir)}}}^{c(-)}(z,t^\prime)\Big\}\;,
\nonumber\\
&&\Big[f_{_{(-+)}}^{R,c}(z,\phi)\Big]\Big[f_{_{(-+)}}^{R,c}(z,\phi^\prime)\Big]
={\sqrt{tt^\prime}\over N_{_{R_{(uv)}}}^{c(-)}(z)N_{_{R_{(ir)}}}^{c(+)}(z)}
\Big\{\theta(t-t^\prime)\varphi_{_{R_{(uv)}}}^{c(-)}(z,t^\prime)\varphi_{_{R_{(ir)}}}^{c(+)}(z,t)
\nonumber\\
&&\hspace{4.8cm}
+\theta(t^\prime-t)\varphi_{_{R_{(uv)}}}^{c(-)}(z,t)\varphi_{_{R_{(ir)}}}^{c(+)}(z,t^\prime)\Big\}\;,
\nonumber\\
&&\Big[f_{_{(-+)}}^{L,c}(z,\phi)\Big]\Big[f_{_{(-+)}}^{L,c}(z,\phi^\prime)\Big]
={\sqrt{tt^\prime}\over N_{_{L_{(uv)}}}^{c(-)}(z)N_{_{L_{(ir)}}}^{c(+)}(z)}
\Big\{\theta(t-t^\prime)\varphi_{_{L_{(uv)}}}^{c(-)}(z,t^\prime)\varphi_{_{L_{(ir)}}}^{c(+)}(z,t)
\nonumber\\
&&\hspace{4.8cm}
+\theta(t^\prime-t)\varphi_{_{L_{(uv)}}}^{c(-)}(z,t)\varphi_{_{L_{(ir)}}}^{c(+)}(z,t^\prime)\Big\}\;,
\nonumber\\
&&\Big[f_{_{(+-)}}^{R,c}(z,\phi)\Big]\Big[f_{_{(+-)}}^{R,c}(z,\phi^\prime)\Big]
={\sqrt{tt^\prime}\over N_{_{R_{(uv)}}}^{c(+)}(z)N_{_{R_{(ir)}}}^{c(-)}(z)}
\Big\{\theta(t-t^\prime)\varphi_{_{R_{(uv)}}}^{c(+)}(z,t^\prime)\varphi_{_{R_{(ir)}}}^{c(-)}(z,t)
\nonumber\\
&&\hspace{4.8cm}
+\theta(t^\prime-t)\varphi_{_{R_{(uv)}}}^{c(+)}(z,t)\varphi_{_{R_{(ir)}}}^{c(-)}(z,t^\prime)\Big\}\;,
\nonumber\\
&&\Big[f_{_{(--)}}^{L,c}(z,\phi)\Big]\Big[f_{_{(--)}}^{L,c}(z,\phi^\prime)\Big]
={\sqrt{tt^\prime}\over N_{_{L_{(uv)}}}^{c(-)}(z)N_{_{L_{(ir)}}}^{c(-)}(z)}
\Big\{\theta(t-t^\prime)\varphi_{_{L_{(uv)}}}^{c(-)}(z,t^\prime)\varphi_{_{L_{(ir)}}}^{c(-)}(z,t)
\nonumber\\
&&\hspace{4.8cm}
+\theta(t^\prime-t)\varphi_{_{L_{(uv)}}}^{c(-)}(z,t)\varphi_{_{L_{(ir)}}}^{c(-)}(z,t^\prime)\Big\}\;,
\nonumber\\
&&\Big[f_{_{(++)}}^{R,c}(z,\phi)\Big]\Big[f_{_{(++)}}^{R,c}(z,\phi^\prime)\Big]
={\sqrt{tt^\prime}\over N_{_{R_{(uv)}}}^{c(+)}(z)N_{_{R_{(ir)}}}^{c(+)}(z)}
\Big\{\theta(t-t^\prime)\varphi_{_{R_{(uv)}}}^{c(+)}(z,t^\prime)\varphi_{_{R_{(ir)}}}^{c(+)}(z,t)
\nonumber\\
&&\hspace{4.8cm}
+\theta(t^\prime-t)\varphi_{_{R_{(uv)}}}^{c(+)}(z,t)\varphi_{_{R_{(ir)}}}^{c(+)}(z,t^\prime)\Big\}\;.
\label{bulkfiles-extension4}
\end{eqnarray}
The above expressions in Eq.(\ref{bulkfiles-extension4}) are valid for $c\neq N+1/2$, one gets the
corresponding expressions for $c=N+1/2$ after replacing $\cos^2c\pi$ with $1$ in the
Eq.(\ref{normalization-z-plane1}).
\section{The four and five dimensional perturbative expansions\label{sec4}}
\indent\indent
The KK excitations affect the theoretical predictions of electroweak scale
although it is very difficult to produce them directly in the colliders running now.
When KK excitations of gauge fields satisfying $(++)$ BCs appear in
the relevant Feynman diagrams as virtual particles in four dimensional effective theory,
the amplitudes possibly contain the factor
\begin{eqnarray}
&&{-i\Big[\chi_{_{(++)}}^G(y_{_{(++)}}^{G(n)},\phi)\Big]\Big[\chi_{_{(++)}}^G
(y_{_{(++)}}^{G(n)},\phi^\prime)\Big]\over p^2-\Lambda_{_{KK}}^2\Big[y_{_{(++)}}^{G(n)}\Big]^2}
\label{propagator-1}
\end{eqnarray}
when we expand them according $\upsilon^2/\Lambda_{_{KK}}^2$, here $\upsilon$ denotes
the nonzero vacuum expectation value (VEV) of neutral Higgs located on IR brane. The denominator
$p^2-\Lambda_{_{KK}}^2\Big[y_{_{(++)}}^{G(n)}\Big]^2$ originates from the four dimensional
propagators of KK excitations for gauge fields in momentum space,
$\Big[\chi_{_{(++)}}^G(y_{_{(++)}}^{G(n)},\phi)\Big]$
and $\Big[\chi_{_{(++)}}^G(y_{_{(++)}}^{G(n)},\phi^\prime)\Big]$ originate from
the neighbor vertices in four dimensional effective theory. Note that
$\pm y_{_{(++)}}^{G(1)}$, $\cdots$, $\pm y_{_{(++)}}^{G(n)}$, $\cdots$ are zeros
of the function $R_{_{(++)}}^{G,\epsilon}(z)$, and the limit
\begin{eqnarray}
&&\lim\limits_{|z|\rightarrow\infty}\Bigg|{-iz\Big[\chi_{_{(++)}}^G(z,\phi)\Big]\Big[\chi_{_{(++)}}^G
(z,\phi^\prime)\Big]\over p^2-\Lambda_{_{KK}}^2z^2}\Bigg|=0
\label{propagator-2}
\end{eqnarray}
when we adopt the analytical extension of the combination of bulk profiles for $(++)$ BCs gauge
fields in Eq.(\ref{bulkfiles-extension3}). Applying Eq.(\ref{lemma3I}), we have
\begin{eqnarray}
&&iD_{_{(++)}}^{G}(p;\phi,\phi^\prime)
=\sum\limits_{n=1}^\infty{-i\Big[\chi_{_{(++)}}^G(y_{_{(++)}}^{G(n)},\phi)\Big]
\Big[\chi_{_{(++)}}^G(y_{_{(++)}}^{G(n)},\phi^\prime)\Big]\over p^2-\Lambda_{_{KK}}^2
\Big[y_{_{(++)}}^{G(n)}\Big]^2}
\nonumber\\
&&\hspace{-0.6cm}=
{i\over p^2}\bigg\{\Big[\chi_{_{(++)}}^G(0,\phi)\Big]
\Big[\chi_{_{(++)}}^G(0,\phi^\prime)\Big]
-\Big[\chi_{_{(++)}}^G({p\over \Lambda_{_{KK}}},\phi)\Big]
\Big[\chi_{_{(++)}}^G({p\over \Lambda_{_{KK}}},\phi^\prime)\Big]\bigg\}
\nonumber\\
&&\hspace{0.0cm}
-{i\over2\Lambda_{_{KK}}p}{\Big[\chi_{_{(++)}}^G({p\over \Lambda_{_{KK}}},\phi)\Big]
\Big[\chi_{_{(++)}}^G({p\over \Lambda_{_{KK}}},\phi^\prime)\Big]\over
R_{_{(++)}}^{G,\epsilon}({p\over \Lambda_{_{KK}}})}
{\partial R_{_{(++)}}^{G,\epsilon}\over\partial z}(z)\bigg|_{z={p\over \Lambda_{_{KK}}}}
\nonumber\\
&&\hspace{-0.6cm}=
-{i\Omega_{_{(++)}}^G(p/\Lambda_{_{KK}})\over\pi\Lambda_{_{KK}}^2}
\bigg\{{\pi tt^\prime\over2R_{_{(++)}}^{G,\epsilon}(p/\Lambda_{_{KK}})}
\Big[\theta(t-t^\prime)\Phi_{_{(uv)}}^{G(+)}(p/\Lambda_{_{KK}},t^\prime)
\Phi_{_{(ir)}}^{G(+)}(p/\Lambda_{_{KK}},t)
\nonumber\\
&&\hspace{0.0cm}
+\theta(t^\prime-t)\Phi_{_{(uv)}}^{G(+)}(p/\Lambda_{_{KK}},t)
\Phi_{_{(ir)}}^{G(+)}(p/\Lambda_{_{KK}},t^\prime)\Big]\bigg\}
+{i\over4\pi p^2}
\label{propagator-3}
\end{eqnarray}
with
\begin{eqnarray}
&&\Omega_{_{(++)}}^G(x)={2R_{_{(++)}}^{G,\epsilon}(x)+x\Big[\Phi_{_{(uv)}}^{G(+)}(x,1)
-\epsilon\Phi_{_{(ir)}}^{G(+)}(x,\epsilon)\Big]\over x^2\Big[N_{_{(uv)}}^{G(+)}(x)\Big]
\Big[N_{_{(ir)}}^{G(+)}(x)\Big]}\;.
\label{propagator-4}
\end{eqnarray}
The factor in the parenthesis $\{\cdots\}$ of the first term in Eq.(\ref{propagator-3})
is the five dimensional mixed position/momentum-space propagator derived in \cite{Randall1},
the factor $\Omega_{_{(++)}}^G/\pi$ originates from the normalization constants of
bulk profiles for gauge fields with $(++)$ BCs in four dimensional effective theory,
the different definitions of couplings between the gauge and fermion fields in five dimensional
full theory and corresponding four dimensional effective theory,
and the difference between the normalization of kinetic terms of relevant fields
in five dimensional full theory and corresponding four dimensional
effective theory, respectively. Additional, $D_{_{(++)}}^{G}(p;\phi,\phi^\prime)$
satisfies the following equation
\begin{eqnarray}
&&\Big\{{1\over r^2}{\partial\over\partial\phi}\Big[e^{-2\sigma(\phi)}
{\partial\over\partial\phi}\Big]+p^2\Big\}D_{_{(++)}}^{G}(p;\phi,\phi^\prime)
=-{i\Omega_{_{(++)}}^G(p/\Lambda_{_{KK}})\over\pi\Lambda_{_{KK}}^2r}\delta(\phi-\phi^\prime)
\label{propagator-5}
\end{eqnarray}
and the corresponding BCs
\begin{eqnarray}
&&{\partial D_{_{(++)}}^{G}\over\partial\phi_{_{(uv)}}}(p;\phi,\phi^\prime)
\bigg|_{\phi_{_{(uv)}}=0}=0\;,
\nonumber\\
&&{\partial D_{_{(++)}}^{G}\over\partial\phi_{_{(ir)}}}(p;\phi,\phi^\prime)
\bigg|_{\phi_{_{(ir)}}=\pi/2}=0\;.
\label{propagator-6}
\end{eqnarray}
For small $z$,
\begin{eqnarray}
&&R_{_{(++)}}^{G,\epsilon}(z)=-{2\ln \epsilon\over\pi}\Big\{1
-{z^2\over4}\Big[1+\epsilon^2+{1-\epsilon^2\over\ln\epsilon}\Big]+{\cal O}(z^4)\Big\}\;,
\nonumber\\
&&{\partial R_{_{(++)}}^{G,\epsilon}\over\partial z}(z)
={z\over\pi}\Big[1-\epsilon^2+(1+\epsilon^2)\ln\epsilon\Big]
\Big\{1-{z^2\over16}{2(1+4\epsilon^2+\epsilon^4)\ln\epsilon+3(1-\epsilon^4)
\over1-\epsilon^2+(1+\epsilon^2)\ln\epsilon}+{\cal O}(z^4)\Big\}\;,
\nonumber\\
&&{t\Phi_{_{(uv)}}^{G(+)}(z,t)\over\Big[N_{_{(uv)}}^{G(+)}
(z)\Big]}={1\over\sqrt{2\pi}}\Big\{1+{z^2\over4}
\Big[t^2(1-2\ln t+2\ln\epsilon)+1+{1-\epsilon^2\over\ln\epsilon}\Big]
\nonumber\\
&&\hspace{2.6cm}
+{z^4\over16}\Big[-{5\over4}t^4+\epsilon^2t^2+(t^4+2\epsilon^2t^2)(\ln t-\ln\epsilon)
\nonumber\\
&&\hspace{2.6cm}
+\Big({1-\epsilon^2\over\ln\epsilon}+1+\epsilon^2\Big)\Big(t^2-2t^2\ln t+2t^2\ln\epsilon\Big)
\nonumber\\
&&\hspace{2.6cm}
+2+{\ln\epsilon\over2}+{47-32\epsilon^2-15\epsilon^4\over16\ln\epsilon}
+{3(1-\epsilon^2)^2\over2\ln^2\epsilon}\Big]+{\cal O}(z^6)\Big\}\;,
\nonumber\\
&&{t\Phi_{_{(ir)}}^{G(+)}(z,t)\over\Big[N_{_{(ir)}}^{G(+)}
(z)\Big]}={1\over\sqrt{2\pi}}\Big\{1+{z^2\over4}
\Big[t^2(1-2\ln t)+\epsilon^2+{1-\epsilon^2\over\ln\epsilon}\Big]
\nonumber\\
&&\hspace{2.6cm}
+{z^4\over16}\Big[-{5\over4}t^4+t^2+(t^4+2t^2)\ln t
+\Big({1-\epsilon^2\over\ln\epsilon}+1+\epsilon^2\Big)\Big(t^2-2t^2\ln t\Big)
\nonumber\\
&&\hspace{2.6cm}
+2\epsilon^4-{\epsilon^4\ln\epsilon\over2}+{15+32\epsilon^2-47\epsilon^4\over16\ln\epsilon}
+{3(1-\epsilon^2)^2\over2\ln^2\epsilon}\Big]+{\cal O}(z^6)\Big\}\;.
\label{propagator-7}
\end{eqnarray}
Inserting the above equations into Eq.(\ref{propagator-3}) and assuming
$p\rightarrow0$, one obtains obviously
\begin{eqnarray}
&&\sum\limits_{n=1}^\infty{\Big[\chi_{_{(++)}}^G(y_{_{(++)}}^{G(n)},\phi)\Big]
\Big[\chi_{_{(++)}}^G(y_{_{(++)}}^{G(n)},\phi^\prime)\Big]\over\Lambda_{_{KK}}^2
\Big[y_{_{(++)}}^{G(n)}\Big]^2}
\nonumber\\
&&\hspace{-0.5cm}=
{1\over8\pi\Lambda_{_{KK}}^2}\Big\{t^2(2\ln t-1)+t^{\prime2}(2\ln t^\prime-1)
\nonumber\\
&&-2\ln\epsilon\Big[t^2\theta(t^\prime-t)+t^{\prime2}\theta(t-t^\prime)\Big]
-{1-\epsilon^2\over\ln\epsilon}\Big\}\;,
\label{propagator-8}
\end{eqnarray}
which is Eq.(34) exactly in Ref.\cite{Casagrande}.
Actually, this result can also be gotten directly by the residue theorem. Applying
Eq.(\ref{lemma3I}) and Eq.(\ref{propagator-7}), we have
\begin{eqnarray}
&&\sum\limits_{n=1}^\infty{\Big[\chi_{_{(++)}}^G(y_{_{(++)}}^{G(n)},\phi)\Big]
\Big[\chi_{_{(++)}}^G(y_{_{(++)}}^{G(n)},\phi^\prime)\Big]\over\Lambda_{_{KK}}^2
\Big[y_{_{(++)}}^{G(n)}\Big]^2}
\nonumber\\
&&\hspace{-0.5cm}=
-{1\over2}{\rm Res}\Big\{\Big[{2\over z^3}+{1\over z^2R_{_{(++)}}^{G,\epsilon}(z)}
{\partial R_{_{(++)}}^{G,\epsilon}\over\partial z}(z)\Big]\Big[\chi_{_{(++)}}^G(z,\phi)\Big]
\Big[\chi_{_{(++)}}^G(z,\phi^\prime)\Big],z=0\Big\}
\nonumber\\
&&\hspace{-0.5cm}=
{1\over8\pi\Lambda_{_{KK}}^2}\Big\{t^2(2\ln t-1)+t^{\prime2}(2\ln t^\prime-1)
\nonumber\\
&&-2\ln\epsilon\Big[t^2\theta(t^\prime-t)+t^{\prime2}\theta(t-t^\prime)\Big]
-{1-\epsilon^2\over\ln\epsilon}\Big\}\;,
\nonumber\\
&&\sum\limits_{n=1}^\infty{\Big[\chi_{_{(++)}}^G(y_{_{(++)}}^{G(n)},\phi)\Big]
\Big[\chi_{_{(++)}}^G(y_{_{(++)}}^{G(n)},\phi^\prime)\Big]\over\Lambda_{_{KK}}^4
\Big[y_{_{(++)}}^{G(n)}\Big]^4}
\nonumber\\
&&\hspace{-0.5cm}=
-{1\over2}{\rm Res}\Big\{\Big[{2\over z^5}+{1\over z^4R_{_{(++)}}^{G,\epsilon}(z)}
{\partial R_{_{(++)}}^{G,\epsilon}\over\partial z}(z)\Big]\Big[\chi_{_{(++)}}^G(z,\phi)\Big]
\Big[\chi_{_{(++)}}^G(z,\phi^\prime)\Big],z=0\Big\}
\nonumber\\
&&\hspace{-0.5cm}=
-{1\over32\pi\Lambda_{_{KK}}^4}\Big\{t^4\Big[\ln t-{5\over4}\Big]+t^2\Big[2+{1-\epsilon^2\over\ln\epsilon}
-{2(1-\epsilon^2)\over\ln\epsilon}\ln t\Big]
\nonumber\\
&&+t^{\prime4}\Big[\ln t^\prime-{5\over4}\Big]+t^{\prime2}\Big[2+{1-\epsilon^2\over\ln\epsilon}
-{2(1-\epsilon^2)\over\ln\epsilon}\ln t^\prime\Big]
\nonumber\\
&&+4t^2t^{\prime2}\Big[\ln t\ln t^\prime-(\ln\epsilon+{1\over2})\ln(tt^\prime)
+{1\over4}+{1\over2}\ln\epsilon\Big]
\nonumber\\
&&-\ln\epsilon\Big[\Big(t^4-4t^2t^{\prime2}\ln t\Big)\theta(t^\prime-t)
+\Big(t^{\prime4}-4t^2t^{\prime2}\ln t^\prime\Big)\theta(t-t^\prime)\Big]
\nonumber\\
&&+{5(1-\epsilon^4)\over8\ln\epsilon}+{(1-\epsilon^2)^2\over\ln^2\epsilon}\Big\}\;,
\label{propagator-9}
\end{eqnarray}
where the second equation is the Eq.(36) exactly in Ref.\cite{Casagrande}.
When the exciting KK modes of the left-handed fields with $(++)$ BCs appear in
relevant Feynman diagrams as virtual particles in four dimensional effective theory,
the amplitudes certainly contain the factor
\begin{eqnarray}
&&{i\Big\{/\!\!\!p\Big[f_{_{(++)}}^{L,c}(y_{_{(\pm\pm)}}^{c(n)},\phi)\Big]
\Big[f_{_{(++)}}^{L,c}(y_{_{(\pm\pm)}}^{c(n)},\phi^\prime)\Big]
+\Lambda_{_{KK}}\Big[y_{_{(\pm\pm)}}^{c(n)}\Big]
\Big[f_{_{(--)}}^{R,c}(y_{_{(\pm\pm)}}^{c(n)},\phi)\Big]
\Big[f_{_{(++)}}^{L,c}(y_{_{(\pm\pm)}}^{c(n)},\phi^\prime)\Big]\Big\}
\over p^2-\Lambda_{_{KK}}^2\Big[y_{_{(\pm\pm)}}^{c(n)}\Big]^2}
\label{propagator-10}
\end{eqnarray}
when we expand them according $\upsilon^2/\Lambda_{_{KK}}^2$. The limit
\begin{eqnarray}
&&\lim\limits_{|z|\rightarrow\infty}\Bigg|{iz/\!\!\!p\Big[f_{_{(++)}}^{L,c}(z,\phi)\Big]
\Big[f_{_{(++)}}^{L,c}(z,\phi^\prime)\Big]
\over p^2-\Lambda_{_{KK}}^2z^2}\Bigg|=0
\label{propagator-11}
\end{eqnarray}
and the function
\begin{eqnarray}
&&\Bigg|{z^2\Lambda_{_{KK}}
\Big[f_{_{(--)}}^{R,c}(z,\phi)\Big]\Big[f_{_{(++)}}^{L,c}(z,\phi^\prime)\Big]
\over p^2-\Lambda_{_{KK}}^2z^2}\Bigg|
\label{propagator-12}
\end{eqnarray}
is uniformly bounded. Since $\pm y_{_{(\pm\pm)}}^{c(1)}$, $\cdots$,
$\pm y_{_{(\pm\pm)}}^{c(n)}$, $\cdots$ are the zeros of the function $R_{_{(\pm\pm)}}^{c,\epsilon}(z)$,
we have the following equation using Eq.(\ref{lemma3I}):
\begin{eqnarray}
&&iD_{_{L(++)}}^{c}(p;\phi,\phi^\prime)
\nonumber\\
&&\hspace{-0.6cm}=
\sum\limits_{n=1}^\infty{i\over p^2-\Lambda_{_{KK}}^2
\Big[y_{_{(\pm\pm)}}^{c(n)}\Big]^2}\bigg\{/\!\!\!p\Big[f_{_{(++)}}^{L,c}(y_{_{(\pm\pm)}}^{c(n)},\phi)\Big]
\Big[f_{_{(++)}}^{L,c}(y_{_{(\pm\pm)}}^{c(n)},\phi^\prime)\Big]
\nonumber\\
&&\hspace{0.0cm}
+\Lambda_{_{KK}}\Big[y_{_{(\pm\pm)}}^{c(n)}\Big]
\Big[f_{_{(--)}}^{R,c}(y_{_{(\pm\pm)}}^{c(n)},\phi)\Big]
\Big[f_{_{(++)}}^{L,c}(y_{_{(\pm\pm)}}^{c(n)},\phi^\prime)\Big]\bigg\}
\nonumber\\
&&\hspace{-0.6cm}=
\sum\limits_{n=1}^\infty{i/\!\!\!p\Big[f_{_{(++)}}^{L,c}(y_{_{(\pm\pm)}}^{c(n)},\phi)\Big]
\Big[f_{_{(++)}}^{L,c}(y_{_{(\pm\pm)}}^{c(n)},\phi^\prime)\Big]\over p^2-\Lambda_{_{KK}}^2
\Big[y_{_{(\pm\pm)}}^{c(n)}\Big]^2}
\nonumber\\
&&\hspace{-0.6cm}=
{i/\!\!\!p\over p^2}\bigg\{\Big[f_{_{(++)}}^{L,c}(0,\phi)\Big]\Big[f_{_{(++)}}^{L,c}(0,\phi^\prime)\Big]
-\Big[f_{_{(++)}}^{L,c}({p\over \Lambda_{_{KK}}},\phi)\Big]
\Big[f_{_{(++)}}^{L,c}({p\over \Lambda_{_{KK}}},\phi^\prime)\Big]\bigg\}
\nonumber\\
&&\hspace{0.0cm}
-{i/\!\!\!p\over2\Lambda_{_{KK}}p}{\Big[f_{_{(++)}}^{L,c}({p\over \Lambda_{_{KK}}},\phi)\Big]
\Big[f_{_{(++)}}^{L,c}({p\over \Lambda_{_{KK}}},\phi^\prime)\Big]\over
R_{_{(\pm\pm)}}^{c,\epsilon}({p\over \Lambda_{_{KK}}})}
{\partial R_{_{(\pm\pm)}}^{c,\epsilon}\over\partial z}(z)\bigg|_{z={p\over \Lambda_{_{KK}}}}
\nonumber\\
&&\hspace{-0.6cm}=
-{i/\!\!\!p\Omega_{_{(++)}}^{L,c}(p/\Lambda_{_{KK}})\over\pi\Lambda_{_{KK}}^2}
\bigg\{{\pi tt^\prime\over2R_{_{(\pm\pm)}}^{c,\epsilon}(p/\Lambda_{_{KK}})}
\Big[\theta(t-t^\prime)\varphi_{_{L_{(uv)}}}^{c(+)}(p/\Lambda_{_{KK}},t^\prime)
\varphi_{_{L_{(ir)}}}^{c(+)}(p/\Lambda_{_{KK}},t)
\nonumber\\
&&\hspace{0.0cm}
+\theta(t^\prime-t)\varphi_{_{L_{(uv)}}}^{c(+)}(p/\Lambda_{_{KK}},t)
\varphi_{_{L_{(ir)}}}^{c(+)}(p/\Lambda_{_{KK}},t^\prime)\Big]\bigg\}
+{i/\!\!\!p\over p^2}\bigg[{-2(1-2c)\epsilon\ln\epsilon\over\pi(1-\epsilon^{1-2c})}\bigg]
{(tt^\prime)^{-c}\over4}
\label{propagator-13}
\end{eqnarray}
with
\begin{eqnarray}
&&\Omega_{_{(++)}}^{L,c}(x)={2R_{_{(++)}}^{G,\epsilon}(x)+x\Big[\varphi_{_{L_{(uv)}}}^{c(+)}(x,1)
-\epsilon\varphi_{_{L_{(ir)}}}^{c(+)}(x,\epsilon)\Big]\over x^2\Big[N_{_{L_{(uv)}}}^{c(+)}(x)\Big]
\Big[N_{_{L_{(ir)}}}^{c(+)}(x)\Big]}\;.
\label{propagator-14}
\end{eqnarray}
Similarly, we also get the summing over the infinite series of exciting KK modes
for right-handed fermion as
\begin{eqnarray}
&&iD_{_{R(++)}}^{c}(p;\phi,\phi^\prime)
\nonumber\\
&&\hspace{-0.6cm}=
\sum\limits_{n=1}^\infty{i\over p^2-\Lambda_{_{KK}}^2
\Big[y_{_{(\mp\mp)}}^{c(n)}\Big]^2}\bigg\{/\!\!\!p\Big[f_{_{(++)}}^{R,c}(y_{_{(\mp\mp)}}^{c(n)},\phi)\Big]
\Big[f_{_{(++)}}^{R,c}(y_{_{(\mp\mp)}}^{c(n)},\phi^\prime)\Big]
\nonumber\\
&&\hspace{0.0cm}
+\Lambda_{_{KK}}\Big[y_{_{(\mp\mp)}}^{c(n)}\Big]
\Big[f_{_{(--)}}^{L,c}(y_{_{(\mp\mp)}}^{c(n)},\phi)\Big]
\Big[f_{_{(++)}}^{R,c}(y_{_{(\mp\mp)}}^{c(n)},\phi^\prime)\Big]\bigg\}
\nonumber\\
&&\hspace{-0.6cm}=
\sum\limits_{n=1}^\infty{i/\!\!\!p\Big[f_{_{(++)}}^{R,c}(y_{_{(\mp\mp)}}^{c(n)},\phi)\Big]
\Big[f_{_{(++)}}^{R,c}(y_{_{(\mp\mp)}}^{c(n)},\phi^\prime)\Big]\over p^2-\Lambda_{_{KK}}^2
\Big[y_{_{(\mp\mp)}}^{c(n)}\Big]^2}
\nonumber\\
&&\hspace{-0.6cm}=
{i/\!\!\!p\over p^2}\bigg\{\Big[f_{_{(++)}}^{R,c}(0,\phi)\Big]\Big[f_{_{(++)}}^{R,c}(0,\phi^\prime)\Big]
-\Big[f_{_{(++)}}^{R,c}({p\over \Lambda_{_{KK}}},\phi)\Big]
\Big[f_{_{(++)}}^{R,c}({p\over \Lambda_{_{KK}}},\phi^\prime)\Big]\bigg\}
\nonumber\\
&&\hspace{0.0cm}
-{i/\!\!\!p\over2\Lambda_{_{KK}}p}{\Big[f_{_{(++)}}^{R,c}({p\over \Lambda_{_{KK}}},\phi)\Big]
\Big[f_{_{(++)}}^{R,c}({p\over \Lambda_{_{KK}}},\phi^\prime)\Big]\over
R_{_{(\mp\mp)}}^{c,\epsilon}({p\over \Lambda_{_{KK}}})}
{\partial R_{_{(\mp\mp)}}^{c,\epsilon}\over\partial z}(z)\bigg|_{z={p\over \Lambda_{_{KK}}}}
\nonumber\\
&&\hspace{-0.6cm}=
-{i/\!\!\!p\Omega_{_{(++)}}^{R,c}(p/\Lambda_{_{KK}})\over\pi\Lambda_{_{KK}}^2}
\bigg\{{\pi tt^\prime\over2R_{_{(\mp\mp)}}^{c,\epsilon}(p/\Lambda_{_{KK}})}
\Big[\theta(t-t^\prime)\varphi_{_{R_{(uv)}}}^{c(+)}(p/\Lambda_{_{KK}},t^\prime)
\varphi_{_{R_{(ir)}}}^{c(+)}(p/\Lambda_{_{KK}},t)
\nonumber\\
&&\hspace{0.0cm}
+\theta(t^\prime-t)\varphi_{_{R_{(uv)}}}^{c(+)}(p/\Lambda_{_{KK}},t)
\varphi_{_{R_{(ir)}}}^{c(+)}(p/\Lambda_{_{KK}},t^\prime)\Big]\bigg\}
+{i/\!\!\!p\over p^2}\bigg[{-2(1+2c)\epsilon\ln\epsilon\over\pi(1-\epsilon^{1+2c})}\bigg]
{(tt^\prime)^{c}\over4}
\label{propagator-15}
\end{eqnarray}
with
\begin{eqnarray}
&&\Omega_{_{(++)}}^{R,c}(x)={2R_{_{(\mp\mp)}}^{c,\epsilon}(x)+x\Big[\varphi_{_{R_{(uv)}}}^{c(+)}(x,1)
-\epsilon\varphi_{_{R_{(ir)}}}^{c(+)}(x,\epsilon)\Big]\over x^2\Big[N_{_{R_{(uv)}}}^{c(+)}(x)\Big]
\Big[N_{_{R_{(ir)}}}^{c(+)}(x)\Big]}\;.
\label{propagator-16}
\end{eqnarray}
For small $z$,
\begin{eqnarray}
&&R_{_{(\pm\pm)}}^{c,\epsilon}(z)={2\epsilon^{c-{1\over2}}(1-\epsilon^{1-2c})
\over(1-2c)\Gamma({1\over2}-c)\Gamma({1\over2}+c)}\Big\{1
-{z^2\over2(1-\epsilon^{1-2c})}\Big[{1-\epsilon^{3-2c}\over3-2c}
+{\epsilon^2-\epsilon^{1-2c}\over1+2c}\Big]
+{\cal O}(z^4)\Big\}\;,
\nonumber\\
&&{\partial R_{_{(\pm\pm)}}^{c,\epsilon}\over\partial z}(z)
=-{2z\epsilon^{c-{1\over2}}\over(1-2c)\Gamma({1\over2}-c)\Gamma({1\over2}+c)}\Big\{
\Big[{1-\epsilon^{3-2c}\over3-2c}+{\epsilon^2-\epsilon^{1-2c}\over1+2c}\Big]
\nonumber\\
&&\hspace{2.4cm}
-{z^2\over2}\Big[{1-\epsilon^{5-2c}\over(5-2c)(3-2c)}
+{2(\epsilon^2-\epsilon^{3-2c})\over(3-2c)(1+2c)}
+{(\epsilon^4-\epsilon^{1-2c})\over(1+2c)(3+2c)}\Big]
+{\cal O}(z^4)\Big\}\;,
\nonumber\\
&&{\sqrt{t}\varphi_{_{L_{(uv)}}}^{c(+)}(z,t)\over\Big[N_{_{L_{(uv)}}}^{c(+)}
(z)\Big]}={t^{-c}\over2}\sqrt{-2(1-2c)\epsilon\ln\epsilon\over\pi(1-\epsilon^{1-2c})}
\bigg\{1-{z^2\over2}\Big[{\epsilon^2\over1+2c}+{t^2\over1-2c}-{2\epsilon^{1-2c}
t^{1+2c}\over(1-2c)(1+2c)}
\nonumber\\
&&\hspace{2.6cm}
-{(1+2c)(1-\epsilon^{3-2c})+(3-2c)(\epsilon^2-\epsilon^{1-2c})\over
(3-2c)(1+2c)(1-\epsilon^{1-2c})}\Big]+{\cal O}(z^4)\bigg\}\;,
\nonumber\\
&&{\sqrt{t}\varphi_{_{L_{(ir)}}}^{c(+)}(z,t)\over\Big[N_{_{L_{(ir)}}}^{c(+)}
(z)\Big]}={t^{-c}\over2}\sqrt{-2(1-2c)\epsilon\ln\epsilon\over\pi(1-\epsilon^{1-2c})}
\bigg\{1-{z^2\over2}\Big[{1\over1+2c}+{t^2\over1-2c}-{2t^{1+2c}\over(1-2c)(1+2c)}
\nonumber\\
&&\hspace{2.6cm}
-{(1+2c)(1-\epsilon^{3-2c})+(3-2c)(\epsilon^2-\epsilon^{1-2c})\over
(3-2c)(1+2c)(1-\epsilon^{1-2c})}\Big]+{\cal O}(z^4)\bigg\}\;.
\label{propagator-17}
\end{eqnarray}
Using Eq.(\ref{propagator-17}), we have
\begin{eqnarray}
&&\sum\limits_{n=1}^\infty{\Big[f_{_{(++)}}^{L,c}(y_{_{(\pm\pm)}}^{c(n)},\phi)\Big]
\Big[f_{_{(++)}}^{L,c}(y_{_{(\pm\pm)}}^{c(n)},\phi^\prime)\Big]\over\Lambda_{_{KK}}^2
\Big[y_{_{(\pm\pm)}}^{c(n)}\Big]^2}
\nonumber\\
&&\hspace{-0.5cm}=
-{1\over2\Lambda_{_{KK}}^2}{\rm Res}\Big\{\Big[{2\over z^3}+{1\over z^2R_{_{(\pm\pm)}}^{c,\epsilon}(z)}
{\partial R_{_{(\pm\pm)}}^{c,\epsilon}\over\partial z}(z)\Big]\Big[f_{_{(++)}}^{L}(z,\phi)\Big]
\Big[f_{_{(++)}}^{L}(z,\phi^\prime)\Big],z=0\Big\}
\nonumber\\
&&\hspace{-0.5cm}=
-{(1-2c)\epsilon\ln\epsilon\over(1-\epsilon^{1-2c})}{(tt^\prime)^{-c}\over4\pi\Lambda_{_{KK}}^2}
\Big\{{2(1-2c)(1-\epsilon^{3-2c})\over(3-2c)(1+2c)(1-\epsilon^{1-2c})}+{t^2+t^{\prime2}\over1-2c}
\nonumber\\
&&-{2\over(1-2c)(1+2c)}
\Big[\theta(t^\prime-t)\Big(t^{\prime1+2c}+\epsilon^{1-2c}t^{1+2c}\Big)
\nonumber\\
&&+\theta(t-t^\prime)\Big(t^{1+2c}+\epsilon^{1-2c}t^{\prime1+2c}\Big)\Big]\Big\}\;,
\nonumber\\
&&\sum\limits_{n=1}^\infty{\Big[f_{_{(++)}}^{R,c}(y_{_{(\mp\mp)}}^{c(n)},\phi)\Big]
\Big[f_{_{(++)}}^{R,c}(y_{_{(\mp\mp)}}^{c(n)},\phi^\prime)\Big]\over\Lambda_{_{KK}}^2
\Big[y_{_{(\mp\mp)}}^{c(n)}\Big]^2}
\nonumber\\
&&\hspace{-0.5cm}=
-{1\over2\Lambda_{_{KK}}^2}{\rm Res}\Big\{\Big[{2\over z^3}+{1\over z^2R_{_{(\mp\mp)}}^{c,\epsilon}(z)}
{\partial R_{_{(\mp\mp)}}^{c,\epsilon}\over\partial z}(z)\Big]\Big[f_{_{(++)}}^{R}(z,\phi)\Big]
\Big[f_{_{(++)}}^{R}(z,\phi^\prime)\Big],z=0\Big\}
\nonumber\\
&&\hspace{-0.5cm}=
-{(1+2c)\epsilon\ln\epsilon\over(1-\epsilon^{1+2c})}{(tt^\prime)^{c}\over4\pi\Lambda_{_{KK}}^2}
\Big\{{2(1+2c)(1-\epsilon^{3+2c})\over(3+2c)(1-2c)(1-\epsilon^{1+2c})}+{t^2+t^{\prime2}\over1+2c}
\nonumber\\
&&-{2\over(1-2c)(1+2c)}
\Big[\theta(t^\prime-t)\Big(t^{\prime1-2c}+\epsilon^{1+2c}t^{1-2c}\Big)
\nonumber\\
&&+\theta(t-t^\prime)\Big(t^{1-2c}+\epsilon^{1+2c}t^{\prime1-2c}\Big)\Big]\Big\}\;.
\label{propagator-18}
\end{eqnarray}
\section{Summing over infinite series of KK modes in unified extra dimension\label{sec5}}
In this section, we depart from the main line above and discuss how to sum over
the infinite series of KK modes in unified extra dimension. In these models all fields
can propagate in all available dimension, the SM particles correspond to
zero modes in the KK decomposition of five dimensional fields with $(++)$ BCs.
The towers of the SM particle KK partners and additional towers of KK modes of
five dimensional fields with $(--)$ BCs do not correspond to any field
in the SM\cite{Arkani-Hamed,Barbieri}. The simplest model of this type
is proposed by Appelquist, Cheng and Dobrescu\cite{ACD} in which the only
additional free parameter relating to the SM is the compactification scale
$1/r$.
Assuming that the topology of the fifth dimension is the orbifold $S^1/Z_2$
and the coordinate $y\equiv x^5$ runs from $0$ to $2\pi r$, one can write
the KK expansions of five dimensional fields respectively as\cite{Buras7}:
\begin{eqnarray}
&&\psi^+(x,y)={1\over\sqrt{2\pi r}}\psi_{_{R(0)}}(x)+{1\over\sqrt{\pi r}}
\sum\limits_{n=1}^\infty\Big\{\psi_{_{R(n)}}(x)\cos{ny\over r}
+\psi_{_{L(n)}}(x)\sin{ny\over r}\Big\}\;,
\nonumber\\
&&\psi^-(x,y)={1\over\sqrt{2\pi r}}\psi_{_{L(0)}}(x)+{1\over\sqrt{\pi r}}
\sum\limits_{n=1}^\infty\Big\{\psi_{_{L(n)}}(x)\cos{ny\over r}
+\psi_{_{R(n)}}(x)\sin{ny\over r}\Big\}\;,
\nonumber\\
&&A^\mu(x,y)={1\over\sqrt{2\pi r}}A^\mu_{_{(0)}}(x)+{1\over\sqrt{\pi r}}
\sum\limits_{n=1}^\infty A^\mu_{_{(n)}}(x)\cos{ny\over r}\;,
\nonumber\\
&&A^5(x,y)={1\over\sqrt{\pi r}}\sum\limits_{n=1}^\infty A^5_{_{(n)}}(x)\sin{ny\over r}\;,
\nonumber\\
&&\phi^+(x,y)={1\over\sqrt{2\pi r}}\phi_{_{(0)}}(x)+{1\over\sqrt{\pi r}}
\sum\limits_{n=1}^\infty\phi_{_{(n)}}^+(x)\cos{ny\over r}\;,
\nonumber\\
&&\phi^-(x,y)={1\over\sqrt{\pi r}}\sum\limits_{n=1}^\infty\phi_{_{(n)}}^-(x)\sin{ny\over r}\;.
\label{unified-1}
\end{eqnarray}
Here, the Dirac spinors $\psi^\pm=(P_R+P_L)\psi^\pm=\psi_L^\pm+\psi_R^\pm$
satisfy the following BCs
\begin{eqnarray}
&&\left\{\begin{array}{l}\partial_y\psi_R^+\Big|_{y=\{0,\pi r\}}=0,\\
\psi_L^+\Big|_{y=\{0,\pi r\}}=0,\end{array}\right.\;\;\; {\rm or}\;\;\;
\left\{\begin{array}{l}\partial_y\psi_L^-\Big|_{y=\{0,\pi r\}}=0,\\
\psi_R^-\Big|_{y=\{0,\pi r\}}=0,\end{array}\right.
\label{unified-2}
\end{eqnarray}
with the chirality projectors $P_{R/L}=(1\pm\gamma_5)/2$. Additional,
the vector fields satisfy the BCs
\begin{eqnarray}
&&\left\{\begin{array}{l}\partial_y A^\mu\Big|_{y=\{0,\pi r\}}=0,\\
A^5\Big|_{y=\{0,\pi r\}}=0,\end{array}\right.
\label{unified-3}
\end{eqnarray}
and the scalar fields $\phi^\pm$ satisfy the BCs
\begin{eqnarray}
&&\left\{\begin{array}{l}\partial_y\phi^+\Big|_{y=\{0,\pi r\}}=0,\\
\phi^-\Big|_{y=\{0,\pi r\}}=0.\end{array}\right.
\label{unified-4}
\end{eqnarray}
In unified extra dimension, the couplings involving KK excitations do not depend
on the bulk profiles since the integral over the 5th coordinate y
can be integrated out explicitly. Correspondingly, the summing over KK excitations
is simplified drastically because it is unnecessary to extend the square
of bulk profile to the complex plane when we apply the residue theorem.
Actually, the authors of Ref.\cite{Buras7} have applied the relation
\begin{eqnarray}
&&\sum\limits_{n=1}^\infty{b\over n^2+c}={b(\sqrt{c}\pi\coth(\sqrt{c}\pi)-1)\over2c}
\label{unified-5}
\end{eqnarray}
to perform the summing over KK excitations in some lower energy processes.
In unified extra dimension, the KK mass $n/r$ is the zero of order one
of the function $\sin(\pi rz)$, and residue of the function $G(z)=\pi r\cos(\pi rz)/\sin(\pi rz)$
at $z=n/r$ is uniform one. Obviously, the function
\begin{eqnarray}
&&f(z)={B\over z^2+C}
\label{unified-6}
\end{eqnarray}
is uniformly bounded because
\begin{eqnarray}
&&\lim\limits_{|z|\rightarrow\infty}\Big|zf(z)\Big|=0.
\label{unified-7}
\end{eqnarray}
Here,
\begin{eqnarray}
&&B={b\over r^2},\;\;C={c\over r^2}\;.
\label{unified-8}
\end{eqnarray}
Using the residue theorem, one obtains obviously
\begin{eqnarray}
&&\sum\limits_{n=1}^\infty{b\over n^2+c}=\sum\limits_{n=1}^\infty{B\over
({n\over r})^2+C}
\nonumber\\
&&\hspace{-0.5cm}=
-{1\over2}{\rm Res}\Big\{{\pi r\cos(\pi rz)\over\sin\pi rz}{B\over
z^2+C},z=0\Big\}
-{1\over2}{\rm Res}\Big\{{\pi r\cos(\pi rz)\over\sin\pi rz}{B\over
z^2+C},z=i\sqrt{C}\Big\}
\nonumber\\
&&-{1\over2}{\rm Res}\Big\{{\pi r\cos(\pi rz)\over\sin\pi rz}{B\over
z^2+C},z=-i\sqrt{C}\Big\}
\nonumber\\
&&\hspace{-0.5cm}=
-{B\over2C}-{B\pi r\cos(i\sqrt{C}\pi r)\over2\sqrt{C}\sin(i\sqrt{C}\pi r)}
={b(\sqrt{c}\pi\coth(\sqrt{c}\pi)-1)\over2c}\;,
\label{unified-9}
\end{eqnarray}
which is the equation presented in Eq.(\ref{unified-5}). Furthermore, the
residue theorem can be applied to sum over the infinite series of KK excitations
in more complicate forms.
\section{The corrections to $B\rightarrow X_s\gamma$ from neutral Higgs in
warped extra dimension\label{sec6}}
\indent\indent
Within the framework of warped extra dimension, the neutral Higgs located on IR brane induces
the mixing between states of the same electric charge, a consequence
of the mixing is that the FCNC transitions are also mediated by neutral
Higgs, the KK excitations of gluon and photon, and the KK excitations of neutral electroweak
gauge bosons besides the charged electroweak gauge bosons $W^\pm$ together with
their KK partners. In this section, we demonstrate that the radiative
correction to the rare decay $b\rightarrow s+\gamma$ from neutral Higgs contains
the suppression factor $m_b^3m_s/m_{_{\rm w}}^4$ comparing with the corrections from other sector.
The effective Hamilton for $B\rightarrow X_s\gamma$ at scales
$\mu={\cal O}(m_b)$ is given by\cite{Buras6}:
\begin{eqnarray}
&&{\cal H}_{eff}(b\rightarrow s\gamma)=-{G_F\over\sqrt{2}}\Big[
\sum\limits_{i=1}^6C_i(\mu){\cal Q}_i(\mu)+C_{7\gamma}(\mu){\cal Q}_{7\gamma}(\mu)
+C_{8G}(\mu){\cal Q}_{8G}(\mu)
\nonumber\\
&&\hspace{3.0cm}+C_{7\gamma}^N(\mu){\cal Q}_{7\gamma}^N(\mu)
+C_{8G}^N(\mu){\cal Q}_{8G}^N(\mu)\Big]\;,
\label{eff_Hamilton}
\end{eqnarray}
with $G_F$ denoting the Fermi constant. The magnetic penguin operators are
\begin{eqnarray}
&&{\cal Q}_{7\gamma}={e\over8\pi^2}m_b\overline{s}_\alpha\sigma^{\mu\nu}
(1+\gamma_5)b_\alpha F_{\mu\nu}\;,
\nonumber\\
&&{\cal Q}_{8G}={g_s\over8\pi^2}m_b\overline{s}_\alpha T^a_{\alpha\beta}
\sigma^{\mu\nu}(1+\gamma_5)b_\beta G_{\mu\nu}^a\;,
\nonumber\\
&&{\cal Q}_{7\gamma}^N={e\over8\pi^2}m_b\overline{s}_\alpha\sigma^{\mu\nu}
(1-\gamma_5)b_\alpha F_{\mu\nu}\;,
\nonumber\\
&&{\cal Q}_{8G}^N={g_s\over8\pi^2}m_b\overline{s}_\alpha T^a_{\alpha\beta}
\sigma^{\mu\nu}(1-\gamma_5)b_\beta G_{\mu\nu}^a\;,
\label{operators}
\end{eqnarray}
and concrete expressions of other dimension six operators can be found in Ref.\cite{Buras6}.
Here $\alpha,\;\beta=1,\;2,\;3$ denote the color indices of quarks, $F_{\mu\nu}$
and $G_{\mu\nu}^a\;(a=1,\;\cdots,\;8)$ are the electromagnetic and strong field
strength tensors, respectively. The magnetic penguin operators ${\cal Q}_{7\gamma},\;
{\cal Q}_{8G},\;{\cal Q}_{7\gamma}^N,\;{\cal Q}_{8G}^N$ in the effective Hamilton
are induced by the virtual heavy freedoms through the one loop diagrams.
As the internal virtual particles are neutral Higgs and electric charge $-1/3$ quarks,
the corresponding corrections to relevant Wilson coefficients can be written as
\begin{eqnarray}
&&C_{7\gamma}(\mu_{_{\rm EW}})={4s_{_{\rm w}}^2Q_f\over e^2}\bigg\{
\sum\limits_{i=1}^3\Big(\xi_{(0)}\Big)_{s,d_i}
\Big(\xi_{(0)}\Big)_{d_i,b}^\dagger F_1(x_{_h},x_{d_i})
\nonumber\\
&&\hspace{2.2cm}
+\sum\limits_{d_i=1}^3{m_{d_i}\over m_b}\Big(\xi_{(0)}\Big)_{s,d_i}^\dagger
\Big(\xi_{(0)}\Big)_{d_i,b}^\dagger F_2(x_{_h},x_{d_i})
\nonumber\\
&&\hspace{2.2cm}
+\sum\limits_{i=1}^3\sum\limits_{n=1}^\infty\Big(\zeta_{(n)}\Big)_{s,D_i}^\dagger
\Big(\zeta_{(n)}\Big)_{D_i,b}F_1(x_{_h},x_{D_i})
\nonumber\\
&&\hspace{2.2cm}
+\sum\limits_{d_i=1}^3\sum\limits_{n=1}^\infty
{m_{D_i}\over m_b}\Big(\zeta_{(n)}\Big)_{s,D_i}
\Big(\zeta_{(n)}\Big)_{D_i,b}F_2(x_{_h},x_{D_i})\bigg\}\;,
\nonumber\\
&&C_{8G}(\mu_{_{\rm EW}})={1\over Q_{_f}}C_{7\gamma}(\mu_{_{\rm EW}})\;,
\nonumber\\
&&C_{7\gamma}^{N}(\mu_{_{\rm EW}})={4s_{_{\rm w}}^2Q_f\over e^2}\bigg\{
\sum\limits_{i=1}^3\Big(\xi_{(0)}\Big)_{s,d_i}^\dagger
\Big(\xi_{(0)}\Big)_{d_i,b}F_1(x_{_h},x_{d_i})
\nonumber\\
&&\hspace{2.2cm}
+\sum\limits_{d_i=1}^3{m_{d_i}\over m_b}\Big(\xi_{(0)}\Big)_{s,d_i}
\Big(\xi_{(0)}\Big)_{d_i,b}F_2(x_{_h},x_{d_i})
\nonumber\\
&&\hspace{2.2cm}
+\sum\limits_{i=1}^3\sum\limits_{n=1}^\infty\Big(\xi_{(n)}\Big)_{s,D_i}^\dagger
\Big(\xi_{(n)}\Big)_{D_i,b}F_1(x_{_h},x_{D_i})
\nonumber\\
&&\hspace{2.2cm}
+\sum\limits_{d_i=1}^3\sum\limits_{n=1}^\infty{m_{D_i}\over m_b}\Big(\xi_{(n)}\Big)_{s,D_i}
\Big(\xi_{(n)}\Big)_{D_i,b}F_2(x_{_h},x_{D_i})
\nonumber\\
&&\hspace{2.2cm}
+\sum\limits_{i=1}^3\sum\limits_{n=1}^\infty\Big(\eta_{(n)}\Big)_{s,D_i}^\dagger
\Big(\eta_{(n)}\Big)_{D_i,b}F_1(x_{_h},x_{D_i})
\nonumber\\
&&\hspace{2.2cm}
+\sum\limits_{d_i=1}^3\sum\limits_{n=1}^\infty
{m_{D_i}\over m_b}\Big(\eta_{(n)}\Big)_{s,D_i}
\Big(\eta_{(n)}\Big)_{D_i,b}F_2(x_{_h},x_{D_i})\bigg\}\;,
\nonumber\\
&&C_{8G}^N(\mu_{_{\rm EW}})={1\over Q_{_f}}C_{7\gamma}^{N}(\mu_{_{\rm EW}})\;,
\label{magnetic_penguin_S1}
\end{eqnarray}
with $Q_{_f}=-1/3,\;x_{d_i}=m_{d_i}^2/m_{_{\rm w}}^2,\;x_{_h}=m_{_h}^2/m_{_{\rm w}}^2$.
The abbreviation $s_{_{\rm w}}=\sin\theta_{_W}$ where $\theta_{_W}$ is the Weinberg angle,
and the functions $F_{1,2}(x,y)$ are defined as
\begin{eqnarray}
&&F_1(x,y)=\Big\{{1\over24}{\partial^3\varrho_{_{3,1}}\over\partial x^3}
-{1\over4}{\partial^2\varrho_{_{2,1}}\over\partial x^2}
+{1\over 4}{\partial\varrho_{_{1,1}}\over\partial x}\Big\}(x,y)\;,
\nonumber\\
&&F_2(x,y)=\Big\{{1\over 4}{\partial^2\varrho_{_{2,1}}\over\partial x^2}
-{1\over2}{\partial\varrho_{_{1,1}}\over\partial x}
+{1\over2}{\partial\varrho_{_{1,1}}\over\partial y}\Big\}(x,y)\;,
\label{magnetic_functions}
\end{eqnarray}
with
\begin{eqnarray}
&&\varrho_{_{m,n}}(x,y)={x^m\ln^nx-y^m\ln^ny\over x-y}\;.
\label{magnetic_function1}
\end{eqnarray}
The couplings $\xi_{(0)}$ depend on the bulk profiles of zero modes
for charge $-1/3$ quarks, and the mixing between zero modes of charge $-1/3$
quarks and corresponding KK partners. To obtain approximately the mixing
between the zero modes of charge $-1/3$ quarks and exciting KK modes, we write
the infinite dimensional column vectors for down type quarks in the chirality
basis as\cite{Buras5}
\begin{eqnarray}
&&\Psi_L(-1/3)=\Big(q_{_{d_L}}^{i(0)}(++),\cdots,q_{_{d_L}}^{i(n)}(++),
D_L^{i(n)}(+-),d_L^{i(n)}(--),\cdots\Big)^T\;,
\nonumber\\
&&\Psi_R(-1/3)=\Big(d_R^{i(0)}(++),\cdots,q_{_{d_R}}^{i(n)}(--),
D_R^{i(n)}(-+),d_R^{i(n)}(++),\cdots\Big)^T\;.
\label{mass1}
\end{eqnarray}
Here, $i=1,\;2,\;3$ is the index of generation, $n=1,\;2,\;\cdots,\;\infty$
is the index of KK modes, the signs in parentheses denote
the BCs satisfied by corresponding fields on UV and IR branes respectively.
We then formally diagonalize the $-1/3$ charge mass matrix and write the mass
eigenvector as
\begin{eqnarray}
&&\Psi_L^{(m)}(-1/3)={\cal D}_L^\dagger\Psi_L(-1/3)\;,
\nonumber\\
&&\Psi_R^{(m)}(-1/3)={\cal D}_R^\dagger\Psi_R(-1/3)\;.
\label{mass2}
\end{eqnarray}
Using above equations, one writes the couplings as
\begin{eqnarray}
&&\Big(\xi_{(0)}\Big)_{d,b}=\sum\limits_{\alpha,\beta=1}^3
\Big[f_{_{(++)}}^{R,c_{_T}^\alpha}(0,{\pi\over2})
f_{_{(++)}}^{L,c_{_B}^\beta}(0,{\pi\over2})\Big]
\Big({\cal D}_R\Big)_{d,\alpha}^\dagger
\Big(\lambda^{d}\Big)_{\alpha\beta}^\dagger\Big({\cal D}_L\Big)_{\beta,b}\;,
\nonumber\\
&&\Big(\xi_{(n)}\Big)_{D_i,b}=\sum\limits_{\alpha,\beta=1}^3\Big[f_{_{(++)}}^{L,c_{_B}^\beta}(0,{\pi\over2})\Big]
\Big({\cal D}_R\Big)_{D_i,9n+\alpha}^\dagger
\Big[f_{_{(++)}}^{R,c_{_T}^\alpha}(y_{_{(\mp\mp)}}^{c_{_T}^\alpha(n)},{\pi\over2})\Big]
\Big(\lambda^{d}\Big)_{\alpha\beta}^\dagger \Big({\cal D}_L\Big)_{\beta,b}\;,
\nonumber\\
&&\Big(\eta_{(n)}\Big)_{D_i,b}=\sum\limits_{\alpha,\beta=1}^3\Big[f_{_{(++)}}^{L,c_{_B}^\beta}(0,{\pi\over2})\Big]
\Big({\cal D}_R\Big)_{D_i,9n-3+\alpha}^\dagger
\Big[f_{_{(-+)}}^{R,c_{_T}^\alpha}(y_{_{(\pm\mp)}}^{c_{_T}^\alpha(n)},{\pi\over2})\Big]
\Big(\lambda^{d}\Big)_{\alpha\beta}^\dagger \Big({\cal D}_L\Big)_{\beta,b}\;,
\nonumber\\
&&\Big(\zeta_{(n)}\Big)_{D_i,b}=\sum\limits_{\alpha,\beta=1}^3\Big[f_{_{(++)}}^{R,c_{_T}^\beta}(0,{\pi\over2})\Big]
\Big({\cal D}_L\Big)_{D_i,9n-6+\alpha}^\dagger
\Big[f_{_{(++)}}^{L,c_{_B}^\alpha}(y_{_{(\pm\pm)}}^{c_{_B}^\alpha(n)},{\pi\over2})\Big]
\Big(\lambda^{d}\Big)_{\alpha\beta}^\dagger \Big({\cal D}_R\Big)_{\beta,b}\;.
\label{mass3}
\end{eqnarray}
For the infinite dimensional column vectors, we have no means of obtaining
the mixing matrices ${\cal D}_{L,R}$ exactly. Adopting the
effective Lagrangian approach\cite{Buras4,Santiago4}, we expand the left-
and right-handed mixing matrices according $\upsilon^2/\Lambda_{_{KK}}^2$
and then approximate the Wilson coefficients $C_{7\gamma}$
in Eq.(\ref{magnetic_penguin_S1}) as
\begin{eqnarray}
&&C_{7\gamma}(\mu_{_{\rm EW}})
={4s_{_{\rm w}}^2Q_f\over e^2}\sum\limits_{i=1}^3\sum\limits_{\alpha=1}^3
\sum\limits_{\beta=1}^3\sum\limits_{\gamma=1}^3\bigg\{
-\Big(\delta_{b,d_i}{m_b^2\over\Lambda_{_{KK}}^2}
+\delta_{s,d_i}{m_s^2\over\Lambda_{_{KK}}^2}\Big)
\nonumber\\
&&\hspace{2.2cm}\times
\Big[f_{_{(++)}}^{R,c_{_T}^\alpha}(0,{\pi\over2})f_{_{(++)}}^{R,c_{_T}^\gamma}(0,{\pi\over2})\Big]
\Big({\cal D}_R^{(0)}\Big)_{s,\alpha}^\dagger\Big(\lambda^d\Big)^\dagger_{\alpha\beta}
\Big(\lambda^d\Big)_{\beta\gamma}\Big({\cal D}_R^{(0)}\Big)_{\gamma,b}
\nonumber\\
&&\hspace{2.2cm}\times
\sum\limits_{n=1}^\infty{1\over\Big[y_{_{(\pm\pm)}}^{c_{_B}^\beta(n)}\Big]^2}
\Big[f_{_{(++)}}^{L,c_{_B}^\beta}(y_{_{(\pm\pm)}}^{c_{_B}^\beta(n)},{\pi\over2})
f_{_{(++)}}^{L,c_{_B}^\beta}(y_{_{(\pm\pm)}}^{c_{_B}^\beta(n)},{\pi\over2})\Big]
\nonumber\\
&&\hspace{2.2cm}
-{m_bm_s\over\Lambda_{_{KK}}^2}\Big(\delta_{b,d_i}+\delta_{s,d_i}\Big)
\Big[f_{_{(++)}}^{L,c_{_B}^\alpha}(0,{\pi\over2})f_{_{(++)}}^{L,c_{_B}^\gamma}(0,{\pi\over2})\Big]
\Big({\cal D}_L^{(0)}\Big)_{s,\alpha}^\dagger\Big(\lambda^d\Big)_{\alpha\beta}
\nonumber\\
&&\hspace{2.2cm}\times
\Big(\lambda^d\Big)_{\beta\gamma}^\dagger
\Big({\cal D}_L^{(0)}\Big)_{\gamma,b}
\sum\limits_{n=1}^\infty\bigg({1\over\Big[y_{_{(\pm\mp)}}^{c_{_T}^\beta(n)}\Big]^2}
\Big[f_{_{(-+)}}^{R,c_{_T}^\beta}(y_{_{(\pm\mp)}}^{c_{_T}^\beta(n)},{\pi\over2})
f_{_{(-+)}}^{R,c_{_T}^\beta}(y_{_{(\pm\mp)}}^{c_{_T}^\beta(n)},{\pi\over2})\Big]
\nonumber\\
&&\hspace{2.2cm}
+{1\over\Big[y_{_{(\mp\mp)}}^{c_{_T}^\beta(n)}\Big]^2}
\Big[f_{_{(++)}}^{R,c_{_T}^\beta}(y_{_{(\mp\mp)}}^{c_{_T}^\beta(n)},{\pi\over2})
f_{_{(++)}}^{R,c_{_T}^\beta}(y_{_{(\mp\mp)}}^{c_{_T}^\beta(n)},{\pi\over2})\Big]\bigg)
\bigg\}F_1(x_{_h},x_{d_i})
\nonumber\\
&&\hspace{2.2cm}
+{4s_{_{\rm w}}^2Q_f\over e^2}\sum\limits_{i=1}^3\sum\limits_{\alpha=1}^3
\sum\limits_{\beta=1}^3\sum\limits_{\gamma=1}^3\bigg\{
-\Big(\delta_{b,d_i}{m_b^2\over\Lambda_{_{KK}}^2}
+\delta_{s,d_i}{m_s^2\over\Lambda_{_{KK}}^2}\Big)
\nonumber\\
&&\hspace{2.2cm}\times
\Big[f_{_{(++)}}^{L,c_{_B}^\alpha}(0,{\pi\over2})f_{_{(++)}}^{L,c_{_B}^\gamma}(0,{\pi\over2})\Big]
\Big({\cal D}_L^{(0)}\Big)_{s,\alpha}^\dagger\Big(\lambda^d\Big)^\dagger_{\alpha\beta}
\Big(\lambda^d\Big)_{\beta\gamma}\Big({\cal D}_L^{(0)}\Big)_{\gamma,b}
\nonumber\\
&&\hspace{2.2cm}\times
\sum\limits_{n=1}^\infty\bigg({1\over\Big[y_{_{(\pm\mp)}}^{c_{_T}^\beta(n)}\Big]^2}
\Big[f_{_{(-+)}}^{R,c_{_T}^\beta}(y_{_{(\pm\mp)}}^{c_{_T}^\beta(n)},{\pi\over2})
f_{_{(-+)}}^{R,c_{_T}^\beta}(y_{_{(\pm\mp)}}^{c_{_T}^\beta(n)},{\pi\over2})\Big]
\nonumber\\
&&\hspace{2.2cm}
+{1\over\Big[y_{_{(\mp\mp)}}^{c_{_T}^\beta(n)}\Big]^2}
\Big[f_{_{(++)}}^{R,c_{_T}^\beta}(y_{_{(\mp\mp)}}^{c_{_T}^\beta(n)},{\pi\over2})
f_{_{(++)}}^{R,c_{_T}^\beta}(y_{_{(\mp\mp)}}^{c_{_T}^\beta(n)},{\pi\over2})\Big]\bigg)
\nonumber\\
&&\hspace{2.2cm}
-\Big(\delta_{b,d_i}{m_bm_s\over\Lambda_{_{KK}}^2}
+\delta_{s,d_i}{m_s^3\over\Lambda_{_{KK}}^2m_b}\Big)\Big[f_{_{(++)}}^{R,c_{_T}^\alpha}(0,{\pi\over2})
f_{_{(++)}}^{R,c_{_T}^\gamma}(0,{\pi\over2})\Big]\Big({\cal D}_R^{(0)}\Big)_{s,\alpha}^\dagger
\Big(\lambda^d\Big)_{\alpha\beta}^\dagger
\nonumber\\
&&\hspace{2.2cm}\times
\Big(\lambda^d\Big)_{\beta\gamma}\Big({\cal D}_R^{(0)}\Big)_{\gamma,b}
\sum\limits_{n=1}^\infty{1\over\Big[y_{_{(\pm\pm)}}^{c_{_B}^\beta(n)}\Big]^2}
\Big[f_{_{(++)}}^{L,c_{_B}^\beta}(y_{_{(\pm\pm)}}^{c_{_B}^\beta(n)},{\pi\over2})
f_{_{(++)}}^{L,c_{_B}^\beta}(y_{_{(\pm\pm)}}^{c_{_B}^\beta(n)},{\pi\over2})\Big]\bigg\}
\nonumber\\
&&\hspace{2.2cm}\times
F_2(x_{_h},x_{d_i})
\nonumber\\
&&\hspace{2.2cm}
+{2s_{_{\rm w}}^2m_{_{\rm w}}^2Q_f\over e^2\Lambda_{_{KK}}^2}\sum\limits_{\alpha=1}^3
\sum\limits_{\beta=1}^3\sum\limits_{\gamma=1}^3\Big[f_{_{(++)}}^{R,c_{_T}^\alpha}(0,{\pi\over2})
f_{_{(++)}}^{R,c_{_T}^\gamma}(0,{\pi\over2})\Big]\Big({\cal D}_R^{(0)}\Big)_{s,\alpha}^\dagger
\Big(\lambda^d\Big)_{\alpha\beta}^\dagger
\nonumber\\
&&\hspace{2.2cm}\times
\Big(\lambda^d\Big)_{\beta\gamma}\Big({\cal D}_R^{(0)}\Big)_{\gamma,b}
\sum\limits_{n=1}^\infty\Big[f_{_{(++)}}^{L,c_{_B}^\beta}(y_{_{(\pm\pm)}}^{c_{_B}^\beta(n)},{\pi\over2})
f_{_{(++)}}^{L,c_{_B}^\beta}(y_{_{(\pm\pm)}}^{c_{_B}^\beta(n)},{\pi\over2})\Big]
\nonumber\\
&&\hspace{2.2cm}\times
F_1({m_{_{\rm w}}^2\over\Lambda_{_{KK}}^2}x_{_h},\Big[y_{_{(\pm\pm)}}^{c_{_B}^\beta(n)}
\Big]^2)+{\cal O}(\upsilon^4/\Lambda_{_{KK}}^4)
\nonumber\\
&&\hspace{1.6cm}=
-{Q_fm_bm_s\over m_{_{\rm w}}^2\Lambda_{_{KK}}^2}\sum\limits_{i=1}^3
\sum\limits_{\beta=1}^3\bigg\{\Big(\delta_{b,d_i}m_b^2+\delta_{s,d_i}m_s^2\Big)
\Big({\cal D}_L^{(0)}\Big)_{s,\beta}^\dagger\Theta_1(c_{_B}^\beta)
\Big({\cal D}_L^{(0)}\Big)_{\beta,b}
\nonumber\\
&&\hspace{2.2cm}
+m_bm_s\Big(\delta_{b,d_i}+\delta_{s,d_i}\Big)
\Big({\cal D}_R^{(0)}\Big)_{s,\beta}^\dagger\Big[\Theta_2(c_{_T}^\beta)
+\Theta_1(-c_{_T}^\beta)\Big]
\Big({\cal D}_R^{(0)}\Big)_{\beta,b}\bigg\}
\nonumber\\
&&\hspace{2.2cm}\times
F_1(x_{_h},x_{d_i})
\nonumber\\
&&\hspace{2.2cm}
-{Q_fm_bm_s\over m_{_{\rm w}}^2\Lambda_{_{KK}}^2}\sum\limits_{i=1}^3\bigg\{
m_bm_s\Big(\delta_{b,d_i}+\delta_{s,d_i}\Big)
\Big({\cal D}_L^{(0)}\Big)_{s,\beta}^\dagger\Theta_1(c_{_B}^\beta)
\Big({\cal D}_L^{(0)}\Big)_{\beta,b}
\nonumber\\
&&\hspace{2.2cm}
+\Big(\delta_{b,d_i}m_b^2+\delta_{s,d_i}m_s^2\Big)
\Big({\cal D}_R^{(0)}\Big)_{s,\beta}^\dagger\Big[\Theta_2(c_{_T}^\beta)
+\Theta_1(-c_{_T}^\beta)\Big]\Big({\cal D}_R^{(0)}\Big)_{\beta,b}\bigg\}
\nonumber\\
&&\hspace{2.2cm}\times
F_2(x_{_h},x_{d_i})+{\cal O}(\upsilon^4/\Lambda_{_{KK}}^4)\;,
\label{mass4}
\end{eqnarray}
Here, ${\cal D}_{L,R}^{(0)}$ are the rotations of zero modes of charge $-1/3$
quarks from chirality eigenvectors to mass eigenvectors in the absence
of the mixing between zero modes and exciting KK modes of quarks.
The nonnegative functions are defined as
\begin{eqnarray}
&&\Theta_1(c)={(1-2c)(1-\epsilon^{3-2c})\over(3-2c)(1+2c)(1-\epsilon^{1-2c})}
+{1\over1-2c}-{1+\epsilon^{1-2c}\over(1-2c)(1+2c)}\;
\nonumber\\
&&\Theta_2(c)=
{1-\epsilon^{3+2c}\over2(3+2c)(1-\epsilon^{1+2c})}
+{\epsilon^2-\epsilon^{1+2c}\over2(1-2c)(1-\epsilon^{1+2c})}
\nonumber\\
&&\hspace{1.6cm}
+{1\over2(1+2c)}+{\epsilon^2\over2(1-2c)}-{\epsilon^{1+2c}\over(1-2c)(1+2c)}\;.
\label{mass5}
\end{eqnarray}
Analogously, we find that the corrections from neutral Higgs to the Wilson coefficients
$C_{8G}(\mu_{_{\rm EW}})$, $C_{7\gamma}^{N}(\mu_{_{\rm EW}})$ and $C_{8G}^N(\mu_{_{\rm EW}})$
all contain the suppression factor $m_b^3m_s/m_{_{\rm w}}^4$
comparing with the corrections from other sectors. The detailed analysis on the radiative
corrections to $b\rightarrow s\gamma$, $(g-2)_\mu$ etc. are presented elsewhere\cite{Feng}.
\section{Summary\label{sec7}}
\indent\indent
In this work, we verify that the eigenvalues of KK excitations in warped extra dimension
are real, and are symmetrically distributed contrasting to the origin in complex plane.
We also present the sufficient condition to judge if the infinite series of KK excitations
is convergent. Applying the residue theorem, we sum over the infinitely series
of KK modes, and analyze the possible relation between summation of the product of
KK mode propagator with the corresponding bulk profiles in four dimensional
effective theory and the propagator of field in five dimensional full theory.
Additional, we also sum over the infinitely series of KK
modes for the gauge boson satisfying the $(++)$ BCs, and recover the results
in literature which are obtained through the equation of motion and completeness
relation of all KK modes. We extend this method to sum over the KK modes
in unified extra dimension, and obtain the equation applied extensively in literature.
As an example, we demonstrate that the radiative correction from the
penguin diagram composed by neutral Higgs and charge $-1/3$ quarks
contains the suppression factor $m_b^3m_s/m_{_{\rm w}}^4$
comparing with the corrections from other sectors.
\begin{acknowledgments}
\indent\indent
The work has been supported by the National Natural Science Foundation of China (NNSFC)
with Grant No. 10975027.
\end{acknowledgments}
|
\section{}
|
\section{Example: generic determinantal varieties}
The symmetric algebra $S(V)$ is Koszul, so we can in particular discuss qoutients of this algebra with almost linear resolutions, and their associated projective schemes.\\
Let $M$ be a matrix of homogeneous polynomials,
\[M=(f_{ij}), \, i=1,\cdots ,n,\,j=1,\cdots ,m.\]
Assume $4\leq n\leq m$.
Let $I$ be the ideal of maximal ({\em i.e.} $n\times n$) minors of $M$.
Consider in particular the case where
\[M=(x_{ij})\]
is a matrix of variables, so that $I$ is the ideal of the generic deteminantal variety in $\mathbb{P}^{nm-1}=\mathbb{P}(V)$, and let $R=k[x_{ij}]/I$ be the homogeneous coordinate ring.\\
The purpose of this section is to present the following example computation using the main theorem.
\begin{thm}
The Ext-algebra of the homogeneous coordinate ring $R$ of the generic determinantal variety in $\mathbb{P}^{nm-1}=\mathbb{P}(V)$ ($4\leq n\leq m$) satisfies
\[\operatorname{Ext}^l_R(k,k)\cong (\Lambda V^\ast \otimes T(\oplus G_i^\vee))_l\]
where $\Lambda V^\ast$ is the exterior algebra and $G_{i+1}$ is the tensor product of the (dual of the) $i$th symmetric power of an $n$-dimensional space and the $(i+n)$th exterior power of an $m$-dimensional space.
The Hilbert series is
\[\operatorname{Hilb} \operatorname{Ext}^\ast_R(k,k)(t) = \frac{(1+t)^{nm}}{1-\binom{m}{n}t^2-n\binom{m}{n+1}t^3-\cdots - \binom{m-2}{m-n-1}\binom{m}{m-1}t^{m-1}-\binom{m-1}{m-n}t^m}.\]
\end{thm}
\begin{proof}
To prove this, we consider the matrix $M$ as a map between free $S(V)$-modules of ranks $m$ and $n$.
Then our algebra $R$ can be considered the cokernel of the $n$th exterior power of this map, and this is resolved by the {\em Eagon-Northcott complex}, see for example Eisenbud's \cite{Eis} Theorem A2.10.
In the same reference, an explicit description of the terms of the Eagon-Northcott complex is given, with $i$th term equal to the tensor product of the (dual of the) $i$th symmetric power of a rank $n$ module and the $(i+n)$th exterior power of a rank $m$ module.
Eisenbud resolves the ideal, not the quotient algebra, so we must change the index by one, and take the generating graded vector space as before.
The differentials in the Eagon-Northcott complex are all linear, so the conditions of Theorem \ref{main} are fulfilled, and the first equality in the theorem follows.\\
The formula for the Hilbert series similarly follows from Corollary \ref{maincor}, by recalling that the Hilbert series of the exterior algebra is
\[\operatorname{Hilb} \Lambda V^\ast=(1+t)^{\operatorname{dim}_k V}=(1+t)^{mn}.\]
\end{proof}
The same proof would give a similar result for other matrices of linear forms, as long as the Eagon-Northcott complex is a resolution.
In particular, consider
\[M=\left(\begin{array}{ccccccc}
x_0 & x_1 & \cdots & x_s & 0 & \cdots & 0\\
0 & x_0 & \cdots & x_{s-1} & x_s & \cdots & 0\\
\vdots & \vdots & & \vdots & \vdots & & \vdots\\
0 & \cdots & x_0 & \cdots & \cdots & x_{s-1} & x_s\\
\end{array}\right),\]
in which case $I$ id the $n$-th power of the irrelevant ideal in $k[x_0,\,\cdots ,x_s]$.
Here $m=n+s$.
In this case, we get that the Hilbert series of the Ext-algebra of the Artin algebra $R=k[x_0,\,\cdots ,x_s]/(x_0,\,x_2,\cdots ,x_s)^n$ is
\[\operatorname{Hilb} \operatorname{Ext}_R^{ast}(k,k)(t) =\frac{(1+t)^s}{1-\binom{n+s}{n}t^2-n\binom{n+s}{n+1}t^3-\cdots - \binom{n+s-2}{s-1}\binom{n+s}{n+s-1}t^{n+s-1}-\binom{n+s-1}{s}t^{n+s}}.\]
See \cite{GKR} for a similar result for the rational normal curve.
\section{Context for the result}
The generalization of the Koszul condition due to Cassidy and Shelton is often easy to check negatively, by considering degrees of Ext-groups (see \cite{CS} page 100, where it is shown that the four-dimensional Artin-Schelter algebras with two generators will not satisfy this condition for degree reasons).
The quotient algebras of Koszul algebras with almost linear resolution pass this test, {\em i.e.} the degrees satisfy the same properties as the degrees of generalized Koszul algebras in this sense.
\begin{question}
Are quotients of Koszul algebras with almost linear resolutions generalized Koszul algebras in the sense of Cassidy and Shelton?
More generally, what is the algebra structure of the Ext-algebra?
\end{question}
The example of the generic determinantal variety studies a commutative algebra, all of whose relations are in a fixed degree.
There is an operadic notion of Koszul algebra over a Koszul operad, and a commutative algebra is Koszul as a commutative algebra if and only if it is Koszul as an associative algebra.
For $N-$Koszul algebras, there aren't any commutative examples (with more than one generator), since the commutativity relation is of degree two.
\begin{question}
Is there an operadic notion of $N-$Koszul algebra over a Koszul operad, with the property that a commutative algebra with all its generetors in degree $N$ is $N-$Koszul if and only if it satisfies the Cassidy-Shelton condition?
\end{question}
If there is such a condition, it might be formulated in terms of operadic cohomology, as the $N-$Koszul condition for associatice algebras can be formulated in terms of Hochschild cohomolgy.
The operadic cohomology for commutative algebras is the Harrison cohomology.
\begin{question}
What is the Harrison cohomology of coordinate ring of the generic determinantal variety?
\end{question}
\section{Introduction}
The Koszul condition for associative algebras was introduced by Priddy \cite{Prid}.
This condition is satisfied by many interesting algebras, and one can often perform explicit computations on algebras with this property.
The condition has been generalized in various ways, including to operads \cite{GK}, to algebras with relations in a fixed higher degree \cite{B01}, and to algebras with relations in any degree \cite{CS}.
Examples of computations in these cases can be found in many articles, including \cite{BF}, \cite{AS87}, \cite{PP}, \cite{FV06}, \cite{BG}, \cite{GMMVZ}, \cite{Val}, \cite{Bar}, \cite{GKR}.
See also \cite{Lod1},\cite{Lod2}.\\
In this article, we will see that it is possible to compute the graded vector space structure on Ext-groups of quotients of Koszul algebras, whose resolution is linear except for the first map, which is homogeneous of a fixed degree $d\geq 4$.
We say that the resolution is {\em almost linear}.
The Ext-groups are isomorphic to graded pieces of the tensor product of the Koszul dual and a tensor algebra based on the resolution, see Theorem \ref{main}.
If $A$ is A Koszul algebra, $R$ the quotient with almost linear resolution, we prove this result by ``reverse engineering'' on the spectral sequence
\[E^{pq}_2=\operatorname{Ext}_R^p(k,\operatorname{Ext}_A^q(R,k))\Rightarrow \operatorname{Ext}_A^{p+q}(k,k).\]
There is one drawback; we have to assume that the degree $d\neq 3$ for our proof to work. Since the case $d=2$ is already known, we will throughout assume that $d\geq 4$.\\
Examples can be found for instance among determinantal varieties, and these examples point towards a possible notion of $N-$Koszul commutative algebra.
\section{Quotients of Koszul algebras}
We fix a Koszul algebra $A$, with a quotient algebra $R$.
\begin{defn}
The resolution of $R$ as a (left) $A$-module is
\[\cdots \rightarrow P_n\rightarrow P_{n-1}\rightarrow \cdots \rightarrow P_1\rightarrow P_0\]
where $P_0=A$.
If the ideal $I=\operatorname{ker}(A\rightarrow R)$ has all its minimal generators in the same degree $d$, and that apart from the first map $P_1\rightarrow P_0$, all differentials are {\em linear}, we say that $R$ has an {\em almost linear resolution.}
Note that this forces the resolution to be minimal.
Thus
\[P_i\cong A(-d-i+1)^{b_i} \text{ for } i\geq 1,\, P_0\cong A.\]
\end{defn}
In this situation, we pick out the generating graded vector space $G_i$ for the $i$th syzygy, so that
\[P_i=G_i \otimes_k A.\]
Thus
\[G_i\cong k(-d-i+1)^{b_i}\]
in its internal grading, but it also has cohomological grading placing it in degree $i+1$. The tensor algebra $T(\oplus G_i^\vee)$, where $()^\vee$ indicates vector space dual, is graded using the cohomological grading with $G_i^\vee$ in degree $i+1$, but also inherits an internal grading from the internal grading of the $G_i^\vee$.
This grading convention will make sense because of the following theorem, and its proof:
\begin{thm}\label{main}
Assume that $R$ is a quotient of a Koszul algebra $A$ with an almost linear resolution, and relations in degree $d\geq 4$.
Then
\[\operatorname{Ext}^m_R(k,k)\cong (A^!\otimes T(\oplus G_i^\vee))_m\]
as (internally) graded vector spaces,
where $A^!$ is the Koszul dual of $A$ and the $G_i$ are the generating graded vector spaces of the syzygies.
\end{thm}
\begin{proof}
We consider the change of rings spectral sequence for $\operatorname{Ext}$ groups:
\[E^{pq}_2=\operatorname{Ext}_R^p(k,\operatorname{Ext}_A^q(R,k))\Rightarrow \operatorname{Ext}_A^{p+q}(k,k).\]
We know the inner Ext-groups from the projective resolution of $R$;
\[\operatorname{Ext}_A^\ast(R,k)\cong\operatorname{Hom}_A(P_\ast,k)\cong G_\ast^\vee.\]
Thus
\begin{equation}\label{produktstruktur}
\operatorname{Ext}_R^\ast(k,\operatorname{Ext}_A^\ast(R,k))\cong\operatorname{Ext}_R^\ast(k, G_\ast^\vee)\cong\operatorname{Ext}_R^\ast(k,k)\otimes G_\ast^\vee.
\end{equation}
In words, the $i$th column in the $E^2$-term is the zeroeth column times the zeroeth row term in the $i$th column.\\
Steps in the proof:
First we consider what is known about the limit and the inital data of this spectral sequence. Our theorem will be proved by identifying the zeroeth row ($q=0$) of the $E_2$-term, since
\[E_2^{p0}=\operatorname{Ext}_R^p(k,\operatorname{Ext}_A^0(R,k))\cong\operatorname{Ext}_R^p(k,k).\]
Second: We show that the differentials $d_m$ are zero whenever the image is not on the zeroeth row, and are injective if the image is on the zeroeth row. We will prove this by using the internal degree only.\\
Third: The results about the differentials imply a recurrence relation on the $\operatorname{Ext}^p_R(k,k)$ which can be solved to give the theorem.\\
First we look at the limit term.
By Koszulity the spectral sequence converges to $A^!$, and the graded pieces of the limit term are concentrated on the zeroeth row of the spectral sequence.
More precisely, $E_\infty^{p0}=A^!_p=\operatorname{Ext}_A^p(k,k)$, with internal degree $-p$.
Since our spectral sequence is in the first quadrant, this limit is achieved for $E_{p+1}$, though the spectral sequence as a whole does not degenerate at any finite step.\\
Then we look at the initial conditions.
The assumption on the resolution of $R$ means that the zeroeth column ($p=0$) in the $E^2$-term is known,
\[E_2^{0q}\cong G_q^\vee\cong k(-q-d+1)^{b_q},\,q>0.\]
Since all generators for the ideal of $R$ in $A$ have degree $d\geq 4$,
\[\operatorname{Ext}_A^1(k,k)\cong\operatorname{Ext}_R^1(k,k)\cong k(-1)^{b_1}.\]
Since there aren't any differentials going into the first column, we get one copy of the zeroeth column, twisted by $(-1)$, for each summand $k(-1)$ in the first Ext-group.
In other words,
\[E_2^{1q}\cong \bigoplus k(-q-d),\,q>0.\]
At this point, we now the two leftmost parts of the $E_2$-term. Writing only the internal degrees, we have that
\[\begin{array}{cccc}
\vdots & \vdots & \vdots & \\
-d-2 & -d-3 & ? &\cdots\\
-d-1 & -d-2 & ? &\cdots\\
-d & -d-1 & ? &\cdots\\
0 & -1 & ? & \cdots
\end{array}\]
We now come to the second stage of the proof.
Consider first the map
\[d_2:E_2^{01}\rightarrow E_2^{20}.\]
This is the only differential going into position $(2,\,0)$, and so the cokernel must be the limit term $A^!_2$. Also, it is the only differential going out of position $(0,\,1)$, so it is necessarily injective:
\[0\rightarrow E_2^{01} \rightarrow E_2^{20} \rightarrow A^!_2 \rightarrow 0.\]
The internal degree of the leftmost term is $-d$, of the rightmost $-2$. Thus the internal degrees of $E_2^{20}$ are $-d$ and $-2$.\\
{\em We will show that the internal degrees at any given point in the spectral sequence are constant modulo d-2}.\\
For the two leftmost columns, this is true (there is only one degree at each point), and at position $(2,\,0)$ the internal degrees are $-d$ and $-2$, both congruent to $-2$ modulo $d-2$. Using Equation \ref{produktstruktur}, we see that
\[E_2^{2q} \text{ has degrees } -d-1-q \text{ and } -2d-q+1,\,\,q\geq 1,\]
so the degrees are constant modulo $d-2$ also in this case, congruent to $-q-3$.
Writing only the internal degrees modulo $d-2$, we now have this image of the three leftmost columns of the $E_2$-term:
\[\begin{array}{ccccc}
\vdots & \vdots & \vdots & \vdots & \\
-d-2 & -d-3 & -d-4,-2d-2 & ? &\cdots\\
-d-1 & -d-2 & -d-3,-2d-1& ? &\cdots\\
-d & -d-1 & -d-2,-2d& ? &\cdots\\
0 & -1 & -2,-d & ? & \cdots
\end{array}\sim \begin{array}{ccccc}
\vdots & \vdots & \vdots & \vdots & \\
-4 & -5 & -6 & ? &\cdots\\
-3 & -4 & -5 & ? &\cdots\\
-2 & -3 & -4 & ? &\cdots\\
0 & -1 & -2 & ? & \cdots
\end{array}\]
At this point, it is clear that the differential $d_2:E_2^{0,q}\rightarrow E_2^{2,\,q-1}$ must be zero for $q>1$, since the source has degree $-q-1$ and the image has degree $-2-q$ modulo $d-2$.\\
We now proceed by induction on the column number $p$, showing three statements at once:
\begin{itemize}
\item[1)] The internal degrees of $E_2^{p0}$ are all congruent to $-p$ modulo $d-2$, the internal degrees of $E_2^{pq}$ are all congruent to $-p-q-1$ modulo $d-2$ for $q\geq 1$ .
\item[2)] The differentials $d_s:E_s^{p-s,s-1}\rightarrow E_s^{p0}$ are injective.
\item[3)] The differentials $d_s:E_s^{p-s,s-1+q}\rightarrow E_s^{pq}$ are zero for $q\geq 1$.
\end{itemize}
The start of the induction is already taken care of; we have seen these properties for $p\leq 2$.\\
Assume that all these properties hold to the left of the $p$-th column.
The possible degrees for $E_2^{p0}$ are $-p$, coming from the $A^!_p$ in the limit, and the degree coming from differentials going into $E_2^{p0}$, namely the degrees of
\[E_2^{p-2,\,1},\,E_2^{p-3,\,2},\,\cdots,\,E_2^{p-k,k-1},\,E_2^{p-p,p-1}=E_2^{0,p-1}.\]
By induction, each of these has internal degrees congruent to
\[-(p-k)-(k-1)-1\cong -p \operatorname{mod} (d-2).\]
Therefore all internal degrees of $E_2^{p0}$ are congruent to $-p$ modulo $d-2$, and by Equation \ref{produktstruktur}, $E_2^{pq}$ has internal degrees congruent to
\[-q-d+1-p\cong-p-q-1 \operatorname{mod} (d-2).\]
1) is therefore true also for column $p$.\\
For property 3), note that the internal degrees of $E_s^{pq}$ must be a subset of the internal degrees of $E_2^{pq}$, which are all congruent to $-p-q-1$ modulo $d-2$ when $q\geq 1$.
The source of the differential $d_s:E_s^{p-s,s-1+q}\rightarrow E_s^{pq}$ likewise has internal degrees congruent to
\[-(p-s)-(s-1+q)-1= -p-q,\]
different from $-p-q-1$ modulo $d-2$ (recall that $d\geq 4$).
Therefore each differential $d_s$ is zero when the target $E_2^{pq}$ has $q\geq 1$, and 3) is proved.\\
Finally, 2) follows since the differential $d_s:E_s^{p-s,s-1}\rightarrow E_s^{p0}$ is the only one starting from position $(p-s,\,s-1)$ that is not zero (by induction), and the limit term $E_\infty^{p-s,\,s-1}$ is zero, so the differential must be injective. This concludes the inductive proof of the three statements.\\
In order to identify to prove that the terms
\[E_2^{p0}\cong \operatorname{Ext}_R^p(k,k)\cong (A^!\otimes T(\oplus G_i^\vee))_p\]
we now use property 2) inductively, to get an equality of graded vector spaces:
\[\begin{array}{rl}
E_2^{p0} \cong & A^!_p\oplus E_2^{p-2,1} \oplus E_2^{p-3,2}\oplus \cdots \oplus E_2^{0,p-1}\\
\cong & A^!_p\oplus (E_2^{p-2,0}\otimes G_1^\vee)\oplus (E_2^{p-3,0}\otimes G_2^\vee)\oplus \cdots \oplus (E_2^{00}\otimes G_{p-1}^\vee)\\
\cong & A^!_p\oplus (A^!\otimes T(\oplus G_i^\vee))_{p-2}\otimes G_1^\vee \oplus (A^!\otimes T(\oplus G_i^\vee))_{p-3}\otimes G_2^\vee \oplus \\
& \oplus \cdots \oplus (A^!\otimes T(\oplus G_i^\vee))_0\otimes G_{p-1}^\vee
\end{array}\]
Since $G_i^\vee$ is in cohomological degree $i+1$, this proves the theorem.
\end{proof}
\begin{cor}\label{maincor}
Under the assumptions of the theorem, the Hilbert series of $\operatorname{Ext}_R^\ast(k,k)$ is the product of the Hilbert series of $A^!$ and $T(\oplus G_i^\vee)$:
\[\operatorname{Hilb}(\operatorname{Ext}_R^\ast(k,k))(t)=\operatorname{Hilb}(A^!)(t)\operatorname{Hilb}(T(\oplus G_i^\vee))(t)=\frac{\operatorname{Hilb}(A^!)(t)}{1-b_1t^2-b_2t^3-b_3t^4-\cdots}.\]
\end{cor}
\begin{remark}
The internal degree can be captured by introducing a second variable $u$, so that we get a two-variable Hilbert series
\[\operatorname{Hilb}(\operatorname{Ext}_R^\ast(k,k))(t,u)=\sum \operatorname{dim}_k\operatorname{Ext}_R^i(k,k)_{-s}t^iu^s,\]
and then the corollary can be refined to
\[\operatorname{Hilb}(\operatorname{Ext}_R^\ast(k,k))(t,u)=\frac{\operatorname{Hilb}(A^!)(tu)}{1-b_1u^{d}t^2-b_2u^{d+1}t^3-b_3u^{d+2}t^4-\cdots}.\]
\end{remark}
\begin{remark}[d=3]
Part 1) of the inductive argument is still true for $d=3$, but being congruent modulo $3-2=1$ is not interesting.
In fact, there will be equal degrees in the source and target of differentials many places in the spectral sequence in case $d=3$, so parts 2) and 3) of the inductive argument can not be proved by looking only on the degrees.\\
On the other hand, the statement of the theorem was found partly by computer experiments in Macaulay2, \cite{M2}, mostly in the case $d=3$.
\end{remark}
\begin{conjecture}
The theorem holds also in case $d=3$.
\end{conjecture}
\begin{remark}[d=2]
The theorem holds for $d=2$, and also the proof, if equality modulo $d-2$ is exchanged by {\em equality}.
This reproves a classical result, see for instance \cite{PP} Chapter 2.5.
\end{remark}
\section{Preliminaries on Koszulity}
We collect a few definitions and results that will be needed later on.
\subsection{Koszul algebras and Koszul duality}
A graded algebra
\[A=\bigoplus_{m\geq 0} A_m\]
with $A_0=k$, the ground field, and generated by the finite dimensional vector space $V=A_1$ is called {\em Koszul} if the minimal resolution of the ground field is linear:
\[ \cdots \rightarrow A(-2)^{a_2}\rightarrow A(-1)^{a_1}\rightarrow A\rightarrow k\]
Under these conditions, we can write $A$ as a quotient of a tensor algebra
\[A=T(A_1)/I,\]
where the ideal $I$ is generated in degree 2, so that
\[I_2\subset A_1\otimes A_1.\]
The Koszul dual algebra is
\[A^!=T(A_1^\ast)/(I_2^\perp),\]
where $V^\ast=\operatorname{Hom}_k(V,k)$ is the vector space dual, and
\[I_2^\perp =\operatorname{ker}(A_1^\ast\otimes A_1^\ast \rightarrow I_2^\ast).\]
When $A$ is Koszul, so is $A^!$. Furthermore,
\[A_m^!\cong \operatorname{Ext}^m_A(k,k),\]
and the algebra structure on $A$ is the same as the Yoneda algebra structure on Ext.
From the grading of $A$, it is clear that the Ext-groups have an internal degree, and in the Koszul case the internal degree of $A_m^!$ is concentrated in degree $-m$ (this is a way to rephrase the linearity).
\subsection{Generalizations of the Koszul condition}
R. Berger introduced the notion of N-Koszul algebras in \cite{B01}. A graded algebra as above, except that the ideal $I$ is generated in degree $N$, is called $N$-Koszul if the graded resolution of the ground field is alternately linear and of degree $N-1$:
\[\cdots \rightarrow A(-2N)^{a_4}\rightarrow A(-N-1)^{a_3} \rightarrow A(-N)^{a_2}\rightarrow A(-1)^{a_1}\rightarrow A\rightarrow k\]
Then the internal grading of the Ext-algebra $\operatorname{Ext}^*_A(k,k)$ places $\operatorname{Ext}_A^i(k,k)$ in a single degree, $-2i$ if $i$ is even, $-2i-1$ if $i$ is odd. In this case, the $A_\infty$-structure on the Ext-algebra, has that the compositions $m_2$ and $m_N$ are the only nonzero compositions, se \cite{FV06}.\\
T. Cassidy and B. Shelton \cite{CS} introduced a more general condition for Koszulity, namely that the associative structure on the Ext-algebra should be generated by $\operatorname{Ext}^1$ and $\operatorname{Ext}^2$.
With the restrictions on degrees of ideal generators for Koszul and $N-$Koszul algebras, this definition exactly captures the Koszul and the $N-$Koszul condition, see for instance \cite{GMMVZ}.\\
The Koszul condition, under its various guises, makes certain computations feasible. See for example \cite{AS87}, \cite{FV06}, \cite{PP}, \cite{BG}, \cite{BF}.
|
\section{Introduction}\label{introduction}
BL Lac objects are characterized by large variability from radio
to TeV frequencies, nearly featureless optical spectra, high
and variable polarization, and in many cases superluminal motion. They
reside in the nuclei of giant elliptical galaxies, where according to
the current paradigm, a supermassive black hole is accreting material
from its surroundings and collimating two relativistic jets in
opposite directions. According to the so-called ``Unified Scheme''
\citep{1995PASP..107..803U}, BL Lac objects are FR I radio galaxies with
one of the jets nearly pointing towards us. Relativistic effects can
boost the jet emission to a level where it almost completely
outshines the host galaxy.
Due to their extreme properties it is not surprising that BL Lacs form
$<$ 1\% of the entire AGN population known today.
More than 100000 confirmed QSO's \citep{2010AJ....139.2360S}
and more than $10^6$ QSO candidates \citep{2009ApJS..180...67R} have been published,
but the number of BL Lac (candidates) hardly exceeds 1000
\citep{2010A&A...518A..10V}.
BL Lac objects are classically
detected in radio- or X-ray surveys and have traditionally been divided
into two groups according to the location of the peak of their
synchrotron emission: the low-energy-peaked BL Lacs (LBL)
and high-energy-peaked BL Lacs (HBL). Until now, the resulting
samples based on single surveys alone only contain a few dozen objects,
irrespective of whether they have been compiled from radio
\citep[e.g. the 1Jy sample by][]{1991ApJ...374..431S} or from
X-ray observations
\citep[e.g. the Einstein Slew Survey sample by][]{1996ApJS..104..251P}.
Samples compiled from a combination of both,
e.g., the ROSAT All-Sky Survey-Green Bank sample by
\citet[][RGB]
{1999ApJ...525..127L} and the Sedentary Survey by
\citet[][]{2005A&A...434..385G} contain a much larger number of objects
(100 - 200). Most important is that they contain
BL Lac objects midway between LBL and HBL.
Regardless of the selection criteria, all samples suffer from a
relatively high number of up to 50\% of
sources with unknown or highly uncertain redshift
and are subject to various biases by their selection criteria.
BL Lacs detected in radio surveys are typically more core-dominated
than those detected in X-rays, and the surveys in different
wavelength regimes have different depths.
Claims have been made that
HBL and LBL may evolve differently
\citep{1991ApJ...374..431S,1991ApJ...380...49M}, but these claims are
subject to low number statistics and to the biases mentioned above,
which cannot be easily corrected for.
Likewise, the general trend in the cosmic evolution of BL
Lacs is not constrained, with claims of negative, positive, or no
evolution \citep[see e.g.][]{2003A&A...401..927B,2007ApJ...662..182P}.
As a consequence, attempts to construct, say, a
luminosity function for BL Lacs and/or their hosts
were limited by small numbers \citep{2007ApJ...662..182P}.
The former is especially important since it allows a
test of the Unified Scheme and can constrain models for jet opening angles as
a function of luminosity by comparing it with the luminosity functions
of other radio-loud AGN.
A viable alternative would be to construct a suitable
optically selected sample of BL Lac objects. Since the optical
region lies between the HBL and LBL peak frequencies,
optically selected samples are
representative of the whole BL Lac population and may be subject
to biases that are easier to control.
However, besides the obvious advantage of optically selected BL Lac samples
over the ones selected from radio and/or X-ray surveys,
only a few attempts have been made to extract an optically
selected sample of BL Lacs. Exactly because BL Lac objects share
optical properties with other sources
\citep[e.g. featureless spectra,
variability, and linear polarization as in the case of magnetic DC white dwarfs,][]
{1978ARA&A..16..487A}, it is very difficult to select and to
confirm candidates from optical data alone.
Early attempts to detect BL Lac objects via optical properties
\citep[e.g.][]{1982MNRAS.201..849I,1984ApJ...276..449B,1993ApJ...404..100J}
have only been moderately successful. Even with the more sophisticated
approach by \citet{2002MNRAS.334..941L}, who
extracted a sample of 56 featureless blue continuum sources with absent
proper motion from the 2-degree field QSO redshift survey
\citep[2QZ,][]{2004MNRAS.349.1397C}, only few a BL Lac objects have been
detected. Follow-up spectroscopy and NIR-imaging has revealed that
most of the sources are
either stellar or extragalactic with faint, but broad emission
features. Only a very few good BL Lac candidates remain
\citep{2005A&A...434..895N,2007MNRAS.374..556L}.
On the other hand,
\citet{2004MNRAS.352..903L} have found an intriguing object within their
sample, which could potentially be a radio-quiet BL Lac object, a class
of objects not believed to exist
\citep[e.g.][]{1990ApJ...348..141S}. This demonstrates the new
discovery space among optically selected samples.
The Sloan Digital Sky Survey (SDSS) offers a unique database
for constructing
such a sample. It is a multi-institutional effort to image
10000 deg$^2$ on the
sky of the north galactic cap in 5 optical filters covering 3800 -
10000 \AA with follow-up moderate-resolution ($\lambda/\Delta \lambda
\sim 1800$) multi-object spectroscopy of about $10^6$ galaxies,
$10^5$ quasars, and a similar number of unusual objects
\citep[see, e.g.][]{2000AJ....120.1579Y}. The SDSS uses a dedicated
2.5m telescope at Apache Point Observatory with a mosaic of 30 CCD cameras
providing a 2\fdg5 field of view \citep[][]{1998AJ....116.3040G}
as well as two multi-object fiber-fed spectrographs allowing to 640 spectra
to be taken simultaneously across a 7\degr field of view
\citep[see, e.g.][]{2002AJ....123..485S}.
Compilations have already been presented by
\citet{2003AJ....126.2209A}, \citet[][A07
hereafter]{2007AJ....133..313A}, and
\citet{2008AJ....135.2453P} but here a cross-correlation with
radio- and/or X-ray properties have been
considered for their derivation.
\citet[][C05 hereafter]{2005AJ....129.2542C} were the first to
present an optically selected sample of 386 BL Lac candidates out of
the SDSS, where
radio- or X-ray properties were a priori not taken into
account\footnote{ ``A priori'' refers to C05 not
cross-correlating his targets with X-ray or radio data bases {\em
before} he extracted the catalog. In fact, 55 of the C05 BL Lac
candidates were
spectroscopically targeted by the SDSS due to their radio/X-ray
properties. The remaining ones are mainly included in the SDSS
spectroscopic database because of their UV excess
\citep[see discussion in
C05, section 5.5 and][section 2]{2007ApJ...663..118S},
so the C05 sample is not completely free of biases.}.
The compilation is divided into a set
of 240 probable and 146 possible candidates each.
Recently, \citet[][P10a hereafter]{2010AJ....139..390P} have presented an
optically selected compilation of 723 BL Lac candidates from the SDSS data
release 7. His approach was similar to the one by C05 (and in
fact, P10a ``recovered'' 226 of the 240 probable BL Lac candidates from
C05).
In order to extract the ``bona fide'' BL Lac objects in the
probable sample of C05, we carried out an extensive program to search for
the two main characteristic properties of BL Lac objects among
the sample, namely variability and polarization.
In a first study of a subset of the sample,
\citet[][S07 hereafter]{2007ApJ...663..118S}
found 24 out of 42 sources to be polarized.
In this paper, we enlarge the study by S07,
present our data set and describe the polarization properties of
182 out of 240 targets of the probable sample of BL Lac candidates of C05.
In combination with the variability characteristics and host galaxy
properties derived from our data, and
broad-band optical-NIR SEDs \citep[using data from the United Kingdom
Infrared Telescope Infrared Deep Sky Survey UKIDSS,][]{2007MNRAS.379.1599L}
all of which
will be determined in a forthcoming paper (Nilsson et al. in prep.),
we will be moving towards the first well-defined optically selected sample
of BL Lac objects unbiased with respect to its radio- or X-ray properties.
In the following sections, we present our observations and describe
the data reduction, followed by analysis and a summary of the results.
We finally end with a discussion, some conclusions and further prospects.
Throughout this paper we use SDSS-magnitudes (which are very
close to AB magnitudes) and a standard cosmology
with $H_0=70$ km/s/Mpc, $\Omega_M=0.3$, and $\Omega_{\Lambda}=0.7$.
\section{Sample selection, observations, and data reduction}
C05 extracted
his sources from a set of over 345,000 individual SDSS spectra
covering 2860 deg$^2$ on the sky \citep[roughly the area covered by the
SDSS Data Release Two, see][]{2004AJ....128..502A}.
For the derivation of the sample, C05 selected quasi-featureless
spectra with the requirement of an S/N of at least $>$ 100 in one
of three spectral regions centered at 4750, 6250, and 7750 \AA. To get
rid of as many likely stellar contaminants as possible (mainly
weak-featured white dwarfs), candidates with significant proper motion
were removed. This resulted in a catalog of 386 BL Lacertae
candidates, which were separated into two subsets with 240
probable and 146 possible
candidates based on their optical colors and other properties, respectively.
``Probable'' signifies a probable
extragalactic nature ($g - r \geq 0.35$ or $r - i \geq 0.13$, or X-ray
or radio counterpart, or measured redshift), while ``possible''
signifies a likely stellar nature ($g - r \leq 0.35$ and $r - i \leq 0.13$
and no indication of extragalactic nature). There are therefore good
reasons to believe that the contamination by stars in the ``probable''
list is very low, while stars are expected to dominate in the
``possible'' list.
For our project, we selected the catalog containing the 240 probable
BL Lacs of C05 since a)
most objects for which radio- and X-ray information is available
cover the region in $\alpha_{\rm ro}$ - $\alpha_{\rm ox}$ space typical of
BL Lacs; b) it contains a
large enough number of candidates for deriving a clean, optically
selected sample, on one hand, but manageable in terms of telescope
time for our polarimetric observations, on the other; c)
it contains a suitable number of radio-weak BL Lac candidates; and d)
only 84 out of the 240 targets in the probable
catalog were listed as BL Lac in NED before C05 published their
sample{\footnote{In the following, ``listed in NED'' refers to
``listed as BL Lac in NED'' when C05 published their sample.}.
Redshifts are available for $>$ 50\% of the
sources (with the majority of them between z = 0 and 1.2).}
The observations were carried out during 16 nights spread over 4 runs
with the ESO-NTT (NTT), the Calar Alto 2.2m (CA), and
Nordic Optical Telescope (NOT) telescopes.
The goal was to observe all 240 sources of the C05 sample, but
since not all nights could be used, we had to prioritize our observations.
Since the observations were scheduled on fixed dates, the selection of
the targets was primarily driven by RA constraints.
Whenever possible, highest priority was given to sources not listed
in NED at a given RA-range before moving to targets without
polarization measurements in the literature. A couple of sources were
observed with two different telescopes to check the reliability of our
analysis.
The observations at the NTT were split into
two separate runs, four nights each, from Oct 2 - 6, 2008 and from Mar 28
- April 1,
2009, respectively. Data could be acquired during 2 1/2 photometric
nights in October and during all four nights (3 of them photometric)
in March/April. Seeing was mostly good (0\farcs6-1\farcs2) during
both runs. We used EFOSC2 attached
to the NTT. The observations were taken through a Gunn-r filter
(\#786), which matched the r'-filter used for the SDSS closest. For the
polarimetric observations, we used the Wollaston prism with a beam
separation of 10\arcsec\ and a
half-wave plate. A 2k Loral-CCD with binning = 2 was employed,
which gave us a field of view of 4\arcmin $\times$ 4\arcmin\
(0\farcs24/pixel). This allowed a suitable number of stars in the field
to be observed at
the same time for characterizing interstellar polarization
along the line of sight. As in all runs, we took care to place the
target at the center of the field of view.
For each target, we took from one to four
sequences at PA = 0, 22.5, 45, and 67.5\degr. Exposure times
ranged from 10 - 1000 sec per PA depending on the brightness of
the source. As in all runs, the goal was to obtain an S/N of $\geq$ 100
to reach an accuracy of 1\% or better. Since EFOSC2 was mounted at the
Nasmyth focus of the NTT, instrumental polarization varying as a
function of the position of the telescope on the sky was
expected. Thus, five to six times per night an unpolarized
standard star from \citet{2007ASPC..364..503F} and three to four times per
night a polarized standard star provided at the ESO-WEB was observed. The
observations of the standard stars were spread homogeneously across the night.
The observations on CA were carried out in service
mode using CAFOS attached to the 2.2m telescope during
five photometric nights with good seeing ($\leq$ 1\farcs5) on
Feb 18 - 24, 2009. Again, a Wollaston prism with a beam separation of
19\arcsec\ and a rotatable $\lambda$/2
plate was used, the observations were taken through a Gunn-r
filter. To save readout time, we only used the
central 1000 $\times$ 1000 pixel of the Site-CCD,
which gave us a field of view of 7\arcmin $\times$ 7\arcmin\
(0\farcs51/pixel).
The layout of the observations was similar to the one
employed at the NTT, except that here only one or two polarized and
unpolarized standards from \citet{1990AJ.....99.1243T} or
\citet{1992AJ....104.1563S}
were observed during each of the nights.
We finally acquired observations of another set of targets using ALFOSC
attached to the NOT, La Palma during three clear nights with mostly
good seeing (0\farcs7 - 1\farcs5) from
April 1 - 4, 2009. Here a calcite plate and a $\lambda$/2
retarder plate were used. The data were taken through an
SDSS-r'filter (NOT Nr. 84). The two beams were separated by 15\arcsec.
The observations were carried out with an E2V-CCD. Due to technical
constraints only the central 1500 $\times$ 650 pixel providing a field
of view of 4\farcm7 $\times$ 2\arcmin\ (0\farcs19/pixel) could be used.
Since both our NTT and CA observations suffered from
substantial instrumental polarization (see section \ref{analysis})
we re-observed 13 sources that had been observed at the NTT and CA as a
``sanity check''. We again used the same observing layout as at the
NTT and CA. Here, one or two polarized and
unpolarized standards were observed each of the night.
The data reduction was similar in all cases. First, the images were
corrected for their bias, and the dark current was proven to be negligible in
all cases. Then we corrected for the pixel-to-pixel variations across the
CCD using either flatfields taken during twilight (NTT and NOT) or
images taken of a homogeneously illuminated screen inside the dome (CA).
The observing log for each source
(telescope used, integration times) is given
in Table \ref{poliresults}. Table \ref{poliresults},
available at the CDS, contains the following information.
Column 1 lists the J2000 coordinates of the source, Column 2 its redshift,
Column 3 whether a source is listed in NED, and Column 4 the SDSS r-mag.
The entries listed in Columns 2-4 are from C05. Column 5 gives the
telescope for the observations used, Column 6 the date of the observations
and Column 7 the exposure time for an individual exposure per
position angle. Columns 8 and 9 give the measured degree of polarization,
as well as the position angle. Finally in Column 10 we give
references to previous measurements of the sources.
\section{\label{analysis}Analysis}
The normalized Stokes parameters $P_Q$ and $P_U$ were computed in the
same way for all three datasets (NTT, Calar Alto, and NOT). We first
used aperture photometry to measure the fluxes in the ordinary and
extraordinary beams in each of the four positions of the Wollaston
prism/calcite. The measurements were made with aperture radii of
1\farcs3 - 3\farcs5 (but mostly between 1\farcs5 and 2\farcs0)
depending on seeing and object brightness.
In addition to the BL Lac candidate, these measurements
also included any sufficiently bright stars present on the CCD
frame, where ``sufficiently bright'' means that the errors of both $P_Q$ and
$P_U$ are smaller than 0.5\%. In this paper we express $P_Q$,
$P_U$ and $P$ in percentage, so a 0.5\% error does not imply an S/N
= 200. We then computed $P_Q$ and $P_U$ using standard
formulae \citep[e.g.][]{2009MNRAS.397.1893V}. In this phase, we checked
that there were no spurious objects inside the measurement aperture.
In the few cases where such an object was detected, the contaminating flux
was measured and subtracted. From $P_Q$ and $P_U$, the degree of
polarization $P$ and polarization position angle PA were then
calculated from $P = \sqrt{P_Q^2 + P_U^2}$ and $PA = 1/2
\tan^{-1}(P_U/P_Q)$.
The errors of $P_Q$ and $P_U$ were computed by propagating the flux
measurement errors through the formulae. Typical errors are $\sim$
0.8\% for $P_Q$, $P_U$ and $P$ and $\sim$ 4$^{\circ}$ for PA. In
addition, small systematic errors are present owing to the correction of
instrumental polarization and a mismatch between the filters used in
this study and the ones used in the literature. As discussed below, the
systematic errors are expected to be smaller that typical errors bars
($<$0.3\% in $P_Q$ and $P_U$ and $<2^{\circ}$ in PA).
Since especially NTT/EFOSC2 was expected to exhibit high
instrumental polarization, we took special care to characterize
possible instrumental effects at all three telescopes. At the NTT we
made observations of zero polarization standards in
\citet{2007ASPC..364..503F} at 62 positions on the sky to map how the
instrumental polarization depends on telescope orientation. In these
measurements the object was placed near the center of the CCD. In
addition, we mapped the instrumental polarization as a function of
position on the CCD by observing three zero polarization standards in
a 5$\times$5 grid over the whole field of view.
Figure \ref{parangplot} shows $P_Q$ and $P_U$ as a function of
parallactic angle of the CCD as measured from zero polarization stars
at the NTT. There is a high instrumental polarization present with an
amplitude of (4.31$\pm$0.02)\%, modulated by the parallactic
angle. We fitted the functions
\begin{eqnarray}
\label{ekay}
P_Q &=& A * \cos(\theta - \theta_0)\\
\label{tokay}
P_U &=& A * \cos(\theta - \theta_0 - \pi/2)
\end{eqnarray}
to the observed data where $\theta$ is the parallactic angle and
$A$ and $\theta_0$ are the fitting variables. The fits are shown as
in Fig. \ref{parangplot}. Subtracting this fit leaves
no significant residuals (Fig. \ref{parangplot}, lower panel).
\begin{figure}
\epsfig{file=16541fg1.eps,width=9cm}
\caption{\label{parangplot} {\em Upper panel}: Dependence of NTT/EFOSC2
instrumental polarization on parallactic angle. Open and closed symbols
refer to the normalized Stokes parameters $P_Q$ and $P_U$, respectively.
The solid and dashed lines are fits to the $P_Q$ and $P_U$
data, respectively (Eqs. \ref{ekay} and \ref{tokay}).
{\em Lower panel}: Residuals after subtracting the fit.}
\end{figure}
After removing the dependence on parallactic angle, we found the
instrumental polarization to also depend on the position on the
CCD. The instrumental polarization was zero at the center of the CCD
and increased to $\sim$0.5\% in the corners. The field dependence was
modeled by fitting two-dimensional polynomials
up to second degree to $P_Q$ and $P_U$ and removed. After this no
significant residuals above the error bars (0.1-0.2\%) were seen.
At Calar Alto we examined the polarization of 360 field stars present
in the CCD frames. These stars are not necessarily unpolarized, but
since they are not at low galactic latitudes, their polarization is
expected to be low, below, or close to our error bars
(0.05-0.5\%). There is a clear dependence of polarization on the
position on the CCD (Fig. \ref{cafieldplot}) reaching 3-4\% near the
corners. As with the NTT data, we fitted up to
second-order, two-dimensional polynomials to $P_Q$ and $P_U$ and
subtracted the fit from the data. After subtraction no residuals above
the rms noise in $P_Q$ and $P_U$ (0.2\%) were seen.
\begin{figure}
\epsfig{file=16541fg2.eps,width=8cm}
\caption{\label{cafieldplot}
Degree of polarization of field stars with the Calar Alto 2.2m
telescope/CAFOS as a function of position on the CCD before (upper
panel) and after (lower panel) applying the corrections. An arrow
with a length of 100 units corresponds to 1\% polarization.}
\end{figure}
The NOT data show no instrumental effects above the rms noise (0.3\%)
of 26 field stars. We can thus say that any remaining instrumental
effects in our whole data set are below 0.2-0.3\% in $P_Q$ and $P_U$,
considerably less than typical error bars in BL Lac candidates
{\bf ($\sim$ 0.8\%).}
To determine whether a target is polarized, we computed the 95\%
confidence limits of the observed degree of polarization $P$ using the
formalism in \citet{1985A&A...142..100S}. If the lower confidence
limit of $P$ is $> 0$, we denote the target as polarized. In this case
the unbiased degree of polarization was computed using the maximum
likelihood estimator in \citet{1985A&A...142..100S}; i.e., $P_{\rm
unbiased} = \sqrt{P^2-1.41*\sigma_P^2}$ where $\sigma_P =
(\sigma_{P_Q} + \sigma_{P_U})/2$. If the target was unpolarized
(i.e. the lower 95\% confidence limit = 0), we used the upper 95\%
confidence limit as the upper limit for the degree of polarization.
The polarization position angle was calibrated by making 21
observations of six highly polarized stars in
\citet{1992AJ....104.1563S} and \citet{2007ASPC..364..503F}. We used
the quoted R-band values to determine the position angle zero
point. The derived zero points have an rms scatter of $\sim$
1$^{\circ}$ internally for each instrument. Since we used filters
that are
slightly different from the R-band and since the polarization position
angle in high polarization standards is typically wavelength-dependent,
though not strongly, a small systematic error in our PA
calibration was expected. We compared the derived r-band zero points to
the ones derived in \cite{2010MNRAS.402.2087V} for the R-band and
found the two to differ by 1.4 $\pm$ 1.3 and by 1.8 $\pm$ 1.1 degrees
for the CA and NOT data, respectively. As a result, any PA offsets due to
filter mismatch are likely to be smaller than 2 degrees.
\section{Results}
The results for each object individually are presented in
Table \ref{poliresults}, while we give
a breakdown of our results in Table \ref{breakdown}
as discussed below.
In total, we have 195
measurements (123 NTT, 47 CA, 25 NOT) of 182 targets. According to
C05, 135 out of 182 targets were not reported in
NED before they published their sample.
Thirteen of our 182 targets were observed twice using the NOT and NTT or CA,
and 10 out of 13 were not listed in NED when C05 published their catalog.
\setcounter{table}{1}
\begin{table*}
\caption{\label{breakdown} Statistics of the observed sample.}
\begin{tabular}{l|c|cc|cc|cc}
\hline
\hline
& N & NED\tablefootmark{1} & not in NED & z\tablefootmark{2} &
no/uncertain z & X or r\tablefootmark{3} & no X or r\\
\hline
Polarized targets& 124 & 39 & 85 & 44 & 80 & 124 & - \\
\hline
Unpolarized targets & 58 & 8 & 50 & 23 & 35 & 39 & 19\\
\hline
\end{tabular}
\tablefoot{
\tablefoottext{1}{NED entry as BL Lac for the target quoted in C05.
\tablefoottext{2}{A reliable redshift exists for the target in}
C05.}
\tablefoottext{3}{X-ray or radio data exist for the target in the literature.}
}
\end{table*}
Out of our 182 targets 124 (68\%) have been found to be polarized and
95 out of the 124 (77\%) polarized objects have been found to be
highly polarized with
P $>$ 4\%\footnote{Following \citet{1993ApJS...85..265J} we set the
border line to distinguish between weakly and highly polarized
objects to 4\%.}.
The average polarization is 7\%, with a substantial tail
of 27 targets whose polarization exceeds 10\% and a maximum
polarization of 22\%. Figure \ref{pdistr} shows the distribution of the
polarization degree of our targets. There is
basically no difference in the distribution of the ``already known''
and really new BL Lac candidates. Indeed, a K-S test confirmed that
the null hypothesis that the two p distributions are drawn from the
same parent population cannot be rejected (significance
0.204).
We also inspected the
distribution of the polarization angles. As expected we do not find
any preferred polarization angle.
\begin{figure}
\epsfig{file=16541fg3.eps,width=8cm}
\caption{\label{pdistr}
Distribution of the degree of polarization of our targets. The
candidates with entries in NED are indicated in black.}
\end{figure}
Only 44 out of 124 (35\%) of our polarized sources have a reliable redshift,
33 have lower limits and/or uncertain redshifts while for 47 targets
no redshift is available at all. The redshifts are taken from the
catalog of C05. Their reliable redshifts are based on at least two spectral
features (mostly but not always from the host galaxy absorption
features), lower limits are derived from intervening absorption
systems (typically $[$ Mg II $]$), while uncertain redshifts
are based on one single
spectral feature.
Our results clearly indicate the need for deep follow-up spectroscopy.
Figure \ref{zpolwoz} shows the distribution of the degree of
polarization for the 44 BL Lac candidates with reliable redshift and the 80
remaining ones in separate panels. Obviously, the two distributions
differ in the sense that the targets with reliable z tend to be less
polarized. The median polarization for the targets with reliable
redshift is 3.8\% and for the remaining ones 7.8\%, respectively.
A KS-test shows that the two distributions are not different at a $<$
0.01\% level. This results can best be explained by contamination of
our polarization measurements by host galaxy light for the sources
with reliable redshifts (see above and discussion in
section \ref{hbllbl}).
In Fig. \ref{z_pol} we show the behavior of our sources in the
polarization - redshift plane. There seems to be a trend toward
sources with higher redshifts to be more polarized roughly up to z
$\sim$ 1. This would indicate
a bias, since at higher redshift only the more beamed sources can be
detected and the dilution by host galaxy light is much lower. (Our
polarization measurements have not yet been corrected for host galaxy
contamination.)
Above z $\sim$ 1, the polarization of the sources seems to drop.
There is one weakly polarized source at a redshift $>$ 3 and a few
more sources where we could derive only upper limits.
These may be high-redshift weak-line QSOs (see discussion in section
\ref{rqbllacs}). However, there are
potentially five more highly and two more weakly polarized sources above
z = 1 in our sample, but their redshifts are marked as uncertain in C05.
To test for a correlation of polarization with redshift,
we calculated the Spearman rank correlation coefficient for all
43 sources with reliable redshifts up to z = 1. We only find $\rho$ =
0.284. A Student's T-test confirmed that $\rho$ is not
significantly different from 0 (${\tau}_{\rho}$ = 1.90). We also tested
whether the polarization properties of the 22 low-z (z $<$ 0.35) and
the 21 high-z (z $>$ 0.35) sources differ. According to a KS-test, the
polarization properties of the two subsamples are not significantly
different. The expected bias is not as pronounced as one may expect. This
could come from the small apertures for our polarimetry used (see
discussion in section \ref{hbllbl}).
\begin{figure}
\epsfig{file=16541fg4.eps,width=8cm}
\caption{\label{zpolwoz}
Distribution of the degree of polarization of our 44 targets with
reliable redshifts (upper panel) and of our 80 targets with either
lower limits, uncertain, or no redshifts at all (lower panel).}
\end{figure}
\begin{figure}
\epsfig{file=16541fg5.eps,width=8cm}
\caption{\label{z_pol} Polarization versus redshift of our targets
with reliable redshifts (black dots), lower limits to the
redshift (right arrows), and upper limits to the polarization (down
arrows). The dashed line marks 4\% polarization. The upper panel
displays the results for the full data set, while the lower panel
shows our results for redshifts up to z = 1 only.}
\end{figure}
All of our 124 polarized sources have a radio detection in either
FIRST
\citep[faint images of the radio sky at twenty centimeters,][]
{1995ApJ...450..559B}
or NVSS \citep[NRAO VLA Sky Survey,][]{1998AJ....115.1693C},
or have recently been detected in deep VLA observations by
\citet[][P10b hereafter]{2010ApJ...721..562P}.
Only 44 (35\%) of our 124 sources are detected in both the radio and X-ray
regime. This is not surprising since the X-ray measurements
were taken from the RASS \citep[ROSAT All-Sky Survey,
e.g.][]{1993Sci...260.1769T}, which is generally much shallover
than the radio surveys used for cross-correlation. Upcoming deep X-ray
surveys like the one planned by eRosita will certainly detect a major
fraction of the sources discussed here. Nevertheless, most
targets that have been found to be polarized are true new ``bona
fide'' BL Lac objects. Only 84 out of 240 sources were listed
as BL Lac in NED at
the time when C05 published his sample.
In the meantime, the number of
NED entries is much larger, but this is exclusively due to newly
presented samples by A07 and P10a.
Remarkable is the breakdown of our 58 unpolarized sources.
Thirtynine of them
(67\%) have either a radio counterpart (25 sources, with one
recently been detected by P10b; SDSS J21155288+000115.5, but still
radio-quiet) or a X-ray detection (1 source), or they have been detected in
radio and X-rays (13 sources).
About 60\% (23 out of 39) of them have a reliable redshift.
The 19 unpolarized sources that have been detected neither at radio nor at
X-ray frequencies
all belong to the potential radio-weak BL Lac candidates
presented by C05. Reliable redshifts are available for 42\% (8 out of 19) of
these sources.
Our data set allows us to inspect some of our targets for
polarization variability. Six targets have been observed with a
separation of a couple of days (NTT and NOT), seven targets with a
separation of about six weeks (CA and NOT), and 21 more targets with a
separation of three to four years (our data and S07). A comparison
of the polarization measurements taken at different epochs is
provided in Fig. \ref{pol_var}. As can be seen, a large fraction
($\sim$ 1/3) of our BL Lac candidates show polarization variability
indeed. Polarization measurements are available in
the literature for six more of our sources.
These are the two 1Jy BL Lacs SDSS J005041.31-092905.1 and
SDSS J014125.83-092843.7 observed by \citet{1986MNRAS.221..739B} and
\citet{1990A&AS...83..183M},
the two EMSS BL Lacs SDSS J020106.18+003400.2
and SDSS J140450.91+040202.2 observed by \citet{1994ApJ...428..130J},
SDSS J105829.62+013358.8 observed by \citet{1988ApJ...333..666I},
and SDSS J121834.93-011954.3 observed by
\citet{2005A&A...433..757S}. For all sources polarization between
5\% and 30\% was detected. Comparing the measurements to ours, all
six sources have shown polarization variability on timescales of years.
\begin{figure}
\epsfig{file=16541fg6a.eps,width=8cm}
\epsfig{file=16541fg6b.eps,width=8cm}
\epsfig{file=16541fg6c.eps,width=8cm}
\caption{\label{pol_var}
Polarization measurements of S07 compared to ours (top), NOT vs. CA
(center) and NOT vs. NTT (bottom). The diagonal line gives the 1:1
correspondence. The arrows indicate upper limits. Polarization
variability for a number of objects is apparent in particular at larger
baselines.
}
\end{figure}
\section{Discussion and conclusions}
\subsection{General comments}
Although BL Lac objects are by definition polarized, have a duty cycle
of at least 40\% to be highly polarized
\citep{1993ApJS...85..265J,1994ApJ...428..130J}, and their polarization
properties can be used to probe radiation processes in their central
engines
\citep[e.g.][]{2008Natur.452..851V,2010MNRAS.402.2087V,2010Natur.463..919F},
polarization measurements have rarely been used to verify them.
So far, only \citet{1982MNRAS.201..849I},
\citet{1984ApJ...276..449B},
\citet{1988ApJ...333..666I},
\citet{1988A&AS...76..145F},
\citet{1990AJ.....99....1K},
\citet{1993ApJ...404..100J},
\citet{1996A&A...307..745K},
\citet{1996MNRAS.281..425M}, and
S07 used polarization measurements
as a diagnostic tool for detecting or confirming BL Lac candidates.
About 124 of the 182 targets observed by us were found to be
polarized, and 95 were highly polarized. S07 observed 21 more sources
not observed by us. All of them were found to be polarized, 15 out of 21
were highly polarized. Since S07 and we concentrated on sources without
an entry in NED, it is not a surprise that the majority of the
remaining sources are already known BL Lacs, e.g., from the RGB sample
by \citet{1999ApJ...525..127L}. Polarization
measurements are published for only eight more
sources\footnote{
SDSS J081815.99+422245.2 \citep{1990AJ.....99....1K},
SDSS J083223.22+491321.0 \citep{1990AJ.....99....1K},
SDSS J105837.74+562811.2 \citep{1996MNRAS.281..425M},
SDSS J123131.40+641418.2 \citep{1993ApJS...85..265J},
SDSS J123739.07+625842.8 \citep{1993ApJS...85..265J},
SDSS J140923.50+593940.7 \citep{1993ApJS...85..265J},
SDSS J150947.97+555617.3 \citep{1996A&A...307..745K},
SDSS J164419.98+454644.4 \citep{1996A&A...307..745K}
},
all of which were found to be polarized and five to be
highly polarized.
In sum, polarization measurements are available for 211 of 240 targets (88\%)
from the C05 sample of probable BL Lac candidates, 153 (64\%) of them
were found to be polarized at least once, and 115 (48\%)
to be highly polarized.
\subsection{\label{hbllbl}HBL versus LBL}
Figure \ref{alphaplot} shows the location of our polarized sources
with both radio and X-ray detection
in the $\alpha_{ox} - \alpha_{ro}$ plane as well as
the degree of polarizarion P as a function of $\alpha_{rx}$.
As in C05, we use $\alpha_{rx}$ = 0.75 as the dividing line between
LBL and HBL.
Obviously, more HBL than LBL are in this figure, which may be because
the radio surveys (NVSS, FIRST) are
much deeper than the X-ray survey (RASS).
Although we are limited by low number statistics here, it seems that the
HBL are only slightly less polarized on average than LBLs.
The average polarization for our
eight LBL in the diagram is 8.7 $\pm$ 1.9\% (mode 6.2\%), while it is
5.6 $\pm$ 0.6\% (mode 5.0\%) for the 37 HBL. As ``errors'' we quote
the standard error of the mean. A K-S test confirmed that our LBL and
HBL do not show a significantly different polarization (significance
0.0758).
The two most highly
polarized LBL are the 1 Jy BL Lac SDSS J005041.31-092905.1 and
SDSS J105829.62+013358.8, both of which were found to be highly polarized
before. On the other hand, a number of strongly ($> 10\%$) polarized sources
are HBL, most of them without any entry in NED.
The object with the highest polarization in our entire data set
(22\%, SDSS J121348.81+642520.2), does not enter the plot here since
only upper limits to its X-ray flux are available.
With $\alpha_{ox} > 1.07
$ and $ \alpha_{ro} = 0.43$ it is very much in the center of the HBL
region. To move it into the LBL region ($\alpha_{ox} \sim 1.4 $ for
$ \alpha_{ro} = 0.43$), its X-ray flux must be a factor 7 below the
upper limit in C05.
Interestingly, 25 of the 37 (68\%) HBL that enter Fig. \ref{alphaplot} were
found to
be $> 4\%$ polarized (and all eight LBL). This seems to be different
from the results obtained
by \citet{1994ApJ...428..130J}, who find a duty cycle
(fraction of time spent with P $>$ 4\% and assuming that the temporal
distribution of polarization is equivalent to the distribution of
polarization measured in all objects of its class)
for HBL to be only about 44\%. Since LBL are stronger
radio emitters than HBL, one would not
expect any strong correlation between optical polarization and radio flux.
As Fig. \ref{radiopol} shows, this is indeed not the case.
The Spearman rank-order correlation coefficient $r_s$
= 0.31; i.e., the two sets of data only show a weak positive correlation.
On the other hand, if we compute an ``error'' for the
duty cycle such that we count the number of of objects that could,
within their 1 $\sigma$ errors, move above and below the 4\% border
as in \citet{1994ApJ...428..130J}, we find a duty cycle for our HBL
of $68^{+2}_{-14}$\%. We note that the average redshift of our
15 HBL with reliable redshift is z = 0.37$\pm$0.17, which
is very similar to the ones obtained for the EMSS or Exosat samples;
i.e., there are good reasons to assume that we are dealing with the
same class of HBL.
Even within the ``error'', we find a higher duty cycle than
\citet{1994ApJ...428..130J}.
There could be several reasons for that.
It might be simply a chance coincidence, but could also be
an effect of large errors in the derivation of the spectral radio -
optical and optical - X-ray indices because the data are not taken
simultaneously.
Most likely, however, it stems from a combination of
the choice of the aperture for measuring the polarimetric fluxes
and of the seeing effects.
\citet{1994ApJ...428..130J} modeled the change in the measured
polarization with respect to the intrinsic polarization as a function
of the Ca II H/K break strength in the spectra of BL Lac objects. For
a reasonable range of break strengths, he predicted a ``depolarization''
of up to 20\%. In Fig. \ref{polihk} we show the Ca II H/K break
strength versus optical polarization for the 50 targets of our sample
(5 LBL, 25 HBL and 20 targets where only upper limits to their X-ray fluxes
are available) with the break strengths published in P10a. Obviously,
highly polarized sources (P $>$ 4\%) can be detected at all break
strengths, but the lower the break strength, the higher the
measured polarization can be. The diagram
basically confirms the prediction by \citet{1994ApJ...428..130J}.
It shows also that HBL do not necessarily populate the righthand side of
the diagram.
\citet{1991AJ....101.1196C}, \citet{2000AJ....119.1534C}, and
\citet{2007A&A...475..199N} have shown that the choice of the aperture
may have a dramatic effect on variability measurements of BL
Lac objects due to the presence of the host galaxy. In addition,
varying seeing while the apterture is kept constant adds further
uncertainties. For their observations \citet{1994ApJ...428..130J}
used mostly a two-holer polarimeter/photometer with aperture
diamaters of 5\arcsec or larger \citep[see Table 4 in][]{1993ApJS...85..265J}.
For the majority of our sources, we instead used apertures with a diameter
of 4\arcsec or smaller (see section
\ref{analysis}). Although our procedure does not completely
remove the host galaxy light (see Fig \ref{polihk}), we expect that the
contamination by stellar light is much less in our case. Thus one
could expect a higher duty cycle for HBL as the one found by
\citet{1994ApJ...428..130J}.
To test this, we redid our polarization analysis with a fixed aperture
diameter of 5\arcsec similar to the one used by
\citet{1994ApJ...428..130J}. Compared to our previous analysis,
the general result did not change much. Now 116 instead
124 out of 182 sources are polarized (64\% instead of 68\%), while 88
out of 116
(76\%) instead of 95 out of 124 (77\%) are highly polarized. Some stronger
differences can be seen when we compare the polarization properties of
our 8 LBL and 36 HBL. Even with a larger aperture, all LBL
are highly polarized. However, of 36 only 29 HBL are polarized and 22 (61\%)
are highly polarized. When we again compute an ``error'' for the duty cycle
as described above, we now find a duty cycle for our HBL of
$61^{+6}_{-14}$\%. This is within the ``error'' very close to the value
derived by \citet{1994ApJ...428..130J}. We are
right now in the process of deblending the polarization measurements
of our BL Lac candidates into the contribution of the AGN and host
galaxy. Potentially, the polarimetry of ``host galaxy free'' LBL and
HBL will show that their duty cycles do not differ. This will be
included in a forthcoming paper.
\begin{figure}
\epsfig{file=16541fg7.eps,width=8cm}
\caption{\label{alphaplot}
Upper panel: Positions of our sources with polarization in the
$\alpha_{ox} - \alpha_{ro}$ plane. The area of the symbol is
proportional to the degree of polarization.
The division line between HBL and LBL is also indicated.
Only the 45 targets with radio and X-ray counterparts are
included here. Lower panel: the degree
of polarizarion P as a function of $\alpha_{rx}$.}
\end{figure}
\begin{figure}
\epsfig{file=16541fg8.eps,width=8cm}
\caption{\label{radiopol} Radio flux versus optical polarization
of our SDSS BL Lac candidates. Open squares denote HBL, crossed
squares HBL, and dots candidates, where only upper limits to
X-ray fluxes exist. Only a weak correlation is apparent. }
\end{figure}
\begin{figure}
\epsfig{file=16541fg9.eps,width=8cm}
\caption{\label{polihk}
Ca II H/K break strength versus optical polarization for 50 of
our BL Lac candidates. The symbols represent the same type of
targets as in Fig. \ref{radiopol}. The dependence of the degree of
polarization on break strength is obvious. HBLs are
distributed across the entire break strength range.}
\end{figure}
\subsection{\label{rqbllacs}Radio-quiet BL Lacs}
The existence of radio-quiet BL Lac objects has a long and
controversial history. While e.g. \citet{1990ApJ...348..141S} did
not found any in the sample of EMSS BL Lac objects,
\citet{2004MNRAS.352..903L} potentially found a candidate from the 2QZ
BL Lac surcey \citep{2002MNRAS.334..941L}. However, as
P10b argued, the investigated samples were
simply too small to find subsets of rarely radio-quiet BL Lac objects.
Alternatively, radio-quiet BL Lac candidates could be low-redshift
counterparts of the (also rare) weak-line QSOs (WLQ), first detected by
\citet{1999ApJ...526L..57F}. These objects have several properties
similar to BL Lac objects, but they do not show strong radio emission and show
very low variability and polarization
\citep[][]{2009ApJ...699..782D}.
They apparently have an
intrinsically weak or absent broad emission line region rather than
being diluted by the beamed emission from a relativistic jet
\citep[e.g.][]{2009ApJ...696..580S,2010ApJ...722L.152S}.
\begin{table}
\caption{\label{rqtab} Properties of radio-quiet BL Lac candidates}
\begin{tabular}{l|c|c|c}
\hline\hline
SDSS J & P & P (S07) & P10b\\
& [\%]& [\%]& \\
\hline
004054.65-091526.8 & $<$ 4.0& & \\
012155.87-102037.2 & $<$ 0.9& & Unknown \\
013408.95+003102.5 & $<$ 2.0& & Star \\
020137.66+002535.1 & $<$ 2.2& $<$ 5.0 & Star \\
024156.38+004351.6 & 2.6 & & Unknown \\
024157.37+000944.1 & 4.0 & & BL Lac\\
025046.48-005449.0 & $<$ 3.0& & Star \\
025612.47-001057.8 & $<$ 2.4& & Unknown? \\
031712.23-075850.4 & $<$ 3.2& (0.7)\tablefootmark{a} & \\
090133.43+031412.5 & $<$ 2.1& & Unknown \\
104833.57+620305.0 & $<$ 1.7& $<$ 2.3 & Unknown? \\
114153.35+021924.4 & $<$ 1.3& (0.9) & \\
121221.56+534128.0 & $<$ 0.7& (1.3) & \\
123743.09+630144.9 & $<$ 1.4& (1.2) & \\
124225.39+642919.1 & $<$ 1.5& & Galaxy \\
133219.65+622715.9 & $<$ 1.7& & \\
142505.61+035336.2 & $<$ 0.6& $<$ 0.9 & \\
150818.97+563611.2 & $<$ 3.7& & Unknown \\
151115.49+563715.4 & $<$ 5.1& & Absorbed AGN \\
154515.78+003235.2 & $<$ 5.6& (1.1) & Unknown \\
165806.77+611858.9 & $<$ 5.6& 1.0 & Absorbed AGN \\
211552.88+000115.5 & $<$ 2.2& & WLQ \\
212019.13-075638.4 & $<$ 3.2& & Star? \\
213950.32+104749.6 & $<$ 2.4& & Galaxy \\
224749.55+134248.2 & & 0.8 & WLQ \\
231000.81-000516.3 & $<$ 3.2 & & Unknown? \\
232428.43+144324.4 & & 0.9 & WLQ \\
\hline
\end{tabular}
\tablefoot{
\tablefoottext{a}{
The measurements of targets polarized according
to S07 but unpolarized
according to our 95\% criterion are given
in parenthesis.
}
}
\end{table}
C05 extracted a set of 27 radio-quiet BL Lac candidates out of his
probable sample of 240 BL Lac objects (a new set of 86 weak-featured
radio-quiet objects has more recently been presented by
P10a). We have polarization measurements for
25 out of 27 of them, which are summarized in Table \ref{rqtab}.
For comparison, the results from S07 are given using our 95\% criterion
of whether a source is polarized or not and from P10b, who presented
new target identifications.
All except SDSS J024156.38+004351.6 (weakly polarized
with P = 2.6\%) and SDSS J024157.37+000944.1 (highly polarized with P =
4.0\%, respectively) were found to be unpolarized by us.
S07 observed 11 out of 27 candidates and
found that three are weakly polarized (SDSS J165806.77+611858.9 with P =
1.0\%, SDSS J224749.55+134248.2 with P = 0.8\% and SDSS
J232428.43+144324.4 with P = 0.9\%, respectively) according to our
95 \% criterion. S07 and we have nine targets in common.
Except SDSS J165806.77+611858.9, which S07 found to be
polarized, while we could only derive upper limits,
the remaining ones did not show any polarization in
either case. Altogether S07 and we observed all 27 radio-quiet BL
Lac candidates and found four to be weakly polarized and one to be
highly polarized. Recently, P10b has
examined 20 candidates from C05 via variability measurements using
SDSS data, new spectral classifications, proper motions, and
new radio and X-ray measurements. They classified three
targets as low-redshift WLQ, two as absorbed AGNs, four as stars, two
as galaxies, eight as unknown, and only one as radio-loud BL Lac.
Our observations confirm that the number of reliable radio-quiet BL Lac
objects in the C05 sample must be low. The only highly polarized
radio-quiet BL Lac candidate
(SDSS J024157.37+000944.1) was classified as
radio-loud BL Lac by P10b based on new VLA
measurements, while the remaining four weakly polarized radio-quiet BL
Lac candidates were classfied as unknown,
absorbed AGN or WLQ.
As already stated by P10b and a few times in the
present paper, besides the lack of
reliable spectroscopic redshifts for many targets, deep X-ray
observations would be very useful for properly evaluating the contents
of the C05 sample.
\subsection{Final comments}
In summary, we have found that
$\bullet$ 124 out of 182 (68\%) of our targets were polarized, and 95 out
of the 124 polarized targets (77\%) to be highly polarized ($>$ 4\%).
$\bullet$ Only 44 out of 124 (35\%) of our polarized
sources have a reliable redshift. There is
a clear need for follow-up high S/N spectroscopy.
$\bullet$ LBL are on average only slighly more strongly
polarized than HBL.
We found a higher duty cycle
of polarization in HBL ($ \sim 66\%$ have polarization $> 4\%$)
than in \citet{1994ApJ...428..130J}. This may be due to the different
apertures used in the analysis.
$\bullet$ Our data do not give any evidence for the presence
of radio-quiet BL Lac objects in the sample of C05.
By just using our (and S07) polarization
measurements, we find strong evidence that the sample selected
by C05 indeed contains a large number of bona fide BL Lacs. At least
70\% of the sources were found to be polarized. Since even very prominent
BL Lacs like OJ 287 are unpolarized from time to time
\citep[e.g.][]{2010MNRAS.402.2087V}, the number of bona fide BL Lac objects
in the sample is presumably even higher. We have further
options for testing this result. First of all, BL Lac objects are variable, with
LBL and HBL having duty cycles of $\sim$ 40 and 80\%, respectively
\citep{1996A&A...305...42H,1998A&A...329..853H}.
Since our data were taken through filters similar to the one
used by the SDSS, we can look for variability in our targets on timescales of
years. In addition, we can use our data to analyze the images for the
presence of a core repesenting an AGN and a host galaxy. This would
demonstrate that these targets where a host galaxy has been found
are extragalactic in nature. Finally, we can use the SDSS and UKIDSS
to construct broad-band SEDs of our targets and to elegantly
identify stars, which may still contaminate the sample,
by their blackbody continuum radiation. All three of the above are in progress.
Along with our polarization analysis, we have four diagnostic tools at
hand to identify the BL Lac content of the
sample of C05. This will be the scope of a forthcoming paper
(Nilsson et al., in prep.).
Like all BL Lac samples, the C05 sample seriously suffers from the
lack of reliable redshifts, which are a prerequisite for the construction
of a lumnosity function of BL Lacs, among other possibilities.
It is thus no surprise that
BL Lac luminosity functions still suffers from low number statistics
\citep[e.g.][]{2007ApJ...662..182P}. The C05 targets
were selected with
the requirement of having S/N $>$ 100 over at least one
of three 500 \AA\ wide spectral bands in the SDSS spectra.
The S/N per resolution
element is thus much lower than 100 for many targets.
\citet{2006AJ....132....1S} has
shown that S/N $>$ 100 per resolution element is required for
detecting the very faint emission lines and/or host galaxy absorption
features. We are now in the process of collecting high S/N spectra for all
C05 sources that we found to be polarized and whose redshift is
uncertain or unknown.
In combination with the results from our analysis, we will
then be able to derive the necessary steps towards constructing
the first luminosity function of an optically selected sample of
BL Lac objects.
\begin{acknowledgements} We would like to thank the anonymous referee
for constructive and detailed suggestions, that improved the
presentation of
our results. We would also like to thank the staff at Calar Alto
for collecting the excellent data in Service Mode, as well as the staff at the NTT
and the NOT for their superb support during the observations. JH acknowledges
support by the Deutsche Forschungsgemeinschaft (DFG) through grant HE 2712/4-1.
\end{acknowledgements}
\bibliographystyle{aa}
|
\section{Introduction}
These notes arise from the attempt to extend the results of \cite{marcello-bolo-conicbd} to a wider class of complex threefolds with
negative Kodaira dimension. If $Y \to S$ is a conic bundle and $S$ is rational, a semiorthogonal decomposition of ${\rm D}^{\rm b}(Y)$
by derived categories of curves and exceptional objects gives a splitting of the intermediate Jacobian as the direct
sum of the Jacobians of the curves \cite[Thm 1.1]{marcello-bolo-conicbd}. This result is based on the relation between fully
faithful functors ${\rm D}^{\rm b}(\Gamma) \to {\rm D}^{\rm b}(Y)$ (where $\Gamma$ is a smooth projective curve) and algebraic cycles on $Y$.
First of all, using the Chow--K\"unneth decompositions of the motives of $\Gamma$ and $Y$, we get from such functor an isogeny
between $J(\Gamma)$ and an abelian subvariety of $J(Y)$. Secondly, using the universal isomorphism between $A^2_{\mathbb{Z}}(Y)$ and a Prym variety
and the incidence property, we prove that such an isogeny is indeed an injective morphism of principally polarized
abelian varieties. Finally, the existence of the mentioned semiorthogonal decomposition assures the splitting of the intermediate Jacobian.
It turns out that the properties needed to prove this result are enjoyed also by threefolds other than conic bundles. One of the aims of this paper is to describe certain varieties satisfying those representability assumptions.
In a generalization attempt, we define a new notion of representability based on semiorthogonal decompositions, which we expect to
carry useful geometrical insights also in higher dimensions.
Let $X$ be a smooth projective variety of dimension $n$. We define \it categorical representability in
(co)dimension $m$ \rm for $X$, roughly by requiring that the derived category ${\rm D}^{\rm b}(X)$ admits a semiorthogonal decomposition
by categories appearing in smooth projective varieties of dimension $m$ (resp. $n-m$).
The easiest case is of course representability in dimension 0. This is equivalent to say that $X$ admits
a full exceptional sequence of a finite number, say $l$, of objects. In this case we have $K_0(X) = \mathbb{Z}^l$. This
is indeed a very strong notion and gives rise to intriguing questions to explore even for surfaces.
Various notions of representability of the group $A^i_{\mathbb{Z}}(X)$ of algebraically trivial cycles of codimension
$i$ on $X$ have appeared throughout the years in the literature, and it seems interesting to understand
their interactions with categorical representability, as our examples suggest.
Roughly speaking, \it weak representability \rm for $A^i_{\mathbb{Z}}(X)$ is given by an algebraic map
$J(\Gamma) \to A^i_{\mathbb{Z}}(X)$ whose kernel is an algebraic group, for an algebraic curve $\Gamma$. Working
with rational coefficients (that is, with $A^i_{\mathbb{Q}}$) gives the notion of \it rational representability.
Algebraic representability \rm
requires the existence of a universal regular isomorphism $A^i_{\mathbb{Z}}(X) \to A$ onto an abelian variety $A$.
Finally, if $\dim(X)=2n+1$ is odd, $A$ is the algebraic representative of $A^n_{\mathbb{Z}}(X)$, and the principal polarization
of $A$ is ``well behaved`` with respect to this regular isomorphism
we say that $A$ carries an \it incidence polarization. \rm
The definition of categorical representability could seem rather disjoint from the classical ones.
It is nevertheless clear that rational representability is strongly related to the structure of the motive of $X$.
For example, if $X$ is a threefold, then rational representability of all the $A^i_{\mathbb{Q}}(X)$ is equivalent to the existence of
a specific Chow--K\"unneth decomposition \cite{gorch-gul-motives-and-repr}.
A first point to note is then that fully faithful functors should hold motivic maps, as stated in the following conjecture by Orlov.
\begin{conjecture}[\cite{orlov-motiv}]\label{orlov-conj}
Let $X$ and $Y$ be smooth projective varieties and $\Phi: {\rm D}^{\rm b}(Y) \to {\rm D}^{\rm b}(X)$ be a fully faithful functor. Then
the motive $h(Y)$ is a direct summand of the motive $h(X)$.
\end{conjecture}
In order to get a link between categorical and rational representability, we should consider the former in
dimension 1. Note that being categorically representable in dimension 1 is equivalent to the existence
of a semiorthogonal decomposition by exceptional objects and derived categories of curves.
Orlov conjecture would then imply that if $X$ is categorically representable in dimension 1, then
its motive is a finite sum of abelian and discrete motives, and this would give informations about rational representability
for $A^i_{\mathbb{Q}}(X)$. Being categorically representable in dimension 1
seems to be in fact a very strong condition. For example a smooth cubic threefold is strongly representable
with incidence property but not categorically representable, otherwise we would have the splitting
of the intermediate Jacobian (see Corollary \ref{cor-no-split-no-fm}).
Notice that in \cite{kuzne-manivel-marku} the study of the Abel--Jacobi map for some hypersurfaces and
its link with categorical constructions were already treated.
Algebraic representability and the incidence property can have deep interactions with
categorical representability, and this is indeed the heart of the proof of Theorem 1.1 in \cite{marcello-bolo-conicbd}.
Consider a smooth projective threefold $X$ and assume it to be
rationally representable, with $h^1(X)=h^5(X)=0$, and with $A^2_{\mathbb{Z}}(X)$ algebraically representable with the incidence property. The arguments
in \cite{marcello-bolo-conicbd} show that if $X$ is categorically representable in dimension 1, then
the intermediate Jacobian $J(X)$ splits into Jacobians of curves, namely of those curves of positive genus appearing in the
semiorthogonal decomposition. This result can then be applied to a large class of complex threefolds with negative
Kodaira dimension (see a list in Remark \ref{remark-list-of-3folds}).
We can then reasonably raise the following question, which also points out how this new definition could be useful
in higher dimensions: is categorical representability in codimension 2 a necessary condition
for rationality? This is true for curves (where we have to replace codimension 2 with dimension 0) and
for surfaces, since any rational smooth projective surface admits a full exceptional sequence.
Remark \ref{list-of-iff} shows that this is true for a wide class of complex threefods with negative Kodaira dimension, but we can only
argue so far by a case by case analysis. Categorical representability should moreover hold the
vanishing of the Clemens--Griffiths component of ${\rm D}^{\rm b}(X)$ mentioned in \cite{kuznetcubicfourfold}.
We can wonder if Kuznetsov's conjecture about rationality of cubic fourfold (\cite[Conj. 1.1]{kuznetcubicfourfold}) could then
be restated as follows: a cubic fourfold is rational if and only if it is categorically representable in
codimension 2. Finally, we can argue some conjectural relation between
categorical representability and the existence of gaps in the Orlov spectrum defined
in \cite{katza-favero-ballard-generation-time}.
\subsection*{Notations}
Any triangulated category is assumed to be essentially small.
Given a smooth projective variety $X$, we denote $\kappa_X$ its Kodaira dimension, ${\rm D}^{\rm b}(X)$ the bounded derived
category of coherent sheaves on it, $K_0(X)$ its Grothendieck group, $CH_{\mathbb{Z}}^d(X)$ the Chow group of
codimension $d$ cycles, and $A_{\mathbb{Z}}^d(X)$ the subgroup of algebraically trivial cycles in $CH_{\mathbb{Z}}^d(X)$.
If is $X$ pure $d$-dimensional, and $Y$ any smooth projective variety,
we denote by ${\rm Corr}^i(X,Y):= CH_{\mathbb{Q}}^{i+d}(X\times Y)$ the group of correspondences with rational
coefficients. If $X=\coprod X_j$, with $X_j$ connected, then
${\rm Corr}^i(X,Y) = \oplus {\rm Corr}^i(X_j,Y)$.
\section{Categorical and classical representabilities for smooth projective varieties}
\subsection{Semiorthogonal decompositions and categorical representability}
We start by recalling some categorical definitions which are necessary to define representability.
Let $K$ be a field and $\cat{T}$ a $K$-linear triangulated category. A full triangulated category $\cat{A}$ of
$\cat{T}$ is called \it admissible \rm if the embedding functor admits a left and a right adjoint.
\begin{definition}[\cite{bondalkap,bondalorlov}]\label{def-semiortho}
A \it semiorthogonal decomposition \rm of $\cat{T}$ is a sequence of admissible subcategories
$\cat{A}_1, \ldots, \cat{A}_l$ of $\cat{T}$ such that ${\rm Hom}_{\cat{T}}(A_i,A_j) = 0$ for all $i>j$ and for
all objects $A_i$ in $\cat{A}_i$ and $A_j$ in $\cat{A}_j$, and for every object $T$ of $\cat{T}$, there is a chain
of morphisms $0=T_n \to T_{n-1} \to \ldots \to T_1 \to T_0 = T$ such that the cone of $T_k \to T_{k-1}$ is
an object of $\cat{A}_k$ for all $k=1,\ldots,l$. Such a decomposition will be written
$$\cat{T} = \langle \cat{A}_1, \ldots, \cat{A}_l \rangle.$$
\end{definition}
\begin{definition}[\cite{bondal}]\label{def-except}
An object $E$ of $\cat{T}$ is called \it exceptional \rm if ${\rm Hom}_{\cat{T}} (E,E) = K$, and ${\rm Hom}_{\cat{T}}(E,E[i])=0$
for all $i \neq 0$. A collection $\{E_1,\ldots,E_l\}$ of exceptional objects is called \it exceptional \rm if
${\rm Hom}_{\cat{T}}(E_j,E_k[i])=0$ for all $j>k$ and for all integer $i$.
\end{definition}
If $E$ in $\cat{T}$ is an exceptional object, the triangulated category generated by $E$ (that is, the smallest full
triangulated subcategory of $\cat{T}$ containing $E$) is equivalent to the derived category of a point, seen as a smooth projective
variety. The equivalence ${\rm D}^{\rm b}(pt) \to \langle E \rangle \subset \cat{T}$ is indeed given by sending ${\cal O}_{pt}$ to $E$.
Given an exceptional collection $\{E_1,\ldots,E_l\}$ in the derived category ${\rm D}^{\rm b}(X)$ of a smooth projective variety,
there is a semiorthogonal decomposition \cite{bondalorlov}
$${\rm D}^{\rm b}(X) = \langle \cat{A}, E_1, \ldots, E_l\rangle,$$
where $\cat{A}$ is the full triangulated subcategory whose objects are all the $A$ satisfying ${\rm Hom}(E_i,A)=0$
for all $i=1,\ldots,l,$ and we denote by $E_i$ the category generated by $E_i$.
We say that the exceptional sequence is \it full \rm if the category $\cat{A}$ is trivial.
\begin{definition}\label{def-rep-for-cat}
A triangulated category $\cat{T}$ is \it representable in dimension $m$ \rm
if it admits a semiorthogonal decomposition
$$\cat{T} = \langle \cat{A}_1, \ldots, \cat{A}_l \rangle,$$
and for all $i=1,\ldots,l$ there exists a smooth projective variety $Y_i$ with $\dim Y_i \leq m$, such that $\cat{A}_i$
is equivalent to an admissible subcategory of ${\rm D}^{\rm b}(Y_i)$.
\end{definition}
\begin{remark}\label{connectd}
Notice that we can assume that the categories $\cat{A}_i$ to be indecomposable, and
then the varieties $Y_i$ in the definition to be connected. Indeed, the derived category
${\rm D}^{\rm b}(Y)$ of a scheme $Y$ is indecomposable if and only if $Y$ is connected (see \cite[Ex. 3.2]{bridg-equiv-and-FM}).
\end{remark}
\begin{definition}
Let $X$ be a smooth projective variety of dimension $n$. We say that $X$ is \it categorically representable \rm
in dimension $m$ (or equivalently in codimension $n-m$) if ${\rm D}^{\rm b}(X)$ is representable in dimension $m$.
\end{definition}
\begin{remark}\label{rem-def-for-non-smooth}
Suppose that $X$ is not smooth. Then to define categorical representability for it, we need to use
categorical resolution of singularities, as defined by Kuznetsov \cite{kuznet-singul}. He constructs,
provided that $X$ has rational singularities, and given a resolution $\widetilde{X} \to X$,
an admissible subcategory $\widetilde{\cat{D}}$ of ${\rm D}^{\rm b}(\widetilde{X})$ which is the \it categorical
resolution of singularities \rm of ${\rm D}^{\rm b}(X)$. Then we can say that $X$ is categorically representable
in dimension $m$ if $\widetilde{\cat{D}}$ is.
\end{remark}
Notice that if any fully faithful functor between smooth projective varieties is of Fourier--Mukai type \cite{orlovequivk3,orlovequivall}.
It is moreover worth noting and recalling the following facts, which are well-known in the derived categorical setting.
\begin{remark}[\cite{beilinson}]
The derived category of $\mathbb{P}^n$ admits a full exceptional sequence.
\end{remark}
\begin{remark}
If $\Gamma$ is a smooth connected projective curve of positive genus, then ${\rm D}^{\rm b}(\Gamma)$ has no proper
admissible subcategory. Indeed any fully faithful functor
$\cat{A} \to {\rm D}^{\rm b}(\Gamma)$ is an equivalence, unless $\cat{A}$ is trivial. Then being categorically representable in dimension
1 is equivalent to admit a semiorthogonal decomposition by exceptional objects and derived
categories of smooth projective curves.
\end{remark}
\begin{remark}\label{rem-grr-decomp}
If $X$ and $Y_i$ are smooth projective and
$${\rm D}^{\rm b}(X) = \langle {\rm D}^{\rm b}(Y_1), \ldots, {\rm D}^{\rm b}(Y_k) \rangle,$$
then
$$K_0(X) = \bigoplus_{i=1}^k K_0(Y_i)$$
and the Grothendieck--Riemann--Roch Theorem gives
$${\rm CH}^*_{\mathbb{Q}}(X) = \bigoplus_{i=1}^k {\rm CH}^*_{\mathbb{Q}}(Y_i).$$
\end{remark}
\begin{proposition}[\cite{orlovprojbund}]\label{blow-up-formula}
Let $X$ be smooth projective and $Z \subset X$ a smooth subvariety of codimension $d > 1$. Denote by
$\varepsilon:\widetilde{X} \to X$ the blow up of $X$ along $Z$. Then
$${\rm D}^{\rm b}(\widetilde{X}) = \langle \varepsilon^* {\rm D}^{\rm b}(X), {\rm D}^{\rm b}(Z)_1, \ldots, {\rm D}^{\rm b}(Z)_{d-1} \rangle,$$
where ${\rm D}^{\rm b}(Z)_i$ is equivalent to ${\rm D}^{\rm b}(Z)$ for all $i=1,\ldots,d-1$.
\end{proposition}
\subsection{Classical representabilities and motives}
In this Section we outline a list of definitions of representabilities for the groups $A^i_{\mathbb{Z}}(X)$. This is
far for being exhaustive, especially in the referencing.
Indeed, giving a faithful list of all contributions
to these questions is out of the aim of these notes. Chow motives and their properties could give, through Conjecture
\ref{orlov-conj}, a way to connect categorical and classical representabilities. We also outline the basic
facts needed to stress the possible interplay between new and old definitions.
Let $X$ as usual be a smooth projective variety over an algebraically closed field $K$.
\begin{definition}[\cite{blochmurreFano}]
The group $A_{\mathbb{Z}}^i(X)$ is said to be \it weakly representable \rm if there exists a smooth projective curve $\Gamma$, a class
$z$ of a cycle in $CH_{\mathbb{Z}}^i(X\times \Gamma)$ and an algebraic subgroup $G \subset J(\Gamma)$ of the Jacobian variety
of $\Gamma$, such that the induced morphism
$$z_*:J(\Gamma) \rightarrow A^i_{\mathbb{Z}}(X)$$
is surjective with kernel $G$.
\end{definition}
When working with coefficients in $\mathbb{Q}$, we have the following definition.
\begin{definition}\label{rat}
The group $A^i_{\mathbb{Q}}(X)$ is \it rationally representable \rm if
there exists a regular surjective morphism
$$z_*:J_{\mathbb{Q}}(\Gamma) \rightarrow A^i_{\mathbb{Q}}(X).$$
\end{definition}
Rational representability is a name that has been used several times in the literature, so it might lead to some misunderstanding.
We underline that Definition \ref{rat} is exactly the one from (\cite{gorch-gul-motives-and-repr}, page 5).
In the complex case, we have also a
stronger notion, which is called the \it Abel--Jacobi property \rm \cite{blochmurreFano}, which requires the existence
of an isogeny (i.e. a regular surjective morphism) $A^i_{\mathbb{Z}}(X) \to J^i(X)$ onto the $i$-th intermediate Jacobian.
The Abel-Jacobi property implies weak representability for smooth projective varieties defined on $\mathbb{C}$.
\begin{definition}[\cite{beauvilleprym}]
An abelian variety $A$ is said to be the algebraic representative of $A^i_{\mathbb{Z}}(X)$ if there exists an isomorphism
$G:A^i_{\mathbb{Z}}(X) \rightarrow A$ which is universal. That is: for any morphism $g$ from $A^i_{\mathbb{Z}}(X)$ to an abelian variety $B$,
there exists a unique morphism $u:A \rightarrow B$ such that $u\circ G=g$. In this case we say that $A^i_{\mathbb{Z}}(X)$ is
\it algebraically representable. \rm
\end{definition}
The first examples of algebraic representatives are the Picard variety ${\rm Pic}^0(X)$ or the Albanese variety
${\rm Alb}(X)$ if $n=1$ or, respectively, $n=\dim(X)$.
\begin{definition}
Let $X$ be a smooth projective variety of odd dimension $2n + 1$
and $A$ the algebraic representative of $A_{\mathbb{Z}}^{n+1}(X)$ via the canonical map $G:
A_{\mathbb{Z}}^{n+1}(X)\rightarrow A$. A polarization of $A$ with associated correspondence $\theta_A$ in ${\rm Corr}(A)$, is the \it incidence polarization \rm with respect
to $X$ if for all algebraic maps $f : T \rightarrow A_{\mathbb{Z}}^{n+1}(X)$ defined by a cycle $z$ in $CH_{\mathbb{Z}}^{n+1}(X\times T)$, we have
$$(G\circ f)^*\theta_A = (-1)^{n+1}I(z);$$
where $I(z)$ in ${\rm Corr}(T)$ is the composition of the correspondences $z \in {\rm Corr}(T,X)$ and $z\in {\rm Corr}(X,T)$.
\end{definition}
There are many complex threefolds $X$ with negative Kodaira dimension, for which $A^2_{\mathbb{Z}}(X)$ is strongly represented by a generalized
Prym with incidence polarization. For these threefolds, we will show how categorical representability in dimension
1 gives a splitting of the intermediate Jacobian. A list of the main cases will be given in Section \ref{section-reconstr}.
\smallskip
A more modern approach to representability questions has to take Chow motives into account.
Let us recall their basic definitions and notations.
The category ${\cal M}_K$ of Chow motives over $K$
with rational coefficients is defined as follows: an object of ${\cal M}_K$ is a triple $(X,p,m)$, where $X$
is a variety, $m$ an integer and $p \in {\rm Corr}^0(X,X)$ an idempotent,
called a \it projector\rm. Morphisms from $(X,p,m)$ to $(Y,q,n)$ are given by elements of ${\rm Corr}^{n-m}(X,Y)$
precomposed with $p$ and composed with $q$.
There is a natural functor $h$ from the category of smooth projective schemes to the category of motives,
defined by $h(X) = (X,{\rm id},0)$, and, for any morphism $\phi: X \to Y$, $h(\phi)$ being the correspondence given
by the graph of $\phi$. We write $\mathbf{1}:=({\rm Spec} \, K, {\rm id}, 0)$ for the unit motive and $\mathbb{L} := ({\rm Spec} \, K, {\rm id}, -1)$
for the Lefschetz motive,
and $M(-i) := M \otimes \mathbb{L}^i$. Moreover, we have
${\rm Hom}(\mathbb{L}^d,h(X))= CH_{\mathbb{Q}}^d(X)$ for all smooth projective schemes $X$ and all integers $d$.
If $X$ is irreducible of dimension $d$, the embedding $\alpha: pt \to X$ of the point defines a motivic map $\mathbf{1} \to h(X)$. We denote
by $h^0(X)$ its image and by $h^{\geq 1}(X)$ the quotient of $h(X)$ via $h^0(X)$. Similarly, $\mathbb{L}^d$ is a quotient of $h(X)$,
and we denote it by $h^{2d}(X)$.
In the case of smooth projective curves of positive genus there exists another factor which corresponds to the Jacobian variety of the
curve. Let $C$ be a smooth projective connected curve, let us define a motive $h^1(C)$ such that we have a direct sum:
$$h(C) = h^0(C) \oplus h^1(C) \oplus h^2(C).$$
The upshot is that the theory of the motives $h^1(C)$ corresponds to that of Jacobian varieties (up to isogeny), in fact we have
$${\rm Hom}(h^1(C),h^1(C')) = {\rm Hom}(J(C),J(C'))\otimes \mathbb{Q}.$$
In particular, the full subcategory of ${\cal M}_K$ whose objects are direct summands of the motive $h^1(C)$ is equivalent
to the category of abelian subvarieties of $J(C)$ up to isogeny. Such motives can be called \it abelian. \rm We will say
that a motive is \it discrete \rm if it is the direct sum of a finite number of Lefschetz motives.
The strict interplay between motives and representability for threefolds is shown by Gorchinskiy and Guletskii.
In this case, the rational representability of $A^i_{\mathbb{Q}}(X)$ for $i \geq 2$ is known (\cite{murre-resultat}).
In \cite{gorch-gul-motives-and-repr} it is proved that
$A^3_{\mathbb{Q}}(X)$ is rationally representable if and only if the Chow motive of $X$ has a given
Chow-K\"{u}nneth decomposition.
\begin{theorem}[\cite{gorch-gul-motives-and-repr}, Thm 8]\label{theo-gor-gul}
Let $X$ be a smooth projective threefold. The group $A^3_{\mathbb{Q}}(X)$ is rationally representable if and only if the motive $h(X)$
has the following Chow-K\"{u}nneth decomposition:
$$h(X) \cong \mathbf{1} \oplus h^1(X) \oplus \mathbb{L}^{\oplus b} \oplus (h^1(J)(-1)) \oplus
(\mathbb{L}^2)^{\oplus b} \oplus h^5(X) \oplus \mathbb{L}^3,$$
where $h^1(X)$ and $h^5(X)$ are the Picard and Albanese motives respectively,
$b = b^2(X) = b^4(X)$ is the Betti number, and $J$ is a certain abelian
variety, which is isogenous to the intermediate Jacobian $J(X)$ if $K = \mathbb{C}$.
\end{theorem}
\section{Interactions between categorical and classical representabilities}\label{tre}
In this section, we will consider varieties defined over the complex numbers. This restriction
is not really necessary, since most of the constructions work over any algebraically closed
field. Anyway, in the complex case, we can simplify our treatment by dealing with
intermediate Jacobians. Moreover, it will be more simple to list examples without the need
to make the choice of the base field explicit for any case.
\subsection{Fully faithful functors and motives}\label{sec-fff-and-mot}
At the end of the last section we have seen that, in the case of threefolds, rational representability of $A^3_{\mathbb{Q}}(X)$ is equivalent
to the existence of some Chow-K\"unneth decomposition.
The first step in relating categorical and rational representability is exploiting an idea of Orlov about the
motivic decomposition which should be induced by a fully faithful functor between the derived categories of
smooth projective varieties. Assuming this conjecture we get that for threefolds categorical representability in
dimension 1 is a stronger notion than rational representability.
Let us sketch Orlov's idea \cite{orlov-motiv}. If $X$ and $Y$ are smooth projective varieties
of dimension respectively $n$ and $m$,
and $\Phi: {\rm D}^{\rm b}(Y) \to {\rm D}^{\rm b}(X)$ is a fully faithful functor, then it is of Fourier--Mukai type \cite{orlovequivk3,orlovequivall}.
Let ${\cal E}$ in ${\rm D}^{\rm b}(X \times Y)$ be its kernel and ${\cal F}$ in ${\rm D}^{\rm b}(X \times Y)$ the kernel
of its right adjoint $\Psi$, we have ${\cal F} \simeq {\cal E}^{\vee} \otimes pr_X^{*}\omega_X [\dim X]$ (see \cite{mukaiabelian}).
Consider $e:= ch({\cal E}) {\rm Td}(X)$ and $f:= ch({\cal F}) {\rm Td}(Y)$, two mixed rational cycles in ${\rm CH}^*_{\mathbb{Q}}(X \times Y)$. We
denote by $e_i$ (resp. $f_i$) the $i$-th codimensional component of $e$ (resp. $f$), that is
$e_i, f_i \in {\rm CH}_{\mathbb{Q}}^i (X \times Y)$. As correspondences they induce
motivic maps $e_i: h(Y) \to h(X)(i-n)$ and $f_j: h(X)(m-j) \to h(Y)$. The Grothendieck--Riemann--Roch
Theorem implies that $f.e := \bigoplus_{i=0}^{n+m} f_{n+m-i} e_i = {\rm id}_{h(Y)}$.
\begin{orlconj}
Let $X$ and $Y$ be smooth projective varieties and $\Phi: {\rm D}^{\rm b}(Y) \to {\rm D}^{\rm b}(X)$ be a fully faithful functor. Then
the motive $h(Y)$ is a direct summand of the motive $h(X)$.
\end{orlconj}
The Conjecture is trivially true for $Y$ a smooth point, in which case $\Phi({\rm D}^{\rm b}(Y))$ is generated
by an exceptional object of ${\rm D}^{\rm b}(X)$.
In \cite{orlov-motiv}, it is proven that the Conjecture holds if $X$ and $Y$ have the same dimension $n$
and ${\cal E}$ is supported in dimension $n$. This already covers some interesting example: if $X$ is a smooth
blow-up of $Y$, or if there is a standard flip from $X$ to $Y$. Using the same methods as in \cite{marcello-bolo-conicbd}
we will show that the conjecture holds (up to restricting to all direct summand of $h(Y)$) for some more examples, namely the
case of $Y$ a curve and $X$ a rationally representable threefold (i.e., $A^i_{\mathbb{Q}}(X)$ is rationally representable for
all $i$) with $h^1(X) = h^5(X) = 0$.
But let us first take a look to the simplest case, that is relating categorical representability in dimension 0 and
discreteness of the motive.
\begin{remark}\label{prop-categorical-repr-in-dim-0}
If a smooth projective variety $X$ is categorically representable in dimension $0$, then the motive
$h(X)$ is discrete.
\end{remark}
\begin{proof}
Being representable in dimension $0$ is equivalent to having a full exceptional sequence. Then the
proof is straightforward, we actually have more, that is $K(X) = \mathbb{Z}^l$, where $l$ is the number of
objects in the sequence.
\end{proof}
A way more interesting case relates categorical representability in dimension 1 and rational representability
for threefolds.
In this case, in light of Theorem \ref{theo-gor-gul}, we have a more specific conjecture.
\begin{conjecture}
If a smooth projective threefold $X$ is categorically representable in dimension $1$, then it is rationally
representable.
\end{conjecture}
If $X$ is a standard conic bundle over a rational surface and $\Gamma$ a smooth projective curve, the Chow-K\"unneth decomposition of $h(X)$
(see \cite{nagel-saito}) can be used to show that a fully faithful functor ${\rm D}^{\rm b}(\Gamma) \to
{\rm D}^{\rm b}(X)$ gives $h^1(\Gamma)(-1)$ as a direct summand of $h(X)$. In particular, this
gives an isogeny between $J(\Gamma)$ and an abelian subvariety of $J(X)$, and proves
(up to codimensional shfit for each direct summand of $h(\Gamma)$) Conjecture \ref{orlov-conj}
in this case. The proof in \cite{marcello-bolo-conicbd} is based on the fact that
the motive $h(X)$ splits into a discrete motive and in a unique abelian motive which
corresponds to $J(X)$. Let us make a first assumption
\begin{itemize}
\item[($\star$)] $X$ is a smooth projective rationally representable threefold with $h^1(X) = 0$ and $h^5(X) = 0$.
\end{itemize}
\begin{theorem}\label{from-Db-to-isogeny}
Suppose $X$ satisfies $\star$. If there is a smooth projective curve $\Gamma$ and
a fully faithful functor ${\rm D}^{\rm b}(\Gamma) \to {\rm D}^{\rm b}(X)$, then there exists an integer $j_i$ such that $h^i(\Gamma)(j_i)$ is
a direct summand of $h(X)$ for $i=0,1,2$, and there is an injective
morphism $J(\Gamma)_{\mathbb{Q}} \to J(X)_{\mathbb{Q}}$, that is an isogeny between $J(\Gamma)$ and an abelian subvariety
of $J(X)$.
\end{theorem}
\begin{proof}
Let ${\cal E}$ and ${\cal F}$, and $e$ and $f$ as before. Consider $h^0(\Gamma)= \mathbf{1}$, we have
$f.e_{\vert h^0(\Gamma)} = {\rm id}_{h^0(\Gamma)}$, which gives the claim, but not an
explicit value of $i_0$. The same argument works for $h^2(\Gamma) = \mathbb{L}$.
For $h^1(\Gamma)$, we only need the case where $g(\Gamma) >0$, and
we can use the same argument as in \cite{marcello-bolo-conicbd} Lemma 4.2 : since
all but one addendum of $h(X)$ are discrete, the map $f.e_{\vert h^1(\Gamma)} = {\rm id}_{h^1(\Gamma)}$ is given
by $f_2.e_2$, which proves that $h^1(\Gamma)(-1)$ is a direct summand of $h^3(X)=M^1(J)(-1)$.
\end{proof}
\begin{corollary}\label{corollary-cat-rep-and-isogeny}
Suppose $X$ satisfies $\star$ and let $\{\Gamma_i\}_{i=1}^k$ be smooth projective curves
of positive genus. If ${\rm D}^{\rm b}(X)$ is categorically representable
in dimension 1 by the categories ${\rm D}^{\rm b}(\Gamma_i)$ and by exceptional objects, then $J(X)$ is isogenous
to $\oplus_{i=1}^k J(\Gamma_i)$.
\end{corollary}
\begin{proof}
Use remark \ref{rem-grr-decomp}.
\end{proof}
\begin{remark}\bf{(Threefolds satisfying $\star$).} \rm
By \cite{gorch-gul-motives-and-repr, nagel-saito} Fano threefolds, threefolds fibered in Del Pezzo or Enriques surfaces over $\mathbb{P}^1$
with discrete Picard group,
and standard conic bundles over rational surfaces satisfy assumptions of Theorem \ref{from-Db-to-isogeny}.
\end{remark}
\subsection{Reconstruction of the intermediate Jacobian}\label{section-reconstr}
The aim of this section is to show how, under appropriate hypothesis,
categorical representability in dimension 1 for a threefold $X$ gives a splitting of the intermediate Jacobian $J(X)$.
Notice that in the case of curves the derived category carries the information about the principal polarization
of the Jacobian \cite{marcellocurves}.
In the case of threefolds, we need first of all the hypothesis of Theorem \ref{from-Db-to-isogeny}.
As we will see, the crucial hypothesis that will allow us to recover also the principal polarization
is that the polarization on $J(X)$ is an \textit{incidence polarization}.
\begin{itemize}
\item[($\natural$)] $X$ is a smooth projective rationally and algebraically representable threefold with $h^1(X) = 0$ and $h^5(X) = 0$
and the algebraic representative of $A^2_{\mathbb{Z}}(X)$ carries an incidence polarization.
\end{itemize}
\begin{theorem}\label{reconstr-of-interm}
Suppose $X$ satisfies $\natural$. Let
$\Gamma$ be smooth projective curve and ${\rm D}^{\rm b}(\Gamma) \to {\rm D}^{\rm b}(X)$ fully faithful. Then there is an injective
morphism $J(\Gamma) \hookrightarrow J(X)$ preserving the principal polarization, that is $J(X) = J(\Gamma) \oplus A$ for
some principally polarized abelian variety $A$.
\end{theorem}
\begin{proof}
From Theorem \ref{from-Db-to-isogeny} we get an isogeny. As in the proof of \cite[Prop. 4.4]{marcello-bolo-conicbd},
the incidence property shows that this isogeny is an injective morphism respecting the principal polarizations.
\end{proof}
\begin{corollary}\label{ortjac}
Suppose $X$ satisfies $\natural$ and let $\{\Gamma_i\}_{i=1}^k$ be smooth projective curves
of positive genus. If ${\rm D}^{\rm b}(X)$ is categorically representable
in dimension 1 by the categories ${\rm D}^{\rm b}(\Gamma_i)$ and by exceptional objects, then $J(X)$ is isomorphic
to $\oplus_{i=1}^k J(\Gamma_i)$ as principally polarized variety.
\end{corollary}
\begin{remark}\label{remark-list-of-3folds}\bf{(Threefolds satisfying $\natural$).} \rm
The assumptions of Theorem \ref{reconstr-of-interm} seem rather restrictive. Anyway, they are satisfied by
a quite big class of smooth projective threefolds with $\kappa_X < 0$. The Chow-K\"unneth decomposition
for the listed varieties
is provided by \cite{nagel-saito} for conic bundles and by \cite{gorch-gul-motives-and-repr} in any other case.
In the following list the references
point out the most general results about strong representability and incidence property.
Giving an exhaustive list of all the results and contributors would be out of reach
(already in the cubic threefold case). We will consider Fano threefolds with Picard number one only. The
interested reader could find an exhaustive treatment in \cite{isko-prok-fano}.
\begin{itemize}
\item[1)] Fano of index $>2$: $X$ is either $\mathbb{P}^3$ or a smooth quadric.
\item[2)] Fano of index $2$: $X$ is a quartic double solid \cite{tihoquarticsolid} , or a smooth cubic in $\mathbb{P}^4$
\cite{clemensgriffiths}, or an intersection of two quadrics in $\mathbb{P}^5$ \cite{reidphd}, or a $V_5$ (in the last case $J(X)$ is trivial).
\item[3)] Fano of index $1$: $X$ is a general sextic double solid \cite{ceresaverra}, or
a smooth quartic in $\mathbb{P}^4$ \cite{blochmurreFano},
or an intersection of a cubic and a quadric in $\mathbb{P}^5$ \cite{blochmurreFano}, or the intersection of three
quadrics in $\mathbb{P}^6$ \cite{beauvilleprym}, or a $V_{10}$ \cite{logachevV10,iliev10}, or a $V_{12}$ \cite{ilievmarkuV12} ($J(X)$
is the jacobian of a genus 7 curve),
or a $V_{14}$ \cite{iliev-marku-v14} (in which case the representability is related to the birational map to a smooth cubic threefold),
or a general $V_{16}$ \cite{ilievV16, mukaigranmisto}, or a general $V_{18}$ \cite{ilievamanovella,isko-prok-fano}
($J(X)$ is the jacobian of a genus 2 curve), or a $V_{22}$ (and the Jacobian is trivial).
\item[4)] Conic bundles: $X \to S$ is a standard conic bundle over a rational surface \cite{beauvilleprym,beltrachow}, this is the case examined in
\cite{marcello-bolo-conicbd}.
\item[5)] Del Pezzo fibrations: $X \to \mathbb{P}^1$ is a Del Pezzo fibration with $2\leq K_X^2 \leq 5$ \cite{kanevdp1,kanevdp2}.
\end{itemize}
\end{remark}
From the unicity of the splitting of the intermediate Jacobian we can easily infer the following.
\begin{corollary}
Suppose $X$ satisfies $\natural$ and is categorically representable in dimension 1, with semiorthogonal decomposition
$${\rm D}^{\rm b}(X)= \langle {\rm D}^{\rm b}(\Gamma_1),\ldots,{\rm D}^{\rm b}(\Gamma_k),E_1,\ldots,E_l\rangle.$$
Then there is no fully faithful functor ${\rm D}^{\rm b}(\Gamma) \to {\rm D}^{\rm b}(X)$ unless $\Gamma \simeq \Gamma_i$ for some $i \in \{ 1,\ldots,k\}$.
Moreover, the semiorthogonal decomposition is essentially unique, that
is any semiorthogonal decomposition of ${\rm D}^{\rm b}(X)$ by smooth projective curves and exceptional objects is
given by all and only the curves $\Gamma_i$ and $l$ exceptional objects.
\end{corollary}
\begin{corollary}\label{cor-no-split-no-fm}
Suppose $X$ satisfies $\natural$, $\Gamma$ is a smooth projective curve of positive genus
and there is no splitting $J(X) = J(\Gamma) \oplus A$. Then there is no fully faithful functor
${\rm D}^{\rm b}(\Gamma) \to {\rm D}^{\rm b}(X)$.
\end{corollary}
The assumptions of Corollary \ref{cor-no-split-no-fm} are trivially satisfied if the
threefold satisfying $\natural$ has $J(X)=0$. A way more interesting case is when
the intermediate Jacobian is not trivial and has no splitting at all, in which case the
variety is not representable in dimension $<2$.
\begin{remark}\label{list-three-not-rep}\bf{(Threefolds not categorically representable in dimension $<2$)}. \rm
The assumptions of Corollary \ref{cor-no-split-no-fm} are satisfied by smooth threefolds with $J(X) \neq 0$
for all curve $\Gamma$ of positive genus in the following cases:
\smallskip
Either $X$ is a smooth cubic \cite{clemensgriffiths}, or a generic quartic threefold \cite{letiziamoratti}, or a generic complete intersection of type $(3,2)$
in $\mathbb{P}^5$ \cite{beauvilleprym} or a symmetric one \cite{bovesym}, or the intersection of three quadrics in $\mathbb{P}^7$ \cite{beauvilleprym},
or a standard conic bundle $X \to \mathbb{P}^2$ degenerating along a curve
of degree $\geq 6$ \cite{beauvilleprym}, or a non-rational standard conic bundle $X \to S$ on a Hirzebruch surface
\cite{shokuprym}, or a non-rational Del Pezzo fibration $X \to \mathbb{P}^1$ of degree four \cite{aleks-dp4}.
There are some other cases of Fano threefolds of specific type satisfying geometric assumptions. For a detailed treatment, see
\cite[Chapt. 8]{isko-prok-fano}.
\end{remark}
Notice that if $X$ is a smooth cubic threefold, the equivalence class of a notable admissible subcategory $\cat{A}_X$
(the orthogonal complement of $\{{\cal O}_X,{\cal O}_X(1)\}$) corresponds to the isomorphism class of $J(X)$ as principally polarized
abelian variety \cite{noicubic}; the proof is based on the reconstruction of the Fano variety and the techniques used
there are far away from the subject of this paper.
A natural question is if, under some hypothesis, one can give the inverse statement of Corollaries \ref{corollary-cat-rep-and-isogeny}
and \ref{ortjac}, that is: suppose that $X$ is a threefold satisfying either $\star$ or $\natural$, such
that $J(X) \simeq \oplus J(\Gamma_i)$. Can one describe a semiorthogonal decomposition of ${\rm D}^{\rm b}(X)$ by
exceptional objects and the categories ${\rm D}^{\rm b}(\Gamma_i)$? Notice that a positive
answer for $X$ implies a positive answer for all the smooth blow-ups of $X$.
\begin{remark}\label{list-of-iff}\bf(Threefolds with $\kappa_X <0$ categorically representable in dimension $\leq 1$). \rm
Let $X$ be a threefold satisfying $\star$ or $\natural$ and with $J(X) = \oplus J(\Gamma_i)$. Then if $X$
is in the following list (or is obtained by a finite number of smooth blow-ups from a variety in the list)
we have a semiorthogonal decomposition
$${\rm D}^{\rm b}(X) = \langle {\rm D}^{\rm b}(\Gamma_1), \ldots, {\rm D}^{\rm b}(\Gamma_k), E_1, \ldots, E_l \rangle,$$
with $E_i$ exceptional objects.
\begin{itemize}
\item[1)] Threefolds with a full exceptional sequence: $X$ is $\mathbb{P}^3$ \cite{beilinson},
or a smooth quadric \cite{kapranovquadric},
or a $\mathbb{P}^1$-bundle over a rational surface or a $\mathbb{P}^2$-bundle over $\mathbb{P}^1$ \cite{orlovprojbund},
or a $V_5$ \cite{orlov-v5}, or a $V_{22}$ Fano threefold \cite{kuznev22}.
\item[2)] Fano threefolds without any full exceptional sequence: $X$ is the complete intersection of two quadrics or a Fano
threefold of type $V_{18}$, and $J(\Gamma) \simeq J(X)$ with $\Gamma$ hyperelliptic.
The semiorthogonal decompositions are described in \cite{bondalorlov,kuznetHyperplane}, and are strikingly
related (as in the cases of $V_5$ and $V_{22}$ and of the cubic and $V_{14}$) by a correspondence in
the moduli spaces, as described in \cite{kuznetfanothreefolds}. $X$ is a $V_{12}$ Fano threefold
\cite{kuznetv12}, or a $V_{16}$ Fano threefold \cite{kuznetHyperplane}.
\item[3)] Conic bundles without any full exceptional sequence: $X \to S$ is a rational conic bundle over a minimal surface
\cite{marcello-bolo-conicbd}. If the degeneration locus of $X$ is either empty or a cubic in $\mathbb{P}^2$, then $X$ is a
$\mathbb{P}^1$-bundle and is listed in 1).
\item[4)] Del Pezzo fibrations: $X \to \mathbb{P}^1$ is a quadric fibration with at most simple degenerations, in which case the
hyperelliptic curve $\Gamma \to \mathbb{P}^1$ ramified along the degeneration appears naturally as the orthogonal complement of
an exceptional sequence of ${\rm D}^{\rm b}(X)$ \cite{kuznetconicbundles}. $X \to \mathbb{P}^1$ is a rational Del Pezzo fibration of degree four.
In this case $X$ is birational to a conic bundle over a Hirzebruch surface
\cite{aleks-dp4} and the semiorthogonal decomposition will be described in a forthcoming paper \cite{asher-marcello-bolo}.
\end{itemize}
Notice that the first two items cover all classes of Fano threefolds with Picard number one whose members are all rational.
\end{remark}
\section[Developments and Questions]{Categorical representability and rationality: \\
further developments and open questions}
This last Section is dedicated to speculations and open questions about categorical representability and rationality.
The baby example of curves is full understood. A smooth projective curve $X$ over a field $K$ is categorically representable in
dimension 0 if and only if it is rational. Indeed, the only case where ${\rm D}^{\rm b}(X)$ has exceptional objects is
$X=\mathbb{P}^1$, and ${\rm D}^{\rm b}(X) = \langle {\cal O}_X,{\cal O}_X(1) \rangle$.
Let us start with a trivial remark: the projective space $\mathbb{P}^n$ over $K$ is categorically representable in dimension 0.
Then if $X$ is given by a finite number of smooth blow-ups of $\mathbb{P}^n$, it is categorically representable in
codimension $\geq 2$. This is easily obtained from Orlov's blow-up formula (see Prop. \ref{blow-up-formula}).
More generally, if a smooth projective variety $X$ of dimension $\geq 2$ is categorically representable in
codimension $m$, then any finite chain of smooth blow-ups of $X$ is categorically representable in
codimension $\geq {\rm min}(2,m)$.
One could naively wonder about the inverse statement: if $X \to Y$
is a finite chain of smooth blow-ups and $X$ is categorically representable in codimension $m$, what can
we say about $Y$? Unfortunately, triangulated categories do not have enough structure to let us
compare different semiorthogonal decomposition. For example, the theory of mutations allows to do this only
in a few very special cases.
In this Section we present some more example to stress how the interaction between categorical representability
and rationality can be devloped further, and we point out some open question.
We deal with surfaces in \ref{surf} and with threefolds in \ref{threefolds}.
Then we will discuss in \ref{noncommutative} how categorical
representability for noncommutative varieties plays an important role in this frame, to deal with
varieties of dimension bigger than 3 in \ref{higherdim}. Finally, we compare in \ref{approaches}
our methods with recent approaches to birationality
problems via derived categories. We will work over the field $\mathbb{C}$ for simplicity, even if many
problems and arguments do not depend on that.
\subsection{Surfaces}\label{surf}
If $X$ is a smooth projective rational surface,
then it is categorically representable in codimension 2. Indeed,
$X$ is the blow-up in a finite number of smooth points of a minimal rational surface, that is either $\mathbb{P}^2$ or
$\mathbb{F}_n$. Are there any other example of surfaces categorically representable in codimension 2? Notice that by
Proposition \ref{prop-categorical-repr-in-dim-0} such a surface would have a discrete motive, and even more:
we would have $K_0(X) = \mathbb{Z}^l$. In particular, if $K_0(X)$ is not locally free, then
$X$ is not categorically representable in dimension 0.
In general, an interesting problem is to construct exceptional sequences on surfaces with $p_g=q=0$, and to study their
orthogonal complement. Suppose for example that $X$ is an Enriques surface: a (non-full) exceptional collection of 10
vector bundles on $X$ is described in \cite{zube-enriques}. Since $K_0(X)$ is not locally free, we do not expect any
full exceptional collection. The orthogonal complement $\cat{A}_X$ turns then out to be a very interesting
object, related also to the geometry of some singular quartic double solid \cite{kuzne-ingalls}.
Using a motivic trick, we can prove that, under some assumption, a surface with $p_g=q=0$ is
either categorically representable in codimension 2 or not categorically representable in positive codimension.
\begin{proposition}\label{prop-surf-p=q=0}
Let $X$ be a surface with $h(X)$ discrete. Then for any curve $\Gamma$ of positive genus, there is no fully faithful
functor ${\rm D}^{\rm b}(\Gamma) \to {\rm D}^{\rm b}(X)$.
\end{proposition}
\begin{proof}
Suppose there is such a curve and such a functor $\Phi: {\rm D}^{\rm b}(\Gamma) \to {\rm D}^{\rm b}(X)$. Let ${\cal E}$ denote the kernel of
$\Phi$ (which has to be of Fourier--Mukai type) and ${\cal F}$ the kernel of its adjoint. Consider the cycles $e$ and $f$ described
in Section \ref{sec-fff-and-mot}, and recall that $f.e = \oplus_{i=0}^3 f_{3-i}.e_i = {\rm id}_{h(\Gamma)}$. Restricting now
to $h^1(\Gamma)$ we would have that ${\rm id}_{h^1(\Gamma)}$ would factor through a discrete motive, which is impossible.
\end{proof}
\begin{corollary}\label{cor-surf-non-rep}
Let $X$ be a surface with $h(X)$ discrete and $K_0(X)$ not locally free. Then $X$ is not categorically representable in codimension
$>0$.
\end{corollary}
\begin{remark}\bf(Surfaces with $p_g=q=0$ not categorically representable in positive codimension). \rm
Proposition \ref{prop-categorical-repr-in-dim-0} could be an interesting tool in the study of derived categories of surfaces
with $p_g=q=0$: notice that many of them have torsion in $H_1(X, \mathbb{Z})$ (for an exhaustive treatment and referencing, see
\cite{bauer-cata-pigna-survey}). Anyway the discreteness of the motive is a rather strong assumption, which for
example implies the Bloch conjecture. There are few cases where this is known.
\begin{itemize}
\item[1)] $X$ is an Enriques surface \cite{coombes}.
\item[2)] $X$ is a Godeaux surface obtained as a quotient of a quintic by an action of $\mathbb{Z}/5\mathbb{Z}$ \cite{gul-pedr-godeaux}.
\end{itemize}
\end{remark}
These observations lead to state some deep question about categorical representability of
surfaces.
\begin{question}\label{exc-obj-on-surf}
Let $X$ be a smooth projective surface with $p_g = q = 0$.
\begin{itemize}
\item[1)] Does ${\rm D}^{\rm b}(X)$ admit an exceptional sequence?
\item[2)] Is the exceptional sequence full? That is, is $X$ representable in codimension 2?
\item[3)] If $X$ is representable in codimension 2, is $X$ rational?
\end{itemize}
\end{question}
\subsection{Threefolds}\label{threefolds}
Remark that there are examples
of smooth projective non-rational threefolds $X$ which are categorically representable in codimension 2:
just consider a rank three vector bundle ${\cal E}$ on a curve $C$ of positive genus and take $X:=\mathbb{P}({\cal E})$.
In \cite[Sect. 6.3]{marcello-bolo-conicbd} we provide a conic bundle example. Anyway, Corollary
\ref{ortjac} somehow suggests that categorical representability in codimension 2 should be a necessary condition
for rationality.
A reasonable idea is to restrict our
attention to minimal threefolds with $\kappa_X <0$ (recall that this is a necessary condition for rationality),
in particular to the ones we expect to satisfy assumption $\natural$, in order to have interesting information about the intermediate Jacobian from semiorthogonal decompositions.
The three big families of such threefolds are:
Fano threefolds, conic bundles over rational surfaces and del Pezzo fibrations over $\mathbb{P}^1$.
Remark \ref{list-of-iff} gives a list of rational threefolds which are categorically representable in codimension 2,
and Remark \ref{remark-list-of-3folds} a list of families whose generic term is non-rational and cannot be categorically
representable in codimension $> 1$.
\begin{question}\label{question-on-3folds}
Let $X$ be a smooth projective threefold with $\kappa_X <0$.
\begin{itemize}
\item[1)] If $X$ is rational, is $X$ categorically representable in codimension 2?
\item[2)] Is $X$ categorically representable in codimension 2 if and only if $X$ is rational?
\end{itemize}
\end{question}
A positive answer to the second question is provided for standard conic bundles over minimal surfaces \cite{marcello-bolo-conicbd},
but it seems to be quite a strong fact to hold in general: recall that having a splitting $J(X) \simeq \oplus J(\Gamma_i)$ is only a necessary
condition for rationality, and Corollary \ref{ortjac} shows that if $X$ satisfies $\natural$, categorical representability
in codimension 2 would give the splitting of the Jacobian.
Remark \ref{list-of-iff} provides a large list of rational threefolds categorically representable in codimension 2.
Is it possible to add examples to this list? In particular in the case of Del Pezzo fibrations over $\mathbb{P}^1$
only the quadric and the degree 4 fibration are described.
A good way to understand these questions is by studying some special rational or non-rational (that is non generic in their
family) threefold. This forces to consider non smooth ones, but we can use Kuznetsov's theory of categorical resolution of singularities \cite{kuznet-singul}
and study the categorical resolution of ${\rm D}^{\rm b}(X)$, as we pointed out in Remark \ref{rem-def-for-non-smooth}.
For example, let $X \subset \mathbb{P}^4$ be nodal cubic threefold with a double point, which is known to be rational.
\begin{proposition}
Let $X \subset \mathbb{P}^4$ be a cubic threefold with a double point and $\widetilde{X} \to X$ the blow-up of the singular point. The categorical
resolution of singularities $\widetilde{\cat{D}} \subset {\rm D}^{\rm b}(\widetilde{X})$ of ${\rm D}^{\rm b}(X)$ is representable in codimension two.
Indeed there is a semiorthogonal decomposition
$$\widetilde{\cat{D}} = \langle {\rm D}^{\rm b}(\Gamma), E \rangle,$$
where $E$ is an exceptional object and $\Gamma$ a complete intersection of a quadric and a cubic in $\mathbb{P}^3$.
\end{proposition}
\begin{proof}
This is shown following step by step \cite[Section 5]{kuznetcubicfourfold}, where the four dimensional case is studied.
Let us give a sketch of the proof. Let $P$ be the singular point of $X$, and $\sigma: \widetilde{X} \to X$ its blow-up.
The exceptional locus $\alpha: Q \hookrightarrow \widetilde{X}$ of $\sigma$ is a quadric surface.
The projection of $\mathbb{P}^4$ to $\mathbb{P}^3$ from the point $P$ restricted to $X$ gives the birational map $X \dashrightarrow \mathbb{P}^3$.
The induced map $\pi: \widetilde{X} \to \mathbb{P}^3$ is the blow-up of a smooth curve $\Gamma$ of genus 4, given by the complete intersection
of a cubic and a quadric surface.
If we denote $h:= \pi^* {\cal O}_{\mathbb{P}^3}(1)$ and $H:= \sigma^* {\cal O}_X (1)$, we have that $Q=2h-D$, $H=3h-D$, then $h=H-Q$ and $D=2H-3Q$
as in \cite[Lemma 5.1]{kuznetcubicfourfold}. The canonical bundle $\omega_{\widetilde{X}} = -4h+D= -2H+Q$ can be calculated
via the blow-up $\pi$.
The same arguments as in \cite{kuznetcubicfourfold} give the decomposition
$${\rm D}^{\rm b}(\widetilde{X}) = \langle \alpha_* {\cal O}_Q (-h), \widetilde{\cat{D}} \rangle.$$
Indeed the conormal bundle of $Q$ is ${\cal O}_Q(h)$ and the Lefschetz decomposition with respect to it is:
$$\langle {\cal A}_1(-h),{\cal A}_0\rangle,$$
where ${\cal A}_1 = \langle{\cal O}_Q\rangle$ and ${\cal A}_0 = \langle {\cal O}_Q, S_1, S_2 \rangle$, with $S_1$ and $S_2$ the two spinor bundles.
We obtain then the semiorthogonal decomposition:
\begin{equation}\label{semiorth1}
{\rm D}^{\rm b}(\widetilde{X}) = \langle \alpha_* {\cal O}_Q (-h), \widetilde{\cat{A}}_X, {\cal O}_{\widetilde{X}}, H \rangle,
\end{equation}
where $\widetilde{\cat{A}}_X$ is the categorical resolution of $\cat{A}_X$, as in \cite[Lemma 5.8]{kuznetcubicfourfold}.
The representability of $\widetilde{\cat{D}}$ relies then on the representability of $\widetilde{\cat{A}}_X$.
On the other side, applying the blow-up formula (see Prop. \ref{blow-up-formula}) to $\pi: \widetilde{X} \to \mathbb{P}^3$,
and choosing $\{ {\cal O}_{\mathbb{P}^3}(-3), \ldots, {\cal O}_{\mathbb{P}^3} \}$ as full exceptional sequence for ${\rm D}^{\rm b}(\mathbb{P}^3)$, we obtain:
$${\rm D}^{\rm b}(\widetilde{X}) = \langle \Phi {\rm D}^{\rm b}(\Gamma), -3h,-2h-h,{\cal O}_{\widetilde{X}} \rangle,$$
where $\Phi: {\rm D}^{\rm b}(\Gamma) \to {\rm D}^{\rm b}(\widetilde{X})$ is fully faithful. Now as in \cite[Lemma 5.3]{kuznetcubicfourfold},
if we mutate $-3h$ and $-2h$ to the left with respect to $\Phi {\rm D}^{\rm b}(\Gamma)$, we get
\begin{equation}\label{semiorth2}
{\rm D}^{\rm b}(\widetilde{X}) = \langle -3h+D, -2h+D, \cat{B}, {\cal O}_{\widetilde{X}} \rangle,
\end{equation}
where $\cat{B} = \langle \Phi {\rm D}^{\rm b}(\Gamma), -h \rangle$ is an admissible subcategory of ${\rm D}^{\rm b}(\widetilde{X})$.
Finally, one can show that $\cat{B}$ and $\widetilde{\cat{A}}_X$ are equivalent,
following exactly the same path of mutations as in \cite[Sect. 5]{kuznetcubicfourfold} to compare the
decompositions (\ref{semiorth1}) and (\ref{semiorth2}). Notice that one can calculate explicitely the
exceptional object $E$ by following the mutations of $-h$.
\end{proof}
Another special very interesting example is described in \cite{kuzne-ingalls}: a singular double solid $X \to \mathbb{P}^3$
ramified along a quartic symmetroid. This threefold is non-rational thanks to \cite{artin-mumford}, because $H^3(X, \mathbb{Z})$
has torsion. A rough account
(skipping the details about the resolution of singularities) of Ingalls and Kuznetsov's result is the following:
if $X'$ is the small resolution of $X$, there is an Enriques surface $S$ and a semiorthogonal decomposition
\begin{equation}\label{decomp-double-solid}
{\rm D}^{\rm b}(X') = \langle \cat{A}_S, E_1,E_2 \rangle,
\end{equation}
where $E_i$ are exceptional objects and $\cat{A}_S$ is the orthogonal complement in ${\rm D}^{\rm b}(S)$ of 10 exceptional vector bundles on $S$
(\cite{zube-enriques}). Then we can apply to this set Corollary \ref{cor-surf-non-rep}.
\begin{corollary}
The threefold $X'$ is not categorically representable in codimension $>1$.
\end{corollary}
\begin{proof}
Consider the Enriques surface $S$ and the semiorthogonal decomposition
$${\rm D}^{\rm b}(S) = \langle \cat{A}_S, E_1, \ldots, E_{10} \rangle,$$
where $E_i$ are the exceptional vector bundles described in \cite{zube-enriques}. Then the non-representability
of $S$ in codimension $>0$ is equivalent to the non-representability of $\cat{A}_S$ in dimension $<2$.
The statement follows then from (\ref{decomp-double-solid}).
\end{proof}
Remark that the
lack of categorical representability of $X'$ (and presumably of $X$, thinking about the categorical resolution of
singularities) is due to the presence of torsion in $K_0(S)$ and in particular in $H_1(S,\mathbb{Z})$, whereas the non-rationality
of $X$ is due to the presence of torsion in $H^3(X,\mathbb{Z})$. The relation between torsion in $H^3(X,\mathbb{Z})$ and categorical
representability needs a further investigation, for example in the case recently described in \cite{katza-iliev-pry-nonrational}.
\subsection{Noncommutative varieties}\label{noncommutative}
The previous speculations and partial results give rise to the hope of extending fruitfully the study of categorical representability
to higher dimensions and to the noncommutative setting. By the latter, we mean, following Kuznetsov
\cite[Sect. 2]{kuznetconicbundles}, an algebraic variety $Y$ with a sheaf ${\cal B}$ of ${\cal O}_Y$-algebras of finite type.
Very roughly, the corresponding noncommutative variety $\bar{Y}$ would have a category of coherent sheaves ${\cat{Coh}}(\bar{Y}) = {\cat{Coh}}(Y,
{\cal B})$ and a bounded derived category ${\rm D}^{\rm b}(\bar{Y}) = {\rm D}^{\rm b}(Y,{\cal B})$. The examples
which appear very naturally in our setting
are the cases where ${\cal B}$ is an Azumaya algebra or the even part of the Clifford algebra associated to some quadratic form over $Y$.
Finally, if a triangulated category $\cat{A}$ has
Serre functor such that $S_{\cat{A}}^m = [n]$, for some integers $n$ and $m$, with $m$ minimal with this property,
we will call it a \it $\frac{n}{m}$-Calabi--Yau category. \rm
If $m=1$, these categories deserve the name of noncommutative Calabi--Yau $n$-folds, even if they are not a priori given
by the derived category of some Calabi--Yau $n$-fold with a sheaf of algebras.
If $S$ is any smooth projective variety, $X \to S$ a Brauer--Severi variety of relative dimension $r$, and ${\cal A}$ the
associated Azumaya algebra in ${\rm Br}(S)$, then \cite{marcellobrauer}
$${\rm D}^{\rm b}(X) = \langle {\rm D}^{\rm b}(S), {\rm D}^{\rm b}(S,{\cal A}^{-1}), \ldots, {\rm D}^{\rm b}(S,{\cal A}^{-r}) \rangle.$$
The categorical representability of $X$ would then
rely on the categorical representability of $(S,{\cal A})$, which is an interesting object in itself. For example,
if $Y$ is a generic cubic fourfold containing a plane, there are a K3 surface $S$ and an Azumaya algebra
${\cal A}$ such that the categorical representability of $(S,{\cal A})$ is the subject of Kuznetsov's Conjecture \cite{kuznetcubicfourfold} about the rationality of cubic fourfolds.
\smallskip
If $S$ is a smooth projective variety and $Q \to S$ a quadric fibration of relative dimension $r$, we
can consider the sheaf ${\cal B}_0$ of the even parts of the Clifford algebra associated to the quadratic form defining
$Q$. There is a semiorthogonal decomposition:
$${\rm D}^{\rm b}(Q) = \langle {\rm D}^{\rm b}(S,{\cal B}_0), {\rm D}^{\rm b}(S)_1, \ldots, {\rm D}^{\rm b}(S)_{r-1}\rangle,$$
where ${\rm D}^{\rm b}(S)_i$ are equivalent to ${\rm D}^{\rm b}(S)$ \cite{kuznetconicbundles}. The categorical representability of $(S,{\cal B}_0)$ should then
be a very important tool in studying birational properties of $Q$. This is indeed the case for conic bundles over rational
surfaces \cite{marcello-bolo-conicbd}.
\smallskip
Finally, let $\cat{A}$ be an $\frac{n}{m}$-Calabi--Yau category. Such categories appear as orthogonal complements
of an exceptional sequence on Fano hypersurfaces in projective spaces \cite[Cor. 4.3]{kuznetv14}. It is then natural
to wonder about their representability. For example, if $X$ is a cubic or a quartic threefold,
it follows from Remark \ref{list-three-not-rep} that these orthogonal complements (which are, respectively, $\frac{5}{3}$ and
$\frac{10}{4}$-Calabi--Yau)
are not representable in dimension 1.
\begin{question}\label{question-cycat}
Let $\cat{A}$ be a $\frac{n}{m}$-Calabi--Yau category.
\begin{itemize}
\item[1)] Is $\cat{A}$ representable in some dimension?
\item[2)] If yes, is there an explicit lower bound for this dimension?
\item[3)] If $m=1$, is $\cat{A}$ representable in dimension $n$ if and only if there exist a smooth
$n$-dimensional variety $X$ and an equivalence ${\rm D}^{\rm b}(X) \simeq \cat{A}$?
\end{itemize}
\end{question}
\subsection{Higher dimensional varieties}\label{higherdim}
Unfortunately, it looks like the techniques used for threefolds in \cite{marcello-bolo-conicbd} hardly extend to
dimension bigger than $3$. The examples and supporting evidences provided so far lead anyway
to suppose that categorical representability can give useful informations on the birational properties of
any projective variety. The main case is a challenging Conjecture by Kuznetsov \cite{kuznetcubicfourfold}. Let $X \subset \mathbb{P}^5$ be
a smooth cubic fourfold, then there is a semiorthogonal decomposition
$${\rm D}^{\rm b}(X) = \langle \cat{A}_X, {\cal O}_X, {\cal O}_X(1), {\cal O}_X(2) \rangle.$$
The category $\cat{A}_X$ is 2-Calabi--Yau.
\begin{conjecture}\label{kuz-conj}\bf(Kuznetsov). \rm
The cubic fourfold $X$ is rational if and only if $\cat{A}_X \simeq {\rm D}^{\rm b}(Y)$ for a smooth projective K3 surface $Y$.
\end{conjecture}
This conjecture has been verified in \cite{kuznetcubicfourfold} for singular cubics, pfaffian cubics and
Hassett's \cite{hassett-special} examples.
When $X$ contains a plane $P$
there is a way more explicit construction: blowing up $P$ we obtain a quadric bundle $\widetilde{X} \to \mathbb{P}^2$ of
relative dimension 2, degenerating along a sextic. If the sextic is smooth, let $S \to \mathbb{P}^2$ be the double
cover, which is a K3 surface. Then
$$\cat{A}_X \simeq {\rm D}^{\rm b}(\mathbb{P}^2,{\cal B}_0) \simeq {\rm D}^{\rm b}(S,{\cal A}),$$
where ${\cal B}_0$ is associated to the quadric fibration and ${\cal A}$ is an Azumaya algebra, obtained lifting ${\cal B}_0$
to $S$. The questions about categorical
representability of noncommutative varieties arise then very naturally in this context.
Notice that if $\cat{A}_X$ is representable in dimension 2, then we know something weaker than Kuznetsov conjecture:
we would have a smooth projective surface $S'$ and a fully faithful functor $\cat{A}_X \to {\rm D}^{\rm b}(S')$. Point
3) of Question \ref{question-cycat} appears naturally in this context.
\begin{question}\label{quest-kuz-conj}
One can then wonder if the Kuzentsov conjecture may be stated in the following form: the cubic fourfold $X$ is rational if and only if
it is categorically representable in codimension 2. This is equivalent to proving that the 2-Calabi--Yau category $\cat{A}_X$
is representable in dimension 2 if and only if there exist a K3 surface $Y$ and an equivalence ${\rm D}^{\rm b}(Y) \simeq \cat{A}_X$.
\end{question}
We can propose some more examples of fourfolds for which a Kuznetsov-type conjecture seems natural: if $X$ is the complete intersection
of three quadrics $Q_1$, $Q_2$, $Q_3$ in $\mathbb{P}^7$, then Homological Projective Duality (\cite{kuznetHPD,kuznetconicbundles}) gives an exceptional
sequence on $X$ and its complement $\cat{A}_X \simeq {\rm D}^{\rm b}(\mathbb{P}^2,{\cal B}_0)$, where ${\cal B}_0$ is the even Clifford algebra associated to the net
of quadrics generated by $Q_1,Q_2,Q_3$. Similarly, if we consider two quadric fibrations $Q_1, Q_2 \to \mathbb{P}^1$ of relative dimension
$4$ and their complete intersection $X$, there is an exceptional sequence
on $X$, and let $\cat{A}_X$ be its orthogonal complement. A realtive version of Homological Projective Duality shows that $\cat{A}_X$
equivalent to ${\rm D}^{\rm b}(S,{\cal B}_0)$, where $S$ is a $\mathbb{P}^1$-bundle over $\mathbb{P}^1$ and ${\cal B}_0$ the
even Clifford algebra associated to the net of quadrics generated by $Q_1$ and $Q_2$. It is natural to wonder if representability in dimension
2 of the noncommutative varieties is equivalent or is a necessary condition for rationality of $X$. A partial answer to the last
example will be provided in a forthcoming paper \cite{asher-marcello-bolo}.
Other examples in dimension 7 are provided in \cite{iliev-manivel-cubic-7-8folds}.
If $X$ is a cubic sevenfold, there is a distinguished subcategory $\cat{A}_X$ of ${\rm D}^{\rm b}(X)$,
namely the orthogonal complement of the exceptional sequence $\{{\cal O}_X, \ldots, {\cal O}_X(5)\}$. This
is a 3-Calabi--Yau category. If $X$ is Pfaffian, it is shown in \cite{iliev-manivel-cubic-7-8folds}
that $\cat{A}_X$ cannot be equivalent to the derived category of any smooth projective variety.
It is also conjectured that $\cat{A}_X$ is equivalent to the orthogonal complement of an exceptional
sequence in the derived category ${\rm D}^{\rm b}(Y)$ of a Fano sevenfold $Y$ of index 3,
birationally equivalent to $X$.
\subsection{Other approaches}\label{approaches}
Of course categorical representability is just one among different approaches to the study
of birational geometry of a variety via derived categories. Nevertheless there is some
common ground.
\smallskip
First of all, Kuznetsov mentions in \cite{kuznetcubicfourfold} the notion of Clemens--Griffiths component of ${\rm D}^{\rm b}(X)$, whose
vanishing would be a necessary condition for rationality. It seems reasonable to expect that categorical representability
in codimension 2 implies the vanishing of the Clemens--Griffiths component.
Another recent theory is based on
Orlov spectra and their gaps \cite{katza-favero-ballard-generation-time}. Let us refrain even to sketch a definition of it,
but just notice that \cite[Conj. 2]{katza-favero-ballard-generation-time} draws a link between categorical
representability and gaps in the Orlov spectrum (see, in particular \cite[Cor. 1.11]{katza-favero-ballard-generation-time}).
Finally, conjectures based on homological mirror symmetry are proposed in
\cite{katza-rat-hms-1,katza-rat-hms-2}, but we cannot state a precise relation with our construction. A careful study
of the example constructed in \cite{katza-iliev-pry-nonrational} would be a good starting point.
\bibliographystyle{amsalpha} |
\section{Introduction}
For a commutative ring $R$ and a finite subset $U\subseteq R$ one may ask how many
among the $\lvert U \rvert^{n^2}$ matrices $A\in U^{[n]^2}$ have $\det(A) = 0$.
Much is known precisely when $R$ is the finite field $\mathbb{F}_{q}$ ($q$ power of a prime)
and $U=R$. For instance, it follows from elementary linear algebra that the number
of singular $n\times n$ matrices with entries from $\mathbb{F}_q$ is
precisely $q^{n^2} - \prod_{0\leq i \leq n-1} (q^n-q^i)$. As an advanced example, very
precise statements can be proved for matrices over finite fields even if the
entries are i.i.d. according to (quite) arbitrary distributions (cf. the work
of Kahn--Koml{\'o}s \cite{MR1833067} and Maples \cite{arXiv:1012.2372v1}.
In stark contrast, if $R=\mathbb{Z}$ and $U=\{-1,+1\}$, the
correct order of decay of the density of singular matrices is still not known,
but there is an old and plausible, yet still unproved conjecture of uncertain
origin, on which the last two decades have brought several remarkable advances:
\begin{conjecture}\label{conj:socalledfolkloreconjecture}
For $n\rightarrow \infty$,
$\frac{\lvert \{A\in \{-1,+1\}^{[n]^2}\colon \det(A) = 0 \in \mathbb{Z}\} \rvert}{2^{n^2}}
\sim (\frac12 + o(1))^n$.
\end{conjecture}
Let us employ the abbreviations $-:=-1$, $+:=+1$,
$\{\pm\}:=\{-1,+1\}$, $\{0,\pm\}:=\{-1,0,+1\}$,
$\textup{\textsf{P}}[\mathrm{Ra}_{<n}(\{\pm\}^{[n]^2})] := \{A\in\{\pm\}^{[n]^2}\colon\det(A) = 0\}$, and
let $\textup{\textsf{P}}[\cdot]$ denote the uniform measure on $\{\pm\}^{[n]^2}$, (so that in
particular Conjecture \ref{conj:socalledfolkloreconjecture} now
reads $\textup{\textsf{P}}[\{A\in\{\pm\}^{[n]^2}\colon\det(A) = 0\}] \sim (\frac12 + o(1))^n$).
Moreover, let us introduce one of the two main protagonists of the present paper:
\begin{definition}[the measure $\textup{\textsf{P}}_{\mathrm{lcf}}$]
\label{def:definitionofthelazycoinflipmeasure}
For $(s,t)\in \mathbb{Z}_{\geq 2}^2$ and $\emptyset\subseteq I\subseteq [s-1]\times [t-1]$
let $\textup{\textsf{P}}_{\mathrm{lcf}}$ denote the \emph{lazy coin flip distribution} on $\{0,\pm\}^I$,
i.e. the probability measure on $\{0,\pm\}^I$ defined by considering the values of
a $B\in \{0,\pm\}^I$ as independent identically distributed random variables, each
governed by the symmetric discrete distribution with values $-1$, $0$, $+1$ and
probabilities $\frac14$, $\frac12$, $\frac14$.
\end{definition}
The name of $\textup{\textsf{P}}_{\mathrm{lcf}}$ may stem from the fact that this is the
distribution obtained when someone sets out to generate the entries of some
$B\in \{0,\pm\}^I$ by performing $\lvert I \rvert$ independent fair coin flips,
but there is a probability of $\frac12$ at every single trial that out of a
fleeting laziness the person decides to simply write $0$ instead of flipping
the coin.
The lazy coin flip distribution $\textup{\textsf{P}}_{\mathrm{lcf}}$ plays a role in the recent article
\cite{MR2557947} of J. Bourgain, V. H. Vu and P. M. Wood in which the authors
set the current record in a chain of successive improvements of upper bounds for
$\textup{\textsf{P}}[\mathrm{Ra}_{<n}(\{\pm\}^{[n]^2})]$ (see Koml{\'o}s \cite{MR0221962},
Kahn--Koml{\'o}s-Szemer{\'e}di \cite{MR1260107}
and Tao--Vu \cite{MR2187480} \cite{MR2291914}):
\begin{theorem}[Bourgain--Vu--Wood \cite{MR2557947}]
\label{thm:bourgainvuwood}
For $n\rightarrow \infty$ it is true that
\begin{align}
\textup{\textsf{P}}[\{ A\in \{\pm\}^{[n]^2}\colon \det(A) = 0 \}] &
\leq (\frac{1}{\sqrt{2}} + o(1))^n \quad ,
\label{thm:bourgainvuwood:uniformdistribution}\\
\textup{\textsf{P}}_{\mathrm{lcf}}[\{ B\in \{0,\pm\}^{[n]^2}\colon \det(B) = 0\}] &
\sim (\frac12 + o(1))^n \label{thm:bourgainvuwood:lazycoinflip} \quad .
\end{align}
\end{theorem}
\begin{proof}[Comments]
The upper bound within $\sim$ in \eqref{thm:bourgainvuwood:lazycoinflip}
is the special case $\mu = \frac12$ of \cite[Corollary 3.1, p. 567]{MR2557947}.
The lower bound within the $\sim$ is obvious: consider the event that the first
column has only zero entries (the lower bound is also explitly stated
in \cite[formula (7), p. 561]{MR2557947}). The upper bound in
\eqref{thm:bourgainvuwood:uniformdistribution} is the special
case $S = \{\pm\}$ and $p=\frac12$ in \cite[Corollary 4.3, p. 576]{MR2557947}.
\end{proof}
So Bourgain--Vu--Wood proved that the correct order of decay of
$\textup{\textsf{P}}_{\mathrm{lcf}}[\{ B\in \{0,\pm\}^{[n]^2}\colon \det(B) = 0\}]$
is $(\frac12 + o(1))^n$---which is also the conjectured one
for $\textup{\textsf{P}}[\{ A\in \{\pm\}^{[n]^2}\colon \det(A) = 0\}]$. It is this latter
achievement, combined with an observation made by the present author,
which spurred the present paper. Note that using the uniform distribution
on $\{\pm\}^{[n]^2}$ is equivalent to considering the $n^2$ entries
as i.i.d. Bernoulli variables with probability $\frac12$. The observation is this:
When we apply one step of Chio condensation (see Definition \ref{def:chiocondensation})
to this Bernoulli matrix, the result is a matrix whose entries are $3$-wise
(and `almost' $6$-wise, see Theorem \ref{thm:comparativecountingtheorem} below)
stochastically independent with $\{-2,0,+2\}$-values which are distributed as if by
the lazy coin flip distribution. Since Bourgain--Vu--Wood demonstrated that for
$\textup{\textsf{P}}_{\mathrm{lcf}}$-distributed entries an asymptotically correct order of decay can be
proved, the observation feels like a hint at deeper connections and makes it seem
imperative to investigate Chio condensation of sign matrices. A first step is taken
in the present paper.
\section{Definitions}
Let $[n]:=\{1,\dotsc,n\}$. For any $A = (a_{i,j})_{(i,j)\in [n]^2}\in \{\pm\}^{[n]^2}$,
any $(i,j)\in [n-1]^2$ and any $\emptyset\subseteq I \subseteq [n-1]^2$ let
$A[i,j] := a_{i,j}$ and $A[I] := (a_{i,j})_{(i,j)\in I}$, hence in particular
$A[\emptyset]$ is the empty function and $A\bigl[[n]^2\bigr] = A$. Let
$\mathfrak{P}(X)$ denote the power set of a set $X$.
For a cartesian product $M\times N$ of two sets $M$ and $N$ let
$\mathrm{p}_1\colon M\times N \rightarrow M$ be the projection onto the first,
and $\mathrm{p}_2\colon M\times N \rightarrow N$ the projection onto the second factor.
If $M$ and $N$ are finite and $\emptyset \subseteq I\subseteq M\times N$ is
some subset, then $I$ is called \emph{rectangular} if and only if
$\lvert I\rvert =\lvert \mathrm{p}_1(I) \rvert \cdot \lvert \mathrm{p}_2(I) \rvert$.
Let us view functions as sets and matrices as functions. If $D$ is a set and
$f\colon D\rightarrow \mathbb{Z}$ is a function let us write
$D=:\mathrm{Dom}(f)\supseteq \mathrm{Supp}(f):=\{ d\in D\colon f(d)\neq 0\}$ for its domain and
support, and let us employ the abbreviations
$\lvert \mathrm{Dom}(f) \rvert=:\mathrm{dom}(f)\geq\mathrm{supp}(f) := \lvert \mathrm{Supp}(f) \rvert$. We have
$\mathrm{Dom}(\emptyset)=\mathrm{Supp}(\emptyset)=\emptyset$ and
therefore $\mathrm{dom}(\emptyset)=\mathrm{supp}(\emptyset)=0$. If $U\subseteq \mathbb{Z}$,
$\emptyset\subseteq I \subseteq [s-1]\times [t-1]$ and $B\in U^I$, then we have
$[s-1]\times [t-1]\supseteq I = \mathrm{Dom}(B) \supseteq \mathrm{Supp}(B) =
\{(i,j)\in [s-1]\times [t-1]\colon B[(i,j)]\neq 0\}$.
For a matrix $M = (m_{i,j})_{(i,j)\in I} \in \mathbb{Q}^I$ and a $q\in \mathbb{Q}$ we define, as usual,
$q\cdot M := (q\cdot m_{i,j})_{(i,j)\in I}$. The symbol $\sqcup$ denotes a set union
$\cup$ and at the same time makes the claim that the union is disjoint. The term
\emph{rank} of matrix has its usual meaning (and we will only use it in the context
of integral domains, so that row-rank, column-rank and determinantal rank are all
the same). For a set $S$, the group of all permutations of $S$ is denoted
by $\mathrm{Sym}(S)$.
The word `graph' without any further qualifications means `finite simple graph'
(i.e. `finite $1$-dimensional simplicial complex'). We will use $\mathrm{V}(X)$
(resp. $\mathrm{E}(X)$) to denote the vertex set (resp. edge set) of a graph $X$, and we
follow \cite{MR1373656} in reserving the more specific term \emph{circuit} for what
is called a \emph{cycle} in \cite{MR2159259} (i. e. closed walk
without self-intersections). Moreover, we will use $f_1(X)$ for the number
of edges of a graph $X$ and $f_0(X)$ for the
number of its of vertices. The \emph{cycle space of $X$} (i.e. $1$-dimensional
cycle group with $\mathbb{Z}/2$-coefficients in the sense of simplicial homology theory)
will be denoted by $\mathrm{Z}_1(X;\mathbb{Z}/2)$ and the \emph{coboundary space of $X$}
by $\mathrm{B}^1(X;\;\mathbb{Z}/2)$ (this is the $1$-dimensional coboundary group
with $\mathbb{Z}/2$-coefficients in the sense of simplicial cohomology theory; a synonym
is `cut space of $X$'). Let $\beta_0(X)$ denote the number of connected components
of a graph $X$ and $\beta_1(X) := \dim_{\mathbb{Z}/2}\ Z_1(X;\mathbb{Z}/2)$ the first Betti
number (a synonym in the graph-theoretical literature is `cyclomatic number'
\cite{MR1847424}). We will (without further notification) use the $1$-dimensional
case of the alternating sum relation between the ranks of the chain groups and
the ranks of the homology groups of a free chain complex,
i.e. $\beta_1(X) - \beta_0(X) = f_1(X) - f_0(X)$ for every graph $X$.
For any two disjoint graphs $X_1$ and $X_2$, the graph obtained by identifying
exactly one vertex of $X_1$ with exactly one vertex of $X_2$ is called
the \emph{(one-point) wedge of $X_1$ and $X_2$} and denoted by $X_1 \vee X_2$.
This is the standard wedge product of pointed topological spaces (but only
vertices of a graph are allowed as basepoints); a synonym within the
graph-theoretical literature is `coalescence' \cite[p. 140]{MR1054137}.
We will use the language of \emph{signed graphs} (see \cite{MR1744869} for a
comprehensive overview). It is customary in signed graph theory to work with
multigraphs (i.e. finite $1$-dimensional CW-complexes) for reasons of higher
flexibility in proofs and applications. However, in the present paper, all we will
need are signed simple graphs, i.e. for us a signed graph $(X,\sigma)$ will
simply consist of a graph $X = (V,E)$ together with an
arbitrary \emph{sign function} $\sigma\colon E\rightarrow \{\pm\}$.
We call \emph{$(+)$-edge} (resp. $(-)$-edge) every $e\in \mathrm{E}(X)$ with
$\sigma(e) = +$ (resp. $\sigma(e) = - $). Define $(+)$-paths (resp. $(-)$-paths)
as paths all of whose edges are $(+)$-edges (resp. $(-)$-edges). For emphasizing the
sign function we employ the notation `($\sigma$, $+$)-edge'.
If $(X,\sigma)$ is a signed graph let
$f_1^{(-)}(X,\sigma):=\lvert \{ e\in \mathrm{E}(X)\colon \sigma(e) = - \}\rvert$ denote
the number of ($\sigma$, $-$)-edges in it. A signed graph $(X,\sigma)$ is
called \emph{balanced}\footnote{The use of this term seems to have been initiated
in \cite{MR0067468}. The notion itself was already studied over seventy years ago by
D. K{\H{o}}nig \cite[p. 149, Paragraph 3]{MR0036989} under the
name `$p$-Teilgraph'.} if and only if $f_1^{(-)}(C,\sigma)$ is even for every
circuit $C$ of $X$. We will denote the set of all balanced signings of $X$ by
$S_{\mathrm{bal}}(X) :=
\{ \sigma\in\{\pm\}^{\mathrm{E}(X)}\colon \text{$(X,\sigma)$ balanced} \}$.
\begin{definition}[{Chio\footnote{In earlier versions I wrote `Chi{\`o}' but this
now seems wrong. All three spellings Chi{\`o}, `Chio' and `Chi{\'o}' are to be
found in the literature. My sole reason for using `{\`o}' was that
in \cite{MR1500275} the authors consistently use the
spelling `Chi{\`o}' and it is said \cite[p. 790]{MR1500275} that a copy of
Chio's original paper had been at the authors' disposal. However,
an original 1853 copy of \cite{chio} which I recently bought from an
antiquarian bookstore in Italy gives strong circumstantial evidence in favour of
the spelling `Chio': on the title page and the inside-cover his given name
`F{\'e}lix' is written with an accent whereas `Chio' does not carry any accent.
Moreover, the title page bears a hand-written dedication to a colleague,
signed `L'autore'. Therefore, to all appearances, Chio signed this title page
himself, 158 years ago. Possibly extant autographs aside, putting this on record
might come as close to a personal statement by Chio as one will ever get nowadays.
Moreover, the spelling is further corroborated by the usage adopted by
Cauchy in \cite{cauchyoeuvrescompletespremieresserietomex}. Cauchy on several
occasions consistently writes `M. F{\'e}lix Chio' \cite[p. 110, pp. 112--113]{cauchyoeuvrescompletespremieresserietomex}.} set}]
Let $(s,t)\in \mathbb{Z}_{\geq 2}^2$ and $I\subseteq [s]\times [t]$. Then $I$ is
called a \emph{Chio set} if and only if $(s,t)\in I$ and
for every $(i,j)\in I$ we have $(i,t)\in I$ and $(s,j)\in I$.
\end{definition}
\begin{definition}[Chio extension\footnote{Due to the change of spelling explained
in the previous footnote I now use a breve instead of a grave accent to denote
Chio extension.}]\label{def:chioextensionofaset}
For every $(s,t)\in\mathbb{Z}_{\geq 2}^2$ and every
$\emptyset \subseteq I \subseteq [s-1]\times [t-1]$,
\begin{equation}\label{eq:definitionofchiodom}
\breve{I}:= \{ (s,t)\} \sqcup \bigcup_{i\in\mathrm{p}_1(I)}\{ (i,t) \}
\sqcup\bigcup_{j\in\mathrm{p}_2(I)} \{ (s,j)\} \sqcup I \quad.
\end{equation}
\end{definition}
Note that $\breve{I} \subseteq [s]\times [t]$ for every
$\emptyset \subseteq I \subseteq [s-1]\times [t-1]$, in particular
$\breve{\emptyset} = \{(s,t)\}$ and $([s-1]\times [t-1])^{\breve{}} = [s]\times [t]$.
Moreover, a set $I'\subseteq [s]\times [t]$ is a Chio set if and only if there
exist an $I\subseteq[s-1]\times [t-1]$ with $I' = \breve{I}$.
Now we come to Chio condensation. In the special (and very common)
case of $s=t$ (hence $[s]\times [t] = [n]^2$) the following definition differs
from the standard convention (as is to be found in \cite{MR1500275} and
\cite{MR649067}) in that the entry $a_{n,n}$ instead of $a_{1,1,}$ is taken to be
the pivot. This seems to be more convenient for handling the indices of a Chio
condensate. There does not appear to be any logical necessity for multiplying
by $\tfrac12$, but the author decided to keep the discussion within the
realm of $\{0,\pm\}$-matrices (instead of $\{-2,0,+2\}$-matrices).
\begin{definition}[Chio condensation, $\tfrac12\mathrm{C}_{(s,t)}^{\breve{I}}$]
\label{def:chiocondensation}
For every $(s,t)\in \mathbb{Z}_{\geq 2}^2$, and every $I\subseteq [s-1]\times [t-1]$ define
the \emph{Chio map with pivot $a_{s,t}$} as
\begin{equation}
\tfrac12\mathrm{C}_{(s,t)}^{\breve{I}}\colon \{\pm\}^{\breve{I}}
\longrightarrow \{0,\pm\}^I \quad , \qquad
A \longmapsto \tfrac12\cdot \mathrm{C}_{(s,t)}(A) \quad ,
\end{equation}
where $\mathrm{C}_{(s,t)}(A) :=\bigr (\det(
\begin{smallmatrix}
a_{i,j} & a_{i,t} \\
a_{s,j} & a_{s,t}
\end{smallmatrix})\bigr)_{(i,j)\in I} \in \{-2,0,+2\}$. An image $\mathrm{C}_{(s,t)}(A)$ of
some $A\in \{\pm\}^{\breve{I}}$ is referred to as the
\emph{Chio condensate of $A$}.
\end{definition}
\begin{definition}[the measure $\textup{\textsf{P}}_{\mathrm{chio}}$]\label{def:measurePchio}
For every $(s,t)\in\mathbb{Z}_{\geq 2}^2$ and every $I\subseteq [s-1]\times [t-1]$ define
\begin{equation}\label{eq:definitionmeasurePchio}
\textup{\textsf{P}}_{\mathrm{chio}} \colon \mathfrak{P}\bigl ( \{0,\pm\}^I \bigr )
\longrightarrow [0,1] \quad , \qquad
\mathcal{B} \longmapsto \frac{1}{2^{\lvert \breve{I} \rvert}}\ \sum_{B\in\mathcal{B}}\;
\lvert(\tfrac12 \mathrm{C}_{(s,t)}^{\breve{I}})^{-1}(B)\rvert \quad .
\end{equation}
\end{definition}
Note that in the special case of $s:=t:=n$ and $I:=[n-1]^2$,
the measure $\textup{\textsf{P}}_{\mathrm{chio}}$ maps a single
$B\in \{0,\pm\}^{[n-1]^2}$ to $\textup{\textsf{P}}_{\mathrm{chio}}[B] := \textup{\textsf{P}}_{\mathrm{chio}}[\{B\}] = 2^{-n^2}\cdot
\bigl \lvert \{ A\in \{\pm\}^{[n]^2}\colon
B = \tfrac12\cdot \mathrm{C}_{(n,n)}(A) \} \bigr \rvert$.
We now define two additional measures. Later we will recognize both of
them as familiar ones:
\begin{definition}[averaged Chio measure]
For every $\emptyset\subseteq I \subseteq [n-1]^2$ define
\begin{equation}
\overline{\textup{\textsf{P}}}_{\mathrm{chio}} \colon \mathfrak{P}(\{0,\pm\}^I)
\longrightarrow [0,1]\quad , \qquad \mathcal{B} \longmapsto
\sum_{B\in\mathcal{B}} \frac{1}{2^{\mathrm{supp} (B)}}
\sum_{\tilde{B} \in \{0,\pm\}^I\colon \mathrm{Supp}(\tilde{B}) = \mathrm{Supp}(B)}
\textup{\textsf{P}}_{\mathrm{chio}}[\tilde{B}]\quad .
\end{equation}
\end{definition}
\begin{definition}[$\lvert \frac12\mathrm{C}_{(s,t)}^{\breve{I}}(\cdot)\rvert$ and
$\textup{\textsf{P}}_{\mathrm{chio}}^{\lvert\cdot\rvert, I}$]
For every $(s,t)\in\mathbb{Z}_{\geq 2}^2$ and every $I\subseteq [s-1]\times [t-1]$ define
a map
\begin{equation}
\lvert\tfrac12\mathrm{C}_{(s,t)}^{\breve{I}}\rvert\colon \{\pm\}^{\breve{I}}
\longrightarrow \{0,1\}^I \quad , \qquad
A \longmapsto \tfrac12\cdot \lvert \mathrm{C}_{(s,t)}(A) \rvert \quad ,
\end{equation}
where $\lvert\mathrm{C}_{(s,t)} (A)\rvert := \bigl (\lvert \det
\left(\begin{smallmatrix}
a_{i,j} & a_{i,t} \\
a_{s,j} & a_{s,t}
\end{smallmatrix}\right)\rvert\bigr)_{(i,j)\in I} \in \{0,2\}^I$.
Furthermore, define
\begin{equation}
\textup{\textsf{P}}_{\mathrm{chio}}^{\lvert\cdot\rvert, I}\colon\mathfrak{P}(\{0,1\}^I)
\longrightarrow [0,1] \cap \mathbb{Q} \quad , \qquad \mathcal{B}
\longmapsto \frac{1}{2^{\lvert\breve{I}\rvert}}\sum_{B\in\mathcal{B}}
\bigl \lvert (\lvert\tfrac12\mathrm{C}_{(s,t)}^{\breve{I}}\rvert)^{-1}(B) \bigr \rvert \quad .
\end{equation}
\end{definition}
\begin{definition}[the entry-specification-events $\mathcal{E}_B^J$]
\label{def:entryspecificationevents}
For $(s,t)\in\mathbb{Z}_{\geq 2}^2$,
$\emptyset \subseteq I \subseteq J \subseteq [s-1]\times[t-1]$,
and $B\in \{0,\pm\}^I$ let $\mathcal{E}_B^J :=
\bigl \{ \tilde{B} \in \{0,\pm\}^J\colon \tilde{B}\mid_{\mathrm{Dom}(B)} = B \bigr \}$.
\end{definition}
Note that $\mathrm{Dom}(B) = I \subseteq J =\mathrm{Dom}(\tilde{B})$, hence
$\tilde{B}\mid_{\mathrm{Dom}(B)}$ is defined. If $\mathrm{Dom}(B)=\emptyset$, i.e. $B=\emptyset$, then
$\mathcal{E}_{\emptyset}^J = \{0,\pm\}^J$, and if $\mathrm{Dom}(B)=J$,
then $\mathcal{E}_B^J = \{B\}$. Furthermore,
$\lvert \mathcal{E}_B^J \rvert = 3^{\lvert J \rvert - \lvert I \rvert}$ for
arbitrary $\emptyset\subseteq I \subseteq J \subseteq [s-1]\times [t-1]$
and $B\in\{0,\pm\}^I$.
In this paper we intend to use graph-theoretical language. For the sake of
specificity and ease of reference, we will explicitly name the set of auxiliary
labelled bipartite graphs that we will talk about (and give it a vertex set which
blends well with the matrix setting).
\begin{definition}
For every $(s,t)\in\mathbb{Z}_{\geq 2}^2$ denote by $\mathrm{BG}_{s,t}$ the $2^{(s-1)\cdot (t-1)}$-element
set of all bipartite graphs $X=(V_1\sqcup V_2,E)$ with
$V_1 =\{(i,t)\colon 1\leq i \leq s-1\}$ and $V_2 =\{(s,j)\colon 1\leq j \leq t-1\}$.
\end{definition}
There is an obvious bijection $\mathrm{BG}_{s,t} \longleftrightarrow \{0,1\}^{[s-1]\times [t-1]}$.
Associating with a (partially specified) $\{0,\pm\}$-matrix the following bipartite
signed graph will be helpful in our study of $\textup{\textsf{P}}_{\mathrm{chio}}$. The definition can be
summarized by saying that $\mathrm{X}$ interprets a $B\in\{0,\pm\}^I$ as a bipartite
adjacency matrix in the natural way (while ignoring the signs), and that $\sigma$ takes
the signs in $B$ as a definition of a sign function.
\begin{definition}[$\mathrm{X}$ and $\sigma$]\label{def:XBandecXB}
For every $(s,t)\in\mathbb{Z}_{\geq 2}^2$ and every $0\leq k \leq (s-1)(t-1)$ define
\begin{equation}\label{eq:def:XBandecXB}
\mathrm{X}^{k,s,t}\colon \bigsqcup_{I \in \binom{[s-1]\times [t-1]}{k}} \{ 0, \pm \}^I
\longrightarrow \mathrm{BG}_{s,t},\quad B \longmapsto \mathrm{X}_B^{k,s,t}
\end{equation}
by letting vertex-set and edge-set be defined as
\begin{align}
\mathrm{V}(\mathrm{X}_B^{k,s,t}) & :=
(\mathrm{Dom}(B))^{\breve{}}\setminus \mathrm{Dom}(B) \setminus \{(s,t)\} \quad ,
\label{def:XBandecXB:definitionofvertexset} \\
\mathrm{E}(\mathrm{X}_B^{k,s,t}) & :=
\bigsqcup_{(i,j) \in \mathrm{Supp}(B)} \bigl \{ \{(i,t),(s,j)\} \bigr \}\quad .
\label{def:XBandecXB:definitionofedgeset}
\end{align}
Define $\sigma_B\colon \mathrm{E}(\mathrm{X}_B) \rightarrow \{\pm\}$ by
$\sigma_B (\{(i,t),(s,j)\}) := b_{i,j} \in \{\pm\}$
for every $\{(i,t),(s,j)\}\in \mathrm{E}(\mathrm{X}_B)$. \\
\end{definition}
If $k=0$, hence $I=\emptyset$, hence $B=\emptyset$ is the empty matrix,
then $\mathrm{X}_B^{k,s,t}$ is the empty graph $(\emptyset,\emptyset)$ and
$\sigma_B = \emptyset$ is the empty function. Note that while for
a $B\in\{0,\pm\}^I$ the set $\mathrm{V}(\mathrm{X}_B)$ depends only on $I=\mathrm{Dom}(B)$,
the set $\mathrm{E}(\mathrm{X}_B)$ depends on $\mathrm{Supp}(B)$ and $\sigma_B$ even depends
on $B$ itself.
When we take the \emph{image} of a matrix $B\in \{0,\pm\}^I$ under $\mathrm{X}^{k,s,t}$,
then usually we will know what $I\in\binom{[s-1]\times [t-1]}{k}$ we are talking
about and then the superscripts $k,s,t$ give redundant information.
Whenever possible we will suppress the superscripts in such a situation
and only write $\mathrm{X}_B$. When we take \emph{preimages} of a graph $X\in\mathrm{BG}_{s,t}$
under $\mathrm{X}^{k,s,t}$, however, the full notation has to be used since in general
it is not possible to tell $k$ from the labelled graph $X$ (let alone from its
isomorphism type). As an example for this, consider the graph $X\in\mathrm{BG}_{4,4}$
with vertex set $\{(1,4),(2,4),(3,4)\}\sqcup\{(4,1),(4,2),(4,3)\}$
and edge set $\{ \{(1,4),(4,1)\},\{(2,4),(4,1)\},\{(2,4),(4,2)\},\{(1,4),(4,2)\}\}$,
which is isomorphic to a $4$-circuit with two additional isolated vertices. Then we
have $\mathrm{X}^{5,4,4}_{B_1} = \mathrm{X}_{B_2}^{6,4,4} = X$ for $B_1 :=\left(\begin{smallmatrix}
1 & 1 & \\ 1 & 1 & \\ & & 0
\end{smallmatrix}\right)\neq B_2 :=\left(\begin{smallmatrix}
1 & 1 & \\ 1 & 1 & 0 \\ & 0 &
\end{smallmatrix}\right)$. Here, $\mathrm{dom}(B_1) = 5 \neq 6 = \mathrm{dom}(B_2)$.
In the following, we deliberately do not define `isomorphism type of a graph' more
precisely. We would not have much use for any of the existing formalizations of an
unlabelled graph.
\begin{definition}[$\mathrm{ul}$, $\beta_1^{\mathrm{ul}}$]
Let $\mathrm{ul}$ be the map which assigns a labelled graph to its isomorphism type.
Let $\beta_1^{\mathrm{ul}}\colon \mathrm{ul}(\mathrm{BG}_{n,n}) \to \mathbb{Z}_{\geq 0}$ be the map which assigns an
unlabelled bipartite graph to its $1$-dimensional Betti number.
\end{definition}
\begin{definition}[$\leftsubscript{\mathrm{ul}}{\upX}^{k,s,t}$]\label{def:upXul}
For $(s,t) \in \mathbb{Z}_{\geq 2}^2$ and $0\leq k \leq (s-1)(t-1)$ let
$\leftsubscript{\mathrm{ul}}{\upX}^{k,s,t} := \mathrm{ul}\circ\mathrm{X}^{k,s,t}$.
\end{definition}
If $\mathfrak{X}$ is some (verbal, pictorial, ...) description of an isomorphism
type of graphs, we can now take its preimage
$(\leftsubscript{\mathrm{ul}}{\upX}^{k,s,t})^{-1}(\mathfrak{X}) \subseteq \bigsqcup_{I\in \binom{[s-1]\times [t-1]}{k}}
\{0,\pm\}^I$. To analyse how $\textup{\textsf{P}}_{\mathrm{chio}}$ and $\textup{\textsf{P}}_{\mathrm{lcf}}$ relate to one another,
it is useful to have the following notations.
\begin{definition}[failure sets]\label{def:parametrizedfailuresets}
For every $k\geq 0$, $n\geq 2$, $\ell\in \mathbb{Q}_{\geq 0}$ and $p \in [0,1]\cap \mathbb{Q}$ let
\begin{enumerate}[label={\rm(\arabic{*})}]
\item $\mathcal{F}^{\mathrm{M}}(k,n)$ $:=$
$\{$ $B\in\{0,\pm\}^I\colon I\in\binom{[n-1]^2}{k},\;
\textup{\textsf{P}}_{\mathrm{chio}}[\mathcal{E}_B^{[n-1]^2}] \neq \textup{\textsf{P}}_{\mathrm{lcf}}[\mathcal{E}_B^{[n-1]^2}]$ $\}$ \quad ,
\item $\mathcal{F}_{\cdot\ell}^{\mathrm{M}}(k,n)$ $:=$
$\{$ $B\in\mathcal{F}^{\mathrm{M}}(k,n)\colon
\textup{\textsf{P}}_{\mathrm{chio}}[\mathcal{E}_B^{[n-1]^2}] = \ell\cdot \textup{\textsf{P}}_{\mathrm{lcf}}[\mathcal{E}_B^{[n-1]^2}]$ $\}$
$\subseteq$ $\mathcal{F}^{\mathrm{M}}(k,n)$ \quad ,
\item $\mathcal{F}_{=p}^{\mathrm{M}}(k,n) := $
$\{$ $B\in \mathcal{F}^{\mathrm{M}}(k,n)\colon
\textup{\textsf{P}}_{\mathrm{chio}}[\mathcal{E}_B^{[n-1]^2}] = p$ $\}$
$\subseteq$ $\mathcal{F}^{\mathrm{M}}(k,n)$ \quad ,
\item\label{def:graphversionsofthefailuresets}
$\mathcal{F}^{\mathrm{G}}(k,n) :=
\leftsubscript{\mathrm{ul}}{\upX}^{k,n,n}(\mathcal{F}^{\mathrm{M}}(k,n))$,
$\mathcal{F}_{\cdot \ell}^{\mathrm{G}}(k,n) :=
\leftsubscript{\mathrm{ul}}{\upX}^{k,n,n}(\mathcal{F}_{\cdot\ell}^{\mathrm{M}}(k,n))$, \\
and $\mathcal{F}_{=p}^{\mathrm{G}}(k,n) :=
\leftsubscript{\mathrm{ul}}{\upX}^{k,n,n}(\mathcal{F}_{=p}^{\mathrm{M}}(k,n))$.
\end{enumerate}
We abbreviate $\mathcal{F}^{\mathrm{M}}(k,n) := \mathcal{F}^{\mathrm{M}}(k,n,n)$, and
analogously for all the other sets just defined.
\end{definition}
Obviously, $\mathcal{F}_{\cdot 1}^{\mathrm{M}}(k,n) = \emptyset$
and $\mathcal{F}_{\cdot 0}^{\mathrm{M}}(k,n) = \mathcal{F}_{= 0}^{\mathrm{M}}(k,n)$ for
all $k$ and $n$. Item \ref{relationbeweenchiomeasureandlazycoinflipmeasuregovernedbyfirstbettinumber} in
Theorem \ref{thm:graphtheoreticalcharacterizationofthechiomeasure} will teach us
that $\mathcal{F}_{\cdot \ell}^{\mathrm{M}}(k,n) = \emptyset$ for every
$\ell\notin \{ 0 \} \sqcup \{2^i\colon i\in\mathbb{Z}_{\geq 0}\}$ (hence in particular
$\textup{\textsf{P}}_{\mathrm{chio}}[\mathcal{E}_B^{[n-1]^2}] \geq \textup{\textsf{P}}_{\mathrm{lcf}}[\mathcal{E}_B^{[n-1]^2}]$ for
every $B\in\mathcal{F}^{\mathrm{M}}(k,n)$ with
$\textup{\textsf{P}}_{\mathrm{chio}}[\mathcal{E}_B^{[n-1]^2}] > 0$).
\begin{definition}[matrix-circuit, $\mathrm{Cir}(s,n)$]
\label{def:circuitofmatrixentrypositions}
For every $(s,t)\in\mathbb{Z}_{\geq 2}^2$ and every $L \subseteq [s-1]\times [t-1]$
with even $l := \lvert L \rvert$, the set $L$ is called a \emph{matrix-$l$-circuit}
if and only if $\mathrm{X}_{\{1\}^L}$ is a graph-theoretical $l$-circuit.
Moreover, $\mathrm{Cir}(l,s,t):=\{ L\subseteq [s-1]\times [t-1]\colon \lvert L \rvert = l,\,
\text{$L$ is a matrix-$l$-circuit}\}$ and $\mathrm{Cir}(l,n):=\mathrm{Cir}(l,n,n)$.
\end{definition}
\begin{definition}[$(-)$-constant, $(+)$-proper vertex $2$-colouring of a
signed graph]
For a graph $X=(V,E)$ and a $\sigma\in \{\pm\}^E$, a function
$\mathrm{c}\in \{\pm\}^V$ is called
\emph{($\sigma$, $-$)-constant, ($\sigma$, $+$)-proper} if and only if
$\mathrm{c}(u) = \mathrm{c}(v)$ for every $e = uv\in \mathrm{E}(X)$ with $\sigma(e) = -$
and
$\mathrm{c}(u) \neq \mathrm{c}(v)$ for every $e = uv\in \mathrm{E}(X)$ with $\sigma(e) = +$.
\end{definition}
\begin{definition}[$\mathrm{Col}(X,\sigma)$]\label{def:ColecX}
For a graph $X=(V,E)$ and a $\sigma\in \{\pm\}^E$ let $\mathrm{Col}(X,\sigma)$ be the set of
all ($\sigma$, $-$)-constant, ($\sigma$, $+$)-proper vertex-$2$-colourings
$\mathrm{c}\in \{\pm\}^V$.
\end{definition}
\begin{definition}[rank-level-sets of matrices]\label{def:ranksets}
For $(s,t)\in\mathbb{Z}_{\geq 2}^2$, $0\leq r \leq \min(s,t)$, $R$ an integral domain,
$U\subseteq R$ and $\mathcal{R}\in\mathfrak{P}(\{0,1,\dotsc,\min(s,t)\})$ let
$\mathrm{Ra}_r(U^{[s]\times[t]} )$ $:=$ $\{ A\in U^{[s]\times[t]}\colon \mathrm{rk}(A) = r\}$,
$\mathrm{Ra}_{\mathcal{R}}(\{\pm\}^{[s]\times [t]}) := \bigsqcup_{r\in\mathcal{R}}
\mathrm{Ra}_r(\{\pm\}^{[s]\times [t]})$ and
$\mathrm{Ra}_{<r}(U^{[s]\times[t]} )$ $:=$ $\mathrm{Ra}_{\{0,1,\dotsc,r-1\}}(\{\pm\}^{[s]\times [t]})$.
\end{definition}
\section{Lemmas}
We will use the following elementary fact:
\begin{lemma}\label{lem:elementarypreimagestatements:finvffinv}
$f^{-1}(f(f^{-1}(U))) = f^{-1}(U)$ for any map $f\colon A\to B$ and any
subset $U\subseteq B$. \hfill $\Box$
\end{lemma}
The following simple statement is essential for the approach developed in the
present paper. More information on this identity can be found in
\cite[last paragraph of Section 9]{MR1500275} and
\cite[Ch. 4, p. 282, Exerc. 43]{MR1181420}. The formulation given here differs from
those in \cite[p. 11]{chio} and \cite{MR1500275} in that $a_{n,n}$ instead of $a_{1,1}$
is taken to be the pivot. This seems more convenient for handling the indices of a
Chio condensate.
\begin{lemma}[Chio's identity]\label{lem:chioidentity}
For $n\geq 2$, $R$ an integral domain and $(a_{i,j}) = A\in R^{[n]^2}$,
\begin{equation}\label{eq:chiocondensation}
\det \bigl ( \mathrm{C}_{(n,n)}(A) \bigr ) = a_{n,n}^{n-2} \cdot \det(A) \quad .
\end{equation}
\end{lemma}
\begin{proof}
This is stated by Chio in \cite[p. 11, Th{\'e}or{\`e}me 4, equation (20)]{chio} and
he proves it on pp. 6--11 (the notation `$\pm a_0b_1$' employed in
\cite[equation (13'')]{chio} is defined at the beginning of p. 6 of \cite{chio}).
To contemporary eyes, this is an easy consequence of the behavior of determinants
under linear transformations, cf. \cite{MR649067} for a direct proof of the version
with pivot $a_{1,1}$. Moreover, this is a special case of
\emph{Sylvester's determinant identity}. To see this, set $k=1$ in formula $(8)$ of
\cite{MR1500275} to get a version of \eqref{eq:chiocondensation} with pivot
$a_{1,1}$. Obvious modifications in the proof in \cite{MR1500275} yield the version
with pivot $a_{n,n}$.
\end{proof}
The following three assertions are obviously true:
\begin{corollary}\label{lem:equivalenceofvanishingofdeterminants}
For every $A\in \{\pm\}^{[n]^2}$, $\det(A) = 0$ if and only if
$\det(\tfrac12\mathrm{C}_{(n,n)}(A)) = 0$. \hfill $\Box$
\end{corollary}
\begin{lemma}[value of lazy coin flip distribution on single matrix]
\label{lem:valueoflazycoinflipdistribution}
For every $\emptyset\subseteq I \subseteq [n-1]^2$ and every
$B \in \{0,\pm\}^{I}$, $\textup{\textsf{P}}_{\mathrm{lcf}}[\mathcal{E}_B] = (\frac12)^{ \mathrm{dom}(B) + \mathrm{supp}(B) }$.
\hfill $\Box$
\end{lemma}
\begin{lemma}\label{lem:onepointwedgepreservesbalancedness}
For any two disjoint graphs $X_1$ and $X_2$ and any two sign functions
$\sigma_{X_1}\in \{\pm\}^{\mathrm{E}(X_1)}$ and $\sigma_{X_2}\in\{\pm\}^{\mathrm{E}(X_2)}$, and for
every graph $X$ obtained by a one-point wedge of $X_1$ and $X_2$ at two arbitrary
vertices, the sign function $\sigma_X\in \{\pm\}^{\mathrm{E}(X)}$ obtained by uniting the
maps $\sigma_{X_1}$ and $\sigma_{X_2}$ is balanced if and only if both
$(X_1,\sigma_{X_1})$ and $(X_2,\sigma_{X_2})$ are balanced. \hfill $\Box$
\end{lemma}
The following will be needed for counting failures of equality
of $\textup{\textsf{P}}_{\mathrm{chio}}$ and $\textup{\textsf{P}}_{\mathrm{lcf}}$.
\begin{lemma}
\label{lem:numberofmatrixcircuitsofgivenlengthwithingivencartesianproduct}
$\lvert \mathrm{Cir}(2j,s,t) \rvert = \binom{s-1}{j}\cdot\binom{t-1}{j}\cdot
\frac{j!(j-1)!}{2}$ for every $n\geq 2$ and every
$1\leq j \leq \min(\lfloor\frac{s}{2}\rfloor, \lfloor\frac{t}{2}\rfloor)$.
\end{lemma}
\begin{proof}
For every $N'\in \binom{[s-1]}{j}$ and $N''\in \binom{[t-1]}{j}$ let
$\mathrm{Cir}(2j,s,t,N',N'') := \{ S\in \mathrm{Cir}(2j,s,t)\colon \mathrm{p}_1(S) = N',\;
\mathrm{p}_2(S) = N'' \}$. Obviously, $\lvert\mathrm{Cir}(2j,s,t)\rvert =
\sum_{N'\in \binom{[s-1]}{j}} \sum_{N''\in \binom{[t-1]}{j}} \lvert\mathrm{Cir}(2j,s,t,N',N'')\rvert$.
To count $\mathrm{Cir}(2j,s,t,N',N'')$, define $\mathrm{Perm}(N',N'')$ $:=$ $\{$ $P$ $\colon$ $P$ is
a permutation matrix, $\mathrm{Supp}(P)$ $\subseteq$ $N'\times N''$ $\}$. Define a binary
relation $\mathfrak{R} \subseteq \mathrm{Cir}(2j,s,t,N',N'')\times \mathrm{Perm}(N',N'')$ by letting
$(S,P) \in \mathfrak{R}$ if and only if $\mathrm{Supp}(S)\supseteq \mathrm{Supp}(P)$. For every
$S\in\mathrm{Cir}(2j,s,t,N',N'')$ we have
$\lvert \{ P\in\mathrm{Perm}(N',N'')\colon (S,P)\in\mathfrak{R} \} \rvert = 2$. On the other
hand, for every $P\in\mathrm{Perm}(N',N'')$ we have
$\lvert \{ S\in\mathrm{Cir}(2j,s,t,N',N'')\colon (S,P)\in\mathfrak{R} \} \rvert = (j-1)!$.
Therefore, $(j-1)!$ $\cdot$ $j!$ $=$ $(j-1)!$ $\cdot$ $\lvert \mathrm{Perm}(N',N'')\rvert$
$=$ $\sum_{P\in\mathrm{Perm}(N',N'')}$ $\lvert \{$ $S\in\mathrm{Cir}(2j,s,t,N',N'')\colon$
$(S,P)$ $\in$ $\mathfrak{R}$ $\} \rvert$ $=$ $\sum_{S\in\mathrm{Cir}(2j,s,t,N',N'')}$
$\lvert \{$ $P$ $\in$ $\mathrm{Perm}(N',N'')\colon$ $(S,P)$ $\in$ $\mathfrak{R}$
$\} \rvert$ $=$ $2$ $\cdot$ $\lvert \mathrm{Cir}(2j,s,t,N',N'') \rvert$.
\end{proof}
The following is contained in K{\H{o}}nig's 1936 classic \cite{MR0036989}.
\begin{lemma}[D. K{\H{o}}nig]\label{lem:equivalenceofexistenceofbconstantrpropervertex2coloringandcyclicallyrevenness}
Let $\mathfrak{X}$ be a labelled or an unlabelled graph. Then:
\begin{enumerate}[label={\rm(K{\H{o}}\arabic{*})}]
\item\label{item:equivalenceofexistenceofbconstantrpropervertex2coloringandcyclicallyrevenness} $S_{\mathrm{bal}}(\mathfrak{X})\neq\emptyset$ if and only if
$\mathrm{Col}(\mathfrak{X},\sigma)\neq\emptyset$.
\item\label{item:cardinalityofsetofbconstantrproper2colorings}
Let $\sigma\colon \{\pm\}^E$ be arbitrary. Then
$\mathrm{Col}(\mathfrak{X},\sigma)\neq \emptyset$ if and only if
$\lvert\mathrm{Col}(\mathfrak{X},\sigma)\rvert = 2^{\beta_0(\mathfrak{X})}$.
\item\label{item:numberofbalancedsignfunctions}
$\lvert \{\sigma\in\{\pm\}^{\mathrm{E}(\mathfrak{X})}\colon
(\mathfrak{X},\sigma)\;\mathrm{balanced}\}\rvert =
2^{f_0(\mathfrak{X}) - \beta_0(\mathfrak{X})}$.
\end{enumerate}
\end{lemma}
\begin{proof}
Modulo terminology a proof for \ref{item:equivalenceofexistenceofbconstantrpropervertex2coloringandcyclicallyrevenness} can be found in \cite[p. 152, Satz 11]{MR0036989}
(for the definition of `$p$-Teilgraph' cf. \cite[p. 149, Paragraph 3]{MR0036989}).
Statement \ref{item:equivalenceofexistenceofbconstantrpropervertex2coloringandcyclicallyrevenness} is also proved in \cite[Theorem 3]{MR0067468}.
Statement \ref{item:cardinalityofsetofbconstantrproper2colorings} is implicit
in the proof of \cite[p. 152, Satz 14]{MR0036989} and can easily be proved
directly by induction on $\lvert E \rvert$. For a proof of
\ref{item:numberofbalancedsignfunctions} cf. e.g. \cite[p. 152, Satz 14]{MR0036989}.
\end{proof}
While for a given graph $X = (V,E)$ and a given sign function
$\sigma\colon E\rightarrow \{\pm\}$, the decision problem of whether $(X,\sigma)$
balanced is trivially in co-NP, the less obvious fact that it is also in NP
follows from \ref{item:equivalenceofexistenceofbconstantrpropervertex2coloringandcyclicallyrevenness}: any ($\sigma$, $-$)-constant, ($\sigma$, $+$)-proper
vertex-$2$-colouring $\mathrm{c}\colon V \rightarrow \{\pm\}$ is a polynomially-sized
certificate for $(X,\sigma)$ being balanced. However, the problem is not only in the
intersection of these two classes but easily seen to be in P:
\begin{corollary}\label{cor:decidingwhetherecXiscyclicallyreven}
For every graph $X = (V,E)$ and every sign function $\sigma\colon E \rightarrow \{\pm\}$, the decision problem whether $(X,\sigma)$ is balanced can be solved in
time $O( f_0(X) + f_1(X))$.
\end{corollary}
\begin{proof}
By \ref{item:equivalenceofexistenceofbconstantrpropervertex2coloringandcyclicallyrevenness}, the question is equivalent to whether there exists a a $(-)$-constant,
$(+)$-proper vertex-$2$-colouring $\mathrm{c}\colon V\rightarrow \{\pm\}$. It is easy to
see that an obvious greedy algorithm via a (e.g.) depth-first search on $X$ succeeds
in finding such a colouring if and only if such a colouring exists. Moreover, the
algorithm requires time $O(f_0(X) + f_1( X ))$.
\end{proof}
The following simple lemma encapsulates a basic mechanism linking Chio
condensation with the auxiliary graph-theoretical viewpoint. For want of topologies
on source or target, `$k$-fold cover' is nothing but shorthand for `surjective map
each of whose fibres has cardinality $k$'.
\begin{lemma}\label{lem:theparametrizedcover}
For every $(s,t)\in\mathbb{Z}_{\geq 2}^2$, arbitrary
$\emptyset \subseteq I \subseteq J \subseteq [s-1]\times [t-1]$, every
$B\in \{0,\pm\}^I$, and with $h(I,J) := \lvert \breve{J} \rvert - \lvert I \rvert
- \lvert \mathrm{p}_1(I)\rvert - \lvert \mathrm{p}_2(I)\rvert
\in \mathbb{Z}_{\geq 1}$, there exists an $2^{h(I,J)}$-fold cover
\begin{equation}
\Phi\colon (\tfrac12\mathrm{C}_{(s,t)}^{\breve{J}})^{-1}(\mathcal{E}_{B}^J) \longrightarrow
\mathrm{Col}(\mathrm{X}_B,\sigma_B)\quad .
\end{equation}
\end{lemma}
\begin{proof}
Let $s$, $t$, $I$, $J$ and $B = (b_{i,j})_{(i,j)\in I}$ be given as stated. The claim
$h(I,J)\in \mathbb{Z}_{\geq 1}$ is true since Definition \ref{def:chioextensionofaset}
implies $\lvert \breve{J} \rvert =
1 + \lvert \mathrm{p}_1(J)\rvert + \lvert \mathrm{p}_2(J)\rvert + \lvert J \rvert$
and because $J\supseteq I$ implies $\lvert J \rvert \geq \lvert I \rvert$,
$\lvert \mathrm{p}_1(J) \rvert\geq\lvert\mathrm{p}_1(I)\rvert$
and $\lvert\mathrm{p}_2(J)\rvert\geq \lvert\mathrm{p}_1(I)\rvert$.
If $\mathrm{Col}(\mathrm{X}_B,\sigma_B) = \emptyset$, the statement of the lemma is vacuously
true (there not being any point of the target for which the definition of a
covering would have to hold). We therefore can assume
that $\mathrm{Col}(\mathrm{X}_B,\sigma_B)\neq \emptyset$. We now first
show that this implies that $(\tfrac12\mathrm{C}_{(s,t)}^{\breve{J}})^{-1}(\mathcal{E}_B^J)
\neq \emptyset$. Then we will construct a cover of the stated kind.
To prove $(\tfrac12\mathrm{C}_{(s,t)}^{\breve{J}})^{-1}(\mathcal{E}_B^J) \neq \emptyset$,
choose an arbitrary $\mathrm{c}\in \mathrm{Col}(\mathrm{X}_B,\sigma_B)$ and define
$A = (a_{i,j})\in\{\pm\}^{\breve{J}}$ by $a_{s,t} := +$,
$a_{i,t} := \mathrm{c}((i,t))$ for every $i\in\mathrm{p}_1(J)$,
$a_{s,j} := \mathrm{c}((s,j))$ for every $j\in\mathrm{p}_2(J)$. Moreover,
for every $(i,j)\in J$ let $a_{i,j} := b_{i,j}$ if $b_{i,j}\neq 0$ and
$a_{i,j}:=\mathrm{c}((i,t))\cdot\mathrm{c}((s,j))$ if $b_{i,j} = 0$. We now show that
$\tfrac12\mathrm{C}_{(s,t)}^{\breve{J}}(A)\mid_{\mathrm{Dom}(B)} = B$. Let $(i,j)\in I = \mathrm{Dom}(B)$ be
arbitrary. If $b_{i,j}=0$, then $\tfrac12\mathrm{C}_{(s,t)}^{\breve{J}}(A)[i,j] =
\tfrac12( a_{i,j} a_{s,t} - a_{i,t} a_{s,j} )
= \tfrac12\bigl ( a_{i,j} - \mathrm{c}((i,t)) \mathrm{c}((s,j)) \bigr ) =
\tfrac12\bigl(\mathrm{c}((i,t)) \mathrm{c}((s,j)) - \mathrm{c}((i,t)) \mathrm{c}((s,j)) \bigr)
= 0 = b_{i,j}$. If $b_{i,j}\neq 0$, then
$\tfrac12\mathrm{C}_{(s,t)}^{\breve{J}}(A)[i,j] = \tfrac12( a_{i,j} a_{s,t} - a_{i,t} a_{s,j} )
= \tfrac12\bigl (b_{i,j} - \mathrm{c}((i,t)) \mathrm{c}((s,j)) \bigr )$.
Now if $b_{i,j} = -$, then $\mathrm{c}\in\mathrm{Col}(\mathrm{X}_B,\sigma_B)$ implies
$\mathrm{c}((i,t)) = \mathrm{c}((s,j))$, hence
$\tfrac12\mathrm{C}_{(s,t)}^{\breve{J}}(A)[i,j] = \tfrac12((-) - (+)) = - = b_{i,j}$,
and if $b_{i,j} = +$, then $\mathrm{c}\in\mathrm{Col}(\mathrm{X}_B,\sigma_B)$ implies
$\mathrm{c}((i,t)) \neq \mathrm{c}((s,j))$, hence
$\tfrac12\mathrm{C}_{(s,t)}^{\breve{J}}(A)[i,j] = \tfrac12((+) - (-)) = + = b_{i,j}$.
We have proved that $A\in(\tfrac12\mathrm{C}_{(s,t)}^{\breve{J}})^{-1}(\mathcal{E}_B^J)$, and
therefore $(\tfrac12\mathrm{C}_{(s,t)}^{\breve{J}})^{-1}(\mathcal{E}_B^J) \neq \emptyset$.
We can therefore define a nonempty map
$\Phi\colon (\tfrac12\mathrm{C}_{(s,t)}^{\breve{J}})^{-1}(\mathcal{E}_B^J)
\rightarrow \mathrm{Col}(\mathrm{X}_B,\sigma_B)$ as follows: for every
$A$ $=$ $(a_{i,j})_{(i,j)\in\breve{J}}$ $\in$
$(\tfrac12\mathrm{C}_{(s,t)}^{\breve{J}})^{-1}(\mathcal{E}_B^J)
\subseteq \{\pm\}^{\breve{J}}$ we let $\Phi(A)$ be the function
$\mathrm{V}(X_B) \rightarrow \{\pm\}$ defined by $\Phi(A)((i,t)) := a_{i,t}$
and $\Phi(A)((s,j)) := a_{s,j}$ for all $i\in\mathrm{p}_1(I)$ and $j\in \mathrm{p}_2(I)$.
Claim 1. $\Phi$ is indeed a map of the stated kind, i.e.
$\Phi(A)\in \mathrm{Col}(\mathrm{X}_B,\sigma_B)$. Proof: Let $\{ (i,t),(s,j) \} \in \mathrm{E}(X_B)$ be
arbitrary. There are two cases. If $\sigma_B(\{ (i,t),(s,j) \}) = -$ then
$b_{i,j} = -$ by Definition \ref{def:XBandecXB}. Moreover, since
$(\tfrac12\mathrm{C}_{(s,t)}^{\breve{J}}(A))\rvert_{\mathrm{Dom}(B)}= B$ by the choice of $A$, it
follows that for every $(i,j)\in I\subseteq J$ we have
$- = b_{i,j} = (\tfrac12\mathrm{C}_{(s,t)}^{\breve{J}}(A))[i,j]
= \tfrac12\cdot (a_{i,j}a_{s,t} - a_{i,t}a_{s,j})$. In view of
$a_{i,j},\ a_{s,t},\ a_{i,t},\ a_{s,j}\in \{\pm\}$, this equation
implies $\Phi(A)((i,t)) = a_{i,t} = a_{s,j} = \Phi(A)((s,j))$. This
proves that $\Phi(A)$ is ($\sigma_B$, $-$)-constant. If
$\sigma_B(\{(i,t),(s,j)\}) = +$ then $b_{i,j} = +$ by Definition
\ref{def:XBandecXB}. Again by the choice of $A$, it is true
that $+ = b_{i,j} = (\tfrac12\mathrm{C}_{(s,t)}^{\breve{J}}(A))[i,j] =
\tfrac12\cdot (a_{i,j}a_{s,t} - a_{i,t}a_{s,j})$. Again in view of
$a_{i,j},\ a_{s,t},\ a_{i,t},\ a_{s,j}\in \{\pm\}$, this equation
implies $\Phi(A)((i,t)) = a_{i,t} \neq a_{s,j} = \Phi(A)((s,j))$. This proves
that $\Phi(A)$ is ($\sigma_B$, $+$)-proper and hence Claim 1.
Claim 2. $\Phi$ is surjective and every fibre under $\Phi$ has cardinality
$2^{h(I,J)}$. Proof: Let an arbitrary $\mathrm{c} \in \mathrm{Col}(\mathrm{X}_B,\sigma_B)$ be
given. We are now looking for those
$A\in (\tfrac12\mathrm{C}_{(s,t)}^{\breve{J}})^{-1}(\mathcal{E}_B^J)$
with $\Phi(A) = \mathrm{c}$. Since the definition of $\Phi$ demands
$\Phi(A)((i,t)) = a_{i,t}$ and $\Phi(A)((s,j)) = a_{s,j}$
for all $(i,t)$ and $(s,j)$ $\in \mathrm{V}(X_B)$, it follows that with regard to the
$\lvert \mathrm{p}_1(I) \rvert + \lvert \mathrm{p}_2(I) \rvert$ different entries $a_{i,j}$
with $(i,j)\in \mathrm{V}(\mathrm{X}_{B})$ we know from the outset that we have no choice
but to define $a_{i,t} := \mathrm{c}((i,t))$ and $a_{s,j} := \mathrm{c}((s,j))$.
Furthermore, since $A$ must be in
$(\tfrac12\mathrm{C}_{(s,t)}^{\breve{J}})^{-1}(\mathcal{E}_B^J)$, there is,
for every $(i,j)\in I\subseteq J$, the condition that
$b_{i,j} = (\tfrac12\mathrm{C}_{(s,t)}^{\breve{J}}(A))[i,j] =
\tfrac12\cdot (\mathrm{C}_{(s,t)}^{\breve{J}}(A)[i,j]) =
\tfrac12\cdot( a_{i,j}a_{s,t} - a_{i,t}a_{s,j} ) =
\tfrac12( a_{i,j}a_{s,t} - \mathrm{c}((i,t))\ \mathrm{c}((s,j)) )$,
where in the last step we have used the information about $A$ that we
already have. Now there are three cases that can occur.
\textsl{Case 1.} $b_{i,j} = -$. Then by Definition \ref{def:XBandecXB} we have
$\{(i,t),(s,j)\} \in \mathrm{E}(X_B)$ and $\sigma_B(\{(i,t),(s,j)\}) = -$. Therefore,
due to the fact that $\mathrm{c} \in \mathrm{Col}(\mathrm{X}_B,\sigma_B)$ is
($\sigma_B$, $-$)-constant, $\mathrm{c}((i,t)) = \mathrm{c}((s,j))$.
Thus, in this case, $- = \tfrac12( a_{i,j}a_{s,t} - 1 )$,
equivalently, $a_{i,j} = - a_{s,t}$.
\textsl{Case 2.} $b_{i,j} = 0$. Then by Definition \ref{def:XBandecXB},
$\{(i,t),(s,j)\} \notin \mathrm{E}(X_B)$, hence $\sigma_B(\{(i,t),(s,j)\})$ is
not defined and therefore the product $\mathrm{c}((i,t))\cdot \mathrm{c}((s,j))$ in the
equation $0 = \tfrac12\cdot( a_{i,j}a_{s,t} - \mathrm{c}((i,t))\cdot \mathrm{c}((s,j)) )$
cannot be simplified further, but the equation itself can: it is equivalent
to $a_{i,j} = \mathrm{c}((i,t))\cdot \mathrm{c}((s,j))\cdot a_{s,t} \in \{\pm\}$
(where we used that $a_{s,t}^{-1} = a_{s,t}$).
\textsl{Case 3.} $b_{i,j} = +$. Then an entirely analogous argument as in
Case 1, but this time using the ($\sigma_B$, $+$)-properness of $\mathrm{c}$, shows
that then there is the equation $a_{i,j} = a_{s,t}$.
We now know what it means to require
$A\in (\tfrac12\mathrm{C}_{(s,t)}^{\breve{J}})^{-1}(\mathcal{E}_B^J)$ in the present situation:
among the $\lvert J \rvert$ entries of $A = (a_{i,j}) \in \{\pm\}^J$, there are the
$\lvert \mathrm{p}_1(I) \rvert + \lvert \mathrm{p}_2(I) \rvert$ `immediately determined' entries
$a_{i,j}$ which have ($i\in \mathrm{p}_1(I)$ and $j=t$) or ($i=s$ and $j\in \mathrm{p}_2(I)$),
and moreover the $\lvert I \rvert$ different entries $a_{i,j}$ with $(i,j)\in I$
which are determined by a system $ \{ a_{i,j} = h_{i,j} \colon (i,j)\in I\}$
of $\lvert I\rvert$ equations where the right-hand sides $h_{i,j}$ are defined by
the Cases 1-3 above. For the remaining $h(I,J)$ different
entries $a_{i,j}\in\{\pm\}$ (note that the pivot $a_{s,t}$ is among them since it
is on the right-hand side in Case 2, hence not determined by the system), the choice
of their value is free; any of the $2^{h(I,J)}$ possible choices gives an
$A\in (\tfrac12\mathrm{C}_{(s,t)}^{\breve{J}})^{-1}(\mathcal{E}_B^J)$. This proves that the
cardinality of the fibre $\Phi^{-1}(\mathrm{c})$ is indeed $2^{h(I,J)}$, and in particular
that $\Phi$ is surjective. Now Claim 2 and Lemma \ref{lem:theparametrizedcover} are
proved.
\end{proof}
Note that in the special case $I=J$, i.e. when all entries are specified, then
$h(I,J)=1$ and the statement says that there is a
double cover
$\Phi\colon (\tfrac12\mathrm{C}_{(s,t)}^{\breve{J}})^{-1}(\{B\}) \rightarrow
\mathrm{Col}(\mathrm{X}_B,\sigma_B)$. This corresponds to the freedom of choosing the sign
of the pivot. Now we can relate Chio condensation to balancedness:
\begin{lemma}\label{lem:graphtheoreticalcharacterizationofchiorealizability}
For every $(s,t)\in \mathbb{Z}_{\geq 2}^2$, every
$\emptyset \subseteq I\subseteq J\subseteq [s-1]\times [t-1]$
and every $B\in \{0,\pm\}^I$, the following statements are equivalent:
\begin{minipage}[b]{0.35\linewidth}
\begin{enumerate}[label={\rm(\arabic{*})}]
\item\label{lem:prop:associatededge2coloringofBiscyclicallyreven}
$(\mathrm{X}_B,\sigma_B)$ is balanced\quad ,
\item\label{lem:prop:thereexistsbconstantrpropervertex2coloring}
$\mathrm{Col}(\mathrm{X}_B,\sigma_B)\neq \emptyset$ \quad ,
\end{enumerate}
\end{minipage}
\begin{minipage}[b]{0.65\linewidth}
\begin{enumerate}[label={\rm(\arabic{*})},start=3]
\item\label{lem:prop:BisChiorealizable}
$(\tfrac12\mathrm{C}_{(s,t)}^{\breve{J}})^{-1}(\mathcal{E}_B^J) \neq \emptyset$ \quad ,
\item\label{lem:prop:cardinalityofinverseimage}
$\bigl \lvert (\tfrac12\mathrm{C}_{(s,t)}^{\breve{J}})^{-1}(\mathcal{E}_B^J)
\bigr \rvert = 2^{ \lvert \breve{J} \rvert - \mathrm{dom}(B) - f_0(\mathrm{X}_B) + \beta_0(\mathrm{X}_B) }$ \quad .
\end{enumerate}
\end{minipage}
\end{lemma}
\begin{proof}
Equivalence \ref{lem:prop:associatededge2coloringofBiscyclicallyreven}
$\Leftrightarrow$ \ref{lem:prop:thereexistsbconstantrpropervertex2coloring}
is true by \ref{item:equivalenceofexistenceofbconstantrpropervertex2coloringandcyclicallyrevenness} with $X:=\mathrm{X}_B$. Equivalence
\ref{lem:prop:thereexistsbconstantrpropervertex2coloring}
$\Leftrightarrow$ \ref{lem:prop:BisChiorealizable} is an immediate consequence of
Lemma \ref{lem:theparametrizedcover} (non-emptiness of the target of a surjective
map implies non-emptiness of its source; non-emptiness of the source
of \emph{any} map implies non-emptiness of its target).
As to \ref{lem:prop:BisChiorealizable}
$\Leftrightarrow$ \ref{lem:prop:cardinalityofinverseimage}, note that
by Lemma \ref{lem:theparametrizedcover}, there is the equation
$\lvert (\tfrac12\mathrm{C}_{(s,t)}^{\breve{J}})^{-1}(\mathcal{E}_B^J)\rvert
= 2^{\lvert \breve{J} \rvert - \mathrm{dom}(B) - f_0(\mathrm{X}_B)}\cdot
\lvert \mathrm{Col}(\mathrm{X}_B,\sigma_B) \rvert$, which may have the form $0=0$.
Now if \ref{lem:prop:BisChiorealizable}, then
$\mathrm{Col}(\mathrm{X}_B,\sigma_B) \neq \emptyset$ by the already proved equivalence
\ref{lem:prop:thereexistsbconstantrpropervertex2coloring}
$\Leftrightarrow$ \ref{lem:prop:BisChiorealizable}, therefore
Lemma \ref{item:cardinalityofsetofbconstantrproper2colorings} implies
$\lvert \mathrm{Col}(\mathrm{X}_B,\sigma_B) \rvert = 2^{\beta_0(\mathrm{X}_B)}$ and hence
\ref{lem:prop:cardinalityofinverseimage} is true. Conversely, if
\ref{lem:prop:cardinalityofinverseimage} is true, then this formula alone implies
\ref{lem:prop:BisChiorealizable}. This completes the proof of
\ref{lem:prop:BisChiorealizable} $\Leftrightarrow$
\ref{lem:prop:cardinalityofinverseimage} and also the proof of Lemma
\ref{lem:graphtheoreticalcharacterizationofchiorealizability}.
\end{proof}
As an example, consider the special case $s:=t:=n$,
$\{(1,1)\} =: I \subseteq J := [n-1]^2$, $B[(1,1)]:=0$,
i.e. $\mathcal{E}_B^J$ is the event that a
$\tilde{B} = (\tilde{b}_{i,j})\in \{0,\pm\}^{[n-1]^2}$ has $\tilde{b}_{1,1} = 0$.
For these data \ref{lem:prop:cardinalityofinverseimage}
in Lemma \ref{lem:graphtheoreticalcharacterizationofchiorealizability}
yields $2^{n^2-1}$. And indeed, it is easy to convince oneself directly that
there are $2^{n^2-1}$ possibilities to realize this event by
Chio condensates of sign matrices $A\in \{\pm\}^{[n]^2}$.
\section{Understanding the Chio measure}
\begin{theorem}[graph-theoretical characterization of the Chio measure
of entry-specification events]
\label{thm:graphtheoreticalcharacterizationofthechiomeasure}
For every $(s,t)\in\mathbb{Z}_{\geq 2}^2$, arbitrary
$\emptyset \subseteq I \subseteq J \subseteq [s-1]\times [t-1]$
and every $B\in \{0,\pm\}^I$:
\begin{enumerate}[label={\rm(C\arabic{*})}]
\item\label{characterizationofwhenchiomeasureispositive}
\textbf{positivity is determined by balancedness:} \\
$\textup{\textsf{P}}_{\mathrm{chio}}[\mathcal{E}_B^J] > 0$ if and only if $(\mathrm{X}_B,\sigma_B)$ is
balanced\quad ,
\item\label{explicitformulaforchiomeasureinthecaseofrealizability}
\textbf{absolute value is determined by the coboundary space:} \\
$\textup{\textsf{P}}_{\mathrm{chio}}[\mathcal{E}_B^J]>0$ if and only if
\begin{equation}
\textup{\textsf{P}}_{\mathrm{chio}}[\mathcal{E}_B^J] = (\tfrac12)^{\mathrm{dom}(B) + f_0(\mathrm{X}_B) - \beta_0(\mathrm{X}_B)}
= \frac{(\tfrac12)^{\mathrm{dom}(B)}}{\lvert \mathrm{B}^1(\mathrm{X}_B;\; \mathbb{Z}/2)\rvert} \quad ,
\end{equation}
\item\label{relationbeweenchiomeasureandlazycoinflipmeasuregovernedbyfirstbettinumber}
\textbf{relative value is determined by the cycle space:} \\
$\textup{\textsf{P}}_{\mathrm{chio}}[\mathcal{E}_B^J]>0$ if and only if
\begin{equation}
\textup{\textsf{P}}_{\mathrm{chio}}[\mathcal{E}_B^J] = 2^{\beta_1(\mathrm{X}_B)}\cdot \textup{\textsf{P}}_{\mathrm{lcf}}[\mathcal{E}_B^J]
= \lvert \mathrm{Z}_1(\mathrm{X}_B;\; \mathbb{Z}/2) \rvert \cdot \textup{\textsf{P}}_{\mathrm{lcf}}[\mathcal{E}_B^J] \quad .
\end{equation}
\end{enumerate}
\end{theorem}
\begin{proof}
As to \ref{characterizationofwhenchiomeasureispositive}, Definition
\ref{def:measurePchio} implies that $\textup{\textsf{P}}_{\mathrm{chio}}[\mathcal{E}_B^J] > 0$
if and only if $(\tfrac12\mathrm{C}_{(s,t)}^{\breve{J}})^{-1}(\mathcal{E}_B^J) \neq \emptyset$,
hence item \ref{characterizationofwhenchiomeasureispositive} follows from
the equivalence \ref{lem:prop:associatededge2coloringofBiscyclicallyreven}
$\Leftrightarrow$ \ref{lem:prop:BisChiorealizable} in
Lemma \ref{lem:graphtheoreticalcharacterizationofchiorealizability}.
As to \ref{explicitformulaforchiomeasureinthecaseofrealizability}, by the just
proved item \ref{characterizationofwhenchiomeasureispositive} we have
$\textup{\textsf{P}}_{\mathrm{chio}}[\mathcal{E}_B^J]>0$ if and only if $(\mathrm{X}_B,\sigma_B)$ is balanced,
and by equivalence \ref{lem:prop:associatededge2coloringofBiscyclicallyreven}
$\Leftrightarrow$ \ref{lem:prop:cardinalityofinverseimage}
in Lemma \ref{lem:graphtheoreticalcharacterizationofchiorealizability}
this is equivalent to
$\bigl \lvert (\tfrac12\mathrm{C}_{(s,t)})^{-1}(\mathcal{E}_B^J)
\bigr \rvert = 2^{\lvert \breve{J} \rvert - \mathrm{dom}(B) - f_0( \mathrm{X}_B ) + \beta_0(\mathrm{X}_B)}$. Dividing
by $2^{\lvert \breve{J} \rvert}$ in accordance with Definition \ref{def:measurePchio}
gives the first equality claimed
in \ref{explicitformulaforchiomeasureinthecaseofrealizability}. As to the
second equality, this is a reformulation not necessary for the equivalence and
is true by the known formula (e.g. \cite[Theorem 14.1.1]{MR1829620}) for the
dimension of the coboundary space of a graph, together with the obvious formula for
the number of elements of a finite-dimensional vector space over a finite field.
As to \ref{relationbeweenchiomeasureandlazycoinflipmeasuregovernedbyfirstbettinumber}, this follows by a simple calculation from \ref{explicitformulaforchiomeasureinthecaseofrealizability}, Lemma \ref{lem:valueoflazycoinflipdistribution}
and $\mathrm{supp}(X_B) = f_1(X_B)$. The second equality in \ref{relationbeweenchiomeasureandlazycoinflipmeasuregovernedbyfirstbettinumber} is true by definition
of $\beta_1(\cdot)$ (and therefore again a reformulation not necessary for the
equivalence). The proof of Theorem \ref{thm:graphtheoreticalcharacterizationofthechiomeasure} is now complete.
\end{proof}
We will now derive several consequences of Theorem \ref{thm:graphtheoreticalcharacterizationofthechiomeasure}. Let us start with:
\begin{corollary}\label{cor:quickconsequencesofthecharacterizations}
Let $(s,t)\in\mathbb{Z}_{\geq 2}^2$, $B\in \{0,\pm\}^{[s-1]\times [t-1]}$,
$\emptyset \subseteq I_1 \subseteq J_1 \subseteq [s-1]\times [t-1]$,
$\emptyset \subseteq I_2 \subseteq J_2 \subseteq [s-1]\times [t-1]$ with
$\lvert I_1 \rvert = \lvert I_2 \rvert$, $B_1\in \{0,\pm\}^{I_1}$ and
$B_2\in\{0,\pm\}^{I_2}$ be arbitrary. Then
\begin{enumerate}[label={\rm(\arabic{*})}]
\item\label{cor:graphfailuresetaspreimage}
$\mathcal{F}^{\mathrm{M}}(k,n) = (\beta_1\circ \mathrm{X}^{k,n,n})^{-1}(\mathbb{Z}_{\geq 1}) =
(\beta_1^{\mathrm{ul}}\circ \leftsubscript{\mathrm{ul}}{\upX}^{k,n,n})^{-1}(\mathbb{Z}_{\geq 1})$ \quad ,
\item\label{cor:chiomeasureofasinglematrix}
$B\in \im (\tfrac12 \mathrm{C}_{(s,t)}\colon
\{\pm\}^{[s]\times [t]} \rightarrow \{0,\pm\}^{[s-1]\times [t-1]})$ if and only if
$\textup{\textsf{P}}_{\mathrm{chio}}[B] = \frac{2\cdot 2^{\beta_0(\mathrm{X}_B)}}{2^{s\cdot t}}$ \quad ,
\item\label{cor:equalityofChiomeasuresofeventswithcoalescedassociatedgraphs}
$\textup{\textsf{P}}_{\mathrm{chio}}[\mathcal{E}_{B_1}^{J_1}] = \textup{\textsf{P}}_{\mathrm{chio}}[\mathcal{E}_{B_2}^{J_2}]$ if
$\mathrm{X}_{B_1}$ is a one-point wedge product of two components of $\mathrm{X}_{B_2}$ \quad .
\end{enumerate}
\end{corollary}
\begin{proof}
As to \ref{cor:graphfailuresetaspreimage}, this is immediate from
\ref{relationbeweenchiomeasureandlazycoinflipmeasuregovernedbyfirstbettinumber} in
Theorem \ref{thm:graphtheoreticalcharacterizationofthechiomeasure}.
As to \ref{cor:chiomeasureofasinglematrix}, this follows by setting
$I:=J:=[s-1]\times [t-1]$ and combining the equivalence
\ref{lem:prop:associatededge2coloringofBiscyclicallyreven} $\Leftrightarrow$
\ref{lem:prop:BisChiorealizable} in Lemma
\ref{lem:graphtheoreticalcharacterizationofchiorealizability} with
\ref{characterizationofwhenchiomeasureispositive} $\Leftrightarrow$
\ref{explicitformulaforchiomeasureinthecaseofrealizability} in Theorem
\ref{thm:graphtheoreticalcharacterizationofthechiomeasure}.
As to \ref{cor:equalityofChiomeasuresofeventswithcoalescedassociatedgraphs}, let us
first note that Lemma \ref{lem:onepointwedgepreservesbalancedness} implies that
either $(\mathrm{X}_{B_1},\sigma_{B_1})$ and $(\mathrm{X}_{B_2},\sigma_{B_2})$ are both not balanced,
or both are. If both are not balanced, then by item
\ref{characterizationofwhenchiomeasureispositive} in Theorem
\ref{thm:graphtheoreticalcharacterizationofthechiomeasure}, the claim is
true in the form of $0=0$. If both are, then by item
\ref{explicitformulaforchiomeasureinthecaseofrealizability} in Theorem
\ref{thm:graphtheoreticalcharacterizationofthechiomeasure}, and using
$\lvert I_1 \rvert = \lvert I_2 \rvert$, the equation
$\textup{\textsf{P}}_{\mathrm{chio}}[\mathcal{E}_{B_1}^{J_1}] = \textup{\textsf{P}}_{\mathrm{chio}}[\mathcal{E}_{B_2}^{J_2}]$ is
equivalent to $f_0(\mathrm{X}_{B_1})-\beta_0(\mathrm{X}_{B_1}) =
f_0(\mathrm{X}_{B_2}) - \beta_0(\mathrm{X}_{B_2})$. Since the one-point wedge product of two
graphs keeps $f_0(\cdot) - \beta_0(\cdot)$ invariant, the equation is true also in
this case and the proof is complete.
\end{proof}
\begin{corollary}[the lazy coin flip measure is an averaged Chio measure]\label{cor:lazycoinflipmeasureisanaveragedchiomeasure}
$\textup{\textsf{P}}_{\mathrm{lcf}}\bigl[B\bigr] = \overline{\textup{\textsf{P}}}_{\mathrm{chio}}\bigl[B\bigr]$ for every
$\emptyset\subseteq I \subseteq [n-1]^2$ and every $B\in\{0,\pm\}^I$.
\end{corollary}
\begin{proof}
It follows from Definition \ref{def:XBandecXB} that $\mathrm{supp}(B) = f_1(\mathrm{X}_B)$
and that $\{ \tilde{B} \in \{0,\pm\}^I\colon \mathrm{Supp}(\tilde{B}) = \mathrm{Supp}(B) \} =
\{ \tilde{B} \in \{0,\pm\}^I\colon \mathrm{X}_{\tilde{B}} = \mathrm{X}_B \}$. Moreover,
by \ref{characterizationofwhenchiomeasureispositive} in
Theorem \ref{thm:graphtheoreticalcharacterizationofthechiomeasure}, every
summand with the property that $(\mathrm{X}_{\tilde{B}},\sigma_{\tilde{B}})$ is not balanced
vanishes. Thus, for every $B\in\{0,\pm\}^I$,
\begin{align}
2^{f_1(\mathrm{X}_{B})}\cdot \overline{\textup{\textsf{P}}}_{\mathrm{chio}}\bigl[B \bigr] & =
\sum_{\substack{\mathrm{all}\; \tilde{B}\in \{0,\pm\}^I\;
\mathrm{with} \\ \mathrm{X}_{\tilde{B}} = \mathrm{X}_{B}\; \mathrm{and}\;
(\mathrm{X}_{\tilde{B}},\sigma_{\tilde{B}})\; \mathrm{balanced} } } \textup{\textsf{P}}_{\mathrm{chio}}[\tilde{B}]
\notag \\
& \By{\ref{relationbeweenchiomeasureandlazycoinflipmeasuregovernedbyfirstbettinumber}}{=} 2^{\beta_1(\mathrm{X}_{\tilde{B}})}\cdot \textup{\textsf{P}}_{\mathrm{lcf}}[B] \cdot
\bigl\lvert \bigl\{ \tilde{B}\in \{0,\pm\}^I\colon
\text{$\mathrm{X}_{\tilde{B}} = \mathrm{X}_{B}$ and $(\mathrm{X}_{\tilde{B}},\sigma_{\tilde{B}})$ balanced}
\bigr \} \bigr \rvert\notag \\
& \By{\ref{item:numberofbalancedsignfunctions}}{=}
2^{\beta_1(\mathrm{X}_{\tilde{B}})}\cdot \textup{\textsf{P}}_{\mathrm{lcf}}[B] \cdot
2^{f_0(\mathrm{X}_{B}) - \beta_0(\mathrm{X}_{B})} = 2^{f_1(\mathrm{X}_{B})}\cdot \textup{\textsf{P}}_{\mathrm{lcf}}[B]\quad . \qedhere
\end{align}
\end{proof}
\begin{corollary}[$\textup{\textsf{P}}_{\mathrm{chio}}^{\lvert\cdot\rvert, I}$ is just the uniform
distribution on $\{0,1\}^I$]\label{cor:signforgettingchiomeasure}
For every $(s,t)\in\mathbb{Z}_{\geq 2}^2$ and every
$\emptyset\subseteq I \subseteq [s-1]\times [t-1]$ let $\textup{\textsf{P}}_{0,1}^I$ denote
the uniform distribution on $\{0,1\}^I$.
Then $\textup{\textsf{P}}_{\mathrm{chio}}^{\lvert\cdot\rvert,I} = \textup{\textsf{P}}_{0,1}^I$.
\end{corollary}
\begin{proof}
This is true since for
$(s,t)\in \mathbb{Z}_{\geq 2}^2$, $\emptyset\subseteq I \subseteq [s-1]\times [t-1]$
and $B\in\{0,1\}^I$ we have
\begin{align}
2^{\lvert\breve{I}\rvert} \cdot \textup{\textsf{P}}_{\mathrm{chio}}^{\lvert\cdot\rvert, I} [B] & =
\bigl\lvert \{ A\in\{\pm\}^{\breve{I}}\colon
\lvert \tfrac12\cdot\mathrm{C}_{(s,t)}(A)\rvert = B \} \bigr\rvert \notag \\
& = \bigl\lvert \{ A\in\{\pm\}^{\breve{I}}\colon
\mathrm{Supp}(\tfrac12 \mathrm{C}_{(s,t)}(A)) = \mathrm{Supp}(B) \} \bigr\rvert \notag \\
\parbox{0.2\linewidth}{\tiny
(using \ref{lem:prop:associatededge2coloringofBiscyclicallyreven}
$\Leftrightarrow$ \ref{lem:prop:BisChiorealizable} in \\
Lemma \ref{lem:graphtheoreticalcharacterizationofchiorealizability})}
& = \sum_{\tilde{B}\in\{0,\pm\}^{I}\colon\mathrm{Supp}(\tilde{B}) = \mathrm{Supp}(B),\
\text{$(\mathrm{X}_{\tilde{B}},\sigma_{\tilde{B}})$ balanced}}
\lvert (\tfrac12\mathrm{C}_{(s,t)}^{\breve{I}})^{-1}
(\mathcal{E}_{\tilde{B}}^{I}) \rvert \notag \\
\parbox{0.2\linewidth}{\tiny (by \ref{lem:prop:cardinalityofinverseimage} in
Lemma \ref{lem:graphtheoreticalcharacterizationofchiorealizability})}
& = \lvert \{ \tilde{B}\in\{0,\pm\}^{I}\colon\text{\scriptsize
$\mathrm{Supp}(\tilde{B}) = \mathrm{Supp}(B)$,\ $\tilde{B}$ balanced} \} \rvert \cdot
2^{\lvert\breve{I}\rvert - \mathrm{dom}(B) - f_0(\mathrm{X}_{\tilde{B}}) + \beta_0(\mathrm{X}_{\tilde{B}})} \notag \\
\parbox{0.2\linewidth}{\tiny (by \ref{item:numberofbalancedsignfunctions}
in Lemma \ref{lem:equivalenceofexistenceofbconstantrpropervertex2coloringandcyclicallyrevenness} )}
& = 2^{f_0(\mathrm{X}_{\tilde{B}}) - \beta_0(\mathrm{X}_{\tilde{B}})} \cdot
2^{\lvert\breve{I}\rvert - \mathrm{dom}(B) - f_0(\mathrm{X}_{\tilde{B}}) + \beta_0(\mathrm{X}_{\tilde{B}})}
= 2^{\lvert\breve{I}\rvert - \mathrm{dom}(B)} \notag \qedhere
\end{align}
\end{proof}
Let us state the special case $s:=t:=n$ and $I:=[n-1]^2$ in
graph-theoretical language:
\begin{corollary}\label{cor:whatrandombipartitegraphisX12CnnAforrandomA}
For random $A\in\{\pm\}^{[n]^2}$, the graph $\mathrm{X}_{\frac12\mathrm{C}_{(n,n)}(A)}$ is a
random bipartite graph with $n-1$ vertices in each class and each edge
chosen i.i.d. with probability $\tfrac12$. \hfill $\Box$
\end{corollary}
Theorem \ref{thm:graphtheoreticalcharacterizationofthechiomeasure} also
teaches us how fast $\textup{\textsf{P}}_{\mathrm{chio}}$ can be computed. In both
Corollary \ref{cor:upperboundontimecomplexityofcomputingPchio} and
\ref{cor:complexityofdecidingwhetherchiomeasureequalslazycoinflipmeasure} the
asymptotic statements are referring to $n\rightarrow\infty$ and to sequences
$I = I(n)$ of index sets with the property that
$\lvert I(n) \rvert \rightarrow \infty$ (and therefore also
$\lvert \mathrm{p}_1(I(n)) \rvert \cdot \lvert \mathrm{p}_2(I(n)) \rvert \rightarrow\infty$)
as $n\rightarrow\infty$.
\begin{corollary}[complexity of computing $\textup{\textsf{P}}_{\mathrm{chio}}$]
\label{cor:upperboundontimecomplexityofcomputingPchio}
For every $\emptyset \subseteq I \subseteq J \subseteq [n-1]^2$ and
every $B\in \{0,\pm\}^I$, the value of $\textup{\textsf{P}}_{\mathrm{chio}}[\mathcal{E}_B^J] \in \mathbb{Q}$ can be
computed exactly in time
$O(\lvert \mathrm{p}_1(I) \rvert + \lvert \mathrm{p}_2(I) \rvert + \lvert I\rvert)
\subseteq O(\lvert \mathrm{p}_1(I) \rvert \cdot \lvert \mathrm{p}_2(I) \rvert)
\subseteq O(n^2)$. However, there does not exist a fixed algorithm
computing $\textup{\textsf{P}}_{\mathrm{chio}}[\mathcal{E}_B^J] \in \mathbb{Q}$ exactly on arbitrary instances
$B\in \{0,\pm\}^{[n-1]^2}$ and $\emptyset \subseteq I \subseteq [n-1]^2$ and taking
time $o(\lvert \mathrm{p}_1(I) \rvert \cdot \lvert \mathrm{p}_2(I) \rvert)$.
\end{corollary}
\begin{proof}
By items \ref{characterizationofwhenchiomeasureispositive}
and \ref{explicitformulaforchiomeasureinthecaseofrealizability} in
Theorem \ref{thm:graphtheoreticalcharacterizationofthechiomeasure}, to compute
$\textup{\textsf{P}}_{\mathrm{chio}}[\mathcal{E}_{B}^J]$ it suffices to first decide whether $\sigma_{B}$
(which in view of Definition \ref{def:XBandecXB} evidently can be read in time
$O(\lvert \mathrm{p}_1(I) \rvert \cdot \lvert \mathrm{p}_2(I) \rvert) \subseteq O(n^2)$) is
balanced, and, if so, to compute $f_0(\mathrm{X}_{B})$ and $\beta_0(\mathrm{X}_{B})$.
By Corollary \ref{cor:decidingwhetherecXiscyclicallyreven}, and since the
depth-first seach mentioned there also computes the numbers $f_0( \mathrm{X}_{B} )$
and $\beta_0(\mathrm{X}_{B})$, both tasks can be accomplished by one depth-first
search in time $O(f_0( \mathrm{X}_{B} ) + f_1( \mathrm{X}_{B} )) \subseteq
O(\lvert \mathrm{p}_1(I) \rvert + \lvert \mathrm{p}_2(I) \rvert + \lvert I\rvert) \subseteq
O(n^2)$. If $(\mathrm{X}_B,\sigma_B)$ is found to be not balanced, then
$\textup{\textsf{P}}_{\mathrm{chio}}[\mathcal{E}_B^J]=0$. Otherwise, the answer is
$(\frac12)^{\lvert I\rvert + f_0( \mathrm{X}_{B} ) - \beta_0(\mathrm{X}_{B})}$.
Since the bitlength of this dyadic fraction is
$\lvert I\rvert + f_0( \mathrm{X}_{B} ) - \beta_0(\mathrm{X}_{B}) =
\lvert I\rvert + f_1( \mathrm{X}_{B} ) - \beta_1(\mathrm{X}_{B}) \leq
\lvert I\rvert + f_1( \mathrm{X}_{B} ) \leq 2\lvert I \rvert \in
O(\lvert \mathrm{p}_1(I)\rvert + \lvert\mathrm{p}_2(I) \rvert + \lvert I \rvert)$ it is
possible to write the output in the time claimed. This proves the first
statement in Corollary \ref{cor:upperboundontimecomplexityofcomputingPchio}.
As to the second statement, notice that any such fixed algorithm could
in particular compute $\textup{\textsf{P}}_{\mathrm{chio}}[\mathcal{E}_{B}^J] \in \mathbb{Q}$ exactly on those
instances $B\in \{0,\pm\}^{[n-1]^2}$ and $\emptyset \subseteq I \subseteq [n-1]^2$
for which $I$ is rectangular. But if $I$ is rectangular, then
$\lvert I\rvert + f_0( \mathrm{X}_{B} ) - \beta_0(\mathrm{X}_{B}) \geq \lvert I\rvert =
\lvert \mathrm{p}_1(I) \rvert \cdot \lvert \mathrm{p}_2(I) \rvert$. Therefore, for these inputs,
the bitlength of the dyadic fraction
$(\frac12)^{\lvert I\rvert + f_0( \mathrm{X}_{B} ) - \beta_0(\mathrm{X}_{B})}$ is at least
$\lvert \mathrm{p}_1(I) \rvert \cdot \lvert \mathrm{p}_2(I) \rvert$. Hence for such inputs the
very task of writing the output takes time
$\Omega(\lvert \mathrm{p}_1(I) \rvert \cdot \lvert \mathrm{p}_2(I) \rvert)$,
which precludes a running time of
$o(\lvert \mathrm{p}_1(I) \rvert \cdot \lvert \mathrm{p}_2(I) \rvert) \subseteq o(n^2)$.
The proof of Corollary \ref{cor:upperboundontimecomplexityofcomputingPchio} is
now complete.
\end{proof}
A priori one might suspect that the task of merely \emph{deciding} whether
$\textup{\textsf{P}}_{\mathrm{chio}}[B] = \textup{\textsf{P}}_{\mathrm{lcf}}[B]$ could be accomplished much faster than the
task of computing the value of $\textup{\textsf{P}}_{\mathrm{chio}}[B]$. Theorem \ref{thm:graphtheoreticalcharacterizationofthechiomeasure} can also tell us that this is not the case.
\begin{corollary}[{complexity of deciding whether $\textup{\textsf{P}}_{\mathrm{chio}}$ and
$\textup{\textsf{P}}_{\mathrm{lcf}}$ agree}]
\label{cor:complexityofdecidingwhetherchiomeasureequalslazycoinflipmeasure}
For every $\emptyset \subseteq I \subseteq J \subseteq [n-1]^2$
and every $B\in \{0,\pm\}^I$, the answer to the decision problem of whether
$\textup{\textsf{P}}_{\mathrm{chio}}[\mathcal{E}_{B}^J] = \textup{\textsf{P}}_{\mathrm{lcf}}[\mathcal{E}_{B}^J]$ can be computed in
time $O(\lvert \mathrm{p}_1(I) \rvert \cdot \lvert \mathrm{p}_2(I) \rvert) \subseteq O(n^2)$.
However, there does not exist a fixed algorithm (having only entry-wise
access to $B$ and no further a priori information) which decides that question on
arbitrary instances $B\in \{0,\pm\}^I$ with
$\emptyset \subseteq I \subseteq [n-1]^2$ in time
$o(\lvert \mathrm{p}_1(I)\rvert \cdot \lvert \mathrm{p}_2(I)\rvert)$.
\end{corollary}
\begin{proof}
Given $\emptyset \subseteq I \subseteq [n-1]^2$ and $B \in\{0,\pm\}^I$, it follows
from item
\ref{relationbeweenchiomeasureandlazycoinflipmeasuregovernedbyfirstbettinumber} in
Theorem \ref{thm:graphtheoreticalcharacterizationofthechiomeasure} that the question
of whether $\textup{\textsf{P}}_{\mathrm{chio}}[\mathcal{E}_{B}^J] = \textup{\textsf{P}}_{\mathrm{lcf}}[\mathcal{E}_{B}^J]$ is
equivalent to asking whether $\mathrm{X}_{B}$ is a forest. The graph $\mathrm{X}_{B}$ can
obviously be computed from $B$ in time
$O(\lvert \mathrm{p}_1(I) \rvert \cdot \lvert \mathrm{p}_2(I) \rvert) \subseteq O(n^2)$, and
deciding whether $\mathrm{X}_{B}$ is a forest, i.e. whether $\mathrm{X}_{B}$ contains a circuit,
can be done by a depth-first search in time $O(f_0(\mathrm{X}_{B}) + f_1( \mathrm{X}_{B} ))
\subseteq O(\lvert \mathrm{p}_1(I) \rvert + \lvert \mathrm{p}_2(I) \rvert +
\lvert I\rvert) \subseteq O(\lvert \mathrm{p}_1(I) \rvert \cdot \lvert \mathrm{p}_2(I) \rvert )$, so the first claim in Corollary \ref{cor:complexityofdecidingwhetherchiomeasureequalslazycoinflipmeasure} is proved.
As to the additional claim, suppose there were a fixed algorithm $A$ with the stated
properties. Let $\mathcal{I}$ be the set of all rectangular
$\emptyset \subseteq I \subseteq [n-1]^2$. By
assumption, the algorithm $A$ is in particular capable of deciding whether
$\textup{\textsf{P}}_{\mathrm{chio}}[\mathcal{E}_{B}^J] = \textup{\textsf{P}}_{\mathrm{lcf}}[\mathcal{E}_{B}^J]$ for each input
$I\in \mathcal{I}$ and for each of them taking time
$o(\lvert \mathrm{p}_1(I)\rvert\cdot\lvert\mathrm{p}_2(I)\rvert)$. However, every bipartite
graph with bipartition sizes of $\lvert \mathrm{p}_1(I) \rvert$ and
$\lvert \mathrm{p}_2(I) \rvert$ can be realized as a $\mathrm{X}_{B}$ with $I\in\mathcal{I}$.
By item
\ref{relationbeweenchiomeasureandlazycoinflipmeasuregovernedbyfirstbettinumber} in
Theorem \ref{thm:graphtheoreticalcharacterizationofthechiomeasure} the property
$\textup{\textsf{P}}_{\mathrm{chio}}[\mathcal{E}_{B}^J] = \textup{\textsf{P}}_{\mathrm{lcf}}[\mathcal{E}_{B}^J]$ is equivalent to
$\mathrm{X}_{B}$ being a forest. Therefore $A$ decides set membership for the set of all
bipartite graphs which have the fixed (that is, fixed for every fixed value of $n$)
bipartition classes $ \mathrm{p}_1(I) $ and $ \mathrm{p}_2(I) $ and do not contain a circuit.
This set is a decreasing (i.e. closed w.r.t. deleting edges) graph property
consisting of bipartite graphs only. Since all graphs in the property have the
same bipartition classes $\mathrm{p}_1(I)$ and $\mathrm{p}_2(I)$ we may appeal to a
theorem of A. C.-C. Yao \cite[p. 518, Theorem 1]{MR941942} which says that every
such property is evasive.\footnote{Due to the fact that the bipartition classes
are the same for all the graphs in the property, it is not necessary to appeal
to the more general theorem of E. Triesch \cite[p. 266, Theorem 4]{MR1401898} in
which the assumption of \emph{fixed} bipartition classes is no longer made. An
earlier version of the present paper stated that one would need this more general
theorem. I was wrong regarding this particular point. Moreover I confused the
adjectives `balanced' and `fixed'. The theorem of Yao suffices.} Hence there exists
at least one $I\in \mathcal{I}$ with the property that $A$ examines every entry of
$B$. This takes time $\Omega(\lvert I\rvert) = \Omega(\lvert \mathrm{p}_1(I)\rvert \cdot
\lvert \mathrm{p}_2(I)\rvert)$, the equality being true because of
$\lvert I\rvert = \lvert \mathrm{p}_1(I) \rvert \cdot \lvert \mathrm{p}_2(I) \rvert$.
This is a contradiction to the assumption about the running time of $A$.
The proof of Corollary \ref{cor:complexityofdecidingwhetherchiomeasureequalslazycoinflipmeasure} is now complete.
\end{proof}
We now take a more quantitative look at the relationship between $\textup{\textsf{P}}_{\mathrm{chio}}$ and
$\textup{\textsf{P}}_{\mathrm{lcf}}$. We start with an enumeration of bipartite nonforests. The fact that
we stop the enumeration at the $f$-vector $(f_0,f_1) = (8,6)$, even though there
are bipartite nonforests with $(f_0,f_1) = (8,7)$, is explained by the application
we have in mind; we will only be concerned with bipartite nonforests having up to
six edges.
\begin{lemma}[bipartite nonforests ordered by their $f$-vectors]\label{lem:bipartitenonforestsorderedbytheirfvectors}
The isomorphism types of bipartite nonforests, ordered lexicographically by
their $f$-vectors up to $(f_0,f_1) = (8,6)$, are:
{\scriptsize
\begin{minipage}[b]{0.45\linewidth}
\begin{enumerate}[label={\rm(t\arabic{*})}]
\item\label{item:comparisonproofcaseLhassize4:4circuit}
$=$ $C^4$
\item\label{item:comparisonproofcaseLhassize6:oneisolatedvertex}
$=$ disjoint union of $C^4$ \\ and one isolated vertex
\item\label{item:comparisonproofcaseLhassize6:oneadditionaledgeintersectingC4}
$=$ $C^4$ intersecting one edge
\item\label{item:comparisonproofcaseLhassize6:XBLisomorphictoK23}
$=$ $K^{2,3}$
\item\label{item:comparisonproofcaseLhassize6:twoisolatedvertices}
$=$ disjoint union of $C^4$ \\ and two isolated vertices
\item\label{item:comparisonproofcaseLhassize6:oneadditionaledgeintersectingC4andoneisolatedvertex} $=$
$C^4$ intersecting one edge, \\ and one extra isolated vertex
\item\label{item:comparisonproofcaseLhassize6:C4withoneadditionaldisjointedge}
$=$ disjoint union of $C^4$ \\ and an isolated edge
\item\label{item:comparisonproofcaseLhassize6:C4intersectingtwoedgesinseparatenonadjacentvertices}
$=$ $C^4$ intersecting two disjoint edges, \\ the intersection set no edge of $C^4$
\item\label{item:comparisonproofcaseLhassize6:C4intersectingtwoedgesinseparateadjacentvertices}
$=$ $C^4$ intersecting two disjoint edges, \\ the intersection set an edge of $C^4$
\item\label{item:comparisonproofcaseLhassize6:C4intersectingatwopathinanendvertex}
$=$ $C^4$ intersecting a $2$-path in an endvertex
\end{enumerate}
\end{minipage}
\begin{minipage}[b]{0.48\linewidth}
\begin{enumerate}[label={\rm(t\arabic{*})},start=11]
\item\label{item:comparisonproofcaseLhassize6:C4intersectingatwopathinitsinnervertex}
$=$ $C^4$ intersecting a $2$-path in its inner vertex
\item\label{item:comparisonproofcaseLhassize6:C6}
$=$ $C^6$
\item\label{item:comparisonproofcaseLhassize6:threeisolatedvertices}
$=$ disjoint union of $C^4$ \\ and three isolated vertices
\item\label{item:comparisonproofcaseLhassize6:oneadditionaledgeintersectingC4andtwoisolatedvertices} $=$ $C^4$ intersecting one edge, \\ and two extra isolated vertices
\item\label{item:comparisonproofcaseLhassize6:oneadditionaldisjointedgeandoneisolatedvertex} $=$ disjoint union of $C^4$ and an edge, \\ and one extra isolated vertex
\item\label{item:comparisonproofcaseLhassize6:twoadditionaledgesonlyoneofthemdisjoint} $=$ $C^4$ intersecting one edge, \\ and one extra isolated edge
\item\label{item:comparisonproofcaseLhassize6:twoadditionalnondisjointedgesdisjointfromC4} $=$ disjoint union of $C^4$ and a $2$-path
\item\label{item:comparisonproofcaseLhassize6:fourisolatedvertices}
$=$ disjoint union of $C^4$ \\ and four isolated vertices
\item\label{item:comparisonproofcaseLhassize6:oneadditionaldisjointedgeandtwoisolatedvertices} $=$ disjoint union of $C^4$ \\ and an edge and two extra isolated vertices
\item\label{item:comparisonproofcaseLhassize6:twoadditionaldisjointedgesdisjointfromC4} $=$ disjoint union of $C^4$ and two disjoint edges
\end{enumerate}
\end{minipage}
}
\end{lemma}
\begin{proof}
Easy to check since the graphs are required to be bipartite and have $f_1\leq 6$.
\end{proof}
\begin{corollary}[isomorphism types for which equality of measures of entry
specification events fails]\label{cor:isomorphismtypesforwhichequalityofmeasuresofentryspecificationeventsfails}
\begin{minipage}[b]{0.4\linewidth}
\begin{enumerate}[label={\rm(Fa\arabic{*})},start=3]
\item\label{item:failuresetofisomorphismtypesk3}
$\mathcal{F}^{\mathrm{G}}(k,n) = \emptyset$ for $0\leq k \leq 3$,
\item\label{item:failuresetofisomorphismtypesk4}
$\mathcal{F}^{\mathrm{G}}(4,n)=\{\ref{item:comparisonproofcaseLhassize4:4circuit}\}$,
\end{enumerate}
\end{minipage}
\begin{minipage}[b]{0.65\linewidth}
\begin{enumerate}[label={\rm(Fa\arabic{*})},start=5]
\item\label{item:failuresetofisomorphismtypesk5}
$\mathcal{F}^{\mathrm{G}}(5,n) = \{
\ref{item:comparisonproofcaseLhassize6:oneisolatedvertex},
\ref{item:comparisonproofcaseLhassize6:oneadditionaledgeintersectingC4},
\ref{item:comparisonproofcaseLhassize6:twoisolatedvertices},
\ref{item:comparisonproofcaseLhassize6:C4withoneadditionaldisjointedge}
\}$,
\item\label{item:failuresetofisomorphismtypesk6}
$\mathcal{F}^{\mathrm{G}}(6,n) = \mathcal{F}^{\mathrm{G}}(5,n)
\sqcup\{ \ref{item:comparisonproofcaseLhassize6:XBLisomorphictoK23}\}
\sqcup \{
\ref{item:comparisonproofcaseLhassize6:oneadditionaledgeintersectingC4andoneisolatedvertex},\dotsc, \ref{item:comparisonproofcaseLhassize6:twoadditionaldisjointedgesdisjointfromC4}\}$.
\end{enumerate}
\end{minipage}
\end{corollary}
\begin{proof}
By \ref{relationbeweenchiomeasureandlazycoinflipmeasuregovernedbyfirstbettinumber}
in Theorem \ref{thm:graphtheoreticalcharacterizationofthechiomeasure} we have
$\textup{\textsf{P}}_{\mathrm{chio}}[\mathcal{E}_B^{[n-1]^2}] \neq \textup{\textsf{P}}_{\mathrm{lcf}}[\mathcal{E}_B^{[n-1]^2}]$ if and
only if $\beta_1(\mathrm{X}_{B}) > 0$. Moreover, directly from
Definition \ref{def:XBandecXB} we have the bound
$f_1(\mathrm{X}_{B}) \leq \lvert I \rvert$. Therefore, for every $k$, we can get a set of
candidates for membership in $\mathcal{F}^{\mathrm{G}}(k,n)$ by collecting all
isomorphism types in Corollary \ref{cor:isomorphismtypesforwhichequalityofmeasuresofentryspecificationeventsfails} having $f_1\leq k$. We then have to decide for each of
these types whether it is possible to realize it as a $\mathrm{X}_{B}$ with
$B\in\{0,\pm\}^I$ and $I\in\binom{[n-1]^2}{k}$.
As to \ref{item:failuresetofisomorphismtypesk3}, this is true since there do not
exist bipartite nonforests with three edges or less.
As to \ref{item:failuresetofisomorphismtypesk4}, i.e. $k=4$, note that the only
isomorphism types in Corollary \ref{cor:isomorphismtypesforwhichequalityofmeasuresofentryspecificationeventsfails} with $f_1 \leq 4$ are \ref{item:comparisonproofcaseLhassize4:4circuit} and \ref{item:comparisonproofcaseLhassize6:oneisolatedvertex}. Because
of $\beta_1(\mathrm{X}_B)\geq 1$ for every $B\in \mathcal{F}^{\mathrm{M}}(4,n)$ the set $I$
must be a matrix-$4$-circuit. This implies $f_0(\mathrm{X}_{B}) = 4$.
Since $f_0\ref{item:comparisonproofcaseLhassize6:oneisolatedvertex} = 5$, it
follows that $\ref{item:comparisonproofcaseLhassize6:oneisolatedvertex}\notin\mathcal{F}^{\mathrm{G}}(4,n)$. Since type \ref{item:comparisonproofcaseLhassize4:4circuit}
obviously can be realized, \ref{item:failuresetofisomorphismtypesk4} is true.
As to \ref{item:failuresetofisomorphismtypesk5}, i.e. $k=5$, note that the only
isomorphism types with $f_1\leq 5$ in Corollary \ref{cor:isomorphismtypesforwhichequalityofmeasuresofentryspecificationeventsfails} are \ref{item:comparisonproofcaseLhassize4:4circuit}, \ref{item:comparisonproofcaseLhassize6:oneisolatedvertex}, \ref{item:comparisonproofcaseLhassize6:oneadditionaledgeintersectingC4}, \ref{item:comparisonproofcaseLhassize6:twoisolatedvertices}, \ref{item:comparisonproofcaseLhassize6:oneadditionaledgeintersectingC4andoneisolatedvertex} and \ref{item:comparisonproofcaseLhassize6:C4withoneadditionaldisjointedge}.
Since $C^4$ is a subgraph of each of these types, it is necessary that there be a
matrix-$4$-circuit $S\subseteq I$. Since the sole non-matrix-circuit entry
must create at least one addition vertex of $\mathrm{X}_B$,
type \ref{item:comparisonproofcaseLhassize4:4circuit} cannot occur. The type
\ref{item:comparisonproofcaseLhassize6:oneadditionaledgeintersectingC4andoneisolatedvertex} cannot occur either since there is only one position
$u\in I\setminus S\in \binom{[n-1]^2}{1}$ left for us to choose freely
and by the choice of $u$ and $B[u]$ we can either create an isolated vertex
in $\mathrm{X}_{B}$ or an edge intersecting the $C^4\hookrightarrow \mathrm{X}_{B}$ but not both.
The remaining types \ref{item:comparisonproofcaseLhassize6:oneisolatedvertex}, \ref{item:comparisonproofcaseLhassize6:oneadditionaledgeintersectingC4}, \ref{item:comparisonproofcaseLhassize6:twoisolatedvertices} and
\ref{item:comparisonproofcaseLhassize6:C4withoneadditionaldisjointedge} can be
indeed be realized, as is proved by the following examples. For all the examples let
$S:=\{(1,1),(1,2),(2,1),(2,2)\}$, $B\mid_{S}:=\{-\}^S$ and $\{u\}:=I\setminus S$.
For \ref{item:comparisonproofcaseLhassize6:oneisolatedvertex} take e.g.
$n:=4$, $u:=(2,3)$ and $B[u] := 0$.
For \ref{item:comparisonproofcaseLhassize6:oneadditionaledgeintersectingC4} take
e.g. $n:=4$, $u:=(2,3)$ and $B[u]:=-$.
For \ref{item:comparisonproofcaseLhassize6:twoisolatedvertices} take
e.g. $n:=4$, $u:=(3,3)$ and $B[u] := 0$.
For \ref{item:comparisonproofcaseLhassize6:C4withoneadditionaldisjointedge} take
e.g. $n:=4$, $u:=(3,3)$ and $B[u]:=-$.
This proves \ref{item:failuresetofisomorphismtypesk5}.
As to \ref{item:failuresetofisomorphismtypesk6}, i.e. $k=6$, as far as only the
necessary condition $f_1(\mathrm{X}_{B}) \leq \lvert I \rvert = k$ is concerned, all
types in Lemma \ref{lem:bipartitenonforestsorderedbytheirfvectors} are candidates.
Type \ref{item:comparisonproofcaseLhassize4:4circuit} cannot be realized
since $f_0\ref{item:comparisonproofcaseLhassize4:4circuit}=4$ but
$f_0(\mathrm{X}_{B}) \geq 5$ for every $I\in \binom{[n-1]^2}{6}$. All others can, as will
now be proved by giving one example for each of them. In all examples again
let $S:=\{(1,1),(1,2),(2,1),(2,2)\}$ and $B\mid_{S}:=\{-\}^S$.
Here, $\{u,v\}:=I\setminus S$.
For \ref{item:comparisonproofcaseLhassize6:oneisolatedvertex} take e.g. $n:=4$,
$u:=(1,3)$, $v:=(2,3)$ and $B[u]:=B[v]:=0$.
For \ref{item:comparisonproofcaseLhassize6:oneadditionaledgeintersectingC4}
take e.g. $n:=4$, $u:=(1,3)$, $v:=(2,3)$, $B[u]:=-$ and $B[v]:=0$.
For \ref{item:comparisonproofcaseLhassize6:XBLisomorphictoK23}
take e.g. $n:=4$, $u:=(1,3)$, $v:=(2,3)$ and $B[u]:=B[v]:=-$.
For \ref{item:comparisonproofcaseLhassize6:twoisolatedvertices}
take e.g. $n:=4$, $u:=(1,3)$, $v:=(3,3)$ and $B[u]:=B[v]:=0$.
For \ref{item:comparisonproofcaseLhassize6:oneadditionaledgeintersectingC4andoneisolatedvertex} take e.g. $n := 4$, $u:=(1,3)$, $v:=(3,3)$ and $B[u]:=-$ and $B[v]:=0$.
For \ref{item:comparisonproofcaseLhassize6:C4withoneadditionaldisjointedge}
take e.g. $n := 4$, $u:=(1,3)$, $v:=(3,3)$, $B[u]:=0$ and $B[v]:=-$.
For \ref{item:comparisonproofcaseLhassize6:C4intersectingtwoedgesinseparatenonadjacentvertices} take e.g. $n := 5$, $u:=(1,3)$, $v:=(2,4)$ and $B[u]:=B[v]:=-$.
For \ref{item:comparisonproofcaseLhassize6:C4intersectingtwoedgesinseparateadjacentvertices} take e.g. $n:=4$, $u:=(1,3)$, $v:=(3,2)$ and $B[u]:=B[v]:=-$.
For \ref{item:comparisonproofcaseLhassize6:C4intersectingatwopathinanendvertex}
take e.g. $n := 4$, $u:=(1,3)$, $v:=(3,3)$ and $B[u]:=B[v]:=-$.
For \ref{item:comparisonproofcaseLhassize6:C4intersectingatwopathinitsinnervertex}
take e.g. $n:=5$, $u:=(1,3)$, $v:=(1,4)$ and $B[u]:=B[v]:=-$.
For \ref{item:comparisonproofcaseLhassize6:C6} we have to make an exception to our
convention that $\{u,v\} = I\setminus S$ with $S$ defined as above, and have to
define the set $I$ in its entirety. We can take e.g.
$n:=4$, $I:=\{(1,1),(1,2),(2,2),(2,3),(3,1),(3,3)\}\in\binom{[n-1]^2}{k}$
and $B=\{-\}^I$.
For \ref{item:comparisonproofcaseLhassize6:threeisolatedvertices} take e.g. $n:=5$,
$u:=(3,3)$, $v:=(3,4)$ and $B[u]:=B[v]:=0$.
For \ref{item:comparisonproofcaseLhassize6:oneadditionaledgeintersectingC4andtwoisolatedvertices}, take e.g. $n:=5$, $u:=(1,3)$, $v:=(3,4)$, $B[u]:=-$ and $B[v]:=0$.
For \ref{item:comparisonproofcaseLhassize6:oneadditionaldisjointedgeandoneisolatedvertex}, take e.g. $n:=5$, $u:=(3,3)$, $v:=(3,4)$, $B[u]:=-$ and $B[v]:=0$.
For \ref{item:comparisonproofcaseLhassize6:twoadditionaledgesonlyoneofthemdisjoint},
take e.g. $n:=5$, $u:=(1,3)$, $v:=(3,4)$ and $B[u]:=B[v]:=-$.
For \ref{item:comparisonproofcaseLhassize6:twoadditionalnondisjointedgesdisjointfromC4}
take e.g. $n:=5$, $u:=(3,3)$, $v:=(3,4)$ and $B[u]:=B[v]:=-$.
For \ref{item:comparisonproofcaseLhassize6:fourisolatedvertices} take e.g. $n:=5$,
$u:=(3,3)$, $v:=(4,4)$ and $B[u]:=B[v]:=0$.
For \ref{item:comparisonproofcaseLhassize6:oneadditionaldisjointedgeandtwoisolatedvertices} take e.g. $n:=5$, $u:=(3,3)$, $v:=(4,4)$, $B[u]:= 0$ and $B[v]:=1$.
For \ref{item:comparisonproofcaseLhassize6:twoadditionaldisjointedgesdisjointfromC4}
take e.g. $n:=5$, $u:=(3,3)$, $v:=(4,4)$ and $B[u]:=B[v]:=-$.
This proves \ref{item:failuresetofisomorphismtypesk6}.
The proof of Corollary \ref{cor:isomorphismtypesforwhichequalityofmeasuresofentryspecificationeventsfails} is now complete.
\end{proof}
\begin{corollary}[ratios and absolute values of $\textup{\textsf{P}}_{\mathrm{Chio}}$ for up to six entry
specifications]
\label{cor:ratiosandvaluesofpchioandplcffortheisomorphismtypesforwhichequalityfails}
{\quad }
\begin{enumerate}[label={\rm(R\arabic{*})}]
\item\label{it:decompositionsofgraphfailuresetsingeneral}
$\mathcal{F}^{\mathrm{G}}(k,n) = \mathcal{F}_{\cdot 0}^{\mathrm{G}}(k,n) =
\bigsqcup_{\beta\in\mathbb{Z}_{\geq 1}}\mathcal{F}_{\cdot 2^\beta}^{\mathrm{G}}(k,n)$\quad ,
\item\label{it:decompositionsofmatrixfailuresetsingeneral}
$\mathcal{F}^{\mathrm{M}}(k,n) = \mathcal{F}_{\cdot 0}^{\mathrm{M}}(k,n) \sqcup
\bigsqcup_{\beta\in\mathbb{Z}_{\geq 1}}\mathcal{F}_{\cdot 2^\beta}^{\mathrm{M}}(k,n)$\quad ,
\item\label{it:decompositionsofgraphfailuresets}
$\mathcal{F}^{\mathrm{G}}(4,n) = \mathcal{F}_{\cdot 2}^{\mathrm{G}}(4,n)$,
$\mathcal{F}^{\mathrm{G}}(5,n) = \mathcal{F}_{\cdot 2}^{\mathrm{G}}(5,n)$ and
$\mathcal{F}^{\mathrm{G}}(6,n) =
\mathcal{F}_{\cdot 2}^{\mathrm{G}}(6,n) \sqcup \mathcal{F}_{\cdot 4}^{\mathrm{G}}(6,n)$\quad ,
\item\label{it:decompositionsofmatrixfailuresets}
$\mathcal{F}^{\mathrm{M}}(4,n) = \mathcal{F}_{\cdot 0}^{\mathrm{M}}(4,n) \sqcup
\mathcal{F}_{\cdot 2}^{\mathrm{M}}(4,n)$, $\mathcal{F}^{\mathrm{M}}(5,n) =
\mathcal{F}_{\cdot 0}^{\mathrm{M}}(5,n) \sqcup \mathcal{F}_{\cdot 2}^{\mathrm{M}}(5,n)$ \\
and $\mathcal{F}^{\mathrm{M}}(6,n) = \mathcal{F}_{\cdot 0}^{\mathrm{M}}(6,n) \sqcup
\mathcal{F}_{\cdot 2}^{\mathrm{M}}(6,n) \sqcup \mathcal{F}_{\cdot 4}^{\mathrm{M}}(6,n)$\quad .
\end{enumerate}
Moreover,
\begin{enumerate}[label={\rm(A\arabic{*})}, start=4]
\item\label{item:absolutevaluesofchiomeasureonfailureentryspecificationevent:4entriesspecified}
$\mathcal{F}^{\mathrm{G}}(4,n) = \mathcal{F}_{=(\frac12)^{7}}^{\mathrm{G}}(4,n)$, \;
with $\mathcal{F}_{=(\frac12)^{7}}^{\mathrm{G}}(4,n) =
\{\ref{item:comparisonproofcaseLhassize4:4circuit}\}$\quad ,
\item\label{item:absolutevaluesofchiomeasureonfailureentryspecificationevent:5entriesspecified}
$\mathcal{F}^{\mathrm{G}}(5,n) = \mathcal{F}_{=(\frac12)^8}^{\mathrm{G}}(5,n) \sqcup
\mathcal{F}_{=(\frac12)^9}^{\mathrm{G}}(5,n)$\quad , \\
with $\mathcal{F}_{=(\frac12)^{8}}^{\mathrm{G}}(5,n) = \{\ref{item:comparisonproofcaseLhassize6:oneisolatedvertex}, \ref{item:comparisonproofcaseLhassize6:twoisolatedvertices}\}$\quad and\quad $\mathcal{F}_{=(\frac12)^9}^{\mathrm{G}}(5,n) = \{\ref{item:comparisonproofcaseLhassize6:oneadditionaledgeintersectingC4}, \ref{item:comparisonproofcaseLhassize6:C4withoneadditionaldisjointedge}\}$\quad ,
\item\label{item:absolutevaluesofchiomeasureonfailureentryspecificationevent:6entriesspecified}
$\mathcal{F}^{\mathrm{G}}(6,n) =
\mathcal{F}_{=(\frac12)^9}^{\mathrm{G}}(6,n)\sqcup
\mathcal{F}_{=(\frac12)^{10}}^{\mathrm{G}}(6,n)\sqcup
\mathcal{F}_{=(\frac12)^{11}}^{\mathrm{G}}(6,n)$, where
\begin{enumerate}[label={\rm(\roman{*})}]
\item\label{item:absolutevaluesofchiomeasureonfailureentryspecificationevent:6entriesspecified:typeswithvalueoneovertwotothenine}
$\mathcal{F}_{=(\frac12)^9}^{\mathrm{G}}(6,n)$
$ = \{
\ref{item:comparisonproofcaseLhassize6:oneisolatedvertex},
\ref{item:comparisonproofcaseLhassize6:twoisolatedvertices},
\ref{item:comparisonproofcaseLhassize6:threeisolatedvertices},
\ref{item:comparisonproofcaseLhassize6:fourisolatedvertices} \}$ \quad ,
\item\label{item:absolutevaluesofchiomeasureonfailureentryspecificationevent:6entriesspecified:typeswithvalueoneovertwototheten}
$\mathcal{F}_{=(\frac12)^{10}}^{\mathrm{G}}(6,n)$
$=\{
\ref{item:comparisonproofcaseLhassize6:oneadditionaledgeintersectingC4},
\ref{item:comparisonproofcaseLhassize6:XBLisomorphictoK23},
\ref{item:comparisonproofcaseLhassize6:oneadditionaledgeintersectingC4andoneisolatedvertex},
\ref{item:comparisonproofcaseLhassize6:C4withoneadditionaldisjointedge},
\ref{item:comparisonproofcaseLhassize6:oneadditionaledgeintersectingC4andtwoisolatedvertices},
\ref{item:comparisonproofcaseLhassize6:oneadditionaldisjointedgeandoneisolatedvertex},
\ref{item:comparisonproofcaseLhassize6:oneadditionaldisjointedgeandtwoisolatedvertices} \}$ \quad ,
\item\label{item:absolutevaluesofchiomeasureonfailureentryspecificationevent:6entriesspecified:typeswithvalueoneovertwototheeleven}
$\mathcal{F}_{=(\frac12)^{11}}^{\mathrm{G}}(6,n)$
$ = \{ \ref{item:comparisonproofcaseLhassize6:C4intersectingtwoedgesinseparatenonadjacentvertices},
\ref{item:comparisonproofcaseLhassize6:C4intersectingtwoedgesinseparateadjacentvertices},
\ref{item:comparisonproofcaseLhassize6:C4intersectingatwopathinanendvertex},
\ref{item:comparisonproofcaseLhassize6:C4intersectingatwopathinitsinnervertex},
\ref{item:comparisonproofcaseLhassize6:C6},
\ref{item:comparisonproofcaseLhassize6:twoadditionaledgesonlyoneofthemdisjoint},
\ref{item:comparisonproofcaseLhassize6:twoadditionalnondisjointedgesdisjointfromC4},
\ref{item:comparisonproofcaseLhassize6:twoadditionaldisjointedgesdisjointfromC4} \}$
\quad .
\end{enumerate}
\end{enumerate}
\end{corollary}
\begin{proof}
As to \ref{it:decompositionsofgraphfailuresetsingeneral}, let us start with
$\mathcal{F}^{\mathrm{G}}(k,n) = \mathcal{F}_{\cdot 0}^{\mathrm{G}}(k,n)$. The inclusion
$\supseteq$ is true directly by \ref{def:graphversionsofthefailuresets} in
Definition \ref{def:parametrizedfailuresets}. Conversely, let
$\mathfrak{X}\in\mathcal{F}^{\mathrm{G}}(k,n)$. Then $\beta_1(\mathfrak{X})\geq 1$ by
Corollary \ref{cor:quickconsequencesofthecharacterizations}.\ref{cor:graphfailuresetaspreimage}, hence
$\lvert \{\pm\}^{\mathrm{E}(\mathfrak{X})}\setminus S_{\mathrm{bal}}(\mathfrak{X})\rvert =
2^{f_1(\mathfrak{X})} - 2^{f_0(\mathfrak{X}) - \beta_0(\mathfrak{X})} =
2^{f_1(\mathfrak{X})} - 2^{f_1(\mathfrak{X}) - \beta_1(\mathfrak{X})} > 0$, hence there exists
$B\in\mathcal{F}_{\cdot 0}^{\mathrm{M}}(k,n)$ with $\mathfrak{X} = \leftsubscript{\mathrm{ul}}{\upX}^{k,n,n}(B)$,
hence $\mathfrak{X}\in\leftsubscript{\mathrm{ul}}{\upX}^{k,n,n}(\mathcal{F}_{\cdot 0}^{\mathrm{M}}(k,n))
\By{Definition
\ref{def:parametrizedfailuresets}.\ref{def:graphversionsofthefailuresets}}{=}
\mathcal{F}_{\cdot 0}^{\mathrm{G}}(k,n)$, proving $\subseteq$.
As to the partition claimed in \ref{it:decompositionsofgraphfailuresetsingeneral},
both claims follow immediately from
\ref{relationbeweenchiomeasureandlazycoinflipmeasuregovernedbyfirstbettinumber}
in Theorem \ref{thm:graphtheoreticalcharacterizationofthechiomeasure} (and the
equality $\mathcal{F}^{\mathrm{G}}(k,n) = \mathcal{F}^{\mathrm{G}}_{\cdot 0}(k,n)$ is the
reason why $\mathcal{F}_{\cdot 0}^{\mathrm{G}}(k,n)$ is missing in the disjoint union
in \ref{it:decompositionsofgraphfailuresetsingeneral}).
As to \ref{it:decompositionsofgraphfailuresets} (respectively
\ref{it:decompositionsofmatrixfailuresets}), this follows by combining
\ref{it:decompositionsofgraphfailuresetsingeneral}
(respectively \ref{it:decompositionsofmatrixfailuresetsingeneral}) with \ref{item:failuresetofisomorphismtypesk4}--\ref{item:failuresetofisomorphismtypesk6} in
Corollary \ref{cor:isomorphismtypesforwhichequalityofmeasuresofentryspecificationeventsfails}. The claims \ref{item:absolutevaluesofchiomeasureonfailureentryspecificationevent:4entriesspecified}--\ref{item:absolutevaluesofchiomeasureonfailureentryspecificationevent:6entriesspecified} will be proved in reverse order.
As to \ref{item:absolutevaluesofchiomeasureonfailureentryspecificationevent:6entriesspecified}, this seems to require some calculations. However,
Corollary \ref{cor:quickconsequencesofthecharacterizations}.\ref{cor:equalityofChiomeasuresofeventswithcoalescedassociatedgraphs} can be used to reduce the amount of
work to be done: if $(a)$ and $(b)$ are isomorphism types of graphs, let us write
$(a) \twoheadrightarrow (b)$ if and only if $(b)$ can be obtained from $(a)$ by a
single one-point wedge of two connected components of $(a)$. Moreover, if $(a)$ is
any of the isomorphism types in $\mathcal{F}^{\mathrm{M}}(6,n)$, let us employ the
abbreviation $\mathcal{E}_B := \mathcal{E}_B^{[n-1]^2}$ and let us write
$\textup{\textsf{P}}_{\mathrm{chio}}[(a)]$ for the number $\textup{\textsf{P}}_{\mathrm{chio}}[\mathcal{E}_B^{[n-1]^2}]$
with $B$ an arbitrary $B\in\{0,\pm\}^I$, $I\in\binom{[n-1]^2}{6}$, $\mathrm{X}_B\in (a)$
and $(\mathrm{X}_B,\sigma_B)$ balanced.
By \ref{explicitformulaforchiomeasureinthecaseofrealizability} in Theorem
\ref{thm:graphtheoreticalcharacterizationofthechiomeasure} we know that
$\textup{\textsf{P}}_{\mathrm{chio}}[(a)]$ then does indeed only depend on $(a)$, not on the choice
of such a $B$.
Since evidently
\ref{item:comparisonproofcaseLhassize6:fourisolatedvertices}
$\twoheadrightarrow$
\ref{item:comparisonproofcaseLhassize6:threeisolatedvertices}
$\twoheadrightarrow$
\ref{item:comparisonproofcaseLhassize6:twoisolatedvertices}
$\twoheadrightarrow$
\ref{item:comparisonproofcaseLhassize6:oneisolatedvertex},
Corollary \ref{cor:quickconsequencesofthecharacterizations}.\ref{cor:equalityofChiomeasuresofeventswithcoalescedassociatedgraphs}
implies
$\textup{\textsf{P}}_{\mathrm{chio}}[\ref{item:comparisonproofcaseLhassize6:oneisolatedvertex}]$
$=$
$\textup{\textsf{P}}_{\mathrm{chio}}[\ref{item:comparisonproofcaseLhassize6:twoisolatedvertices}]$
$=$
$\textup{\textsf{P}}_{\mathrm{chio}}[\ref{item:comparisonproofcaseLhassize6:threeisolatedvertices}]$
$=$
$\textup{\textsf{P}}_{\mathrm{chio}}[\ref{item:comparisonproofcaseLhassize6:fourisolatedvertices}]$.
Since evidently
\ref{item:comparisonproofcaseLhassize6:oneadditionaldisjointedgeandtwoisolatedvertices} $\twoheadrightarrow$ \ref{item:comparisonproofcaseLhassize6:oneadditionaldisjointedgeandoneisolatedvertex} $\twoheadrightarrow$ \ref{item:comparisonproofcaseLhassize6:oneadditionaledgeintersectingC4andoneisolatedvertex} $\twoheadrightarrow$
\ref{item:comparisonproofcaseLhassize6:oneadditionaledgeintersectingC4}, Corollary
\ref{cor:quickconsequencesofthecharacterizations}.\ref{cor:equalityofChiomeasuresofeventswithcoalescedassociatedgraphs} implies
$\textup{\textsf{P}}_{\mathrm{chio}}[\ref{item:comparisonproofcaseLhassize6:oneadditionaldisjointedgeandtwoisolatedvertices}]$
$=$
$\textup{\textsf{P}}_{\mathrm{chio}}[\ref{item:comparisonproofcaseLhassize6:oneadditionaldisjointedgeandoneisolatedvertex}]$
$=$
$\textup{\textsf{P}}_{\mathrm{chio}}[\ref{item:comparisonproofcaseLhassize6:oneadditionaledgeintersectingC4andoneisolatedvertex}]$
$=$
$\textup{\textsf{P}}_{\mathrm{chio}}[\ref{item:comparisonproofcaseLhassize6:oneadditionaledgeintersectingC4}]$. Since also
\ref{item:comparisonproofcaseLhassize6:oneadditionaledgeintersectingC4andtwoisolatedvertices} $\twoheadrightarrow$ \ref{item:comparisonproofcaseLhassize6:oneadditionaledgeintersectingC4andoneisolatedvertex}, Corollary
\ref{cor:quickconsequencesofthecharacterizations}.\ref{cor:equalityofChiomeasuresofeventswithcoalescedassociatedgraphs} implies
$\textup{\textsf{P}}_{\mathrm{chio}}[\ref{item:comparisonproofcaseLhassize6:oneadditionaledgeintersectingC4andtwoisolatedvertices}]$
$=$
$\textup{\textsf{P}}_{\mathrm{chio}}[\ref{item:comparisonproofcaseLhassize6:oneadditionaledgeintersectingC4andoneisolatedvertex}]$. Since moreover \ref{item:comparisonproofcaseLhassize6:C4withoneadditionaldisjointedge} $\twoheadrightarrow$ \ref{item:comparisonproofcaseLhassize6:oneadditionaledgeintersectingC4}, Corollary
\ref{cor:quickconsequencesofthecharacterizations}.\ref{cor:equalityofChiomeasuresofeventswithcoalescedassociatedgraphs} implies
$\textup{\textsf{P}}_{\mathrm{chio}}[\ref{item:comparisonproofcaseLhassize6:C4withoneadditionaldisjointedge}]$ $=$ $\textup{\textsf{P}}_{\mathrm{chio}}[\ref{item:comparisonproofcaseLhassize6:oneadditionaledgeintersectingC4}]$.
These equations together imply
$\textup{\textsf{P}}_{\mathrm{chio}}[\ref{item:comparisonproofcaseLhassize6:oneadditionaledgeintersectingC4}]$
$=$
$\textup{\textsf{P}}_{\mathrm{chio}}[\ref{item:comparisonproofcaseLhassize6:oneadditionaledgeintersectingC4andoneisolatedvertex}]$
$=$
$\textup{\textsf{P}}_{\mathrm{chio}}[\ref{item:comparisonproofcaseLhassize6:C4withoneadditionaldisjointedge}]$
$=$
$\textup{\textsf{P}}_{\mathrm{chio}}[\ref{item:comparisonproofcaseLhassize6:oneadditionaledgeintersectingC4andtwoisolatedvertices}]$
$=$
$\textup{\textsf{P}}_{\mathrm{chio}}[\ref{item:comparisonproofcaseLhassize6:oneadditionaldisjointedgeandoneisolatedvertex}]$
$=$
$\textup{\textsf{P}}_{\mathrm{chio}}[\ref{item:comparisonproofcaseLhassize6:oneadditionaldisjointedgeandtwoisolatedvertices}]$.
Since evidently
\ref{item:comparisonproofcaseLhassize6:twoadditionaldisjointedgesdisjointfromC4}
$\twoheadrightarrow$ \ref{item:comparisonproofcaseLhassize6:twoadditionalnondisjointedgesdisjointfromC4} $\twoheadrightarrow$ \ref{item:comparisonproofcaseLhassize6:C4intersectingatwopathinanendvertex}, Corollary
\ref{cor:quickconsequencesofthecharacterizations}.\ref{cor:equalityofChiomeasuresofeventswithcoalescedassociatedgraphs} implies
$\textup{\textsf{P}}_{\mathrm{chio}}[\ref{item:comparisonproofcaseLhassize6:twoadditionaldisjointedgesdisjointfromC4}]$
$=$
$\textup{\textsf{P}}_{\mathrm{chio}}[\ref{item:comparisonproofcaseLhassize6:twoadditionalnondisjointedgesdisjointfromC4}]$
$=$
$\textup{\textsf{P}}_{\mathrm{chio}}[\ref{item:comparisonproofcaseLhassize6:C4intersectingatwopathinanendvertex}]$.
Since evidently
\ref{item:comparisonproofcaseLhassize6:twoadditionaldisjointedgesdisjointfromC4}
$\twoheadrightarrow$ \ref{item:comparisonproofcaseLhassize6:twoadditionaledgesonlyoneofthemdisjoint} $\twoheadrightarrow$ \ref{item:comparisonproofcaseLhassize6:C4intersectingatwopathinanendvertex}, Corollary
\ref{cor:quickconsequencesofthecharacterizations}.\ref{cor:equalityofChiomeasuresofeventswithcoalescedassociatedgraphs} implies
$\textup{\textsf{P}}_{\mathrm{chio}}[\ref{item:comparisonproofcaseLhassize6:twoadditionaldisjointedgesdisjointfromC4}]$
$=$
$\textup{\textsf{P}}_{\mathrm{chio}}[\ref{item:comparisonproofcaseLhassize6:twoadditionaledgesonlyoneofthemdisjoint}]$
$=$
$\textup{\textsf{P}}_{\mathrm{chio}}[\ref{item:comparisonproofcaseLhassize6:C4intersectingatwopathinanendvertex}]$.
Since evidently
\ref{item:comparisonproofcaseLhassize6:twoadditionalnondisjointedgesdisjointfromC4} $\twoheadrightarrow$ \ref{item:comparisonproofcaseLhassize6:C4intersectingatwopathinitsinnervertex}, Corollary
\ref{cor:quickconsequencesofthecharacterizations}.\ref{cor:equalityofChiomeasuresofeventswithcoalescedassociatedgraphs} implies
$\textup{\textsf{P}}_{\mathrm{chio}}[\ref{item:comparisonproofcaseLhassize6:twoadditionalnondisjointedgesdisjointfromC4}]$
$=$
$\textup{\textsf{P}}_{\mathrm{chio}}[\ref{item:comparisonproofcaseLhassize6:C4intersectingatwopathinitsinnervertex}]$.
Since evidently
\ref{item:comparisonproofcaseLhassize6:twoadditionaledgesonlyoneofthemdisjoint} $\twoheadrightarrow$ \ref{item:comparisonproofcaseLhassize6:C4intersectingtwoedgesinseparateadjacentvertices}, Corollary
\ref{cor:quickconsequencesofthecharacterizations}.\ref{cor:equalityofChiomeasuresofeventswithcoalescedassociatedgraphs} implies
$\textup{\textsf{P}}_{\mathrm{chio}}[\ref{item:comparisonproofcaseLhassize6:twoadditionaledgesonlyoneofthemdisjoint}]$
$=$
$\textup{\textsf{P}}_{\mathrm{chio}}[\ref{item:comparisonproofcaseLhassize6:C4intersectingtwoedgesinseparateadjacentvertices}]$.
Since evidently
\ref{item:comparisonproofcaseLhassize6:twoadditionaledgesonlyoneofthemdisjoint} $\twoheadrightarrow$ \ref{item:comparisonproofcaseLhassize6:C4intersectingtwoedgesinseparatenonadjacentvertices}, Corollary
\ref{cor:quickconsequencesofthecharacterizations}.\ref{cor:equalityofChiomeasuresofeventswithcoalescedassociatedgraphs} implies
$\textup{\textsf{P}}_{\mathrm{chio}}[\ref{item:comparisonproofcaseLhassize6:twoadditionaledgesonlyoneofthemdisjoint}]$
$=$
$\textup{\textsf{P}}_{\mathrm{chio}}[\ref{item:comparisonproofcaseLhassize6:C4intersectingtwoedgesinseparatenonadjacentvertices}]$. These equations together imply that
$\textup{\textsf{P}}_{\mathrm{chio}}[\ref{item:comparisonproofcaseLhassize6:C4intersectingtwoedgesinseparatenonadjacentvertices}]$
$=$
$\textup{\textsf{P}}_{\mathrm{chio}}[\ref{item:comparisonproofcaseLhassize6:C4intersectingtwoedgesinseparateadjacentvertices}]$
$=$
$\textup{\textsf{P}}_{\mathrm{chio}}[\ref{item:comparisonproofcaseLhassize6:C4intersectingatwopathinanendvertex}]$
$=$
$\textup{\textsf{P}}_{\mathrm{chio}}[\ref{item:comparisonproofcaseLhassize6:C4intersectingatwopathinitsinnervertex}]$
$=$
$\textup{\textsf{P}}_{\mathrm{chio}}[\ref{item:comparisonproofcaseLhassize6:twoadditionaledgesonlyoneofthemdisjoint}]$
$=$
$\textup{\textsf{P}}_{\mathrm{chio}}[\ref{item:comparisonproofcaseLhassize6:twoadditionalnondisjointedgesdisjointfromC4}]$
$=$
$\textup{\textsf{P}}_{\mathrm{chio}}[\ref{item:comparisonproofcaseLhassize6:twoadditionaldisjointedgesdisjointfromC4}]$.
This proves that it suffices (note that of the nineteen elements of
$\mathcal{F}^{\mathrm{M}}(6,n)$ exactly
$\ref{item:comparisonproofcaseLhassize6:XBLisomorphictoK23}$
and $\ref{item:comparisonproofcaseLhassize6:C6}$
have not been part of one of the equality chains) to calculate only
$\textup{\textsf{P}}_{\mathrm{chio}}[\ref{item:comparisonproofcaseLhassize6:oneisolatedvertex}]$,
$\textup{\textsf{P}}_{\mathrm{chio}}[\ref{item:comparisonproofcaseLhassize6:oneadditionaledgeintersectingC4}]$, $\textup{\textsf{P}}_{\mathrm{chio}}[\ref{item:comparisonproofcaseLhassize6:XBLisomorphictoK23}]$,
$\textup{\textsf{P}}_{\mathrm{chio}}[\ref{item:comparisonproofcaseLhassize6:C4intersectingtwoedgesinseparatenonadjacentvertices}]$ and $\textup{\textsf{P}}_{\mathrm{chio}}[\ref{item:comparisonproofcaseLhassize6:C6}]$. With the formulas in \ref{explicitformulaforchiomeasureinthecaseofrealizability}
of Theorem \ref{thm:graphtheoreticalcharacterizationofthechiomeasure} and
in Lemma \ref{lem:valueoflazycoinflipdistribution}, this can be done as
follows (keep in mind that, being within item \ref{item:absolutevaluesofchiomeasureonfailureentryspecificationevent:6entriesspecified}, $\mathrm{dom}(B) = \lvert I \rvert = 6$
in each calculation):
If $\mathfrak{X}=\ref{item:comparisonproofcaseLhassize6:oneisolatedvertex}$, then
$f_0( \mathfrak{X} ) = 5$, $\beta_0(\mathfrak{X}) = 2$, hence
$\textup{\textsf{P}}_{\mathrm{chio}}[\mathcal{E}_B]$ $=$ $(\frac12)^{\lvert I \rvert+5-2}$ $=$ $(\frac12)^9$. \\
If $\mathfrak{X}\in\{\ref{item:comparisonproofcaseLhassize6:oneadditionaledgeintersectingC4}, \ref{item:comparisonproofcaseLhassize6:XBLisomorphictoK23}\}$, then
$f_0( \mathfrak{X} ) = 5$, $\beta_0(\mathfrak{X}) = 1$, hence
$\textup{\textsf{P}}_{\mathrm{chio}}[\mathcal{E}_B]$ $=$ $(\frac12)^{\lvert I \rvert+5-1}$ $=$ $(\frac12)^{10}$.\\
If $\mathfrak{X}\in \{\ref{item:comparisonproofcaseLhassize6:C4intersectingtwoedgesinseparatenonadjacentvertices}, \ref{item:comparisonproofcaseLhassize6:C6}\}$, then
$f_0( \mathfrak{X} ) = 6$, $\beta_0(\mathfrak{X}) = 1$, hence
$\textup{\textsf{P}}_{\mathrm{chio}}[\mathcal{E}_B]$ $=$ $(\frac12)^{\lvert I \rvert+6-1}$ $=$
$(\frac12)^{11}$.
As to \ref{item:absolutevaluesofchiomeasureonfailureentryspecificationevent:5entriesspecified}, it follows by an entirely analogous (but much shorter) argument as the
one given for \ref{item:absolutevaluesofchiomeasureonfailureentryspecificationevent:5entriesspecified} that it suffices to calculate only $\textup{\textsf{P}}_{\mathrm{chio}}[\ref{item:comparisonproofcaseLhassize6:oneisolatedvertex}]$ and $\textup{\textsf{P}}_{\mathrm{chio}}[\ref{item:comparisonproofcaseLhassize6:oneadditionaledgeintersectingC4}]$, and these calculations are
identical to the ones made for $\textup{\textsf{P}}_{\mathrm{chio}}[\ref{item:comparisonproofcaseLhassize6:oneisolatedvertex}]$ and $\textup{\textsf{P}}_{\mathrm{chio}}[\ref{item:comparisonproofcaseLhassize6:oneadditionaledgeintersectingC4}]$ in the preceding paragraph,
except that now $\lvert I \rvert = 5$.
As to \ref{item:absolutevaluesofchiomeasureonfailureentryspecificationevent:4entriesspecified}, in view of \ref{item:failuresetofisomorphismtypesk4} in
Corollary \ref{cor:isomorphismtypesforwhichequalityofmeasuresofentryspecificationeventsfails}, we only have to deal with the single type \ref{item:comparisonproofcaseLhassize4:4circuit} where $f_0\ref{item:comparisonproofcaseLhassize4:4circuit} = 4$,
$\beta_0\ref{item:comparisonproofcaseLhassize4:4circuit} = 1$ and therefore
$\textup{\textsf{P}}_{\mathrm{chio}}[\ref{item:comparisonproofcaseLhassize4:4circuit}] =
(\tfrac12)^{\lvert I \rvert + 4 - 1} = (\tfrac12)^{4+4-1} = (\tfrac12)^7$.
\end{proof}
The results obtained so far can be turned into a set of instructions of how to
tell the measure of $\textup{\textsf{P}}_{\mathrm{chio}}[\mathcal{E}_B^{[n-1]^2}]$ from a
given $B\in \{0,\pm\}^I$ provided that $\lvert I \rvert \leq 6$. We formulate the
instructions exclusively in terms of those data, avoiding any mention
of the associated signed graph ($\mathrm{X}_B$,$\sigma_B$).
\begin{corollary}[how to find the measure under $\textup{\textsf{P}}_{\mathrm{chio}}$ of large
entry-specification events]
\label{cor:recipe}
For every $\emptyset \subseteq I \subseteq [n-1]^2$ with $\lvert I \rvert \leq 6$
and every $B\in \{0,\pm\}^I $, the following instructions lead to the
correct Chio measure of $\mathcal{E}_B := \mathcal{E}_B^{[n-1]^2}$:
\begin{enumerate}[label={\rm(H\arabic{*})}, start=3]
\item\label{comparisonthm:item:uptto3entries} If
$0\leq \lvert I\rvert \leq 3$, then
$\textup{\textsf{P}}_{\mathrm{chio}}[\mathcal{E}_B] = \textup{\textsf{P}}_{\mathrm{lcf}} [\mathcal{E}_B]
= (\frac12)^{\mathrm{dom}(B)+\mathrm{supp}(B)}$ .
\item\label{thm:it:fourentriesspecified} If $\lvert I\rvert = 4$,
then check whether $I$ is a matrix-$4$-circuit \emph{such that}
$B\in \{\pm\}^I$. If not, then
$\textup{\textsf{P}}_{\mathrm{chio}}[\mathcal{E}_B] = \textup{\textsf{P}}_{\mathrm{lcf}} [\mathcal{E}_B]
= (\frac12)^{\mathrm{dom}(B)+\mathrm{supp}(B)}$. If $B$ has this property, then check whether
an odd number of the four nonzero values $B[i,j]$ with $(i,j)\in I$ are $+$.
If so, $\textup{\textsf{P}}_{\mathrm{chio}}[\mathcal{E}_B]=0$. If not, then
$\textup{\textsf{P}}_{\mathrm{chio}}[\mathcal{E}_B] = (\frac12)^7 = 2\cdot \textup{\textsf{P}}_{\mathrm{lcf}}[\mathcal{E}_B]$.
\item\label{thm:it:fiveentriesspecified}
If $\lvert I\rvert = 5$, then check whether there exists within $I$ a
matrix-$4$-circuit $S\subseteq I$ \emph{such that} $B\mid_{S}\in \{\pm\}^S$.
If not, then
$\textup{\textsf{P}}_{\mathrm{chio}}[\mathcal{E}_B] = \textup{\textsf{P}}_{\mathrm{lcf}} [\mathcal{E}_B]
= (\frac12)^{\mathrm{dom}(B)+\mathrm{supp}(B)}$. If so, then check whether an odd number of
the four nonzero values $B[i,j]$ with $(i,j)\in S$ are $+$.
If so, $\textup{\textsf{P}}_{\mathrm{chio}}[\mathcal{E}_B]=0$. If not, then check whether the
entry $B[i,j]$ with $(i,j)\in I\setminus S$ is zero. If it is, then
$\textup{\textsf{P}}_{\mathrm{chio}}[\mathcal{E}_B] = (\frac12)^8 =
2 \cdot \textup{\textsf{P}}_{\mathrm{lcf}}[\mathcal{E}_B]$. If it is nonzero, then
$\textup{\textsf{P}}_{\mathrm{chio}}[\mathcal{E}_B] = (\frac12)^9 =
2\cdot \textup{\textsf{P}}_{\mathrm{lcf}}[\mathcal{E}_B]$.
\item\label{thm:it:sixentriesspecified}
If $\lvert I\rvert = 6$, then check whether at least four entries in $B$
are nonzero. If not, then $\textup{\textsf{P}}_{\mathrm{chio}}[\mathcal{E}_B] = \textup{\textsf{P}}_{\mathrm{lcf}} [\mathcal{E}_B]
= (\frac12)^{\mathrm{dom}(B)+\mathrm{supp}(B)}$. If so, then check whether $I$ contains a
matrix-$4$-circuit $S\subseteq I$ with $B\mid_{S} \in \{\pm\}^S$.
\begin{enumerate}[label={\rm(\roman{*})}]
\item\label{item:thmonfindingPchio:sizeLis6no4circuit} If $I$ does not
contain such a matrix-$4$-circuit $S$, then check whether $I$ is a
matrix-$6$-circuit such that $B \in \{\pm\}^I$. If not, then
$\textup{\textsf{P}}_{\mathrm{chio}}[\mathcal{E}_B] = \textup{\textsf{P}}_{\mathrm{lcf}} [\mathcal{E}_B]
= (\frac12)^{\mathrm{dom}(B)+\mathrm{supp}(B)}$. If so, then check
whether an odd number of these six entries are $+$. If so, then
$\textup{\textsf{P}}_{\mathrm{chio}}[\mathcal{E}_B] = 0$. If not, then
$\textup{\textsf{P}}_{\mathrm{chio}}[\mathcal{E}_B] = (\frac12)^{11} =
2\cdot \textup{\textsf{P}}_{\mathrm{lcf}}[\mathcal{E}_B]$.
\item\label{item:thmonfindingPchio:sizeLis6thereis4circuit} If $I$ does indeed
contain such a matrix-$4$-circuit, then check whether an odd number of the four
nonzero values $B[i,j]$ with $(i,j)\in S$ are $+$.
If so, then $\textup{\textsf{P}}_{\mathrm{chio}}[\mathcal{E}_B] = 0$. Else, there are three
further cases:
\begin{enumerate}[label={\rm(\alph{*})}]
\item
\label{item:thmonfindingPchio:sizeLis6thereis4circuit:bothnoncircuitentrieszero}
If both entries indexed by $I\setminus S$ are zero, then
$\textup{\textsf{P}}_{\mathrm{chio}}[\mathcal{E}_B] = (\frac12)^9 = 2\cdot
\textup{\textsf{P}}_{\mathrm{lcf}} [\mathcal{E}_B]$.
\item
\label{item:thmonfindingPchio:sizeLis6thereis4circuit:onenoncircuitentrieszero}
If exactly one of the two entries indexed by $I\setminus S$ is zero, then
$\textup{\textsf{P}}_{\mathrm{chio}}[\mathcal{E}_B] = (\frac12)^{10} = 2\cdot
\textup{\textsf{P}}_{\mathrm{lcf}} [\mathcal{E}_B]$.
\item
\label{item:thmonfindingPchio:sizeLis6thereis4circuit:bothnoncircuitentriesnonzero}
If both entries indexed by $I\setminus S$ are nonzero, then the positions of
these two nonzero entries must be taken into account: if there do \emph{neither}
exist $1\leq i<i'<i'' \leq n-1$ and $1 \leq j < j' \leq n-1$ such
that $I$ $=$ $\{$ $(i,j),$ $(i',j),$ $(i'',j),$ $(i,j'),$ $(i',j'),$ $(i'',j')$
$\}$ \emph{nor} $1\leq i < i' \leq n-1$ and $1\leq j < j' < j'' \leq n-1$ such
that $I$ $=$ $\{$ $(i,j),$ $(i,j'),$ $(i,j''),$ $(i',j),$ $(i',j'),$
$(i',j'')$ $\}$, then---whatever the $B$-values indexed by the two elements in
$I\setminus S$ may be---you know that
$\textup{\textsf{P}}_{\mathrm{chio}}[\mathcal{E}_B] = (\frac12)^{11} = 2\cdot
\textup{\textsf{P}}_{\mathrm{lcf}} [\mathcal{E}_B]$. Else, check whether in any one of the then
existing two additional matrix-$4$-circuits in $I$ the number of $+$ entries
is odd. If so, then $\textup{\textsf{P}}_{\mathrm{chio}}[\mathcal{E}_B]=0$. If not, then
$\textup{\textsf{P}}_{\mathrm{chio}}[\mathcal{E}_B] = (\frac12)^{10}
= 4\cdot \textup{\textsf{P}}_{\mathrm{lcf}} [\mathcal{E}_B]$. \hfill $\Box$
\end{enumerate}
\end{enumerate}
\end{enumerate}
\end{corollary}
If $U\subseteq \mathrm{Dom}(\mathrm{X}^{k,n,n}) = \bigsqcup_{I\in\binom{[s-1]\times [t-1]}{k}}\{0,\pm\}^I$
is an arbitrary subset, then on the abstract set-theoretical level all we know is
$(\mathrm{X}^{k,n,n})^{-1}(\mathrm{X}^{k,n,n}(U))\supseteq U$. For the specific subsets
$U = \mathcal{F}^{\mathrm{M}}(k,n)$, however, the inclusion is an equality:
\begin{corollary}\label{cor:graphmapcanbeinvertedsetwiseonthefailuresets}
$(\mathrm{X}^{k,n,n})^{-1}(\mathrm{X}^{k,n,n}(\mathcal{F}^{\mathrm{M}}(k,n))) =
\mathcal{F}^{\mathrm{M}}(k,n)$
\end{corollary}
\begin{proof}
Since $(\mathrm{X}^{k,n,n})^{-1}(\mathrm{X}^{k,n,n}(\mathcal{F}^{\mathrm{M}}(k,n)))$
$\By{Corollary \ref{cor:quickconsequencesofthecharacterizations}.\ref{cor:graphfailuresetaspreimage}}{=}$
$(\mathrm{X}^{k,n,n})^{-1}(\mathrm{X}^{k,n,n}( (\mathrm{X}^{k,n,n})^{-1}(\beta_1^{-1}(\mathbb{Z}_{\geq 1}))))$ \\
$\By{Lemma \ref{lem:elementarypreimagestatements:finvffinv}}{=}$
$(\mathrm{X}^{k,n,n})^{-1} (\beta_1^{-1}(\mathbb{Z}_{\geq 1}))$
$\By{again Corollary \ref{cor:quickconsequencesofthecharacterizations}.\ref{cor:graphfailuresetaspreimage}}{=}$
$\mathcal{F}^{\mathrm{M}}(k,n)$.
\end{proof}
Corollaries \ref{cor:isomorphismtypesforwhichequalityofmeasuresofentryspecificationeventsfails} and \ref{cor:graphmapcanbeinvertedsetwiseonthefailuresets} allow us to
express the failure sets $\mathcal{F}^{\mathrm{M}}(k,n)$ as partitions indexed by
isomorphism types of bipartite graphs:
\begin{corollary}\label{cor:partitionsoffailuresets}
For every $n$,
\begin{enumerate}[label={\rm(M\arabic{*})}, start=4]
\item\label{it:partitionoffailuresets:kequals4} $\mathcal{F}^{\mathrm{M}}(4,n) =
(\leftsubscript{\mathrm{ul}}{\upX}^{4,n,n})^{-1}\ref{item:comparisonproofcaseLhassize4:4circuit}$\quad ,
\item\label{it:partitionoffailuresets:kequals5} $\mathcal{F}^{\mathrm{M}}(5,n) =
(\leftsubscript{\mathrm{ul}}{\upX}^{5,n,n})^{-1}\ref{item:comparisonproofcaseLhassize6:oneisolatedvertex} \sqcup
(\leftsubscript{\mathrm{ul}}{\upX}^{5,n,n})^{-1}\ref{item:comparisonproofcaseLhassize6:oneadditionaledgeintersectingC4} \sqcup (\leftsubscript{\mathrm{ul}}{\upX}^{5,n,n})^{-1}\ref{item:comparisonproofcaseLhassize6:twoisolatedvertices}\sqcup (\leftsubscript{\mathrm{ul}}{\upX}^{5,n,n})^{-1}\ref{item:comparisonproofcaseLhassize6:C4withoneadditionaldisjointedge}$\quad ,
\item\label{it:partitionoffailuresets:kequals6} $\mathcal{F}^{\mathrm{M}}(6,n) = (\leftsubscript{\mathrm{ul}}{\upX}^{6,n,n})^{-1}\ref{item:comparisonproofcaseLhassize6:oneisolatedvertex} \sqcup (\leftsubscript{\mathrm{ul}}{\upX}^{6,n,n})^{-1}\ref{item:comparisonproofcaseLhassize6:oneadditionaledgeintersectingC4} \sqcup (\leftsubscript{\mathrm{ul}}{\upX}^{6,n,n})^{-1}\ref{item:comparisonproofcaseLhassize6:twoisolatedvertices} \sqcup (\leftsubscript{\mathrm{ul}}{\upX}^{6,n,n})^{-1}\ref{item:comparisonproofcaseLhassize6:C4withoneadditionaldisjointedge} \\
\sqcup (\leftsubscript{\mathrm{ul}}{\upX}^{6,n,n})^{-1}
\ref{item:comparisonproofcaseLhassize6:XBLisomorphictoK23}
\sqcup \bigsqcup_{6\leq k \leq 20} (\leftsubscript{\mathrm{ul}}{\upX}^{6,n,n})^{-1}(\mathrm{t}k)$\quad .
\end{enumerate}
\end{corollary}
\begin{proof}
In general we have
$
\mathcal{F}^{\mathrm{M}}(k,n)
\By{\text{\tiny Corollary \ref{cor:graphmapcanbeinvertedsetwiseonthefailuresets}}}{=}
(\leftsubscript{\mathrm{ul}}{\upX}^{k,n,n})^{-1}(\leftsubscript{\mathrm{ul}}{\upX}^{k,n,n}(\mathcal{F}^{\mathrm{M}}(k,n)))
\By{\text{\tiny (by Definition \ref{def:parametrizedfailuresets}.\ref{def:graphversionsofthefailuresets})}}{=} (\leftsubscript{\mathrm{ul}}{\upX}^{k,n,n})^{-1}( \mathcal{F}^{\mathrm{G}}(k,n) )
\By{\text{\tiny (for every map)}}{=}
\bigsqcup_{\mathfrak{X}\in\mathcal{F}^{\mathrm{G}}(k,n)} (\leftsubscript{\mathrm{ul}}{\upX}^{k,n,n})^{-1}(\mathfrak{X})$,
and for the specific values $4\leq k \leq 6$ we can use Corollary
\ref{cor:isomorphismtypesforwhichequalityofmeasuresofentryspecificationeventsfails}
to obtain the claimed partitions.
\end{proof}
While having the aim of explicitly determining the
numbers $\lvert(\leftsubscript{\mathrm{ul}}{\upX}^{k,n,n})^{-1}(\mathfrak{X})\rvert$ for
certain $k$ and $\mathfrak{X}$ which interest us, we will start slowly by first
formulating some linear relations among $\lvert(\leftsubscript{\mathrm{ul}}{\upX}^{5,n,n})^{-1}\ref{item:comparisonproofcaseLhassize6:oneisolatedvertex}\rvert$, $\dotsc$, $\lvert(\leftsubscript{\mathrm{ul}}{\upX}^{6,n,n})^{-1}\ref{item:comparisonproofcaseLhassize6:twoadditionaldisjointedgesdisjointfromC4}\rvert$ which will later serve as a check for the formulas given in
Theorem \ref{thm:countingmatrixrealizations}.
\begin{lemma}[linear relations among
$\lvert (\leftsubscript{\mathrm{ul}}{\upX}^{k,n,n})^{-1}(\mathfrak{X}) \rvert$ for $5\leq k \leq 6$]
\label{lem:relationsamongthenumbersofmatrixrealizations}
$\quad$
\begin{enumerate}[label={\rm(l\arabic{*})}]
\item\label{linearrelationbetweent2andt3:kequals5}
$(3^1-1)\cdot \lvert (\leftsubscript{\mathrm{ul}}{\upX}^{5,n,n})^{-1}\ref{item:comparisonproofcaseLhassize6:oneisolatedvertex} \rvert
= \lvert (\leftsubscript{\mathrm{ul}}{\upX}^{5,n,n})^{-1}\ref{item:comparisonproofcaseLhassize6:oneadditionaledgeintersectingC4}\rvert$\quad ,
\item\label{linearrelationbetweent5andt7:kequals5}
$(3^1-1)\cdot \lvert (\leftsubscript{\mathrm{ul}}{\upX}^{5,n,n})^{-1}\ref{item:comparisonproofcaseLhassize6:twoisolatedvertices} \rvert
= \lvert (\leftsubscript{\mathrm{ul}}{\upX}^{5,n,n})^{-1}\ref{item:comparisonproofcaseLhassize6:C4withoneadditionaldisjointedge}\rvert$\quad ,
\item\label{linearrelationbetweent4tot10:kequals6}
$
(3^2 -1) \cdot \lvert(\leftsubscript{\mathrm{ul}}{\upX}^{6,n,n})^{-1}\ref{item:comparisonproofcaseLhassize6:twoisolatedvertices}\rvert
=
\lvert(\leftsubscript{\mathrm{ul}}{\upX}^{6,n,n})^{-1}\ref{item:comparisonproofcaseLhassize6:oneadditionaledgeintersectingC4andoneisolatedvertex}\rvert
+ \dotsm +
\lvert(\leftsubscript{\mathrm{ul}}{\upX}^{6,n,n})^{-1}\ref{item:comparisonproofcaseLhassize6:C4intersectingatwopathinitsinnervertex}\rvert$\quad ,
\item\label{linearrelationbetweent11tot15:kequals6}
$
(3^2 - 1) \cdot \lvert(\leftsubscript{\mathrm{ul}}{\upX}^{6,n,n})^{-1}\ref{item:comparisonproofcaseLhassize6:threeisolatedvertices}\rvert
=
\lvert(\leftsubscript{\mathrm{ul}}{\upX}^{6,n,n})^{-1}\ref{item:comparisonproofcaseLhassize6:oneadditionaledgeintersectingC4andtwoisolatedvertices}\rvert
+
\dotsm
+
\lvert(\leftsubscript{\mathrm{ul}}{\upX}^{6,n,n})^{-1}\ref{item:comparisonproofcaseLhassize6:twoadditionalnondisjointedgesdisjointfromC4}\rvert$\quad ,
\item\label{linearrelationbetweent16tot18:kequals6}
$
(3^2 - 1) \cdot \lvert(\leftsubscript{\mathrm{ul}}{\upX}^{6,n,n})^{-1}\ref{item:comparisonproofcaseLhassize6:fourisolatedvertices}\rvert
=
\lvert(\leftsubscript{\mathrm{ul}}{\upX}^{6,n,n})^{-1}\ref{item:comparisonproofcaseLhassize6:oneadditionaldisjointedgeandtwoisolatedvertices}\rvert
+
\lvert(\leftsubscript{\mathrm{ul}}{\upX}^{6,n,n})^{-1}\ref{item:comparisonproofcaseLhassize6:twoadditionaldisjointedgesdisjointfromC4}\rvert$\quad .
\end{enumerate}
\end{lemma}
\begin{proof}
It follows from Definition \ref{def:XBandecXB} that
$\mathrm{X}_{B}\in\ref{item:comparisonproofcaseLhassize6:oneadditionaledgeintersectingC4}$ if and only if equation \eqref{eq:conditioninproofofnumberofrealizationsoftype:oneisolatedvertex:kequals5} is true and $B[u]\in\{\pm\}$. This implies
$\lvert(\leftsubscript{\mathrm{ul}}{\upX}^{5,n,n})^{-1}\ref{item:comparisonproofcaseLhassize6:oneadditionaledgeintersectingC4}\rvert =
2\cdot \lvert(\leftsubscript{\mathrm{ul}}{\upX}^{5,n,n})^{-1}\ref{item:comparisonproofcaseLhassize6:oneisolatedvertex}\rvert$, proving \ref{linearrelationbetweent2andt3:kequals5}.
It also follows from Definition \ref{def:XBandecXB} that $\mathrm{X}_{B}\in\ref{item:comparisonproofcaseLhassize6:C4withoneadditionaldisjointedge}$ if and only if equation \eqref{eq:conditioninproofofnumberofrealizationsoftype:twoisolatedvertices:kequals5} is true and $B[u]\in\{\pm\}$. This implies $\lvert(\leftsubscript{\mathrm{ul}}{\upX}^{5,n,n})^{-1}\ref{item:comparisonproofcaseLhassize6:C4withoneadditionaldisjointedge}\rvert= 2\cdot \lvert(\leftsubscript{\mathrm{ul}}{\upX}^{5,n,n})^{-1}\ref{item:comparisonproofcaseLhassize6:twoisolatedvertices}\rvert$, proving \ref{linearrelationbetweent5andt7:kequals5}.
The isomorphism types \ref{item:comparisonproofcaseLhassize6:twoisolatedvertices}--\ref{item:comparisonproofcaseLhassize6:C4intersectingatwopathinitsinnervertex} are all the isomorphism types of bipartite nonforests with six vertices and exactly one
copy of $C^4$. Therefore
$\lvert(\leftsubscript{\mathrm{ul}}{\upX}^{6,n,n})^{-1}\ref{item:comparisonproofcaseLhassize6:twoisolatedvertices}\rvert + \dotsm + \lvert(\leftsubscript{\mathrm{ul}}{\upX}^{6,n,n})^{-1}\ref{item:comparisonproofcaseLhassize6:C4intersectingatwopathinitsinnervertex}\rvert
$ is the number of all $B\in\{0,\pm\}^I$ with $I\in\binom{[n-1]^2}{6}$ such that
$\mathrm{X}_{B}$ contains exactly one $C^4$ and $f_0(\mathrm{X}_{B}) = 6$. Imagine counting these
$B$ by partitioning the set of all such $B$ according to the copy of $C^4$, and
for each such copy, further partitioning the $B$ according to the mandatory
$\pm$-values on the edges of the $C^4$, and then further partitioning according to
the \emph{positions} of the two elements of $I$ which are not responsible for the
copy of $C^4$. When partitioning in that way, the number of blocks of the partition
obtained so far equals $\lvert(\leftsubscript{\mathrm{ul}}{\upX}^{6,n,n})^{-1}\ref{item:comparisonproofcaseLhassize6:twoisolatedvertices}\rvert$. The reason for this is that to realize the
type $\ref{item:comparisonproofcaseLhassize6:twoisolatedvertices}$ there is no
choice for the values indexed by the positions which are not responsible for
the $C^4$, both must be zero. In the enumeration we are currently carrying out,
however, there \emph{is} still complete freedom left on how to choose any one of the
$\lvert \{0,\pm\}^{[2]} \rvert = 3^2$ values which can be indexed by these two
positions, in other words, each of the blocks has size $3^2$. Therefore
$
\lvert(\leftsubscript{\mathrm{ul}}{\upX}^{6,n,n})^{-1}\ref{item:comparisonproofcaseLhassize6:twoisolatedvertices}\rvert
+ \dotsm +
\lvert(\leftsubscript{\mathrm{ul}}{\upX}^{6,n,n})^{-1}\ref{item:comparisonproofcaseLhassize6:C4intersectingatwopathinitsinnervertex}\rvert
=
3^2\cdot \lvert(\leftsubscript{\mathrm{ul}}{\upX}^{6,n,n})^{-1}\ref{item:comparisonproofcaseLhassize6:twoisolatedvertices}\rvert$, which proves \ref{linearrelationbetweent4tot10:kequals6}. Equations \ref{linearrelationbetweent11tot15:kequals6} and \ref{linearrelationbetweent16tot18:kequals6} are true for an entirely analogous reason.
\end{proof}
We will now quantify the claims in Corollary \ref{cor:isomorphismtypesforwhichequalityofmeasuresofentryspecificationeventsfails} by determining
$\lvert(\leftsubscript{\mathrm{ul}}{\upX}^{k,n,n})^{-1}(\mathfrak{X})\rvert$ for each $k$ and each isomorphism
type $\mathfrak{X}$ mentioned there. A few preparatory comments seem in order. The
behaviour of $\lvert(\leftsubscript{\mathrm{ul}}{\upX}^{k,n,n})^{-1}(\mathfrak{X})\rvert$ as a function of $k$
for a given isomorphism type $\mathfrak{X}$ is a little subtle. For example, note that
Theorem \ref{thm:countingmatrixrealizations} tells us that
\begin{equation}\label{eq:exampleforcounterintuitivepointinenumerationofrealizations}
\lvert(\leftsubscript{\mathrm{ul}}{\upX}^{5,n,n})^{-1}\ref{item:comparisonproofcaseLhassize6:oneisolatedvertex}\rvert > \lvert(\leftsubscript{\mathrm{ul}}{\upX}^{6,n,n})^{-1}\ref{item:comparisonproofcaseLhassize6:oneisolatedvertex}\rvert
\end{equation}
in spite of the fact that in the case of
$\lvert(\leftsubscript{\mathrm{ul}}{\upX}^{6,n,n})^{-1}\ref{item:comparisonproofcaseLhassize6:oneisolatedvertex}\rvert$ we have one matrix entry more at our disposal to realize \ref{item:comparisonproofcaseLhassize6:oneisolatedvertex}. The reason for this could be summarized thus:
when wanting to keep the number of isolated vertices in $\leftsubscript{\mathrm{ul}}{\upX}_{B}$ at one, the additional matrix entry curtails our freedom more than it
adds to it---after having chosen a position for one of the
non-matrix-circuit-entries which `hides' one of its two `shadows' in one
of the four shadows of the matrix-circuit-entries, we then have to position the
second non-matrix-circuit-entry so as to hide \emph{both} of its two
shadows in already existing shadows, and this determines it position completely.
Moreover, since \ref{item:comparisonproofcaseLhassize6:oneisolatedvertex} is an
isomorphism type in which there do not exist edges outside the $4$-circuit, the
non-matrix-circuit positions must index the value $0$. The net result of these
rigid requirements are (since in effect for $\lvert(\leftsubscript{\mathrm{ul}}{\upX}^{6,n,n})^{-1}\ref{item:comparisonproofcaseLhassize6:oneisolatedvertex}\rvert$ we are counting the
possible \emph{$2$-sets} of non-circuit positions while for $\lvert(\leftsubscript{\mathrm{ul}}{\upX}^{5,n,n})^{-1}\ref{item:comparisonproofcaseLhassize6:oneisolatedvertex}\rvert$ we counted the
possible $1$-sets of such positions) \emph{less} possibilities. For other types it
can happen that the mechanism just described is counterbalanced by the additional
possibilities of indexing different values. This is the essential reason why
$\lvert(\leftsubscript{\mathrm{ul}}{\upX}^{5,n,n})^{-1}\ref{item:comparisonproofcaseLhassize6:oneadditionaledgeintersectingC4}\rvert
=
\lvert(\leftsubscript{\mathrm{ul}}{\upX}^{6,n,n})^{-1}\ref{item:comparisonproofcaseLhassize6:oneadditionaledgeintersectingC4}\rvert$, despite \eqref{eq:exampleforcounterintuitivepointinenumerationofrealizations} and despite the fact that the set of all \emph{domains} in the
preimages in question are the same as in
\eqref{eq:exampleforcounterintuitivepointinenumerationofrealizations}, i.e.
{
\scriptsize
\begin{equation}
\mathrm{Dom}((\leftsubscript{\mathrm{ul}}{\upX}^{5,n,n})^{-1}\ref{item:comparisonproofcaseLhassize6:oneisolatedvertex})
= \mathrm{Dom}((\leftsubscript{\mathrm{ul}}{\upX}^{5,n,n})^{-1}\ref{item:comparisonproofcaseLhassize6:oneadditionaledgeintersectingC4}) \quad , \qquad
\mathrm{Dom}((\leftsubscript{\mathrm{ul}}{\upX}^{6,n,n})^{-1}\ref{item:comparisonproofcaseLhassize6:oneisolatedvertex})
= \mathrm{Dom}((\leftsubscript{\mathrm{ul}}{\upX}^{6,n,n})^{-1}\ref{item:comparisonproofcaseLhassize6:oneadditionaledgeintersectingC4}) \quad .
\end{equation}
}
Since biadjacency matrices are quite a fundamental topic, it would be of interest to
treat these phenomena in more generality. It seems advisable to do this with a
view towards the theory of $\{0,1\}$-matrices with given row and column sums (for a
start, cf. e.g. \cite{MR878703}, \cite{arXiv:1010.5706v1} and \cite{MR2600999}).
However, so far the author could not harness the literature on this topic in any way
which would lessen the burden of proving the following theorem:
\begin{theorem}[cardinality of preimages of $\leftsubscript{\mathrm{ul}}{\upX}^{k,n,n}$ on bipartite
nonforests for $4\leq k \leq 6$]\label{thm:countingmatrixrealizations}
The claims \ref{item:failuresetofisomorphismtypesk4}---\ref{item:failuresetofisomorphismtypesk6} can be quantified as follows
(with $\xi_n:=2^4\cdot \lvert \mathrm{Cir}(4,n)\rvert = 2^4\cdot\binom{n-1}{2}^2$),
\begin{enumerate}[label={\rm(QFa\arabic{*})}, start=4]
\item\label{item:numberofrealizationsofnonforestswhenkequals4} For every $n\geq 3$,
$\lvert(\leftsubscript{\mathrm{ul}}{\upX}^{4,n,n})^{-1}\ref{item:comparisonproofcaseLhassize4:4circuit}\rvert = \xi_n$\quad .
\item\label{item:numberofrealizationsofnonforestswhenkequals5} For every $n\geq 3$,
{\scriptsize
\begin{minipage}[b]{0.45\linewidth}
\begin{enumerate}[label={\rm(m5.t\arabic{*})}, start=2]
\item\label{numberofmatrixrealizationsoftype:oneisolatedvertex:kequals5}
$\lvert(\leftsubscript{\mathrm{ul}}{\upX}^{5,n,n})^{-1}\ref{item:comparisonproofcaseLhassize6:oneisolatedvertex}
\rvert = 4\cdot(n-3) \cdot \xi_n$
\item\label{numberofmatrixrealizationsoftype:oneadditionaledgeintersectingC4:kequals5}
$\lvert(\leftsubscript{\mathrm{ul}}{\upX}^{5,n,n})^{-1}\ref{item:comparisonproofcaseLhassize6:oneadditionaledgeintersectingC4}\rvert = 8\cdot(n-3) \cdot \xi_n$
\end{enumerate}
\end{minipage}
\begin{minipage}[b]{0.45\linewidth}
\begin{enumerate}[label={\rm(m5.t\arabic{*})}, start=5]
\item\label{numberofmatrixrealizationsoftype:twoisolatedvertices:kequals5}
$\lvert(\leftsubscript{\mathrm{ul}}{\upX}^{5,n,n})^{-1}\ref{item:comparisonproofcaseLhassize6:twoisolatedvertices}\rvert = 1\cdot (n-3)^2 \cdot \xi_n$
\end{enumerate}
\begin{enumerate}[label={\rm(m5.t\arabic{*})}, start=7]
\item\label{numberofmatrixrealizationsoftype:C4withoneadditionaldisjointedge:kequals5}
$\lvert(\leftsubscript{\mathrm{ul}}{\upX}^{5,n,n})^{-1}\ref{item:comparisonproofcaseLhassize6:C4withoneadditionaldisjointedge}\rvert = 2\cdot (n-3)^2 \cdot \xi_n$\quad .
\end{enumerate}
\end{minipage}
}
\item\label{item:numberofrealizationsofnonforestswhenkequals6}
For every $n\geq 3$,
{\scriptsize
\begin{minipage}[b]{0.5\linewidth}
\begin{enumerate}[label={\rm(m6.t\arabic{*})}, start=2]
\item\label{numberofmatrixrealizationsoftype:oneisolatedvertex}
$\lvert(\leftsubscript{\mathrm{ul}}{\upX}^{6,n,n})^{-1}\ref{item:comparisonproofcaseLhassize6:oneisolatedvertex}\rvert = 2 (n-3) \xi_n$
\item\label{numberofmatrixrealizationsoftype:oneadditionaledgeintersectingC4}
$\lvert(\leftsubscript{\mathrm{ul}}{\upX}^{6,n,n})^{-1}\ref{item:comparisonproofcaseLhassize6:oneadditionaledgeintersectingC4}\rvert = 8 (n-3) \xi_n$
\item\label{numberofmatrixrealizationsoftype:XBLisomorphictoK23}
$\lvert(\leftsubscript{\mathrm{ul}}{\upX}^{6,n,n})^{-1}\ref{item:comparisonproofcaseLhassize6:XBLisomorphictoK23}\rvert = 2^7 \binom{n-1}{2} \binom{n-1}{3}$
\item\label{numberofmatrixrealizationsoftype:twoisolatedvertices}
$\lvert(\leftsubscript{\mathrm{ul}}{\upX}^{6,n,n})^{-1}\ref{item:comparisonproofcaseLhassize6:twoisolatedvertices}\rvert = (8 (n-3)^2 + 8 \binom{n-3}{2}) \xi_n$
\item\label{numberofmatrixrealizationsoftype:oneadditionaledgeintersectingC4andoneisolatedvertex}
$\lvert(\leftsubscript{\mathrm{ul}}{\upX}^{6,n,n})^{-1}\ref{item:comparisonproofcaseLhassize6:oneadditionaledgeintersectingC4andoneisolatedvertex}\rvert = (24 (n-3)^2 + 32 \binom{n-3}{2}) \xi_n$
\item\label{numberofmatrixrealizationsoftype:C4withoneadditionaldisjointedge}
$\lvert(\leftsubscript{\mathrm{ul}}{\upX}^{6,n,n})^{-1}\ref{item:comparisonproofcaseLhassize6:C4withoneadditionaldisjointedge}\rvert = 8 (n-3)^2 \xi_n$
\item\label{numberofmatrixrealizationsoftype:C4intersectingtwoedgesinseparatenonadjacentvertices}
$\lvert(\leftsubscript{\mathrm{ul}}{\upX}^{6,n,n})^{-1}\ref{item:comparisonproofcaseLhassize6:C4intersectingtwoedgesinseparatenonadjacentvertices}\rvert = 16 \binom{n-3}{2} \xi_n$
\item\label{numberofmatrixrealizationsoftype:C4intersectingtwoedgesinseparateadjacentvertices}
$\lvert(\leftsubscript{\mathrm{ul}}{\upX}^{6,n,n})^{-1}\ref{item:comparisonproofcaseLhassize6:C4intersectingtwoedgesinseparateadjacentvertices}\rvert = 16 (n-3)^2 \xi_n$
\item\label{numberofmatrixrealizationsoftype:C4intersectingatwopathinanendvertex}
$\lvert(\leftsubscript{\mathrm{ul}}{\upX}^{6,n,n})^{-1}\ref{item:comparisonproofcaseLhassize6:C4intersectingatwopathinanendvertex}\rvert = 16 (n-3)^2 \xi_n$
\item\label{numberofmatrixrealizationsoftype:C4intersectingatwopathinitsinnervertex}
$\lvert(\leftsubscript{\mathrm{ul}}{\upX}^{6,n,n})^{-1}\ref{item:comparisonproofcaseLhassize6:C4intersectingatwopathinitsinnervertex}\rvert = 16 \binom{n-3}{2} \xi_n$
\end{enumerate}
\end{minipage}
\begin{minipage}[b]{0.5\linewidth}
\begin{enumerate}[label={\rm(m6.t\arabic{*})}, start=12]
\item\label{numberofmatrixrealizationsoftype:C6}
$\lvert(\leftsubscript{\mathrm{ul}}{\upX}^{6,n,n})^{-1}\ref{item:comparisonproofcaseLhassize6:C6}\rvert =
2^6 \lvert \mathrm{Cir}(6,n) \rvert$
\item\label{numberofmatrixrealizationsoftype:threeisolatedvertices}
$\lvert(\leftsubscript{\mathrm{ul}}{\upX}^{6,n,n})^{-1}\ref{item:comparisonproofcaseLhassize6:threeisolatedvertices}\rvert = 10 (n-3) \binom{n-3}{2} \xi_n$
\item\label{numberofmatrixrealizationsoftype:oneadditionaledgeintersectingC4andtwoisolatedvertices}
$\lvert(\leftsubscript{\mathrm{ul}}{\upX}^{6,n,n})^{-1}\ref{item:comparisonproofcaseLhassize6:oneadditionaledgeintersectingC4andtwoisolatedvertices}\rvert = 16 (n-3) \binom{n-3}{2} \xi_n$
\item\label{numberofmatrixrealizationsoftype:oneadditionaldisjointedgeandoneisolatedvertex}
$\lvert(\leftsubscript{\mathrm{ul}}{\upX}^{6,n,n})^{-1}\ref{item:comparisonproofcaseLhassize6:oneadditionaldisjointedgeandoneisolatedvertex}\rvert = 24 (n-3) \binom{n-3}{2} \xi_n$
\item\label{numberofmatrixrealizationsoftype:twoadditionaledgesonlyoneofthemdisjoint}
$\lvert(\leftsubscript{\mathrm{ul}}{\upX}^{6,n,n})^{-1}\ref{item:comparisonproofcaseLhassize6:twoadditionaledgesonlyoneofthemdisjoint}\rvert = 32 (n-3)\binom{n-3}{2} \xi_n$
\item\label{numberofmatrixrealizationsoftype:twoadditionalnondisjointedgesdisjointfromC4}
$\lvert(\leftsubscript{\mathrm{ul}}{\upX}^{6,n,n})^{-1}\ref{item:comparisonproofcaseLhassize6:twoadditionalnondisjointedgesdisjointfromC4}\rvert = 8 (n-3)\binom{n-3}{2} \xi_n$
\item\label{numberofmatrixrealizationsoftype:fourisolatedvertices}
$\lvert(\leftsubscript{\mathrm{ul}}{\upX}^{6,n,n})^{-1}\ref{item:comparisonproofcaseLhassize6:fourisolatedvertices}\rvert = 2 \binom{n-3}{2} \binom{n-3}{2} \xi_n$
\item\label{numberofmatrixrealizationsoftype:oneadditionaldisjointedgeandtwoisolatedvertices}
$\lvert(\leftsubscript{\mathrm{ul}}{\upX}^{6,n,n})^{-1}\ref{item:comparisonproofcaseLhassize6:oneadditionaldisjointedgeandtwoisolatedvertices}\rvert = 8 \binom{n-3}{2} \binom{n-3}{2} \xi_n$
\item\label{numberofmatrixrealizationsoftype:twoadditionaldisjointedgesdisjointfromC4}
$\lvert(\leftsubscript{\mathrm{ul}}{\upX}^{6,n,n})^{-1}\ref{item:comparisonproofcaseLhassize6:twoadditionaldisjointedgesdisjointfromC4}\rvert = 8 \binom{n-3}{2} \binom{n-3}{2} \xi_n$\quad .
\end{enumerate}
\end{minipage}
}
\end{enumerate}
\end{theorem}
\begin{proof}[Proof of \ref{item:numberofrealizationsofnonforestswhenkequals4}]
We have $\mathrm{X}_B \cong C^4$ if and only if $I$ is a matrix-$4$-circuit and
$\mathrm{Supp}(B) = I$. By Lemma \ref{lem:numberofmatrixcircuitsofgivenlengthwithingivencartesianproduct} there exist $\binom{n-1}{2}^2$ possible matrix-$4$-circuits $I$ and
for each of them there are $2^4$ possibilities for a $B\in \{0,\pm\}^I$
with $\mathrm{Supp}(B)=I$.
\end{proof}
Let us now prepare for the rest of the proof of
Theorem \ref{thm:countingmatrixrealizations} with some observations and definitions.
Inspecting the isomorphism types in $\mathcal{F}^{\mathrm{G}}(6,n)
\setminus \{ \ref{item:comparisonproofcaseLhassize6:XBLisomorphictoK23},
\ref{item:comparisonproofcaseLhassize6:C6} \}$ (the types \ref{item:comparisonproofcaseLhassize6:XBLisomorphictoK23} and \ref{item:comparisonproofcaseLhassize6:C6} are
exceptions whose preimages are also exceptionally easy to count) we see that
in each of them the graph contains exactly one $C^4$. We therefore know that for
every $\mathfrak{X}\in \mathcal{F}^{\mathrm{G}}(6,n)$ (hence in particular for every
$\mathfrak{X}\in \mathcal{F}^{\mathrm{G}}(5,n)$ since
$\mathcal{F}^{\mathrm{G}}(5,n)\subseteq \mathcal{F}^{\mathrm{G}}(6,n)$
by \ref{item:failuresetofisomorphismtypesk6} in Corollary \ref{cor:isomorphismtypesforwhichequalityofmeasuresofentryspecificationeventsfails}), and for every
$I\in \binom{[n-1]^2}{6}$ it is necessary that there exist a matrix-$4$-circuit
$S\subseteq I$ with $B\mid_{S}\in \{\pm\}^S$. For this there are
$2^4\cdot \lvert \mathrm{Cir}(4,n) \rvert$ possibilities. A priori it could be that the
number of possibilities to realize an isomorphism type depends on the choice of
this necessary $S\subseteq I$. However, since we will take this $S$ to be arbitrary
in the proofs to follow, and since we
will get results which do not depend on $S$, it follows as a byproduct that they
are not, more precisely that for each $\mathfrak{X}\in \mathcal{F}^{\mathrm{G}}(6,n)
\setminus \{ \ref{item:comparisonproofcaseLhassize6:C6},
\ref{item:comparisonproofcaseLhassize6:XBLisomorphictoK23} \}$
the values of $\lvert(\leftsubscript{\mathrm{ul}}{\upX}^{k,n,n})^{-1}(\mathfrak{X})\rvert$ are equal to the
product of $2^4\cdot \lvert \mathrm{Cir}(4,n)\rvert$ and the number of possibilities to choose
$B\mid_{I\setminus S}\in \{0,\pm\}^{I\setminus S}$ in such a way that
$\mathrm{X}_{B} \in \mathfrak{X}$. By determining the latter number for each of the
isomorphism types, we will prove all of the formulas
\ref{numberofmatrixrealizationsoftype:oneisolatedvertex:kequals5}---\ref{numberofmatrixrealizationsoftype:twoadditionaldisjointedgesdisjointfromC4}, except, as
already mentioned, \ref{numberofmatrixrealizationsoftype:XBLisomorphictoK23}
and \ref{numberofmatrixrealizationsoftype:C6}, which do not fit into the overall
plan of the proof (in the case of
\ref{numberofmatrixrealizationsoftype:XBLisomorphictoK23} we would be overcounting
the number of realizations since $K^{2,3}$ contains three copies of $C^4$) but which
are easy to count directly.
Let $\prec$ denote the lexicographic ordering on $[n-1]^2$. Throughout the proof,
we use the following conventions: we consider $I\supseteq S\in \binom{[n-1]^2}{4}$
and $B\mid_{S}\in\{\pm\}^S$ to be arbitrary.
We set $\{a,b,c,d\} := S$, $a_1 := \mathrm{p}_1(a)$, $a_2 := \mathrm{p}_2(a)$ and analogously
for $b_1$, $b_2$, $c_1$, $c_2$, $d_1$ and $d_2$. Since $\prec$ is a total
order, we may assume $a\prec b \prec c \prec d$ which combined with the fact
that $S$ is a matrix-$4$-circuit implies $a_1 = b_1$, $c_1=d_1$, $a_2 = c_2$
and $b_2 = d_2$. The cardinality of $I\setminus S$ depends on whether we are proving
formulas from \ref{item:numberofrealizationsofnonforestswhenkequals5} or \ref{item:numberofrealizationsofnonforestswhenkequals6}. In the former case we set
$\{u\}:=I\setminus S$, in the latter $\{u,v\}:=I\setminus S$ with the
assumption that $u\prec v$. Moreover, $u_1:=\mathrm{p}_1(u)$, $u_2:=\mathrm{p}_2(u)$,
$v_1:=\mathrm{p}_1(v)$ and $v_2:=\mathrm{p}_2(v)$. Finally, let us use the abbreviation
$\mathrm{p}(S) := \mathrm{p}_1(S)\sqcup\mathrm{p}_2(S)$.
\begin{proof}[Proof of \ref{item:numberofrealizationsofnonforestswhenkequals5}]
As to \ref{numberofmatrixrealizationsoftype:oneisolatedvertex:kequals5}, we start
by noting that it follows directly from Definition \ref{def:XBandecXB} that
$\mathrm{X}_{B} \in \ref{item:comparisonproofcaseLhassize6:oneisolatedvertex}$
if and only if $B[u]=0$ and
\begin{equation}\label{eq:conditioninproofofnumberofrealizationsoftype:oneisolatedvertex:kequals5}
\lvert \{u_1\}\setminus \mathrm{p}_1(S)\rvert + \lvert \{u_2\}\setminus \mathrm{p}_2(S)\rvert = 1
\quad .
\end{equation}
We distinguish cases according to how \eqref{eq:conditioninproofofnumberofrealizationsoftype:oneisolatedvertex:kequals5} is satisfied.
\begin{enumerate}[label={\rm(C.\ref{numberofmatrixrealizationsoftype:oneisolatedvertex:kequals5}.\arabic{*})}]
\item\label{numberofmatrixrealizationsoftype:oneisolatedvertex:kequals5:case:lvertu1rvertbackslashp1Sequals0} $\lvert \{u_1\}\setminus \mathrm{p}_1(S)\rvert = 0$,
i.e. $u_1\in\mathrm{p}_1(S)$. Then \eqref{eq:conditioninproofofnumberofrealizationsoftype:oneisolatedvertex:kequals5} implies that $u_2\notin \mathrm{p}_2(S)$.
Since there are $2$ different $u_1$ with $u_1\in\mathrm{p}_1(S)$ and for each of them there
are $((n-1)-2)=(n-3)$ different $u_2$ with $u_2\notin\mathrm{p}_2(S)$ it follows that
if \ref{numberofmatrixrealizationsoftype:oneisolatedvertex:kequals5:case:lvertu1rvertbackslashp1Sequals0}, then there are $2(n-3)$ realizations of type \ref{item:comparisonproofcaseLhassize6:oneisolatedvertex} by $B[u]$.
\item\label{numberofmatrixrealizationsoftype:oneisolatedvertex:kequals5:case:lvertu1rvertbackslashp1Sequals1} $\lvert \{u_1\}\setminus \mathrm{p}_1(S)\rvert = 1$. This case is
easily seen to be symmetric to \ref{numberofmatrixrealizationsoftype:oneisolatedvertex:kequals5:case:lvertu1rvertbackslashp1Sequals0} w.r.t. swapping the subscripts $1$ and
$2$. Therefore, if \ref{numberofmatrixrealizationsoftype:oneisolatedvertex:kequals5:case:lvertu1rvertbackslashp1Sequals1}, then there are also $2(n-3)$ realizations of
type \ref{item:comparisonproofcaseLhassize6:oneisolatedvertex} by $B[u]$.
\end{enumerate}
It follows that there are $2(n-3)+2(n-3)=4(n-3)$ realizations of type
\ref{item:comparisonproofcaseLhassize6:oneisolatedvertex} by $B[u]$, proving
\ref{numberofmatrixrealizationsoftype:oneisolatedvertex:kequals5}.
As to \ref{numberofmatrixrealizationsoftype:oneadditionaledgeintersectingC4:kequals5},
this follows from \ref{numberofmatrixrealizationsoftype:oneisolatedvertex:kequals5}
and Lemma \ref{lem:relationsamongthenumbersofmatrixrealizations}.\ref{linearrelationbetweent2andt3:kequals5}.
As to \ref{numberofmatrixrealizationsoftype:twoisolatedvertices:kequals5}, it follows
from Definition \ref{def:XBandecXB} that $\mathrm{X}_{B} \in \ref{item:comparisonproofcaseLhassize6:twoisolatedvertices}$ if and only if $B[u]=0$ and
\begin{equation}\label{eq:conditioninproofofnumberofrealizationsoftype:twoisolatedvertices:kequals5}
\lvert \{u_1\}\setminus \mathrm{p}_1(S)\rvert + \lvert\{u_2\}\setminus \mathrm{p}_2(S)\rvert = 2
\quad .
\end{equation}
Property \eqref{eq:conditioninproofofnumberofrealizationsoftype:twoisolatedvertices:kequals5} is equivalent to $u_1\notin\mathrm{p}_1(S)$ and $u_2\notin\mathrm{p}_2(S)$, and there are
obviously $((n-1)-2)^2=(n-3)^2$ different $u\in [n-1]^2$ satisfying this. Therefore,
\ref{numberofmatrixrealizationsoftype:twoisolatedvertices:kequals5} is correct.
As to \ref{numberofmatrixrealizationsoftype:C4withoneadditionaldisjointedge:kequals5}, this follows from \ref{numberofmatrixrealizationsoftype:twoisolatedvertices:kequals5}
and Lemma \ref{lem:relationsamongthenumbersofmatrixrealizations}.\ref{linearrelationbetweent5andt7:kequals5}. This completes the proof of \ref{item:numberofrealizationsofnonforestswhenkequals5}.
\end{proof}
We now take on the task of proving \ref{item:numberofrealizationsofnonforestswhenkequals6}, which will take some effort. We prepare by proving four lemmas characterizing
the realizations of the types \ref{item:comparisonproofcaseLhassize6:C4intersectingtwoedgesinseparatenonadjacentvertices}--\ref{item:comparisonproofcaseLhassize6:C4intersectingatwopathinitsinnervertex}.
\begin{lemma}\label{lem:characterizationofisomorphismtype:C4intersectingtwoedgesinseparatenonadjacentvertices}
For every $B\in \{0,\pm\}^I$ with $I\in \binom{[n-1]^2}{6}$,
$I = S \sqcup \{u,v\}$ and $\mathrm{X}_{B\mid_{S}}\cong C^4$ we have $\mathrm{X}_{B} \cong \ref{item:comparisonproofcaseLhassize6:C4intersectingtwoedgesinseparatenonadjacentvertices}$ if and only if
{\scriptsize
\begin{minipage}[b]{0.35\linewidth}
\begin{enumerate}[label={\rm(P.\ref{item:comparisonproofcaseLhassize6:C4intersectingtwoedgesinseparatenonadjacentvertices}.\arabic{*})}]
\item\label{characterizationoftype:C4intersectingtwoedgesinseparatenonadjacentvertices:property1}
$\mathrm{X}_{\{0\}^{\{u,v\}} \sqcup B\mid_{S}} \cong
\ref{item:comparisonproofcaseLhassize6:twoisolatedvertices}$,
\item\label{characterizationoftype:C4intersectingtwoedgesinseparatenonadjacentvertices:property2}
$B[u]\in\{\pm\}$ and $B[v]\in \{\pm\}$,
\end{enumerate}
\end{minipage}
\begin{minipage}[b]{0.65\linewidth}
\begin{enumerate}[label={\rm(P.\ref{item:comparisonproofcaseLhassize6:C4intersectingtwoedgesinseparatenonadjacentvertices}.\arabic{*})},start=3]
\item\label{characterizationoftype:C4intersectingtwoedgesinseparatenonadjacentvertices:property3}
$\{u_1,u_2\}\cap\{v_1,v_2\} = \emptyset$,
\item\label{characterizationoftype:C4intersectingtwoedgesinseparatenonadjacentvertices:property4}
($u_1\in\mathrm{p}_1(S)$ and $v_1\in\mathrm{p}_1(S)$)
or
($u_2\in\mathrm{p}_2(S)$ and $v_2\in\mathrm{p}_2(S)$).
\end{enumerate}
\end{minipage}
}
\end{lemma}
\begin{proof}
First suppose that $\mathrm{X}_{B} \cong \ref{item:comparisonproofcaseLhassize6:C4intersectingtwoedgesinseparatenonadjacentvertices}$. Then Definition \ref{def:XBandecXB}
implies that both \ref{characterizationoftype:C4intersectingtwoedgesinseparatenonadjacentvertices:property1} and \ref{characterizationoftype:C4intersectingtwoedgesinseparatenonadjacentvertices:property2} are true. To prove \ref{characterizationoftype:C4intersectingtwoedgesinseparatenonadjacentvertices:property3} and
\ref{characterizationoftype:C4intersectingtwoedgesinseparatenonadjacentvertices:property4}, let $e\neq f\in \mathrm{E}(\mathrm{X}_{B})$ denote the two edges in $\mathrm{X}_{B\mid_{\{u,v\}}}$,
where $\{e\}$ $:=$ $\mathrm{E}(\mathrm{X}_{B[u]})$ $=$ $\bigl\{\{(u_1,n), (n,u_2)\}\bigr\}$ and
$\{f\}$ $:=$ $\mathrm{E}(\mathrm{X}_{B[v]})$ $=$ $\bigl\{ \{(v_1,n), (n,v_2)\}\bigr\}$. By
hypothesis, $e\cap f = \emptyset$ and this implies that \ref{characterizationoftype:C4intersectingtwoedgesinseparatenonadjacentvertices:property3} is true. Moreover, again
by hypothesis, both $e$ and $f$ intersect $\mathrm{X}_{B\mid_S}\cong C^4$ and the
intersection set is \emph{not} an edge of it.
If $u_1\in\mathrm{p}_1(S)$, then there are still two possibilities for the intersection set
$f\cap \mathrm{V}(\mathrm{X}_{B\mid_S})$, namely $f\cap \mathrm{V}(\mathrm{X}_{B\mid_S})=\{(v_1,n)\}$
(equivalently, $v_1\in\mathrm{p}_1(S)$) or
$f\cap \mathrm{V}(\mathrm{X}_{B\mid_S}) = \{(n,v_2)\}$ (equivalently, $v_2\in\mathrm{p}_2(S)$). It follows from
Definition \ref{def:XBandecXB} that the vertex in the intersection set
$e\cap \mathrm{V}(\mathrm{X}_{B\mid_S}) = \{ (u_1,n)\}$ is \emph{not} adjacent to the vertex in $f\cap \mathrm{V}(\mathrm{X}_{B\mid_S})$
if and only if the first possibility is true, i.e. $f\cap \mathrm{V}(\mathrm{X}_{B\mid_S}) = \{(v_1,n)\}$, i.e.
$v_1\in\mathrm{p}_1(S)$. This proves that the first
clause of \ref{characterizationoftype:C4intersectingtwoedgesinseparatenonadjacentvertices:property4},
and hence \ref{characterizationoftype:C4intersectingtwoedgesinseparatenonadjacentvertices:property4} itself, is true.
If $u_2\in\mathrm{p}_2(S)$, then an entirely analogous argument as the one in the preceding
paragraph shows that the second clause of \ref{characterizationoftype:C4intersectingtwoedgesinseparatenonadjacentvertices:property4},
hence again \ref{characterizationoftype:C4intersectingtwoedgesinseparatenonadjacentvertices:property4} itself, is true. This proves
that $\mathrm{X}_{B} \cong \ref{item:comparisonproofcaseLhassize6:C4intersectingtwoedgesinseparatenonadjacentvertices}$ implies that
\ref{characterizationoftype:C4intersectingtwoedgesinseparatenonadjacentvertices:property1}--\ref{characterizationoftype:C4intersectingtwoedgesinseparatenonadjacentvertices:property4}
are true.
Conversely, assume \ref{characterizationoftype:C4intersectingtwoedgesinseparatenonadjacentvertices:property1}--\ref{characterizationoftype:C4intersectingtwoedgesinseparatenonadjacentvertices:property4}. Then \ref{characterizationoftype:C4intersectingtwoedgesinseparatenonadjacentvertices:property2} implies
that $f_1( \mathrm{X}_{B}) = 6$ and \ref{characterizationoftype:C4intersectingtwoedgesinseparatenonadjacentvertices:property3} implies that the two edges in $\mathrm{E}(\mathrm{X}_{B\mid_{\{u,v\}}})$ do
not intersect. Let $e$ and $f$ be defined as in the preceding proof of the other
implication. It remains to show that
$(e\cap \mathrm{V}(\mathrm{X}_{B\mid_{S}}))\cup (f\cap \mathrm{V}(\mathrm{X}_{B\mid_{S}})) \notin \mathrm{E}(\mathrm{X}_{B\mid_{S}})$. By
definition of $e$, either
$e\cap \mathrm{V}(\mathrm{X}_{B\mid_{S}}) = \{ (u_1,n)\}$ or $e\cap \mathrm{V}(\mathrm{X}_{B\mid_{S}}) = \{(n,u_2)\}$.
In the former case we have $u_1\in\mathrm{p}_1(S)$, hence the first clause
of \ref{characterizationoftype:C4intersectingtwoedgesinseparatenonadjacentvertices:property4} implies $v_1\in\mathrm{p}_1(S)$, hence $(v_1,n)\in f\cap \mathrm{V}(\mathrm{X}_{B})$ by
definition of $f$, hence $f\cap \mathrm{V}( \mathrm{X}_{B}) = \{(v_1,n)\}$ since
$f\cap \mathrm{V}( \mathrm{X}_{B})$ is a singleton by construction. In view of
Definition \ref{def:XBandecXB} this implies that indeed $(e\cap \mathrm{V}(\mathrm{X}_{B\mid_{S}}))\cup (f\cap \mathrm{V}(\mathrm{X}_{B\mid_{S}})) = \{ (u_1,n) , (v_1,n)\}\notin \mathrm{E}(\mathrm{X}_{B\mid_{S}})$.
In the latter case we have $u_2\in\mathrm{p}_2(S)$, hence the second clause
of \ref{characterizationoftype:C4intersectingtwoedgesinseparatenonadjacentvertices:property4} implies $v_2\in\mathrm{p}_2(S)$, hence $(n,v_2)\in f\cap \mathrm{V}(\mathrm{X}_{B})$ by
definition of $f$, hence $f\cap \mathrm{V}( \mathrm{X}_{B}) = \{(n,v_2)\}$ since
$f\cap \mathrm{V}( \mathrm{X}_{B})$ is a singleton by construction. In view of
Definition \ref{def:XBandecXB} this implies that indeed $(e\cap \mathrm{V}(\mathrm{X}_{B\mid_{S}}))\cup (f\cap \mathrm{V}(\mathrm{X}_{B\mid_{S}})) = \{ (n,u_2) , (n,v_2)\}\notin \mathrm{E}(\mathrm{X}_{B\mid_{S}})$.
This completes the proof that \ref{characterizationoftype:C4intersectingtwoedgesinseparatenonadjacentvertices:property1}--\ref{characterizationoftype:C4intersectingtwoedgesinseparatenonadjacentvertices:property4} imply $\mathrm{X}_{B} \cong \ref{item:comparisonproofcaseLhassize6:C4intersectingtwoedgesinseparatenonadjacentvertices}$.
\end{proof}
\begin{lemma}\label{lem:characterizationofisomorphismtype:C4intersectingtwoedgesinseparateadjacentvertices}
For every $B\in \{0,\pm\}^I$ with $I\in\binom{[n-1]^2}{6}$,
$I = S \sqcup \{u,v\}$ and $\mathrm{X}_{B\mid_{S}}\cong C^4$ we have $\mathrm{X}_{B} \cong \ref{item:comparisonproofcaseLhassize6:C4intersectingtwoedgesinseparateadjacentvertices}$
if and only if
{\scriptsize
\begin{minipage}[b]{0.35\linewidth}
\begin{enumerate}[label={\rm(P.\ref{item:comparisonproofcaseLhassize6:C4intersectingtwoedgesinseparateadjacentvertices}.\arabic{*})}]
\item\label{characterizationoftype:C4intersectingtwoedgesinseparateadjacentvertices:property1}
$\mathrm{X}_{\{0\}^{\{u,v\}} \sqcup B\mid_{S}} \cong
\ref{item:comparisonproofcaseLhassize6:twoisolatedvertices}$,
\item\label{characterizationoftype:C4intersectingtwoedgesinseparateadjacentvertices:property2}
$B[u]\in\{\pm\}$ and $B[v]\in \{\pm\}$,
\end{enumerate}
\end{minipage}
\begin{minipage}[b]{0.65\linewidth}
\begin{enumerate}[label={\rm(P.\ref{item:comparisonproofcaseLhassize6:C4intersectingtwoedgesinseparateadjacentvertices}.\arabic{*})},start=3]
\item\label{characterizationoftype:C4intersectingtwoedgesinseparateadjacentvertices:property3}
$\{u_1,u_2\}\cap\{v_1,v_2\} = \emptyset$,
\item\label{characterizationoftype:C4intersectingtwoedgesinseparateadjacentvertices:property4}
($u_1\in\mathrm{p}_1(S)$ and $v_2\in\mathrm{p}_2(S)$)
or
($u_2\in\mathrm{p}_2(S)$ and $v_1\in\mathrm{p}_1(S)$).
\end{enumerate}
\end{minipage}
}
\end{lemma}
\begin{proof}
First suppose that $\mathrm{X}_{B}\cong \ref{item:comparisonproofcaseLhassize6:C4intersectingtwoedgesinseparateadjacentvertices}$. Then Definition \ref{def:XBandecXB} implies
that both \ref{characterizationoftype:C4intersectingtwoedgesinseparateadjacentvertices:property1} and \ref{characterizationoftype:C4intersectingtwoedgesinseparateadjacentvertices:property2} are true.
To prove \ref{characterizationoftype:C4intersectingtwoedgesinseparateadjacentvertices:property3} and
\ref{characterizationoftype:C4intersectingtwoedgesinseparateadjacentvertices:property4}, let $e\neq f\in \mathrm{E}(\mathrm{X}_{B})$ denote the two edges in $\mathrm{X}_{B\mid_{\{u,v\}}}$, where
$\{e\}$ $:=$ $\mathrm{E}(\mathrm{X}_{B[u]})$ $=$ $\bigl\{\{(u_1,n), (n,u_2)\}\bigr\}$ and
$\{f\}$ $:=$ $\mathrm{E}(\mathrm{X}_{B[v]})$ $=$ $\bigl\{ \{(v_1,n), (n,v_2)\}\bigr\}$.
By hypothesis $e\cap f = \emptyset$ and this implies that \ref{characterizationoftype:C4intersectingtwoedgesinseparateadjacentvertices:property3} is true. Moreover, again
by hypothesis, both $e$ and $f$ intersect $\mathrm{X}_{B\mid_S}$ $\cong$ $C^4$ and the
intersection set is an edge of it.
If $u_1\in \mathrm{p}_1(S)$, then there are still two possibilities for the intersection
set $f\cap \mathrm{V}(\mathrm{X}_{B\mid_S})$, namely $f\cap \mathrm{V}(\mathrm{X}_{B\mid_S}) = \{(v_1,n)\}$
(equivalently, $v_1\in \mathrm{p}_1(S)$) or $f\cap \mathrm{V}(\mathrm{X}_{B\mid_S}) = \{(n,v_2)\}$
(equivalently, $v_2\in \mathrm{p}_2(S)$). It is evident from Definition \ref{def:XBandecXB}
that the vertex in the intersection set $e\cap \mathrm{V}(\mathrm{X}_{B\mid_S}) = \{(u_1,n)\}$ is
adjacent to the vertex in $f\cap \mathrm{V}(\mathrm{X}_{B\mid_S})$ if and only if the second
possibility is true, i.e. $f\cap \mathrm{V}(\mathrm{X}_{B\mid_S}) = \{(n,v_2)\}$,
i.e. $v_2\in\mathrm{p}_2(S)$. This proves that the first clause of \ref{characterizationoftype:C4intersectingtwoedgesinseparateadjacentvertices:property4}, and hence
\ref{characterizationoftype:C4intersectingtwoedgesinseparateadjacentvertices:property4} itself, is true.
If $u_2\in \mathrm{p}_2(S)$, an entirely analogous argument as the one in the
preceding paragraph shows that then the second clause
of \ref{characterizationoftype:C4intersectingtwoedgesinseparateadjacentvertices:property4}, hence again \ref{characterizationoftype:C4intersectingtwoedgesinseparateadjacentvertices:property4} itself is true. This completes the proof that $\mathrm{X}_{B} \cong \ref{item:comparisonproofcaseLhassize6:C4intersectingtwoedgesinseparateadjacentvertices}$
implies properties \ref{characterizationoftype:C4intersectingtwoedgesinseparateadjacentvertices:property1}--\ref{characterizationoftype:C4intersectingtwoedgesinseparateadjacentvertices:property4}.
Conversely, suppose that \ref{characterizationoftype:C4intersectingtwoedgesinseparateadjacentvertices:property1}--\ref{characterizationoftype:C4intersectingtwoedgesinseparateadjacentvertices:property4} are true. Then \ref{characterizationoftype:C4intersectingtwoedgesinseparateadjacentvertices:property2} implies
that $f_1( \mathrm{X}_{B} ) = 6$ and \ref{characterizationoftype:C4intersectingtwoedgesinseparateadjacentvertices:property3} implies that the two edges in
$\mathrm{E}(\mathrm{X}_{B\mid_{\{u,v\}}})$ do not intersect. Let $e$ and $f$ be defined as in the
preceding proof of the other implication. It remains to show that
$(e\cap \mathrm{V}(\mathrm{X}_{B\mid_{S}})) \cup (f\cap \mathrm{V}(\mathrm{X}_{B\mid_{S}})) \in \mathrm{E}(\mathrm{X}_{B\mid_{S}})$. By definition of $e$, either $e\cap \mathrm{V}(\mathrm{X}_{B\mid_{S}}) = \{(u_1,n)\}$ or
$e\cap \mathrm{V}(\mathrm{X}_{B\mid_{S}}) = \{(n,u_2)\}$.
In the former case we have $u_1\in \mathrm{p}_1(S)$, hence
the first clause of \ref{characterizationoftype:C4intersectingtwoedgesinseparateadjacentvertices:property4} implies that $v_2\in \mathrm{p}_2(S)$, hence
$(n,v_2) \in f\cap \mathrm{V}(\mathrm{X}_{B\mid_{S}})$ by definition of $f$, hence
$f\cap \mathrm{V}(\mathrm{X}_{B\mid_{S}}) = \{ (n,v_2) \}$ since $f\cap \mathrm{V}(\mathrm{X}_{B\mid_{S}})$ is a
singleton by construction. In view of Definition \ref{def:XBandecXB} this implies
that indeed $(e\cap \mathrm{V}(\mathrm{X}_{B\mid_{S}})) \cup (f\cap \mathrm{V}(\mathrm{X}_{B\mid_{S}})) =
\{ (u_1,n), (n,v_2)\} \in \mathrm{E}(\mathrm{X}_{B\mid_{S}})$.
In the latter case we have $u_2\in \mathrm{p}_2(S)$, hence the second clause of
\ref{characterizationoftype:C4intersectingtwoedgesinseparateadjacentvertices:property4}
implies that $v_1\in \mathrm{p}_1(S)$, hence $(v_1,n) \in f\cap \mathrm{V}(\mathrm{X}_{B\mid_{S}})$ by
definition of $f$, hence $f\cap \mathrm{V}(\mathrm{X}_{B\mid_{S}}) = \{ (v_1,n) \}$ since
$f\cap \mathrm{V}(\mathrm{X}_{B\mid_{S}})$ is a singleton by construction. In view of
Definition \ref{def:XBandecXB} this implies that indeed
$(e\cap \mathrm{V}(\mathrm{X}_{B\mid_{S}})) \cup (f\cap \mathrm{V}(\mathrm{X}_{B\mid_{S}})) = \{ (n,u_2), (v_1,n)\} \in \mathrm{E}(\mathrm{X}_{B\mid_{S}})$. This completes the proof that \ref{characterizationoftype:C4intersectingtwoedgesinseparateadjacentvertices:property1}--\ref{characterizationoftype:C4intersectingtwoedgesinseparateadjacentvertices:property4} imply $\mathrm{X}_{B} \cong \ref{item:comparisonproofcaseLhassize6:C4intersectingtwoedgesinseparateadjacentvertices}$.
\end{proof}
\begin{lemma}\label{lem:characterizationofisomorphismtype:C4intersectingatwopathinanendvertex}
For every $B\in \{0,\pm\}^I$ with $I\in\binom{[n-1]^2}{6}$,
$I = S \sqcup \{u,v\}$ and $\mathrm{X}_{B\mid_{S}}\cong C^4$ we have $\mathrm{X}_{B} \cong \ref{item:comparisonproofcaseLhassize6:C4intersectingatwopathinanendvertex}$ if and only if
{\scriptsize
\begin{minipage}[b]{0.35\linewidth}
\begin{enumerate}[label={\rm(P.\ref{item:comparisonproofcaseLhassize6:C4intersectingatwopathinanendvertex}.\arabic{*})}]
\item\label{characterizationoftype:C4intersectingatwopathinanendvertex:property1}
$\mathrm{X}_{\{0\}^{\{u,v\}} \sqcup B\mid_{S}} \cong
\ref{item:comparisonproofcaseLhassize6:twoisolatedvertices}$,
\item\label{characterizationoftype:C4intersectingatwopathinanendvertex:property2}
$B[u]\in\{\pm\}$ and $B[v]\in \{\pm\}$,
\end{enumerate}
\end{minipage}
\begin{minipage}[b]{0.65\linewidth}
\begin{enumerate}[label={\rm(P.\ref{item:comparisonproofcaseLhassize6:C4intersectingatwopathinanendvertex}.\arabic{*})},start=3]
\item\label{characterizationoftype:C4intersectingatwopathinanendvertex:property3}
$\{u_1,u_2\}\cap\{v_1,v_2\} \neq \emptyset$,
\item\label{characterizationoftype:C4intersectingatwopathinanendvertex:property4}
$\{u_1,u_2\}\cap \mathrm{p}(S) = \emptyset$
or $\{v_1,v_2\}\cap \mathrm{p}(S) = \emptyset$.
\end{enumerate}
\end{minipage}
}
\end{lemma}
\begin{proof}
First suppose that $\mathrm{X}_{B}\cong \ref{item:comparisonproofcaseLhassize6:C4intersectingatwopathinanendvertex}$. Then Definition \ref{def:XBandecXB} implies that both \ref{characterizationoftype:C4intersectingatwopathinanendvertex:property1} and \ref{characterizationoftype:C4intersectingatwopathinanendvertex:property2} are true. To prove
\ref{characterizationoftype:C4intersectingatwopathinanendvertex:property3} and \ref{characterizationoftype:C4intersectingatwopathinanendvertex:property4},
let $e\in \mathrm{E}(\mathrm{X}_{B})$ denote the unique edge which does not intersect
$\mathrm{X}_{B\mid_S}\cong C^4$, and let $f\in \mathrm{E}(\mathrm{X}_{B})$ denote the unique edge
which intersects $\mathrm{X}_{B\mid_S}\cong C^4$.
In view of Definition \ref{def:XBandecXB},
{\scriptsize
\begin{equation}\label{eq:propertyinthepreparationofanalysisoftypet14}
\text{($e = \{ (u_1,n), (n,u_2) \} $ and $f=\{(v_1,n),(n,v_2)\}$)
or
($e = \{(v_1,n), (n,v_2)\}$ and $f=\{(u_1,n),(n,u_2)\}$).
}
\end{equation}
}
By definition of $e$ and $f$ we have $e\cap f\neq \emptyset$, hence whatever of the
two clauses of \eqref{eq:propertyinthepreparationofanalysisoftypet14} is true, either
$u_1=v_1$ or $u_2=v_2$. Therefore property \ref{characterizationoftype:C4intersectingatwopathinanendvertex:property3}
is true.
By definition of $e$, if $e = \{ (u_1,n), (n,u_2)\}$, then $u_1\notin\mathrm{p}_1(S)$
and $u_2\notin\mathrm{p}_2(S)$, hence the first clause
of \ref{characterizationoftype:C4intersectingatwopathinanendvertex:property4} is true, and if $e=\{(v_1,n),(n,v_2)\}$,
then $v_1\notin \mathrm{p}_1(S)$ and $v_2\notin\mathrm{p}_2(S)$, hence then the second clause of
\ref{characterizationoftype:C4intersectingatwopathinanendvertex:property4} is true.
Conversely, suppose that properties \ref{characterizationoftype:C4intersectingatwopathinanendvertex:property1}--\ref{characterizationoftype:C4intersectingatwopathinanendvertex:property4} are true.
Then \ref{characterizationoftype:C4intersectingatwopathinanendvertex:property2}
implies $f_1( \mathrm{X}_{B}) = 6$ and \ref{characterizationoftype:C4intersectingatwopathinanendvertex:property3}
implies that the two edges corresponding to $u\neq v$ intersect.
It remains to prove that the $2$-path consisting of these edges intersects
$\mathrm{X}_{B\mid_S} \cong C^4$ with one of its endvertices. To see this, note first
that \ref{characterizationoftype:C4intersectingatwopathinanendvertex:property1} implies \eqref{eq:proofofnumberofeventsinducing:item:comparisonproofcaseLhassize6:twoisolatedvertices} which combined with $u\neq v$ implies that
\begin{equation}\label{eq:preparationofanalysisoftypet14:the2pathintersectstheC4}
\{u_1,u_2,v_1,v_2\}\cap \mathrm{p}(S) \neq \emptyset\quad .
\end{equation}
Moreover, we know from \eqref{eq:propertyinthepreparationofanalysisoftypet14} together
with $e\cap f\neq \emptyset$ that $u_1=v_1$ or $u_2=v_2$.
If $u_1=v_1$, then \eqref{eq:preparationofanalysisoftypet14:the2pathintersectstheC4}
cannot be true by virtue of $u_1=v_1\in \mathrm{p}_1(S)$ since then \ref{characterizationoftype:C4intersectingatwopathinanendvertex:property4} would become false. Therefore, if $u_1=v_1$, then
$u_1=v_1\notin\mathrm{p}_1(S)$, and \eqref{eq:preparationofanalysisoftypet14:the2pathintersectstheC4} implies $\{u_2,v_2\}\cap\mathrm{p}_2(S) \neq \emptyset$. It is impossible
that $\{u_2,v_2\}\subseteq \mathrm{p}_2(S)$ for this combined with $u_1=v_1$ would imply
$\lvert \{u_1,v_1\}\setminus\mathrm{p}_1(S)\rvert
+ \lvert \{u_2,v_2\}\setminus\mathrm{p}_2(S)\rvert = 1$, hence
contradict \eqref{eq:proofofnumberofeventsinducing:item:comparisonproofcaseLhassize6:twoisolatedvertices}. Therefore, $\lvert \{u_2,v_2\}\cap\mathrm{p}_2(S)\rvert = 1$, and
since $u_1=v_1$ implies $u_2\neq v_2$, this is what we wanted to prove: exactly one
of the two endvertices $(n,u_2), (n,v_2)\in \mathrm{V}(\mathrm{X}_{B})$ intersects
$\mathrm{X}_{B\mid_S}\cong C^4$.
If $u_2=v_2$, then an entirely analogous argument as in the preceding paragraph
shows that exactly one of the two endvertices $(u_1,n), (v_1,n)\in \mathrm{V}(\mathrm{X}_{B})$
intersects the $\mathrm{X}_{B\mid_S}\cong C^4$.
The proof that properties \ref{characterizationoftype:C4intersectingatwopathinanendvertex:property1}--\ref{characterizationoftype:C4intersectingatwopathinanendvertex:property4} imply $\mathrm{X}_{B} \cong \ref{item:comparisonproofcaseLhassize6:C4intersectingatwopathinanendvertex}$ is now complete.
\end{proof}
\begin{lemma}\label{lem:characterizationofisomorphismtype:C4intersectingatwopathinitsinnervertex}
For every $B\in \{0,\pm\}^I$ with $I\in \binom{[n-1]^2}{6}$,
$I = S \sqcup \{u,v\}$ and $\mathrm{X}_{B\mid_{S}}\cong C^4$ we have
$\mathrm{X}_{B} \cong \ref{item:comparisonproofcaseLhassize6:C4intersectingatwopathinitsinnervertex}$ if and only if
{\scriptsize
\begin{minipage}[b]{0.35\linewidth}
\begin{enumerate}[label={\rm(P.\ref{item:comparisonproofcaseLhassize6:C4intersectingatwopathinitsinnervertex}.\arabic{*})}]
\item\label{characterizationoftype:C4intersectingatwopathinitsinnervertex:property1}
$\mathrm{X}_{\{0\}^{\{u,v\}} \sqcup B\mid_{S}} \cong
\ref{item:comparisonproofcaseLhassize6:twoisolatedvertices}$,
\item\label{characterizationoftype:C4intersectingatwopathinitsinnervertex:property2}
$B[u]\in\{\pm\}$ and $B[v]\in \{\pm\}$,
\end{enumerate}
\end{minipage}
\begin{minipage}[b]{0.65\linewidth}
\begin{enumerate}[label={\rm(P.\ref{item:comparisonproofcaseLhassize6:C4intersectingatwopathinitsinnervertex}.\arabic{*})}, start=3]
\item\label{characterizationoftype:C4intersectingatwopathinitsinnervertex:property3}
$\{u_1,u_2\}\cap\{v_1,v_2\} \neq \emptyset$,
\item\label{characterizationoftype:C4intersectingatwopathinitsinnervertex:property4}
$\{u_1,u_2\}\cap \mathrm{p}(S)\neq\emptyset$ and $\{v_1,v_2\}\cap \mathrm{p}(S) \neq \emptyset$.
\end{enumerate}
\end{minipage}
}
\end{lemma}
\begin{proof}
First suppose that $\mathrm{X}_{B}\cong \ref{item:comparisonproofcaseLhassize6:C4intersectingatwopathinitsinnervertex}$. Then Definition \ref{def:XBandecXB} implies that
both \ref{characterizationoftype:C4intersectingatwopathinitsinnervertex:property1}
and \ref{characterizationoftype:C4intersectingatwopathinitsinnervertex:property2}
are true. To prove
\ref{characterizationoftype:C4intersectingatwopathinitsinnervertex:property3} and
\ref{characterizationoftype:C4intersectingatwopathinitsinnervertex:property4}, let
$e\neq f\in \mathrm{E}(\mathrm{X}_{B})$ denote the two edges in $\mathrm{E}(\mathrm{X}_{B})$ forming
the $2$-path which intersects $\mathrm{X}_{B\mid_S}\cong C^4$ with its inner vertex.
As in the proof of Lemma \ref{lem:characterizationofisomorphismtype:C4intersectingatwopathinanendvertex}, we know that \eqref{eq:propertyinthepreparationofanalysisoftypet14} is true. By definition of $e$ and $f$ we have $e\cap f\neq \emptyset$, hence
whatever of the two clauses of \eqref{eq:propertyinthepreparationofanalysisoftypet14} is true, either $u_1=v_1$ or $u_2=v_2$. Therefore property
\ref{characterizationoftype:C4intersectingatwopathinitsinnervertex:property3} is
true. By definition of $e$ and $f$, both $e$ and $f$ intersect
$\mathrm{X}_{B\mid_S}$ $\cong$ $C^4$. If the first clause
in \eqref{eq:propertyinthepreparationofanalysisoftypet14} is true then
$e$ intersecting $\mathrm{X}_{B\mid_S}\cong C^4$ is equivalent
to ($u_1\in \mathrm{p}_1(S)$ or $u_2\in \mathrm{p}_2(S)$)
and $f$ intersecting $\mathrm{X}_{B\mid_S}\cong C^4$ is equivalent
to ($v_1\in \mathrm{p}_1(S)$ or $v_2\in \mathrm{p}_2(S)$). Then \ref{characterizationoftype:C4intersectingatwopathinitsinnervertex:property4} is indeed true. If the second clause
in \eqref{eq:propertyinthepreparationofanalysisoftypet14} is true,
interchanging `$u$' and `$v$' in the preceding sentence shows that then
\ref{characterizationoftype:C4intersectingatwopathinitsinnervertex:property4} is true as well. This completes the proof that $\mathrm{X}_{B}\cong \ref{item:comparisonproofcaseLhassize6:C4intersectingatwopathinitsinnervertex}$ implies
\ref{characterizationoftype:C4intersectingatwopathinitsinnervertex:property1}--\ref{characterizationoftype:C4intersectingatwopathinitsinnervertex:property4}.
Conversely, suppose that properties \ref{characterizationoftype:C4intersectingatwopathinitsinnervertex:property1}--\ref{characterizationoftype:C4intersectingatwopathinitsinnervertex:property4} are true.
Then \ref{characterizationoftype:C4intersectingatwopathinitsinnervertex:property2}
implies $f_1( \mathrm{X}_{B}) = 6$ and \ref{characterizationoftype:C4intersectingatwopathinitsinnervertex:property3} implies that the two edges corresponding to $u\neq v$
intersect. It remains to prove that the $2$-path consisting of these edges
intersects $\mathrm{X}_{B\mid_S}\cong C^4$ with its inner vertex. Similar to the proof of
Lemma \ref{lem:characterizationofisomorphismtype:C4intersectingatwopathinanendvertex}
we know that \eqref{eq:preparationofanalysisoftypet14:the2pathintersectstheC4} and
that $u_1=v_1$ or $u_2=v_2$.
If $u_1=v_1$, then $u_2\in \mathrm{p}_2(S)$ is impossible since this together
with $u\neq v$ would imply $u_1=v_1\notin \mathrm{p}_1(S)$ which due to the second
clause of \ref{characterizationoftype:C4intersectingatwopathinitsinnervertex:property4} would imply $v_2\in\mathrm{p}_2(S)$; but
$u_2\in\mathrm{p}_2(S)$, $u_1=v_1$ and $v_2\in \mathrm{p}_2(S)$ combined imply
$\lvert \{u_1,v_1\}\setminus\mathrm{p}_1(S)\rvert + \lvert \{u_2,v_2\}\setminus \mathrm{p}_2(S)\rvert =1$, a contradiction to \eqref{eq:proofofnumberofeventsinducing:item:comparisonproofcaseLhassize6:twoisolatedvertices}. For an entirely analogous reason
$v_2\in \mathrm{p}_2(S)$ is impossible, too. Since both $u_2\notin\mathrm{p}_2(S)$
and $v_2\notin\mathrm{p}_2(S)$, it follows from
\eqref{eq:preparationofanalysisoftypet14:the2pathintersectstheC4} that
$u_1=v_1\in \mathrm{p}_1(S)$. This is what we wanted to prove: the common
vertex $(u_1,n)=(v_1,n)$ of $e$ and $f$ (i.e. the inner vertex
of the $2$-path formed by $e$ and $f$) is also the unique vertex of intersection
with $\mathrm{X}_{B\mid_S}\cong C^4$.
If $u_2=v_2$, then an entirely analogous argumentation as in the preceding
paragraph shows that the common vertex $(n,u_2)=(n,v_2)$ of $e$ and $f$ (i.e. the
inner vertex of the $2$-path which is formed by $e$ and $f$) is also the unique
vertex of intersection with $\mathrm{X}_{B\mid_S}\cong C^4$.
The proof that properties \ref{characterizationoftype:C4intersectingatwopathinitsinnervertex:property1}--\ref{characterizationoftype:C4intersectingatwopathinitsinnervertex:property4} imply $\mathrm{X}_{B} \cong \ref{item:comparisonproofcaseLhassize6:C4intersectingatwopathinitsinnervertex}$ is now complete.
\end{proof}
\begin{proof}[Proof of \ref{item:numberofrealizationsofnonforestswhenkequals6}]
As to \ref{numberofmatrixrealizationsoftype:oneisolatedvertex}, it follows
from Definition \ref{def:XBandecXB} that
$\mathrm{X}_{B} \cong \ref{item:comparisonproofcaseLhassize6:oneisolatedvertex}$
if and only if
\begin{equation}\label{eq:proofofnumberofeventsinducing:item:comparisonproofcaseLhassize6:oneisolatedvertex}
\lvert \{u_1,v_1\}\setminus\mathrm{p}_1(S) \rvert +
\lvert\{ u_2,v_2 \} \setminus \mathrm{p}_2(S) \rvert = 1 \quad .
\end{equation}
\begin{enumerate}[label={\rm(C.\ref{numberofmatrixrealizationsoftype:oneisolatedvertex}.\arabic{*})}]
\item\label{numberofmatrixrealizationsoftype:oneisolatedvertex:kequals6:case:lvertu1v1rvertbackslashp1Sequals0} $\lvert\{u_1,v_1\}\setminus\mathrm{p}_1(S)\rvert = 0$.
Then \eqref{eq:proofofnumberofeventsinducing:item:comparisonproofcaseLhassize6:oneisolatedvertex} implies that $\lvert\{u_2,v_2\}\setminus\mathrm{p}_2(S)\rvert = 1$ which is
equivalent to \eqref{eq:proofofcomparativecountingtheoremcase3secondeq}. The
property defining Case 1 is equivalent to $\{u_1,v_1\}\subseteq\mathrm{p}_1(S)$. Property
\eqref{eq:proofofcomparativecountingtheoremcase3secondeq} implies two cases:
\begin{enumerate}[label={\rm(\arabic{*})}]
\item\label{numberofmatrixrealizationsoftype:twoisolatedvertices:kequals6:case:lvertu1v1rvertbackslashp1Sequals0:case:u2equalsv2} $u_2=v_2$ and
$\{u_2,v_2\}\cap\mathrm{p}_2(S)=\emptyset$. Then $u_2=v_2$ and $u\prec v$ imply that
$u_1 < v_1$. This together with $\{u_1,v_1\} \subseteq\mathrm{p}_1(S)$ implies
$u_1=a_1=c_1$ and $v_1=b_1=d_1$. Therefore, it is $u_2=v_2$ alone which determines
the two pairs $u$ and $v$. The property $u_2=v_2$ and
$\{u_2,v_2\}\cap\mathrm{p}_2(S)=\emptyset$ is equivalent to $u_2=v_2\notin \mathrm{p}_2(S)$. It
follows that if \ref{numberofmatrixrealizationsoftype:oneisolatedvertex:kequals6:case:lvertu1v1rvertbackslashp1Sequals0}.\ref{numberofmatrixrealizationsoftype:twoisolatedvertices:kequals6:case:lvertu1v1rvertbackslashp1Sequals0:case:u2equalsv2}, then there
are $(n-1)-2$ realizations of type \ref{item:comparisonproofcaseLhassize6:oneisolatedvertex} by $u$ and $v$.
\item\label{numberofmatrixrealizationsoftype:twoisolatedvertices:kequals6:case:lvertu1v1rvertbackslashp1Sequals0:case:u2notequaltov2} $u_2\neq v_2$ and
$\lvert \{u_2,v_2\}\cap\mathrm{p}_2(S) \rvert = 1$. Then either $u_2\in \mathrm{p}_2(S)$ or
$v_2\in \mathrm{p}_2(S)$. If $u_2\in \mathrm{p}_2(S)$, then because of
$\{u_1,v_1\}\subseteq\mathrm{p}_1(S)$ it follows that $u\in \{a,b,c,d\}$, a contradiction
to $I\setminus S = \{ u, v\}$. Similarly, if $v_2\in \mathrm{p}_2(S)$, then the same
contradiction arises with regard to $v$. Therefore, the case \ref{numberofmatrixrealizationsoftype:oneisolatedvertex:kequals6:case:lvertu1v1rvertbackslashp1Sequals0}.\ref{numberofmatrixrealizationsoftype:twoisolatedvertices:kequals6:case:lvertu1v1rvertbackslashp1Sequals0:case:u2notequaltov2} cannot occur.
\end{enumerate}
It follows that if \ref{numberofmatrixrealizationsoftype:oneisolatedvertex:kequals6:case:lvertu1v1rvertbackslashp1Sequals0}, then there are exactly $(n-1)-2$ realizations
of type \ref{item:comparisonproofcaseLhassize6:oneisolatedvertex} by $B\mid_{\{u,v\}}$.
\item\label{numberofmatrixrealizationsoftype:oneisolatedvertex:kequals6:case:lvertu1v1rvertbackslashp1Sequals1} $\lvert\{u_1,v_1\}\setminus\mathrm{p}_1(S)\rvert = 1$. Then
\eqref{eq:proofofnumberofeventsinducing:item:comparisonproofcaseLhassize6:oneisolatedvertex} implies $\lvert\{u_2,v_2\}\setminus\mathrm{p}_2(S)\rvert = 0$ which is
equivalent to $\{u_2,v_2\} \subseteq\mathrm{p}_2(S)$. The property defining
\ref{numberofmatrixrealizationsoftype:oneisolatedvertex:kequals6:case:lvertu1v1rvertbackslashp1Sequals1} is equivalent
to \eqref{eq:proofofcomparativecountingtheoremcase2secondeq}. Now an argument
entirely analogous to the one given for \ref{numberofmatrixrealizationsoftype:oneisolatedvertex:kequals6:case:lvertu1v1rvertbackslashp1Sequals0} shows that if
\ref{numberofmatrixrealizationsoftype:oneisolatedvertex:kequals6:case:lvertu1v1rvertbackslashp1Sequals1}, then there are exactly $(n-1)-2$ realizations of type
\ref{item:comparisonproofcaseLhassize6:oneisolatedvertex} by $B\mid_{\{u,v\}}$.
\end{enumerate}
It follows that there are exactly $2\cdot ((n-1)-2)$ different
$I\setminus S = \{u,v\}$ with $\mathrm{X}_{B} \cong \ref{item:comparisonproofcaseLhassize6:oneisolatedvertex}$. This completes the proof of \ref{numberofmatrixrealizationsoftype:oneisolatedvertex}.
As to \ref{numberofmatrixrealizationsoftype:oneadditionaledgeintersectingC4},
notice that a necessary condition for type \ref{item:comparisonproofcaseLhassize6:oneadditionaledgeintersectingC4} is that
$\lvert \mathrm{V}(\mathrm{X}_{B})\setminus \mathrm{V}(\mathrm{X}_{B\mid_{S}}) \rvert = 1$. Therefore
the set of all suitable $I\in\binom{[n-1]^2}{6}$ is a subset (possibly nonproper) of
those which are suitable for
type \ref{item:comparisonproofcaseLhassize6:oneisolatedvertex}. We may therefore
reexamine the analysis carried out for
\ref{numberofmatrixrealizationsoftype:oneisolatedvertex} and in each of the
cases count the number of $B\in\{0,\pm\}^I$ with $\mathrm{X}_{B}\cong \ref{item:comparisonproofcaseLhassize6:oneadditionaledgeintersectingC4}$.
If \ref{numberofmatrixrealizationsoftype:oneisolatedvertex:kequals6:case:lvertu1v1rvertbackslashp1Sequals0}.\ref{numberofmatrixrealizationsoftype:twoisolatedvertices:kequals6:case:lvertu1v1rvertbackslashp1Sequals0:case:u2equalsv2}, then properties
$u_1=a_1=c_1$ and $v_1=b_1=d_1$ show that both $u$ and $v$ have the property that if
one of them indexes a nonzero value of $B$, then there is an edge intersecting
$\mathrm{X}_{B\mid_S}\cong C^4$. Since otherwise we would have $K^{2,3}$, \emph{exactly} one
of them must be nonzero. This implies exactly $4$ possibilities to realize type
\ref{item:comparisonproofcaseLhassize6:oneadditionaledgeintersectingC4} for
each of the $(n-1)-2$ realizations of type
\ref{item:comparisonproofcaseLhassize6:oneisolatedvertex} which were offered
in \ref{numberofmatrixrealizationsoftype:oneisolatedvertex:kequals6:case:lvertu1v1rvertbackslashp1Sequals0}.\ref{numberofmatrixrealizationsoftype:twoisolatedvertices:kequals6:case:lvertu1v1rvertbackslashp1Sequals0:case:u2equalsv2}. Therefore, if
\ref{numberofmatrixrealizationsoftype:oneisolatedvertex:kequals6:case:lvertu1v1rvertbackslashp1Sequals0}.\ref{numberofmatrixrealizationsoftype:twoisolatedvertices:kequals6:case:lvertu1v1rvertbackslashp1Sequals0:case:u2equalsv2}, then there are exactly
$4\cdot ((n-1)-2)$ realizations of type \ref{item:comparisonproofcaseLhassize6:oneadditionaledgeintersectingC4} by $B\mid_{I\setminus S}$. Since the case \ref{numberofmatrixrealizationsoftype:oneisolatedvertex:kequals6:case:lvertu1v1rvertbackslashp1Sequals0}.\ref{numberofmatrixrealizationsoftype:twoisolatedvertices:kequals6:case:lvertu1v1rvertbackslashp1Sequals0:case:u2notequaltov2} is as impossible now as it was back then, it
follows that this is also the number of realizations for the entire \ref{numberofmatrixrealizationsoftype:oneisolatedvertex:kequals6:case:lvertu1v1rvertbackslashp1Sequals0}.
If \ref{numberofmatrixrealizationsoftype:oneisolatedvertex:kequals6:case:lvertu1v1rvertbackslashp1Sequals1}, then by interchanging the subscripts $1$
and $2$ we may use the same analysis as for the case \ref{numberofmatrixrealizationsoftype:oneisolatedvertex:kequals6:case:lvertu1v1rvertbackslashp1Sequals0} to reach the
conclusion that there are exactly $4\cdot ((n-1)-2)$ realizations of type
\ref{item:comparisonproofcaseLhassize6:oneadditionaledgeintersectingC4} by
$B\mid_{\{u,v\}} = B\mid_{I\setminus S}$.
It follows that there are are exactly $4\cdot ((n-1)-2) + 4\cdot ((n-1)-2) =
8\cdot (n-3)$ realizations of type \ref{item:comparisonproofcaseLhassize6:oneadditionaledgeintersectingC4} by $B\mid_{I\setminus S}$. This proves \ref{numberofmatrixrealizationsoftype:oneadditionaledgeintersectingC4}.
As to \ref{numberofmatrixrealizationsoftype:XBLisomorphictoK23}, it is
evident that the number of possibilities to realize a $K^{2,3}$ is
$2\cdot 2^6 \cdot \binom{n-1}{2}\cdot \binom{n-1}{3}$, the first factor accouting
for the two possibilities of either choosing two of the first indices and three of
the last, or vice versa.
As to \ref{numberofmatrixrealizationsoftype:twoisolatedvertices}, note that
$f_1\ref{item:comparisonproofcaseLhassize6:twoisolatedvertices} = 4$, hence it is
necessary that $B[u]=B[v]=0$. Therefore, the number of
$B\mid_{\{u,v\}}\in\{0,\pm\}^{I\setminus S}$ with
$\mathrm{X}_{B} \cong \ref{item:comparisonproofcaseLhassize6:twoisolatedvertices}$ equals
the number of $\{u,v\}\in \binom{[n-1]^2\setminus S}{2}$ such that
\begin{equation}\label{eq:proofofnumberofeventsinducing:item:comparisonproofcaseLhassize6:twoisolatedvertices}
\lvert \{ u_1, v_1\} \setminus \mathrm{p}_1(S) \rvert +
\lvert \{ u_2, v_2\} \setminus \mathrm{p}_2(S) \rvert = 2\quad .
\end{equation}
We now distinguish cases according to how \eqref{eq:proofofnumberofeventsinducing:item:comparisonproofcaseLhassize6:twoisolatedvertices} is satisfied.
\begin{enumerate}[label={\rm(C.\ref{numberofmatrixrealizationsoftype:twoisolatedvertices}.\arabic{*})}]
\item\label{numberofmatrixrealizationsoftype:twoisolatedvertices:case:lvertu1v1rvertbackslashp1Sequals0} $\lvert \{ u_1, v_1\} \setminus \mathrm{p}_1(S) \rvert = 0$. Then \eqref{eq:proofofnumberofeventsinducing:item:comparisonproofcaseLhassize6:twoisolatedvertices} implies $\lvert \{u_2, v_2 \} \setminus \mathrm{p}_2(S) \rvert = 2$, which is equivalent to \eqref{eq:proofofcomparativecountingtheoremcase2firsteq}. The property
defining Case \ref{numberofmatrixrealizationsoftype:twoisolatedvertices:case:lvertu1v1rvertbackslashp1Sequals0} is equivalent to $\{u_1,v_1\}\subseteq \mathrm{p}_1(S)$.
There are now two further cases:
\begin{enumerate}[label={\rm(\arabic{*})}]
\item\label{numberofmatrixrealizationsoftype:twoisolatedvertices:case:lvertu1v1rvertbackslashp1Sequals0:case:u1equalsv1} $u_1=v_1$. Then there are
exactly $2$ possible set inclusions $\{u_1,v_1\} = \{u_1\} \subseteq \mathrm{p}_1(S)
= \{a_1,c_1\}$. For each of them, there are exactly $\binom{(n-1)-2}{2}$ different
sets $\{u_2,v_2\}$ with property
\eqref{eq:proofofcomparativecountingtheoremcase2firsteq}. Since
$u_1=v_1$ and $u\prec v$ imply $u_2<v_2$, each of these sets determines the two
pairs $u$ and $v$.
Therefore, if \ref{numberofmatrixrealizationsoftype:twoisolatedvertices:case:lvertu1v1rvertbackslashp1Sequals0}.\ref{numberofmatrixrealizationsoftype:twoisolatedvertices:case:lvertu1v1rvertbackslashp1Sequals0:case:u1equalsv1},
then there are exactly $2\cdot \binom{(n-1)-2}{2}$
realizations of type \ref{item:comparisonproofcaseLhassize6:twoisolatedvertices}
by $B\mid_{\{u,v\}}$.
\item\label{numberofmatrixrealizationsoftype:twoisolatedvertices:case:lvertu1v1rvertbackslashp1Sequals0:case:u1doesnotequalv1} $u_1 \neq v_1$. Then $u\prec u$ implies
$u_1 < u_1$. Now there is exactly one possible set inclusion
$\{u_1,v_1\}\subseteq \mathrm{p}_1(S)=\{a_1,c_1\}$. When this inclusion holds, there are
exactly $\binom{(n-1)-2}{2}$ different sets $\{u_2,v_2\}$ with
property \eqref{eq:proofofcomparativecountingtheoremcase2firsteq}. Each of them can
be realized in exactly $2$ ways by $u$ and $v$, either by $u_2<v_2$ or by $v_2<u_2$.
Therefore, if \ref{numberofmatrixrealizationsoftype:twoisolatedvertices:case:lvertu1v1rvertbackslashp1Sequals0}.\ref{numberofmatrixrealizationsoftype:twoisolatedvertices:case:lvertu1v1rvertbackslashp1Sequals0:case:u1doesnotequalv1}, then there are again
(with a qualitatively different reason for the factor $2$)
exactly $2\cdot \binom{(n-1)-2}{2}$ realizations of
type \ref{item:comparisonproofcaseLhassize6:twoisolatedvertices} by $B\mid_{\{u,v\}}$.
\end{enumerate}
It follows that if \ref{numberofmatrixrealizationsoftype:twoisolatedvertices:case:lvertu1v1rvertbackslashp1Sequals0}, then there are exactly $4\cdot \binom{(n-1)-2}{2}$
different realizations of type \ref{item:comparisonproofcaseLhassize6:twoisolatedvertices} by $B\mid_{\{u,v\}}$.
\item\label{numberofmatrixrealizationsoftype:twoisolatedvertices:case:lvertu1v1rvertbackslashp1Sequals1} $\lvert \{ u_1, v_1\} \setminus \mathrm{p}_1(S) \rvert = 1$. Then \eqref{eq:proofofnumberofeventsinducing:item:comparisonproofcaseLhassize6:twoisolatedvertices}
implies $\lvert \{u_2, v_2 \} \setminus \mathrm{p}_2(S) \rvert = 1$. Hence, in the present
situation, the equations \eqref{eq:proofofcomparativecountingtheoremcase2secondeq}
and \eqref{eq:proofofcomparativecountingtheoremcase3secondeq} are simultaneously
true. There are now two further cases depending on the manner in
which \eqref{eq:proofofcomparativecountingtheoremcase2secondeq} is true:
\begin{enumerate}[label={\rm(\arabic{*})}]
\item\label{numberofmatrixrealizationsoftype:twoisolatedvertices:case:lvertu1v1rvertbackslashp1Sequals1:u1equalsv1andbracesu1v1bracesintersectp1Sisempty}
$u_1=v_1$ and $\{u_1,v_1\}\cap\mathrm{p}_1(S)=\emptyset$. There are $((n-1)-2)$
possibilities for this. For each of them the clause $u_2=v_2$ and
$\{u_2,v_2\}\cap\mathrm{p}_1(S)=\emptyset$ in
\eqref{eq:proofofcomparativecountingtheoremcase3secondeq} cannot be true because it
would imply $u=v$ and therefore for each of them $u_2\neq v_2$ and
$\lvert \{u_2,v_2\}\cap\mathrm{p}_2(S)\rvert = 1$ must be true. In the following, keep in
mind that $u_1=v_1$ and $u\prec v$ implies $u_2<v_2$. If $u_2 = a_2 = c_2$, then
$v_2\neq b_2=d_2$, hence there are exactly $n-1 - a_2 - 1$ different $v_2$, and
therefore as many different realizations of type \ref{item:comparisonproofcaseLhassize6:twoisolatedvertices} by $u$ and $v$.
If $u_2 = b_2 = d_2$, then there are exactly $n-1 - b_2$ different $v_2$, and
therefore as many different realizations of type \ref{item:comparisonproofcaseLhassize6:twoisolatedvertices} by $u$ and $v$.
If $v_2 = a_2 = c_2$, then there are exactly $a_2 - 1$ different
$u_2$, and therefore as many different realizations of type \ref{item:comparisonproofcaseLhassize6:twoisolatedvertices} by $u$ and $v$.
If $v_2 = b_2 = d_2$, then $u_2\neq a_2=c_2$, hence there are
exactly $b_2 - 1 - 1$ different $u_2$ and therefore as many different
realizations of type \ref{item:comparisonproofcaseLhassize6:twoisolatedvertices} by $u$ and $v$.
Therefore, if \ref{numberofmatrixrealizationsoftype:twoisolatedvertices:case:lvertu1v1rvertbackslashp1Sequals1}.\ref{numberofmatrixrealizationsoftype:twoisolatedvertices:case:lvertu1v1rvertbackslashp1Sequals1:u1equalsv1andbracesu1v1bracesintersectp1Sisempty}, then there are
$
\bigl( (n-1-a_2-1)+(n-1-b_2)+(a_2-1)+(b_2-1-1)\bigr) \cdot ((n-1)-2)
= 2\cdot ((n-1)-2)^2
$
realizations of type \ref{item:comparisonproofcaseLhassize6:twoisolatedvertices}
by $B\mid_{I\setminus S} = B\mid_{\{u,v\}}$.
\item\label{numberofmatrixrealizationsoftype:twoisolatedvertices:case:lvertu1v1rvertbackslashp1Sequals1:u1doesnotequalv1andbracesu1v1bracesintersectp1Sisasingleton}
$u_1 \neq v_1$ and $\lvert \{u_1,v_1\} \cap \mathrm{p}_1(S) \rvert = 1$.
Then $u\prec v$ implies $u_1 < v_1$ and there are two further cases:
\begin{enumerate}[label={\rm(\arabic{*})}]
\item\label{numberofmatrixrealizationsoftype:twoisolatedvertices:case:lvertu1v1rvertbackslashp1Sequals1:u1doesnotequalv1andbracesu1v1bracesintersectp1Sisasingletoncontainingu1} $\lvert \{u_1,v_1\} \cap \mathrm{p}_1(S) \rvert = 1$ is true due
to $u_1 \in \mathrm{p}_1(S) = \{a_1,c_1\}$. Then $v_1\notin \{a_1,c_1\}=\mathrm{p}_1(S)$ and there
are two further cases:
\begin{enumerate}[label={\rm(\arabic{*})}]
\item\label{numberofmatrixrealizationsoftype:twoisolatedvertices:case:lvertu1v1rvertbackslashp1Sequals1:u1doesnotequalv1andbracesu1v1bracesintersectp1Sisasingletoncontainingu1:u1equaltoa1} $u_1 = a_1 = b_1$. Since $u_1 < v_1$ and $v_1\notin \{a_1,c_1\}$,
there are then exactly $n-1 - a_1 - 1$ different $v_1$. For each of them, there are
two cases. If the first clause of
\eqref{eq:proofofcomparativecountingtheoremcase3secondeq} is true, then there are
exactly $(n-1)-2$ different such $u_2=v_2$, and each such
$u_2=v_2$ \emph{determines} the two pairs $u$ and $v$. Therefore in this case there
are exactly $(n-1-a_1-1)\cdot ((n-1)-2)$ realizations of type
\ref{item:comparisonproofcaseLhassize6:twoisolatedvertices} by $B\mid_{\{u,v\}}$.
If the second clause of \eqref{eq:proofofcomparativecountingtheoremcase3secondeq} is
true, then since $u_1=a_1=b_1$ combined with $u\neq a$ and $u\neq b$ implies
$u_2\notin \{a_2=c_2,b_2=d_2\} = \mathrm{p}_2(S)$, it follows that
$\lvert \{u_2, v_2\} \cap \mathrm{p}_2(S)\rvert =1$ is true as $v_2\in \mathrm{p}_2(S)$.
Therefore in this case there are exactly $(n-1)-2$ different $u_2$ and
exactly $2$ different $v_2$ for each of them and hence exactly
$(n-1-a_1-1)\cdot 2\cdot ((n-1)-2)$ realizations of type
\ref{item:comparisonproofcaseLhassize6:twoisolatedvertices} by $B\mid_{\{u,v\}}$.
Therefore, if \ref{numberofmatrixrealizationsoftype:twoisolatedvertices:case:lvertu1v1rvertbackslashp1Sequals1}.\ref{numberofmatrixrealizationsoftype:twoisolatedvertices:case:lvertu1v1rvertbackslashp1Sequals1:u1doesnotequalv1andbracesu1v1bracesintersectp1Sisasingleton}.\ref{numberofmatrixrealizationsoftype:twoisolatedvertices:case:lvertu1v1rvertbackslashp1Sequals1:u1doesnotequalv1andbracesu1v1bracesintersectp1Sisasingletoncontainingu1}.\ref{numberofmatrixrealizationsoftype:twoisolatedvertices:case:lvertu1v1rvertbackslashp1Sequals1:u1doesnotequalv1andbracesu1v1bracesintersectp1Sisasingletoncontainingu1:u1equaltoa1}, then there are exactly $3\cdot (n-1-a_1-1)\cdot ((n-1)-2)$
different realizations of type \ref{item:comparisonproofcaseLhassize6:twoisolatedvertices} by $B\mid_{\{u,v\}}$.
\item\label{numberofmatrixrealizationsoftype:twoisolatedvertices:case:lvertu1v1rvertbackslashp1Sequals1:u1doesnotequalv1andbracesu1v1bracesintersectp1Sisasingletoncontainingu1:u1equaltoc1} $u_1 = c_1 = d_1$. Since $u_1< v_1$, there are then exactly
$n-1 - c_1$ different $v_1$. For each of them, there are two cases. If the first
clause of \eqref{eq:proofofcomparativecountingtheoremcase3secondeq} is true, then
there are exactly $(n-1 - c_1)\cdot ((n-1)-2)$ realizations of type
\ref{item:comparisonproofcaseLhassize6:twoisolatedvertices} by $B\mid_{\{u,v\}}$. If
the second clause of \eqref{eq:proofofcomparativecountingtheoremcase3secondeq} is
true, then since $u_1=c_1=d_1$ combined with $u\neq c$ and $u\neq d$ implies
$u_2\notin \{a_2=c_2,b_2=d_2\}= \mathrm{p}_2(S)$, it follows that $\lvert \{u_2, v_2\} \cap \mathrm{p}_2(S)\rvert =1$ is true as $v_2\in \mathrm{p}_2(S)$. Therefore in this case there are
exactly $(n-1)-2$ different $u_2$ and for each of them exactly $2$ different $v_2$,
thus exactly $(n-1-c_1) \cdot 2\cdot ((n-1)-2)$ realizations of type \ref{item:comparisonproofcaseLhassize6:twoisolatedvertices} by $B\mid_{\{u,v\}}$.
Therefore, if \ref{numberofmatrixrealizationsoftype:twoisolatedvertices:case:lvertu1v1rvertbackslashp1Sequals1}.\ref{numberofmatrixrealizationsoftype:twoisolatedvertices:case:lvertu1v1rvertbackslashp1Sequals1:u1doesnotequalv1andbracesu1v1bracesintersectp1Sisasingleton}.\ref{numberofmatrixrealizationsoftype:twoisolatedvertices:case:lvertu1v1rvertbackslashp1Sequals1:u1doesnotequalv1andbracesu1v1bracesintersectp1Sisasingletoncontainingu1}.\ref{numberofmatrixrealizationsoftype:twoisolatedvertices:case:lvertu1v1rvertbackslashp1Sequals1:u1doesnotequalv1andbracesu1v1bracesintersectp1Sisasingletoncontainingu1:u1equaltoc1}, then there are exactly $3\cdot (n-1-c_1)\cdot ((n-1)-2)$ realizations of
type \ref{item:comparisonproofcaseLhassize6:twoisolatedvertices} by $B\mid_{\{u,v\}}$.
\end{enumerate}
It follows that if \ref{numberofmatrixrealizationsoftype:twoisolatedvertices:case:lvertu1v1rvertbackslashp1Sequals1}.\ref{numberofmatrixrealizationsoftype:twoisolatedvertices:case:lvertu1v1rvertbackslashp1Sequals1:u1doesnotequalv1andbracesu1v1bracesintersectp1Sisasingleton}.\ref{numberofmatrixrealizationsoftype:twoisolatedvertices:case:lvertu1v1rvertbackslashp1Sequals1:u1doesnotequalv1andbracesu1v1bracesintersectp1Sisasingletoncontainingu1}, then there are exactly
$3\cdot (n-1-a_1-1)\cdot ((n-1)-2) + 3\cdot (n-1-c_1)\cdot ((n-1)-2)
= 3\cdot (2n-a_1-c_1-3)\cdot ((n-1)-2)$ different realizations of
type \ref{item:comparisonproofcaseLhassize6:twoisolatedvertices} by $B\mid_{\{u,v\}}$.
\item\label{numberofmatrixrealizationsoftype:twoisolatedvertices:case:lvertu1v1rvertbackslashp1Sequals1:u1doesnotequalv1andbracesu1v1bracesintersectp1Sisasingletoncontainingv1} $\lvert \{u_1,v_1\} \cap \mathrm{p}_1(S) \rvert = 1$ is true due to
$v_1 \in \mathrm{p}_1(S) = \{a_1, c_1\}$. Then $u_1\notin \{a_1,c_1\}$ and there are two
further cases:
\begin{enumerate}[label={\rm(\arabic{*})}]
\item\label{numberofmatrixrealizationsoftype:twoisolatedvertices:case:lvertu1v1rvertbackslashp1Sequals1:u1doesnotequalv1andbracesu1v1bracesintersectp1Sisasingletoncontainingu1:v1equaltoa1} $v_1 = a_1 = b_1$. Since $u_1<v_1$, there are then exactly $a_1-1$
different $u_1$. For each of them, there are two cases. If the first clause of
\eqref{eq:proofofcomparativecountingtheoremcase3secondeq} is true, then there are
exactly $(n-1)-2$ different $u_2=v_2$ and each such $u_2=v_2$ determines the
pair $(u,v)$. Hence in this case there are exactly $(a_1-1)\cdot ((n-1)-2)$
realizations of type \ref{item:comparisonproofcaseLhassize6:twoisolatedvertices} by
$B\mid_{\{u,v\}}$. If the second clause of
\eqref{eq:proofofcomparativecountingtheoremcase3secondeq} is true, then since
$v_1=a_1=b_1$ combined with $v\neq a$ and $v\neq b$ implies
$v_2\notin\{a_2=c_2,b_2=d_2\} = \mathrm{p}_2(S)$, we know that
$\lvert \{u_2, v_2\} \cap \mathrm{p}_2(S)\rvert =1$ must be true as $u_2\in \mathrm{p}_2(S)$,
hence there are $2$ different $u_2$ and for each of them exactly $(n-1)-2$ different
$v_2$, hence in this case there are exactly $(a_1-1)\cdot 2\cdot ((n-1)-2)$
realizations of type \ref{item:comparisonproofcaseLhassize6:twoisolatedvertices}
by $B\mid_{\{u,v\}}$.
Therefore, if \ref{numberofmatrixrealizationsoftype:twoisolatedvertices:case:lvertu1v1rvertbackslashp1Sequals1}.\ref{numberofmatrixrealizationsoftype:twoisolatedvertices:case:lvertu1v1rvertbackslashp1Sequals1:u1doesnotequalv1andbracesu1v1bracesintersectp1Sisasingleton}.\ref{numberofmatrixrealizationsoftype:twoisolatedvertices:case:lvertu1v1rvertbackslashp1Sequals1:u1doesnotequalv1andbracesu1v1bracesintersectp1Sisasingletoncontainingv1}.\ref{numberofmatrixrealizationsoftype:twoisolatedvertices:case:lvertu1v1rvertbackslashp1Sequals1:u1doesnotequalv1andbracesu1v1bracesintersectp1Sisasingletoncontainingu1:v1equaltoa1}, then there are exactly $3\cdot (a_1-1)\cdot ((n-1)-2)$ different
realizations of type \ref{item:comparisonproofcaseLhassize6:twoisolatedvertices}
by $B\mid_{\{u,v\}}$.
\item\label{numberofmatrixrealizationsoftype:twoisolatedvertices:case:lvertu1v1rvertbackslashp1Sequals1:u1doesnotequalv1andbracesu1v1bracesintersectp1Sisasingletoncontainingu1:v1equaltoc1} $v_1 = c_1 = d_1$. Since $u_1 < v_1$ and $u_1\notin \{a_1,c_1\}$,
there are then exactly $c_1-1-1$ different $u_1$. For each of them, there are two
cases. If the first clause of
\eqref{eq:proofofcomparativecountingtheoremcase3secondeq} is true, then there are
exactly $(c_1-1-1)\cdot ((n-1)-2)$ realizations of type
\ref{item:comparisonproofcaseLhassize6:twoisolatedvertices} by $B\mid_{\{u,v\}}$. If
the second clause of \eqref{eq:proofofcomparativecountingtheoremcase3secondeq} is
true, then since $v_1 = c_1 = d_1$ combined with $v\neq c$ and $v\neq d$ implies
$v_2\notin\{a_2=c_2,b_2=d_2\}=\mathrm{p}_2(S)$, it follows that
$\lvert\{u_2,v_2 \}\cap\mathrm{p}_2(S)\rvert=1$ is true as $u_2\in \mathrm{p}_2(S)$. Therefore
in this case there are exactly $(n-1)-2$ different $v_2$ and for each of them
exactly $2$ different $u_2$, hence exactly $(c_1-1-1)\cdot 2\cdot ((n-1) - 2)$
realizations of type \ref{item:comparisonproofcaseLhassize6:twoisolatedvertices} by
$B\mid_{\{u,v\}}$.
Therefore, if \ref{numberofmatrixrealizationsoftype:twoisolatedvertices:case:lvertu1v1rvertbackslashp1Sequals1}.\ref{numberofmatrixrealizationsoftype:twoisolatedvertices:case:lvertu1v1rvertbackslashp1Sequals1:u1doesnotequalv1andbracesu1v1bracesintersectp1Sisasingleton}.\ref{numberofmatrixrealizationsoftype:twoisolatedvertices:case:lvertu1v1rvertbackslashp1Sequals1:u1doesnotequalv1andbracesu1v1bracesintersectp1Sisasingletoncontainingv1}.\ref{numberofmatrixrealizationsoftype:twoisolatedvertices:case:lvertu1v1rvertbackslashp1Sequals1:u1doesnotequalv1andbracesu1v1bracesintersectp1Sisasingletoncontainingu1:v1equaltoc1}, then there are exactly $3\cdot (c_1-1-1)\cdot ((n-1)-2)$ realizations
of type \ref{item:comparisonproofcaseLhassize6:twoisolatedvertices} by $B\mid_{\{u,v\}}$.
\end{enumerate}
It follows that if \ref{numberofmatrixrealizationsoftype:twoisolatedvertices:case:lvertu1v1rvertbackslashp1Sequals1}.\ref{numberofmatrixrealizationsoftype:twoisolatedvertices:case:lvertu1v1rvertbackslashp1Sequals1:u1doesnotequalv1andbracesu1v1bracesintersectp1Sisasingleton}.\ref{numberofmatrixrealizationsoftype:twoisolatedvertices:case:lvertu1v1rvertbackslashp1Sequals1:u1doesnotequalv1andbracesu1v1bracesintersectp1Sisasingletoncontainingv1}, then there are exactly
$3\cdot (a_1-1)\cdot ((n-1)-2) + 3\cdot (c_1-1-1)\cdot ((n-1)-2)
= 3\cdot (a_1+c_1-3)\cdot ((n-1)-2)$ realizations of
type \ref{item:comparisonproofcaseLhassize6:twoisolatedvertices} by $B\mid_{\{u,v\}}$.
\end{enumerate}
It follows that if \ref{numberofmatrixrealizationsoftype:twoisolatedvertices:case:lvertu1v1rvertbackslashp1Sequals1}.\ref{numberofmatrixrealizationsoftype:twoisolatedvertices:case:lvertu1v1rvertbackslashp1Sequals1:u1doesnotequalv1andbracesu1v1bracesintersectp1Sisasingleton}, then there are exactly
$3\cdot (2n-a_1-c_1-3)\cdot ((n-1)-2) + 3\cdot (a_1+c_1-3)\cdot ((n-1)-2)
= 6\cdot ((n-1)-2)^2$ realizations of type
\ref{item:comparisonproofcaseLhassize6:twoisolatedvertices} by $B\mid_{\{u,v\}}$.
\end{enumerate}
It follows that if \ref{numberofmatrixrealizationsoftype:twoisolatedvertices:case:lvertu1v1rvertbackslashp1Sequals1}, then there are exactly
$2\cdot ((n-1)-2)^2 + 6\cdot ((n-1)-2)^2 = 8\cdot ((n-1)-2)^2$
realizations of type \ref{item:comparisonproofcaseLhassize6:twoisolatedvertices}
by $B\mid_{\{u,v\}}$.
\item\label{numberofmatrixrealizationsoftype:twoisolatedvertices:case:lvertu1v1rvertbackslashp1Sequals2} $\lvert \{u_1,v_1\}\setminus\mathrm{p}_1(S)\rvert = 2$. This is
equivalent to \eqref{eq:proofofcomparativecountingtheoremcase3firsteq}. Moreover
\eqref{eq:proofofnumberofeventsinducing:item:comparisonproofcaseLhassize6:twoisolatedvertices} implies $\lvert \{u_2,v_2\}\setminus\mathrm{p}_2(S)\rvert = 0$, which is
equivalent to $\{u_2,v_2\}\subseteq\mathrm{p}_2(S)$. Swapping the subscripts $1$ and
$2$ in the analysis of \ref{numberofmatrixrealizationsoftype:twoisolatedvertices:case:lvertu1v1rvertbackslashp1Sequals0} shows that if \ref{numberofmatrixrealizationsoftype:twoisolatedvertices:case:lvertu1v1rvertbackslashp1Sequals2}, then there are exactly
$4\cdot\binom{(n-1)-2}{2}$ realizations of type
\ref{item:comparisonproofcaseLhassize6:twoisolatedvertices} by $B\mid_{\{u,v\}}$.
\end{enumerate}
It follows that for the fixed $S$, there are exactly
$4\cdot \binom{(n-1)-2}{2} + 8\cdot((n-1)-2)^2 + 4\cdot \binom{(n-1)-2}{2}
= 8\cdot(n-3)^2 + 8\cdot \binom{n-3}{2}$ different $I\setminus S = \{u,v\}$ with
the property $\mathrm{X}_{B} \cong \ref{item:comparisonproofcaseLhassize6:twoisolatedvertices}$. This completes the proof of
\ref{numberofmatrixrealizationsoftype:twoisolatedvertices}.
As to \ref{numberofmatrixrealizationsoftype:oneadditionaledgeintersectingC4andoneisolatedvertex} and \ref{numberofmatrixrealizationsoftype:C4withoneadditionaldisjointedge},
let us first note that both for \ref{item:comparisonproofcaseLhassize6:oneadditionaledgeintersectingC4andoneisolatedvertex} and for \ref{item:comparisonproofcaseLhassize6:C4withoneadditionaldisjointedge} a necessary condition is that $\mathrm{X}_{B}$ contain
exactly two vertices not in $\mathrm{X}_{B\mid_S}$ $\cong$ $C^4$.
Therefore, the set of all $I\setminus S = \{u,v\}$ with
$\mathrm{X}_B \cong \ref{item:comparisonproofcaseLhassize6:oneadditionaledgeintersectingC4andoneisolatedvertex}$ is a subset of the set of those $\{u,v\}$ with
$\mathrm{X}_{\{0\}^{\{u,v\}}} \cong \ref{item:comparisonproofcaseLhassize6:twoisolatedvertices}$,
and likewise for \ref{item:comparisonproofcaseLhassize6:C4withoneadditionaldisjointedge}. We may therefore determine both \ref{numberofmatrixrealizationsoftype:oneadditionaledgeintersectingC4andoneisolatedvertex} and \ref{numberofmatrixrealizationsoftype:C4withoneadditionaldisjointedge} by a single reexamination of the analysis given for type
\ref{item:comparisonproofcaseLhassize6:twoisolatedvertices}. In each of the cases
which we distinguished there we now have to count the number of those
$B\mid_{\{u,v\}}\in\{0,\pm\}^{\{u,v\}}$ with $\mathrm{X}_B \cong \ref{item:comparisonproofcaseLhassize6:oneadditionaledgeintersectingC4andoneisolatedvertex}$ and also of
those $B\mid_{\{u,v\}}\in\{0,\pm\}^{\{u,v\}}$ with $\mathrm{X}_{B} \cong \ref{item:comparisonproofcaseLhassize6:C4withoneadditionaldisjointedge}$.
We can prepare for this as follows. Consider the properties
{\small
\begin{enumerate}[label={\rm(i\arabic{*})}]
\item
\label{item:casesfortype8casethatexactlyoneofuandvhasaprojectionmeetingaprojectionofS}
($\{u_1,u_2\}\cap \mathrm{p}(S) \neq \emptyset$ and
$\{v_1,v_2\}\cap \mathrm{p}(S) = \emptyset$) or
($\{u_1,u_2\}\cap \mathrm{p}(S) = \emptyset$ and
$\{v_1,v_2\}\cap \mathrm{p}(S) \neq \emptyset$)\quad ,
\item
\label{item:casesfortype8casethatbothuandvhaveaprojectionmeetingaprojectionofS}
($\{u_1,u_2\}\cap \mathrm{p}(S) \neq \emptyset$ and
$\{v_1,v_2\}\cap \mathrm{p}(S) \neq \emptyset$)\quad .
\end{enumerate}
}
In each of the cases to be reexamined, these properties alone determine how many
$B\mid_{\{u,v\}}\in \{0,\pm\}^{\{u,v\}}$ realize \ref{item:comparisonproofcaseLhassize6:oneadditionaledgeintersectingC4andoneisolatedvertex} or \ref{item:comparisonproofcaseLhassize6:C4withoneadditionaldisjointedge}.
Let us first focus on \ref{item:comparisonproofcaseLhassize6:oneadditionaledgeintersectingC4andoneisolatedvertex}. In \ref{item:casesfortype8casethatexactlyoneofuandvhasaprojectionmeetingaprojectionofS}, each of the two clauses of that disjunction has the
property that if it is true, then there are exactly $2$ possibilities for a
$B\mid_{\{u,v\}}$ with $\mathrm{X}_B \cong \ref{item:comparisonproofcaseLhassize6:oneadditionaledgeintersectingC4andoneisolatedvertex}$. For the first clause these
are ($B[u]\in \{\pm\}$ and $B[v]=0$), for the second
clause ($B[u] = 0$ and $B[v]\in\{\pm\}$). Moreover, the disjunction is evidently
exclusive. Therefore, if property \ref{item:casesfortype8casethatexactlyoneofuandvhasaprojectionmeetingaprojectionofS} is true, then the number
of $B\mid_{\{u,v\}}\in \{0,\pm\}^{\{u,v\}}$ with $\mathrm{X}_B \cong \ref{item:comparisonproofcaseLhassize6:oneadditionaledgeintersectingC4andoneisolatedvertex}$
is exactly $2$-times as large as the number of
$\{u,v\}\in\binom{[n-1]^2}{2}$ with $\mathrm{X}_{\{0\}^{\{u,v\}}} \cong \ref{item:comparisonproofcaseLhassize6:twoisolatedvertices}$ which was determined in the proof
of \ref{numberofmatrixrealizationsoftype:twoisolatedvertices}.
If property \ref{item:casesfortype8casethatbothuandvhaveaprojectionmeetingaprojectionofS} is true, then there are exactly $4$ possibilities for a $B\mid_{\{u,v\}}$ with
$\mathrm{X}_B \cong \ref{item:comparisonproofcaseLhassize6:oneadditionaledgeintersectingC4andoneisolatedvertex}$:
($B[u]\in \{\pm\}$ and $B[v]=0$) or ($B[u] = 0$ and $B[v] \in \{\pm\}$). Therefore,
if property \ref{item:casesfortype8casethatbothuandvhaveaprojectionmeetingaprojectionofS} is true, then there are exactly $4$-times as many realizations of isomorphism type
\ref{item:comparisonproofcaseLhassize6:oneadditionaledgeintersectingC4andoneisolatedvertex} by $B\mid_{I\setminus S} = B\mid_{\{u,v\}}$ as there had been for type
\ref{item:comparisonproofcaseLhassize6:twoisolatedvertices}.
Let us now turn to \ref{item:comparisonproofcaseLhassize6:C4withoneadditionaldisjointedge}. In case \ref{item:casesfortype8casethatexactlyoneofuandvhasaprojectionmeetingaprojectionofS} there are (just as for type \ref{item:comparisonproofcaseLhassize6:oneadditionaledgeintersectingC4andoneisolatedvertex}) exactly $2$ possibilities for
a $B\mid_{\{u,v\}}$ with $\mathrm{X}_{B} \cong \ref{item:comparisonproofcaseLhassize6:C4withoneadditionaldisjointedge}$. This time, these are ($B[u]=0$ and $B[v]\in\{\pm\}$) for
the first clause of \ref{item:casesfortype8casethatexactlyoneofuandvhasaprojectionmeetingaprojectionofS}, and ($B[u]\in\{\pm\}$ and $B[v]=0$) for the second clause.
Again, due to the mutual exclusiveness of the clauses, it follows that whenever
case \ref{item:casesfortype8casethatexactlyoneofuandvhasaprojectionmeetingaprojectionofS} is true (no matter by way of which clause), there are exactly $2$-times as many
$B\mid_{\{u,v\}}\in \{0,\pm\}^{\{u,v\}}$ with $\mathrm{X}_B \cong \ref{item:comparisonproofcaseLhassize6:C4withoneadditionaldisjointedge}$ as there are
$\{u,v\}\in \binom{[n-1]^2}{2}$ with $\mathrm{X}_{\{0\}^{\{u,v\}}} \cong \ref{item:comparisonproofcaseLhassize6:twoisolatedvertices}$. Concerning property \ref{item:casesfortype8casethatbothuandvhaveaprojectionmeetingaprojectionofS}, however, there is a genuine
difference: when this property is true, there is \emph{no} possibility to
choose $B\mid_{\{u,v\}}$ so as to create exactly one edge disjoint from
the $\mathrm{X}_{B\mid_S}\cong C^4$. We can now begin inspecting the cases.
If \ref{numberofmatrixrealizationsoftype:twoisolatedvertices:case:lvertu1v1rvertbackslashp1Sequals0}, then the inclusion $\{u_1,v_1\}\subseteq \mathrm{p}_1(S)$ alone, no matter
whether $u_1=v_1$ or not, implies that property \ref{item:casesfortype8casethatbothuandvhaveaprojectionmeetingaprojectionofS} is true and without going any deeper we know
that there are exactly
$4\cdot 4 \cdot \binom{(n-1)-2}{2} = 16 \cdot \binom{(n-1)-2}{2}$
realizations of type \ref{item:comparisonproofcaseLhassize6:oneadditionaledgeintersectingC4andoneisolatedvertex} and $0$ realizations of type \ref{item:comparisonproofcaseLhassize6:C4withoneadditionaldisjointedge} by $B\mid_{\{u,v\}}$.
If \ref{numberofmatrixrealizationsoftype:threeisolatedvertices:case:lvertu1v1rvertbackslashp1Sequals1}, we have to descend one level deeper. If \ref{numberofmatrixrealizationsoftype:threeisolatedvertices:case:lvertu1v1rvertbackslashp1Sequals1}.\ref{numberofmatrixrealizationsoftype:twoisolatedvertices:case:lvertu1v1rvertbackslashp1Sequals1:u1equalsv1andbracesu1v1bracesintersectp1Sisempty} then it is known that
$u_1=v_1$, $\{u_1,v_1\}\cap\mathrm{p}_1(S) = \emptyset$,
$u_2<v_2$ and $\lvert \{ u_2,v_2\}\cap\mathrm{p}_2(S) \rvert = 1$ and obviously this implies
that property \ref{item:casesfortype8casethatexactlyoneofuandvhasaprojectionmeetingaprojectionofS} is true. Therefore without having to reexamine further subcases
we then know that if \ref{numberofmatrixrealizationsoftype:threeisolatedvertices:case:lvertu1v1rvertbackslashp1Sequals1}.\ref{numberofmatrixrealizationsoftype:twoisolatedvertices:case:lvertu1v1rvertbackslashp1Sequals1:u1equalsv1andbracesu1v1bracesintersectp1Sisempty}, then there are exactly $2\cdot 2\cdot ((n-1)-2)^2 = 4\cdot ((n-1)-2)^2$
realizations of type \ref{item:comparisonproofcaseLhassize6:oneadditionaledgeintersectingC4andoneisolatedvertex} and also exactly $2 \cdot 2 \cdot ((n-1)-2)^2
= 4\cdot((n-1)-2)^2$ realizations of type \ref{item:comparisonproofcaseLhassize6:C4withoneadditionaldisjointedge} by $B\mid_{\{u,v\}}$.
If \ref{numberofmatrixrealizationsoftype:threeisolatedvertices:case:lvertu1v1rvertbackslashp1Sequals1}.\ref{numberofmatrixrealizationsoftype:twoisolatedvertices:case:lvertu1v1rvertbackslashp1Sequals1:u1doesnotequalv1andbracesu1v1bracesintersectp1Sisasingleton}, however, then we have to go deeper still. Although we then already know
that $u_1<v_1$ and $\lvert \{u_1,v_1\} \cap \mathrm{p}_1(S) \rvert = 1$, and therefore
know that
\begin{equation}\label{partialknowledgeintheanalysisoftype8}
\text{$\{u_1,u_2\} \cap \mathrm{p}(S)\neq\emptyset$ or
$\{v_1,v_2\} \cap \mathrm{p}(S)\neq\emptyset$ \quad , }
\end{equation}
at the present stage of our knowledge this latter property is compatible with both
\ref{item:casesfortype8casethatexactlyoneofuandvhasaprojectionmeetingaprojectionofS}
and \ref{item:casesfortype8casethatbothuandvhaveaprojectionmeetingaprojectionofS}
(i.e., we do not know yet whether the `or' in
\eqref{partialknowledgeintheanalysisoftype8} is true as an `and'). The
reason is that we do not yet have any knowledge about $u_2$ and $v_2$. Therefore,
neither descending down to \ref{numberofmatrixrealizationsoftype:threeisolatedvertices:case:lvertu1v1rvertbackslashp1Sequals1}.\ref{numberofmatrixrealizationsoftype:twoisolatedvertices:case:lvertu1v1rvertbackslashp1Sequals1:u1doesnotequalv1andbracesu1v1bracesintersectp1Sisasingleton}.\ref{numberofmatrixrealizationsoftype:twoisolatedvertices:case:lvertu1v1rvertbackslashp1Sequals1:u1doesnotequalv1andbracesu1v1bracesintersectp1Sisasingletoncontainingu1} nor to \ref{numberofmatrixrealizationsoftype:threeisolatedvertices:case:lvertu1v1rvertbackslashp1Sequals1}.\ref{numberofmatrixrealizationsoftype:twoisolatedvertices:case:lvertu1v1rvertbackslashp1Sequals1:u1doesnotequalv1andbracesu1v1bracesintersectp1Sisasingleton}.\ref{numberofmatrixrealizationsoftype:twoisolatedvertices:case:lvertu1v1rvertbackslashp1Sequals1:u1doesnotequalv1andbracesu1v1bracesintersectp1Sisasingletoncontainingu1}.\ref{numberofmatrixrealizationsoftype:twoisolatedvertices:case:lvertu1v1rvertbackslashp1Sequals1:u1doesnotequalv1andbracesu1v1bracesintersectp1Sisasingletoncontainingu1:u1equaltoa1} is sufficient for us to know whether \ref{item:casesfortype8casethatexactlyoneofuandvhasaprojectionmeetingaprojectionofS} or \ref{item:casesfortype8casethatbothuandvhaveaprojectionmeetingaprojectionofS} is true. We therefore have to go all the way down to the two (anonymous) subcases of maximal depth within the case \ref{numberofmatrixrealizationsoftype:threeisolatedvertices:case:lvertu1v1rvertbackslashp1Sequals1}.\ref{numberofmatrixrealizationsoftype:twoisolatedvertices:case:lvertu1v1rvertbackslashp1Sequals1:u1doesnotequalv1andbracesu1v1bracesintersectp1Sisasingleton}.\ref{numberofmatrixrealizationsoftype:twoisolatedvertices:case:lvertu1v1rvertbackslashp1Sequals1:u1doesnotequalv1andbracesu1v1bracesintersectp1Sisasingletoncontainingu1}.\ref{numberofmatrixrealizationsoftype:twoisolatedvertices:case:lvertu1v1rvertbackslashp1Sequals1:u1doesnotequalv1andbracesu1v1bracesintersectp1Sisasingletoncontainingu1:u1equaltoa1}. In the first of the two subcases we know that
$\{u_2,v_2\}\cap \mathrm{p}_2(S) = \emptyset$ and combining this with our knowledge
of $\lvert \{u_1,v_1\} \cap \mathrm{p}_1(S) \rvert = 1$ we may conclude that exactly one
of the two clauses in \eqref{partialknowledgeintheanalysisoftype8}, and hence
property \ref{item:casesfortype8casethatexactlyoneofuandvhasaprojectionmeetingaprojectionofS} is true. Therefore, in the present subcase there are exactly
$2\cdot (n-1-a_1-1)\cdot ((n-1)-2)$ realizations of type \ref{item:comparisonproofcaseLhassize6:oneadditionaledgeintersectingC4andoneisolatedvertex} and also $2\cdot (n-1-a_1-1)\cdot ((n-1)-2)$ realizations of type \ref{item:comparisonproofcaseLhassize6:C4withoneadditionaldisjointedge} by $B\mid_{\{u,v\}}$.
In the second of the two subcases we know that $u_2\neq v_2$ and
$\lvert \{u_2,v_2\}\cap\mathrm{p}_2(S) \rvert = 1$. Recall that at present we also know
that $u_1<v_1$ and $\lvert \{u_1,v_1\} \cap \mathrm{p}_1(S) \rvert = 1$.
Keeping in mind the fact that because of $u\notin S$ at most one projection of $u$
can be contained in $\mathrm{p}(S)$ (and the analogous fact about $v$),
we may argue that if
$\lvert \{u_1,v_1\} \cap \mathrm{p}_1(S) \rvert = 1$ is true as
($u_1\in \mathrm{p}_1(S)$ and $v_1\notin\mathrm{p}_1(S)$), then
$\lvert \{u_2,v_2\}\cap\mathrm{p}_2(S) \rvert = 1$ must be true
as ($u_2\notin\mathrm{p}_2(S)$ and $v_2\in\mathrm{p}_2(S)$), and if
$\lvert \{u_1,v_1\} \cap \mathrm{p}_1(S) \rvert = 1$ is true as
($u_1\notin\mathrm{p}_1(S)$ and $v_1\in \mathrm{p}_1(S)$ ),
then $\lvert \{u_2,v_2\}\cap\mathrm{p}_2(S) \rvert = 1$ must be true as
($u_2\in\mathrm{p}_2(S)$ and $v_2\notin\mathrm{p}_2(S)$). Since in both
cases both clauses of \eqref{partialknowledgeintheanalysisoftype8} are true,
it follows that \ref{item:casesfortype8casethatbothuandvhaveaprojectionmeetingaprojectionofS} is true. Therefore in the present subcase there are exactly
$4\cdot (n-1-a_1-1)\cdot 2\cdot ((n-1)-2) = 8\cdot (n-1-a_1-1)\cdot ((n-1)-2)$
realizations of type \ref{item:comparisonproofcaseLhassize6:oneadditionaledgeintersectingC4andoneisolatedvertex} and $0$ realizations of type \ref{item:comparisonproofcaseLhassize6:C4withoneadditionaldisjointedge} by $B\mid_{\{u,v\}}$.
Adding up our findings, it follows that if \ref{numberofmatrixrealizationsoftype:threeisolatedvertices:case:lvertu1v1rvertbackslashp1Sequals1}.\ref{numberofmatrixrealizationsoftype:twoisolatedvertices:case:lvertu1v1rvertbackslashp1Sequals1:u1doesnotequalv1andbracesu1v1bracesintersectp1Sisasingleton}.\ref{numberofmatrixrealizationsoftype:twoisolatedvertices:case:lvertu1v1rvertbackslashp1Sequals1:u1doesnotequalv1andbracesu1v1bracesintersectp1Sisasingletoncontainingu1}.\ref{numberofmatrixrealizationsoftype:twoisolatedvertices:case:lvertu1v1rvertbackslashp1Sequals1:u1doesnotequalv1andbracesu1v1bracesintersectp1Sisasingletoncontainingu1:u1equaltoa1}, then there are exactly
$2\cdot (n-1-a_1-1)\cdot ((n-1)-2) + 8\cdot (n-1-a_1-1)\cdot ((n-1)-2)
=
10\cdot (n-1-a_1-1)\cdot ((n-1)-2)$ realizations of type \ref{item:comparisonproofcaseLhassize6:oneadditionaledgeintersectingC4andoneisolatedvertex} but merely
$2\cdot (n-1-a_1-1)\cdot ((n-1)-2)$ realizations of type \ref{item:comparisonproofcaseLhassize6:C4withoneadditionaldisjointedge} by $B\mid_{\{u,v\}}$.
The case \ref{numberofmatrixrealizationsoftype:threeisolatedvertices:case:lvertu1v1rvertbackslashp1Sequals1}.\ref{numberofmatrixrealizationsoftype:twoisolatedvertices:case:lvertu1v1rvertbackslashp1Sequals1:u1doesnotequalv1andbracesu1v1bracesintersectp1Sisasingleton}.\ref{numberofmatrixrealizationsoftype:twoisolatedvertices:case:lvertu1v1rvertbackslashp1Sequals1:u1doesnotequalv1andbracesu1v1bracesintersectp1Sisasingletoncontainingu1}.\ref{numberofmatrixrealizationsoftype:twoisolatedvertices:case:lvertu1v1rvertbackslashp1Sequals1:u1doesnotequalv1andbracesu1v1bracesintersectp1Sisasingletoncontainingu1:u1equaltoc1} is again not sufficient for us to know whether \ref{item:casesfortype8casethatexactlyoneofuandvhasaprojectionmeetingaprojectionofS} or \ref{item:casesfortype8casethatbothuandvhaveaprojectionmeetingaprojectionofS} is true and we again have to consider its anonymous subcases. In the first of them, an argument entirely analogous to
the one given for the first subcase of \ref{numberofmatrixrealizationsoftype:threeisolatedvertices:case:lvertu1v1rvertbackslashp1Sequals1}.\ref{numberofmatrixrealizationsoftype:twoisolatedvertices:case:lvertu1v1rvertbackslashp1Sequals1:u1doesnotequalv1andbracesu1v1bracesintersectp1Sisasingleton}.\ref{numberofmatrixrealizationsoftype:twoisolatedvertices:case:lvertu1v1rvertbackslashp1Sequals1:u1doesnotequalv1andbracesu1v1bracesintersectp1Sisasingletoncontainingu1}.\ref{numberofmatrixrealizationsoftype:twoisolatedvertices:case:lvertu1v1rvertbackslashp1Sequals1:u1doesnotequalv1andbracesu1v1bracesintersectp1Sisasingletoncontainingu1:u1equaltoa1} proves that then property \ref{item:casesfortype8casethatexactlyoneofuandvhasaprojectionmeetingaprojectionofS} is true and therefore we know that there are exactly $2\cdot (n-1-c_1)\cdot ((n-1)-2)$ realizations of
type \ref{item:comparisonproofcaseLhassize6:oneadditionaledgeintersectingC4andoneisolatedvertex} and also exactly $2\cdot (n-1-c_1)\cdot ((n-1)-2)$ realizations of
type \ref{item:comparisonproofcaseLhassize6:C4withoneadditionaldisjointedge}
by $B\mid_{\{u,v\}}$. In the second of them, analogously to the second subcase
of \ref{numberofmatrixrealizationsoftype:threeisolatedvertices:case:lvertu1v1rvertbackslashp1Sequals1}.\ref{numberofmatrixrealizationsoftype:twoisolatedvertices:case:lvertu1v1rvertbackslashp1Sequals1:u1doesnotequalv1andbracesu1v1bracesintersectp1Sisasingleton}.\ref{numberofmatrixrealizationsoftype:twoisolatedvertices:case:lvertu1v1rvertbackslashp1Sequals1:u1doesnotequalv1andbracesu1v1bracesintersectp1Sisasingletoncontainingu1}.\ref{numberofmatrixrealizationsoftype:twoisolatedvertices:case:lvertu1v1rvertbackslashp1Sequals1:u1doesnotequalv1andbracesu1v1bracesintersectp1Sisasingletoncontainingu1:u1equaltoa1} proves that then property \ref{item:casesfortype8casethatbothuandvhaveaprojectionmeetingaprojectionofS} is true and therefore there are exactly
$4\cdot (n-1-c_1)\cdot 2\cdot ((n-1)-2) = 8\cdot (n-1-c_1)\cdot ((n-1)-2)$
realizations of type \ref{item:comparisonproofcaseLhassize6:oneadditionaledgeintersectingC4andoneisolatedvertex} and $0$ realizations of type \ref{item:comparisonproofcaseLhassize6:C4withoneadditionaldisjointedge} by $B\mid_{\{u,v\}}$.
It follows that if \ref{numberofmatrixrealizationsoftype:threeisolatedvertices:case:lvertu1v1rvertbackslashp1Sequals1}.\ref{numberofmatrixrealizationsoftype:twoisolatedvertices:case:lvertu1v1rvertbackslashp1Sequals1:u1doesnotequalv1andbracesu1v1bracesintersectp1Sisasingleton}.\ref{numberofmatrixrealizationsoftype:twoisolatedvertices:case:lvertu1v1rvertbackslashp1Sequals1:u1doesnotequalv1andbracesu1v1bracesintersectp1Sisasingletoncontainingu1}.\ref{numberofmatrixrealizationsoftype:twoisolatedvertices:case:lvertu1v1rvertbackslashp1Sequals1:u1doesnotequalv1andbracesu1v1bracesintersectp1Sisasingletoncontainingu1:u1equaltoc1}, then there are exactly
$2\cdot (n-1-c_1)\cdot ((n-1)-2) + 8\cdot (n-1-c_1)\cdot ((n-1)-2)
=
10\cdot (n-1-c_1)\cdot ((n-1)-2)$
realizations of type \ref{item:comparisonproofcaseLhassize6:oneadditionaledgeintersectingC4andoneisolatedvertex} but only $2\cdot (n-1-c_1)\cdot ((n-1)-2)$ realizations of
type \ref{item:comparisonproofcaseLhassize6:C4withoneadditionaldisjointedge} by $B\mid_{\{u,v\}}$.
It now follows that if \ref{numberofmatrixrealizationsoftype:threeisolatedvertices:case:lvertu1v1rvertbackslashp1Sequals1}.\ref{numberofmatrixrealizationsoftype:twoisolatedvertices:case:lvertu1v1rvertbackslashp1Sequals1:u1doesnotequalv1andbracesu1v1bracesintersectp1Sisasingleton}.\ref{numberofmatrixrealizationsoftype:twoisolatedvertices:case:lvertu1v1rvertbackslashp1Sequals1:u1doesnotequalv1andbracesu1v1bracesintersectp1Sisasingletoncontainingu1}, then there are exactly
$10\cdot (n-1-a_1-1)\cdot ((n-1)-2) + 10\cdot (n-1-c_1)\cdot ((n-1)-2)
=
10\cdot (2n-a_1-c_1-3) \cdot ((n-1)-2)$
realizations of type \ref{item:comparisonproofcaseLhassize6:oneadditionaledgeintersectingC4andoneisolatedvertex} but only $2\cdot( 2n-a_1-c_1-3 )\cdot( (n-1)-2 )$ of type
\ref{item:comparisonproofcaseLhassize6:C4withoneadditionaldisjointedge} by
$B\mid_{\{u,v\}}$.
The case \ref{numberofmatrixrealizationsoftype:threeisolatedvertices:case:lvertu1v1rvertbackslashp1Sequals1}.\ref{numberofmatrixrealizationsoftype:twoisolatedvertices:case:lvertu1v1rvertbackslashp1Sequals1:u1doesnotequalv1andbracesu1v1bracesintersectp1Sisasingleton}.\ref{numberofmatrixrealizationsoftype:twoisolatedvertices:case:lvertu1v1rvertbackslashp1Sequals1:u1doesnotequalv1andbracesu1v1bracesintersectp1Sisasingletoncontainingv1} will now be treated analogously to \ref{numberofmatrixrealizationsoftype:threeisolatedvertices:case:lvertu1v1rvertbackslashp1Sequals1}.\ref{numberofmatrixrealizationsoftype:twoisolatedvertices:case:lvertu1v1rvertbackslashp1Sequals1:u1doesnotequalv1andbracesu1v1bracesintersectp1Sisasingleton}.\ref{numberofmatrixrealizationsoftype:twoisolatedvertices:case:lvertu1v1rvertbackslashp1Sequals1:u1doesnotequalv1andbracesu1v1bracesintersectp1Sisasingletoncontainingu1}.
In the first subcase of \ref{numberofmatrixrealizationsoftype:threeisolatedvertices:case:lvertu1v1rvertbackslashp1Sequals1}.\ref{numberofmatrixrealizationsoftype:twoisolatedvertices:case:lvertu1v1rvertbackslashp1Sequals1:u1doesnotequalv1andbracesu1v1bracesintersectp1Sisasingleton}.\ref{numberofmatrixrealizationsoftype:twoisolatedvertices:case:lvertu1v1rvertbackslashp1Sequals1:u1doesnotequalv1andbracesu1v1bracesintersectp1Sisasingletoncontainingv1}.\ref{numberofmatrixrealizationsoftype:twoisolatedvertices:case:lvertu1v1rvertbackslashp1Sequals1:u1doesnotequalv1andbracesu1v1bracesintersectp1Sisasingletoncontainingu1:v1equaltoa1} we find that property \ref{item:casesfortype8casethatexactlyoneofuandvhasaprojectionmeetingaprojectionofS} is true and therefore there are exactly
$2\cdot (a_1-1)\cdot((n-1)-2)$ realizations of type \ref{item:comparisonproofcaseLhassize6:oneadditionaledgeintersectingC4andoneisolatedvertex} and also
$2\cdot (a_1-1)\cdot((n-1)-2)$ realizations of type \ref{item:comparisonproofcaseLhassize6:C4withoneadditionaldisjointedge} by $B\mid_{\{u,v\}}$. In the
second subcase of \ref{numberofmatrixrealizationsoftype:threeisolatedvertices:case:lvertu1v1rvertbackslashp1Sequals1}.\ref{numberofmatrixrealizationsoftype:twoisolatedvertices:case:lvertu1v1rvertbackslashp1Sequals1:u1doesnotequalv1andbracesu1v1bracesintersectp1Sisasingleton}.\ref{numberofmatrixrealizationsoftype:twoisolatedvertices:case:lvertu1v1rvertbackslashp1Sequals1:u1doesnotequalv1andbracesu1v1bracesintersectp1Sisasingletoncontainingv1}.\ref{numberofmatrixrealizationsoftype:twoisolatedvertices:case:lvertu1v1rvertbackslashp1Sequals1:u1doesnotequalv1andbracesu1v1bracesintersectp1Sisasingletoncontainingu1:v1equaltoa1} we find that property \ref{item:casesfortype8casethatbothuandvhaveaprojectionmeetingaprojectionofS} is true and therefore there are exactly
$4\cdot (a_1-1)\cdot 2\cdot ((n-1)-2) = 8\cdot (a_1-1)\cdot ((n-1)-2)$
realizations of type \ref{item:comparisonproofcaseLhassize6:oneadditionaledgeintersectingC4andoneisolatedvertex} but $0$ realizations of type \ref{item:comparisonproofcaseLhassize6:C4withoneadditionaldisjointedge} by $B\mid_{\{u,v\}}$. Therefore, if \ref{numberofmatrixrealizationsoftype:threeisolatedvertices:case:lvertu1v1rvertbackslashp1Sequals1}.\ref{numberofmatrixrealizationsoftype:twoisolatedvertices:case:lvertu1v1rvertbackslashp1Sequals1:u1doesnotequalv1andbracesu1v1bracesintersectp1Sisasingleton}.\ref{numberofmatrixrealizationsoftype:twoisolatedvertices:case:lvertu1v1rvertbackslashp1Sequals1:u1doesnotequalv1andbracesu1v1bracesintersectp1Sisasingletoncontainingv1}.\ref{numberofmatrixrealizationsoftype:twoisolatedvertices:case:lvertu1v1rvertbackslashp1Sequals1:u1doesnotequalv1andbracesu1v1bracesintersectp1Sisasingletoncontainingu1:v1equaltoa1}, then there
are exactly $2\cdot (a_1-1)\cdot((n-1)-2) + 8\cdot (a_1-1)\cdot ((n-1)-2)
= 10\cdot (a_1-1)\cdot((n-1)-2)$ realizations of type \ref{item:comparisonproofcaseLhassize6:oneadditionaledgeintersectingC4andoneisolatedvertex} but only
$2\cdot (a_1-1)\cdot ( (n-1)-2)$ realizations of type \ref{item:comparisonproofcaseLhassize6:C4withoneadditionaldisjointedge} by $B\mid_{\{u,v\}}$.
In the first subcase of \ref{numberofmatrixrealizationsoftype:threeisolatedvertices:case:lvertu1v1rvertbackslashp1Sequals1}.\ref{numberofmatrixrealizationsoftype:twoisolatedvertices:case:lvertu1v1rvertbackslashp1Sequals1:u1doesnotequalv1andbracesu1v1bracesintersectp1Sisasingleton}.\ref{numberofmatrixrealizationsoftype:twoisolatedvertices:case:lvertu1v1rvertbackslashp1Sequals1:u1doesnotequalv1andbracesu1v1bracesintersectp1Sisasingletoncontainingv1}.\ref{numberofmatrixrealizationsoftype:twoisolatedvertices:case:lvertu1v1rvertbackslashp1Sequals1:u1doesnotequalv1andbracesu1v1bracesintersectp1Sisasingletoncontainingu1:v1equaltoc1} we conclude that property \ref{item:casesfortype8casethatexactlyoneofuandvhasaprojectionmeetingaprojectionofS} is true and therefore there exist
exactly $2\cdot (c_1-1-1)\cdot ((n-1)-2)$ realizations of type \ref{item:comparisonproofcaseLhassize6:oneadditionaledgeintersectingC4andoneisolatedvertex} and also
$2\cdot (c_1-1-1)\cdot ((n-1)-2)$ realizations of type \ref{item:comparisonproofcaseLhassize6:C4withoneadditionaldisjointedge} by $B\mid_{\{u,v\}}$.
In the second subcase of \ref{numberofmatrixrealizationsoftype:threeisolatedvertices:case:lvertu1v1rvertbackslashp1Sequals1}.\ref{numberofmatrixrealizationsoftype:twoisolatedvertices:case:lvertu1v1rvertbackslashp1Sequals1:u1doesnotequalv1andbracesu1v1bracesintersectp1Sisasingleton}.\ref{numberofmatrixrealizationsoftype:twoisolatedvertices:case:lvertu1v1rvertbackslashp1Sequals1:u1doesnotequalv1andbracesu1v1bracesintersectp1Sisasingletoncontainingv1}.\ref{numberofmatrixrealizationsoftype:twoisolatedvertices:case:lvertu1v1rvertbackslashp1Sequals1:u1doesnotequalv1andbracesu1v1bracesintersectp1Sisasingletoncontainingu1:v1equaltoc1} we conclude that property \ref{item:casesfortype8casethatbothuandvhaveaprojectionmeetingaprojectionofS} is true and therefore there are exactly
$4\cdot (c_1-1-1)\cdot 2 \cdot ((n-1)-2) = 8\cdot (c_1-1-1)\cdot ((n-1)-2)$
realizations of type \ref{item:comparisonproofcaseLhassize6:oneadditionaledgeintersectingC4andoneisolatedvertex} and $0$ realizations of type \ref{item:comparisonproofcaseLhassize6:C4withoneadditionaldisjointedge} by $B\mid_{\{u,v\}}$. Therefore, if \ref{numberofmatrixrealizationsoftype:threeisolatedvertices:case:lvertu1v1rvertbackslashp1Sequals1}.\ref{numberofmatrixrealizationsoftype:twoisolatedvertices:case:lvertu1v1rvertbackslashp1Sequals1:u1doesnotequalv1andbracesu1v1bracesintersectp1Sisasingleton}.\ref{numberofmatrixrealizationsoftype:twoisolatedvertices:case:lvertu1v1rvertbackslashp1Sequals1:u1doesnotequalv1andbracesu1v1bracesintersectp1Sisasingletoncontainingv1}.\ref{numberofmatrixrealizationsoftype:twoisolatedvertices:case:lvertu1v1rvertbackslashp1Sequals1:u1doesnotequalv1andbracesu1v1bracesintersectp1Sisasingletoncontainingu1:v1equaltoc1}, then there are exactly
$2\cdot (c_1-1-1)\cdot ((n-1)-2) + 8\cdot (c_1-1-1)\cdot ((n-1)-2)
= 10\cdot (c_1-1-1)\cdot ((n-1)-2)$ realizations of
type \ref{item:comparisonproofcaseLhassize6:oneadditionaledgeintersectingC4andoneisolatedvertex} but only $2\cdot (c_1-1-1)\cdot ((n-1)-2)$ realizations of type
\ref{item:comparisonproofcaseLhassize6:C4withoneadditionaldisjointedge} by
$B\mid_{\{u,v\}}$.
It now follows that if \ref{numberofmatrixrealizationsoftype:threeisolatedvertices:case:lvertu1v1rvertbackslashp1Sequals1}.\ref{numberofmatrixrealizationsoftype:twoisolatedvertices:case:lvertu1v1rvertbackslashp1Sequals1:u1doesnotequalv1andbracesu1v1bracesintersectp1Sisasingleton}.\ref{numberofmatrixrealizationsoftype:twoisolatedvertices:case:lvertu1v1rvertbackslashp1Sequals1:u1doesnotequalv1andbracesu1v1bracesintersectp1Sisasingletoncontainingv1}, then there are exactly
$10\cdot (a_1-1)\cdot((n-1)-2) + 10\cdot (c_1-1-1)\cdot ((n-1)-2)
= 10\cdot (a_1+c_1-3) \cdot ((n-1)-2)$
realizations of type \ref{item:comparisonproofcaseLhassize6:oneadditionaledgeintersectingC4andoneisolatedvertex} but merely
$2 \cdot (a_1-1)\cdot((n-1)-2) + 2 \cdot (c_1-1-1)\cdot ((n-1)-2)
= 2 \cdot (a_1+c_1-3) \cdot ((n-1)-2)$ realizations of type \ref{item:comparisonproofcaseLhassize6:C4withoneadditionaldisjointedge} by $B\mid_{\{u,v\}}$. Moreover we may now
conclude that if \ref{numberofmatrixrealizationsoftype:threeisolatedvertices:case:lvertu1v1rvertbackslashp1Sequals1}.\ref{numberofmatrixrealizationsoftype:twoisolatedvertices:case:lvertu1v1rvertbackslashp1Sequals1:u1doesnotequalv1andbracesu1v1bracesintersectp1Sisasingleton}, then there are exactly
$10\cdot (2n-a_1-c_1-3) \cdot ((n-1)-2) + 10\cdot (a_1+c_1-3) \cdot ((n-1)-2)
=
20\cdot ((n-1)-2)^2$ realizations of type \ref{item:comparisonproofcaseLhassize6:oneadditionaledgeintersectingC4andoneisolatedvertex} but only
$2 \cdot (2n-a_1-c_1-3) \cdot ((n-1)-2) + 2 \cdot (a_1+c_1-3) \cdot ((n-1)-2)
=
4 \cdot ((n-1)-2)^2$ realizations of type \ref{item:comparisonproofcaseLhassize6:C4withoneadditionaldisjointedge} by $B\mid_{\{u,v\}}$. Finally we can conclude that
if \ref{numberofmatrixrealizationsoftype:threeisolatedvertices:case:lvertu1v1rvertbackslashp1Sequals1}, then there are exactly
$4\cdot ((n-1)-2)^2 + 20\cdot ((n-1)-2)^2 = 24\cdot ((n-1)-2)^2$ realizations
of type \ref{item:comparisonproofcaseLhassize6:oneadditionaledgeintersectingC4andoneisolatedvertex} but only $4\cdot ((n-1)-2)^2 + 4\cdot((n-1)-2)^2 = 8\cdot ((n-1)-2)^2$
realizations of type \ref{item:comparisonproofcaseLhassize6:C4withoneadditionaldisjointedge} by $B\mid_{\{u,v\}}$.
If \ref{numberofmatrixrealizationsoftype:twoisolatedvertices:case:lvertu1v1rvertbackslashp1Sequals2}, then the inclusion $\{u_2,v_2\}\subseteq\mathrm{p}_2(S)$ alone implies that
property \ref{item:casesfortype8casethatbothuandvhaveaprojectionmeetingaprojectionofS}
is true and therefore there are exactly
$4\cdot 4\cdot \binom{(n-1)-2}{2} = 16\cdot \binom{(n-1)-2}{2}$
realizations of type \ref{item:comparisonproofcaseLhassize6:oneadditionaledgeintersectingC4andoneisolatedvertex} but $0$ realizations of type \ref{item:comparisonproofcaseLhassize6:C4withoneadditionaldisjointedge} by $B\mid_{\{u,v\}}$.
Summing up, it follows that for each fixed $S$ there are exactly
$2\cdot 16\cdot \binom{(n-1)-2}{2} + 24\cdot ((n-1)-2)^2 = 24\cdot(n-3)^2 + 32\cdot\binom{n-3}{2}$ realizations of type \ref{item:comparisonproofcaseLhassize6:oneadditionaledgeintersectingC4andoneisolatedvertex} but only $8\cdot ((n-1)-2)^2$ realizations of type \ref{item:comparisonproofcaseLhassize6:C4withoneadditionaldisjointedge} by $B\mid_{\{u,v\}}$. This completes the proof of both \ref{numberofmatrixrealizationsoftype:oneadditionaledgeintersectingC4andoneisolatedvertex} and \ref{numberofmatrixrealizationsoftype:C4withoneadditionaldisjointedge}.
We can now turn to counting the realizations of \ref{item:comparisonproofcaseLhassize6:C4intersectingtwoedgesinseparatenonadjacentvertices}--\ref{item:comparisonproofcaseLhassize6:C4intersectingatwopathinitsinnervertex}, i.e. to proving \ref{numberofmatrixrealizationsoftype:C4intersectingtwoedgesinseparatenonadjacentvertices}--\ref{numberofmatrixrealizationsoftype:C4intersectingatwopathinitsinnervertex}
By \ref{characterizationoftype:C4intersectingtwoedgesinseparatenonadjacentvertices:property1}, \ref{characterizationoftype:C4intersectingtwoedgesinseparateadjacentvertices:property1}, \ref{characterizationoftype:C4intersectingatwopathinanendvertex:property1} and \ref{characterizationoftype:C4intersectingatwopathinitsinnervertex:property1}, for each of the four types \ref{item:comparisonproofcaseLhassize6:C4intersectingtwoedgesinseparatenonadjacentvertices}, \ref{item:comparisonproofcaseLhassize6:C4intersectingtwoedgesinseparateadjacentvertices}, \ref{item:comparisonproofcaseLhassize6:C4intersectingatwopathinanendvertex} and \ref{item:comparisonproofcaseLhassize6:C4intersectingatwopathinitsinnervertex} it is necessary that $\mathrm{X}_{\{0\}^{u,v}\sqcup B\mid_{S}} \cong \ref{item:comparisonproofcaseLhassize6:twoisolatedvertices}$.
We may therefore determine
each of the four functions \ref{numberofmatrixrealizationsoftype:C4intersectingtwoedgesinseparatenonadjacentvertices}, \ref{numberofmatrixrealizationsoftype:C4intersectingtwoedgesinseparateadjacentvertices}, \ref{numberofmatrixrealizationsoftype:C4intersectingatwopathinanendvertex}, \ref{numberofmatrixrealizationsoftype:C4intersectingatwopathinitsinnervertex} in the course of one reexamination of the proof of
\ref{numberofmatrixrealizationsoftype:twoisolatedvertices}.
We consider each of the cases in turn, each time descending down just deep enough
until we are able to decide which of the four isomorphism types \ref{item:comparisonproofcaseLhassize6:C4intersectingtwoedgesinseparatenonadjacentvertices}, \ref{item:comparisonproofcaseLhassize6:C4intersectingtwoedgesinseparateadjacentvertices}, \ref{item:comparisonproofcaseLhassize6:C4intersectingatwopathinanendvertex} can be realized in
that case.
Since by \ref{characterizationoftype:C4intersectingtwoedgesinseparatenonadjacentvertices:property2}, \ref{characterizationoftype:C4intersectingtwoedgesinseparateadjacentvertices:property2}, \ref{characterizationoftype:C4intersectingatwopathinanendvertex:property2} and \ref{characterizationoftype:C4intersectingatwopathinitsinnervertex:property2} the property $B[u]\in\{\pm\}$ and $B[v]\in \{\pm\}$ is necessary for each of
the four types, the positions $u$ and $v$ alone, not $B\mid_{\{u,v\}}$ itself, decide
about which type can be realized. Therefore, if a decision is reached about which
of the four types can be realized in a case, then we obtain the number of
realizations by multiplying the number of realizations of
type \ref{item:comparisonproofcaseLhassize6:twoisolatedvertices} in that particular
case by $4$.
We now reexamine \ref{numberofmatrixrealizationsoftype:twoisolatedvertices:case:lvertu1v1rvertbackslashp1Sequals0}.
If \ref{numberofmatrixrealizationsoftype:twoisolatedvertices:case:lvertu1v1rvertbackslashp1Sequals0}, then the property $\lvert \{u_1,v_1\}\setminus \mathrm{p}_1(S)\rvert = 0$
makes \ref{characterizationoftype:C4intersectingatwopathinanendvertex:property4}
impossible, hence in the entire case \ref{numberofmatrixrealizationsoftype:twoisolatedvertices:case:lvertu1v1rvertbackslashp1Sequals0} the type \ref{item:comparisonproofcaseLhassize6:C4intersectingatwopathinanendvertex} is impossible.
If \ref{numberofmatrixrealizationsoftype:twoisolatedvertices:case:lvertu1v1rvertbackslashp1Sequals0}.\ref{numberofmatrixrealizationsoftype:twoisolatedvertices:case:lvertu1v1rvertbackslashp1Sequals0:case:u1equalsv1}, then we have $u_1=v_1$, which
makes both \ref{characterizationoftype:C4intersectingtwoedgesinseparateadjacentvertices:property3} and \ref{characterizationoftype:C4intersectingtwoedgesinseparatenonadjacentvertices:property3} impossible. The only type remaining is \ref{item:comparisonproofcaseLhassize6:C4intersectingatwopathinitsinnervertex} (and all the properties \ref{characterizationoftype:C4intersectingatwopathinitsinnervertex:property1}--\ref{characterizationoftype:C4intersectingatwopathinitsinnervertex:property4} are indeed satisfied). It
follows that if \ref{numberofmatrixrealizationsoftype:twoisolatedvertices:case:lvertu1v1rvertbackslashp1Sequals0}.\ref{numberofmatrixrealizationsoftype:twoisolatedvertices:case:lvertu1v1rvertbackslashp1Sequals0:case:u1equalsv1}, then there are exactly
$8\cdot\binom{(n-1)-2}{2}$ realizations of type \ref{item:comparisonproofcaseLhassize6:C4intersectingatwopathinitsinnervertex} by $B\mid_{\{u,v\}}$.
If \ref{numberofmatrixrealizationsoftype:twoisolatedvertices:case:lvertu1v1rvertbackslashp1Sequals0}.\ref{numberofmatrixrealizationsoftype:twoisolatedvertices:case:lvertu1v1rvertbackslashp1Sequals0:case:u1doesnotequalv1}, then we have $u_1\neq v_1$.
Since we also have $u_2\neq v_2$ throughout \ref{numberofmatrixrealizationsoftype:twoisolatedvertices:case:lvertu1v1rvertbackslashp1Sequals0}, this
makes \ref{characterizationoftype:C4intersectingatwopathinitsinnervertex:property3}
impossible. Moreover, since throughout \ref{numberofmatrixrealizationsoftype:twoisolatedvertices:case:lvertu1v1rvertbackslashp1Sequals0} we also have \eqref{eq:proofofcomparativecountingtheoremcase2firsteq}, in particular
$\{ u_2, v_2 \} \cap \mathrm{p}_2(S) = \emptyset$, it follows that \ref{characterizationoftype:C4intersectingtwoedgesinseparateadjacentvertices:property4} is impossible. The only
type remaining is \ref{item:comparisonproofcaseLhassize6:C4intersectingtwoedgesinseparatenonadjacentvertices} (and all the properties \ref{characterizationoftype:C4intersectingtwoedgesinseparatenonadjacentvertices:property1}--\ref{characterizationoftype:C4intersectingtwoedgesinseparatenonadjacentvertices:property4} are indeed satisfied;
notice in particular that in \ref{characterizationoftype:C4intersectingtwoedgesinseparatenonadjacentvertices:property4} the first of the two mutually exclusive clauses
is true). If follows that if \ref{numberofmatrixrealizationsoftype:twoisolatedvertices:case:lvertu1v1rvertbackslashp1Sequals0}.\ref{numberofmatrixrealizationsoftype:twoisolatedvertices:case:lvertu1v1rvertbackslashp1Sequals0:case:u1doesnotequalv1}, then
there are exactly $8\cdot\binom{(n-1)-2}{2}$ realizations of
type \ref{item:comparisonproofcaseLhassize6:C4intersectingtwoedgesinseparatenonadjacentvertices} by $B\mid_{\{u,v\}}$. This completes our reexamination of \ref{numberofmatrixrealizationsoftype:twoisolatedvertices:case:lvertu1v1rvertbackslashp1Sequals0}.
We now reexamine \ref{numberofmatrixrealizationsoftype:twoisolatedvertices:case:lvertu1v1rvertbackslashp1Sequals1}. The information defining \ref{numberofmatrixrealizationsoftype:twoisolatedvertices:case:lvertu1v1rvertbackslashp1Sequals1} is by itself
not yet sufficient to rule out any of the four types \ref{item:comparisonproofcaseLhassize6:C4intersectingtwoedgesinseparatenonadjacentvertices}--\ref{item:comparisonproofcaseLhassize6:C4intersectingatwopathinitsinnervertex}. The information defining \ref{numberofmatrixrealizationsoftype:twoisolatedvertices:case:lvertu1v1rvertbackslashp1Sequals1}.\ref{numberofmatrixrealizationsoftype:twoisolatedvertices:case:lvertu1v1rvertbackslashp1Sequals1:u1equalsv1andbracesu1v1bracesintersectp1Sisempty}, more
specificly $u_1=v_1$, makes both \ref{characterizationoftype:C4intersectingtwoedgesinseparatenonadjacentvertices:property3} and \ref{characterizationoftype:C4intersectingtwoedgesinseparateadjacentvertices:property3} impossible, still leaving two types.
We argued in \ref{numberofmatrixrealizationsoftype:twoisolatedvertices:case:lvertu1v1rvertbackslashp1Sequals1}.\ref{numberofmatrixrealizationsoftype:twoisolatedvertices:case:lvertu1v1rvertbackslashp1Sequals1:u1equalsv1andbracesu1v1bracesintersectp1Sisempty} that we have $u_2\neq v_2$ and $\lvert\{u_2,v_2\}\cap\mathrm{p}_2(S)\rvert = 1$ in this
case. If the latter is true as $u_2\in\mathrm{p}_2(S)$, then $v_2\notin\mathrm{p}_2(S)$, and
combining this information with $v_1\notin\mathrm{p}_1(S)$ (which we know since we are
in \ref{numberofmatrixrealizationsoftype:twoisolatedvertices:case:lvertu1v1rvertbackslashp1Sequals1}.\ref{numberofmatrixrealizationsoftype:twoisolatedvertices:case:lvertu1v1rvertbackslashp1Sequals1:u1equalsv1andbracesu1v1bracesintersectp1Sisempty}) makes the
second clause of the conjunction \ref{characterizationoftype:C4intersectingatwopathinitsinnervertex:property4} impossible. If on the other hand it is true
as $v_2\in\mathrm{p}_2(S)$, then $u_2\notin\mathrm{p}_2(S)$, and combining this
with $u_1\notin\mathrm{p}_1(S)$ (which again we know since we are in \ref{numberofmatrixrealizationsoftype:twoisolatedvertices:case:lvertu1v1rvertbackslashp1Sequals1}.\ref{numberofmatrixrealizationsoftype:twoisolatedvertices:case:lvertu1v1rvertbackslashp1Sequals1:u1equalsv1andbracesu1v1bracesintersectp1Sisempty}) makes \ref{characterizationoftype:C4intersectingatwopathinitsinnervertex:property4} impossible (this time, the first clause). This rules out type \ref{item:comparisonproofcaseLhassize6:C4intersectingatwopathinitsinnervertex}.
The only type remaining is \ref{item:comparisonproofcaseLhassize6:C4intersectingatwopathinanendvertex} (and all the properties \ref{characterizationoftype:C4intersectingatwopathinanendvertex:property1}--\ref{characterizationoftype:C4intersectingatwopathinanendvertex:property4} are indeed satisfied; note that due
to $\lvert \{u_2,v_2\}\cap\mathrm{p}_2(S)\rvert = 1$ the two clauses of \ref{characterizationoftype:C4intersectingatwopathinanendvertex:property4} are mutually exclusive, the
second being true if $u_2\in\mathrm{p}_2(S)$ and the first if $v_2\in\mathrm{p}_2(S)$).
It follows that in case \ref{numberofmatrixrealizationsoftype:twoisolatedvertices:case:lvertu1v1rvertbackslashp1Sequals1:u1equalsv1andbracesu1v1bracesintersectp1Sisempty} of \ref{numberofmatrixrealizationsoftype:twoisolatedvertices:case:lvertu1v1rvertbackslashp1Sequals1} there are exactly $8\cdot ((n-1)-2)^2$ realizations of type \ref{item:comparisonproofcaseLhassize6:C4intersectingatwopathinanendvertex} by $B\mid_{\{u,v\}}$.
The information defining \ref{numberofmatrixrealizationsoftype:twoisolatedvertices:case:lvertu1v1rvertbackslashp1Sequals1}.\ref{numberofmatrixrealizationsoftype:twoisolatedvertices:case:lvertu1v1rvertbackslashp1Sequals1:u1doesnotequalv1andbracesu1v1bracesintersectp1Sisasingleton} is not enough to rule out any of the four types
\ref{item:comparisonproofcaseLhassize6:C4intersectingtwoedgesinseparatenonadjacentvertices}--\ref{item:comparisonproofcaseLhassize6:C4intersectingatwopathinitsinnervertex},
and descending one level deeper to \ref{numberofmatrixrealizationsoftype:twoisolatedvertices:case:lvertu1v1rvertbackslashp1Sequals1}.\ref{numberofmatrixrealizationsoftype:twoisolatedvertices:case:lvertu1v1rvertbackslashp1Sequals1:u1doesnotequalv1andbracesu1v1bracesintersectp1Sisasingleton}.\ref{numberofmatrixrealizationsoftype:twoisolatedvertices:case:lvertu1v1rvertbackslashp1Sequals1:u1doesnotequalv1andbracesu1v1bracesintersectp1Sisasingletoncontainingu1} does not change this.
If \ref{numberofmatrixrealizationsoftype:twoisolatedvertices:case:lvertu1v1rvertbackslashp1Sequals1}.\ref{numberofmatrixrealizationsoftype:twoisolatedvertices:case:lvertu1v1rvertbackslashp1Sequals1:u1doesnotequalv1andbracesu1v1bracesintersectp1Sisasingleton}.\ref{numberofmatrixrealizationsoftype:twoisolatedvertices:case:lvertu1v1rvertbackslashp1Sequals1:u1doesnotequalv1andbracesu1v1bracesintersectp1Sisasingletoncontainingu1}.\ref{numberofmatrixrealizationsoftype:twoisolatedvertices:case:lvertu1v1rvertbackslashp1Sequals1:u1doesnotequalv1andbracesu1v1bracesintersectp1Sisasingletoncontainingu1:u1equaltoa1}, then the decision still cannot be made and depends on the (anonymous) subcases
which we distinguished in that case, namely whether the first or the second clause of
\eqref{eq:proofofcomparativecountingtheoremcase3secondeq} is true:
If \ref{numberofmatrixrealizationsoftype:twoisolatedvertices:case:lvertu1v1rvertbackslashp1Sequals1}.\ref{numberofmatrixrealizationsoftype:twoisolatedvertices:case:lvertu1v1rvertbackslashp1Sequals1:u1doesnotequalv1andbracesu1v1bracesintersectp1Sisasingleton}.\ref{numberofmatrixrealizationsoftype:twoisolatedvertices:case:lvertu1v1rvertbackslashp1Sequals1:u1doesnotequalv1andbracesu1v1bracesintersectp1Sisasingletoncontainingu1}.\ref{numberofmatrixrealizationsoftype:twoisolatedvertices:case:lvertu1v1rvertbackslashp1Sequals1:u1doesnotequalv1andbracesu1v1bracesintersectp1Sisasingletoncontainingu1:u1equaltoa1}, and the first clause of \eqref{eq:proofofcomparativecountingtheoremcase3secondeq}
is true, then in particular we know that
$v_1\notin\mathrm{p}_1(S)$ and $u_2=v_2\notin\mathrm{p}_2(S)$. The latter contradicts \ref{characterizationoftype:C4intersectingtwoedgesinseparatenonadjacentvertices:property3} and \ref{characterizationoftype:C4intersectingtwoedgesinseparateadjacentvertices:property3}. Moreover,
$v_1\notin\mathrm{p}_1(S)$ and $v_2\notin\mathrm{p}_2(S)$ combined render the second clause of the
conjunction \ref{characterizationoftype:C4intersectingatwopathinitsinnervertex:property4} false. Note that for each of three discarded types we used in whole or in part
the information $u_2=v_2\notin\mathrm{p}_2(S)$, which defines the present subcase.
Hence deferring any decision about the types for so so long was necessary.
We are now left with only the type
\ref{item:comparisonproofcaseLhassize6:C4intersectingatwopathinanendvertex} (and indeed the properties \ref{characterizationoftype:C4intersectingatwopathinanendvertex:property1}--\ref{characterizationoftype:C4intersectingatwopathinanendvertex:property4} are all satisfied). Since within \ref{numberofmatrixrealizationsoftype:twoisolatedvertices:case:lvertu1v1rvertbackslashp1Sequals1}.\ref{numberofmatrixrealizationsoftype:twoisolatedvertices:case:lvertu1v1rvertbackslashp1Sequals1:u1doesnotequalv1andbracesu1v1bracesintersectp1Sisasingleton}.\ref{numberofmatrixrealizationsoftype:twoisolatedvertices:case:lvertu1v1rvertbackslashp1Sequals1:u1doesnotequalv1andbracesu1v1bracesintersectp1Sisasingletoncontainingu1}.\ref{numberofmatrixrealizationsoftype:twoisolatedvertices:case:lvertu1v1rvertbackslashp1Sequals1:u1doesnotequalv1andbracesu1v1bracesintersectp1Sisasingletoncontainingu1:u1equaltoa1} we found that in this
situation there are exactly $(n-1-a_1-1)\cdot ((n-1)-2)$ realizations of type
\ref{item:comparisonproofcaseLhassize6:twoisolatedvertices} by $B\mid_{\{u,v\}}$, it follows
that here there are exactly $4\cdot (n-1-a_1-1)\cdot ((n-1)-2)$ realizations of type
\ref{item:comparisonproofcaseLhassize6:C4intersectingatwopathinanendvertex}.
If \ref{numberofmatrixrealizationsoftype:twoisolatedvertices:case:lvertu1v1rvertbackslashp1Sequals1}.\ref{numberofmatrixrealizationsoftype:twoisolatedvertices:case:lvertu1v1rvertbackslashp1Sequals1:u1doesnotequalv1andbracesu1v1bracesintersectp1Sisasingleton}.\ref{numberofmatrixrealizationsoftype:twoisolatedvertices:case:lvertu1v1rvertbackslashp1Sequals1:u1doesnotequalv1andbracesu1v1bracesintersectp1Sisasingletoncontainingu1}.\ref{numberofmatrixrealizationsoftype:twoisolatedvertices:case:lvertu1v1rvertbackslashp1Sequals1:u1doesnotequalv1andbracesu1v1bracesintersectp1Sisasingletoncontainingu1:u1equaltoa1} and the second clause of \eqref{eq:proofofcomparativecountingtheoremcase3secondeq}
is true, then we know that $u_1\neq v_1$, $u_1 \in\mathrm{p}_1(S)$,
$v_1\notin\mathrm{p}_1(S)$, $u_2\neq v_2$, $u_2\notin\mathrm{p}_2(S)$ and $v_2\in\mathrm{p}_2(S)$.
We can now rule out three types: properties $v_1\notin\mathrm{p}_1(S)$ and $u_2\notin\mathrm{p}_2(S)$ combined
render both clauses of the disjunction \ref{characterizationoftype:C4intersectingtwoedgesinseparatenonadjacentvertices:property4} false.
Properties $u_1\in\mathrm{p}_1(S)$ and $v_2\in\mathrm{p}_2(S)$ combined render both clauses of
the disjunction \ref{characterizationoftype:C4intersectingatwopathinanendvertex:property4} false.
Properties $u_1\neq v_1$ and $u_2\neq v_2$ proves \ref{characterizationoftype:C4intersectingatwopathinitsinnervertex:property3} to be false.
Again note that in all three decisions we used the information defining the present
subcase. The only type remaining now is \ref{item:comparisonproofcaseLhassize6:C4intersectingtwoedgesinseparateadjacentvertices}, and indeed all properties \ref{characterizationoftype:C4intersectingtwoedgesinseparateadjacentvertices:property1}--\ref{characterizationoftype:C4intersectingtwoedgesinseparateadjacentvertices:property4} are
satisfied (in \ref{characterizationoftype:C4intersectingtwoedgesinseparateadjacentvertices:property4} only the first clause of the disjunction). Since in \ref{numberofmatrixrealizationsoftype:twoisolatedvertices:case:lvertu1v1rvertbackslashp1Sequals1}.\ref{numberofmatrixrealizationsoftype:twoisolatedvertices:case:lvertu1v1rvertbackslashp1Sequals1:u1doesnotequalv1andbracesu1v1bracesintersectp1Sisasingleton}.\ref{numberofmatrixrealizationsoftype:twoisolatedvertices:case:lvertu1v1rvertbackslashp1Sequals1:u1doesnotequalv1andbracesu1v1bracesintersectp1Sisasingletoncontainingu1}.\ref{numberofmatrixrealizationsoftype:twoisolatedvertices:case:lvertu1v1rvertbackslashp1Sequals1:u1doesnotequalv1andbracesu1v1bracesintersectp1Sisasingletoncontainingu1:u1equaltoa1} we found that in this situation
there are exactly $(n-1-a_1-1)\cdot 2\cdot ((n-1)-2)$ realizations of
type \ref{item:comparisonproofcaseLhassize6:twoisolatedvertices} by $B\mid_{\{u,v\}}$,
it follows that here there are exactly $8\cdot (n-1-a_1-1)\cdot ((n-1)-2)$
realizations of type \ref{item:comparisonproofcaseLhassize6:C4intersectingtwoedgesinseparateadjacentvertices}.
If \ref{numberofmatrixrealizationsoftype:twoisolatedvertices:case:lvertu1v1rvertbackslashp1Sequals1}.\ref{numberofmatrixrealizationsoftype:twoisolatedvertices:case:lvertu1v1rvertbackslashp1Sequals1:u1doesnotequalv1andbracesu1v1bracesintersectp1Sisasingleton}.\ref{numberofmatrixrealizationsoftype:twoisolatedvertices:case:lvertu1v1rvertbackslashp1Sequals1:u1doesnotequalv1andbracesu1v1bracesintersectp1Sisasingletoncontainingu1}.\ref{numberofmatrixrealizationsoftype:twoisolatedvertices:case:lvertu1v1rvertbackslashp1Sequals1:u1doesnotequalv1andbracesu1v1bracesintersectp1Sisasingletoncontainingu1:u1equaltoc1}, then again we have to distinguish whether the first or the second clause of
\eqref{eq:proofofcomparativecountingtheoremcase3secondeq} is true to reach a
conclusion:
If \ref{numberofmatrixrealizationsoftype:twoisolatedvertices:case:lvertu1v1rvertbackslashp1Sequals1}.\ref{numberofmatrixrealizationsoftype:twoisolatedvertices:case:lvertu1v1rvertbackslashp1Sequals1:u1doesnotequalv1andbracesu1v1bracesintersectp1Sisasingleton}.\ref{numberofmatrixrealizationsoftype:twoisolatedvertices:case:lvertu1v1rvertbackslashp1Sequals1:u1doesnotequalv1andbracesu1v1bracesintersectp1Sisasingletoncontainingu1}.\ref{numberofmatrixrealizationsoftype:twoisolatedvertices:case:lvertu1v1rvertbackslashp1Sequals1:u1doesnotequalv1andbracesu1v1bracesintersectp1Sisasingletoncontainingu1:u1equaltoc1} and the first clause of \eqref{eq:proofofcomparativecountingtheoremcase3secondeq}
is true, then again we in particular know that
$v_1\notin\mathrm{p}_1(S)$ and $u_2=v_2\notin\mathrm{p}_2(S)$, and therefore an argument
analogous to the one given for the first subcase of \ref{numberofmatrixrealizationsoftype:twoisolatedvertices:case:lvertu1v1rvertbackslashp1Sequals1}.\ref{numberofmatrixrealizationsoftype:twoisolatedvertices:case:lvertu1v1rvertbackslashp1Sequals1:u1doesnotequalv1andbracesu1v1bracesintersectp1Sisasingleton}.\ref{numberofmatrixrealizationsoftype:twoisolatedvertices:case:lvertu1v1rvertbackslashp1Sequals1:u1doesnotequalv1andbracesu1v1bracesintersectp1Sisasingletoncontainingu1}.\ref{numberofmatrixrealizationsoftype:twoisolatedvertices:case:lvertu1v1rvertbackslashp1Sequals1:u1doesnotequalv1andbracesu1v1bracesintersectp1Sisasingletoncontainingu1:u1equaltoa1} shows that in the present
situation there are exactly $4\cdot (n-1-c_1)\cdot((n-1)-2)$ realizations of
type \ref{item:comparisonproofcaseLhassize6:C4intersectingatwopathinanendvertex}
by $B\mid_{\{u,v\}}$ and no other type possible.
If \ref{numberofmatrixrealizationsoftype:twoisolatedvertices:case:lvertu1v1rvertbackslashp1Sequals1}.\ref{numberofmatrixrealizationsoftype:twoisolatedvertices:case:lvertu1v1rvertbackslashp1Sequals1:u1doesnotequalv1andbracesu1v1bracesintersectp1Sisasingleton}.\ref{numberofmatrixrealizationsoftype:twoisolatedvertices:case:lvertu1v1rvertbackslashp1Sequals1:u1doesnotequalv1andbracesu1v1bracesintersectp1Sisasingletoncontainingu1}.\ref{numberofmatrixrealizationsoftype:twoisolatedvertices:case:lvertu1v1rvertbackslashp1Sequals1:u1doesnotequalv1andbracesu1v1bracesintersectp1Sisasingletoncontainingu1:u1equaltoc1} and the second clause of \eqref{eq:proofofcomparativecountingtheoremcase3secondeq}
is true, then again we know that $u_1\neq v_1$, $u_1\in\mathrm{p}_1(S)$,
$v_1\notin\mathrm{p}_1(S)$ and $u_2\neq v_2$,
but this time we have $u_2\notin \mathrm{p}_2(S)$ and $v_2\in\mathrm{p}_2(S)$. We can now
rule out three types: properties
$v_1\notin\mathrm{p}_1(S)$ and $u_2\notin\mathrm{p}_2(S)$ combined
render both clauses of the disjunction \ref{characterizationoftype:C4intersectingtwoedgesinseparatenonadjacentvertices:property4} false.
Properties $u_1\in\mathrm{p}_1(S)$ and $v_2\in\mathrm{p}_2(S)$ combined render both clauses of
the disjunction \ref{characterizationoftype:C4intersectingatwopathinanendvertex:property4} false. Properties $u_1\neq v_1$ and $u_2\neq v_2$ contradict \ref{characterizationoftype:C4intersectingatwopathinitsinnervertex:property3}. What
remains is type \ref{item:comparisonproofcaseLhassize6:C4intersectingtwoedgesinseparateadjacentvertices}, and all properties \ref{characterizationoftype:C4intersectingtwoedgesinseparateadjacentvertices:property1}--\ref{characterizationoftype:C4intersectingtwoedgesinseparateadjacentvertices:property4} are satisfied (in \ref{characterizationoftype:C4intersectingtwoedgesinseparateadjacentvertices:property4} only the first clause of the disjunction). Since in \ref{numberofmatrixrealizationsoftype:twoisolatedvertices:case:lvertu1v1rvertbackslashp1Sequals1}.\ref{numberofmatrixrealizationsoftype:twoisolatedvertices:case:lvertu1v1rvertbackslashp1Sequals1:u1doesnotequalv1andbracesu1v1bracesintersectp1Sisasingleton}.\ref{numberofmatrixrealizationsoftype:twoisolatedvertices:case:lvertu1v1rvertbackslashp1Sequals1:u1doesnotequalv1andbracesu1v1bracesintersectp1Sisasingletoncontainingu1}.\ref{numberofmatrixrealizationsoftype:twoisolatedvertices:case:lvertu1v1rvertbackslashp1Sequals1:u1doesnotequalv1andbracesu1v1bracesintersectp1Sisasingletoncontainingu1:u1equaltoc1} we found that in this
situation there are exactly $2\cdot (n-1-c_1)\cdot ((n-1)-2)$ realizations of
type \ref{item:comparisonproofcaseLhassize6:twoisolatedvertices} by $B\mid_{\{u,v\}}$,
it follows that here there are exactly $8\cdot (n-1-c_1)\cdot ((n-1)-2)$
realizations of type \ref{item:comparisonproofcaseLhassize6:C4intersectingtwoedgesinseparateadjacentvertices}.
The next case to reexamine is \ref{numberofmatrixrealizationsoftype:twoisolatedvertices:case:lvertu1v1rvertbackslashp1Sequals1}.\ref{numberofmatrixrealizationsoftype:twoisolatedvertices:case:lvertu1v1rvertbackslashp1Sequals1:u1doesnotequalv1andbracesu1v1bracesintersectp1Sisasingleton}.\ref{numberofmatrixrealizationsoftype:twoisolatedvertices:case:lvertu1v1rvertbackslashp1Sequals1:u1doesnotequalv1andbracesu1v1bracesintersectp1Sisasingletoncontainingv1} which again does not give enough information to decide about
the types.
If \ref{numberofmatrixrealizationsoftype:twoisolatedvertices:case:lvertu1v1rvertbackslashp1Sequals1}.\ref{numberofmatrixrealizationsoftype:twoisolatedvertices:case:lvertu1v1rvertbackslashp1Sequals1:u1doesnotequalv1andbracesu1v1bracesintersectp1Sisasingleton}.\ref{numberofmatrixrealizationsoftype:twoisolatedvertices:case:lvertu1v1rvertbackslashp1Sequals1:u1doesnotequalv1andbracesu1v1bracesintersectp1Sisasingletoncontainingv1}.\ref{numberofmatrixrealizationsoftype:twoisolatedvertices:case:lvertu1v1rvertbackslashp1Sequals1:u1doesnotequalv1andbracesu1v1bracesintersectp1Sisasingletoncontainingu1:v1equaltoa1}, then the decision still cannot be made and once more depends on the nameless
subcases that were distinguished in that case, namely whether the first or the
second clause of \eqref{eq:proofofcomparativecountingtheoremcase3secondeq} is true:
If \ref{numberofmatrixrealizationsoftype:twoisolatedvertices:case:lvertu1v1rvertbackslashp1Sequals1}.\ref{numberofmatrixrealizationsoftype:twoisolatedvertices:case:lvertu1v1rvertbackslashp1Sequals1:u1doesnotequalv1andbracesu1v1bracesintersectp1Sisasingleton}.\ref{numberofmatrixrealizationsoftype:twoisolatedvertices:case:lvertu1v1rvertbackslashp1Sequals1:u1doesnotequalv1andbracesu1v1bracesintersectp1Sisasingletoncontainingv1}.\ref{numberofmatrixrealizationsoftype:twoisolatedvertices:case:lvertu1v1rvertbackslashp1Sequals1:u1doesnotequalv1andbracesu1v1bracesintersectp1Sisasingletoncontainingu1:v1equaltoa1}, and the first clause of \ref{eq:proofofcomparativecountingtheoremcase3secondeq} is
true, then we know that $u_1\neq v_1$, $u_1\notin\mathrm{p}_1(S)$, $v_1\in\mathrm{p}_1(S)$ and
$u_2 = v_2 \notin \mathrm{p}_2(S)$. The latter, more specificly $u_2 = v_2$, contradicts
both \ref{characterizationoftype:C4intersectingtwoedgesinseparatenonadjacentvertices:property3} and \ref{characterizationoftype:C4intersectingtwoedgesinseparateadjacentvertices:property3}. Combining $u_1\notin\mathrm{p}_1(S)$ and $u_2\notin \mathrm{p}_2(S)$ proves the
first clause of the conjunction \ref{characterizationoftype:C4intersectingatwopathinitsinnervertex:property4} to be false. The only type remaining is \ref{item:comparisonproofcaseLhassize6:C4intersectingatwopathinanendvertex}, and all properties
\ref{characterizationoftype:C4intersectingatwopathinanendvertex:property1}--\ref{characterizationoftype:C4intersectingatwopathinanendvertex:property4} are indeed satisfied (with only the first clause of the disjunction \ref{characterizationoftype:C4intersectingatwopathinanendvertex:property4} being true). Since in \ref{numberofmatrixrealizationsoftype:twoisolatedvertices:case:lvertu1v1rvertbackslashp1Sequals1}.\ref{numberofmatrixrealizationsoftype:twoisolatedvertices:case:lvertu1v1rvertbackslashp1Sequals1:u1doesnotequalv1andbracesu1v1bracesintersectp1Sisasingleton}.\ref{numberofmatrixrealizationsoftype:twoisolatedvertices:case:lvertu1v1rvertbackslashp1Sequals1:u1doesnotequalv1andbracesu1v1bracesintersectp1Sisasingletoncontainingv1}.\ref{numberofmatrixrealizationsoftype:twoisolatedvertices:case:lvertu1v1rvertbackslashp1Sequals1:u1doesnotequalv1andbracesu1v1bracesintersectp1Sisasingletoncontainingu1:v1equaltoa1} we found that there are exactly
$(a_1-1)\cdot ((n-1)-2)$ realizations of type \ref{item:comparisonproofcaseLhassize6:twoisolatedvertices} by $B\mid_{\{u,v\}}$, it follows that in the present situation there are
exactly $4\cdot (a_1-1)\cdot ((n-1)-2)$ realizations of type \ref{item:comparisonproofcaseLhassize6:C4intersectingatwopathinanendvertex}.
If \ref{numberofmatrixrealizationsoftype:twoisolatedvertices:case:lvertu1v1rvertbackslashp1Sequals1}.\ref{numberofmatrixrealizationsoftype:twoisolatedvertices:case:lvertu1v1rvertbackslashp1Sequals1:u1doesnotequalv1andbracesu1v1bracesintersectp1Sisasingleton}.\ref{numberofmatrixrealizationsoftype:twoisolatedvertices:case:lvertu1v1rvertbackslashp1Sequals1:u1doesnotequalv1andbracesu1v1bracesintersectp1Sisasingletoncontainingv1}.\ref{numberofmatrixrealizationsoftype:twoisolatedvertices:case:lvertu1v1rvertbackslashp1Sequals1:u1doesnotequalv1andbracesu1v1bracesintersectp1Sisasingletoncontainingu1:v1equaltoa1}, and the second clause of \eqref{eq:proofofcomparativecountingtheoremcase3secondeq}
is true, then we know that $u_1\neq v_1$, $u_1\notin\mathrm{p}_1(S)$, $v_1\in\mathrm{p}_1(S)$,
$u_2\in\mathrm{p}_2(S)$ and $v_2\notin\mathrm{p}_2(S)$. Since then $u_2\neq v_2$
and $u_1\neq v_1$, both \ref{characterizationoftype:C4intersectingatwopathinanendvertex:property3} and \ref{characterizationoftype:C4intersectingatwopathinitsinnervertex:property3} are impossible. Moreover, combining
$u_1\notin\mathrm{p}_1(S)$ and $v_2\notin\mathrm{p}_2(S)$ shows that both clauses of the
disjunction \ref{characterizationoftype:C4intersectingtwoedgesinseparatenonadjacentvertices:property4}. The remaining type is \ref{item:comparisonproofcaseLhassize6:C4intersectingtwoedgesinseparateadjacentvertices} and all the properties
\ref{characterizationoftype:C4intersectingtwoedgesinseparateadjacentvertices:property1}--\ref{characterizationoftype:C4intersectingtwoedgesinseparateadjacentvertices:property4} are indeed satisfied (as to the disjunction \ref{characterizationoftype:C4intersectingtwoedgesinseparateadjacentvertices:property4}, only its second clause is true).
Since in \ref{numberofmatrixrealizationsoftype:twoisolatedvertices:case:lvertu1v1rvertbackslashp1Sequals1}.\ref{numberofmatrixrealizationsoftype:twoisolatedvertices:case:lvertu1v1rvertbackslashp1Sequals1:u1doesnotequalv1andbracesu1v1bracesintersectp1Sisasingleton}.\ref{numberofmatrixrealizationsoftype:twoisolatedvertices:case:lvertu1v1rvertbackslashp1Sequals1:u1doesnotequalv1andbracesu1v1bracesintersectp1Sisasingletoncontainingv1}.\ref{numberofmatrixrealizationsoftype:twoisolatedvertices:case:lvertu1v1rvertbackslashp1Sequals1:u1doesnotequalv1andbracesu1v1bracesintersectp1Sisasingletoncontainingu1:v1equaltoa1} we found exactly $2\cdot (a_1-1)\cdot ((n-1)-2)$ realizations of type \ref{item:comparisonproofcaseLhassize6:twoisolatedvertices} by $B\mid_{\{u,v\}}$, it follows that
right now there are exactly $8\cdot (a_1-1)\cdot ((n-1)-2)$ realizations of
type \ref{item:comparisonproofcaseLhassize6:C4intersectingtwoedgesinseparateadjacentvertices}.
If \ref{numberofmatrixrealizationsoftype:twoisolatedvertices:case:lvertu1v1rvertbackslashp1Sequals1}.\ref{numberofmatrixrealizationsoftype:twoisolatedvertices:case:lvertu1v1rvertbackslashp1Sequals1:u1doesnotequalv1andbracesu1v1bracesintersectp1Sisasingleton}.\ref{numberofmatrixrealizationsoftype:twoisolatedvertices:case:lvertu1v1rvertbackslashp1Sequals1:u1doesnotequalv1andbracesu1v1bracesintersectp1Sisasingletoncontainingv1}.\ref{numberofmatrixrealizationsoftype:twoisolatedvertices:case:lvertu1v1rvertbackslashp1Sequals1:u1doesnotequalv1andbracesu1v1bracesintersectp1Sisasingletoncontainingu1:v1equaltoc1}, then one more time we have to distinguish the anonymous subcases.
If \ref{numberofmatrixrealizationsoftype:twoisolatedvertices:case:lvertu1v1rvertbackslashp1Sequals1}.\ref{numberofmatrixrealizationsoftype:twoisolatedvertices:case:lvertu1v1rvertbackslashp1Sequals1:u1doesnotequalv1andbracesu1v1bracesintersectp1Sisasingleton}.\ref{numberofmatrixrealizationsoftype:twoisolatedvertices:case:lvertu1v1rvertbackslashp1Sequals1:u1doesnotequalv1andbracesu1v1bracesintersectp1Sisasingletoncontainingv1}.\ref{numberofmatrixrealizationsoftype:twoisolatedvertices:case:lvertu1v1rvertbackslashp1Sequals1:u1doesnotequalv1andbracesu1v1bracesintersectp1Sisasingletoncontainingu1:v1equaltoc1}, and the first clause of \eqref{eq:proofofcomparativecountingtheoremcase3secondeq} is true,
then we know $u_1\neq v_1$, $u_1\notin\mathrm{p}_1(S)$, $v_1\in\mathrm{p}_1(S)$ and
$u_2 = v_2 \notin \mathrm{p}_2(S)$, with $u_2 = v_2$ ruling out both \ref{item:comparisonproofcaseLhassize6:C4intersectingtwoedgesinseparatenonadjacentvertices} and \ref{item:comparisonproofcaseLhassize6:C4intersectingtwoedgesinseparateadjacentvertices}.
Moreover, combining $u_1\notin\mathrm{p}_1(S)$ with $u_2\notin\mathrm{p}_2(S)$ proves the first
clause of the conjuction \ref{characterizationoftype:C4intersectingatwopathinitsinnervertex:property4} to be false. Again, for each decision the information in the
first clause of \eqref{eq:proofofcomparativecountingtheoremcase3secondeq} was used.
Now only \ref{item:comparisonproofcaseLhassize6:C4intersectingatwopathinanendvertex}
is left and indeed all properties \ref{characterizationoftype:C4intersectingatwopathinanendvertex:property1}--\ref{characterizationoftype:C4intersectingatwopathinanendvertex:property4} are true (with the disjunction \ref{characterizationoftype:C4intersectingatwopathinanendvertex:property4} satisfied only by way of its second clause). Since in \ref{numberofmatrixrealizationsoftype:twoisolatedvertices:case:lvertu1v1rvertbackslashp1Sequals1}.\ref{numberofmatrixrealizationsoftype:twoisolatedvertices:case:lvertu1v1rvertbackslashp1Sequals1:u1doesnotequalv1andbracesu1v1bracesintersectp1Sisasingleton}.\ref{numberofmatrixrealizationsoftype:twoisolatedvertices:case:lvertu1v1rvertbackslashp1Sequals1:u1doesnotequalv1andbracesu1v1bracesintersectp1Sisasingletoncontainingv1}.\ref{numberofmatrixrealizationsoftype:twoisolatedvertices:case:lvertu1v1rvertbackslashp1Sequals1:u1doesnotequalv1andbracesu1v1bracesintersectp1Sisasingletoncontainingu1:v1equaltoc1} we found that in the first subcase there are exactly $(c_1-2)\cdot ((n-1)-2)$ realizations of
type \ref{item:comparisonproofcaseLhassize6:twoisolatedvertices} by $B\mid_{\{u,v\}}$,
it follows for our present situation that there are
exactly $4\cdot (c_1-2)\cdot ((n-1)-2)$ realizations of \ref{item:comparisonproofcaseLhassize6:C4intersectingatwopathinanendvertex} by $B\mid_{\{u,v\}}$.
If \ref{numberofmatrixrealizationsoftype:twoisolatedvertices:case:lvertu1v1rvertbackslashp1Sequals1}.\ref{numberofmatrixrealizationsoftype:twoisolatedvertices:case:lvertu1v1rvertbackslashp1Sequals1:u1doesnotequalv1andbracesu1v1bracesintersectp1Sisasingleton}.\ref{numberofmatrixrealizationsoftype:twoisolatedvertices:case:lvertu1v1rvertbackslashp1Sequals1:u1doesnotequalv1andbracesu1v1bracesintersectp1Sisasingletoncontainingv1}.\ref{numberofmatrixrealizationsoftype:twoisolatedvertices:case:lvertu1v1rvertbackslashp1Sequals1:u1doesnotequalv1andbracesu1v1bracesintersectp1Sisasingletoncontainingu1:v1equaltoc1}, and the second clause of \eqref{eq:proofofcomparativecountingtheoremcase3secondeq}
is true, then we know that $u_1\neq v_1$, $u_1\notin\mathrm{p}_1(S)$, $v_1\in\mathrm{p}_1(S)$
and $u_2\in\mathrm{p}_2(S)$ and $v_2\notin\mathrm{p}_2(S)$. Since then $u_2\neq v_2$ it follows
$\{u_1,u_2\}\cap \{v_1,v_2\} = \emptyset$, contradicting both
\ref{characterizationoftype:C4intersectingatwopathinanendvertex:property3} and
\ref{characterizationoftype:C4intersectingatwopathinitsinnervertex:property3}.
Combining $u_1\notin\mathrm{p}_1(S)$ and $v_2\notin\mathrm{p}_2(S)$ we see that both clauses
of the disjunction \ref{characterizationoftype:C4intersectingtwoedgesinseparatenonadjacentvertices:property4} are false. We are left with type \ref{item:comparisonproofcaseLhassize6:C4intersectingtwoedgesinseparateadjacentvertices} and indeed, all properties
\ref{characterizationoftype:C4intersectingtwoedgesinseparateadjacentvertices:property1}--\ref{characterizationoftype:C4intersectingtwoedgesinseparateadjacentvertices:property4} are satisfied (for the disjunction \ref{characterizationoftype:C4intersectingtwoedgesinseparateadjacentvertices:property4} it is only the second clause, which is).
Since in \ref{numberofmatrixrealizationsoftype:twoisolatedvertices:case:lvertu1v1rvertbackslashp1Sequals1}.\ref{numberofmatrixrealizationsoftype:twoisolatedvertices:case:lvertu1v1rvertbackslashp1Sequals1:u1doesnotequalv1andbracesu1v1bracesintersectp1Sisasingleton}.\ref{numberofmatrixrealizationsoftype:twoisolatedvertices:case:lvertu1v1rvertbackslashp1Sequals1:u1doesnotequalv1andbracesu1v1bracesintersectp1Sisasingletoncontainingv1}.\ref{numberofmatrixrealizationsoftype:twoisolatedvertices:case:lvertu1v1rvertbackslashp1Sequals1:u1doesnotequalv1andbracesu1v1bracesintersectp1Sisasingletoncontainingu1:v1equaltoc1} we found that there exist exactly $2\cdot (c_1-2)\cdot ((n-1)-2)$ realizations
of type \ref{item:comparisonproofcaseLhassize6:twoisolatedvertices} by $B\mid_{\{u,v\}}$,
it follows that in the present situation there are exactly
$8\cdot (c_1-2)\cdot ((n-1)-2)$ realizations of type \ref{item:comparisonproofcaseLhassize6:C4intersectingtwoedgesinseparateadjacentvertices} by $B\mid_{\{u,v\}}$.
The next case to reexamine is \ref{numberofmatrixrealizationsoftype:twoisolatedvertices:case:lvertu1v1rvertbackslashp1Sequals2}. This case is symmetric to
\ref{numberofmatrixrealizationsoftype:twoisolatedvertices:case:lvertu1v1rvertbackslashp1Sequals0} under swapping the subscripts $1$ and $2$. Therefore, we can analyse its
subcases by reexaming the analysis of \ref{numberofmatrixrealizationsoftype:twoisolatedvertices:case:lvertu1v1rvertbackslashp1Sequals0} with this swap in mind.
First of all, if \ref{numberofmatrixrealizationsoftype:twoisolatedvertices:case:lvertu1v1rvertbackslashp1Sequals2}, then $\{u_2,v_2\}\subseteq \mathrm{p}_2(S)$, and this
renders both clauses of the disjunction \ref{characterizationoftype:C4intersectingatwopathinanendvertex:property4} false.
Reading \ref{numberofmatrixrealizationsoftype:twoisolatedvertices:case:lvertu1v1rvertbackslashp1Sequals0}.\ref{numberofmatrixrealizationsoftype:twoisolatedvertices:case:lvertu1v1rvertbackslashp1Sequals0:case:u1equalsv1} this way implies that we know
$u_2=v_2 \in \mathrm{p}_2(S)$, and $u_2=v_2$ contradicts both \ref{characterizationoftype:C4intersectingtwoedgesinseparatenonadjacentvertices:property3} and \ref{characterizationoftype:C4intersectingtwoedgesinseparateadjacentvertices:property3}. The only type
remaining is \ref{item:comparisonproofcaseLhassize6:C4intersectingatwopathinitsinnervertex} and indeed, all properties \ref{characterizationoftype:C4intersectingatwopathinitsinnervertex:property1}--\ref{characterizationoftype:C4intersectingatwopathinitsinnervertex:property4} are satisfied. Since in \ref{numberofmatrixrealizationsoftype:twoisolatedvertices:case:lvertu1v1rvertbackslashp1Sequals0}.\ref{numberofmatrixrealizationsoftype:twoisolatedvertices:case:lvertu1v1rvertbackslashp1Sequals0:case:u1equalsv1} we found
that there are exactly $2\cdot \binom{(n-1)-2}{2}$ realizations of type \ref{item:comparisonproofcaseLhassize6:twoisolatedvertices} by $B\mid_{\{u,v\}}$,
it follows that here there exist exactly $8\cdot \binom{(n-1)-2}{2}$ realizations
of type \ref{item:comparisonproofcaseLhassize6:C4intersectingatwopathinitsinnervertex} by $B\mid_{\{u,v\}}$.
Reading \ref{numberofmatrixrealizationsoftype:twoisolatedvertices:case:lvertu1v1rvertbackslashp1Sequals0}.\ref{numberofmatrixrealizationsoftype:twoisolatedvertices:case:lvertu1v1rvertbackslashp1Sequals0:case:u1doesnotequalv1} this way implies that we know
$u_2\neq v_2$, $\{u_2,v_2\}\cap\mathrm{p}_2(S) = \emptyset$, $u_1\neq v_1$
and $\{u_1,v_1\}\subseteq \mathrm{p}_1(S)$. The properties $u_1\neq v_1$ and $u_2\neq v_2$
taken together contradict both \ref{characterizationoftype:C4intersectingatwopathinanendvertex:property3} and \ref{characterizationoftype:C4intersectingatwopathinitsinnervertex:property3}. The property $\{ u_2 , v_2\}\cap \mathrm{p}_2(S) = \emptyset$ alone renders
both clauses of the disjunction \ref{characterizationoftype:C4intersectingtwoedgesinseparateadjacentvertices:property4} false. What we are left with is type \ref{item:comparisonproofcaseLhassize6:C4intersectingtwoedgesinseparatenonadjacentvertices} and
indeed all properties \ref{characterizationoftype:C4intersectingtwoedgesinseparatenonadjacentvertices:property1}--\ref{characterizationoftype:C4intersectingtwoedgesinseparatenonadjacentvertices:property4} are satisfied. Since in \ref{numberofmatrixrealizationsoftype:twoisolatedvertices:case:lvertu1v1rvertbackslashp1Sequals0}.\ref{numberofmatrixrealizationsoftype:twoisolatedvertices:case:lvertu1v1rvertbackslashp1Sequals0:case:u1doesnotequalv1} we found that there are exactly
$2\cdot \binom{(n-1)-2}{2}$ realizations of type
\ref{item:comparisonproofcaseLhassize6:twoisolatedvertices} by $B\mid_{\{u,v\}}$, it follows
by symmetry that here, too, there exist exactly $8\cdot \binom{(n-1)-2}{2}$
realizations of type \ref{item:comparisonproofcaseLhassize6:C4intersectingtwoedgesinseparatenonadjacentvertices} by $B\mid_{\{u,v\}}$.
We have now completed reexamining the analysis of \ref{numberofmatrixrealizationsoftype:twoisolatedvertices} and we may now add up (separately for each of the four types
\ref{item:comparisonproofcaseLhassize6:C4intersectingtwoedgesinseparatenonadjacentvertices}--\ref{item:comparisonproofcaseLhassize6:C4intersectingatwopathinitsinnervertex}
the number of realizations we found during the reexamination.
For \ref{item:comparisonproofcaseLhassize6:C4intersectingtwoedgesinseparatenonadjacentvertices} we found exactly $8\cdot\binom{(n-1)-2}{2} + 8\cdot\binom{(n-1)-2}{2}
= 16 \cdot \binom{(n-1)-2}{2}$ realizations by $B\mid_{\{u,v\}}$. Now \ref{numberofmatrixrealizationsoftype:C4intersectingtwoedgesinseparatenonadjacentvertices} is proved.
For \ref{item:comparisonproofcaseLhassize6:C4intersectingtwoedgesinseparateadjacentvertices} we found exactly
$
8\cdot (n-1-a_1-1)\cdot ((n-1)-2)
+ 8 \cdot (n-1-c_1)\cdot ((n-1)-2)
+ 8\cdot (a_1-1)\cdot ((n-1)-2)
+ 8\cdot (c_1-2)\cdot ((n-1)-2)
=
16\cdot ((n-1)-2)^2
$ realizations by $B\mid_{\{u,v\}}$. Now \ref{numberofmatrixrealizationsoftype:C4intersectingtwoedgesinseparateadjacentvertices} is proved.
For \ref{item:comparisonproofcaseLhassize6:C4intersectingatwopathinanendvertex}
we found exactly $
8 \cdot ((n-1)-2)^2
+ 4 \cdot (n-1-a_1-1)\cdot ((n-1)-2)
+ 4 \cdot (n-1-c_1)\cdot ((n-1)-2)
+ 4 \cdot (a_1-1)\cdot ((n-1)-2)
+ 4 \cdot (c_1-2)\cdot ((n-1)-2)
=
8\cdot ((n-1)-2)^2
+ 8\cdot ((n-1)-2)^2
= 16 \cdot ((n-1)-2)^2
$ realizations by $B\mid_{\{u,v\}}$. Now
\ref{numberofmatrixrealizationsoftype:C4intersectingatwopathinanendvertex} is proved.
For \ref{item:comparisonproofcaseLhassize6:C4intersectingatwopathinitsinnervertex} we
found exactly
$8\cdot\binom{(n-1)-2}{2} + 8\cdot\binom{(n-1)-2}{2} = 16 \cdot \binom{(n-1)-2}{2}$
realizations by $B\mid_{\{u,v\}}$. Now \ref{numberofmatrixrealizationsoftype:C4intersectingatwopathinitsinnervertex} is proved.
As to \ref{numberofmatrixrealizationsoftype:threeisolatedvertices}, let us first
note that Definition \ref{def:XBandecXB} implies:
\begin{lemma}
For every $B\in\{0,\pm\}^I$ with $I\in\binom{[n-1]^2}{6}$, $I = S\sqcup \{u,v\}$ and
$\mathrm{X}_{B\mid_{S}} \cong C^4$ we have $\mathrm{X}_{B} \cong \ref{item:comparisonproofcaseLhassize6:threeisolatedvertices}$ if and only if
{\scriptsize
\begin{minipage}[b]{0.3\linewidth}
\begin{enumerate}[label={\rm(P.\ref{item:comparisonproofcaseLhassize6:threeisolatedvertices}.\arabic{*})}]
\item\label{characterizationoftype:threeisolatedvertices:property1}
$B[u]=B[v]=0$,
\end{enumerate}
\end{minipage}
\begin{minipage}[b]{0.6\linewidth}
\begin{enumerate}[label={\rm(P.\ref{item:comparisonproofcaseLhassize6:threeisolatedvertices}.\arabic{*})},start=2]
\item\label{characterizationoftype:threeisolatedvertices:property2}
$\lvert \{ u_1, v_1\} \setminus \mathrm{p}_1(S) \rvert +
\lvert \{ u_2, v_2\} \setminus \mathrm{p}_2(S) \rvert = 3$\quad .
\end{enumerate}
\end{minipage}
}
\end{lemma}
We now distinguish cases according to how \ref{characterizationoftype:threeisolatedvertices:property2} is satisfied.
\begin{enumerate}[label={\rm(C.\ref{numberofmatrixrealizationsoftype:threeisolatedvertices}.\arabic{*})}]
\item\label{numberofmatrixrealizationsoftype:threeisolatedvertices:case:lvertu1v1rvertbackslashp1Sequals0} $\lvert \{ u_1, v_1\} \setminus \mathrm{p}_1(S) \rvert = 0$. Then
$\lvert \{ u_2, v_2\} \setminus \mathrm{p}_2(S)\rvert = 3$ by
\ref{characterizationoftype:threeisolatedvertices:property2}, which is impossible.
Hence Case 1 does not occur.
\item\label{numberofmatrixrealizationsoftype:threeisolatedvertices:case:lvertu1v1rvertbackslashp1Sequals1} $\lvert \{ u_1, v_1\} \setminus \mathrm{p}_1(S) \rvert = 1$. Then
$\lvert \{ u_2, v_2\} \setminus \mathrm{p}_2(S)\rvert = 2$ by
\ref{characterizationoftype:threeisolatedvertices:property2},
which is equivalent to
\begin{equation}\label{eq:proofofcomparativecountingtheoremcase2firsteq}
\text{$u_2\neq v_2$ and $\{ u_2, v_2 \} \cap \mathrm{p}_2(S) = \emptyset$} \quad .
\end{equation}
Since the condition defining \ref{numberofmatrixrealizationsoftype:threeisolatedvertices:case:lvertu1v1rvertbackslashp1Sequals1} is equivalent to
\begin{equation}\label{eq:proofofcomparativecountingtheoremcase2secondeq}
\text{
($u_1 = v_1$ and $\{u_1,v_1\} \cap \mathrm{p}_1(S) = \emptyset$) or
($u_1 \neq v_1$ and $\lvert \{u_1,v_1\} \cap \mathrm{p}_1(S) \rvert = 1$)
}\quad ,
\end{equation}
there are two further cases.
\begin{enumerate}[label={\rm(\arabic{*})}]
\item\label{numberofmatrixrealizationsoftype:threeisolatedvertices:case:lvertu1v1rvertbackslashp1Sequals1:u1equalsv1andbracesu1v1bracesintersectp1Sisempty}
$u_1 = v_1$ and $\{u_1,v_1\} \cap \mathrm{p}_1(S) = \emptyset$. Then there are exactly
$(n-1)-2$ such $u_1=v_1$. Combining \eqref{eq:proofofcomparativecountingtheoremcase2firsteq} with $u\prec v$ it follows that $u_2<v_2$, therefore in the present case each
of the $\binom{(n-1)-2}{2}$ different sets $\{u_2,v_2\}$ satisfying
\eqref{eq:proofofcomparativecountingtheoremcase2firsteq} determines the two
pairs $u$ and $v$. Therefore there are exactly $((n-1)-2) \cdot \binom{(n-1)-2}{2}$
realizations of \ref{numberofmatrixrealizationsoftype:threeisolatedvertices:case:lvertu1v1rvertbackslashp1Sequals1:u1equalsv1andbracesu1v1bracesintersectp1Sisempty}.
\item\label{numberofmatrixrealizationsoftype:threeisolatedvertices:case:lvertu1v1rvertbackslashp1Sequals1:u1doesnotequalv1andbracesu1v1bracesintersectp1Sisasingleton}
$u_1 \neq v_1$ and $\lvert \{u_1,v_1\} \cap \mathrm{p}_1(S) \rvert = 1$. From
$u_1 \neq v_1$ and the assumption $u\prec v$ it follows that $u_1 < v_1$. By
\eqref{eq:proofofcomparativecountingtheoremcase2firsteq}, $u_2<v_2$ or $v_2<u_2$,
but nothing more is known about $u_2$ and $v_2$. Therefore, both possibilities must
be taken into account. Because of $\mathrm{p}_1(S) = \{a_1,b_1,c_1,d_1\} = \{a_1,c_1\}$
and $u_1<v_1$ there are exactly four possibilities for
$\lvert \{u_1,v_1\} \cap \mathrm{p}_1(S) \rvert = 1$ to be true:
\begin{enumerate}[label={\rm(\arabic{*})}]
\item\label{numberofmatrixrealizationsoftype:threeisolatedvertices:case:lvertu1v1rvertbackslashp1Sequals1:u1doesnotequalv1andbracesu1v1bracesintersectp1Sisasingletonu1equalsa1} $u_1 = a_1 = b_1$. Then because of $u_1 < v_1$ and $v_1 \neq c_1 = d_1$ it
follows that there are exactly $n-1 - a_1 - 1$ different $v_1$ with
$v_1\notin \mathrm{p}_1(S)$ in this case. For each of them, there exist exactly
$\binom{(n-1)-2}{2}$ different \emph{sets} $\{u_2,v_2\}$ satisfying
\eqref{eq:proofofcomparativecountingtheoremcase2firsteq}. Now the two pairs $u$ and
$v$ are not determined by them: each of the sets can be realized in exactly two
ways, both by $u_2<v_2$ and by $v_2<u_2$. Therefore, there are exactly
$(n-1-a_1-1)\cdot 2\cdot \binom{(n-1)-2}{2}$ realizations of
\ref{numberofmatrixrealizationsoftype:threeisolatedvertices:case:lvertu1v1rvertbackslashp1Sequals1:u1doesnotequalv1andbracesu1v1bracesintersectp1Sisasingletonu1equalsa1}
by $u$ and $v$.
\item\label{numberofmatrixrealizationsoftype:threeisolatedvertices:case:lvertu1v1rvertbackslashp1Sequals1:u1doesnotequalv1andbracesu1v1bracesintersectp1Sisasingletonu1equalsc1} $u_1 = c_1 = d_1$. Then because of $u_1 < v_1$ it follows that there are
exactly $n-1 - c_1$ different $v_1$ with $v_1\notin \mathrm{p}_1(S)$. As in the preceding
case, for each of these $v_1$ there exist exactly $\binom{(n-1)-2}{2}$ different
sets $\{u_2,v_2\}$ satisfying
\eqref{eq:proofofcomparativecountingtheoremcase2firsteq}, hence exactly
$2\cdot\binom{(n-1)-2}{2}$ different $u$ and $v$. Therefore there exist exactly
$(n-1 - c_1) \cdot 2 \cdot \binom{(n-1)-2}{2}$ different realizations of type
\ref{item:comparisonproofcaseLhassize6:threeisolatedvertices} by $B\mid_{\{u,v\}}$.
\item\label{numberofmatrixrealizationsoftype:threeisolatedvertices:case:lvertu1v1rvertbackslashp1Sequals1:u1doesnotequalv1andbracesu1v1bracesintersectp1Sisasingletonv1equalsa1} $v_1 = a_1 = b_1$. Then because of $u_1 < v_1$ it follows that there are
exactly $a_1 - 1$ different $u_1$ with $u_1\notin \mathrm{p}_1(S)$ in this case. For the
same reasons as in the preceding two cases we know that here there exist exactly
$(a_1 - 1) \cdot 2 \cdot \binom{(n-1)-2}{2}$ different realizations of type
\ref{item:comparisonproofcaseLhassize6:threeisolatedvertices} by $B\mid_{\{u,v\}}$.
\item\label{numberofmatrixrealizationsoftype:threeisolatedvertices:case:lvertu1v1rvertbackslashp1Sequals1:u1doesnotequalv1andbracesu1v1bracesintersectp1Sisasingletonv1equalsc1} $v_1 = c_1 = d_1$. Then because of $u_1 < v_1$ and $u_1 \neq a_1 = b_1$ it
follows that there are $c_1 - 1 - 1$ different $u_1$ with $u_1\notin \mathrm{p}_1(S)$ in
this case. For the same reasons as in the preceding three cases we know that here
there exist exactly $(c_1 - 1 - 1)\cdot 2 \cdot \binom{(n-1)-2}{2}$ different
realizations of type \ref{item:comparisonproofcaseLhassize6:threeisolatedvertices}
by $B\mid_{\{u,v\}}$.
\end{enumerate}
It follows that if \ref{numberofmatrixrealizationsoftype:threeisolatedvertices:case:lvertu1v1rvertbackslashp1Sequals1}.\ref{numberofmatrixrealizationsoftype:threeisolatedvertices:case:lvertu1v1rvertbackslashp1Sequals1:u1doesnotequalv1andbracesu1v1bracesintersectp1Sisasingleton}, then there exist exactly
$
\bigl ( (n-1-a_1-1) + (n-1-c_1) + (a_1-1) + (c_1-1-1)\bigr) \cdot
2\cdot \binom{(n-1)-2}{2} = 4\cdot ((n-1)-2) \cdot \binom{(n-1)-2}{2}
$
different realizations of type \ref{item:comparisonproofcaseLhassize6:threeisolatedvertices} by $B\mid_{\{u,v\}}$.
\end{enumerate}
It follows that if \ref{numberofmatrixrealizationsoftype:threeisolatedvertices:case:lvertu1v1rvertbackslashp1Sequals1}, then there are exactly
$( (n-1)-2 + 4\cdot((n-1)-2))\cdot \binom{(n-1)-2}{2} =
5\cdot ((n-1)-2) \cdot \binom{(n-1)-2}{2}$ realizations of type \ref{item:comparisonproofcaseLhassize6:threeisolatedvertices} by $B\mid_{\{u,v\}}$.
\item\label{numberofmatrixrealizationsoftype:threeisolatedvertices:case:lvertu1v1rvertbackslashp1Sequals2} $\lvert \{ u_1, v_1\} \setminus \mathrm{p}_1(S) \rvert = 2$. This is
equivalent to
\begin{equation}\label{eq:proofofcomparativecountingtheoremcase3firsteq}
\text{$u_1\neq v_1$ and $\{ u_1, v_1 \} \cap \mathrm{p}_1(X) = \emptyset$}\quad .
\end{equation}
Equation \ref{characterizationoftype:threeisolatedvertices:property2} implies $\lvert \{ u_2, v_2\} \setminus \mathrm{p}_2(S)\rvert = 1$,
which is equivalent to
\begin{equation}\label{eq:proofofcomparativecountingtheoremcase3secondeq}
\text{
($u_2 = v_2$ and $\{u_2,v_2\} \cap \mathrm{p}_2(S) = \emptyset$) or
($u_2 \neq v_2$ and $\lvert \{u_2,v_2\} \cap \mathrm{p}_2(S) \rvert = 1$)
}\quad .
\end{equation}
By swapping the subscripts $1$ and $2$ in the argument given for Case 2 it now
follows that if \ref{numberofmatrixrealizationsoftype:threeisolatedvertices:case:lvertu1v1rvertbackslashp1Sequals2}, then there are
exactly $5\cdot ((n-1)-2) \cdot \binom{(n-1)-2}{2}$ different realizations of
type \ref{item:comparisonproofcaseLhassize6:threeisolatedvertices} by $B\mid_{\{u,v\}}$.
\item\label{numberofmatrixrealizationsoftype:threeisolatedvertices:case:lvertu1v1rvertbackslashp1Sequals3} $\lvert \{ u_1, v_1\} \setminus \mathrm{p}_1(S) \rvert = 3$. This is
impossible, hence \ref{numberofmatrixrealizationsoftype:threeisolatedvertices:case:lvertu1v1rvertbackslashp1Sequals3} does not occur.
\end{enumerate}
It follows that for every fixed $S$ there are exactly
$10\cdot ((n-1)-2) \cdot \binom{(n-1)-2}{2}$ possibilities to position the two zeros
indexed by $I\setminus S$ such that $\mathrm{X}_{B} \cong \ref{item:comparisonproofcaseLhassize6:threeisolatedvertices}$. This completes the proof of \ref{numberofmatrixrealizationsoftype:threeisolatedvertices}.
As to \ref{numberofmatrixrealizationsoftype:oneadditionaledgeintersectingC4andtwoisolatedvertices}--\ref{numberofmatrixrealizationsoftype:twoadditionalnondisjointedgesdisjointfromC4} we begin by noting that for each of the four isomorphism types
\ref{item:comparisonproofcaseLhassize6:oneadditionaledgeintersectingC4andtwoisolatedvertices}--\ref{item:comparisonproofcaseLhassize6:twoadditionalnondisjointedgesdisjointfromC4}, a necessary condition is that $\lvert \mathrm{V}(\mathrm{X}_{B})\setminus \mathrm{V}(\mathrm{X}_{B\mid_{S}}) \rvert = 3$.
We can therefore prove \ref{numberofmatrixrealizationsoftype:oneadditionaledgeintersectingC4andtwoisolatedvertices}--\ref{numberofmatrixrealizationsoftype:twoadditionalnondisjointedgesdisjointfromC4} during one reexamination of the proof of \ref{numberofmatrixrealizationsoftype:threeisolatedvertices}.
Since \ref{numberofmatrixrealizationsoftype:threeisolatedvertices:case:lvertu1v1rvertbackslashp1Sequals0} and \ref{numberofmatrixrealizationsoftype:threeisolatedvertices:case:lvertu1v1rvertbackslashp1Sequals3} are impossible, we only have to
consider \ref{numberofmatrixrealizationsoftype:threeisolatedvertices:case:lvertu1v1rvertbackslashp1Sequals1} and \ref{numberofmatrixrealizationsoftype:threeisolatedvertices:case:lvertu1v1rvertbackslashp1Sequals2}.
If \ref{numberofmatrixrealizationsoftype:threeisolatedvertices:case:lvertu1v1rvertbackslashp1Sequals1}, we know \eqref{eq:proofofcomparativecountingtheoremcase2firsteq} but
this is not sufficient to rule out any of the types \ref{item:comparisonproofcaseLhassize6:oneadditionaledgeintersectingC4andtwoisolatedvertices}--\ref{item:comparisonproofcaseLhassize6:twoadditionalnondisjointedgesdisjointfromC4}.
If \ref{numberofmatrixrealizationsoftype:threeisolatedvertices:case:lvertu1v1rvertbackslashp1Sequals1}.\ref{numberofmatrixrealizationsoftype:threeisolatedvertices:case:lvertu1v1rvertbackslashp1Sequals1:u1equalsv1andbracesu1v1bracesintersectp1Sisempty} we
know that
\begin{equation}\label{eq:knowledgeincase:numberofmatrixrealizationsoftype:threeisolatedvertices:case:lvertu1v1rvertbackslashp1Sequals1:u1equalsv1andbracesu1v1bracesintersectp1Sisempty}
u_1 = v_1,\; \{u_1,v_1\}\cap\mathrm{p}_1(S) = \emptyset,\;
u_2\neq v_2,\; \{u_2,v_2\}\cap\mathrm{p}_2(S) = \emptyset\quad ,
\end{equation}
and will now consider the consequences of this for \ref{numberofmatrixrealizationsoftype:oneadditionaledgeintersectingC4andtwoisolatedvertices}--\ref{numberofmatrixrealizationsoftype:twoadditionalnondisjointedgesdisjointfromC4}.
\begin{enumerate}[label={\rm(\arabic{*})}]
\item Concerning contributions to \ref{numberofmatrixrealizationsoftype:oneadditionaledgeintersectingC4andtwoisolatedvertices}, note that properties $\{u_1,v_1\}\cap\mathrm{p}_1(S) = \emptyset$ and $\{u_2,v_2\}\cap\mathrm{p}_2(S) = \emptyset$ make an edge intersecting $\mathrm{X}_{B\mid_{S}} \cong C^4$ impossible, hence the case \ref{numberofmatrixrealizationsoftype:threeisolatedvertices:case:lvertu1v1rvertbackslashp1Sequals1}.\ref{numberofmatrixrealizationsoftype:threeisolatedvertices:case:lvertu1v1rvertbackslashp1Sequals1:u1equalsv1andbracesu1v1bracesintersectp1Sisempty} does not contribute\footnote{The fact that neither \ref{numberofmatrixrealizationsoftype:threeisolatedvertices:case:lvertu1v1rvertbackslashp1Sequals1}.\ref{numberofmatrixrealizationsoftype:threeisolatedvertices:case:lvertu1v1rvertbackslashp1Sequals1:u1equalsv1andbracesu1v1bracesintersectp1Sisempty} nor the corresponding subcase of \ref{numberofmatrixrealizationsoftype:threeisolatedvertices:case:lvertu1v1rvertbackslashp1Sequals2} (which due to symmetry was not spelled out in the proof of \ref{numberofmatrixrealizationsoftype:threeisolatedvertices} and therefore does not have a name)
contribute to $\lvert(\leftsubscript{\mathrm{ul}}{\upX}^{6,n,n})^{-1}\ref{item:comparisonproofcaseLhassize6:oneadditionaledgeintersectingC4andtwoisolatedvertices}\rvert$ is a reason why $\lvert(\leftsubscript{\mathrm{ul}}{\upX}^{6,n,n})^{-1}\ref{item:comparisonproofcaseLhassize6:oneadditionaledgeintersectingC4andtwoisolatedvertices}\rvert$ is larger but not twice as large as
$\lvert(\leftsubscript{\mathrm{ul}}{\upX}^{6,n,n})^{-1}\ref{item:comparisonproofcaseLhassize6:threeisolatedvertices}\rvert$ even though in the cases where \ref{item:comparisonproofcaseLhassize6:oneadditionaledgeintersectingC4andtwoisolatedvertices} \emph{can} be realized the number of realizations is twice as large as for \ref{item:comparisonproofcaseLhassize6:threeisolatedvertices}.} to $\lvert(\leftsubscript{\mathrm{ul}}{\upX}^{6,n,n})^{-1}\ref{item:comparisonproofcaseLhassize6:oneadditionaledgeintersectingC4andtwoisolatedvertices}\rvert$.
\item Concerning contributions to \ref{numberofmatrixrealizationsoftype:oneadditionaldisjointedgeandoneisolatedvertex}, note that \eqref{eq:knowledgeincase:numberofmatrixrealizationsoftype:threeisolatedvertices:case:lvertu1v1rvertbackslashp1Sequals1:u1equalsv1andbracesu1v1bracesintersectp1Sisempty} implies that $\mathrm{X}_{B} \cong \ref{item:comparisonproofcaseLhassize6:oneadditionaldisjointedgeandoneisolatedvertex}$ if and only if
either ($B[u]\in \{\pm\}$ and $B[v] = 0$) or ($B[u] = 0$ and $B[v] \in \{\pm\}$).
Each of these clauses corresponds to $2$ different $B$. It follows that if
\ref{numberofmatrixrealizationsoftype:threeisolatedvertices:case:lvertu1v1rvertbackslashp1Sequals1}.\ref{numberofmatrixrealizationsoftype:threeisolatedvertices:case:lvertu1v1rvertbackslashp1Sequals1:u1equalsv1andbracesu1v1bracesintersectp1Sisempty}, then
there are $4$-times as many realizations of type
\ref{item:comparisonproofcaseLhassize6:oneadditionaldisjointedgeandoneisolatedvertex} as there are of type
\ref{item:comparisonproofcaseLhassize6:threeisolatedvertices}. Therefore,
if \ref{numberofmatrixrealizationsoftype:threeisolatedvertices:case:lvertu1v1rvertbackslashp1Sequals1}.\ref{numberofmatrixrealizationsoftype:threeisolatedvertices:case:lvertu1v1rvertbackslashp1Sequals1:u1equalsv1andbracesu1v1bracesintersectp1Sisempty},
there are exactly $4\cdot ((n-1)-2)\cdot\binom{(n-1)-2}{2}$ realizations of type
\ref{item:comparisonproofcaseLhassize6:oneadditionaldisjointedgeandoneisolatedvertex}
by $B\mid_{\{u,v\}}$.
\item Concerning contributions to \ref{numberofmatrixrealizationsoftype:twoadditionaledgesonlyoneofthemdisjoint}, since properties $\{u_1,v_1\}\cap\mathrm{p}_1(S) = \emptyset$ and $\{u_2,v_2\}\cap\mathrm{p}_2(S) = \emptyset$ make an edge intersecting $\mathrm{X}_{B\mid_{S}} \cong C^4$ impossible, the case \ref{numberofmatrixrealizationsoftype:threeisolatedvertices:case:lvertu1v1rvertbackslashp1Sequals1}.\ref{numberofmatrixrealizationsoftype:threeisolatedvertices:case:lvertu1v1rvertbackslashp1Sequals1:u1equalsv1andbracesu1v1bracesintersectp1Sisempty} does not contribute to \ref{numberofmatrixrealizationsoftype:twoadditionaledgesonlyoneofthemdisjoint}.
\item Concerning contributions to \ref{numberofmatrixrealizationsoftype:twoadditionalnondisjointedgesdisjointfromC4}, we see from \eqref{eq:knowledgeincase:numberofmatrixrealizationsoftype:threeisolatedvertices:case:lvertu1v1rvertbackslashp1Sequals1:u1equalsv1andbracesu1v1bracesintersectp1Sisempty} that $\mathrm{X}_{B} \cong \ref{item:comparisonproofcaseLhassize6:twoadditionalnondisjointedgesdisjointfromC4}$ if and only if
($B[u]\in \{\pm\}$ and $B[v]\in\{\pm\}$), and there are $4$ different
$B\mid_{\{u,v\}}\in\{0,\pm\}^{\{u,v\}}$ satisfying this. Therefore, if
\ref{numberofmatrixrealizationsoftype:threeisolatedvertices:case:lvertu1v1rvertbackslashp1Sequals1}.\ref{numberofmatrixrealizationsoftype:threeisolatedvertices:case:lvertu1v1rvertbackslashp1Sequals1:u1equalsv1andbracesu1v1bracesintersectp1Sisempty},
there are exactly $4\cdot ((n-1)-2)\cdot\binom{(n-1)-2}{2}$ realizations of type
\ref{item:comparisonproofcaseLhassize6:twoadditionalnondisjointedgesdisjointfromC4}
by $B\mid_{\{u,v\}}$.
\end{enumerate}
If \ref{numberofmatrixrealizationsoftype:threeisolatedvertices:case:lvertu1v1rvertbackslashp1Sequals1}.\ref{numberofmatrixrealizationsoftype:threeisolatedvertices:case:lvertu1v1rvertbackslashp1Sequals1:u1doesnotequalv1andbracesu1v1bracesintersectp1Sisasingleton}, then we know
\begin{equation}\label{eq:knowledgeincase:numberofmatrixrealizationsoftype:threeisolatedvertices:case:lvertu1v1rvertbackslashp1Sequals1:u1doesnotequalv1andbracesu1v1bracesintersectp1Sisasingleton}
u_1\neq v_1,\; \lvert \{u_1,v_1\}\cap\mathrm{p}_1(S) \rvert = 1,\;
u_2\neq v_2,\; \{ u_2,v_2\}\cap\mathrm{p}_2(S) = \emptyset\quad .
\end{equation}
and will now consider the consequences of this for \ref{numberofmatrixrealizationsoftype:oneadditionaledgeintersectingC4andtwoisolatedvertices}--\ref{numberofmatrixrealizationsoftype:twoadditionalnondisjointedgesdisjointfromC4}.
\begin{enumerate}[label={\rm(\arabic{*})}]
\item Concerning contributions to \ref{numberofmatrixrealizationsoftype:oneadditionaledgeintersectingC4andtwoisolatedvertices}, we can argue as follows:
If $\lvert \{u_1,v_1\}\cap\mathrm{p}_1(S) \rvert = 1$ is true as $u_1\in\mathrm{p}_1(S)$,
then there are exactly two $B\mid_{\{u,v\}}\in\{0,\pm\}^{\{u,v\}}$ with $\mathrm{X}_{B}\cong \ref{item:comparisonproofcaseLhassize6:oneadditionaledgeintersectingC4andtwoisolatedvertices}$,
namely those which satisfy ($B[u]\in\{\pm\}$ and $B[v] = 0$). If it is true
as $v_1\in\mathrm{p}_1(S)$, then again there are exactly two such $B\mid_{\{u,v\}}$, namely those
which satisfy ($B[u] = 0$ and $B[v] \in \{\pm\}$). It follows that
without having to reexamine the subcases \ref{numberofmatrixrealizationsoftype:threeisolatedvertices:case:lvertu1v1rvertbackslashp1Sequals1}.\ref{numberofmatrixrealizationsoftype:threeisolatedvertices:case:lvertu1v1rvertbackslashp1Sequals1:u1doesnotequalv1andbracesu1v1bracesintersectp1Sisasingleton}.\ref{numberofmatrixrealizationsoftype:threeisolatedvertices:case:lvertu1v1rvertbackslashp1Sequals1:u1doesnotequalv1andbracesu1v1bracesintersectp1Sisasingletonu1equalsa1}--\ref{numberofmatrixrealizationsoftype:threeisolatedvertices:case:lvertu1v1rvertbackslashp1Sequals1}.\ref{numberofmatrixrealizationsoftype:threeisolatedvertices:case:lvertu1v1rvertbackslashp1Sequals1:u1doesnotequalv1andbracesu1v1bracesintersectp1Sisasingleton}.\ref{numberofmatrixrealizationsoftype:threeisolatedvertices:case:lvertu1v1rvertbackslashp1Sequals1:u1doesnotequalv1andbracesu1v1bracesintersectp1Sisasingletonv1equalsc1} we know that there are twice as many realizations of \ref{item:comparisonproofcaseLhassize6:oneadditionaledgeintersectingC4andtwoisolatedvertices} by $B\mid_{\{u,v\}}$ in the case
\ref{numberofmatrixrealizationsoftype:threeisolatedvertices:case:lvertu1v1rvertbackslashp1Sequals1}.\ref{numberofmatrixrealizationsoftype:threeisolatedvertices:case:lvertu1v1rvertbackslashp1Sequals1:u1doesnotequalv1andbracesu1v1bracesintersectp1Sisasingleton}
than of \ref{numberofmatrixrealizationsoftype:threeisolatedvertices}. Therefore,
if \ref{numberofmatrixrealizationsoftype:threeisolatedvertices:case:lvertu1v1rvertbackslashp1Sequals1}.\ref{numberofmatrixrealizationsoftype:threeisolatedvertices:case:lvertu1v1rvertbackslashp1Sequals1:u1doesnotequalv1andbracesu1v1bracesintersectp1Sisasingleton},
then there are exactly $8\cdot ((n-1)-2)\cdot \binom{(n-1)-2}{2}$ realizations
of \ref{item:comparisonproofcaseLhassize6:oneadditionaledgeintersectingC4andtwoisolatedvertices} by $B\mid_{\{u,v\}}$.
\item Concerning contributions to \ref{numberofmatrixrealizationsoftype:oneadditionaldisjointedgeandoneisolatedvertex}, we have to distinguish in what way property
$\lvert \{ u_1,v_1 \} \cap \mathrm{p}_1(S) \rvert = 1 $ in
\eqref{eq:knowledgeincase:numberofmatrixrealizationsoftype:threeisolatedvertices:case:lvertu1v1rvertbackslashp1Sequals1:u1doesnotequalv1andbracesu1v1bracesintersectp1Sisasingleton} is satisfied. If $u_1\in\mathrm{p}_1(S)$ but $v_1\notin\mathrm{p}_1(S)$, then
$\mathrm{X}_{B} \cong \ref{numberofmatrixrealizationsoftype:oneadditionaldisjointedgeandoneisolatedvertex}$ if and only if $B[u]=0$ and $B[v]\in\{\pm\}$, hence in this case
there exist $2$ different $B\mid_{\{u,v\}}$ with $\mathrm{X}_{B} \cong \ref{numberofmatrixrealizationsoftype:oneadditionaldisjointedgeandoneisolatedvertex}$. If $u_1\notin\mathrm{p}_1(S)$ but $v_1\in\mathrm{p}_1(S)$, then $\mathrm{X}_{B} \cong \ref{numberofmatrixrealizationsoftype:oneadditionaldisjointedgeandoneisolatedvertex}$ if and only if $B[u]\in\{\pm\}$
and $B[v] = 0$, hence in this case there again exist $2$ different $B\mid_{\{u,v\}}$
with $\mathrm{X}_{B} \cong \ref{numberofmatrixrealizationsoftype:oneadditionaldisjointedgeandoneisolatedvertex}$. It follows that if \ref{numberofmatrixrealizationsoftype:threeisolatedvertices:case:lvertu1v1rvertbackslashp1Sequals1}.\ref{numberofmatrixrealizationsoftype:threeisolatedvertices:case:lvertu1v1rvertbackslashp1Sequals1:u1doesnotequalv1andbracesu1v1bracesintersectp1Sisasingleton}, then there there are $2$-times as many realizations of \ref{item:comparisonproofcaseLhassize6:oneadditionaldisjointedgeandoneisolatedvertex} than of \ref{item:comparisonproofcaseLhassize6:threeisolatedvertices} by $B\mid_{\{u,v\}}$. Therefore, if \ref{numberofmatrixrealizationsoftype:threeisolatedvertices:case:lvertu1v1rvertbackslashp1Sequals1}.\ref{numberofmatrixrealizationsoftype:threeisolatedvertices:case:lvertu1v1rvertbackslashp1Sequals1:u1doesnotequalv1andbracesu1v1bracesintersectp1Sisasingleton}, then there are exactly
$2\cdot 4\cdot ((n-1)-2)\cdot \binom{(n-1)-2}{2} = 8\cdot (n-3)\cdot \binom{n-3}{2}$
realizations of \ref{item:comparisonproofcaseLhassize6:oneadditionaldisjointedgeandoneisolatedvertex} by $B\mid_{\{u,v\}}$.
\item Concerning contributions to \ref{numberofmatrixrealizationsoftype:twoadditionaledgesonlyoneofthemdisjoint}, note that no matter how \eqref{eq:knowledgeincase:numberofmatrixrealizationsoftype:threeisolatedvertices:case:lvertu1v1rvertbackslashp1Sequals1:u1doesnotequalv1andbracesu1v1bracesintersectp1Sisasingleton} is satisfied, we have
$\mathrm{X}_{B} \cong \ref{numberofmatrixrealizationsoftype:twoadditionaledgesonlyoneofthemdisjoint}$ if and only if ($B[u]\in\{\pm\}$ and $B[v]\in\{\pm\}$). Hence, if \ref{numberofmatrixrealizationsoftype:threeisolatedvertices:case:lvertu1v1rvertbackslashp1Sequals1}.\ref{numberofmatrixrealizationsoftype:threeisolatedvertices:case:lvertu1v1rvertbackslashp1Sequals1:u1doesnotequalv1andbracesu1v1bracesintersectp1Sisasingleton},
then there are $4$-times as many realizations of \ref{numberofmatrixrealizationsoftype:twoadditionaledgesonlyoneofthemdisjoint} than there are of \ref{numberofmatrixrealizationsoftype:threeisolatedvertices}, that is, if \ref{numberofmatrixrealizationsoftype:threeisolatedvertices:case:lvertu1v1rvertbackslashp1Sequals1}.\ref{numberofmatrixrealizationsoftype:threeisolatedvertices:case:lvertu1v1rvertbackslashp1Sequals1:u1doesnotequalv1andbracesu1v1bracesintersectp1Sisasingleton}, then there are exactly
$4\cdot 4\cdot ( (n-1)-2)\cdot \binom{(n-1)-2}{2} = 16\cdot (n-3)\cdot \binom{n-3}{2}$
realizations of type \ref{numberofmatrixrealizationsoftype:twoadditionaledgesonlyoneofthemdisjoint} by $B\mid_{\{u,v\}}$.
\item Concerning contributions to \ref{numberofmatrixrealizationsoftype:twoadditionalnondisjointedgesdisjointfromC4}, note that \eqref{eq:knowledgeincase:numberofmatrixrealizationsoftype:threeisolatedvertices:case:lvertu1v1rvertbackslashp1Sequals1:u1doesnotequalv1andbracesu1v1bracesintersectp1Sisasingleton} says that
$u_1\neq v_1$ and $u_2\neq v_2$, and this makes it impossible to create a
$2$-path outside of $\mathrm{X}_{B\mid_S}\cong C^4$. Therefore, if \ref{numberofmatrixrealizationsoftype:threeisolatedvertices:case:lvertu1v1rvertbackslashp1Sequals1}.\ref{numberofmatrixrealizationsoftype:threeisolatedvertices:case:lvertu1v1rvertbackslashp1Sequals1:u1doesnotequalv1andbracesu1v1bracesintersectp1Sisasingleton}, there is no contribution to \ref{numberofmatrixrealizationsoftype:twoadditionalnondisjointedgesdisjointfromC4}.
\end{enumerate}
We now take stock of what we found in the subcases
\ref{numberofmatrixrealizationsoftype:threeisolatedvertices:case:lvertu1v1rvertbackslashp1Sequals1}.\ref{numberofmatrixrealizationsoftype:threeisolatedvertices:case:lvertu1v1rvertbackslashp1Sequals1:u1equalsv1andbracesu1v1bracesintersectp1Sisempty}
and \ref{numberofmatrixrealizationsoftype:threeisolatedvertices:case:lvertu1v1rvertbackslashp1Sequals1}.\ref{numberofmatrixrealizationsoftype:threeisolatedvertices:case:lvertu1v1rvertbackslashp1Sequals1:u1doesnotequalv1andbracesu1v1bracesintersectp1Sisasingleton}
in order to know what the entire case \ref{numberofmatrixrealizationsoftype:threeisolatedvertices:case:lvertu1v1rvertbackslashp1Sequals1} contributes to \ref{numberofmatrixrealizationsoftype:oneadditionaledgeintersectingC4andtwoisolatedvertices}--\ref{numberofmatrixrealizationsoftype:twoadditionalnondisjointedgesdisjointfromC4}.
Since \ref{numberofmatrixrealizationsoftype:threeisolatedvertices:case:lvertu1v1rvertbackslashp1Sequals1}.\ref{numberofmatrixrealizationsoftype:threeisolatedvertices:case:lvertu1v1rvertbackslashp1Sequals1:u1equalsv1andbracesu1v1bracesintersectp1Sisempty}
did not contribute to $\lvert(\leftsubscript{\mathrm{ul}}{\upX}^{6,n,n})^{-1}\ref{item:comparisonproofcaseLhassize6:oneadditionaledgeintersectingC4andtwoisolatedvertices}\rvert$ but \ref{numberofmatrixrealizationsoftype:threeisolatedvertices:case:lvertu1v1rvertbackslashp1Sequals1}.\ref{numberofmatrixrealizationsoftype:threeisolatedvertices:case:lvertu1v1rvertbackslashp1Sequals1:u1doesnotequalv1andbracesu1v1bracesintersectp1Sisasingleton} did contribute
$8\cdot (n-3)\cdot \binom{n-3}{2}$, it follows that if \ref{numberofmatrixrealizationsoftype:threeisolatedvertices:case:lvertu1v1rvertbackslashp1Sequals1}, then there are
exactly $8\cdot (n-3)\cdot \binom{n-3}{2}$ realizations of type \ref{item:comparisonproofcaseLhassize6:oneadditionaledgeintersectingC4andtwoisolatedvertices}
by $B\mid_{\{u,v\}}$.
Since \ref{numberofmatrixrealizationsoftype:threeisolatedvertices:case:lvertu1v1rvertbackslashp1Sequals1}.\ref{numberofmatrixrealizationsoftype:threeisolatedvertices:case:lvertu1v1rvertbackslashp1Sequals1:u1equalsv1andbracesu1v1bracesintersectp1Sisempty}
contributed $4\cdot (n-3) \cdot\binom{n-3}{2}$ to $\lvert(\leftsubscript{\mathrm{ul}}{\upX}^{6,n,n})^{-1}\ref{item:comparisonproofcaseLhassize6:oneadditionaldisjointedgeandoneisolatedvertex}\rvert$ while \ref{numberofmatrixrealizationsoftype:threeisolatedvertices:case:lvertu1v1rvertbackslashp1Sequals1}.\ref{numberofmatrixrealizationsoftype:threeisolatedvertices:case:lvertu1v1rvertbackslashp1Sequals1:u1doesnotequalv1andbracesu1v1bracesintersectp1Sisasingleton} contributed
$8\cdot (n-3)\cdot\binom{n-3}{2}$, it follows that if \ref{numberofmatrixrealizationsoftype:threeisolatedvertices:case:lvertu1v1rvertbackslashp1Sequals1}, then there are
exactly $12\cdot (n-3) \cdot\binom{n-3}{2}$ realizations of type \ref{item:comparisonproofcaseLhassize6:oneadditionaldisjointedgeandoneisolatedvertex} by $B\mid_{\{u,v\}}$.
Since \ref{numberofmatrixrealizationsoftype:threeisolatedvertices:case:lvertu1v1rvertbackslashp1Sequals1}.\ref{numberofmatrixrealizationsoftype:threeisolatedvertices:case:lvertu1v1rvertbackslashp1Sequals1:u1equalsv1andbracesu1v1bracesintersectp1Sisempty}
did not contribute to $\lvert(\leftsubscript{\mathrm{ul}}{\upX}^{6,n,n})^{-1}\ref{item:comparisonproofcaseLhassize6:twoadditionaledgesonlyoneofthemdisjoint}\rvert$ but
\ref{numberofmatrixrealizationsoftype:threeisolatedvertices:case:lvertu1v1rvertbackslashp1Sequals1}.\ref{numberofmatrixrealizationsoftype:threeisolatedvertices:case:lvertu1v1rvertbackslashp1Sequals1:u1doesnotequalv1andbracesu1v1bracesintersectp1Sisasingleton} did contribute $16 \cdot (n-3) \cdot \binom{n-3}{2}$, it follows that if
\ref{numberofmatrixrealizationsoftype:threeisolatedvertices:case:lvertu1v1rvertbackslashp1Sequals1}, then there are exactly $16 \cdot (n-3) \cdot \binom{n-3}{2}$ realizations of type \ref{item:comparisonproofcaseLhassize6:twoadditionaledgesonlyoneofthemdisjoint} by $B\mid_{\{u,v\}}$.
Since \ref{numberofmatrixrealizationsoftype:threeisolatedvertices:case:lvertu1v1rvertbackslashp1Sequals1}.\ref{numberofmatrixrealizationsoftype:threeisolatedvertices:case:lvertu1v1rvertbackslashp1Sequals1:u1equalsv1andbracesu1v1bracesintersectp1Sisempty}
contributed $4\cdot (n-3)\cdot\binom{n-3}{2}$ to $\lvert(\leftsubscript{\mathrm{ul}}{\upX}^{6,n,n})^{-1}\ref{item:comparisonproofcaseLhassize6:twoadditionalnondisjointedgesdisjointfromC4}\rvert$ while \ref{numberofmatrixrealizationsoftype:threeisolatedvertices:case:lvertu1v1rvertbackslashp1Sequals1}.\ref{numberofmatrixrealizationsoftype:threeisolatedvertices:case:lvertu1v1rvertbackslashp1Sequals1:u1doesnotequalv1andbracesu1v1bracesintersectp1Sisasingleton} did not contribute
anything, it follows that if \ref{numberofmatrixrealizationsoftype:threeisolatedvertices:case:lvertu1v1rvertbackslashp1Sequals1}, then there are exactly
$4\cdot (n-3)\cdot\binom{n-3}{2}$ realizations of type \ref{item:comparisonproofcaseLhassize6:twoadditionalnondisjointedgesdisjointfromC4} by $B\mid_{\{u,v\}}$.
Since the case \ref{numberofmatrixrealizationsoftype:threeisolatedvertices:case:lvertu1v1rvertbackslashp1Sequals2} is symmetric to the case \ref{numberofmatrixrealizationsoftype:threeisolatedvertices:case:lvertu1v1rvertbackslashp1Sequals1} via interchanging
the subscripts $1$ and $2$, we will get the same contributions to \ref{numberofmatrixrealizationsoftype:oneadditionaledgeintersectingC4andtwoisolatedvertices}--\ref{numberofmatrixrealizationsoftype:twoadditionalnondisjointedgesdisjointfromC4} as in the
case \ref{numberofmatrixrealizationsoftype:threeisolatedvertices:case:lvertu1v1rvertbackslashp1Sequals1}. We therefore have to double each of the four results found for
\ref{numberofmatrixrealizationsoftype:threeisolatedvertices:case:lvertu1v1rvertbackslashp1Sequals1} to get the correct numbers of realizations of
types \ref{item:comparisonproofcaseLhassize6:oneadditionaledgeintersectingC4andtwoisolatedvertices}--\ref{item:comparisonproofcaseLhassize6:twoadditionalnondisjointedgesdisjointfromC4}. This proves \ref{numberofmatrixrealizationsoftype:oneadditionaledgeintersectingC4andtwoisolatedvertices}--\ref{numberofmatrixrealizationsoftype:twoadditionalnondisjointedgesdisjointfromC4}.
As to \ref{numberofmatrixrealizationsoftype:fourisolatedvertices}, it suffices
to note that the two values of $B\mid_{\{u,v\}}$ are determined: since there does
not exist an edge outside $\mathrm{X}_{B\mid_S}$ $\cong$ $C^4$, they both must be zero.
Therefore \ref{numberofmatrixrealizationsoftype:fourisolatedvertices} is the number
of $\{u,v\}\in\binom{[n-1]^2}{2}$ such that $\mathrm{X}_{B\mid_{S}\sqcup \{0\}^{\{u,v\}}} \cong
\ref{item:comparisonproofcaseLhassize6:fourisolatedvertices}$. By definition of $S$
the latter is equivalent to saying that $\mathrm{X}_{B\mid_{S}\sqcup \{0\}^{\{u,v\}}}$ has exactly
eight vertices. It follows from Definition \ref{def:XBandecXB} that this is the case
if and only if simultaneously
\begin{equation}\label{eq:conditioninproofof:numberofmatrixrealizationsoftype:fourisolatedvertices}
\lvert \{ u_1, v_1 \} \setminus \mathrm{p}_1(S) \rvert = 2
\quad \text{and}\quad
\lvert \{ u_2, v_2 \} \setminus \mathrm{p}_2(S) \rvert = 2\quad .
\end{equation}
Due to $u_1<v_1$, the number of $\{u_1,v_1\}\subseteq [n-1]$
with $\lvert \{ u_1, v_1 \}\setminus \mathrm{p}_1(S) \rvert = 2$ is $\binom{n-3}{2}$.
Since we only assume $u\prec v$ and hence both $u_2<v_2$ and $u_2>v_2$ are possible,
for each of these $\binom{n-3}{2}$ different $\{u_1,v_1\}$ there are
$2\cdot \binom{n-3}{2}$ different $\{u_2,v_2\}$ with
$\lvert \{ u_2, v_2 \} \setminus \mathrm{p}_2(S) \rvert = 2$. This proves
\ref{numberofmatrixrealizationsoftype:fourisolatedvertices}.
As to \ref{numberofmatrixrealizationsoftype:oneadditionaldisjointedgeandtwoisolatedvertices} and \ref{numberofmatrixrealizationsoftype:twoadditionaldisjointedgesdisjointfromC4}, note that for both ismorphism types \ref{item:comparisonproofcaseLhassize6:oneadditionaldisjointedgeandtwoisolatedvertices} and \ref{item:comparisonproofcaseLhassize6:twoadditionaldisjointedgesdisjointfromC4} it is necessary
that $X_{ B\mid_{S} \sqcup \{0\}^{\{u,v\}} } \cong \ref{item:comparisonproofcaseLhassize6:fourisolatedvertices}$. We can therefore prove \ref{numberofmatrixrealizationsoftype:oneadditionaldisjointedgeandtwoisolatedvertices} and \ref{numberofmatrixrealizationsoftype:twoadditionaldisjointedgesdisjointfromC4} by reexamining the proof of \ref{numberofmatrixrealizationsoftype:fourisolatedvertices}. In whatever way
\eqref{eq:conditioninproofof:numberofmatrixrealizationsoftype:fourisolatedvertices}
is satisfied, there are exactly $4$ different $B\mid_{\{u,v\}}$ with
$\mathrm{X}_{B} \cong \ref{item:comparisonproofcaseLhassize6:oneadditionaldisjointedgeandtwoisolatedvertices}$, namely those satisfying
\begin{equation}\label{eq:conditioninproofof:numberofmatrixrealizationsoftype:oneadditionaldisjointedgeandtwoisolatedvertices}
(B[u]\in\{\pm\}\; \text{and}\; B[v]=0) \quad \text{or}\quad
(B[u]=0\; \text{and}\; B[v]\in\{\pm\})\quad ,
\end{equation}
but there are also $4$ different $B\mid_{\{u,v\}}$ with
$\mathrm{X}_{B} \cong \ref{item:comparisonproofcaseLhassize6:twoadditionaldisjointedgesdisjointfromC4}$, namely those satisfying
\begin{equation}\label{eq:conditioninproofof:numberofmatrixrealizationsoftype:numberofmatrixrealizationsoftype:twoadditionaldisjointedgesdisjointfromC4}
B[u]\in\{\pm\}\; \text{and}\; B[v]\in\{\pm\} \quad .
\end{equation}
This proves both $\lvert(\leftsubscript{\mathrm{ul}}{\upX}^{6,n,n})^{-1}\ref{item:comparisonproofcaseLhassize6:oneadditionaldisjointedgeandtwoisolatedvertices}\rvert =\lvert(\leftsubscript{\mathrm{ul}}{\upX}^{6,n,n})^{-1}\ref{item:comparisonproofcaseLhassize6:twoadditionaldisjointedgesdisjointfromC4}\rvert = 4 \cdot \lvert(\leftsubscript{\mathrm{ul}}{\upX}^{6,n,n})^{-1}\ref{item:comparisonproofcaseLhassize6:fourisolatedvertices}\rvert$, and therefore both \ref{numberofmatrixrealizationsoftype:oneadditionaldisjointedgeandtwoisolatedvertices} and \ref{numberofmatrixrealizationsoftype:twoadditionaldisjointedgesdisjointfromC4}. The proof of \ref{item:numberofrealizationsofnonforestswhenkequals6} is now complete.
\end{proof}
The relations \ref{linearrelationbetweent2andt3:kequals5}--\ref{linearrelationbetweent16tot18:kequals6} in Lemma \ref{lem:relationsamongthenumbersofmatrixrealizations}
give us a plausibility check (i.e. necessary conditions) for the explicit formulas
$\lvert(\leftsubscript{\mathrm{ul}}{\upX}^{6,n,n})^{-1}\ref{item:comparisonproofcaseLhassize6:oneisolatedvertex}\rvert$, $\dotsc$, $\lvert(\leftsubscript{\mathrm{ul}}{\upX}^{6,n,n})^{-1}\ref{item:comparisonproofcaseLhassize6:twoadditionaldisjointedgesdisjointfromC4}\rvert$ that we found in \ref{numberofmatrixrealizationsoftype:oneisolatedvertex:kequals5}--\ref{numberofmatrixrealizationsoftype:twoadditionaldisjointedgesdisjointfromC4}. For brevity let $x:= n-3$ and
$y := \binom{n-3}{2}$. Then, indeed, the explicit
formulas that we found in \ref{item:numberofrealizationsofnonforestswhenkequals5}
and \ref{item:numberofrealizationsofnonforestswhenkequals6} pass the test: the
formulas in \ref{item:numberofrealizationsofnonforestswhenkequals5} evidently
satisfy \ref{linearrelationbetweent2andt3:kequals5} and \ref{linearrelationbetweent5andt7:kequals5}. Moreover, since $(3^2 - 1) \cdot \bigl ( 8 x^2 + 8 y \bigr )
= 24 x^2 + 32 y + 8 x^2 + 16 y + 16 x^2 + 16 x^2 + 16 y$, the formulas
\ref{numberofmatrixrealizationsoftype:twoisolatedvertices}--\ref{numberofmatrixrealizationsoftype:C4intersectingatwopathinitsinnervertex} satisfy \ref{linearrelationbetweent4tot10:kequals6}. Since $(3^2 - 1) \cdot 10 xy = 16 x y + 24 x y + 32 x y + 8 x y$,
the formulas in \ref{numberofmatrixrealizationsoftype:threeisolatedvertices}--\ref{numberofmatrixrealizationsoftype:twoadditionalnondisjointedgesdisjointfromC4}
satisfy \ref{linearrelationbetweent11tot15:kequals6}. Since
$(3^2-1)\cdot 2y^2 = 8y^2 + 8y^2$, the formulas in \ref{numberofmatrixrealizationsoftype:fourisolatedvertices}--\ref{numberofmatrixrealizationsoftype:twoadditionaldisjointedgesdisjointfromC4} satisfy \ref{linearrelationbetweent16tot18:kequals6}.
\subsection{Counting failures of equality of $\textup{\textsf{P}}_{\mathrm{chio}}$ and $\textup{\textsf{P}}_{\mathrm{lcf}}$}
While determining an absolute cardinality
$\lvert(\leftsubscript{\mathrm{ul}}{\upX}^{k,n,n})^{-1}(\mathfrak{X})\rvert$ seems to necessitate work
specifically depending on the isomorphism type $\mathfrak{X}$, the ratio of
all \emph{balanced} matrix realizations to \emph{all} realizations
is easy to compute since it is determined by the Betti number of $\mathfrak{X}$
alone. This is the content of \ref{eq:ratioofbalancedmatrixrealizationsofanisomorphismtypetoallrealizations} in the following lemma:
\begin{lemma}\label{lem:forfixedisomorphismtyperatioofbalancedrealizationsdependsonthebettinumberalone}
For every $(s,t)\in\mathbb{Z}_{\geq 2}^2$, every $0\leq k \leq (s-1)(t-1)$, every unlabelled
bipartite graph $\mathfrak{X}$ and every $\beta\in \mathbb{Z}_{\geq 1}$,
\begin{enumerate}[label={\rm(E\arabic{*})}]
\item\label{eq:ratioofbalancedmatrixrealizationsofanisomorphismtypetoallrealizations} $\lvert\{ B\in(\leftsubscript{\mathrm{ul}}{\upX}^{k,s,t})^{-1}( \mathfrak{X}) \colon
(\mathrm{X}_B,\sigma_B)\;\mathrm{balanced}\} \rvert = (\tfrac12)^{\beta_1(\mathfrak{X})}
\cdot\lvert(\leftsubscript{\mathrm{ul}}{\upX}^{k,s,t})^{-1} (\mathfrak{X})\rvert$ \quad ,
\item\label{eq:expansionofamatrixfailuresetwithapositiveratiobyisomorphismtypes}
$\lvert\mathcal{F}_{\cdot 2^\beta}^{\mathrm{M}}(k,s,t)\rvert =
\sum_{\mathfrak{X}\in\im(\leftsubscript{\mathrm{ul}}{\upX}^{k,s,t})\colon \beta_1(\mathfrak{X}) = \beta}\; (\tfrac12)^\beta \cdot
\bigl \lvert \bigl( \leftsubscript{\mathrm{ul}}{\upX}^{k,s,t}\bigr)^{-1} (\mathfrak{X})\bigr \rvert$\quad ,
\item\label{eq:expansionofamatrixfailuresetwithzeroreatiobyisomorphismtypes}
$\lvert\mathcal{F}_{\cdot 0}^{\mathrm{M}}(k,s,t)\rvert =
\sum_{\mathfrak{X}\in\im(\leftsubscript{\mathrm{ul}}{\upX}^{k,s,t})\colon \beta_1(\mathfrak{X}) \geq 1}\;
\bigl (1-(\tfrac12)^{\beta_1(\mathfrak{X})} \bigr )
\cdot\lvert(\leftsubscript{\mathrm{ul}}{\upX}^{k,s,t})^{-1} (\mathfrak{X})\rvert$ \quad .
\end{enumerate}
\end{lemma}
\begin{proof}
If $\mathcal{M}$ is a set of matrices, let us define
$\mathrm{Dom}(\mathcal{M}) := \{ \mathrm{Dom}(B)\colon B\in\mathcal{M}\}$ and
$\mathrm{Supp}(\mathcal{M}) := \{ \mathrm{Supp}(B)\colon B\in\mathcal{M}\}$. Moreover,
if $S$ is a set, $\mathcal{S}\subseteq \mathfrak{P}(S)$ a set of subsets and
$U\in \mathfrak{P}(S)$ a subset, then $U\cap\mathcal{S} :=
\{ U\cap S\colon S\in\mathcal{S}\}$. Using these notations, we can prove
\ref{eq:ratioofbalancedmatrixrealizationsofanisomorphismtypetoallrealizations} by
the following calculation: for every unlabelled $\mathfrak{X}$ we have
$\lvert\{$ $B$ $\in$ $(\leftsubscript{\mathrm{ul}}{\upX}^{k,s,t})^{-1}$ $(\mathfrak{X})\colon$
$(\mathrm{X}_B,\sigma_B)$ balanced $\}\rvert$ $=$
$\sum_{J\in \mathrm{Supp}( (\leftsubscript{\mathrm{ul}}{\upX}^{k,s,t})^{-1}(\mathfrak{X}))}$
$\lvert \{$ $B$ $\in$ $(\leftsubscript{\mathrm{ul}}{\upX}^{k,s,t})^{-1}(\mathfrak{X})\colon$
$(\mathrm{X}_B,\sigma_B)$ balanced, $\mathrm{Supp}(B)$ $=$ $J$ $\} \rvert$
$=$ (\ref{item:numberofbalancedsignfunctions} in Lemma \ref{lem:equivalenceofexistenceofbconstantrpropervertex2coloringandcyclicallyrevenness}) $=$
$\sum_{J\in\mathrm{Supp}( (\leftsubscript{\mathrm{ul}}{\upX}^{k,s,t})^{-1}(\mathfrak{X}))}$
$2^{\lvert J \rvert - \beta_1(\mathfrak{X})}$ $=$ $(\tfrac12)^{\beta_1(\mathfrak{X})}$
$\sum_{I\in\mathrm{Dom}((\leftsubscript{\mathrm{ul}}{\upX}^{k,s,t})^{-1}(\mathfrak{X}))}$
$\sum_{J\in I\cap\mathrm{Supp}( (\leftsubscript{\mathrm{ul}}{\upX}^{k,s,t})^{-1}(\mathfrak{X}))}$ $2^{\lvert J \rvert}$ $=$ (directly
from the definitions) $=$
$(\tfrac12)^{\beta_1(\mathfrak{X})}$ $\lvert (\leftsubscript{\mathrm{ul}}{\upX}^{k,s,t})^{-1}(\mathfrak{X})\rvert$.
As to \ref{eq:expansionofamatrixfailuresetwithapositiveratiobyisomorphismtypes},
this is true since $\lvert\mathcal{F}_{\cdot 2^\beta}^{\mathrm{M}}(k,s,t)\rvert$
$=$ (\ref{relationbeweenchiomeasureandlazycoinflipmeasuregovernedbyfirstbettinumber} in Theorem \ref{thm:graphtheoreticalcharacterizationofthechiomeasure}) $=$
$\sum_{\mathfrak{X}\in\im(\leftsubscript{\mathrm{ul}}{\upX}^{k,s,t})\colon\beta_1(\mathfrak{X})=\beta}$
$\lvert \{$ $B\in(\leftsubscript{\mathrm{ul}}{\upX}^{k,s,t})^{-1}(\mathfrak{X})\colon$ $(\mathrm{X}_B,\sigma_B)$
$\mathrm{balanced}$ $\}\rvert$ $=$ (by \ref{eq:ratioofbalancedmatrixrealizationsofanisomorphismtypetoallrealizations}) $=$
$\sum_{\mathfrak{X}\in\im(\leftsubscript{\mathrm{ul}}{\upX}^{k,s,t})\colon\beta_1(\mathfrak{X})=\beta}$
$(\tfrac12)^\beta\cdot\bigl\lvert(\leftsubscript{\mathrm{ul}}{\upX}^{k,s,t}\bigr)^{-1}(\mathfrak{X})\bigr\rvert$.
As to \ref{eq:expansionofamatrixfailuresetwithzeroreatiobyisomorphismtypes}, note
that $\lvert\mathcal{F}_{\cdot 0}^{\mathrm{M}}(k,s,t)\rvert$ $=$
(\ref{characterizationofwhenchiomeasureispositive} in Theorem
\ref{thm:graphtheoreticalcharacterizationofthechiomeasure}) $=$
$\sum_{\mathfrak{X}\in\im(\leftsubscript{\mathrm{ul}}{\upX}^{k,s,t})\colon\beta_1(\mathfrak{X})\geq 1}$ $\lvert\{$
$B$ $\in$ $(\leftsubscript{\mathrm{ul}}{\upX}^{k,s,t})^{-1}$ $(\mathfrak{X})\colon$ $(\mathrm{X}_B,$ $\sigma_B)$
$\mathrm{not}$ $\mathrm{balanced}$ $\}\rvert$ $=$
(using \ref{eq:ratioofbalancedmatrixrealizationsofanisomorphismtypetoallrealizations}) $=$ $\sum_{\mathfrak{X}\in\im(\leftsubscript{\mathrm{ul}}{\upX}^{k,s,t})\colon\beta_1(\mathfrak{X})\geq 1}$
$\bigl (1-(\tfrac12)^{\beta_1(\mathfrak{X})}\bigr) \cdot \lvert(\leftsubscript{\mathrm{ul}}{\upX}^{k,s,t})^{-1} (\mathfrak{X})\rvert$.
\end{proof}
The fewer the number $\mathrm{dom}(B)$ of entries specified, the larger an
entry-specification event $\mathcal{E}_B^{[n-1]^2}$ is (as a set). Any two
probability measures by definition agree on the largest possible event, the
entire sample space. The following theorem explores to what extent
$\textup{\textsf{P}}_{\mathrm{chio}}$ and $\textup{\textsf{P}}_{\mathrm{lcf}}$ agree on successively smaller entry-specification
events, descending down to as much as six specifications.
\begin{theorem}[number of exceptions to equality of
$\textup{\textsf{P}}_{\mathrm{chio}}$ and $\textup{\textsf{P}}_{\mathrm{lcf}}$ on large entry-specification events]
\label{thm:comparativecountingtheorem}
In the following statements let $\emptyset\subseteq I \subseteq [n-1]^2$,
$B\in \{0,\pm\}^I$ and $\mathcal{E}_B:=\mathcal{E}_B^{[n-1]^2}$.
\begin{enumerate}[label={\rm(Ex\arabic{*})}, start=3]
\item\label{comparativecountingtheorem:item:uptto3entries} For each of
the $\sum_{0\leq k \leq 3} 3^k \cdot \binom{(n-1)^2}{k}\sim \frac92 \cdot n^6$
possible events $\mathcal{E}_B$ with $0\leq \mathrm{dom}(B) \leq 3$, it
is true that $\textup{\textsf{P}}_{\mathrm{chio}}[\mathcal{E}_B] = \textup{\textsf{P}}_{\mathrm{lcf}} [\mathcal{E}_B]
= (\frac12)^{\mathrm{dom}(B)+\mathrm{supp}(B)}$. \\
\item\label{comparativecountingtheorem:item:fourentriesspecified}
Among the $3^4 \cdot \binom{(n-1)^2}{4} \sim \frac{27}{8} \cdot n^8$ possible events
$\mathcal{E}_B$ with $\mathrm{dom}(B)=4$, there are precisely
$\lvert \mathcal{F}^{\mathrm{M}} (4,n) \rvert
= 2^4 \cdot \binom{n-1}{2}\cdot \binom{n-1}{2} \sim 4\cdot n^4$ events
for which $\textup{\textsf{P}}_{\mathrm{chio}}[\mathcal{E}_B] = \textup{\textsf{P}}_{\mathrm{lcf}} [\mathcal{E}_B]$ does not
hold. Of these, we have $\tfrac{\lvert\mathcal{F}_{\cdot 0}^{\mathrm{M}}(4,n)\rvert}{\lvert\mathcal{F}^{\mathrm{M}}(4,n)\rvert} = \tfrac{\lvert\mathcal{F}_{\cdot 2}^{\mathrm{M}}(4,n)\rvert}{\lvert\mathcal{F}^{\mathrm{M}}(4,n)\rvert} = \tfrac12$.
\item\label{comparativecountingtheorem:item:fiveentriesspecified}
Among the $3^5 \cdot \binom{(n-1)^2}{5} \sim \frac{81}{40} \cdot n^{10}$
different events $\mathcal{E}_B$ with $\mathrm{dom}(B) = 5$, there are precisely
$\lvert \mathcal{F}^{\mathrm{M}}(5,n) \rvert =
48\cdot ((n-1)^2 - 4) \cdot \binom{n-1}{2}\cdot \binom{n-1}{2}\sim 12\cdot n^6$
events for which $\textup{\textsf{P}}_{\mathrm{chio}}[\mathcal{E}_B] = \textup{\textsf{P}}_{\mathrm{lcf}} [\mathcal{E}_B]$ does
not hold. Of these, we have
$\tfrac{\lvert \mathcal{F}_{\cdot 0}^{\mathrm{M}}(5,n) \rvert}{\lvert \mathcal{F}^{\mathrm{M}}(5,n) \rvert} = \tfrac{\lvert \mathcal{F}_{\cdot 2}^{\mathrm{M}}(5,n) \rvert}{\lvert \mathcal{F}^{\mathrm{M}}(5,n) \rvert} = \tfrac12$.
\item\label{comparativecountingtheorem:item:sixentriesspecified} Among the
$3^6 \cdot \binom{(n-1)^2}{6} \sim \frac{81}{80} n^{12}$ different events
$\mathcal{E}_B$ with $\mathrm{dom}(B) = 6$, there are precisely
$\lvert \mathcal{F}^{\mathrm{M}}(6,n) \rvert =
18 n^8 - 180 n^7 + \frac{1868}{3} n^6 - \frac{2176}{3} n^5 -
\frac{754}{3} n^4 + \frac{428}{3} n^3 + \frac{8144}{3} n^2 -
\frac{11536}{3} n + 1504 \sim 18n^8$ events for which
$\textup{\textsf{P}}_{\mathrm{chio}}[\mathcal{E}_B] = \textup{\textsf{P}}_{\mathrm{lcf}} [\mathcal{E}_B]$ does not
hold. Of these, we have $\lvert \mathcal{F}_{\cdot 0}^{\mathrm{M}}(6,n) \rvert =
9 n^8 - 90 n^7 + \tfrac{934}{3} n^6 - 360 n^5 - \tfrac{449}{3} n^4 + 154 n^3 +
\tfrac{3664}{3} n^2 - 1816 n + 720 \sim 9n^8$,
$\lvert \mathcal{F}_{\cdot 2}^{\mathrm{M}}(6,n) \rvert =
9 n^8 - 90 n^7 + \tfrac{934}{3} n^6 - 368 n^5 - \tfrac{233}{3} n^4 - 94 n^3 +
\tfrac{4888}{3} n^2 - 2136 n + 816 \sim 9n^8$, and
$\lvert \mathcal{F}_{\cdot 4}^{\mathrm{M}}(6,n) \rvert$ $=$
$\tfrac83 n^5 - 24 n^4 + \tfrac{248}{3} n^3 - 136 n^2 + \tfrac{320}{3} n - 32 \sim
\tfrac83 n^5$, hence in particular
$\lim_{n\rightarrow\infty} \tfrac{\lvert \mathcal{F}_{\cdot 0}^{\mathrm{M}}(6,n) \rvert}{\lvert \mathcal{F}^{\mathrm{M}}(6,n) \rvert} =
\lim_{n\rightarrow\infty} \tfrac{\lvert \mathcal{F}_{\cdot 2}^{\mathrm{M}}(6,n) \rvert}{\lvert \mathcal{F}^{\mathrm{M}}(6,n) \rvert} = \tfrac12$ and
$\lim_{n\rightarrow\infty} \tfrac{\lvert \mathcal{F}_{\cdot 4}^{\mathrm{M}}(6,n) \rvert}{\lvert \mathcal{F}^{\mathrm{M}}(6,n) \rvert} = 0$.
\end{enumerate}
\end{theorem}
\begin{proof}
The total numbers of entry specification events mentioned at the beginning of
\ref{comparativecountingtheorem:item:fourentriesspecified}--\ref{comparativecountingtheorem:item:sixentriesspecified}, and all the asymptotic equalities are easily
checked, so we do not have to say more about them.
As to \ref{comparativecountingtheorem:item:uptto3entries}, this follows
immediately from \ref{item:failuresetofisomorphismtypesk3} in Corollary
\ref{cor:isomorphismtypesforwhichequalityofmeasuresofentryspecificationeventsfails}.
As to \ref{comparativecountingtheorem:item:fourentriesspecified}, the claimed
value of $\lvert \mathcal{F}^{\mathrm{M}}(4,n)\rvert$ is true by
\ref{it:partitionoffailuresets:kequals4} in Corollary
\ref{cor:partitionsoffailuresets} combined with Lemma \ref{lem:numberofmatrixcircuitsofgivenlengthwithingivencartesianproduct} and the obvious fact that
$\lvert(\leftsubscript{\mathrm{ul}}{\upX}^{4,n,n})^{-1}\ref{item:comparisonproofcaseLhassize4:4circuit}\rvert =
2^4\cdot \lvert \mathrm{Cir}(4,n)\rvert$. The claimed ratios can be deduced as follows: note
that $\{ \mathfrak{X}\in\im(\leftsubscript{\mathrm{ul}}{\upX}^{4,n,n})\colon\beta_1(\mathfrak{X})\geq 1\} =
\{ \ref{item:comparisonproofcaseLhassize4:4circuit} \}$ by Corollary \ref{cor:isomorphismtypesforwhichequalityofmeasuresofentryspecificationeventsfails}, hence $\lvert\mathcal{F}_{\cdot 0}^{\mathrm{M}}(4,n)\rvert$ $=$ (by \ref{eq:expansionofamatrixfailuresetwithzeroreatiobyisomorphismtypes}) $=$ $\bigl(1-(\tfrac12)^{\beta_1\ref{item:comparisonproofcaseLhassize4:4circuit}}\bigr) \cdot
\lvert (\leftsubscript{\mathrm{ul}}{\upX}^{4,n,n})^{-1}\ref{item:comparisonproofcaseLhassize4:4circuit} \rvert$
$=$ $\tfrac12$ $\cdot$ $\lvert (\leftsubscript{\mathrm{ul}}{\upX}^{4,n,n})^{-1}\ref{item:comparisonproofcaseLhassize4:4circuit} \rvert$ $=$ (by \ref{it:partitionoffailuresets:kequals4} in
Corollary \ref{cor:partitionsoffailuresets}) $=$
$\tfrac12\cdot \lvert\mathcal{F}^{\mathrm{M}}(4,n)\rvert$, which together with
the equation $\mathcal{F}^{\mathrm{M}}(4,n) = \mathcal{F}_{\cdot 0}^{\mathrm{M}}(4,n)
\sqcup \mathcal{F}_{\cdot 2}^{\mathrm{M}}(4,n)$ from
\ref{it:decompositionsofmatrixfailuresets} in Corollary \ref{cor:ratiosandvaluesofpchioandplcffortheisomorphismtypesforwhichequalityfails} proves both
$\lvert\mathcal{F}_{\cdot 0}^{\mathrm{M}}(4,n)\rvert / \lvert\mathcal{F}^{\mathrm{M}}(4,n)\rvert$ $=$ $\tfrac12$ and $\lvert\mathcal{F}_{\cdot 2}^{\mathrm{M}}(4,n)\rvert /
\lvert\mathcal{F}^{\mathrm{M}}(4,n)\rvert$ $=$ $\tfrac12$.
As to \ref{comparativecountingtheorem:item:fiveentriesspecified}, the number
stated first is obvious. The claimed value of
$\lvert \mathcal{F}^{\mathrm{G}}(5,n)\rvert$ can be deduced as follows: by
\ref{relationbeweenchiomeasureandlazycoinflipmeasuregovernedbyfirstbettinumber} in
Theorem \ref{thm:graphtheoreticalcharacterizationofthechiomeasure} we have
$\textup{\textsf{P}}_{\mathrm{chio}}[\mathcal{E}_B] \neq \textup{\textsf{P}}_{\mathrm{lcf}}[\mathcal{E}_B]$ if and only if
$\beta_1(\mathrm{X}_{B}) > 0$. Since due to Definition \ref{def:XBandecXB} we have
$0\leq f_1(\mathrm{X}_{B}) \leq \mathrm{dom}(B) = \lvert I \rvert = 5$, it is
easy to see that $\beta_1(\mathrm{X}_B) \leq 1$. Therefore
$\textup{\textsf{P}}_{\mathrm{chio}}[\mathcal{E}_B] \neq \textup{\textsf{P}}_{\mathrm{lcf}}[\mathcal{E}_B]$ if and only if
$C^4\hookrightarrow \mathrm{X}_{B}$. The latter property is equivalent to the existence
of a matrix-$4$-circuit $S\subseteq I$ with $S\subseteq\mathrm{Supp}(B)$. Note that every
$I\in\binom{[n-1]^2}{5}$ contains at most one matrix-$4$-circuit. Therefore, the
number of all $I\in\binom{[n-1]^2}{5}$ with
$\textup{\textsf{P}}_{\mathrm{chio}}[\mathcal{E}_B] \neq \textup{\textsf{P}}_{\mathrm{lcf}}[\mathcal{E}_B]$ is equal to the
number of all matrix-$4$-circuits $S\in \binom{[n-1]^2}{4}$ with
$S\subseteq\mathrm{Supp}(B)$, multiplied by the number of possibilities to choose an
arbitrary position $u\in [n-1]^2\setminus S$ and an arbitrary $B[u]\in \{0,\pm\}$,
i.e. $2^4\cdot \binom{n-1}{2}^2 \cdot 3 \cdot ((n-1)^2-4)
= 48 \cdot ((n-1)^2-4)\cdot\binom{n-1}{2}^2$. This proves the second claim
in \ref{comparativecountingtheorem:item:fiveentriesspecified}. The claimed ratios
can be deduced as follows: note that
$\{ \mathfrak{X}\in\im(\leftsubscript{\mathrm{ul}}{\upX}^{5,n,n})\colon\beta_1(\mathfrak{X})\geq 1\} =
\{ \ref{item:comparisonproofcaseLhassize6:oneisolatedvertex},
\ref{item:comparisonproofcaseLhassize6:oneadditionaledgeintersectingC4},
\ref{item:comparisonproofcaseLhassize6:twoisolatedvertices},
\ref{item:comparisonproofcaseLhassize6:C4withoneadditionaldisjointedge}
\}$ by Corollary \ref{cor:isomorphismtypesforwhichequalityofmeasuresofentryspecificationeventsfails}, hence $\lvert\mathcal{F}_{\cdot 0}^{\mathrm{M}}(5,n)\rvert$ $=$ (by \ref{eq:expansionofamatrixfailuresetwithzeroreatiobyisomorphismtypes}) $=$
$\bigl(1-(\tfrac12)^{\beta_1\ref{item:comparisonproofcaseLhassize6:oneisolatedvertex}}\bigr) \cdot
\lvert (\leftsubscript{\mathrm{ul}}{\upX}^{5,n,n})^{-1}\ref{item:comparisonproofcaseLhassize6:oneisolatedvertex} \rvert$ $+$
$\bigl(1-(\tfrac12)^{\beta_1\ref{item:comparisonproofcaseLhassize6:oneadditionaledgeintersectingC4}}\bigr) \cdot \lvert (\leftsubscript{\mathrm{ul}}{\upX}^{5,n,n})^{-1}\ref{item:comparisonproofcaseLhassize6:oneadditionaledgeintersectingC4} \rvert$
$+$
$\bigl(1-(\tfrac12)^{\beta_1\ref{item:comparisonproofcaseLhassize6:twoisolatedvertices}}\bigr) \cdot \lvert (\leftsubscript{\mathrm{ul}}{\upX}^{5,n,n})^{-1}\ref{item:comparisonproofcaseLhassize6:twoisolatedvertices} \rvert$ $+$
$\bigl(1-(\tfrac12)^{\beta_1\ref{item:comparisonproofcaseLhassize6:C4withoneadditionaldisjointedge}}\bigr) \cdot \lvert (\leftsubscript{\mathrm{ul}}{\upX}^{5,n,n})^{-1}\ref{item:comparisonproofcaseLhassize6:C4withoneadditionaldisjointedge} \rvert$
$=$
$\tfrac12$ $\bigl ($ $\lvert (\leftsubscript{\mathrm{ul}}{\upX}^{5,n,n})^{-1}\ref{item:comparisonproofcaseLhassize6:oneisolatedvertex} \rvert$ $+$
$\lvert (\leftsubscript{\mathrm{ul}}{\upX}^{5,n,n})^{-1}\ref{item:comparisonproofcaseLhassize6:oneadditionaledgeintersectingC4} \rvert$ $+$
$\lvert (\leftsubscript{\mathrm{ul}}{\upX}^{5,n,n})^{-1}\ref{item:comparisonproofcaseLhassize6:twoisolatedvertices}\rvert$ $+$
$\lvert (\leftsubscript{\mathrm{ul}}{\upX}^{5,n,n})^{-1}\ref{item:comparisonproofcaseLhassize6:C4withoneadditionaldisjointedge} \rvert$ $\bigr)$ $=$ (by \ref{it:partitionoffailuresets:kequals5} in
Corollary \ref{cor:partitionsoffailuresets}) $=$
$\tfrac12\cdot \lvert\mathcal{F}^{\mathrm{M}}(5,n)\rvert$, which together with
the equation $\mathcal{F}^{\mathrm{M}}(5,n) = \mathcal{F}_{\cdot 0}^{\mathrm{M}}(5,n)
\sqcup \mathcal{F}_{\cdot 2}^{\mathrm{M}}(5,n)$ from
\ref{it:decompositionsofmatrixfailuresets} in Corollary \ref{cor:ratiosandvaluesofpchioandplcffortheisomorphismtypesforwhichequalityfails} proves both
$\lvert\mathcal{F}_{\cdot 0}^{\mathrm{M}}(5,n)\rvert / \lvert\mathcal{F}^{\mathrm{M}}(5,n)\rvert$ $=$ $\tfrac12$ and $\lvert\mathcal{F}_{\cdot 2}^{\mathrm{M}}(5,n)\rvert /
\lvert\mathcal{F}^{\mathrm{M}}(5,n)\rvert$ $=$ $\tfrac12$.
As to \ref{comparativecountingtheorem:item:sixentriesspecified}, the claimed
value of $\lvert \mathcal{F}^{\mathrm{G}}(6,n)\rvert$ can be deduced as follows:
using \ref{item:failuresetofisomorphismtypesk6} in Corollary
\ref{cor:isomorphismtypesforwhichequalityofmeasuresofentryspecificationeventsfails},
and inspecting the list of isomorphism types in Lemma
\ref{lem:bipartitenonforestsorderedbytheirfvectors}, we know that equality of the
measures fails if and only $C^4\hookrightarrow \mathrm{X}_B$ or
$C^6\hookrightarrow \mathrm{X}_B$. To count the events for which this is true let us define
$h_{C^6}(n):=\lvert \{ B \in \{0,\pm\}^I\colon
I\in \binom{[n-1]^2}{6},\quad C^6 \hookrightarrow \mathrm{X}_B\} \rvert$,
$h_{K^{2,3}}(n) := \lvert \{ B \in \{0,\pm\}^I\colon
I\in \binom{[n-1]^2}{6},\quad K^{2,3}\hookrightarrow \mathrm{X}_{B}\}\rvert$ and
$h_{C^4, \neg K^{2,3}}(n) := \lvert \{ B \in \{0,\pm\}^I\colon
I\in \binom{[n-1]^2}{6},\quad C^4\hookrightarrow \mathrm{X}_{B},\quad
K^{2,3} \not\hookrightarrow \mathrm{X}_{B}\}\rvert$.
Since for a graph $X$ with at most six edges the properties
$C^6\hookrightarrow X$, $K^{2,3}\hookrightarrow X$, and $C^4\hookrightarrow X$
but $K^{2,3} \not\hookrightarrow X$ are mutually exclusive, it follows that
\begin{equation}
\label{eq:numberofentryspecificationeventswithunequalmeasuressizeLis6}
\lvert \mathcal{F}^{\mathrm{M}}(6,n)\rvert :=
h_{C^6}(n) + h_{C^4,\ \neg K^{2,3}}(n) + h_{K^{2,3}}(n)\quad .
\end{equation}
Lemma \ref{lem:numberofmatrixcircuitsofgivenlengthwithingivencartesianproduct}
implies that $h_{C^6}(n) = 2^6\cdot 3!\cdot \binom{n-1}{3}^2$, the factor of $2^6$
accounting for the fact that the property $C^6\hookrightarrow \mathrm{X}_{B}$ is
indifferent to the choice of the six signs in $B$. Moreover, evidently,
$h_{K^{2,3}}(n) = 2\cdot 2^6 \cdot \binom{n-1}{3} \cdot \binom{n-1}{2}$,
where the first $2$ accounts for the two possibilities
$\lvert \mathrm{p}_1(I) \rvert = 2$ and $\lvert \mathrm{p}_2(I) \rvert = 3$
or
$\lvert \mathrm{p}_1(I) \rvert = 3$ and $\lvert \mathrm{p}_2(I) \rvert = 2$.
In order to compute $h_{C^4,\ \neg K^{2,3}}(n)$, we will employ a simple
inclusion-exclusion-argument: for any of the $\binom{n-1}{2}^2$ choices
$1\leq i_1 < i_2 \leq n-1$ and $1\leq j_1 < j_2 \leq n-1$ for the position of
a matrix-$4$-circuit $S = \{(i_1,j_1), (i_1,j_2), (i_2,j_1), (i_2,j_2)\}$, the number
of distinct $I\subseteq [n-1]^2$ with $\lvert I\rvert =6$ such that $B$ has an
matrix-$4$-circuit \emph{at least} at the positions $(i_1,j_1)$, $(i_1,j_2)$,
$(i_2,j_1)$ and $(i_2,j_2)$ is $ 2^4 \cdot 3^2 \cdot \binom{(n-1)^2-4}{2}$,
where the factor $2^4$ accounts for the mandatory ($\pm$)-values of the $4$-circuit
entries, the factor $3^2$ accounts for the arbitrary $\{0,\pm\}$-values of the
two non-$4$-circuit-entries and the factor $\binom{(n-1)^2-4}{2}$ accounts for the
arbitrary positions of the two non-$4$-circuit-entries in $[n-1]^2\setminus S$.
Summing this expression over all $\binom{n-1}{2}^2$ possible choices of $S$, we
obtain $h_{\geq}(n) := 2^4 \cdot \binom{n-1}{2}^2 \cdot 3^2 \cdot
\binom{(n-1)^2-4}{2}$. But this is not $h_{C^4,\ \neg K^{2,3}}(n)$ yet: from the list in
Lemma \ref{lem:bipartitenonforestsorderedbytheirfvectors} we see that there is
precisely one type which contains more than one copy of $C^4$, namely $K^{2,3}$,
which contains exactly three copies. Therefore, in $h_{\geq}(n)$ every $I$ with
$\mathrm{X}_{B} \cong K^{2,3}$ has been counted exactly three times, and we did not
overcount any of the realizations of the other isomorphism types. Thus, in order to
arrive at $h_{C^4,\ \neg K^{2,3}}(n)$, we have to subtract three times $h_{K^{2,3}}(n)$.
Hence, according to \cite{mathematica6.0},
{\scriptsize
\begin{equation}\label{eq:numberofmatriceswith6elementdomainandassociatedgraphwithC4butnotK23:viaexclusioninclusion}
h_{C^4,\ \neg K^{2,3}}(n)
= h_{\geq}(n) - 3\cdot h_{K^{2,3}}(n) = 18 n^8 - 180 n^7 + 612 n^6 - 608 n^5
- 774 n^4 + 1348 n^3 + 1200 n^2 - 2864 n + 1248.
\end{equation}
}
Substituting this into
\eqref{eq:numberofentryspecificationeventswithunequalmeasuressizeLis6}, one indeed
arrives at the claimed value of $\lvert \mathcal{F}^{\mathrm{M}}(6,n)\rvert$.
As to $\lvert\mathcal{F}_{\cdot 0}^{\mathrm{M}}(6,n)\rvert$, we can use
\ref{eq:expansionofamatrixfailuresetwithapositiveratiobyisomorphismtypes}
in Lemma \ref{lem:forfixedisomorphismtyperatioofbalancedrealizationsdependsonthebettinumberalone} to calculate $\lvert \mathcal{F}_{\cdot 0}^{\mathrm{M}}(6,n)\rvert$ $=$
$\sum_{\mathfrak{X}\in \im(\leftsubscript{\mathrm{ul}}{\upX}^{6,n,n})\colon \beta_1(\mathfrak{X}) = 1}$
$(1-(\tfrac12)^1)$ $\cdot$ $\lvert (\mathrm{X}^{6,n,n})^{-1}(\mathfrak{X}) \rvert$ $+$
$\sum_{\mathfrak{X}\in \im(\leftsubscript{\mathrm{ul}}{\upX}^{6,n,n})\colon\beta_1(\mathfrak{X}) = 2}$
$(1-(\tfrac12)^2)$ $\cdot$ $\lvert (\mathrm{X}^{6,n,n})^{-1}(\mathfrak{X}) \rvert$
$=$ (from the list in Lemma \ref{lem:bipartitenonforestsorderedbytheirfvectors}) $=$
$\tfrac12$ $\cdot$ $\sum_{\mathfrak{X}\in \{\ref{item:comparisonproofcaseLhassize6:oneisolatedvertex},\dotsc,\ref{item:comparisonproofcaseLhassize6:twoadditionaldisjointedgesdisjointfromC4}\}\setminus\{\ref{item:comparisonproofcaseLhassize6:XBLisomorphictoK23}\}}$ $\lvert (\mathrm{X}^{6,n,n})^{-1}(\mathfrak{X}) \rvert$ $+$ $\tfrac34$ $\cdot$ $\lvert (\mathrm{X}^{6,n,n})^{-1}\ref{item:comparisonproofcaseLhassize6:XBLisomorphictoK23} \rvert$ $=$ $\tfrac12 \cdot(h_{C^4,\, \neg K^{2,3}}(n) +
h_{C^6}(n)) + \tfrac34\cdot h_{K^{2,3}}(n)$, and it can be checked that this equals
the claimed value of $\lvert \mathcal{F}_{\cdot 0}^{\mathrm{M}}(6,n)\rvert$.
Similarly, $\lvert \mathcal{F}_{\cdot 2}^{\mathrm{M}}(6,n)\rvert$ $=$
$\sum_{\mathfrak{X}\in \im(\leftsubscript{\mathrm{ul}}{\upX}^{6,n,n})\colon \beta_1(\mathfrak{X}) = 1}$ $(\tfrac12)^1$ $\cdot$
$\lvert (\mathrm{X}^{6,n,n})^{-1}(\mathfrak{X}) \rvert$ $=$ (from the list in
Lemma \ref{lem:bipartitenonforestsorderedbytheirfvectors}) $=$
$(\tfrac12)$ $\cdot$ $\sum_{\mathfrak{X}\in \{\ref{item:comparisonproofcaseLhassize6:oneisolatedvertex},\dotsc,\ref{item:comparisonproofcaseLhassize6:twoadditionaldisjointedgesdisjointfromC4}\}\setminus\{\ref{item:comparisonproofcaseLhassize6:XBLisomorphictoK23}\}}$ $\lvert (\mathrm{X}^{6,n,n})^{-1}(\mathfrak{X}) \rvert$ $=$ $\tfrac12$ $($ $h_{C^4,\, \neg K^{2,3}}(n)$ $+$ $h_{C^6}(n)$ $)$,
and it can be checked that this is equal to the value of
$\lvert \mathcal{F}_{\cdot 2}^{\mathrm{M}}(6,n)\rvert$ which is claimed in
\ref{comparativecountingtheorem:item:sixentriesspecified}. Finally,
$\lvert\mathcal{F}_{\cdot 4}^{\mathrm{M}}(6,n)\rvert$ $=$
$\sum_{\mathfrak{X}\in \im(\leftsubscript{\mathrm{ul}}{\upX}^{6,n,n})\colon\beta_1(\mathfrak{X})=2)}$ $(\tfrac12)^2$ $\cdot$
$\lvert (\mathrm{X}^{6,n,n})^{-1}(\mathfrak{X}) \rvert$ $=$ (from the list in
Lemma \ref{lem:bipartitenonforestsorderedbytheirfvectors}) $=$
$(\tfrac14)$ $\cdot$ $\lvert (\mathrm{X}^{6,n,n})^{-1}\ref{item:comparisonproofcaseLhassize6:XBLisomorphictoK23} \rvert$ $=$ $\tfrac14\cdot h_{K^{2,3}}$, and this equals the
value of $ \lvert \mathcal{F}_{\cdot 4}^{\mathrm{M}}(6,n)\rvert$ claimed in
\ref{comparativecountingtheorem:item:sixentriesspecified}.
\end{proof}
\subsubsection{Alternative checks using Theorem \ref{thm:countingmatrixrealizations}}\label{subsubsec:alternativechecks}
We did not need Theorem \ref{thm:countingmatrixrealizations} in our proof of
Theorem \ref{thm:comparativecountingtheorem}. It can, nevertheless, provide
additional security since via Corollary \ref{cor:partitionsoffailuresets} and
Theorem \ref{thm:countingmatrixrealizations} one may take an
inclusion-exclusion-free (but, all told, much more laborious) alternative route
to the claimed values of $\lvert \mathcal{F}^{\mathrm{M}}(5,n)\rvert$ and
$\lvert \mathcal{F}^{\mathrm{M}}(6,n)\rvert$. As to the claimed value
of $\lvert\mathcal{F}^{\mathrm{G}}(5,n)\rvert$, by
\ref{it:partitionoffailuresets:kequals5} in Corollary
\ref{cor:partitionsoffailuresets} combined with Theorem
\ref{thm:countingmatrixrealizations} we have
$\lvert\mathcal{F}^{\mathrm{G}}(5,n)\rvert$ $=$
\ref{numberofmatrixrealizationsoftype:oneisolatedvertex:kequals5} $+$
\ref{numberofmatrixrealizationsoftype:oneadditionaledgeintersectingC4:kequals5}
$+$ \ref{numberofmatrixrealizationsoftype:twoisolatedvertices:kequals5} $+$
\ref{numberofmatrixrealizationsoftype:C4withoneadditionaldisjointedge:kequals5}
$=$ $2^4\cdot \binom{n-1}{2}^2\cdot
\bigl ( 4\cdot (n-3) + 8 \cdot (n-3) + 1\cdot (n-3)^2 + 2\cdot (n-3)^2\bigr)$ $=$
$48\cdot ((n-1)^2 - 4)\cdot \binom{n-1}{2} \cdot \binom{n-1}{2}$.
As to the claimed value of $\lvert\mathcal{F}^{\mathrm{G}}(6,n)\rvert$, by
\ref{it:partitionoffailuresets:kequals6} in Corollary \ref{cor:partitionsoffailuresets}, the function $\lvert\mathcal{F}^{\mathrm{G}}(6,n)\rvert$ equals the sum of the
nineteen functions which were found in \ref{numberofmatrixrealizationsoftype:oneisolatedvertex}--\ref{numberofmatrixrealizationsoftype:twoadditionaldisjointedgesdisjointfromC4} of \ref{item:numberofrealizationsofnonforestswhenkequals6} in Theorem \ref{thm:countingmatrixrealizations}, and
one can check (e.g. with \cite{mathematica6.0}) that indeed
$\lvert\mathcal{F}^{\mathrm{G}}(6,n)\rvert =
\sum_{2\leq k \leq 20} (\mathrm{m}6.\mathrm{t}k) =
18n^8 - 180n^7 + \tfrac{1868}{3}n^6 - \tfrac{2176}{3}n^5 - \tfrac{754}{3}n^4 +
\tfrac{428}{3}n^3 - \tfrac{8144}{3}n^2 - \tfrac{11536}{3}n + 1504$.
Incidentally, let us note that by summing all functions in
\ref{numberofmatrixrealizationsoftype:oneisolatedvertex}--\ref{numberofmatrixrealizationsoftype:twoadditionaldisjointedgesdisjointfromC4} except
\ref{numberofmatrixrealizationsoftype:XBLisomorphictoK23} and
\ref{numberofmatrixrealizationsoftype:threeisolatedvertices}
(i.e. by summing seventeen functions) one may also check the equation
$h_{C^4,\ \neg K^{2,3}}(n) = h_{\geq}(n) - 3\cdot h_{K^{2,3}}(n) =
18 n^8 - 180 n^7 + 612 n^6 - 608 n^5 - 774 n^4 + 1348 n^3 + 1200 n^2 - 2864 n + 1248$
claimed in the proof above.
\subsubsection{Quantitatively dominant graph-theoretical reasons for
$\textup{\textsf{P}}_{\mathrm{chio}}\neq\textup{\textsf{P}}_{\mathrm{lcf}}$}
Let us note that Theorem \ref{thm:countingmatrixrealizations}
tells us that of the $\lvert \mathcal{F}^{\mathrm{M}}(6,n) \rvert\in
\Omega_{n\to\infty} (n^8)$ six-element-entry-specifications which cause non-agreement
of $\textup{\textsf{P}}_{\mathrm{chio}}$ and $\textup{\textsf{P}}_{\mathrm{lcf}}$, most of the failures are concentrated at only
three out of the nineteen isomorphism types in
\ref{item:numberofrealizationsofnonforestswhenkequals6}: only the types \ref{item:comparisonproofcaseLhassize6:fourisolatedvertices}, \ref{item:comparisonproofcaseLhassize6:oneadditionaldisjointedgeandtwoisolatedvertices} and
\ref{item:comparisonproofcaseLhassize6:twoadditionaldisjointedgesdisjointfromC4}
have a preimage under $\leftsubscript{\mathrm{ul}}{\upX}^{6,n,n}$ which is of size $\Omega_{n\to\infty}(n^8)$.
The quantitative domination of these isomorphism types is, however, a rather slow
one in that
\[ \tfrac{\sum_{2\leq k \leq 17} \lvert(\leftsubscript{\mathrm{ul}}{\upX}^{6,n,n})^{-1}(\mathrm{t}k)\rvert}{\lvert(\leftsubscript{\mathrm{ul}}{\upX}^{6,n,n})^{-1}\ref{item:comparisonproofcaseLhassize6:fourisolatedvertices}\rvert + \lvert(\leftsubscript{\mathrm{ul}}{\upX}^{6,n,n})^{-1}\ref{item:comparisonproofcaseLhassize6:oneadditionaldisjointedgeandtwoisolatedvertices}\rvert + \lvert(\leftsubscript{\mathrm{ul}}{\upX}^{6,n,n})^{-1}\ref{item:comparisonproofcaseLhassize6:twoadditionaldisjointedgesdisjointfromC4}\rvert} \in \Theta(n^{-1})\quad .\]
\subsubsection{Estimate of the number of failures of equality of $\textup{\textsf{P}}_{\mathrm{chio}}$ and
$\textup{\textsf{P}}_{\mathrm{lcf}}$ for events $\mathcal{E}_B$ with $B\in\{0,\pm\}^I$ and
$I\in\binom{[n-1]^2}{k}$ and $k$ general}
\begin{proposition}[for fixed $k$ the measures
$\textup{\textsf{P}}_{\mathrm{chio}}$ and $\textup{\textsf{P}}_{\mathrm{lcf}}$ agree for almost all entry-specifications]
\label{prop:roughestimationofnumberoffailureevents}
For every fixed $k\geq 1$ we have
$\lvert \mathcal{F}^{\mathrm{M}}(k,n)\rvert$ $/$ $\lvert \{ B\in \{0,\pm\}^I,\;
I\in\binom{[n-1]^2}{k}\}\rvert$ $\in$
$\mathcal{O}_{n\to\infty}(n^{-2})$\quad .
\end{proposition}
\begin{proof}
We will estimate numerator and denominator of this fraction separately. The
denominator is equal to $3^k\cdot \binom{(n-1)^2}{k} \in \Omega_{n\to\infty}(n^{2k})$.
Moreover, a very rough estimate suffices to obtain a bound on the numerator which
nevertheless is sufficiently small to prove that the ratio vanishes:
$\lvert \mathcal{F}^{\mathrm{M}}(k,n) \rvert$
$\By{Theorem \ref{thm:graphtheoreticalcharacterizationofthechiomeasure}.\ref{relationbeweenchiomeasureandlazycoinflipmeasuregovernedbyfirstbettinumber}}{=}$
$\lvert \{$ $B$ $\in$ $\{0,\pm\}^I\colon$
$I\in\binom{[n-1]^2}{k}$, $B$ contains a matrix-circuit $\} \rvert$
$=$ $\bigl \lvert$ $\bigcup_{1\leq j \leq \lfloor \tfrac{k}{2} \rfloor}$
$\bigcup_{L\in\mathrm{Cir}(2j,n)}$ $\{$ $B$ $\in$ $\{0,\pm\}^I\colon$ $I\in\binom{[n-1]^2}{k}$,
$L\subseteq \mathrm{Supp}(B)$ $\}$ $\bigr \rvert$ $\leq$
$\sum_{1\leq j \leq \lfloor \tfrac{k}{2} \rfloor}$ $\sum_{L\in\mathrm{Cir}(2j,n)}$
$\lvert \{$ $B$ $\in$ $\{0,\pm\}^I\colon$ $I$ $\in$ $\binom{[n-1]^2}{k}$,
$L$ $\subseteq$ $\mathrm{Supp}(B)$ $\} \rvert$ $=$ $\sum_{1\leq j \leq \lfloor \tfrac{k}{2} \rfloor}$
$2^{2j}$ $\cdot$ $3^{k-2j}$ $\cdot$ $\binom{(n-1)^2-2j}{k-2j}$ $\cdot$
$\lvert \mathrm{Cir}(2j,n) \rvert$ $\By{Lemma \ref{lem:numberofmatrixcircuitsofgivenlengthwithingivencartesianproduct}}{=}$ $\sum_{1\leq j \leq \lfloor \tfrac{k}{2} \rfloor}$ $2^{2j}$
$\cdot$ $3^{k-2j}$ $\cdot$ $\binom{(n-1)^2-2j}{k-2j}$ $\cdot$ $\binom{n-1}{j}^2$
$\cdot$ $\frac{j!(j-1)!}{2}$ $\in$ $\sum_{1\leq j \leq \lfloor \tfrac{k}{2} \rfloor}$
$\mathcal{O}_{n\to\infty}(1)$ $\cdot$ $\mathcal{O}_{n\to\infty}(n^{2k-4j})$ $\cdot$
$\mathcal{O}_{n\to\infty}(n^{2j})$ $\cdot$ $\mathcal{O}_{n\to\infty}(1)$ $\subseteq$
$\sum_{1\leq j \leq \lfloor \tfrac{k}{2} \rfloor}$ $\mathcal{O}_{n\to\infty}(n^{2k-2j})$
$\subseteq$ $\mathcal{O}_{n\to\infty}(n^{2k-2})$.
\end{proof}
\section{Connection to counting singular $\{\pm\}$-matrices}
\subsection{Basic connections}\label{subsec:basicconnections}
The Lemmas \ref{lem:chiocondensationaffectsrankintheleastpossibleway} and
\ref{lem:uniformmeasureofranksetaschiomeasureofranksetoneranklower}, which are
consequences of Chio's identity \ref{lem:chioidentity}, are the basic reason why
the measure $\textup{\textsf{P}}_{\mathrm{chio}}$ is relevant for the study of singular $\pm$-matrices.
\begin{lemma}[Chio condensation affects rank to the least possible degree]
\label{lem:chiocondensationaffectsrankintheleastpossibleway}
For every integral domain $R$, every $(s,t)\in \mathbb{Z}_{\geq 2}^2$ and
every $A\in R^{[s]\times [t]}$ with $a_{s,t}\neq 0$ we have
$\mathrm{rk}(\tfrac12\mathrm{C}_{(s,t)}(A)) = \mathrm{rk}(A)-1$.
\end{lemma}
\begin{proof}
If $\mathrm{rk}(A) = 1$, then obviously $\mathrm{C}_{(s,t)}(A) = \{0\}^{[s-1]\times [t-1]}$ and the
claim is true. We may therefore assume that $r:=\mathrm{rk}(A)\geq 2$. By the equality
of rank and determinantal rank over integral domains
(cf. e.g. \cite[Corollary 2.29(2)]{MR1181420}) there
exists $S\in\binom{[s]}{r}$ and $T\in \binom{[t]}{r}$ such that
$\det(A\mid_{S\times T}) \neq 0$. If $s\notin S$, then by temporarily passing to the
field of fractions of $R$ we may appeal to Steinitz' exchange lemma for vector spaces
to prove the existence of at least one $i_0\in S$ such that
$\det\bigl ( A\mid_{((S\setminus \{i_0\})\sqcup \{s\})\times T} \bigr ) \neq 0$.
Analogously for $t\notin T$. Therefore we may assume that $s\in S$ and $t\in T$. Hence
$S\times T = ((S\setminus\{s\})\times (T\setminus\{t\}))^{\breve{}}$ and therefore
$\mathrm{C}_{(s,t)}(A\mid_{S\times T})$ is defined. By Lemma \ref{lem:chioidentity} we know
that $\det(\mathrm{C}_{(s,t)}(A\mid_{S\times T})) =
a_{s,t}^{r-2} \cdot \det(A\mid_{S\times T}) \neq 0$, the latter since $R$ is an
integral domain and $a_{s,t}\neq 0$ by assumption. Since $\mathrm{C}_{(s,t)}(A\mid_{S\times T}) =
\mathrm{C}_{(s,t)}(A)\mid_{(S\setminus\{s\})\times(T\setminus\{t\})} \in R^{(r-1)\times (r-1)}$, and
by the equality of rank and determinantal rank, this implies
$\mathrm{rk}(\mathrm{C}_{(s,t)}(A)) \geq r-1$. On the other hand we also have
$\mathrm{rk}(\mathrm{C}_{(s,t)}(A)) \leq r-1$. To see this, it suffices to note that every
$r\times r$ submatrix of $\mathrm{C}_{(s,t)}(A)$ is the Chio condensate of an
$(r+1)\times(r+1)$ submatrix of $A$, hence by Chio's identity a
nonvanishing $r\times r$ minor of $\mathrm{C}_{(s,t)}(A)$ would imply a nonvanishing
$(r+1)\times (r+1)$ minor of $A$, contrary to the assumption of $\mathrm{rk}(A) = r$.
\end{proof}
\begin{lemma}\label{lem:uniformmeasureofranksetaschiomeasureofranksetoneranklower}
$\textup{\textsf{P}}\bigl [ \mathrm{Ra}_r(\{\pm\}^{[s]\times[t]})\bigr ] =
\textup{\textsf{P}}_{\mathrm{chio}} \bigl [ \mathrm{Ra}_{r-1}(\{0,\pm\}^{[s-1]\times[t-1]})\bigr ]$ for every
$(s,t)\in\mathbb{Z}_{\geq 2}^2$ and $1\leq r \leq \min(s,t)$.
\end{lemma}
\begin{proof}
This follows from the calculation
$\textup{\textsf{P}}$ $\bigl [$ $\mathrm{Ra}_r($ $\{\pm\}^{[s]\times[t]})$ $\bigr ]$ $=$ $\frac{1}{2^{s\cdot t}}$
$\lvert\{$ $A$ $\in$ $\{\pm\}^{[s]\times [t]}\colon$ $\mathrm{rk}(A)$ $=$ $r$ $\}\rvert$
$\By{Lemma \eqref{lem:chiocondensationaffectsrankintheleastpossibleway}}{=}$
$\frac{1}{2^{s\cdot t}}$ $\sum_{B\in\{0,\pm\}^{[s-1]\times[t-1]}\colon\mathrm{rk}(B) = r-1}$ $\lvert ($
$\tfrac12\mathrm{C}_{(s,t)}$ $)^{-1}$ $(B)$ $\rvert$
$\By{Definition \ref{eq:definitionmeasurePchio}}{=}$
$\textup{\textsf{P}}_{\mathrm{chio}}$ $[$ $\mathrm{Ra}_{r-1}($ $\{$ $0,$ $\pm$ $\}^{[s-1]\times [t-1]}$ $)$ $]$. Note,
incidentally, that with the third equality sign, many zero-summands are introduced.
\end{proof}
\begin{corollary}\label{cor:uniformmeasureofranksublevelsetaschiomeasureofranksublevelsetoneranklower}
$\textup{\textsf{P}}\bigl [ \mathrm{Ra}_{\mathcal{R}}(\{\pm\}^{[s]\times[t]})\bigr ] =
\textup{\textsf{P}}_{\mathrm{chio}} \bigl [ \mathrm{Ra}_{\mathcal{R}}(\{0,\pm\}^{[s-1]\times[t-1]})\bigr ]$ for every
$(s,t)\in\mathbb{Z}_{\geq 2}^2$ and every $\mathcal{R}\in\mathfrak{P}([\min(s,t)]\sqcup\{0\})$.
\end{corollary}
\begin{proof}
Immediate from Lemma \ref{lem:uniformmeasureofranksetaschiomeasureofranksetoneranklower} and Definition \ref{def:ranksets}.
\end{proof}
\begin{corollary}\label{cor:implicationofbourgainvuwoodtheoremonchiomeasure}
$\textup{\textsf{P}}_{\mathrm{chio}} [ \mathrm{Ra}_{<n-1}(\{0,\pm\}^{[n-1]^2}) ] \leq (1/\sqrt{2} + o(1))^n$
for $n\rightarrow \infty$.
\end{corollary}
\begin{proof}
By combining Lemma \ref{cor:uniformmeasureofranksublevelsetaschiomeasureofranksublevelsetoneranklower} with \eqref{thm:bourgainvuwood:uniformdistribution} in
Theorem \ref{thm:bourgainvuwood}.
\end{proof}
\subsection{Sign functions which are both singular and balanced}
\label{subsec:countingsingularandcyclicallyrevenedgecolourings}
\begin{definition}[$G_{s,t}$]
For every $(s,t)\in\mathbb{Z}_{\geq 2}^2$ define $G_{s,t} := \bigl(\bigoplus_{1\leq i \leq s-1}\mathbb{Z}/2\bigr) \oplus \bigl(\bigoplus_{1\leq j\leq t-1}\mathbb{Z}/2\bigr)$.
\end{definition}
We will use the following group actions. Informally, $\alpha_X$ is the action by
switching signs of edges simultaneously in all `stars' centered at those vertices
for which $g$ has nonzero components.
\begin{definition}\label{def:thetwoswitchinggroupactions}
For every $(s,t)\in\mathbb{Z}_{\geq 2}^2$ define the group action
$\alpha_{s,t}\colon G_{s,t} \longrightarrow \mathrm{Sym}(\{0,\pm\}^{[s-1]\times[t-1]})$,
$((g_i)_{i\in [s-1]} , (g_j)_{j\in [t-1]} ) \longmapsto
\binom{\{0,\pm\}^{[s-1]\times [t-1]}\to\{0,\pm\}^{[s-1]\times [t-1]}}{(b_{i,j})_{(i,j)\in {[s-1]\times [t-1]}} \mapsto (-1)^{g_i}\cdot (-1)^{g_j} \cdot b_{i,j}}$. For every $X\in \mathrm{BG}_{s,t}$
define the group action $\alpha_X\colon G_{s,t} \longrightarrow
\mathrm{Sym}( \{\pm\}^{\mathrm{E}(X)})$ which is defined by $(\alpha_X(g)(\sigma))(e) :=
(-1)^{g_i}\cdot (-1)^{g_j} \cdot \sigma(e)$ for every $\sigma\in\{\pm\}^{\mathrm{E}(X)}$
and every $e = \{ (i,t), (s,j)\} \in \mathrm{E}(X)$.
\end{definition}
Note that neither $\alpha_{s,t}$ nor $\alpha_X$ are faithful group actions. More
precisely, not only is both $\ker(\alpha_{s,t})$ and $\ker(\alpha_X)$ a
$2$-element set, but $\alpha_{s,t}$ and $\alpha_X$ are both double-covers onto their
images. We could construct a faithful action by making an arbitrary choice of
a $\iota\in [s-1]\cup [t-1]$ and then refraining from switching at this index
(analogously, by making an arbitrary choice of a star in $X$ and then refraining
from switching that particular star).
Balancedness is a very `rigid' property of an edge-signing in that it is
determined by the signing of an arbitrary spanning tree:
\begin{lemma}[rigidity of balanced edge signings]\label{lem:setofbalancedsigningsdeterminedbysigningsofarbitraryspanningtree}
For every connected graph $X$ and every spanning tree $T$ of $X$, there is a
bijection $\{\pm\}^{\mathrm{E}(T)} \leftrightarrow
\{ \sigma\in \{\pm\}^{\mathrm{E}(X)}\colon \text{$(X,\sigma)$ balanced}\}$.
\end{lemma}
\begin{proof}[Sketch of proof]
Since the balance-preserving sign of every edge $e\in\mathrm{E}(X)\setminus\mathrm{E}(T)$ is
determined by the unique circuit in $\mathrm{E}(T)\cup\{e\}$, for every given
$\sigma\in \{\pm\}^{\mathrm{E}(T)}$, there is at most one balanced extension of $\sigma$,
i.e. at most one $\tilde{\sigma}\in \{\pm\}^{\mathrm{E}(X)}$ with
$\sigma = \tilde{\sigma}\mid_{\mathrm{E}(T)}$ and $(X,\tilde{\sigma})$ balanced. Moreover,
this extension can be constructed in the obvious `greedy' way by
successively adding in the elements of
$\mathrm{E}(X)\setminus\mathrm{E}(T)$ in an arbitrary order while at each step of the
construction choosing the sign of the added edge so as to avoid non-balanced
circuits. That this is indeed possible can be proved by an induction on the
number $\lvert\mathrm{E}(X)\setminus\mathrm{E}(T)\rvert$ of edges to be added. A key
observation (routine to prove and known since at least \cite[Theorem 2]{MR0067468})
is that at each step of the construction, for each pair of vertices
either \emph{all} paths within the partially constructed graphs
which have these two vertices as endvertices have sign $(-)$ or \emph{all} such
paths have sign $(+)$, and therefore the greedy construction never stalls.
\end{proof}
\begin{lemma}\label{lem:starswitchingistransitiveonbalancedsignings}
For every graph $X$ the restriction $\im(\alpha_X)\mid_{S_{\mathrm{bal}}(X)}$ is a
transitive permutation group on $S_{\mathrm{bal}}(X)$.
\end{lemma}
\begin{proof}[Sketch of proof]
One way to look at this is as `making use of the rigidity of balanced signings': we
can choose an arbitrary spanning tree $T_i$ for each connected component $X_i$
of $X$, then show that $\im(\alpha_X)\mid_{S_{\mathrm{bal}}(X)}$ is transitive on the set
$\{\pm\}^{\mathrm{E}(T_i)}$ of all edge-signings of $T_i$, and then appeal to
Lemma \ref{lem:setofbalancedsigningsdeterminedbysigningsofarbitraryspanningtree}
which says that this transitivity already implies transitivity on the full
set $S_{\mathrm{bal}}(X)$.
\end{proof}
Given a $\{0,1\}$-matrix, it can be possible to increase its $\mathbb{Z}$-rank by choosing
signs for the entries. If we require the signed matrix to be \emph{balanced},
however, the rank must stay the same. This follows quickly from the
graph-theoretical considerations above:
\begin{proposition}[all balanced signings of a $\{0,1\}$-matrix have the
same rank]\label{prop:allbalancedsigningshavethesamerank}
Let $B\in \{0,1\}^{[s-1]\times [t-1]}$. Let $\tilde{B}$ be an arbitrary
`balanced signing of $B$', i.e. $\tilde{B} \in\{0,\pm\}^{[s-1]\times [t-1]}$,
$\mathrm{Supp}(\tilde{B}) = \mathrm{Supp}(B)$ and
$(\mathrm{X}_{\tilde{B}},\sigma_{\tilde{B}}) = (\mathrm{X}_{B},\sigma_{\tilde{B}})$ is a balanced
signed graph. Then $\mathrm{rk}(\tilde{B}) = \mathrm{rk}(B)$.
\end{proposition}
\begin{proof}
Since both $(\mathrm{X}_B,\sigma_B)$ and $(\mathrm{X}_{\tilde{B}},\sigma_{\tilde{B}})$ are balanced,
by Lemma \ref{lem:starswitchingistransitiveonbalancedsignings} there exists
$g\in G_{s,t}$ such that $\alpha_X(g)(\sigma_{\tilde{B}}) = \sigma_B$. In view of
Definition \ref{def:thetwoswitchinggroupactions} and
Definition \ref{def:XBandecXB}, this implies $\alpha_{s,t}(g)(\tilde{B}) = B$.
Since $\alpha_{s,t}$ obviously keeps the rank invariant, the claim is proved.
\end{proof}
We will now use the knowledge established so far to analyse the
tempting `absolute' route of using Corollary \ref{cor:uniformmeasureofranksublevelsetaschiomeasureofranksublevelsetoneranklower} and then partitioning according to isomorphism type of the associated bipartite graph. The conclusion is that this will lead us
onto a well-beaten path (counting singular $\{0,1\}$-matrices):
\begin{proposition}[on rank-level-sets, the Chio measure agrees with the uniform
measure after forgetting the signs]\label{prop:chiomeasureasuniformmeasureonchiosets}
Let $(s,t)\in\mathbb{Z}_{\geq 2}^2$ and $\mathcal{R}\in\mathfrak{P}(\{1,\dotsc,\min(s,t)\})$.
Then
\begin{equation}
\textup{\textsf{P}}_{\mathrm{chio}}[\mathrm{Ra}_{\mathcal{R}}(\{0,\pm\}^{[s-1]\times [t-1]})] =
\textup{\textsf{P}}[\mathrm{Ra}_{\mathcal{R}}(\{0,1\}^{[s-1]\times [t-1]})] \quad .
\end{equation}
\end{proposition}
\begin{proof}
This follows from the calculation
\begin{align}
\textup{\textsf{P}}_{\mathrm{chio}}[\mathrm{Ra}_{\mathcal{R}}(\{0,\pm\}^{[s-1]\times [t-1]})]
& \By{\text{\tiny (\ref{characterizationofwhenchiomeasureispositive} in
Theorem \ref{thm:graphtheoreticalcharacterizationofthechiomeasure})}}{=}
\textup{\textsf{P}}_{\mathrm{chio}} \bigl [\bigl \{ B\in \{0,\pm\}^{[s-1]\times [t-1]}\colon
\parbox{0.15\linewidth}{\tiny $\mathrm{rk}(B) \in \mathcal{R}$,
$(\mathrm{X}_B,\sigma_B)$ balanced} \bigr \} \bigr ] \notag \\
& = \sum_{\mathfrak{X}\in\mathrm{ul}(\mathrm{BG}_{s,t})} \textup{\textsf{P}}_{\mathrm{chio}} \bigl [\bigl
\{ B\in \{0,\pm\}^{[s-1]\times [t-1]}\colon
\parbox{0.16\linewidth}{\tiny $\mathrm{rk}(B) \in \mathcal{R}$,
$\mathrm{X}_B\cong \mathfrak{X}$, \\ $(\mathrm{X}_B,\sigma_B)$ balanced} \bigr \} \bigr ] \notag \\
\text{\tiny (by \ref{cor:chiomeasureofasinglematrix} in Corollary \ref{cor:quickconsequencesofthecharacterizations})}
& = \sum_{\mathfrak{X}\in\mathrm{ul}(\mathrm{BG}_{s,t})} 2^{-st + \beta_0(\mathfrak{X}) + 1} \cdot
\lvert \{ \text{ \tiny $B\in \{0,\pm\}^{[s-1]\times [t-1]}\colon$ }
\parbox{0.16\linewidth}{\tiny $\mathrm{rk}(B)\in\mathcal{R}$, $\mathrm{X}_B\cong \mathfrak{X}$, \\
$(\mathrm{X}_B,\sigma_B)$ balanced} \bigr \} \rvert \notag \\
\parbox{0.2\linewidth}{\tiny (by \ref{item:numberofbalancedsignfunctions} in
Lemma \ref{lem:equivalenceofexistenceofbconstantrpropervertex2coloringandcyclicallyrevenness} \\ and Proposition \ref{prop:allbalancedsigningshavethesamerank})} & =
\sum_{\mathfrak{X}\in\mathrm{ul}(\mathrm{BG}_{s,t})} 2^{f_0(\mathfrak{X})-st + 1} \cdot \lvert
\{ \text{ \scriptsize $B\in \{0,1\}^{[s-1]\times [t-1]}\colon
\mathrm{rk}(B)\in\mathcal{R},\; \mathrm{X}_B\cong \mathfrak{X}$ }\} \rvert \notag \\
\parbox{0.2\linewidth}{\tiny ( since $f_0(\mathfrak{X})$ is equal \\
to $(s-1)+(t-1)$ for \\
every $\mathfrak{X}\in\mathrm{ul}(\mathrm{BG}_{s,t})$ )} & = \sum_{\mathfrak{X}\in\mathrm{ul}(\mathrm{BG}_{s,t})}
2^{-(s-1)(t-1)} \cdot \lvert \{ \text{ \scriptsize $B\in \{0,1\}^{[s-1]\times [t-1]}\colon
\mathrm{rk}(B)\in\mathcal{R},\; \mathrm{X}_B\cong \mathfrak{X}$ }\} \rvert \notag \\
& = (\tfrac12)^{(s-1)(t-1)}\cdot \lvert \{ B\in \{0,1\}^{[s-1]\times[t-1]}\colon
\mathrm{rk}(B)\in\mathcal{R} \}\rvert \notag \\
& = \textup{\textsf{P}}[\mathrm{Ra}_{\mathcal{R}}(\{0,1\}^{[s-1]\times[t-1]})]\quad .\notag
\end{align}
The proof of Proposition \ref{prop:chiomeasureasuniformmeasureonchiosets} is now
complete.
\end{proof}
It should be noted that the special case
$\textup{\textsf{P}}[\mathrm{Ra}_{<n}(\{\pm\}^{[n]^2})] = \textup{\textsf{P}}[\mathrm{Ra}_{<n-1}(\{0,1\}^{[n-1]^2})]$ seems well-known
(the author does not have an explicit reference corroborating this, but there are
articles in which this is implicit (e.g. \cite{MR2216479}).
\subsection{A relative point of view}
What really appears to promise progress on Conjecture \ref{conj:socalledfolkloreconjecture} is to use the theorem of Bourgain--Vu--Wood to reach a more relative
vantage point:
\begin{proposition}[relative formulations of Conjecture \ref{conj:socalledfolkloreconjecture}]\label{prop:equivalentnewformulationofoldconjecture}
The following statements are equivalent:
{\scriptsize
\begin{enumerate}[label={\rm(Q\arabic{*})}]
\item\label{lem:equivalentproblem:it:folkloreconjecture}
$\textup{\textsf{P}}\bigl [ \mathrm{Ra}_{<n}(\{\pm\}^{[n]^2}) \bigr ]
\leq (\frac12+o_{n\to\infty}(1))^n$
\item\label{eq:equivalentformulationoftheconjecturewithabbreviations}
$\textup{\textsf{P}}_{\mathrm{chio}}\bigl [ \mathrm{Ra}_{<n-1} ( \{0,\pm\}^{[n-1]^2} )\bigr ]
\leq (\frac12+o_{n\to\infty}(1)) \cdot \textup{\textsf{P}}_{\mathrm{lcf}} \bigl [ \mathrm{Ra}_{<n-1}( \{0,\pm\}^{[n-1]^2} )\bigr ]$
\item\label{conj:relativizedconjecture}
$\sum_{ B'\ \in \ \mathrm{Ra}_{<n-1}( \{0,\pm\}^{[n-1]^2} ) } \textup{\textsf{P}}_{\mathrm{chio}}[B'] \leq
\bigl(\frac12+o_{n\to\infty}(1)\bigr) \cdot \sum_{ B''\ \in \ \mathrm{Ra}_{<n-1}( \{0,\pm\}^{[n-1]^2} ) }
\textup{\textsf{P}}_{\mathrm{lcf}} \bigl [ B'' \bigr ]$
\item\label{arelativizedformulation}
$\lvert \{ B'\in\{0,1\}^{[n-1]^2}\colon \mathrm{rk}(B') < n-1 \} \rvert \leq
(\tfrac12 + o_{n\to\infty}(1)) \cdot \sum_{B\in\{0,1\}^{[n-1]^2}} (\tfrac12)^{\mathrm{supp}(B)} \cdot
\Biggl \lvert \Biggl \{ \parbox{0.2\linewidth}{$B''\in\{0,\pm\}^{[n-1]^2}\colon$ \\
$\mathrm{Supp}(B'')=\mathrm{Supp}(B)$, $\mathrm{rk}(B'') < n-1$ } \Biggr \} \Biggr \rvert$
\end{enumerate}
}
\end{proposition}
\begin{proof}
As to the equivalence
\ref{lem:equivalentproblem:it:folkloreconjecture} $\Leftrightarrow$
\ref{eq:equivalentformulationoftheconjecturewithabbreviations},
if \ref{lem:equivalentproblem:it:folkloreconjecture}, then
$\textup{\textsf{P}}_{\mathrm{chio}}\bigl [ \mathrm{Ra}_{<n-1} ( \{0,\pm\}^{[n-1]^2} )\bigr ]$
$\By{Corollary \eqref{cor:uniformmeasureofranksublevelsetaschiomeasureofranksublevelsetoneranklower}}{=}$
$\textup{\textsf{P}} \bigl [ \mathrm{Ra}_{<n} ( \{\pm\}^{[n]^2} )\bigr ]$
$\By{\ref{lem:equivalentproblem:it:folkloreconjecture}}{\leq} $
$(\frac12 + o_{n\to\infty}(1))^n$ $=$ $(\frac12 + o_{n\to\infty}(1))\cdot$
$(\frac12 + o_{n\to\infty}(1))^{n-1}$ $\By{Theorem \ref{thm:bourgainvuwood}}{\sim}$
$(\frac12 + o_{n\to\infty}(1))$ $\cdot$ $\textup{\textsf{P}}_{\mathrm{lcf}} \bigl[\mathrm{Ra}_{<n-1}( \{0,\pm\}^{[n-1]^2} )\bigr ]$,
which is \ref{eq:equivalentformulationoftheconjecturewithabbreviations}. As to the
converse, \ref{eq:equivalentformulationoftheconjecturewithabbreviations} implies
$\textup{\textsf{P}} \bigl [ \mathrm{Ra}_{<n} ( \{\pm\}^{[n]^2} )\bigr ]$
$\By{Corollary \eqref{cor:uniformmeasureofranksublevelsetaschiomeasureofranksublevelsetoneranklower}}{=}$ $\textup{\textsf{P}}_{\mathrm{chio}}\bigl [ \mathrm{Ra}_{<n-1} ( \{0,\pm\}^{[n-1]^2} )\bigr ]$
$\By{\ref{eq:equivalentformulationoftheconjecturewithabbreviations}}{\leq}$
$(\frac12 + o_{n\to\infty}(1))$ $\cdot$ $\textup{\textsf{P}}_{\mathrm{lcf}} \bigl [ \mathrm{Ra}_{<n-1}( \{0,\pm\}^{[n-1]^2} )\bigr]$
$\By{Theorem \ref{thm:bourgainvuwood}}{\sim}$ $(\frac12 + o_{n\to\infty}(1))\cdot$
$(\frac12 + o_{n\to\infty}(1))^{n-1}$ $=$ $(\frac12 + o_{n\to\infty}(1))^n$, which is \ref{lem:equivalentproblem:it:folkloreconjecture}.
The equivalence \ref{eq:equivalentformulationoftheconjecturewithabbreviations}
$\Leftrightarrow$ \ref{conj:relativizedconjecture} is obvious. As to
\ref{conj:relativizedconjecture} $\Leftrightarrow$ \ref{arelativizedformulation},
note that
\ref{conj:relativizedconjecture} $\By{Proposition \ref{prop:chiomeasureasuniformmeasureonchiosets}}{\Leftrightarrow}$
$\sum$ $\llbracket$ $B'$ $\in$ $\{0,1\}^{[n-1]^2}\colon$ $\mathrm{rk}(B')$ $<$ $n-1$
$\rrbracket$ $\textup{\textsf{P}}[B']$ $\leq$ $\bigl(\frac12+o_{n\to\infty}(1)\bigr)$ $\cdot$
$\sum$ $\llbracket$ $B''$ $\in$ $\mathrm{Ra}_{<n-1}( \{0,\pm\}^{[n-1]^2} )$ $\rrbracket$
$\textup{\textsf{P}}_{\mathrm{lcf}}$ $\bigl [$ $B''$ $\bigr ]$
$\Leftrightarrow$ $\lvert \{$ $B'$ $\in$ $\{0,1\}^{[n-1]^2}\colon$ $\mathrm{rk}(B')$
$<$ $n-1$ $\} \rvert$
$\leq$ $(\tfrac12 + o_{n\to\infty}(1))$ $\cdot$ $\sum_{B''\in\{0,\pm\}^{[n-1]^2}\colon\mathrm{rk}(B')<n-1}$
$(\tfrac12)^{\mathrm{supp}(B'')}$
$\Leftrightarrow$
$\lvert \{$ $B'$ $\in$ $\{0,1\}^{[n-1]^2}\colon$ $\mathrm{rk}(B')$ $<$ $n-1$ $\} \rvert$ $\leq$
$(\tfrac12 + o_{n\to\infty}(1))$ $\cdot$ $\sum_{B\in\{0,1\}^{[n-1]^2}}$
$\sum$ $\llbracket$ $B''$ $\in$ $\{0,\pm\}^{[n-1]^2}\colon$ $\mathrm{Supp}(B'')$ $=$
$\mathrm{Supp}(B)$, $\mathrm{rk}(B'')$ $<$ $n-1$ $\rrbracket$ $(\tfrac12)^{\mathrm{supp}(B'')}$
$\Leftrightarrow$
$\lvert\{$ $B'$ $\in$ $\{0,1\}^{[n-1]^2}\colon$ $\mathrm{rk}(B') < n-1$ $\}$ $\rvert$
$\leq$ $(\tfrac12 + o_{n\to\infty}(1))$ $\cdot$ $\sum_{B\in\{0,1\}^{[n-1]^2}}$
$(\tfrac12)^{\mathrm{supp}(B)}$
$\cdot$ $\lvert$ $\{$ $B''$ $\in$ $\{0,\pm\}^{[n-1]^2}\colon$ $\mathrm{Supp}(B'')=\mathrm{Supp}(B)$,
$\mathrm{rk}(B'')$ $<$ $n-1$ $\}\rvert$ $\Leftrightarrow$ \ref{arelativizedformulation}.
\end{proof}
Note the `relativizing' effect of having two sums over the same index set on either
side of an (conjectured) inequality: thanks to commutativity of addition one may
go about pitting (collections of) unequally indexed summands on both sides of
\ref{conj:relativizedconjecture} against one another, in the hope of finding a
rearragement that allows one to prove the inequality without any a priori knowledge
about the size of the index set of the sums. Of course, \emph{if}
\ref{lem:equivalentproblem:it:folkloreconjecture} is true, \emph{then} the
inequality is true for \emph{every} permutation of the summands but the point is
that this is not known and that it would suffice to prove the existence of only
one suitable rearrangement of the summands to prove (or maybe disprove)
Conjecture \ref{lem:equivalentproblem:it:folkloreconjecture}.
\subsubsection{The inequality \ref{eq:equivalentformulationoftheconjecturewithabbreviations} must fail on entry-specification events}
Let us remark that in view of the formula
\ref{relationbeweenchiomeasureandlazycoinflipmeasuregovernedbyfirstbettinumber} in
Theorem \ref{thm:graphtheoreticalcharacterizationofthechiomeasure} we find ourselves
in a slightly ironic situation: while
\ref{eq:equivalentformulationoftheconjecturewithabbreviations}, which speaks about
the $\textup{\textsf{P}}_{\mathrm{chio}}$-measure of the (rather mysterious) event
$\mathrm{Ra}_{<n-1}(\{0,\pm\}^{[n-1]^2})$, may well be true, it cannot possibly be true in a
non-trivial way (left-hand side nonzero) on any of the (rather simple)
entry-specification events.
\subsubsection{Worst possible failure of \ref{eq:equivalentformulationoftheconjecturewithabbreviations} for singleton events}\label{worstpossiblefailure}
Already in Corollary \ref{cor:recipe} we have seen examples that
the inequality \ref{eq:equivalentformulationoftheconjecturewithabbreviations} can fail when the
event $\mathrm{Ra}_{<n-1}(\{0,\pm\}^{[n-1]^2})$ is replaced by other events---the failure
seeming more likely and more severe as the events get smaller. We will now see
that \ref{eq:equivalentformulationoftheconjecturewithabbreviations} fails arbitrarily
badly on every atom of the measure space we are dealing with (i.e. a singleton
event $\{B\}$ with $B\in \{0,\pm\}^{[n-1]^2}$ and $\textup{\textsf{P}}_{\mathrm{chio}}[B] > 0$). For such
events the ratio of $\textup{\textsf{P}}_{\mathrm{chio}}$ and $\textup{\textsf{P}}_{\mathrm{lcf}}$ diverges as quickly
as $2^{n^2}$ when $n\rightarrow \infty$,
while \ref{eq:equivalentformulationoftheconjecturewithabbreviations} asserts a
bounded ratio as $n\rightarrow \infty$.
By \ref{relationbeweenchiomeasureandlazycoinflipmeasuregovernedbyfirstbettinumber}
in Theorem \ref{thm:graphtheoreticalcharacterizationofthechiomeasure}, we know
$\textup{\textsf{P}}_{\mathrm{chio}}[ \mathcal{E}_B^J ]/\textup{\textsf{P}}_{\mathrm{lcf}} [\mathcal{E}_B^J] = 2^{\beta_1( \mathrm{X}_B )}$.
Therefore, to determine the maximum of $\textup{\textsf{P}}_{\mathrm{chio}}[ \mathcal{E}_B^J ]/\textup{\textsf{P}}_{\mathrm{lcf}} [\mathcal{E}_B^J]$ over all entry specification events $\mathcal{E}_B^J$ it suffices to
determine the maximum of $\beta_1( \mathrm{X}_B )$ over all bipartite graphs
$\mathrm{X}_B\in \mathrm{BG}_{n,n}$ with $B\in \{0,\pm\}^{[n-1]^2}$. The Betti number
$\beta_1(X) = f_1(X) - f_0(X) + \beta_0(X)$ as a function of $X\in\mathrm{BG}_{n,n}$ attains
a unique maximum at $X = K^{n-1,n-1}$. The corresponding value is
$(n-1)^2 - 2(n-1) + 1 = (n-2)^2$. Since $K^{n-1,n-1}$ can indeed occur as $\mathrm{X}_B$ with
$B$ $\in$ $\mathrm{Ra}_{<n-1}(\{0,\pm\}^{[n-1]^2})$ $\cap$
$\im( \tfrac12 \mathrm{C}_{(n,n)}\colon$ $\{\pm\}^{[n]^2}$ $\rightarrow$
$\{0,\pm\}^{[n-1]^2} )$, it follows that for every fixed $n$, the maximum of
$\textup{\textsf{P}}_{\mathrm{chio}}[ \mathcal{E}_B^J ] / \textup{\textsf{P}}_{\mathrm{lcf}} [\mathcal{E}_B^J]$ over all
$\mathcal{E}_B^J$ with $\emptyset\neq I \subseteq J \subseteq [n-1]^2$ and
$B\in\{0,\pm\}^I \cap \im( \tfrac12 \mathrm{C}_{(n,n)}\colon
\{\pm\}^{[n]^2} \rightarrow \{0,\pm\}^{[n-1]^2} )$ is $2^{(n-2)^2}\sim 2^{n^2}$.
Since Proposition \ref{prop:allbalancedsigningshavethesamerank} implies that every
$B\in\{0,\pm\}^I \cap \im( \tfrac12 \mathrm{C}_{(n,n)}\colon
\{\pm\}^{[n]^2} \rightarrow \{0,\pm\}^{[n-1]^2} )$ which realizes the maximum,
i.e. $\mathrm{X}_B\cong K^{n-1,n-1}$, has rank $1$, it follows that $2^{(n-2)^2}$ is also the
maximum of $\textup{\textsf{P}}_{\mathrm{chio}}[ \mathcal{E}_B^J ] / \textup{\textsf{P}}_{\mathrm{lcf}} [\mathcal{E}_B^J]$ over all
$\mathcal{E}_B^J$ with $\emptyset\neq I \subseteq J \subseteq [n-1]^2$ and
$B\in\mathrm{Ra}_{<n-1}(\{0,\pm\}^{[n-1]^2}) \cap \im( \tfrac12 \mathrm{C}_{(n,n)}\colon
\{\pm\}^{[n]^2} \rightarrow \{0,\pm\}^{[n-1]^2} )$.
\subsubsection{Extent of failure of \ref{eq:equivalentformulationoftheconjecturewithabbreviations} on those $B$
which are Chio condensates of random $A\in\{\pm\}^{[n]^2}$}
Let us have a quick informal look at the typical value of the ratio
$\textup{\textsf{P}}_{\mathrm{chio}}[B]$ and $\textup{\textsf{P}}_{\mathrm{lcf}}[B]$ if $\{0,\pm\}^{[n-1]^2}\ni B := \mathrm{C}_{(n,n)}(A)$
with $A\in\{\pm\}^{[n]^2}$ chosen uniformly at random. Of course, such a $B$ is
(by Theorem \ref{thm:bourgainvuwood} and Lemma \ref{lem:chiocondensationaffectsrankintheleastpossibleway}) asymptotically almost surely \emph{not} an element of
$\mathrm{Ra}_{<n-1}(\{0,\pm\}^{[n-1]^2})$.
By Corollary \ref{cor:whatrandombipartitegraphisX12CnnAforrandomA},
for $A\in\{\pm\}^{[n]^2}$ chosen uniformly at random, the graph $\mathrm{X}_B$ is a
random bipartite graph with partition classes of $n-1$ vertices on either side
and having i.i.d. edges with probability $\frac12$. Since this is very far above
the threshold proved by Frieze \cite{MR829352} for hamiltonicity of a random
bipartite graph, it follows that (for \emph{very} strong reasons) a.a.s.
$\beta_0(\mathrm{X}_B) = 1$. As to $f_1( \mathrm{X}_B )$, a standard argument using Chernoff's
bound shows that for every $\epsilon>0$ a.a.s. (and approaching $1$ exponentially
fast)
$(\frac12-\epsilon)\cdot (n-1)^2 \leq f_1( \mathrm{X}_B ) \leq (\frac12 + \epsilon) \cdot
(n-1)^2$. For simplicity let us pretend that $f_1( \mathrm{X}_B ) = \frac12 (n-1)^2$
exactly. Then $\textup{\textsf{P}}_{\mathrm{chio}}[ B ] / \textup{\textsf{P}}_{\mathrm{lcf}} [B] = 2^{\beta_1(\mathrm{X}_B)} =
2^{f_1(\mathrm{X}_B) - f_0(\mathrm{X}_B) + \beta_0(\mathrm{X}_B)} = 2^{\frac12 n^2 - 3n + \frac32} \sim 2^{\frac12 n^2} =
\sqrt{2^{n^2}} \rightarrow \infty$ as $n\rightarrow \infty$, and we have learned that
the worst-case failure-ratio of \ref{eq:equivalentformulationoftheconjecturewithabbreviations} found
in \ref{worstpossiblefailure} arises roughly by squaring the failure-ratio for a
$B = \tfrac12\mathrm{C}_{(n,n)}(A)$ with $A\in\{\pm\}^{[n]^2}$ random.
\section{Concluding questions}
Let us close with three questions:
\subsection{What to make of the $k$-wise independence?}
Note that Theorem \ref{thm:comparativecountingtheorem} in particular says that
one application of $\tfrac12 \mathrm{C}_{(n,n)}$ to a $\{\pm\}$-valued $n\times n$-matrix
yields a $\{0,\pm\}$-matrix whose entries are distributed
as $\frac14,\frac12,\frac14$ but are merely $3$-wise stochastically independent
(in the sense of \cite[Definition 2.4, p.209]{MR2410304}) (and `almost' $k$-wise,
the `almost' quantified exactly for $k\in \{4,5,6\}$ in Theorem \ref{thm:comparativecountingtheorem} and quantified roughly for general $k$ in
Proposition \ref{prop:roughestimationofnumberoffailureevents}). There appears to be
ongoing research on how partial independence relates to full independence, a keyword
being `$k$-wise independence'. To give only
one very recent (the topic has been studied at least since the 1980s) example, which
appears to summarize well the overall spirit, here is a quote from the abstract
of \cite{BenjaminiGurelGurevichPeled}:
{\scriptsize
\begin{quote}
We pursue a systematic study of the following problem. Let
$f\colon \{0,1\}^n\rightarrow \{0,1\}$ be a (usually monotone) boolean function
whose behaviour is well understood when the input bits are identically independently
distributed. What can be said about the behaviour of the function when the input
bits are not completely independent, but only $k$-wise independent [...]? How high
should $k$ be so that any $k$-wise independent distribution ``fools'' the function,
i.e. causes it to behave nearly the same as when the bits are completely independent?
\end{quote}
}
The author wonders whether the theory of $k$-wise stochastical independence has any
bearing on the present problem. In particular, can the proof of Bourgain--Vu--Wood
for $\textup{\textsf{P}}_{\mathrm{lcf}}$ be deconstructed and somehow reassembled for $\textup{\textsf{P}}_{\mathrm{chio}}$, aided
by knowledge about $k$-wise independence?
\subsection{More accurate estimations?}
Note that Theorem \ref{thm:comparativecountingtheorem}.\ref{comparativecountingtheorem:item:fourentriesspecified}---\ref{comparativecountingtheorem:item:sixentriesspecified} teaches us that, asymptotically, the ratio
$\lvert \mathcal{F}^{\mathrm{M}}(k,n)\rvert$ $/$ $\lvert \{ B\in \{0,\pm\}^I,\;
I\in\binom{[n-1]^2}{k}\}\rvert$ for $k\in \{4,5,6\}$ takes values
$4 n^4 / \frac{27}{8} n^8$ $=$ $\frac{32}{27}$ $n^{-4}$ $\in$
$[$ $1.1 n^{-4},$ $1.2 n^{-4}$ $]$,
$12 n^6 / \frac{81}{40} n^{10}$ $=$ $\frac{480}{81}$ $n^{-4}$ $\in$ $[$ $5.9 n^{-4},$
$6.0$ $n^{-4}$ $]$ and
$18 n^8 / \frac{81}{80} n^{12}$ $=$ $\frac{1440}{81}$ $n^{-4}$ $\in$ $[$ $17.7$
$n^{-4},$ $17.8$ $n^{-4}$ $]$ which are all---although of course growing
with $k$---vanishing with the same speed
$\mathcal{O}_{n\to\infty}(n^{-4})$. This shows that the rough bound in Proposition
\ref{prop:roughestimationofnumberoffailureevents} is not an asymptotically tight
one. This of course raises two questions: \emph{Is it true that
$\lvert \mathcal{F}^{\mathrm{M}}(k,n)\rvert$ $/$ $\lvert \{ B\in \{0,\pm\}^I,\;
I\in\binom{[n-1]^2}{k}\}\rvert$ $\in\mathcal{O}_{n\to\infty}(n^{-4})$ for
every fixed $k$?} Moreover, note that while Proposition
\ref{prop:roughestimationofnumberoffailureevents} shows that
$\lvert \mathcal{F}^{\mathrm{M}}(k,n)\rvert$ $/$ $\lvert \{ B\in \{0,\pm\}^I,\;
I\in\binom{[n-1]^2}{k}\}\rvert$ vanishes as $n\to\infty$ for every fixed $k$,
the discussion in \ref{worstpossiblefailure} shows that for the extreme case
of $k=(n-1)^2$, i.e. if the entire matrix $B\in \{0,\pm\}^{[n-1]^2}$ is specified,
$\textup{\textsf{P}}_{\mathrm{chio}}[\mathcal{E}_B^{[n-1]^2}] \neq \textup{\textsf{P}}_{\mathrm{lcf}}[\mathcal{E}_B^{[n-1]^2}]$ for the
vast majority of $B\in \{0,\pm \}^{[n-1]}$. In between these two
extremes, i.e. almost sure agreement as apposed to almost sure non-agreement of
$\textup{\textsf{P}}_{\mathrm{chio}}$ and $\textup{\textsf{P}}_{\mathrm{lcf}}$, there should be a tipping point. This raises the
question: \emph{For what order of growth $k=k(n)$ does
$\lvert \mathcal{F}^{\mathrm{M}}(k,n)\rvert$ $/$ $\lvert \{ B\in \{0,\pm\}^I,\;
I\in\binom{[n-1]^2}{k}\}\rvert$ first become bounded away from zero? And for what
order does it first tilt in favour of the non-agreement events?}
\subsection{Hidden connections to the Guralnick--Mar{\'o}ti-theorem?}
If $\sigma\colon G\rightarrow \Aut_K(V)$ is a representation of a group $G$ on
a $K$-vector space $V$, then for every $g\in G$ let $\mathrm{Fix}_V(g)$ denote the
\emph{fixed-point space of $g$}, i.e. the $K$-linear subspace
$\{v\in V\colon \sigma(g)(v) = v \}$.
In recent times there have been advances (\cite{MR1726791}, \cite{MR2231893},
\cite{MR2669683}) concerning the problem of bounding averages of dimensions of
fixed-point spaces by a fraction of the dimension of the representation, leading
to a full proof (and in more general form) by R. M. Guralnick and
A. Mar{\'o}ti \cite{MR2735760} of a 1966 conjecture of P. M. Neumann:
\begin{theorem}[Guralnick--Mar{\'o}ti \cite{MR2735760}, Theorem 1.1]
\label{thm:guralnickmarotitheorem}
For every finite group $G$ with smallest prime factor of $\lvert G \rvert$
denoted by $p$, every field $K$, every finite-dimensional $K$-vector space $V$,
every homomorphism $\sigma\colon G\rightarrow \Aut_K(V)$, and
every normal subgroup $N$ of $G$ which does not have a trivial composition
factor on $V$,
\begin{equation}\label{eq:guralnickmarotibound}
\frac{1}{\lvert N \rvert} \ \sum_{\tilde{g} \in N \cdot g}\ \dim_K(\mathrm{Fix}_V(\tilde{g}))
\leq \frac{1}{p}\ \dim_K(V) \quad .
\end{equation}
\end{theorem}
Although the resemblance is likely to be merely superficial, the author cannot
help being intrigued by the similarity of this inequality to
\ref{conj:relativizedconjecture}, together with the fact that both in
\ref{conj:relativizedconjecture} and in \eqref{eq:guralnickmarotibound} there can be
zero-summands on the left-hand side. Moreover, when trying to combine Chio
condensation of sign matrices with group actions, one gets the impression that
groups of even order (i.e. $p=2$) play a natural role. One goal along these lines
is to discover a vector space avatar of the lazy coin flip measure. Via
Theorem \ref{thm:graphtheoreticalcharacterizationofthechiomeasure} the author
found the following formula (which is of course easy to check directly): for
every $B\in \{0,\pm\}^{[n-1]^2}$ we have
\begin{equation}
\textup{\textsf{P}}_{\mathrm{lcf}}[B] = (\tfrac12)^{(n-1)^2}\cdot
(\tfrac12)^{\dim_{\mathbb{Z}/2}( \mathrm{B}^1(\mathrm{X}_B;\; \mathbb{Z}/2) \oplus \mathrm{Z}_1(\mathrm{X}_B;\; \mathbb{Z}/2) ) } \quad .
\end{equation}
This suggests studying group actions on the direct sum
$\mathrm{B}^1(\mathrm{X}_{\frac12\mathrm{C}_{(n,n)}(A)};\mathbb{Z}/2) \oplus \mathrm{Z}_1(\mathrm{X}_{\frac12\mathrm{C}_{(n,n)}(A)};\mathbb{Z}/2)$.
A sensible first choice are those actions which are induced by the `natural'
$\lvert \det(\cdot)\rvert$-preserving (hence intransitive) group actions on
$\{\pm\}^{[n]^2}$ (a merit of those `standard actions' is that Chio condensation
commutes with the corresponding actions on $\{0,\pm\}^{[n-1]^2}$). So far the
author did not detect much resonance with Theorem \ref{thm:guralnickmarotitheorem}.
Mirage or more?
\bibliographystyle{amsplain}
|
\section{Introduction}
For most problems of practical interest, the Schr\"{o}dinger
equation describing the evolution of a quantum system is too
complicated to be solved exactly. Numerical simulations are
therefore commonly employed for this purpose. However, simulating
quantum systems on a classical computer is a hard problem. The
dimension of the Hilbert space of the system increases
exponentially with the size of the system~(e.g., the number of
particles in the system). On a quantum computer, on the other
hand, the number of qubits required to simulate the system
increases linearly with the size of the system. As a result, the
simulation of quantum systems is more efficient on a quantum
computer~(see, e.g.,~\cite{bul, kas, somma1, somma2, somma3, aa,
whf1, temme}) than on a classical computer.
In general, quantum systems are never perfectly isolated from their
environments, and in some cases they must be treated as open systems.
Understanding the dynamics of open quantum systems is therefore important
for studying various quantum phenomena~\cite{gardiner, stolze, schl, weiss,
breuer1, cohen}. However, the simulation of open quantum systems on a
classical computer is also a hard task. In addition to the exponential
increase in the size of the Hilbert space of the open system, one also has
to consider the effect of the environment, which adds even more degrees of
freedom to the problem.
The Hamiltonian for an open quantum system coupled to an environment can be
expressed as
\begin{equation}
H=H_{S}+H_{B}+H_{I},
\end{equation
where $H_{S}$, $H_{B}$ and $H_{I}$ represent the Hamiltonians of the open
system, the environment and the interaction between the open system and the
environment, respectively. The interaction Hamiltonian $H_{I}$ can usually
be written in the form
\begin{equation}
H_{I}=\sum_{i}A_{i}\otimes B_{i},
\end{equation
where the operators $A_{i}$ and $B_{i}$ act in the state space of the open
system and the environment, respectively. In general, the environment has a
large number of degrees of freedom. Therefore the Hamiltonian in Eq.~($1$)
for the global system becomes too complicated to be solvable. However, one
is usually only interested in the evolution of the open system, and not the
environment. It turns out that for environments that have short correlation
times~(i.e., no memory effects) and induce slow decoherence in the system,
the microscopic details of the environment are not important. For purposes
of analyzing the dynamics of the system, one only needs to know a quantity
called the spectral density of the environment. The spectral density
characterizes the frequency distribution of the noise from the environment
on the open system. It combines the density of the environment modes and the
strength of the coupling between the environment modes and the system. For a
small set of discrete modes, the spectral density would consist of a
sequence of peaks. However, the frequencies of the environment modes are
usually so dense that the spectral density is a smooth function of frequency
\cite{weiss}.
Then, the task for the study of the dynamics of open systems can be
formulated as follows: Given the Hamiltonian of the open system, the
operators through which the open system interacts with the environment, the
spectral density of the environment and the temperature, how can we obtain
the dynamics of the open system? And how can we obtain the evolution of the
various physical properties of the open system?
In the case where the environment has no memory effects~(also referred to as
Markovian dynamics), the evolution of the system can be obtained by solving
the so-called master equation, which in the interaction picture has the form
\cite{breuer1}
\begin{widetext}
\begin{eqnarray}
\frac{d}{dt}\rho _{S}(t)\!&=&\!\sum_{\omega }\sum_{ij}\Biggl\{\!
-iS_{ij}\!\left(\omega \right)\!\left[A_{i}^{\dag }(\omega
)A_{j}(\omega),\rho_{S}\!(t)\right]\!+\!\gamma_{ij}(\omega
)\biggl\{A_{j}(\omega )\rho_{S}(t)A_{i}^{\dag }(\omega
)\!-\!\frac{1}{2}\left[A_{i}^{\dag }(\omega )A_{j}(\omega ), \rho
_{S}(t)\right]_{+}\biggr\}\Biggr\}.
\end{eqnarray}
\end{widetext} where $\rho _{S}(t)$ represents the density matrix of the
open system, and $A_{i}(\omega )$ is the decomposition of the operator
A_{i} $ into eigen-projectors of the Hamiltonian $H_{S}$. The operator
A_{i} $ represents the $i$-th operator that acts in the state space of the
open system in the interaction Hamiltonian as shown in Eq.~($2$). The
operators $A_{i}(\omega )$ can be very complicated in matrix form depending
on the details of the open system. The coefficients $S_{ij}\left( \omega
\right) $ are given~\cite{breuer1} by
\begin{widetext}
\begin{equation}
S_{ij}\left( \omega \right) =\frac{1}{2i}\left[\int_{0}^{\infty
}dte^{i\omega t}\langle B_{i}^{\dag }\left( t\right) B_{j}\left(
0\right) \rangle -\int_{0}^{\infty }dte^{-i\omega t}\langle
B_{i}^{\dag }\left( 0\right) B_{j}\left( t\right) \rangle \right],
\end{equation}
\end{widetext}where
\begin{equation}
B_{i}\left( t\right) =\exp \left( iH_{B}t\right) B_{i}\exp \left(
-iH_{B}t\right),
\end{equation}
and we have set $\hbar =1$. The operator $B_{i}$ represents the $i$-th
operator that acts in the state space of the environment in the interaction
Hamiltonian as shown in Eq.~($2$). The non-negative quantities $\gamma
_{ij}\left( \omega \right) $ play the role of decay rates for different
decay channels of the open system and are given in terms of certain
correlation functions of the environment~\cite{breuer1}
\begin{equation}
\gamma _{ij}\left( \omega \right) =\int_{-\infty }^{+\infty
}\!\!dt\;e^{i\omega t}\langle B_{i}^{\dag }\!\left( t\right) B_{j}\!\left(
0\right) \rangle .
\end{equation}
If one were to try and simulate the dynamics of a large open quantum system
on a classical computer, one would be faced with the problem that the number
of basis states grows exponentially with the size of the open system. The
master equation that describes the dynamics of the open system becomes too
complicated to be exactly solvable, and sometimes it is even practically
impossible to derive the master equation. One natural possibility for
tackling such problems is therefore to use quantum simulation.
There has been some work on the quantum simulation of open systems in the
literature. In Ref.~\cite{bacon}, the authors suggested an approach for
simulating the Markovian dynamics of an open quantum system on a quantum
computer. They showed that the simulation of the Markovian dynamics can be
reduced to building generators for a Markovian semigroup. However, as
mentioned above, the generator for the Markovian semigroup can be difficult
to obtain for a large system. In Ref.~\cite{terhal}, an approach for
preparing the thermal equilibrium state of an open quantum system was
suggested. Small-scale open system quantum simulators have also has been
demonstrated experimentally~\cite{bar, ryan}
In this paper, by extending the approach of Ref.~\cite{terhal}, we present a
quantum algorithm for simulating the Markovian dynamics of an open system
given the following information: the Hamiltonian of the open system, the
operators through which the open system interacts with the environment, the
spectral density of the environment and the temperature. This information
forms the input of the simulation.
The structure of this work is as follows: In Sec.~\ref{alg}, we present an
algorithm for simulating the dynamics of an open quantum system. In Sec.~\re
{res}, we estimate the resource requirements for the algorithm. In Sec.~\re
{eg}, we provide an example for the algorithm. We close with a conclusion
section.
\section{Theoretical description of the algorithm}
\label{alg}
In general, there are three steps in simulating the dynamics of an open
system on a quantum computer: first, preparing the initial state~(see, e.g.,
\cite{shende, berg, sok, nico, whf, bil}) of the open system and the
environment; second, implementing the dynamics on the open system and
finally reading out the state of the open system. We will focus on the step
of implementing the dynamics of the open system, and briefly describe the
other two steps of the algorithm in subsection II.C since they have been
discussed in the literature.
\subsection{Constructing a model Hamiltonian for the global system}
In general, the details related to the environment in Eq.~($1$)
are unknown. However, as mentioned above, one does not need to
know all these details in order to obtain the system dynamics. One
therefore has a good amount of freedom in constructing a model
Hamiltonian for the environment. One could therefore say that
under the Born-Markov approximation the spectral density is the
only piece of information that one needs to know about the
environment~(see e.g. Ref.~\cite{weiss, cohen}). Therefore, as
long as the spectral density of the model environment matches that
of the real environment, the effect on the system will be the
same.
In theoretical studies it is common to model the environment by a bath of
harmonic oscillators. For an open system in such an environment, the
environment Hamiltonian takes the form
\begin{equation}
H_{B}\!=\!\sum_{k}\omega _{k}\left( b_{k}^{\dag }b_{k}+\frac{1}{2}\right) ,
\end{equation
and the interaction Hamiltonian can be expressed as
\begin{equation}
H_{I}\!=\!\sum_{k}c_{k}\;\widetilde{A}_{k}\otimes \left(
b_{k}^{\dag }+b_{k}\right),
\end{equation}
where $b_{k}^{\dag }$~($b_{k}$) are the creation~(annihilation) operators of
the environment modes; $\omega _{k}$ are the frequencies of the environment
modes and $c_{k}$ are the coupling coefficients between the open system and
the environment modes; $\widetilde{A}_{k}$ are operators that act in the
state space of the open system and depend on the coupling mechanism between
the open system and the environment. For a set of discrete modes, the
spectral density is usually written as~\cite{weiss}
\begin{equation}
J\left( \omega \right) =\frac{\pi }{2}\sum_{k}\frac{c_{k}^{2}}{m_{k}\omega
_{k}}\;\delta \!\left( \omega -\omega _{k}\right) ,
\end{equation
where $m_{k}$ is the mass of the $k$-th oscillator. The $\delta $ in Eq.~($9
) is not restricted to infinitely-sharp $\delta $-functions, but to $\delta
-function approximants.
For purposes of simulating the dynamics of an open quantum system on a
digital quantum computer~(which is based on two-state qubits), it is
probably more natural to model the environment as a bath of two-level
systems or spin-$1/2$ particles. Such models are also sometimes used in
theoretical studies~(see, e.g.,~\cite{hanggi, lidar}). For an open system in
a spin bath, the environment Hamiltonian and the interaction Hamiltonian can
be expressed in the form
\begin{equation}
H_{B}=\frac{1}{2}\sum_{k}\omega _{k}\;\sigma _{k}^{z},
\end{equation
and
\begin{equation}
H_{I}=\frac{1}{2}\sum_{k}c_{k}\;\widetilde{A}_{k}\otimes \left(
g_{r}\,\sigma _{k}^{x}+g_{\varphi }\,\sigma _{k}^{z}\right) ,
\end{equation
respectively, where $\sigma _{k}^{x}$ and $\sigma _{k}^{z}$ are the Pauli
operators, $g_{r}$ and $g_{\varphi }$ are coefficients that describe the
relative size of the transverse coupling and the longitudinal coupling to
the open system. The transverse component induces relaxation, whereas the
longitudinal component induces pure dephasing. As will be explained in Sec.
II.D, there is a simple alternative method for simulating pure dephasing. We
therefore take $g_{r}=1$ and $g_{\varphi }=0$. The spectral density can then
be expressed as~\cite{hanggi}
\begin{equation}
J\left( \omega \right) =\pi \sum_{k}c_{k}^{2}\;\delta \!\left( \omega
-\omega _{k}\right) .
\end{equation
The difference between Eq.~($9$) and Eq.~($12$) is mostly a matter of
convention.
The environment mode frequencies $\omega _{k}$ and the coupling coefficients
$c_{k}$ can be determined as follows: We discretize the frequency spectrum
of the environment modes in the full frequency range from $\omega _{\min }$
to $\omega _{\max }$ into $d$ elements where each element has a width
\Delta \omega $:
\begin{equation}
\Delta \omega =\frac{\omega _{\max }-\omega _{\min }}{d}.
\end{equation
Correspondingly, the spectral density of the environment, $J\left( \omega
\right) $, is discretized and has the value $J\left( \omega _{k}\right) $
for the $k$-th element, where $\omega _{k}$ is the frequency of the $k$-th
element that represents the $k$-th environment mode. For a given $\omega
_{k} $, by making the following approximation
\begin{equation}
\int_{\omega _{k}-\Delta \omega /2}^{\omega _{k}+\Delta \omega /2}J\left(
\omega \right) d\omega \;\approx \;J\left( \omega _{k}\right) \Delta \omega
\;=\;\pi c_{k}^{2},
\end{equation
the corresponding coupling coefficient $c_{k}$ between the $k$-th element
and the open system can be obtained. Then the model Hamiltonian is
constructed based on the given information of the global system and will be
used in the implementation of the algorithm.
\subsection{Simulating the Markovian dynamics of an open system}
The dynamics of the open system is described by the evolution of the reduced
density matrix obtained by tracing out the environment degrees of freedom
from the density matrix of the global system:
\begin{equation}
\rho _{S}(t)=\text{Tr}_{B}[\rho (t)],
\end{equation
where $\rho _{S}(t)$ and $\rho (t)$ are the density matrices of the open
system and the global system, respectively. The density matrix $\rho (t)$
undergoes unitary evolution
\begin{equation}
\rho (t)=U(t,t_{0})\rho (t_{0})U^{\dag }(t,t_{0}),
\end{equation
where
\begin{equation}
U(t,t_{0})=\exp \left[ -iH(t-t_{0})\right] .
\end{equation}
Ideally, the dynamics of the open system can be obtained by coupling the
open system to an environment that has a large number of particles, letting
it evolve and reading out the state of the open system. In practice, on a
digital quantum computer that has a limited number of qubits, it is
impossible to represent all the particles of a typical environment.
Therefore we have to use an alternative technique to simulate the dynamics
of the open system in a large environment.
\begin{figure}[tbp]
\includegraphics[width=0.9\columnwidth, clip]{fig1.eps}
\caption{(a)~Quantum circuit for simulating the Markovian dynamics of an
open quantum system coupled to a time-independent environment. The first
register represents the open system and the second register represents the
environment. (b)~Quantum circuit for simulating the Markovian dynamics of an
open system coupled to a time-dependent environment.}
\end{figure}
The large size of the environment plays a crucial role in justifying the
Markovian approximation. Under this approximation, the typical time during
which the internal correlations of the environment related to the effects of
the open system exist, $\tau _{R}$, is much shorter than the characteristic
relaxation time of the open system, $\tau _{S}$; as a result, the influence
of the open system on the environment is small and can be ignored.
Therefore, the state of the global system at any time $t$ can be
approximately described by the tensor product~\cite{breuer1}
\begin{equation}
\rho (t)\approx \rho _{S}(t)\otimes \rho _{B}^{\text{th}}.
\end{equation
where $\rho _{B}^{\text{th}}$ is the thermal equilibrium state of the
environment and can be written as
\begin{equation}
\rho _{B}^{\text{th}}=\sum_{j}P_{j}|j\rangle \langle j|,j=1,\cdots ,L=2^{d},
\end{equation
with
\begin{equation}
P_{j}=\frac{e^{-\beta E_{j}}}{Z},
\end{equation
and $|j\rangle $~($E_{j}$) are the eigenvectors~(eigenvalues) of the
environment Hamiltonian $H_{B}$, $Z$ is the partition function
\begin{equation}
Z=\sum_{j=1}^{L}e^{-\beta E_{j}},
\end{equation
where $\beta =1/k_{B}T$, $k_{B}$ is the Boltzmann constant and $T$ is the
temperature. Note that in the absence of interactions between the
environment qubits, the eigenstates $|j\rangle $ can be written as a tensor
product of the states of the environment qubits
\begin{equation}
|j\rangle =|j_{1}\rangle \otimes |j_{2}\rangle \otimes \cdots \otimes
|j_{d}\rangle ,
\end{equation
and the eigenvalues $E_{j}$ can be written as a sum of the eigenvalues of
the Hamiltonians of the environment qubits for the corresponding eigenstate
|j\rangle
\begin{equation}
E_{j}=E\left( j_{1}\right) +E\left( j_{2}\right) +\cdots +E\left(
j_{d}\right) .
\end{equation
As a result, the thermal-equilibrium state of the environment is a tensor
product of the thermal-equilibrium states of the individual qubits
\begin{equation}
\rho _{B}^{\text{th}}=\rho _{1}^{\text{th}}\otimes \rho _{2}^{\text{th
}\otimes \cdots \otimes \rho _{d}^{\text{th}},
\end{equation
where
\begin{equation}
\rho _{k}^{\text{th}}=\left( 1-p_{k}\right) |0\rangle \langle
0|+p_{k}|1\rangle \langle 1|,
\end{equation
denotes the thermal equilibrium state of the $k$-th environment qubit an
\begin{equation}
p_{k}=\frac{1}{1+e^{\beta \omega _{k}}}.
\end{equation}
An alternative method that can be used to obtain the Markovian approximation
for a relatively small environment is to frequently force it back into its
thermal-equilibrium state. This process can be implemented relatively easily
on a quantum computer, where one has full access to all the qubits. The
procedure for simulating the Markovian dynamics of the open system is
therefore as follows: Set the environment to its thermal equilibrium state,
couple the open system to the environment modes and let the global system
evolve for some time, then reset the state of the environment to its thermal
equilibrium state. Repeat this process many times and then read out the
state of the open system. Mathematically, the evolution of the open system
in one step is expressed as
\begin{equation}
\rho _{S}^{\left( j+1\right) }\left[ (j+1)\tau \right] =\text{Tr}_{B}\left[
U(\tau )\rho ^{j}(j\tau )U^{\dag }(\tau )\right] ,j=0,1,2,\dots
\end{equation
where $\rho ^{j}(j\tau )=\rho _{S}^{j}\otimes \rho _{B}^{th}$.
Based on the above analysis, the evolution of the open system can be
simulated as follows on a quantum computer $\left[ \text{see Fig.}~1\left(
a\right) \right] $: prepare two quantum registers $R_{S}$ and $R_{B}$ to
represent the open system and the environment, respectively;
$\left( i\right) $ on the quantum register $R_{S}$, prepare the initial
state of the open system;
$\left( ii\right) $ on the quantum register $R_{B}$, prepare the thermal
equilibrium state of the environment;
$\left( iii\right) $ implement the unitary operation $U(\tau )=\exp (-iH\tau
)$ on the registers $R_{S}$ and $R_{B}$, where $H$ is the model Hamiltonian;
$\left( iv\right) $ repeat steps $\left( ii\right) -\left( iii\right) $ a
number of times.
$\left( v\right) $ read out the state, or the desired observable, of the
register $R_{S}$.
In general, the different terms in the model Hamiltonian do not commute. We
therefore employ the Trotter-Suzuki formula~\cite{nc} for implementing the
unitary operation $U(\tau )=\exp (-iH\tau )$
\begin{widetext}
\begin{eqnarray}
U(\tau ) &=&\exp \left[ -iH\tau \right] \notag \\
&=&\lim_{n\rightarrow \infty }\left[ e^{-iH_{S}\tau /n}e^{-iH_{B_{1}}\tau
/n}e^{-iH_{I_{1}}\tau /n}\cdots e^{-iH_{B_{d}}\tau /n}e^{-iH_{I_{d}}\tau /n
\right] ^{n} \notag \\
&=&\lim_{n\rightarrow \infty }\left[ U_{S}\left( \tau /n\right)
U_{B_{1}}\left( \tau /n\right) U_{I_{1}}\left( \tau /n\right)
\cdots U_{B_{d}}\left( \tau /n\right) U_{I_{d}}\left( \tau
/n\right) \right] ^{n} \notag \\
&\approx&\left[ U_{S}\left( \tau /n_0\right) U_{B_{1}}\left( \tau
/n_0\right) U_{I_{1}}\left( \tau /n_0\right) \cdots
U_{B_{d}}\left( \tau /n_0\right) U_{I_{d}}\left( \tau /n_0\right)
\right] ^{n_0},
\end{eqnarray}
\end{widetext}where $n_{0}$ is a finite but large number.
In many cases, one is interested in the value of some physical properties of
the open system in the thermal-equilibrium state, such as various
correlation functions, the partition function, etc. The thermal equilibrium
state of the open system can be obtained by repeating the steps $\left(
ii\right) -\left( iii\right) $ until the open system reaches its thermal
equilibrium state~\cite{terhal}.
Note that the steps followed in implementing the algorithm explained above
are the same as those used in the algorithm of Ref.~\cite{terhal}. The goal
of that work, however, is different from the goal of our work. Our algorithm
simulates the dynamics whereas the algorithm of Ref.~\cite{terhal} aims to
prepare the thermal equilibrium state of the system. This difference in the
purpose of the algorithm leads to a number of further differences. For
example, in our case, we have to design the interaction Hamiltonian and the
parameters of the environment such that we accurately reproduce the spectral
density of the environment. In Ref.~\cite{terhal}, one only requires that
certain inequalities are satisfied in order for the environment register to
act as a good environment. In other words, the exact speed of reaching
thermal equilibrium is not a crucial issue in that work, as long as the
thermal-equilibrium state is reached in polynomial time. In contrast, if for
example there is some symmetry in the simulated system that prevents it from
reaching the thermal-equilibrium state, then our algorithm would still be
considered to work successfully if it produces the correct dynamics, even
though the thermal-equilibrium state is never reached.
\subsubsection{Timescales}
At this point we should make a few comments regarding the time scales
involved in applying the algorithm. The time interval $\tau $ should be very
short compared with $\tau _{S}$~(so that the state of the open system
changes slightly during $\tau $). The time $\tau $ can be considered the
memory time of the environment since the environment is reset to its thermal
equilibrium state at every time interval $\tau $. Since the timescale for
dynamics involving the system and a resonant mode of the environment is
given by the inverse of $c_{i}$ times a matrix element between energy
eigenstates of the system, the Markovian condition would require that $\tau $
must be small compared to this timescale, such that that the change in the
system's density matrix is small during the time $\tau $. When constructing
the model Hamiltonian, the frequency spectrum of the environment is
discretized into many elements where each element has a width $\Delta \omega
$. This width $\Delta \omega $ must be at most on the order of $1/\tau $ to
make sure that the different $\delta $-peaks have large overlap and produce
a smooth spectral density since the width of the $\delta $-peaks is on the
order of $1/\tau $.
Note that our algorithm can be straightforwardly generalized to simulate the
Markovian dynamics of an open system coupled to a time-dependent
environment~[see Fig.~$1\left( b\right) $]. In each iteration, the state of
the environment register is reset to the time-dependent thermal equilibrium
state of the time-dependent environment. The algorithm can also be
generalized for the case of an open system that is simultaneously coupled to
two or more different environments with different temperatures.
\subsubsection{Sequential application of the different dissipation channels}
In the above procedure for simulating the dynamics of an open quantum
system, we have assumed that each mode in the environment is represented by
one qubit. In the quantum circuit shown in Fig.~$1\left( a\right) $, the
environment quantum register $R_{B}$ is prepared in the thermal equilibrium
state of all the environment qubits with all the different frequencies. The
operator $U(\tau )$ acts in the state space of the system qubits and all the
environment qubits.
In this subsection, we show that one can reduce the number of qubits
required to represent the environment by having each qubit represent
multiple environment modes. In this approach each \textquotedblleft
evolve-reset\textquotedblright step in the algorithm above is split into
multiple evolve-reset steps, and in each one of those steps a qubit
represents one environment mode. However, the qubit is reset to different
frequencies in the different steps such that it produces the effect of
multiple environment modes on the system. Under certain conditions even a
single qubit can be used to represent the entire environment: the parameters
of this qubit are sequentially alternated between several different settings
such that the sequence covers all the dissipation channels of the
environment.
In the master equation shown in Eq.~($3$), the derivative of the reduced
density matrix of the open system is described by a sum over the decay
processes of the open system through all the different dissipation channels.
In the simulation algorithm, the decay of the open system in one step of the
evolution can be expressed as
\begin{eqnarray}
\rho _{S}^{\text{int}}\bigl[(j+1)\tau \bigr] &=&\text{Tr}_{B}\Bigl\{U^{\text
int}}(\tau )\bigl[\rho _{S}(j\tau )\otimes \rho _{B}^{th}\bigr]U^{\text{int
\dag }(\tau )\Bigr\} \notag \\
&=&\left( 1-\tau \varGamma\right) \rho _{S}^{\text{int}}(j\tau ),
\end{eqnarray
where $\varGamma$ is a superoperator that describes the decay of the open
system and it is a sum over the contributions from all the different
environment modes~(the superscript \textquotedblleft int\textquotedblright\
indicates the interaction picture). Let $\varGamma_{j}$ denote the
superoperator that describes the decay of the open system caused by the
coupling of the open system to only the $j$-th environment mode. If the open
system undergoes a small change during the time $\tau $, such that $\tau
\varGamma\ll 1$~(i.e., all the eigenvalues of $\varGamma$ times $\tau $ are
much smaller than $1$), then the evolution of the open system can be
approximated as
\begin{equation}
1-\tau \varGamma\approx \left( 1-\tau \varGamma_{1}\right) \left( 1-\tau
\varGamma_{2}\right) \cdots \left( 1-\tau \varGamma_{d}\right) .
\end{equation
Therefore, the decay of the open system can be simulated by applying the
environment modes to the open system sequentially.
As a result, an alternative approach for simulating the dynamics of an open
system goes as follows: We divide the environment modes into a few sets and
let the different sets interact with the open system sequentially. In this
way, we effectively simulate the interaction between the open system and all
the different modes in the environment. The Hamiltonian for the open system
interacting with the $i$-th set of the environment modes is given by
\begin{equation}
H^{\left( i\right) }=H_{S}+\sum_{k=1}^{d_{i}}H_{B_{k}}^{\left( i\right)
}+\sum_{k=1}^{d_{i}}H_{I_{k}}^{\left( i\right) },
\end{equation
where $i=1,2,\cdots ,d/d_{i}$ and $d_{i}$ is the number of the environment
modes in each set.
One point that requires a little bit extra care here is that each mode in
the environment is coupled to the system in only one step out of $d/d_{i}$
steps, whereas the system Hamiltonian $H_{S}$ is applied in all of the
steps. One therefore needs to be careful how to calculate the elapsed time
in the simulated system. If the Hamiltonian in Eq.~($31$) is applied with a
given value of $\tau $, then after covering all the $d/d_{i}$ sets of
environment modes the system Hamiltonian would have induced a change
corresponding to a time $\tau \times d/d_{i}$, while each environment mode
would have induced a change corresponding to a time $\tau $. In order to
avoid any problems arising from this inconsistency, one can note that
decoherence rates are proportional to the spectral density~(and therefore
proportional to $c_{i}^{2}$) from the interaction Hamiltonian. One can then
use a rescaled Hamiltonian
\begin{equation}
\widetilde{H}^{^{\left( i\right)
}}=H_{S}+\sum_{j=1}^{d_{i}}H_{B_{j}}^{\left( i\right) }+\sqrt{\frac{d}{d_{i}
}\sum_{j=1}^{d_{i}}H_{I_{j}}^{\left( i\right) }.
\end{equation
If this Hamiltonian is now applied for a time $\tau $, then after covering
all the different environment modes both the system Hamiltonian and the
coupling to the environment would have induced changes that correspond to a
time $\tau \times d/d_{i}$, removing the inconsistency in the elapsed time.
Note that in order to guarantee the validity of the Markovian approximation,
the rescaled coupling strengths $c_{i}\times \sqrt{d/d_{i}}$ must still be
small compared to $1/\tau $.
The new procedure for simulating the Markovian dynamics of open systems is
implemented on a quantum computer as follows:
$i)$ on the quantum register $R_{S}$, prepare the initial state of the open
system;
$ii)$ on the quantum register $R_{B}$, prepare the thermal equilibrium state
of the qubits representing a subset of the environment modes;
$iii)$ implement the unitary operation $U_{i}(\tau )=\exp \left[ -
\widetilde{H}^{^{\left( i\right) }}\tau \right] $ on the registers $R_{S}$
and $R_{B}$;
$iv)$ perform steps $ii)-iii)$ for another set of environment modes, and
keep repeating this process until the algorithm runs over all the sets of
the environment modes;
$v)$ repeat steps $ii)-iv)$ a number of times;
$vi)$ read out the state, or the desired observable, of the register $R_{S}$;
In Eq.~($30$), the error in the decay of the open system introduced by
sequentially applying the dissipation channels, up to second order in $\tau
, is $\tau ^{2}\sum_{i,j}\varGamma_{i}\varGamma_{j}$. In the case of using
only a single qubit to represent the environment, the algorithm will have an
error of $d^{2}\tau ^{2}\overline{\varGamma}_{0}^{2}$, where $\overline
\varGamma}_{0}$ denotes the overall scale of the decay rate of the open
system caused by the coupling of the open system to a single environment
mode. When $d$ is large, this approximation can introduce a large error. On
the other hand, if we use more qubits to represent the environment, the
errors will be smaller because there will be fewer terms in the sum for the
error.
\subsection{State preparation and readout}
In the algorithm, the initial state of the global system is set to $\rho
_{S}(0)\otimes \rho _{B}^{\text{th}}$, where $\rho _{S}(0)$ is the initial
state of the open system and $\rho _{B}^{\text{th}}$ represents the thermal
equilibrium state of the environment. To prepare the initial state of the
global system on a quantum computer, we prepare two quantum registers $R_{S}$
and $R_{B}$ to represent the open system and the environment, respectively.
In general, the thermal equilibrium state of the environment is a mixed
state. In order to prepare the mixed state $\rho _{B}^{\text{th}}$ as shown
in Eq.~($19$), we can generate a random integer $j$, where $j\in \left[
\text{, }L\right] $, with probability $P_{j}$. Then we prepare the
corresponding state $|j\rangle $ on the quantum register $R_{B}$. We repeat
this step many times. This procedure produces an ensemble~(in time) of
states $|j\rangle $ with the corresponding probabilities $P_{j}$. This
ensemble gives the same effect as the case where the quantum register $R_{B}$
is prepared in a thermal equilibrium state in every time step.
As discussed in subsection II.A, in the absence of interactions between the
environment qubits, the thermal-equilibrium state of the environment is a
tensor product of the thermal-equilibrium states of the individual qubits as
shown in Eq.~($24$). In such cases, the thermal-equilibrium state of the
environment can be prepared in a simpler way: randomly generating $0$ or $1$
with respective probabilities $\left( 1-p_{k}\right) $ and $p_{k}$, then
preparing the corresponding states $|0\rangle $ or $|1\rangle $ on the
environment qubit, and repeating this step many times, the mixed state $\rho
_{B}^{\text{th}}$ can be prepared. In this procedure, an ensemble~(in time)
of states $|0\rangle $ and $|1\rangle $ with the corresponding probabilities
$p_{k}$ and $\left( 1-p_{k}\right) $ is produced, and it gives the same
effect as the case where the environment qubit is prepared in a thermal
equilibrium state in every time step.
\begin{figure}[tbp]
\includegraphics[width=0.9\columnwidth, clip]{fig2.eps}
\caption{Quantum circuit for a quantum estimator. $H$ represents the
Hadamard gate and SWAP represents the swap gate.}
\end{figure}
Observing the decoherence dynamics can be performed in a number of different
ways. For example, the phase estimation procedure~\cite{nc} can be used to
measure the energy of the open system, and one can then monitor the dynamics
of the energy distribution.
Alternatively, any matrix element in the density matrix of the open system
can be obtained using a quantum estimator~\cite{ekert, horodecki} as shown
in Fig. $2$. In the circuit for the quantum estimator, if we prepare the
second register in the state $|\varphi \rangle $ and the third register in
the state $|\psi \rangle $, then the value of $|\langle \varphi |\psi
\rangle |$ can be estimated by performing single-qubit measurements on the
index qubit. Through some derivation~\cite{ekert} we have
\begin{equation}
P\left( 0\right) =\frac{1}{2}\left( 1+|\langle \varphi |\psi \rangle
|^{2}\right) ,
\end{equation
where $P\left( 0\right) $ is the probability for obtaining the state
|0\rangle $ in the index qubit. If the third register is in a mixed state
\rho _{S}$, then we have
\begin{equation}
P\left( 0\right) =\frac{1}{2}\left( 1+\langle \varphi |\rho _{S}|\varphi
\rangle \right) .
\end{equation}
For the readout of the state of the open system in some chosen basis
\{|n\rangle \}$, any diagonal element $\rho _{nn}^{S}=\langle n|\rho
_{S}|n\rangle $ can be estimated by preparing the second register in the
state $|n\rangle $ and the third register in the state $\rho _{S}$. For the
off-diagonal elements $\rho _{mn}^{S}=\langle m|\rho _{S}|n\rangle $, the
real part of $\rho _{mn}^{S}$ can be estimated by preparing the second
register in the state $|\varphi \rangle =\left( |m\rangle +|n\rangle \right)
/\sqrt{2}$. The imaginary part of $\rho _{mn}^{S}$ can be estimated by
setting $|\varphi \rangle =\left( |m\rangle +i|n\rangle \right) /\sqrt{2}$.
The circuit shown in Fig.~$2$ can also be used for evaluating the
expectation values of arbitrary observables. For an operator $F$, by
applying the technique developed in Refs.~\cite{ekert, horodecki}, one can
obtain the value of $\langle F\rangle _{\rho _{S}}$. Then combined with the
quantum circuit for simulating the Markovian dynamics of open systems, one
can simulate the evolution of the expectation values of the physical
observables. Various correlation functions can also be obtained using this
technique.
\subsection{Simulating pure dephasing of a quantum system}
So far we have concentrated on the case where the nonunitary part of the
dynamics is caused by the environment modes that are resonant with the open
system. This picture is valid for energy relaxation. Pure dephasing, on the
other hand, is caused by low-frequency noise. Such low-frequency noise can
also be generated using the algorithm explained in the previous subsections:
whenever an environment qubit's state is changed from the ground to the
excited state or vice versa, the open system feels a telegraph-noise-like
change. This telegraph noise then causes~(mostly) pure dephasing in the
system. Although this effect can be induced using environment qubits, there
is a simpler method to generate telegraph noise. The telegraph noise is
essentially a classical noise signal affecting the open system. There is
therefore no need to use qubits in order to produce this classical signal.
It can be generated using a classical algorithm and added to the system
Hamiltonian $H_{S}$. If different noise signals are used in the different
runs of the algorithm, the density matrix of the system~(averaged over the
different realizations) will exhibit dephasing dynamics as a function of
time.
For an open system coupled to many fluctuators, the Hamiltonian for the
system can be expressed as~\cite{gal}~[see Eq.~($11$) for comparison
\begin{equation}
H=H_{S}+\sum_{k}\chi \!_{k}\left( t\right) \widetilde{A}_{k},
\end{equation
where
\begin{equation}
\ \chi _{k}\left( t\right) =\sum_{i}v_{ik}\;\xi _{ik}\left( t\right) ,
\end{equation
the random functions $\xi _{i}\left( t\right) $ characterize the
fluctuators' state, instantly switching between $\pm 1/2$ at random times.
Therefore, the simulation of the open system coupled to a bath of many
fluctuators is reduced to simulating a closed quantum system, which will
reduce the resources required for performing the simulation.
In simulating the dynamics of a quantum system coupled to many fluctuators,
we prepare the initial state of the system on a quantum register, then
implement the unitary operation $U=\exp \left( -iH\tau \right) $. Since the
signal $\chi \left( t\right) $ is random, to obtain the evolution of the
quantum system, we have to run the algorithm many times with a different
signal $\chi \left( t\right) $ in each run.
The noise spectrum of a noise signal is defined as~\cite{ymg
\begin{equation}
s\left( \omega \right) =\frac{1}{\pi }\int_{0}^{\infty }\!\!dt\;\cos \omega
t\langle \chi \left( t\right) \chi \left( 0\right) \rangle .
\end{equation
In order to generate the telegraph-noise signal~\cite{kog}, we
assume that the environment contains a number of fluctuators. Each
fluctuator switches between two possible configurations with
switching rate $\gamma _{i}$, and couples to the system with a
coupling strength $v_{i}$. The corresponding contribution of a
fluctuator to the noise spectrum is a Lorentzian~\cite{ymg},
\begin{equation}
s_{i}\left( \omega \right) =\frac{v_{i}^{2}\gamma _{i}}{4\pi \left( \omega
^{2}+\gamma _{i}^{2}\right) }.
\end{equation}
The~(low-frequency) noise spectrum of the entire environment is
the sum of the noise spectra of all the fluctuators
\begin{equation}
S\left( \omega \right) =\sum_{i}\frac{v_{i}^{2}\gamma _{i}}{\omega
^{2}+\gamma _{i}^{2}}.
\end{equation}
By adjusting the parameters $\gamma _{i}$ and $v_{i}$, one can produce a
variety of noise spectra. A typical example in practical situations is $1/f$
noise. If the number and density of fluctuators is sufficiently large, and
the distribution of the switching rates of the fluctuators $D\left( \gamma
\right) \propto \gamma ^{-1}$, and is independent of the distribution of the
coupling strengths between the fluctuators and the open system, the sum over
the fluctuators produces the $1/f$ noise spectrum~\cite{dutta}
\begin{equation}
S\left( \omega \right) =\frac{G}{\omega },
\end{equation
where $G$ is a constant.
\subsection{Simulating the non-Markovian dynamics of an open system}
The quantum algorithm presented above for simulating the Markovian dynamics
of an open system can also be used for simulating a class of non-Markovian
dynamics of open systems. A common situation where non-Markovian dynamics
occurs is the case where a small number of degrees of freedom in the
environment are coherent enough that they have non-negligible memory effects
in their interaction with the open system~\cite{breuer1}. Examples of this
situation include the relaxation dynamics of an atom through a cavity that
has a high quality factor~(see e.g., Ref.~\cite{breuer1}) and the relaxation
dynamics of a superconducting qubit close to resonance with a coherent
two-level defect~(see e.g. Ref.~\cite{ashhab}).
When a small number of degrees of freedom in the environment are responsible
for the non-Markovian dynamics, and assuming that one has sufficient
understanding of these degrees of freedom~(i.e. their intrinsic Hamiltonian,
their coupling to the system and their decoherence mechanisms and rates), it
becomes relatively straightforward to include them in the algorithm. One now
adds the appropriate number of ancilla qubits needed to describe these
degrees of freedom. When implementing the evolve-reset part of the
algorithm, one treats the additional degrees of freedom as part of the
system, i.e. they are not reset to their initial state. When the measurement
is performed at the end of the algorithm, however, only the system qubits
are used and the additional degrees of freedom are ignored. This step
corresponds to taking the trace over the state of these degrees of freedom.
The introduction of the additional degrees of freedom increases the resource
requirements~(which will be the subject of Sec.~III) as follows: the number
of qubits used in implementing the unitary operation $U(\tau )$ is increased
by $\log (D_{\text{ext}})$, where $D_{\text{ext}}$ is the number of degrees
of freedom in the environment that are responsible for the non-Markovian
dynamics. Since the interactions in the expanded system should still be
local, the scaling of resources will still be polynomial in system+ancilla
size and therefore efficient.
\subsection{Implementing environments other than spin baths}
Our algorithm simulates the effect of the environment using a set of ancilla
qubits that each represents one spin in a bath of independent spins~\cit
{pro}. This type of environment is rather straightforward to implement,
since the properties and manipulation of the environment are done by
considering the ancilla qubits one at a time. For a variety of purposes,
this spin bath is sufficient for the implementation of the desired quantum
simulation. However, the spin bath has certain limitations. For example, the
two-level nature of the environment elements~(i.e., spins) means that
increasing the temperature will reduce the number of spins that are in their
ground states, thus reducing the relaxation rate induced by this environment
\cite{lidar}. This behavior contrasts with the case of a bath of harmonic
oscillators, where both relaxation and excitation rates increase with
increasing temperature. Therefore, if one is interested in the temperature
dependence of the dynamics, one needs to be careful about the differences
between different types of environments.
One possible technique to use a spin bath in order to simulate an oscillator
bath and obtain the correct temperature dependence is to calculate a
modified~(temperature-dependent) spectral density for the spin bath and use
this spectral density in the simulation. Alternatively, different types of
environments can be simulated by modifying the algorithm such that each
element in the environment is encoded into multiple qubits that are treated
as a single physical object. For example, one could use $n$ environment
qubits to represent the lowest $2^{n}$ energy levels of a harmonic
oscillator and then design a Hamiltonian where this harmonic oscillator
represents one mode in the environment. Thus, one can simulate the dynamics
of an open quantum system interacting with a bath of harmonic oscillators.
Note that the state preparation and the form of the Hamiltonian become more
complicated in this case than in the case of a spin bath.
\section{Resource estimation}
\label{res}
In this section, we discuss the resources including the number of qubits and
the operations needed for implementing the algorithm.
As shown in Fig.~$1\left( a\right) $, the number of qubits required for
simulating the Markovian dynamics of the open system is $\lceil \log
_{2}N\rceil +d$~(here $\lceil x\rceil $ provides the smallest integer larger
than $x$, or equal to $x$ if $x$ is an integer), where $N$ is the dimension
of the Hilbert space of the open system and $d$ is the total number of
qubits representing the environment. In the quantum estimator for the
readout of the state of the open system, $\lceil \log _{2}N\rceil +1$
additional qubits are needed as shown in Fig.~$2$. Therefore the total
number of qubits for simulating the Markovian dynamics of open systems is
2\lceil \log _{2}N\rceil +d+1$. One could also use the ancilla qubits that
are used in representing the environment in the readout of the state of the
open system, which would reduce the number of qubits used in the algorithm
to $\max \left[ \lceil \log _{2}N\rceil +d,2\lceil \log _{2}N\rceil +1\right]
$.
\begin{table}[tbp]
\caption{Resource needed for implementing the algorithm. Here $\lceil
x\rceil $ provides the smallest integer larger than $x$, or equal to $x$ if
x$ is an integer. $N$ is the dimension of the Hilbert space of the open
system, $d$ is the total number of qubits needed in representing the
environment in one time, $d_{i}$ is the number of qubits used in
representing the environment in approach $2$. $m$ is the number of times the
unitary operation $U(\protect\tau )=\exp (-iH\protect\tau )$ is implemented,
and $n_{0}$ is the parameter for dividing the time in the Trotter expansion
in Eq.~($28$).}
\begin{center}
\begin{tabular}{ccccc}
\hline
Algorithm & & \# of qubits & & \# of operations \\ \hline
Approach~$1$ & & $2\lceil \log _{2}N\rceil \!+\!d\!+\!\!1$ & & $O(m
\!\times \! \left(2d+1\right)^{n_{0}})$ \\ \hline
Approach~$2$ & & $2\lceil \log _{2}N\rceil \!+\!d_{i}\!+\!\!1$ & &
O(m\!\times\! d/d_{i}\!\times \!\left( 2d_{i}+1\right) ^{n_{0}})$ \\ \hline
\end{tabular
\end{center}
\end{table}
The unitary operation $U(\tau )=\exp (-iH\tau )$, where $H$ is the model
Hamiltonian, is implemented a finite number $m$ of times. To implement
U(\tau )$ on the quantum circuit, we employ the Trotter-Suzuki formula as
shown in Eq.~($28$), in which $U(\tau )$ is approximated by the product of
\left( 2d+1\right) ^{n_{0}}$ unitary transformations, where $n_{0}$ is the
parameter for dividing the time in the Trotter expansion. Therefore the
number of unitary operations needed in the quantum circuit shown in Fig.~
1\left( a\right) $ is $m\times \left( 2d+1\right) ^{n_{0}}$. Note one could
obtain higher efficiency using higher order Suzuki-Trotter formulas, as
discussed in Ref.~\cite{childs}
In the second approach we presented, the environment is represented using a
few or a single qubit. The number of qubits required for simulating the
dynamics of open systems is: $2\lceil \log _{2}N\rceil +d_{i}+1$, where
d_{i}$ is the number of qubits representing the environment.
The unitary operations implemented for each set of qubits are $U_{i}(\tau
)=\exp \left[ -i\widetilde{H}^{^{\left( i\right) }}\tau \right] $, where
\widetilde{H}^{^{\left( i\right) }}$ is given in Eq.~($32$). Thus
implementing $U_{i}(\tau )$ on the circuits by employing the Trotter-Suzuki
formula requires $\left( 2d_{i}+1\right) ^{n_{0}}$ unitary operations. For
d/d_{i}$ sets of the environment elements where each set of the elements is
implemented $m$ times, the total number of unitary operations that need to
be implemented is $m\times d/d_{i}\times \left( 2d_{i}+1\right) ^{n_{0}}$.
In preparing the initial state of the global system, we have to prepare the
thermal equilibrium state of the environment, which is a mixed state, a
number of times. To prepare the thermal equilibrium state $\rho _{B}^{th}$
as shown in Eq.~($19$), we prepare the state $|j\rangle $ with the
corresponding probability $P_{j}$, and repeat this step many times. In
running the algorithm, these basis states are fed to the register $R_{B}$ in
the quantum circuit in Fig.~$1\left( a\right) $ one at a time, and run the
algorithm many times with the the register $R_{B}$ prepared in the basis
states. For each basis state $|j\rangle $, we need to reset the state on the
register $R_{B}$ $m$ times in the algorithm. To do this, we need to erase
the state on the register $R_{B}$ and then prepare $R_{B}$ in state
|j\rangle $. We can first perform a measurement on the register $R_{B}$, the
state on $R_{B}$ collapse to a basis state, then we can perform a unitary
operation to rotate this state into the state $|j\rangle $.
In the readout of the state of the open system, we employ the quantum
estimator, in which the information of the open system is obtained by
performing single-qubit measurements on the index qubit and taking the
average. We have to prepare many copies of the state of the open system in
order to obtain accurate results. Both this procedure and the procedure for
preparing the thermal equilibrium state of the environment require running
the algorithm many times. The number of times for running the algorithm,
however, does not depend on the dimension of the Hilbert space of the open
system. Therefore the algorithm can be implemented efficiently using $O\left[
m\times \left( 2d+1\right) ^{n_{0}}\right] $ (or $O\left[ m\times
d/d_{i}\times \left( 2d_{i}+1\right) ^{n_{0}}\right] $, if the environment
is represented using $d_{i}$ qubits) unitary operations. All these results
are summarized in Table~I.
\section{Example: a two-level system in a spin bath}
\label{eg}
In this subsection, we consider the example of simulating the Markovian
dynamics of a two-level system that is immersed in a thermal bath of
independent two-level systems. The Hamiltonian of the global system is given
b
\begin{equation}
H=-\frac{1}{2}\,\omega _{s}\,\sigma ^{z}-\frac{1}{2}\sum_{k}\omega
_{k}\,\sigma _{k}^{z}+\frac{1}{2}\sigma ^{x}\otimes \sum_{k}c_{k}\,\sigma
_{k}^{x},
\end{equation
where $\sigma ^{x}$ and $\sigma ^{z}$ are the Pauli operators, and $\omega
_{s}$ and $\omega _{k}$ are the frequencies of the two-level system and the
environment modes, respectively. The first term is the Hamiltonian of the
open system, the second term is the Hamiltonian of the environment and the
third term describes the interaction between the open system and the
environment.
In this example, assuming Ohmic dissipation, the spectral density of the
spin bath is expressed as~\cite{hanggi}
\begin{equation}
J\left( \omega \right) =2\pi \alpha \omega \exp \left( -\omega /\omega
_{c}\right) ,
\end{equation
where $\alpha $ is the dissipation coefficient and $\omega _{c}$ denotes the
cutoff frequency~($\omega _{c}/\omega _{s}\gg 1$). Below we specify
frequencies, temperatures and times using a standard unit frequency $\Delta
_{0}$. We first set $\omega _{s}/\Delta _{0}=1$, $\alpha =2\times 10^{-4}$,
\omega _{c}/\Delta _{0}=100$, and $\beta \Delta _{0}=1$. The frequency
spectrum in the region $\omega /\Delta _{0}\in \left[ 0.8\text{, }1.15\right]
$ is discretized into eight elements at frequencies $\omega _{k}/\Delta
_{0}=\left( 0.80+0.05k\right) $, with $k=0$, $1$, $\cdots 7$. The width of
each element is $\Delta \omega /\Delta _{0}=0.05$. The coupling coefficients
$c_{k}$ are determined using Eq.~($14$).
For this example, the analytical results for the relaxation rate and the
dephasing rate, $1/T_{1}$ and $1/T_{2}$, can be derived under the Markovian
approximation~\cite{cohen} as
\begin{equation}
\frac{1}{T_{1}}=\frac{1}{2}J\left( \omega =\omega _{s}\right) ,\text{ \ \ \
\ }\frac{1}{T_{2}}=\frac{1}{2T_{1}}.
\end{equation
The Markovian dynamics of the two-level system is simulated with the initial
state of the open system being set to the excited state $|e\rangle $ of the
two-level system. We set the unitary evolution time $\Delta _{0}\tau =30$,
and we obtain the evolution of the state of the open system. The relaxation
dynamics of the two-level system is shown in Fig.~$3$.
\begin{figure}[tbp]
\includegraphics[width=0.9\columnwidth, clip]{fig3.eps}
\caption{(Color online)~The evolution of the matrix element $\protect\rho
_{ee}$ of a two-level system in a spin bath, where $\protect\rho _{ee}$
denotes the diagonal element of the density matrix that describes the
population of the excited state. The frequency of the environment mode in
the region $\protect\omega /\Delta _{0}\in \left[ 0.8\text{, }1.15\right] $
is divided into $8$ elements with equal width and each element is
represented by a qubit. The Hamiltonian for the global system is given by
Eq.~($41$). The unitary evolution time $\Delta _{0}\protect\tau =30$. The
red solid line represents the analytical result for the evolution of
\protect\rho _{ee}$ with the initial state $|e\rangle $. The black square
dots represent the simulated results for the evolution of $\protect\rho
_{ee} $.}
\end{figure}
\begin{figure}[tbp]
\includegraphics[width=0.9\columnwidth, clip]{fig4.eps}
\caption{(Color online)~The relaxation rate of the two-level system as a
function of its frequency. The black squares represent the results obtained
from the simulation of the algorithm; the blue dotted line represents the
relaxation rate obtained by using the spectral density that is spanned by
eight $\protect\delta $-peaks in Eq.~($44$); and the red solid line
represents the relaxation rate obtained from the analytical spectral density
in Eq.~($42$). Obviously only the first two ones agree well with each other.}
\end{figure}
\begin{figure}[tbp]
\includegraphics[width=0.9\columnwidth, clip]{fig5.eps}
\caption{(Color online)~The relaxation rate of the two-level system as a
function of the frequency of the two-level system. The black square dots
represent the results obtained from the simulation of the algorithm using
the improved coupling coefficients as discussed in the text; the blue dotted
line represents the relaxation rate obtained by using the spectral density
that is produced by eight $\protect\delta $-peaks in Eq.~($44$) using the
improved coupling coefficients; and the red solid line represents the
relaxation rate obtained from the analytical spectral density in Eq.~($42$).}
\end{figure}
From Fig.~$3$, we can see that there is a clear discrepancy between the
relaxation rates obtained from the numerical simulation and the exact
results given by Eq.~($43$). This discrepancy is due to the fact that we
used Eq.~($14$) to determine the coupling coefficients $c_{k}$ even though
only eight $\delta $-peaks are used to represent the spectral density:
\begin{equation}
J\left( \omega \right) =\pi \sum_{k=1}^{k=8}c_{k}^{2}\;\delta \left( \omega
-\omega _{k}\right) .
\end{equation
Each $\delta $-peak has the analytical form~\cite{terhal
\begin{equation}
\delta \left( \frac{\omega -\omega _{0}}{\Delta _{0}},\tau \Delta
_{0}\right) =\frac{1-\cos [\tau \left( \omega -\omega _{0}\right) ]}{\pi
\tau \Delta _{0}\left[ (\omega -\omega _{0})/\Delta _{0}\right] ^{2}}.
\end{equation
In Eq.~($14$), a good approximation can be achieved when many $\delta
-peaks are used to represent the spectral density and the peaks cover a
sufficiently wide range of frequencies. The eight peaks used in our
simulation do not satisfy these conditions.
In order to further demonstrate the above explanation of the discrepancy, in
Fig.~$4$ we show the relaxation rate as a function of the frequency $\omega
_{s}$ of the two-level system. There we compare the numerical results
obtained from the simulation, the analytical results obtained from Eq.~($43
) with the spectral density given by Eqs.~($44$) and ($45$) and the coupling
coefficients determined through Eq.~($14$); and the exact results obtained
from Eq.~($43$) with the spectral density given by Eq.~($42$). From Fig.~$4$
we can see that the numerical results for the relaxation rate obtained from
the simulation fit very well the analytical results of the eight-peak
spectral density, while both have a systematic deviation from the exact
results.
\begin{table}[tbp]
\caption{Results for simulating the Markovian dynamics of a two level system
using different number of qubits to represent the environment to
sequentially simulate a total of eight dissipation channels. $N$ denotes the
number of qubits representing the environment modes. The exact result for
T_{1}$ is $T_{1}^{\text{exact}}=2/J(\protect\omega =\Delta _{0})$, and
T_{2}^{\text{exact}}=2T_{1}$. The results here are obtained using the
improved choice of the coupling coefficients as discussed in the text. The
results are essentially independent of $N$.}
\begin{center}
\begin{tabular}{ccccccccc}
\hline
$N$ & & $1$ & & $2$ & & $4$ & & $8$ \\ \hline
$T_{1}/T_{1}^{\text{exact}}$ & & $0.998$ & & $0.998$ & & $0.998$ & &
0.996$ \\ \hline
$T_{2}/T_{1}^{\text{exact}}$ & & $1.994$ & & $1.990$ & & $1.991$ & &
1.991$ \\ \hline
\end{tabular
\end{center}
\end{table}
This systematic deviation from the exact results can be eliminated by
plugging in the analytical form of the $\delta $-peaks as shown in Eq.~($45
) into Eq.~($44$) for the eight elements and then obtaining an improved
approximation for the coupling coefficients $c_{k}$. In Fig.~$5$, we show
the relaxation rate as a function of the frequency of the two-level system
using the improved choice of the coupling coefficients. One can see that in
the region around $\omega _{s}/\Delta _{0}=1.0$, the numerical results are
now in good agreement with the exact results.
We also perform the same simulation with the sequential application of the
different dissipation channels. In different simulations we use different
numbers of qubits to represent the environment. We use $1,2,4,$ or $8$
qubits to represent the environment using the improved choice of $c_{k}$.
The ratios $T_{1}/T_{1}^{\text{exact}}$ and $T_{2}/T_{1}^{\text{exact}}$ are
shown in Table~II. One can see that they are in good agreement with the
exact results and that the different simulations give essentially the same
results.
We do not perform any numerical calculations for the simulation of pure
dephasing using telegraph noise here, because such calculations would follow
closely similar calculations that have been performed in the literature in
theoretical studies of telegraph-noise-induced dephasing~(see, e.g., Ref.
\cite{ymg, fal}).
\section{Conclusion}
\label{con}
In this paper, we have presented an algorithm for simulating the Markovian
dynamics of an open quantum system. The algorithm takes as an input the
Hamiltonian of the open system, the operators through which the open system
interacts with the environment, the spectral density of the environment and
temperature. One therefore does not explicitly deal with the master equation
describing the dynamics. In the simulation, the environment is represented
by a set of ancilla qubits that are designed to have the same effect on the
open system as the simulated environment. We have also shown that different
dissipation channels can be implemented sequentially, thus reducing the
number of qubits needed to represent the environment. Pure dephasing also
allows a reduction in the number of needed qubits, since it can be induced
by a properly designed classical noise signal. The algorithm can also be
used to simulate non-Markovian dynamics.
In the present algorithm, the ancilla qubits play a rather passive role in
the sense that they only facilitate the dissipative dynamics of the system.
These ancilla qubits could be used in a more active role as probes or
actuators for the open quantum system. By monitoring the response of these
ancilla qubits as they interact with the open system, one could obtain the
energy spectrum of the system. Once the spectrum is known, the ancilla
qubits can also be used to provide or absorb any given amount of energy and
guide the system to any desired energy eigenstate. The details of this
algorithm will be presented elsewhere.
\begin{acknowledgements}
We thank L.-A. Wu for helpful discussions. We acknowledge partial
support from DARPA, Air Force Office for Scientific Research, the
Laboratory of Physical Sciences, National Security Agency, Army
Research Office, National Science Foundation grant No.~0726909,
JSPS-RFBR contract No.~09-02-92114, Grant-in-Aid for Scientific
Research~(S), MEXT Kakenhi on Quantum Cybernetics, and Funding
Program for Innovative R\&D on S\&T~(FIRST).
\end{acknowledgements}
|
\section{Introduction}
During the last few years there has been a growing body of literature targeting the algebraic structure of quantum field theory. A purely algebraic interpretation of Wick ordering \cite{fauser01} was provided giving insight to one of the most intriguing aspects of theoretical physics. On the other hand, this work provides more insight at the mathematical front to the structure of the tensor algebra of a vector space. Our interest lies on the mathematical basis of the approach and not its physical meaning. Circle products were introduced by Rota and Stein in \cite{rota94} in the context of super-symmetric Hopf algebras. Brouder in \cite{brouder09} carried out the same construction to derive similar quantities for the symmetric algebra. The above work leaves out the important question of whether or not this construction is carried out for the symmetric tensor algebra in an analogous manner, like Rota's and Stein's construction, in the context of the antisymmetric algebra. Though the two constructions are similar in spirit there is no clear indication of an underlying more general construction. Since both algebras are quotient algebras of the tensor algebra the most natural generalization is to create a unified construction to the tensor algebra and then to specialize to subspaces. More specifically, one expects to see a circle product and a Laplace pairing on the tensor algebra that specializes to the above constructions when transferred to the quotient algebras. In this work we achieve exactly this. We carry out this construction on the tensor algebra and re-derive the two specialized constructions. In this interpretation, our work is complementary to both these works and unifies them. We re-derive the symmetric algebra as a necessary ingredient of this approach. We also re-derive the circle products of Brouder and Rota as consequences of more general constructions.
\section{Definitions, notation and basic results}
In this section we fix notation and review some basic facts on the Hopf algebraic structure of the tensor algebra of a finite dimensional vector space which will be used in the next sections. We also define the joint tensor algebra of two vector spaces in duality and describe its Hopf algebraic structure. Finally, we revisit the definition of a circle product defined on the symmetric algebra generated by a finite dimensional vector space equipped with an inner product. For basic definitions of a Hopf algebra, reference \cite{podles98} provides an excellent short introduction. See also \cite{majid02,kassel95} which are the standard textbooks on the subject among others. For the definition of the tensor algebra of a vector space and symmetric tensors we refer to \cite{cartier07}.
\subsection{Review of basic facts on Hopf algebras}
Let $A$ be a unital algebra over the field $\mathbb{K}$. In general its dual is not an algebra. If there is a special additional structure that we will outline below then not only its dual becomes a unital algebra but this additional structure is also present to the dual.
\begin{definition}
Let $A$ be a unital algebra with unit $\mathbf{1}$. Suppose that following mappings are defined
\begin{enumerate}
\renewcommand{\labelenumi}{(\roman{enumi})}
\item $\Delta : A \to A \otimes A$ (the co-multiplication, homomorphism),
\item $S : A \to A \otimes A$ (the antipode, anti-homomorphism),
\item $\epsilon : A \to \mathbb{K}$ (the co-unit, homomorphism)
\end{enumerate}
with the additional properties
\begin{enumerate}
\renewcommand{\labelenumi}{(\roman{enumi})}
\item $(I \otimes \Delta) \Delta= (\Delta \otimes I) \Delta$ (the co-associativity),
\item $(\epsilon \otimes \Delta) \Delta= (\Delta \otimes \epsilon) \Delta = I$,
\item $(S \otimes I) \Delta= (I \otimes S) \Delta = \epsilon(\cdot) \mathbf{1}$ (the co-unit)
\end{enumerate}
where $I$ is the identity operator. We call the quadruple $\langle A, \Delta, S, \epsilon \rangle$ a \emph{Hopf algebra}.
\end{definition}
We use the well known Sweedler notation, (see \cite{kassel95}), for the co-multiplication of an element $a$ of the Hopf algebra $A$ as
\begin{equation*}
\Delta(a) = \sum_{(a)} a_{(1)} \otimes a_{(2)}\\
\end{equation*}
We call a Hopf algebra co-commutative when
\begin{equation*}
\Delta(a) = \sum_{(a)} a_{(1)} \otimes a_{(2)} = \sum_{(a)} a_{(2)} \otimes a_{(1)} \,.
\end{equation*}
For repeated application of $\Delta$, we use the standard notation,
\begin{equation*}
(I\bigotimes \Delta) \Delta(a)= \sum_{(a)} a_{(1)} \otimes a_{(2,1)} \otimes a_{(2,2)}\,
\end{equation*}
and
\begin{equation*}
(\Delta\bigotimes I) \Delta(a)= \sum_{(a)} a_{(1,1)} \otimes a_{(1,2)} \otimes a_{(2)}\,.
\end{equation*}
We call a Hopf algebra co-associative when the two quantities above are equal.
A standard result \cite{majid02} states that given two Hopf algebras $A_{1},A_{2}$ we can turn also $A_{1} \bigotimes A_{2}$ into a Hopf algebra. We summarize now the details. Since $A_{1},A_{2}$ are unital algebras their tensor product is again an algebra with product (defined for elementary tensors)
\begin{equation*}
(a_{1} \otimes a_{2} ) (a^{'}_{1} \otimes a^{'}_{2} )= a_{1} a^{'}_{1} \otimes a_{2}a^{'}_{2}\,.
\end{equation*}
Let us define for $a \in A_{1}$ and $b \in A_{2}$
\begin{enumerate}
\renewcommand{\labelenumi}{(\roman{enumi})}
\item $(\Delta_{1} \otimes \Delta_{2})(a_{1} \otimes a_{2} ) = \sum_{(a),(b)} (a_{(1)i} \otimes b_{(1)}) \otimes (a_{(2)} \otimes b_{(2)})$\,,
\item $(S_{1} \otimes S_{2}) (a_{1} \otimes b ) = S_{1}(a)S_{2}(b)$\,,
\item $(\epsilon_{1} \otimes \epsilon_{2}) (a \otimes b ) = \epsilon_{1}(a) \epsilon_{2}(b)$\,,
\end{enumerate}
\begin{theorem}\label{thm:tensorhopf}
The quadruple $\langle A_{1}\bigotimes A_{2},\Delta_{1} \otimes \Delta_{2}, S_{1} \otimes S_{2}, \epsilon_{1} \otimes \epsilon_{2} \rangle$ is a Hopf algebra.
\end{theorem}
\subsection{Review of basic facts on tensor algebra}
Let $U$ be a vector space. The letter $x$ possibly indexed by some subscript
will usually denote a generic element of U. Its tensor algebra $\mathcal{T}(U)$ is defined as the direct sum of the $n$-fold tensor products of $U$ with itself $\bigotimes^{n}U$, namely
\[
\mathcal{T}(U) = \bigoplus_{n=0}^{\infty} \bigotimes ^{n} U\,.
\]
We define $\bigotimes^{0}U =\mathbb{K}$, the field of $U$. The unit element of the field when we view it as a member of the tensor algebra will be denoted by ${\bm 1}$. For our purposes we restrict our attention to the complex field $\mathbb{C}$ or to the real field $\mathbb{R}$. One can also view the tensor algebra as the free non-commutative algebra generated by the elements of $U$. Since $\bigotimes^{n}U$ is the span of elementary tensors $x_{1}\otimes \cdots \otimes x_{n}$ where $x_{i} \in U$, we define the product in the algebra between two elements of $\bigotimes^{n}U$ and $\bigotimes^{m}U$ as
\begin{equation*}
(x_{1}\otimes \cdots \otimes x_{n}) \cdot (x^{'}_{1}\otimes \cdots \otimes x^{'}_{m})=x_{1}\otimes \cdots \otimes x_{n} \otimes x^{'}_{1}\otimes \cdots \otimes x^{'}_{m}\,,
\end{equation*}
and extend with linearity to the whole algebra. It can be proved that with this multiplication the tensor algebra is a unital associative algebra. Because of the direct sum construction of the tensor algebra and the definition of the product, it is a graded algebra.
Let $U$ and $V$ be two vector spaces in duality via the bilinear form $\langle \cdot , \cdot \rangle$ , both over the same field $\mathbb{K}$. The letter $x$ possibly indexed by some subscript will usually denote a generic element of U while the letter $y$ indexed by some subscript will denote a generic element of V. Their {\it joint tensor algebra} is defined as the pair consisting of the duality and the tensor product
\[
\mathcal{T} (U, V)= \mathcal{T}(U) \bigotimes \mathcal{T}(U)\,.
\]
The term ``{\it joint tensor algebra}" is not used in the literature. However, since our work pertains to the properties of this specific tensor product we decided to give it a name encoding its special nature. It is easy to see that we can equip it with a suitable product to become a unital non-commutative associative algebra. We define it first for elementary tensors, that is
\begin{equation}
(u_{1} \otimes v_{1}) \cdot (u_{2} \otimes v_{2}) = (u_{1} \cdot u_{2}) \otimes (v_{1} \cdot v_{2})\,,
\end{equation}
where $u_{i} \in \bigotimes^{n_{i}}U$ and $v_{j} \in \bigotimes^{m_{j}}V$. We can extend it with linearity to the joint tensor algebra. It is well-known that the bilinear form can be extended to the joint tensor algebra.
The extension takes place by first defining it for elementary tensors
\begin{equation*}
\langle (x_{1}\otimes \cdots \otimes x_{n}) , (y_{1}\otimes \cdots \otimes y_{m})\rangle =\delta_{m,n} \langle x_{1}, y_{1} \rangle \cdots \langle x_{n}, y_{m} \rangle\,.
\end{equation*}
A typical element of the joint tensor algebra can be written as
\begin{equation*}
a= \sum_{i} u_{i} \otimes v_{i}\,,
\end{equation*}
where $u_{i} \in \bigotimes^{n_{i}} U$ and $v_{i} \in \bigotimes^{m_{i}} V$.
For this reason we define the bilinear form for two elements in the joint tensor algebra $a,b$ as
\begin{equation*}
\langle a , b \rangle := \langle \sum_{i} u_{i} \otimes v_{i} , \sum_{j} u^{'}_{j} \otimes v^{'}_{j} \rangle=\sum_{i} \sum_{j} \langle u_{i} , v^{'}_{j} \rangle \langle u^{'}_{j} , v_{i} \rangle\,.
\end{equation*}
We will specialize the previous definitions to an important quotient algebra of the tensor algebra, namely the symmetric tensor algebra. On the tensor algebra of the vector space $U$ we can define the symmetrization operator first for elementary tensors as
\begin{equation}\label{symm:def}
{\rm Symm}( x_{1} \otimes \cdots \otimes x_{n}) := \frac{1}{n!} \sum_{p \in Per_{n}} x_{p(1)} \otimes \cdots \otimes \cdots x_{p(n)}
\end{equation}
and extend it by linearity on the whole tensor algebra. We collect in the next theorem some useful facts on the symmetrization operator \cite{rodrigues07,ryan80}.
\begin{theorem}\label{thm:symm}
Let $U$ be a vector space. Then, the symmetrization operator defined by (\ref{symm:def}) is a projection. The projection of the tensor algebra is denoted by $\mathcal{T}_{S}(U)={\rm Symm}(\mathcal{T}(U))$ and is called {\it the symmetric algebra generated by $U$}. It is itself a commutative unital algebra with multiplication defined as
\begin{equation*}
u \cdot v = {\rm Symm}(u \otimes v)\,.
\end{equation*}
It can be written as the span of the elementary symmetric tensors
\begin{equation*}
x_{1} \cdots x_{n}={\rm Symm}( x_{1} \otimes \cdots \otimes x_{n} ) = \frac{1}{n!} \sum_{p \in Per_{n}} x_{p(1)} \otimes \cdots \otimes \cdots x_{p(n)}\,.
\end{equation*}
The algebra can be viewed as the free commutative algebra generated by the vector space $U$. It can also be viewed as the quotient of the tensor algebra by the two-sided ideal $I$ generated by the set
\begin{equation*}
S= \{ (x \otimes y - y \otimes x ) \}_{x,y \in U}\,.
\end{equation*}
\end{theorem}
The symmetric algebra comes with a natural grading:
\[
\mathcal{T}_{S}(U) = \bigoplus _{t=0}^{\infty} \bigodot^{t} U\qquad\text{where ${\rm Symm}(\bigotimes^{t} U)=\bigodot^{t} U$}.
\]
Important elements of the symmetric tensor algebra are the \emph{homogeneous} elements, namely symmetric tensor powers of a single vector. A crucial algebraic result for symmetric tensors is the polarization formula \cite{ryan80} that allows us to rewrite each grade as the span of homogeneous elements. This means
\begin{equation*}
\bigodot^{t} U = span (x^{t} )_{x \in U}\,.
\end{equation*}
The polarization formula is very useful and for this reason we repeat it here for completeness. The formula reads as follows
\begin{equation}\label{polar:ref}
x_{1} x_{2} \cdots x_{t} = \frac{1}{2^{t} t! } \sum_{\epsilon_{i}=\pm 1} \epsilon_{1} \cdots \epsilon_{t}(\epsilon_{1}x_{1}+\epsilon_{2}x_{2} + \cdots + \epsilon_{t}x_{t})^{t}\,.
\end{equation}
We note here that this formula is required when proving that the space of homogeneous polynomials over $U$ is isomorphic to the space of symmetric multi-linear forms of the same order. The symmetric algebra is not the only one with the above properties. The anti-symmetrization operator ${\rm ASymm}$ is defined on the tensor algebra $\mathcal{T}(U)$ as
\begin{equation}\label{asymm:def}
{\rm ASymm}( x_{1} \otimes \cdots \otimes x_{n}) := \frac{1}{n!} \sum_{p \in Per_{n}} sign(p) (x_{p(1)} \otimes \cdots \otimes \cdots x_{p(n)})\,,
\end{equation}
where $sign(p)$ is the sign of the permutation, see \cite{rodrigues07}. We can extend the mapping via linearity to the whole tensor algebra.
We collect in the next theorem some useful facts on the anti-symmetrization operator \cite{rodrigues07}.
\begin{theorem}\label{thm:asymm}
Let $U$ be a vector space. The anti-symmetrization operator defined by (\ref{asymm:def}) is a projection. The projection of the tensor algebra is denoted by $\mathcal{T}_{A}(U)={\rm ASymm}(\mathcal{T}(U))$ and is called the
{\it antisymmetric algebra generated by $U$}. It is itself a non-commutative unital algebra with multiplication defined as
\begin{equation*}
u \wedge v = {\rm ASymm}(u \otimes v)\,.
\end{equation*}
It can be written as the span of the elementary antisymmetric tensors
\begin{equation*}
x_{1} \wedge \cdots \wedge x_{n}={\rm ASymm}( x_{1} \otimes \cdots \otimes x_{n} ) = \frac{1}{n!} \sum_{p \in Per_{n}} sign(p) (x_{p(1)} \otimes \cdots \otimes \cdots x_{p(n)})\,.
\end{equation*}
The algebra can be viewed as the quotient of the tensor algebra by the two-sided ideal $I$ generated by the set
\begin{equation*}
S= \{ (x \otimes y + y \otimes x ) \}_{x,y \in U}\,.
\end{equation*}
For this reason $u \wedge v = (-1)^{pq} v \wedge u$, when both $u,v$ are elementary antisymmetric tensors of non-zero grades $p$ and $q$ respectively.
\end{theorem}
The anti-symmetric algebra comes with a natural grading:
\[
\mathcal{T}_{A}(U) = \bigoplus _{t=0}^{\infty} (\bigwedge^{t} U)\qquad \text{where $A{\rm Symm}(\bigotimes^{t} U)=\bigwedge^{t} U$}.
\]
Contrary to the symmetric algebra, the anti-symmetric algebra is finite dimensional when the vector space is finite dimensional. It also inherits the inner product from the tensor algebra when it is equipped with one.
\subsection{Review of the Hopf algebraic structure of tensor algebra}
The tensor algebra has more structure than its grading. This structure displays itself with the ability to equip it with suitable operations to turn it to a Hopf algebra. To this end we first define the co-multiplication $\Delta : \mathcal{T}(U) \to \mathcal{T}(U) \bigotimes \mathcal{T}(U)$ of the tensor algebra. The co-multiplication should be a homomorphism and because our algebra is freely generated by the elements of $U$ it is enough to define it on the elements of $U$. For this we make the following two definitions
\begin{enumerate}
\renewcommand{\labelenumi}{(\roman{enumi})}
\item $\Delta(\mathbf{1}) = \mathbf{1} \otimes \mathbf{1}$,
\item $\Delta(x) = \mathbf{1} \otimes x + x \otimes \mathbf{1}$\qquad where $x\in U$.
\end{enumerate}
For later convenience let $T_{n}$ be the set $\{ 1,\cdots , n \}$. For a subset $S$ of $T_{n}$ with $m$ elements and an elementary tensor $u_{n}=x_{1} \otimes \cdots \otimes x_{n}$ of grade $n$ we write $u_{S}=x_{s_{1}} \otimes \cdots \otimes x_{s_{m}}$, where the elements of $S$ are arranged in increasing order. We can easily see that
\begin{equation} \label{tree:delta}
\Delta(x_{1} \otimes \cdots \otimes x_{n}) = \sum_{S \subseteq T_{n}} u_{S} \otimes u_{T_{n} \setminus S}\,.
\end{equation}
For the definition of the antipode $S : \mathcal{T}(U) \to \mathcal{T}(U)$ we similarly define (because it is a homomorphism)
\begin{enumerate}
\renewcommand{\labelenumi}{(\roman{enumi})}
\item $S(\mathbf{1}) = \mathbf{1}$,
\item $S(x) = -x$\qquad where $x\in U$.
\end{enumerate}
Finally, due to its homomorphic property the co-unit of the algebra $\epsilon : \mathcal{T}(U) \to \mathbf{K}$ can be defined on the elements of $U$ as
\begin{enumerate}
\renewcommand{\labelenumi}{(\roman{enumi})}
\item $\epsilon(\mathbf{1}) = 1$,
\item $\epsilon(x) = 0$\qquad where $x \in U$.
\end{enumerate}
Now we are ready to state a useful well-known theorem concerning the Hopf algebra structure of the tensor algebra \cite{cartier07}.
\begin{theorem}
Given a vector space $U$ with corresponding tensor algebra $\mathcal{T}(U)$, the quadruple $\langle \mathcal{T}(U) , \Delta, S , \epsilon \rangle $ is a co-commutative and co-associative Hopf algebra.
\end{theorem}
The Hopf algebra structure of the tensor algebra carries over to its symmetric algebra.
\begin{theorem}
Let $U$ be a vector space. The symmetric tensor algebra $\mathcal{T}_{S}(U)$ generated by $U$ can be equipped with the structure of a co-commutative and co-associative Hopf algebra as
\begin{enumerate}
\renewcommand{\labelenumi}{(\roman{enumi})}
\item $\Delta_{symm} (a) = ({\rm Symm} \otimes {\rm Symm}) \Delta(a)$\,,
\item $S_{symm}(a) = {\rm Symm}(S(A)) $\,,
\item $\epsilon_{symm} (a) = \epsilon(a)$\,,
\end{enumerate}
where $a \in \mathcal{T}_{S}(U)$.
\end{theorem}
For homogeneous elements the co-multiplication takes an elegant form
\begin{equation} \label{delta:homog}
\Delta(x^{t}) = \sum_{k=0}^{t} {t \choose k} x^{t-k} \otimes x^{k}\,.
\end{equation}
The Hopf algebra structure of the tensor algebra also carries over to its antisymmetric algebra.
\begin{theorem}
Let $U$ be a vector space. The antisymmetric tensor algebra $\mathcal{T}_{A}(U)$ generated by $U$ can be equipped with the structure of a co-commutative and co-associative Hopf algebra as
\begin{enumerate}
\renewcommand{\labelenumi}{(\roman{enumi})}
\item $\Delta_{{\rm asymm}} (a) = ({\rm ASymm} \otimes {\rm ASymm}) \Delta(a)$,
\item $S_{{\rm asymm}}(a) = {\rm ASymm}(S(A)) $,
\item $\epsilon_{{\rm asymm}} (a) = \epsilon(a)$,
\end{enumerate}
where $a \in \mathcal{T}_{A}(U)$.
\end{theorem}
As far as the joint algebra is concerned, observe that by theorem \ref{thm:tensorhopf} it can be equipped with a Hopf algebra structure. The joint algebra and its Hopf algebraic properties will be our focus in the next section. We conclude this section with some remarks.
\begin{remark}
The element $\Delta{a}$, where $a \in \mathcal{T}(U)$, can be identified with an element of $\mathcal{T}(U)$ as follows
\begin{equation}
\Delta(a) = \sum_{(a)} a_{(1)} \bigotimes a_{(2)}= \sum_{(a)} a_{(1)} \otimes a_{(2)}\,.
\end{equation}
\end{remark}
\begin{remark}
When $U$ and $V$ two vector spaces in duality. Let $a \in \mathcal{T}(U)$ and $b \in \mathcal{T}(V)$. We have the duality relations
\begin{enumerate}
\renewcommand{\labelenumi}{(\roman{enumi})}
\item $\langle {\rm Symm}(a) , b \rangle = \langle a , {\rm Symm}(b) \rangle = \langle {\rm Symm}(a) ,
{\rm Symm}(b) \rangle$,
\item $\langle {\rm ASymm}(a) , b \rangle = \langle a , {\rm ASymm}(b) \rangle=\langle {\rm ASymm}(a) ,
{\rm ASymm}(b) \rangle$.
\end{enumerate}
\end{remark}
\section{The square product}
In this section we aim to define a Laplace pairing on the joint tensor algebra. We need it in order to define a circle product in the sense of \cite{brouder09} on the joint tensor algebra, which we call it {\it square product}.
Once our goal is completed, we will proceed to derive the ordinary circle product of the symmetric algebra defined in \cite{brouder09} as a symmetrization of the square product of the tensor algebra. A Laplace pairing on the joint tensor algebra is a symmetric bilinear form $( \cdot | \cdot )$ which should satisfy the important identity (see \cite{brouder09})
\begin{equation}\label{laplace:prop}
( a | b c) = \sum_{(a)} (a_{(1)} | b ) ( a_{(2)} | c)\qquad \text{where $a, b, c\in\mathcal{T}(U) \bigotimes \mathcal{T}(V)$}.
\end{equation}
Due to its recursive definition, in order to make it compatible with the duality already defined on the joint tensor algebra we require
\begin{enumerate}
\renewcommand{\labelenumi}{(\roman{enumi})}
\item $(\mathbf{1} | \mathbf{1} ) =1$
\item $( 1 \otimes y | x \otimes 1 ) = ( x \otimes 1 | 1 \otimes y) = \langle x , y \rangle$\qquad where $x \in U$ and $y \in V$
\item $ (x_{1} \otimes 1 | x_{2} \otimes 1 ) = ( 1 \otimes y_{1} | 1 \otimes y_{2}) = 0$\,\qquad where $x_{1},x_{2} \in U$ and $y_{1},y_{2} \in V$
\item $(x \otimes 1 | 1 \otimes 1) = (1 \otimes 1 | x \otimes 1) = (1 \otimes 1 | 1 \otimes y)= (1 \otimes y | 1 \otimes 1) = 0$\qquad\qquad where $x \in U$ and $y \in V $.
\end{enumerate}
One important question is whether fixing these `` initial conditions" or the recursion, is enough to define the Laplace pairing. Moreover, we can also question the existence of the Laplace pairing. Our next theorem gives a positive answer.
\begin{theorem}\label{laplace:exists}
A Laplace pairing defined on the joint tensor algebra generated by two vector spaces $U$ and $V$ in duality compatible with the duality exists and is unique. Moreover, it is symmetric
\[
( a | b ) = (b | a) \quad\text{where $a, b\in\mathcal{T}(U)\bigotimes \mathcal{T}(V)$}.
\]
\end{theorem}
We split the proof in several steps.
\begin{lemma}
Under the hypotheses of theorem \ref{laplace:exists}, if such a Laplace pairing exists then for $m\neq n$,
\begin{equation}\label{tensor:orthog}
( x_{1} \otimes \cdots \otimes x_{n} | y_{1} \otimes \cdots \otimes y_{m} ) =0\,.
\end{equation}
\end{lemma}
\begin{proof}
We will proceed with double induction. We prove first the important fact that for $n > 0$
\begin{equation}\label{one:ortho}
( x_{1} \otimes \cdots \otimes x_{n} | \mathbf{1} ) =0\,.
\end{equation}
We argue by induction. For $n=1$ it is true because of the compatibility conditions. Suppose now that this property holds for $n$. Then, by the splitting property in equation (\ref{laplace:prop}) we have
\[
( x_{1} \otimes \cdots \otimes x_{n} \otimes x_{n+1} | \mathbf{1} ) = ( x_{1} \otimes \cdots \otimes x_{n} | \mathbf{1} ) ( x_{n+1} | \mathbf{1} ) = 0
\]
and so the property holds for $n+1$. In a similar fashion we prove that
\[
( \mathbf{1} | y_{1} \otimes \cdots \otimes y_{m} ) =0\,.
\]
We have proved relation (\ref{tensor:orthog}) when $m=0 < n \leq 1$. Suppose now that the orthogonality holds for every pair $(n, m)$, $N\geq n > m$. We need to show the orthogonality for every pair $(n, m)$, $N+1\geq n > m$. By the induction hypothesis it is enough to prove the assertion for the pairs $(N+1, m)$ with $N+1> m$. By the properties of the Laplace pairing (\ref{laplace:prop}) and the co-multiplication (\ref{tree:delta}),
\[
( x_{1} \otimes \cdots \otimes x_{N+1} | y_{1} \otimes \cdots \otimes y_{m} ) =
\sum_{Q \subseteq T_{m}} ( x_{1} \otimes \cdots \otimes x_{N} | y_{Q} )( x_{N+1} | y_{T_{m} \setminus Q} )\,.
\]
Observe that the only surviving terms occur when $|Q|=N$ and $|T_{m} \setminus Q| = 1$. But then $m=N+1$ which is not possible.
\end{proof}
\begin{lemma}
Under the hypotheses of theorem \ref{laplace:exists}, if such a Laplace pairing exists
\begin{equation}
( x_{1} \otimes \cdots \otimes x_{n} | y_{1} \otimes \cdots \otimes y_{n} ) = ( x_{1} \otimes \cdots \otimes x_{n} | \sum_{p \in Per(n)} y_{p(1)} \otimes \cdots \otimes y_{p(n)})\,.
\end{equation}
\end{lemma}
\begin{proof}
Once more we argue by induction. For $n=0, 1$ the theorem obviously holds. If the theorem holds for $n$, we need prove it for $n+1$. First observe that by using (\ref{tree:delta}) we have
\[
( x_{1} \otimes \cdots \otimes x_{n+1} | y_{1} \otimes \cdots \otimes y_{n+1} ) = ( x_{1} \otimes \cdots \otimes x_{n} | \sum_{Q \subseteq T_{n+1} } y_{Q} ) ( x_{n+1} | y_{T_{n+1} \setminus Q})\,.
\]
The RHS can be written as
\[
\sum_{k=1}^{n+1} ( x_{1} \otimes \cdots \otimes x_{n} | \sum_{ Q \subseteq ( T_{n+1}\setminus \{ k\}) } y_{Q} ) ( x_{n+1} | y_{k} )\,.
\]
Let $M_{k}$ be the totality of ``$1-1$" onto functions $g : T_{n+1}\setminus\{ k\} \to T_{n}$. By the induction hypothesis we can rewrite the above summations as
\[
\sum_{k=1}^{n+1} ( x_{1} \otimes \cdots \otimes x_{n} \otimes x_{n+1} | \sum_{ g \in M_{k} } y_{g^{-1}(T_{n})} \otimes y_{k} )\,.
\]
But every such $g$ is derived uniquely by restricting a $T_{n+1}$ permutation on $T_{n+1}\setminus\{ k\}$. For this reason
\[
( x_{1} \otimes \cdots \otimes x_{n} | y_{1} \otimes \cdots \otimes y_{n} ) = ( x_{1} \otimes \cdots \otimes x_{n+1} | \sum_{ g \in Per(n+1) } y_{g^{-1}(T_{n+1})} )
\]
which proves the lemma.
\end{proof}
\begin{lemma}
We assume the hypotheses of theorem \ref{laplace:exists}. Suppose that such a Laplace pairing exists. Moreover, let $u_{i} \in \bigotimes^{n_{i}}U$ and $v_{j} \in \bigotimes^{m_{j}}V$. Then,
\begin{equation} \label{prod:prod}
( u_{1} \otimes v_{1} | u_{2} \otimes v_{2}) = ( u_{1} | u_{2}) (v_{1} | v_{2})\,.
\end{equation}
\end{lemma}
\begin{proof}
By the splitting property (\ref{laplace:prop}) and the fact that $\mathcal{T}(U)\bigotimes\mathcal{V}$ is a tensor product of Hopf algebras we have
\begin{align*}
( u_{1} \otimes v_{1} | u_{2} \otimes v_{2}) &= \sum_{(u_{1}), (v_{1})} ( u_{1,(1)} \otimes v_{1,(1)} | u_{2}) ( u_{1,(2)} \otimes v_{1,(2)} | v_{2})\\
&= \sum_{ \substack { (u_{1}), (v_{1}) , \\ (u_{2}), (v_{2})} } ( u_{1,(1)} | u_{2,(1)}) )(v_{1,(1)} | u_{2,(2)}) ( u_{1,(2)} | v_{2,(1)} ( v_{1,(2)} | v_{2,(2)})\,.
\end{align*}
By the previous discussion $u_{1,(1)} = u_{2,(1)}=1$ and $v_{1,(2)} = v_{2,(2)}=1$. But then
$u_{1,(2)}=u_{1}$ and $u_{2,(2)} = u_{2}$. In the same way $v_{1,(1)}=v_{1}$ and $v_{2,(1)} = v_{2}$. For this reason (\ref{prod:prod}) holds.
\end{proof}
\begin{lemma}
We assume the hypotheses of theorem \ref{laplace:exists} and the existnce of the aforementioned Laplace pairing. Suppose that $a,b \in \mathcal{T}(U) \bigotimes \mathcal{T}(V)$. We can write
$a=\sum_{i,j} u_{i} \otimes v_{j} $, with $u_{i} \in \bigotimes^{i}U , v_{j} \in \bigotimes^{j}V$
and $b=\sum_{k,l} u^{'}_{k} \otimes v^{'}_{l} $, with $u^{'}_{k} \in \bigotimes^{k}U , v^{'}_{l} \in \bigotimes^{l}V$. Then, we have
\[
(a | b )= \sum_{i,j} \sum_{k,l} i! k! \langle {\rm Symm}(u_{i}), {\rm Symm}(v^{'}_{l})\rangle \langle
{\rm Symm}(u^{'}_{k}), {\rm Symm}(v_{j})\rangle\,.
\]
\end{lemma}
\begin{proof}
This is the result of the previous discussion and of theorem ~\ref{thm:symm}.
\end{proof}
Now we have the required form of the pairing if it exists. In the course of the previous discussion we have also proven its uniqueness. In order to conclude the proof of our main theorem (theorem \ref{laplace:exists}) we have to verify that the above defined pairing has the splitting property.
{\it Proof of Theorem \ref{laplace:exists}}. Because of the previous discussion, it suffices to prove the theorem when
$a=x_{1} \otimes \cdots \otimes x_{n}$, $b=y_{1} \otimes \cdots \otimes y_{m}$ and $c=y^{'}_{1} \otimes \cdots \otimes y^{'}_{m^{'}}$. Then,
\[
(x_{1} \otimes \cdots \otimes x_{n} | y_{1} \otimes \cdots \otimes y_{m} \otimes y^{'}_{1} \otimes \cdots \otimes y^{'}_{m^{'}} )= n! \langle x_{1} \cdots x_{n} , y_{1} \cdots y_{m} y^{'}_{1} \cdots y^{'}_{m^{'}} \rangle\,.
\]
Because of the symmetrization and polarization formula (\ref{polar:ref}), it is enough to prove the theorem when $a=x^{n}$, $b=y^{m}$ and $c=y^{' m^{'}}$. Then, we have
\[
(x^{n} | y^{m} \otimes y^{'}_{1} \otimes y^{' m^{'}} )= n! \langle x , y \rangle ^{m} \langle x , y^{'} \rangle ^{m^{'}}\,.
\]
On the other hand, by (\ref{delta:homog})
\[
\sum_{(a)} (a_{(1)} | b ) (a_{(2)} | c) = \sum_{k} {n \choose k} k! (n-k)! \langle x , y \rangle ^{m} \langle x , y^{'} \rangle ^{m^{'}}\,.
\]
The symmetric nature of the pairing follows from its definition and the self-adjointness properties of the symmetrization operator. The proof of Theorem \ref{laplace:exists} is complete.
\begin{remark}
The above theorem proves the striking fact that when $U=V$ with a self-duality, the derived pairing is the same as the one derived for the symmetric algebra in \cite{brouder09}.
\end{remark}
\begin{remark}
Another striking fact is that the Laplace pairing works only with symmetric tensors. Without defining a symmetric tensor algebra we re-derived the symmetrization operator only from pure algebraic facts.
\end{remark}
Now we are in a position to present the definition of the square product.
\begin{definition}
Let $U$ and $V$ be two vector spaces in duality via the bi-linear form
$\langle \cdot , \cdot \rangle$. We define the square product for two elements
$a, b\in\mathcal{T}(U)\bigotimes\mathcal{T}(V)$ as
\begin{equation}\label{square:def}
a \square b = \sum_{(a),(b)} ( a_{(1)} | b_{(1)}) a_{(1)} \otimes b_{(2)}\,.
\end{equation}
\end{definition}
\begin{remark}
If the vector space $U$ is already in duality with itself, the joint tensor algebra is the ordinary tensor algebra with square product defined exactly as above.
\end{remark}
\begin{remark}
If the vector space $U$ is already in duality with itself, and if we restrict the product on elements of the symmetric tensor algebra, as we shall see later
\[
{\rm Symm}(a \square b )=a \circ b\qquad\text{where $a, b\in \mathcal{T}_{S}(U)$}.
\]
In the above expression $\circ$ is the circle product defined for the specific Laplace pairing (the bilinear form).
\end{remark}
It would be convenient if the joint tensor algebra could be turned into a unital associative algebra like in the case of the circle product. We cannot require commutativity for obvious reasons. The content of the next theorem fills this gap.
\begin{theorem}
Let $U$ and $V$ be two vector spaces in duality via the bi-linear form $\langle \cdot , \cdot \rangle$. The joint tensor algebra equiped with the square product is a unital associative algebra.
\end{theorem}
\begin{proof}
Let $a,b,c \in \mathcal{T}(U,V)$. We want to prove that
\[
(a \square b) \square c= a \square ( b \square c).
\]
The LHS expression can be computed to be
\[
(a \square b) \square c= \sum_{(a),(b),(c)} (a_{(1)} | b_{(1)}) ( a_{(2,1)} b_{(2,1)} | c_{(1)}) a_{(2,2)} b_{(2,2)} c_{(2)}
\]
while the RHS is equal to (due to co-associativity)
\[
a \square ( b \square c)= \sum_{(a),(b),(c)} ( a_{(1)} | b_{(2,1)} c_{(2,1)} ) (b_{(1)} | c_{(1)}) a_{(2)} b_{(2,2)} c_{(2,2)}\,.
\]
We now use the spliting property of the pairing (\ref{laplace:prop}). The above two expressions can be written as
\begin{equation*}
(a \square b) \square c= \sum_{(a),(b),(c)} (a_{(1)} | b_{(1)}) ( a_{(2,1)} | c_{(1,1)}) ( b_{(2,1)} | c_{(1,2)}) a_{(2,2)} b_{(2,2)} c_{(2)}
\end{equation*}
and
\[
a \square ( b \square c)= \sum_{(a),(b),(c)} ( a_{(1,1)} | b_{(2,1)} ) ( a_{(1,2)} | c_{(2,1)} ) (b_{(1)} | c_{(1)}) a_{(2)} b_{(2,2)} c_{(2,2)} .
\]
respectively. Because of the co-associativity and co-commutativity of the Hopf algebras we have used, the above expressions are the same. Finally, as we can easily show, the unit of the algebra is the ordinary unit.
\end{proof}
A weak commutativity property that generalizes the one given in \cite{brouder09}, is stated in the next lemma.
\begin{lemma}
Let $a,b,c \in \mathcal{T}(V,U)$ and let the assumptions of the main theorem be fulfilled. Then, the following identity holds
\begin{equation} \label{laplace:perm}
( a \square b | c )= ( a | b \square c )\,.
\end{equation}
\end{lemma}
\begin{proof}
The proof is a matter of simple algebraic manipulations. We have
\begin{align*}
(a\square b | c ) &= \sum_{(a),(b)} ( (a_{(1)} | b_{(1)}) a_{(2)} \otimes b_{(2)} | c)\\
&= \sum_{(a),(b)} (a_{(1)} | b_{(1)}) ( a_{(2)} \otimes b_{(2)} | c)\\
&= \sum_{(a),(b),(c)} (a_{(1)}| b_{(1)}) (a_{(2)} | c_{(1)}) ( b_{(2)} | c_{(2)})\\
&= \sum_{(b),(c)} (a | b_{(1)} \otimes c_{(1)}) (b_{(2)} | c_{(2)})\\
&= ( a | b \square c )\,.
\end{align*}
\end{proof}
Finally, we can write the square product via the ordinary product just like in \cite{brouder09}.
\begin{theorem}
Let $a,b,c \in \mathcal{T}(V,U)$ and let the assumptions of the main theorem be fulfilled. Then, the following identity holds
\begin{equation}
a \otimes b = \sum_{(a),(b)} ( S(a_{(1)}) | b_{(1)} ) a_{(2)} \square b_{(2)}\,,
\end{equation}
where $S$ is the antipode.
\end{theorem}
\begin{proof}
The proof is the same as in \cite{brouder09} for the circle product.
\end{proof}
Now we restrict our attention to the case where $U$ is equipped with an inner product and $V=U$. We also restrict our attention to the symmetric tensor algebra. In this case the above complex calculations can be considerably simplified. We will use the notation in \cite{brouder09}.
\begin{lemma}
Let $u,v \in \mathcal{T}_{S}(V)$, then
\begin{equation}
{\rm Symm}( u \square v ) = u \circ v\,.
\end{equation}
\end{lemma}
\begin{proof}
By the definition of the square product we have
\[
{\rm Symm}( u \square v ) = \sum_{(u),(v)} {\rm Symm}( (u_{(1)} | v_{(1)}) u_{(2)} \otimes v_{(2)}
= \sum_{(u),(v)} (u_{(1)} | v_{(1)}) {\rm Symm}(u_{(2)} \otimes v_{(2)})\,.
\]
By the definition of the circle product observe that the RHS is equal to
\begin{equation}\label{symmcirc:def}
\sum_{(u),(v)} (u_{(1)} | v_{(1)}) {\rm Symm}(u_{(2)} \otimes v_{(2)} )=u \circ v\,.
\end{equation}
\end{proof}
\section{Automorphic property of the circle and square products}
We restrict ourselves to a vector space $U$ equipped with a self-duality. We will prove that the symmetric algebra equipped with the circle product in (\ref{symmcirc:def}) is homomorphically equivalent to the symmetric algebra equipped with the ordinary product.This is not very clearly stated in \cite{brouder09} and we aim to fill this gap. We prove this property and use it as a model to extend it to the case of the square product. Let us define the mapping $\phi$ with domain and codomain equal to the symmetric tensor algebra $ \mathcal{T}_{S} (V)$ first on elementary symmetric tensors as
\begin{enumerate}
\renewcommand{\labelenumi}{(\roman{enumi})}
\item $\phi(\mathbf{1} )=\mathbf{1}$
\item $\phi(x^{n})=x \circ \cdots \circ x$\qquad\text{where $x \in U$}
\end{enumerate}
and then extend it over the symmetric algebra by linearity.
\begin{lemma}
For $u,v \in \mathcal{T}_{S}(V)$, the above mapping $\phi$ is a homomorphism as the property
\begin{equation*}
\phi(uv)=\phi(u) \circ \phi(v)
\end{equation*}
holds. Moreover, the homomorphism is injective and surjective.
\end{lemma}
\begin{proof}
The mapping is clearly linear. We shall prove the homomorphic property for elementary homogeneous tensors. For this
we use the polarization formula to prove that for $n$ vectors
$x_{1}, x_{2}, \ldots , x_{n}$ of $U$ we have
\[
\phi(x_{1} x_{2} \cdots x_{n})= x_{1} \circ x_{2} \circ \cdots \circ x_{n}\,.
\]
Since both RHS and LHS are symmetric multilinear functions of the vectors, by the polarization formula (\ref{polar:ref}) it is enough to prove that given a linear functional $\Omega$ on the symmetric tensor algebra,
\[
\Omega ( \phi(x^{n}) ) = \Omega(x \circ x \circ \cdots \circ x )\,.
\]
But, this is true by the very definition of the homomorphism. For this reason
\[
\Omega( \phi(x_{1} x_{2} \cdots x_{n}) )= \Omega ( x_{1} \circ x_{2} \circ \cdots \circ x_{n} )\,,
\]
and since $\Omega$ is arbitrary the above identity is valid. Observe now that
\begin{align*}
\phi(x_{1} x_{2} \cdots x_{n} y_{1} y_{2} \cdots y_{m}) &= x_{1} \circ x_{2} \circ \cdots \circ x_{n} \circ y_{1} \circ y_{2} \circ \cdots \circ y_{m}\\
&= \phi(x_{1} x_{2} \cdots x_{n}) \circ \phi( y_{1} y_{2} \cdots y_{m})
\end{align*}
and by linearity
\[
\phi(uv)=\phi(u) \circ \phi(v)\qquad\text{for every $u\in \bigodot^{t} U$ and every $v \in \bigodot^{s} U$}.
\]
By linear extension we have shown that $\phi$ is homomorphic. We prove now that $\phi$ is injective. For this suppose that there exists an element $u$ of the symmetric algebra with
\[
\phi(u)= 0\,.
\]
First observe that
\[
\phi(x^{t})= x^{t}+ \text{lower grade terms}\,.
\]
By linearity for an element $u \in \bigodot^{t} U$ we have
\[
\phi(u)= u+ \text{lower grade terms}\,.
\]
For a general element $u \in \mathcal{T}_{S}(U)$ lying in the kernel of $\phi$, we have by the natural grading
\[
u= \sum_{k=0}^{m} u_{k}\qquad \text{with $u_{k}\in\bigodot^{k} U$}.
\]
Obviously $u_{k} \neq 0$. By the homomorphic property
\[
\phi(u)= \sum_{k=0}^{m} \phi(u_{k}) = u_{m} + \text{lower grade terms}=0
\]
and we conclude that $u_{k}=0$ which is a contradiction. This proves the injectivity.
It remains to show that $\phi$ is surjective. It is enough to prove the theorem for the elementary homogeneous elements . We already know that
\[
x \circ \cdots \circ x= x^{t}+ \text{lower grade terms} \iff x^{t}= x \circ \cdots \circ x + \text{lower grade terms}\,.
\]
By iterating this procedure for the low order terms we arrive at the result that
\[
x^{t}= \sum_{k=0}^{t} a_{k} x^{\circ k} = \phi ( \sum_{k=0}^{t} a_{k} x^{ k} )\,.
\]
This ends the proof of the theorem.
\end{proof}
Now we have a model to prove the same property for the square product. Namely, the fact that the tensor algebra with this square product is homomorphically equivalent to the tensor algebra with the ordinary product. Let us define the mapping $\phi$ with domain and codomain equal to the tensor algebra $ \mathcal{T} (V)$ first on elementary tensors as
\begin{enumerate}
\renewcommand{\labelenumi}{(\roman{enumi})}
\item $\phi(\mathbf{1} )=\mathbf{1}$
\item $\phi(x_{1} \otimes\cdots \otimes x_{n} ) = x_{1} \square \cdots \square x_{n}$\,,
\end{enumerate}
and then over the whole tensor algebra by linearity.
\begin{lemma}
For $u,v \in \mathcal{T} (V)$, the above mapping $\phi$ is a homomorphism as the property
\[
\phi(uv)=\phi(u) \square \phi(v)
\]
holds. Moreover, the homomorphism is injective and surjective.
\end{lemma}
The proof is analogous to the one for the circle product.
\section{The circle product for the anti-symmetric algebra}
In this section we aim to carry out the rest of the construction we mentioned in the introduction. We aim to prove that if we define
\[
u \circ v = {\rm ASymm}( u \square v)\qquad\text{where $u, v\in \mathcal{T}_{A}(U)$}
\]
then we can create yet another product on the antisymmetric algebra. As a first step we observe that
\[
x \circ y = (x | y) + x \wedge y \qquad\text{where $x, y\in V$}
\]
and
\[
x \wedge y \circ z \wedge w = (x | z) y \wedge w + (x | w) y \wedge z + (y | z) x \wedge w + (y | w) x \wedge z + x \wedge y \wedge z \wedge w\,.
\]
Obviously the new product is neither commutative , nor antisymmetric. One also observes that the pairings operate on terms of grade 0 or 1 because of anti-symmetry. We prove that the new circle product is associative.
\begin{lemma}
For elements $u, v, w\in\mathcal{T}_{A}(U)$ we have the following relation
\[
(u \circ v ) \circ w= u \circ (v \circ w)\,.
\]
\end{lemma}
\begin{proof}
It is enough to prove the theorem for elementary antisymmetric tensors. By the properties of the antisymmetric algebra outlined in theorem \ref{thm:asymm}, it is not difficult to see that
\[
(u \circ v ) \circ w= \sum_{(u),(v),(w)} (u_{(1)} | v_{(1)} ) (u_{(2,1)} \wedge v_{(2,1)} | w_{(1)} ) u_{(2,2)} \wedge v_{(2,2)} \wedge w_{(2)}\,,
\]
while
\[
u \circ (v \circ w) = \sum_{(u),(v),(w)} (v_{(1)} | w_{(1)} ) ( u_{(1)} | v_{(2,1)} \wedge w_{(2,1)} )
u_{(2)} \wedge v_{(2,2)} \wedge w_{(2,2)}\,.
\]
Since the pairing is symmetric the terms inside it antisymmetric, they must be of grade at most 1 otherwise the pairing is 0. For this reason the anti-symmetrization restricted in these cases is self-adjoint because the pairing degenerates to the ordinary duality. Then we have by the co-commutativity
\[
(u \circ v ) \circ w= \sum_{(u),(v),(w)} (u_{(1)} | v_{(1)} ) (u_{(2,1)} \otimes v_{(2,1)} | w_{(1)} ) u_{(2,2)} \wedge v_{(2,2)} \wedge w_{(2)}
\]
and
\[
u \circ (v \circ w) =
u \circ (v \circ w) = \sum_{(u),(v),(w)} (v_{(1)} | w_{(1)} ) ( u_{(1)} | v_{(2,1)} \otimes w_{(2,1)} )
u_{(2)} \wedge v_{(2,2)} \wedge w_{(2,2)}\,.
\]
By the definition of the square product (\ref{square:def}) and co-associativity,
\[
(u \circ v ) \circ w= \sum_{(u),(v),(w)} (u_{(1)} \square v_{(1)} | w_{(1)} ) u_{(2)} \wedge v_{(2)} \wedge w_{(2)}
\]
while
\[
u \circ (v \circ w) =\sum_{(u),(v),(w)} ( u_{(1)} | v_{(1)} \square w_{(1)} ) u_{(2)} \wedge v_{(2)} \wedge w_{(2)}\,.
\]
The lemma follows from (\ref{laplace:perm}), the permutation of the square product inside the Laplace pairing.
\end{proof}
It is a matter of very simple steps to prove that the new circle product is in fact an algebra product. We record in the next theorem the analogous properties of the circle product for the anti-symmetric algebra. Let us define the mapping $\phi$ with domain and codomain equal to the antisymmetric tensor algebra $ \mathcal{T}_{A} (V)$ first on elementary tensors as
\begin{enumerate}
\renewcommand{\labelenumi}{(\roman{enumi})}
\item $\phi(\mathbf{1} )=\mathbf{1}$
\item $\phi(x_{1} \wedge \cdots \wedge x_{n} ) = x_{1} \circ \cdots \circ x_{n}$\,,
\end{enumerate}
and then we extend it over the antisymmetric algebra by linearity.
\begin{theorem}
Let $U$ be a vector space in self-duality via the bi-linear form $\langle \cdot , \cdot \rangle$. The anti-symmetric tensor algebra $\mathcal{T}_{A}(U)$ equipped with the circle product is a unital associative algebra. For $u,v \in \mathcal{T}_{A}(U)$, the above mapping $\phi$ is a homomorphism as the property
\begin{equation*}
\phi(u \wedge v)=\phi(u) \circ \phi(v)
\end{equation*}
holds. Moreover, the homomorphism is injective and surjective.
\end{theorem}
The proof is analogous to the one for the circle product on symmetric tensors. The reader can see that this circle product is indeed the product presented in \cite{rota94}.
\section{Conclusion}
In this paper we have proved that the circle product on the symmetric tensor algebra, introduced in \cite{brouder09}, can be derived as a restriction of a more general product on the tensor algebra. This general product shares many commonalities like the circle product. It is defined via a more general Laplace pairing. For symmetric tensors this Laplace pairing makes the circle multiplication by a symmetric tensor self-adjoint. For general tensors the adjoint map of left square multiplication is the left square multiplication in the opposite algebra. Finally, we have proved the important fact that the new algebraic structures are homomorphically equivalent with the traditional structures. As a by-product we have shown that the symmetric tensor algebra arises naturally via algebraic considerations, namely via the Laplace pairing properties. We have also extended our results to the anti-symmetric algebra to re-derive the product defined by Rota and Stein \cite{rota94}. This step gives us a unified picture of the circle products defined in the literature as specializations of a general structure.
\bibliographystyle{amsplain}
|
\section{Introduction}
\par The internal shock collisions version of the fireball model provides an adequate description of the origin of GRBs \citep{RM94}. A collapsing massive star \citep{Woos1993, WB2006} or a binary merger \citep{Narayan92} followed by a strong relativistic explosion are considered to be the progenitors of these violent events. Internal collisions inside the fireball of relativistic shells departing from the progenitor with different velocities, give rise to the GRB \citep{Piran04}.
The same model attributes the afterglow emission to synchrotron radiation which is emitted during the deceleration of the external shock in the interstellar medium (ISM) \citep{Sari98, ZM04}. This behaviour can last for several days or even months after the burst covering a wide range of the spectrum. As afterglow observations improved, however, certain questions were raised that could not be answered with the standard model (\cite{Zhang07}, for a detailed discussion). Recent observations in the optical \citep{Stanek2006}, radio as well as the X-ray band \citep{Burrows2005, Nousek2006}, show a strong variability in the afterglow phase for a large proportion of the bursts, which can not be reproduced by the standard external shock model.
It has been proposed that a bump in the afterglow light curve may result when the forward shock propagating in the ISM encounters a density jump, caused by an inhomogeneity of the surrounding medium generated by interstellar turbulence or by anisotropy in a precursor wind from the GRB progenitor \citep{Wang2000,Lazzati2002}. However, numerical simulations of a spherical explosion exhibit a rather canonical behaviour and even for a sharp and large increase in the external density this model does not produce sharp features in the light curve and cannot account for significant temporal variability in GRB afterglows \citep{NG2006,HvE09}. It has been suggested that a late activity of the central engine could explain the observed variability \citep{Falcone2006, Zang06, Romano2006, Ioka2005}.
In the late activity scenario, the central engine produces consecutive explosions after the initial burst which collide when a slow shell is followed by a faster one. This late activity of the source could be explained from a two-stage collapse in the central object. As proposed by \citet{King05} a collapsing core which has enough angular momentum to fragment will leave behind a second compact ``star'' in the form of a self-gravitating neutron lump. The fallback of this ``star'' at later times on the initial compact object can restart the central engine. Other theories suggest that the viscous hyperaccreting accretion disk around a black hole which fragments at large radii becomes dynamically unstable on different timescales and thus collapses at different times \citep{Perna2006}. It is proposed that the region at the vicinity of the accretor can play an important role in determining the accretion rate and therefore the energy output of the explosion \citep{Proga2006}.
According to late activity models the second blast wave continuously supplies the system with energy while colliding with the initially ejected material, producing in that way the rebrightening observed in the afterglow.
The role of magnetic fields in GRBs is still arguable. In the fireball model the presence of a magnetic field is not dynamically important for the evolution of the flow but plays an important role for the emission during the interaction of the flow with the external medium. Although the early afterglow emission strongly depends on the magnetization of the flow, in the late stages of the afterglow where the shells experience strong deceleration, the evolution of strongly magnetized shells resembles that of hydrodynamic shells and can be described by the self-similar Blandford-Mckee (BM) \citep{BM} approximation \citep{Mimica09}. At this stage of the afterglow the emission no longer contains information about the initial magnetization of the flow.
In section 2 we describe the high resolution numerical simulations we performed of late collisions of two ultra-relativistic shells during the afterglow phase.
We claim that differences in the flow must have an impact on the resulting light curves and perform four simulations with varying Lorentz factor and energy content of the second shell in order to investigate the effect of these parameters. The adaptive mesh refinement (AMR) technique enables us to use high resolution in long term, 1D relativistic hydro simulations, in order to capture the forward and reverse shock formation on the second shell and study in detail the stages before, during and after the merger of the two shells.
The effects of the collision between the two shells in the light curves is described in section 3. We study both spherical explosions as well as a hard-edged jet scenario where no lateral spreading has occured (numerical simulations in two dimensions have shown that only very modest lateral spreading occurs while the jet is relativistic \citep{ZMf09,Granot2001,Meliani2010}. Optical and radio on-axis light curves are calculated for different opening angles and the strength of the occuring flare or rebrightening is found to depend on this opening angle.
We also note a clear difference in the shape between optical and radio light curves as well as a difference in the time of the appearance of the flare between the two frequencies. We explain this chromatic behaviour in terms of the synchrotron self-absorption mechanism and the different main contributing regions of the jet to the emission.
We construct emission images for the different stages of the merger and connect the dynamical characteristics of the flow at each stage of the collision to the features in the light curves.
We will discuss and summarize our results in section 4.
For the dynamical simulations we are using the Adaptive Mesh Refinement version of the Versatile Advection Code (AMRVAC) \citep{Kep03,Mel07}, and for the light curves and emission images calculations the radiation code of van Eerten \& Wijers (2009).
\section{Modeling of the multi-shell dynamics}
When the initially ultra-relativistic shell ejected from the central source starts to decelerate in the interstellar medium, a forward shock is created separating the shocked ISM from the ambient ISM. As mass is swept up, the kinetic energy of the shell is transformed into kinetic and thermal energy of the shocked matter. At the same time a reverse shock is formed which crosses the shell leading to conversion of the shell's kinetic energy into thermal. The resulting shocked ISM matter ultimately follows the self-similar BM analytical solution \citep{Mel07}. The afterglow is nowadays widely recognized as synchrotron radiation emitted during this phase of the propagation of the shell \citep{MR97}.
In our model we consider that the central engine remains active even after the initial ejection of the first shell resulting in a delayed second explosion. The produced blast wave will now travel with a steady velocity into an empty medium, since most of the matter has been swept up, until it reaches the termination shock of the first shell. In this paper we reproduce the collision process of these two shells and claim that for small opening angles the heating of the matter that happens during this phase is responsible for the appearance of the flares observed in the light curves.
\subsection{Special relativistic hydrodynamic equations}
We perform the dynamical simulations using the 1D special relativistic hydrodynamic equations in spherical coordinates and the code AMRVAC.
The equations describing the motion of a relativistic fluid are given by the five conservation laws
\begin{equation}
(\rho u^{\mu})_{;\mu}=0, \ \ \ \ (T^{\mu\nu})_{;\nu}=0
\end{equation}
where $\mu , \nu = 0,1,2,3$ are the indices running over the 4-dimensional spacetime, $\rho$ is the proper rest mass density of the fluid, $u^{\mu}$ is the four-velocity and $T^{\mu\nu}$ is the stress-energy tensor given by $T^{\mu\nu}=\rho h u^{\mu}u^{\nu} + pg^{\mu\nu}$. Here with $p$ we denote the fluid rest frame pressure, while $g^{\mu\nu}$ is the Minkowski metric tensor, as we will consider a flat spacetime at distances far from the central engine. The specific enthalpy $h$ of the fluid is given by $h=1+ \varepsilon +p/\rho$ where $\varepsilon$ is the specific internal energy. Rewriting the conservation equations in vector form we have in a familiar 3+1 split the conservation laws
\begin{equation}
\frac{\partial U}{\partial t} + \frac{\partial F^{i}(U)}{\partial x^{i}}=0, \ \ \ \mbox{with} \ \ i=1,2,3.
\end{equation}
The vector $U$ is defined by the conserved variables as
\begin{equation}
U= [D=\rho \gamma, {\bf{S}}=\rho h \gamma^{2} {\bf{v}}, \tau = \rho h \gamma^{2} -p -D]^{T},
\end{equation}
and the fluxes are given by
\begin{equation}
F=[\rho\gamma{\bf{v}}, \rho h \gamma^{2} {\bf{v}}{\bf{v}} + p{\bf{I}}, \rho h \gamma^{2}{\bf{v}}-\rho \gamma {\bf{v}}]^{T},
\end{equation}
where ${\bf{v}}$ is the three-velocity and $\bf{I}$ is the $3\times3$ identity matrix. The system of equations is closed by using the equation of state
\begin{equation}
p=(\Gamma_{p} -1)\rho \varepsilon.
\end{equation}
In our simulations we choose the polytropic index to be $\Gamma_{p}=4/3$. This is a good approximation since most of the shocks in the cases described below are mainly relativistic or near-relativistic.
\subsection{Initial setting}
For the dynamics of the first shell we consider that the reverse shock has already crossed the shell which now decelerates in the external medium. For the purpose of this simulation we will use the Blandford \& McKee (BM) approximation to describe this phase. This is a self-similar solution of a relativistic blast wave expanding in a uniform or radially varying medium. We consider the case where the explosion is assumed to be spherically symmetric and adiabatic.
In the BM model the density of the circumburst medium scales as a power law with distance $\rho_{1}(r)\propto r^{-k}$. For all the simulations in the present paper we will consider that the density of the circumburst medium is constant $(k=0)$, with particle number density $n_{1}=1\,\rm{cm^{-3}}$ and cold, with the pressure given by $p_{1}=10^{-5}n_{1}m_{p}c^{2}$ chosen such that it does not dynamically affect the system. We set the Lorentz factor of the BM shock at $\Gamma_{0}=23$ at the start of the simulation placed in distance $R_{0}\simeq2.04\times10^{17}$ cm. Considering the decelerating radius of the BM shock $R_{dec}=(3E_{0}/(4\pi n_{0}m_{p}c^{2}\Gamma^{2}))^{1/3}$, after which the Lorentz factor of the shock starts decreasing with distance as a power law, the initial distance $R_{0}$ of the shock corresponds to a distance 3.7 times greater than the deceleration radius $R_{dec}$ of a jet with initial Lorentz factor 100 and 7.8 times greater than the deceleration radius of a jet with initial Lorentz factor 300. The energy content of the shell is $E=10^{52}\,\rm{erg}$. According to the BM model in the ultra-relativistic case the jump conditions at the BM shock are given by
\begin{equation}
p_{2}=\frac{2}{3} \Gamma^{2}\rho_{1},
\end{equation}
\begin{equation}
n_{2}=\frac{2\Gamma^{2}}{\gamma_{2}}n_{1},
\end{equation}
\begin{equation}
\gamma_{2}^{2}=\frac{1}{2}\Gamma^{2}.
\end{equation}
where index 2 denotes the shocked medium and $\Gamma$ is the Lorentz factor of the shock. According to the BM model in the ultra-relativistic case, the radius of the shock at time $t$ is given up to order $O(\Gamma^{-2})$ by
\begin{equation}
R(t) = ct\left(1-\frac{1}{8\Gamma^{2}}\right).
\end{equation}
From the jump conditions and by choosing the similarity variable to be $\chi=\left[1+8\Gamma^{2}\right]\left(1-r/t\right)$, we obtain the properties of the shocked medium
\begin{equation}
p_{2}\left(r,t\right)=\frac{2}{3}\rho_{1}\Gamma^{2}\left[\left(1+8\Gamma^{2}\right)\left(1-\frac{r}{t}\right)\right]^{-17/12},
\end{equation}
\begin{equation}
\gamma_{2}\left(r,t\right)=\frac{1}{2}\Gamma^{2}\left[\left(1+8\Gamma^{2}\right)\left(1-\frac{r}{t}\right)\right]^{-1},
\end{equation}
\begin{equation}
\rho_{2}\left(r,t\right)=\frac{2\rho_{1}
\Gamma^{2}}{\gamma_{2}}\left[\left(1+8\Gamma^{2}\right)\left(1-\frac{r}{t}\right)\right]^{-7/4}.
\end{equation}
The total energy is then given by $E=8\pi \rho_{1}\Gamma_{0}^{2} c^{5}t_{0}^{3}/17$. If the initial Lorentz factor of the simulation is fixed at $\Gamma_{0}$, the duration $t_{0}$ of the shock so far then follows from this equation. The initial pressure and density jumps between the BM shell and the ISM at the position of the shock are $p_{2}/p_{1}= 10^{7}$ and $\rho_{2}/\rho_{1}=10^{2}$.
The second shell is uniform, cold and ultra-relativistic and placed at distance $\Delta R = 10^{14}$ cm behind the BM shell. It is therefore assumed that the shell has moved freely up to this point. Considering a duration of the second ejection event of $\Delta t = 1000$ s, the initial thickness of the shell will be $\delta = c\Delta t = 3\times10^{13}$ cm. The energy of the second shell is given by
\begin{equation}\label{E_sh}
E_{sh}=4\pi\Gamma_{sh}^{2}R_{in}^{2}\delta\rho_{sh}c^{2},
\end{equation}
where $R_{in}$ denotes the initial distance of the shell and $\rho_{sh}$ and $\Gamma_{sh}$ the initial density and Lorentz factor. The initial pressure is chosen as $p_{sh}=5\times10^{-2}\rho_{sh}$. In our initial conditions, we vary from case to case the given parameters $\Gamma_{sh}$ and $E_{sh}$ which then serve to specify the shell density $\rho_{sh}$. The energy provided here refers to the isotropic-equivalent energy. For all cases described in this paper we consider emission along the rotation axis of the system and that the two ejecta have the same opening angle. We also neglect the effects of lateral spreading in both dynamical and radiation calculations. As shown in \citet{ZMf09}, sideways expansion can be a very slow process and definitely negligible for times under consideration in this paper.
We perform four simulations with varying Lorentz factor and energy for the second shell. In this simulation we are using a domain of size $[0.01, 10]\times10^{18}$ cm and 240 cells at the coarsest level of refinement. The physical properties of the afterglow shock collision model require a large domain and a very thin second shell which demands very high resolution in order to be resolved. The maximum level of refinement is 22, leading to an effective resolution of $5.03316\times10^{8}$ cells. We force the front and the back of the uniform shell to be refined in order to avoid numerical diffusion which would cause an artificial spreading of the shell. A summary of the simulation parameters can be found in Table 1. The top left figures in Figs. \ref{c9sn} and \ref{c12sn} show the initial conditions of case 1 and 4 respectively.
\begin{table}
\centering
\caption{Properties of the second shell for each case. $\Gamma=23$ and $E=10^{52}$ ergs are the Lorentz factor and energy of the BM shell.}
\begin{tabular}{cccr}
\hline
\hline
case 1 & case 2 & case 3 & case 4 \\
\hline
\hline
$\Gamma/\sqrt{2}$ & $2\Gamma/\sqrt{2}$ & $\Gamma/\sqrt{2}$ & $2\Gamma/\sqrt{2}$ \\
$E$ & $E$ & $2E$ & $2E$ \\
\hline
\end{tabular}
\label{table1}
\end{table}
\subsection{Interaction between two shells.}
\subsubsection{Jump conditions.}
The evolution of the two-shell system is shown in figures \ref{c9sn} and \ref{c12sn} for case 1 and case 4 respectively. In case 1 the two shells initially have the same energy and Lorentz factor. As the BM shell decelerates in the interstellar medium the second shell catches up. While the matter from the BM shell is swept up by the second shell, a forward shock is created separating the shocked matter from the unshocked. At the same time a reverse shock crosses the second shell and a contact discontinuity appears in between both shocks. At this stage the front of the second shell has split into four regions. In region 1 there is the BM matter. In region 2 there is the BM matter which has been heated by the forward shock. In region 3 there is the matter of the uniform shell that has been heated from the propagation of the reverse shock and in region 4 there is the unshocked matter of the second shell. This is also shown for all 4 cases in Fig. \ref{etrans} where a close-up of the second shell at early times is given.
The jump conditions at the second shell as given by \citet{BM} for an arbitrary strong shock ($p_{2}/n_{2} \gg p_{1}/n_{1}$) are
\begin{equation}
e_{2}=\overline{\gamma}_{2}\frac{n_{2}}{n_{1}}w_{1}, \;\;\;\;\;\;\;\;\;\ \frac{n_{2}}{n_{1}}=4\overline{\gamma}_{2}+3,
\end{equation}
\begin{equation}
\frac{e_{3}}{n_{3}m_{p}c^{2}}=\overline{\gamma}_{3}-1, \;\;\ \frac{n_{3}}{n_{4}}=4\overline{\gamma}_{3}+3,
\end{equation}
where $\overline{\gamma}_{2}$ is the Lorentz factor of the fluid in region 2 relative to region 1 and $\overline{\gamma}_{3}$ is the Lorentz factor of the fluid in region 3 relative to region 4. The primitive variables $n_{i}$, $p_{i}$ as well as the internal energy density $e_{i}$ and enthalpy $w_{i}$, are measured in the fluid frame, while the Lorentz factors of the several regions, $\overline{\gamma}_{i}$, are measured in respect to the ISM which is considered to be at rest. In our case the forward shock of the second shell is moving in a hot medium which has already been heated by the BM shock. Therefore $w_{1}=e_{1}+p_{1}\simeq 4p_{1}$, assuming the ultra-relativistic equation of state $e_{1} \simeq 3p_{1}$.
\subsubsection{Dynamics of the different cases.}
After the initial explosion and acceleration of the uniform shell, the swept up BM matter starts playing an effective role in the deceleration of this shell. In Fig. \ref{etrans} we show snapshots of all 4 simulations taken at emission time $t_{e}= 6.98 \times 10^{6}$ s after the explosion when the forward shock, the contact discontinuity and the reverse shock are fully developed. At this time the BM matter is continuously heated by the forward shock resulting in an increase in the density of the shocked matter, $n_{2}/n_{1} \simeq 7.10$ (case 1). The relative Lorentz factor is given by
$\overline{\gamma}_{2}=\gamma_{2}\gamma_{1}(1-\sqrt{(1-\gamma_{2}^{-2})(1-\gamma_{1}^{-2})}) \sim 1.46$. A small inconsistency is observed between the simulation results and the jump conditions which derives from the fact that in all our simulations the random kinetic energy per particle at the two sides of the shock is $p_{2}/n_{2}>p_{1}/n_{1}$ rather than $p_{2}/n_{2} \gg p_{1}/n_{1}$.
Similarly the propagation of the forward shock results in an increase of the internal energy of the shocked matter which now is
$e_{2}\simeq 0.61 \sim \overline{\gamma}_{2}w_{1}n_{2}/n_{1}$.
At the same time the reverse shock crosses the shell while heating the uniform shell matter and transforming the kinetic energy of the shell into thermal. The contact discontinuity separating the two regions appears as a density jump between the two shocked fluids, $n_{3}/n_{2}\simeq 10^{2}$ while the Lorentz factor and pressure remain continuous, $\gamma_{2}=\gamma_{3}$ and $p_{2}=p_{3}$. The efficiency of this energy transformation mechanism is highly dependent on the reverse shock itself.
According to Sari \& Piran (1995) the reverse shock depends only on two parameters, the Lorentz factor of the unshocked shell material relative to the BM unshocked matter, $\overline{\gamma}_{4}$, and the density jump, $n_{4}/n_{1}$. For $\overline{\gamma}_{4}^{2}\gg n_{4}/n_{1}$ the reverse shock will be relativistic, $\overline{\gamma}_{3} \gg 1$. For case 1, at emission time $t_{e}=6.98\times10^{6}$ s. (Fig. \ref{etrans}), the reverse shock remains Newtonian ($\overline{\gamma}_{3} = 1.00484$) while propagating into the shell and thus insufficient to heat the shell effectively.
Following this behaviour, by the time the reverse shock reaches the back of the shell, the shell is still dominated by its kinetic energy although significantly decelerated (see also Fig. \ref{c9sn}).
\begin{figure*}
\centering
\includegraphics[scale=0.4]{c9sn0.eps}
\includegraphics[scale=0.4]{c9sn100.eps}
\includegraphics[scale=0.4]{c9sn280.eps}
\includegraphics[scale=0.4]{c9sn1000.eps}
\caption{Snapshots of the dynamics for case 1, taken at local emission times $t_{e}=6.81\times10^{6}$ s, $t_{e}=8.48\times10^{6}$ s, $t_{e}=1.15\times10^{7}$ s and $t_{e}=2.35\times10^{7}$ s after the initial explosion (top left - bottom right). Lorentz factor, density and pressure are indicated as the dotted, solid, and dashed line. The top-left figure indicates the initial setting of our simulation.}
\label{c9sn}
\end{figure*}
We follow the same analysis for case 2 in which we double the Lorentz factor of the second shell compared to case 1. Since we choose to maintain constant the energy and the thickness of the second shell, the change in the Lorentz factor affects the density of the second shell, which is now smaller by a factor of 4 compared to case 1 (eq. \ref{E_sh}). Comparing the properties of the flow for the two cases at similar times after the explosion we notice that the forward shock is stronger in case 2. Specifically at emission time $t_{e}= 6.98 \times 10^{6}$ s after the explosion the forward shock has a Lorentz factor relative to the BM matter $\overline{\gamma}_{2} = 1.99$ and the number density satisfies $n_{2}/n_{1} \simeq 7.67$. That shows that the forward shock is more efficient in the second case as it compresses the matter of the BM shell to a higher degree and to higher Lorentz factor than in case 1.
At the same time the reverse shock is stronger than in case 1, $\overline{\gamma}_{3} = 1.16$ and although it is not ultrarelativistic it is more efficient in converting the kinetic energy of the second shell into thermal. That appears clearly as a difference in the pressure between the shocked and unshocked shell matter, $p_{3}/p_{4}=12.27$ for case 2 compared to $p_{3}/p_{4}=1.78$ for case 1.
In case 3 we double the energy of the second shell. The relevant Lorentz factor between the shocked and unshocked BM matter is $\overline{\gamma}_{2} \simeq 1.75$ and the density jump satisfies $n_{2}/n_{1}\simeq 6.69$. Compared to case 1 the reverse shock for this case remains very weak, $\overline{\gamma}_{3}\simeq 1.00013$, while propagating in the uniform shell since it has to traverse a denser medium than before, leading to a compression ratio between the shocked and the unshocked shell matter $p_{3}/p_{4}=1.09$. In case 4 we double both the energy and Lorentz factor of the second shell. The initially very fast shell leads to a strong forward shock, $\overline{\gamma}_{2}\simeq 2.26$. The reverse shock is now stronger compared to case 3, $\overline{\gamma}_{3}\simeq 1.15$, compressing the shocked shell matter to higher pressure, $p_{3}/p_{4}=7.98$.
In figure \ref{etrans} we plot the thermal to mass energy ratio together with Lorentz factor, density and pressure for fixed early time for all cases. We observe that as matter from the BM shell is swept up from the forward shock of the second shell, the thermal energy of the fluid increases compared to the mass energy (which appears as a small bump in the thermal to mass energy curve at the position of the forward shock).
While the shell gets traversed by the reverse shock part of the kinetic energy of the shell is transformed into thermal as a result of the propagation of the reverse shock. We notice that the relativistic reverse shock in case 2 is a lot more efficient in raising the thermal energy component than in case 1. In case 3 and 4 a similar behaviour is observed. By increasing the energy in case 3 and with the Lorentz factor being the same as in case 1, the density becomes higher in the shell (eq. 13). As a result the reverse shock is very weak and highly inefficient in thermalizing the cold shell. In case 4 both the energy and the Lorentz factor of the second shell have twice the value compared to case 1. In this case the reverse shock is propagating in a less dense medium compared to case 3 and is now more efficient in transforming the kinetic energy of the second shell into thermal (Fig. \ref{etrans}). The Lorentz factor of the reverse shock is slightly overestimated as a result of our use of a fixed adiabatic index of 4/3 throughout the domain. Nevertheless, as we see from Fig. \ref{etrans}, the fluid is highly relativistic from the contact discontinuity of the second shell onward and therefore the regions that mostly contribute to the observed flux (see next section) are appropriately described by a fixed adiabatic index.
\begin{figure*}
\centering
\includegraphics[scale=0.4]{c120sn.eps}
\includegraphics[scale=0.4]{c12100snb.eps}
\includegraphics[scale=0.4]{c12280snb.eps}
\includegraphics[scale=0.4]{c121000sn.eps}
\caption{Snapshots of the dynamics for case 4, taken at local emission times $t_{e}=6.81\times10^{6}$ s, $t_{e}=8.48\times10^{6}$ s, $t_{e}=1.15\times10^{7}$ s and $t_{e}=2.35\times10^{7}$ s after the initial explosion (top left - bottom right). Lorentz factor, density and pressure are indicated as the dotted, solid, and dashed line. The top-left figure indicates the initial setting of our simulation.}
\label{c12sn}
\end{figure*}
Later, at emission time $t_{e}=1.513\times10^{7}$ s (see Figs. \ref{c9sn} and \ref{c12sn}), the forward shock will catch up and overcome the BM shell as the latter one decelerates in the ISM. As the forward shocks merge and the reverse shock has crossed the back of the shell, a dense but slow and underpressured region is left behind unable to follow the forward shock as it has lost almost all of its kinetic energy reaching a near-equilibrium state with the surrounding matter.
\begin{figure*}
\centering
\includegraphics[scale=0.4]{c9sn10energyfinalB.eps}
\includegraphics[scale=0.4]{c10sn10energyfinalB.eps}
\includegraphics[scale=0.4]{c11sn10energyfinalB.eps}
\includegraphics[scale=0.4]{c12sn10energyfinalB.eps}
\caption{Ratio of thermal to mass energy ($E_{th}/(\rho c^{2})$) and normalized Lorentz factor, density and pressure for case 1 to case 4 (top left to bottom right) for the second shell at emission time $t_{e}=6.98 \times 10^{6}$ s. Distance is normalized to $10^{17}$ cm. For cases 1 and 2 the transformation from thermal to kinetic energy is clear at the forward shock as well as the kinetic to thermal energy transformation at the position of the reverse shock. The ratio depends strongly on the properties of the second shell. The four regions of interaction between the two shells are indicated on the figure for case 1. Region (1) consists of the unshocked BM matter, region (2) of the shocked BM matter, region (3) of the shocked uniform shell matter and region (4) of the unshocked uniform shell matter. We focus on the front of the second shell, the dynamical evolution of which plays an important role on the light curves. }
\label{etrans}
\end{figure*}
\section{Radiation calculations}
In this section we describe the numerical calculations we performed in order to construct the light curves at the afterglow phase. The following calculations were carried out with the radiation code introduced in \citet{HvE09} and \citet{HvE10}. The main process contributing to the afterglow is synchrotron radiation. In both external and internal shock collision models, a magnetic field is required in order to fit the observational data. This magnetic field is most likely generated by instabilities forming in the shock such as the relativistic two-stream instability or the Weibel instability \citep{Med99, Weibel59}. In this case a fraction of the total thermal energy behind the shock goes to particle acceleration and another one to the generation of the magnetic field. Assuming that the fractions of the thermal energy density $e_{th}$, contributing to the magnetic energy density, $\epsilon_{B}$, and electron energy density, $\epsilon_{E}$, have a fixed value, we can calculate the afterglow emission.
In our calculations we neglect the effect of the magnetic fields and radiative losses on the dynamics and do not take into account the effects of Compton scattering and electron cooling (that play a negligible role at times and frequencies under consideration).
We do however include the effect of synchrotron self-absorption (ssa) due to the re-absorption of the radiation from the synchrotron electrons by solving simultaneously the linear radiative transfer equations for a large number of rays through the evolving fluid. An analytical formula for obtaining the self-absorption frequency $\nu_{sa}$ can be found in \citet{Granot1999b}.
The radiation code is specifically written to include the snapshots produced by the dynamical simulations performed with AMRVAC. In all our calculations the values of $\epsilon_{B}$ and $\epsilon_{E}$ are fixed to 0.01 and 0.1 respectively and we assume a power law distribution for the accelerated electrons with $p=2.5$. We also fix the fraction of the electrons accelerated to this power law distribution $\xi_{N}$ equal to 0.1. Assuming isotropic radiation in the comoving frame, the observer's flux calculated by the radiation code is given by
\begin{equation}
F_{\nu}= \frac{1+z}{d_{L}^{2}}\int \frac{d^{2}P_{V}}{d\nu d\Omega}(1 - \beta \mu)cdAdt_{e},
\end{equation}
in the optically thin limit, where for the purpose of our simulations the redshift z is chosen to be zero. Here $D_{L}$ denotes the observer luminosity distance, $\beta$ the fluid velocity in units of $c$, $d^{2}P_{V}/d\nu d\Omega$ is the received power per unit volume, frequency and solid angle and $\mu=\cos\theta$, with $\theta$ being the angle between the fluid velocity and the line of sight. The surface element $dA$ corresponds to an equidistant surface \textit{A}, which is a surface intersecting the fluid grid from which the radiation arrives at the observer at time $t_{obs}$. Each surface corresponds to a specific emission time $t_{e}$. For a photon emitted from a location (r,$\theta$) at emission time $t_{e}$, the observer time is given by
\begin{equation}
t_{obs}= t_{e} - \frac{r\mu}{c}.
\end{equation}
When electron cooling does not play a role, the shape of the observed spectrum follows directly from the dimensionless function $Q(\nu / \nu_{m})$ which has the limiting behaviour $Q \propto (\nu/\nu_{m})^{1/3}$ for small $(\nu/ \nu_{m})$ and $Q\propto(\nu/\nu_{m})^{(1-p)/2}$ for large $(\nu/\nu_{m})$. The received power depends on this shape and on the local fluid quantities via
\begin{equation}
\frac{d^{2}P_{V}}{d\nu d\Omega} \propto \frac{\xi_{N}nB'}{\gamma^{3}(1-\beta\mu)^{3}}Q\left(\frac{\nu}{\nu_{m}}\right),
\end{equation}
where $n$ is the lab frame number density of the electrons and $B'$ is the comoving magnetic field strength. The synchrotron peak frequency $\nu_{m}$ corresponds to the Lorentz factor of the lower cut-off of the accelerated electron's power-law distribution $\gamma_{m}$, assuming that the Lorentz factor for the upper cut-off goes to infinity. Then the lower cut-off for the electrons will relate to the comoving number density $n'$ and thermal energy density $e'_{th}$ via
\begin{equation}\label{gammam}
\gamma_{m}\propto\left(\frac{p-2}{p-1}\right)\frac{\epsilon_{E}e'_{th}}{\xi_{N}n'}.
\end{equation}
The particle distribution and its lower cut-off are set at the shock front. However, subsequent evolution of $\gamma_{m}$ is dictated by adiabatic expansion of the fluid rather than synchrotron radiative losses. Therefore, in a relativistic fluid, eq. (\ref{gammam}) also holds further from the shock front with the same value of $\epsilon_{E}$.
The temporal behaviour of several GRBs contradicts the spherical explosion senario. A rapidly decaying afterglow emission suggests that the flow must be collimated rather than spherical. This is vital for the GRB mechanism since a spherical expansion would require a total energy budget $\sim 10^{54}$ ergs which is hard to produce from a stellar mass progenitor, while a jet shaped explosion can have the same result with less energy, $\sim 10^{51}-10^{52}$ ergs. This model suggests that when the jet decelerates to Lorentz factors such that $\gamma \sim \theta_{h}^{-1}$ is satisfied, with $\theta_{h}$ being the half jet opening angle, the flux that the observer receives will start decreasing resulting to a break in the afterglow light curves. In contrast to AGN jets that can be directly observed, GRB jets are only implicitly assumed from this break.
In our simulations we construct optical and radio light curves and emission images at various times for the 4 cases described above. We try different opening angles of the jet and associate the flare characteristics with the jet opening angle.
The main characteristic of our approach is the separation of the dynamics from the radiation simulation. The outputs from the dynamical simulations are used as an input to the radiation code in order to calculate synchrotron emission. When absorption plays a role, the code solves a series of linear radiative transfer equations (rather than eq. 16 directly) along the light rays starting from the back of the jet and passing through the jet towards the observer. For given emission time $t_{e}$ a surface on the radiative volume of the jet exists from which emission arrives at the observer at time $t_{obs}$.
A summation over the light rays along this \textit{equidistant surface}, and an integration over all the \textit{equidistant surfaces} is applied in order to calculate the contribution to the emission of every part of the jet. The observed flux is obtained by summing over the rays emerging from the jet. When the rays are not summed over, a spatially resolved emission image of the two shell system is obtained. An adaptive procedure similar to the one used in AMRVAC is employed when more rays need to be calculated in order to adequately capture the emission from the underlying fluid profile.
More details on the radiation code algorithmic strategy can be found in \citet{HvE09, HvE10}.
\subsection{Early afterglow and jet break estimation.}
The dynamical simulations cover a timescale starting from 0.072 days and ending 10 days after the initial explosion in the observer's time frame. In order to include the initial stages of the afterglow in our simulation we assume that prior to the simulation the outer shock has evolved according to BM and the second has moved with a constant velocity while retaining its initial shape.
In this way the deceleration of the initial explosion is taken into account resulting in the appearance of the jet break in the light curve.
Due to relativistic beaming the emission that reaches the observer is limited to emission angles $\theta<\gamma^{-1}$. During the afterglow phase however the flow is significantly decelerated and thus larger emission angles can be observed. When the Lorentz factor becomes small enough so that the condition $\theta_{h}\sim\gamma^{-1}$ is satisfied, where $\theta_{h}$ is the half opening angle of the jet, the observation cone becomes big enough for the edges of the jet to become visible to the observer. When there is no significant spreading of the matter at the edges only part of the visible region is occupied by the jet and thus from this point after the flux will start decaying faster. This transition to the faster decaying part of the light curve is often referred to as the jet break.
An analytical formula given in \citet{HvE10b} gives an estimation of the observed jet break time, $t_{obs,br}$ depending on the total energy of the explosion $E$, the half opening angle $\theta_{h}$ of the jet, the circumburst density $n_{0}$ and profile $k$. The contributing area to the emission in this case is not only the shock front itself but also the area behind the shock. For that purpose the radial profile from the BM model has been used, and the formula reads
\begin{eqnarray}\label{tbreak}
\lefteqn{t_{obs,br}=\theta_{h}^{2+2/(3-k)}\left(\frac{A_{1}}{2(1+2(4-k))}\right)^{1/(3-k)}\times} \nonumber \\
& & \left(\frac{1}{2}+\frac{1}{2(1+2(4-k))} + \frac{1}{2(4-k)2(1+2(4-k))}\right)
\end{eqnarray}
where $k$ corresponds to the parameter defining the power-law of the circumburst medium density, which for our case is always set to 0, and $A_{1}$ a parameter of the fluid depending on the energy $E$ of the explosion and the number density at the position of the shock, $A_{1}=E(17-4k)/(8\pi m_{p} n_{0} R_{0}^{k} c^{5-k})$ with $n_{0}$ being the proton number density at distance $R_{0}$.
\begin{table}
\centering
\caption{Observer time estimation of the jet break (estimation made in days after the explosion) for the four simulated cases assuming two hard-edged jet scenarios. }
\begin{tabular}{c|cccr}
\hline
\hline
\multicolumn{5}{c}{\bf{Jet break estimation}} \\
\hline
& case 1 & case 2 & case 3 & case 4 \\
\hline
$2\theta_{h}=2 $ & $0.0028$ & $0.0028$ & $0.0035$ & $0.0035$ \\
$2\theta_{h}=5 $ & $0.032$ & $0.032$ & $0.041$ & $0.041$ \\
\hline
\hline
\end{tabular}
\label{table2}
\end{table}
The jet break observer time estimation $t_{obs,br}$, is presented in Table \ref{table2} for all the cases we simulate and for two different opening angles of the jet. As expected, when the total energy of the explosion remains the same, the jet break time remains the same even if the Lorentz factor of the second shell is different. For larger opening angles of the jet the jet break is observed at later observer times as it takes longer for the shell to decelerate to Lorentz factors such that $\theta_{h} \sim \gamma^{-1}$ condition is satisfied. We discuss the jet break characteristics further in section 3.4.
In our model the interaction between the two shells starts happening after the jet break has occured, leading to a rebrightening of the afterglow and a sudden increase in the flux. The shape of the light curve and the characteristics of the emission are subjected to the dynamics of the flow and present qualitative and quantitative differences from case to case.
\begin{figure*}
\centering
\includegraphics[scale=0.4]{c9finalopt.eps}
\includegraphics[scale=0.4]{c10finalopt.eps}
\includegraphics[scale=0.4]{c9finalradio.eps}
\includegraphics[scale=0.4]{c10finalradio.eps}
\caption{Optical (top) and radio (bottom) light curves for case 1 (left) and case 2 (right) for different values of the jet opening angle $2\theta_{h}$. In all cases an increase of the flux due to energy injection from the second shell is observed. For small opening angles a flare appears in all four simulated cases which differs in shape according to the frequency. The peak flux is greater the higher the Lorentz factor and energy of the second shell are. The jet break estimation is denoted with a vertical line for both 2 and 5 degree jets. For case 1 we overplot the light curves produced for a single BM shell for a spherical explosion and a 2 degrees hard-edged jet. The analytical estimation of the slope is shown above the optical and radio light curves.}
\label{c910lc}
\end{figure*}
\begin{figure*}
\centering
\includegraphics[scale=0.4]{c11optlcD.eps}
\includegraphics[scale=0.4]{c12optlcD.eps}
\includegraphics[scale=0.4]{c11radiolcD.eps}
\includegraphics[scale=0.4]{c12radiolcD.eps}
\caption{Optical (top) and radio (bottom) light curves for case 3 (left) and case 4 (right) for different values of the jet opening angle $2\theta_{h}$. In all cases an increase of the flux due to energy injection from the second shell is observed. For small opening angles a flare appears in all four simulated cases which differs in shape according to the frequency. The peak flux is greater the higher the Lorentz factor and energy of the second shell are. The jet break estimation is denoted with a vertical line for both 2 and 5 degree jets.}
\label{c1112lc}
\end{figure*}
\subsection{Optical and radio light curves}
For the four cases described in section 2.3 we produce optical and radio light curves. We examine each case assuming different opening angles of the jet. In Figure \ref{c910lc} and \ref{c1112lc} we present the light curves produced by the radiation code for spherical and collimated expansion of the two colliding shells.
A rebrightening of the optical light curve is observed for $t_{obs}=0.2$ days for the spherical explosion in case 1 (Fig. \ref{c910lc}), which is attributed to the interaction of the two shells when the second one reaches the back of the BM shell. We find that the smaller the opening angle of the jet, the steeper the rise of the flare. This can be understood from the fact that for a jet with small opening angle there is, at a given observer time, less emission still arriving from earlier emission times and higher emission angles. As we show in section 3.5 (Figures \ref{eiflarec10}, \ref{eiflarec12}) early emission time contribution comes from high emission angles on the jet, thus the features of the light curve are less smeared out for a collimated outflow.
This can also explain the difference we observe in the decreasing rate of the flux at late observer times. Since there is no contribution from early emission times, the flux for a jetted outflow will decline faster compared to a spherical explosion. We also observe that the flare is sharper in case 2 compared to case 1 and in case 4 compared to all four cases (Fig. \ref{c910lc} and \ref{c1112lc}). This confirms our hypothesis that the flare is strongly dependent on the dynamical properties of the collision. For case 2, where the forward and reverse shock of the second shell are significantly stronger compared to case 1 (Fig. \ref{etrans}), the change in the flux during the collision of the two shells appears substantially stronger. In cases 3 and 4 we observe the same behaviour in the light curves. In case 4, the higher Lorentz factor together with the higher energy imposed in the second shell, lead to a flare clearly stronger compared to all the other cases for a 2 degrees opening angle of the jet. In that case the timescale of the variation for the flare is $\Delta T/T = 1.08$ for the optical and $\Delta T/T = 1.90$ for the radio, where $\Delta T$ is calculated as the full width at half maximum of the flare and \textit{T} is the observer time of the peak. The relative flux increase with respect to the underlying afterglow is $\Delta F/F = 3.95$ for the optical and $\Delta F/F = 1.04$ for the radio which is significantly reduced due to ssa mechanism. Although extensive comparison to observational data is beyond the scope of this work, we note that this case resembles the afterglow of GRB 060206 which shows an increase in brightness by $\sim 1$ mag 1 hr after the burst, followed by a typical broken power-law decay \citep{Stanek2006,Wozniak2006b}.
Before and after the flares the resulting light curves follow the analytically predicted slopes for a single forward shock as shown in Fig \ref{c910lc}. The light curves at a given frequency depend on the temporal evolution of the characteristic frequencies of the system, that is in our case the synchrotron peak frequency $\nu_{m}$ and the self-absorption frequency $\nu_{sa}$. As described in \citet{Granot2002} these slopes read $t^{1/2}$ for times before the break frequency $\nu_{m}$, crosses the observed frequency $\nu$ and $t^{3(1-p)/4}$ for the time after.
In addition to this behaviour the slope of the light curve steepens by $t^{-3/4}$ after the jet break, in both optical and radio frequency, for small opening angles of the jet. This steepening appears clearer in the radio ($t_{obs,br}=0.0054$ days for 2 deg. jet and $t_{obs,br}=0.068$ days for 5 deg. jet) where the emission comes from late emission times and lower emission angles close to the jet axis. In the optical, where earlier emission times also contribute to the observed flux (as discussed in section 3.5 and shown in Fig. \ref{eiflarec10} and \ref{eiflarec12}), this steepening is delayed and appears at $t_{obs}=0.075$ days almost at the same time as the break frequency $\nu_{m}$ crosses the optical band.
Throughout our calculations and for all cases the critical frequencies satisfy $\nu_{c}>\nu_{obs}$, $\nu_{c}>\nu_{m}$ and $\nu_{c}\gg\nu_{sa}$ at the BM shock and the forward shock of the second shell, where $\nu_{c}$ is the cooling frequency (all critical frequencies are calculated using the formulas found in Table 2 of \citet{Granot2002} for a BM solution. Between $t_{obs}$ of 0.001 and 10 days, $\nu_{c}$ decreases from $10^{20}$ Hz to $10^{17}$ Hz, while $\nu_{m}$ decreases from $10^{16}$ Hz to $10^{10}$ Hz for the BM shock of case 1). Hence, for the frequencies discussed in the present work we are always in the slow cooling regime. In addition to the fact that the shock regions forming at the second shell are a lot thinner than a typical BM profile, this allows us to neglect, with small error, the effect of synchrotron cooling on the electron distribution for the times under consideration and consequently any changes on the self-absorption coefficient.
Furthermore our assumption that inverse Compton (IC) doesn't influence the observed spectrum is confirmed as follows. There are two ways in which IC scattering can change the overall synchrotron spectral component. First by producing an additional emission component at high frequencies and second by dominating the electron cooling and thus reducing the available energy for synchrotron radiation. As discussed in \citet{Sari2001}, the first is estimated by considering the ratio of specific fluxes measured at the peak of the respective flux components, $f_{max}$ for the synchrotron and $f_{max}^{IC}$ for the IC respectively. Then this ratio is $f_{max}^{IC}/f_{max} \sim 1/3 \sigma_{T}nR$, where $\sigma_{T}$ is the Thompson cross-section and $n$ the electron density at distance $R$. This ratio remains well below unity for all the shock surfaces in all four simulations which are taken into consideration. The second factor can be estimated directly from the ratio $\eta\epsilon_{E}/\epsilon_{B}$. As shown by Sari \& Esin, the IC cooling rate will be unimportant compared to synchrotron if $\eta\epsilon_{E}/\epsilon_{B}\ll1$, where $\eta=(\gamma_{c}/\gamma_{m})^{2-p}=(\nu_{c}/\nu_{m})^{-(p-2)/2}$. In our case this ratio is close to unity for the early afterglow but decreases rapidly and remains below unity during our simulation covering the pre and post flaring activity period.
\begin{figure}
\centering
\includegraphics[scale=0.43]{lcvssingleB.eps}
\caption{Optical light curve and single emission images for several snapshots of the two shell system for case 4. The emission time of the snapshots range from $t_{e}=7.64\times10^{6}$ s for snapshot 50 to $t_{e}=1.26\times10^{7}$ s for snapshot 350.}
\label{single}
\end{figure}
\subsection{The shape of the optical flare.}
The flaring activity as observed in the optical follows three stages. At the first stage the Lorentz factor and the thermal energy of the BM matter in front of the second shell increase due to the propagation of the forward shock leading to the sudden rise of the flux. The flare observed in Fig. \ref{c910lc} and \ref{c1112lc}, is attributed to that increase.
This behaviour continues until $t_{obs}=0.23$ days when the reverse shock crosses the back of the second shell. At that time the forward shock starts decelerating and that can be seen as a reduction of the flux, after the initial peak, in the optical light curves. At time $t_{obs}=0.25$ days the shocked BM matter has separated from the shell and now propagates in the BM shell while constantly heating the swept up matter. The flux for that stage of the motion remains almost constant in the optical light curve as the density in front of the forward shock continuously increases and the heated matter compensates in the flux for the deceleration of the forward shock. This remains until the forward shock of the second shell and the BM shell merge at time $t_{obs}=0.31$ days. At that time the optical light curve presents a change in the slope and the flux starts decreasing as the emission now originates from the merged shell while propagating into the ISM.
In figure \ref{single} we plot the optical light curve for case 4 considering a hard-edged jet with opening angle $2\theta_{h}=2$ during the period of the flare and the emission from several snapshots before, after and during the collision of the two shells. The emission from each snapshot consists of a BM shell contribution which peaks at early observer times and the contribution from the second shell which peaks during the flare and at later times. A comparison between the light curve and the emission snapshots shows that immediately after the forward shock of the second shell is created the flare is observed ($t_{e}=7.64\times10^{6}$ s). The stage of the forward shock deceleration and the flux reduction can be seen from $t_{e}=7.64\times10^{6}$ to $t_{e}=1.01\times10^{7}$ s. As soon as the forward shock separates from the second shell, the shape of the emission arising from the forward shock becomes sharper ($t_{e}=1.05\times10^{7}$ s). The third stage of the flare where the forward shock propagates inside the BM matter shapes the plateau which is observed from $t_{obs}=0.25$ to $t_{obs}=0.31$ days and can be seen from $t_{e}=1.05\times10^{7}$ to $t_{e}=1.26\times10^{7}$ s. At that emission time the merger of the two shells has almost completed and the flux starts decaying.
\subsection{Time delay observed between optical and radio flares.}
Comparing the optical to the radio light curves, we notice that the flaring activity occurs with a distinct time delay (approximately 0.1 days) for the latter ones (Fig. \ref{optradiocompare}).
The reason for this is the ssa mechanism. For optical emission the jet is optically thin and the contribution from the second shell is obtained as soon as the forward shock is created while propagating in the BM shell. Below the self-absorption frequency though the jet behaves differently. In the radio the jet is optically thick due to ssa mechanism and the merger becomes visible only after the collision has nearly completed. This results in observing the plateau at the second stage of the optical flare and a sharp peak in the radio flare.
This highlights the significance of taking into account ssa when calculating radio light curves.
In fact any variability resulting from changes in the fluid conditions may manifest in a chromatic fashion \citep{HvE10b}. For that reason the jet break is significantly postponed in the radio light curves (Figs. \ref{c910lc} and \ref{c1112lc}).
\subsection{Emission images}
By directly plotting the relative contributions to the light curve from different parts of the fluid, the reasons for the differences between radio and optical light curves become even more obvious.
The flux at a given observer time is obtained by solving the linear radiative transfer equation through the evolving fluid for a large set of rays. By separately storing the local contributions to the emerging rays, we have created emission images showing exactly the relative contributions of different parts of the jet.
We produce emission images for different observer time and opening angle, for optical ($5\times10^{14}$ Hz) and radio ($10^{8}$ Hz) frequencies covering the time before, during and after the flaring activity is observed. In all the images the jet axis is aligned to the horizontal direction and the observer is at the right end of the horizontal axis.
Which area of the jet is the main contributing area to the emission is strongly frequency dependent. When the frequency of observation lies well above the self absorption frequency, the system is optically thin and the main contribution to the emission is from early emission times and from high emission angles on the jet.
For lower frequencies the system becomes optically thick due to self absorption, hence the main contribution to the emission is from later emission times and the emitting region shifts to lower emission angles closer to the jet axis.
This behaviour is observed throughout the figures \ref{eic12}-\ref{eiflarec12}, where we plot the ring-integrated, absorption-corrected local emission coefficients, as a difference in the contrast between optical and radio emission.
\begin{figure}
\centering
\includegraphics[scale=0.43]{compareD.eps}
\caption{Optical and radio light curve comparison for case 4, assuming a jet opening angle $2\theta_{h}=2$. The optical flare precedes the radio one by 0.1 days. The three different stages of the flare evolution can be seen on the optical flare.}
\label{optradiocompare}
\end{figure}
\begin{figure}
\centering
\includegraphics[scale=0.4, angle=270]{multiplotc12t10_backup.ps}
\caption{Optical $5\times10^{14}$ Hz, (left) and radio $10^{8}$ Hz (right) emission image in case 4 at time $t_{obs}=10$ days. The upper pictures correspond to a spherical explosion and the lower ones to a jet with opening angle $2\theta_{h}=2$. The horizontal axis scales from $2\times10^{17}$ to $1.5\times10^{18}$ cm and the vertical from $5\times10^{15}$ to $3.5\times10^{16}$ cm. The main contribution area for the optical image comes from higher emission angles while for the radio image lower emission angles contribute the most. For a hard-edged jet (lower graphs) the radio image appears stronger, since the main contribution area to the optical at the back of the jet is not taken into account. The remainder of the second shell significantly heated after the traverse of the reverse shock reveals its contribution to the radio image.}
\label{eic12}
\end{figure}
\begin{figure}
\centering
\includegraphics[scale=0.42, angle=270]{multiplotc10_backup.ps}\vspace{5pt}
\includegraphics[scale=0.42, angle=270]{multiplotc10oa20test_backup.ps}\vspace{5pt}
\caption{Optical ($5\times10^{14}$ Hz) and radio ($10^{8}$ Hz) emission images of the two-shell system during the flaring activity for case 2 and jet opening angle 2 degrees (top figures) and 20 degrees (bottom figures). The horizontal axis scales from $2\times10^{17}$ to $5\times10^{17}$ cm and the vertical from $2\times10^{13}$ to $8.5\times10^{15}$ cm for the 2 degrees jet and from $2\times10^{13}$ to $2.5\times10^{16}$ cm for the 20 degrees jet. For each subfigure the bottom images correspond to $t_{obs}=0.23$ days, which is the time the sudden rise of the flux is observed, the middle ones to the weak decay at $t_{obs}=0.28$ days corresponding to the propagation of the forward shock into the BM medium, and the top ones to the fast decay at $t_{obs}=0.35$ days, once the merger has completed.}
\label{eiflarec10}
\end{figure}
\begin{figure}
\centering
\includegraphics[scale=0.42, angle=270]{multiplotc12_backup.ps}\vspace{5pt}
\includegraphics[scale=0.42, angle=270]{multiplotc12oa20_backup.ps}\vspace{5pt}
\caption{Optical ($5\times10^{14}$ Hz) and radio ($10^{8}$ Hz) emission images of the two-shell system during the flaring activity for case 4 and jet opening angle 2 degrees (top figures) and 20 degrees (bottom figures). The horizontal axis scales from $2\times10^{17}$ to $5\times10^{17}$ cm and the vertical from $2\times10^{13}$ to $8.5\times10^{15}$ cm for the 2 degrees jet and from $2\times10^{13}$ to $2.5\times10^{16}$ cm for the 20 degrees jet. For each subfigure the bottom images correspond to $t_{obs}=0.23$ days, which is the time the sudden rise of the flux is observed, the middle ones to the weak decay at $t_{obs}=0.28$ days corresponding to the propagation of the forward shock into the BM medium, and the top ones to the fast decay at $t_{obs}=0.35$ days, once the merger has completed.}
\label{eiflarec12}
\end{figure}
In Fig. \ref{eic12} we plot the optical and radio emission images at observer time $t_{obs}=10$ days, long after the merger has completed, for case 4 for a spherical explosion and for a hard-edged jet of opening angle $2\theta_{h}=2$. We observe that different parts of the merged shell contribute to different parts of the spectrum. The optical contribution emerges mainly from high emission angles of the merged shell whereas radio emission comes mainly from lower emission angles close to the jet axis. This behaviour is also encountered in the emission images derived for the hard-edged jet. In this case, high emission angles are excluded from the calculation of the emission, which affects the optical more strongly than the radio emission.
A contribution from the remainder of the second shell is observed at the radio emission image of the spherical explosion. Although significantly reduced due to ssa and almost two orders of magnitude less than the emission originating from the merged shell, this contribution owns its existence to the traverse of the reverse shock (Fig. \ref{eic12}). As the reverse shock crosses the second shell the Lorentz factor decreases, and by time $t_{obs}=10$ days, when the shell has been significantly decelerated, the peak frequency of the synchrotron spectrum originating from that region is shifted to lower frequencies.
The stronger the reverse shock the stronger the deceleration of the shell. The contribution of the second shell is therefore expected to be higher.
In Fig. \ref{eiflarec10} we show the evolution of the two shells before, during and after the collision phase for case 2, for both radio and optical frequencies assuming opening angles of the jet 2 and 20 degrees (rather than 5, to better capture high-angle emission and illustrate the contrast between narrow and wide jets). The collision appears between observer times $t_{obs}=0.28$ and $t_{obs}=0.35$ days which verifies that the break in the light curves occurs at the same time as the collision of the two shells.
From the emission images it becomes clear that the smearing out of the flare from early emission time contribution becomes less the smaller the opening angle of the jet is. Comparing the two different cases in figure \ref{eiflarec10} (case 2) and figure \ref{eiflarec12} (case 4) it is evident that in the optical case and for large opening angles, early time contribution dominates the emission and drowns out the flaring activity. For small opening angles this effect is suppressed revealing this way the contribution of the second shell while it collides with the BM shell.
In the radio emission images, where the main contributing area to the emission is transferred to lower emission angles, we do not observe this behaviour. Instead as described in section 3.2, in the radio frequency the jet is optically thick due to ssa mechanism and the merger can be seen only when the collision has almost completed. This difference in behaviour between optical and radio, manifests itself as a sharp rise in the radio light curve and a plateau at the flare in the optical light curve (see also Fig. \ref{optradiocompare}).
From the emission images in fig. \ref{eiflarec10} and for a jet with an opening angle 20 degrees we derive the conclusion that for higher frequencies (optical, X-rays), for which the main contributing region to the observed flux shifts to higher emission angles on the jet, a double ring shaped image should appear in GRB late afterglow observations. In this type of image the inner ring carries information from the forward shock emission of the second shell, whereas the outer ring is determined from the emission at earlier stages of the forward shell. For radio frequencies where the observed flux arrives mainly from emission angles close to the jet axis, the distinct contribution of the two shells can only be observed after the merger has completed.
\section{Discussion and Conclusions}
We performed high resolution numerical simulations of late collisions between two ultra relativistic shells and produced optical and radio light curves and emission images for a spherical explosion case and different opening angles of a hard-edged jet. The AMR technique allowed us to reach high resolution and properly resolve the shocks developed during the merger of the shells.
The simulations have shown that different values of the Lorentz factor and energy of the second shell can significantly change the characteristics of the variability on the light curves. We demonstrate that for small opening angles, for which the flux is not smeared out by early emission time contribution, a flare appears at the light curve. The onset of the flare is found to be strongly dependent on the strength of the forward and reverse shock of the second shell and to become stronger the higher the Lorentz factor and energy of the second shell are. For case 4 in which the Lorentz factor and energy of the second shell are the highest in our simulations, the relative increase of the flux in respect to the underlying afterglow is $\Delta F/F = 3.95$ for the optical and $\Delta F/F = 1.04$ for the radio. The timescale of the variation of the flare is $\Delta T/T = 1.08$ and $\Delta T/T = 1.90$ respectively. The shape of the flare is understood through the dynamical simulations and the different stages at the light curve are associated with different parts of the collision process. We show that the difference in shape between the optical and radio flare as well as the time difference observed between them is a direct result of the ssa mechanism and the angle dependence of the emission.
We predict that this type of behaviour should appear in late afterglow observations as a two-ring feature in spatially resolved optical emission images, although for the time being the resolution required would be unrealistic.
Athough a detailed numerical approach is required to fit observational data with the numerical results, a straightforward comparison was made between case 4 of our simulations and the optical flare observed in the afterglow of GRB 060206 showing strong resemblance in the magnitude and time variation of the flare. New low-frequency facilities such as ALMA are expected to provide enough data to compare with the radio light curves as well.
The details which determine the jet collimation as the jet propagates into the circumburst medium are currently not well understood. In this work we have used the same opening angle for both shells and we defer analysis of a scenario with different opening angles to future work. Such a study would require detailed simulations in 2D.
We have not explicitly discussed X-ray flares, although for these the largest amount of observational data is available. The X-ray emission is influenced by electron cooling (i.e. lies above the cooling break in the synchrotron spectrum) and is therefore also more sensitive to the details of particle acceleration than optical and radio emission. Especially in the case of multiple shocks, it becomes difficult to implement an approach to radiation that correctly captures all relevant physics in a simple parametrization.
Many afterglows show X-ray flares superimposed on a broken power-law where a single slope describes the region before and after the flare and are often characterized by small timescale variations, $\Delta T/T \ll 1$ (e.g. Burrows et al. 2007). However, there are occasions, such as the GRB 060206, where the afterglow light curve presents a clear difference between the pre and post flare region with a significantly increased timescale variation, indicating an energy injection to the external shock as a probable cause for the generation of the flare \citep{Monf2006}.
\section{Acknowledgments}
We thank R.A.M.J. Wijers and Petar Mimica for useful discussion. Computations performed on JADE (CINES) in DARI project c2010046216 and on HPC VIC3 at K.U.Leuven. These results were obtained in the framework of the Project GOA/2009/009 (K.U.Leuven). ZM acknowledges support from HPC Europa, project 228398. HJvE is supported by NASA under Grant No. 09-ATP09-0190 issued through the Astrophysics Theory Program (ATP).
|
\section{Introduction}
\label{sec:introduction}
\IEEEPARstart{I}{n} view of the decentralized nature of future and emerging wireless networks, non-cooperative game theory has become an important tool to analyze distributed problems in networks whose nodes cannot be assumed to adhere to centrally controlled protocols. The main goal has been to develop policies and algorithms that nodes can use to optimize their resources (power, bandwidth, etc.) on their own, so, following \cite{LDA09}, the questions that arise are
\begin{inparaenum}[\itshape a\upshape)]
\item whether there exist ``equilibrial'' policies which are stable against unilateral deviations;
\item whether these (Nash) equilibria are unique; and
\item whether they can be reached by distributed learning algorithms that require only local information.
\end{inparaenum}
Accordingly, an important paradigm which has attracted significant interest in the wireless communications literature concerns the allocation of power over orthogonal communication channels \cite{LDA09,GT06,MPS07}.
From a centralized viewpoint, this is a relatively well-studied subject, especially with respect to optimal power allocation schemes which allow users to reach the boundary of the rate region assuming full channel knowledge and central control \cite{CV93, TH98}. On the other hand, more recent examinations \cite{ABSA02,SPB08-jsac,SPB09-sp} focus on socially stable power allocation policies because, even if the globally optimal, capacity-achieving power profile is known and used, it might be unstable under deviations by selfish users (and thus useless in a decentralized setting).
In this paper, we consider the problem of uplink communication in multi-user networks consisting of several receivers that operate on distinct, non-interfering channels, and we focus on giving definitive answers to points (a)\textendash(c) above, analyzing the equilibrial structure of the problem and its convergence aspects. Despite its apparent simplicity, this \ac{PMAC} model has several relevant applications such as, for instance, in 802.11-based \acp{WLAN} with non-overlapping channels \cite{CHD10,LSdLD08}, distributed soft handoffs in cellular systems \cite{PBLD09}, distributed power allocation in \acp{DSL} \cite{YGC02}, and, finally, in throughput-maximizing power control in \ac{MC-CDMA} systems \cite{MCPS06}.
Our analysis will focus on the \ac{SUD} scheme where the transmitted signal of each user is decoded separately by the receiver(s) who treat the incoming signal of other users as additive (Gaussian) noise. The main reason for using \ac{SUD} instead of \ac{SIC} is that the former is known to have lower decoding complexity and signalling overhead than the latter \textendash\ a consequence of \ac{SUD} not having to broadcast the decoding order to the transmitters \cite{BelmegaThesis}. As a result, \ac{SIC}-based schemes suffer from scalability issues, especially when there are several receivers and/or the channel is highly time-varying.
In this context, non-cooperative power allocation games for static Gaussian \acp{IC} have been studied in a series of related papers \cite{SPB08-jsac,SPB09-sp}. There, the existence of a Nash equilibrium is a consequence of the convexity properties of the users' achievable rates and follows from the general theory of \cite{Ro65}. In fact, under suitable (but stringent) conditions on the channel matrices, it was shown that this equilibrium is unique and that iterative water-filling algorithms converge to it.
Formally, the static \ac{PMAC} is a special case of this \ac{IC} framework, but the conditional analysis of \cite{SPB08-jsac,SPB09-sp} almost always fails for static \ac{PMAC} models. Thus, although the (global) capacity region of this channel is well-understood \cite{CV93,TH98,GT06}, the channel's distributed version remains unresolved. A first attempt to remedy this was carried out in \cite{PBLD09} where it was shown that an associated power allocation game admits an exact potential function \cite{MS96} whose minimum corresponds to the game's Nash equilibria. However, this potential is \emph{not} strictly convex, so Nash equilibrium uniqueness might fail along with the uniqueness conditions of \cite{SPB08-jsac,SPB09-sp}. Rather surprisingly, it turns out that this is \emph{not} the case:
even though these conditions do not hold in the static \ac{PMAC} context, the Nash equilibrium of the static \ac{PMAC} game \emph{is} unique (Theorem~\ref{thm:uniqueness}).
In itself, uniqueness allows us to characterize the system's behavior at equilibrium, but it does not provide a way of actually getting there. Regarding such convergence issues, the authors of \cite{ABSA02} considered a single channel with pricing and exhibited power control algorithms which converge to equilibrium under ``mild-interference'' conditions. Similarly, one of the main results of \cite{YGC02} was to show that if the transmitters know the local channel state and the overall inteference-plus-noise covariance matrix, then, subject to similar ``mild-interference'' conditions, iterative water-filling converges to the equilibrium set of the game (a result which was then enhanced in \cite{YRBC04} by dropping this condition for a modified water-filling scheme).
Instead of taking a water-filling approach, we present a simpler learning scheme based on the replicator dynamics of evolutionary game theory \cite{TJ78} which involves the same (often less) information from the side of the players, and which does not require them to solve a nonlinear water-filling problem. Dynamics of this type have been studied extensively in finite \cite{FL98} and continuous population games \cite{Sandholm11}, but, in \emph{nonlinear games} (such as the one we have here), their properties are not as well understood. Nonetheless, by taking a modified version of the users' utilities, we show that the replicator dynamics converge to an $\varepsilon$-neighborhood of the game's (a.s.) unique equilibrium exponentially fast, i.e. in time $\bigoh(\log(1/\varepsilon))$ (Theorem \ref{thm:convergence}).
In the context of fading however, the static game and the corresponding replicator dynamics lose much of their relevance because variations due to fading open the door to stochasticity. To account for this randomness, we study both fast and block-fading models. Using techniques from the theory of stochastic approximation \cite{Borkar08}, we show that by properly adjusting their learning scheme, users converge to the (unique) equilibrium of an averaged game whose payoff functions correspond to the users' achievable ergodic rates (Theorems \ref{thm:stochconvergence} and \ref{thm:ergconvergence}).
Our convergence analysis is based on a novel game-theoretic result which is of independent interest: {\em in games which admit a (star-)convex potential function, the replicator dynamics converge to (the game's unique) equilibrium at an exponential rate} (Theorem \ref{thm:convspeed}). To the best of our knowledge, this is the fastest convergence rate that has been established for the replicator dynamics in the current state of the art \cite{Sandholm11}.
\paragraph*{Notational Conventions}
If ${\mathbb R}^{\mathcal{S}}$ is the vector space spanned by the set $\mathcal{S}= \{s_{\alpha}\}_{\alpha=1}^{S}$ and $\{e_{\alpha}\}_{\alpha=1}^{S}$ denotes its canonical basis, we will use $\alpha$ to refer interchangeably to either $s_{\alpha}$ or $e_{\alpha}$, and we will identify the set $\Delta(\mathcal{S})$ of probability measures on $\mathcal{S}$ with the standard $(S-1)$-dimensional simplex of ${\mathbb R}^{\mathcal{S}}$: $\Delta(\mathcal{S}) \equiv \{x\in{\mathbb R}^{\mathcal{S}}: \sum\nolimits_{\alpha} x_{\alpha} =1 \text{ and } x_{\alpha}\geq0\}$. Finally, we will employ Latin indices for players ($k,\ell,\dotso$), while reserving Greek ones ($\alpha,\beta,\dotso$) for their (``pure'') strategies; and when summing over $\alpha\in\mathcal{A}_{k}$, we will simply write $\sum^{k}_{\alpha} \equiv \sum_{\alpha\in\mathcal{A}_{k}}$.
\section{System Model}
\label{sec:model}
Following \cite{PBLD09}, the basic setup of our model is as follows: we consider a finite set $\mathcal{K}=\ito{K}$ of wireless single-antenna transmitters (the {\em players} of the game) who wish to transmit to a group of single-antenna receivers (possibly clustered as a single receiver). Each of these receivers operates on a given channel $\alpha\in\mathcal{A}\equiv\ito{A}$ (assumed to be orthogonal, typically in the frequency domain), and each user $k\in\mathcal{K}$ may transmit over a subset $\mathcal{A}_{k}\subseteq\mathcal{A}$ of these channels (with $A_{k}\equiv \card(\mathcal{A}_{k})\geq2$).
In particular, if $x_{k\alpha}\sim\mathcal{CN}(0,p_{k\alpha})$ is the transmitted message of user $k$ on channel $\alpha\in\mathcal{A}_{k}$ and $h_{k\alpha}$ denotes the respective channel coefficient, then the received signal on channel $\alpha$ will be $y_{\alpha} = \sum\nolimits_{k} h_{k\alpha} x_{k\alpha} + z_{\alpha}$, where $z_{\alpha} \sim \mathcal{CN}(0,\sigma_{\alpha}^{2})$ denotes the thermal noise. Accordingly, user $k\in\mathcal{K}$ can split his transmitting power among the channels $\alpha\in\mathcal{A}_{k}$ subject to the constraint:
\begin{equation}
\label{eq:powerconstraint}
\textstyle
\sum\nolimits^{k}_{\alpha} p_{k\alpha} \leq P_{k},
\end{equation}
where $p_{k\alpha} = \ex\big[|x_{k\alpha}|^{2}\big]$ represents the power with which the user transmits on channel $\alpha$, and $P_{k}$ is his maximum power. As a result, the {\em power allocation} of user $k$ will be given by the point $p_{k} = \sum^{k}_{\alpha} p_{k\alpha} e_{k\alpha}\in{\mathbb R}^{\mathcal{A}_{k}}$, and, analogously, the {\em power profile} which collects all users' power allocations will be re\-pre\-sented by $p = (p_{1},\dotsc,p_{K})\in\prod_{k}{\mathbb R}^{\mathcal{A}_{k}}={\mathbb R}^{Q}$ with $Q \equiv \sum\nolimits_{k} A_{k}$.
In this context, our performance metric will be the users' achievable transmission rates, which depend on their \ac{SINR}:
\begin{equation}
\label{eq:sinr}
\sinr_{k\alpha}(p) = \frac{g_{k\alpha} p_{k\alpha}}{\sigma_{\alpha}^{2} + \sum\nolimits_{\ell\neq k} g_{\ell\alpha} p_{\ell\alpha}},
\end{equation}
with $g_{k\alpha}=|h_{k\alpha}|^{2}$ denoting the channel gain coefficient of user $k$ in channel $\alpha\in\mathcal{A}_{k}$. Clearly, the users' achievable rates will depend on their power allocation policies through their \ac{SINR}, but the exact dependence hinges on the time-variability of the channel gain coefficients $g_{k\alpha}$.
At one end, we will study \emph{static} channels, i.e. channels whose coherence time is much larger than both the self-decodable block duration and the power updating period. At the other extreme, we will also consider \emph{fast-fading} channels where the coherence time is much shorter than those characteristic times; here, what matters is the ergodic value of the \ac{SINR} and the corresponding rate. Finally, we will also analyze the more interesting intermediate case, where the coherence time is greater than the block length but comparable to the update time \textendash\ hence allowing blocks to be decoded using the instantaneous channel values in (\ref{eq:sinr}), but also introducing stochasticity in the game.
\subsection{Static Channels}
\label{subsec:static.game}
We will start with the case of static channels employing the \acf{SUD}. In this case, the spectral efficiency of user $k$ in the power profile $p$ will be given by \cite{PBLD09,ABSA02}:
\begin{equation}
\label{eq:payoff}
\textstyle
u_{k}(p)
= \sum\nolimits_{\alpha} b_{\alpha} \log\Big(1 + \sinr_{k\alpha}(p)\Big)
\end{equation}
where $b_{\alpha}>0$ represents the bandwidth of channel $\alpha\in\mathcal{A}_{k}$ and the channel gains $g_{k\alpha}$ are drawn once and for all from a continuous distribution on $[0,\infty)$ at the outset of the game and remain fixed for the duration of the transmission \cite{SPB08-jsac,PBLD09}.
Then, to maximize their spectral efficiency (\ref{eq:payoff}), users saturate (\ref{eq:powerconstraint}) by transmitting at the highest possible power \cite{PBLD09}, so we are led to the \emph{static} \ac{PMAC} game $\mathfrak{G}\equiv\mathfrak{G}\big(\mathcal{K},\{\Delta_{k}\},\{u_{k}\}\big)$ where:
\begin{enumerate}
\item The {\em players} of $\mathfrak{G}$ are the transmitters $\mathcal{K} = \ito{K}$.
\item The {\em strategy space} of player $k$ is the (scaled) simplex $\Delta_{k} \equiv P_{k}\,\Delta(\mathcal{A}_{k}) = \left\{p_{k}\in{\mathbb R}^{\mathcal{A}_{k}}: p_{k\alpha}\geq 0 \text{ and }\sum_{\alpha} p_{k\alpha}=P_{k}\right\}$ of power allocation vectors; the game's space of strategy profiles $p=(p_{1},\dotsc,p_{K})$ will then be $\Delta\equiv\prod_{k}\Delta_{k}$.
\item The players' {\em payoffs} (or {\em utilities}) are given by the spectral efficiencies $u_{k}\colon\Delta\to{\mathbb R}$ of (\ref{eq:payoff}).
\end{enumerate}
Of course, the game $\mathfrak{G}$ defined in this way does not adhere to the original normal form of Nash in the sense that
\begin{inparaenum}[\itshape a\upshape)]
\item players are not mixing probabilities over a finite set of possible actions, and
\item even though the players' strategy spaces are simplices, their payoffs are not (multi)linear.
\end{inparaenum}
On the other hand, with $u_{k}$ being concave in $p_{k}$, it is easy to see that $\mathfrak{G}$ is itself {\em concave} in the sense of Rosen \cite{Ro65}. Also, as was shown in \cite{PBLD09}, $\mathfrak{G}$ possesses an \emph{exact potential} \cite{MS96}, i.e. a function $\Phi\colon\Delta\to{\mathbb R}$ such that:
\begin{equation}
\label{eq:potentialdef}
u_{k}(p_{-k};p_{k}') - u_{k}(p_{-k};p_{k}) = \Phi(p_{-k};p_{k}) - \Phi(p_{-k};p_{k}'),
\end{equation}
for all users $k\in\mathcal{K}$, and for all power allocations $p_{k},p_{k}'\in\Delta_{k}$ of user $k$ and $p_{-k}\in\Delta_{-k}\equiv\prod_{\ell\neq k}\Delta_{\ell}$ of $k$'s opponents;
in fact, the analysis of \cite{PBLD09} also provides the expression:
\begin{equation}
\label{eq:potential}
\textstyle
\Phi(p) = -\sum\nolimits_{\alpha} b_{\alpha}
\log\left(1 + \sum\nolimits_{k} g_{k\alpha} p_{k\alpha}\big/ \sigma_{\alpha}^{2}\right).
\end{equation}
\subsection{Fading Channels}
\label{subsec:fading.game}
As discussed above, the non-static models that we will examine are block-fading and fast-fading channels.
\subsubsection{Block-fading channels}
In this case, the coefficients $g_{k\alpha} \equiv g_{k\alpha}(t)$ remain constant over an entire transmission block, so, assuming the transmitter knows (\ref{eq:sinr}) for each block through feedback, the users' utilities are still given by (\ref{eq:payoff}), with different gains $g_{k\alpha}$ at each self-decodable block.\footnote{Note that we do not assume any delay constraints at the receiver \cite{OSW94}; in this way, reliable information (a la Shannon) can be transmitted over each block.} As such, (\ref{eq:potential}) is still a potential for the (now evolving) game $\mathfrak{G}(t)$, the only difference being that $\Phi$ will evolve over time following the channels and the game.
\subsubsection{Fast-fading channels}
In this regime, the coefficients $h_{k\alpha}\equiv h_{k\alpha}(t)$ evolve ergodically at a rate which is much faster than the characteristic length of a transmission block, so the ``instantaneous'' utilities (\ref{eq:payoff}) lose their relevance. Instead, and assuming for simplicity that users saturate their power constraints, their utilities will be given by the \emph{ergodic rates} of \cite{GV97}:
\begin{equation}
\label{eq:ergpayoff}
\textstyle
\overline u_{k}(p) = \sum\nolimits_{\alpha} b_{\alpha} \ex_{g}
\left[\log
\left(1+\sinr_{k\alpha}(p)\right)
\right].
\end{equation}
We thus obtain the {\em ergodic game} $\overline\mathfrak{G}\equiv\big(\mathcal{K},\{\Delta_{k}\},\{\overline u_{k}\}\big)$, which has the same strategic structure as its static counterpart $\mathfrak{G}$ but payoffs given by (\ref{eq:ergpayoff}) instead of (\ref{eq:payoff}).
In fact, as in the static case, $\overline\mathfrak{G}$ admits the exact potential:
\begin{equation}
\label{eq:ergpotential}
\textstyle
\overline\Phi(p)
\equiv \ex_{g}[\Phi(p)]
=-\sum\nolimits_{\alpha} b_{\alpha} \ex_{g}\big[
\log\big(1+ \sum\nolimits_{k} g_{k\alpha} p_{k\alpha}\big/\sigma_{\alpha}^{2}\big)
\big],
\end{equation}
whose form depends on the law of the $g_{k\alpha}$. Thus, with $h_{k\alpha}\sim\mathcal{CN}(0,\sqrt{\gamma_{k\alpha}})$, $\gamma_{k\alpha}\geq0$, the coefficients $g_{k\alpha} = |h_{k\alpha}|^{2}$ will be $\chi^{2}$-distributed, and the calculations of \cite[eq. (11)]{MS03} yield:
\begin{proposition}
\label{prop:gaussianpotential}
In i.i.d. Gaussian fast-fading channels with $h_{k\alpha}\sim\mathcal{CN}(0,\sqrt{\gamma_{k\alpha}})$, the ergodic potential $\overline\Phi$ is:
\begin{equation}
\label{eq:Gaussianpotential}
\overline\Phi (p)=
-\sum\nolimits_{k,\alpha} b_{\alpha} \zeta(r_{k\alpha}^{-1})
\prod\nolimits_{\ell\neq k}(1-r_{\ell\alpha}/r_{k\alpha})^{-1}
\end{equation}
where $r_{k\alpha} \!=\! \gamma_{k\alpha}p_{k\alpha}\big/\sigma_{\alpha}^{2}$ and $\zeta(x) \!\equiv \!\int_{0}^{\infty}\!(x+t)^{-1} e^{-t} \,d t \!=\! - e^{x} \Ei(-x)$.
\end{proposition}
This proposition will be crucial in the numerical calculations of Section~\ref{sec:numerics}. For posterity, we only note here that (\ref{eq:Gaussianpotential}) implies that $\overline\Phi$ is strictly convex \cite{BLDH10}, even though, in general, $\Phi$ is not.
\section{Equilibrium Analysis}
\label{sec:equilibrium}
We begin with the notion of \emph{Nash equilibrium}:
\begin{definition}
A power profile $q\in\Delta$ will be a {\em Nash equilibrium} of the game $\mathfrak{G}$ (resp. $\overline\mathfrak{G}$) when:
\begin{equation}
\label{eq:Nash}
u_{k}(q) \geq u_{k}(q_{-k};q_{k}'),\quad
\text{(resp. $\overline u_{k}(q) \geq \overline u_{k}(q_{-k};q_{k}')$)}
\end{equation}
for all $k\in\mathcal{K}$ and for all $q_{k}'\in\Delta_{k}$. In particular, if $q$ satisfies the strict version of (\ref{eq:Nash}) for all $q_{k}'\neq q_{k}$, it will be called \emph{strict}.
\end{definition}
Given that $\mathfrak{G}$ (resp. $\overline\mathfrak{G}$) admits a convex potential, \emph{its equilibrium set will coincide with the minimum set of $\Phi$} (resp.~$\overline\Phi$) \cite{Neyman97}. As such, the existence of an equilibrium is guaranteed, and this is already important from a practical point of view because learning protocols would never converge otherwise. Our goal in this section will be to show that these equilibria are essentially \emph{unique}, thus ensuring the system's \emph{predictability} \textendash\ a crucial feature for performance evaluation, QoS guarantees, etc.
\subsection{Static Channels}
\label{subseq:static.equilibrium}
With regards to the static potential $\Phi$, it is easy to see that two power profiles $p,p'\in\Delta$ will have $\Phi(p) = \Phi(p')$ whenever
\begin{equation}
\label{eq:degconstraint}
\textstyle
\sum\nolimits_{k} g_{k\alpha} (p_{k\alpha}' - p_{k\alpha}) = 0\quad
\text{for all $\alpha\in\mathcal{A}$}.
\end{equation}
In that case, $\Phi$ will not be strictly convex, so its minimum set might fail to be a singleton as well. More precisely, if we set $z=p'-p$, we will also have $\sum_{\alpha}^{k} z_{k\alpha}=0$ for all $k\in\mathcal{K}$, so, on account of (\ref{eq:degconstraint}) above, $\Phi$ will not be strictly convex if the following linear system admits a non-zero solution in $z$:
\begin{align}
\label{eq:constraints}
\textstyle
\sum\nolimits_{k} g_{k\alpha} z_{k\alpha} = 0,\quad \alpha\in\mathcal{A};
&&
\textstyle
\sum\nolimits_{\alpha} z_{k\alpha} = 0,\quad k\in\mathcal{K}.
\end{align}
Since $z\in{\mathbb R}^{Q}$, $Q=\sum_{k}A_{k}$, and the above $A+K$ constraints are independent (a.s.), we see that if $Q- A - K >0$, then $\Phi$ cannot be strictly convex \textendash\ see \cite{ValueTools11b} for more details. In fact, the quantity $\ind(\mathfrak{G}) \equiv Q - A -K$ will be called the \emph{degeneracy index} of the game $\mathfrak{G}$, and the condition $\ind(\mathfrak{G}) >0$ means that \emph{if the number of links $Q$ exceeds the number of channels plus transmitters $A+K$, then the game's potential is not strictly convex}.
In typical uplink scenarios of practical interest (e.g. single-receiver \acs{OFDM}), each user can access all channels, so $A_{k}=A$ for all $k$, and, hence $Q = AK > A + K$ (except in small $2\times2$ systems). This implies that degeneracy appears almost always, so in the absence of strict convexity, a promising way to determine whether the \ac{PMAC} game admits a \emph{unique} equilibrium would be to use the conditions of \cite{SPB08-jsac,SPB09-sp,LP06} where one constructs a certain matrix $\mathbf{S}$ from the channel gain coefficients and tries to show that said matrix has a spectral radius $\rho(\mathbf{S})<1$. However, as is shown in \cite{ValueTools11b}, this spectral radius exceeds $1$ (a.s.), so the results of \cite{SPB08-jsac,SPB09-sp,LP06} do not apply to our problem. Still, we have:
\begin{theorem}
\label{thm:uniqueness}
The static game $\mathfrak{G}$ admits a unique Nash equilibrium (a.s.).
\end{theorem}
\begin{IEEEproof}[Sketch of Proof]
Let $p\in\Delta$ and consider the \emph{(multi)graph} $\mathcal{G}(p) = (\mathcal{A}, \mathcal{E}(p))$ whose vertices are the network's receivers and whose edge set $\mathcal{E}(p)$ is the \emph{multiset sum} $\mathcal{E}(p) = \biguplus_{k} \mathcal{E}_{k}(p_{k})$ where each $\mathcal{E}_{k}(p_{k})$ is a star graph on the nodes $\alpha\in\mathcal{A}$ to which $p_{k}$ assigns positive power $p_{k\alpha}>0$ (i.e. all channels to which $k$ transmits with positive power are star-connected and these graphs are su\-per\-imposed for all $k\in\mathcal{K}$). If $p$ is equilibrial, $\mathcal{G}(p)$ has to be acyclic \cite{ValueTools11b}, so $p$ must lie in the interior of an at most $(A-1)$-dimensional face of $\Delta$ (a.s.); our assertion then follows from a dimension-counting argument (see \cite{ValueTools11b} for details).
\end{IEEEproof}
\subsection{Fading Channels}
\label{subsec:ergodic.equilibrium}
\subsubsection{Block-fading channels}
As we discussed in Section \ref{subsec:fading.game}, the time-varying version of (\ref{eq:potential}) which corresponds to block-fading channels $g_{k\alpha} \equiv g_{k\alpha}(t)$ is a potential for the block-fading game $\mathfrak{G}(t)$. Accordingly, Theorem \ref{thm:uniqueness} implies:
\begin{corollary}
At each channel realization, the block-fading game $\mathfrak{G}(t)$ admits a unique Nash equilibrium (a.s.).
\end{corollary}
\subsubsection{Fast-fading channels}
On the other hand, the averaging effect in the ergodic rates (\ref{eq:ergpayoff}) can be used to show that the ergodic potential $\overline\Phi$ is, in fact, \emph{strictly} convex. This gives:
\begin{theorem}[\cite{BLDH10}]
\label{thm:erguniqueness}
The ergodic game $\overline\mathfrak{G}$ admits a unique Nash equilibrium.
\end{theorem}
\section{Learning Dynamics and Convergence to Equilibrium}
\label{sec:convergence}
Although Theorems \ref{thm:uniqueness} and \ref{thm:erguniqueness} guarantee equilibrium uniqueness, it is not at all clear whether users will be able to calculate this equilibrium in decentralized environments where only partial/local information is available at the terminal (e.g., as in distributed or partially distributed cognitive radio networks). Consequently, our goal in this section will be to present a simple distributed learning scheme which allows users to converge to equilibrium, and to determine the speed of this convergence.
This question has attracted considerable interest from the point of view of learning, and two of the most well-studied paradigms are \ac{BR} algorithms and \ac{RL} \cite{BS97,FL98,Sandholm11}. In standard \ac{BR} schemes \cite{Sandholm11}, players are assumed to monitor their opponents' power allocation policies and respond optimally to them (with respect to their individual utilities). Unfortunately (and in addition to the ``perfect monitoring'' requirement), it is quite hard to calculate these best responses in large games, so the applicability of this approach to large decentralized networks is quite limited. To circumvent these limitations (in static channels at least), a promising solution lies in the water-filling approach of \cite{YGC02,SPB08-jsac,SPB09-sp,YRBC04} where users only need their local channel and overall noise-plus-interference covariance matrix. In that case however
\begin{inparaenum}[\itshape a\upshape)]
\item users must solve a non-convex fixed point problem at each step; and
\item convergence is conditional on the interference being low enough (except in \cite{YRBC04}).
\end{inparaenum}
In fact, the conditions of \cite{SPB08-jsac,SPB09-sp} do \emph{not} hold in the \ac{PMAC} case \cite{ValueTools11b}, while the approach of \cite{YGC02} breaks down for large numbers of users \cite{HH10}.
On the other hand, \ac{RL} algorithms (such as regret-matching \cite{HMC00}) rely on the players knowing their (possibly fictitious) payoffs. Thanks to this information (which, however, is often hard to come by), these algorithms enjoy strong convergence properties in potential games. However, such learning algorithms have been designed for discrete action sets, so it is very hard to adapt them to games with continuous action spaces (such as the ones we are considering here).
To overcome these limitations, our starting point will be the replicator dynamics of evolutionary game theory \cite{TJ78,FL98,Sandholm11}. The reinforcement aspect of these dynamics does suffer from the same drawback as most \ac{RL} algorithms (i.e. it applies only to \emph{finite} action sets), but, by exploiting the simplicial structure of the game and its potential, we derive a learning scheme which applies to \emph{continuous} action spaces and which allows users to converge to equilibrium \emph{unconditionally} and \emph{exponentially quickly} (Theorems \ref{thm:convergence} and \ref{thm:ergconvergence}).
\subsection{Static Channels}
\label{subsec:static.convergence}
Since the replicator equation applies to discrete sets (such as $\mathcal{A}_{k}$), a reasonable channel-specific utility would be:
\begin{equation}
\label{eq:nodepayoffs}
u_{k\alpha}(p)
= b_{\alpha}\log\Big(1+\sinr_{k\alpha}(p)\Big)
\end{equation}
which leads to the replicator equation:
\begin{equation}
\label{eq:naiveRD}
\frac{dp_{k\alpha}}{dt}
= p_{k\alpha}(t) \left(u_{k\alpha}(p(t)) - P_{k}^{-1}\sum\nolimits^{k}_{\beta} p_{k\beta}(t) u_{k\beta}(p(t))\right),
\end{equation}
whose second term ensures that $p(t)\in\Delta$ for all $t\geq0$. Unfortunately, the utility $u_{k}$ of eq.~(\ref{eq:payoff}) is \emph{not} a convex combination of the $u_{k\alpha}$, so (\ref{eq:naiveRD}) is not well-behaved w.r.t. the game $\mathfrak{G}$ either \textendash\ for instance, Nash equilibria are not stationary.
Instead, given that each user invariably seeks to unilaterally increase his utility, we will consider the \emph{marginal utilities}:
\begin{equation}
\label{eq:derpayoff}
v_{k\alpha}(p)
\equiv \frac{\partial u_{k}}{\partial p_{k\alpha}}
= \frac{b_{\alpha}g_{k\alpha}}{\sigma_{\alpha}^{2} + \sum\nolimits_{\ell} g_{\ell\alpha} p_{\ell\alpha}}.
\end{equation}
Since player $k$ can calculate $v_{k\alpha}(p)$ by means of $\sinr_{k\alpha}$ and $g_{k\alpha}$ alone (the bandwidths $b_{\alpha}$ are assumed fixed and known), any dynamics based on the $v_{k\alpha}$'s will be inherently distributed (and simpler than solving a water-filling problem to boot). We will thus consider the replicator equation:
\begin{equation}
\label{eq:RD}
\frac{dp_{k\alpha}}{dt}
\textstyle= p_{k\alpha}(t) \big(v_{k\alpha}(p(t)) - v_{k}(p(t))\big),
\end{equation}
where $v_{k}$ denotes the user average $v_{k}(p) = P_{k}^{-1}\sum\nolimits^{k}_{\beta} p_{k\beta} v_{k\beta}(p)$.
\begin{remark}[Comparison to other power updating schemes]
The replicator equation (\ref{eq:RD}) is clearly quite unlike the water-filling schemes of \cite{YGC02,SPB08-jsac,SPB09-sp,YRBC04}. Closer in spirit to (\ref{eq:RD}) are the algorithms developed in the 90's with the goal of minimizing transmitting power by comparing the instantaneous \ac{SINR} to a target value and iteratively updating the power proportionally to this difference \cite{FM93,BCP94,Yates95}; still, there is little overlap with these algorithms, both in terms of setup and convergence.
\end{remark}
\begin{remark}
The marginal utilities $v_{k\alpha}$ are similar but {\em not} equal to the \ac{SINR} (\ref{eq:sinr}) of user $k$ at a given channel $\alpha\in\mathcal{A}_{k}$, and they do not coincide with the popular metric of total rate per unit power either \cite{MCPS06,MPS07}. These metrics can all be calculated based on the same feedback (and might appear more appealing than $v_{k\alpha}$), but we shall see that it is precisely the (perhaps unconventional) choice of the marginal utilities $v_{k\alpha}$ that leads to convergence.
\end{remark}
\begin{figure*}
\centering
\subfigure[Convergence of the replicator dynamics in static channels.]{
\label{subfig:portrait}
\includegraphics[width=0.825\columnwidth]{PhasePortrait.pdf}}
\hspace{30pt}
\subfigure[Spectral efficiency over time for different initial configurations.]{
\label{subfig:speed.power}
\includegraphics[width=.97\columnwidth]{PowerRates.pdf}}
\caption{Convergence to equilibrium in a $2\times2$ game with static channels. The dashed contours in Fig.~\ref{subfig:portrait} are the level sets of the K-L divergence w.r.t. the game's equilibrium (red dot), while $p_{1}, p_{2}$ represent the normalized power allocation of each user. In Fig.~\ref{subfig:speed.power}, we plot the spectral efficiency $u_{k}$ of each user as a function of time for three randomly drawn initial configurations; as can be seen in the inlay, the equilibration rate $\lambda(t)$ coincides with the value predicted by Theorem \ref{thm:convergence} (solid black line).
\vspace{-15pt}}
\label{fig:convergence}
\end{figure*}
The first important property of (\ref{eq:RD}) is that its rest points satisfy the (waterfilling) condition $v_{k\alpha}(p) = v_{k\beta}(p)$ for all nodes $\alpha,\beta\in\supp(p_{k})$ to which user $k$ transmits with positive power. Hence, by the \ac{KKT} conditions of \cite{PBLD09}, we see that \emph{Nash equilibria of $\mathfrak{G}$ are stationary in (\ref{eq:RD})}.
The converse of this statement is not true: every vertex of $\Delta$ is stationary \emph{without} necessarily being a Nash equilibrium. Nonetheless, the game's (unique) Nash equilibrium is the {\em only} attracting state of the dynamics (see Appendix~\ref{apx:convspeed} for the proof):
\begin{theorem}
\label{thm:convergence}
Let $q\in\Delta$ be the (a.s.) unique equilibrium of $\mathfrak{G}$. Then, every interior solution orbit of the replicator dynamics (\ref{eq:RD}) converges to $q$; moreover, there exists $c>0$ such that
\begin{equation}
\label{eq:convspeed}
D_{\textup{KL}}\left(q \,\|\, p(t)\right) \leq D_{\textup{KL}}\left(q\,\|\, p(0)\right)\,e^{-ct}\,\,
\text{for all $t\geq0$,}
\end{equation}
where $D_{\textup{KL}}$ is the Kullback-Leibler divergence. In other words, replicator trajectories converge to an $\varepsilon$-neigh\-borhood of a Nash equilibrium in time $\bigoh(\log(1/\varepsilon))$.
\end{theorem}
\setcounter{remark}{0}
\begin{remark}
To the best of our knowledge, the exponential convergence rate of Theorem \ref{thm:convergence} (see also Theorem \ref{thm:convspeed} in App.~\ref{apx:convspeed} and Fig.~\ref{fig:convergence}) is the fastest known estimate for the replicator dynamics (see \cite{Sandholm11} for a review of the state of the art).
\end{remark}
\begin{remark}
The {\em Kullback-Leibler divergence} is defined as \cite{Sandholm11}:
\begin{equation}
D_{\textup{KL}}(q\,\|\, p)
= \sum\nolimits_{k,\alpha:q_{k\alpha}>0} q_{k\alpha} \log\left({q_{k\alpha}}\big/{p_{k\alpha}}\right).
\end{equation}
Clearly, $D_{\textup{KL}}(q_{k} \,\|\, p_{k})$ is finite if and only if $p_{k}$ allocates positive power $p_{k\alpha}>0$ to all channels $\alpha\in\supp(q)$ which are present in $q_{k}$. Thus, in particular, Theorem \ref{thm:convergence} guarantees that uniform initial power allocations equilibrate exponentially quickly.
\end{remark}
\begin{remark}
In the evolutionary analysis of \cite{Sa01}, the fitness of a species (the number of descendants in the unit of time) might be nonlinear, but it is still a convex combination of each phenotype's fitness. In this special case, the dynamics of \cite{Sa01} are formally equivalent to (\ref{eq:RD}), and it is shown therein that their limit points are Nash equilibria. Theorem \ref{thm:convergence} (see also Theorem \ref{thm:convspeed}) extends this analysis by demonstrating that the replicator dynamics really \emph{do} converge to Nash equilibrium, and that the rate of this convergence is exponential.
\end{remark}
\begin{remark}
It was shown in \cite{SPB08-jsac} that iterative water-filling algorithms converge exponentially when the water-filling operator is a contraction. However, given that the sufficient conditions which guarantee the contraction property fail in the \ac{PMAC} case \cite{ValueTools11b}, the analysis of \cite{SPB08-jsac} does not apply here.
\end{remark}
Of course, the value of the exponent of (\ref{eq:convspeed}) is critical because it controls how fast users converge to equilibrium. Thus, if we consider the ``instantaneous'' convergence exponents:
\begin{equation}
\label{eq:exponent}
\lambda_{k}(t) \equiv -\frac{1}{t} \log \frac{D_{\textup{KL}}(q_{k}\,\|\, p_{k}(t))}{D_{\textup{KL}}(q_{k}\,\|\, p_{k}(0))},
\end{equation}
then Theorem \ref{thm:convergence} simply states that the total equilibration rate $\lambda(t) \equiv \min_{k}\{\lambda_{k}(t)\}$ is at least $c$. For the sake of simplicity, we will only present here an analytic expression for the value of $c$ for \emph{strict} equilibria (for the full analysis including non-strict equilibria, see Appendix \ref{apx:convspeed}). In this case, if user $k$ transmits with full power to channel $\alpha_{k}$ at equilibrium, we will have:
\begin{equation}
\label{eq:strictspeed}
\textstyle
c_{k} \equiv \liminf\nolimits_{t}\{\lambda_{k}(t)\}
=\gamma_{k}^{-1}\!\left(1-e^{-\gamma_{k}}\right) \Delta v_{k}
\text{ and }
c=\min\nolimits_{k}\{c_{k}\}
\end{equation}
where $\Delta v_{k} = \min\big\{v_{k,\alpha_{k}}(q) - v_{k\beta}(q):\beta\neq \alpha_{k}\big\}$ is $k$'s minimum deviation cost and $\gamma_{k} = D_{\textup{KL}}(q\,\|\, p(0))/P_{k}$. We thus obtain:
\begin{proposition}
\label{prop:powerconvergence}
If $q=\sum_{k}P_{k} e_{k,\alpha_{k}}$ is a strict equilibrium of the static game $\mathfrak{G}$, the power of user $k$ on channel $\alpha_{k}$ grows as:
\begin{equation}
p_{k,\alpha_{k}}(t)\sim P_{k}\left(1-e^{-\Delta v_{k}\,t}\right).
\end{equation}
\end{proposition}
\begin{IEEEproof}
From (\ref{eq:strictspeed}), we have $c_{k}\to\Delta v_{k}$ as $\gamma_{k}\to 0$. However, since $D_{\textup{KL}}(q_{k}\,\|\, p_{k}(t))\to 0$ as $t\to\infty$ (Theorem \ref{thm:convergence}), we will have $\gamma_{k}\to0$ by definition, and our assertion follows.
\end{IEEEproof}
\subsection{Fading Channels}
\label{subsec:ergodic.convergence}
\subsubsection{Block-fading channels}
In this case, the replicator equation (\ref{eq:RD}) becomes non-deterministic because the coefficients $g_{k\alpha}$ evolve stochastically over time. To account for this, we will rewrite the replicator dynamics (\ref{eq:RD}) in discrete time as:
\begin{equation}
\label{eq:stochRD}
\Delta p_{k\alpha}(n+1) = \delta(n) p_{k\alpha}(n) \, \big[v_{k\alpha}(p(n),g(n)) - v_{k}(p(n),g(n))\big],
\end{equation}
where $\Delta p_{k\alpha}(n+1) \equiv p_{k\alpha}(n+1) - p_{k\alpha}(n)$ and the ``step'' $\delta(n)$ is a (possibly time-dependent) learning parameter.
For simplicity, we will concentrate here on the case where the temporal variations of the channels are uncorrelated. In this case, if we set $\overline v_{k\alpha} = \ex[v_{k\alpha}]$ and $\eta_{k\alpha} = v_{k\alpha} -\overline v_{k\alpha}$, we obtain:
\begin{flalign}
\label{eq:stochapproxRD}
p_{k\alpha}(n+1)
&= p_{k\alpha}(n)
+ \delta(n) p_{k\alpha}(n) \Big(\overline v_{k\alpha}(p(n)) - \overline v_{k}(p(n))\Big)\notag\\
&+ \delta(n) p_{k\alpha}(n) \Big(\eta_{k\alpha}(p(n),g(n)) - \eta_{k}(p(n),g(n))\Big),
\end{flalign}
where the $\overline v_{k\alpha}$ are deterministic and the $\eta_{k\alpha}$ are zero-mean. In fact, if we interchange expectation and differentiation, we get:
\begin{equation}
\label{eq:ergderpayoff}
\overline v_{k\alpha}(p)
= \ex\left[\frac{\partial u_{k}}{\partial p_{k\alpha}}\right]
= \frac{\partial}{\partial p_{k\alpha}}\ex[u_{k}]
= \frac{\partial \overline u_{k}}{\partial p_{k\alpha}},
\end{equation}
so the mean utilities of (\ref{eq:stochapproxRD}) are the gradients of the ergodic rates (\ref{eq:ergpayoff}). Thus, if we remove the noise $\eta_{k\alpha}$ from (\ref{eq:stochapproxRD}), the general theory of \cite{Borkar08} shows that (\ref{eq:stochapproxRD}) will track the {\em mean-field equation}:
\begin{equation}
\label{eq:meanRD}
\frac{dp_{k\alpha}}{dt} = p_{k\alpha}(t)\,\big[\overline v_{k\alpha}(p(t)) - \overline v_{k}(p(t))\big],
\end{equation}
so the asymptotic properties of (\ref{eq:stochapproxRD}) will follow those of (\ref{eq:meanRD}).
We thus see that the asymptotic behavior of the replicator algorithm (\ref{eq:stochRD}) for block-fading channels is intimately linked to the ergodic game $\overline\mathfrak{G}$. In itself, this is quite natural because the ergodic equilibrium of $\overline\mathfrak{G}$ represents the only reasonable time-invariant equilibrial notion for the block-fading game $\mathfrak{G}(t)$ with temporally uncorrelated channels (see also \cite{MM10}). More to the point, we have (see App.~\ref{apx:stochastic} for the proof):
\begin{theorem}
\label{thm:stochconvergence}
If the learning parameters $\delta(n)$ of (\ref{eq:stochRD}) satisfy $\sum_{n=1}^{\infty} \delta(n) = \infty$ and $\sum_{n=1}^{\infty} \delta^{2}(n) < \infty$, then the algorithm (\ref{eq:stochRD}) for temporally uncorrelated block-fading channels converges (a.s) to the (unique) Nash equilibrium of the ergodic game $\overline\mathfrak{G}$.
\end{theorem}
\begin{remark*}
The most usual choice for the parameters $\delta(n)$ is $\delta(n) = 1/n$. These variable rates can be interpreted either as the actual time step of the algorithm, or as a discount that users apply to their updating scheme at every tick of a timer. This last interpretation is crucial for practical purposes because there are hard limits to how fast a device can update its policy. For constant $\delta(n)\equiv\delta$, the dynamics (\ref{eq:stochRD}) evolve faster, but convergence to equilibrium is in the distribution sense of \cite{Borkar08}.
\end{remark*}
\subsubsection{Fast-fading channels}
As we have already noted in Section \ref{subsec:fading.game}, the users' (ergodic) rates (\ref{eq:ergpayoff}) in the fast-fading regime depend on the channels' statistics, so instantaneous channel information obtained when updating their powers is of little use. Because of this, the system becomes effectively deterministic, so, similarly to the static case, users may base their learning on the (mean) replicator learning dynamics (\ref{eq:meanRD}) \textendash\ i.e. the discrete-time learning scheme (\ref{eq:stochRD}) without the noise term and with a constant step $\delta(n) = \delta$. We thus obtain (see App. \ref{apx:convspeed}):
\begin{theorem}
\label{thm:ergconvergence}
The mean dynamics (\ref{eq:meanRD}) converge to the (unique) Nash equilibrium of the ergodic game $\overline\mathfrak{G}$, and this convergence is exponential: interior orbits $\varepsilon$-equilibrate in time $\bigoh(\log(1/\varepsilon))$.
\end{theorem}
\begin{remark*}
If the ergodic equilibrium is strict, the convergence exponent (\ref{eq:exponent}) of the mean dynamics (\ref{eq:meanRD}) is just (App.~\ref{apx:convspeed}):
\begin{equation}
\label{eq:strictstochspeed}
\textstyle
\overline c=\min\nolimits_{k}\left\{\overline c_{k}\right\}
\quad\text{with}\quad
\overline c_{k} =\gamma_{k}^{-1}\left(1-e^{-\gamma_{k}}\right) \Delta \overline v_{k},
\end{equation}
where $\gamma_{k} = D_{\textup{KL}}(q\,\|\, p(0))/P_{k}$ is as in (\ref{eq:strictspeed}), but now $\Delta \overline v_{k} = \min\big\{\overline v_{k,\alpha_{k}}(q) - \overline v_{k\beta}(q):\beta\neq \alpha_{k}\big\}$ is the minimum deviation cost of user $k$ for the \emph{mean} marginal utilities $\overline v_{k\alpha}$.
\end{remark*}
\section{Numerical Simulations}
\label{sec:numerics}
\begin{figure*}
\centering
\subfigure[Equilibrium \acl{SRE} for static channels.]{
\label{subfig:staticSRE}
\includegraphics[height=165pt,width=.9\columnwidth]{SRE_cdf_3x2and10.pdf}}
\qquad
\subfigure[Equilibrium \acl{SRE} for ergodic channels.]{
\label{subfig:ergodicSRE}
\includegraphics[height=165pt,width=.9\columnwidth]{SRE_ergodic_varKvarA.pdf}}
\caption{The \ac{CDF} of the equilibrium \acl{SRE} for static channels (Fig.~\ref{subfig:staticSRE}) and the equilibrium \ac{SRE} for ergodic Gaussian i.i.d. channels as a function of the thermal \ac{SNR} parameter $\rho = P_{\textit{max}}/\sigma^{2}$ (Fig.~\ref{subfig:ergodicSRE}) for different numbers of channels $A$ and users $K$ (all with similar maximum power constraints).}
\label{fig:SRE}
\end{figure*}
\begin{figure*}
\vspace{-10pt}
\centering
\subfigure[Sum-rate efficiency and equilibration over time for static channels.]{
\label{subfig:staticEQL}
\includegraphics[width=0.9575\columnwidth]{static_figure.pdf}}
\subfigure[Long-term equilibration in block-fading channels.]{
\label{subfig:ergodicEQL}
\includegraphics[width=\columnwidth]{erg_figure.pdf}}
\caption{Equilibration (and its efficiency) for different numbers of users and channels. In static channels, the replicator dynamics equilibrate extremely fast, even for a large number of users; in temporally uncorrelated block-fading channels, the dynamics still converge, but slower \textendash\ due to the discounting $\delta(n) = 1/n$ in (\ref{eq:stochRD}).}
\label{fig:EQL}
\end{figure*}
\begin{figure*}
\centering
\subfigure{
\label{subfig:Jakes.follow.slow}
\includegraphics[width=\columnwidth]{JakesFollow-v5f.pdf}}
\hfill
\subfigure{
\label{subfig:Jakes.follow.fast}
\includegraphics[width=.95\columnwidth]{JakesFollow-v15f.pdf}}
\caption{
The Nash power level (dashed blue) and the actual power level learned by a user (solid red) in a $2\times2$ Jakes-fading game. As can be seen by the cross-correlograms of the two time-series (inlays), users are away from equilibrium for only a very short amount of time (light blue shading): we observe a 9~ms tracking delay for user velocities in the 5~km/h range and 6~ms for 15~km/h, meaning that the replicator dynamics converge within 10-15\% of the system's coherence time (108~ms and 36~ms respectively).}
\label{fig:Jakes}
\end{figure*}
In this section, our aim is to validate our theoretical results by means of numerical simulations. We begin by introducing the \acfi{SRE} of a power profile $p$:
\begin{equation}
\label{eq:SRE}
\sre(p) = \frac{\sum_{k} u_{k}(p)}{C_{sum}},\quad
\text{(resp. $\displaystyle \overline\sre(p) = \frac{\sum_{k} \overline u_{k}(p)}{\overline C_{sum}}$ for $\overline\mathfrak{G}$)}
\end{equation}
i.e. the ratio of the sum of achievable rates in the power profile $p$ over the maximum achievable aggregate sum-rate under \ac{SIC} (which is the sum-capacity of the \ac{MAC}); interestingly, if $q$ is the game's (unique) equilibrium, then $C_{sum} = -\Phi(q)$ (and similarly for the ergodic case).
In Fig.~\ref{subfig:staticSRE}, we plot the \ac{SRE} at Nash equilibrium for randomly drawn static channels. While the equilibrium \ac{SRE} can deviate significantly from its maximum value (unity) for $A<K$, in the $A\geq K$ regime, the \ac{SRE} is typically close to 100\% (and, in fact, equal to 100\% with positive probability. We may attribute this to the fact that for $A\geq K$ there is a finite probability that the system's equilibrium is at a vertex of $\Delta$ where each user is alone on a single channel, inducing optimal performance. Thus, with fair probability (close to $1/2$ based on our simulations) the complex \ac{SIC} scheme yields no performance benefit over the much simpler \ac{SUD} approach. Analogously, in Fig.~\ref{subfig:ergodicSRE}, we plot the \ac{SRE} at equilibrium for ergodic channels by using Proposition \ref{prop:gaussianpotential} to evaluate the maximum sum-rate under \ac{SIC}; in that case, while the \ac{SRE} is nearly optimal for small \ac{SNR}, it deviates strongly from its maximum value for larger \ac{SNR} values.
Furthermore, to test the equilibration rate of (\ref{eq:stochRD}) for different values of $K$ and $A$, we introduce the \ac{EQL}:
\begin{equation}
\label{eq:EQL}
\eql(p) = \Phi(p)\big/\Phi(q),\,\,
\text{(resp. $\displaystyle \overline\eql(p) = \overline\Phi(p)\big/\overline\Phi(q)$ for $\overline{\mathfrak{G}}$),}
\end{equation}
with $q$ being the equilibrium of $\mathfrak{G}$ (resp.~$\overline\mathfrak{G}$) \textendash\ so an \ac{EQL} of $1$ implies that the system has reached its equilibrium.
Beginning with the static case, in Fig.~\ref{subfig:staticEQL} we drew $N=50$ channel realizations and ran the discrete-time learning scheme (\ref{eq:stochRD}) with constant $g(n)$ and $\delta(n)$ for $A=20$ channels and $K=5,\,10,\,20$ and $30$ users. Then, by plotting the average \ac{SRE} and \ac{EQL} over time, we see that even for $30$ users, the system equilibrates within a few tens of iterations. On the other hand, for the ergodic block-fading scenario of Fig.~\ref{subfig:ergodicEQL}, we used $A=10$ channels and plotted the system's {\footnotesize $\overline\eql$} over time for $K=2, 4$ and $8$ users learning with $\delta(n) = 1/n$. As predicted by Theorem \ref{thm:stochconvergence}, the system converges to the game's ergodic equilibrium \textendash\ but slower due to the discounting $\delta(n) = 1/n$. Finally, for fast-fading users following the discrete-time version of (\ref{eq:meanRD}), the \ac{EQL} and \ac{SRE} plots were virtually identical to the static case (to be expected since both dynamical systems are deterministic), so they have been omitted for space considerations.
Finally, to test the convergence rate of the replicator dynamics in more realistic (non-ergodic) fading conditions that do not possess a long-term stationary equilibrium, we also simulated in Fig.~\ref{fig:Jakes} channels that follow the well-known Jakes model for Rayleigh fading \cite{Calcev++07}. Specifically, we considered a $2\times2$ game with user velocities $v=5$~km/h and $15$~km/h (Figs.~\ref{subfig:Jakes.follow.slow} and \ref{subfig:Jakes.follow.fast} respectively) transmitting at a carrier frequency of $\nu = 2\,\mathrm{GHz}$. We then ran the learning scheme (\ref{eq:stochRD}) with a constant update period of $\delta=3\,\mathrm{ms}$, and we plotted the (normalized) power level of a single user against the evolving equilibrium power level (calculated at each step based on the instantaneous channel coefficients). Then, to quantify how well the users follow the system's evolving equilibrium, we calculated the cross-correlation of the two processes and the users' \emph{tracking delay} (defined as the point of maximum cross-correlation).
Remarkably, Fig.~\ref{fig:Jakes} shows that the dynamics track the game's evolving equilibrium extremely closely. On average, users equilibrate within 9~ms for 5~km/h fading velocities, and within 6~ms for 15~km/h \textendash\ meaning that users converge within 10-15\% of the system's coherence time (108~ms and 36~ms respectively).
\begin{remark*}
For larger numbers of users (or channels), the results observed are similar; for instance, if the users in the previous game are increased to a few tens (we went up to $N=50$), their tracking delay becomes longer but never exceeds 30-35\% of the channel coherence time. However, due to space limitations, we opted to present here only the $2\times2$ case for simplicity.
\end{remark*}
\section{Conclusions and Future Directions}
\label{sec:conclusions}
In this paper, we studied the distributed power allocation problem for orthogonal uplink channels by introducing a game which admits a convex potential function. For both static and fading channels, we found that the associated game admits a unique Nash equilibrium and we showed that a simple distributed learning scheme based on the replicator dynamics converges to equilibrium from (almost) any initial condition. In fact, by proving a general result for convex potential games, we showed that the speed of this convergence is exponential: users converge to an $\varepsilon$-neighborhood of an equilibrium in time which is at most of order $\bigoh(\log(1/\varepsilon))$.
There is a number of important extensions of this work which demonstrate the strength of the replicator dynamics in continuous nonlinear games of this sort. First off, instead of the achievable rates $u_{k}$, one could consider energy-efficient metrics where users do not saturate their power constraints \textendash\ e.g., when the price of transmission power might restrain users from transmitting at maximum power. More importantly, these techniques can be extended even to non-orthogonal channel models such as the \ac{MIMO} \ac{MAC} case where the game's strategy space consists of all positive-definite precoding matrices with constrained trace. Because of this nonlinear structure, the form (\ref{eq:RD}) of the replicator dynamics no longer applies, but one can still write down a suitably modified {\em matrix-valued} replicator equation which allows users to converge to equilibrium.
\appendices
\section{Convergence Speed of the Replicator Dynamics}
\label{apx:convspeed}
Recall first that the \emph{solid tangent cone} to $\Delta$ at $q$ is the set of rays starting at $q$ and in\-ter\-secting $\Delta$ in at least one other point, i.e.: $T^{c}_{q}\Delta \equiv \big\{z\in {\mathbb R}^{Q}: z_{k\alpha}\geq0 \text{ for all $\alpha\in\mathcal{A}_{k}$ with $q_{k\alpha}=0$}\big\}$.
With this in mind, we have the following generalization of convexity:
\begin{definition}
\label{def:starconvex}
A function $F\colon\Delta\to{\mathbb R}$ will be called \emph{star-convex} w.r.t. $q\in\Delta$ if $f(\theta)\equiv F(q+\theta z)$ is convex and increasing for all $z\inT^{c}_{q}\Delta$ and for all $\theta>0$ s.t. $q+\theta z\in\Delta$.
\end{definition}
Star-convex functions need not be convex, but strictly convex functions are star-convex w.r.t. their global minimum and weakly convex functions with a unique minimum are also star-convex \textendash\ in particular, both $\Phi$, $\overline\Phi$ are star-convex. For games with star-convex potentials, we then have:
\begin{theorem}
\label{thm:convspeed}
Let $\mathfrak{Q}\equiv\mathfrak{Q}\left(\mathcal{K},\{\Delta_{k}\}, \{\phi_{k}\}\right)$ be a game with a star-convex potential $F$. Then, the replicator dynamics (\ref{eq:RD}) for the marginal utilities $\phi_{k\alpha} = \frac{\partial \phi_{k}}{\partial p_{k\alpha}}$ converge to $q$ for any initial condition that starts at finite K-L divergence $h_{0}\equiv D_{\textup{KL}}(q\,\|\, p(0))$ from $q$. Moreover, there exists $c>0$ such that:
\begin{equation}
\label{eq:genconvspeed}
D_{\textup{KL}}(q\,\|\, p(t)) \leq h_{0}\,e^{-ct}\quad
\text{for all $t\geq0$}.
\end{equation}
\end{theorem}
Our proof strategy will be to establish an inequality of the form $\frac{d}{dt} H_{q}(p(t)) \leq -c H_{q}(p(t))$ and then employ Gr\"onwall's lemma. To that end, we first define the ``evolutionary index'':
\begin{equation}
\label{eq:evindex}
L_{q}(p) = - \textstyle\sum\nolimits_{k,\alpha} \left(p_{k\alpha} - q_{k\alpha}\right) \phi_{k\alpha}(p),
\end{equation}
so named because $q$ is evolutionarily stable iff $L_{q}(p)>0$ near $q$. In fact, if we set $H_{q}(p) = D_{\textup{KL}}(q\,\|\, p)$, an easy calculation shows that $L_{q}$ is just (the negative of) the time derivative of $H_{q}$ w.r.t. the replicator dynamics (\ref{eq:RD}): $\frac{d}{dt} H_{q}(p(t)) = \sum\nolimits_{k,\alpha} \frac{\partial H_{q}}{\partial p_{k\alpha}} \dot p_{k\alpha} =-L_{q}(p(t))$.
We will therefore begin by showing that $L_{q}(p)>0$ for all $p\in\Delta\exclude\{q\}$, implying that $H_{q}$ is Lyapunov for the replicator dynamics (\ref{eq:RD}), and proving the convergence part of Theorem \ref{thm:convspeed}. Indeed, if we fix some $z\inT^{c}_{q}\Delta$, (\ref{eq:evindex}) may be rewritten as $f'(\theta) = \sum\nolimits_{k,\alpha} \left.\frac{\partial F}{\partial p_{k\alpha}}\right|_{q+\theta z} z_{k\alpha} = \theta^{-1} L_{q}(q+\theta z)$, for all $\theta>0$ such that $q+\theta z\in\Delta$. Then, with $f(\theta)$ convex and increasing (Definition \ref{def:starconvex}), we obtain the estimate $L_{q}(p) = \theta f'(\theta) \geq f(\theta) - f(0) = F(p) - F(q)$, which shows that $L_{q}(p)>0$ for all $p\neq q$.
To prove the convergence time estimate (\ref{eq:genconvspeed}) we will need to show that $L_{q}$ grows linearly along directions which are not supported in $q$, and quadratically along those which {\em are} supported in $q$. To be specific, let $V_{q} = \{x\in{\mathbb R}^{Q}: \text{$x_{k\alpha}=0$ if $q_{k\alpha}=0$}\}$ be the subspace of directions of ${\mathbb R}^{Q}$, $Q = \sum_{k} A_{k}$, which are supported in $q$, and let $V_{q}^{\perp}$ be its orthocomplement in ${\mathbb R}^{Q}$. Then, by decomposing $z\in{\mathbb R}^{Q}$ as $z=z_{\varparallel}+z_{\perp}$ with $z_{\varparallel}\in V_{q}$ and $z_{\perp}\in V_{q}^{\perp}$, we define the seminorms $\|\cdot\|_{\varparallel}$ and $|\cdot|_{\perp}$ as:
\begin{align}
\label{eq:seminorms}
\textstyle
\|z\|_{\varparallel}^{2} \equiv \|z_{\varparallel}\|_{2}^{2} = \sum\nolimits_{k,\alpha}^{\varparallel} z_{k\alpha}^{2},
&\quad
&
\textstyle
|z|_{\perp} \equiv \|z_{\perp}\|_{1} = \sum\nolimits_{k,\alpha}^{\perp} |z_{k\alpha}|,
\end{align}
where the notation $\sum\nolimits_{k,\alpha}^{\varparallel}$, $\sum\nolimits_{k,\alpha}^{\perp}$ is shorthand for summing over the directions of $V_{q}$ and $V_{q}^{\perp}$ respectively.
We thus get:
\begin{lemma}
\label{lem:evindexestimate}
Let $F\colon\Delta\to{\mathbb R}$ be star-convex w.r.t. $q\in\Delta$. Then:
\begin{equation}
\label{eq:evindexestimate}
L_{q}(p) \geq F(p) - F(q) \geq m\,|p-q|_{\perp} + \tfrac{1}{2}r\,\|p-q\|_{\varparallel}^{2},
\end{equation}
where $m = \min_{k}\{\phi_{k\alpha}(q) - \phi_{k\mu}(q): q_{k\mu}=0,\, q_{k\alpha}>0\}$, and $r$ is the minimum of the Rayleigh quotient $\langle z,\mathbf M(q+z) z\rangle/\|z\|^{2}$ for the Hessian $\mathbf M(p) = \frac{\partial^{2}F}{\partial p_{k\alpha}\partial p_{\ell\beta}}$ of $F$, restricted over $T^{c}_{q}\Delta$.
\end{lemma}
\begin{IEEEproof}
Since $q$ minimizes $F$, the \ac{KKT} conditions give $\phi_{k\alpha}(q) = -\left.\frac{\partial F}{\partial p_{k\alpha}}\right|_{q} = -\lambda_{k}$ for all $\alpha\in\mathcal{A}_{k}$ such that $q_{k\alpha}>0$ and $\phi_{k\alpha}(q)<-\lambda_{k}$ otherwise (where $\lambda_{k}$ denotes the complementary slackness Lagrange multiplier of $F$ over $\Delta$). Thus, a first order Taylor estimate with Lagrange remainder readily yields:
\begin{equation}
\label{eq:fTaylor}
f(\theta) = f(0) + f'(0) \theta + \tfrac{1}{2} f''(\xi) \theta^{2}
\end{equation}
for some $\xi\in(0,\theta)$, so (\ref{eq:evindexestimate}) will follow once we properly estimate the linear and quadratic terms of (\ref{eq:fTaylor}).
As far as the linear term of (\ref{eq:fTaylor}) is concerned, we will have $f'(0)
= \sum\nolimits_{k,\alpha} z_{k\alpha} \left.\frac{\partial F}{\partial p_{k\alpha}}\right|_{q}
= \sum\nolimits_{k,\alpha}^{\varparallel} z_{k\alpha} \left.\frac{\partial F}{\partial p_{k\alpha}}\right|_{q}
+ \sum\nolimits_{k,\alpha}^{\perp} z_{k\alpha} \left.\frac{\partial F}{\partial p_{k\alpha}}\right|_{q}
= \sum\nolimits_{k,\alpha}^{\perp} z_{k\alpha} \left(\left.\frac{\partial F}{\partial p_{k\alpha}}\right|_{q} - \lambda_{k}\right)
\geq m\,|z|_{\perp}$, where the last equality holds because $\sum\nolimits_{\alpha}^{\perp} z_{k\alpha} = - \sum\nolimits_{\alpha}^{\varparallel} z_{k\alpha}$ (recall that $z\inT^{c}_{q}\Delta$) and the last inequality is just the definition of $m$. Similarly, for any $\xi\in(0,\theta)$ and $z\inT^{c}_{q}\Delta$, we get $f''(\xi)= \Big\langle z, \mathbf M(q+\xi z) z\Big\rangle=R_{q+\xi z}(\xi z)\,\|z\|^{2}$, where $R_{p}(w) = \big\langle w,\mathbf M(p) w\big\rangle$, $p\in\Delta$, $w\in T_{p}\Delta$, denotes the Rayleigh quotient of the Hessian $\mathbf M$ of $F$. Hence, if $r$ is the minimum of $R_{q+w}(w)$ over the set $B_{q} = \{w\inT^{c}_{q}\Delta: q+w\in\Delta\}$, we will also have $f''(\xi)\geq r \|z\|^{2}$, and (\ref{eq:evindexestimate}) follows by plugging the above into (\ref{eq:fTaylor}) and noting that $\|z\|\geq\|z\|_{\varparallel}$.
\end{IEEEproof}
Obtaining similar estimates for the relative entropy function $H_{q}$ is harder (after all, $H_{q}$ blows up near the boundary of $\Delta$), so we will need two more auxiliary lemmas:
\begin{lemma}
\label{lem:uniqueroot}
For all $z\inT^{c}_{q}\Delta\exclude\{0\}$ and for all $a>1$, the equation
\begin{equation}
\label{eq:uniqueroot}
\textstyle
H_{q}(q+\theta z) = a\,|z|_{\perp} \theta + \tfrac{1}{2} a \sum\nolimits_{k,\beta}^{\varparallel} z_{k\beta}^{2}\big/ q_{k\beta} \, \theta^{2},
\end{equation}
admits a unique positive root $\theta_{a}\equiv\theta_{a}(z)$. Consequently:
\begin{equation}
\label{eq:directedestimate}
\textstyle
H_{q}(q+\theta z) \leq a\,|z|_{\perp} \theta + \tfrac{1}{2} a \sum\nolimits_{k,\beta}^{\varparallel} z_{k\beta}^{2}\big/ q_{k\beta} \, \theta^{2}\,\,
\text{for all $\theta\leq\theta_{a}(z)$.}
\end{equation}
\end{lemma}
\begin{IEEEproof}
Let $h(\theta) \equiv H_{q}(q+\theta z)$ be the LHS of (\ref{eq:uniqueroot}), and denote its RHS by $ag(\theta)$. Then, if we set $w(\theta) = h(\theta) - ag(\theta)$, we readily obtain $w(0)=0$, $w'(0) = |z|_{\perp} (1-a)\leq 0$, and $w''(0) = \sum\nolimits_{k,\beta}^{\varparallel} z_{k\beta}^{2}/q_{k\beta} (1-a)<0$, and the result follows by simple arguments relying on the mean value theorem.
\end{IEEEproof}
\begin{lemma}
\label{lem:entropyestimate}
Let $F\colon\Delta\to{\mathbb R}$ be star-convex w.r.t. $q\in\Delta$ and let $p(t)$ be a solution orbit of the replicator dynamics with initial relative entropy $h_{0} = H_{q}(p(0))$. Then, there exists $b>1$ s.t.:
\begin{equation}
\label{eq:entropyestimate}
\textstyle
H_{q}(p(t)) \leq b\,|p(t)-q|_{\perp} + \frac{b}{2q_{0}} \|p(t) - q\|_{\varparallel}^{2},\quad
\end{equation}
where $q_{0} = \min_{k,\alpha}\{q_{k\alpha}: q_{k\alpha}>0\}$.
\end{lemma}
\begin{IEEEproof}
Fix some $a>1$. Then, by Lemma \ref{lem:uniqueroot}, we know that (\ref{eq:uniqueroot}) admits a unique positive root $\theta_{a}(z)$, so let $h_{a}(z) = H_{q}(q+\theta_{a}(z) z)$ and set $h_{a} = \max\{h_{a}(z): z\in S_{q}\}$, where $S_{q} = \{z\inT^{c}_{q}\Delta:\text{$z+q\in\Delta$ but $q+(1+\varepsilon)z\notin\Delta$ for any $\varepsilon>0$}\}$. Moreover, set $h_{c} = \max\{h_{0},h_{a}\}$, let $\theta_{c}(z)$ be the unique positive root of the equation $H_{q}(q+\theta_{c}(z) z) = h_{c}$, and define $b(z) = g(\theta_{c}(z))/h_{c}$ with $g(\theta) = |z|_{\perp} \theta + \tfrac{1}{2} \sum_{k,\beta}^{\varparallel} z_{k\beta}^{2}/q_{k\beta} \theta^{2}$ (as in the proof of Lemma \ref{lem:uniqueroot}). We will then have $b(z)\geq a$ since, otherwise, (\ref{eq:directedestimate}) would yield the contradiction $h_{c} = b(z) g(\theta_{c}(z)) < a g(\theta_{c}(z)) < h(\theta_{c}(z)) = h_{c}$.
With $b(z)>1$, a second application of Lemma \ref{lem:uniqueroot} yields $H_{q}(q+\theta z) \leq b(z) \left(|z|_{\perp} \theta + \tfrac{1}{2} \sum\nolimits_{k,\beta}^{\varparallel} z_{k\beta}^{2}/q_{k\beta} \theta^{2}\right)$ for all $\theta\leq\theta_{c}(z)$. Thus, if we decompose $p(t)$ as $p(t) = q + \theta(t) z(t)$ with $\theta>0$ and $z(t)\in S_{q}$, we will have $\theta(t) \leq \theta_{c}(z(t))$; indeed, should this ever fail, we would have $H_{q}(p(t)) > b(z(t)) g(\theta(t)) > b(z(t)) g(\theta_{c}(t)) = h_{c} \geq h_{0}$ which contradicts the fact that $H_{q}$ is Lyapunov. Hence, with $\theta(t)\leq\theta_{c}(z(t))$ for all $t\geq0$, we get $H_{q}(p(t)) \leq b(z(t)) \left(|z(t)|_{\perp} \theta(t) + \tfrac{1}{2} \sum\nolimits_{k,\beta}^{\varparallel} z_{k\beta}^{2}(t)/q_{k\beta} \theta^{2}(t)\right)$, and (\ref{eq:entropyestimate}) follows by taking $b = \max\{b(z):z\in S_{q}\}$.
\end{IEEEproof}
\begin{IEEEproof}[Proof of Theorem \ref{thm:convspeed}]
With notation as in Lemmas \ref{lem:evindexestimate} and \ref{lem:entropyestimate}, let $c=\min\{m/b,rq_{0}/b\}$. We then get $L_{q}(p(t)) \geq m\,|p(t)-q|_{\perp} + \tfrac{1}{2}r\,\|p(t)-q\|_{\varparallel}^{2} \geq c H_{q}(p(t))$ and Gr\"onwall's lemma yields $H_{q}(p(t)) \leq h_{0} e^{-ct}$. Since the \ac{KKT} inequalities for $F$ are strict along any direction of ${\mathbb R}^{Q}$ which is not supported in $q$, we will have $m>0$ and, consequently, $c>0$ as well.
\end{IEEEproof}
\begin{IEEEproof}[Proof of Theorems \ref{thm:convergence} and \ref{thm:ergconvergence}]
The potentials $\Phi$ and $\overline\Phi$ are star-convex, so both theorems follow from Theorem \ref{thm:convspeed}.
\end{IEEEproof}
All that remains is to calculate the value of $c$ when $q$ is strict. In that case, given that the intersection of $V_{q}$ with $T^{c}_{q}\Delta$ is trivial, the quadratic term of (\ref{eq:evindexestimate}) can be ignored and we get $L_{q}(p) \geq \frac{1}{2} \sum\nolimits_{k} \|p_{k}-q_{k}\|_{1} \Delta \phi_{k}$, where $\Delta \phi_{k} = \min_{\mu\neq\alpha_{k}}\{\phi_{k,\alpha_{k}}(q) - \phi_{k\mu}(q)\}>0$. As for (\ref{eq:entropyestimate}), we may decompose $p_{k}\in \Delta_{k}\exclude\{q_{k}\}$ as $p_{k} = q_{k} + \theta_{k} z_{k}$ where $z_{k}\in T^{c}_{q_{k}}\Delta_{k}$ has $z_{k,\alpha_{k}} = - P_{k}$. Thus, with $p_{k,\alpha_{k}} = P_{k}(1-\theta_{k})$, we readily obtain $H_{q}(p) = -\sum\nolimits_{k} P_{k} \log(1-\theta_{k})$. Now, let $\theta_{k}^{*}$ be defined by the equation $h_{0} = H_{q_{k}}(q_{k}+\theta_{k}z_{k})$, i.e., $\theta_{k}^{*} = 1 -\exp(-h_{0}/P_{k})$, implying that $-P_{k}\log(1-\theta_{k}) \leq h_{0} \theta_{k}/\theta_{k}^{*}$ iff $0\leq\theta_{k}\leq\theta_{k}^{*}$ (because of convexity). We then claim that $H_{q}(p(t)) = -\sum\nolimits_{k}P_{k} \log(1-\theta_{k}(t)) \leq h_{0} \sum\nolimits_{k} \theta_{k}(t)/\theta_{k}^{*}$, where $\theta_{k}(t)$ is defined via the decomposition $p_{k}(t) = q_{k} + \theta_{k}(t) z_{k}(t)$.
However, if $\theta_{k}(t)>\theta_{k}^{*}$ for some $t\geq0$, then we would have $H_{q_{k}}(p_{k}(t)) > h_{0}$, and, hence, $H_{q}(p(t)) > H_{q}(p(0))$ as well, a contradiction \textendash\ recall that $H_{q}(p(t))$ is decreasing. Thus, combining all the above, we only need pick $c$ such that $P_{k}\,\Delta\phi_{k}\geq c h_{0}/\theta_{k}^{*}$, and the sharpest such choice is:
\begin{equation}
\label{eq:strictexponent}
c=\min\nolimits_{k}\left\{P_{k}/h_{0}\left(1-e^{-h_{0}/P_{k}}\right)\Delta \phi_{k}\right\}.
\end{equation}
\section{Stochastic Approximation of the Replicator Dynamics}
\label{apx:stochastic}
\begin{proof}[Proof of Theorem \ref{thm:stochconvergence}]
Note first that $\Delta$ is invariant under the dynamics (\ref{eq:stochapproxRD}) if the $\delta(n)$ are chosen small enough. To see this, we will restrict ourselves w.l.o.g. to a game with one user and two choices, A and B (the general argument being similar). Thus, if we let $p_{A}(n) \equiv p_{1,A}(n)$ be the power that the user sends to channel A at the $n$-th iteration of the dynamics, we must find $\delta(n)$ such that $0\leq p(n) \leq 1$ for all $n\geq 0$ and all possible $g_{A,B}(n)\geq0$. So, assuming this holds for some $n\geq0$, we get:
\begin{multline}
\label{eq:diffRD}
p_{A}(n+1) - p_{A}(n)
= \delta(n) \,p_{A}(n) (1-p_{A}(n))\\
\times\left(
\frac{g_{A}(n)}{\sigma_{A}^{2} + g_{A}(n) p_{A}(n)} - \frac{g_{B}(n)}{\sigma_{B}^{2} + g_{B}(n) (1-p_{A}(n))}
\right).
\end{multline}
The first term of the LHS of (\ref{eq:diffRD}) is positive and the second is uniformly bounded, say by $M$, so $\delta(n)\leq M$ yields $p_{A}(n+1)\geq0$. The complementary inequality $p_{A}(n+1)\leq 1$ then follows similarly, so, with $p(n)\in\Delta$ for all $n$, our theorem follows from Theorem~2 and Corollary~4 in Chap.~2 of \cite{Borkar08}.
\end{proof}
\bibliographystyle{ieeetran}
\footnotesize
|
\section{Introduction}
This work is devoted to the investigation of Leibniz $n$-algebras.
In 1985, Filippov \cite{Fil} introduced a notion of Lie
$n$-algebra with an n-ary skew-symmetric multiplication which satisfies
the identity
\[
[[x_1,x_2, \dots ,x_n],y_2,\dots ,y_n]=\sum_{i=1}^n \ [x_1,\dots ,
x_{i-1},[x_i,y_2,\dots,y_n],x_{i+1},\dots,x_n] \eqno(1)
\]
bearing in mind the general notion of $\Omega$-algebra considered
by Kurosh \cite{Kur}.
Earlier in 1973, Nambu
\cite{Nambu} had constructed an example of 3-Lie algebra, where the multiplication for a triple of classical observables on
the three-dimensional phase space $\mathbb{R}^3$ was given by the Jacobian.
This bracket naturally generalizes the usual Poisson bracket from a
binary to a ternary operation.
In 1993, Loday \cite{Lo1,Lo2} introduced a non skew-symmetric
version of Lie algebras, the so-called Leibniz algebras. As a
generalization of Leibniz algebras and n-Lie algebras, in 2002, Casas, Loday and
Pirashvili \cite{Cas1} defined n-Leibniz algebras as a non
skew-symmetric version of Lie $n$-algebras. They also presented
constructions between the varieties of Leibniz algebras and Leibniz
$n$-algebras $(n \geq 3)$ which are not invertible.
In the present work, in Section \ref{prel}, we introduce the Frattini subalgebra of a Leibniz $n$-algebra and establish properties extending some results of
Leibniz algebras and of Lie $n$-algebras. Frattini theory was originally discovered in group
theory and further have been studied in Lie algebras
in \cite{Mars,barnes,Stit}, in Lie $n$-algebras in \cite{Bai1,Will2}
and in Leibniz algebras in \cite{barnes2,barnes3}. Here we show that
many results concerning Frattini subalgebras and Frattini
ideals from the theory of Lie $n$-algebras remain true when we omit
the skew-symmetrical property of the $n$-ary multiplication.
In Section \ref{rmo}, we study the right multiplication operators in a Leibniz $n$-algebra.
Filippov \cite{Fil} noted that the so-called right
multiplication operators play the same crucial role in the theory of Lie
$n$-algebras as in Lie algebras since they form a Lie algebra
with respect to the commutator. The space of the right multiplication
operators in Leibniz $n$-algebras also forms an ideal in the Lie
algebra of derivations. However, in the case $n\geq 3$, some well-known
properties of the right multiplication operators do not hold in
general; for instance, in \cite{cartan} it was given an example of a
Leibniz $n$-algebra which admits a non-degenerate right multiplication operator. Because of that curious properties of these
operators, to obtain some results on right
multiplication operators which are valid for Leibniz and Lie $n$-algebras we
must consider them with additional conditions.
In Section \ref{invar}, we study solvability and nilpotency in
Leibniz $n$-algebras and show that the solvable and nilpotent radicals are invariant under all derivations. Since multiplication in Leibniz $n$-algebras is not anti-symmetric
in all the variables, notions such as nilpotency and solvability may
be introduced in different ways depending on the position of the
multiplicand. The product in the definition of the corresponding
series is not necessarily an ideal and this makes some
arguments difficult to prove. Hence, we
introduce special notions, as $k$-solvability, nilpotency and
$K_1$-nilpotency of Leibniz $n$-algebras. Most of them agree
with the corresponding notions on particular cases: Lie $n$-algebras
\cite{Kasymov} and Leibniz algebras. We establish some properties
of $k$-solvable (nilpotent) ideals, as well.
Finally, in Section \ref{no_conj}, we construct examples that show the
non-conjugacy of Cartan
subalgebras for Leibniz $n$-algebras.
In \cite{cartan}, it was proved that the null root subspace of the right
multiplication operators with respect to a regular element is a
nilpotent subalgebra. Here we obtain that this subalgebra under
some restriction is a Cartan subalgebra. Moreover, a classical
result about conjugacy of Cartan subalgebras in Lie algebras that
was extended to the general cases - Leibniz algebras \cite{Omirov2}
and Lie $n$-algebras \cite{Kasymov}, unfortunately does not hold in the
case of Leibniz $n$-algebras ($n\geq 2$). Starting with a particular Lie $n$-algebra,
we construct Leibniz $n$-algebras which factored out by the ideal
$I$ generated by the elements
$[x_1,\dots,x_i,\dots,x_j,\dots,x_n]$, where $x_i=x_j$ for some
$1\leq i \neq j \leq n$, are isomorphic to the given Lie $n$-algebra
under some conditions. These Leibniz $n$-algebras have Cartan subalgebras of different
dimensions and therefore they are not conjugated (see Example \ref{cartan_examp}).
\section{Preliminaries}\label{prel}
\begin{defn}[\cite{Cas1}]
A vector space $L$ with an $n$-ary multiplication
$[-,-,\dots,-]: L^{\otimes n} \to L$ is called a Leibniz $n$-algebra
if it satisfies the following identity
\begin{equation}\label{FI}
[[x_1,x_2, \dots ,x_n],y_2,\dots ,y_n]=\sum_{i=1}^n \ [x_1,\dots ,
x_{i-1},[x_i,y_2,\dots,y_n],x_{i+1},\dots,x_n]
\end{equation}
\end{defn}
It should be noted that if the product $[-,-,\dots,-]$ is
skew-symmetric in each pair of variables, i.e.
\[[x_1, x_2, \dots, x_i,
\dots, x_j, \dots, x_n]= - [x_1, x_2, \dots, x_j, \dots, x_i,
\dots, x_n] \, ,\] then this Leibniz $n$-algebra becomes a Lie
$n$-algebra.
Since in Leibniz $n$-algebras the $n$-ary multiplication is not
necessarily skew-symmetrical, basic notions such as ideals have to be
considered with additional conditions.
\begin{defn} A subspace $I$ of a Leibniz $n$-algebra $L$ is called an $s$-sided ideal of
$L$, if \[[\underbrace{L,\dots,L}_{s-1},I,\underbrace{L,\dots,
L}_{n-s}]\subseteq I.\] If $I$ is $s$-ideal for all $1\leq s \leq
n$, then $I$ is called an ideal.
\end{defn}
\begin{defn}
A proper subalgebra $M$ of a Leibniz $n$-algebra $L$ is called
maximal if the only subalgebra properly containing $M$ is $L$.
\end{defn}
\begin{defn}
The intersection of all maximal subalgebras of a Leibniz $n$-algebra $L$
is a subalgebra denoted by $F(L)$ and it is called the Frattini
subalgebra.
The maximal ideal of $L$ that is contained in $F(L)$ is called the Frattini ideal and it is denoted by $\phi(L)$.
\end{defn}
The following statements which hold for Lie $n$-algebras
\cite{Bai1} can be extended in a similar way to the case of
Leibniz $n$-algebras.
\begin{pr}
Let $L$ be a Leibniz $n$-algebra. Then the following statements
hold:
\begin{enumerate}
\item If $B$ is a subalgebra of $L$ such that $B+F(L)=L$, then $B=L$.
\item If $B$ is a subalgebra of $L$ such that $B+\phi(L)=L$, then
$B=L$.
\end{enumerate}
\end{pr}
\begin{pr}\label{F(L)}
Let $L$ be a Leibniz $n$-algebra and $B$ an ideal of $L$.
Then there exists a proper subalgebra $C$ of $L$ such that $L=B+C$
iff $B \not \subseteq F(L)$.
\end{pr}
Moreover, the assertion of Proposition \ref{F(L)} holds if we
substitute $\phi(L)$ for $F(L)$.
\begin{pr}
Let $C$ be a subalgebra of $L$ and $B$ an ideal of $L$ such that $B \subseteq F(C) \ \big(B \subseteq \phi(C) \big) $.
Then $ B\subseteq F(L)$ \ \big($B\subseteq \phi(L)$, respectively\big).
\end{pr}
\begin{cor}
Let $L$ be a Leibniz $n$-algebra and $B$ a subalgebra of $L$
such that $F(B) \ \big( \phi(B)\big )$ is an ideal of $ L$. Then
$F(B) \subseteq F(L)$ \ \big($\phi(B) \subseteq \phi(L)$, respectively\big).
\end{cor}
\begin{pr}
Let $L$ be a Leibniz $n$-algebra and $B$ an ideal
of $L$. Then the following statements hold:
\begin{enumerate}
\item $ (F(L)+B)/B \subseteq F(L/B)$, \ \big($(\phi(L)+B)/B \subseteq
\phi(L/B)$\big);
\item If $ B \subseteq F(L)$ then $F(L)/B=F(L/B), \ \ \phi(L)/B=\phi(L/B)$;
\item If $ F(L/B)=0 \ \ (\phi(L/B)=0)$, then $ F(L) \subseteq B \ \
(\phi(L)) \subseteq B$.
\end{enumerate}
\end{pr}
\begin{thm}
If a Leibniz $n$-algebra $L$ has a decomposition \[ L=L_1\oplus
L_2\oplus \dots\oplus L_m,\] where $L_i \ (1 \leq i \leq m)$ are
ideals of $L$, then
\begin{enumerate}
\item $ F(L) \subseteq F(L_1) + \dots + F(L_m)$;
\item $\phi(L) = \phi(L_1)+ \dots + \phi(L_m)$.
\end{enumerate}
\end{thm}
Given an arbitrary Leibniz $n$-algebra $L$ consider the following
sequences ($s$ is a fixed natural number, $1\leq s\leq n$):
\begin{align*} L^{<1>_s}=L,\qquad & L^{<k+1>_s}=
\displaystyle[\underbrace{L,\ldots,L}_{(s-1)-\textrm{times}},L^{<k>_s},
\underbrace{L,\ldots,L}_{(n-s)-\textrm{times}}], \\
L^1=L, \qquad & L^{k+1}=\sum\limits_{i=1}^n \ \displaystyle[\underbrace{L,\ldots,L}_{(i-1)-
\textrm{times}},L^k,
\underbrace{L,\ldots,L}_{(n-i)-\textrm{times}}].
\end{align*}
\begin{defn} A Leibniz $n$-algebra $L$ is said to be $s$-nilpotent
(nilpotent) if there exists a natural number $k\in
\mathbb{N}$ ($l\in \mathbb{N}$) such that
$L^{<k>_s}=0$ ($L^l=0$, respectively).
\end{defn}
It should be noticed that for Lie $n$-algebras the above notions of
$s$-nilpotency and nilpotency coincide. Recall also that for
Leibniz algebras (i.e. Leibniz $2$-algebras) the notions of
$1$-nilpotency and nilpotency also coincide \cite{Ayup}.
In \cite[Example 2.2]{cartan}, it is shown that the $s$-nilpotency
property for Leibniz $n$-algebra ($n\geq 3$) essentially depends
on $s$.
Let $H$ be an ideal of a Leibniz $n$-algebra $L$. Put
$H^{(1)_k}=H$ and
\[H^{(m+1)_k}
=\sum_{i_1+\dots+i_k=0}^{n-k}
[\underbrace{L,\dots,L}_{i_1},H^{(m)_k},\underbrace{L,\dots,L}_{i_2},H^{(m)_k},\dots,
\underbrace{L,\dots,L}_{i_k}H^{(m)_k},\underbrace{L,\dots,L}_{n-i_1-\dots-i_k}]\]
for all $1\leq k \leq n$ and $m\geq 1$.
\begin{defn} An $n$-sided ideal $H$ of Leibniz $n$-algebra is said
to be $k$-solvable with index of $k$-solvability equal to $m$ if
there exists $ m\in \mathbb{N}$ such that $H^{(m)_k}=0$ and
$H^{(m-1)_k}\neq 0$.
When $L = H$, $L$ is called a $k$-solvable Leibniz $n$-algebra.
\end{defn}
Notice that this definition agrees with the definition of
$k$-solvability of Lie $n$-algebras given in \cite{Kasymov}.
\begin{defn} We say that a subalgebra $U$ of a Leibniz $n$-algebra $L$
is left subnormal if there exists a chain of subalgebras
$U=U_k\subseteq\cdots\subseteq U_1\subseteq U_0=L$ with each
$U_{i+1}$ an $r$-ideal $(r\neq 1)$ in $U_i$.
\end{defn}
\begin{thm} Let $U$ be a left subnormal subalgebra of Leibniz
$n$-algebra $L$ and $V$ an ideal in $U$ such that $V\subseteq
F(L)$. If $U/V$ is $1$-nilpotent, then $U$ is $1$-nilpotent.
\end{thm}
\begin{proof}
Similar to the proof of \cite[Theorem 3.6]{barnes2}.
\end{proof}
The following statements hold for Lie $n$-algebras
\cite{Bai1} and are also true for Leibniz $n$-algebras.
\begin{cor} If $I \subseteq F(L)$ is an $r$-ideal $(r\neq
1)$ of $L$, then $I$ is $1$-nilpotent. Particularly, $\phi(L)$ is a
$1$-nilpotent ideal of $L$.
\end{cor}
\begin{defn}
In a Leibniz $n$-algebra $L$ the intersection of all maximal
ideals of $L$ is called the Jacobson radical and it is denoted by
$J(L)$.
\end{defn}
\begin{pr}\label{onesidewilliams}
Let $L$ be a finite dimensional Leibniz $n$-algebra. Then
\[F(L)\subseteq[L,L,\dots,L]\textrm{ and } J(L)\subseteq [L,L,\dots,L].\]
Moreover, if $L$ is a $k$-solvable Leibniz $n$-algebra, then
\[J(L)=[L,L,\dots,L].\]
\end{pr}
\begin{thm} Let $L$ be a finite dimensional nilpotent Leibniz $n$-algebra. Then the following statements hold:
\begin{enumerate}
\item Any maximal subalgebra $M$ of $L$ is an ideal of $L$;
\item $F(L)=\phi(L)=J(L)=[L,L,\dots,L]$.
\end{enumerate}
\end{thm}
\section{Right multiplication Operators}\label{rmo}
\begin{defn} A linear map $d$ defined on a Leibniz $n$-algebra $L$ is called a derivation if
\[d([x_1,x_2,\dots , x_n])=\sum_{i=1}^n \ [x_1,\dots d(x_i),\dots , x_n].\]
The space of all derivations of a given Leibniz $n$-algebra $L$ is
denoted by $\operatorname{Der}(L)$.
\end{defn}
The space $\operatorname{Der}(L)$ forms a Lie algebra with respect to the
commutator \cite{cartan}.
Set $A^{\times k}=\underbrace{A\times A \times \dots \times
A}_{k-times}$.
Given an arbitrary element $x=(x_2,\dots,x_n) \in L^{\times(n-1)}$
consider the operator $R(x): L \to L$ of right multiplication
defined by \[R(x)(z)=[z,x_2,\dots,x_n].\]
Any right multiplication operator is a derivation and the space
$R(L)$ of all right multiplication operators forms a Lie ideal of
$\operatorname{Der}(L)$ \cite{cartan}.
\begin{thm}[\cite{cartan} Engel's theorem] A Leibniz $n$-algebra
$L$ is 1-nilpotent if and only if $R(x)$ is nilpotent for all
$x\in L^{\times(n-1)}$.
\end{thm}
In \cite{cartan} it was given an example of a Leibniz $n$-algebra
which admits a non-degenerated right multiplication operator.
This is the significant difference between Leibniz $n$-algebras ($n\geq
3$) on one hand and Leibniz algebras and Lie $n$-algebras on the other.
Below we assume that all right multiplication operators are
degenerated.
The following lemma yields a decomposition of a given vector space
into a direct sum of two subspaces which are invariant with
respect to a given linear transformation.
\begin{lm}[Fitting Lemma] Let $V$ be a vector space and
$A:V\to V$ be a linear transformation. Then $V= V_{0A}\oplus
V_{1A}$, where $A(V_{0A})\subseteq V_{0A}$, $A(V_{1A})\subseteq
V_{1A}$ and $V_{0A}=\{ v \in V|\ A^i(v)=0 \ \mbox{for some} \ i \}$ and
$V_{1A}=\bigcap\limits_{i=1}^\infty A^i(V)$. Moreover, $A_{|
V_{0A}}$ is a nilpotent transformation and $A_{|V_{1A}}$ is an
automorphism. $V_{0A}$ is called the Fitting null-component of $V$ with
respect to $A$.
\end{lm}
\begin{proof}
See \cite[Chapter II, \S 4]{Jac}.
\end{proof}
\begin{defn} An element $h \in L^{\times (n-1)}$ is said to be regular
for the algebra $L$ if the dimension of the Fitting null-component of the space $L$ with respect to $R(h)$ is minimal.
\end{defn}
\begin{lm}[\cite{onnilpotent}]\label{decomposition} Let $L$ be a finite dimensional complex Leibniz n-algebra with
given derivation $d$, and let $L=L_{\alpha}\oplus
L_{\beta}\oplus\cdots\oplus L_{\gamma}$ be the decomposition of
the algebra $L$ into root spaces with respect to the derivation
$d$ (i.e. $L_{\alpha}=\{x\in L | \ (d-\alpha I)^k x=0 \ \mbox{for
some} \ k \}$). Then
\[[L_{\alpha_1},L_{\alpha_2},\dots,L_{\alpha_n}]\subseteq\left\{\begin{array}{ll}
0 & \mbox{if } \alpha_1+\alpha_2+\cdots+\alpha_n \mbox{ is not a root of d}\\
L_{\alpha_1+\alpha_2+\cdots+\alpha_n} & \mbox{if }
\alpha_1+\alpha_2+\cdots+\alpha_n \mbox{ is a root of d} \, .
\end{array}\right.\]
\end{lm}
\begin{pr} \label{zerosum}In a Leibniz $n$-algebra $L$ any right multiplication operator
$ R(a_2,\dots,a_n)$ is a sum of right multiplication operators with zero
root space with respect to $ R(a_2,\dots,a_n)$.
\end{pr}
\begin{proof}
Let $ \alpha_0=0, \alpha_1, \dots, \alpha_k$ be the eigenvalues of
$ R(a_2,\dots,a_n)$. Then $L$ is decomposed into a direct sum
\[L=L_0 \oplus L_{\alpha_1} \oplus \dots \oplus L_{\alpha_k},\] where $
L_{\alpha_i}=\{x\, | \,(R(a_2,\dots,a_n)-\alpha_i I)^m(x)=0 \mbox{
for some } m \in \mathbb{N}\}$.
Consider $ a_i=a_{0}^i+a_{\alpha_1}^i+\dots+a_{\alpha_k}^i,
\,a_{\alpha_m}^i \in L_m, \, 2 \leq i \leq k$. Then for all $ x
\in L$, we have
\begin{multline*}
R(a_2,\dots,a_n)(x)=[x,a_2,\dots , a_n]=
[x,a_{0}^2+a_{\alpha_1}^2+\dots+a_{\alpha_k}^2,\dots ,
a_{0}^n+a_{\alpha_1}^n+\dots+a_{\alpha_k}^n] \\
=[x,a_{0}^2,\dots,a_{0}^n]+[x,a_{\alpha_1}^2,a_0^3\dots,a_{0}^n]+\dots
+[x,a_{\alpha_k}^2,a_{\alpha_k}^3, \dots,a_{\alpha_k}^n]\\
=R(a_{0}^2,a_{0}^3,\dots,a_{0}^n)(x)+R(a_{\alpha_1}^2,a_{0}^3\dots,a_{0}^n)(x)+
\dots + R(a_{\alpha_k}^2,a_{\alpha_k}^3,
\dots,a_{\alpha_k}^n)(x) \, .
\end{multline*}
By Lemma \ref{decomposition}, we obtain that $R(a_2,\dots,a_n)(x)=B(x)
+C(x)$, where $B$ is a sum of right multiplication operators
with zero weight and $C$ is a sum of right multiplication operators with nonzero weights. Then for any $ x\in
L_{\alpha_i}$, we have
\[C(x)=(R(a_2,\dots,a_n)-B)(x) =R(a_2,\dots,a_n)(x)-B(x) \subseteq
L_{\alpha_i} \, ,\] which holds only if $ C(x)=0$ since $C$ adds a
weight. Therefore, $C$ is a zero operator on $L_{\alpha_i}$. Since
$ L=L_0 \oplus L_{\alpha_1}\oplus \dots \oplus L_{\alpha_k}$, we
obtain $C=0$ on $L$.
So, $R(a_2,\dots,a_n)=B$, i.e. is a sum of right multiplication operators with zero weight with respect to
$R(a_2,\dots,a_n)$.
\end{proof}
In \cite{barnes2} it was proved the following result for left
Leibniz algebras which is also valid for right Leibniz algebras,
i.e. Leibniz 2-algebras.
\begin{lm}[\cite{barnes2}] In a Leibniz algebra $L$ for any $a\in L$ there exists
$b\in L_0(R_a)$ such that $L_0(R_b)=L_0(R_a)$.
\end{lm}
Concerning this lemma we establish the following result for the
case $n\geq 3$.
\begin{cor}\label{eigen} If the nonzero eigenvalues $\alpha_1, \dots ,
\alpha_k$ of the right multiplication operator $R(a_2,\dots, a_n)$ in a
Leibniz $n$-algebra $(n\geq 3)$ satisfy
\[\mu_1\alpha_1+\mu_2 \alpha_2+\dots +\mu_k \alpha_k\neq
0,\] for all non-negative integers $ \mu_1,\dots, \mu_k$ such that
\[0<\mu_1+\dots+\mu_k \leq n-1,\]
then there exist $ b_2,b_3,\dots, b_n \in L_0(R(a_2,\dots, a_n))$
such that
\[L_0(R(b_2,\dots,
b_n))=L_0(R(a_2,\dots, a_n)).\]
\end{cor}
\begin{proof}
From Proposition \ref{zerosum} we obtain that $ R(a_2,\dots, a_n) = B$.
From the condition on the eigenvalues we conclude that $B$ consists of
just one right multiplication operator, namely
$B=R(a_{0}^2,a_{0}^3,\dots,a_{0}^n)$. So, if we take $b_i= a_0^i$,
we obtain $ L_0(R(b_2,\dots,b_n))=L_0(R(a_2,\dots, a_n))$.
\end{proof}
A Leibniz $n$-algebra satisfying the conditions of Corollary
\ref{eigen} is given in the following
\begin{exam}[\cite{cartan}] Consider a Leibniz $n$-algebra $L=\langle e_1,e_2,\dots,e_n\rangle$
with the following multiplication:
\[[e_k,e_1,\dots,e_1]=e_k \,\,\,\,\, (2 \leq k \leq m).\]
The right multiplication operator $R(e_1,\dots,e_1)$ has only two
eigenvalues: $0$ and $1$. It is easy to see that the conditions of
Corollary \ref{eigen} are satisfied and $e_1\in
L_0(R(e_1,\dots,e_1))$.
\end{exam}
Below, we present an example which shows the sufficiency of the
condition in Corollary \ref{eigen}.
\begin{exam} Consider an $m$ dimensional Leibniz $n$-algebra $L$ with
the following multiplication:
\[\begin{array}{rcl}
[e_k,e_1,e_2\dots,e_{n-1}] & = & \alpha_k e_k \\
\, [e_{k+1},e_1,e_2,\dots, e_{n-1}] & = & \alpha_{k+1} e_{k+1} \\
& \vdots & \\
\, [e_m,e_1,e_2, \dots, e_{n-1}] & = & \alpha_m e_m \\
\end{array}\]
where $ \{e_1,\dots, e_m\}$ is a basis, $ k < n-1$ and
\ $\displaystyle \sum_{i=k}^{n-1}\alpha_i=0, \alpha_k\cdots \alpha_m \neq 0$.
Then $ L_0(R(e_1,\dots, e_{n-1}))=\{e_1,\dots, e_{k-1}\}$. Since
any other right multiplication operator either coincides with
$R(e_1,\dots, e_{n-1})$ or is identically zero, there does not
exist $b_2,\dots , b_n \in L_0(R(e_1,\dots, e_{n-1}))$ such that $
L_0(R(b_2, \dots, b_n))=L_0(R(e_1,\dots, e_{n-1}))$.
\end{exam}
\begin{defn}[\cite{cartan}] Given a subset $X$ in a Leibniz $n$-algebra, the
$s$-normalizer of $X$ is the set \[N_s(X)=\{a\in L \,|\,
[x_1,\dots,x_{s-1},a,x_{s+1},\dots,x_n]\in X \textrm{ for all }
x_i\in X\}.\] The set $\displaystyle N(X)=\bigcap_{s=1}^n N_s(X)$
is called the normalizer of $X$.
\end{defn}
Notice that, if $X$ is a subalgebra of $L$, then $N(X),
N_s(X)\supseteq X$.
\begin{lm}[\cite{cartan}] \label{desc_lem} Let $M$ be an invariant subspace of a vector space $L$
with respect to a linear transformation $Q:L\to L$. Let
$x=x_0+x_{\alpha}+x_{\beta}+\dots+x_{\gamma}$ be any decomposition of an element $x$
into a sum of characteristic vectors from the corresponding
characteristic spaces $L_\xi (\xi \in \{0,\alpha,\beta,\dots,
\gamma\})$. If $Q(x)\in M$, then $x-x_0 \in M$.
\end{lm}
The following lemma is an extension of
\cite[Lemma 3.2]{barnes2} under the condition $a_2,\dots, a_n \in
L_0(R(a_2,\dots,a_n))$.
\begin{lm} Let $L$ be a Leibniz $n$-algebra and
$R(a_2,\dots,a_n): L \to L$ a right multiplication operator such that $a_2,\dots, a_n \in
L_0(R(a_2,\dots,a_n))$. Then for any subalgebra $U$ containing
$L_0(R(a_2,\dots,a_n))$ the equality $ N(U)=U$ holds.
\end{lm}
\begin{proof}
Let $z\in N_1(U)$. Then $ [z,U,\dots, U] \subseteq U$. Denote
$L_0= L_0(R(a_2,\dots,a_n))$. Then
\[R(a_2,\dots,a_n)(z)=[z,a_2,\dots,a_n]\in [z,L_0,\dots, L_0]
\subseteq [z,U,\dots, U] \subseteq U.\] Hence $ R(a_2,\dots,
a_n)(z) \in U$.
Notice that $ R(a_2,\dots, a_n) (U) =[U,a_2,\dots, a_n]\subseteq
[U,L_0,\dots, L_0]\subseteq [U,U,\dots, U] \subseteq U$ since $U$
is a subalgebra. Therefore, the conditions of Lemma \ref{desc_lem} are
satisfied. Thus $ z-z_0 \in U$. Then $ z\in U$. So we have proved
$ N_1(U)=U$.
Since $U$ is a subalgebra, $N_s(U)\supseteq U$ for all $2\leq s
\leq n$. Then $N(U)=N_1 \bigcap \big(\cap_{s=2}^n
N_s(U)\big)=U$.
\end{proof}
\begin{pr} Let $a_2,\dots, a_n $ be elements of a Leibniz $n$-algebra $L$ such that
$a_2,\dots, a_n \in L_0(R(a_2,\dots,a_n))$. If every maximal
subalgebra is an $i$- and a $j$-ideal $(1\leq i\neq j\leq n)$ in $L$,
then $R(a_2,\dots,a_n)$ is nilpotent.
\end{pr}
\begin{proof} Assume that $L_0(R(a_2,\dots,a_n)) \neq L$.
Then there exists maximal algebra $M$ such that
$L_0(R(a_2,\dots,a_n)) \subseteq M$. Then by previous lemma we
have $N(M)=M$.
Since $M$ is an $i$- and a $j$-ideal $(i\neq j)$, we have
\[[\underbrace{M,\dots M}_{s-1},L,M,\dots, M] \subseteq M\] for
all $1\leq s \leq n$. Thus, $ L=N_s(M)$ for all $1\leq s \leq n$
and $L=N(M)$. Contradiction.
Therefore $L=L_0(R(a_2,\dots,a_n))$ and $R(a_2,\dots,a_n)$ is a
nilpotent operator.
\end{proof}
In \cite[Theorem 2.2]{Will1} there were given several statements
equivalent to nilpotency of the finite dimensional Lie
$n$-algebras. For Leibniz $n$-algebras Proposition
\ref{onesidewilliams} verifies the statement in one direction. The
other direction of the statement in our
case is not true in general. However we establish the following result.
\begin{pr}Let $L$ be a finite dimensional Leibniz $n$-algebra with
condition $a_i\in L_0(R(a_2,\dots,a_n))$ for some $2\leq i \leq n$
for an arbitrary $(a_2,\dots, a_n)\in L^{\times (n-1)}$. If any
maximal subalgebra $M$ of $L$ is an ideal of $L$ then
$L$ is $1$-nilpotent.
\end{pr}
\begin{proof} Assume that $L$ is not $1$-nilpotent. Then there
exists a non-nilpotent right multiplication operator
$R(a_2,\dots,a_n)$. Since $R(a_2,\dots,a_n)$ is non-nilpotent,
the Fitting null-component $L_0(R(a_2,\dots,a_n))\neq L$.
Let $M$ be a maximal subalgebra of $L$ containing
$L_0(R(a_2,\dots,a_n))$. Then $a_i\in
L_0(R(a_2,\dots,a_n))\subseteq M$ for some $2\leq i \leq n$ by
assumption of the proposition. Since $M$ is a maximal subalgebra, it
is also an ideal of $L$. Then
$R(a_2,\dots,a_n)(L)\subseteq M$.
Since $R(a_2,\dots,a_n)$ is an automorphism on
$L_1(R(a_2,\dots,a_n))$, we obtain that $L_1=R(L_1)=L_1\cap M$. Hence
$L_1 \subseteq M$.
Then $L=L_0\oplus L_1\subseteq M \neq L$. This is a contradiction.
Hence, all right multiplication operators are nilpotent.
Therefore, by Engel's theorem $L$ is $1$-nilpotent.
\end{proof}
\section{Invariance of some radicals under derivation}\label{invar}
In the following section we establish some classical results from the
theory of Lie algebras concerning solvability and nilpotency which
are also true in Leibniz algebras and Lie $n$-algebras.
\begin{pr}\label{equality} For an ideal $H$ of a Leibniz $n$-algebra $L$
the equality $(H^{(m)_k})^{(r)_k} =H^{(m+r-1)_k}$ holds for all
$m,r \in \mathbb{N}$.
\end{pr}
\begin{proof} Using induction on $r$ one can easily prove the
assertion of the proposition.
\end{proof}
Even though we can not state that $H^{m_k}$ is an $s$-sided ideal
for all $1\leq s \leq n$, we establish the following result.
\begin{pr} For an ideal $H$ of a Leibniz $n$-algebra $L$,
$H^{(m)_k}$ is a $1$-ideal of $L$ for all $m,k \in \mathbb{N}$.
\end{pr}
\begin{proof} Let $k$ be an arbitrary fixed natural number.
For $m=1$ we have $[H^{(1)_k},L,\dots,L] \subseteq [H,L,\dots,L]
\subseteq H=H^{(1)_k}$ since $H$ is an ideal.
Let $H^{(m)_k}$ be a $1$-ideal, i.e. $[H^{(m)_k},L,\dots,L]\subseteq
H^{(m)_k}$.
Then
\begin{multline*}
[H^{(m+1)_k},L,\dots,L] \\
=\left[\sum_{i_1+\dots+i_k=0}^{n-k} \
[\underbrace{L,\dots,L}_{i_1},H^{(m)_k},\dots,
\underbrace{L,\dots,L}_{i_k},H^{(m)_k},\underbrace{L,\dots,L}_{n-i_1-\dots-i_k}],L,\dots,L\right] \\
=\sum_{i_1+\dots+i_k=0}^{n-k}
\Big[[\underbrace{L,\dots,L}_{i_1},H^{(m)_k},\dots,
\underbrace{L,\dots,L}_{i_k},H^{(m)_k},\underbrace{L,\dots,L}_{n-i_1-\dots-i_k}],L,\dots,L\Big] \, .
\end{multline*}
Since $H^{(m)_k}$ is a $1$-ideal by induction hypothesis, using
identity \eqref{FI} we obtain that
\begin{multline*}
\Big[[\underbrace{L,\dots,L}_{i_1},H^{(m)_k},\dots,
\underbrace{L,\dots,L}_{i_k},H^{(m)_k},\underbrace{L,\dots,L}_{n-i_1-\dots-i_k}],L,\dots,L\Big]\\
\subseteq [\underbrace{L,\dots,L}_{i_1},H^{(m)_k},\dots,
\underbrace{L,\dots,L}_{i_k},H^{(m)_k},\underbrace{L,\dots,L}_{n-i_1-\dots-i_k}] \, .
\end{multline*}
Therefore
\begin{multline*}
[H^{(m+1)_k},L,\dots,L] \\
\subseteq \sum_{i_1+\dots+i_k=0}^{n-k}
[\underbrace{L,\dots,L}_{i_1},H^{(m)_k},\dots,
\underbrace{L,\dots,L}_{i_k},H^{(m)_k},\underbrace{L,\dots,L}_{n-i_1-\dots-i_k}]=H^{(m+1)_k}
\end{multline*}
and $H^{(m+1)_k}$ is a $1$-ideal of $L$.
\end{proof}
\begin{pr}\label{k-sol} Let $I$ be a $k$-solvable ideal of a Leibniz $n$-algebra
$L$ such that $L/I$ is also $k$-solvable. Then $L$ is
$k$-solvable.
\end{pr}
\begin{proof}Let $\phi:L \to L/I$ be the natural homomorphism. Since $L/I$ is $k$-solvable, we have
$0=( L/I)^{(m)_k}=\big(\phi(L)\big)^{(m)_k}=\phi(L^{(m)_k})$ for some
$m\in \mathbb{N}$. Thus $L^{(m)_k} \subseteq I$. Since $I$ is
$k$-solvable, there exists $p\in \mathbb{N}$ such that
$I^{(p)_k}=0$. Therefore by Proposition \ref{equality} we have
$L^{(m+p-1)_k}=(L^{(m)_k})^{(p)_k}\subseteq I^{(p)_k}=0$ and so $L$
is $k$-solvable.
\end{proof}
By induction it is easy to prove that if $I$ is a $k$-solvable
ideal of a Leibniz $n$-algebra $L$, then $I$ is also
$(k+p)$-solvable for all $p\in \mathbb{N}$.
Using standard methods and Proposition \ref{k-sol} we obtain that
the sum of $k$-solvable ideals is also $k$-solvable. Now
let $H$ be a maximal $k$-solvable ideal in a finite
dimensional Leibniz $n$-algebra $L$ and let $K$ be an arbitrary
$k$-solvable ideal of $L$. Then $H+K$ is also
$k$-solvable and $H+K \supseteq H$. Since $H$ is a maximal
$k$-solvable ideal, we obtain that $H+K=H$. Therefore we can
define the maximal $k$-solvable ideal as the sum of all the
$k$-solvable ideals in $L$ and call it the \emph{$k$-solvable radical}.
The following formula for a
derivation $d:L \to L$ of a Leibniz $n$-algebra $L$ over a field
$\mathbb{K}$ of characteristic zero, for any $k \in \mathbb{N}$, was given in \cite{onnilpotent}:
\begin{equation} \label{der}
d^k([x_1,\dots,x_n]) = \sum_{i_1+i_2+\cdots+i_n=k} \ \frac{k!}{i_1!i_2!\dots
i_n!} \ [d^{i_1}(x_1),d^{i_2}(x_2),\dots,d^{i_n}(x_n)] \, .
\end{equation}
\begin{pr} \label{inclusion} Let $I$ be an ideal of a
Leibniz $n$-algebra $L$ and $d \in \operatorname{Der}(L)$. Then \[\big(d(I)\big)^{(m)_k}
\subseteq I+d^{k^{m-1}}\big(I^{(m)_k}\big)\] for all $m\in \mathbb{N}$ and
$1\leq k \leq n$.
\end{pr}
\begin{proof} For $m=1$ we have $d(I)\subseteq
I+d(I)$ which obviously holds.
Assume that $\big(d(I)\big)^{(m)_k} \subseteq I+d^{k^{m-1}}\big(I^{(m)_k}\big)$.
Using formula \eqref{der} we verify the inclusion for $m+1:$
\begin{multline*}\big(d(I)\big)^{(m+1)_k}\\
= \sum_{i_1+\dots+i_k=0}^{n-k}
[\underbrace{L,\dots,L}_{i_1},d(I)^{(m)_k},\underbrace{L,\dots,L}_{i_2},d(I)^{(m)_k},\dots,
\underbrace{L,\dots,L}_{i_k},d(I)^{(m)_k},\underbrace{L,\dots,L}_{n-i_1-\dots-i_k}]\\
\subseteq \sum_{i_1+\dots+i_k=0}^{n-k}
[\underbrace{L,\dots,L}_{i_1},I+d^{k^{m-1}}(I^{(m)_k}),\dots,
\underbrace{L,\dots,L}_{i_k},I+d^{k^{m-1}}(I^{(m)_k}),\underbrace{L,\dots,L}_{n-i_1-\dots-i_k}]\\
\subseteq I+d^{k^m}\left(\sum_{i_1+\dots+i_k=0}^{n-k}
[\underbrace{L,\dots,L}_{i_1},I^{(m)_k},\dots,
\underbrace{L,\dots,L}_{i_k},I^{(m)_k},\underbrace{L,\dots,L}_{n-i_1-\dots-i_k}]\right)
=I+d^{k^m}\big(I^{(m+1)_k}\big) \, .
\end{multline*}
Therefore, the assertion of the proposition is true.
\end{proof}
Also, in \cite{onnilpotent}, it was shown that for any ideal I
of $L$ and $d\in \operatorname{Der}(L)$ the $I+d(I)$ is also an ideal
of $L$.
\begin{thm}\label{sol_rad} Let $J$ be the $k$-solvable radical of a finite
dimensional Leibniz $n$-algebra $L$ over a field $\mathbb{K}$
of characteristic zero. Then $d(J)\subseteq J$ for any $d\in
\operatorname{Der}(L)$.
\end{thm}
\begin{proof}
Let $s\in \mathbb{N}$ be such $J^{(s)_k}=0$. Then by Proposition
\ref{inclusion} we have $\big(d(J)\big)^{(s)_k} \subseteq
J+d^{k^{s-1}}\big(J^{(s)_k}\big)=J$.
Using formula \eqref{der}, we obtain that $\big(J+d(J)\big)^{(s)_k}\subseteq
J+\big(d(J)\big)^{(s)_k} \subseteq J$. Now by Proposition \ref{equality} we
have $\big(J+d(J)\big)^{(2s-1)_k}=\Big(\big(J+d(J)\big)^{(s)_k}\Big)^{(s)_k}\subseteq
J^{(s)_k}=0$. But this means that $J+d(J)$ is a $k$-solvable ideal.
Since $J$ is a $k$-solvable radical, we obtain that $J+d(J)\subseteq J$
and therefore $d(J) \subseteq J$.
\end{proof}
Similarly as in \cite{Bai2} we introduce the following series
for a $1$-sided ideal $I$ of a Leibniz $n$-algebra $L:$
\[I^{[1]}=I, \quad I^{[k+1]}=[I^{[k]},I,L,\dots,L] \quad (k\geq 1).\]
By a simple induction using identity \eqref{FI} it can be proved that for
any $1$-sided ideal $I$ and for all $n\in\mathbb{N}$, $I^{[n]}$ is a
$1$-sided ideal.
\begin{defn} A $1$-sided ideal $I$ is called $K_1$-nilpotent, if
there exists $k\in\mathbb{N}$ such that $I^{[k]}=0$.
\end{defn}
The introduced type of nilpotency is also known as nilpotency in
the sense of Kuzmin for Lie $n$-algebras. Identity \eqref{FI} is
organized in such way, that the elements of the above introduced
series are $1$-ideals. However, if we change the position of
$I^{[k]}$ in the product defining $I^{[k+1]}$ from the first to
any other, we are not able to state that the elements of the obtained
series will be $s$-ideals of $L$ for any $2\leq s \leq n$.
\begin{pr} Let $I$ and $J$ be $K_1$-nilpotent $1$-sided ideals. Then
$I+J$ is also a $K_1$-nilpotent $1$-sided ideal.
\end{pr}
\begin{proof} First, observe that
\[[I^{[p]}\cap J^{[q]},I,L\dots,L]\subseteq [I^{[p]},I,L,\dots,L]=I^{[p+1]} \, ,\]
and since $J^{[q]}$ is a $1$-ideal, we get
\[ [I^{[p]}\cap J^{[q]},I,L\dots,L]\subseteq
[J^{[q]},I,L\dots,L]\subseteq J^{[q]}.\]
Therefore,
\[[I^{[p]}\cap J^{[q]},I,L\dots,L]\subseteq I^{[p+1]}\cap
J^{[q]}.\] Analogously,
\[[I^{[p]}\cap J^{[q]},J,L\dots,L]\subseteq I^{[p]}\cap
J^{[q+1]}.\]
We have $(I+J)^{[1]}=I+J=I^{[1]}+J^{[1]}$.
Now assume that
\[(I+J)^{[k]}\subseteq I^{[k]}+ \big(I^{[k-1]}\cap J^{[1]}\big) +\dots +
\big( I^{[1]}\cap J^{[k-1]} \big)+J^{[k]}.\]
Then
\begin{multline*}
(I+J)^{[k+1]}=[(I+J)^{[k]},I+J,L,\dots,L]\\
\subseteq [(I+J)^{[k]},I,L,\dots,L]+[(I+J)^{[k]},J,L,\dots,L]\\
\subseteq [I^{[k]},I,L,\dots,L]+\sum_{r=1}^{k-1} \ [I^{[k-r]}\cap
J^{[r]},I,L,\dots,L]+[J^{[k]},I,L,\dots,L]\\
+ [I^{[k]},J,L,\dots,L]+\sum_{r=1}^{k-1} \ [I^{[k-r]}\cap
J^{[r]},J,L,\dots,L]+[J^{[k]},J,L,\dots,L]\\
\subseteq I^{[k+1]}+\Big(\sum_{r=1}^{k-1} \ I^{[k-r+1]}\cap
J^{[r]}\Big)+\big(I^{[1]}\cap J^{[k]}\big) \\
+\big(I^{[k]}\cap J^{[1]}\big)+\Big(\sum_{r=1}^{k-1} \
I^{[k-r]}\cap J^{[r+1]}\Big)+J^{[k+1]}\\
\subseteq I^{[k+1]}+\big( I^{[k]}\cap J^{[1]}\big)+\dots +\big(I^{[1]}\cap J^{[k]}\big)+J^{[k+1]} \,.
\end{multline*}
Hence, for any $n\in\mathbb{N}$ we have \[(I+J)^{[n]}\subseteq
I^{[n]}+\big(I^{[n-1]}\cap J^{[1]}\big)+\dots +\big(I^{[1]}\cap J^{[n-1]}\big)+
J^{[n]}.\]
So if $I^{[n_1]}=0$ and $J^{[n_2]}=0$, then for $n=n_1+n_2$ every
summand in the above sum is zero. Therefore $(I+J)$ is also
$K_1$-nilpotent.
\end{proof}
\begin{cor} Let $I$ and $J$ be $K_1$-nilpotent ideals. Then
$I+J$ is also a $K_1$-nilpotent ideal.
\end{cor}
Let $I$ be a maximal $K_1$-nilpotent ideal in a finite
dimensional Leibniz $n$-algebra $L$ and let $J$ be an arbitrary
$K_1$-nilpotent ideal of $L$. Then $I+J$ is also
$K_1$-nilpotent and $I+J \supseteq I$. Since $I$ is a maximal
$K_1$-nilpotent ideal, we obtain that $I+J=I$. Therefore we
can define the maximal $K_1$-nilpotent ideal as the sum of
all the $K_1$-nilpotent ideals in $L$ and call it the \emph{$K_1$-nilradical}. Notice that, the $K_1$-nilradical do not possess the
properties of the radical in the sense of Kurosh.
\\
Using the same argumentation as in the proof of Proposition
\ref{inclusion} and Theorem \ref{sol_rad} the following statements
can be established.
\begin{pr} Let $I$ be an ideal of a Leibniz
$n$-algebra $L$. Then for any $d\in \operatorname{Der}(L)$ we have
$\big(d(I)\big)^{[n]}\subseteq I+d^n(I^{[n]})$ for all $n\in\mathbb{N}$.
\end{pr}
\begin{thm} Let $J$ be the $K_1$-nilradical of a Leibniz
$n$-algebra $L$. Then for any $d\in \operatorname{Der}(L)$ we have $d(J)\subseteq
J$.
\end{thm}
Analogously, we can establish similar results concerning the
nilpotency and $s$-nilpotency.
By induction it is not difficult to show that the sum of
$s$-nilpotent (nilpotent) ideals of Leibniz $n$-algebra
$L$ is also $s$-nilpotent (nilpotent, respectively)
ideal of $L$.
Now let $N$ be a maximal $s$-nilpotent (nilpotent)ideal
in a finite dimensional Leibniz $n$-algebra $L$ and let $M$ be an
arbitrary $s$-nilpotent ideal of $L$. Then $N+M$ is
also $s$-nilpotent (nilpotent, respectively) and $N+M \supseteq
N$. Since $N$ is maximal $s$-nilpotent (nilpotent, respectively)
ideal, we obtain $N+M=N$. Therefore we can define the
maximal $s$-nilpotent (nilpotent, respectively) ideal as
the sum of all the $s$-nilpotent (nilpotent, respectively) ideals in $L$
and call it the \emph{$s$-nilradical} (\emph{nilradical}, respectively).
\begin{pr}Let $J$ be the $s$-nilradical (nilradical) of a finite
dimensional Leibniz $n$-algebra $L$ over a field $\mathbb{K}$
of characteristic zero. Then $\big(J+d(J)\big)^{<m>_s} \subseteq
J^{<m>_s}+\big(d(J)\big)^{<m>_s}$ \ \Big($\big(J+d(J)\big)^{m} \subseteq
J^{m}+\big(d(J)\big)^{m}$, respectively\Big).
\end{pr}
\begin{proof} Analogous to the proof of Proposition \ref{inclusion}.
\end{proof}
\begin{thm}Let $J$ be the $s$-nilradical (nilradical) of a finite
dimensional Leibniz $n$-algebra $L$ over a field $\mathbb{K}$ of
characteristic zero. Then $d(J)\subseteq J$ for any $d\in
\operatorname{Der}(L)$.
\end{thm}
\begin{proof} Analogous to the proof of Theorem \ref{sol_rad}.
\end{proof}
\section{Non-conjugacy of Cartan Subalgebras}\label{no_conj}
In this section we consider Cartan and Frattini subalgebras of
Leibniz $n$-algebras.
\begin{defn}[\cite{cartan}] A subalgebra $C$ of a Leibniz $n$-algebra $L$ is said to be Cartan subalgebra if
\begin{itemize}
\item[a)] $C$ is 1-nilpotent;
\item[b)] $C=N_1(C)$.
\end{itemize}
\end{defn}
The importance of considering 1-normalizer in the definition of
Cartan subalgebras was shown in \cite{Omirov}.
\begin{pr}[\cite{cartan}] Let $C$ be a nilpotent subalgebra of a Leibniz $n$-algebra $L$.
Then $C$ is a Cartan subalgebra if and only if it coincides with $L_0$
in the Fitting decomposition of the algebra $L$ with respect to $R(C)$.
\end{pr}
Similarly as in \cite{Kasymov}, if $L$ is a direct sum of Leibniz
$n$-algebras $L_i,\, 1\leq i \leq k$, and $C_i$ are Cartan
subalgebras of $L_i$, then $C= \oplus_{i=1}^k C_i$ is a Cartan subalgebra of
$L$ and any Cartan subalgebra of $L$ has the same form.
The following result
concerning the regular elements of a Leibniz $n$-algebra was established in \cite{cartan}:
\begin{thm}[\cite{cartan}] Let $L$ be a Leibniz $n$-algebra over an infinite
field and let $x$ be a regular element for $L$. Then the Fitting
null-component $L_0$ with respect to the operator $R(x)$ is a
1-nilpotent subalgebra of $L$.
\end{thm}
In Leibniz algebras and Lie $n$-algebras the corresponding theorem states
that $L_0$ is a Cartan subalgebra. However, in \cite{cartan} we give
an example of a Leibniz $n$-algebra in which this result is not
true. Now we establish this result under some restrictions.
\begin{pr}Let $L$ be a Leibniz $n$-algebra over an infinite
field and let $x=(x_2,\dots,x_n)\in L^{\times (n-1)}$ be a regular
element for $L$ such that $x_2,\dots,x_n \in
L_0(R(x_2,\dots,x_n))$. Then the Fitting null-component $L_0$
with respect to operator $R(x)$ is a Cartan subalgebra of $L$.
\end{pr}
\begin{proof} Due to the previous theorem, we need to prove $N_1(L_0)=L_0$.
Let $y\in N_1(L_0)$. Then $[y,x_2,\dots,x_n]\in
[y,L_0,\dots,L_0]\subseteq L_0$. Hence $y\in L_0$. Therefore,
$N(L_0)\subseteq L_0$ and since $L_0$ is a subalgebra $N_1(L_0)\supseteq
L_0$. Thus, $L_0=N_1(L_0)$ and $L_0$ is a Cartan subalgebra of $L$.
\end{proof}
Now let us construct a Leibniz $n$-algebra $L$ such that the quotient $n$-algebra $L/I$ is a
simple Lie $n$-algebra, where
\[ I = \mbox{ideal} \ \langle [x_1,\dots,x_i, \dots
, x_j, \dots , x_n] \ | \ \exists \, i, j: x_i=x_j \rangle\] is an
ideal of $L$.
\begin{exam}\label{examp_simple}
Let $\{e_1,\dots, e_{n+1},x_1,\dots, x_m\}$ be a basis of $L$.
Consider an algebra with the following multiplication:
\begin{align*}[e_{1}, \dots, e_{i-1},e_{i+1},\dots,
e_{n+1}] \ = & \ e_{i} \\
[x_k,e_j,\dots, e_j] \ = & \ \alpha_{kj} x_k \, ,
\end{align*}
where $ 1 \leq i,j\leq n+1,\, 1\leq k \leq
m, \ |\alpha_{k1}|^2+\dots+|\alpha_{k \, n+1}|^2 \neq 0$ for all $k$,
and the multiplication is skew symmetric in all the variables on $\langle
e_1,\dots,e_{n+1}\rangle$.
Then this algebra is a Leibniz $n$-algebra and $I=\langle x_1,
\dots, x_m\rangle $.
\end{exam}
Note that $L/I$ is a simple Lie $n$-algebra and by \cite[Theorem 2.2]{Bai1}
we have that $F(L/I)=0$. Hence $F(L) \subseteq I$.
\begin{pr} In Example \ref{examp_simple}, $F(L)=0$.
\end{pr}
\begin{proof}
Consider the subspaces \[L_k=\langle e_1,\dots, e_{n+1}, x_1,\dots,
x_{k-1}, x_{k+1}, \dots, x_m \rangle ,\, 1\leq k \leq m.\] From
the multiplication table we get that they are subalgebras. Since
the dimension of these subalgebras is $n+m=\dim L -1$, they are
maximal subalgebras.
Hence, $\displaystyle F(L) \subseteq \bigcap_{k=1}^m L_k =\langle
e_1,\dots, e_{n+1}\rangle$. But $F(L) \subseteq I=\langle x_1,
\dots, x_m \rangle$. Thus $F(L)=0$.
\end{proof}
Below, we present a more general construction.
Let us consider an arbitrary Lie $n$-algebra with the basis
$e_1,\dots e_{n+1}$ and the conditions
\[[e_i,f_2,\dots, f_n]\in \langle e_1,\dots, e_{i-1},
e_{i+1},\dots, e_{n+1}\rangle,\] for all $f_2,\dots, f_n \in
\{e_1,\dots, e_{n+1}\}, \, 1\leq i \leq n+1$.
One of the Lie $n$-algebras with these conditions is a simple
Lie $n$-algebra. Complement this algebra with independent vectors
$x_1,\dots,x_m$ with the following multiplication
\[[x_k,e_p,\dots, e_p]=\alpha_{kp}^1x_1+\alpha_{kp}^2 x_2 +\dots +
\alpha_{kp}^m x_m\] for all $1\leq k \leq m, 1\leq p \leq n+1$.
Checking identity \eqref{FI} we will find restrictions on the
coefficients $\alpha_{ij}^k:$
\[\sum_{i=1}^m \alpha_{kp}^i \alpha_{iq}^j =\sum_{i=1}^m
\alpha_{kq}^i \alpha_{ip}^j\] for all $1\leq k,j\leq m,\, 1\leq p,
q \leq n+1$.
Hence, the satisfaction of the above condition guaranties that the
supplemented algebra is a Leibniz $n$-algebra.
Particularly, in this way, one can supplement simple Lie
$n$-algebras till Leibniz $n$-algebras.
On the ground of Example \ref{examp_simple} we give the following
\begin{exam} \label{cartan_examp} Let $L_s \ (1\leq s \leq n+1)$ be a
Leibniz $n$-algebra with the basis $\langle e_1,e_2,\dots,
e_{n+1},x_1,\dots,x_m\rangle$ and the following multiplication:
\begin{align*}
[e_1,\dots,e_{p-1},e_{p+1},\dots,e_{n+1}] \ = & \ e_p\, , & 1\leq p \leq n+1,\\
[x_k,e_k,e_k,\dots,e_k] \ = & \ x_k \,, & 1\leq k \leq s,\\
[x_{s+i},e_s,e_s,\dots,e_s] \ = & \ x_{s+i}\, , & 1\leq i \leq m-s \, ,
\end{align*}
where the multiplication is skew symmetric in all the variables on
$\langle e_1,e_2,\dots, e_{n+1}\rangle$.
Then \[\begin{array}{rl}
H_1= & \langle e_1,e_2,\dots,e_s, e_{s+1},\dots,e_{n-1}\rangle \\
H_2= & \langle e_1,e_2,\dots,e_s, e_{s+2},\dots,e_{n-1},e_n \rangle \\
H_3= & \langle e_1,e_2,\dots,e_s, e_{s+3},\dots,e_{n-1},e_{n+1} \rangle \\
\end{array}\]
are $n-1$ dimensional Cartan subalgebras.
The subalgebras
\begin{align*}
N_1 \ = & \ \langle x_1,e_2,e_3,\dots,e_n \rangle \\
N_2 \ = & \ \langle e_1,x_2,e_3,\dots,e_n \rangle\\
\vdots & \\
N_{s-1} \ = & \ \langle e_1,\dots,e_{s-2},x_{s-1},e_s,\dots,e_n \rangle
\end{align*}
are $n$ dimensional Cartan subalgebras.
The subalgebras
\begin{align*}
M_1 \ = & \ \langle e_1,e_2,\dots,e_{s-1},e_{s+1},e_{s+2},\dots,e_n,x_s,x_{s+1},\dots, x_m \rangle \\
M_2 \ = & \ \langle e_1,e_2,\dots,e_{s-1},e_{s+2},e_{s+3},\dots,e_{n+1},x_s,x_{s+1},\dots, x_m \rangle
\end{align*}
are $m+n-s$ dimensional Cartan subalgebras.
\begin{align*}
C_1 \ = & \ \langle x_1,e_2,\dots,e_{s-1},e_{s+1},\dots,e_{n+1},x_s,x_{s+1},\dots, x_m \rangle \\
C_2 \ = & \ \langle e_1,x_2,\dots,e_{s-1},e_{s+1},\dots,e_{n+1},x_s,x_{s+1},\dots, x_m \rangle \\
\vdots & \\
C_{s-1} \ = & \ \langle e_1,\dots,e_{s-2},x_{s-1},e_{s+1},\dots,e_{n+1},x_s,x_{s+1},\dots, x_m \rangle
\end{align*}
are $m+n-s+1$ dimensional Cartan subalgebras.
\end{exam}
In the considered Leibniz $n$-algebra we found Cartan subalgebras of dimensions
$n-1, n , n+m-s$ and $n+m-s+1$. Hence, in general, Cartan subalgebras of a given
Leibniz $n$-algebra are not conjugated.
Here we give a theorem that establishes the conjugacy of Cartan
subalgebras under some restrictions on the Leibniz $n$-algebra.
\begin{thm} \label{condition} Let $L$ be a finite dimensional Leibniz $n$-algebra and $H$
be a Cartan subalgebra of $L$. Suppose that
\begin{itemize}
\item[(i)] the multiplication is skew symmetric in the first two variables; and that
\item[(ii)] for any element $h=(h_1,\dots,h_{n-1})\in H^{\times (n-1)}$, we
have $h_i\in \operatorname{Ker} R(h)$ for all $1\leq i \leq n-1$.
\end{itemize}
Then there is a regular element $h\in H^{\times (n-1)}$ such that
$H=L_0(R(h))$.
\end{thm}
\begin{proof} Suppose that $H$ is a Cartan subalgebra of a Leibniz
$n$-algebra and $L=L_0\oplus L_{\alpha_1}\oplus\dots\oplus
L_{\alpha_s}$ is the decomposition of $L$ into a direct sum of
root subspaces with respect to $H$ and $\Delta =\{ \alpha_1,\dots,
\alpha_s\}$ is the set of non-zero roots of $H$ in $L$. Then the
functions $\alpha_i$ are multilineal and, in particular,
polynomial. Since $H^{\times (n-1)}$ is an irreducible variety,
it follows that $\alpha_1\alpha_2\cdots \alpha_s$ is also a
non-zero polynomial function from $H^{n-1}$ to the ground field of the
Leibniz $n$-algebra. Hence, $\alpha_1(h^0)\alpha_2(h^0)\cdots
\alpha_s(h^0)\neq 0$ for some $h^0=(h_0,h^0_1,\dots,h^0_{n-2})$.
This means that the characteristic roots of the restriction
$\overline{R}(h^0)$ of the endomorphism $R(h^0)$ to
$L_1\big(R(h^0)\big)=\sum_{\alpha\in \Delta} L_{\alpha}$ are all nonzero,
and hence $\overline{R}(h^0)$ is a non-degenerate operator.
The proof of the theorem is based on the proof of conjugacy of Cartan
subalgebras in Lie $n$-algebras given by Kasymov \cite{Kasymov}.
Similarly, we define a polynomial function $P$ on $L$ by
\[P(x)=\exp R(x_1,h^0_1,\dots, h^0_{n-2})\cdots \exp R(x_s,h^0_1,\dots,
h^0_{n-2})(h),\] where $x=h+x_1+\dots+x_s, h\in H=L_0\big(R(h^0)\big), \ x_i
\in L_{\alpha_i}$.
Notice that, if a right multiplication operator $R(x)$ is
nilpotent, then $\exp R(x)$ is an inner automorphism of the
Leibniz $n$-algebra $L$. Automorphisms of this kind generate a
certain subgroup $G_0$ in the group $G=\operatorname{Aut} L$. Elements of $G_0$
are called \emph{special (invariant) automorphisms}.
Using the skew symmetrical property of the multiplication in the first two
variables, we establish that the differential $d_{h^0}P$ of $P$ at
a point $h^0$ is an epimorphism. Hence, by facts from algebraic
geometry in \cite{Kasymov}, this polynomial function $P$ is
dominating, i.e. for any non-zero polynomial function $f$ on $L$ there
exists a non-zero polynomial function $g$ on $L$ such that every
$y\in L$ with $g(y)\neq 0$ is represented as $y=P(x)$, where
$f(x)\neq 0$.
Assuming that for any regular element $h=(h_1,\dots,h_{n-1})\in
H^{\times (n-1)}$ we have $h_i\in \operatorname{Ker} R(h)$ for all $1\leq i \leq
n-1$, then
\[P(h_i)=\left(\prod_{j=1}^s \exp
R(x_j,h_1,\dots,h_i,\dots, h_{n-2})\right) (h_i)=h_i.\] Hence, we
can use similar induction as in \cite{Kasymov} to prove the
existence of a regular element $h\in H^{\times (n-1)}$ such that
$H=L_0\big(R(h)\big)$.
\end{proof}
Under the conditions of Theorem \ref{condition} the following
theorem can be proved similarly as in the case of Lie
$n$-algebras \cite{Kasymov}.
\begin{thm} Let $L$ be a finite-dimensional Leibniz $n$-algebra
which satisfies the conditions (i)-(ii) of Theorem
\ref{condition}. If $H$ and $K$ are Cartan subalgebras of $L$,
then there exists a special automorphism $\delta$ of $L$ such that
$H=\delta(K)$.
\end{thm}
\section*{Acknowledgments}
The second author was supported by MICINN grant MTM 2009-14464-C02 (European
FEDER support included) and by Xunta de Galicia grant Incite 09207215 PR. The third named author would like to
acknowledge the hospitality of the University of Santiago de
Compostela (Spain). He was supported by Grant NATO-Reintegration
ref. CBP.EAP.RIG. 983169.
The last named author would like to acknowledge ACTP OEA-AC-84 for
a given support.
|
\section{Introduction}
The notions of quantum liquids and their instabilities are
paradigmatic for condensed matter physics~\cite{Leggett2006}. For
multicomponent fluids, an important set of instabilities is associated
with interactions between components. A classic example is the Cooper
instability of a spin-$1/2$ Fermi liquid~: even an infinitesimal
attractive coupling between fermions of opposite spins drives a phase
transition into the Bardeen-Cooper-Schrieffer
superconductor~\cite{Schrieffer1999}. A one-dimensional (1D)
counterpart of the Fermi liquid, the spinful Luttinger liquid, has a
similar instability, where an attractive inter-spin coupling opens a
gap in the spin channel~\cite{Giamarchi1995, Giamarchi2004}.
Traditionally, the bulk of the discussion on two-species liquids
assumed the SU(2) spin symmetry. The recent years have witnessed a growing availability of experimental
studies of mixtures of unlike particles. This includes loading
ultracold atoms to spin-dependent optical lattices~\cite{Mandel2003}, and
trapping atoms of different masses~\cite{Wille2008}
or even different statistics~\cite{Giorgini2008}. While most of
experimental progress so far is in the domain of ultracold atoms, we
stress that the relevance of such \emph{asymmetric} mixtures is not
confined to the realm of cold gases~: dealing with more traditional
solid state systems, one faces an asymmetric mixture situation as soon
as the Fermi level spans several bands (which \textit{a priori} need
not be equivalent). This setup is typical for such diverse materials
as semi-metallic compounds, mixed-valence materials, organic
superconductors~\cite{Penc1990}, small radius nanotubes
\cite{Carpentier2006}, and even graphene-based
heterostructures~\cite{Martin2008}.
A generic question immediately arises~: given a two-component mixture,
what is the role of (the lack of) SU(2) symmetry? Or, more precisely,
does the symmetry between components limit the set of instabilities of
a liquid? Clearly, the answer might depend on the universality
class of the liquid, and on the particular way the symmetry is
broken. The simplest albeit non-trivial way of breaking the symmetry
is to assume species-dependent masses of the particles. Even if we
consider the few-body problem, this is known to bring new physics like
the Efimov phenomenon~: while for an equal-mass Fermi liquid the only
allowed bound state is a Cooper pair, three-body bound states
(trimers) appear once the mass ratio exceeds a certain
threshold~\cite{Braaten2006}.
The atom-dimer scattering is strongly affected by
the mass asymmetry ~\cite{Petrov2003}
and the ultimate fate of a
Fermi liquid in presence of the Efimov effect is currently an open
question being actively investigated~\cite{Levinsen2009}.
All these theoretical considerations are strongly motivated by cold-atom experiments which have recently achieved
degeneracy of Fermi gases with different masses~\cite{Wille2008}
and spin-imbalanced two-component fermionic gases~\cite{Partridge2006}.
\begin{figure}[b]
\centering
\includegraphics[width=\columnwidth,clip]{fig1}
\caption{Qualitative formation of trimers in the weak-coupling picture
of the bosonization (on the left) and in the strong-coupling picture for
large interactions (on the right).}
\label{fig:trimers}
\end{figure}
The physics of 1D quantum many-body systems offers powerful methods~\cite{Giamarchi2004},
both analytical and numerical, to have quantitative predictions on the fate of the
Luttinger liquid in the presence of perturbations.
The role of mass asymmetry for a two-component Luttinger liquid has
been investigated in the renormalization group (RG) framework
originally in the context of solid state physics~\cite{Muttalib1986, Penc1990, Loss1994},
and recently revisited mostly in the context of cold atoms~\cite{Cazalilla2003, Mathey2004, Cazalilla2005, Mathey2007,
Mathey2007a, Lu2009, Crepin2010, Orignac2010, Tsvelik2010}, and
supplemented by numerical investigations~\cite{Pollet2006, Mering2008, Batrouni2009, Wang2009}.
Overall the consensus was that the only new instability arising
due to asymmetry is the collapse (demixing) instability for large
asymmetry and/or strong interspecies attraction (repulsion). Recently,
a novel family of instabilities was predicted~\cite{Burovski2009} to
exist due to the interplay between \emph{polarization and asymmetry}~:
These instabilities only take place for \emph{polarized} mixtures of
either statistics, and are characterized by the locking of the ratio
of the densities to a \textit{rational} value. Subsequent work in
Ref.~\onlinecite{Orso2010} elucidated the relation of these
instabilities and existence of few-body bound states. A qualitative
picture of the mode-locking mechanism and the strong-coupling limit of
the trimer formation is given in Fig.~\ref{fig:trimers}. The latter
regime recalls another approach to multi-particle bound-states which
is the use of many-colors ($N$-component) fermions~\cite{Wu2005, Capponi2008, Roux2009}
with which the physics of the trimers share
qualitative features.
This paper is divided in two main parts : the first one investigates
in detail the bosonization approach and the mode-locking mechanism
mentioned above, while the second is dedicated to the specific but
important example of the 1D asymmetric Hubbard model
using the density-matrix renormalization group (DMRG)
technique~\cite{White1992}. The
predictions of the first part account for most of the numerical data, but a
more phenomenological Bose-Fermi picture is proposed as a
complementary analysis.
Other important questions such as the effect
of a trapping potential or the emergence of crystal phases are
eventually addressed.
\section{Bosonization analysis}
\label{sec:boso}
In this section, we describe the salient features of the effective
bosonic field theory appropriate to a 1D mixture of two distinct
fermionic (or bosonic) atoms. The aim of this section is to give a
bosonization interpretation for the formation of few-body bound states
and their effective behavior through a mode-locking mechanism between
the two species. Predictions on the nature of the resulting phase are
then made. The theory is a priori valid for models in the continuum or
the continuous version of lattice models at generic
(i.e. non-commensurate) densities. The effects of the presence of the
lattice on certain commensurate densities will be briefly discussed in
section~\ref{sec:lattice}. Notation conventions are standard and taken from
Ref.~\onlinecite{Giamarchi2004}.
\subsection{Mode locking mechanism}
\label{sec:modecoupling}
The two species are labeled by a pseudo-spin index $\sigma =
{\up,\down}$ and their corresponding densities $n_{\sigma}$ such that
$n=n_{\up}+n_\down$ is the total density. Each species can be
described by a scalar field $\phi_\sigma$ and its dual
$\theta_\sigma$. The creation operators can be expressed
as a function of these fields, with, for fermions
\begin{equation}
\label{eq:fermions}
\Psi^\dagger_\sigma(x) \sim \Big( n_{\sigma}- \frac{1}{\pi} \dx\phi_\sigma
\Big)^{1/2} \sum_{p} e^{ i (2p+1)(k_{\sigma} x -\phi_\sigma)} \, e^{-i\theta_\sigma}\,,
\end{equation}
and, for bosons,
\begin{equation}
\label{eq:bosons}
b^\dagger_\sigma(x) \sim \Big( n_{\sigma}- \frac{1}{\pi} \dx\phi_\sigma
\Big)^{1/2} \sum_{p} e^{ i 2p (k_{\sigma} x-\phi_\sigma) } \, e^{-i\theta_\sigma}\;.
\end{equation}
We have included all higher harmonics : as a consequence, the summation
is over all integers~\cite{Haldane1981}. The ``Fermi momenta'',
$k_\sigma=\pi n_\sigma$ are a priori not equal to each other,
corresponding to a spin-imbalanced situation. The
density operators $\hat{n}_\sigma$ read
\begin{equation}
\label{HaldaneN}
\hat{n}_\sigma(x) \sim \Big( n_{\sigma}- \frac{1}{\pi} \dx\phi_\sigma\Big)
\sum_p e^{ i2p(k_\sigma x -\phi_\sigma)} \; .
\end{equation}
The effective low-energy Hamiltonian can be written in terms of the
fields $\phi_\sigma$ and their canonically conjugate momentum
$\Pi_\sigma = \dx \theta_\sigma/\pi$. In the case of absence of
inter-species interactions the effective bosonic theory is given by
$\mathcal{H}_0(\phi_\up) +\mathcal{H}_0(\phi_\down)$, where
\begin{equation}
\mathcal{H}_0(\phi_\sigma) = \dfrac{v_\sigma}{2\pi} \int\! dx\,\left[
K_\sigma(\pi\Pi_\sigma)^2 +K_\sigma^{-1} \left(\dx\phi_\sigma\right)^2 \right]\; ,
\label{H0}
\end{equation}
where $v_\sigma$ is the sound velocity and $K_\sigma$ the so-called
Luttinger parameter, which is equal to one in the free fermions or
free hard-core bosons cases. Taking into account density-density
interactions between species, of the generic form $\int\! dx dx'\,
U(x-x') \hat{n}_\up(x) \hat{n}_{\down}(x')$, changes the effective
theory and brings new kinds of terms : zero-momentum terms in the
density representation \eqref{HaldaneN} couples the two spin species
through a bilinear operator
\begin{equation}
\label{H1}
\mathcal{H}_1 = \frac{g}{2\pi} \int\! dx \, (\dx \phi_\up)(\dx \phi_\down) \;,
\end{equation}
where $g$ is a forward scattering constant, and higher harmonics terms
involving multiples of the spatial frequencies $k_\sigma$ :
\begin{align}
\label{H2}
\sum_{p,q>0}G_{pq}^{\pm}\!\int\!dx\,\cos{\big[2\pi(pn_\up \pm qn_\down )x - 2(p\phi_\up \pm q\phi_\down)\big]}
\end{align}
where $G_{pq}^{\pm}$ are non-universal coupling constants. Clearly,
if the generalized commensurability condition
\begin{equation}
\label{pq}
p\,n_\up - q n_\down =0 \;,
\end{equation}
is satisfied (with $p$ and $q$ coprime integers), and provided these
terms are relevant, they will tend to lock the up and down fields
together. When the densities are fine-tuned to the definite
commensurability \eqref{pq}, then all other cosine operators in the
sum are oscillating, in which case they don't contribute in the
continuum limit (or they are less relevant for multiples of $p$ and
$q$). The remaining important operator in the sum \eqref{H2} is thus
the sine-Gordon term
\begin{equation}
\label{H2pq}
\mathcal{H}_2 = G \int\! dx \, \cos{\sqrt{8}\phi_a} \; ,
\end{equation}
with the combination
\begin{equation}
\label{eq:phia}
\phi_a=\frac{1}{\sqrt{2}}(p\phi_{\up}-q\phi_{\down})\;.
\end{equation}
For attractive interactions $G<0$ (we will argue below that this
choice favors the relevance of the term), energy will be minimized
when the field is pinned to $\moy{\phi_a} = 0$. Notice again that the above
argument on the mode-locking mechanism does not rely on the presence
of a lattice. Lastly, the cosine locks a combination of the bosonic
modes but at a generic total density $n$, there remains another bosonic
mode leaving the full excitation spectrum gapless. We will see that the
latter describes the effective behavior of the bound-states. In the
following, we dub $\phi_b$ this massless bosonic mode.
We can draw a last remark on the operators \eqref{H2} : they
have high scaling dimensions near the free fermion fixed point
and are expected to be irrelevant apart from some special circumstances
which are the object of this work. In the fermionic language, they involve
$(p+q)$-body interactions of the form:
\begin{equation}
\sum_{\{k\}} \prod_{i=1}^q \psi^\dagger_{R\down}(k_i) \psi_{L\down}(k_i') \,
\prod_{j=1}^{p} \psi^\dagger_{L\up}(k_j'') \psi_{R\up}(k_j''') + \mathrm{h.c.} \; ,
\label{H2pqFERM}
\end{equation}
where the summation over $\{k\}$ runs over all combinations of
$2(p+q)-1$ momenta due to the total momentum conservation law~:
$\sum_{i=1}^q ( k_i - k_i') + \sum_{j=1}^p (k_j'' - k_j''') =0$.
Such interactions appear at high order in perturbation theory in a Hubbard model
for example or after several steps of a RG treatment.
For practical purposes it is simpler to work with the bosonic formulation
given by Eq.~\eqref{H2pq}, and this is what we do from now on.
In the following, we assume that the densities are commensurate via the condition
\eqref{pq}, and analyze the simplified effective theory written in terms of the $\up$ and $\down$ fields
\begin{equation}
\mathcal{H}= \mathcal{H}_0(\phi_\up) + \mathcal{H}_0(\phi_\down) + \mathcal{H}_1 + \mathcal{H}_2 \; ,
\label{fullmodel}
\end{equation}
where the velocities $v_\sigma$ and Luttinger parameters $K_\sigma$
are determined by the intra-species interactions.
The quadratic part $\mathcal{H}_0(\phi_\up) + \mathcal{H}_0(\phi_\down) + \mathcal{H}_1$
can be diagonalized by a Bogoliubov transformation~\cite{Engelsberg1964}
which could give a starting point for a perturbative RG calculation~\cite{Penc1990,
Cazalilla2003, Cazalilla2005, Mathey2007, Crepin2010}. Due to the
velocity asymmetry, additional couplings are generated and velocities
are renormalized along the flow. The discussion of the nature of the
gapped phases and their correlations remains unclear.
In particular, diagonalizing the
quadratic part $\mathcal{H}_0(\phi_\up) + \mathcal{H}_0(\phi_\down) +
\mathcal{H}_1$ of the Hamiltonian \eqref{fullmodel} does not give,
apart from special choice of the parameters,
the combination \eqref{eq:phia} that appears in the
cosine term $\mathcal{H}_2$. In the next section, we take the following
strategy : we look for the conditions under which the
quadratic part and the cosine term are simultaneously
diagonalizable. At the price of a restriction on the parameters range,
the analysis can be done safely both for the criteria of
relevance of the cosine and for the correlation functions in the single-mode
phase. In spite of the limitation of the approach, we believe the scenario
does occur without this restriction :
as shown numerically in Sec.~\ref{sec:trimers} on a realistic model,
the single-mode multimer phase can span a wide region of the phase diagram.
Lastly, we notice that, similarly to the phase separation
criteria in two-component mixtures (when one of the mode velocities vanishes),
the single-mode phase will undergo a phase separation instability
when the gapless mode velocity $v_b$ vanishes. We thus expect to find
the single-mode phase surrounded with the two-mode phase and a
demixed phase.
\subsection{Field transformation}
\label{sec:fieldtransfo}
We have qualitatively discussed the fact that the physics should
generically be described by two fields
$\phi_s$ where $s=a,b$, with $\phi_a =
(p\phi_{\up}-q\phi_{\down}) / \sqrt{2}$ being the one entering in the
cosine term \eqref{H2pq}. In general, it is hard to have a complete
form for the transformation between the $\phi_s$ and the
$\phi_{\sigma}$. Such a transformation is important
both for the RG analysis and the calculation of physical correlators
which are naturally expressed in terms of the $\phi_{\sigma},
\theta_{\sigma}$ fields. Below, we discuss a special case where the
transformation can be performed and its range of validity.
The simplest transformation, and yet rather general, one can work with
is a linear combination of the fields with coefficients that are
independent of the position~:
\begin{align}
\phi_{\up} &= \mathfrak{p}_{a\up} \phi_a + \mathfrak{p}_{b\up} \phi_b\,,
& \theta_{\up} &= \mathfrak{t}_{a\up} \theta_a + \mathfrak{t}_{b\up} \theta_b\,, \\
\phi_{\down} &= \mathfrak{p}_{a\down} \phi_a + \mathfrak{p}_{b\down} \phi_b\,,
& \theta_{\down} &= \mathfrak{t}_{a\down} \theta_a + \mathfrak{t}_{b\down} \theta_b\,.
\end{align}
When $p \neq q$, excitations corresponding to the eigenmodes
$\phi_{a,b}$ carry both spin and charge modes which are respectively
the sum and the difference of the $\up$ and $\down$ modes. As the
transformation must preserve the commutation relations
\begin{align}
[ \theta_{\sigma}(x), \nabla \phi_{\sigma'}(x') ] &= i\pi\delta_{\sigma\sigma'} \delta(x-x') \,,\\
[ \phi_{\sigma}(x), \nabla \theta_{\sigma'}(x') ] &= i\pi\delta_{\sigma\sigma'} \delta(x-x') \,,
\end{align}
we get that $\mathfrak{p}_{a\sigma} \mathfrak{t}_{a\sigma'} +
\mathfrak{p}_{b\sigma} \mathfrak{t}_{b\sigma'} =
\delta_{\sigma\sigma'}$. Then, we have
\begin{align}
\mathfrak{t}_{a\up} &= \mathfrak{p}_{b\down}/D\,, & \mathfrak{t}_{a\down} &= -\mathfrak{p}_{b\up}/D\,,\\
\mathfrak{t}_{b\up} &=-\mathfrak{p}_{a\down}/D\,, & \mathfrak{t}_{b\down} &= \mathfrak{p}_{a\up}/D\,,
\end{align}
with the determinant
\begin{equation}
D = \mathfrak{p}_{a\up} \mathfrak{p}_{b\down} -
\mathfrak{p}_{a\down} \mathfrak{p}_{b\up} =
(\mathfrak{t}_{a\up} \mathfrak{t}_{b\down} -\mathfrak{t}_{a\down} \mathfrak{t}_{b\up})^{-1}\,.
\end{equation}
In a shortened version, we have $\phi_{\sigma} = \mathcal{P} \phi_s$,
where $\mathcal{P}$ is the matrix of the $\mathfrak{p}$, and
$\theta_{\sigma} = (\mathcal{P}^{-1})^{t} \theta_s$. If $\mathcal{P}$
is unitary, the $\theta$ and the $\phi$ undergo the same
transformation.
We now impose that $\phi_a = (p\phi_{\up}-q\phi_{\down}) / \sqrt{2}$
which gives
\begin{align}
\label{eq:tas}
\mathfrak{t}_{a\up} &= \frac{p}{\sqrt{2}}\;, & \mathfrak{t}_{a\down} &= -\frac{q}{\sqrt{2}}\,.
\end{align}
As we want to cancel the cross-terms in Eq.~\eqref{fullmodel}, we
require that~:
\begin{align}
v_{\up} K_{\up} \mathfrak{t}_{a\up} \mathfrak{t}_{b\up} &= -v_{\down}
K_{\down} \mathfrak{t}_{a\down} \mathfrak{t}_{b\down} \,,\\
\frac{v_{\up}}{K_{\up}} \mathfrak{p}_{a\up} \mathfrak{p}_{b\up} +
\frac{v_{\down}}{K_{\down}} \mathfrak{p}_{a\down}
\mathfrak{p}_{b\down} &= -g(\mathfrak{p}_{a\down} \mathfrak{p}_{b\up}
+ \mathfrak{p}_{a\up} \mathfrak{p}_{b\down})\,,
\end{align}
which can be rewritten as~:
\begin{eqnarray}
p v_{\up} K_{\up} \mathfrak{t}_{b\up}\quad -& q v_{\down} K_{\down} &\mathfrak{t}_{b\down} =0 \\
-\Big(p\frac{v_{\down}}{K_{\down}}+gq\Big)\mathfrak{t}_{b\up}\quad + &\displaystyle
\Big(q \frac{v_\up}{K_\up}+gp\Big)&\mathfrak{t}_{b\down} = 0\;.
\end{eqnarray}
There exists a non-zero solution only if the condition~:
\begin{equation}
\label{eq:condition}
v_{\up}\Big(v_{\up}+\frac{gp}{q}K_\up\Big) = v_{\down}\Big(v_{\down}+\frac{gq}{p}K_\down\Big)
\end{equation}
is satisfied. When this condition is satisfied, we have a
one-parameter family of transformations with the desirable property of
having only one eigenmode in the argument of the cosine operator. The
parameter is just the choice of scale of the field $\phi_b$~: in 1D, we
can change the scale of the Bose field provided we change accordingly
its Luttinger parameter $K_b$. Here, we choose the scale of $\phi_b$ so
that~:
\begin{align}
\mathfrak{t}_{b\up} &= \frac{q}{\sqrt{2}}\sqrt{\frac{v_{\down}}{v_\up}} K_{\down}\;, &
\mathfrak{t}_{b\down} &= \frac{p}{\sqrt{2}}\sqrt{\frac{v_{\up}}{v_\down}}K_{\up}\,,
\end{align}
The condition \eqref{eq:condition} strongly reduces the range of
applicability of the transformation : for a given coupling $g$, the
Luttinger parameters and velocities of each species must satisfy the
above relation.
When the transformation can be used, \eqref{fullmodel} splits into a
free boson field for $b$ and a sine-Gordon model for $a$ :
$\mathcal{H} = \mathcal{H}_0(\phi_b) +
\mathcal{H}_{\text{sG}}(\phi_a)$ with $\mathcal{H}_{\text{sG}} =
\mathcal{H}_0 + \mathcal{H}_2$. In this case, the new velocities and
Luttinger parameters associated with the $a,b$ modes are given by the
following relations~:
\begin{align}
\nonumber
v_a^2 &= \frac{\sqrt{v_{\up}v_{\down}}}{2\mathcal{K}} \Big(p^2 v_\up
K_{\up}\frac{v_\up}{v_\down} + q^2 v_\down K_{\down}
\frac{v_\down}{v_\up} - gpq K_{\up}K_{\down}\Big) \,,\\
\label{eq:Ka}
K_a &= \mathcal{K}\frac{\sqrt{v_{\up}v_{\down}}}{v_a} \,, \\
\nonumber
v_b^2 &= \frac{\sqrt{v_{\up}v_{\down}}}{2\mathcal{K}}\left(p^2 v_{\down}
K_{\up} + q^2 v_{\up} K_{\down} + gpq K_{\up}K_{\down}\right)\,, \\
\nonumber
K_b &= K_{\up}K_{\down}\mathcal{K}\frac{\sqrt{v_{\up}v_{\down}}}{v_b}\,.
\end{align}
where we have defined~:
\begin{equation}
\mathcal{K} = \frac{p^2 v_{\up} K_{\up} + q^2 v_{\down} K_{\down}}{2\sqrt{v_{\up}v_{\down}}}\,.
\end{equation}
With our definition of $\phi_a$ and provided the sine-Gordon
description is applicable, the requirement for the cosine to be
relevant, and thus to enter the single-mode phase, is simply
\begin{equation}
K_a < 1\;.
\end{equation}
One qualitatively observes that a velocity much smaller than the
other favors a small $K_a$ and that
large attractive interactions $g<0$ will help increase $v_a$ and
reduce $K_a$. In the following, we consider limiting cases in which
the discussion simplifies in order to identify how the parameters
would favor the formation of a gap in the $a$ sector.
The limit $v_b^2=0$ (attained with attractive interactions) signals the
transition to the phase-separated or Falicov-Kimball regime from the
multimer phase.
\subsubsection{The case of equal velocities}
When $v_\up=v_\down=v_0$, the condition \eqref{eq:condition} imposes
that either (i) $g=0$ or (ii) $\displaystyle \frac{K_\up}{K_\down} =
\frac{q^2}{p^2}$. The transformation and new velocities and Luttinger
parameters then take a simple form : in case (i), we have
$v_a=v_b=v_0$ and~:
\begin{align}
\nonumber
\phi_a &=\frac{1}{\sqrt{2}}(p\phi_\up - q\phi_\down)\,, &\phi_b &=\frac{1}{\sqrt{2}}\left(qK_\down
\phi_\up + pK_\up\phi_\down\right)\,, \\
\label{eq:eqv(i)}
K_a &= \frac{p^2 K_\up +q^2 K_\down}{2}\,, & K_b &= K_\up K_\down \frac{p^2 K_\up +q^2 K_\down}{2}\,.
\end{align}
while in case (ii), we have $\mathcal{K} = p^2K_\up$ and
\begin{align}
\nonumber
\phi_a &=\frac{1}{\sqrt{2}}(p\phi_\up - q\phi_\down)\,, & \phi_b &=\frac{pK_\up}{q\sqrt{2}}
\left(p \phi_\up + q\phi_\down\right)\,, \\
\nonumber
v_a^2 &=v_0^2\Big(1-\frac{gp}{2qv_0}K_\up\Big)\,, & v_b^2 &= v_0^2\Big(1+\frac{gp}{2qv_0}K_\up\Big)\,,\\
\nonumber
K_a &= p^2 K_\up & K_b &= \frac{p^4K_\up^3 }{q^2\sqrt{1+\frac{gp}{2qv_0}K_\up}}\,.
\end{align}
In both cases, having $K_a<1$ would require a very small $K_{\up}$
(assuming $q=1$ for example). This could be realized with long-range
intra-species interactions but may not be easily achievable. Notice a
peculiarity of the formula for the massive mode $\phi_b$ in
\eqref{eq:eqv(i)}: while $K_\up$ and $K_\down$ are length-scale
dependent (in the RG sense), the expression in \eqref{eq:eqv(i)} holds
on all length-scales.
\subsubsection{The limit of large asymmetry}
In order to identify the influence of the velocities ratio on $K_a$, one can introduce the dimensionless
quantities $\nu = v_\down/v_\up$, $\rho = q/p$ and $\gamma =
g/v_\up$. Then, \eqref{eq:condition} and \eqref{eq:Ka} are rewritten
as
\begin{align}
\label{eq:condition-simple}
1+\rho \gamma K_\up &= \nu(\nu+\gamma \rho^{-1} K_\down) \,, \\
\label{eq:Ka-simple}
K_a &= \nu \frac{K_\up + \rho\nu K_\down}{K_\up - \rho\gamma K_\up K_\down \nu + \rho^2K_\down \nu^3 }\,.
\end{align}
If one takes into account \eqref{eq:Ka-simple} only, $K_a$ vanishes in
the limit of large velocity ratio $\nu \rightarrow 0$ or $\infty$
and passes through a maximum in between so that there are two windows
of $\nu$ such that $K_a<1$. The smaller the maximum, the wider these
windows are so, clearly, negative and large interactions ($\gamma <
0$) favor the mode-locking mechanism. Yet, \eqref{eq:condition-simple}
imposes another constraint and we just consider the $\nu \rightarrow
0$ limit for simplicity. There, this limit is possible provided $K_\up
\simeq -1/\gamma \rho$, i.e. in the case of attractive interaction
only. As a consequence, this analysis shows that we should expect the
formation of multimer in the attractive and large interaction regime,
favored by large asymmetry.
\subsubsection{In the single-mode Luttinger liquid phase}
Deep in the massive-$a$ phase, one can make a crude quadratic
approximation to the cosine operator in \eqref{H2pq} by replacing it
with a mass term $\propto (p\phi_\up-q\phi_\down)^2$. This leads to
approximate expressions for the velocity and Luttinger parameter of
the remaining mode~$b$\ :
\begin{align*}
v_b^2 & = v_\up v_\down \frac{p^2 K_\up v_\down + q^2 K_\down v_\up}{p^2
K_\up v_\up + q^2 K_\down v_\down}\; , \\
K_b & = \frac{1}{2}K_\up K_\down \frac{(p^2 K_\up + q^2 K_\down)^2} {\sqrt{p^2
\frac{K_\up}{v_\up}+q^2 \frac{K_\down}{v_\down}} \sqrt{p^2
{K_\up}{v_\down}+q^2 {K_\down}{v_\up}}} \;,
\end{align*}
which reduces to the correct result for equal velocities.
\subsection{Correlation functions and the nature of the phases}
\label{subsec:corr}
In one-dimensional models, the classification of the groundstates is
determined by their dominant correlations. One can break discrete
symmetries (for instance translational symmetry on a lattice model)
but order parameters associated with continuous symmetries are always
zero. The naming of a phase then corresponds to the connected
equal-time correlator with the slowest decay in space. Quite
generally, these correlators are asymptotically decaying either
algebraically or exponentially. Such algebraic correlations are
usually referred to as a quasi long-range order (QLRO). The slowest
decay (or smallest exponent of algebraic correlations) criteria is
based on a RPA argument by considering a set of weakly coupled
Luttinger liquids~\cite{Giamarchi2004} which shows that order will
build up provided the exponent of the correlator is smaller than two,
and that the main instability is associated with the smallest
exponent. However, if the Green's function, associated with
$\Psi_\sigma$ which is not an order parameter, has the slowest
decaying exponent, a RG analysis shows that
coupling the Luttinger liquids yield a Fermi liquid phase (provided
that the decay exponent is smaller than two again). If all physical
correlators are exponentially decaying (apart from the density one
which always keep, at least, a quadratic decay), the term liquid is often
used. This approach yet remains phenomenological as the
higher-dimension situation and is much more involved.
In this section, we follow the standard practice and consider the
correlation functions of various observables to discuss the nature of
the phases that are realized in the single-mode and two-mode
regimes. The asymptotic decay of the connected correlation functions
associated with the order parameter $\mathcal{O}(x)$ typically reads
$\langle \mathcal{O}(0) \mathcal{O}^\dagger(x)\rangle_c \propto
x^{-\alpha_\mathcal{O}}$ with some exponent $\alpha_\mathcal{O}$. In
order to compute the correlators, we only keep the first harmonics in
Eq.~\eqref{eq:fermions} and begin with the richer case of fermions where
we use the representation in terms of right and left movers:
\begin{equation}
\Psi_{\sigma}(x) \sim e^{i k_{\sigma}x} e^{i(\theta_{\sigma}-\phi_{\sigma})}
+ e^{-ik_{\sigma}x} e^{i(\theta_{\sigma}+\phi_{\sigma})} \,.
\end{equation}
We use the results that when a field $\phi_a$ is pinned,
$\moy{f(\phi_a)} = f(\moy{\phi_a})$ and its dual $\theta_a$ is
disordered, leading to an exponential decay. In the case of algebraic
correlations, the decay exponents are obtained using the result that,
for a field $\phi$ described by $\mathcal{H}_0$, the equal-time
correlator associated with $\displaystyle A_{m,n}(x) =
e^{i[m\phi(x)+n\theta(x)]}$ behaves asymptotically as
\begin{equation}
\moy{A_{m,n}(x)A_{-m,-n}(0)} \propto x^{-(m^2 K + n^2/K)/2} \;.
\end{equation}
We now give the leading contributions of the order parameters as a
function of the $a,b$ fields, assuming general
transformation coefficients of the $\up$ and $\down$ modes~:
\begin{widetext}
\begin{align}
\label{eq:green}
\Psi_\sigma(x) & \sim e^{i k_{\sigma}x} e^{-i [\mathfrak{p}_{a\sigma}\phi_a +
\mathfrak{p}_{b\sigma}\phi_b - \mathfrak{t}_{a\sigma}\theta_{a} -
\mathfrak{t}_{b\sigma}\theta_{b}]} & \text{Green's function}\\
\label{eq:density}
\hat{n}_{\sigma}(x) & \sim -\mathfrak{p}_{a\sigma}\nabla\phi_a -
\mathfrak{p}_{b\sigma}\nabla\phi_b + \Lambda^{-1}\cos(2k_{\sigma}x-2(\mathfrak{p}_{a\sigma}\phi_a
+ \mathfrak{p}_{b\sigma}\phi_b)) & \text{density}\\
\label{eq:singlet}
\Psi_\up(x) \Psi_\down(x) & \sim e^{i(k_{\up}-k_{\down})x} e^{-i[(\mathfrak{p}_{a\down}-
\mathfrak{p}_{a\up})\phi_a + (\mathfrak{p}_{b\down}-\mathfrak{p}_{b\up})\phi_b+
(\mathfrak{t}_{a\down}+\mathfrak{t}_{a\up})\theta_{a}+(\mathfrak{t}_{b\down}+
\mathfrak{t}_{b\up})\theta_{b}]} & \text{singlet pairing}\\
\label{eq:triplet}
\Psi_\sigma(x) \Psi_\sigma(x) & \sim e^{2i[\mathfrak{t}_{a\sigma}\theta_{a}+
\mathfrak{t}_{b\sigma}\theta_{b}]} &\text{triplet pairing}
\end{align}
\end{widetext}
where $\Lambda$ is a short-range cutoff. Among the multiple
combinations of right and left movers, we have chosen the ones which
should lead to the lowest decay exponents, by having the lowest $m$
and $n$ constant. They usually correspond to the smallest wave-vector.
\subsubsection{The two-mode Luttinger liquid (2M-LL) phase}
In the two mode-regime, all correlators are algebraic
and the leading one will strongly depend on the actual coefficients of
the transformation. The expression of
Eqs.~\eqref{eq:green}--\eqref{eq:triplet} are here understood with general
transformation coefficients of the $\up$ and $\down$ modes as one does
not necessarily have to impose the restriction \eqref{eq:tas} since
$\phi_a$ does not here identify with \eqref{eq:phia}. The transformation
coefficients can be computed exactly~\cite{Mathey2004, Mathey2007a}
in the absence of the cosine term \eqref{H2pq}. The correlation functions can as well be computed
directly using a Green's function approach~\cite{Orignac2010}.
In the presence of \eqref{H2pq}, the coefficients will be renormalized
in this two-mode phase to unknown values. This regime is rather generic
and, depending on the interaction and densities,
with many competing orders among which are a Fermi liquid-like phase, a superconducting singlet
or triplet FFLO phase~\cite{Fulde1964} (pairing correlations
displaying the typical $k_{\down}-k_{\up}$), a
spin-density wave (SDW) or charge-density wave (CDW) phase. The case
of equal densities, $p = q = 1$ has the dominant
channels~\cite{Giamarchi2004} among the superconducting, CDW, and SDW
fluctuations. In the cases where spin and charge degrees of freedom
separate, CDW and SDW states are mutually exclusive. Furthermore, for
SU(2)-symmetric models, $x$-, $y$- and $z$-components of the SDW order
parameter are degenerate. These last remarks are no longer valid in
our situation.
\subsubsection{The single-mode Luttinger (1M-LL) multimer phase}
Another regime corresponds to the case where the cosine in Eq.~\eqref{H2pq}
is relevant in the RG sense. Then, the system has a
massive mode $\phi_a$ given by \eqref{eq:phia}, and a massless mode $\phi_b$.
The massless mode is described in the
low-energy limit by a free bosonic with a velocity $v_{b}$ and a
Luttinger parameter $K_{b}$. In this single-mode
Luttinger liquid, algebraic decays will be ruled by this $K_b$
Luttinger parameter when they occur. When the parameters of the problem satisfy
Eq.~\eqref{eq:condition}, then the massless mode can be found
explicitly. In this section we use these results to discuss
in details the behavior of the correlation functions.
When $\phi_a$ gets pinned, we see that the above
correlators \eqref{eq:green}--\eqref{eq:triplet} are all exponential
because the presence of $\theta_a$ in their expression,
with the exception of the density one. In particular, all two-body pairing
channels are suppressed, even in the presence of attractive
interactions. In order to construct an operator which has algebraic
correlations, the prefactor in front of $\theta_a$ must vanish. This
is realized by taking the $(p+q)$-mer combination $\Psi_\up^q(x)
\Psi_\down^p(x)$ (bound states of $p$ $\down$-fermions with $q$
$\up$-fermions) which has the prefactor $q\mathfrak{t}_{a\up} +
p\mathfrak{t}_{a\down}$ which is clearly zero from \eqref{eq:tas}:
\begin{widetext}
\begin{align}
\label{eq:pqmer}
\Psi_\up^q(x) \Psi_\down^p(x) & \sim e^{iQ_{qp}x} e^{i[(q\mathfrak{t}_{a\up}
+ p\mathfrak{t}_{a\down})\theta_{a}+(q\mathfrak{t}_{b\up} + p\mathfrak{t}_{b\down})\theta_{b}
- (s_q\mathfrak{p}_{a\up}-s_p\mathfrak{p}_{a\down})\phi_a - (s_q\mathfrak{p}_{b\up}-
s_p\mathfrak{p}_{b\down})\phi_b]} \;,\\
\intertext{and, in the special case of trimers,}
\label{eq:trimer}
\Psi_\up(x)\Psi_\down(x)\Psi_\down(x) & \sim e^{ik_{\up}x} e^{i[(\mathfrak{t}_{a\up}+
2\mathfrak{t}_{a\down})\theta_{a} + (\mathfrak{t}_{b\up}+2\mathfrak{t}_{b\down})\theta_{b}
- \mathfrak{p}_{a\up}\phi_a - \mathfrak{p}_{b\up}\phi_b]}\;,
\end{align}
\end{widetext}
in which $Q_{qp} = s_qk_\up - s_pk_\down$ and $s_p = p,
p-2,\ldots,(\text{0 or 1})$, $s_q = q,q-2,\ldots,(\text{0 or 1})$ are
integers accounting for the combination of left and right movers. We
have used a somewhat symbolic notation: by $\Psi^p(x)$, we mean
$\Psi(x+\delta_1)\Psi_(x+\delta_2)\cdots \Psi(x+\delta_p)$, where
$|\delta_i| < \Lambda$ where $\Lambda$ is the short-range cutoff. We stress that
the family of operators \eqref{eq:pqmer} is different from the
``polaronic'' operators introduced in Ref.~\onlinecite{Mathey2004}:
the latter are constructed specifically for minimizing the decay
exponents in the massless phase of \eqref{fullmodel}. On the contrary,
the family \eqref{eq:pqmer} arises naturally in the massive phase of
Eq.\ \eqref{fullmodel}, as a many-body consequence of a existence of
$(p+q)$-body bound states in the microscopic counterpart of
\eqref{fullmodel}.
The effective theory of this $(p+q)$-mer object is then governed by
the gapless mode $b$. Remarkably, as $q\mathfrak{t}_{b\up} +
p\mathfrak{t}_{b\down} = q/\mathfrak{p}_{b\up} = \sqrt{2}\mathcal{K}$,
the exponent is parametrized only by $K_b$, $\mathcal{K}$ and
$p,q$. In order to have the smallest exponent, we have to select the
combination $(s_q,s_p)$ which minimizes the coefficient in front of
$\phi_b$ (one cannot have the combination $p\mathfrak{p}_{b\up} -
q\mathfrak{p}_{b\down}=0$) and which is proportional to $C_{qp} = s_pp
- s_qq$. We list below the coefficients and corresponding wave-vectors
for the simplest commensurabilities:\\
\begin{center}
\begin{tabular}{|c|c|c|c|}
\hline\hline
$(q,p)$ & $(s_q,s_p)$ & $C_{qp}$ & $Q_{qp}$ \\
\hline\hline
$(1,2)$ & $(1,0)$ & 1 & $k_\up$ \\
$(1,3)$ & $(1,1)$ & 2 & $k_\up-k_\down$ \\
$(1,4)$ & $(1,0)$ & 1 & $k_\up$ \\
$(1,5)$ & $(1,1)$ & 4 & $k_\up-k_\down$ \\
$(3,2)$ & $(1,1)$ & 1 & $k_\up$ \\
$(3,4)$ & $(1,1)$ & 1 & $k_\up$ \\
$(3,5)$ & $(1,1)$ & 2 & $k_\up-k_\down$ \\
$(3,7)$ & $(3,1)$ & 2 & $3k_\up-k_\down$ \\
$(5,7)$ & $(1,1)$ & 2 & $k_\up-k_\down$ \\
\hline\hline
\end{tabular}\\
\end{center}
The exponent of the propagator of the $(p+q)$-mer then reads
$\displaystyle \frac 1 2 \left(K_{\text{eff}}^{-1} + C_{qp}^2
K_{\text{eff}} \right)$ with the effective Luttinger parameter
\begin{equation}
\label{eq:Keff}
K_{\text{eff}} = \frac{K_b}{2\mathcal{K}^2}\;.
\end{equation}
In this phase, the connected density correlations $\mathcal{N}_{\sigma
\sigma'}(x)=\langle n_\sigma(0) n_{\sigma'}(x) \rangle - \langle
n_\sigma(0) \rangle \langle n_{\sigma'}(x) \rangle$ remain algebraic
with the following dominant contributions~:
\begin{gather}
\mathcal{N}_{\up\up}(x) = -\frac{K_\text{eff}}{2 \pi^2}\frac{q^2}{x^2} +
A_{\up\up}\frac{\cos(2k_\up x)}{x^{2q^2K_\text{eff}}} \;, \label{eq:nupnup} \\
\mathcal{N}_{\down\down}(x) = -\frac{K_\text{eff}}{2 \pi^2}\frac{p^2}{x^2}
+ A_{\down\down}\frac{\cos(2k_\down x)}{x^{2p^2K_\text{eff}}}\;,\label{eq:dndn}
\end{gather}
where $A_{\sigma\sigma'}$ are non-universal amplitudes. The main
remarks are that (i) the ratio of the zero-momentum fluctuations is
exactly $(q/p)^2$ while the ratio of the density is $q/p$ and (ii) the
wave-vectors are different since $k_\up = \pi n\frac{q}{p+q}$ and
$k_\down = \pi n\frac{p}{p+q}$ as well as their exponents which ratio
should be $(q/p)^2$ exactly. Notice that for the sine-Gordon model,
the ratio of the amplitudes $A_{\up\up} / A_{\down\down}$ are
\emph{exponentially} small in $p-q$ \cite{Lukyanov1997}.
When $C_{qp}=1$, we see that the multimer is effectively behaving as
a spinless fermion (as expected from the combination of a total odd
number of fermions) which Fermi level is $k_{\up}$ and Luttinger
exponent $K_{\text{eff}}$. For instance, trimers belong to this
ensemble. The effective interaction between these spinless fermions,
which are spatially extended objects, is highly non-trivial and
certainly depends on the distance, density and microscopic
parameters (a discussion of such interactions in the case of a boson
mixture can be found in Ref.~\onlinecite{Soyler2009}). However, its
overall effect can be captured by $K_{\text{eff}}$ with effective
repulsion expected when $K_{\text{eff}}<1$ (dominant
CDW fluctuations), and effective attraction expected if
$K_{\text{eff}}>1$ (dominant trimer-pairing
fluctuations). The latter turns out to be a superfluid phase of trimers.
By associating an even total number of fermions, one should
effectively expect build a bosonic-like multimer. Yet, we see that, in
the propagator of the multimer, one cannot suppress the contribution
from the $\phi_b$ field (as $C_{qp} \neq 0$) and the exponent is not
simply $1/2K_{\text{eff}}$ and thus not simply related to the one of
the density correlations as one would get for a simple bosonic propagator.
Furthermore, while the momentum distribution
of a boson would usually have a peak at zero-momentum, we see that this
observable will be here diverging at $Q_{qp} \neq 0$.
\subsubsection{The case of a bosonic mixture in the single-mode phase}
As previously mentioned, the effective theory under study can be as
well applied to the situation where the particles are bosons. In the
single-mode phase, a bosonic multimer phase will emerge under the
mode-coupling mechanism and the motivation of this small section is to
discuss the form of the corresponding correlators. We assume repulsive
interactions for the intra-species channels (for stability reasons and
also to lower the $K_{\sigma}$ to be able to fulfill the $K_a<1$
requirement) but attractive interactions in the inter-species channel
(as for the fermions). The boson creator operators are bosonized as
$b_{\sigma} \sim e^{i\theta_\sigma}$ (dropping the higher harmonics
term of Eq.~\eqref{eq:bosons}) which immediately yields
\begin{equation*}
b_\up^q(x) b_\down^p(x) \sim e^{i[(q\mathfrak{t}_{a\up} + p\mathfrak{t}_{a\down})\theta_{a}+(q\mathfrak{t}_{b\up} + p\mathfrak{t}_{b\down})\theta_{b}]}\;.
\end{equation*}
The $(p+q)$-mer is then a true bosonic molecule with an effective
Luttinger parameter which is exactly given by \eqref{eq:Keff}. The
density correlations do not depend on the statistics and still have
the form of \eqref{eq:nupnup}--\eqref{eq:dndn}.
\subsection{Lattice commensurability effects}
\label{sec:lattice}
So far, we have only considered two-component fluids in the continuum
limit which is expected at generic densities on a lattice or in
continuum space. In this section, we briefly discuss the additional
effects arising from the presence of a lattice~\footnote{We assume that
the field theory description is appropriate --- for too strong
interactions and/or too large asymmetry it breaks down and the
lattice model falls into the Falicov-Kimball universality class, see
Sec~\ref{sec:trimers}.}. An underlying lattice with period $a_0$ can
be viewed as a periodic external potential, in which particles have a
momentum being only defined modulo the reciprocal lattice vector
$2\pi/a_0$. Therefore, umklapp processes with momentum transfer of a
multiple of $2\pi/a_0$ are allowed at low energy. If a Fermi momentum $k_\sigma$
of a species $\sigma$, is itself a multiple of $2\pi/a_0$,
i.e. if a density of species $\sigma$ is commensurate with the
lattice, $s\, n_\sigma = \text{integer}/a_0$, with an integer $s$, an
additional term $\cos(2s\phi_\sigma)$ appears in the low-energy
Hamiltonian. The effects stemming from such a cosine operator
\textit{alone} are well known~: for $K_\sigma \leqslant 2/s^2$ the
cosine is relevant in the RG sense and the system undergoes a Mott
transition into a density wave state with the unit cell of $s$ lattice
sites. In a two-component system, it is possible to have two operators
of this sort, one for each species. Furthermore, if the densities are
such that $s n_\up + s' n_\down$ is an integer (we set $a_0=1$ from
now on) for some integers $s$ and $s'$, there is yet another term in
the low-energy Hamiltonian, namely $\cos{ 2(s\phi_\up + s'\phi_\down)
}$ (cf. Eq.~\eqref{H2}).
Here, we analyze a simple special case where
\begin{align}
p\,n_\up - q\, n_\down &= 0 \;, \label{latt:pqrl1}\\
r\,n_\up + l\, n_\down &= 1 \; ,\label{latt:pqrl2}
\end{align}
or $n_\up = q/(pl+qr)$ and $n_\down = p/(pl+qr)$ with the integers
$p$, $q$, $r$ and $l$. Given \eqref{latt:pqrl1}--\eqref{latt:pqrl2},
Eq.~\eqref{HaldaneN} yields the Hamiltonian in the form
$\mathcal{H}_0(\phi_\up) +\mathcal{H}_0(\phi_\down) + \mathcal{H}_1+
\mathcal{H}_\mathrm{cos}$ with
\begin{eqnarray}
\mathcal{H}_\mathrm{cos} =&& G_1 \int\! dx \, \cos{ 2(p\,\phi_\up - q\,\phi_\down) } \label{latt:H1} \\
&+& G_2 \int\! dx \, \cos{ 2(r\,\phi_\up + l\,\phi_\down) } \label{latt:H2} \\
&+& G_3 \int\! dx \, \cos{ 2(pl+qr)\,\phi_\down } \label{latt:H3} \\
&+& G_4 \int\! dx \, \cos{ 2(pl+qr)\,\phi_\up } \label{latt:H4} \; .
\end{eqnarray}
where $G_1$,\dots,$G_4$ are non-universal amplitudes. Interpretation
of Eqs.\ \eqref{latt:H1}--\eqref{latt:H4} is straightforward: Eq.\
\eqref{latt:H1} stems from the condition \eqref{latt:pqrl1} and is
thus insentive to the presence of the lattice (cf. Sec.\
\ref{sec:modecoupling}); Eqs.\ \eqref{latt:H3} and \eqref{latt:H4}
favor the Mott localization of the species $\down$ and $\up$,
respectively. On the other hand, operator \eqref{latt:H2} is unique to
two-component lattice systems and owes its existence to the peculiar
commensurability condition \eqref{latt:pqrl2}. The physical meaning of
\eqref{latt:pqrl2} is clear: by analogy with Sec.\ \ref{subsec:corr},
it favors the quasi long-range ordering of the operator
$\mathcal{O}_{r+l} = \Psi_\down^r(\Psi_\up^\dagger)^l$.
\begin{figure}[t]
\centering
\includegraphics[width=\columnwidth,keepaspectratio=true,clip]{fig2}
\caption{(Color online). Diagram showing the effect of commensurate
densities (see text for discussion) in the special case of
$n_\up = 1/3$, $n_\down=2/3$, i.e. $p=2$ and
$q=r=l=1$. For $2-2/9 < K_\down < 2/9$ the interaction with the
lattice leads to a formation of a 'trimer crystal' state. For larger(smaller)
values of $K_\down$ the system has a phase transition from a
massless phase into a Mott insulator of $\up$($\down$)-component.
The 'trimer' operator $\cos{2(2\phi_\up-\phi_\down)}$ is always
subdominant.}
\label{latt:fig_tommaso}
\end{figure}
In the following, for sake of simplicity, we assume equal velocities
of the two-components and drop the $\mathcal{H}_1$ term.
The dominant instability of the massless theory
$\mathcal{H}_0(\phi_\up) + \mathcal{H}_0(\phi_\down)$ is due to the
operator with largest positive scaling dimension. Depending on the
values of $K_\up$ and $K_\down$, the following inequalities define
which of the operators \eqref{latt:H1}--\eqref{latt:H4} is relevant:
\begin{align}
p^2 K_\up + q^2 K_\down &\leqslant 2 \label{latt:dim_pq} \; , \\
r^2 K_\up + l^2 K_\down &\leqslant 2 \label{latt:dim_rl} \; , \\
(pl+qr)^2 K_\down & \leqslant 2 \label{latt:dim_down} \;, \\
(pl+qr)^2 K_\up & \leqslant 2 \label{latt:dim_up} \;,
\end{align}
respectively. In Figs.\ \ref{latt:fig_tommaso} and
\ref{latt:fig_nup15ndown25} we plot the $(K_\up,K_\down)$ diagrams
corresponding to Eqs.\ \eqref{latt:dim_pq}--\eqref{latt:dim_up} for
two values of the densities. We see that which instability takes
place depends on the values of the bare Luttinger parameters $K_\up$
and $K_\down$, and thus on microscopic details of an underlying
lattice model. Numerically, a crystal phase has been reported~\cite{Keilmann2009}
in a two-component bosonic Hubbard model and a similar result
is presented in the fermionic counterpart in Sec.~\ref{sec:crystal}
for the commensurabilities discussed in Fig.~\ref{latt:fig_tommaso}.
These phases do correspond to the locking of several combinations of
the modes according to Eqs.\ \eqref{latt:H1}--\eqref{latt:H4} but
they are achieved for very large asymmetry. Consequently,
the above criteria \eqref{latt:dim_pq}--\eqref{latt:dim_up} determined
for equal velocities are not directly applicable in these situations.
The quantitative predictions of \eqref{latt:dim_pq}--\eqref{latt:dim_up}
could be relevant to the case of strongly renormalized $K_{\sigma}$,
for instance with long-range intra-species interactions.
A striking feature of the phase diagrams \ref{latt:fig_tommaso} and
\ref{latt:fig_nup15ndown25} is the appearance of the multicritical
points where several instabilities compete. In the above treatment we
have only considered an effect of various operators
\eqref{latt:H1}--\eqref{latt:H4} \emph{alone}. An interplay between
different operators is non-trivial and may lead to consequences not
captured by the simple power counting of
Eqs.~\eqref{latt:dim_pq}--\eqref{latt:dim_up}. Hence, applicability of
the above analysis in the vicinities of the multicritical points is
not granted. There are several possible scenarios of the phase
transitions at such multicritical points. For one thing, it is easy to
construct fine-tuned theories where two continuous transitions occur
simultaneously. An other possibility is a first order transition, as
been observed in numerical simulations of higher-dimensional bosonic
systems \cite{Kuklov2004}. Detailed analysis of these
multicritical points is beyond the scope of the present paper.
\begin{figure}[t]
\includegraphics[width=\columnwidth,keepaspectratio=true,clip]{fig3}
\caption{(Color online). Same as Fig.~\ref{latt:fig_tommaso} for
$n_\up = 1/5$, $n_\down=2/5$. In this case, Eqs.\
\eqref{latt:pqrl1}--\eqref{latt:pqrl2} allow \emph{two} sets of
solutions: (A) $l=1$ and $r=3$, and (B) $l=2$ and $r=1$, with $p=2$
and $q=1$ in both cases. Solution (A) is always subdominant, while
(B) dominates in the window $2/5<K_\down<2/25$. For
$48/25<K_\down<2/5$, the dominant instability is the formation of a
Luttinger liquid of trimers.}
\label{latt:fig_nup15ndown25}
\end{figure}
\section{Trimer formation in the 1D asymmetric Hubbard model}
\label{sec:trimers}
In this second part, we study the emergence of a trimer phase
on a particular microscopic model :
the 1D asymmetric attractive Hubbard model. After defining the model and providing
its phase diagram as a function of the parameters, we discuss some
limitations of the bosonization approach to this model and an
alternative phenomenological description that completes
the interpretation of the obtained data.
\subsection{Model and qualitative aspects}
We consider two species of fermions which internal degree of freedom
is denoted by a spin index $\sigma$. They hop on a lattice with a
spin-dependent amplitudes $t_{\sigma}$ (which would experimentally
corresponds to different optical lattices for each species) and
interact locally only in the inter-species channel with a Hubbard term
$U$ which we take negative, as suggested by the arguments of
Sec.~\ref{sec:boso} and as a natural choice to favor bonding between
particles. The Hamiltonian is then :
\begin{equation}
\label{eq:hubbard}
\mathcal{H} = -\sum_{i,\sigma=\up,\down} t_{\sigma}
[c^{\dag}_{i+1,\sigma}c_{i,\sigma} + \text{h.c.}] + U\sum_i n_{i,\up}n_{i,\down}\;.
\end{equation}
One of the key parameter for the physics is the ratio between the
hoppings $\eta = t_{\down}/t_{\up}$. In order to have the possibility
of forming trimers, we take the commensurate condition $2n_{\up} =
n_{\down}$ but the total density $n$ varies freely and is another
important parameter of the physics. Using the notations of
Sec.~\ref{sec:boso}, we thus have $p=2$ and $q=1$ (the simplest new
combination one can have). The Fermi momenta are $k_{\sigma} = \pi
n_{\sigma}$ and free fermions Fermi velocities read $v_{\sigma} =
2t_{\sigma} \sin(\pi n_{\sigma})$. Since $n_\down \leq 1$, the maximum
total density one can have for this commensurability is $n=3/2$.
The above Hamiltonian has been widely studied in the case of
balanced~\cite{Giamarchi1995} and imbalanced densities~\cite{Yang2001}
but the special commensurability where trimers emerge
has only been investigated for one set of data in
Ref.~\onlinecite{Burovski2009}, showing that the pairing correlations
were indeed suppressed, in agreement with the bosonization
approach. When the asymmetry is very large, one species behaves
quasi-classically (they get localized) and the model is in the regime
of the Falicov-Kimball (FK) model~\cite{Falicov1969} where there exists
a lot of quasi-degenerate states at low-energies, analogue to a phase
separation regime. We expect generically a first-order transition to
this segregated (or demixed) phase when lowering $\eta$ in the phase
diagrams. The FK regime can display rather rich physics recently
investigated in Ref.~\onlinecite{Barbiero2010} and which will not be
analyzed here : our aim is only to draw the boundary of this regime.
Numerically, the transition to the FK is rather sharp and all
observables clearly display segregation.
For $\eta=1$, the arguments of Sec.~\ref{sec:boso} suggest that the
two-mode regime will be generically realized. Qualitatively, in a
strong-coupling picture where two spin-$\down$ fermions are localized
on neighboring sites, the delocalization of a spin-$\up$ electron on
these sites will be favored by attractive interactions, forming a very
local trimer state. This picture will be correct at small enough
densities and actually not too large $U$ and too small $\eta$,
otherwise such bound states will agglomerate with other spin-$\up$ and
$\down$ fermions, leading to the FK regime. We thus expect the
formation of the trimer phase in the vicinity of the FK but at both
finite $U$ and finite $\eta$. Within the framework of
Sec.~\ref{sec:boso} and considering that the starting point of
bosonization are free fermions, the ratio between the velocities $\nu
= v_\down/v_\up = 2 \eta \cos(\pi n /3)$ supports that small $\eta$
clearly favors the formation of trimers while small densities should
not.
\begin{figure}[t]
\centering
\includegraphics[width=0.95\columnwidth,clip]{fig4}
\caption{Opening of the trimer gap increasing mass asymmetry (lowering
$\eta= t_\down/t_\up$) for a fixed interaction and density. The
magnitude of the gap (in units of $t_\up$) is small in comparison to $U$ and $t_\up$. The
grey areas are estimates of the transition points. \emph{Inset:} finite
size extrapolations of the gap. The upper dashed curve shows the
behavior for $\eta=0.2$ when entering in the FK regime.}
\label{fig:gap}
\end{figure}
\subsection{Phase diagrams}
The phase diagrams of model~\eqref{eq:hubbard} are numerically
determined using standard DMRG with
open-boundary conditions (OBC) and keeping up to $M = 2000$ states. In
order to discriminate between the different possible regimes, we use
both ``global'' probes and local observables and correlation
functions. Among ``global'' probes, one can use the trimer gap
$\Delta_{\text{t}}$ associated with the formation of the bound
state. It can be defined following Ref.~\onlinecite{Orso2010} as:
\begin{equation}
\begin{split}
\Delta_{\text{t}} = &E_0(N_\up+1,N_\down+1) + E_0(N_\up,N_\down+1) \\
&- E_0(N_\up+1,N_\down+2) - E_0(N_\up,N_\down)\;,
\end{split}
\end{equation}
with $E_0(N_\up,N_\down)$ the ground-state energy with $N_\up,N_\down$
fermions. Results as the function of the asymmetry $\eta$ for an
incommensurate density $n=3/7$ and large interaction $U=-4t_\up$ have
been extrapolated to the thermodynamical limit and are given in
Fig.~\ref{fig:gap}. The slow opening of the trimer gap is
\emph{qualitatively} compatible with the sine-Gordon behavior of
Sec.~\ref{sec:boso} although the transformation is not directly
applicable for any $\eta$. Notice that the whole system remains
gapless. The slow opening of the gap makes it difficult to precisely
locate the transition point. In such situation, a usual approach would
be to use the prediction on the critical Luttinger parameter
$K_a=K_a^c$ at the transition point. Furthermore, the determination of
$K_a$ using correlators in the two-mode phase is very
difficult as it would require to know, and then to disentangle,
the complicated expression of the exponents as a function of
$K_a$ and $K_b$ to extract them independently.
\begin{figure}[t]
\centering
\includegraphics[width=0.8\columnwidth,clip]{fig5}
\caption{(Color online). Examples of fits of the entanglement entropy
in the two-mode and single-mode phases using Eq.~\eqref{eq:EE}. It
shows a clear quantitative difference with respectively $c\simeq 2$
and $c\simeq 1$, as expected.}
\label{fig:EEexamples}
\end{figure}
\begin{figure}[t]
\centering
\includegraphics[width = 0.85\columnwidth,clip]{fig6}
\caption{(Color online). Central charge $c$ obtained from fits as in
Fig.~\ref{fig:EEexamples} as a function of the asymmetry $\eta$ for
$U=-4t_\up$ and $n=3/7$. The stair-like behavior with increasing
system size $L$ allows an efficient determination of the
transition. In the Falicov-Kimball regime (left), fits give
$c\simeq 0$ or irrelevant numbers. \textbf{(a)} Zoom of the
$c(L)$ curves in the crossing region illustrating the extraction of
crossing points between successive sizes. \textbf{(b)} Tentative
finite-size extrapolation of the crossing points $\eta_c(L)$.}
\label{fig:centralchargeU-4}
\end{figure}
Therefore, we use another global approach to the distinction between
the two-mode and single-mode phases, which is particularly well-suited
for this model, and more generally in a similar context.
Using universal results on the entanglement entropy
(EE), the central charge $c$ of the model can be extracted which
directly gives access to the number of bosonic modes, without further
information on their nature. Hence, we expect $c=2$ in the two-mode
regime while $c=1$ in the single-mode trimer phase. This stair-like
expectation in the thermodynamical limit will be smoothed out by
finite-size effects. The central charge is obtained on finite-systems
using the following ansatz for the EE between a left block of size $x$
and the right block of length $L+1-x$ with OBC:
\begin{equation}
\label{eq:EE}
S(x) = \frac c 6 \ln d(x|L+1) + A\,t(x) + B\;,
\end{equation}
where $d(x|L)$ is the cord function
\begin{equation}
\label{eq:cord}
d(x|L) = \frac L \pi \sin\left(\frac{\pi x}{L}\right)\;,
\end{equation}
and $t(x)$ is the local kinetic energy on bound $(x,x+1)$ (obtained
numerically), and $A,B$ are fitting parameters. The first log term is
the leading and universal one~\cite{Calabrese2004}
while the second accounts for finite-size oscillations due to OBC and
which can have a significant magnitude~\cite{Laflorencie2006, Roux2009}. It is thus essential to take them into
account to improve the quality of the fits. In the end, there are only
three parameters in the procedure and typical examples in both the
two-mode and single-mode phases are given in
Fig.~\ref{fig:EEexamples}. Systematic fits on finite-size systems
provide an estimate of $c$ as a function of the parameters. As seen in
Fig.~\ref{fig:centralchargeU-4}, the $c(L)$ curves crosses around the
transition point. Although we do not have any quantitative prediction
for the finite-size corrections of $c(L)$ obtained in this way, we can
argue that if $L$ is smaller than the correlation length associated
with the trimer gap, $c(L)$ will be larger than one as the system is
effectively in a two-mode regime. Thus, $c(L)$ should decrease with $L$
towards one in the single-mode phase, as observed. In the two-mode
phase, there is no obvious discussion : we only expect that the larger
the system, the better the agreement with the continuous limit. One
can also check the effect of the number of kept states on the fits and
see that they don't have the dominant effect in this model which
converges well numerically. We have estimated the transition point by
extrapolating the crossing points between successive sizes (see
Fig.~\ref{fig:centralchargeU-4}\textbf{(a)}) as a function of the
inverse size (see Fig.~\ref{fig:centralchargeU-4}\textbf{(b)}). From
this approach and the opening of the gap, we get a critical value
$\eta_c \simeq 0.54\pm 0.02$ for the mass asymmetry on this
cut. Although the gap is rather small, the trimer region appears to be
rather wide.
\begin{figure}[t]
\includegraphics[width=\columnwidth,clip]{fig7}
\caption{(Color online). Maps of the central charge $c$ vs interaction $U$ and asymmetry $\eta$
for a system with $L=112$ at four different densities. For
$n=3/7\simeq 0.428$, the lines with error bars are the ones
estimated from Figs.~\ref{fig:gap}
and~\ref{fig:centralchargeU-4}. The $\eta=0$ cuts correspond to
data obtained with a very low but non-zero value $\eta=0.005$.}
\label{fig:PhaseDiagN}
\end{figure}
Using the central charge calculation, one can map out the phase
diagram in the $(\eta,U)$ plane for a fixed density, or in the
$(\eta,n)$ plane for a fixed interaction $U$. Results are gathered in
Fig.~\ref{fig:PhaseDiagN} and \ref{fig:PhaseDiagU} respectively.
These diagrams display bare data for a
given system with a rather large size $L=112$ and the previous
estimate of the cut is given as error bars. These diagrams show that a
wide trimer phase can be achieved at large enough interactions, small
enough $\eta$, as expected, and also that low densities strongly
favors their formation. At large densities $n\simeq 1.3$, the trimer
region vanishes within our grid resolution so that it is at most confined to a
very tiny region between the two-mode phase and the FK regime. While
the large-$|U|$ situation is rather clear, the competition between the
three regimes at small $U$ is more involved. Indeed, two scenarios can
occur in the $(\eta,U)$ plane: either the trimer phase always
separates the FK and two-mode regimes, corresponding to two boundaries
starting from the $(\eta=0,U=0)$ corner, or there is a critical
$\vert{U}\vert$ above which the trimer phase emerges, corresponding to
a tricritical point $(\eta_c,U_c)$. We could not numerically
discriminate between both scenarios, but we do find a small trimer
region at relatively small $U$s ($U\simeq -1\,t_\up,-2\,t_\up$) for most
densities : we do not have evidence for a tricritical point with a
large $U_c$. As the density plays a central role in the stabilization
of the trimer phase, we give in Fig.~\ref{fig:PhaseDiagU} the central
charge map for a fixed interaction $U=-4t_\up$ as a function of the total density
$n$ and mass asymmetry. A similar question about a intervening trimer
phase between the two-mode and the FK regimes can be raised. While the
two-mode and FK are clearly separated at small densities, we found
that if a trimer intermediate phase exists at large densities up to
the $(\eta=0,n=1.5)$ point, its extension will be particularly small
(not seen within our numerical calculations). In addition to
the three main phases, commensurability effects are also present in
this diagram. When the maximum density $n=1.5$ is reached, the
$\down$-band is completely filled while the $\up$-band is half-filled,
leading to a single-mode phase well-captured by the central charge
approach. Lastly, as it will be discussed in Sec.~\ref{sec:crystal}, a
crystal phase (fully gapped) exists for the commensurate density $n=1$
at very small $\eta$ and is pointed on
Fig.~\ref{fig:PhaseDiagU}. Other commensurabilities could yield
additional crystal-like phases in this diagram but this is beyond the
scope of this study.
\begin{figure}[t]
\centering
\includegraphics[width=\columnwidth,clip]{fig8}
\caption{Map of the central charge $c$ vs asymmetry $\eta$ and total density $n$ for fixed
interaction $U=-4t_\up$ on a system with $L=128$. The lines with
error bars are the ones estimated from Figs.~\ref{fig:gap}
and~\ref{fig:centralchargeU-4}.}
\label{fig:PhaseDiagU}
\end{figure}
\subsection{Observables and effective behavior of the trimer liquid}
In this section, we give the behavior of several observables in order
to see how they are affected by the entrance into the trimer phase or
the FK regime and, as well, to investigate the effective behavior of
the trimer fermion.
\begin{figure}[t]
\centering
\includegraphics[width=\columnwidth,clip]{fig9}
\caption{(Color online). Maps of averaged local quantities in a system
with $L=56$ at total density $n=3/7\simeq 0.428$. The non-interacting
expectation ($U=0$ line) has been subtracted in order to unveil the
effect of the interaction.}
\label{fig:local_maps}
\end{figure}
\subsubsection{Local observables}
First, we select a set of local correlators
(living on sites or on bonds) which illustrate the phenomenological
picture of the different parts of the phase diagram. We compute the
local double occupancy $\moy{n_{i,\up}n_{i,\down}}$, the trimer local
operator as $\mathcal{T}_i = \frac 1 2
(\moy{n_{i,\up}n_{i,\down}n_{i+1,\down}}+
\moy{n_{i+1,\up}n_{i,\down}n_{i+1,\down}})$ (since the light particle
is in principle delocalized above two heavier), and the density
correlators $\moy{n_{i,\up}n_{i+1,\up}}$ and
$\moy{n_{i,\down}n_{i+1,\down}}$. This choice of local correlators is
well suited to a strong-coupling picture as pairs or trimers should in
principle correspond to a narrow bound-state, spread over only a few
lattice sites. These local correlators should then pick up a
reasonable weight of the local bound-state. The results are averaged
over all lattice sites and plotted in Fig.~\ref{fig:local_maps}. The
expectation value at $U=0$ has been subtracted so that the reference
state is the free fermions limit at a given $\eta$ (the pairing or trimer local
correlators defined above are obviously non-zero even in the free
fermions limit). Fig.~\ref{fig:local_maps} display behaviors in
qualitative agreement with the picture we have on the trimer formation : the
${\up\up}$ density correlator is nearly zero everywhere but in the FK
regime, signaling phase separation. On the contrary ${\down\down}$
density correlator increases significantly in the region corresponding
to the trimer phase, surrounding the FK pocket, and together with the
local trimer density $\mathcal{T}_i$. Lastly, we see that the double
occupancies, or pairs, acquire a strong weight with negative $U$
everywhere in two-mode and single-mode regions : they are either
``independent'' or embedded in the trimer bound-state. Their coherence
can yet be probed only by measuring correlations as discussed below.
\begin{figure}[t]
\centering
\includegraphics[width=0.9\columnwidth,clip]{fig10}
\caption{(Color online). Behavior of \textbf{(a)} pairing and
\textbf{(b)} trimer correlations when lowering $\eta$ and
entering the trimer phase along a cut at $U=-4t_\up$ in the phase
diagram (absolute values are displayed). The Inset of \textbf{(a)}
shows the same data but in log-linear scale to highlight the
exponential decay.}
\label{fig:TPcorrelations}
\end{figure}
\subsubsection{Pairing and trimer correlations, effective behavior of the trimers}
We now turn to the behavior of correlation functions across the phase
diagram. From Sec.~\ref{subsec:corr}, and as already observed for a
particular point in Refs.~\onlinecite{Burovski2009, Orso2010}, the
pairing correlations change from algebraic to exponential decay when
entering into the trimer phase. These correlations are here computed
in the singlet channel and for local pairs $\hat{P}_i =
c_{i,\up}c_{i,\down}$. In addition, we compute the trimer correlator
using the local trimer operator $\hat{T}_i =
c_{i,\up}c_{i,\down}c_{i+1,\down}$ defined on neighboring sites. The
associated correlation functions $P(x) =
\moy{\hat{P}^\dag_i\hat{P}_{i+x}}$ and $T(x) =
\moy{\hat{T}^{\dag}_i\hat{T}_{i+x}}$ are computed with $i$ taken at
the center of the chain. Increasing the mass asymmetry along the same
cut at $U=-4t_\up$ as in previous figures, the suppression of pairing correlations is
clearly seen in Fig.~\ref{fig:TPcorrelations}\textbf{(a)}. On the
contrary, trimer correlations, which are subdominant in the two-mode
regime, are boosted by smaller $\eta$, both in amplitude (as for the
local correlators previously evoked) and in the decay exponent which
gets smaller (see Fig.~\ref{fig:TPcorrelations}\textbf{(b)}). Notice
that the wave-vector is the same for both correlators since $k_\down -
k_\up=k_\up=\pi n /3$ for this commensurability. We have tentatively
extracted the decay exponents of both correlators by fitting the
functions using a power-law modulated by a cosine oscillations. The
correlation length $\xi$ of the pairing correlator in the trimer phase
is obtained using an exponential envelope $e^{-x/\xi}$. The results are gathered in
Fig.~\ref{fig:exponent}, showing the evolution in both phases. We must stress
that the data computed on a finite-size system display a transition at
a lower $\eta$ than in the thermodynamical limit. From
Sec.~\ref{subsec:corr}, we expect that the decay exponent of the
trimer propagator is of the form
$(K_{\text{eff}}+K_{\text{eff}}^{-1})/2$ while the $\up$-density
correlations have a decay exponent of $2K_{\text{eff}}$. A first
consequence is that the trimer exponent should always be larger than
one which is not reproduced for the lowest $\eta$s, and which we
attribute to numerical inaccuracies of the fits of trimer correlations.
Besides, inverting
$(K_{\text{eff}}+K_{\text{eff}}^{-1})/2$ to get $K_{\text{eff}}$ is
subjected to strong errors when $K_{\text{eff}}\simeq 1$ and does not
tell whether $K_{\text{eff}}>1$ or $K_{\text{eff}}<1$ which is
essential for the effective behavior of the trimers. In order to get
a better estimate of $K_{\text{eff}}$, we rather use Friedel oscillations of
the $\up$-density operator which decay exponent $\alpha$ is equal to
$K_{\text{eff}}$ in the trimer phase according to Sec.~\ref{subsec:corr}.
\begin{figure}[t]
\centering
\includegraphics[width=0.9\columnwidth,clip]{fig11}
\caption{(Color online). Comparison of the decay exponents of pairing
and trimer correlations. The correlation length of pairing
correlation (in units of the lattice spacing) is given in the trimer
phase. The strong grey areas are the previous estimates of the
transition points while the light maroon area illustrates the
location of the transition on the $L=147$ finite-size system under
study. The exponent of the Friedel oscillations of $n_\up$ is also
displayed, together with the expected trimer exponent derived from it
(see text for discussion).}
\label{fig:exponent}
\end{figure}
Even though the approach of Sec.~\ref{subsec:corr} is not
applicable for most parameters, the fact that there exists an
effective Luttinger exponent $K_{\text{eff}}$ describing the physics
of the fermionic trimer with a propagator
$(K_{\text{eff}}+K_{\text{eff}}^{-1})/2$ and with Friedel
oscillations with $K_{\text{eff}}$ is more general : the limitation of
the bosonization approach is rather that $K_{\text{eff}}$ will not take the form of
Eq.~\eqref{eq:Keff}. Local observables are believed to have less
numerical errors associated with a finite number of kept states than
correlations~\cite{White2007}. Thus, we fit the Friedel oscillations
of the $\up$-density using the following symmetric ansatz
\begin{equation}
\label{eq:friedel}
n_{i,\uparrow} = n_0 + A\frac{\cos\,q(i -\frac{L+1}{2})}{[d(i|L+1)]^{\alpha}}
\end{equation}
with $1 \leq x \leq L$ and only four fitting parameters $n_0$, $A$,
$q$ and $\alpha$~\footnote{We expect that $q = 2\pi n_{\uparrow} = 2\pi n/3$
and $n_0 = n_{\uparrow}$ but at low densities, it is usually better to take $q$
and $n_0$ as free independent fitting parameters due to the depletion
of the density at the edges which effectively increases it in the
bulk. For instance, one has $n_0=(N+1/2)/(L+1)$ for free spinless
fermions on a finite system with $N$ fermions, which is not exactly
the average density $N/L$, particularly at small $N/L$.}.
In the trimer phase, we expect $\alpha = K_{\text{eff}}$. Some typical fits are
plotted in Fig.~\ref{fig:density}\textbf{(a)}. From them, we
extract the decay exponent $\alpha$ and plot it on
Fig.~\ref{fig:density}\textbf{(b)} as a function of the total
density. A cusp is found around $n\simeq 0.6$ signaling the transition
from the two-mode regime to the trimer phase. We have seen that a low
density favors the formation of the trimer phase. This figure shows
that, in the trimer
phase, we have $K_{\text{eff}}<1$ for the larger densities,
corresponding to a repulsive effective interaction between trimers
(dominant CDW order of trimers). We observe that the
exponent increases with decreasing density, compatible with the fact
that at low densities in the trimer regime, the trimers should be
close to free spinless fermions having $K_{\text{eff}}\simeq 1$.
Bare data for the smallest density on a system with $L=144$ even
display an exponent $K_{\text{eff}}\simeq 1.2$ larger than one.
\begin{figure}[t]
\centering
\includegraphics[width=\columnwidth,clip]{fig12}
\caption{(Color online). \textbf{(a)} Typical Friedel oscillations of
the $\up$-density for $U=-4t_\up$, $\eta=0.3$ and $L=144$ for
various densities $n$. Full lines are fits using
Eq.~\eqref{eq:friedel}. \textbf{(b)} Decay exponents obtained from
the fits as a function of the density. The cusp at $n \simeq 0.6$
roughly corresponds to the transition from the two-mode to the
single-mode regime. \textbf{(c)} Large exponents at low densities,
close to the FK regime when lowering $\eta$ : increasing the size
tends to reduce the exponent below one.}
\label{fig:density}
\end{figure}
Interestingly, the trimers in this model are necessarily objects
with a finite extension of at least two sites and two close trimers
may have the possibility to overlap by delocalizing their $\up$
electrons. The distance dependence and sign of the effective
interaction between trimers is non-trivial -- a perturbation
theory to derive it looks challenging as it
involves many sites and degrees of freedom. Yet, since the
trimer phase is found close to the FK regime, we can expect the effective interaction to become
attractive close to this boundary, leading to $K_{\text{eff}}>1$. Such
physics would correspond to a superfluid phase of trimers. This would
be physically very remarkable since the microscopic
Hamiltonian~\eqref{eq:hubbard} would contain both the formation of
bound-states, or molecules, and their effective superfluid
behavior. However, the behavior close to a phase separation at small
densities is numerically involved. Indeed, increasing the system size
$L$ shows that $\alpha$ actually tends to decrease below one, as
reported in Fig.~\ref{fig:density}\textbf{(c)}, or one enters the FK
regime for larger sizes. We did not find clear evidence of a
stabilization of $K_{\text{eff}}>1$ in the thermodynamical limit and
understand the observed $K_{\text{eff}}>1$ as finite size effects. A
superfluid droplet picture can be qualitatively put forward. Starting
from the FK regime and looking at the local density pattern, one sees
that the fermions are clustered into droplets while other parts of the
box are empty. Approaching the trimer phase from the FK regime tends
to increase the size of these droplets to gain kinetic energy. When a
box confinement is present (finite system with OBC), it naturally favors the
overlap between trimers, by depleting the edges, and can prevent the
droplet to form (for instance if their typical size is larger than the
box size). Increasing further the size of the box (at constant
density) can lead to droplets formation.
This is a possible interpretation of the data
observed in Fig.~\ref{fig:density}\textbf{(c)}.
In addition, we must stress
that there is a lot of competing low-energy states in the FK regimes
so DMRG, as an essentially variational methods, can be trapped into
metastable states. Even though the
thermodynamical limit is unclear, it is experimentally motivating to
have signatures of superfluidity on mesoscopic confined systems as one
has in cold-atoms setups. We further mention that a recent careful
study of the t-J model on a chain~\cite{Moreno2010} which could
qualitatively contain a similar phenomenon as pair-clustering did not
find evidences for such clustering.
\begin{figure}[t]
\centering
\includegraphics[width=\columnwidth,clip]{fig13}
\caption{(Color online). Density correlations in \textbf{(a)} the
trimer phase and \textbf{(b)} the two-mode phase with $n=3/7$.}
\label{fig:densityCorr}
\end{figure}
\subsubsection{Locking between density correlations}
Lastly, the comparison of density-density correlations in the $\up$
and $\down$ channels is another interesting point of this model. In
fact, the bosonization results of Sec.~\ref{subsec:corr} predicts that
the exponent of $\mathcal{N}_{\up\up}(x)$ should be four times larger
than the exponent of $\mathcal{N}_{\down\down}(x)$ (if both remains
smaller than two), and that the dominant wave-vector should differ by
a factor two. Numerically, the typical behavior for a rather large
interaction $U=-4t_\up$ is given on Fig.~\ref{fig:densityCorr} for two
values of the asymmetry $\eta$ in the single-mode and two-mode regimes.
We see
that in the two-mode phase (Fig.~\ref{fig:densityCorr}\textbf{(b)}),
the two fluctuations have a slightly different exponent, that the
amplitude are quite different (including the natural factor four). Yet,
the dominant wave-vectors are both $2k_\up=k_\down$. In the trimer
phase, the disagreement with the bosonization picture is even worse
since the two densities are locked together, and nearly identical
(Fig.~\ref{fig:densityCorr}\textbf{(a)}). This latter fact cannot be
explained by the $1/x^2$ decay since the leading term is the
oscillating one, with an exponent clearly smaller than two. It is yet
physically not surprising in the strong coupling picture of
Fig.~\ref{fig:trimers} : trimers are local bound-states separated by
the typical distance $1/n_\up = 2\pi/2k_\up$ which does correspond to
the $2k_\up$ fluctuations but cannot be accounted by any of the
harmonics for the $\down$-density operator of Eq.~\eqref{HaldaneN} (we
work at an incommensurate filling). This short-distance binding cannot
be captured by the bosonization results of Sec.~\ref{sec:boso} but a
phenomenological Bose-Fermi approach described in the next section can
account for this strong-coupling regime. Lastly, the same comment can
be made on Friedel oscillations on the $\down$-component : they are
locked to the $\up$-component in the strong-coupling picture. One
might argue that there could be a crossover from the weak-coupling to
strong-coupling picture of Fig.~\ref{fig:trimers} as $|U|$ increases,
so that the bosonization results could be valid in the small $U$s
region. However, the trimer region is very sharp at small $U$s and we
could not find evidence for such a weak-coupling behavior in our
numerical data, although we cannot exclude such a possibility.
\subsection{Phenomenological Bose-Fermi picture at large $\vert{U}\vert$}
\label{subsec:bosefermi}
We here propose a simple picture that reconciles the numerical
observation and a bosonization approach at the prize of a strong
assumption, physically reasonable at large negative $U$, but difficult
to justify rigorously starting from the microscopic model.
This picture has been for instance proposed at large
interaction and low-density limit~\cite{Iskin2008}.
Bose-Fermi mixtures of 1D models have extensively studied in recent
years~\cite{Cazalilla2003,
Pollet2006, Mathey2007, Mathey2007a, Hebert2008, Barillier-Pertuisel2008, Rizzi2008, Mering2008,
Marchetti2009, Crepin2010, Orignac2010} and a similar picture emerges in certain
regimes of three-component Fermi-gases~\cite{Luscher2009}. When
$\vert{U}\vert$ is large, $\up$ and $\down$ fermions naturally form
onsite pairs which are effectively hard-core bosons which we
label $b$. We phenomenologically assume that the system is equivalent
to a Luttinger liquid of hard-core bosons with density $n_b = n_\up$,
an effective velocity $v_b$ and Luttinger parameter $K_b$, while the
remaining unpaired $\down$ fermions behave as a Luttinger liquid of
fermions labeled by $f$ and with parameters $n_f$, $v_f$ and
$K_f$. These two Luttinger liquids interact through an
effective interaction which will contain terms such as
\begin{equation}
\label{eq:BFcosine}
\int\! dx \, \cos{\left[2\pi(n_f-n_b)x - 2(\phi_f-\phi_b)\right]} \;,
\end{equation}
which have the tendency to lock the fields $\phi_f$ and $\phi_b$
together (with $\moy{\phi_f}=\moy{\phi_b}$ for attractive interaction), provided $n_f=n_b$. Such an effect has already been
discussed in the context of Bose-Fermi mixtures~\cite{Rizzi2008,
Marchetti2009}. Clearly, the latter relation is the same as the
trimer commensurability condition $n_\down = 2n_\up$ (because bosons carry two particles)
so that the
formation of trimer is now interpreted as a bound-state between the
bosons and the fermions. Following the same reasoning as in
Sec.~\ref{sec:fieldtransfo}, we can introduce a general transformation
of the $b/f$ fields into two new fields $c/d$ where $\phi_c =
(\phi_f-\phi_b)/\sqrt{2}$. Writing the matrix transformation from the
$r=c,d$ to the $s=b,f$ as $\mathfrak{p}_{rs}$ for the $\phi$s and
$\mathfrak{t}_{rs}$ for the dual $\theta$s, we have
\begin{align}
\label{eq:trs}
\mathfrak{t}_{cf} &= \frac{1}{\sqrt{2}}\;, & \mathfrak{t}_{cb} &= -\frac{1}{\sqrt{2}}\,.
\end{align}
which is slightly different from Eq.~\eqref{eq:tas}. Yet, the
2$k_F$-like fluctuating part of the density correlators for the
fermions and the bosons will have the leading contributions (dropping
the $x^{-2}$ terms) :
\begin{gather}
\mathcal{N}_{ff}(x) \sim \frac{\cos(2k_f x)}{x^{2\mathfrak{p}_{df}^2K_d}} \label{eq:fnupnup}\;, \\
\mathcal{N}_{bb}(x) \sim \frac{\cos(2k_b x)}{x^{2\mathfrak{p}_{db}^2K_d}} \label{eq:bfdndn}\;,
\end{gather}
which have both the same wave-vector associated with the Fermi levels
$k_f = k_b = k_\up$ and same decay exponents since $\mathfrak{p}_{df}
= \mathfrak{p}_{db}$ from the canonical transformation relations. In this picture, the
trimer is simply a bound-state between the bosons and the fermions so
its propagator reads
\begin{align}
T(x) &\sim e^{ik_f x}e^{i\theta_b}e^{i(\theta_f-\phi_f)} \\
&\sim e^{ik_\up x}e^{i[(\mathfrak{t}_{df}+\mathfrak{t}_{db})\theta_d
+ (\mathfrak{t}_{cf}+\mathfrak{t}_{cb})\theta_c
- \mathfrak{p}_{df}\phi_d - \mathfrak{p}_{cf}\phi_c]}\,.
\end{align}
From Eq.~\eqref{eq:trs} and the determinant of the transformation
matrix, which gives that $\mathfrak{t}_{df}+\mathfrak{t}_{db} =
1/\mathfrak{p}_{db} = 1/\mathfrak{p}_{df}$, we obtain that the
propagator is of the spinless fermionic type with an effective
Luttinger parameter $K'_{\text{eff}} = \mathfrak{p}_{df}^2K_d$. Clearly, both
the $f$-fermions and $b$-bosons propagators become short-range, the
latter corresponding to the pairing correlations in the native
fermionic model. Consequently, we recover the physics of the trimer
phase developed in Sec.~\ref{sec:boso}, with a better agreement with
the numerical observations in the strong coupling regime. However,
splitting the initial gas of $\down$ fermions into two parts can only
be done phenomenologically and could be questionable in a microscopic
derivation. This highlights the limitation of the bosonization
approach of Sec.~\ref{sec:boso} at short distances (high energies).
\subsection{Possible observation of the trimer phase in the presence of parabolic confinement}
\label{subsec:trapped}
\begin{figure}[b]
\centering
\includegraphics[width=\columnwidth,clip]{fig14}
\caption{(Color online). Realization of the trimer phase in a trapped
configuration with parameters $U=-4t_\up$, $\eta=0.3$,
$\omega=0.002\sqrt{t_\up}$, and $N=51$ fermions. \textbf{(a)} Local densities
profiles. \textbf{(b)} Correlation functions from the center of the
trap. \emph{Inset}: same in log-linear scale.}
\label{fig:trapped}
\end{figure}
In this section, we briefly discuss the condition to favor the trimer
phase in the presence of a parabolic confinement, as used in
cold-atoms experiments. Our goal is only to exhibit some parameters
for which the trimer phase is stabilized and to give some
qualitative comments. The trapping potential is taken into account
by adding the quadratic term
\begin{equation}
\mathcal{H}_{\text{trap}} = \frac 1 2 \omega^2 \sum_i (i-i_0)^2
\end{equation}
to Eq.~\ref{eq:hubbard}, with the trapping frequency $\omega$ and the
center of the lattice $i_0 = (L+1)/2$. According to a local-density
approximation (LDA) picture and using the phase diagram of
Fig.~\ref{fig:PhaseDiagU}, the trimer phase is likely to be found at
small enough densities and not too small $\eta$ to prevent the
occurrence of the FK regime. However, we find that the average density
of the trapped system is strongly dependent on the Hamiltonian
parameters. At fixed number of particles $N$ and trap size $\omega$,
changing $U$ and $\eta$ strongly affects the radius of the cloud and
the density in the middle. We only exhibit in Fig.~\ref{fig:trapped}
parameters for which the main features of the trimer phase are
reproduced in the presence of a parabolic confinement. The density
profiles of Fig.~\ref{fig:trapped}\textbf{(a)} illustrate the locking
of the $\up$ and $\down$ densities (up to exactly a factor two), and
the emergence of an appreciable density of local trimers (each local
maximum roughly corresponding to a trimer). In
Fig.~\ref{fig:trapped}\textbf{(b)}, the pairing and trimer
correlations are strongly different from that of a superfluid phase :
we have dominating trimer correlations and exponential pairing
correlations as in the homogeneous counterpart. In agreement with a LDA picture, since we have seen
that $K_{\text{eff}}$ decreases with density, the trimer correlations are boosted
at long distances. Similarly, the pairing correlations decrease
slightly faster then an exponential close to the edge of the
cloud. These results are encouraging in the perspective of a possible
achievement of the trimer phase in actual experiments.
\subsection{Observation of a crystal of trimers}
\label{sec:crystal}
\begin{figure}[t]
\centering
\includegraphics[width=\columnwidth,clip]{fig15}
\caption{(Color online). Observation of the crystallization of trimers
in the commensurability situation of
Fig.~\ref{latt:fig_tommaso} ($n=1$). \textbf{(a)} the short-range behavior of
both pairing and trimer correlations. \textbf{(b)} ordering of the
local density $n_\up$ with the expected period of three sites.}
\label{fig:crystal}
\end{figure}
According to the analysis of Sec.~\ref{sec:lattice}, a crystalline
phase of trimers can occur in this lattice model when the total
density $n$ is commensurate. Evidence of this scenario together with a
phase diagram for $n=1$ has been proposed in
Ref.~\onlinecite{Keilmann2009} in the case of a mixture of
two-component bosons, for large enough asymmetry. As the order
parameter (the density) associated with this transition is independent
of the statistics, we expect a similar scenario (see
Sec.~\ref{sec:lattice}) and a similar location of the transition in
the fermionic version of the model under study. Indeed, we give in
Fig.~\ref{fig:crystal} an example of the crystal phase. Notice that
very small $\eta$ are required to stabilize such a phase. We have not
investigated the extension of the phase which should rather be small
on the scale of the phase diagram of Fig.~\ref{fig:PhaseDiagU} and its
neighboring phases which could be either the two-mode LL or the
trimer phase. A crude argument can be proposed to understand this
crystallization within the Bose-Fermi picture of
Sec.~\ref{subsec:bosefermi} : when the mass asymmetry is very large
(very small $\eta$), it is reasonable to say that the mass of the
boson will be essentially the one of the heaviest particle, which is
the same as the unpaired fermions so that $v_f \simeq v_b$. In terms
of commensurability effects, one has $n_f = n_b = n/3$ so that
standard umklapp terms at $2\pi(n_f+n_b) = 4\pi/3$ do not account for
the crystallization. One rather has to look for higher order terms
with commensurabilities such as $2n_f+n_b = n_f+2n_b = n = 1$, which
are typically associated with terms like
\begin{equation}
\label{eq:BFcosinePlus}
\int\! dx \, \cos{2(2\phi_f+\phi_b)}\;,
\end{equation}
in addition to the one of Eq.~\eqref{eq:BFcosine}. Such term can lock
the field $2\phi_f+\phi_b$ and make the system fully gapped. Lastly,
we would like to stress that such commensurabilities are rather
surprising in terms of the initial fermion densities as they belong to
\emph{odd} filling fractions $n_\up = 1/3$ and $n_\down=2/3$.
\section{Conclusions}
In summary, the consideration of unusual commensurability conditions
in density-density interactions for 1D two-component
gases leads to a very rich physics with the possibility of building
bound-states of $(p+q)$-particles as the leading order. Such a
mode-locking mechanism can be described within the framework of
Luttinger liquid theory which reveals the main ingredients to
stabilize such a new phase. In particular, mass or velocity asymmetry
is shown to drive efficiently the transition into the multimer phase.
Novel fully gapped phases are proposed
when taking into account umklapp couplings specific to lattice models
at commensurate densities. These ideas are illustrated and confronted with the
asymmetric 1D attractive Hubbard model for the special commensurability $2n_\up =
n_\down$ for which the formation of trimers is found. The features of
the phase diagram are computed, displaying the important role of the
density in favoring the trimer phase. The effective behavior of
trimers, which are effectively spinless fermionic objects, is very
sensitive to the density and mass asymmetry. Although the model seems
to have promising features to sustain a superfluid phase of trimers,
we did not find clear evidence for it in the thermodynamical limit,
while finite-size systems display a ``superfluid droplets'' physics.
Notice that superfluidity of bound-states made of four fermions
(quartets) can be achieved reliably in 1D with a four-color Hubbard
model~\cite{Capponi2008, Roux2009}. There, the bound-states are bosons
which natural ``free'' regime (attained in the low-density limit) is a
superfluid phase. A superfluid phase of trimers would in this respect
be even more exotic but is in strong competition with phase
separation. Lastly, we found that a trapping confinement supports the
trimer phase for reasonably high densities and that surprising crystal
phases can emerge at commensurate densities.
\begin{acknowledgments}
We would like to thank Giuliano Orso for earlier collaborations on related
problems. G.~R. thanks Fran{\c c}ois Cr{\'e}pin, Fabian Heidrich-Meisner
and Alexei Kolezhuk for fruitful discussions.
E.~B. gratefully acknowledges the hospitality of LPTMS, where the
majority of this work was done. We have benefited from the supports of the \textit{Institut
Francilien de Recherche sur les Atomes Froids} (IFRAF) and ANR
under grant 08-BLAN-0165-01.
\end{acknowledgments}
|
\section{Introduction}
\label{Introduction}
The past few years have witnessed major strides in our efforts to
understand the atmospheric circulation of short-period
exoplanets---both gas giants (hot Jupiters) and smaller terrestrial
planets. Infrared photometry, spectra, and
light curves from the {\it Spitzer} and {\it Hubble} Space Telescopes now
provide constraints on the three-dimensional temperature structure of
several hot Jupiters, which hint at a vigorous atmospheric circulation on
these bodies \citep[e.g.,][]{knutson-etal-2007b, knutson-etal-2009a,
charbonneau-etal-2008, harrington-etal-2006,
cowan-etal-2007, swain-etal-2009, crossfield-etal-2010}. These
observations have motivated a growing effort to model the atmospheric
circulation on these objects: to date, many three-dimensional atmospheric circulation models
of hot Jupiters have been published \citep{showman-guillot-2002,
cooper-showman-2005, cooper-showman-2006, showman-etal-2008a,
showman-etal-2009, dobbs-dixon-lin-2008, menou-rauscher-2009,
rauscher-menou-2010, dobbs-dixon-etal-2010, thrastarson-cho-2010,
lewis-etal-2010, perna-etal-2010, heng-etal-2010}. These models have
emphasized synchronously rotating hot Jupiters in circular,
approximately 2--5-day orbits.
Just as the last decade witnessed the first characterization of hot Jupiters,
the next decade will see a shift toward characterizing
``super Earths'' (planets of 1--10 Earth
masses) and terrestrial planets. To date, roughly 30 super Earths have been
discovered, including several that transit their host stars
\citep{charbonneau-etal-2009, leger-etal-2009, batalha-etal-2011}
with hundreds of additional candidates recently announced from
the NASA {\it Kepler} mission
\citep{borucki-etal-2011}. Attempts to observationally characterize
their atmospheres have already begun \citep{bean-etal-2010}. In
anticipation of this vanguard, several three-dimensional circulation
models of tidally locked, short-period terrestrial exoplanets have
been published \citep{joshi-etal-1997, joshi-2003, merlis-schneider-2011,
heng-vogt-2011}.
Intriguingly, the flows in most of these three-dimensional
models---both hot Jupiters and terrestrial planets---develop
a fast eastward, or {\it superrotating}, jet stream at the equator, with
westward flow typically occurring at deeper levels and/or higher latitudes.
In hot-Jupiter models,
the superrotating jet extends from the equator to latitudes of typically
20--$60^{\circ}$ and is perhaps the dominant dynamical feature
of the modeled flows. In some cases (depending on the strength of
the imposed stellar heating and other factors), this jet causes an eastward
displacement of the hottest regions
from the substellar point by $\sim$$10^{\circ}$ to $60^{\circ}$ longitude.
\citet{showman-guillot-2002} first predicted this feature and suggested
that, if it existed on hot Jupiters, it would have important implications for
infrared spectra and light curves. This prediction has been confirmed
in Spitzer infrared observations of HD 189733b
\citep{knutson-etal-2007b, knutson-etal-2009a}, suggesting that this
planet may indeed exhibit a superrotating jet.
However, despite the ubiquity of equatorial superrotation in three-dimensional
models of synchronously rotating exoplanets---and its
relevance for observations---the mechanisms that produce
this superrotation have yet to be identified. As demonstrated in
a theorem due to \citet{hide-1969}, such superrotation cannot result
from atmospheric circulations that are longitudinally
symmetric or that conserve angular momentum per mass about the planetary
rotation axis. The equatorial atmosphere is the region of the planet
farthest from the planetary rotation axis, and a superrotating equatorial
jet therefore corresponds to a local maximum in the angular momentum per
mass about the planetary rotation axis. Thus, any angular-momentum-conserving
circulation that moves air to the equatorial atmosphere from higher
latitudes or deeper levels tends to produce {\it westward} equatorial flow.
Equivalently, Coriolis forces
always induce westward acceleration for air moving equatorward or upward,
so an eastward equatorial jet cannot result from Coriolis forces
acting on air that moves into the equatorial atmosphere from higher latitudes
or deeper levels. In Earth's equatorial troposphere, for example,
the flow is westward, which results from the tropospheric Hadley
cell on Earth (a regime where a mean overturning circulation and its
Coriolis accelerations plays the defining role \citep{held-hou-1980}).
To maintain equatorial superrotation, a mechanism is needed that
pumps angular momentum per mass from regions where it
is {\it small} (outside the jet) to regions where it is {\it large} (within
the jet)---a so-called ``up-gradient''
momentum transport. According to Hide's theorem, this transport can
only be accomplished by waves or eddies.
Equatorial superrotation exists in several atmospheres of the
Solar System---the equatorial atmospheres of Venus, Titan, Jupiter, and
Saturn all superrotate. Even localized layers within Earth's equatorial
stratosphere exhibit superrotation, part of a phenomenon called
the ``Quasi-Biennial Oscillation'' or QBO \citep{andrews-etal-1987}.
The mechanisms for driving the equatorial superrotation on
these planets are diverse. Possible mechanisms include eddy transport
associated with baroclinic instabilities, barotropic instabilities, turbulence,
and the absorption/radiation of various types of atmospheric waves
\citep[e.g.,][]{williams-2003a, williams-2003b,
lian-showman-2008, lian-showman-2010, schneider-liu-2009, delgenio-etal-1993,
delgenio-zhou-1996, andrews-etal-1987, mitchell-vallis-2010}.
Here, we demonstrate how the equatorial superrotation in three-dimensional
models of synchronously rotating, short-period exoplanets can result from the
existence of standing, planetary-scale Rossby waves;
such waves are naturally excited by the longitudinally dependent
heating patterns---dayside heating and nightside cooling---that
accompany the photospheric regions of short-period, synchronously rotating
planets. Section 2 provides background. In Section 3, we present an analytic
theory demonstrating the mechanism, and we systematically explore its behavior
in idealized, nonlinear one-layer models. In Section 4 we extend the
analysis to three-dimensional circulation models. In Section 5, we
summarize and discuss implications.
\section{Background Theory}
\label{background}
The ability of Rossby waves to accelerate jets can be schematically
illustrated using the
two-dimensional non-divergent barotropic vorticity equation,
which is the simplest model for the global-scale flow of a
planetary atmosphere \citep[see discussion in][]{vallis-2006, showman-etal-2010}.
The equation reads
\begin{equation}
{d(\zeta + f)\over dt} = F
\label{barotropic-vorticity-eq}
\end{equation}
where $\zeta\equiv {\bf k}\cdot\nabla\times {\bf v}$ is the relative
vorticity, ${\bf v}$ is the horizontal wind velocity, ${\bf k}$ is
the local upward unit vector, $f\equiv 2\Omega\sin\phi$ is the
Coriolis parameter, $\Omega$ is the planetary rotation rate ($2\pi$
over the rotation period), $\phi$ is latitude, and $d/dt = \partial/\partial t +
{\bf v}\cdot\nabla$ is the material derivative (i.e., the derivative
following the flow). The equation states that individual fluid parcels
conserve the absolute vorticity, $\zeta + f$, save for vorticity
sources/sinks, which are represented by the term $F$. The equation
can equivalently be written
\begin{equation}
{\partial\zeta\over\partial t} + {\bf v}\cdot \nabla\zeta + v\beta = F
\label{barotropic-vorticity-eq2}
\end{equation}
where $v$ is the meridional (northward) wind speed and $\beta=df/dy$
is the gradient of the Coriolis parameter with northward distance
$y$. Because the flow in this simple model is horizontally non-divergent,
we can define a streamfunction $\psi$ such that $u=-\partial\psi/\partial y$
and $v=\partial \psi/\partial x$, where $x$ is the eastward
coordinate and $u$ is the zonal (eastward) wind speed. This allows
the equation to be written as a function of one variable, $\psi$.
For purposes of illustration, consider the linearized version of
Eq.~(\ref{barotropic-vorticity-eq2}) with no sources and sinks.
The solutions to this linearized, unforced equation are Rossby waves.
For simplicity, we
consider Cartesian geometry with constant $\beta$, representing a local
region on the sphere. Decomposing variables into zonal means (denoted
with overbars) and deviations therefrom (denoted with primes), and
assuming that the mean flow is zero, leads to a solution
given by $\psi'=\hat\psi \exp[i(kx+ ly)]$,
where $i$ is the imaginary number and $k$ and $l$ are the zonal and meridional
wavenumbers. The dispersion relation is
\begin{equation}
\omega = -{\beta k\over k^2 + l^2}.
\end{equation}
These waves propagate meridionally with a group velocity given by
\begin{equation}
{\partial\omega\over\partial l}={2\beta kl\over (k^2+l^2)^2}.
\label{group-velocity}
\end{equation}
A simple argument, first clearly presented by \citet{thompson-1971}
and reviewed in \citet{held-2000} and
\citet{vallis-2006}, shows how these waves can produce an east-west
acceleration of the zonal-mean flow. The latitudinal transport of eastward
eddy momentum per mass is $\overline{u'v'}$, where $u'$ and $v'$ are
the deviations of the zonal and meridional winds from their zonal means,
respectively,
and the overbar denotes a zonal average. Using the solutions for the
wave-induced zonal and meridional wind,
$u'=-il\hat\psi\exp[i(kx+ly)]$ and $v'=ik\hat\psi\exp[i(kx +ly)]$,
yields a momentum flux
\begin{equation}
\overline{u'v'}=-{1\over 2}\hat\psi^2 kl.
\label{uprime-vprime}
\end{equation}
Since the group velocity must point away from the region of wave
generation (which we call the ``wave source''), and
since $\beta$ is positive, we
must have $kl>0$ north of the source and $kl<0$ south of the source.
Thus, {\it north} of the source,
$\overline{u'v'}$ is negative, implying southward transport of
eastward momentum. But {\it south} of the source, $\overline{u'v'}$
is positive, implying northward transport of eastward momentum.
Rossby waves therefore transport eastward momentum into the wave source
region. An eastward acceleration must therefore occur in the wave
source region and a westward acceleration must occur in the region
of wave breaking or dissipation. This can lead to the formation
of zonal (east-west) jet streams.\footnote{The dynamical picture outlined
above is not limited to small-amplitude
disturbances, as can be shown with a simple argument described for
example in \citet{held-2000} and \citet{vallis-2006}. Imagine
an initially undisturbed latitude, where the absolute-vorticity contour
initially aligns with the latitude circle, and suppose a disturbance---of any
amplitude---propagates into that latitude from elsewhere. The disturbance
will perturb the absolute vorticity contours, causing northward transport
of air in some regions and southward transport in others. Because
absolute vorticity generally increases northward,
the northward advection carries with it air of low absolute vorticity,
whereas the southward advection carries with it air of high absolute
vorticity. Thus, this process will generally cause a southward flux of
absolute vorticity, thereby decreasing the areal integral of the absolute
vorticity over the polar cap bounded by the latitude circle in question.
By Stokes' theorem,
this implies that the zonal-mean zonal wind
{\it decelerates} (i.e., accelerates westward) because of this vorticity
flux. In absence of dissipative processes, this deceleration would reverse
if the disturbance exited the region. However, when mixing occurs
(e.g., if the wave breaks), or if the disturbance is damped before
air parcels can return to their original latitudes, then the areal
integral of the vorticity inside the latitude circle has been irreversibly
decreased, and the westward impulse cannot be undone. Thus, we
again recover the result that westward acceleration occurs in the
region of wave dissipation; if momentum is conserved, eastward acceleration
would then occur in the wave-source region.}
Rossby waves correspond to latitudinal oscillations in
surfaces of constant potential vorticity\footnote{Potential vorticity
is a quantity related to vorticity
that is conserved in adiabatic, frictionless, stratified
flow. For the barotropic system it is simply the absolute
vorticity $\zeta+f$; for the shallow-water system it is absolute
vorticity over layer thickness $(\zeta+f)/h$, and for a
three-dimensional stratified atmosphere is it given by
$\rho^{-1}(\nabla\times{\bf v}+2{\bf \Omega})\cdot\nabla\theta$,
where $\rho$ is density, $\Omega$ is the planetary rotation vector,
and $\theta$ is potential temperature. For discussion of the
conservation of potential vorticity and its uses in dynamics, see
\citet{pedlosky-1987} or \citet{vallis-2006}.}; thus, any process
that triggers such oscillations at large scales will tend to excite
Rossby waves. In Earth's atmosphere, one of the predominant sources
is baroclinic instability, which occurs
in the mid-latitude troposphere where latitudinal temperature gradients
are large. Spatially varying tropospheric heating and cooling
(e.g., due to land-sea contrasts)
or flow over topography also perturb the potential vorticity contours
and can therefore trigger Rossby waves. In the atmospheres of tidally
locked, hot exoplanets,
on the other hand, the day-night heating pattern constitutes the
overriding dynamical forcing. For such planets, we expect this
heating/cooling pattern to trigger Rossby waves at low latitudes
(Fig.~\ref{qualitative-scenario}).
The above theory is for free waves. Consider now the extension
to an atmosphere forced by vorticity sources/sinks and damped by frictional
drag. The zonal-mean zonal momentum equation of the barotropic system
reads
\begin{equation}
{\partial \overline{u}\over\partial t}=-{\partial (\overline{u'v'})\over
\partial y} - {\overline{u}\over\tau_{\rm drag}}
\label{barotropic-zonal-mean-momentum}
\end{equation}
where overbars denote zonal means and primes denote deviations
therefrom. The equation states that accelerations of the zonal-mean zonal flow result
from convergences of the latitudinal eddy momentum flux and from drag,
which we have parameterized as a term that relaxes the
zonal-mean zonal wind toward zero over a drag time constant $\tau_{\rm drag}$.
The relationship between the eddy acceleration in
Eq.~(\ref{barotropic-zonal-mean-momentum}) and the vorticity
sources/sinks can be made in two steps. First, we note that
the definition of vorticity implies that $\overline{v'\zeta'}
= - \partial(\overline{u'v'})/\partial y$. Second, we multiply
the linearized version of Eq.~(\ref{barotropic-vorticity-eq2})
by $\zeta'$ and zonally average. This leads to an equation
for the budget of the so-called ``pseudomomentum''
\citep[][p.~493]{vallis-2006}:
\begin{equation}
{\partial{\cal A}\over\partial t} + \overline{v'\zeta'} =
{\overline{\zeta'F'}\over
2(\beta-{\partial^2 \overline{u}\over\partial y^2})}.
\label{pseudomomentum}
\end{equation}
For the two-dimensional non-divergent model,
${\cal A}=(\beta - \partial^2 \overline{u}/\partial y^2)^{-1}
\overline{\zeta'^2}/2$ is the pseudomomentum, which is a measure
of wave activity. By combining
Eqs.~(\ref{barotropic-zonal-mean-momentum})--(\ref{pseudomomentum})
and supposing that the wave amplitudes and zonal-mean zonal wind
are statistically steady, i.e., $\partial {\cal A}/\partial t \approx 0$ and
$\partial\overline{u}/\partial t \approx 0$, we obtain
\begin{equation}
{\overline{u}\over\tau_{\rm drag}}= {\overline{\zeta'F'}\over
2(\beta-{\partial^2 \overline{u}\over\partial y^2})}.
\label{barotropic-steady}
\end{equation}
This equation relates the vorticity sources/sinks and drag to
the zonal-mean zonal wind, $\overline{u}$.
When eddy sources/sinks of relative vorticity on average
exhibit the same sign as the vorticity itself (i.e. $\overline{\zeta'F'}>0$),
the eddy acceleration is eastward, and in steady state results in an
eastward zonal-mean zonal wind. When sources/sink of relative vorticity tend to exhibit
the opposite sign as the vorticity ($\overline{\zeta'F'}<0$), the eddy
acceleration is westward, and in steady state results in westward
zonal-mean zonal wind.\footnote{These arguments assume that
$\beta-\partial^2\overline{u}/\partial y^2 > 0$, which is generally the case.}
In analogy with the free solutions, this behavior is typically interpreted in
terms of the generation, latitudinal propagation, and dissipation
of Rossby waves.
This mechanism is thought to be responsible for the eddy-driven
jet streams (and the associated eastward surface winds) in
Earth's midlatitudes: baroclinic instability
generates Rossby waves that radiate away from the midlatitudes,
causing eastward eddy acceleration there and leading to eastward
surface flow in steady state \citep{held-2000, vallis-2006}. At
the equator, Earth's troposphere does not superrotate;
nevertheless, idealized Earth general circulation models (GCMs) have
shown that the presence of strong zonally varying heating and cooling in the tropics
can cause equatorial superrotation to emerge \citep{suarez-duffy-1992,
saravanan-1993, kraucunas-hartmann-2005, norton-2006}. In analogy
with the theory described above, \citet{held-1999b} suggested heuristically that the
superrotation in these models results from the generation and poleward
propagation of Rossby waves by the tropical heating sources.
Still, application of this barotropic theory to tidally locked
exoplanets is problematic.
Most models of atmospheric circulation on synchronously
rotating, zero-eccentricity hot Jupiters exhibit relatively steady
circulation patterns whose velocity and temperature patterns are
approximately symmetric about the equator \citep{showman-guillot-2002,
cooper-showman-2005, cooper-showman-2006, showman-etal-2008a,
showman-etal-2009, dobbs-dixon-lin-2008, rauscher-menou-2010}.
In such models, the relative vorticity is approximately antisymmetric
about, and zero at, the equator.
Eq.~(\ref{barotropic-steady}) predicts that $\overline{u}=0$ at the
equator for this situation. Thus, while attractive, this theory fails
to explain the equatorial superrotation in these hot Jupiter models.
\begin{figure}
\vskip 10pt
\includegraphics[scale=0.5, angle=0]{f1.pdf}
\caption{Illustration of the dynamical mechanism for generating equatorial
superrotation on tidally locked short-period exoplanets, including hot
Jupiters and super Earths. The intense day-night heating gradient
generates standing, planetary-scale Rossby and Kelvin waves. These
waves develop a structure with velocities tilting northwest-to-southeast
in the northern hemisphere and southwest-to-northeast in the southern
hemisphere (yellow and red ovals). In turn, these patterns
transport eddy momentum from high latitudes to the equator (dashed arrows).
Equatorial superrotation therefore emerges (thick, right-pointing arrow). }
\label{qualitative-scenario}
\end{figure}
There are other difficulties. First, the theory presented here assumes
the flow is barotropic---i.e. that it can
be described by Eq.~(\ref{barotropic-vorticity-eq})---and thereby
ignores the role of finite Rossby deformation radius in shaping
the wave properties. Second, vertical motions normally accompany
regions of heating and cooling, and these vertical motions lead
to nonzero horizontal divergence. The
low-latitude flow is thus inherently divergent in the presence
of heating/cooling, in conflict with the stated assumptions.
Moreover, the theory cannot account in any rigorous way
for the generation of the Rossby waves by thermal forcing. The above
equations include
no thermodynamics and only the effect of such forcing in producing
relative vorticity is represented. Away from the equator, one expects
that heating produces rising motion and horizontal divergence aloft, which
generates anticylonic relative vorticity by the action of Coriolis
forces. At the equator, however, this vorticity source is small,
leaving unclear the applicability of this barotropic theory in
producing eastward flow {\it at the equator}.\footnote{Within
the context of barotropic theory, one can relax the nondivergent assumption
by resolving the horizontal velocity into rotational and divergent
components, specifying the horizontal divergence field, and then
solving Eq.~(\ref{barotropic-vorticity-eq2}) for the rotational component
of the flow \citep[see][]{sardeshmukh-hoskins-1985, sardeshmukh-hoskins-1988}.
In this context, the specified divergence field represents the spatial field
of heating and cooling. The vorticity source in
Eq.~(\ref{barotropic-vorticity-eq2}) would then be $F=-(\zeta + f)
\nabla\cdot{\bf v_{\chi}} - {\bf v_{\chi}}\cdot\nabla(\zeta +f)$, where
${\bf v_{\chi}}$ is the specified divergent component of the flow field
and the velocity and vorticity on the lefthand side of
Eq.~(\ref{barotropic-vorticity-eq2}) represent only the rotational
component. One can then rework
Eqs.~(\ref{barotropic-zonal-mean-momentum})--(\ref{barotropic-steady})
under the usual barotropic assumptions that vertical momentum/vorticity
transport and vortex tilting
are negligible. However, such a
divergent barotropic model suffers from the same
failing as the nondivergent version: when the pattern of heating/cooling
and flowfield are mirror symmetric about the equator---as in most hot-Jupiter
models---the vorticity source $F$ is zero at the equator,
leading to a Rossby-wave source
$\overline{\zeta'F'}$ that is likewise zero at the equator. In
the presence of frictional drag, one again obtains
the result that $\overline{u}=0$
at the equator. Thus, even such an extended barotropic treatment is
insufficient to explain the equatorial
superrotation in hot-Jupiter models.}
When the geostrophic assumption is relaxed and finite deformation
radius is included,
analytic solutions of freely propagating equatorial waves show
that such waves tend to be trapped within an equatorial waveguide
and cannot propagate away from the equator \citep{holton-2004,
andrews-etal-1987}; such wave solutions
involve no net meridional (north-south) momentum transports, thus
raising the question of whether the above mechanism is viable at the
equator in the presence of finite deformation radius. Finally,
for hot Jupiters, the Rossby waves are expected to be global in scale,
and it is not clear {\it a priori} whether there is {\it room} for them
to propagate poleward from the equatorial regions.
We present a sequence of models in the following sections that
overcome these obstacles and provide a theoretical foundation
for understanding equatorial superrotation on tidally locked
exoplanets.
\section{Shallow-water model of equatorial superrotation}
\label{sw}
Full GCM solutions, although useful, involve so many interacting
processes that it is often difficult to cleanly identify specific
dynamical mechanisms from such solutions \citep[see, e.g.,][]
{showman-etal-2010}. Simplified models therefore play an important
role in atmospheric dynamics. Here, we present a highly idealized
model intended to capture the mechanism in the simplest possible context.
We adopt a two-layer model, with constant densities in each
layer, where the upper layer represents the meteorologically active
atmosphere and the lower layer represents the quiescent deep atmosphere and
interior. In the limit where the lower layer becomes infinitely
deep and the lower-layer winds and pressure gradients remain steady
in time (which requires the upper layer to be isostatically balanced),
this two-layer system reduced to the shallow-water equations
for the flow in the upper layer \citep[e.g.,][p.~129-130]{vallis-2006}:
\begin{equation}
{d{\bf v}\over dt}+g\nabla h + f{\bf k}\times {\bf v} = {\bf R}
- {{\bf v}\over{\tau_{\rm drag}}}
\label{momentum}
\end{equation}
\begin{equation}
{\partial h\over\partial t} + \nabla\cdot ({\bf v}h) =
{h_{\rm eq}(\lambda,\phi) - h\over\tau_{\rm rad}}
\equiv Q
\label{continuity}
\end{equation}
where ${\bf v}(\lambda,\phi,t)$ is horizontal velocity, $h(\lambda,\phi,t)$
is the upper layer thickness,
$t$ time, $g$ is the (reduced) gravity\footnote{$g$ in
Eq.~(\ref{momentum}) is the actual gravity times the fractional
density difference between the layers, and is therefore called the
``reduced gravity.'' When interpreting this shallow-water model
in the context of a three-dimensional atmosphere, this density
difference should be interpreted as (for example) the fractional
change in potential temperature across a scale height. This is of
order unity for a hot Jupiter.}, $f=2\Omega\sin\phi$
is Coriolis parameter, ${\bf k}$ is the upward unit vector,
$\Omega$ is planetary rotation rate, and $\phi$ is latitude.
Again, $d/dt=\partial/\partial t + {\bf v}\cdot\nabla$ is the
material derivative.
The boundary between the layers represents an atmospheric isentrope,
across which mass is exchanged in the presence of heating or cooling.
Heating and cooling are therefore represented in the shallow-water
system using mass sources and sinks, represented here as a Newtonian
relaxation of the height toward a specified radiative-equilibrium height
$h_{\rm eq}$---thick on the dayside and thin on the nightside---over
a radiative time scale $\tau_{\rm rad}$.
The momentum equations (\ref{momentum}) include drag with a timescale
$\tau_{\rm drag}$, which could represent the potentially
important effects of magnetohydrodynamic friction \citep{perna-etal-2010},
vertical turbulent mixing \citep[e.g.][]{li-goodman-2010}, or
momentum transport by breaking gravity waves \citep{lindzen-1981, watkins-cho-2010}.
The term ${\bf R}$ in Eq.~(\ref{momentum}) represents the effect on the
upper layer of momentum advection from the lower layer, and takes the
same form as in \citet{shell-held-2004} and \citet{showman-polvani-2010}:
\begin{equation}
{\bf R}(\lambda,\phi,t)=
\begin{cases}
-{Q{\bf v}\over h}, &Q>0;\\
0, &Q<0
\end{cases}
\label{R}
\end{equation}
where $\lambda$ and $\phi$ are longitude and latitude.
Air moving out of the upper layer $(Q<0)$ does not locally affect the upper
layer's specific angular momentum or wind speed, hence ${\bf R}=0$
for that case. But air transported into the upper layer carries
lower-layer momentum with it and thus alters the local specific
angular momentum and zonal wind in the upper layer. For the
simplest case where the lower layer winds are assumed to be zero, this
process preserves the column-integrated ${\bf v}h$ of the upper layer,
leading to the expression in (\ref{R}) for $Q>0$.
Importantly, the expression for ${\bf R}$ follows directly from
the momentum budget and contains no free parameters.\footnote{The
condition that air moving out of the upper layer induces no change in the
upper layer's specific momentum requires that such momentum advection
cause no accelerations, implying that ${\bf R}=0$ for $Q<0$.
To derive the expression for the case $Q>0$, add ${\bf v}$ times
the continuity equation to $h$ times the momentum equation,
thus yielding an equation for the time rate of change of ${\bf v}h$.
Terms involving heating/cooling are $h{\bf R} + {\bf v}Q$. In
the special case where the lower-layer winds are zero, mass transport
into the upper layer does not change the column-integrated horizontal
wind, ${\bf v}h$, of the upper layer. This
implies that $h{\bf R} + {\bf v}Q=0$, thus yielding Eq.~(\ref{R}).}
The relative roles of the two terms on the right side of
the momentum equation can be clarified by rewriting it
as follows:
\begin{equation}
{d{\bf v}\over dt}+g\nabla h + f{\bf k}\times {\bf v} =
-{\bf v}\left[{1\over{\tau_{\rm drag}}} +
{1\over \tau_{\rm rad}}\left({h_{\rm eq}-h\over h}\right)
{\cal H}(h_{\rm eq}-h)\right]
\label{momentum2}
\end{equation}
where ${\cal H}(h_{\rm eq}-h)$ is the Heaviside
step function, defined as 1 when $h_{\rm eq}-h>0$ and 0 otherwise.
Dynamically, it should be clear that both terms on the right side
play a role analogous
to drag; one can define the entire quantity in square braces as one
over an {\it effective} drag time constant. Still, the second term
(${\bf R}$) is spatially heterogeneous and only exists in regions of heating,
and we will show that its effect on the zonal-mean flow is qualitatively
different than that of the first term (frictional drag).
For a strongly irradiated
hot Jupiter, we might expect $(h_{\rm eq}-h)/h \sim 0.01$--1, and if so
then the first term would dominate if $\tau_{\rm drag} \ll \tau_{\rm rad}$
whereas the second term would dominate if $\tau_{\rm drag} \gg \tau_{\rm rad}$.
We present linear, analytic solutions and fully nonlinear, numerically
determined solutions of Eqs.~(\ref{momentum})--(\ref{R}) in the next two
subsections.
\subsection{Linear solutions}
\label{linear-sw}
To enable analytic solutions, we solve Eqs.~(\ref{momentum})--(\ref{R}) in
Cartesian geometry assuming that the Coriolis parameter can be
approximated as $f=\beta y$, where $y$ is northward distance from the
equator and $\beta$ (the gradient of Coriolis parameter with northward
distance) is assumed constant. This approximation, called the
``equatorial $\beta$-plane,'' is strictly valid only at low latitudes,
but we will see in \S\ref{nonlinear-sw} that the qualitative features
of these solutions are recovered by the full solutions in spherical
geometry.
We now linearize Eqs.~(\ref{momentum})--(\ref{R}) about a state of
rest. By definition, all the terms in
the linearized equations have magnitudes that scale with the forcing
amplitude. Note that the term ${\bf R}$
involves the product of the velocity and forcing amplitude,
and therefore is quadratic in the forcing amplitude and does
not appear in the linearized equations.
(We will come back to it when evaluating the implications of the
solutions for the zonal-mean zonal wind). The linearized
equations read
\begin{eqnarray}
\label{u-mom-linear}
{\partial u\over\partial t} + g {\partial \eta\over\partial x} -\beta y v
= -{u\over\tau_{\rm drag}} \\
\label{v-mom-linear}
{\partial v\over\partial t} + g{\partial\eta\over\partial y} + \beta y u
= -{v \over\tau_{\rm drag}} \\
\label{h-linear}
{\partial \eta\over\partial t} +
H\left({\partial u\over\partial x}+{\partial v\over\partial y}\right)
= S(x,y) - {\eta\over \tau_{\rm rad}}
\end{eqnarray}
where $x$ is eastward distance and $\eta$ is the deviation of the
thickness from its constant reference value $H$, such that $h=H + \eta$.
The quantity $S\equiv (h_{\rm eq}-H)/\tau_{\rm rad}$ is the forcing,
which can also be expressed as $S\equiv \eta_{\rm eq}/\tau_{\rm rad}$,
where $\eta_{\rm eq}\equiv h_{\rm eq}-H$ is the deviation of the
radiative-equilibrium height from $H$.
The free solutions to these equations (i.e., when
the right-hand sides are set to zero) are the well-known equatorially
trapped wave modes, described for example in \citet{holton-2004}
and \citet{andrews-etal-1987}. Given the intense heating and cooling
experienced by hot, tidally locked exoplanets, however,
we seek solutions to the forced problem.
Most three-dimensional dynamical models of hot Jupiters exhibit
relatively steady circulation patterns \citep{showman-guillot-2002,
cooper-showman-2005, cooper-showman-2006, dobbs-dixon-lin-2008,
dobbs-dixon-etal-2010, showman-etal-2008a, showman-etal-2009,
rauscher-menou-2010}, and so we seek steady solutions in the
presence of forcing and damping.
We nondimensionalize Eqs.~(\ref{u-mom-linear})--(\ref{h-linear}) with
a length scale $L=(\sqrt{gH}/\beta)^{1/2}$, a velocity scale
$U=\sqrt{gH}$, and a time scale ${\cal T}=(\sqrt{gH}\beta)^{-1/2}$,
which correspond respectively to the equatorial Rossby deformation
radius, the gravity wave speed, and the time for a gravity wave
to cross a deformation radius in the shallow-water system. The
thickness is nondimensionalized with $H$, the drag and thermal
time constants with ${\cal T}$, and the forcing with
$H/{\cal T}$. This yields, for steady flows,
\begin{eqnarray}
\label{u-mom-nd}
{\partial \eta\over\partial x} -yv = - {u\over\tau_{\rm drag}}\\
\label{v-mom-nd}
{\partial\eta\over\partial y} + yu = - {v\over\tau_{\rm drag}}\\
\label{h-nd}
\left({\partial u\over\partial x} + {\partial v\over\partial y}\right)
= S(x,y) - {\eta\over\tau_{\rm rad}}
\end{eqnarray}
where all quantities, including $\tau_{\rm rad}$ and $\tau_{\rm drag}$,
are now nondimensional.
In pioneering investigations, \citet{matsuno-1966} and \citet{gill-1980}
obtained analytic solutions to Eqs.~(\ref{u-mom-nd})--(\ref{h-nd})
for the special case where the drag and
radiative time constants are equal and drag is
neglected from the meridional momentum equation (\ref{v-mom-nd});
the latter assumption is called the ``longwave approximation'' because it
is valid in the limit where zonal length scales greatly exceed
meridional ones. On tidally locked exoplanets, however, the drag and radiative
time scales can differ greatly, and the longwave approximation may
not apply, because the flow exhibits comparable zonal and meridional scales.
We therefore retain the full form of Eqs.~(\ref{u-mom-nd})--(\ref{h-nd}).
Equations~(\ref{u-mom-nd})--(\ref{h-nd}) can be combined to
yield a single differential equation for the meridional velocity
$v$ \citep[e.g.,][]{wu-etal-2001}:
\begin{eqnarray}
\nonumber
{1\over\tau_{\rm drag}}
\left({\partial^2v \over\partial y^2} + {\partial^2 v\over\partial x^2}\right) +
{\partial v\over\partial x} - {1\over\tau_{\rm rad}}\left(y^2 + {1\over\tau_{\rm drag}^2}\right)v=\\
\left(-y{\partial S\over\partial x}
+{1\over\tau_{\rm drag}}{\partial S\over\partial y}\right).
\label{v-equation}
\end{eqnarray}
If one seeks separable solutions, then, as described by
\citet{gill-1980} and \citet{wu-etal-2001},
the meridional structure of the solutions to this equation with finite
$\tau_{\rm drag}$ are
the parabolic cylinder functions $\psi_n(y)$, which are simply Gaussians
times Hermite polynomials\footnote{The first few Hermite polynomials are
$H_0(\xi)=1$, $H_1(\xi)=2\xi$, $H_2(\xi)=4\xi^2 -2$, and
$H_3(\xi)=8\xi^3-12\xi$.}:
\begin{equation}
\label{parabolic-cylinder-functions}
\psi_n(y)=\exp\left(-{y^2\over 2 {\cal P}^2}\right) H_n\left({y\over {\cal P}}\right)
\end{equation}
where ${\cal P}\equiv (\tau_{\rm rad}/\tau_{\rm drag})^{1/4}$ is the fourth root of
a Prandtl number.
Our goal is to specify the thermal forcing, $S(x,y)$, and solve for the
unknowns $u$, $v$, and $\eta$.
In general, any desired pattern of thermal forcing can be represented
as
\begin{equation}
S(x,y) = \sum_{n=0}^{\infty} S_n(x) \psi_n(y).
\end{equation}
For tidally locked exoplanets, we expect this pattern to consist
of a day-night variation in heating/cooling whose amplitude peaks
at low latitudes and diminishes near the poles. We take the
forcing to be symmetric about the equator (appropriate for
a planet with zero obliquity) and, to keep the mathematics tractable,
retain solely the term $S_0$, corresponding to pattern of heating
and cooling that is a Gaussian, centered about the equator,
with a latitudinal half-width of the equatorial
Rossby radius of deformation modified by frictional and radiative
effects. While the full solution would require consideration
of $S_n$ for all $n\ge0$, the first term, $S_0$, will be the dominant
term for cases where the deformation radius is similar to a planetary
radius, as is the case on typical hot Jupiters.
Consideration of this term alone will therefore suffice to illustrate the
qualitative features relevant for inducing an equatorially
superrotating jet on tidally locked exoplanets.
Appendix \ref{gill-solutions} describes the solution method of
Eqs.~(\ref{u-mom-nd})--(\ref{h-nd}) and presents the
solution for the specific case where the forcing consists solely
of the $S_0$ term varying sinusoidally in longitude, i.e.,
$S(x,y)=\hat S_0 e^{ikx}\psi_0(y)$, where $\hat S_0$ is a constant.
Figure~\ref{matsuno}
shows an example for parameter values typical of a hot Jupiter or
hot super Earth (zonal wavelengths associated with the day-night heating
contrast of a planetary circumference and
radiative time constants of order $10^5\rm\,sec$). For this example,
the drag time constant is taken equal to the radiative time constant.
Figure~\ref{matsuno}a shows the radiative-equilibrium height field and
Fig.~\ref{matsuno}b presents the steady state height and velocity fields.
\begin{figure}
\vskip 10pt
\includegraphics[scale=0.6, angle=0]{f2.pdf}
\caption{An example linear, analytic solution for parameters relevant to
hot, tidally locked exoplanets. ({\it a})
Spatial structure of radiative-equilibrium height field, $h_{\rm eq}$
(orangescale and contours).
({\it b}) Height field (orangescale) and horizontal wind velocities
(arrows) for the linear, analytic solution forced by relaxation to the
$h_{\rm eq}$ profile shown in panel ({\it a}) and with
nondimensional zonal wavenumber $k=0.5$ and nondimensional radiative
and drag times $\tau_{\rm rad}=\tau_{\rm drag}=5$. For a hot Jupiter
or hot super Earth,
these correspond to dimensional zonal wavelengths of a planetary
circumference and dimensional radiative and drag time constants
of $\sim$$10^5\rm\,sec$ (see Appendix~\ref{nondim}). In ({\it a}) and
({\it b}), the horizontal and vertical axes are dimensionless
eastward and northward distance, respectively; one unit of distance
corresponds to a dimensional distance of one Rossby deformation radius,
$(\sqrt{gH}/\beta)^{1/2}$. The $\times$ marks the longitude along
the equator where $h$ reaches a maximum and the eddy zonal wind
changes sign. ({\it c}) Zonal (east-west)
accelerations of the zonal-mean flow implied by the linear solution.
The black and dark blue curves give the accelerations due to
horizontal and vertical eddy transport (terms II and III, respectively,
in Eq.~\ref{tem}). The light-green and cyan curves show friction (term IV)
and the effect of the mean-meridional circulation (term I), respectively.
The red curve shows the sum of all terms. The numerical
values adopt a forcing amplitude $\Delta h_{\rm eq}/H=1$. For
this value, the nondimensional peak winds approach 0.5, corresponding
to speeds of $\sim$$1\rm \, km\rm \, s^{-1}$ on a hot Jupiter.}
\label{matsuno}
\end{figure}
The solutions exhibit several important features. Although
the radiative-equilibrium height field
is symmetric in longitude about the substellar point (Fig.~\ref{matsuno}a),
the actual height field deviates significantly from
radiative equilibrium and exhibits considerable dynamical structure
(Fig.~\ref{matsuno}b). Two fundamental types of behavior are
present. First, at mid-to-high latitudes
($|y|\sim1$--3 in the figure), the flow exhibits vortical behavior.
The dayside contains an anticyclone in each hemisphere, manifesting
as a pressure high (i.e., local maximum of the height) around which winds flow
clockwise in the northern hemisphere and counterclockwise in the
southern hemisphere; the nightside contains a cyclone
in each hemisphere, manifesting as a pressure low around which winds
flow counterclockwise in the northern hemisphere and clockwise
in the southern hemisphere. Second, at low latitudes ($|y| \lesssim 1$),
the flows are nearly east-west; they diverge from a point east of the
substellar longitude (marked with a cross in Fig.~\ref{matsuno}b)
and converge toward a point east of the antistellar
longitude.
As discussed by \citet{gill-1980}, these features can be interpreted
in terms of forced, damped, steady equatorial wave modes. The mid-to-high
latitude feature described above is dynamically analogous to that of
an $n=1$ equatorially trapped Rossby wave, which
exhibits cyclones and anticyclones---alternating in longitude---that
peak off the equator \citep[see][Fig.~4c for an example of the flowfield
in this mode]{matsuno-1966}. The low-latitude feature discussed above
is dynamically analogous to a superposition of the $n=1$ Rossby
wave and the equatorial Kelvin wave, which is a fundamental equatorially
trapped wave mode with strong zonal winds but very weak meridional winds and
whose amplitude is symmetric about, and peaks at, the equator
(see, e.g., \citet{holton-2004} or \citet{andrews-etal-1987}).
Both of these wave modes exhibit winds that are primarily
east-west at the equator; in the example shown in Fig.~\ref{matsuno},
the Kelvin component dominates over the $n=1$ Rossby component
at the equator.
\begin{figure*}
\vskip -15pt
\includegraphics[scale=0.9, angle=0]{f3.pdf}
\caption{Analytic solutions of linearized shallow-water equations
(Eqs.~\ref{u-mom-nd}--\ref{h-nd}), as presented in
Appendix~\ref{gill-solutions}, for dimensionless zonal
wavenumber $k=0.5$ and dimensionless radiative and drag
time constants of 1, 10, and 100, corresponding roughly
to dimensional time constants of 6 hours, 3 Earth days,
and one month for parameters appropriate to hot Jupiters
(see Appendix~\ref{nondim}). All of these cases are forced by Newtonian
relaxation of the height field toward a distribution analogous
to that in Fig.~\ref{matsuno}a.}
\label{cases}
\end{figure*}
Equations~(\ref{u-mom-nd})--(\ref{h-nd}) indicate that this is
a problem governed by two parameters: the radiative time constant and
the drag time constant. We now examine how the behavior depends on
their values. Figure~\ref{cases} shows linear
solutions, as presented in Appendix~\ref{gill-solutions},
for dimensionless radiative
time constants of 1, 10, and 100 (top, middle, and bottom rows, respectively)
and drag time constants of 1, 10 and 100 (left, middle, and right
columns, respectively). For parameters appropriate to hot Jupiters
(rotation period of 3 Earth days and $gH\approx 4\times10^6\rm \,m^2\rm \, s^{-2}$),
these dimensionless values correspond to dimensional
time constants of $\sim$$3\times10^4$, $3\times10^5$, and
$3\times10^6\rm\,sec$, respectively (see Appendix~\ref{nondim}).
When the radiative and drag time constants are short (upper left
corner of Fig.~\ref{cases}), the maximum and minimum thermal (height)
perturbations lie on the equator and are close to the substellar
and antistellar points; in this limit, the height field is close
to radiative equilibrium (compare top left of Fig.~\ref{cases}
with Fig.~\ref{matsuno}a)\footnote{Appendix~\ref{gill-solutions}
demonstrates formally that, in the limit of either time constant
going to zero, the height field converges to the radiative-equilibrium
height field.}, and distinct Rossby-wave gyres do not appear.
When the radiative and drag time constants have intermediate
values (middle of Fig.~\ref{cases}), cyclones and anticyclones
become visible and---as in Fig.~\ref{matsuno}b---exhibit height
extrema that are phase shifted westward of the extrema in radiative
equilibrium. Similarly, the height extrema along the equator
become phase shifted eastward relative to radiative equilibrium;
thickness variations along the equator become modest relative
to those in midlatitudes. When the radiative and drag time constants are
long (lower right corner of Fig.~\ref{cases}), the height field
becomes dominated by the off-equatorial anticyclones and cyclones,
with minimal variation of height at the equator.
In the limits $\tau_{\rm rad}\to\infty$ and $\tau_{\rm drag}\to\infty$,
the solution becomes flat at the equator and is
symmetric in longitude about the $x=0$ axis
(a point demonstrated explicitly in Appendix~\ref{gill-nodrag});
Fig.~\ref{cases} shows that this limit is almost reached even for
$\tau_{\rm rad}$ and $\tau_{\rm drag}$ of 100.
Much of the behavior in Fig.~\ref{cases} can be understood
in terms of the zonal propagation of equatorially trapped
Rossby and Kelvin modes. Kelvin waves exhibit eastward group
propagation while long-wavelength, equatorially trapped
Rossby waves exhibit westward group propagation. When $\tau_{\rm rad}$
and $\tau_{\rm drag}$ are very short (upper left corner of Fig.~\ref{cases}),
the damping is so strong that the waves are unable to propagate zonally.
As a result, the height is close to the radiative equilibrium height
field. When the two time constants have intermediate
values, the propagation produces an eastward phase
shift of the height field at the equator (the Kelvin component) and a
westward phase shift of the height field in the off-equatorial cyclones
and anticyclones (the Rossby component)---exactly as seen in Fig.~\ref{matsuno}b
and the middle of Fig.~\ref{cases}. As the two time constants become
very long, the westward phase offset of the off-equatorial cyclones
and anticyclones achieves maximal values of $90^{\circ}$. At the equator,
however, the height variations go to zero; this is explained by the
fact that Coriolis forces are zero at the equator, so the linearized
force balance is between pressure-gradient forces and drag. Weak
drag requires weak pressure-gradient forces and hence a flat layer
at the equator.
Now, the key point of our paper is that these linear solutions
have major implications for the development of equatorial superrotation
on tidally locked exoplanets. As can be seen in Figs.~\ref{matsuno}b and
\ref{cases}, the wind vectors exhibit an
overall tilt from northwest-to-southeast in the northern hemisphere
and southwest-to-northeast in the southern hemisphere. This pattern,
which resembles a chevron centered at the equator and pointing east, is
particularly strong when the radiative and drag time constants are
short, but occurs in all the cases shown. This structure implies
that, on average, equatorward moving air has faster-than-average eastward
wind speed while poleward moving air has slower-than-average eastward
wind speed, so that
$\overline{u'v'} < 0$ in the northern hemisphere and $\overline{u'v'} > 0$
in the southern hemisphere. As shown schematically in
Fig.~\ref{qualitative-scenario}, this is exactly the
type of pattern that causes a flux of eastward eddy momentum
to the equator and can induce equatorial
superrotation. Since momentum is being removed from the mid-latitudes,
one would expect westward zonal-mean flow to develop there.
The physical mechanism responsible for producing these phase
tilts are twofold. First, the differential wave propagation discussed
above: this propagation causes
an eastward phase shift of the height field in the Kelvin waves
and a westward shift of the height field in the Rossby waves
relative to the radiative-equilibrium height field.
Because the Rossby wave lies on the poleward flanks of the
Kelvin wave, the result is a chevron pattern
where the height contours tilt northwest-southeast
in the northern hemisphere and southwest-northeast in the
southern hemisphere. To the extent that velocity vectors
approximately parallel the geopotential contours (as they
tend to do away from the equator when drag is weak or moderate),
this will generate tilts in the velocities such
that $\overline{u'v'}<0$ in the northern hemisphere and
$\overline{u'v'}>0$ in the southern hemisphere.
The second mechanism for generating the velocity tilts
needed for equatorial superrotation is simply the three-way
force balance between Coriolis, drag, and pressure-gradient
forces. Because drag acts opposite to the velocity, and
Coriolis forces are perpendicular to the velocity, this
three-way force balance requires the velocities to be
rotated clockwise of $-\nabla\eta$ in the northern hemisphere
and counterclockwise of $-\nabla\eta$ in the southern hemisphere.
Given the expected day-night gradients in $\eta$, this
balance implies that the velocities will tend to tilt
northwest-southeast in the northern hemisphere and southwest-northeast
in the northern hemisphere. We demonstrate this fact
explicitly with an analytic solution in the limit of $\tau_{\rm rad}\to0$
in Appendix~\ref{zero-taurad}; even when the height field is
nearly in radiative equilibrium and hence exhibits no overall phase
tilts, the velocities themselves develop tilts such that
$\overline{u'v'}<0$ in the northern hemisphere and $\overline{u'v'}>0$
in the southern hemisphere (see Fig.~\ref{analyt-zero-taurad}).
The calculation in the limit $\tau_{\rm rad}\to0$ is particularly
interesting because, in this limit, there is no zonal propagation
of the Kelvin and Rossby waves: the radiative damping is infinitely
strong and the zonal phase shift of the height field (relative
to radiative equilibrium) is zero. This is
the dominant mechanism for the velocity tilts
in the top-left panel of Fig.~\ref{cases}.
To demonstrate explicitly how superrotation would emerge from
these standing-wave patterns, we analyze the zonal
accelerations associated with these linear solutions.
Decomposing variables into their zonal means
(denoted by overbars) and deviations therefrom (denoted with primes)
and zonally averaging the zonal-momentum equation (Eq.~\ref{momentum})
leads to \citep[e.g.,][]{thuburn-lagneau-1999}
\begin{eqnarray}
\nonumber
{\partial\overline{u}\over\partial t}=\underbrace{\overline{v}^*\left[f -
{\partial \overline{u}\over\partial y}\right]}_{I}
\nonumber
\underbrace{-{1\over \overline{h}}{\partial\over\partial y}
[\overline{(hv)'u'}]}_{II}\\
+ \underbrace{\left[{1\over\overline{h}}
\overline{u'Q'} + \overline{R_u}^*\right]}_{III}
\underbrace{-{\overline{u}^*\over \tau_{\rm drag}}}_{IV}
- {1\over\overline{h}}{\partial(\overline{h'u'})\over\partial t}
\label{tem}
\end{eqnarray}
where $a$ is the planetary radius and $\overline{A}^*\equiv \overline{hA}/
\overline{h}$ denotes the thickness-weighted zonal average of any quantity $A$.
Eq.~(\ref{tem}) is the shallow-water version of the
Transformed Eulerian Mean (TEM) momentum equation, analogous to that in
the isentropic-coordinate form of the primitive equations
\citep[see][Section 3.9]{andrews-etal-1987}. On the right-hand side,
terms I, II, and III represent accelerations due to (i)
momentum advection by the mean-meridional circulation,
(ii) the convergence of the meridional flux of zonal eddy momentum, and (iii)
correlations between the regions of eddy zonal flow and eddy mass source
(essentially vertical eddy-momentum transport). Within this term,
the quantity $R_u$ is the zonal component of ${\bf R}$
(equal to $-Qu/h$ when $Q>0$ and 0 when $Q<0$). Term IV is
frictional drag. The final term represents the time rate of
change of the eddy momentum. In the linear limit,
all the terms on the right side of Eq.~(\ref{tem}) have vanishingly
small amplitude and, in this case, the solutions in
Figs.~\ref{matsuno}--\ref{cases}
represent true steady states. At any finite amplitude, however,
terms I--IV are nonzero and would cause generation of a zonal-mean
zonal flow.
Figure~\ref{matsuno}c depicts these terms for the example solution
presented in Fig.~\ref{matsuno}b. As expected, horizontal convergence
of eddy momentum, term II, causes a strong eastward acceleration at the
equator and westward acceleration in the midlatitudes ({\it black
curve}). On the other hand, the acceleration associated with
vertical eddy-momentum transport, term III, is strong and westward
at the equator ({\it blue}), implying downward transport of eddy
momentum at the equator. The remaining terms---the mean-meridional
circulation (term I, {\it cyan}) and mass-weighted friction
(term IV, {\it light green})---are small at the equator.
The two eddy terms partially
cancel at the equator, but the acceleration due to horizontal eddy
momentum convergences exceeds that due to vertical eddy momentum
convergences, leading to a net eastward acceleration at the equator
and westward acceleration in midlatitudes ({\it red curve}).
Remarkably, despite the wide
range of morphologies that occur when $\tau_{\rm rad}$ and
$\tau_{\rm drag}$ are varied (Fig.~\ref{cases}), all the solutions exhibit
an equatorward flux of eddy momentum and a net eastward acceleration
at the equator. This is shown in Fig.~\ref{cases-accel}, which
presents the two eddy acceleration terms from Eq.~(\ref{tem}) for each of
the cases shown in Fig.~\ref{cases}.
These solutions therefore suggest that superrotation
at the equator and westward mean flow in the midlatitudes should
occur at essentially any value of the control parameters.
\begin{figure*}
\vskip 10pt
\includegraphics[scale=0.9, angle=0]{f4.pdf}
\caption{Accelerations of the zonal-mean zonal flow for the same
solutions as in Fig.~\ref{cases}. Black curve is acceleration
due to horizontal eddy-momentum convergence (term II in Eq.~\ref{tem}), and
blue curve is acceleration due to vertical eddy-momentum
transport (term III). Red curve is sum of terms I, II, III, and IV.
Numerical values plotted here assume that $\hat S_0/\tau_{\rm rad}$ (which
is essentially equivalent to $\Delta h_{\rm eq}/H$) equals one.}
\label{cases-accel}
\end{figure*}
The patterns of spatial velocity and mass source/sink illuminate
the physical origin of the westward equatorial acceleration caused by the
vertical eddy exchange. The solutions show that the
longitudes of zero zonal wind at the equator lie east of the
mass-source extrema (Fig.~\ref{matsuno}b), a feature also clearly visible
in the steady, linear calculations of \citet[][Fig.~9]{matsuno-1966}
and \citet[][Fig.~1]{gill-1980}. Because of this shift,
equatorial mass sources (sinks) occur predominantly in regions of
westward (eastward) eddy zonal flow. On average, therefore,
the mass sink regions transport air with eastward column-integrated
eddy momentum out of the layer. The mass source regions
transport air with no relative zonal momentum from the quiescent
abyssal layer into the upper layer; this process conserves the
local, column-integrated relative momentum ${\bf v}h$ of the upper layer.
Thus, when zonally averaged, vertical exchange at the equator removes
momentum from the layer, leading to $(\overline{u'Q'} + \overline{R_u}^*)<0$
and contributing a westward acceleration (blue curve in Fig.~\ref{matsuno}c).
The above argument, however, does not determine which of
the two eddy terms (II and III in Eq.~\ref{tem}) dominates.
To determine which is larger---and hence whether the net equatorial
eddy acceleration is eastward or westward---we write the zonally
averaged zonal momentum equation in the form
\begin{equation}
{\partial\overline{u}\over \partial t}=\overline{v'\zeta'} +
\overline{v}(f + \overline{\zeta}) - {\overline{u}\over\tau_{\rm drag}}
+ \overline{R_u}.
\label{eulerian-mean}
\end{equation}
where $\zeta$ is the relative vorticity.
For the case where the forcing is symmetric about the equator,
the solutions are symmetric about the equator in $u$ and $h$ but
antisymmetric about the equator in $v$ and $\zeta$. As a result, the meridional velocity
and relative vorticity are zero at the equator, so the terms
$\overline{v}(f+\zeta)$ and $\overline{v'\zeta'}$ vanish there.
Therefore,
\begin{equation}
{\partial\overline{u}\over\partial t}=-{\overline{u}\over\tau_{\rm drag}}
+\overline{R_u}\qquad\qquad {\rm at}\;y=0.
\label{balance}
\end{equation}
Essentially, at the equator,
$\overline{R_u}$ is the mismatch between the accelerations
caused by horizontal and vertical eddy-momentum fluxes. The
analytic solutions, which assume $\overline{u}=0$,
show that $u$ is predominantly westward in regions
where $Q>0$, which therefore implies that $\overline{R_u}>0$.
From Eq.~(\ref{balance}), the net eddy-induced acceleration
is therefore eastward.
This explains, in a general way, the sign of the net eddy
accelerations
at the equator in Fig.~\ref{cases-accel}. Of course, once a
zonal-mean flow ($\overline{u}\neq 0$)
develops, the magnitude of $\overline{R_u}$ changes and
the friction term becomes important in Eq.~(\ref{balance}); eventually
these terms balance and allow a steady state to be achieved. We
discuss the possible steady states in light of this equation in
\S\ref{nonlinear-sw}.
We have so far emphasized the {\it spatial patterns} of the circulation,
but it is also interesting to examine the {\it magnitudes} of
the velocities predicted by our linear solutions. When the day-night
difference in the
radiative-equilibrium height is comparable to the mean value
and the radiative time constant is a few days or less (as expected for
the strongly forced conditions on hot Jupiters),
the winds shown in Fig.~\ref{matsuno}b and Fig.~\ref{cases}
reach nondimensional speeds of order unity. For a hot Jupiter,
with typical $g=20\rm \,m\rm \, s^{-2}$ and $H=200\rm \, km$, this corresponds to dimensional
speeds of $\sim$$2\rm\,km\,sec^{-1}$.
To within a factor of a few, this is similar to the speeds obtained
in fully nonlinear three-dimensional atmospheric circulation models of
hot Jupiters \citep{showman-guillot-2002, cooper-showman-2005,
showman-etal-2008a, showman-etal-2009, dobbs-dixon-lin-2008,
dobbs-dixon-etal-2010, menou-rauscher-2009, rauscher-menou-2010,
thrastarson-cho-2010}. For a tidally locked, Earth-like
planet in the habitable zone of an M-dwarf, with $g=10\rm \,m\rm \, s^{-2}$,
$H=10\rm \, km$, and an Earth-like radiative time constant of $\sim$10 days
(corresponding to dimensionless time constants of 10--100), the solutions
then yield nondimensional speeds of $\sim0.02$--0.1. This corresponds
to dimensional speeds of up to a few tens of $\rm \,m\rm \, s^{-1}$, similar to
speeds obtained in models of tidally locked
terrestrial planets \citep{joshi-etal-1997, heng-vogt-2011,
merlis-schneider-2011}.
\subsection{Nonlinear solutions}
\label{nonlinear-sw}
Next, we relax the small-amplitude and Cartesian constraints to
demonstrate how nonlinearity and full spherical geometry
affect the solutions,
and we show how the wave-induced accelerations
interact with the mean flow to generate an equilibrated
state exhibiting equatorial superrotation. To do so, we solve
the fully nonlinear forms of Eqs.~(\ref{momentum})--(\ref{R})
in global, spherical geometry, using a radiative-equilibrium
thickness given by
\begin{equation}
h_{\rm eq}=H + \Delta h_{\rm eq}\cos\lambda\cos\phi
\label{heq}
\end{equation}
where $H$ is the mean thickness, $\Delta h_{\rm eq}$ is the day-night
contrast in radiative-equilibrium thickness, and the substellar
point is at longitude $0^{\circ}$ and latitude $0^{\circ}$. The
planet is assumed to be synchronously rotating, so that the pattern
of $h_{\rm eq}(\lambda,\phi)$ remains fixed in time. For concreteness,
we adopt planetary parameters appropriate to a hot Jupiter, although we
expect qualitatively similar solutions to apply to super Earths.
For a typical gravity of $20\rm\,m\,sec^{-2}$ and scale height of
$200\rm\,km$ appropriate to hot Jupiters, we might expect
$gH=4\times10^6\rm\,m^2\,sec^{-2}$, and we adopt this value for all
our runs. (Note that $g$ and $H$ do not need to be specified
independently.) We also take $\Omega=3.2\times10^{-5}\rm\,sec^{-1}$
and $a=8.2\times10^7\rm\,m$, corresponding to rotation period
and planetary radius of 2.3 Earth days and 1.15 Jupiter radii,
respectively, similar to the values for HD 189733b.
We reiterate that the equations represent a two layer system
with an active layer overlying a quiescent, infinitely deep
lower layer. Because of coupling between the layers (specifically,
mass exchange in the presence of heating/cooling), the solutions
readily reach a steady state for any value of the
drag time constant, including the limit where drag is excluded
entirely in the upper layer ($\tau_{\rm drag}\to\infty$).
This in fact is a simple
representation of the situation in many full 3D GCMs of Solar System
atmospheres, including Earth, which often have strong frictional drag near the
surface, little-to-no friction in the upper layers, and yet
easily reach a steady configuration throughout
{\it all} the model layers. In our case, we find that,
when drag is strong, the solutions reach steady states in
runtimes $\lesssim 10\tau_{\rm drag}$. In the case where
drag is turned off, the time to reach steady state is
determined by the magnitude of momentum and energy exchange
between the layers (e.g., by the magnitude of the ${\bf R}$ term),
and is generally $\lesssim 10\tau_{\rm rad}
|H/\Delta h|$, where $|\Delta h|/H$ is a characteristic
value of the fractional height variations in the active layer.
All solutions shown here are equilibrated and steady.
We solve Eqs.~(\ref{momentum})--(\ref{R}) using the Spectral Transform
Shallow Water Model (STSWM) of \citet{hack-jakob-1992}. Rather
than integrating the equations for $u$ and $v$, the code
solves the momentum equations in a vorticity-divergence form. The
initial condition is a flat layer of geopotential $gH$ at rest;
the equations are integrated using a spectral truncation of
T170, corresponding to a resolution of $0.7^{\circ}$ in longitude
and latitude (i.e., a global grid of $512\times256$ in longitude
and latitude). A $\nabla^6$ hyperviscosity is applied to each
of the dynamical variables to maintain numerical stability.
The code adopts the leapfrog timestepping scheme and applies
an Asselin filter at each timestep to suppress the computational
mode. These methods are standard practice; for further
details, the reader is referred to \citet{hack-jakob-1992}.
To facilitate comparison with the analytic theory in \S\ref{linear-sw},
we first describe the solutions at very low amplitude where the
behavior is linear. Figure~\ref{stswm-equal-tau} shows
the geopotential (i.e., $gh$) for equal radiative and drag time constants
of 0.1, 1, and 10 (Earth) days, respectively. Qualitatively,
the numerical solutions in spherical geometry bear a striking resemblance
to the analytic solutions on a $\beta$ plane. At time constants
of a fraction of day, the geopotential maxima occur on the equator, and
for time constants of 0.1 day ({\it a}), the geopotential
resembles the radiative-equilibrium solution, with wind
flowing from the substellar point to the antistellar point.
Longer time constants (1 day, panel {\it b}) allow zonal
energy propagation of the Kelvin and Rossby waves, leading to an
eastward phase shift of the geopotential at the equator
and a westward phase shift at high latitudes ($\sim$40--$90^{\circ}$).
The result is contours of geopotential that develop
northwest-southeast tilts in the northern hemisphere and
southwest-northeast tilts in the southern hemisphere.
When the time constants are
long (10 days, panel {\it c}) off-equatorial cyclones and
anticyclones dominate the geopotential, with only weak
geopotential variations along the equator. These vortices
are oval in shape, exhibiting no overall phase tilt, though
the regions close to the equator do develop phase tilts
(westward-poleward to easward-equatorward). The momentum
fluxes cause a prograde eddy acceleration (and superrotation)
at the equator for all these cases. All of these features are
also shared by the analytic solutions (Fig.~\ref{cases}).
\begin{figure}
\vskip 10pt
\includegraphics[scale=0.45, angle=0]{f5a.pdf}\\
\includegraphics[scale=0.45, angle=0]{f5b.pdf}\\
\includegraphics[scale=0.45, angle=0]{f5c.pdf}
\caption{Geopotential $gh$ (orangescale and contours) for the equilibrated
(steady-state) solutions
to the shallow-water equations (Eqs.~\ref{momentum}--\ref{R})
in full spherical geometry assuming equal radiative and drag
time constants of 0.1, 1, and 10 days in {\it (a), (b),}
and {\it (c)}, respectively. White is thick and dark is thin.
Although the equations solved are fully nonlinear, the
forcing amplitude is small here ($\Delta h_{\rm eq}/H=0.01$)
so that the solutions in these panels are essentially linear.}
\label{stswm-equal-tau}
\end{figure}
We now explore how the solutions change when the radiative and
frictional time scales are different.
In the linear limit, the latitudinal width of the region exhibiting
prograde phase tilts (i.e., northwest-to-southeast in the northern
hemisphere and southwest-to-northeast in the southern hemisphere)
contracts toward the equator when the drag time constant greatly
exceeds the radiative time constant. This is illustrated in
Fig.~\ref{stswm-vary-drag}, which shows the equilibrated (steady-state)
solutions for $\tau_{\rm rad}=1\rm\,day$ and $\tau_{\rm drag}/\tau_{\rm rad}
=1$
({\it top}), 10 ({\it middle}), and infinite ({\it bottom}).
When the time constants are equal, the entire northern (southern)
hemisphere exhibits northwest-to-southeast (southwest-to-northeast)
phase tilts. When $\tau_{\rm drag}/\tau_{\rm rad}=10$, these
phase tilts are confined within $\sim$$20^{\circ}$ latitude of
the equator, and for $\tau_{\rm drag}\to\infty$, the width shrinks toward
zero. This behavior is explained by the analytic
theory in \S\ref{linear-sw}. As shown in
Eq.~(\ref{parabolic-cylinder-functions}),
the parabolic cylinder functions comprising the latitudinal structure
exhibit a characteristic latitudinal width of $L(\tau_{\rm rad}/
\tau_{\rm drag})^{1/4}$, where $L=(\sqrt{gH}/\beta)^{1/2}$ is
the equatorial Rossby deformation radius; these functions thus
collapse toward the equator as $\tau_{\rm drag}/\tau_{\rm rad}$ becomes
infinite.\footnote{The numerical solutions show that the region of prograde
phase tilts does not become precisely zero as $\tau_{\rm drag}$ becomes
infinite because of the ${\bf R}$ term. As shown in Eq.~(\ref{momentum2}),
${\bf R}$ plays a role analogous to drag, and the {\it effective}
drag time constant (one over the quantity in square brackets in
Eq.~\ref{momentum2}) has a characteristic
magnitude $\tau_{\rm rad} h/(h_{\rm eq}-h)$. This suggests that,
in the absence of drag, the region of prograde phase tilts exhibits
a latitudinal width of order $L[(h_{\rm eq}-h)/h]^{1/4}$. This goes
to zero in the limit of zero amplitude but is nonzero at any finite
amplitude.} Poleward of this region, the solutions exhibit
phase tilts of the opposite direction
(northeast-to-southwest in the northern hemisphere
and southeast-to-northwest in the southern hemisphere).
Appendix~\ref{gill-nodrag} gives the explanation for this reversal in
phase tilts; the low-amplitude, full spherical numerical solutions
at $\tau_{\rm drag}/\tau_{\rm rad}\gg1$ strongly resemble analytic
solutions in the absence of drag, presented in Appendix~\ref{gill-nodrag}
(compare Figs.~\ref{stswm-vary-drag}c and \ref{analyt-nodrag}).
\begin{figure}
\vskip 10pt
\includegraphics[scale=0.45, angle=0]{f6a.pdf}\\
\includegraphics[scale=0.45, angle=0]{f6b.pdf}\\
\includegraphics[scale=0.45, angle=0]{f6c.pdf}
\caption{Geopotential $gh$ (orangescale and contours) for the equilibrated
(steady-state) solutions to the shallow-water equations
(Eqs.~\ref{momentum}--\ref{R}) in full spherical geometry illustrating
the effect of varying $\tau_{\rm drag}$ in the linear limit. All of the
cases depicted have $\tau_{\rm rad}=1\,\rm day$. The drag time constant
is 1 day, 10 days, and infinite in the top, middle, and bottom panels,
respectively. White is thick and dark is thin.
As in Fig.~\ref{stswm-equal-tau}, $\Delta h_{\rm eq}/H=0.01$ here.}
\label{stswm-vary-drag}
\end{figure}
Nonlinearity alters the solutions in several important ways,
which we illustrate in Fig.~\ref{stswm-nonlinearity}, showing
a sequence of solutions
for $\tau_{\rm rad}=0.1\rm\,day$, $\tau_{\rm drag}=10\rm\,days$
and, from top to bottom, amplitudes $\Delta h_{\rm eq}/H$ of 0.01,
0.1, and 0.5, respectively. This choice of time constants
is representative of the regime of strong radiative forcing
and weak drag that may be appropriate to typical hot Jupiters.
Increasing the forcing amplitude
(i.e., increasing $\Delta h_{\rm eq}/H$ while holding $\tau_{\rm rad}$
and $\tau_{\rm drag}$ constant) of course leads to increased wind speeds and
day-night geopotential variations; for the parameters in
Fig.~\ref{stswm-nonlinearity}, the zonal-mean zonal wind speed at
the equator ranges from $\sim$$10\rm\,m\,sec^{-1}$ at the lowest
amplitude to almost $1000\rm\,m\,sec^{-1}$ for the highest amplitude shown.
Moreover, beyond a critical value of $\Delta h_{\rm eq}/H$ (depending on
the values of $\tau_{\rm rad}$ and $\tau_{\rm drag}$), the solutions begin
to deviate {\it qualitatively} from the linear solutions.
First, nonlinearity allows greater geopotential variations to
occur along the equator, such that at extreme forcing amplitude
the geopotential extrema can in some cases occur
along the equator when they otherwise would not. In the linear
limit, the zonal force balance at the equator in the steady state
is between the pressure-gradient force and drag (cf Eq.~\ref{u-mom-linear});
therefore, when drag is weak, the pressure-gradient force must likewise
be small, implying that minimal variations of geopotential occur
along the equator. This restriction does not apply at higher latitudes
(where the Coriolis force can balance the pressure-gradient force),
so for very weak drag the thickness extrema generally occur off the
equator (as can be seen in the lower right portion of Fig.~\ref{cases};
Fig.~\ref{stswm-equal-tau}{\it b} and {\it c}; Fig.~\ref{stswm-vary-drag},
and Fig.~\ref{stswm-nonlinearity}{\it a}). At large forcing amplitude,
however, the momentum advection term ${\bf v}\cdot{\nabla h}$ and the
${\bf R}$ terms become
important and can balance the pressure-gradient force, allowing
significant zonal pressure gradients---and hence significant variations
in thickness---to occur along the equator. For the parameters
in Fig.~\ref{stswm-nonlinearity}, the thickness variations peak at
the equator when the forcing amplitude is sufficiently large ({\it bottom
panel}).
Second, at high amplitude, the phase tilts of wind and geopotential
tend to be from northwest-to-southeast (southwest-to-northeast) throughout
much of the northern (southern) hemisphere---as in
Fig.~\ref{stswm-nonlinearity}{\it b} and \ref{stswm-nonlinearity}{\it c}---
even when the phase tilts are in the opposite direction at low
amplitude (as in Fig.~\ref{stswm-nonlinearity}{\it a}). This effect can
be directly attributed to the term ${\bf R}$ in the momentum equations. As
shown in Eq.~(\ref{momentum2}), ${\bf R}$ plays a role analogous to
drag. When true drag is weak or absent, the {\it effective} drag time
constant (one over the quantity in square brackets in Eq.~\ref{momentum2})
has a characteristic magnitude $\tau_{\rm rad}h/(h_{\rm eq}-h)$. At
large forcing amplitude, $h/(h_{\rm eq}-h) \sim 1$, and in that case
the {\it effective} drag time constant is comparable to $\tau_{\rm rad}$.
The linear solutions show that prograde phase tilts dominate over
much of the globe when the radiative and drag time constants are comparable,
but when the drag time constant greatly exceeds the radiative time constant,
the phase tilts are in the opposite direction
(see Fig.~\ref{stswm-vary-drag}). In Fig.~\ref{stswm-nonlinearity},
$\tau_{\rm drag}/\tau_{\rm rad}=100$, but the ratio of the {\it effective}
drag time constant to the radiative time constant decreases from
top to bottom and reaches $\sim$1 in the bottom panel, explaining
the transition in the phase tilts from Fig.~\ref{stswm-nonlinearity}{\it a}
through \ref{stswm-nonlinearity}{\it c}. Through
the momentum fluxes that accompany these phase tilts, the equatorial
jet becomes broader and more dominant with increasing nonlinearity.
\begin{figure}
\vskip 10pt
\includegraphics[scale=0.53, angle=0]{f7a.pdf}\\
\includegraphics[scale=0.53, angle=0]{f7b.pdf}\\
\includegraphics[scale=0.53, angle=0]{f7c.pdf}
\caption{Geopotential $gh$ (orangescale and contours) for the equilibrated
solutions to the shallow-water equations
(Eqs.~\ref{momentum}--\ref{R}) in full spherical geometry illustrating
the effect of nonlinearity. All of the
cases depicted have $\tau_{\rm rad}=0.1\,$ day and $\tau_{\rm drag}=10\,$
days. The forcing amplitude is $\Delta h_{\rm eq}/H=0.01$ 0.1, and 0.5
in ({\it a}), ({\it b}), and ({\it c}), respectively.
White is thick and dark is thin. Range of $gh$ values is
3.97--$4.03\times10^6\rm \,m^2\rm \, s^{-2}$, 3.6--$4.3\times10^6\rm \,m^2\rm \, s^{-2}$, and
2.1--$5.5\times10^6\rm \,m^2\rm \, s^{-2}$ from top to bottom, respectively.
The solution in ({\it a}) is steady;
those in ({\it b}) and ({\it c}) exhibit small-scale eddy generation
and time variability (most evident in panel {\it b}), although the
global properties (total kinetic energy, potential energy, and equatorial
jet speed) are essentially constant in time.}
\label{stswm-nonlinearity}
\end{figure}
The momentum balance in the equatorial jet can achieve steady
state in two ways. In steady state, Eq.~(\ref{balance}) becomes
\begin{equation}
{\overline{u}\over\tau_{\rm drag}}=\overline{R_u} \qquad\qquad
({\rm at \;the\; equator}).
\label{steady-state}
\end{equation}
As described previously,
when the zonal-mean zonal winds are weak, the zonal wind $u$
is predominantly westward
in regions where $Q>0$, so that $\overline{R_u}>0$ (\S\ref{linear-sw}
and Fig.~\ref{cases-accel}). This implies an eastward eddy acceleration of
the zonal-mean zonal winds at the equator, which induces equatorial
superrotation. In the first type of
steady state, corresponding to a regime of strong friction (short
drag time constant), this superrotation implies a strong westward
acceleration due to friction ($-\overline{u}/\tau_{\rm drag}$). Steady
state occurs when the zonal-mean equatorial jet becomes strong enough for
the friction to balance the eastward eddy-induced acceleration at
the equator. We call this the ``high Prandtl number'' regime.
In the second type of steady state, which we call the ``low Prandtl number''
regime, the friction is
sufficiently weak that the term
$-\overline{u}/\tau_{\rm drag}$ is unimportant in the momentum balance.
Because of the eastward eddy acceleration, the zonal-mean zonal winds
can build to high speed. Once they do, they change the nature of
$\overline{R_u}$---the larger $\overline{u}$ becomes, the smaller the
extent to which $u<0$ in the region $Q>0$, as necessary
for $\overline{R_u}>0$. Eventually, for sufficiently
large $\overline{u}$, the quantity $\overline{R_u}$ goes to zero
at the equator. The equatorial jet thus achieves steady state.
Figure~\ref{stswm-diag}
shows examples of each of these regimes illustrating how the
momentum balance occurs. The left column presents an example with strong
drag ($\tau_{\rm rad}=\tau_{\rm drag}=1\rm\,day$) and the right column
presents an example with weak drag
($\tau_{\rm rad}=0.1\rm\,day$ and $\tau_{\rm drag}=\infty$); both are
equilibrated and steady.
These are high-amplitude cases, so the thickness
({\it top row, a} and {\it e}) exhibits large fractional
variations, and the phase tilts exhibit an overall trend of
northwest-to-southeast (southwest-to-northeast) in the northern
(southern) hemisphere, as explained in previous discussion
(cf Fig.~\ref{stswm-nonlinearity}). These tilts indicate transport of
eddy momentum from midlatitudes to the equator; as a result,
the zonal-mean zonal winds are eastward at the equator and westward
in the midlatitudes ({\it b} and {\it f}). Interestingly, however,
the relative strengths of the equatorial
and midlatitude jets differ and reflect the range of possible variation.
Panels ({\it c}) and ({\it g}) show the terms in the zonal-mean momentum
equation, just the spherical equivalent of Eq.~(\ref{tem}):
\begin{eqnarray}
\nonumber
{\partial\overline{u}\over\partial t}=\underbrace{\overline{v}^*\left[f -
{1\over a\cos\phi}{\partial (\overline{u}\cos\phi)\over\partial \phi}\right]}_{I}\\
\nonumber
\underbrace{-{1\over \overline{h}a\cos^2\phi}{\partial\over\partial\phi}
[\overline{(hv)'u'}\cos^2\phi]}_{II}\\
+ \underbrace{\left[{1\over\overline{h}}
\overline{u'Q'} + \overline{R_u}^*\right]}_{III}
\underbrace{-{\overline{u}^*\over \tau_{\rm drag}}}_{IV}
- {1\over\overline{h}}{\partial(\overline{h'u'})\over\partial t}.
\label{tem2}
\end{eqnarray}
As expected, horizontal convergence of eddy
momentum, term II, causes a strong eastward acceleration at the
equator and westward acceleration in midlatitudes ({\it black curves}).
The vertical eddy-momentum transport,
term III ({\it dark blue curves}), causes a westward acceleration
at the equator that counteracts the eastward acceleration
due to horizontal eddy-momentum convergence. In the case of
strong drag (Fig.~\ref{stswm-diag}{\it c}), the cancellation is
imperfect, leading to a net eddy-induced acceleration that is
eastward at the equator---as predicted by the linear, analytic theory
in \S\ref{linear-sw} (compare to Fig.~\ref{matsuno}c).
A superrotating equatorial jet therefore
emerges and only reaches steady state when the jet becomes
sufficiently strong that the zonal-mean drag on the jet, $-\overline{u}^*/
\tau_{\rm drag}$, balances the eastward acceleration at the equator
(Fig.~\ref{stswm-diag}{\it c}). On the other hand, when drag is
absent, the superrotation induced by the eddy fluxes
becomes quite strong (Fig.~\ref{stswm-diag}{\it f}).
This mean flow alters the eddy fluxes, causing them to self-adjust
to an equilibrium where the accelerations at the equator due to horizontal
and vertical momentum fluxes cancel, leading to no net eddy-induced
acceleration at the equator in steady state (Fig.~\ref{stswm-diag}{\it g}).
\begin{figure*}
\vskip 10pt
\includegraphics[scale=0.9, angle=0]{f8.pdf}
\caption{Fully nonlinear shallow-water solutions with
$\tau_{\rm rad}=1\rm\,day$, $\tau_{\rm drag}=1\rm
\,day$, and $\Delta h_{\rm eq}/H=1.25$ ({\it left column}) and
$\tau_{\rm rad}=0.1\rm\,day$, no drag ($\tau_{\rm drag}\to \infty$),
and $\Delta h_{\rm eq}/H=0.2$ ({\it right column}). ({\it a, e}) Geopotential
$gh$ (orangescale and contours, with thick and thin regions in white
and black, respectively). Scale runs from $1.1$--$5.4\times10^6\rm\,m^2
\,sec^{-2}$ in top panel and
$3.3$--$4.6\times10^6\rm\,m^2\,sec^{-2}$ in bottom panel.
({\it b, f}) Zonal-mean zonal wind $\overline{u}$.
({\it c, g}) Accelerations of the zonal-mean zonal wind in the equilibrated
state. From Eq.~(\ref{tem2}), curves are terms I ({\it cyan}), II ({\it black}), III ({\it dark blue}),
IV ({\it light green}). ({\it d, h}) Mass source $gQ$ at the equator (orangescale, white being
positive and dark being negative values) and winds (arrows). Mass source scale runs from $-41.3$
to $45.1\rm \,m^2\rm \, s^{-3}$ (left) and $-38.1$ to $33.2\rm \,m^2\rm \, s^{-3}$ (right).}
\label{stswm-diag}
\end{figure*}
We find that equatorial superrotation occurs at
all forcing amplitudes, even arbitrarily small amplitudes where the
solutions behave linearly. This
is illustrated in Fig.~\ref{stswm-ueq}, which shows the equilibrated
equatorial zonal-mean zonal wind versus forcing amplitude
for solutions with a range of $\tau_{\rm rad}$ and $\tau_{\rm drag}$
combinations. We therefore conclude that the mechanism
for generating equatorial superrotation described here
has no inherent threshold. Nevertheless, other processes---not included in
the shallow-water model---can in some cases overwhelm the desire of
the day-night forcing to trigger superrotation, particularly when the
day-night forcing is weak. This occurs for example in the cases
examined by \citet{suarez-duffy-1992} and \citet{saravanan-1993},
where superrotation only developed for forcing amplitudes exceeding
a threshold value. In their case, the tropical wave forcing only
triggers superrotation when it attains sufficiently great amplitudes
to overcome the westward torques provided by midlatitude eddies
propagating into the tropics. These issues are discussed further in
\S\ref{discussion}.
In all the cases shown in Fig.~\ref{stswm-ueq} where
$\tau_{\rm drag}$ is finite, the zonal-mean speed of equatorial
superrotation scales
with the square of the forcing amplitude when the forcing amplitude
is sufficiently small. This
in fact is the expected low-amplitude behavior in the high-Prandtl-number
regime described above: at low amplitude, the solutions become linear,
such that the velocities,
height perturbations, and mass source/sink scale with the
forcing amplitude. Because $\overline{R_u}$ scales as
the product of the mass source/sink and the velocities at low
amplitude, it is therefore quadratic in the forcing amplitude.
In the frictional regime, Eq.~(\ref{steady-state}) implies that
$\overline{u}$ at the equator is simply $\tau_{\rm rad}\overline{R_u}$,
and therefore $\overline{u}$ itself is quadratic in the forcing
amplitude. This behavior breaks down when the solutions become
sufficiently high amplitude, as can be seen in Fig.~\ref{stswm-ueq}.
The low-Prandtl-number regime is more complex and can lead to a variety
of scaling behaviors depending on the parameters.
\begin{figure}
\vskip 10pt
\includegraphics[scale=0.6, angle=0]{f9.pdf}
\caption{Equilibrated (steady-state) equatorial jet speed from the
nonlinear shallow-water solutions for a variety of forcing amplitudes,
radiative time constants, and drag time constants. The jet is eastward
(i.e., superrotating) in all cases. In most
cases, the zonal-mean jet speed scales as forcing amplitude squared
at low amplitude but exhibits a flatter dependence at high amplitude.
Here, forcing amplitude is defined as $g\Delta h_{\rm eq}/\tau_{\rm rad}$.}
\label{stswm-ueq}
\end{figure}
The flow in the shallow-water models differ from that in
3D models in one major respect. In many three-dimensional
models of hot Jupiters, eastward equatorial flow occurs not
only in the zonal mean but
at {\it all longitudes}, at least over some range of pressures.
In contrast, although the shallow-water models described here
all exhibit eastward {\it zonal-mean} flow at the equator, the
zonal wind at the equator is always westward over some range
of longitudes. This can be seen as follows: $\overline{R_u}$ is
essentially the mismatch in equatorial zonal acceleration between
horizontal and vertical eddy-momentum transport and in steady state,
when $\overline{u}>0$,
will be greater than or equal to zero. From the definition of
$R_u$, this implies westward equatorial flow at some longitudes.
This trait probably arises because the meteorologically active
atmosphere has here been resolved with only one layer overlying
a deep interior; in future work, it would be interesting
to explore models that represent the flow with two or
more layers overlying
a quiescent interior to see whether they can develop
equatorial flow that is eastward at all longitudes.
\section{Three-dimensional model of equatorial superrotation}
\label{3d}
Here, we show how
the basic mechanism for generating equatorial
superrotation identified in \S\ref{sw} occurs also in three dimensions
under realistic conditions. To do so, we analyze the three-dimensional
model of HD 189733b presented in \citet{showman-etal-2009}.
\citet{showman-etal-2009} coupled the dynamical core of the MITgcm
\citep{adcroft-etal-2004},
which solves the primitive equations of meteorology in global, spherical
geometry, using pressure as a vertical coordinate,
to the state-of-the-art, non-gray radiative transfer scheme of
\citet{marley-mckay-1999}, which solves the multi-stream radiative
transfer equations using the correlated-k method to treat the
wavelength dependence of the opacities. This coupled model, dubbed the
Substellar and Planetary Atmospheric Circulation and Radiation
(SPARC) model, is to date the only GCM to include realistic radiative
transfer for hot Jupiters. Gaseous opacities were calculated assuming local
chemical equilibrium for
a specified atmospheric metallicity, assuming rainout of any condensates
(i.e. ignoring cloud opacity). \citet{showman-etal-2009} presented
synchronously rotating models of HD 189733b with one, five, and ten times
solar metallicity and of HD 209458b with solar metallicity, along with
several models with non-synchronous rotation. Their HD 189733b
models in particular compare favorably with a variety of observational
constraints \citep{showman-etal-2009, agol-etal-2010}, and here we
focus on their solar-metallicity, synchronously rotating HD 189733b
case. This model adopts planetary radius and gravity
of $8.24\times10^7\rm\,m$ and $9.36\rm \,m\,sec^{-2}$. The
rotation rate is $3.3\times10^{-5}\rm\,sec$, corresponding to a rotation
period of 2.2 Earth days.
\begin{figure*}
\centering
\vskip 10pt
\includegraphics[scale=0.7, angle=0]{f10.pdf}
\caption{Temperature (colorscale, in K) and winds (arrows) during the spin-up
phase of the three-dimensional, solar-metallicity
model of HD 189733b from \citet{showman-etal-2009}. This shows
the state at the 30-mbar pressure level---near
the mean infrared photosphere---at a time of one Earth day, after
the day-night forcing has generated a global wave response but before
strong zonal-mean jets have developed. Substellar point
is at $0^{\circ}$ latitude, $0^{\circ}$ longitude.
Note the development of a ``Matsuno-Gill''-type pattern,
which leads to an equatorward flux of eddy momentum
that pumps the equatorial jet.
}
\label{3d-spinup-gill}
\end{figure*}
Figure~\ref{3d-spinup-gill} shows the velocity and temperature
structure at the 30-mbar level
during the spin-up phase of this model---after the
forcing has had sufficient time to trigger a global wave response but
before the equatorial jet has spun up to high speed.
The velocity pattern in the three-dimensional model (Fig.~\ref{3d-spinup-gill})
strongly resembles the standing
Kelvin and Rossby-wave pattern described in \S\ref{sw}.
The flow clearly exhibits the east-west divergence along the equator, emanating
from a point near the substellar longitude, identified in \S\ref{sw}
as the standing Kelvin-wave response. The longitude of peak divergence
(i.e., the longitude at the equator where the zonal velocity changes sign)
lies east of the substellar longitude, as expected from the
analytic theory and nonlinear shallow-water runs in \S\ref{sw}.
Moreover, the flow exhibits the broad gyres in each hemisphere,
anticyclonic on the dayside and cyclonic on the nightside, identified
in \S\ref{sw} as the standing Rossby-wave response. As predicted
analytically, the velocities in these gyres exhibit a
northwest-to-southeast (southwest-to-northeast) phase tilt in the
northern (southern) hemisphere. These phase tilts
imply that $\overline{u'v'}$ is negative in the northern
hemisphere and positive in the southern hemisphere. Eddy momentum
therefore fluxes from the midlatitudes to the equator, and it is
this flux that produces the superrotating equatorial jet
(see Fig.~\ref{qualitative-scenario}). The overall
qualitative resemblance to the analytic calculation in Fig.~\ref{matsuno} is
striking.
As in the shallow-water solutions, the three-dimensional
models exhibit a net downward eddy momentum flux at the equator
throughout the upper atmosphere where the radiative heating/cooling
is strong. This momentum flux results from the fact that, at
the equator, (i) the Matsuno-Gill-type standing-wave patterns
lead to net zonal eddy
velocities that are predominantly westward on the dayside and
eastward on the nightside (see Fig.~\ref{3d-spinup-gill}),
and (ii) net radiative heating occurs on much of the dayside, leading
to net upward velocities, whereas net radiative cooling occurs on the
nightside, leading to net downward velocities. Thus, at the equator,
upward velocities tend to be correlated with westward eddy velocities
and vice versa. This transports eastward momentum downward and
causes a westward acceleration at the equator throughout the upper
atmosphere, which counteracts the eastward equatorial acceleration caused by
latitudinal eddy-momentum transport---just as predicted by the
analytic and numerical shallow-water solutions in \S\ref{sw} (see
Figs.~\ref{matsuno}, \ref{cases-accel}, and \ref{stswm-diag}).
To quantify the accelerations resulting from these momentum fluxes,
we consider the Eulerian-mean zonal-momentum equation in pressure coordinates.
By expanding the dynamical variables into
zonal-mean and deviation (eddy) components, and zonally averaging the
zonal-momentum equation, and adopting pressure as the vertical coordinate,
we obtain\footnote{An equation analogous to this, except using
log-pressure rather than pressure itself as the coordinate, can
be found in \citet[][Eq.~3.3.2a]{andrews-etal-1987}.}
\begin{eqnarray}
\nonumber
{\partial\overline{u}\over\partial t}=\overline{v}\left[f
- {1\over a\cos\phi}{\partial(\overline{u}\cos\phi)\over
\partial \phi}\right] - \overline{\omega}{\partial\overline{u}\over\partial p}\\
-{1\over a\cos^2\phi}{\partial(\overline{u'v'}\cos^2\phi)\over\partial \phi}
-{\partial(\overline{u'\omega'})\over\partial p} + \overline{X}
\label{eulerian-mean}
\end{eqnarray}
On the righthand side, the terms describe the meridional momentum
advection by the zonal-mean circulation, vertical momentum advection
by the zonal-mean circulation, the meridional eddy-momentum convergence, the
vertical eddy-momentum convergence, and friction (represented
generically by $\overline{X}$), respectively.
At the equator, the Coriolis term is zero. Because of the approximate
symmetry of the flow about the equator, $\overline{v}$ and the meridional
gradient of $\overline{u}$ are small there, so the mean-meridional
advection term is small at the equator. The mean vertical-advection
term also tends to be weak for the flow considered here, and the
net zonal acceleration at the equator is
then determined primarily by a competition between the horizontal and vertical
eddy-momentum convergence terms (the analogs of terms
II and III in Eq.~(\ref{tem2}) for the shallow-water system).
\begin{figure}
\vskip 10pt
\includegraphics[scale=0.7, angle=0]{f11.pdf}
\caption{Time evolution of the upper-atmospheric circulation in the three-dimensional
model of HD 189733b by \citet{showman-etal-2009}. ({\it Top:})
Zonal-mean zonal wind at the equator. ({\it Bottom:}) Zonal
accelerations due to latitudinal eddy-momentum convergence,
$-(a\cos^2\phi)^{-1}\partial(\overline{u'v'}\cos^2\phi)/\partial\phi$
({\it black curve}) and vertical eddy-momentum
convergence, $-\partial(\overline{u'\omega'})/\partial p$
({\it blue curve}). All quantities are shown in the upper atmosphere, averaged
vertically from 30 mbar to the top of the model.}
\label{3d-time-evol}
\end{figure}
Figure~\ref{3d-time-evol} shows the time evolution of
the zonal-mean zonal wind and the two eddy acceleration terms
at the equator for the solar-metallicity model of HD 189733b
from \citet{showman-etal-2009}. These are vertical averages
through the top portion of the atmosphere where the radiative heating/cooling
is strong. The zonal-mean zonal wind accelerates rapidly from the
initial rest state and approaches an equilibrium within $\sim$100 days
({\it top}). As expected, the acceleration due to horizontal
eddy transport is eastward, while that due to
vertical eddy transport is westward ({\it bottom}). Moreover,
as suggested by the linear and nonlinear shallow-water calculations,
the magnitude of the horizontal momentum convergence exceeds that
of the vertical momentum convergence during spin-up,
so the net acceleration is eastward
at early times. A superrotating equatorial jet therefore develops.
As the jet speed builds,
the two acceleration terms weaken significantly, and the ratio of their
magnitudes approaches one. As a result, the {\it net} acceleration
drops to zero, allowing the jet to equilibrate to a constant speed
({\it top}). This model is in the same regime as the shallow-water
calculation presented in the right column of Fig.~\ref{stswm-diag}.
\begin{figure*}[ht]
\centering
\subfigure{
\includegraphics[scale=0.6]{f12a.pdf}
}
\subfigure{
\includegraphics[scale=0.6]{f12b.pdf}
}
\caption{{\it Top:} Temperature (colorscale, in K) and winds (arrows) for
the solar-metallicity model of HD 189733b from \citet{showman-etal-2009}.
{\it Bottom:} Eddy temperature $T'$ (colorscale, K) and eddy winds
$(u',v')$ (arrows) for the same model. Both are shown at 30 mbar pressure,
near the mean infrared photosphere, after the winds at these levels
have reached steady state.}
\label{equil}
\end{figure*}
The weakening in time of the eddy accelerations seen in
Fig.~\ref{3d-time-evol} indicates that the mean flow, once
it forms, exerts a back-reaction on the eddies that alters their
structure. The nature of these changes are illustrated in
Fig.~\ref{equil}. The top panel shows the temperatures
and winds at 30 mbar pressure after the flow at this altitude has
become steady; the superrotating equatorial jet, eastward offset
of the hottest region from the substellar point, and other features
are evident as detailed in \citet{showman-etal-2009}. The bottom
panel depicts the {\it eddy} temperature and {\it eddy} winds for the
same pressure and time---that is, $T'$ in colorscale and ($u'$,$v'$)
as arrows. Several features are similar to those in
Fig.~\ref{3d-spinup-gill}: the eddy flow near the equator is approximately
zonal and exhibits a Kelvin-wave-like character, with predominantly
eastward flow at some longitudes and westward flow at others; the
midlatitudes contain broad Rossby-wave gyres in each hemisphere,
anticyclonic on the dayside and cyclonic on the nightside.
Interestingly, however, the Kelvin-wave structure is shifted eastward,
and the midlatitude velocity structure differs significantly,
relative to that with weak mean flow (compare
Fig.~\ref{equil}b to Figs.~\ref{matsuno}b and
\ref{3d-spinup-gill}). From longitudes of about
$-75^{\circ}$ to $+40^{\circ}$, the midlatitude velocity structure induces
equatorward momentum flux (i.e., $u'v'$ negative in the northern
hemisphere and positive in the southern hemisphere), but at longitudes
$\sim80$ to $150^{\circ}$ the flux is reversed (i.e., $u'v'$ positive
in the northern hemisphere and negative in the southern hemisphere).
Due to this cancellation, the magnitude of the zonally averaged flux
$\overline{u'v'}$ is significantly weaker in the equilibrated state
than during the spin-up phase, when the signs of the midlatitude
$u'v'$ add coherently at most longitudes (see Fig.~\ref{3d-spinup-gill}).
The latitudinal pattern of zonal-mean
eddy accelerations in the upper atmosphere of the 3D
model, shown in Fig.~\ref{3d-acceleration}, exhibit a
strong relationship to those from
the shallow-water calculations. The comparison is most
apt to shallow-water calculations with short
radiative time constant ($\tau_{\rm rad}\sim0.1$--1 day), weak
frictional drag ($\tau_{\rm drag}\to\infty$), and large
amplitude, as depicted for example in the right
column of Fig.~\ref{stswm-diag}.
In the 3D model (Fig.~\ref{3d-acceleration}),
the acceleration due to horizontal eddy-momentum
convergence is eastward at the equator and westward in
midlatitudes, in agreement with analytic theory and non-linear shallow-water
solutions (compare black curve in
Fig.~\ref{3d-acceleration} to Figs.~\ref{cases-accel}
and \ref{stswm-diag}g). Near the poles,
the situation is more complex. Poleward of $\sim$$50^{\circ}$ latitude,
the acceleration in the upper atmosphere of the 3D run is eastward.
In the shallow-water solutions, the pattern of eddy
accelerations at high latitude depend on $\tau_{\rm rad}$, $\tau_{\rm drag}$,
and the forcing amplitude (e.g., compare Figs.~\ref{stswm-diag}c and
\ref{stswm-diag}g
poleward of $\sim$$60^{\circ}$ latitude), but for the shallow-water
cases most relevant to the 3D run shown here---such as the
right column in Fig.~\ref{stswm-diag}---the acceleration
due to horizontal eddy convergence
becomes eastward at high latitudes (Fig.~\ref{stswm-diag}g), like
that in the 3D run. In the 3D model, the acceleration due to
vertical eddy-momentum convergence ({\it blue curve} in
Fig.~\ref{3d-acceleration}) is westward at the equator
and eastward in the midlatitudes, again like that arising
in the shallow-water solutions, although a significant difference
is that the midlatitude eastward acceleration is weak in
the shallow-water runs but strong in the three-dimensional
run (relative to the magnitude of acceleration at the equator).
\begin{figure}
\vskip 10pt
\includegraphics[scale=0.6, angle=0]{f13.pdf}
\caption{Zonal-mean zonal accelerations from the solar-metallicity
model of HD 189733b from \citet{showman-etal-2009}. Black and
blue show accelerations due to latitudinal and vertical
convergence of eddy momentum, respectively. Both are vertically
averaged from 30 mbar to the top of the model at 0.2 mbar and temporally
averaged from 11 to 870 days.}
\label{3d-acceleration}
\end{figure}
Of course, the standing eddy patterns and resulting zonal-wind
accelerations in 3D models depend on the strength
of radiative heating/cooling and drag, just as they do in the
shallow-water models. For example, in the shallow-water solutions,
the Kelvin-wave structure and Rossby gyres are spatially distinct
when the radiative and/or drag time constants are long and the
forcing amplitude is small but not when the time constants
are short or the forcing amplitude is large (e.g., compare the
upper left versus lower right of Fig.~\ref{cases}, the top
versus the bottom of Fig.~\ref{stswm-equal-tau}, and the
top versus the bottom of Fig.~\ref{stswm-nonlinearity}).
The 3D models shown here lie at an intermediate position along
this continuum, with Rossby and Kelvin-wave structures that
are visibly distinct, analogous for example to the shallow-water case in
Fig.~\ref{matsuno}b. 3D models with very strong heating rates,
however, seem to exhibit eddy patterns lacking distinct Rossby
wave gyres, more analogous to the top-left case in Fig.~\ref{cases}
and the top case in Fig.~\ref{stswm-equal-tau}.
Examples of models in this regime include the topmost part of the
atmosphere in the models of \citet{cooper-showman-2005, cooper-showman-2006},
\citet{koskinen-etal-2007},
\citet{rauscher-menou-2010} and the HD 209458b model of
\citet{showman-etal-2009}. In contrast, cases in the literature with
more modest heating rates tend to exhibit distinct standing Rossby and
Kelvin-wave structures; examples include \citet[][Fig.~5]{showman-guillot-2002},
\citet[][Figs.~1 and 12]{heng-vogt-2011}, the lower portion of
some of the models of \citet[][Fig.~3b]{koskinen-etal-2007},
and several of the runs in \citet{thrastarson-cho-2010},
which exhibit a planetary-scale cyclone and anticyclone in each
hemisphere.
Despite differences of detail, the overall broad similarities
described here between the
3D and shallow-water models argues strongly that the mechanism
for equatorial jet maintenance that we have identified occurs
in both the shallow-water and 3D models.
\section{Discussion}
\label{discussion}
The development of an eastward equatorial
jet---that is, equatorial superrotation---is a common feature emerging
from three-dimensional models of synchronously rotating hot Jupiters
and extrasolar terrestrial planets \citep{showman-guillot-2002,
cooper-showman-2005, cooper-showman-2006,
showman-etal-2008a, showman-etal-2009, dobbs-dixon-lin-2008,
menou-rauscher-2009, rauscher-menou-2010, perna-etal-2010,
heng-etal-2010, joshi-etal-1997, merlis-schneider-2011, heng-vogt-2011}.
\citet{showman-guillot-2002} first pointed out that, when
the radiative and advective time constants are similar,
this superrotation causes an eastward displacement of the hottest
regions from the substellar point---a phenomenon discovered on HD 189733b
five years later \citep{knutson-etal-2007b, knutson-etal-2009a}.
Despite its relevance, however, the dynamial mechanisms responsible for
generating the equatorial superrotation on tidally locked exoplanets
have not been previously identified.
Here, we have shown that the equatorial superrotating jet
results from an interaction of the mean flow
with standing, planetary-scale Rossby and Kelvin waves generated by
the day-night thermal forcing. The strong longitudinal variations
in radiative heating---namely
intense dayside heating and nightside cooling---trigger the formation
of standing, planetary-scale equatorial Rossby and Kelvin waves;
this is essentially a linear response when wind speeds are modest,
although nonlinearities affect the wave structure at high
amplitude. The Kelvin waves straddle the equator while the
Rossby waves lie on their poleward flanks. As a result of the
differential zonal propagation---Kelvin waves propagating to the east and
long-wavelength Rossby waves to the west---as well as the multi-way
force balance between pressure-gradient, Coriolis, advective,
and drag forces, the velocities develop tilts that resemble
an eastward-pointing chevron centered at the equator. These velocity tilts
pump eastward momentum from high latitudes to the equator, thereby
inducing equatorial superrotation. In steady state, the zonal-mean
equatorial jet speed near the photosphere is determined by a balance
between this eastward, wave-induced acceleration and westward
equatorial acceleration resulting from vertical eddy-momentum transport
and/or drag. We demonstrated the mechanism in a hierarchy of dynamical
models---including linear, analytic shallow-water models,
fully nonlinear shallow-water models, and state-of-the-art
three-dimensional GCMs. For conditions relevant
to hot Jupiters, such equatorial superrotation
occurs over a wide range of radiative heating rates and drag time
constants. The consistency of the picture emerging
from this sequence of models with widely varying complexity is encouraging
and suggests that the mechanism is robust.
The mechanism identified here has several implications:
\begin{itemize}
\item It implies that the equatorial jet
results from a direct, essentially weakly nonlinear interaction
between the thermally forced waves and the mean flow at the planetary scale.
Eddy-eddy interactions, including the possibility of inverse
or forward energy cascades or other turbulent interactions,
may occur but are not essential to the basic
mechanism. (This is analogous to the situation suggested by
\citet{ogorman-schneider-2007} for interaction of baroclinic
midlatitude eddies with the mean flow on Earth.)
\item The wave-mean-flow interaction produces an equatorial jet whose
latitudinal width is comparable to that of the Rossby waves, namely
the equatorial Rossby deformation radius modified by radiative and
frictional effects. For conditions typical of synchronously rotating
hot Jupiters, this length is comparable to a planetary radius, explaining
the broad scale of the equatorial jet obtained in most hot Jupiter models.
\item When acting in isolation, this mechanism for generating
superrotation has no inherent forcing-amplitude threshold; it operates at
any forcing amplitude, unlike the behavior reported in some
earlier studies in the terrestrial context \citep{suarez-duffy-1992,
saravanan-1993}.
\item For parameter combinations appropriate to hot exoplanets,
our models yield flows whose hottest regions along the equator
lie to the east of the substellar point. This property
results from the eastward group propagation of
Kelvin waves. The development of a strong mean flow (equatorial
superrotation) further enhances the offset by its eastward advection
of the temperature pattern. Together, these effects
can explain the offsets observed on HD 189733b
\citep{knutson-etal-2007b, knutson-etal-2009a}.
\end{itemize}
Despite the ubiquity of superrotation in our models,
the flow could shift regimes to one with westward
zonal-mean equatorial flow if westward equatorial accelerations caused
by other processes---not considered here---become important.
For example, if baroclinic instabilities occur in midlatitudes, they
could cause Rossby wave radiation at midlatitudes, potentially allowing
absorption of Rossby waves near the equator. This would contribute
a westward wave-induced acceleration near the equator.
If this westward acceleration dominates over the eastward equatorial
acceleration caused by
the day-night heating contrast, then the net wave-induced
acceleration at the equator could be westward and equatorial
superrotation would not occur.
This could occur on hot Jupiters if the planetary rotation rate
is sufficiently fast and heating rates are sufficiently low.
Note that baroclinic instabilities cannot occur in a one-layer model such
as the shallow-water model, helping to explain why no such transitions
to westward zonal-mean equatorial flow were seen in the shallow-water models
presented here. In some cases, Hadley cells may also force the
circulation into a regime of westward equatorial flow, particularly if
the planetary obliquity is non-zero; this is the case in Earth's
troposphere. On the other hand,
if the planetary rotation is sufficiently slow, additional mechanisms
for generating equatorial superrotation become possible, even when the
forcing is axisymmetric \citep{delgenio-zhou-1996, mitchell-vallis-2010}.
Exploring the combinations of orbital semimajor axes,
stellar fluxes, planetary rotation rates, and atmospheric compositions
under which such regime transitions can occur in 3D models is an
important goal for future work, as such transitions could have
major implications for visible and infrared light curves.
It is worth mentioning that the presence of a physical surface---and
entropy gradients on that surface---promote the
existence of baroclinic instabilities in midlatitudes
\citep[see for example][Chapter 6]{vallis-2006}, so tidally locked
terrestrial planets may be more prone than hot Jupiters to experience
a regime transition to a flow containing midlatitude eastward jets in
addition to, or instead of, a flow dominated by equatorial superrotation.
In the geophysical-fluid-dynamics (GFD) literature, the generation of
midlatitude eastward zonal jets is often described theoretically
using simple barotropic considerations analogous to those summarized in
\S\ref{background}, involving the propagation of waves.
While such barotropic considerations
seem to work reasonably well for midlatitude jets,
the barotropic framework fundamentally fails to explain the emergence
of equatorial superrotation
in 3D models of synchronously rotating exoplanets, where the flow
is often steady and symmetric about the equator.
The theory presented here overcomes this obstacle and
represents an extension of the barotropic
framework to a thermally forced, stratified, ageostrophic flow with
finite Rossby deformation radius.
Specifically, the generation of eastward jets is often
interpreted in terms of the meridional propagation of Rossby waves
away from a source region and their dissipation at other latitudes
\citep[e.g.,][]{held-2000, vallis-2006}. As described in
\S\ref{background}, this interpretation seems to flow naturally
from barotropic theory, in which free Rossby waves, even at the
equator, can propagate not only in longitude but also in latitude.
In contrast, although our work clearly shows how superrotation can emerge
on tidally locked planets,
meridional wave propagation plays no obvious role in our theory.
Unlike barotropic Rossby waves, the baroclinic Rossby
waves in our theory
are equatorially trapped, confined to an equatorial waveguide:
they can propagate in longitude and potentially height but not
latitude.\footnote{Although the theory presented here is for steady,
forced conditions, this key
distinction holds even for freely propagating, unforced waves:
barotropic Rossby waves can generally propagate in latitude---even
at the equator---while baroclinic equatorial Rossby waves tend to
be confined to an equatorial waveguide. For a discussion of the
trapping of equatorial waves, see \citet[][pp.~394-400, 429-432]{holton-2004}
or \citet[][pp.~200-208]{andrews-etal-1987}. }
Moreover, under conditions appropriate to typical hot Jupiters,
these waves exhibit meridional scales typically stretching from the
equator to the pole. It is not at all clear that the paradigm of waves
propagating from one latitude to another applies in this context.
Rather, the velocity tilts that allow the meridional momentum
fluxes to generate superrotation
appear to be explained by the differential {\it zonal}---rather
than meridional---propagation of equatorially trapped Kelvin and
Rossby waves. Future work may further clarify the issue.
|
\section{Introduction}\label{sec:introduction}
This paper has two seemingly different motivations and,
correspondingly, can be read from two different points of view,
a more algebraic and a more geometric one. Both in the introduction and
the main body of the paper, we try to separate these two points of view
as much as possible, in the hope to keep the paper accessible
for readers with specific interests.
\subsection{Algebraic motivation: equivariant modular categories}
\label{intro:1.1}
Among tensor categories, modular tensor categories
are of particular interest for representation theory and
mathematical physics.
The representation categories of several algebraic
structures give examples of semisimple modular tensor categories:
\\[-1.66em]\begin{enumerate}\addtolength\itemsep{-2pt}
\item
Left modules over connected factorizable ribbon weak Hopf algebras
with Haar integral over an algebraically closed field {\rm \cite{nitv}}.
\item
Local sectors of a finite $\mu$-index net of von Neumann algebras
on $\mathbb R$, if the net is strongly additive and split {\rm \cite{KLM}}.
\item
Representations of selfdual $C_2$-cofinite vertex algebras
with an additional finiteness condition on the homogeneous
components and which have semisimple representation categories {\rm \cite{Hu}}.
\end{enumerate}
Despite this list and the rather different fields in which modular
tensor categories arise, it is fair to say that modular
tensor categories are rare mathematical objects.
Arguably, the simplest incarnation of the first
algebraic structure in the list is the Drinfel'd double ${\mathcal D}(G)$
of a finite group $G$.
Bantay \cite{bantay} has suggested a more general source for
modular tensor categories: a pair, consisting of a finite group $H$ and
a normal subgroup $G\triangleleft H$. (In fact, Bantay has suggested general
finite crossed modules, but for this paper, only the case of a normal
subgroup is relevant.) In this situation, Bantay constructs a
ribbon category which is, in a natural
way, a representation category of a ribbon Hopf algebra
${\mathcal B}(G\triangleleft H)$. Unfortunately, it turns out that,
for a proper subgroup inclusion,
the category ${\mathcal B}(G\triangleleft H)\text{-}\mathrm{mod}$ is only premodular and
not modular.
Still, the category ${\mathcal B}(G\triangleleft H)\text{-}\mathrm{mod}$ is modularizable in the
sense of Brugui\`eres \cite{brug}, and the next
candidate for new modular tensor categories is the modularization
of ${\mathcal B}(G\triangleleft H)\text{-}\mathrm{mod}$.
However, it has been shown \cite{maier} that this modularization
is equivalent to the representation category of
the Drinfel'd double ${\mathcal D}(G)$.
The modularization procedure of Brugui\`eres is based
on the observation that the violation of modularity of a
modularizable tensor category ${\mathcal C}$ is captured in terms
of a canonical Tannakian subcategory of ${\mathcal C}$. For
the category ${\mathcal B}(G\triangleleft H)\text{-}\mathrm{mod}$, this subcategory can
be realized as the representation category of the
the quotient group $J:= H/G$ \cite{maier}. The modularization functor
$$ {\mathcal B}(G\triangleleft H)\text{-}\mathrm{mod} \to {\mathcal D}(G)\text{-}\mathrm{mod} $$
is induction along the commutative Frobenius algebra given by the
regular representation of $J$. This has the important
consequence that the modularized category ${\mathcal D}(G)$
is endowed with a $J$-action.
Experience with orbifold constructions, see \cite{kirI17,turaev2010}
for a categorical formulation, raises the question
of whether the category ${\mathcal D}(G)\text{-}\mathrm{mod}$ with this $J$-action
can be seen in a natural way as
the neutral sector of a $J$-modular tensor category.
We thus want to complete the following square of tensor categories
\begin{equation}\label{int:square}
\xymatrix{
{\mathcal D}(G)\text{-}\mathrm{mod} \ar@<0.7ex>[d]^{\text{orbifold}} \ar@(lu,ur)[]^J\ar@{^{(}->}[r]
& ??? \ar@(lu,ur)[]^J \ar@<0.7ex>[d]^{\text{orbifold}} \\
{\mathcal B}(G\triangleleft H)\text{-}\mathrm{mod} \ar[u]^{\text{modularization}}\ar@{^{(}->}[r]
& ??? \ar[u]
}
\end{equation}
Here vertical arrows pointing upwards stand for
induction functors along the commutative
algebra given by the regular representation of $J$,
while downwards pointing arrows indicate orbifoldization.
In the upper right corner, we wish to place a $J$-modular category,
and in the lower right corner its $J$-orbifold which, on general
grounds \cite{kirI17}, has to be a modular tensor category.
Horizontal arrows indicate the inclusion of neutral sectors.
In general, such a completion need not exist. Even if it exists,
there might be inequivalent choices of $J$-modular tensor categories
of which a given modular tensor category with $J$-action is the neutral
sector \cite{ENO2009}.
\subsection{Geometric motivation: equivariant extended TFT}
Topological field theory is a mathematical structure that has been inspired
by physical theories \cite{witten} and which has developed
into an important tool in low-dimensional topology. Recently, these theories have received increased attention due
to the advent of {\em extended} topological field theories
\cite{Lurie,priesPhD}. The present paper focuses on three-dimensional
topological field theory.
Dijkgraaf-Witten theories provide a class of extended topological
field theories. They can be seen as discrete variants of Chern-Simons
theories, which provide invariants of three-manifolds and
play an important role in knot theory \cite{witten}.
Dijkgraaf-Witten theories have the advantage of being
particularly tractable and admitting a very conceptual geometric construction.
A Dijkgraaf-Witten theory is based on a finite group $G$; in this case the
'field configurations' on a manifold $M$ are given by $G$-bundles over $M$,
denoted by $\mathcal A_G(M)$. Furthermore, one has to choose a suitable
action functional $S: \mathcal A_G(M) \to \mathbb{C}$
(which we choose here in fact to be trivial) on field configurations;
this allows to make the structure suggested by formal path integration
rigorous and to obtain a topological field theory. A conceptually
very clear way
to carry this construction out rigorously is described in \cite{FQ93} and
\cite{Morton}, see section \ref{sec:DW} of this paper for a review.
Let us now assume that as a further input datum we have another finite group
$J$ which acts on $G$. In this situation, we get an action of $J$ on the
Dijkgraaf-Witten theory based on $G$. But it turns out that this
topological field
theory together with the $J$-action does not fully reflect
the equivariance of the situation: it has been an important insight that
the right notion is the one of equivariant topological field theories,
which have been another point of recent interest \cite{kirI17,turaev2010}.
Roughly speaking, equivariant topological field theories require that
all geometric objects
(i.e.\ manifolds of different dimensions) have to be decorated by a
$J$-cover (see definitions \ref{defcobj} and \ref{defjtft} for details).
Equivariant field theories also provide a conceptual setting for the
orbifold construction, one of the standard tools for model building in
conformal field theory and string theory.
Given the action of a finite group $J$ on a finite group $G$, these
considerations lead to the question of whether Dijkgraaf-Witten theory
based on $G$ can be enlarged to a $J$-equivariant topological field theory.
Let us pose this question more in detail:
\begin{itemize}
\item What exactly is the right notion of an action of $J$ on $G$
that leads to interesting theories? To keep equivariant Dijkgraaf-Witten
theory as explicit as the non-equivariant theory, one needs notions to keep
control of this action as explicitly as possible.
\item
Ordinary Dijkgraaf-Witten theory is mainly determined by the choice of
field configurations $\mathcal A_G(M)$ to be $G$-bundles. As mentioned before,
for $J$-equivariant theories, we should replace manifolds
by manifolds with $J$-covers. We thus need a geometric notion of a
$G$-bundle that is 'twisted' by this $J$-cover in order to develop the
theory parallel to the non-equivariant one.
\end{itemize}
Based on an answer to these two points, we wish to construct
equivariant Dijkgraaf-Witten theory as explicitly as possible.
\subsection{Summary of the results}
This paper solves both the algebraic and the geometric problem we have
just described. In fact, the two problems turn out to be closely related.
We first solve the problem of explicitly constructing equivariant
Dijkgraaf-Witten and then use our solution to construct the relevant
modular categories that complete the square \eqref{int:square}.
Despite this strong mathematical interrelation, we have taken some
effort to write the paper in such a way that it is accessible to
readers sharing only a geometric or algebraic interest. The geometrically
minded reader might wish to restrict his attention to section 2 and 3,
and only take notice of the result about $J$-modularity
stated in theorem \ref{J-modular}. An algebraically oriented
reader, on the other hand, might simply accept the categories
described in proposition \ref{sectors} together with the structure
described in propositions \ref{prop:fusionproduct},
\ref{prop:action}
and \ref{prop:braiding} and then directly delve into section 4.
\medskip
For the benefit of all readers, we present here an outline of all our findings.
In section 2, we review the pertinent aspects of Dijkgraaf-Witten theory and
in particular the specific construction given in \cite{Morton}.
Section 3 is devoted to the equivariant case: we observe that the correct
notion of $J$-action on $G$ is what we call a weak
action of the group $J$ on the group $G$; this notion is
introduced in definition \ref{Def:action}. Based on this notion,
we can very explicitly construct for every $J$-cover $P\to M$ a category
$\mathcal A_G(P \to M)$ of $P$-twisted $G$-bundles. For the definition and
elementary properties of twisted bundles, we refer to section
\ref{sec:twisted} and for a local description to appendix
\ref{cech}. We are then ready to construct equivariant Dijkgraaf
Witten theory along the lines of the construction described in
\cite{Morton}. This is carried out in section \ref{subsec:eqDW} and
\ref{subsec:spans}. We obtain a construction of equivariant Dijkgraaf-Witten
theory that is so explicit that we can read off the category ${\mathcal C}^J(G)$ it
assigns to the circle $\mathbb{S}^1$. The equivariant topological field
theory induces additional structure on this category, which can also be
computed by geometric methods due to the explicit control of the theory,
and part of which we compute in section \ref{twisted_sectors_and_fusion}.
This finishes the geometric part of our work. It remains to show
that the category ${\mathcal C}^J(G)$ is indeed $J$-modular.
\medskip
To establish the $J$-modularity of the category
${\mathcal C}^J(G)$, we have to resort to algebraic tools.
Our discussion is based on the appendix 6 of \cite{turaev2010}
by A.\ Vir\'elizier. At the same time,
we explain the solution of the algebraic problems described in
section \ref{intro:1.1}. The Hopf algebraic notions we encounter
in section 4, in particular Hopf algebras with a weak group
action and their
orbifold Hopf algebras might be of independent algebraic interest.
In section 4, we introduce the notion of a $J$-equivariant ribbon
Hopf algebra. It turns out that it is natural to relax
some strictness requirements on the $J$-action on such a Hopf algebra.
Given a weak action of a finite group $J$ on a
finite group $G$, we describe in proposition \ref{double-J-ribbon} a specific
ribbon Hopf algebra which we call the equivariant Drinfel'd double ${\mathcal D}^J(G)$.
This ribbon Hopf algebra is designed in such a way that its representation
category is equivalent to the geometric category ${\mathcal C}^J(G)$
constructed in section 3, compare
proposition \ref{J-fusion}.
The $J$-modularity of ${\mathcal C}^J(G)$ is established via the modularity of
its orbifold category. The corresponding notion of an orbifold algebra
is introduced in subsection 4.4. In the case of the equivariant
Drinfel'd double ${\mathcal D}^J(G)$, this orbifold algebra is shown to be
isomorphic, as a ribbon Hopf algebra, to a Drinfel'd double. This implies
modularity of the orbifold theory and, by a result of \cite{kirI17},
$J$-modularity of the category ${\mathcal C}^J(G)$, cf.\ theorem \ref{J-modular}.
In the course of our construction, we develop
several notions of independent interest.
In fact, our paper might be seen as a study of the
geometry of chiral backgrounds. It allows for various
generalizations, some of which are briefly sketched in the
conclusions. These generalizations include in
particular twists by 3-cocycles in group cohomology
and, possibly, even the case of non-semi simple chiral
backgrounds.
\bigskip
\noindent{\bf Acknowledgements.}
We thank Urs Schreiber for helpful discussions and Ingo
Runkel for a careful reading of the manuscript.
TN and CS are partially supported by the Collaborative
Research Centre 676
``Particles, Strings and the Early Universe - the Structure of Matter
and Space-Time'' and the cluster of excellence
``Connecting particles with the cosmos''.
JM and CS are partially supported by the
Research priority program SPP 1388
``Representation theory''.
JM is partially supported by the Marie Curie Research Training Network
MRTN-CT-2006-031962 in Noncommutative Geometry, EU-NCG.
\section{Dijkgraaf-Witten theory and Drinfel'd double}\label{sec:DW}
This section contains a short review of Dijkgraaf-Witten
theory as an extended three-dimensional topological
field theory, covering the contributions of many authors,
including in particular the work of Dijkgraaf-Witten
\cite{DW90}, of Freed-Quinn \cite{FQ93}
and of Morton \cite{Morton}. We explain how these extended
3d TFTs give rise to modular tensor categories.
These specific modular tensor categories are the representation
categories of a well-known class of quantum groups, the
Drinfel'd doubles of finite groups.
While this section does not contain original material,
we present the ideas in such a way that equivariant generalizations
of the theories can be conveniently discussed. In this section, we also
introduce some categories and functors that we need for later sections.
\subsection{Motivation for Dijkgraaf-Witten theory}
\label{sec:motivation}
We start with a brief motivation for Dijkgraaf-Witten
theory from physical principles. A reader already familiar
with Dijkgraaf-Witten theory might wish to take
at least notice
of the definition \ref{def:2.2} and of proposition \ref{DW}.
It is an old, yet successful idea to extract invariants of
manifolds from quantum field theories, in particular
from quantum field theories for which the fields are $G$-bundles
with connection, where $G$ is some group. In this paper we
mostly consider the case of a finite group and only occasionally
make reference to the case of a compact Lie group.
Let $M$ be a compact oriented manifold of dimension 1,2 or 3, possibly
with boundary. As the \textit{`space' of field configurations},
we choose $G$ bundles with connection,
$$ \mathcal A_G(M) := \mathcal{B}{\!un}^{\nabla}_G(M). $$
In this way, we really assign to a manifold a groupoid,
rather than an actual space.
The morphisms of the category take gauge transformations
into account.
We will nevertheless keep on calling it 'space' since
the correct framework to handle $\mathcal A_G(M)$ is
as a stack on the category of smooth manifolds.
Moreover, another piece of data specifying the model
is a function
defined on manifolds of a specific dimension,
$$ S: \mathcal A_G(M) \to \mathbb{C} $$
called the \textit{action}. In the simplest case, when $G$ is a finite group,
a field configuration is given by a $G$-bundle, since all bundles are
canonically flat and no connection data are involved.
Then, the simplest action is given by $S[P] := 0$ for all $P$.
In the case of a compact, simple, simply connected Lie group $G$,
consider a 3-manifold $M$. In this situation, each $G$-bundle $P$ over $M$
is globally of the form $P \cong G \times M$, because
$\pi_1(G) = \pi_2(G) = 0$. Hence a field configuration is given by a connection on the
trivial bundle which is a 1-form $A \in \Omega^1(M,\mathfrak g)$ with values
in the Lie algebra of $G$. An example of an action yielding
a topological field theory that can be defined in this
situation is the Chern-Simons action
$$ S[A] := \int_M \langle A \wedge dA \rangle
- \frac 1 6 \langle A \wedge A \wedge A \rangle $$
where $\langle \cdot , \cdot \rangle$ is the basic invariant inner product
on the Lie algebra $\mathfrak g$.
The heuristic idea is then to introduce an
invariant $Z(M)$ for a 3-manifold $M$ by integration over all field
configurations:
$$ Z(M) := " \int_{\mathcal A_G(M)} d\phi ~ e^{i S[\phi]} ~". $$
\begin{Warning} In general, this path integral has only a heuristic
meaning. In the case of a finite group, however, one can choose a
counting measure $d\phi$ and thereby reduce the integral to a well-defined
finite sum. The definition of Dijkgraaf-Witten theory \cite{DW90}
is based on this idea.
\end{Warning}
Instead of giving a well-defined meaning to the invariant $Z(M)$ as
a path-integral, we exhibit some formal properties these
invariants are expected to satisfy. To this end, it is crucial to allow
for manifolds that are not closed, as well. This allows to cut a
three-manifold into several simpler three-manifolds with boundaries
so that the computation of the invariant can be reduced to the computation
of the invariants of simpler pieces.
Hence, we consider a 3-manifold $M$ with a 2-dimensional boundary
$\partial M$. We fix boundary values $\phi_1 \in \mathcal A_G(\partial M)$
and consider the space $\mathcal A_G(M,\phi_1)$ of all fields $\phi$ on $M$
that restrict to the given boundary values $\phi_1$. We then introduce, again
at a heuristic level, the quantity
\begin{equation}\label{vektor}
Z(M)_{\phi_1} := " \int_{\mathcal A_G(M,\phi_1)} d\phi ~ e^{i S[\phi]} ~". \end{equation}
The assignment $\phi_1 \mapsto Z(M)_{\phi_1}$ could be called a
`wave function' on the space $\mathcal A_G(\partial M)$
of boundary values of fields.
These `wave functions' form a vector space
$\mathcal H_{\partial M}$, the \textit{state space}
$$ \mathcal H_{\partial M} := "L^2 \big(\mathcal A_G(\partial M),\mathbb{C}\big)~" $$
that we assign to the boundary $\partial M$. The transition to wave functions amounts to a
linearization. The notation $L^2$ should be taken with a grain
of salt and should indicate the choice of an appropriate
vector space for the category $\mathcal A_G(\partial M)$;
it should not suggest the
existence of any distinguished measure on the category.
In the case of Dijkgraaf-Witten theory based on a finite group $G$,
the space of states has a basis consisting
of $\delta$-functions on the set of isomorphism classes of
field configurations on the boundary $\partial M$:
$$ \mathcal H_{\partial M} = \mathbb{C} \big <\delta_{\phi_1} \mid \phi_1 \in Iso \mathcal A_G(\partial M) \big>.$$
In this way, we associate finite dimensional vector spaces
$\mathcal H_{\Sigma}$ to compact oriented 2-manifolds
$\Sigma$. The heuristic path integral in equation
\eqref{vektor} suggests to associate to a
3-manifold $M$ with boundary $\partial M$
an element
$$ Z(M) \in \mathcal H_{\partial M} \,\, , $$
or, equivalently, a linear map $\mathbb{C} \to
\mathcal H_{\partial M}$.
A natural generalization of this situation are
cobordisms $M:
\Sigma \to \Sigma'$, where $\Sigma$ and $\Sigma'$ are
compact oriented
2-manifolds. A cobordism is a compact oriented 3-manifold $M$ with
boundary $\partial M \cong \bar{\Sigma} \sqcup \Sigma'$
where $\bar \Sigma$
denotes $\Sigma$, with the opposite orientation.
To a cobordism, we wish to associate a linear map
$$ Z(M) : \mathcal H_{\Sigma} \to \mathcal H_{\Sigma'} $$
by giving its matrix elements in terms of the path
integral
$$ Z(M)_{\phi_0,\phi_1} := " \int_{\mathcal A_G(M,\phi_0,\phi_1)} d\phi
~ e^{i S[\phi]} ~" \,\,\, $$
with fixed boundary values
$\phi_0 \in \mathcal A_G(\Sigma)$ and
$\phi_1 \in \mathcal A_G(\Sigma')$.
Here $\mathcal A_G(M,\phi_0,\phi_1)$ is the space of field
configurations on $M$ that
restrict to the field configuration $\phi_0$ on the ingoing
boundary $\Sigma$ and to the field configuration $\phi_1$
on the outgoing boundary $\Sigma'$. One can now show that the linear maps $Z(M)$
are compatible with gluing of cobordisms along boundaries.
(If the group $G$ is not finite, additional subtleties arise; e.g.\
$Z(M)_{\phi_0,\phi_1}$ has to be interpreted as an integral kernel.)
Atiyah \cite{At88} has given a definition of
a topological field theory that formalizes these
properties: it describes a topological field theory as
a symmetric monoidal functor from a
geometric tensor category to an algebraic category.
To make this definition explicit, let $\mathfrak{C}\mathrm{ob}(2,3)$ be the category which has 2-dimensional compact
oriented smooth manifolds as objects. Its morphisms
$M:\Sigma\to \Sigma'$ are given by (orientation preserving)
diffeomorphism
classes of 3-dimensional, compact oriented cobordism
from $\Sigma$ to $\Sigma'$ which we write as
$$ \Sigma\hookrightarrow M \hookleftarrow \Sigma'.$$
Composition of morphisms is given by gluing cobordisms together along the
boundary. The disjoint union of 2-dimensional manifolds and cobordisms equips this category
with the structure of a symmetric monoidal category. For the
algebraic category, we choose the symmetric tensor category
$\textnormal{Vect}_{{\mathbb K}}$ of finite dimensional vector spaces over an algebraically
closed field ${\mathbb K}$ of characteristic zero.
\begin{Definition}[Atiyah] \label{def:2.2}
A \textit{3d TFT} is a symmetric monoidal functor
$$ Z: \mathfrak{C}\mathrm{ob}(2,3) \to \textnormal{Vect}_{{\mathbb K}}. $$
\end{Definition}
\medskip
Let us set up such a functor for Dijkgraaf-Witten theory,
i.e.\ fix a finite group $G$ and choose the
trivial action $S: A_G(M) \to \mathbb{C}$, i.e. $S[P] = 0$
for all $G$-bundles $P$ on $M$.
Then the path integrals reduce to finite sums over $1$
hence simply count
the number of elements in the category $\mathcal A_G$.
Since we are counting objects in a category, the
stabilizers have to be taken
appropriately into account, for details see e.g.\
\cite[Section 4]{Morton08}. This is achieved by
the \textit{groupoid
cardinality} (which is sometimes also called the
Euler-characteristic of the groupoid $\Gamma$)
$$ |\Gamma| := \sum_{[g] \in Iso(\Gamma)} \frac 1 {|Aut(g)|}. $$
A detailed discussion of groupoid cardinality can be found in \cite{BD00}
and \cite{Leinster08}.
We summarize the discussion:
\begin{Proposition}[\cite{DW90},\cite{FQ93}] \label{DW}
Given a finite group $G$, the following assignment $Z_G$ defines
a 3d TFT: to a closed, oriented 2-manifold $\Sigma$, we assign
the vector space freely generated by the isomorphism
classes of $G$-bundles on $\Sigma$,
$$ \Sigma \quad \longmapsto \quad \mathcal H_\Sigma
:= {\mathbb K} \big< \delta_P \mid P \in
Iso \mathcal A_G(\Sigma) \big> \,\, . $$
To a 3 dimensional cobordism $M$, we associate the linear map
$$ Z_G\Big(\Sigma\hookrightarrow M \hookleftarrow \Sigma'\Big) : \quad \mathcal H_\Sigma \to \mathcal H_{\Sigma'}$$
with matrix elements given by the groupoid cardinality of
the categories $\mathcal A_G(M,P_0,P_1)$:
$$ Z_G(M)_{P_0,P_1} := \big|\mathcal A_G(M,P_0,P_1)\big| \,\,\, . $$
\end{Proposition}
\begin{Remark}
\begin{enumerate}
\item
In the original paper \cite{DW90}, a generalization of the trivial action
$S[P] = 0$, induced by an element $\eta$ in the group cohomology
$\H^3_{Gp}\big(G,\mathrm U(1)\big)$ with values in $\mathrm U(1)$, has been studied. We postpone
the treatment of this generalization to a separate paper: in the present paper,
the term \textit{Dijkgraaf-Witten theory} refers to the 3d TFT
of proposition \ref{DW} or its extended version.
\item
In the case of a compact, simple, simply-connected Lie group $G$,
a definition of a 3d TFT by a path integral is not available. Instead,
the combinatorial definition of Reshetikin-Turaev \cite{RT} can be used
to set up a 3d TFT which has the properties expected
for Chern-Simons theory.
\item
The vector spaces $\mathcal H_\Sigma$ can be described rather explicitly.
Since every compact, closed, oriented 2-manifold is given by a disjoint union
of surfaces $\Sigma_g$ of genus $g$, it suffices to compute the dimension of
$\mathcal H_{\Sigma_g}$. This can be done using the well-known description of
moduli spaces of flat $G$-bundles in terms of homomorphisms
from the fundamental group $\pi_1(\Sigma_g)$ to the
group $G$, modulo conjugation,
$$ Iso \mathcal A_G(\Sigma_g) \cong \Hom( \pi_1(\Sigma_g) , G) /G $$
which can be combined with the usual description of the fundamental group
$\pi_1(\Sigma_g)$ in terms of generators and relations. In this way, one finds
that the space is one-dimensional
for surfaces of genus 0. In the case of surfaces of genus $1$,
it is generated by pairs of commuting group elements, modulo simultaneous conjugation.
\item
Following the same line of argument, one can show that for a closed
3-manifold $M$, one has
$$ \big|\mathcal A_G(M)\big| = \big|\Hom(\pi_1(M),G)\big| ~/~ |G| \,\, . $$
This expresses the 3-manifold invariants in terms of the
fundamental group of $M$.
\end{enumerate}
\end{Remark}
\subsection{Dijkgraaf-Witten theory as an extended TFT}
Up to this point, we have considered a version
of Dijkgraaf-Witten theory which assigns invariants
to closed 3-manifolds $Z(M)$ and vector spaces to
2-dimensional manifolds
$\Sigma$. Iterating the argument that has lead us to consider
three-manifolds with boundaries, we might wish to cut the
two-manifolds into smaller pieces as well, and thereby introduce
two-manifolds with boundaries into the picture.
Hence, we drop the requirement on the 2-manifold $\Sigma$
to be closed and allow $\Sigma$ to be
a compact, oriented 2-manifold with 1-dimensional boundary
$\partial \Sigma$. Given a field configuration
$\phi_1 \in \mathcal A_G(\partial \Sigma)$ on the boundary
of the surface $\Sigma$, we consider the space of all field
configurations
$\mathcal A_G(\Sigma,\phi_1)$ on $\Sigma$ that restrict
to the given field configuration $\phi_1$ on
the boundary $\partial \Sigma$. Again, we linearize the
situation and consider for each field configuration
$\phi_1$ on the 1-dimensional boundary $\partial \Sigma$
the vector space freely generated by the isomorphism classes of field
configurations on $\Sigma$,
$$ \mathcal H_{\Sigma, \phi_1} := " L^2\big(\mathcal A_G(\Sigma,\phi_1)\big)
" = \mathbb{C} \big<\delta_\phi \mid \phi \in Iso \mathcal A_G(\Sigma,\phi_1)\big>.$$
The object we associate to the 1-dimensional boundary
$\partial\Sigma$ of a 2-manifold $\Sigma$ is thus a map
$\phi_1 \mapsto \mathcal H_{\Sigma, \phi_1}$ of
field configurations to vector spaces, i.e.\
a complex vector bundle over the space of all fields on the
boundary. In the case of a
finite group $G$, we prefer to see these vector bundles
as objects of the functor category from the essentially
small category $\mathcal A_G(\partial \Sigma)$ to
the category $\textnormal{Vect}_\mathbb{C}$ of finite-dimensional
complex vector spaces, i.e.\ as an element of
$$ \textnormal{Vect}(\mathcal A_G(\partial \Sigma)) = \big[ \mathcal A_G(\partial \Sigma) , \textnormal{Vect}_\mathbb{C} \big]. $$
Thus the extended version of the theory assigns the category
$Z(S) =[\mathcal A_G(S),\textnormal{Vect}_\mathbb{C}]$ to a
one dimensional, compact oriented manifold $S$. These categories possess
certain additional properties which can be summarized by saying that they
are 2-vector spaces in the sense of {\cite{KV94}}:
\begin{Definition}\label{twovect}
\begin{enumerate}
\item A \textit{2-vector space} (over a field ${\mathbb K}$) is a ${\mathbb K}$-linear, abelian,
finitely semi-simple category. Here finitely semi-simple means that
the category has finitely many isomorphism classes of simple objects and
each object is a finite direct sum of simple objects.
\item
Morphisms between 2-vector spaces are ${\mathbb K}$-linear functors and 2-morphisms
are natural transformations. We denote the 2-category of 2-vector spaces
by $2\textnormal{Vect}_{\mathbb K}$
\item
The \textit{Deligne tensor product} $~ \boxtimes ~$ endows $2\textnormal{Vect}_{\mathbb K}$
with the structure of a symmetric monoidal 2-category.
\end{enumerate}
\end{Definition}
For the Deligne tensor product, we refer to \cite[Sec. 5]{de90} or
\cite[Def. 1.1.15]{BKLec}. The definition and the properties of symmetric
monoidal bicategories (resp.\ 2-categories) can be found in
\cite[ch. 3]{priesPhD}.
In the spirit of definition \ref{def:2.2}, we formalize the properties of
the extended theory $Z$
by describing it as a functor from a cobordism 2-category to the algebraic category
$2\textnormal{Vect}_{\mathbb K}$. It remains to state the formal definition of the relevant geometric
category. Here, we ought to be a little bit more careful, since we expect
a 2-category and hence can not identify diffeomorphic 2-manifolds.
For precise statements on how to address the difficulties in gluing smooth
manifolds with corners, we refer to \cite[4.3]{Morton06}; here, we
restrict ourselves to the following short definition:
\begin{Definition}\label{cob123}
$\mathfrak{C}\mathrm{ob}(1,2,3)$ is the following symmetric monoidal bicategory:
\begin{itemize}\setlength{\itemsep}{-0.5ex}
\item { Objects} are compact, closed, oriented 1-manifolds $S$.
\item {1-Morphisms} are
2-dimensional, compact, oriented collared cobordisms
$S \times I \hookrightarrow \Sigma \hookleftarrow
S' \times I$.
\item {2-Morphisms} are generated by diffeomorphisms of cobordisms
fixing the collar and 3-dimensional collared, oriented cobordisms with
corners $M$, up to diffeomorphisms preserving the orientation and boundary.
\item Composition is by gluing along collars.
\item The monoidal structure is given by disjoint union
with the empty set $\emptyset$ as the monoidal unit.
\end{itemize}
\end{Definition}
\begin{Remark}\label{smoothgluing}
The 1-morphisms are defined as collared surfaces, since
in the case of extended cobordism categories, we consider
surfaces rather than diffeomorphism classes of surfaces.
A choice of collar is always possible, but not unique.
The choice of collars ensures that the glued surface has
a well-defined smooth structure. Different choices for the
collars yield equivalent 1-morphisms in $\mathfrak{C}\mathrm{ob}(1,2,3)$.
\end{Remark}
Obviously, extended cobordism categories can be defined in
dimensions different from three as well. We are now ready to give
the definition of an extended TFT which goes essentially
back to Lawrence
\cite{Lawrence}:
\begin{Definition}\label{ext}
An \textit{extended 3d TFT} is a weak symmetric monoidal 2-functor
$$ Z: \quad \mathfrak{C}\mathrm{ob}(1,2,3) \to 2\textnormal{Vect}_{{\mathbb K}} \,\,\, . $$
\end{Definition}
We pause to explain
in which sense extended TFTs extend the TFTs defined in
definition \ref{def:2.2}. To this end, we note that the monoidal
2-functor $Z$ has to send the monoidal unit in
$\mathfrak{C}\mathrm{ob}(1,2,3)$ to the monoidal
unit in $2\textnormal{Vect}_{{\mathbb K}}$. The monoidal unit in $\mathfrak{C}\mathrm{ob}(1,2,3)$
is the empty set
$\emptyset$, and the unit in $2\textnormal{Vect}_{\mathbb K}$ is the category
$\textnormal{Vect}_{\mathbb K}$. The functor $Z$
restricts to a functor $Z|_\emptyset$ from the
endomorphisms of $\emptyset$
in $\mathfrak{C}\mathrm{ob}(1,2,3)$ to the endomorphisms of $\textnormal{Vect}_{\mathbb K}$ in $2\textnormal{Vect}_{\mathbb K}$.
It follows directly from the definition that
$\text{End}_{\mathfrak{C}\mathrm{ob}(1,2,3)}\big(\emptyset\big) \cong \mathfrak{C}\mathrm{ob}(2,3)$.
Using the fact that
the morphisms in $2\textnormal{Vect}_{\mathbb K}$ are additive (which follows from
$\mathbb{C}$-linearity of functors in the definition
of 2-vector spaces),
it is also easy to see that the equivalence of categories
$\text{End}_{2\textnormal{Vect}_{\mathbb K}}\big(\textnormal{Vect}_{\mathbb K}\big) \cong \textnormal{Vect}_{\mathbb K}$
holds. Hence we have deduced:
\begin{Lemma}\label{restriction}
Let $Z$ be an extended 3d TFT. Then $Z|_{\emptyset}$ is a
3d TFT in the sense of definition \ref{def:2.2}.
\end{Lemma}
At this point, the question arises whether a given
(non-extended) 3d TFT can be
extended. In general, there is no reason for this to be true.
For Dijkgraaf-Witten theory, however, such an extension can be
constructed based on ideas which we described at the beginning of this section.
A very conceptual presentation of this
this construction based on important ideas of \cite{Freed92}
and \cite{FQ93} can be found in
\cite{Morton}.
Before we describe this construction in more detail in
subsection \ref{lin}, we first state the result:
\begin{Proposition}\cite{Morton}\label{exDW}
Given a finite group $G$, there exists an extended 3d TFT
$Z_G$ which assigns the categories
$$ \big[ \mathcal A_G(S) , \textnormal{Vect}_{\mathbb K} \big] $$
to 1-dimensional, closed oriented manifolds $S$ and whose restriction
$Z_G|_\emptyset$ is (isomorphic to) the Dijkgraaf-Witten TFT
described in proposition \ref{DW}.
\end{Proposition}
\begin{Remark}
One can iterate the procedure of extension and introduce
the notion of a fully extended TFT which also assigns
quantities to points rather than just 1-manifolds.
It can be shown that Dijkgraaf-Witten
theory can be
turned into a fully extended TFT, see \cite{FHLT}. The full
extension will not be needed in the present article.
\end{Remark}
\subsection{Construction via 2-linearization}\label{lin}
In this subsection, we describe in detail
the construction of the extended 3d TFT of
proposition \ref{exDW}. An impatient reader may skip
this subsection and should still be able to understand
most of the paper.
He might, however, wish to take notice of the technique
of 2-linearization in proposition \ref{lin2} which is also an
essential ingredient in our
construction of equivariant Dijkgraaf-Witten theory in
sequel of this paper.
As emphasized in particular by Morton \cite{Morton}, the
construction of the extended TFT is
naturally split into two steps, which have already been
implicitly present
in preceding sections. The first step is to assign to manifolds and
cobordisms the configuration spaces $\mathcal A_G$ of $G$
bundles. We now restrict ourselves to the case when $G$ is a
finite group. The following fact is standard:
\begin{itemize}
\item
The assignment $M \mapsto \mathcal A_G(M) := \mathcal{B}{\!un}_G$ is a contravariant
2-functor from the category of manifolds to the 2-category of
groupoids.
Smooth maps between manifolds are mapped to the corresponding pullback
functors on categories of bundles.
\end{itemize}
A few comments are in order: for a connected manifold $M$, the category
$\mathcal A_G(M)$ can be replaced by the equivalent category
given by the action groupoid $\Hom\big(\pi_1(M),G\big)//G$
where $G$ acts by conjugation. In particular, the category
$\mathcal A_G(M)$ is essentially finite,
if $M$ is compact. It should be
appreciated that at this stage no restriction is imposed on
the dimension of the manifold $M$.
The functor $\mathcal A_G(-)$ can be evaluated on a 2-dimensional cobordism
$S \hookrightarrow \Sigma \hookleftarrow S'$ or a 3-dimensional cobordism
$\Sigma \hookrightarrow M \hookleftarrow \Sigma'$. It then yields diagrams of the form
\begin{eqnarray*}
\mathcal A_G(S) \longleftarrow ~\mathcal A_G(\Sigma) \longrightarrow \mathcal A_G(S') \\
\mathcal A_G(\Sigma) \longleftarrow \mathcal A_G(M) \longrightarrow \mathcal A_G(\Sigma').
\end{eqnarray*}
Such diagrams are called spans. They are the morphisms of
a symmetric monoidal bicategory $\mathfrak{S}\mathrm{pan}$ of spans of groupoids as follows (see e.g. \cite{Daw04} or \cite{Morton06}):
\begin{itemize}\setlength{\itemsep}{-1ex}
\item Objects are (essentially finite) groupoids.
\item Morphisms are spans of essentially finite groupoids.
\item 2-Morphisms are isomorphism classes of spans of span-maps.
\item Composition is given by forming weak fiber products.
\item The monoidal structure is given by the cartesian product $\times$ of groupoids.
\end{itemize}
\begin{Proposition}[\cite{Morton}]\label{conf}
$\mathcal A_G$ induces a symmetric monoidal 2-functor
$$\widetilde{\mathcal A_G}: \mathfrak{C}\mathrm{ob}(1,2,3) \to \mathfrak{S}\mathrm{pan}.$$
This functor assigns to a 1-dimensional manifold $S$
the groupoid $\mathcal A_G(S)$, to a 2-dimensional
cobordism $S \hookrightarrow \Sigma \hookleftarrow S'$
the span $\mathcal A_G(S) \longleftarrow ~\mathcal A_G(\Sigma) \longrightarrow \mathcal A_G(S')$ and to a 3-cobordism with corners a span of
span-maps.
\end{Proposition}
\begin{proof}
It only remains to be shown that composition of morphisms
and the monoidal structure is respected. The first assertion
is shown in \cite[theorem 2]{Morton} and the second
assertion follows immediately from the fact that bundles
over disjoint unions are given by pairs of bundles over the
components, i.e.\
$\mathcal A_G(M \sqcup M') = \mathcal A_G(M) \times \mathcal A_G(M')$.
\end{proof}
The second step in the construction of extended Dijkgraaf-Witten theory is the 2-linearization
of \cite{Morton08}. As we have explained in section
\ref{sec:motivation}, the idea is to associate to a
groupoid $\Gamma$ its category of vector bundles
$\textnormal{Vect}_{\mathbb K}(\Gamma)$. If $\Gamma$ is essentially finite, the category of vector bundles is
conveniently defined as the functor category $\big[\Gamma,\textnormal{Vect}_{\mathbb K}\big]$.
If ${\mathbb K}$ is algebraically closed of characteristic zero, this category is a 2-vector space,
see \cite[Lemma 4.1.1]{Morton08}.
\begin{itemize}
\item
The assignment $\Gamma \mapsto \textnormal{Vect}_{\mathbb K}\big(\Gamma\big) := \big[\Gamma,\textnormal{Vect}_{\mathbb K}\big]$ is a
contravariant 2-functor from the bicategory of (essentially finite) groupoids to the 2-category
of 2-vector spaces. Functors between groupoids are sent to pullback functors.
\end{itemize}
We next need to explain what 2-linearization assigns to spans of groupoids.
To this end, we use the following lemma due to
\cite[4.2.1]{Morton08}:
\begin{Lemma}
Let $f: \Gamma \to \Gamma'$ be a functor between essentially finite groupoids. Then the pullback functor
$ f^*: \textnormal{Vect}\big(\Gamma'\big) \to \textnormal{Vect}\big(\Gamma\big) $
admits a 2-sided adjoint $ f_*: \textnormal{Vect}\big(\Gamma\big) \to \textnormal{Vect}\big(\Gamma'\big)$, called the
\textit{pushforward}.
\end{Lemma}
Two-sided adjoints are also called `ambidextrous' adjoint,
see \cite[ch. 5]{bartlett} for a discussion.
We use this pushforward to associate to a span $$ \xymatrix{
\Gamma & \Lambda \ar_-{~p_0}[l] \ar^-{p_1}[r] & \Gamma' } $$
of (essentially finite) groupoids the \textit{`pull-push'-functor}
$$ (p_1)_* \circ (p_0)^* : \quad \textnormal{Vect}_{\mathbb K}\big(\Gamma\big) \longrightarrow \textnormal{Vect}_{\mathbb K}\big(\Gamma'\big).$$
A similar construction \cite{Morton08} associates to spans of span-morphisms a natural transformation.
Altogether we have:
\begin{Proposition}[\cite{Morton08}]\label{lin2}
The functor $\Gamma \mapsto \textnormal{Vect}_{\mathbb K}(\Gamma)$ can be
extended to a symmetric monoidal 2-functor on the category of spans of groupoids
$$ \widetilde{{\mathcal V}_{\mathbb K}}: \mathfrak{S}\mathrm{pan} \to 2\textnormal{Vect}_{{\mathbb K}}.$$
This 2-functor is called \emph{2-linearization}.
\end{Proposition}
\begin{proof}
The proof that $\widetilde{{\mathcal V}_{\mathbb K}} $ is a 2-functor is in
\cite{Morton08}. The fact that $\widetilde{{\mathcal V}_{\mathbb K}} $ is
monoidal follows from the fact that
$\textnormal{Vect}_{\mathbb K}\big(\Gamma \times \Gamma'\big) \cong
\textnormal{Vect}_{\mathbb K}\big(\Gamma\big) \boxtimes
\textnormal{Vect}_{\mathbb K}\big(\Gamma'\big)$ for a product
$\Gamma \times \Gamma'$ of essentially finite groupoids.
\end{proof}
Arguments similar to the ones in \cite[prop 1.10]{Daw04}
which are based on the universal property of the span
category can be used to show that such an extension is
essentially unique.
We are now in a position to give the functor $Z_G$
described in proposition \ref{exDW} which is
Dijkgraaf-Witten theory as an extended 3d TFT
as the composition of functors
$$ Z_G := \widetilde{{\mathcal V}_{\mathbb K}} \circ \widetilde{\mathcal A_G} :
\quad \mathfrak{C}\mathrm{ob}(1,2,3) \longrightarrow 2\textnormal{Vect}_{\mathbb K}.\qquad$$
It follows from propositions \ref{conf} and \ref{lin2} that $Z_G$
is an extended 3d TFT in the sense of definition \ref{ext}.
For the proof of proposition \ref{exDW}, it remains to be shown
that $Z_G|_\emptyset$ is the Dijkgraaf-Witten 3d TFT from
proposition \ref{DW}; this follows from a
calculation which can be found in \cite[Section 5.2]{Morton}.
\subsection{Evaluation on the circle}\label{evaluation}
The goal of this subsection is a more detailed discussion
of extended Dijkgraaf-Witten theory $Z_G$ as
described in proposition \ref{exDW}. Our focus is
on the object assigned to the 1-manifold $\mathbb{S}^1$ given
by the circle with its standard orientation.
We start our discussion by evaluating an arbitrary
extended 3d TFT $Z$ as in definition \ref{ext}
on certain manifolds of different dimensions:
\begin{enumerate}
\item
To the circle $\mathbb{S}^1$, the extended TFT assigns a ${\mathbb K}$-linear, abelian finitely semisimple category ${\mathcal C}_Z:=Z(\mathbb{S}^1)$.
\item
To the two-dimensional sphere with three boundary components,
two incoming and one outgoing, also known as the \emph{pair of pants},
\begin{center}
\includegraphics{Hose.pdf}
\end{center}
the TFT associates a functor
$$\ \otimes:\,\, {\mathcal C}_Z\boxtimes{\mathcal C}_Z \to{\mathcal C}_Z\ \,\, , $$
which turns out to provide a tensor product on the
category ${\mathcal C}_Z$.
\item The figure
\begin{center}
\includegraphics{Dose.pdf}
\end{center}
shows a 2-morphism between two
three-punctured spheres, drawn as the upper and lower lid.
The outgoing circle is drawn as the boundary of the big disk.
To this cobordism, the TFT associates a natural
transformation
$$\otimes \Rightarrow \otimes^{{\rm opp}} $$
which turns out to be a braiding.
\end{enumerate}
Moreover, the TFT provides coherence cells, in particular associators and relations
between the given structures. This endows the category ${\mathcal C}_Z$ with much additional
structure. This structure can be summarized as follows:
\begin{Proposition}
For $Z$ an extended 3d TFT, the category ${\mathcal C}_Z := Z(\mathbb{S}^1)$ is naturally endowed with
the structure of a braided tensor category.
\end{Proposition}
For details, we refer to \cite{Freed92} \cite{freed94} \cite{freed99} and \cite{CY99}.
This is not yet the complete structure that can be extracted:
from the braiding-picture above it is intuitively clear that the braiding is not symmetric;
in fact, the braiding is `maximally non-symmetric' in a precise sense that is explained in
definition \ref{modularity}. We discuss this in the next section for the category obtained
from the Dijkgraaf-Witten extended TFT.
\medskip
We now specialize to the case of extended Dijkgraaf-Witten
TFT $Z_G$. We first determine
the category ${\mathcal C}(G) := {\mathcal C}_Z$; it is by definition
$$ {\mathcal C}(G) = \big[ \mathcal A_G(\mathbb{S}^1) , \textnormal{Vect}_{\mathbb K} \big]. $$
It is a standard result in the theory of coverings that
$G$-covers on $\mathbb{S}^1$ are described by group homomorphisms $\pi_1(\mathbb{S}^1)
\to G$ and their morphisms by group elements acting by
conjugation. Thus the category $\mathcal A_G(\mathbb{S}^1)$
is equivalent to the action groupoid
$G//G$ for the conjugation action. As a consequence,
we obtain the abelian category
${\mathcal C}(G) \cong [G // G, \textnormal{Vect}_{\mathbb K} ]$.
We spell out this functor category explicitly:
\begin{Proposition}\label{DWcat}
For the extended Dijkgraaf-Witten 3d TFT, the category
${\mathcal C}(G)$ associated to the circle $\mathbb{S}^1$ is given by
the category of $G$-graded vector spaces $V = \bigoplus_{g \in G} V_g$
together with a $G$-action on $V$ such that for all
$x,y\in G$
$$x . V_g \subset V_{xgx^{-1}}\,\,\, .$$
\end{Proposition}
As a next step we determine the tensor product on ${\mathcal C}(G)$.
Since the fundamental group of the pair
of pants is the free group on two generators, the relevant
category of $G$-bundles is
equivalent to the action groupoid $(G \times G) // G$ where $G$
acts by simultaneous conjugation on the two copies of $G$.
The 2-linearization $\widetilde{{\mathcal V}_{\mathbb K}}$ on the span
$$(G//G) \times (G//G) \leftarrow (G\times G) //G \rightarrow G//G.$$
is treated in detail in \cite[rem.\ 5]{Morton}; the result
of this calculation yields the following tensor product:
\begin{Proposition}\label{prop:2.17}
The tensor product of $V$ and $W$ is given by the $G$-graded vector space
$$ (V \otimes W)_g = \bigoplus_{st =g} V_s \otimes W_t $$
together with the $G$-action $g. (v,w) = (gv, gw)$. The associators are the
obvious ones induced by the tensor product in $\textnormal{Vect}_{\mathbb K}$.
\end{Proposition}
In the same vein, the braiding can be calculated:
\begin{Proposition} \label{prop:2.18}
The braiding $V \otimes W \to W \otimes V$ is for $v \in V_g$ and $w \in W$ given by
$$ v \otimes w \mapsto gw \otimes v.$$
\end{Proposition}
\subsection{Drinfel'd double and modularity}\label{dd}
The braided tensor category ${\mathcal C}(G)$ we just computed
from the last section has a well-known description as the
category of modules over
a braided Hopf-algebra ${\mathcal D}(G)$,
the Drinfel'd double ${\mathcal D}(G):={\mathcal D}({\mathbb K}[G])$ of the group algebra
${\mathbb K}[G]$ of $G$, see e.g.\ \cite[Chapter 9.4]{kassel1995quantum}
The Hopf-algebra ${\mathcal D}(G)$ is defined as follows:
As a vector space, ${\mathcal D}(G)$ is the tensor product ${\mathbb K}(G) \otimes {\mathbb K}[G]$ of the algebra of functions on $G$
and the group algebra of $G$, i.e.\ we have the canonical basis $(\delta_g \otimes h)_{g,h\in G}$. The algebra structure can be described as a
smash product (\cite{montgomery1993hopf}),
an analogue of the semi-direct product for groups: in the canonical basis, we have
\[(\delta_{g} \otimes h)(\delta_{g'}\otimes h')
= \left\{\begin{array}{l} \delta_{g}\otimes hh'
\text{ for } g=hg'h^{-1} \\[.2em]
0 \text{ else.}
\end{array}\right.
\]
where the unit is given by the tensor product of the two units:
$\sum_{g \in G} \delta_g \otimes 1$.
The coalgebra structure of ${\mathcal D}(G)$ is given by the
tensor product of the coalgebras ${\mathbb K}(G)$
and ${\mathbb K}[G]$, i.e.\ the coproduct reads
\[\Delta(\delta_g \otimes h) = \sum_{g'g'' = g} (\delta_{g'}\otimes h) \otimes (\delta_{g''}\otimes h)\]
and the counit is given by
$\epsilon(\delta_1 \otimes h) = 1$ and
$\epsilon(\delta_g \otimes h) = 0$ for $g\neq 1$ for all
$h\in G$.
It can easily be checked that this defines a bialgebra structure on ${\mathbb K}(G) \otimes {\mathbb K}[G]$ and that furthermore the linear map
\[S:(\delta_g \otimes h) \mapsto (\delta_{h^{-1}g^{-1}h} \otimes h^{-1})\]
is an antipode for this bialgebra so that ${\mathcal D}(G)$ is
a Hopf algebra. Furthermore, the element
\[R := \sum_{g,h\in G} (\delta_g \otimes 1)\otimes (\delta_h \otimes g) \in {\mathcal D}(G)\otimes {\mathcal D}(G)\]
is a universal R-matrix, which fulfills the defining identities of a braided bialgebra and corresponds to the braiding in
proposition \ref{prop:2.18}. At last, the element
\[\theta := \sum_{g \in G} (\delta_g \otimes g^{-1}) \in {\mathcal D}(G)\]
is a ribbon-element in ${\mathcal D}(G)$, which gives ${\mathcal D}(G)$
the structure of a
ribbon Hopf-algebra (as defined in \cite[Definition 14.6.1]{kassel1995quantum}).
Comparison with propositions \ref{prop:2.17} and
\ref{prop:2.18} shows
\begin{Proposition}
The category ${\mathcal C}(G)$ is isomorphic, as a braided tensor category, to the category ${\mathcal D}(G)\text{-}\mathrm{mod}$.
\end{Proposition}
The category ${\mathcal D}(G)\text{-}\mathrm{mod}$ is actually endowed with more structure than the one of a braided monoidal
category. Since ${\mathcal D}(G)$ is a ribbon Hopf-algebra, the category of representations ${\mathcal D}(G)\text{-}\mathrm{mod}$ has
also dualities and a compatible twist, i.e. has the structure of a ribbon
category (see \cite[Proposition 16.6.2]{kassel1995quantum}
or \cite[Def. 2.2.1]{BKLec} for the notion of a ribbon category). Moreover, the category ${\mathcal D}(G)\text{-}\mathrm{mod}$ is a
2-vector space over ${\mathbb K}$ and thus, in particular, finitely semi-simple. We finally make explicit the
non-degeneracy condition on the braiding that was mentioned in the last subsection.
\begin{Definition}\label{modularity}
\begin{enumerate}
\item
Let ${\mathbb K}$ be an algebraically closed field of characteristic zero.
A \emph{premodular tensor category} over ${\mathbb K}$ is a
${\mathbb K}$-linear, abelian, finitely semisimple category
${\mathcal C}$ which has the structure of a ribbon category
such that the tensor product is linear in each variable and the tensor unit is absolutely
simple, i.e.\ $\mathrm{End}(\mathbf{1}) = {\mathbb K}$.
\item
Denote by $\Lambda_{{\mathcal C}}$ a set of representatives for the
isomorphism classes of simple objects. The braiding on ${\mathcal C}$
allows to define the
\emph{S-matrix} with entries in the field ${\mathbb K}$
\[s_{XY} := tr(R_{YX} \circ R_{XY}) \,\, ,\]
where $X,Y \in \Lambda_{{\mathcal C}}$.
A premodular category is called \emph{modular}, if the S-matrix is invertible.
\end{enumerate}
\end{Definition}
In the case of the Drinfel'd double, the S-matrix can be
expressed explicitly in terms of characters of finite
groups \cite[Section 3.2]{BKLec}. Using orthogonality relations, one shows:
\begin{Proposition}\label{modular-Drinfeld}
The category ${\mathcal C}(G) \cong {\mathcal D}(G)\text{-}\mathrm{mod}$ is modular.
\end{Proposition}
The notion of a modular tensor category first arose as a formalization of
the Moore-Seiberg data of a two-dimensional rational conformal field theory.
They are the input for the Turaev-Reshetikhin construction
of three-dimensional topological field theories.
\section{Equivariant Dijkgraaf-Witten theory}
We are now ready to turn to the construction of equivariant
generalization of the results of section \ref{sec:DW}.
We denote again by $G$ a finite group.
Equivariance will be
with respect to another finite group $J$ that acts on $G$ in a way
we will have to explain. As usual, `twisted sectors'
\cite{VW95} have to be taken into account for a consistent equivariant
theory. A description of these twisted sectors in terms
of bundles twisted by $J$-covers is one important
result of this section.
\subsection{Weak actions and extensions}
Our first task is to identify the appropriate definition of a $J$-action.
The first idea that comes to mind -- a genuine action of
the group $J$ acting on $G$ by group automorphisms --
turns out to need a modification. For reasons that will become apparent
in a moment, we only require an action up to inner automorphism.
\begin{Definition}\label{Def:action}
\begin{enumerate}
\item
A \emph{weak action} of a group $J$ on a group $G$ consists
of a collection of group automorphisms $\rho_j: G \to G$,
one automorphism for each $j \in J$, and a collection of group
elements $c_{i,j} \in G$, one group element for each pair
of elements $i,j \in J$. These data are required to obey
the relations:
$$
\rho_i \circ \rho_j = \mathrm{Inn}_{c_{i,j}} \circ \rho_{ij} \qquad
\rho_i(c_{j,k}) \cdot c_{i,jk} = c_{i,j} \cdot c_{ij,k} \quad \text{ and } \quad c_{1,1} = 1
$$
for all $i,j,k \in J$. Here $\mathrm{Inn}_g$ denotes the inner
automorphism $G \to G$ associated to an element $g \in G$.
We will also use the short hand notation $ ~^j\!g := \rho_j(g)$.
\item Two weak actions $\big(\rho_j,c_{i,j})$
and $\big(\rho'_j,c'_{i,j})$ of a group $J$ on a group $G$
are called isomorphic, if there is a collection of group elements
$h_j \in G$, one group element for each $j \in J$, such that
$$ \rho_j' = \mathrm{Inn}_{h_j} \circ \rho_j \quad \text{ and } c'_{ij} \cdot h_{ij} = h_i \cdot \rho_i(h_j) \cdot c_{ij}$$
\end{enumerate}
\end{Definition}
\begin{Remark}\label{rem:3.2}
\begin{enumerate}
\item
If all group elements $c_{i,j}$ equal the neutral element,
$c_{i,j} = 1$, the weak action reduces to a strict action
of $J$ on $G$ by group automorphisms.
\item A weak action induces a strict action of $J$ on the
group $\mathrm{Out}(G)={\rm Aut}(G)/\mathrm{Inn}(G)$ of outer automorphisms.
\item
In more abstract terms,
a weak action amounts to a (weak) 2-group homomorphism
$J \to {\rm AUT}(G)$. Here
${\rm AUT}(G)$ denotes the automorphism
2-group of $G$. This automorphism 2-group can be described
as the monoidal category of endofunctors of the
one-object-category with morphisms $G$. The group $J$
is considered as a discrete 2-group with only identities
as morphisms. For more details
on 2-groups, we refer to \cite{baez04}.
\end{enumerate}
\end{Remark}
Weak actions are also known under the name \textit{Dedecker cocycles},
due to the work \cite{Ded60}.
The correspondence between weak actions and extensions
of groups is also termed \textit{Schreier theory},
with reference to \cite{Schreier}. Let us briefly sketch
this correspondence:
\begin{itemize}
\item
Let $\big(\rho_j, c_{i,j}\big)$ be a weak action of $J$ on $G$.
On the set $H := G \times J$, we define a multiplication by
\begin{equation} \label{mult}
(g,i) \cdot (g',j) := \big(g \cdot ~^i (g') \cdot c_{i,j}~, ~ij\big).
\end{equation}
One can check that this turns $H$ into a group in such
a way that the sequence $G \to H \to J$ consisting of the
inclusion $g \mapsto (g,1)$ and the projection
$(g,j) \mapsto j$ is exact.
\item
Conversely, let $G \longrightarrow H \stackrel{\pi}{\longrightarrow} J$
be an extension of groups. Choose a set theoretic section
$s: J \to H$ of $\pi$ with $s(1) = 1$. Conjugation with
the group element $s(j)\in H$ leaves the normal
subgroup $G$ invariant. We thus obtain for
$j\in J$\ the automorphism
$\rho_j(g) := s(j)~ g ~s(j)^{-1}$ of $G$.
Furthermore, the element $c_{i,j} := s(i) s(j) s(ij)^{-1}$
is in the kernel of $\pi$ and thus actually contained in
the normal subgroup $G$. It is then straightforward to
check that $\big(\rho_j,c_{i,j}\big)$ defines a weak
action of $J$ on $G$.
\item Two different set-theoretic sections $s$ and $s'$ of the extension
$G \to H \to J$ differ by a map $J \to G$. This map
defines an isomorphism of the induced weak actions
in the sense of definition \ref{Def:action}.2.
\end{itemize}
We have thus arrived at the
\begin{Proposition}[Dedecker, Schreier]
\label{dedeschreier}
There is a 1-1 correspondence between isomorphism classes of weak actions
of $J$ on $G$ and isomorphism classes of group extensions $G \to H \to J$.
\end{Proposition}
\begin{Remark}
\begin{enumerate}
\item
One can easily turn this statement into an
equivalence of categories. Since we do not
need such a statement in this paper,
we leave a precise formulation to the reader.
\item
Under this correspondence,
\emph{strict actions} of $J$
on $G$ correspond to \emph{split extensions}. This can be easily seen as follows: given a \emph{split} extension $G \to H \to J$, one can choose
the section $J\to H$ as a group homomorphism and thus
obtains a \emph{strict} action of $J$ on $G$. Conversely for a strict action of $J$ on $G$ it is easy to see that the group
constructed in equation (\ref{mult}) is a semidirect
product and thus the sequence of groups splits.
To cover all extensions, we thus really
need to consider weak actions.
\end{enumerate}
\end{Remark}
\subsection{Twisted bundles}\label{sec:twisted}
It is a common lesson from field theory that in an equivariant
situation, one has to include ``twisted sectors'' to
obtain a complete theory.
Our next task is to construct the parameters labeling
twisted sectors for a given weak action of a
finite group $J$ on $G$, with corresponding extension
$G \to H \to J$ of groups and chosen set-theoretic section
$J\to H$. We will adhere to a two-step procedure as
outlined after proposition \ref{lin2}. To this end,
we will first construct for any smooth manifold a
category of twisted bundles. Then, the linearization functor
can be applied to spans of such categories.
We start our discussion of twisted $G$-bundles with
the most familiar case of the circle, $M=\mathbb{S}^1$.
The isomorphism classes of $G$-bundles on $\mathbb{S}^1$ are in bijection to connected
components of the free loop space ${\mathcal L} BG$ of the classifying space $BG$:
$$ Iso\big(\mathcal A_G(\mathbb{S}^1)\big) = \Hom_{\mathrm{Ho(Top)}}(\mathbb{S}^1,BG) = \pi_0({\mathcal L} BG). $$
Given a (weak) action of $J$ on $G$, one can introduce
twisted loop spaces. For any element $j\in J$, we have a group automorphism $j: G \to G$
and thus a homeomorphism $j: BG \to BG$. The $j$-twisted loop space is then defined to be
$$ {\mathcal L}^j BG :=\big\{ f:[0,1]\to BG \mid f(0)=j \cdot f(1) \big\}. $$
Our goal is to introduce for every group element $j\in J$
a category $\mathcal A_G(\mathbb{S}^1,j)$ of $j$-twisted $G$-bundles on $\mathbb{S}^1$ such that
$$ Iso\big(\mathcal A_G(\mathbb{S}^1,j)\big) = \pi_0({\mathcal L}^j BG) \,\,\, . $$
In the case of the circle $\mathbb{S}^1$, the twist parameter was a group element
$j\in J$. A more geometric description uses a family of $J$-covers $P_j$
over $\mathbb{S}^1$, with $j \in J$. The cover $P_j$ is uniquely determined by
its monodromy $j$
for the base point $1 \in \mathbb{S}^1$ and a fixed point in the fiber over $1$.
A concrete construction of the cover $P_j$ is given by the quotient
$P_j := [0,1] \times J / \sim$ where $(0,i) \sim (1,ji)$ for all $i \in J$.
In terms of these $J$-covers, we can write
$$ {\mathcal L}^j BG = \big\{ f: P_j \to BG \mid f \text{ is $J$-equivariant} \big\}.$$
This description generalizes to an arbitrary smooth manifold $M$.
The natural twist parameter in the general case is
a $J$-cover $P\stackrel J\to M$.
Suppose, we have a weak $J$-action on $G$ and construct
the corresponding extension $G\to H \stackrel\pi\to J$.
The category of bundles we
need are $H$-lifts of the given $J$-cover:
\begin{Definition}\label{def:twisted}
Let $J$ act weakly on $G$.
Let $P\stackrel J\to M$ be a $J$-cover over $M$.
\begin{itemize}
\item
A $P$-twisted $G$-bundle over $M$ is a pair $(Q,\varphi)$, consisting of an
$H$-bundle $Q$ over $M$
and a smooth map $\varphi: Q \to P$ over $M$ that is required to obey
$$ \varphi(q\cdot h) = \varphi(q) \cdot \pi(h) \,\,\, $$
for all $q\in Q$ and $h\in H$. Put differently, a
$P\stackrel J\to M$-twisted $G$-bundle is
a lift of the $J$-cover $P$
reduction along the group homomorphism
$\pi:\ H\to J$.
\item
A morphism of $P$-twisted bundles $(Q,\varphi)$ and $(Q',\varphi')$ is a morphism
$f: Q \to Q'$ of $H$-bundles such that $\varphi' \circ f = \varphi$.
\item We denote the category of $P$-twisted $G$-bundles by $\mathcal A_G\big(P \to M\big)$.
For $M= \mathbb{S}^1$, we introduce the abbreviation $\mathcal A_G\big(\mathbb{S}^1,j) := \mathcal A_G\big(P_j \to \mathbb{S}^1\big)$ for the standard covers of the circle.
\end{itemize}
\end{Definition}
\begin{Remark}
There is an alternative point of view on a $P$-twisted bundle $(Q,\varphi)$:
the subgroup $G\subset H$ acts on the total space $Q$ in such a way that the map $\varphi: Q \to P$
endows $Q$ with the structure of a $G$-bundle on $P$. Both the structure group
$H$ of the bundle $Q$ and the bundle $P$ itself carry an action of $G$; for twisted
bundles, an equivariance condition on this action has to be imposed.
Unfortunately this equivariance property is relatively involved; therefore, we have opted
for the definition in the form given above.
\end{Remark}
A morphism $f: P \to P'$ of $J$-covers over the same manifold induces a functor
$f_*: \mathcal A_G\big(P \to M\big) \to \mathcal A_G\big(P' \to M\big)$ by $f_*(Q,\varphi) := (Q,f \circ \varphi)$.
Furthermore, for a smooth map $f: M \to N$, we can pull back the twist data $P\to M$ and
get a \textit{pullback functor} of twisted $G$-bundles:
$$f^*: \mathcal A_G\big(P \to N\big) \to \mathcal A_G\big(f^*P \to M\big)$$
by $f^*(Q,\varphi)= (f^*Q,f^*\varphi)$.
Before we discuss more sophisticated properties of twisted
bundles, we have to
make sure that our definition is consistent with `untwisted' bundles:
\begin{Lemma} \label{trivbundle}
Let the group $J$ act weakly on the group $G$. For $G$-bundles
twisted by the trivial $J$-cover $M\!\times\!J \to M$, we
have a canonical equivalence of categories
$$
\mathcal A_G\big(M\!\times\!J \to M\big)\cong \mathcal A_G(M) .
$$
\end{Lemma}
\begin{proof}
We have to show that for an element $(Q,\varphi) \in
\mathcal A_G\big(M\!\times\!J \to M\big)$ the $H$-bundle $Q$
can be reduced to a $G$-bundle. Such a reduction is the
same as a section of the associated fiber bundle
$\pi_*(Q) \in \mathcal{B}{\!un}_J(M)$
see e.g. \cite[Satz 2.14]{baum}).
Now $\varphi: Q\to M\times J$ induces an isomorphism
of $J$-covers $Q\times_H\ J\cong (M\times J)\times_H\ J
\cong M\times J$ so that the bundle $Q\times_H\ J$
is trivial as a $J$-cover and in particular admits
global sections.
Since morphisms of twisted bundles have to commute with these
sections, we obtain in that way a functor
$\mathcal A_G\big(M\!\times\!J \to M\big) \to \mathcal A_G(M)$. Its inverse
is given by extension of $G$-bundles on $M$ to $H$-bundles on $M$.
\end{proof}
We also give a description of twisted bundles using
standard covering theory; for an alternative description
using \v{C}ech-cohomology, we refer to appendix \ref{cech}.
We start by recalling the following standard fact
from covering theory, see e.g.\ \cite[1.3]{hatcher}
that has already been used to prove proposition \ref{DWcat}:
for a finite group $J$, the category of $J$-covers
is equivalent to the action groupoid
$\Hom(\pi_1(M),J) //J$. (Note that this equivalence
involves choices and is not canonical.)
To give a similar description of twisted bundles, fix
a $J$-cover $P$. Next, we choose
a basepoint $m \in M$ and a point $p$ in the fiber $P_m$
over $m$. These data determine a unique group
morphism $\omega: \pi_1(M,m) \to J$ representing $P$.
\begin{Proposition}\label{covering}
Let $J$ act weakly on $G$.
Let $M$ be a connected manifold and $P$ be a $J$-cover
over $M$ represented after the choices just indicated
by the group homomorphism $\omega: \pi_1(M) \to J$.
Then there is a (non-canonical) equivalence of categories
$$ \mathcal A_G\big(P \to M \big) \cong \Hom^\omega\big(\pi_1(M),H\big) // G $$
where we consider group homomorphisms
$$ \Hom^\omega\big(\pi_1(M),H\big) := \big\{\mu: \pi_1(M) \to H \mid \pi \circ \mu = \omega \big\} $$
whose composition restricts to the group
homomorphism $\omega$ describing the $J$-cover $P$. The group
$G$ acts on $\Hom^\omega\big(\pi_1(M),H\big)$
via pointwise conjugation using the inclusion
$G \to H$.
\end{Proposition}
\begin{proof}
Let $m\in M$ and $p\in P$ over $m$ be the choices of
base point in the $J$-cover $P\to M$ that lead to
the homomorphism $\omega$. Consider a $(P\to M)$ twisted
bundle $Q\to M$. Since $\varphi:\ Q\to P$ is
surjective, we can choose a base point $q$ in the fiber
of $Q$ over $m$ such that $\varphi(q)=p$. The
group homomorphism $\pi_1(M)\to H$ describing the
$H$-bundle $Q$
is obtained by lifting closed paths in $M$ starting in
$m$ to paths in $Q$ starting
in $q$. They are mapped under $\varphi$ to
lifts of the same path to $P$ starting at $p$, and
these lifts are just described by the group homomorphism
$\omega: \pi_1(M)\to J$ describing the cover $P$.
If the end point of the path in $Q$ is
$qh$ for some $h\in H$, then by the defining property
of $\varphi$, the lifted path in $P$ has endpoint
$\varphi(qh)=\varphi(q)\pi(h)= p\pi(h)$. Thus
$\pi\circ\mu=\omega$.
\end{proof}
\begin{Remark}
For non-connected manifolds, a description as in proposition
\ref{covering} can be obtained for every component.
Again the equivalence involves choices of base points on
$M$ and in the fibers over the base points. This could
be fixed by working with pointed manifolds, but pointed
manifolds cause problems when we consider cobordisms.
Alternatively, we could use the fundamental groupoid
instead of the fundamental group, see e.g. \cite{may}.
\end{Remark}
\begin{Example}\label{ex_twisted}
We now calculate the categories of twisted bundles over
certain manifolds using proposition \ref{covering}.
\begin{enumerate}\setlength{\itemsep}{-0.5ex}
\item For the circle $\mathbb{S}^1$,
$\omega\in\Hom(\pi_1(\mathbb{S}^1),J)=\Hom(\mathbb Z,J)$ is determined
by an element $j\in J$ and the condition $\pi\circ\mu=
\omega$ requires $\mu(1)\in H$ to be in the preimage
$H_j:=\pi^{-1}(j)$ of $j$. Thus, we have $\mathcal A_G(\mathbb{S}^1,j)
\cong H_j // G$.
\item For the 3-Sphere $\mathbb{S}^3$, all twists $P$ and
all $G$-bundles are trivial. Thus, we have
$\mathcal A_G(P \to \mathbb{S}^3) \cong \mathcal A_G(\mathbb{S}^3)
\cong pt // G$.
\end{enumerate}
\end{Example}
\subsection{Equivariant Dijkgraaf-Witten theory}\label{subsec:eqDW}
The key idea in the construction of equivariant
Dijkgraaf-Witten theory is to take twisted bundles
$\mathcal A_G(P \to M)$ as the field configurations, taking
the place of $G$-bundles in section \ref{sec:DW}.
We cannot expect to get then invariants of closed
3-manifolds $M$, but rather invariants of 3-manifolds $M$
together with a twist datum, i.e.\ a $J$-cover
$P$ over $M$. Analogous statements apply to
manifolds with boundary and cobordisms. Therefore we
need to introduce extended cobordism-categories as
$\mathfrak{C}\mathrm{ob}(1,2,3)$ in definition \ref{cob123}, but endowed
with the extra datum of a $J$-cover over each manifold.
\begin{Definition}\label{defcobj}
$\mathfrak{C}\mathrm{ob}^J(1,2,3)$ is the following symmetric monoidal bicategory:
\begin{itemize}\setlength{\itemsep}{-0.5ex}
\item { Objects} are compact, closed, oriented 1-manifolds $S$, together with a $J$-cover $P_S \stackrel{J}{\to} S$.
\item {1-Morphisms} are collared cobordisms
$$S \times I \hookrightarrow \Sigma \hookleftarrow S' \times I$$
where $\Sigma$ is a 2-dimensional, compact, oriented cobordism, together
with a $J$-cover $P_\Sigma \to \Sigma$ and isomorphisms
$$ P_\Sigma|_{(S \times I)} \stackrel{\sim}{\longrightarrow} P_S \times I \qquad \text{ and } \qquad P_\Sigma|_{(S' \times I)} \stackrel{\sim}{\longrightarrow} P_{S'} \times I.$$
over the collars.
\item {2-Morphisms} are generated by
\begin{enumerate}[-]\setlength{\itemsep}{-0.5ex}
\item orientation preserving diffeomorphisms $\varphi: \Sigma \to \Sigma'$ of cobordisms fixing the collar together with an
isomorphism $\widetilde\varphi: P_\Sigma \to P_{\Sigma'}$ covering $\varphi$.
\item 3-dimensional collared, oriented cobordisms with corners $M$ with
cover $P_M \to M$ together with covering isomorphisms over the collars
(as before) up
to diffeomorphisms preserving the orientation and boundary.
\end{enumerate}
\item Composition is by gluing cobordisms and covers along collars.
\item The monoidal structure is given by disjoint union.
\end{itemize}
\end{Definition}
\begin{Remark}
In analogy to remark \ref{smoothgluing}, we point out
that the isomorphisms of covers are defined
over the collars, rather than only over the the boundaries.
This endows the glued cover with a well-defined smooth structure.
\end{Remark}
\begin{Definition}\label{defjtft}
An extended 3d $J$-TFT is a symmetric monoidal 2-functor
$$ Z: \mathfrak{C}\mathrm{ob}^J(1,2,3) \to 2\textnormal{Vect}_{\mathbb K}.$$
\end{Definition}
Just for the sake of completeness, we will also give a
definition of non-extended $J$-TFT. Therefore define the
symmetric monoidal category
$\mathfrak{C}\mathrm{ob}^J(2,3)$ to be the endomorphism category of the
monoidal unit $\emptyset$ in $\mathfrak{C}\mathrm{ob}(1,2,3)$. More concretely, this
category has as
objects closed, oriented 2-manifolds with $J$-cover and as
morphisms $J$-cobordisms between them.
\begin{Definition}
A (non-extended) 3d $J$-TFT is a symmetric monoidal 2-functor
$$ \mathfrak{C}\mathrm{ob}^J(2,3) \to \textnormal{Vect}_{\mathbb K}.$$
\end{Definition}
Similarly as in the non-equivariant case (lemma \ref{restriction}),
we get
\begin{Lemma}
Let $Z$ be an extended 3d $J$-TFT. Then $Z|_{\emptyset}$ is a
(non-extended) 3d $J$-TFT.
\end{Lemma}
Now we can state the main result of this section:
\begin{Theorem} \label{Thm1}
For a finite group $G$ and a weak $J$-action on $G$, there is
an extended 3d $J$-TFT called $Z_G^J$
which assigns the categories
$$ \textnormal{Vect}_{\mathbb K}\big(\mathcal A_G(P \to S)\big) =
\big[ \mathcal A_G(P \to S) , \textnormal{Vect}_{\mathbb K} \big] $$
to 1-dimensional, closed oriented manifolds $S$ with $J$-cover $P \to S$.
\end{Theorem}
We will give a proof of this theorem in the next sections.
Having twisted bundles at our disposal,
the main ingredient will again be the 2-linearization
described in section \ref{lin}.
\subsection{Construction via spans}\label{subsec:spans}
As in the case of ordinary Dijkgraaf-Witten theory, cf.\ section
\ref{lin}, equivariant Dijkgraaf-Witten $Z_G^J$
theory is constructed as the composition of the symmetric monoidal 2-functors
$$\widetilde{\mathcal A_G}: \mathfrak{C}\mathrm{ob}^J(1,2,3) \to \mathfrak{S}\mathrm{pan} \quad \text{ and }
\quad \widetilde{{\mathcal V}_{\mathbb K}} : \mathfrak{S}\mathrm{pan} \to 2\textnormal{Vect}_{\mathbb K}.$$
The second functor will be exactly the 2-linearization
functor of proposition \ref{lin2}.
Hence we can limit our discussion to the construction of the first
functor $\widetilde{\mathcal A_G}$.
As it will turn out, our definition of twisted bundles is
set up precisely in such a way that the
construction of the corresponding functor in proposition \ref{conf} can be generalized. \\
Our starting point is the following observation:
\begin{itemize}
\item
The assignment $(P_M \stackrel{J}{\to} M) \longmapsto
\mathcal A_G(P_M \stackrel J\to M)$ of
twisted bundles to a twist datum $P_M\to M$ constitutes
a contravariant 2-functor from the category of manifolds with
$J$-cover to the 2-category of groupoids. Maps between manifolds with cover are mapped to
the corresponding pullback functors of bundles.
\end{itemize}
From this functor which is defined on manifolds of any dimension, we
construct a functor $\widetilde{\mathcal A_G}$ on $J$-cobordisms
with values in the 2-category $\mathfrak{S}\mathrm{pan}$ of spans of groupoids,
where the category $\mathfrak{S}\mathrm{pan}$ is defined in section \ref{lin}.
To an object in $\mathfrak{C}\mathrm{ob}^J(1,2,3)$, i.e.\
to a $J$-cover $P_S\to M$, we assign
the category $\mathcal A_G(P_S \to S)$ of $J$-covers. To a 1-morphism
$ P_S \hookrightarrow P_\Sigma \hookleftarrow P_S'$ in
$\mathfrak{C}\mathrm{ob}^J(1,2,3)$, we associate the span
\begin{equation}\label{span1}
\mathcal A_G(P_S \to S) \leftarrow \mathcal A_G(P_\Sigma \to \Sigma) \rightarrow \mathcal A_G(P_{S'} \to S')
\end{equation}
and to a 2-morphism of the type $P_\Sigma \hookrightarrow P_M \hookleftarrow P_{\Sigma'}$ the span
\begin{equation}\label{span2}
\mathcal A_G(P_\Sigma \to \Sigma) \leftarrow \mathcal A_G(P_M \to M) \rightarrow \mathcal A_G(P_{\Sigma'} \to \Sigma')\text{.}
\end{equation}
We have to show that this defines a symmetric monoidal functor
$\widetilde{\mathcal A_G}: \mathfrak{C}\mathrm{ob}^J(1,2,3) \to \mathfrak{S}\mathrm{pan}$.
In particular, we have to show that the composition of morphisms is respected.
\begin{Lemma}\label{kleben}
Let $P_\Sigma \to \Sigma$ and $P_{\Sigma'} \to \Sigma'$ be two 1-morphisms in $\mathfrak{C}\mathrm{ob}^J(1,2,3)$
which can be composed at the object $P_S \to S$ to get
the 1-morphism
$$P_\Sigma \circ P_{\Sigma'}:=
\big(P_\Sigma \sqcup_{P_S\times I} P_{\Sigma'} \to \Sigma \sqcup_{S \times I} \Sigma'\big)
\,\, ,$$
where $I = [0,1]$ is the standard interval. (Recall that
we are gluing over collars.)
Then the category $ \mathcal A_G\big(P_\Sigma \circ P_{\Sigma'}\big)$ is the weak pullback of $\mathcal A_G(P_\Sigma \to \Sigma)$ and $\mathcal A_G(P_{\Sigma'} \to \Sigma')$ over $\mathcal A_G(P_S \to S)$.
\end{Lemma}
\begin{proof}
By definition the category $$ \mathcal A_G\big(P_\Sigma \circ P_{\Sigma'}\big)$$ has as objects twisted $G$-bundles over the 2-manifold
$\Sigma \sqcup_{S \times I} \Sigma' =: N$. The manifold $N$ admits an open covering $N = U_0 \cup U_1$ with $U_0 = \Sigma \setminus S$ and $U_1 = \Sigma'\setminus S$ where the intersection is the cylinder
$U_0 \cap U_1 = S \times (0,1)$.
By construction, the restrictions of the glued bundle $P_N \to N$ to $U_0$ and $U_1$ are given by $P_\Sigma \setminus P_S$ and
$P_{\Sigma'} \setminus P_S$.
The natural inclusions $U_0 \to \Sigma$ and $U_1 \to \Sigma'$ induce equivalences
\begin{eqnarray*}
\mathcal A_G(P_\Sigma \to \Sigma) \stackrel{\sim}{\longrightarrow} \mathcal A_G( P_N|_{U_0} \to U_0) \\
\mathcal A_G(P_{\Sigma'} \to \Sigma') \stackrel{\sim}{\longrightarrow} \mathcal A_G( P_N|_{U_1} \to U_1)
\end{eqnarray*}
Analogously, we have an equivalence
$$
\mathcal A_G\big(P_N|_{U_0 \cap U_1} \to U_0 \cap U_1\big) \stackrel{\sim}{\longrightarrow}
\mathcal A_G(P_S \to S) \,\,\, .
$$
At this point, we have reduced the claim to an
assertion about descent of twisted bundles which we will
prove in corollary \ref{desc_pullback}.
This corollary implies that $\mathcal A_G(P_N \to N)$ is
the weak pullback of
$\mathcal A_G(P_N|_{U_0} \to U_0)$ and $\mathcal A_G(P_N|_{U_1} \to U_1)$ over $\mathcal A_G\big(P_N|_{U_0 \cap U_1}\big)$. Since weak pullbacks are invariant under
equivalence of groupoids, this shows the claim.
\end{proof}
We now turn to the promised results about descent of
twisted bundles.
Let $P \to M$ be a $J$-cover over a manifold $M$ and
$\{U_\alpha\}$ be an open covering of $M$, where for the
sake of generality we allow for arbitrary open coverings.
We want to show that twisted bundles can be glued together
like ordinary bundles; while the precise meaning of this statement
is straightforward, we briefly summarize the relevant
definitions for the sake of completeness:
\begin{Definition}
Let $P \to M$ be a $J$-cover over a manifold $M$ and $\{U_\alpha\}$ be an
open covering of $M$. The descent category $\mathcal{D}\!esc(U_\alpha,P)$ has
\begin{itemize}
\item \underline{Objects:} families of $P|_{U_\alpha}$-twisted bundles
$Q_\alpha$ over $U_\alpha$, together with isomorphisms of twisted bundles
$\varphi_{\alpha\beta} : Q_\alpha|_{U_\alpha \cap U_\beta} \stackrel{\sim}{\longrightarrow} Q_\beta|_{U_\alpha \cap U_\beta}$ satisfying the cocycle condition $\varphi_{\alpha\beta} \circ \varphi_{\beta\gamma} = \varphi_{\alpha\gamma}$.
\item \underline{Morphisms:}
families of morphisms $f_\alpha: Q_\alpha \to Q'_\alpha$ of twisted
bundles such that over $U_{\alpha\beta}$ we have
$\varphi'_{\alpha\beta} \circ (f_\alpha)|_{U_{\alpha\beta}} = (f_\beta)|_{U_{\alpha\beta}} \circ \varphi_{\alpha\beta}$.
\end{itemize}
\end{Definition}
\begin{Proposition}[Descent for twisted bundles]
Let $P \to M$ be a $J$-cover over a manifold $M$ and
$\{U_\alpha\}$ be an open covering of $M$. Then the groupoid $\mathcal A_G(P \to M)$
is equivalent to the descent category $\mathcal{D}\!esc(U_\alpha,P)$.
\end{Proposition}
\begin{proof}
Note that the corresponding statements are true for
$H$-bundles and for $J$-covers. Then the description in definition \ref{def:twisted}
of a twisted bundle as an $H$-bundle together with a
morphism of the associated $J$-cover
immediately implies the claim.
\end{proof}
\begin{corollary}\label{desc_pullback}
For an open covering of $M$ by two open sets $U_0$ and $U_1$ the category $\mathcal A_G(P \to M)$ is the weak pullback of $\mathcal A_G(P|_{U_0}\to U_0)$ and $\mathcal A_G(P|_{U_1}\to U_1)$ over $\mathcal A_G(P|_{U_0 \cap U_1}\to U_0 \cap U_1)$.
\end{corollary}
In order to prove that the assignment \eqref{span1} and \eqref{span2} really promotes $\mathcal A_G$ to a symmetric monoidal functor $\widetilde{\mathcal A_G}: \mathfrak{C}\mathrm{ob}^J(1,2,3) \to \mathfrak{S}\mathrm{pan}$, it remains to show that $\mathcal A_G$ preserves the monoidal structure.
Now a bundle over a disjoint union is given by a pair of
bundles over each component. Thus, for a disjoint union
of $J$-manifolds $P \to M = (P_1 \sqcup P_2)
\to (M_1 \sqcup M_2)$, we have
$\mathcal A_G(P \to M) \cong \mathcal A_G(P_1 \to M_1) \times
\mathcal A_G(P_2 \to M_2)$. Note that the manifolds
$M,M_1$ and $M_2$ can also be cobordisms.
The isomorphism of categories is clearly associative and
preserves the symmetric structure. Together
with lemma \ref{kleben}, this proves the next proposition.
\begin{Proposition}
$\mathcal A_G$ induces a symmetric monoidal functor
$$\widetilde{\mathcal A_G}: \mathfrak{C}\mathrm{ob}^J(1,2,3) \to \mathfrak{S}\mathrm{pan}$$
which assigns the spans \eqref{span1} and \eqref{span2} to 2 and 3-dimensional cobordisms with $J$-cover.
\end{Proposition}
\subsection{Twisted sectors and fusion}\label{twisted_sectors_and_fusion
We next proceed to evaluate the
$J$-equivariant TFT $Z_G^J$ constructed in the last section
on the circle, as we did in section \ref{evaluation}
for the non-equivariant TFT.
We recall from section \ref{sec:twisted} the fact
that over the circle $\mathbb{S}^1$ we have for each $j \in J$
a standard cover $P_j$. The associated category
$$ {\mathcal C}(G)_j := Z_G^J\big(P_j \to \mathbb{S}^1\big) $$
is called the $j$-twisted sector of the theory; the sector
${\mathcal C}(G)_1$ is called the neutral sector. By lemma
\ref{trivbundle}, we have an equivalence
$\mathcal A_G(P_1\to \mathbb{S}^1)\cong \mathcal A_G(\mathbb{S}^1)$; hence we
get an equivalence of categories ${\mathcal C}(G)_1\cong{\mathcal C}(G)$,
where ${\mathcal C}(G)$ is the category arising in the
non-equivariant Dijkgraaf-Witten model, we discussed
in section \ref{evaluation}. We
have already computed the twisted sectors as
abelian categories in example
\ref{ex_twisted} and note the result for future reference:
\begin{Proposition}\label{sectors}
For the $j$-twisted sector of equivariant Dijkgraaf-Witten theory,
we have an equivalence of abelian categories
$${\mathcal C}(G)_j \cong [H_j //G,\textnormal{Vect}_{\mathbb K}] \,\, , $$
where $H_j //G$ is the action groupoid given by the conjugation action of $G$ on $H_j := \pi^{-1}(j)$. More concretely, the
category ${\mathcal C}(G)_j$ is equivalent to the category of $H_j$-graded
vector spaces $V = \bigoplus_{h \in H_j} V_h$ together with a
$G$-action on $V$ such that
$$ g.V_h \subset V_{ghg^{-1}}\text{.}$$
\end{Proposition}
As a next step, we want to make explicit additional
structure on the categories ${\mathcal C}(G)_j$ coming from certain
cobordisms. Therefore, consider the pair of pants ${\Sigma(2,1)}$:
\begin{center}
\includegraphics{Hose.pdf}
\end{center}
The fundamental group of ${\Sigma(2,1)}$ is the free group on two
generators. Thus, given a pair of group elements
$j,k\in J$, there is a $J$-cover $P^{\Sigma(2,1)}_{j,k} \to {\Sigma(2,1)}$
which restricts to the standard covers $P_j$ and $P_k$
on the two ingoing boundaries and to the standard cover
$P_{jk}$ on the outgoing boundary circle. (To find a
concrete construction, one should fix a parametrization of
the pair of pants ${\Sigma(2,1)}$.)
The cobordism $P^{\Sigma(2,1)}_{j,k}$ is a morphism
\begin{equation} \label{fusionproduct}
P^{\Sigma(2,1)}_{j,k}: \big(P_j \to \mathbb{S}^1\big) \sqcup \big(P_k \to
\mathbb{S}^1\big) \longrightarrow \big(P_{jk} \to \mathbb{S}^1\big)
\end{equation}
in the category $\mathfrak{C}\mathrm{ob}^J(1,2,3)$. Applying the equivariant
TFT-functor $Z_G^J$ yields a functor
$$ \otimes_{jk} :\quad {\mathcal C}(G)_j \boxtimes {\mathcal C}(G)_k \longrightarrow {\mathcal C}(G)_{jk}.$$
We describe this functor in terms of the equivalent categories
of graded vector spaces as a functor
$$H_j//G\text{-}\mathrm{mod} \times H_k//G\text{-}\mathrm{mod} \to H_{jk}// G\text{-}\mathrm{mod} \,\, .$$
\begin{Proposition}\label{prop:fusionproduct}
For $V= \bigoplus_{h\in H_j} V_h \in H_j//G\text{-}\mathrm{mod} $ and
$W = \bigoplus W_h \in H_k//G\text{-}\mathrm{mod}$ the product
$V \otimes_{jk} W \in H_{jk}//G\text{-}\mathrm{mod}$ is given by
$$ (V \otimes_{jk} W)_{h} = \bigoplus_{st = h} V_s \otimes W_t $$
together with the action $g .(v \otimes w) = g.v \otimes g.w$.
\end{Proposition}
\begin{proof}
As a first step we have to compute the span
$\widetilde{\mathcal A_G}(P^{\Sigma(2,1)}_{j,k})$
associated to the cobordism $P^H_{j,k}$. From
the description of twisted bundles in proposition
\ref{covering} and the fact that the fundamental group of ${\Sigma(2,1)}$ is the
free group on two generators, we derive the following
equivalence of categories:
$$ \mathcal A_G\big(P^{\Sigma(2,1)}_{jk} \to {\Sigma(2,1)}\big) \cong (H_j \times H_k) // G
\,\,\, . $$
Here we have
$H_j \times H_k = \{(h,h') \in H \times H \mid \pi(h)
= j,~ \pi(h') = k\}$, on which $G$ acts by simultaneous
conjugation.
This leads to the span of action groupoids
$$ H_j //G \times H_k //G \longleftarrow (H_j \times H_k) // G
\longrightarrow H_{jk} //G $$
where the left map is given be projection to the factors
and the right hand map by multiplication. Applying the
2-linearization functor
$\widetilde{ {\mathcal V}_{\mathbb K} }$ from proposition \ref{lin2}
amounts to computing the corresponding pull-push
functor. This yields the result.
\end{proof}
Next, we consider the 2-manifold ${\Sigma(1,1)}$ given by the cylinder
over $\mathbb{S}^1$, i.e. ${\Sigma(1,1)} = \mathbb{S}^1 \times I$:
\begin{center}
\includegraphics{Zylinder.pdf}
\end{center}
There exists a cover $P^{\Sigma(1,1)}_{j,x} \to {\Sigma(1,1)}$ for $j,x \in J$
that restricts to $P_{j}$ on the ingoing circle and to
$P_{xjx^{-1}}$ on the outgoing circle. The simplest
way to construct such a cover is to consider the
cylinder $P_{xjx^{-1}} \times I \to \mathbb{S}^1 \times I$ and
to use the identification of
$P^{\Sigma(1,1)}_{j,x}$ over (a collaring neighborhood of) the outgoing
circle by the identity and over the ingoing circle
the identification by the morphism
$P_{\Sigma(1,1)}|_{\mathbb{S}^1 \times {1}} = P_j \to P_{x^{-1}jx}$ given by conjugation with $x$. In this way, we obtain a cobordism that
is a 1-morphism
\begin{equation} \label{jaction}
P^{\Sigma(1,1)}_{j,x}: (P_{j} \to \mathbb{S}^1) \longrightarrow (P_{xjx^{-1} } \to \mathbb{S}^1)
\end{equation}
in the category $\mathfrak{C}\mathrm{ob}^J(1,2,3)$ and hence induces a functor
$$ \phi_x: {\mathcal C}(G)_{j} \to {\mathcal C}(G)_{xjx^{-1}}.$$
We compute the functor on the equivalent action groupoids
explicitly:
\begin{Proposition} \label{prop:action}
The image under $\phi_x$ of an object
$V = \bigoplus V_h \in H_{j}//G\text{-}\mathrm{mod}$ is
the graded vector space with homogeneous component
$$ \phi_x(V)_h = V_{s(x^{-1}) h s\left(x^{-1}\right)^{-1}} $$
for $h\in H_{xjx^{-1}}$ and with $G$-action on $v \in V_h$
given by $s(x^{-1}) g s\left(x^{-1}\right)^{-1} \cdot v$.
\end{Proposition}
\begin{proof}
As before we compute the span $\widetilde{\mathcal A_G}(P^{\Sigma(1,1)}_{j,x})$.
Using explicitly the equivalence given in the proof of proposition
\ref{covering}, we obtain the span of action groupoids
$$ H_{j}//G \leftarrow H_{xjx^{-1}}//G \rightarrow H_{xjx^{-1}}//G $$
where the right-hand map is the identity and the left-hand map is given by
$$ (h,g) \mapsto \big(s(x^{-1}) h s(x^{-1})^{-1}, s(x^{-1}) g s(x^{-1})^{-1}\big)
\,\, . $$
Computing the corresponding pull-push functor, which here in fact only consists of a pullback,
shows the claim.
\end{proof}
Finally we come to the structure corresponding to the
braiding of section \ref{evaluation}. Note that the
cobordism that interchanges the two ingoing circles of
the pair of pants ${\Sigma(2,1)}$, as in the following picture,
\begin{center}
\includegraphics{Dose.pdf}
\end{center}
can also be realized as the diffeomorphism $F: {\Sigma(2,1)} \to {\Sigma(2,1)}$
of the pair of pants
that rotates the ingoing circles counterclockwise around
each other and leaves the outgoing circle fixed.
In this picture, we think of the cobordism as the cylinder
${\Sigma(2,1)} \times I$ where the identification with ${\Sigma(2,1)}$ on the top
is the identity and on the bottom is given by the
diffeomorphism $F$. More explicitly, denote by
$\tau:\ \mathbb{S}^1\times\mathbb{S}^1\to\mathbb{S}^1\times\mathbb{S}^1$ the map that interchanges
the two copies. We then consider the following diagram
in the 2-category $\mathfrak{C}\mathrm{ob}(1,2,3)$:
$$ \xymatrix{
&&{\Sigma(2,1)}\ar@{=>}^F[dd]&& \\
\mathbb{S}^1\times \mathbb{S}^1\ar_\tau[dr]\ar^\iota[urr]&&&&\ar[ull]\ar[dl]\mathbb{S}^1 \\
&\mathbb{S}^1\times\mathbb{S}^1\ar^\iota[rr] &\ &{\Sigma(2,1)}
}$$
where $\iota:\mathbb{S}^1\times\mathbb{S}^1\to {\Sigma(2,1)}$ is the standard inclusion of
the two ingoing boundary circles into the trinion ${\Sigma(2,1)}$.
Our next task is to lift this situation to manifolds
with $J$-covers. On the ingoing trinion, we take the $J$ cover
$P^{\Sigma(2,1)}_{jk}$. We denote the symmetry isomorphism in $\mathfrak{C}\mathrm{ob}^J(1,2,3)$
by $\tau$ as well. Applying the diffeomorphism of the
trinion explicitly, one sees that the outgoing trinion will
have monodromies $jkj^{-1}$ and $j$ on the ingoing circles.
Hence we have to apply a $J$-cover $P_{j,k}^{\Sigma(1,1)}$
of the cylinder ${\Sigma(1,1)}$ first to one insertion.
The next lemma asserts that then the 2-morphism
in $\mathfrak{C}\mathrm{ob}^J(1,2,3)$ is fixed:
\begin{Lemma}\label{braiding-morphism}
In the 2-category $\mathfrak{C}\mathrm{ob}^J(1,2,3)$, there
is a unique 2-morphism
$$\hat F:\quad P^{\Sigma(2,1)}_{j,k} \Longrightarrow \big(P^{\Sigma(2,1)}_{jkj^{-1},j}\big)
\circ \tau \circ \big(\mathrm{id} \sqcup P^{\Sigma(1,1)}_{j,k}\big) $$
that covers the 2-morphism $F$ in $\mathfrak{C}\mathrm{ob}(1,2,3)$.
\end{Lemma}
\begin{proof}
First we show that a morphism
$\tilde F: P^{\Sigma(2,1)}_{jkj^{-1},j} \to P^{\Sigma(2,1)}_{j,k}$
can be found that covers the diffeomorphism $F: {\Sigma(2,1)} \to {\Sigma(2,1)}$.
This morphism is most easily described using the
action of $F$ on the fundamental group $\pi_1({\Sigma(2,1)})$ of
the pair of pants. The latter is a free group with two
generators which can be chosen as the paths $a,b$ around the
two ingoing circles,
$\pi_1({\Sigma(2,1)}) = \mathbb{Z}* \mathbb{Z} = \langle a,b \rangle$.
Then the induced action of $F$\ on the generators is
$\pi_1(F)(a) = aba^{-1}$ and $\pi_1(F)(b) = a$.
Hence, we find on the covers $F^*P_{j,k} \cong P_{jkj^{-1},j}$.
This implies that we have a diffeomorphism
$\tilde F: P_{jkj^{-1},j} \to P_{j,k}$ covering $F$.
To extend $\tilde F$ to a 2-morphism in $\mathfrak{C}\mathrm{ob}^J(1,2,3)$, we have
to be a bit careful about how we consider
the cover $P^{\Sigma(2,1)}_{jkj^{-1},j} \to {\Sigma(2,1)}$ of the trinion as a
1-morphism. In fact, it has to be considered as a morphism
$(P_j \to \mathbb{S}^1) \sqcup (P_k \to \mathbb{S}^1) \longrightarrow (P_{jk} \to \mathbb{S}^1)$
where the ingoing components are first exchanged and then the
identification of $P_k \to \mathbb{S}^1$ and $P_{jkj^{-1}} \to \mathbb{S}^1$
via the conjugation isomorphisms $P^{\Sigma(1,1)}_{j,k}$ induced by
covers of the cylinders is
used first, compare the lower arrows in the preceding
commuting diagram. This yields the composition
$ \big(P^{\Sigma(2,1)}_{jkj^{-1},j}\big) \circ \tau \circ
\big(\mathrm{id} \sqcup P^{\Sigma(1,1)}_{j,k}\big) $
on the right hand side of the diagram.
\end{proof}
The next step is to apply the TFT functor $Z_G^J$ to
the 2-morphism $\hat F$. The target 1-morphism of
$\hat F$ can be computed using the fact that $Z_G^J$
is a symmetric monoidal 2-functor; we find the
following functor ${\mathcal C}(G)_j\otimes {\mathcal C}(G)_k
\to {\mathcal C}(G)_{jk}$:
$$Z_G^J\Big(\big(P^{\Sigma(2,1)}_{jkj^{-1},j}\big) \circ \tau
\circ \big(\mathrm{id} \sqcup P^{\Sigma(1,1)}_{j,k}\big)\Big)
= (-)^j \otimes_{jkj^{-1},j}^{op} (-)$$
We thus have the functor which acts on objects as
$(V,W) \mapsto \phi_j(W) \otimes V$
for $V \in {\mathcal C}(G)_j$ and $W \in {\mathcal C}(G)_k$.
Then $c := Z_G^J(\hat F)$ is a natural transformation
$ (-) \otimes_{j,k}(-) \Longrightarrow (-)^j
\otimes_{jkj^{-1},j}^{op}(-) $
i.e. a family of isomorphisms
\begin{equation}
c_{V,W}:V \otimes_{j,k} W \stackrel{\sim}{\longrightarrow} \quad \phi_j(W) \otimes_{jkj^{-1},j} V
\end{equation}
in ${\mathcal C}(G)_{jk}$ for $V \in {\mathcal C}(G)_j$ and $W \in {\mathcal C}(G)_k$.
We next show how this natural transformation is expressed
when we use the equivalent description of the categories
${\mathcal C}(G)_j$ as vector bundles on action groupoids:
\begin{Proposition}\label{prop:braiding}
For $V= \bigoplus V_h \in H_j//G\text{-}\mathrm{mod}$ and
$W= \bigoplus W_h \in H_k//G\text{-}\mathrm{mod}$ the natural isomorphism
$c_{V,W}: V \otimes W \to \phi_j(W) \otimes V$ is given by
$$ v \otimes w \mapsto (s(j^{-1})h).w \otimes v$$
for $v \in V_h$ with $h \in H_{j}$ and $w \in W$.
\end{Proposition}
\begin{proof}
We first compute the 1-morphism in the category $\mathfrak{S}\mathrm{pan}$
of spans of finite groupoids that corresponds to
the target 1-morphism $\big(P^{\Sigma(2,1)}_{jkj^{-1},j}\big) \circ \tau
\circ \big(id \sqcup P^{\Sigma(1,1)}_{j,k}\big)$. From the previous
proposition, we obtain the following zig-zag diagram:
$$ H_j//G \times H_k //G \leftarrow H_{jkj^{-1}}//G \times H_j//G
\leftarrow (H_{jkj^{-1}} \times H_k) //G \to H_{jk} //G
\,\, . $$
The first morphism is given by the morphisms implementing
the $J$-action that has been computed in the
proof of proposition \ref{prop:action}, composed
with the exchange of factors. The second 1-morphism
is obtained from the two projections and the last 1-morphism
is the product in the group $H$.
Thus, the 2-morphism $\hat F$ from Lemma \ref{braiding-morphism} yields a 2-morphism $\hat F_G$
in the diagram
$$\xymatrix{
& & H_j \times H_k //G \ar[lld]\ar[rrd]\ar@{=>}[dd]^{\hat F_G} & \\
H_j//G \times H_k //G & & & & H_{jk} //G \\
& H_{jkj^{-1}}//G \times H_j//G\ar[lu] & (H_{jkj^{-1}} \times H_j) //G \ar[l]\ar[rru] &
}$$
where $\hat F_G$ is induced by the equivariant map
$(h,h') \mapsto (h h' h^{-1},h)$. Once the situation
is presented in this way, one can carry out explicitly
the calculation along the lines described in
\cite[Section 4.3]{Morton} and obtain the result.
\end{proof}
A similar discussion can in principle be carried out to compute the
associators. More generally, structural morphisms
on $H//G\text{-}\mathrm{mod}$ can be derived from suitable 3-cobordisms. The relevant
computations become rather involved. On the other hand,
the category $H//G\text{-}\mathrm{mod}$ also inherits structural morphisms
from the underlying category of vector spaces. We
will use in the sequel the latter type of structural
morphism.
\section{Equivariant Drinfel'd double}
The goal of this section is to show that the category
$\calc^J(G) := \bigoplus_{j \in J} {\mathcal C}(G)_j$
comprising the categories
we have constructed in proposition \ref{sectors}
has a natural structure of a $J$-modular category.
Very much like ordinary modularity, $J$-modularity
is a completeness requirement for the relevant
tensor category that is suggested by
principles of field theory. Indeed,
it ensures that one can construct a $J$-equivariant
topological field theory, see \cite{turaev2010}.
For the definition of
$J$-modularity we refer to \cite[Definition 10.1]{kirI17}.
To establish the structure of a modular tensor category
on the category found in the previous sections, we
realize this category
as the representation category of a finite-dimensional
algebra, more precisely of a $J$-Hopf algebra.
This section is organized as follows: we first
recall the notions of equivariant ribbon categories and
of equivariant ribbon algebras, taking into account
a suitable form of weak actions. In section 4.3,
we then present the appropriate generalization of the
Drinfel'd double that describes the category $\calc^J(G)$.
We then describe its orbifold category as the
category of representations of a braided Hopf algebra,
which allows us to establish the modularity of the orbifold
category. We then apply a result of
\cite{kirI17} to deduce that the structure with which
we have endowed $\calc^J(G)$ is the one of a $J$-modular
tensor category.
The Hopf algebraic structures endowed with weak actions
we introduce in this section might be of independent interest.
\subsection{Equivariant braided categories.}
Let $1\to G\to H\stackrel\pi\to J\to 1$ be an exact sequence of finite
groups. The normal subgroup $G$ acts on $H$ by
conjugation; denote by $H//G$ the corresponding action
groupoid. We consider the
functor category $H//G\text{-}\mathrm{mod}:=[H//G,\textnormal{Vect}_{\mathbb K}]$, where
${\mathbb K}$ is an algebraically closed field of characteristic
zero. The category $H//G\text{-}\mathrm{mod}$ is
the category of $H$-graded vector spaces,
endowed with an action of the subgroup $G$ such that
$g.V_h \subset V_{ghg^{-1}}$ for all $g\in G, h\in H$.
An immediate corollary of proposition \ref{sectors}
is the following description of the category
$\calc^J(G) := \bigoplus_{j \in J} {\mathcal C}(G)_j$
as an abelian category:
\begin{Proposition}\label{HG-Mod}
The category $\calc^J(G)$ is equivalent, as an
abelian category, to the category $H//G\text{-}\mathrm{mod}$.
In particular, the category ${\mathcal C}^J(G)$ is a 2-vector space
in the sense of definition \ref{twovect}.
\end{Proposition}
\begin{proof}
With $H_j:=\pi^{-1}(j)$, proposition \ref{sectors}
gives the equivalence ${\mathcal C}(G)_j\cong\H_j//G\text{-}\mathrm{mod}$
of abelian categories. The equivalence of categories
$\calc^J(G)\cong H//G\text{-}\mathrm{mod}$ now follows
from the decomposition $H = \bigsqcup_{j \in J} H_j$.
By \cite[Lemma 4.1.1]{Morton08}, the representation
category of a finite groupoid is a 2-vector space.
\end{proof}
Representation categories of finite groupoids are very
close in structure to representation categories of finite
groups. In particular, there is a complete character
theory that describes the simple objects,
see appendix \ref{groupoidrep}.
We next introduce equivariant categories.
\begin{Definition}
Let $J$ be a finite group and ${\mathcal C}$ a category.
\begin{enumerate}
\item
A \emph{categorical action} of the group $J$ on
the category ${\mathcal C}$ consists of the
following data:
\begin{itemize}
\item[--] A functor $\phi_j:{\mathcal C} \to {\mathcal C}$ for
every group element $j \in J$.
\item[--] A functorial isomorphism
$\alpha_{i,j}:\phi_i \circ \phi_j \stackrel\sim\to \phi_{ij}$, called \emph{compositors}, for every
pair of group elements $i,j \in J$
\end{itemize} such that the coherence conditions
\begin{equation}\label{coherence-compositor}
\alpha_{ij,k} \circ \alpha_{i,j} = \alpha_{i,jk}
\circ \phi_i(\alpha_{j,k}) \quad \text{ and } \quad \phi_1 = \mathrm{id}
\end{equation}
hold.
\item
If ${\mathcal C}$ is a monoidal category, we only consider
actions by monoidal functors $\phi_j$ and
require the natural
transformations to be monoidal natural transformations.
In particular, for each group element $j\in J$,
we have the additional
datum of a natural isomorphism
$$\gamma_j(U,V): \phi_j(U)\otimes\phi_j(V)\stackrel\sim\to\phi_j(U\otimes V)$$
for each pair of objects $U,V$ of ${\mathcal C}$ such that
the following diagrams commute:
$$ \xymatrix{
^{jk}X\otimes {} ^{jk}Y \ar^{\gamma_{jk}(X,Y)}[rrrr]
\ar_{\alpha_{jk}(X)\otimes\alpha_{jk}(Y)}[d]&&&&
^{jk}(X\otimes Y)\ar^{\alpha_{jk}(X\otimes Y)}[d]\\
^j(^k{} (X))\otimes {}^j(^k{} ( Y))
\ar_{^j\gamma_k(X,Y)\circ\gamma_j(^kX,^k Y)}[rrrr]
&&&&
^j(^k{} (X\otimes Y))
}$$
(The data of a monoidal functor includes an isomorphism $\phi_j(1)\to 1$ in principal, but in this paper, the isomorphism will be the identity and therefore we will suppress it in our discussion.)
We use the notation $^jU := \phi_j(U)$ for the image of
an object $U \in {\mathcal C}$ under the functor $\phi_j$.
\item
A \emph{$J$-equivariant} category ${\mathcal C}$ is a category
with a decomposition ${\mathcal C} = \bigoplus_{j \in J} {\mathcal C}_j$
and a categorical action of $J$, subject to the
compatibility requirement
\[\phi_{i}{\mathcal C}_j \subset {\mathcal C}_{iji^{-1}}\]
with the grading.
\item
A \emph{$J$-equivariant tensor category} is a $J$-equivariant
monoidal category ${\mathcal C}$, subject
to the compatibility requirement that
the tensor product of two homogeneous elements
$U \in {\mathcal C}_i, V \in {\mathcal C}_j$ is again homogeneous,
$U \otimes V \in {\mathcal C}_{ij}$.
\end{enumerate}
\end{Definition}
\begin{Remark}
We remark that the condition $\phi_1 = \mathrm{id}$ \ref{coherence-compositor} should in general be replaced by an extra datum, an isomorphism $\eta:\mathrm{id} \stackrel\sim\to \phi_1$ and two coherence conditions which involve the compositors $\alpha_{i,j}$. The diagrams can be found as follows:
For any category ${\mathcal C}$,
consider the category ${\rm AUT}({\mathcal C})$ whose objects
are automorphisms of ${\mathcal C}$ and whose morphisms are
natural isomorphisms.
The composition of functors and natural transformations
endow ${\rm AUT}({\mathcal C})$ with the natural structure of a strict
tensor category.
A categorical action of a finite group $J$ on a category
${\mathcal C}$ then amounts to a tensor functor
$\phi:J \to {\rm AUT}({\mathcal C})$, where $J$ is seen as a tensor category
with only identity morphisms, compare also remark
\ref{rem:3.2}.3.
The condition $\phi_1 = \mathrm{id}$ holds in the categories we are interested in, hence we impose it.
Similarly, we consider
for a monoidal category ${\mathcal C}$ the category
${\rm AUT} mon({\mathcal C})$ whose objects are monoidal
automorphisms of ${\mathcal C}$ and whose morphisms are monoidal
natural automorphisms. The categorical actions we consider
for monoidal categories are then tensor
functors $\phi:J \to {\rm AUT} mon({\mathcal C})$
For more details, we
refer to \cite{turaev2010} Appendix 5.
\end{Remark}
The category $H//G\text{-}\mathrm{mod}$ has a natural structure of a
monoidal category: the tensor product of two objects
$V=\oplus_{h\in H} V_h$ and $W=\oplus_{h\in H} W_h$
is the vector space $V\otimes W$ with $H$ grading
given by $(V\otimes W)_h := \oplus_{h_1h_2=h}
V_{h_1}\otimes W_{h_2}$ and $G$ action given by
$g.(v\otimes w)= g.v\otimes g.w$. The associators
are inherited from the underlying category of
vector spaces.
\begin{Proposition}
Consider an exact sequence of groups
$1\to G\to H \to J\to 1$. Any choice of a
a set-theoretic section $s: J\to H$ allows us to endow the
abelian category $H//G\text{-}\mathrm{mod}$ with the structure of a
$J$-equivariant tensor category as follows: the functor
$\phi_j$ is given by shifting the grading from
$h$ to $s(j)hs(j)^{-1}$ and replacing the action by
$g$ by the action of $s(j)gs(j)^{-1}$.
The isomorphism
$\alpha_{i,j}:\phi_i\circ\phi_k\to\phi_{ij}$
is given by the left action action of the element
$$ \alpha_{i,j} = s(i)s(j)s(ij)^{-1} \,\,.$$
\end{Proposition}
The fact that the action is only a weak action
thus accounts for the failure of $s$ to be a section in the
category of groups.
\begin{proof}
Only the coherence conditions
$\alpha_{ij,k} \circ \alpha_{i,j} = \alpha_{i,jk}
\circ \phi_i(\alpha_{j,k})$ remain to be checked.
By the results of Dedecker and Schreier, cf.\ proposition
\ref{dedeschreier}, the group elements
$s(i)s(j)s(ij)^{-1}\in G$ are the coherence cells of
a weak group action of $J$ on $H$. By
definition \ref{Def:action}, this implies the coherence
identities, once one takes into account that that
composition of functors is written in different order
than group multiplication.
\end{proof}
We have derived in section
\ref{twisted_sectors_and_fusion} from the geometry
of extended cobordism categories
more structure on the geometric category $\calc^J(G) = \bigoplus_{j \in J} \mathcal{C}(G)_j$.
In particular, we collect the functors $\otimes_{jk}: {\mathcal C}(G)_j
\boxtimes {\mathcal C}(G)_k \to {\mathcal C}(G)_{jk}$ from
proposition \ref{prop:fusionproduct} into a functor
\begin{equation}
\otimes : {\mathcal C} \boxtimes {\mathcal C} \to {\mathcal C}\text{.}
\label{eq:grouptogether}
\end{equation}
Another structure are the isomorphisms
$V \otimes W \to \phi_j(W) \otimes V$
for $V \in {\mathcal C}(G)_j$, described in proposition
\ref{prop:braiding}. Together with the associators,
this suggests to endow the category $\calc^J(G)$
with a structure of a braided $J$-equivariant tensor
category:
\begin{Definition}\label{J-equ.tc}
A \emph{braiding} on a $J$-equivariant tensor category is
a family
\[c_{U,V}:U\otimes V \to {}^jV \otimes U\]
of isomorphisms, one for every pair of objects $U \in {\mathcal C}_i, V \in {\mathcal C}_j$,
which are natural in $U$ and $V$.
Moreover, a braiding is required to satisfy
an analogue of the hexagon axioms (see
\cite[appendix A5]{turaev2010})
and to be preserved under the action of $J$,
i.e.\ the following diagram commutes for all objects
$U,V$ with $U\in{\mathcal C}_j$ and $i\in J$
\begin{equation}\label{action-braiding-compatibility}
\xymatrix{
^i(U\otimes V)\ar_{\gamma_i}[d]\ar^{^i(c_{U,V})}[rr]
&& ^i(^jV \otimes U) \ar^{\gamma_i}[rr]&& ^i(^jV)\otimes \, ^iU
\ar^{\alpha_{ij}(V)\otimes\mathrm{id}}[d]\\
^iU\otimes ^iV\ar_{c_{^iU,^iV}}[rr] &&
^{ij^{-1}}(^iV)\otimes\, ^iU
\ar_{\alpha_{iji^{-1},i(V)}\otimes\mathrm{id}}[rr]
&& ^{ij}V\otimes \, ^i U
}
\end{equation}
\end{Definition}
\begin{Remark}
\begin{enumerate}
\item
It should be appreciated that a braided $J$-equivariant
category is not, in general, a braided
category. Its {\em neutral component} ${\mathcal C}_1$ with
$1\in J$ the neutral element, is a braided tensor category.
\item
By replacing the underlying category by an equivalent
category, one can replace a weak action by a strict
action, compare \cite[Appendix A5]{turaev2010}.
In our case, weak actions actually lead to simpler
algebraic structures.
\item
The $J$-equivariant monoidal category $H//G\text{-}\mathrm{mod}$ has a
natural braiding isomorphism that has been
described in proposition \ref{prop:braiding}
\end{enumerate}
\end{Remark}
We use the equivalence of abelian categories between
$\calc^J(G) = \bigoplus_{j \in J} \mathcal{C}(G)_j$ and $H//G\text{-}\mathrm{mod}$ to endow the category
$\calc^J(G) = \bigoplus_{j \in J} \mathcal{C}(G)_j$ with associators. The category has now enough
structure that we can state our next result:
\begin{Proposition}\label{prop:Jtensor}
The category $\calc^J(G) = \bigoplus_{j \in J} \mathcal{C}(G)_j$, with the tensor product functor
from (\ref{eq:grouptogether}), can be endowed with
the structure of a braided $J$-equivariant tensor
category such that the isomorphism $\calc^J(G) = \bigoplus_{j \in J} \mathcal{C}(G)_j
\cong H//G\text{-}\mathrm{mod}$ becomes an isomorphism of
braided $J$-equivariant tensor categories.
\end{Proposition}
\begin{proof}
The compatibility with the grading is implemented by definition via the graded components $\otimes_{jk}$ of $\otimes$ and the graded components of $c_{V,W}$. It remains to check that the action is by tensor functors and that the braiding satisfies the hexagon axiom. The second boils down to a simple calculation and the first is seen by noting that the action is essentially an index shift which is preserved by tensoring together the respective components.
\end{proof}
\subsection{Equivariant ribbon algebras}
In the following, let $J$ again be a finite group.
To identify the structure of a $J$-modular tensor
category on the geometric category $\calc^J(G) = \bigoplus_{j \in J} \mathcal{C}(G)_j$, we need
dualities. This will lead us
to the discussion of (equivariant) ribbon algebras.
Apart from strictness issues, our discussion closely follows
\cite{turaev2010}.
We start our discussion with the relevant category-theoretic
structures, which generalize premodular categories (cf. definition \ref{modularity}) to the equivariant setting.
\begin{Definition}\label{ribbon_def}
\begin{enumerate}
\item
A \emph{$J$-equivariant ribbon category} is a $J$-braided
category with dualities and a family of isomorphisms
$\theta_V:V \to{} ^jV$ for all $j \in J, V \in {\mathcal C}_j$,
such that $\theta$ is compatible with duality and the action of $J$ (see \cite[VI.2.3]{turaev2010} for the identities).
In contrast to \cite{turaev2010}, we allow weak $J$-actions
and thus require the diagram
\begin{align}\label{com.diagram-twist}
\begin{xy}
\xymatrix{
U\otimes V \ar[rrrr]^{\theta_{U\otimes V}} \ar[d]_{\theta_U \otimes \theta_V} &&&& ^{ij}(U\otimes V) \\
^iU\otimes ^jV \ar[rd]_{c_{{}^iU, ^jV}} & &&& ^{iji^{-1}}(^i U) \otimes ^{ij}V\ar[u]_{\gamma_{ij} \circ (\alpha_{iji^{-1},i} \otimes \mathrm{id})} \\
&^i(^jV) \otimes {}^i U\ar[rr]_{\alpha_{i,j}\otimes \mathrm{id}} & &^{ij}V \otimes {}^i U\ar[ur]_{c_{{}^{ij}V,^{i}U}} &
}
\end{xy}
\end{align}
involving compositors,
to commute for $U \in {\mathcal C}_i$ and $V \in {\mathcal C}_j$.
\item A \emph{$J$-premodular category} is a ${\mathbb K}$-linear, abelian, finitely semi-simple $J$-equivariant ribbon category such that the tensor product is a ${\mathbb K}$-bilinear functor and the tensor unit is absolutely simple.
\end{enumerate}
\end{Definition}
\begin{Remark}
The following facts directly follow from the
definition
of $J$-equivariant ribbon category:
the neutral component ${\mathcal C}_1$ is itself a braided
tensor category. In particular, it contains the
tensor unit of the $J$-equivariant tensor category.
The dual object of an object $V\in {\mathcal C}_j$ is in
the category ${\mathcal C}_{j^{-1}}$.
\end{Remark}
We will not be able to directly endow the geometric
category $\calc^J(G) = \bigoplus_{j \in J} \mathcal{C}(G)_j$ with the structure of a $J$-equivariant ribbon category.
Rather, we will realize an equivalent category as the
category of modules over a suitable algebra.
To this end, we introduce in several steps the notions of a
$J$-ribbon algebra and analyze the extra structure
induced on its representation category.
\begin{Definition}
Let $A$ be an (associative, unital) algebra over a
field ${\mathbb K}$. A \emph{weak $J$-action} on $A$
consists of algebra automorphisms $\varphi_j \in {\rm Aut}(A)$,
one for every element $j\in J$, and invertible
elements $c_{ij} \in A$, one for every pair of elements
$i,j \in J$, such that for all $i,j,k \in J$
the following conditions hold:
\begin{equation}\label{coherence-compositors2}
\varphi_i \circ \varphi_j = \mathrm{Inn}_{c_{i,j}} \circ \varphi_{ij}\qquad
\varphi_i(c_{j,k}) \cdot c_{i,jk} = c_{i,j} \cdot
c_{ij,k} \quad \text{ and } \quad c_{1,1} = 1
\end{equation}
Here $\mathrm{Inn}_x$ with $x$ an invertible element of $A$ denotes
the algebra automorphism $a \mapsto xax^{-1}$.
A weak action of a group $J$ is called
strict, if $c_{i,j} = 1$ for all pairs $i,j\in J$.
\end{Definition}
\begin{Remark}
As discussed for weak actions on groups in remark
\ref{rem:3.2}, a weak action
on a ${\mathbb K}$-algebra $A$ can be seen as a categorical action
on the category
which has one object and the elements
of $A$ as endomorphisms.
\end{Remark}
We now want to relate a weak action $(\varphi_j,c_{i,j})$
of a group $J$ on an algebra $A$ to a categorical action
on the representation category $A\text{-}\mathrm{mod}$. To this end,
we define for each element $j\in J$ a functor
on objects by
$$^j(M,\rho) := (M,\rho \circ (\varphi_{j^{-1}}\otimes \mathrm{id}_M))$$
and on morphisms by $^jf = f$. For the functorial
isomorphisms, we take
\begin{equation*}
\alpha_{i,j}(M,\rho):= \rho\big((c_{j^{-1},i^{-1}})^{-1}\otimes \mathrm{id}_M\big).
\end{equation*}
The inversions in the above formulas make sure that the action on the level of categories really becomes a left action.
\begin{Lemma}\label{action}
Given a weak action of $J$ on a ${\mathbb K}$-algebra $A$,
these data define a categorical action on
the category $A\text{-}\mathrm{mod}$.
\end{Lemma}
\begin{proof}
Let $V$ be an $A$-Modul. We first show that $\alpha_{i,j}$ is a morphism $^{i}(^{j}V)\to ^{ij}V$ in $A\text{-}\mathrm{mod}$:
Let $V\in A\text{-}\mathrm{mod}$, then for every $v \in V, a \in A$, we have:
\begin{align*}
\alpha_{i,j}(V) \circ \rho_{^{i}(^jV)}(a\otimes v)& = (c_{i^{-1},j^{-1}})^{-1}\varphi_{j^{-1}} \circ \varphi_{i^{-1}}(a).v\\
& = \varphi_{(ij)^{-1}}(a)(c_{i^{-1},j^{-1}})^{-1}.v\\
& = \rho_{(^{ij}V)} \circ (\mathrm{id}_A \otimes \alpha_{i,j}(V))(a \otimes v),
\end{align*}
where we used the abbreviation $a.v := \rho(a\otimes v)$.
The validity of the coherence condition \eqref{coherence-compositor} for the $\alpha_{i,j}$ follows from the second equation in \eqref{coherence-compositors2}, since the right side of \eqref{coherence-compositor} evaluated on an element $v \in V$ reads
\[ \alpha_{i,jk} \circ \phi_i(\alpha_{j,k}) (v) = (c_{(jk)^{-1},i^{-1}})^{-1}(c_{k^{-1},j^{-1}})^{-1}.v \]
and the left one is:
\[\alpha_{ij,k} \circ \alpha_{i,j}(v) = (c_{k^{-1},(ij)^{-1}})^{-1} \varphi_{k^{-1}} (c_{j^{-1},i^{-1}})^{-1} .v\]
\end{proof}
We next turn to an algebraic structure that yields
$J$-equivariant tensor categories.
\begin{Definition}\label{JHopf}
A \emph{$J$-Hopf algebra} over ${\mathbb K}$ is a Hopf algebra
$A$ with a $J$-grading $A = \bigoplus_{j\in J} A_{j}$
and a weak $J$-action such that:
\begin{itemize}
\item
The algebra structure of $A$ restricts to
the structure of an associative algebra on
each homogeneous component so that $A$ is the direct
sum of the components $A_j$ as an algebra.
\item
$J$ acts by homomorphisms of Hopf algebras.
\item The action of $J$ is compatible with the grading,
i.e.\ $\varphi_i(A_j) \subset A_{iji^{-1}}$
\item
The coproduct $\Delta:\ A\to A\otimes A$ respects
the grading, i.e.\
$$\ \Delta (A_j) \subset \bigoplus_{p,q\in J,pq=j} A_p\otimes
A_q \,\, . $$
\item The elements $(c_{i,j})_{i,j\in J}$ are group-like, i.e $\Delta(c_{i,j}) = c_{i,j}\otimes c_{i,j}$.
\end{itemize}
\end{Definition}
\begin{Remark
\begin{enumerate}
\item
For the counit $\epsilon$ and the antipode $S$ of
a $J$-Hopf algebra, the compatibility relations
with the grading $\epsilon(A_j) = 0$ for $j \ne 1$ and
$S(A_j) \subset A_{j^{-1}}$ are immediate consequences of
the definitions.
\item The restrictions of the structure maps
endow the homogeneous component $A_1$ of $A$
with the structure of a Hopf algebra with a weak
$J$-action.
\item $J$-Hopf algebras with strict $J$-action
have been considered under the name ``$J$-crossed Hopf coalgebra''
in \cite[Chapter VII.1.2]{turaev2010}.
\end{enumerate}
\end{Remark}
The category $A\text{-}\mathrm{mod}$ of finite-dimensional modules
over a $J$-Hopf algebra inherits a natural duality
from the duality of the underlying category of
${\mathbb K}$-vector spaces. The weak action described in
Lemma \ref{action} is even a monoidal action,
since $J$ acts by Hopf algebra morphisms.
A grading on $A\text{-}\mathrm{mod}$ can be given by taking
$(A\text{-}\mathrm{mod})_j = A_j\text{-}\mathrm{mod}$ as the $j$-homogeneous component.
From the properties of a $J$-Hopf algebra one can finally deduce
that the tensor product, duality and grading are compatible with
the $J$-action. We have thus arrived at the following statement:
\begin{Lemma}\label{J-equivariant}
The category of representations of a $J$-Hopf algebra has a
natural structure of a ${\mathbb K}$-linear, abelian
$J$-equivariant tensor category with compatible duality
as introduced in definition \ref{J-equ.tc}.
\end{Lemma}
\begin{proof}
We show, that the grading and the action on $A\text{-}\mathrm{mod}$ are compatible. Let $
V \in A_j\text{-}\mathrm{mod}$, then the $j$-component $1_j$ of the unit in A acts as the identity on $V$. We have to check, that $^{i}V \in A_{iji^{-1}}\text{-}\mathrm{mod}$, i.e that $1_{iji^{-1}}$ acts as the identity on $^iV$. For $v\in ^iV$, we have:
\[\rho_i(1_{iji^{-1}}\otimes v) = \varphi_{i^{-1}}(1_{iji^{-1}}).v = 1_j.v = v\]
This shows $^i V \in A_{iji^{-1}}\text{-}\mathrm{mod}$.
\end{proof}
The representation category of a braided Hopf algebra
is a braided tensor category. If the Hopf algebra has, moreover, a
twist element, its representation category is even
a ribbon category. We now present $J$-equivariant
generalizations of these structures.
\begin{Definition}\label{def:jribbon}
Let $A$ be a $J$-Hopf algebra.
\begin{enumerate}
\item
A $J$-equivariant $R$-matrix is an element $R = R_{(1)}\otimes R_{(2)} \in A\otimes A$ such that for $V\in (A\text{-}\mathrm{mod})_i$, $W\in A\text{-}\mathrm{mod}$, the map
\begin{align*}
c_{VW}:& V\otimes W \to ^{i}W \otimes V\\
&v\otimes w \mapsto R_{(2)}.w\otimes R_{(1)}.v
\end{align*}
is a $J$-braiding on the category $A\text{-}\mathrm{mod}$ according to definition \ref{J-equ.tc}.
\item A $J$-twist is an invertible element $\theta \in A$ such that for every object $V\in (A\text{-}\mathrm{mod})_i$ the induced map
\begin{align*}
\theta_V: &V\to ^{i}V\\
& v \mapsto \theta^{-1}.v
\end{align*}
is a $J$-twist on $A\text{-}\mathrm{mod}$ as defined in \ref{ribbon_def}.
\end{enumerate}
If $A$ has an $R$-matrix and a twist, we call it a $J$-\emph{ribbon-algebra}.
\end{Definition}
\begin{Remark}
\begin{itemize}
\item
A $J$-ribbon algebra is not, in general,
a ribbon Hopf algebra.
\item
The component $A_1$ with the obvious restricitions of $R$ and $\theta$ is a ribbon algebra.
\item The conditions that the category $A\text{-}\mathrm{mod}$ is braided resp. ribbon can be translated into algebraic coniditions on the elements $R$ and $\theta$. Since we are mainly interested in the categorical structure we refrain from doing that here.
\end{itemize}
\end{Remark}
Taking all the introduced algebraic strucutre into account we get the following proposition.
\begin{Proposition}\label{category-J-ribbon}
The representation category of a $J$-ribbon algebra is a $J$-ribbon category.
\end{Proposition}
\begin{Remark}
In \cite{turaev2010}, Hopf algebras and ribbon Hopf algebras
with strict $J$-action have been considered.
The next subsection will give an illustrative
example where the natural action is not strict.
\end{Remark}
\subsection{Equivariant Drinfel'd Double}
The goal of this subsection is to construct a $J$-ribbon algebra, given a finite group $G$
with a weak $J$-action. As explained in subsection
\ref{Def:action}, such a weak $J$-action amounts to
a group extension
\begin{equation}\label{groupext}
1\to G \longrightarrow H \stackrel{\pi}{\longrightarrow} J
\to 1
\end{equation}
with a set-theoretical splitting $s:J \to H$.
We start from the well-known fact reviewed
in subsection \ref{dd} that the Drinfel'd double
${\mathcal D}(H)$ of the finite group $H$ is a ribbon Hopf
algebra. The double ${\mathcal D}(H)$ has a canonical basis
$\delta_{h_1}\otimes h_2$ indexed by pairs
$h_1,h_2$ of elements of $H$. Let $G\subset H$ be
a subgroup. We are interested in
the vector subspace ${\mathcal D}^J(G)$ spanned by the
basis vectors $\delta_h\otimes g$
with $h\in H$ and $g\in G$.
\begin{Lemma}
The structure maps of the Hopf algebra
${\mathcal D}(H)$ restrict to the vector subspace
${\mathcal D}^J(G)$ in such a way that the
latter is endowed with the structure of a Hopf subalgebra.
\end{Lemma}
\begin{Remark}
The induced algebra structure on ${\mathcal D}^J(G)$ is the
one of the groupoid algebra of the action groupoid
$H//G$.
\end{Remark}
The Drinfel'd double ${\mathcal D}(H)$ of a group $H$
has also the structure of a ribbon algebra. However,
neither the R-matrix nor the the ribbon element yield
an R-matrix or a ribbon element of ${\mathcal D}^J(G)\subset{\mathcal D}(H)$.
Rather, this Hopf subalgebra can be endowed with the
structure of a $J$-ribbon Hopf algebra as in definition
\ref{def:jribbon}.
To this end, consider the partition of the group $H$
into the subsets $H_j:=\pi^{-1}(j) \subset H$. It gives
a $J$-grading of the algebra $A$ as a direct sum
of subalgebras:
\[A_j := \langle \delta_h\otimes g\rangle_{h\in H_j,g\in G}
\,\,\, . \]
The set-theoretical section $s$ gives a weak action of $J$
on $A$ that can be described by its action
on the canonical basis of $A_j$:
\begin{eqnarray*}
\varphi_j(\delta_h \otimes g) :=& (\delta_{s(j)hs(j)^{-1}} \otimes s(j)gs(j)^{-1})
\,\, ;
\end{eqnarray*}
the coherence elements are
\[c_{ij} := \sum_{h\in H} \delta_h \otimes
s(i)s(j)s(ij)^{-1} \,\,\,\, .\]
\begin{Proposition}
The Hopf algebra ${\mathcal D}^J(G)$, together with the grading
and weak $J$-action derived from the weak $J$-action
on the group $G$, has the structure of a $J$-Hopf algebra.
\end{Proposition}
\begin{proof}
It only remains to check the compatibility relations of grading and weak $J$-action
with the Hopf algebra structure that have been formulated in
definition \ref{JHopf}.
The fact that $\varphi_i(A_j) \subset A_{iji^{-1}}$ is immediate, since $s(i)H_js(i^{-1}) \subset H_{iji^{-1}}$. The axioms follow are essentially equivalent to the property that conjugation commutes with the product and coproduct of the Drinfel'd double.
\end{proof}
We now turn to the last piece of structure,
an $R$-matrix and twist element in ${\mathcal D}^J(G)$.
Consider the element
$R= \sum_{ij}R_{i,j} \in {\mathcal D}^J(G) \otimes {\mathcal D}^J(G)$
with homogeneous elements $R_{i,j}$ defined as
\begin{eqnarray}\label{JR-matrix}
R_{i,j} := \sum_{h_1 \in H_i, h_2 \in H_{j}}
(\delta_{h_1} \otimes 1)\otimes(\delta_{h_2} \otimes s(i^{-1}) h_1).
\end{eqnarray}
(Note that $\pi(s(i^{-1})h_1) = 1$ for $h_1 \in H_i$ and thus $s(i^{-1})h_1 \in G$.)\\
The element $R_{i,j}$ is invertible with inverse
\begin{eqnarray*}
R_{i,j}^{-1} = \sum_{h_1 \in H_i, h_2 \in H_{j}}
(\delta_{h_1} \otimes 1)\otimes(\delta_{h_2}
\otimes h_1^{-1} s(i^{-1})^{-1}).
\end{eqnarray*}
We also introduce a twist element
$\theta = \sum_{j \in J} \theta_j \in {\mathcal D}^J(G)$
with
\begin{align}\label{J-twist}
\theta_j^{-1} := \sum_{h\in H_j}
\delta_h \otimes s\big(j^{-1}\big)h \in A_j
\end{align}
for every
element $j \in J$.
\begin{Proposition}\label{double-J-ribbon}
The elements $R$ and $\theta$ endow the $J$-Hopf algebra
${\mathcal D}^J(G)$ with the structure of a $J$-ribbon algebra
that we call the $J$-Drinfel'd double of $G$.
\end{Proposition}
\begin{proof}
By Definition \ref{def:jribbon} we have to check that the induced tranformations on the level of categories satisify the axioms for a braiding and a twist. We first compute the induced braiding by using the R-matrix given in \eqref{JR-matrix}:
For $V\in ({\mathcal D}^J(G)\text{-}\mathrm{mod})_j = {\mathcal D}^J(G)_j\text{-}\mathrm{mod}$ and $W\in {\mathcal D}^J(G)\text{-}\mathrm{mod}$, we get the linear map
\begin{align*}
c_{V,W}:V\otimes W &\to ^jW \otimes V\\
v\otimes w & \mapsto s(j^{-1})h.w\otimes v \quad\text{for} \quad v \in V_h
\end{align*}
First of all, we show that this is a morphism in ${\mathcal D}^J(G)\text{-}\mathrm{mod}$.
Let $v\in V_h$ and $w \in W_{h'}$. We have $v\otimes w \in (V\otimes W)_{hh'}$ and
$c_{V,W}(v\otimes w) = s(j^{-1})h.w\otimes v \in (^jW\otimes V)_{hh'}$, since the element $s(j^{-1})h.w$ is in the component $\left(s(j^{-1})hh'h^{-1}s(j^{-1})^{-1}\right)$ of $W$ which is the component $hh'h^{-1}$ of $^jW$. So $c_{V,W}$ respects the grading.
As for the action of $G$, we observe that for $g \in G$, the element $g.v$ lies in the component $V_{ghg^{-1}}$, and so
\begin{align*}
c_{V,W}(g.(v\otimes w)) = \left(s(j^{-1})ghg^{-1}\right)g.w\otimes g.v = s(j^{-1})gh.w\otimes g.v\\
g.c_{V,W}(v\otimes w) = \left(s(j^{-1})gs(j^{-1})^{-1}\right)s(j^{-1})h.w \otimes g.v = s(j^{-1})gh.w\otimes g.v
\end{align*}
which shows that $c_{V,W}$ commutes with the action of $G$.\\
Furthermore, it needs to be checked, that $c_{V,W}$ satisfies the hexagon axioms, which in our case reduce to the two diagrams
\begin{align*}
\xymatrix{
U\otimes V \otimes W\ar[rr]^{c_{U, V\otimes W}}\ar[d]_{c_{U,V}\otimes \mathrm{id}_W} && ^{i}V\otimes ^{i}W\otimes U\\
^{i}V \otimes U \otimes W\ar[rru]_{\mathrm{id}_{^{i}V} \otimes c_{U,W}}&&
}
\end{align*}
and
\begin{align*}
\xymatrix{
U\otimes V \otimes W\ar[rr]^{c_{U\otimes V,W}}\ar[d]_{\mathrm{id}_U \otimes c_{V,W}} &&^{ij}W \otimes U \otimes V\\
U \otimes ^{j}W \otimes V\ar[rr]_{c_{U,^{j}W} \otimes \mathrm{id}_{V}} && ^{i}(^{j}W) \otimes U \otimes V
\ar[u]_{\alpha_{i,j}\otimes \mathrm{id}_U}}
\end{align*}
where $U \in {\mathcal C}_i, V \in {\mathcal C}_j$ and we suppressed the associators in $A\text{-}\mathrm{mod}$.\\
And at last we need to check the compatibility of the braiding and the $J$-action, i.e. diagram \eqref{action-braiding-compatibility}.
All of that can be proven by straightforward calculations, which show, that the morphism induced by the element $R$ is really a braiding in the module category.\\
We now compute the morphism given by the action with the inverse twist element \eqref{J-twist}. For $V\in ({\mathcal D}^J(G)\text{-}\mathrm{mod})_j = {\mathcal D}^J(G)_j\text{-}\mathrm{mod}$ we get the linear map:
\begin{align*}
\theta_V:V &\to ^{j}V\\
v&\mapsto s(j^{-1})h.v
\end{align*}
for $v \in V_h$. This map is compatible with the $H$-grading since for $v \in V_h$, we have $s(j^{-1})h.v \in V_{s(j^{-1})hs(j^{-1})^{-1}} = (^j V)_h$ and it is also compatible with the $G$-action, since
\begin{align*}
\theta_V(g.v) &= s(j^{-1})(ghg^{-1})g.v = s(j^{-1})gh.v\\
g.\theta_V(v) &= \left(s(j^{-1})gs(j^{-1})^{-1}\right)s(j^{-1})h.v = s(j^{-1})gh.v
\end{align*}
So the induced map is a morphism in the category ${\mathcal D}^J(G)$. In order to proof that it is a twist, we have to check the commutativity of diagram \eqref{com.diagram-twist} as well as for every $i,j \in J$ and object $V$ in the component $j$, the commutativity of the diagrams
\begin{align*}
\xymatrix{
^{j}V^{\vee}\ar[r]^{(\theta_V)^{\vee}}\ar[d]_{\theta_{^{j}V^{\vee}}} &V^{\vee}\\
^{j^{-1}}(^{j}V^{\vee})\ar[ru]_{\alpha_{j^{-1},j}}&}
\end{align*}
(which displays the compatibility of the twist with the duality),
and
\begin{align*}
\xymatrix{
^{i}V\ar[rr]^{ ^{i}\theta_V}\ar[d]_{\theta_{^{i}V}} && ^{i}(^{j}V)\ar[d]^{\alpha_{i,j}}\\
^{iji^{-1}}(^{i}V) \ar[rr]_{\alpha_{iji^{-1},i}}&& ^{ij}V}
\end{align*}
(which is the compatibility of the twist with the $J$-action). This again can be checked by straightforward calculations.
\end{proof}
We are now ready to come back to the $J$-equivariant
tensor category $\calc^J(G) = \bigoplus_{j \in J} \mathcal{C}(G)_j$ described in proposition
\ref{prop:Jtensor}.
From this proposition, we know that the category
$\calc^J(G)$ is equivalent to $H//G\text{-}\mathrm{mod}\cong{\mathcal D}^J(G)\text{-}\mathrm{mod}$ as a $J$-equivariant
tensor category.
Also $J$-action and tensor product coincide with
the ones on ${\mathcal D}^J(G)\text{-}\mathrm{mod}$.
Moreover, the equivariant braiding of $\calc^J(G)$ computed in proposition
\ref{prop:braiding} coincides with the braiding on ${\mathcal D}^J(G)$ computed in the proof of the last proposition
\ref{double-J-ribbon}.
This allows us to
transfer also the other structure on the representation category
of the $J$-Drinfel'd double ${\mathcal D}^J(G)$ described
in proposition \ref{double-J-ribbon} to the category
$\calc^J(G)$:
\begin{Proposition}\label{J-fusion}
The $J$-equivariant tensor category $\calc^J(G) = \bigoplus_{j \in J} \mathcal{C}(G)_j$ described
in proposition \ref{prop:Jtensor} can be endowed with
the structure of a $J$-premodular category
such that it is equivalent, as a $J$-premodular
category, to the category ${\mathcal D}^J(G)\text{-}\mathrm{mod}$.
\end{Proposition}
\begin{Remark}
At this point, we have constructed in paticular a $J$-equivariant
braided fusion category $\calc^J(G) = \bigoplus_{j \in J} \mathcal{C}(G)_j$ (see \cite{ENO2009}) with neutral component
${\mathcal C}(G)_1\cong{\mathcal D}(G)\text{-}\mathrm{mod}$
from a weak action of the group $J$ on the group $G$,
or in different words, from a 2-group homomorphisms
$J \to {\rm AUT}(G)$ with ${\rm AUT}(G)$ the automorphism 2-group
of $G$.
In this remark, we very briefly sketch the relation
to the description of $J$-equivariant braided fusion categories
with given neutral sector ${\mathcal B}$ in terms of 3-group
homomorphisms $J \to \underline{\text{Pic}}({\mathcal B})$ given in \cite{ENO2009}.
Here $\underline{\text{Pic}}({\mathcal B})$ denotes the so called Picard 3-group
whose objects are invertible module-categories of the
category ${\mathcal B}$. The group structure comes from
the tensor product of module categories which can be
defined since the braiding on ${\mathcal B}$ allows to turn module categories
into bimodule categories.
Using this setting, we give a description of our
$J$-equivariant braided fusion category ${\mathcal D}^J(G)\text{-}\mathrm{mod}$
in terms of a functor $\Xi: J \to \underline{\text{Pic}}({\mathcal D}(G))$.
To this end, we construct a 3-group homomorphism
${\rm AUT}(G) \to \underline{\text{Pic}}({\mathcal D}(G))$ and write $\Xi$ as the
composition of this functor and the functor
$J \to {\rm AUT}(G)$ defining the weak $J$-action.
The 3-group homomorphism ${\rm AUT}(G) \to \underline{\text{Pic}}({\mathcal D}(G))$
is given as follows: to an object $\varphi \in {\rm AUT}(G)$
we associate the twisted conjugation groupoid
$G//^{\varphi}G$, where $G$ acts on itself by
twisted conjugation, $g.x := g x \varphi(g)^{-1}$.
This yields the category
$G//^{\varphi}G\text{-}\mathrm{mod} := [ G//^{\varphi}G , \textnormal{Vect}_{\mathbb K} ]$
which is naturally a module category over ${\mathcal D}(G) \text{-}\mathrm{mod}$.
Morphisms $\varphi \to \psi$ in ${\rm AUT}(G)$
are given by group elements $g \in G$ with
$g \varphi g^{-1} = \psi$; to such a morphism we associate
the functor $L_g: G//^{\varphi}G \to G//^{\psi}G$ given by
conjugating with $g\in\ G$ on objects and morphisms.
This induces functors of module categories
$G//^{\varphi}G\text{-}\mathrm{mod} \to G//^{\psi}G\text{-}\mathrm{mod}$. Natural
coherence data exist; one then shows
that this really establishes the desired 3-group homomorphism.
\end{Remark}
\subsection{Orbifold category and orbifold algebra}
It remains to show that the $J$-equivariant ribbon
category $\calc^J(G) = \bigoplus_{j \in J} \mathcal{C}(G)_j$ described in \ref{double-J-ribbon}
is $J$-modular. To this end, we will use
the orbifold category of the $J$-equivariant category:
\begin{Definition}
Let ${\mathcal C}$ be a $J$-equivariant category. The
\emph{orbifold category} ${\mathcal C}^J$ of ${\mathcal C}$ has:
\begin{itemize}
\item as objects pairs $(V,(\psi_j)_{j \in J})$
consisting of an object $V\in{\mathcal C}$ and a family
of isomorphisms $\psi_j:{}^jV \to V$ with $j\in J$
such that $\psi_i\,\circ\, {}^i\psi_j = \psi_{ij}\circ \alpha_{i,j}$.
\item as morphisms $f:(V,\psi^V_j) \to (W,\psi^W_j)$
those morphisms $f:V \to W$ in ${\mathcal C}$ for which
$\psi_j \circ {}^j(f) = f \circ \psi_j$ holds for
all $j \in J$.
\end{itemize}
\end{Definition}
In \cite{kirI17}, it has been shown that the orbifold
category of a $J$-ribbon category is an ordinary,
non-equivariant ribbon category:
\begin{Proposition}\label{kirillov1}
\begin{enumerate}
\item
Let ${\mathcal C}$ be a $J$-ribbon category. Then the orbifold
category ${\mathcal C}^J$ is naturally endowed with the
structure of a ribbon category by the following data:
\begin{itemize}
\item The tensor product of the objects $(V,(\psi^V_j))$
and $(W,(\psi^W_j))$ is defined as the object
$(V\otimes W, (\psi^V_j\otimes \psi^W_j))$.
\item The tensor unit for this tensor product
is $\mathbf{1} = (\mathbf{1},(\mathrm{id}))$
\item The dual object of $(V,(\psi_j))$ is
the object $(V^{\ast},(\psi_j^{\ast})^{-1})$,
where $V^{\ast}$ denotes the dual object in ${\mathcal C}$.
\item The braiding of the two objects $(V,(\psi^V_j))$ and
$(W,(\psi^W_j))$ with $V \in {\mathcal C}_j$ is given by the
isomorphism
$(\psi_j \otimes \mathrm{id}_V)\circ c_{V,W}$, where
$c_{V,W}:V \otimes W \to ^jW \otimes V$ is the $J$-braiding
in ${\mathcal C}$.
\item The twist on an object $(V,(\psi_j))$ is
$\psi_j \circ \theta$, where $\theta:V \to {}^jV$ is the twist in ${\mathcal C}$.
\end{itemize}
\item
If ${\mathcal C}$ is a $J$-premodular category, then the
orbifold category ${\mathcal C}^J$ is even a premodular category.
\end{enumerate}
\end{Proposition}
It has been shown in \cite{kirI17} that the $J$-modularity
of a $J$-premodular category is equivalent to
the modularity as in definition \ref{modularity} of its
orbifold category. Our problem is thus reduced to showing
modularity of the orbifold category of ${\mathcal D}^J(G)\text{-}\mathrm{mod}$.
To this end, we describe orbifoldization on
the level of (Hopf-)algebras: given a $J$-equivariant
algebra $A$, we introduce an orbifold algebra $\widehat{A}^J$
such that its representation category $\widehat{A}^J\text{-}\mathrm{mod}$
is isomorphic to the orbifold category of $A\text{-}\mathrm{mod}$.
\begin
{Definition}\label{Def:o.algebra}
Let $A$ be an algebra with a weak $J$-action
$(\varphi_j,c_{ij})$.
We endow the vector space
$\widehat{A}^J:=A \otimes {\mathbb K}[J]$
with a unital associative multiplication which is defined on an element
of the form $(a\otimes j)$ with $a \in A$ and $j \in J$
by\[(a\otimes i)(b\otimes j) := a \varphi_i(b)c_{ij}\otimes ij
\,\, . \]
This algebra is called the
\emph{orbifold algebra} $\widehat{A}^J$ of the
$J$-equivariant
algebra $A$ with respect to the weak $J$-action.
\end{Definition}
If $A$ is even a $J$-Hopf algebra, it is possible
to endow the orbifold algebra with even more structure.
To define the coalgebra structure on the orbifold algebra,
we use the standard coalgebra structure on the
group algebra ${\mathbb K}[J]$ with coproduct
$\Delta_J(j) = j\otimes j$ and counit $\epsilon_J(j) = 1$
on the canonical basis $(j)_{j\in J}$. The
tensor product coalgebra on $A\otimes {\mathbb K}[J]$ has
the coproduct and counit
\begin{equation}
\Delta(a\otimes j) = (\mathrm{id}_A \otimes \tau \otimes
\mathrm{id}_{{\mathbb K}[J]})(\Delta_A(a) \otimes j \otimes j ),\quad\text{ and }\quad
\epsilon(a\otimes j) = \epsilon_A(a)
\end{equation}
which is clearly coassociative and counital.
To show that this endows the orbifold algebra with the
structure of a bialgebra, we have first to show that
the coproduct $\Delta$ is a unital algebra morphism.
This follows from the fact, that $\Delta_A$ is already
an algebra morphism and that the action of $J$ is by
coalgebra morphisms. Next, we have to show that
the counit $\epsilon$ is a unital algebra morphism as well.
This follows from the fact that the action of $J$
commutes with the counit and from the fact that the elements $c_{i,j}$ are group-like.
The compatibility of $\epsilon$ with the unit is obvious.
In a final step, one verifies that the endomorphism
\[S(a\otimes j) = (c_{j^{-1}, j})^{-1}\varphi_{j^{-1}}(S_A(a))\otimes j^{-1}\]
is an antipode. Altogether, one arrives at
\begin{Proposition}
If $A$ is a $J$-Hopf-algebra, then the
orbifold algebra $\widehat{A}^J$ has a natural
structure of a Hopf algebra.
\end{Proposition}
\begin{Remark}
\begin{enumerate}
\item
The algebra $\widehat A^J$ is not the fixed point
subalgebra $A^J$ of $A$; in general, the categories
$A^J\text{-}\mathrm{mod}$ and $\widehat A^J\text{-}\mathrm{mod}$ are inequivalent.
\item
Given any Hopf algebra $A$ with weak $J$-action,
we have an exact sequence of Hopf algebras
\begin{align}\label{Hopf-algebra-sequ.}
A \to \widehat{A}^J \to {\mathbb K}[J] \,\,\, .
\end{align}
In particular, $A$ is a sub-Hopf algebra of $\widehat{A}^J$. In general, there is no inclusion of ${\mathbb K}[J]$ into $\widehat{A}^J$
as a Hopf algebra.
\item
If the action of $J$ on the algebra $A$ is strict, then
the algebra $A$ is a module algebra over the Hopf algebra
${\mathbb K}[J]$ (i.e.\ an algebra in the tensor category
${\mathbb K}[J]\text{-}\mathrm{mod}$). Then the orbifold algebra is the smash
product $A\# {\mathbb K}[J]$
(see \cite[Section 4]{montgomery1993hopf}
for the definitions). The situation described occurs,
if and only if the exact sequence (\ref{Hopf-algebra-sequ.})
splits.
\end{enumerate}
\end{Remark}
The next proposition justifies the name ``orbifold algebra''
for $\widehat{A}^J$:
\begin{Proposition}\label{orbifoldization}
Let $A$ be a $J$-Hopf algebra. Then there is an equivalence
of tensor categories
\[\widehat{A}^J\text{-}\mathrm{mod} \cong (A\text{-}\mathrm{mod})^J \,\, . \]
\end{Proposition}
\begin{proof}
\begin{itemize}
\item
An object of $(A\text{-}\mathrm{mod})^J$ consists of a ${\mathbb K}$-vector
space $M$, an $A$-action $\rho: A\to \text{End}(M)$
and a family of $A$-module morphisms $(\psi_j)_{j\in J}$.
We define on the same ${\mathbb K}$-vector space $M$ the
structure of an $\widehat{A}^J$ module
by $\tilde\rho: \widehat{A}^J \to\text{End}(M)$ with
$\tilde\rho(a\otimes j) := \rho(a)\circ(\psi_{j^{-1}})^{-1}$.
One next checks that, given two objects
$(M,\rho,\psi)$ and $(M',\rho',\psi')$
in $(A\text{-}\mathrm{mod})^J$, a ${\mathbb K}$-linear map
$f\in \Hom_{\mathbb K}(M,M')$ is in the subspace
$\Hom_{(A\text{-}\mathrm{mod})^J}(M,M')$ if and only if it is
in the subspace
$\Hom_{\widehat{A}^J\text{-}\mathrm{mod}}(M,\tilde\rho),(M',\tilde\rho'))$.
We can thus consider a ${\mathbb K}$-linear functor
\begin{equation}\label{orbifold-equivalence}
F: (A\text{-}\mathrm{mod})^J \to \widehat{A}^J\text{-}\mathrm{mod}
\end{equation}
which maps on objects by $(M,\rho,\psi)\mapsto (M,\tilde
\rho)$ and on morphisms as the identity. This functor
is clearly fully faithful.
To show that the functor is also essentially surjective,
we note that for any object $(M,\tilde\rho)$ in
$\widehat{A}^J\text{-}\mathrm{mod}$, an object in $(A\text{-}\mathrm{mod})^J$ can
be obtained as follows: on the underlying vector space,
we have the structure of an $A$-module by restriction,
$\rho(a):=\tilde\rho(a\otimes 1_J)$. A family of
equivariant morphisms is given by
$\psi_j :=(\tilde\rho(\mathbf{1}\otimes j^{-1}))^{-1}$. Clearly
its image under $F$ is isomorphic to $(M,\tilde\rho)$.
This shows that the functor $F$ is an equivalence of
categories, indeed even an isomorphism of categories.
\item The functor $F$ is also a strict tensor functor:
consider two objects $F(M,\rho,\psi)$ and $F(M',\rho',\psi')$
in $(A\text{-}\mathrm{mod})^J$. The functor
$F$ yields the
following action of the orbifold Hopf algebra
$\widehat{A}^J$ on the ${\mathbb K}$-vector space $M\otimes_{\mathbb K} M'$:
$$ \tilde\rho_{M\otimes M'}(a\otimes j)
= \rho\otimes \rho'(\Delta(a)) \circ ((\psi_{j^{-1}})^{-1}
\otimes (\psi'_{j^{-1}})^{-1})\,\, .$$
Since the coproduct on $\widehat{A}^J$ was just given
by the tensor product of coproducts on
$A$ and ${\mathbb K}[J]$, this coincides with the tensor product
of $F(M,\rho,\psi)$ and $F(M',\rho',\psi')$
in $\widehat{A}^J\text{-}\mathrm{mod}$.
\end{itemize}
\end{proof}
In a final step, we assume that the $J$-equivariant algebra $A$
has the additional structure of a $J$-ribbon algebra.
Then, by proposition \ref{category-J-ribbon},
the category $A\text{-}\mathrm{mod}$ is a $J$-ribbon category
and by proposition \ref{kirillov1} the orbifold
category $(A\text{-}\mathrm{mod})^J$ is a ribbon category.
The strict isomorphism (\ref{orbifold-equivalence})
of tensor categories allows us to transport both the
braiding and the ribbon structure to the representation category
of the orbifold Hopf algebra $\widehat{A}^J$.
General results \cite[Proposition 16.6.2]{kassel1995quantum}
assert that this amounts to a natural structure of
a ribbon algebra on $\widehat{A}^J$. In fact, we directly
read off the R-matrix and the ribbon element. For example,
the R-matrix $\hat R$ of $\hat A^J$ equals
$$\ \hat R= \hat\tau \,\,
c_{\widehat{A}^J,\widehat{A}^J}
(1_{\widehat{A}^J}\otimes 1_{\widehat{A}^J})
\in \widehat{A}^J\otimes \widehat{A}^J \,\,\, , $$
where the linear map $\hat\tau$
flips the two components of the tensor product
$\widehat{A}^J\otimes \widehat{A}^J$. This expression can be
explicitly evaluated, using the fact that
$A \otimes {\mathbb K}[J]$ is an object in $(A\text{-}\mathrm{mod})^J$ with
$A$-module structure given by left action
on the first component and that the morphisms $\psi_j$
are given by left multiplication on the second component.
We find for the R-matrix of $\widehat{A}^J$
\begin{align*}
\hat R & = \sum_{i,j \in J}(\mathrm{id} \otimes \psi_j)(\rho\otimes \rho)(R_{ij})
((1_A \otimes 1_J) \otimes 1_A \otimes 1_J)\\
& = \sum_{i,j \in J}((R_{i,j})_{1}\otimes 1_J)
\otimes (1_A \otimes j^{-1})^{-1}((R_{i,j})_{2} \otimes 1_J))
\end{align*}
where $R$ is the R-matrix of $A$.
The twist element of $\widehat{A}^J$ can be computed
similarly; one finds
\[\theta^{-1} = \sum_{j\in J}\psi_j\circ
\rho(\theta_j)(1_{A_j} \otimes 1_J)
=\sum_{j\in J}(1_A \otimes j^{-1})^{-1} (\theta_j \otimes 1_J)\]
We summarize our findings:
\begin{corollary}\label{ribbon-ribbon}
If $A$ is a $J$-ribbon algebra, then the orbifold algebra
$\widehat{A}^J$ inherits a natural structure of a ribbon
algebra such that the equivalence of tensor categories in
proposition \ref{orbifoldization} is an equivalence of
ribbon categories.
\end{corollary}
\subsection{Equivariant modular categories}
In this subsection, we show that the orbifold category
of the $J$-equivariant ribbon
category $\calc^J(G)\text{-}\mathrm{mod}$ is $J$-modular. A theorem of
Kirillov \cite[Theorem 10.5]{kirI17} then immediately implies
that the category $\calc^J(G)\text{-}\mathrm{mod}$ is $J$-modular.
Since we have already seen in corollary \ref{ribbon-ribbon}
that the orbifold category is equivalent, as a
ribbon category, to the representation category of the
orbifold Hopf algebra, it suffices the compute
this Hopf algebra explicitly. Our final result asserts
that this Hopf algebra is an ordinary
Drinfel'd double:
\begin{Proposition}\label{J-Drinfeld-orbifold}
The ${\mathbb K}$-linear map
\begin{equation}
\begin{array}{rll}
\Psi:\qquad
\widehat{{\mathcal D}^J(G)}^J &\to & {\mathcal D}(H)\\[.3em]
(\delta_h \otimes g \otimes j) & \mapsto &(\delta_h \otimes gs(j))
\end{array}
\end{equation}
is an isomorphism of ribbon algebras, where the Drinfel'd
double ${\mathcal D}(H)$ is taken with the standard ribbon
structure introduced in subsection 2.5.
\end{Proposition}
This result immediately implies the equivalence
\[(\widehat{{\mathcal D}^J(G)}\text{-}\mathrm{mod})^J \cong {\mathcal D}(H)\text{-}\mathrm{mod}\]
of ribbon categories and thus, by proposition \ref{modular-Drinfeld},
the modularity of the orbifold category, so that we have
finally proven:
\begin{Theorem}\label{J-modular}
The category $\calc^J(G) = \bigoplus_{j \in J} \mathcal{C}(G)_j$ has a natural structure of a
$J$-modular tensor category.
\end{Theorem}
\begin{proof}[Proof of proposition \ref{J-Drinfeld-orbifold}]
We show by direct computations that the linear map
$\Psi$ preserves product, coproduct, R-matrix and
twist element:
\begin{itemize}
\item
Compatibility with the product:
\begin{align*}
\Psi((\delta_h \otimes g \otimes j)&(\delta_h' \otimes g' \otimes j'))
= \Psi((\delta_h \otimes g)\cdot {}^j(\delta_h' \otimes g')c_{jj'} \otimes jj') \\[.2em]
& = \Psi\left((\delta_h \otimes g)\cdot (\delta_{s(j)h's(j)^{-1}} \otimes s(j)g's(j)^{-1})\cdot\sum_{h''\in H} (\delta_{h''} \otimes s(j)s(j')s(jj')^{-1})\otimes jj'\right) \\[.2em]
& = \Psi\left(\delta(h,gs(j)hs(j)^{-1}g^{-1})(\delta_h \otimes gs(j)g's(j')s(jj')^{-1})\otimes jj'\right) \\[.2em]
& = \delta(h,gs(j)hs(j)^{-1}g^{-1})(\delta_h \otimes gs(j)g's(j')) \\[.2em]
& = (\delta_{h}\otimes gs(j))\cdot (\delta_{h'}\otimes g's(j'))\\[.2em]
& = \Psi(\delta_h \otimes g \otimes j)\Psi(\delta_{h'} \otimes g' \otimes j')
\end{align*}
\item
Compatibility with the coproduct:
\begin{align*}
(\Psi \otimes \Psi)\Delta(\delta_h \otimes g \otimes j)
&= \sum_{h'h'' = h} \Psi(\delta_{h'} \otimes g \otimes j) \otimes \Psi(\delta_{h''} \otimes g \otimes j)\\[.2em]
&= \sum_{h'h'' = h} (\delta_{h'} \otimes gs(j)) \otimes (\delta_{h''} \otimes gs(j))\\[.2em]
&= \Delta\left(\Psi(\delta_h \otimes g \otimes j)\right)
\end{align*}
\item
The R-matrix of the orbifold algebra $\widehat{{\mathcal D}^J(G)}^J $ can
be determined using the lines preceding
corollary \ref{ribbon-ribbon} and the definition of the R-Matrix of ${\mathcal D}^J(G)$ given in \eqref{JR-matrix}:
$$R = \sum_{j,j'\in J}\sum_{h \in H_j, h' \in H_{j'}} (\delta_{h}
\otimes 1_G \otimes 1_J) \otimes (1 \otimes 1_G \otimes j^{-1})^{-1}(\delta_{h'} \otimes s(j^{-1})h
\otimes 1_J) $$
This implies
\begin{align*}
(\Psi \otimes \Psi)(R)
&= \sum_{j,j'\in J}\sum_{h \in H_j, h' \in H_{j'}} \Psi(\delta_{h} \otimes 1_G \otimes 1_J) \otimes \Psi(1 \otimes 1_G \otimes j^{-1})^{-1} \cdot \Psi(\delta_{h'} \otimes s(j^{-1})h \otimes 1_J)\\[.2em]
&=\sum_{j,j'\in J}\sum_{h \in H_j, h' \in H_{j'}} (\delta_{h} \otimes 1_H) \otimes (1 \otimes s(j^{-1})^{-1}) \cdot (\delta_{h'} \otimes s(j^{-1})h)\\[.2em]
&=\sum_{h \ h' \in H}(\delta_{h} \otimes 1) \otimes (\delta_{h'} \otimes h),\\[.2em]
\end{align*}
which is the standard R-matrix of the Drinfel'd double ${\mathcal D}(H)$.
\item The twist in $\widehat{{\mathcal D}^J(G)}^J $ is by corollary \ref{ribbon-ribbon} equal to
\begin{equation*}
\theta^{-1} = \sum_{j\in J}\sum_{h\in H_j} (\delta_h \otimes s(j^{-1})h\otimes 1_J)
\end{equation*}
and thus it gets mapped to the element
\begin{align*}
\Psi(\theta^{-1})
&= \sum_{j\in J}\sum_{h\in H_j} \Psi(1 \otimes 1_G \otimes j^{-1})^{-1} \Psi(\delta_h \otimes s(j^{-1})h\otimes 1_J)\\[.2em]
&= \sum_{j\in J}\sum_{h\in H_j} (1 \otimes s(j^{-1})^{-1}) \cdot (\delta_h \otimes s(j^{-1})h)\\[.2em]
&=\sum_{h\in H} (\delta_h \otimes h)
\end{align*}
which is the inverse of the twist element in ${\mathcal D}(H)$.
\end{itemize}
\end{proof}
\subsection{Summary of all tensor categories involved}
We summarize our findings by discussing again the
four tensor categories mentioned in the introduction,
in the square of equation (\ref{int:square}),
thereby presenting the
explicit solution of the algebraic problem described in section
\ref{intro:1.1}.
Given a finite group $G$
with a weak action of a finite group $J$, we get an
extension $1\to G \to H \to J\to 1$ of finite groups,
together with a set-theoretic section $s:J\to H$.
\begin{Proposition}\mbox{} \\
We have the following natural realizations of the categories
in question in terms of categories of finite-dimensional representations
over finite-dimensional ribbon algebras:
\begin{enumerate}
\item The premodular category introduced in \cite{bantay} is
${\mathcal B}(G\triangleleft H)\text{-}\mathrm{mod}$. As an abelian category, it is equivalent
to the representation category $G//H\text{-}\mathrm{mod}$ of the action groupoid $G//H$, i.e.\
to the category of $G$-graded ${\mathbb K}$-vector spaces with compatible
action of $H$.
\item The modular category obtained by modularization
is ${\mathcal D}(G)\text{-}\mathrm{mod}$. As an abelian category, it is equivalent to
$G//G\text{-}\mathrm{mod}$.
\item
The $J$-modular category constructed in this paper is
${\mathcal D}^J(G)\text{-}\mathrm{mod}$. As an abelian category, it is equivalent to
$H//G\text{-}\mathrm{mod}$.
\item The modular category obtained by orbifoldization
from the $J$-modular category ${\mathcal D}(G)\text{-}\mathrm{mod}$ is equivalent to
${\mathcal D}(H)\text{-}\mathrm{mod}$. As an abelian category, it is equivalent to
$H//H\text{-}\mathrm{mod}$.
\end{enumerate}
Equivalently, the diagram in equation (\ref{int:square}), has
the explicit realization:
\begin{equation}\label{square49}
\xymatrix{
{\mathcal D}(G)\text{-}\mathrm{mod} \ar@<0.7ex>[d]^{\text{orbifold}} \ar@(lu,ur)[]^J\ar@{^{(}->}[r]
& {\mathcal D}^J(G)\text{-}\mathrm{mod} \ar@(lu,ur)[]^J \ar@<0.7ex>[d]^{\text{orbifold}} \\
{\mathcal B}(G\triangleleft H)\text{-}\mathrm{mod} \ar[u]^{\text{modularization}}\ar@{^{(}->}[r]
& {\mathcal D}(H)\text{-}\mathrm{mod} \ar[u]
}
\end{equation}
\end{Proposition}
We could have chosen the inclusion in the lower line as an
alternative starting point for the solution of the algebraic
problem presented in introduction \ref{intro:1.1}. Recall
from the introduction that the category ${\mathcal B}(G\triangleleft H)\text{-}\mathrm{mod}$
contains a Tannakian subcategory that can be identified with the
category of representations of the quotient group $J=H/G$.
The Tannakian subcategory and thus the category ${\mathcal B}(G\triangleleft H)\text{-}\mathrm{mod}$
contain a commutative Frobenius algebra given by the algebra of functions
on $J$; recall that the modularization function was just induction along
this algebra. The image of this
algebra under the inclusion in the lower line yields
a commutative Frobenius algebra in the category ${\mathcal D}(H)\text{-}\mathrm{mod}$.
In a next step, one can consider induction along this algebra to
obtain another tensor category which, by general results
\cite[Theorem 4.2]{kirI17} is a $J$-modular category.
In this approach, it remains to show that this $J$-modular
tensor category is equivalent, as a $J$-modular tensor category,
to ${\mathcal D}^J(G)\text{-}\mathrm{mod}$ and, in a next step that the modularization
${\mathcal D}(G)\text{-}\mathrm{mod}$ can be naturally identified with the neutral sector
of the $J$-modular category. This line of thought has been discussed in
\cite[Lemma 2.2]{kirI15} including the square \eqref{square49} of Hopf algebras. Our results directly lead to a natural
Hopf algebra ${\mathcal D}^J(G)$ and additionally show how
the various categories arise from extended topological field theories
which are built on clear geometric principles and through which
all additional structure of the algebraic categories become explicitly
computable.
\section{Outlook}
Our results very explicitly provide an interesting class
$J$-modular tensor categories. All data of these theories,
including the representations of
the modular group $SL(2,\mathbb Z)$ on the vector spaces assigned to the
torus, are directly accessible in terms of representations of
finite groups. Also series of examples exist in which closed
formulae for all quantities can be derived, e.g.\ for
the inclusion of the alternating group in the symmetric group.
Our results admit generalizations in various directions.
In fact, in this paper, we have only studied a
subclass of Dijkgraaf-Witten theories. The general
case requires, apart from the choice of a finite group $G$,
the choice of an element of
$$\ H^3_{Gp}(G,U(1))= H^4(\mathcal A_G,\mathbb Z)\, . $$
This element can be interpreted \cite{Wi} geometrically
as a 2-gerbe on $\mathcal A_G$. It is known that in this
case a quasi-triangular Hopf algebra can be extracted that
is exactly the one discussed in \cite{dipr}. Indeed,
our results can also be generalized by including
the additional choice of a non-trivial element
$$\omega\in H^4_J(\mathcal A_G,\mathbb Z)\equiv H^4(\mathcal A_G//J,\mathbb Z)
\,\, . $$
Only all these data together
allow to investigate in a similar manner the
categories constructed by Bantay \cite{bantay}
for crossed modules with a boundary map that is not necessarily
injective any longer. We plan to explain this general case
in a subsequent publication.
\renewcommand{\theDef}{\Alph{section}.\arabic{subsection}.\arabic{Def}}
\renewenvironment{Definition}{ \refstepcounter{Def}\mbox{}\\
\noindent\sl\textbf{Definition
\Alph{section}.\arabic{subsection}.\arabic{Def}} }
{\vspace{0.3cm}}
|
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.